From: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Date: Mon, 24 Apr 2017 01:38:36 +0000 (-0700)
Subject: i965: perf: add support for Kabylake
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=9a50fc7cfcc0b084f3644ce79bb5c16505cfb404;p=mesa.git

i965: perf: add support for Kabylake

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
---

diff --git a/src/mesa/drivers/dri/i965/Makefile.am b/src/mesa/drivers/dri/i965/Makefile.am
index 3e09fef2d21..62c8fb162ef 100644
--- a/src/mesa/drivers/dri/i965/Makefile.am
+++ b/src/mesa/drivers/dri/i965/Makefile.am
@@ -110,6 +110,8 @@ EXTRA_DIST = \
 	brw_oa_sklgt3.xml \
 	brw_oa_sklgt4.xml \
 	brw_oa_bxt.xml \
+	brw_oa_kblgt2.xml \
+	brw_oa_kblgt3.xml \
 	brw_oa.py
 
 # Note: we avoid using a multi target rule here and outputting both the
diff --git a/src/mesa/drivers/dri/i965/Makefile.sources b/src/mesa/drivers/dri/i965/Makefile.sources
index 2905a392133..08ab16146bd 100644
--- a/src/mesa/drivers/dri/i965/Makefile.sources
+++ b/src/mesa/drivers/dri/i965/Makefile.sources
@@ -172,4 +172,8 @@ i965_oa_GENERATED_FILES = \
 	brw_oa_sklgt4.h \
 	brw_oa_sklgt4.c \
 	brw_oa_bxt.h \
-	brw_oa_bxt.c
+	brw_oa_bxt.c \
+	brw_oa_kblgt2.h \
+	brw_oa_kblgt2.c \
+	brw_oa_kblgt3.h \
+	brw_oa_kblgt3.c
diff --git a/src/mesa/drivers/dri/i965/brw_oa_kblgt2.xml b/src/mesa/drivers/dri/i965/brw_oa_kblgt2.xml
new file mode 100644
index 00000000000..ed68948584e
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/brw_oa_kblgt2.xml
@@ -0,0 +1,10455 @@
+<?xml version="1.0"?>
+<metrics version="1491576018" merge_md5="">
+  <set name="Render Metrics Basic Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_basic"
+       hw_config_guid="f8d677e9-ff6f-4df1-9310-0334c6efacce"
+       chipset="KBLGT2"
+       symbol_name="RenderBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Misses 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Cache Misses"
+             description="The total number of sampler cache misses in all LODs in all sampler units."
+             data_type="uint64"
+             equation="B 4 READ 8 UMUL"
+             underscore_name="sampler_l1_misses"
+             units="messages"
+             symbol_name="SamplerL1Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler 1 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 1 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler1_bottleneck"
+             units="percent"
+             symbol_name="Sampler1Bottleneck"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$SamplerL1Misses 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Sampler 0 Busy"
+             description="The percentage of time in which Sampler 0 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler0_busy"
+             units="percent"
+             symbol_name="Sampler0Busy"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler 1 Busy"
+             description="The percentage of time in which Sampler 1 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler1_busy"
+             units="percent"
+             symbol_name="Sampler1Busy"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Samplers Busy"
+             description="The percentage of time in which samplers have been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="$Sampler0Busy $Sampler1Busy FMAX"
+             underscore_name="samplers_busy"
+             units="percent"
+             symbol_name="SamplersBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI Fixed Pipe Throughput"
+             description="The total number of GPU memory bytes transferred between 3D Pipeline (Command Dispatch, Input Assembly and Stream Output) and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="B 6 READ B 7 READ UADD 64 UMUL"
+             underscore_name="gti_vf_throughput"
+             units="bytes"
+             symbol_name="GtiVfThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/3D Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler 0 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 0 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler0_bottleneck"
+             units="percent"
+             symbol_name="Sampler0Bottleneck"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="L3 Lookup Accesses w/o IC"
+             description="The total number of L3 cache lookup accesses w/o IC."
+             data_type="uint64"
+             equation="$SamplerL1Misses $ShaderMemoryAccesses UADD"
+             underscore_name="l3_lookups"
+             units="messages"
+             symbol_name="L3Lookups"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Depth Throughput"
+             description="The total number of GPU memory bytes transferred between depth caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 0 READ C 1 READ UADD 64 UMUL"
+             underscore_name="gti_depth_throughput"
+             units="bytes"
+             symbol_name="GtiDepthThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Depth Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Samplers Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which samplers have been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="$Sampler0Bottleneck $Sampler1Bottleneck FMAX"
+             max_equation="100"
+             underscore_name="sampler_bottleneck"
+             units="percent"
+             symbol_name="SamplerBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Indicate System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI HDC TLB Lookup Throughput"
+             description="The total number of GPU memory bytes transferred between GTI and HDC, when HDC is doing TLB lookups."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_hdc_lookups_throughput"
+             units="bytes"
+             symbol_name="GtiHdcLookupsThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI RCC Throughput"
+             description="The total number of GPU memory bytes transferred between render color caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 2 READ C 3 READ UADD 64 UMUL"
+             underscore_name="gti_rcc_throughput"
+             units="bytes"
+             symbol_name="GtiRccThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Color Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x166C01E0" />
+        <register type="NOA" address="0x00009888" value="0x12170280" />
+        <register type="NOA" address="0x00009888" value="0x12370280" />
+        <register type="NOA" address="0x00009888" value="0x11930317" />
+        <register type="NOA" address="0x00009888" value="0x159303DF" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x1A4E0080" />
+        <register type="NOA" address="0x00009888" value="0x0A6C0053" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x0A1B4000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0001" />
+        <register type="NOA" address="0x00009888" value="0x002F1000" />
+        <register type="NOA" address="0x00009888" value="0x042F1000" />
+        <register type="NOA" address="0x00009888" value="0x004C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C8400" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F6600" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x162C2200" />
+        <register type="NOA" address="0x00009888" value="0x062D8000" />
+        <register type="NOA" address="0x00009888" value="0x082D8000" />
+        <register type="NOA" address="0x00009888" value="0x00133000" />
+        <register type="NOA" address="0x00009888" value="0x08133000" />
+        <register type="NOA" address="0x00009888" value="0x00170020" />
+        <register type="NOA" address="0x00009888" value="0x08170021" />
+        <register type="NOA" address="0x00009888" value="0x10170000" />
+        <register type="NOA" address="0x00009888" value="0x0633C000" />
+        <register type="NOA" address="0x00009888" value="0x0833C000" />
+        <register type="NOA" address="0x00009888" value="0x06370800" />
+        <register type="NOA" address="0x00009888" value="0x08370840" />
+        <register type="NOA" address="0x00009888" value="0x10370000" />
+        <register type="NOA" address="0x00009888" value="0x0D933031" />
+        <register type="NOA" address="0x00009888" value="0x0F933E3F" />
+        <register type="NOA" address="0x00009888" value="0x01933D00" />
+        <register type="NOA" address="0x00009888" value="0x0393073C" />
+        <register type="NOA" address="0x00009888" value="0x0593000E" />
+        <register type="NOA" address="0x00009888" value="0x1D930000" />
+        <register type="NOA" address="0x00009888" value="0x19930000" />
+        <register type="NOA" address="0x00009888" value="0x1B930000" />
+        <register type="NOA" address="0x00009888" value="0x1D900157" />
+        <register type="NOA" address="0x00009888" value="0x1F900158" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x2B908000" />
+        <register type="NOA" address="0x00009888" value="0x2D908000" />
+        <register type="NOA" address="0x00009888" value="0x2F908000" />
+        <register type="NOA" address="0x00009888" value="0x31908000" />
+        <register type="NOA" address="0x00009888" value="0x15908000" />
+        <register type="NOA" address="0x00009888" value="0x17908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1190001F" />
+        <register type="NOA" address="0x00009888" value="0x51904400" />
+        <register type="NOA" address="0x00009888" value="0x41900020" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C21" />
+        <register type="NOA" address="0x00009888" value="0x47900061" />
+        <register type="NOA" address="0x00009888" value="0x57904440" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x59900004" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+        <register type="NOA" address="0x00009888" value="0x53904444" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Basic Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_basic"
+       hw_config_guid="e17fc42a-e614-41b6-90c4-1074841a6c77"
+       chipset="KBLGT2"
+       symbol_name="ComputeBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 6 READ B 7 READ C 0 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_read"
+             units="bytes"
+             symbol_name="UntypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Bytes Written"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 3 READ B 4 READ B 5 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_written"
+             units="bytes"
+             symbol_name="TypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Writes"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 1 READ C 2 READ C 3 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_written"
+             units="bytes"
+             symbol_name="UntypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="Typed Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 0 READ B 1 READ B 2 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_read"
+             units="bytes"
+             symbol_name="TypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x104F00E0" />
+        <register type="NOA" address="0x00009888" value="0x124F1C00" />
+        <register type="NOA" address="0x00009888" value="0x106C00E0" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4E0820" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x064F0900" />
+        <register type="NOA" address="0x00009888" value="0x084F0032" />
+        <register type="NOA" address="0x00009888" value="0x0A4F1891" />
+        <register type="NOA" address="0x00009888" value="0x0C4F0E00" />
+        <register type="NOA" address="0x00009888" value="0x0E4F003C" />
+        <register type="NOA" address="0x00009888" value="0x004F0D80" />
+        <register type="NOA" address="0x00009888" value="0x024F003B" />
+        <register type="NOA" address="0x00009888" value="0x006C0002" />
+        <register type="NOA" address="0x00009888" value="0x086C0100" />
+        <register type="NOA" address="0x00009888" value="0x0C6C000C" />
+        <register type="NOA" address="0x00009888" value="0x0E6C0B00" />
+        <register type="NOA" address="0x00009888" value="0x186C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x001B4000" />
+        <register type="NOA" address="0x00009888" value="0x081B8000" />
+        <register type="NOA" address="0x00009888" value="0x0C1B4000" />
+        <register type="NOA" address="0x00009888" value="0x0E1B8000" />
+        <register type="NOA" address="0x00009888" value="0x101C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C8000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0024" />
+        <register type="NOA" address="0x00009888" value="0x065B8000" />
+        <register type="NOA" address="0x00009888" value="0x085B4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5BC000" />
+        <register type="NOA" address="0x00009888" value="0x0C5B8000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B4000" />
+        <register type="NOA" address="0x00009888" value="0x005B8000" />
+        <register type="NOA" address="0x00009888" value="0x025B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C6000" />
+        <register type="NOA" address="0x00009888" value="0x1C5C001B" />
+        <register type="NOA" address="0x00009888" value="0x125C8000" />
+        <register type="NOA" address="0x00009888" value="0x145C8000" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C2000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0208" />
+        <register type="NOA" address="0x00009888" value="0x000DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x020D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5500" />
+        <register type="NOA" address="0x00009888" value="0x100F0155" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2CC000" />
+        <register type="NOA" address="0x00009888" value="0x162CFB00" />
+        <register type="NOA" address="0x00009888" value="0x182C00BE" />
+        <register type="NOA" address="0x00009888" value="0x022CC000" />
+        <register type="NOA" address="0x00009888" value="0x042CC000" />
+        <register type="NOA" address="0x00009888" value="0x19900157" />
+        <register type="NOA" address="0x00009888" value="0x1B900158" />
+        <register type="NOA" address="0x00009888" value="0x1D900105" />
+        <register type="NOA" address="0x00009888" value="0x1F900103" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x11900FFF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900821" />
+        <register type="NOA" address="0x00009888" value="0x47900802" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900802" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900002" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900422" />
+        <register type="NOA" address="0x00009888" value="0x53904444" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Render Metrics for 3D Pipeline Profile Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_pipe_profile"
+       hw_config_guid="d7a17a3a-ca71-40d2-a919-ace80d50633f"
+       chipset="KBLGT2"
+       symbol_name="RenderPipeProfile"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which vertex shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_bottleneck"
+             units="percent"
+             symbol_name="VsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Hi-Depth Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which early hierarchical depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hi_depth_bottleneck"
+             units="percent"
+             symbol_name="HiDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which geometry shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_bottleneck"
+             units="percent"
+             symbol_name="GsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="BC Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which barycentric coordinates calculation pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="bc_bottleneck"
+             units="percent"
+             symbol_name="BcBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Barycentric Calc"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Stall"
+             description="The percentage of time in which hull stall pipeline stage was stalled."
+             data_type="float"
+             equation="C 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_stall"
+             units="percent"
+             symbol_name="HsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Hull Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="VF Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which vertex fetch pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vf_bottleneck"
+             units="percent"
+             symbol_name="VfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Input Assembler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Strip-Fans Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which strip-fans pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="10"
+             equation="B 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_bottleneck"
+             units="percent"
+             symbol_name="SfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SF Stall"
+             description="The percentage of time in which strip-fans pipeline stage was stalled."
+             data_type="float"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_stall"
+             units="percent"
+             symbol_name="SfStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Bottleneck"
+             low_watermark="3"
+             description="The percentage of time in which hull shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="9"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_bottleneck"
+             units="percent"
+             symbol_name="HsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Hull Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CL Stall"
+             description="The percentage of time in which clipper pipeline stage was stalled."
+             data_type="float"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_stall"
+             units="percent"
+             symbol_name="ClStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Clipper"
+             />
+    <counter name="SO Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which stream output pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_bottleneck"
+             units="percent"
+             symbol_name="SoBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Stream Output"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="DS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which domain shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_bottleneck"
+             units="percent"
+             symbol_name="DsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Domain Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Clipper Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which clipper pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_bottleneck"
+             units="percent"
+             symbol_name="ClBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Clipper"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Stall"
+             description="The percentage of time in which domain shader pipeline stage was stalled."
+             data_type="float"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_stall"
+             units="percent"
+             symbol_name="DsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Domain Shader"
+             />
+    <counter name="Early Depth Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which early depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="early_depth_bottleneck"
+             units="percent"
+             symbol_name="EarlyDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SO Stall"
+             description="The percentage of time in which stream-output pipeline stage was stalled."
+             data_type="float"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_stall"
+             units="percent"
+             symbol_name="SoStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Stream Output"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x0C0E001F" />
+        <register type="NOA" address="0x00009888" value="0x0A0F0000" />
+        <register type="NOA" address="0x00009888" value="0x10116800" />
+        <register type="NOA" address="0x00009888" value="0x178A03E0" />
+        <register type="NOA" address="0x00009888" value="0x11824C00" />
+        <register type="NOA" address="0x00009888" value="0x11830020" />
+        <register type="NOA" address="0x00009888" value="0x13840020" />
+        <register type="NOA" address="0x00009888" value="0x11850019" />
+        <register type="NOA" address="0x00009888" value="0x11860007" />
+        <register type="NOA" address="0x00009888" value="0x01870C40" />
+        <register type="NOA" address="0x00009888" value="0x17880000" />
+        <register type="NOA" address="0x00009888" value="0x022F4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0040" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x040D4000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x020E5400" />
+        <register type="NOA" address="0x00009888" value="0x000E0000" />
+        <register type="NOA" address="0x00009888" value="0x080F0040" />
+        <register type="NOA" address="0x00009888" value="0x000F0000" />
+        <register type="NOA" address="0x00009888" value="0x100F0000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0040" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x06104000" />
+        <register type="NOA" address="0x00009888" value="0x06110012" />
+        <register type="NOA" address="0x00009888" value="0x06131000" />
+        <register type="NOA" address="0x00009888" value="0x01898000" />
+        <register type="NOA" address="0x00009888" value="0x0D890100" />
+        <register type="NOA" address="0x00009888" value="0x03898000" />
+        <register type="NOA" address="0x00009888" value="0x09808000" />
+        <register type="NOA" address="0x00009888" value="0x0B808000" />
+        <register type="NOA" address="0x00009888" value="0x0380C000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A0075" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0000" />
+        <register type="NOA" address="0x00009888" value="0x118A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A4000" />
+        <register type="NOA" address="0x00009888" value="0x138A8000" />
+        <register type="NOA" address="0x00009888" value="0x1D81A000" />
+        <register type="NOA" address="0x00009888" value="0x15818000" />
+        <register type="NOA" address="0x00009888" value="0x17818000" />
+        <register type="NOA" address="0x00009888" value="0x0B820030" />
+        <register type="NOA" address="0x00009888" value="0x07828000" />
+        <register type="NOA" address="0x00009888" value="0x0D824000" />
+        <register type="NOA" address="0x00009888" value="0x0F828000" />
+        <register type="NOA" address="0x00009888" value="0x05824000" />
+        <register type="NOA" address="0x00009888" value="0x0D830003" />
+        <register type="NOA" address="0x00009888" value="0x0583000C" />
+        <register type="NOA" address="0x00009888" value="0x09830000" />
+        <register type="NOA" address="0x00009888" value="0x03838000" />
+        <register type="NOA" address="0x00009888" value="0x07838000" />
+        <register type="NOA" address="0x00009888" value="0x0B840980" />
+        <register type="NOA" address="0x00009888" value="0x03844D80" />
+        <register type="NOA" address="0x00009888" value="0x11840000" />
+        <register type="NOA" address="0x00009888" value="0x09848000" />
+        <register type="NOA" address="0x00009888" value="0x09850080" />
+        <register type="NOA" address="0x00009888" value="0x03850003" />
+        <register type="NOA" address="0x00009888" value="0x01850000" />
+        <register type="NOA" address="0x00009888" value="0x07860000" />
+        <register type="NOA" address="0x00009888" value="0x0F860400" />
+        <register type="NOA" address="0x00009888" value="0x09870032" />
+        <register type="NOA" address="0x00009888" value="0x01888052" />
+        <register type="NOA" address="0x00009888" value="0x11880000" />
+        <register type="NOA" address="0x00009888" value="0x09884000" />
+        <register type="NOA" address="0x00009888" value="0x1B931001" />
+        <register type="NOA" address="0x00009888" value="0x1D930001" />
+        <register type="NOA" address="0x00009888" value="0x19934000" />
+        <register type="NOA" address="0x00009888" value="0x1B958000" />
+        <register type="NOA" address="0x00009888" value="0x1D950094" />
+        <register type="NOA" address="0x00009888" value="0x19958000" />
+        <register type="NOA" address="0x00009888" value="0x09E58000" />
+        <register type="NOA" address="0x00009888" value="0x0BE58000" />
+        <register type="NOA" address="0x00009888" value="0x03E5C000" />
+        <register type="NOA" address="0x00009888" value="0x0592C000" />
+        <register type="NOA" address="0x00009888" value="0x0B928000" />
+        <register type="NOA" address="0x00009888" value="0x0D924000" />
+        <register type="NOA" address="0x00009888" value="0x0F924000" />
+        <register type="NOA" address="0x00009888" value="0x11928000" />
+        <register type="NOA" address="0x00009888" value="0x1392C000" />
+        <register type="NOA" address="0x00009888" value="0x09924000" />
+        <register type="NOA" address="0x00009888" value="0x01985000" />
+        <register type="NOA" address="0x00009888" value="0x07988000" />
+        <register type="NOA" address="0x00009888" value="0x09981000" />
+        <register type="NOA" address="0x00009888" value="0x0B982000" />
+        <register type="NOA" address="0x00009888" value="0x0D982000" />
+        <register type="NOA" address="0x00009888" value="0x0F989000" />
+        <register type="NOA" address="0x00009888" value="0x05982000" />
+        <register type="NOA" address="0x00009888" value="0x13904000" />
+        <register type="NOA" address="0x00009888" value="0x21904000" />
+        <register type="NOA" address="0x00009888" value="0x23904000" />
+        <register type="NOA" address="0x00009888" value="0x25908000" />
+        <register type="NOA" address="0x00009888" value="0x27904000" />
+        <register type="NOA" address="0x00009888" value="0x29908000" />
+        <register type="NOA" address="0x00009888" value="0x2B904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1190C080" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900440" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900400" />
+        <register type="NOA" address="0x00009888" value="0x47900C21" />
+        <register type="NOA" address="0x00009888" value="0x57900400" />
+        <register type="NOA" address="0x00009888" value="0x49900042" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900024" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900841" />
+        <register type="NOA" address="0x00009888" value="0x53900400" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFEA" />
+        <register type="OA" address="0x00002774" value="0x00007FFC" />
+        <register type="OA" address="0x00002778" value="0x0007AFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000F5FD" />
+        <register type="OA" address="0x00002780" value="0x00079FFA" />
+        <register type="OA" address="0x00002784" value="0x0000F3FB" />
+        <register type="OA" address="0x00002788" value="0x0007BF7A" />
+        <register type="OA" address="0x0000278C" value="0x0000F7E7" />
+        <register type="OA" address="0x00002790" value="0x0007FEFA" />
+        <register type="OA" address="0x00002794" value="0x0000F7CF" />
+        <register type="OA" address="0x00002798" value="0x00077FFA" />
+        <register type="OA" address="0x0000279C" value="0x0000EFDF" />
+        <register type="OA" address="0x000027A0" value="0x0006FFFA" />
+        <register type="OA" address="0x000027A4" value="0x0000CFBF" />
+        <register type="OA" address="0x000027A8" value="0x0003FFFA" />
+        <register type="OA" address="0x000027AC" value="0x00005F7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Reads Distribution Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_reads"
+       hw_config_guid="57b59202-172b-477a-87de-33f85572c589"
+       chipset="KBLGT2"
+       symbol_name="MemoryReads"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank0Reads"
+             description="The total number of GTI memory reads from L3 Bank 0 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_reads"
+             units="messages"
+             symbol_name="GtiL3Bank0Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all accesses from GTI to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiL3Bank3Reads"
+             description="The total number of GTI memory reads from L3 Bank 3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_reads"
+             units="messages"
+             symbol_name="GtiL3Bank3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiRsMemoryReads"
+             description="The total number of GTI memory reads from Resource Streamer."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_rs_memory_reads"
+             units="messages"
+             symbol_name="GtiRsMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Resource Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiHizMemoryReads"
+             description="The total number of GTI memory reads from Hierarchical Depth Cache (Hi-Depth Cache misses)."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_reads"
+             units="messages"
+             symbol_name="GtiHizMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="GtiRccMemoryReads"
+             description="The total number of GTI memory reads from Render Color Cache (Render Color Cache misses)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_reads"
+             units="messages"
+             symbol_name="GtiRccMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank1Reads"
+             description="The total number of GTI memory reads from L3 Bank 1 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_reads"
+             units="messages"
+             symbol_name="GtiL3Bank1Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiCmdStreamerMemoryReads"
+             description="The total number of GTI memory reads from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_reads"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="GtiL3Bank2Reads"
+             description="The total number of GTI memory reads from L3 Bank 2 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_reads"
+             units="messages"
+             symbol_name="GtiL3Bank2Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiMemoryReads"
+             description="The total number of GTI memory reads."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_reads"
+             units="messages"
+             symbol_name="GtiMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GtiRczMemoryReads"
+             description="The total number of GTI memory reads from Render Depth Cache (Render Depth Cache misses)."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_reads"
+             units="messages"
+             symbol_name="GtiRczMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="GtiMscMemoryReads"
+             description="The total number of GTI memory reads from Multisampling Color Cache (Multisampling Color Cache misses)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_reads"
+             units="messages"
+             symbol_name="GtiMscMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiVfMemoryReads"
+             description="The total number of GTI memory reads from Vertex Fetch."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="gti_vf_memory_reads"
+             units="messages"
+             symbol_name="GtiVfMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Vertex Fetch"
+             />
+    <counter name="GtiStcMemoryReads"
+             description="The total number of GTI memory reads from Stencil Cache (Stencil Cache misses)."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_reads"
+             units="messages"
+             symbol_name="GtiStcMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Reads"
+             description="The total number of GTI memory reads from L3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Reads $GtiL3Bank1Reads $GtiL3Bank2Reads $GtiL3Bank3Reads UADD UADD UADD"
+             underscore_name="gti_l3_reads"
+             units="messages"
+             symbol_name="GtiL3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F900064" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930055" />
+        <register type="NOA" address="0x00009888" value="0x03E58000" />
+        <register type="NOA" address="0x00009888" value="0x05E5C000" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x13900150" />
+        <register type="NOA" address="0x00009888" value="0x21900151" />
+        <register type="NOA" address="0x00009888" value="0x23900152" />
+        <register type="NOA" address="0x00009888" value="0x25900153" />
+        <register type="NOA" address="0x00009888" value="0x27900154" />
+        <register type="NOA" address="0x00009888" value="0x29900155" />
+        <register type="NOA" address="0x00009888" value="0x2B900156" />
+        <register type="NOA" address="0x00009888" value="0x2D900157" />
+        <register type="NOA" address="0x00009888" value="0x2F90015F" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C60" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C00" />
+        <register type="NOA" address="0x00009888" value="0x47900C63" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900C63" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x0000272C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002728" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x0000271C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002718" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F872" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Writes Distribution Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_writes"
+       hw_config_guid="3addf8ef-8e9b-40f5-a448-3dbb5d5128b0"
+       chipset="KBLGT2"
+       symbol_name="MemoryWrites"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiMemoryWrites"
+             description="The total number of GTI memory writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_writes"
+             units="messages"
+             symbol_name="GtiMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all GTI accesses to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiMscMemoryWrites"
+             description="The total number of GTI memory writes from Multisampling Color Cache (Multisampling Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_writes"
+             units="messages"
+             symbol_name="GtiMscMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiCmdStreamerMemoryWrites"
+             description="The total number of GTI memory writes from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_writes"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiL3Bank0Writes"
+             description="The total number of GTI memory writes from L3 Bank 0 (L3 Bank 0 invalidations)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_writes"
+             units="messages"
+             symbol_name="GtiL3Bank0Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank1Writes"
+             description="The total number of GTI memory writes from L3 Bank 1 (L3 Bank 1 invalidations)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_writes"
+             units="messages"
+             symbol_name="GtiL3Bank1Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank2Writes"
+             description="The total number of GTI memory writes from L3 Bank 2 (L3 Bank 2 invalidations)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_writes"
+             units="messages"
+             symbol_name="GtiL3Bank2Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank3Writes"
+             description="The total number of GTI memory writes from L3 Bank 3 (L3 Bank 3 invalidations)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_writes"
+             units="messages"
+             symbol_name="GtiL3Bank3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Writes"
+             description="The total number of GTI memory writes from L3 (L3 invalidations)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Writes $GtiL3Bank1Writes $GtiL3Bank2Writes $GtiL3Bank3Writes UADD UADD UADD"
+             underscore_name="gti_l3_writes"
+             units="messages"
+             symbol_name="GtiL3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiRccMemoryWrites"
+             description="The total number of GTI memory writes from Render Color Cache (Render Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_writes"
+             units="messages"
+             symbol_name="GtiRccMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiSoMemoryWrites"
+             description="The total number of GTI memory writes from Stream Output."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_so_memory_writes"
+             units="messages"
+             symbol_name="GtiSoMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Stream Output"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiStcMemoryWrites"
+             description="The total number of GTI memory writes from Stencil Cache."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_writes"
+             units="messages"
+             symbol_name="GtiStcMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GtiRczMemoryWrites"
+             description="The total number of GTI memory writes from Render Depth Cache."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_writes"
+             units="messages"
+             symbol_name="GtiRczMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="GtiHizMemoryWrites"
+             description="The total number of GTI memory writes from Hierarchical Depth Cache."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_writes"
+             units="messages"
+             symbol_name="GtiHizMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F901000" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930055" />
+        <register type="NOA" address="0x00009888" value="0x03E58000" />
+        <register type="NOA" address="0x00009888" value="0x05E5C000" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x13900160" />
+        <register type="NOA" address="0x00009888" value="0x21900161" />
+        <register type="NOA" address="0x00009888" value="0x23900162" />
+        <register type="NOA" address="0x00009888" value="0x25900163" />
+        <register type="NOA" address="0x00009888" value="0x27900164" />
+        <register type="NOA" address="0x00009888" value="0x29900165" />
+        <register type="NOA" address="0x00009888" value="0x2B900166" />
+        <register type="NOA" address="0x00009888" value="0x2D900167" />
+        <register type="NOA" address="0x00009888" value="0x2F900150" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C60" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C00" />
+        <register type="NOA" address="0x00009888" value="0x47900C63" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900C63" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x0000272C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002728" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x0000271C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002718" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F822" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extended Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extended"
+       hw_config_guid="4af0400a-81c3-47db-a6b6-deddbd75680e"
+       chipset="KBLGT2"
+       symbol_name="ComputeExtended"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Writes 0"
+             description="The subslice 0 typed writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="typed_writes0"
+             units="messages"
+             symbol_name="TypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuTypedAtomics0"
+             description="The subslice 0 EU Typed Atomics subslice 0."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="eu_typed_atomics0"
+             units="messages"
+             symbol_name="EuTypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Atomics 0"
+             description="The subslice 0 typed atomics."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="typed_atomics0"
+             units="messages"
+             symbol_name="TypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedAtomicsPerCacheLine"
+             description="The ratio of EU typed atomics requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedAtomics0 $TypedAtomics0 FDIV"
+             underscore_name="typed_atomics_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedAtomicsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedReads0"
+             description="The subslice 0 EU Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="eu_untyped_reads0"
+             units="messages"
+             symbol_name="EuUntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Writes 0"
+             description="The subslice 0 untyped writes (including SLM writes)."
+             data_type="uint64"
+             equation="C 1 READ"
+             underscore_name="untyped_writes0"
+             units="messages"
+             symbol_name="UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedAtomics0"
+             description="The subslice 0 EU Untyped Atomics subslice 0."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="eu_untyped_atomics0"
+             units="messages"
+             symbol_name="EuUntypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedWrites0"
+             description="The subslice 0 EU Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="eu_untyped_writes0"
+             units="messages"
+             symbol_name="EuUntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedWrites0"
+             description="The subslice 0 EU A64 Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="eu_a64_untyped_writes0"
+             units="messages"
+             symbol_name="EuA64UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedWritesPerCacheLine"
+             description="The ratio of EU untyped write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuUntypedWrites0 $EuA64UntypedWrites0 UADD $UntypedWrites0 FDIV"
+             underscore_name="untyped_writes_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="UntypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedWrites0"
+             description="The subslice 0 EU Typed Writes subslice 0."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="eu_typed_writes0"
+             units="messages"
+             symbol_name="EuTypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedWritesPerCacheLine"
+             description="The ratio of EU typed write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedWrites0 $TypedWrites0 FDIV"
+             underscore_name="typed_writes_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Reads 0"
+             description="The subslice 0 typed reads."
+             data_type="uint64"
+             equation="C 2 READ"
+             underscore_name="typed_reads0"
+             units="messages"
+             symbol_name="TypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Reads 0"
+             description="The subslice 0 untyped reads (including SLM reads)."
+             data_type="uint64"
+             equation="C 3 READ"
+             underscore_name="untyped_reads0"
+             units="messages"
+             symbol_name="UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedReads0"
+             description="The subslice 0 EU A64 Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="eu_a64_untyped_reads0"
+             units="messages"
+             symbol_name="EuA64UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedReads0"
+             description="The subslice 0 EU Typed Reads subslice 0."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="eu_typed_reads0"
+             units="messages"
+             symbol_name="EuTypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedReadsPerCacheLine"
+             description="The ratio of EU untyped read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuUntypedReads0 $EuA64UntypedReads0 UADD $UntypedReads0 FDIV"
+             underscore_name="untyped_reads_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="UntypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedReadsPerCacheLine"
+             description="The ratio of EU typed read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuTypedReads0 $TypedReads0 FDIV"
+             underscore_name="typed_reads_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x106C00E0" />
+        <register type="NOA" address="0x00009888" value="0x141C8160" />
+        <register type="NOA" address="0x00009888" value="0x161C8015" />
+        <register type="NOA" address="0x00009888" value="0x181C0120" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4E8000" />
+        <register type="NOA" address="0x00009888" value="0x184E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4EAAA0" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x024E8000" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E6C0B01" />
+        <register type="NOA" address="0x00009888" value="0x006C0200" />
+        <register type="NOA" address="0x00009888" value="0x026C000C" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x001B8000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x001C0041" />
+        <register type="NOA" address="0x00009888" value="0x061C4200" />
+        <register type="NOA" address="0x00009888" value="0x081C4443" />
+        <register type="NOA" address="0x00009888" value="0x0A1C4645" />
+        <register type="NOA" address="0x00009888" value="0x0C1C7647" />
+        <register type="NOA" address="0x00009888" value="0x041C7357" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0030" />
+        <register type="NOA" address="0x00009888" value="0x101C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0000" />
+        <register type="NOA" address="0x00009888" value="0x121C8000" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4CAA2A" />
+        <register type="NOA" address="0x00009888" value="0x0C4C02AA" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x000DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5515" />
+        <register type="NOA" address="0x00009888" value="0x100F0155" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162CAA00" />
+        <register type="NOA" address="0x00009888" value="0x182C00AA" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x11907FFF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900040" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900802" />
+        <register type="NOA" address="0x00009888" value="0x47900842" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900842" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900800" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FC2A" />
+        <register type="OA" address="0x00002774" value="0x0000BF00" />
+        <register type="OA" address="0x00002778" value="0x0007FC6A" />
+        <register type="OA" address="0x0000277C" value="0x0000BF00" />
+        <register type="OA" address="0x00002780" value="0x0007FC92" />
+        <register type="OA" address="0x00002784" value="0x0000BF00" />
+        <register type="OA" address="0x00002788" value="0x0007FCA2" />
+        <register type="OA" address="0x0000278C" value="0x0000BF00" />
+        <register type="OA" address="0x00002790" value="0x0007FC32" />
+        <register type="OA" address="0x00002794" value="0x0000BF00" />
+        <register type="OA" address="0x00002798" value="0x0007FC9A" />
+        <register type="OA" address="0x0000279C" value="0x0000BF00" />
+        <register type="OA" address="0x000027A0" value="0x0007FE6A" />
+        <register type="OA" address="0x000027A4" value="0x0000BF00" />
+        <register type="OA" address="0x000027A8" value="0x0007FE7A" />
+        <register type="OA" address="0x000027AC" value="0x0000BF00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics L3 Cache Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_l3_cache"
+       hw_config_guid="0e22f995-79ca-4f67-83ab-e9d9772488d8"
+       chipset="KBLGT2"
+       symbol_name="ComputeL3Cache"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 03 Accesses"
+             description="The total number of accesses to L3 Bank 03."
+             data_type="uint64"
+             equation="B 3 READ 2 UMUL"
+             underscore_name="l3_bank03_accesses"
+             units="messages"
+             symbol_name="L3Bank03Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="L3 Accesses"
+             description="The total number of L3 accesses from all entities."
+             data_type="uint64"
+             equation="C 0 READ C 1 READ B 2 READ B 3 READ UADD UADD UADD 2 UMUL"
+             underscore_name="l3_accesses"
+             units="messages"
+             symbol_name="L3Accesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 29 READ 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="EU FPU0 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu0_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ C 5 READ UADD"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="L3 Bank 00 Accesses"
+             description="The total number of accesses to L3 Bank 00."
+             data_type="uint64"
+             equation="C 0 READ 2 UMUL"
+             underscore_name="l3_bank00_accesses"
+             units="messages"
+             symbol_name="L3Bank00Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="EU FPU0 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 19 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu0_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 14 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu1_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ C 5 READ UADD 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 00 IC Accesses"
+             description="The total number of accesses to L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 0 READ B 1 READ UADD 2 UMUL $L3Bank00Accesses UMIN"
+             underscore_name="l3_bank00_ic_accesses"
+             units="messages"
+             symbol_name="L3Bank00IcAccesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="L3 Bank 00 IC Hits"
+             description="The total number of hits in L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 1 READ 2 UMUL $L3Bank00IcAccesses UMIN"
+             underscore_name="l3_bank00_ic_hits"
+             units="messages"
+             symbol_name="L3Bank00IcHits"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="Sampler Accesses"
+             description="The total number of messages send to samplers."
+             data_type="uint64"
+             equation="A 28 READ"
+             underscore_name="sampler_accesses"
+             units="messages"
+             symbol_name="SamplerAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler"
+             />
+    <counter name="L3 Bank 01 Accesses"
+             description="The total number of accesses to L3 Bank 01."
+             data_type="uint64"
+             equation="C 1 READ 2 UMUL"
+             underscore_name="l3_bank01_accesses"
+             units="messages"
+             symbol_name="L3Bank01Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 20 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu1_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="L3 Bank 02 Accesses"
+             description="The total number of accesses to L3 Bank 02."
+             data_type="uint64"
+             equation="B 2 READ 2 UMUL"
+             underscore_name="l3_bank02_accesses"
+             units="messages"
+             symbol_name="L3Bank02Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="L3 Total Throughput"
+             description="The total number of GPU memory bytes transferred via L3."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Accesses 64 UMUL"
+             underscore_name="l3_total_throughput"
+             units="bytes"
+             symbol_name="L3TotalThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="EU FPU1 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x166C0760" />
+        <register type="NOA" address="0x00009888" value="0x1593001E" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4E8000" />
+        <register type="NOA" address="0x00009888" value="0x184E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4E8020" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x006C0051" />
+        <register type="NOA" address="0x00009888" value="0x066C5000" />
+        <register type="NOA" address="0x00009888" value="0x086C5C5D" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5E5F" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x186C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x001B4000" />
+        <register type="NOA" address="0x00009888" value="0x061B8000" />
+        <register type="NOA" address="0x00009888" value="0x081BC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x101C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1CE000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0030" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C2A00" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0280" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F1500" />
+        <register type="NOA" address="0x00009888" value="0x100F0140" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162C0A00" />
+        <register type="NOA" address="0x00009888" value="0x182C00A0" />
+        <register type="NOA" address="0x00009888" value="0x03933300" />
+        <register type="NOA" address="0x00009888" value="0x05930032" />
+        <register type="NOA" address="0x00009888" value="0x11930000" />
+        <register type="NOA" address="0x00009888" value="0x1B930000" />
+        <register type="NOA" address="0x00009888" value="0x1D900157" />
+        <register type="NOA" address="0x00009888" value="0x1F900158" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1190030F" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900000" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900021" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x53904444" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFFA" />
+        <register type="OA" address="0x00002774" value="0x0000FEFE" />
+        <register type="OA" address="0x00002778" value="0x0007FFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000FEFD" />
+        <register type="OA" address="0x00002790" value="0x0007FFFA" />
+        <register type="OA" address="0x00002794" value="0x0000FBEF" />
+        <register type="OA" address="0x00002798" value="0x0007FFFA" />
+        <register type="OA" address="0x0000279C" value="0x0000FBDF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00101100" />
+        <register type="FLEX" address="0x0000E45C" value="0x00201200" />
+        <register type="FLEX" address="0x0000E55C" value="0x00301300" />
+        <register type="FLEX" address="0x0000E65C" value="0x00401400" />
+    </register_config>
+  </set>
+
+  <set name="Metric set HDCAndSF"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="hdc_and_sf"
+       hw_config_guid="bc2a00f7-cb8a-4ff2-8ad0-e241dad16937"
+       chipset="KBLGT2"
+       symbol_name="HDCAndSF"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Polygon Data Ready"
+             description="The percentage of time in which geometry pipeline output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="poly_data_ready"
+             units="percent"
+             symbol_name="PolyDataReady"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="HDC stalled by L3 (s0.ss1)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss1)"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ C 4 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader01_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader01AccessStalledOnL3"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss2)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss2)"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ C 2 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader02_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader02AccessStalledOnL3"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss0)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss0)"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ C 6 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader00_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader00AccessStalledOnL3"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x104F0232" />
+        <register type="NOA" address="0x00009888" value="0x124F4640" />
+        <register type="NOA" address="0x00009888" value="0x106C0232" />
+        <register type="NOA" address="0x00009888" value="0x11834400" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4E8000" />
+        <register type="NOA" address="0x00009888" value="0x004F1880" />
+        <register type="NOA" address="0x00009888" value="0x024F08BB" />
+        <register type="NOA" address="0x00009888" value="0x044F001B" />
+        <register type="NOA" address="0x00009888" value="0x046C0100" />
+        <register type="NOA" address="0x00009888" value="0x066C000B" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x041B8000" />
+        <register type="NOA" address="0x00009888" value="0x061B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C1800" />
+        <register type="NOA" address="0x00009888" value="0x005B8000" />
+        <register type="NOA" address="0x00009888" value="0x025BC000" />
+        <register type="NOA" address="0x00009888" value="0x045B4000" />
+        <register type="NOA" address="0x00009888" value="0x125C8000" />
+        <register type="NOA" address="0x00009888" value="0x145C8000" />
+        <register type="NOA" address="0x00009888" value="0x165C8000" />
+        <register type="NOA" address="0x00009888" value="0x185C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C00A0" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x022CC000" />
+        <register type="NOA" address="0x00009888" value="0x042CC000" />
+        <register type="NOA" address="0x00009888" value="0x062CC000" />
+        <register type="NOA" address="0x00009888" value="0x082CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0F828000" />
+        <register type="NOA" address="0x00009888" value="0x0F8305C0" />
+        <register type="NOA" address="0x00009888" value="0x09830000" />
+        <register type="NOA" address="0x00009888" value="0x07830000" />
+        <register type="NOA" address="0x00009888" value="0x1D950080" />
+        <register type="NOA" address="0x00009888" value="0x13928000" />
+        <register type="NOA" address="0x00009888" value="0x0F988000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900040" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x10800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000FDFF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_1"
+       hw_config_guid="d2bbe790-f058-42d9-81c6-cdedcf655bc2"
+       chipset="KBLGT2"
+       symbol_name="L3_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank1 Active"
+             description="The percentage of time in which slice0 L3 bank1 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_active"
+             units="percent"
+             symbol_name="L30Bank1Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 L3 Bank1 Stalled"
+             description="The percentage of time in which slice0 L3 bank1 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_stalled"
+             units="percent"
+             symbol_name="L30Bank1Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Active"
+             description="The percentage of time in which slice0 L3 bank0 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_active"
+             units="percent"
+             symbol_name="L30Bank0Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Stalled"
+             description="The percentage of time in which slice0 L3 bank0 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_stalled"
+             units="percent"
+             symbol_name="L30Bank0Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C7B40" />
+        <register type="NOA" address="0x00009888" value="0x166C0020" />
+        <register type="NOA" address="0x00009888" value="0x0A603444" />
+        <register type="NOA" address="0x00009888" value="0x0A613400" />
+        <register type="NOA" address="0x00009888" value="0x1A4EA800" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x024E8000" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x064F4000" />
+        <register type="NOA" address="0x00009888" value="0x0C6C5327" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5425" />
+        <register type="NOA" address="0x00009888" value="0x006C2A00" />
+        <register type="NOA" address="0x00009888" value="0x026C285B" />
+        <register type="NOA" address="0x00009888" value="0x046C005C" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0800" />
+        <register type="NOA" address="0x00009888" value="0x0C1BC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x001B8000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C003C" />
+        <register type="NOA" address="0x00009888" value="0x121C8000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0800" />
+        <register type="NOA" address="0x00009888" value="0x065B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C1000" />
+        <register type="NOA" address="0x00009888" value="0x10600000" />
+        <register type="NOA" address="0x00009888" value="0x04600000" />
+        <register type="NOA" address="0x00009888" value="0x0C610044" />
+        <register type="NOA" address="0x00009888" value="0x10610000" />
+        <register type="NOA" address="0x00009888" value="0x06610000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C02A8" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C002A" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F0154" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x182C00AA" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2CC000" />
+        <register type="NOA" address="0x00009888" value="0x1190FFC0" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900420" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900021" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900400" />
+        <register type="NOA" address="0x00009888" value="0x43900421" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_2"
+       hw_config_guid="2f8e32e4-5956-46e2-af31-c8ea95887332"
+       chipset="KBLGT2"
+       symbol_name="L3_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank2 Stalled"
+             description="The percentage of time in which slice0 L3 bank2 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_stalled"
+             units="percent"
+             symbol_name="L30Bank2Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank2 Active"
+             description="The percentage of time in which slice0 L3 bank2 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_active"
+             units="percent"
+             symbol_name="L30Bank2Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C02E0" />
+        <register type="NOA" address="0x00009888" value="0x146C0001" />
+        <register type="NOA" address="0x00009888" value="0x0A623400" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x064F4000" />
+        <register type="NOA" address="0x00009888" value="0x026C3324" />
+        <register type="NOA" address="0x00009888" value="0x046C3422" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0800" />
+        <register type="NOA" address="0x00009888" value="0x065B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C1000" />
+        <register type="NOA" address="0x00009888" value="0x06614000" />
+        <register type="NOA" address="0x00009888" value="0x0C620044" />
+        <register type="NOA" address="0x00009888" value="0x10620000" />
+        <register type="NOA" address="0x00009888" value="0x06620000" />
+        <register type="NOA" address="0x00009888" value="0x084C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C002A" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F4000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2CC000" />
+        <register type="NOA" address="0x00009888" value="0x1190F800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00028002" />
+        <register type="OA" address="0x0000277C" value="0x000087FF" />
+        <register type="OA" address="0x00002780" value="0x00020002" />
+        <register type="OA" address="0x00002784" value="0x00008FFF" />
+        <register type="OA" address="0x00002788" value="0x00008002" />
+        <register type="OA" address="0x0000278C" value="0x0000A7FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_3"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_3"
+       hw_config_guid="ca046aad-b5fb-4101-adce-6473ee6e5b14"
+       chipset="KBLGT2"
+       symbol_name="L3_3"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 L3 Bank3 Stalled"
+             description="The percentage of time in which slice0 L3 bank3 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_stalled"
+             units="percent"
+             symbol_name="L30Bank3Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank3 Active"
+             description="The percentage of time in which slice0 L3 bank3 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_active"
+             units="percent"
+             symbol_name="L30Bank3Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C4E80" />
+        <register type="NOA" address="0x00009888" value="0x146C0000" />
+        <register type="NOA" address="0x00009888" value="0x0A633400" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4E8000" />
+        <register type="NOA" address="0x00009888" value="0x026C3321" />
+        <register type="NOA" address="0x00009888" value="0x046C342F" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C2000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x061B4000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C1800" />
+        <register type="NOA" address="0x00009888" value="0x06604000" />
+        <register type="NOA" address="0x00009888" value="0x0C630044" />
+        <register type="NOA" address="0x00009888" value="0x10630000" />
+        <register type="NOA" address="0x00009888" value="0x06630000" />
+        <register type="NOA" address="0x00009888" value="0x084C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C00AA" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F4000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x1190F800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900002" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00028002" />
+        <register type="OA" address="0x0000277C" value="0x000087FF" />
+        <register type="OA" address="0x00002780" value="0x00020002" />
+        <register type="OA" address="0x00002784" value="0x00008FFF" />
+        <register type="OA" address="0x00002788" value="0x00008002" />
+        <register type="OA" address="0x0000278C" value="0x0000A7FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set RasterizerAndPixelBackend"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="rasterizer_and_pixel_backend"
+       hw_config_guid="605f388f-24bb-455c-88e3-8d57ae0d7e9f"
+       chipset="KBLGT2"
+       symbol_name="RasterizerAndPixelBackend"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 Pixel Values Ready"
+             description="The percentage of time in which slice0 pixel values are ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_values0_ready"
+             units="percent"
+             symbol_name="PixelValues0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Slice0 Rasterizer Input Available"
+             description="The percentage of time in which slice0 rasterizer input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_input_available"
+             units="percent"
+             symbol_name="Rasterizer0InputAvailable"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice0 Post-EarlyZ Pixel Data Ready"
+             description="The percentage of time in which slice0 post-EarlyZ pixel data is ready (after early Z tests have been applied)"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_data0_ready"
+             units="percent"
+             symbol_name="PixelData0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Rasterizer Output Ready"
+             description="The percentage of time in which slice0 rasterizer output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_output_ready"
+             units="percent"
+             symbol_name="Rasterizer0OutputReady"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Slice0 PS Output Available"
+             description="The percentage of time in which slice0 PS output is available"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_output0_available"
+             units="percent"
+             symbol_name="PSOutput0Available"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x102F3800" />
+        <register type="NOA" address="0x00009888" value="0x144D0500" />
+        <register type="NOA" address="0x00009888" value="0x120D03C0" />
+        <register type="NOA" address="0x00009888" value="0x140D03CF" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0004" />
+        <register type="NOA" address="0x00009888" value="0x0C4E4000" />
+        <register type="NOA" address="0x00009888" value="0x042F0480" />
+        <register type="NOA" address="0x00009888" value="0x082F0000" />
+        <register type="NOA" address="0x00009888" value="0x022F0000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0090" />
+        <register type="NOA" address="0x00009888" value="0x064D0027" />
+        <register type="NOA" address="0x00009888" value="0x004D0000" />
+        <register type="NOA" address="0x00009888" value="0x000D0D40" />
+        <register type="NOA" address="0x00009888" value="0x020D803F" />
+        <register type="NOA" address="0x00009888" value="0x040D8023" />
+        <register type="NOA" address="0x00009888" value="0x100D0000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x020F0010" />
+        <register type="NOA" address="0x00009888" value="0x000F0000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0050" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41901400" />
+        <register type="NOA" address="0x00009888" value="0x43901485" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900001" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000EFFF" />
+        <register type="OA" address="0x00002778" value="0x00006000" />
+        <register type="OA" address="0x0000277C" value="0x0000F3FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set Sampler"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="sampler"
+       hw_config_guid="31dd157c-bf4e-4bab-bf2b-f5c8174af1af"
+       chipset="KBLGT2"
+       symbol_name="Sampler"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice2 Input Available"
+             description="The percentage of time in which slice0 subslice2 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_input_available"
+             units="percent"
+             symbol_name="Sampler02InputAvailable"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice0 Input Available"
+             description="The percentage of time in which slice0 subslice0 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_input_available"
+             units="percent"
+             symbol_name="Sampler00InputAvailable"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 Subslice2 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice2 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_output_ready"
+             units="percent"
+             symbol_name="Sampler02OutputReady"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Slice0 Subslice1 Input Available"
+             description="The percentage of time in which slice0 subslice1 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_input_available"
+             units="percent"
+             symbol_name="Sampler01InputAvailable"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Slice0 Subslice0 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice0 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_output_ready"
+             units="percent"
+             symbol_name="Sampler00OutputReady"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Slice0 Subslice1 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice1 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_output_ready"
+             units="percent"
+             symbol_name="Sampler01OutputReady"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x14152C00" />
+        <register type="NOA" address="0x00009888" value="0x16150005" />
+        <register type="NOA" address="0x00009888" value="0x121600A0" />
+        <register type="NOA" address="0x00009888" value="0x14352C00" />
+        <register type="NOA" address="0x00009888" value="0x16350005" />
+        <register type="NOA" address="0x00009888" value="0x123600A0" />
+        <register type="NOA" address="0x00009888" value="0x14552C00" />
+        <register type="NOA" address="0x00009888" value="0x16550005" />
+        <register type="NOA" address="0x00009888" value="0x125600A0" />
+        <register type="NOA" address="0x00009888" value="0x062F6000" />
+        <register type="NOA" address="0x00009888" value="0x022F2000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0050" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0010" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F0350" />
+        <register type="NOA" address="0x00009888" value="0x0C0FB000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F00DA" />
+        <register type="NOA" address="0x00009888" value="0x182C0028" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x022DC000" />
+        <register type="NOA" address="0x00009888" value="0x042D4000" />
+        <register type="NOA" address="0x00009888" value="0x0C138000" />
+        <register type="NOA" address="0x00009888" value="0x0E132000" />
+        <register type="NOA" address="0x00009888" value="0x0413C000" />
+        <register type="NOA" address="0x00009888" value="0x1C140018" />
+        <register type="NOA" address="0x00009888" value="0x0C157000" />
+        <register type="NOA" address="0x00009888" value="0x0E150078" />
+        <register type="NOA" address="0x00009888" value="0x10150000" />
+        <register type="NOA" address="0x00009888" value="0x04162180" />
+        <register type="NOA" address="0x00009888" value="0x02160000" />
+        <register type="NOA" address="0x00009888" value="0x04174000" />
+        <register type="NOA" address="0x00009888" value="0x0233A000" />
+        <register type="NOA" address="0x00009888" value="0x04333000" />
+        <register type="NOA" address="0x00009888" value="0x14348000" />
+        <register type="NOA" address="0x00009888" value="0x16348000" />
+        <register type="NOA" address="0x00009888" value="0x02357870" />
+        <register type="NOA" address="0x00009888" value="0x10350000" />
+        <register type="NOA" address="0x00009888" value="0x04360043" />
+        <register type="NOA" address="0x00009888" value="0x02360000" />
+        <register type="NOA" address="0x00009888" value="0x04371000" />
+        <register type="NOA" address="0x00009888" value="0x0E538000" />
+        <register type="NOA" address="0x00009888" value="0x00538000" />
+        <register type="NOA" address="0x00009888" value="0x06533000" />
+        <register type="NOA" address="0x00009888" value="0x1C540020" />
+        <register type="NOA" address="0x00009888" value="0x12548000" />
+        <register type="NOA" address="0x00009888" value="0x0E557000" />
+        <register type="NOA" address="0x00009888" value="0x00557800" />
+        <register type="NOA" address="0x00009888" value="0x10550000" />
+        <register type="NOA" address="0x00009888" value="0x06560043" />
+        <register type="NOA" address="0x00009888" value="0x02560000" />
+        <register type="NOA" address="0x00009888" value="0x06571000" />
+        <register type="NOA" address="0x00009888" value="0x1190FF80" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900060" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900060" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x70800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x0000C000" />
+        <register type="OA" address="0x00002774" value="0x0000E7FF" />
+        <register type="OA" address="0x00002778" value="0x00003000" />
+        <register type="OA" address="0x0000277C" value="0x0000F9FF" />
+        <register type="OA" address="0x00002780" value="0x00000C00" />
+        <register type="OA" address="0x00002784" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_1"
+       hw_config_guid="105db928-5542-466b-9128-e1f3c91426cb"
+       chipset="KBLGT2"
+       symbol_name="TDL_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x12120000" />
+        <register type="NOA" address="0x00009888" value="0x12320000" />
+        <register type="NOA" address="0x00009888" value="0x12520000" />
+        <register type="NOA" address="0x00009888" value="0x002F8000" />
+        <register type="NOA" address="0x00009888" value="0x022F3000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0015" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F03A0" />
+        <register type="NOA" address="0x00009888" value="0x0C0FF000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0095" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2D4000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x02108000" />
+        <register type="NOA" address="0x00009888" value="0x0410C000" />
+        <register type="NOA" address="0x00009888" value="0x02118000" />
+        <register type="NOA" address="0x00009888" value="0x0411C000" />
+        <register type="NOA" address="0x00009888" value="0x02121880" />
+        <register type="NOA" address="0x00009888" value="0x041219B5" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x02134000" />
+        <register type="NOA" address="0x00009888" value="0x04135000" />
+        <register type="NOA" address="0x00009888" value="0x0C308000" />
+        <register type="NOA" address="0x00009888" value="0x0E304000" />
+        <register type="NOA" address="0x00009888" value="0x06304000" />
+        <register type="NOA" address="0x00009888" value="0x0C318000" />
+        <register type="NOA" address="0x00009888" value="0x0E314000" />
+        <register type="NOA" address="0x00009888" value="0x06314000" />
+        <register type="NOA" address="0x00009888" value="0x0C321A80" />
+        <register type="NOA" address="0x00009888" value="0x0E320033" />
+        <register type="NOA" address="0x00009888" value="0x06320031" />
+        <register type="NOA" address="0x00009888" value="0x00320000" />
+        <register type="NOA" address="0x00009888" value="0x0C334000" />
+        <register type="NOA" address="0x00009888" value="0x0E331000" />
+        <register type="NOA" address="0x00009888" value="0x06331000" />
+        <register type="NOA" address="0x00009888" value="0x0E508000" />
+        <register type="NOA" address="0x00009888" value="0x00508000" />
+        <register type="NOA" address="0x00009888" value="0x02504000" />
+        <register type="NOA" address="0x00009888" value="0x0E518000" />
+        <register type="NOA" address="0x00009888" value="0x00518000" />
+        <register type="NOA" address="0x00009888" value="0x02514000" />
+        <register type="NOA" address="0x00009888" value="0x0E521880" />
+        <register type="NOA" address="0x00009888" value="0x00521A80" />
+        <register type="NOA" address="0x00009888" value="0x02520033" />
+        <register type="NOA" address="0x00009888" value="0x0E534000" />
+        <register type="NOA" address="0x00009888" value="0x00534000" />
+        <register type="NOA" address="0x00009888" value="0x02531000" />
+        <register type="NOA" address="0x00009888" value="0x1190FF80" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900062" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x00007FFF" />
+        <register type="OA" address="0x00002778" value="0x00000000" />
+        <register type="OA" address="0x0000277C" value="0x00009FFF" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000EFFF" />
+        <register type="OA" address="0x00002788" value="0x00000000" />
+        <register type="OA" address="0x0000278C" value="0x0000F3FF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FDFF" />
+        <register type="OA" address="0x00002798" value="0x00000000" />
+        <register type="OA" address="0x0000279C" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_2"
+       hw_config_guid="03db94d2-b37f-4c58-a791-0d2067b013bb"
+       chipset="KBLGT2"
+       symbol_name="TDL_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort0"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort1"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort1"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort0"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort0"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort1"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x12124D60" />
+        <register type="NOA" address="0x00009888" value="0x12322E60" />
+        <register type="NOA" address="0x00009888" value="0x12524D60" />
+        <register type="NOA" address="0x00009888" value="0x022F3000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0014" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0FE000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0097" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x002D8000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x0410C000" />
+        <register type="NOA" address="0x00009888" value="0x0411C000" />
+        <register type="NOA" address="0x00009888" value="0x04121FB7" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x04135000" />
+        <register type="NOA" address="0x00009888" value="0x00308000" />
+        <register type="NOA" address="0x00009888" value="0x06304000" />
+        <register type="NOA" address="0x00009888" value="0x00318000" />
+        <register type="NOA" address="0x00009888" value="0x06314000" />
+        <register type="NOA" address="0x00009888" value="0x00321B80" />
+        <register type="NOA" address="0x00009888" value="0x0632003F" />
+        <register type="NOA" address="0x00009888" value="0x00334000" />
+        <register type="NOA" address="0x00009888" value="0x06331000" />
+        <register type="NOA" address="0x00009888" value="0x0250C000" />
+        <register type="NOA" address="0x00009888" value="0x0251C000" />
+        <register type="NOA" address="0x00009888" value="0x02521FB7" />
+        <register type="NOA" address="0x00009888" value="0x00520000" />
+        <register type="NOA" address="0x00009888" value="0x02535000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x43900063" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extra Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extra"
+       hw_config_guid="aa7a3fb9-22fb-43ff-a32d-0ab6c13bbd16"
+       chipset="KBLGT2"
+       symbol_name="ComputeExtra"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active including Ext Math"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing including Extended Math processing"
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ  C 7 READ C 6 READ FADD C 5 READ FADD 8 FMUL FADD 100 FMUL $EuCoresTotalCount FDIV $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active_adjusted"
+             units="percent"
+             symbol_name="Fpu1ActiveAdjusted"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x121203E0" />
+        <register type="NOA" address="0x00009888" value="0x123203E0" />
+        <register type="NOA" address="0x00009888" value="0x125203E0" />
+        <register type="NOA" address="0x00009888" value="0x022F4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0040" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F006C" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x042D8000" />
+        <register type="NOA" address="0x00009888" value="0x06104000" />
+        <register type="NOA" address="0x00009888" value="0x06114000" />
+        <register type="NOA" address="0x00009888" value="0x06120033" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x06131000" />
+        <register type="NOA" address="0x00009888" value="0x04308000" />
+        <register type="NOA" address="0x00009888" value="0x04318000" />
+        <register type="NOA" address="0x00009888" value="0x04321980" />
+        <register type="NOA" address="0x00009888" value="0x00320000" />
+        <register type="NOA" address="0x00009888" value="0x04334000" />
+        <register type="NOA" address="0x00009888" value="0x04504000" />
+        <register type="NOA" address="0x00009888" value="0x04514000" />
+        <register type="NOA" address="0x00009888" value="0x04520033" />
+        <register type="NOA" address="0x00009888" value="0x00520000" />
+        <register type="NOA" address="0x00009888" value="0x04531000" />
+        <register type="NOA" address="0x00009888" value="0x1190E000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x43900C00" />
+        <register type="NOA" address="0x00009888" value="0x45900002" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00001000" />
+        <register type="FLEX" address="0x0000E558" value="0x00003002" />
+        <register type="FLEX" address="0x0000E658" value="0x00005004" />
+        <register type="FLEX" address="0x0000E758" value="0x00011010" />
+        <register type="FLEX" address="0x0000E45C" value="0x00050012" />
+        <register type="FLEX" address="0x0000E55C" value="0x00052051" />
+        <register type="FLEX" address="0x0000E65C" value="0x00000008" />
+    </register_config>
+  </set>
+
+  <set name="Media Vme Pipe Gen9"
+       mdapi_supported_apis="MEDIA IO BB"
+       underscore_name="vme_pipe"
+       hw_config_guid="398a4268-ef6f-4ffc-b55f-3c7b5363ce61"
+       chipset="KBLGT2"
+       symbol_name="VMEPipe"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL  GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="8 A 10 READ FMUL $EuThreadsCount FDIV $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VME Busy"
+             description="The percentage of time in which VME (IME or CRE) was actively processing data."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ B 3 READ FADD 2 FDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vme_busy"
+             units="percent"
+             symbol_name="VMEBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Overview System Batch Tier2"
+             mdapi_group="VME Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x141A5800" />
+        <register type="NOA" address="0x00009888" value="0x161A00C0" />
+        <register type="NOA" address="0x00009888" value="0x12180240" />
+        <register type="NOA" address="0x00009888" value="0x14180002" />
+        <register type="NOA" address="0x00009888" value="0x143A5800" />
+        <register type="NOA" address="0x00009888" value="0x163A00C0" />
+        <register type="NOA" address="0x00009888" value="0x12380240" />
+        <register type="NOA" address="0x00009888" value="0x14380002" />
+        <register type="NOA" address="0x00009888" value="0x002F1000" />
+        <register type="NOA" address="0x00009888" value="0x022F8000" />
+        <register type="NOA" address="0x00009888" value="0x042F3000" />
+        <register type="NOA" address="0x00009888" value="0x004C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C1500" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F9500" />
+        <register type="NOA" address="0x00009888" value="0x100F002A" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162C0A00" />
+        <register type="NOA" address="0x00009888" value="0x0A2DC000" />
+        <register type="NOA" address="0x00009888" value="0x0C2DC000" />
+        <register type="NOA" address="0x00009888" value="0x04193000" />
+        <register type="NOA" address="0x00009888" value="0x081A28C1" />
+        <register type="NOA" address="0x00009888" value="0x001A0000" />
+        <register type="NOA" address="0x00009888" value="0x00133000" />
+        <register type="NOA" address="0x00009888" value="0x0613C000" />
+        <register type="NOA" address="0x00009888" value="0x0813F000" />
+        <register type="NOA" address="0x00009888" value="0x00172000" />
+        <register type="NOA" address="0x00009888" value="0x06178000" />
+        <register type="NOA" address="0x00009888" value="0x0817A000" />
+        <register type="NOA" address="0x00009888" value="0x00180037" />
+        <register type="NOA" address="0x00009888" value="0x06180940" />
+        <register type="NOA" address="0x00009888" value="0x08180000" />
+        <register type="NOA" address="0x00009888" value="0x02180000" />
+        <register type="NOA" address="0x00009888" value="0x04183000" />
+        <register type="NOA" address="0x00009888" value="0x06393000" />
+        <register type="NOA" address="0x00009888" value="0x0C3A28C1" />
+        <register type="NOA" address="0x00009888" value="0x003A0000" />
+        <register type="NOA" address="0x00009888" value="0x0A33F000" />
+        <register type="NOA" address="0x00009888" value="0x0C33F000" />
+        <register type="NOA" address="0x00009888" value="0x0A37A000" />
+        <register type="NOA" address="0x00009888" value="0x0C37A000" />
+        <register type="NOA" address="0x00009888" value="0x0A380977" />
+        <register type="NOA" address="0x00009888" value="0x08380000" />
+        <register type="NOA" address="0x00009888" value="0x04380000" />
+        <register type="NOA" address="0x00009888" value="0x06383000" />
+        <register type="NOA" address="0x00009888" value="0x119000FF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900040" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900800" />
+        <register type="NOA" address="0x00009888" value="0x47901000" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900844" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00100030" />
+        <register type="OA" address="0x00002774" value="0x0000FFF9" />
+        <register type="OA" address="0x00002778" value="0x00000002" />
+        <register type="OA" address="0x0000277C" value="0x0000FFFC" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000FFF3" />
+        <register type="OA" address="0x00002788" value="0x00100180" />
+        <register type="OA" address="0x0000278C" value="0x0000FFCF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00000002" />
+        <register type="OA" address="0x0000279C" value="0x0000FF3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00008003" />
+    </register_config>
+  </set>
+
+  <set name="MDAPI testing set Gen9.5"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="test_oa"
+       hw_config_guid="baa3c7e4-52b6-4b85-801e-465a94b746dd"
+       chipset="KBLGT2"
+       symbol_name="TestOa"
+       >
+    <counter name="TestCounter7"
+             description="HW test counter 7. Factor: 0.666"
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="counter7"
+             units="events"
+             symbol_name="Counter7"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TestCounter8"
+             description="HW test counter 8. Should be equal to 1."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="counter8"
+             units="events"
+             symbol_name="Counter8"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter4"
+             description="HW test counter 4. Factor: 0.333"
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="counter4"
+             units="events"
+             symbol_name="Counter4"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter5"
+             description="HW test counter 5. Factor: 0.333"
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="counter5"
+             units="events"
+             symbol_name="Counter5"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter6"
+             description="HW test counter 6. Factor: 0.166"
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="counter6"
+             units="events"
+             symbol_name="Counter6"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter3"
+             description="HW test counter 3. Factor: 0.5"
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="counter3"
+             units="events"
+             symbol_name="Counter3"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter0"
+             description="HW test counter 0. Factor: 0.0"
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="counter0"
+             units="events"
+             symbol_name="Counter0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter1"
+             description="HW test counter 1. Factor: 1.0"
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="counter1"
+             units="events"
+             symbol_name="Counter1"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter2"
+             description="HW test counter 2. Factor: 1.0"
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="counter2"
+             units="events"
+             symbol_name="Counter2"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810000" />
+        <register type="NOA" address="0x00009888" value="0x07810013" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930040" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x00000004" />
+        <register type="OA" address="0x00002774" value="0x00000000" />
+        <register type="OA" address="0x00002778" value="0x00000003" />
+        <register type="OA" address="0x0000277C" value="0x00000000" />
+        <register type="OA" address="0x00002780" value="0x00000007" />
+        <register type="OA" address="0x00002784" value="0x00000000" />
+        <register type="OA" address="0x00002788" value="0x00100002" />
+        <register type="OA" address="0x0000278C" value="0x0000FFF7" />
+        <register type="OA" address="0x00002790" value="0x00100002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00100082" />
+        <register type="OA" address="0x0000279C" value="0x0000FFEF" />
+        <register type="OA" address="0x000027A0" value="0x001000C2" />
+        <register type="OA" address="0x000027A4" value="0x0000FFE7" />
+        <register type="OA" address="0x000027A8" value="0x00100001" />
+        <register type="OA" address="0x000027AC" value="0x0000FFE7" />
+    </register_config>
+  </set>
+
+</metrics>
diff --git a/src/mesa/drivers/dri/i965/brw_oa_kblgt3.xml b/src/mesa/drivers/dri/i965/brw_oa_kblgt3.xml
new file mode 100644
index 00000000000..b4ee2315703
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/brw_oa_kblgt3.xml
@@ -0,0 +1,10500 @@
+<?xml version="1.0"?>
+<metrics version="1491576028" merge_md5="">
+  <set name="Render Metrics Basic Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_basic"
+       hw_config_guid="0286c920-2f6d-493b-b22d-7a5280df43de"
+       chipset="KBLGT3"
+       symbol_name="RenderBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Misses 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Cache Misses"
+             description="The total number of sampler cache misses in all LODs in all sampler units."
+             data_type="uint64"
+             equation="B 4 READ B 5 READ UADD 8 UMUL"
+             underscore_name="sampler_l1_misses"
+             units="messages"
+             symbol_name="SamplerL1Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler 1 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 1 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler1_bottleneck"
+             units="percent"
+             symbol_name="Sampler1Bottleneck"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$SamplerL1Misses 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Sampler 0 Busy"
+             description="The percentage of time in which Sampler 0 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler0_busy"
+             units="percent"
+             symbol_name="Sampler0Busy"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler 1 Busy"
+             description="The percentage of time in which Sampler 1 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler1_busy"
+             units="percent"
+             symbol_name="Sampler1Busy"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Samplers Busy"
+             description="The percentage of time in which samplers have been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="$Sampler0Busy $Sampler1Busy FMAX"
+             underscore_name="samplers_busy"
+             units="percent"
+             symbol_name="SamplersBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI Fixed Pipe Throughput"
+             description="The total number of GPU memory bytes transferred between 3D Pipeline (Command Dispatch, Input Assembly and Stream Output) and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="B 6 READ B 7 READ UADD 64 UMUL"
+             underscore_name="gti_vf_throughput"
+             units="bytes"
+             symbol_name="GtiVfThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/3D Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler 0 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 0 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler0_bottleneck"
+             units="percent"
+             symbol_name="Sampler0Bottleneck"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="L3 Lookup Accesses w/o IC"
+             description="The total number of L3 cache lookup accesses w/o IC."
+             data_type="uint64"
+             equation="$SamplerL1Misses $ShaderMemoryAccesses UADD"
+             underscore_name="l3_lookups"
+             units="messages"
+             symbol_name="L3Lookups"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Depth Throughput"
+             description="The total number of GPU memory bytes transferred between depth caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 0 READ C 1 READ UADD 64 UMUL"
+             underscore_name="gti_depth_throughput"
+             units="bytes"
+             symbol_name="GtiDepthThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Depth Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Samplers Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which samplers have been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="$Sampler0Bottleneck $Sampler1Bottleneck FMAX"
+             max_equation="100"
+             underscore_name="sampler_bottleneck"
+             units="percent"
+             symbol_name="SamplerBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Indicate System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI HDC TLB Lookup Throughput"
+             description="The total number of GPU memory bytes transferred between GTI and HDC, when HDC is doing TLB lookups."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_hdc_lookups_throughput"
+             units="bytes"
+             symbol_name="GtiHdcLookupsThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI RCC Throughput"
+             description="The total number of GPU memory bytes transferred between render color caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 2 READ C 3 READ UADD 64 UMUL"
+             underscore_name="gti_rcc_throughput"
+             units="bytes"
+             symbol_name="GtiRccThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Color Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x166C01E0" />
+        <register type="NOA" address="0x00009888" value="0x12170280" />
+        <register type="NOA" address="0x00009888" value="0x12370280" />
+        <register type="NOA" address="0x00009888" value="0x16EC01E0" />
+        <register type="NOA" address="0x00009888" value="0x11930317" />
+        <register type="NOA" address="0x00009888" value="0x159303DF" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x1A4E0380" />
+        <register type="NOA" address="0x00009888" value="0x0A6C0053" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x0A1B4000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0001" />
+        <register type="NOA" address="0x00009888" value="0x002F1000" />
+        <register type="NOA" address="0x00009888" value="0x042F1000" />
+        <register type="NOA" address="0x00009888" value="0x004C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C8400" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0002" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F6600" />
+        <register type="NOA" address="0x00009888" value="0x100F0001" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x162CA200" />
+        <register type="NOA" address="0x00009888" value="0x062D8000" />
+        <register type="NOA" address="0x00009888" value="0x082D8000" />
+        <register type="NOA" address="0x00009888" value="0x00133000" />
+        <register type="NOA" address="0x00009888" value="0x08133000" />
+        <register type="NOA" address="0x00009888" value="0x00170020" />
+        <register type="NOA" address="0x00009888" value="0x08170021" />
+        <register type="NOA" address="0x00009888" value="0x10170000" />
+        <register type="NOA" address="0x00009888" value="0x0633C000" />
+        <register type="NOA" address="0x00009888" value="0x0833C000" />
+        <register type="NOA" address="0x00009888" value="0x06370800" />
+        <register type="NOA" address="0x00009888" value="0x08370840" />
+        <register type="NOA" address="0x00009888" value="0x10370000" />
+        <register type="NOA" address="0x00009888" value="0x1ACE0200" />
+        <register type="NOA" address="0x00009888" value="0x0AEC5300" />
+        <register type="NOA" address="0x00009888" value="0x10EC0000" />
+        <register type="NOA" address="0x00009888" value="0x1CEC0000" />
+        <register type="NOA" address="0x00009888" value="0x0A9B8000" />
+        <register type="NOA" address="0x00009888" value="0x1C9C0002" />
+        <register type="NOA" address="0x00009888" value="0x0CCC0002" />
+        <register type="NOA" address="0x00009888" value="0x0A8D8000" />
+        <register type="NOA" address="0x00009888" value="0x108F0001" />
+        <register type="NOA" address="0x00009888" value="0x16AC8000" />
+        <register type="NOA" address="0x00009888" value="0x0D933031" />
+        <register type="NOA" address="0x00009888" value="0x0F933E3F" />
+        <register type="NOA" address="0x00009888" value="0x01933D00" />
+        <register type="NOA" address="0x00009888" value="0x0393073C" />
+        <register type="NOA" address="0x00009888" value="0x0593000E" />
+        <register type="NOA" address="0x00009888" value="0x1D930000" />
+        <register type="NOA" address="0x00009888" value="0x19930000" />
+        <register type="NOA" address="0x00009888" value="0x1B930000" />
+        <register type="NOA" address="0x00009888" value="0x1D900157" />
+        <register type="NOA" address="0x00009888" value="0x1F900158" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x2B908000" />
+        <register type="NOA" address="0x00009888" value="0x2D908000" />
+        <register type="NOA" address="0x00009888" value="0x2F908000" />
+        <register type="NOA" address="0x00009888" value="0x31908000" />
+        <register type="NOA" address="0x00009888" value="0x15908000" />
+        <register type="NOA" address="0x00009888" value="0x17908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1190003F" />
+        <register type="NOA" address="0x00009888" value="0x51902240" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x55900242" />
+        <register type="NOA" address="0x00009888" value="0x45900084" />
+        <register type="NOA" address="0x00009888" value="0x47901400" />
+        <register type="NOA" address="0x00009888" value="0x57902220" />
+        <register type="NOA" address="0x00009888" value="0x49900C60" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900002" />
+        <register type="NOA" address="0x00009888" value="0x43900C63" />
+        <register type="NOA" address="0x00009888" value="0x53902222" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Basic Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_basic"
+       hw_config_guid="9823aaa1-b06f-40ce-884b-cd798c79f0c2"
+       chipset="KBLGT3"
+       symbol_name="ComputeBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 6 READ B 7 READ C 0 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_read"
+             units="bytes"
+             symbol_name="UntypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Bytes Written"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 3 READ B 4 READ B 5 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_written"
+             units="bytes"
+             symbol_name="TypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Writes"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 1 READ C 2 READ C 3 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_written"
+             units="bytes"
+             symbol_name="UntypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="Typed Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 0 READ B 1 READ B 2 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_read"
+             units="bytes"
+             symbol_name="TypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x104F00E0" />
+        <register type="NOA" address="0x00009888" value="0x124F1C00" />
+        <register type="NOA" address="0x00009888" value="0x106C00E0" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4E0820" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x064F0900" />
+        <register type="NOA" address="0x00009888" value="0x084F0032" />
+        <register type="NOA" address="0x00009888" value="0x0A4F1891" />
+        <register type="NOA" address="0x00009888" value="0x0C4F0E00" />
+        <register type="NOA" address="0x00009888" value="0x0E4F003C" />
+        <register type="NOA" address="0x00009888" value="0x004F0D80" />
+        <register type="NOA" address="0x00009888" value="0x024F003B" />
+        <register type="NOA" address="0x00009888" value="0x006C0002" />
+        <register type="NOA" address="0x00009888" value="0x086C0100" />
+        <register type="NOA" address="0x00009888" value="0x0C6C000C" />
+        <register type="NOA" address="0x00009888" value="0x0E6C0B00" />
+        <register type="NOA" address="0x00009888" value="0x186C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x001B4000" />
+        <register type="NOA" address="0x00009888" value="0x081B8000" />
+        <register type="NOA" address="0x00009888" value="0x0C1B4000" />
+        <register type="NOA" address="0x00009888" value="0x0E1B8000" />
+        <register type="NOA" address="0x00009888" value="0x101C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C8000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0024" />
+        <register type="NOA" address="0x00009888" value="0x065B8000" />
+        <register type="NOA" address="0x00009888" value="0x085B4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5BC000" />
+        <register type="NOA" address="0x00009888" value="0x0C5B8000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B4000" />
+        <register type="NOA" address="0x00009888" value="0x005B8000" />
+        <register type="NOA" address="0x00009888" value="0x025B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C6000" />
+        <register type="NOA" address="0x00009888" value="0x1C5C001B" />
+        <register type="NOA" address="0x00009888" value="0x125C8000" />
+        <register type="NOA" address="0x00009888" value="0x145C8000" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C2000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0208" />
+        <register type="NOA" address="0x00009888" value="0x000DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x020D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5500" />
+        <register type="NOA" address="0x00009888" value="0x100F0155" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2CC000" />
+        <register type="NOA" address="0x00009888" value="0x162CFB00" />
+        <register type="NOA" address="0x00009888" value="0x182C00BE" />
+        <register type="NOA" address="0x00009888" value="0x022CC000" />
+        <register type="NOA" address="0x00009888" value="0x042CC000" />
+        <register type="NOA" address="0x00009888" value="0x19900157" />
+        <register type="NOA" address="0x00009888" value="0x1B900158" />
+        <register type="NOA" address="0x00009888" value="0x1D900105" />
+        <register type="NOA" address="0x00009888" value="0x1F900103" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x11900FFF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900821" />
+        <register type="NOA" address="0x00009888" value="0x47900802" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900802" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900002" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900422" />
+        <register type="NOA" address="0x00009888" value="0x53904444" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Render Metrics for 3D Pipeline Profile Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_pipe_profile"
+       hw_config_guid="c7c735f3-ce58-45cf-aa04-30b183f1faff"
+       chipset="KBLGT3"
+       symbol_name="RenderPipeProfile"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which vertex shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_bottleneck"
+             units="percent"
+             symbol_name="VsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Hi-Depth Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which early hierarchical depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hi_depth_bottleneck"
+             units="percent"
+             symbol_name="HiDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which geometry shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_bottleneck"
+             units="percent"
+             symbol_name="GsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="BC Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which barycentric coordinates calculation pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="bc_bottleneck"
+             units="percent"
+             symbol_name="BcBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Barycentric Calc"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Stall"
+             description="The percentage of time in which hull stall pipeline stage was stalled."
+             data_type="float"
+             equation="C 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_stall"
+             units="percent"
+             symbol_name="HsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Hull Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="VF Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which vertex fetch pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vf_bottleneck"
+             units="percent"
+             symbol_name="VfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Input Assembler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Strip-Fans Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which strip-fans pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="10"
+             equation="B 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_bottleneck"
+             units="percent"
+             symbol_name="SfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SF Stall"
+             description="The percentage of time in which strip-fans pipeline stage was stalled."
+             data_type="float"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_stall"
+             units="percent"
+             symbol_name="SfStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Bottleneck"
+             low_watermark="3"
+             description="The percentage of time in which hull shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="9"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_bottleneck"
+             units="percent"
+             symbol_name="HsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Hull Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CL Stall"
+             description="The percentage of time in which clipper pipeline stage was stalled."
+             data_type="float"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_stall"
+             units="percent"
+             symbol_name="ClStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Clipper"
+             />
+    <counter name="SO Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which stream output pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_bottleneck"
+             units="percent"
+             symbol_name="SoBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Stream Output"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="DS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which domain shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_bottleneck"
+             units="percent"
+             symbol_name="DsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Domain Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Clipper Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which clipper pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_bottleneck"
+             units="percent"
+             symbol_name="ClBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Clipper"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Stall"
+             description="The percentage of time in which domain shader pipeline stage was stalled."
+             data_type="float"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_stall"
+             units="percent"
+             symbol_name="DsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Domain Shader"
+             />
+    <counter name="Early Depth Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which early depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="early_depth_bottleneck"
+             units="percent"
+             symbol_name="EarlyDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SO Stall"
+             description="The percentage of time in which stream-output pipeline stage was stalled."
+             data_type="float"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_stall"
+             units="percent"
+             symbol_name="SoStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Stream Output"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x0C0E001F" />
+        <register type="NOA" address="0x00009888" value="0x0A0F0000" />
+        <register type="NOA" address="0x00009888" value="0x10116800" />
+        <register type="NOA" address="0x00009888" value="0x178A03E0" />
+        <register type="NOA" address="0x00009888" value="0x11824C00" />
+        <register type="NOA" address="0x00009888" value="0x11830020" />
+        <register type="NOA" address="0x00009888" value="0x13840020" />
+        <register type="NOA" address="0x00009888" value="0x11850019" />
+        <register type="NOA" address="0x00009888" value="0x11860007" />
+        <register type="NOA" address="0x00009888" value="0x01870C40" />
+        <register type="NOA" address="0x00009888" value="0x17880000" />
+        <register type="NOA" address="0x00009888" value="0x022F4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0040" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x040D4000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x020E5400" />
+        <register type="NOA" address="0x00009888" value="0x000E0000" />
+        <register type="NOA" address="0x00009888" value="0x080F0040" />
+        <register type="NOA" address="0x00009888" value="0x000F0000" />
+        <register type="NOA" address="0x00009888" value="0x100F0000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0040" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x06104000" />
+        <register type="NOA" address="0x00009888" value="0x06110012" />
+        <register type="NOA" address="0x00009888" value="0x06131000" />
+        <register type="NOA" address="0x00009888" value="0x01898000" />
+        <register type="NOA" address="0x00009888" value="0x0D890100" />
+        <register type="NOA" address="0x00009888" value="0x03898000" />
+        <register type="NOA" address="0x00009888" value="0x09808000" />
+        <register type="NOA" address="0x00009888" value="0x0B808000" />
+        <register type="NOA" address="0x00009888" value="0x0380C000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A0075" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0000" />
+        <register type="NOA" address="0x00009888" value="0x118A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A4000" />
+        <register type="NOA" address="0x00009888" value="0x138A8000" />
+        <register type="NOA" address="0x00009888" value="0x1D81A000" />
+        <register type="NOA" address="0x00009888" value="0x15818000" />
+        <register type="NOA" address="0x00009888" value="0x17818000" />
+        <register type="NOA" address="0x00009888" value="0x0B820030" />
+        <register type="NOA" address="0x00009888" value="0x07828000" />
+        <register type="NOA" address="0x00009888" value="0x0D824000" />
+        <register type="NOA" address="0x00009888" value="0x0F828000" />
+        <register type="NOA" address="0x00009888" value="0x05824000" />
+        <register type="NOA" address="0x00009888" value="0x0D830003" />
+        <register type="NOA" address="0x00009888" value="0x0583000C" />
+        <register type="NOA" address="0x00009888" value="0x09830000" />
+        <register type="NOA" address="0x00009888" value="0x03838000" />
+        <register type="NOA" address="0x00009888" value="0x07838000" />
+        <register type="NOA" address="0x00009888" value="0x0B840980" />
+        <register type="NOA" address="0x00009888" value="0x03844D80" />
+        <register type="NOA" address="0x00009888" value="0x11840000" />
+        <register type="NOA" address="0x00009888" value="0x09848000" />
+        <register type="NOA" address="0x00009888" value="0x09850080" />
+        <register type="NOA" address="0x00009888" value="0x03850003" />
+        <register type="NOA" address="0x00009888" value="0x01850000" />
+        <register type="NOA" address="0x00009888" value="0x07860000" />
+        <register type="NOA" address="0x00009888" value="0x0F860400" />
+        <register type="NOA" address="0x00009888" value="0x09870032" />
+        <register type="NOA" address="0x00009888" value="0x01888052" />
+        <register type="NOA" address="0x00009888" value="0x11880000" />
+        <register type="NOA" address="0x00009888" value="0x09884000" />
+        <register type="NOA" address="0x00009888" value="0x1B931001" />
+        <register type="NOA" address="0x00009888" value="0x1D930001" />
+        <register type="NOA" address="0x00009888" value="0x19934000" />
+        <register type="NOA" address="0x00009888" value="0x1B958000" />
+        <register type="NOA" address="0x00009888" value="0x1D950094" />
+        <register type="NOA" address="0x00009888" value="0x19958000" />
+        <register type="NOA" address="0x00009888" value="0x09E58000" />
+        <register type="NOA" address="0x00009888" value="0x0BE58000" />
+        <register type="NOA" address="0x00009888" value="0x03E5C000" />
+        <register type="NOA" address="0x00009888" value="0x0592C000" />
+        <register type="NOA" address="0x00009888" value="0x0B928000" />
+        <register type="NOA" address="0x00009888" value="0x0D924000" />
+        <register type="NOA" address="0x00009888" value="0x0F924000" />
+        <register type="NOA" address="0x00009888" value="0x11928000" />
+        <register type="NOA" address="0x00009888" value="0x1392C000" />
+        <register type="NOA" address="0x00009888" value="0x09924000" />
+        <register type="NOA" address="0x00009888" value="0x01985000" />
+        <register type="NOA" address="0x00009888" value="0x07988000" />
+        <register type="NOA" address="0x00009888" value="0x09981000" />
+        <register type="NOA" address="0x00009888" value="0x0B982000" />
+        <register type="NOA" address="0x00009888" value="0x0D982000" />
+        <register type="NOA" address="0x00009888" value="0x0F989000" />
+        <register type="NOA" address="0x00009888" value="0x05982000" />
+        <register type="NOA" address="0x00009888" value="0x13904000" />
+        <register type="NOA" address="0x00009888" value="0x21904000" />
+        <register type="NOA" address="0x00009888" value="0x23904000" />
+        <register type="NOA" address="0x00009888" value="0x25908000" />
+        <register type="NOA" address="0x00009888" value="0x27904000" />
+        <register type="NOA" address="0x00009888" value="0x29908000" />
+        <register type="NOA" address="0x00009888" value="0x2B904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1190C080" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900440" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900400" />
+        <register type="NOA" address="0x00009888" value="0x47900C21" />
+        <register type="NOA" address="0x00009888" value="0x57900400" />
+        <register type="NOA" address="0x00009888" value="0x49900042" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900024" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900841" />
+        <register type="NOA" address="0x00009888" value="0x53900400" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFEA" />
+        <register type="OA" address="0x00002774" value="0x00007FFC" />
+        <register type="OA" address="0x00002778" value="0x0007AFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000F5FD" />
+        <register type="OA" address="0x00002780" value="0x00079FFA" />
+        <register type="OA" address="0x00002784" value="0x0000F3FB" />
+        <register type="OA" address="0x00002788" value="0x0007BF7A" />
+        <register type="OA" address="0x0000278C" value="0x0000F7E7" />
+        <register type="OA" address="0x00002790" value="0x0007FEFA" />
+        <register type="OA" address="0x00002794" value="0x0000F7CF" />
+        <register type="OA" address="0x00002798" value="0x00077FFA" />
+        <register type="OA" address="0x0000279C" value="0x0000EFDF" />
+        <register type="OA" address="0x000027A0" value="0x0006FFFA" />
+        <register type="OA" address="0x000027A4" value="0x0000CFBF" />
+        <register type="OA" address="0x000027A8" value="0x0003FFFA" />
+        <register type="OA" address="0x000027AC" value="0x00005F7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Reads Distribution Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_reads"
+       hw_config_guid="96ec2219-040b-428a-856a-6bc03363a057"
+       chipset="KBLGT3"
+       symbol_name="MemoryReads"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank0Reads"
+             description="The total number of GTI memory reads from L3 Bank 0 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_reads"
+             units="messages"
+             symbol_name="GtiL3Bank0Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all accesses from GTI to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiL3Bank3Reads"
+             description="The total number of GTI memory reads from L3 Bank 3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_reads"
+             units="messages"
+             symbol_name="GtiL3Bank3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiRsMemoryReads"
+             description="The total number of GTI memory reads from Resource Streamer."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_rs_memory_reads"
+             units="messages"
+             symbol_name="GtiRsMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Resource Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiHizMemoryReads"
+             description="The total number of GTI memory reads from Hierarchical Depth Cache (Hi-Depth Cache misses)."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_reads"
+             units="messages"
+             symbol_name="GtiHizMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="GtiRccMemoryReads"
+             description="The total number of GTI memory reads from Render Color Cache (Render Color Cache misses)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_reads"
+             units="messages"
+             symbol_name="GtiRccMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank1Reads"
+             description="The total number of GTI memory reads from L3 Bank 1 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_reads"
+             units="messages"
+             symbol_name="GtiL3Bank1Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiCmdStreamerMemoryReads"
+             description="The total number of GTI memory reads from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_reads"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="GtiL3Bank2Reads"
+             description="The total number of GTI memory reads from L3 Bank 2 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_reads"
+             units="messages"
+             symbol_name="GtiL3Bank2Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiMemoryReads"
+             description="The total number of GTI memory reads."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_reads"
+             units="messages"
+             symbol_name="GtiMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GtiRczMemoryReads"
+             description="The total number of GTI memory reads from Render Depth Cache (Render Depth Cache misses)."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_reads"
+             units="messages"
+             symbol_name="GtiRczMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="GtiMscMemoryReads"
+             description="The total number of GTI memory reads from Multisampling Color Cache (Multisampling Color Cache misses)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_reads"
+             units="messages"
+             symbol_name="GtiMscMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiVfMemoryReads"
+             description="The total number of GTI memory reads from Vertex Fetch."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="gti_vf_memory_reads"
+             units="messages"
+             symbol_name="GtiVfMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Vertex Fetch"
+             />
+    <counter name="GtiStcMemoryReads"
+             description="The total number of GTI memory reads from Stencil Cache (Stencil Cache misses)."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_reads"
+             units="messages"
+             symbol_name="GtiStcMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Reads"
+             description="The total number of GTI memory reads from L3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Reads $GtiL3Bank1Reads $GtiL3Bank2Reads $GtiL3Bank3Reads UADD UADD UADD"
+             underscore_name="gti_l3_reads"
+             units="messages"
+             symbol_name="GtiL3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F900064" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930055" />
+        <register type="NOA" address="0x00009888" value="0x03E58000" />
+        <register type="NOA" address="0x00009888" value="0x05E5C000" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x13900150" />
+        <register type="NOA" address="0x00009888" value="0x21900151" />
+        <register type="NOA" address="0x00009888" value="0x23900152" />
+        <register type="NOA" address="0x00009888" value="0x25900153" />
+        <register type="NOA" address="0x00009888" value="0x27900154" />
+        <register type="NOA" address="0x00009888" value="0x29900155" />
+        <register type="NOA" address="0x00009888" value="0x2B900156" />
+        <register type="NOA" address="0x00009888" value="0x2D900157" />
+        <register type="NOA" address="0x00009888" value="0x2F90015F" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C60" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C00" />
+        <register type="NOA" address="0x00009888" value="0x47900C63" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900C63" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x0000272C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002728" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x0000271C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002718" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F872" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Writes Distribution Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_writes"
+       hw_config_guid="03372b64-4996-4d3b-aa18-790e75eeb9c2"
+       chipset="KBLGT3"
+       symbol_name="MemoryWrites"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiMemoryWrites"
+             description="The total number of GTI memory writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_writes"
+             units="messages"
+             symbol_name="GtiMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all GTI accesses to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiMscMemoryWrites"
+             description="The total number of GTI memory writes from Multisampling Color Cache (Multisampling Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_writes"
+             units="messages"
+             symbol_name="GtiMscMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiCmdStreamerMemoryWrites"
+             description="The total number of GTI memory writes from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_writes"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiL3Bank0Writes"
+             description="The total number of GTI memory writes from L3 Bank 0 (L3 Bank 0 invalidations)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_writes"
+             units="messages"
+             symbol_name="GtiL3Bank0Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank1Writes"
+             description="The total number of GTI memory writes from L3 Bank 1 (L3 Bank 1 invalidations)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_writes"
+             units="messages"
+             symbol_name="GtiL3Bank1Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank2Writes"
+             description="The total number of GTI memory writes from L3 Bank 2 (L3 Bank 2 invalidations)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_writes"
+             units="messages"
+             symbol_name="GtiL3Bank2Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank3Writes"
+             description="The total number of GTI memory writes from L3 Bank 3 (L3 Bank 3 invalidations)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_writes"
+             units="messages"
+             symbol_name="GtiL3Bank3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Writes"
+             description="The total number of GTI memory writes from L3 (L3 invalidations)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Writes $GtiL3Bank1Writes $GtiL3Bank2Writes $GtiL3Bank3Writes UADD UADD UADD"
+             underscore_name="gti_l3_writes"
+             units="messages"
+             symbol_name="GtiL3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiRccMemoryWrites"
+             description="The total number of GTI memory writes from Render Color Cache (Render Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_writes"
+             units="messages"
+             symbol_name="GtiRccMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiSoMemoryWrites"
+             description="The total number of GTI memory writes from Stream Output."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_so_memory_writes"
+             units="messages"
+             symbol_name="GtiSoMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Stream Output"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiStcMemoryWrites"
+             description="The total number of GTI memory writes from Stencil Cache."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_writes"
+             units="messages"
+             symbol_name="GtiStcMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GtiRczMemoryWrites"
+             description="The total number of GTI memory writes from Render Depth Cache."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_writes"
+             units="messages"
+             symbol_name="GtiRczMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="GtiHizMemoryWrites"
+             description="The total number of GTI memory writes from Hierarchical Depth Cache."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_writes"
+             units="messages"
+             symbol_name="GtiHizMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F901000" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930055" />
+        <register type="NOA" address="0x00009888" value="0x03E58000" />
+        <register type="NOA" address="0x00009888" value="0x05E5C000" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x13900160" />
+        <register type="NOA" address="0x00009888" value="0x21900161" />
+        <register type="NOA" address="0x00009888" value="0x23900162" />
+        <register type="NOA" address="0x00009888" value="0x25900163" />
+        <register type="NOA" address="0x00009888" value="0x27900164" />
+        <register type="NOA" address="0x00009888" value="0x29900165" />
+        <register type="NOA" address="0x00009888" value="0x2B900166" />
+        <register type="NOA" address="0x00009888" value="0x2D900167" />
+        <register type="NOA" address="0x00009888" value="0x2F900150" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C60" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C00" />
+        <register type="NOA" address="0x00009888" value="0x47900C63" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900C63" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x0000272C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002728" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x0000271C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002718" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F822" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extended Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extended"
+       hw_config_guid="31b4ce5a-bd61-4c1f-bb5d-f2e731412150"
+       chipset="KBLGT3"
+       symbol_name="ComputeExtended"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Writes 0"
+             description="The subslice 0 typed writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="typed_writes0"
+             units="messages"
+             symbol_name="TypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuTypedAtomics0"
+             description="The subslice 0 EU Typed Atomics subslice 0."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="eu_typed_atomics0"
+             units="messages"
+             symbol_name="EuTypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Atomics 0"
+             description="The subslice 0 typed atomics."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="typed_atomics0"
+             units="messages"
+             symbol_name="TypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedAtomicsPerCacheLine"
+             description="The ratio of EU typed atomics requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedAtomics0 $TypedAtomics0 FDIV"
+             underscore_name="typed_atomics_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedAtomicsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedReads0"
+             description="The subslice 0 EU Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="eu_untyped_reads0"
+             units="messages"
+             symbol_name="EuUntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Writes 0"
+             description="The subslice 0 untyped writes (including SLM writes)."
+             data_type="uint64"
+             equation="C 1 READ"
+             underscore_name="untyped_writes0"
+             units="messages"
+             symbol_name="UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedAtomics0"
+             description="The subslice 0 EU Untyped Atomics subslice 0."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="eu_untyped_atomics0"
+             units="messages"
+             symbol_name="EuUntypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedWrites0"
+             description="The subslice 0 EU Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="eu_untyped_writes0"
+             units="messages"
+             symbol_name="EuUntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedWrites0"
+             description="The subslice 0 EU A64 Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="eu_a64_untyped_writes0"
+             units="messages"
+             symbol_name="EuA64UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedWritesPerCacheLine"
+             description="The ratio of EU untyped write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuUntypedWrites0 $EuA64UntypedWrites0 UADD $UntypedWrites0 FDIV"
+             underscore_name="untyped_writes_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="UntypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedWrites0"
+             description="The subslice 0 EU Typed Writes subslice 0."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="eu_typed_writes0"
+             units="messages"
+             symbol_name="EuTypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedWritesPerCacheLine"
+             description="The ratio of EU typed write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedWrites0 $TypedWrites0 FDIV"
+             underscore_name="typed_writes_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Reads 0"
+             description="The subslice 0 typed reads."
+             data_type="uint64"
+             equation="C 2 READ"
+             underscore_name="typed_reads0"
+             units="messages"
+             symbol_name="TypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Reads 0"
+             description="The subslice 0 untyped reads (including SLM reads)."
+             data_type="uint64"
+             equation="C 3 READ"
+             underscore_name="untyped_reads0"
+             units="messages"
+             symbol_name="UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedReads0"
+             description="The subslice 0 EU A64 Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="eu_a64_untyped_reads0"
+             units="messages"
+             symbol_name="EuA64UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedReads0"
+             description="The subslice 0 EU Typed Reads subslice 0."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="eu_typed_reads0"
+             units="messages"
+             symbol_name="EuTypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedReadsPerCacheLine"
+             description="The ratio of EU untyped read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuUntypedReads0 $EuA64UntypedReads0 UADD $UntypedReads0 FDIV"
+             underscore_name="untyped_reads_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="UntypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedReadsPerCacheLine"
+             description="The ratio of EU typed read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuTypedReads0 $TypedReads0 FDIV"
+             underscore_name="typed_reads_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x106C00E0" />
+        <register type="NOA" address="0x00009888" value="0x141C8160" />
+        <register type="NOA" address="0x00009888" value="0x161C8015" />
+        <register type="NOA" address="0x00009888" value="0x181C0120" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4E8000" />
+        <register type="NOA" address="0x00009888" value="0x184E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4EAAA0" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x024E8000" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E6C0B01" />
+        <register type="NOA" address="0x00009888" value="0x006C0200" />
+        <register type="NOA" address="0x00009888" value="0x026C000C" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x001B8000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x001C0041" />
+        <register type="NOA" address="0x00009888" value="0x061C4200" />
+        <register type="NOA" address="0x00009888" value="0x081C4443" />
+        <register type="NOA" address="0x00009888" value="0x0A1C4645" />
+        <register type="NOA" address="0x00009888" value="0x0C1C7647" />
+        <register type="NOA" address="0x00009888" value="0x041C7357" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0030" />
+        <register type="NOA" address="0x00009888" value="0x101C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0000" />
+        <register type="NOA" address="0x00009888" value="0x121C8000" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4CAA2A" />
+        <register type="NOA" address="0x00009888" value="0x0C4C02AA" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x000DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5515" />
+        <register type="NOA" address="0x00009888" value="0x100F0155" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162CAA00" />
+        <register type="NOA" address="0x00009888" value="0x182C00AA" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x11907FFF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900040" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900802" />
+        <register type="NOA" address="0x00009888" value="0x47900842" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900842" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900800" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FC2A" />
+        <register type="OA" address="0x00002774" value="0x0000BF00" />
+        <register type="OA" address="0x00002778" value="0x0007FC6A" />
+        <register type="OA" address="0x0000277C" value="0x0000BF00" />
+        <register type="OA" address="0x00002780" value="0x0007FC92" />
+        <register type="OA" address="0x00002784" value="0x0000BF00" />
+        <register type="OA" address="0x00002788" value="0x0007FCA2" />
+        <register type="OA" address="0x0000278C" value="0x0000BF00" />
+        <register type="OA" address="0x00002790" value="0x0007FC32" />
+        <register type="OA" address="0x00002794" value="0x0000BF00" />
+        <register type="OA" address="0x00002798" value="0x0007FC9A" />
+        <register type="OA" address="0x0000279C" value="0x0000BF00" />
+        <register type="OA" address="0x000027A0" value="0x0007FE6A" />
+        <register type="OA" address="0x000027A4" value="0x0000BF00" />
+        <register type="OA" address="0x000027A8" value="0x0007FE7A" />
+        <register type="OA" address="0x000027AC" value="0x0000BF00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics L3 Cache Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_l3_cache"
+       hw_config_guid="2ce0911a-27fc-4887-96f0-11084fa807c3"
+       chipset="KBLGT3"
+       symbol_name="ComputeL3Cache"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 03 Accesses"
+             description="The total number of accesses to L3 Bank 03."
+             data_type="uint64"
+             equation="B 3 READ 2 UMUL"
+             underscore_name="l3_bank03_accesses"
+             units="messages"
+             symbol_name="L3Bank03Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="L3 Accesses"
+             description="The total number of L3 accesses from all entities."
+             data_type="uint64"
+             equation="C 0 READ C 1 READ B 2 READ B 3 READ UADD UADD UADD 2 UMUL"
+             underscore_name="l3_accesses"
+             units="messages"
+             symbol_name="L3Accesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 29 READ 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="EU FPU0 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu0_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ C 5 READ UADD"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="L3 Bank 00 Accesses"
+             description="The total number of accesses to L3 Bank 00."
+             data_type="uint64"
+             equation="C 0 READ 2 UMUL"
+             underscore_name="l3_bank00_accesses"
+             units="messages"
+             symbol_name="L3Bank00Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="EU FPU0 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 19 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu0_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 14 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu1_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ C 5 READ UADD 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 00 IC Accesses"
+             description="The total number of accesses to L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 0 READ B 1 READ UADD 2 UMUL $L3Bank00Accesses UMIN"
+             underscore_name="l3_bank00_ic_accesses"
+             units="messages"
+             symbol_name="L3Bank00IcAccesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="L3 Bank 00 IC Hits"
+             description="The total number of hits in L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 1 READ 2 UMUL $L3Bank00IcAccesses UMIN"
+             underscore_name="l3_bank00_ic_hits"
+             units="messages"
+             symbol_name="L3Bank00IcHits"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="Sampler Accesses"
+             description="The total number of messages send to samplers."
+             data_type="uint64"
+             equation="A 28 READ"
+             underscore_name="sampler_accesses"
+             units="messages"
+             symbol_name="SamplerAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler"
+             />
+    <counter name="L3 Bank 01 Accesses"
+             description="The total number of accesses to L3 Bank 01."
+             data_type="uint64"
+             equation="C 1 READ 2 UMUL"
+             underscore_name="l3_bank01_accesses"
+             units="messages"
+             symbol_name="L3Bank01Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 20 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu1_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="L3 Bank 02 Accesses"
+             description="The total number of accesses to L3 Bank 02."
+             data_type="uint64"
+             equation="B 2 READ 2 UMUL"
+             underscore_name="l3_bank02_accesses"
+             units="messages"
+             symbol_name="L3Bank02Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="L3 Total Throughput"
+             description="The total number of GPU memory bytes transferred via L3."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Accesses 64 UMUL"
+             underscore_name="l3_total_throughput"
+             units="bytes"
+             symbol_name="L3TotalThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="EU FPU1 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x166C0760" />
+        <register type="NOA" address="0x00009888" value="0x1593001E" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4E8000" />
+        <register type="NOA" address="0x00009888" value="0x184E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4E8020" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x006C0051" />
+        <register type="NOA" address="0x00009888" value="0x066C5000" />
+        <register type="NOA" address="0x00009888" value="0x086C5C5D" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5E5F" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x186C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x001B4000" />
+        <register type="NOA" address="0x00009888" value="0x061B8000" />
+        <register type="NOA" address="0x00009888" value="0x081BC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x101C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1CE000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0030" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C2A00" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0280" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F1500" />
+        <register type="NOA" address="0x00009888" value="0x100F0140" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162C0A00" />
+        <register type="NOA" address="0x00009888" value="0x182C00A0" />
+        <register type="NOA" address="0x00009888" value="0x03933300" />
+        <register type="NOA" address="0x00009888" value="0x05930032" />
+        <register type="NOA" address="0x00009888" value="0x11930000" />
+        <register type="NOA" address="0x00009888" value="0x1B930000" />
+        <register type="NOA" address="0x00009888" value="0x1D900157" />
+        <register type="NOA" address="0x00009888" value="0x1F900158" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1190030F" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900000" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900021" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x53904444" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFFA" />
+        <register type="OA" address="0x00002774" value="0x0000FEFE" />
+        <register type="OA" address="0x00002778" value="0x0007FFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000FEFD" />
+        <register type="OA" address="0x00002790" value="0x0007FFFA" />
+        <register type="OA" address="0x00002794" value="0x0000FBEF" />
+        <register type="OA" address="0x00002798" value="0x0007FFFA" />
+        <register type="OA" address="0x0000279C" value="0x0000FBDF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00101100" />
+        <register type="FLEX" address="0x0000E45C" value="0x00201200" />
+        <register type="FLEX" address="0x0000E55C" value="0x00301300" />
+        <register type="FLEX" address="0x0000E65C" value="0x00401400" />
+    </register_config>
+  </set>
+
+  <set name="Metric set HDCAndSF"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="hdc_and_sf"
+       hw_config_guid="546c4c1d-99b8-42fb-a107-5aaabb5314a8"
+       chipset="KBLGT3"
+       symbol_name="HDCAndSF"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Polygon Data Ready"
+             description="The percentage of time in which geometry pipeline output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="poly_data_ready"
+             units="percent"
+             symbol_name="PolyDataReady"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="HDC stalled by L3 (s0.ss1)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss1)"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ C 4 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader01_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader01AccessStalledOnL3"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss2)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss2)"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ C 2 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader02_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader02AccessStalledOnL3"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss0)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss0)"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ C 6 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader00_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader00AccessStalledOnL3"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x104F0232" />
+        <register type="NOA" address="0x00009888" value="0x124F4640" />
+        <register type="NOA" address="0x00009888" value="0x106C0232" />
+        <register type="NOA" address="0x00009888" value="0x11834400" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4E8000" />
+        <register type="NOA" address="0x00009888" value="0x004F1880" />
+        <register type="NOA" address="0x00009888" value="0x024F08BB" />
+        <register type="NOA" address="0x00009888" value="0x044F001B" />
+        <register type="NOA" address="0x00009888" value="0x046C0100" />
+        <register type="NOA" address="0x00009888" value="0x066C000B" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x041B8000" />
+        <register type="NOA" address="0x00009888" value="0x061B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C1800" />
+        <register type="NOA" address="0x00009888" value="0x005B8000" />
+        <register type="NOA" address="0x00009888" value="0x025BC000" />
+        <register type="NOA" address="0x00009888" value="0x045B4000" />
+        <register type="NOA" address="0x00009888" value="0x125C8000" />
+        <register type="NOA" address="0x00009888" value="0x145C8000" />
+        <register type="NOA" address="0x00009888" value="0x165C8000" />
+        <register type="NOA" address="0x00009888" value="0x185C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C00A0" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x022CC000" />
+        <register type="NOA" address="0x00009888" value="0x042CC000" />
+        <register type="NOA" address="0x00009888" value="0x062CC000" />
+        <register type="NOA" address="0x00009888" value="0x082CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0F828000" />
+        <register type="NOA" address="0x00009888" value="0x0F8305C0" />
+        <register type="NOA" address="0x00009888" value="0x09830000" />
+        <register type="NOA" address="0x00009888" value="0x07830000" />
+        <register type="NOA" address="0x00009888" value="0x1D950080" />
+        <register type="NOA" address="0x00009888" value="0x13928000" />
+        <register type="NOA" address="0x00009888" value="0x0F988000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900040" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x10800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000FDFF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_1"
+       hw_config_guid="4e93d156-9b39-4268-8544-a8e0480806d7"
+       chipset="KBLGT3"
+       symbol_name="L3_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank1 Active"
+             description="The percentage of time in which slice0 L3 bank1 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_active"
+             units="percent"
+             symbol_name="L30Bank1Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 L3 Bank1 Stalled"
+             description="The percentage of time in which slice0 L3 bank1 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_stalled"
+             units="percent"
+             symbol_name="L30Bank1Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Active"
+             description="The percentage of time in which slice0 L3 bank0 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_active"
+             units="percent"
+             symbol_name="L30Bank0Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Stalled"
+             description="The percentage of time in which slice0 L3 bank0 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_stalled"
+             units="percent"
+             symbol_name="L30Bank0Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C7B40" />
+        <register type="NOA" address="0x00009888" value="0x166C0020" />
+        <register type="NOA" address="0x00009888" value="0x0A603444" />
+        <register type="NOA" address="0x00009888" value="0x0A613400" />
+        <register type="NOA" address="0x00009888" value="0x1A4EA800" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x024E8000" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x064F4000" />
+        <register type="NOA" address="0x00009888" value="0x0C6C5327" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5425" />
+        <register type="NOA" address="0x00009888" value="0x006C2A00" />
+        <register type="NOA" address="0x00009888" value="0x026C285B" />
+        <register type="NOA" address="0x00009888" value="0x046C005C" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0800" />
+        <register type="NOA" address="0x00009888" value="0x0C1BC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x001B8000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C003C" />
+        <register type="NOA" address="0x00009888" value="0x121C8000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0800" />
+        <register type="NOA" address="0x00009888" value="0x065B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C1000" />
+        <register type="NOA" address="0x00009888" value="0x10600000" />
+        <register type="NOA" address="0x00009888" value="0x04600000" />
+        <register type="NOA" address="0x00009888" value="0x0C610044" />
+        <register type="NOA" address="0x00009888" value="0x10610000" />
+        <register type="NOA" address="0x00009888" value="0x06610000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C02A8" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C002A" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F0154" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x182C00AA" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2CC000" />
+        <register type="NOA" address="0x00009888" value="0x1190FFC0" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900420" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900021" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900400" />
+        <register type="NOA" address="0x00009888" value="0x43900421" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_2"
+       hw_config_guid="de1bec86-ca92-4b43-89fa-147653221cc0"
+       chipset="KBLGT3"
+       symbol_name="L3_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank2 Stalled"
+             description="The percentage of time in which slice0 L3 bank2 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_stalled"
+             units="percent"
+             symbol_name="L30Bank2Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank2 Active"
+             description="The percentage of time in which slice0 L3 bank2 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_active"
+             units="percent"
+             symbol_name="L30Bank2Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C02E0" />
+        <register type="NOA" address="0x00009888" value="0x146C0001" />
+        <register type="NOA" address="0x00009888" value="0x0A623400" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x064F4000" />
+        <register type="NOA" address="0x00009888" value="0x026C3324" />
+        <register type="NOA" address="0x00009888" value="0x046C3422" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0800" />
+        <register type="NOA" address="0x00009888" value="0x065B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C1000" />
+        <register type="NOA" address="0x00009888" value="0x06614000" />
+        <register type="NOA" address="0x00009888" value="0x0C620044" />
+        <register type="NOA" address="0x00009888" value="0x10620000" />
+        <register type="NOA" address="0x00009888" value="0x06620000" />
+        <register type="NOA" address="0x00009888" value="0x084C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C002A" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F4000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2CC000" />
+        <register type="NOA" address="0x00009888" value="0x1190F800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00028002" />
+        <register type="OA" address="0x0000277C" value="0x000087FF" />
+        <register type="OA" address="0x00002780" value="0x00020002" />
+        <register type="OA" address="0x00002784" value="0x00008FFF" />
+        <register type="OA" address="0x00002788" value="0x00008002" />
+        <register type="OA" address="0x0000278C" value="0x0000A7FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_3"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_3"
+       hw_config_guid="e63537bb-10be-4d4a-92c4-c6b0c65e02ef"
+       chipset="KBLGT3"
+       symbol_name="L3_3"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 L3 Bank3 Stalled"
+             description="The percentage of time in which slice0 L3 bank3 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_stalled"
+             units="percent"
+             symbol_name="L30Bank3Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank3 Active"
+             description="The percentage of time in which slice0 L3 bank3 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_active"
+             units="percent"
+             symbol_name="L30Bank3Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C4E80" />
+        <register type="NOA" address="0x00009888" value="0x146C0000" />
+        <register type="NOA" address="0x00009888" value="0x0A633400" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4E8000" />
+        <register type="NOA" address="0x00009888" value="0x026C3321" />
+        <register type="NOA" address="0x00009888" value="0x046C342F" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C2000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x061B4000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C1800" />
+        <register type="NOA" address="0x00009888" value="0x06604000" />
+        <register type="NOA" address="0x00009888" value="0x0C630044" />
+        <register type="NOA" address="0x00009888" value="0x10630000" />
+        <register type="NOA" address="0x00009888" value="0x06630000" />
+        <register type="NOA" address="0x00009888" value="0x084C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C00AA" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F4000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x1190F800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900002" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00028002" />
+        <register type="OA" address="0x0000277C" value="0x000087FF" />
+        <register type="OA" address="0x00002780" value="0x00020002" />
+        <register type="OA" address="0x00002784" value="0x00008FFF" />
+        <register type="OA" address="0x00002788" value="0x00008002" />
+        <register type="OA" address="0x0000278C" value="0x0000A7FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set RasterizerAndPixelBackend"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="rasterizer_and_pixel_backend"
+       hw_config_guid="7a03a9f8-ec5e-46bb-8b67-1f0ff1476281"
+       chipset="KBLGT3"
+       symbol_name="RasterizerAndPixelBackend"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 Pixel Values Ready"
+             description="The percentage of time in which slice0 pixel values are ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_values0_ready"
+             units="percent"
+             symbol_name="PixelValues0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Slice0 Rasterizer Input Available"
+             description="The percentage of time in which slice0 rasterizer input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_input_available"
+             units="percent"
+             symbol_name="Rasterizer0InputAvailable"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice0 Post-EarlyZ Pixel Data Ready"
+             description="The percentage of time in which slice0 post-EarlyZ pixel data is ready (after early Z tests have been applied)"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_data0_ready"
+             units="percent"
+             symbol_name="PixelData0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Rasterizer Output Ready"
+             description="The percentage of time in which slice0 rasterizer output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_output_ready"
+             units="percent"
+             symbol_name="Rasterizer0OutputReady"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Slice0 PS Output Available"
+             description="The percentage of time in which slice0 PS output is available"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_output0_available"
+             units="percent"
+             symbol_name="PSOutput0Available"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x102F3800" />
+        <register type="NOA" address="0x00009888" value="0x144D0500" />
+        <register type="NOA" address="0x00009888" value="0x120D03C0" />
+        <register type="NOA" address="0x00009888" value="0x140D03CF" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0004" />
+        <register type="NOA" address="0x00009888" value="0x0C4E4000" />
+        <register type="NOA" address="0x00009888" value="0x042F0480" />
+        <register type="NOA" address="0x00009888" value="0x082F0000" />
+        <register type="NOA" address="0x00009888" value="0x022F0000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0090" />
+        <register type="NOA" address="0x00009888" value="0x064D0027" />
+        <register type="NOA" address="0x00009888" value="0x004D0000" />
+        <register type="NOA" address="0x00009888" value="0x000D0D40" />
+        <register type="NOA" address="0x00009888" value="0x020D803F" />
+        <register type="NOA" address="0x00009888" value="0x040D8023" />
+        <register type="NOA" address="0x00009888" value="0x100D0000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x020F0010" />
+        <register type="NOA" address="0x00009888" value="0x000F0000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0050" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41901400" />
+        <register type="NOA" address="0x00009888" value="0x43901485" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900001" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000EFFF" />
+        <register type="OA" address="0x00002778" value="0x00006000" />
+        <register type="OA" address="0x0000277C" value="0x0000F3FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set Sampler"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="sampler"
+       hw_config_guid="b25d2ebf-a6e0-4b29-96be-a9b010edeeda"
+       chipset="KBLGT3"
+       symbol_name="Sampler"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice2 Input Available"
+             description="The percentage of time in which slice0 subslice2 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_input_available"
+             units="percent"
+             symbol_name="Sampler02InputAvailable"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice0 Input Available"
+             description="The percentage of time in which slice0 subslice0 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_input_available"
+             units="percent"
+             symbol_name="Sampler00InputAvailable"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 Subslice2 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice2 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_output_ready"
+             units="percent"
+             symbol_name="Sampler02OutputReady"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Slice0 Subslice1 Input Available"
+             description="The percentage of time in which slice0 subslice1 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_input_available"
+             units="percent"
+             symbol_name="Sampler01InputAvailable"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Slice0 Subslice0 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice0 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_output_ready"
+             units="percent"
+             symbol_name="Sampler00OutputReady"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Slice0 Subslice1 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice1 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_output_ready"
+             units="percent"
+             symbol_name="Sampler01OutputReady"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x14152C00" />
+        <register type="NOA" address="0x00009888" value="0x16150005" />
+        <register type="NOA" address="0x00009888" value="0x121600A0" />
+        <register type="NOA" address="0x00009888" value="0x14352C00" />
+        <register type="NOA" address="0x00009888" value="0x16350005" />
+        <register type="NOA" address="0x00009888" value="0x123600A0" />
+        <register type="NOA" address="0x00009888" value="0x14552C00" />
+        <register type="NOA" address="0x00009888" value="0x16550005" />
+        <register type="NOA" address="0x00009888" value="0x125600A0" />
+        <register type="NOA" address="0x00009888" value="0x062F6000" />
+        <register type="NOA" address="0x00009888" value="0x022F2000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0050" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0010" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F0350" />
+        <register type="NOA" address="0x00009888" value="0x0C0FB000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F00DA" />
+        <register type="NOA" address="0x00009888" value="0x182C0028" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x022DC000" />
+        <register type="NOA" address="0x00009888" value="0x042D4000" />
+        <register type="NOA" address="0x00009888" value="0x0C138000" />
+        <register type="NOA" address="0x00009888" value="0x0E132000" />
+        <register type="NOA" address="0x00009888" value="0x0413C000" />
+        <register type="NOA" address="0x00009888" value="0x1C140018" />
+        <register type="NOA" address="0x00009888" value="0x0C157000" />
+        <register type="NOA" address="0x00009888" value="0x0E150078" />
+        <register type="NOA" address="0x00009888" value="0x10150000" />
+        <register type="NOA" address="0x00009888" value="0x04162180" />
+        <register type="NOA" address="0x00009888" value="0x02160000" />
+        <register type="NOA" address="0x00009888" value="0x04174000" />
+        <register type="NOA" address="0x00009888" value="0x0233A000" />
+        <register type="NOA" address="0x00009888" value="0x04333000" />
+        <register type="NOA" address="0x00009888" value="0x14348000" />
+        <register type="NOA" address="0x00009888" value="0x16348000" />
+        <register type="NOA" address="0x00009888" value="0x02357870" />
+        <register type="NOA" address="0x00009888" value="0x10350000" />
+        <register type="NOA" address="0x00009888" value="0x04360043" />
+        <register type="NOA" address="0x00009888" value="0x02360000" />
+        <register type="NOA" address="0x00009888" value="0x04371000" />
+        <register type="NOA" address="0x00009888" value="0x0E538000" />
+        <register type="NOA" address="0x00009888" value="0x00538000" />
+        <register type="NOA" address="0x00009888" value="0x06533000" />
+        <register type="NOA" address="0x00009888" value="0x1C540020" />
+        <register type="NOA" address="0x00009888" value="0x12548000" />
+        <register type="NOA" address="0x00009888" value="0x0E557000" />
+        <register type="NOA" address="0x00009888" value="0x00557800" />
+        <register type="NOA" address="0x00009888" value="0x10550000" />
+        <register type="NOA" address="0x00009888" value="0x06560043" />
+        <register type="NOA" address="0x00009888" value="0x02560000" />
+        <register type="NOA" address="0x00009888" value="0x06571000" />
+        <register type="NOA" address="0x00009888" value="0x1190FF80" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900060" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900060" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x70800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x0000C000" />
+        <register type="OA" address="0x00002774" value="0x0000E7FF" />
+        <register type="OA" address="0x00002778" value="0x00003000" />
+        <register type="OA" address="0x0000277C" value="0x0000F9FF" />
+        <register type="OA" address="0x00002780" value="0x00000C00" />
+        <register type="OA" address="0x00002784" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_1"
+       hw_config_guid="469a05e5-e299-46f7-9598-7b05f3c34991"
+       chipset="KBLGT3"
+       symbol_name="TDL_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x12120000" />
+        <register type="NOA" address="0x00009888" value="0x12320000" />
+        <register type="NOA" address="0x00009888" value="0x12520000" />
+        <register type="NOA" address="0x00009888" value="0x002F8000" />
+        <register type="NOA" address="0x00009888" value="0x022F3000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0015" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F03A0" />
+        <register type="NOA" address="0x00009888" value="0x0C0FF000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0095" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2D4000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x02108000" />
+        <register type="NOA" address="0x00009888" value="0x0410C000" />
+        <register type="NOA" address="0x00009888" value="0x02118000" />
+        <register type="NOA" address="0x00009888" value="0x0411C000" />
+        <register type="NOA" address="0x00009888" value="0x02121880" />
+        <register type="NOA" address="0x00009888" value="0x041219B5" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x02134000" />
+        <register type="NOA" address="0x00009888" value="0x04135000" />
+        <register type="NOA" address="0x00009888" value="0x0C308000" />
+        <register type="NOA" address="0x00009888" value="0x0E304000" />
+        <register type="NOA" address="0x00009888" value="0x06304000" />
+        <register type="NOA" address="0x00009888" value="0x0C318000" />
+        <register type="NOA" address="0x00009888" value="0x0E314000" />
+        <register type="NOA" address="0x00009888" value="0x06314000" />
+        <register type="NOA" address="0x00009888" value="0x0C321A80" />
+        <register type="NOA" address="0x00009888" value="0x0E320033" />
+        <register type="NOA" address="0x00009888" value="0x06320031" />
+        <register type="NOA" address="0x00009888" value="0x00320000" />
+        <register type="NOA" address="0x00009888" value="0x0C334000" />
+        <register type="NOA" address="0x00009888" value="0x0E331000" />
+        <register type="NOA" address="0x00009888" value="0x06331000" />
+        <register type="NOA" address="0x00009888" value="0x0E508000" />
+        <register type="NOA" address="0x00009888" value="0x00508000" />
+        <register type="NOA" address="0x00009888" value="0x02504000" />
+        <register type="NOA" address="0x00009888" value="0x0E518000" />
+        <register type="NOA" address="0x00009888" value="0x00518000" />
+        <register type="NOA" address="0x00009888" value="0x02514000" />
+        <register type="NOA" address="0x00009888" value="0x0E521880" />
+        <register type="NOA" address="0x00009888" value="0x00521A80" />
+        <register type="NOA" address="0x00009888" value="0x02520033" />
+        <register type="NOA" address="0x00009888" value="0x0E534000" />
+        <register type="NOA" address="0x00009888" value="0x00534000" />
+        <register type="NOA" address="0x00009888" value="0x02531000" />
+        <register type="NOA" address="0x00009888" value="0x1190FF80" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900062" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x00007FFF" />
+        <register type="OA" address="0x00002778" value="0x00000000" />
+        <register type="OA" address="0x0000277C" value="0x00009FFF" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000EFFF" />
+        <register type="OA" address="0x00002788" value="0x00000000" />
+        <register type="OA" address="0x0000278C" value="0x0000F3FF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FDFF" />
+        <register type="OA" address="0x00002798" value="0x00000000" />
+        <register type="OA" address="0x0000279C" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_2"
+       hw_config_guid="52f925c6-786a-4ec6-86ce-cba85c83453a"
+       chipset="KBLGT3"
+       symbol_name="TDL_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort0"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort1"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort1"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort0"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort0"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort1"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x12124D60" />
+        <register type="NOA" address="0x00009888" value="0x12322E60" />
+        <register type="NOA" address="0x00009888" value="0x12524D60" />
+        <register type="NOA" address="0x00009888" value="0x022F3000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0014" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0FE000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0097" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x002D8000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x0410C000" />
+        <register type="NOA" address="0x00009888" value="0x0411C000" />
+        <register type="NOA" address="0x00009888" value="0x04121FB7" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x04135000" />
+        <register type="NOA" address="0x00009888" value="0x00308000" />
+        <register type="NOA" address="0x00009888" value="0x06304000" />
+        <register type="NOA" address="0x00009888" value="0x00318000" />
+        <register type="NOA" address="0x00009888" value="0x06314000" />
+        <register type="NOA" address="0x00009888" value="0x00321B80" />
+        <register type="NOA" address="0x00009888" value="0x0632003F" />
+        <register type="NOA" address="0x00009888" value="0x00334000" />
+        <register type="NOA" address="0x00009888" value="0x06331000" />
+        <register type="NOA" address="0x00009888" value="0x0250C000" />
+        <register type="NOA" address="0x00009888" value="0x0251C000" />
+        <register type="NOA" address="0x00009888" value="0x02521FB7" />
+        <register type="NOA" address="0x00009888" value="0x00520000" />
+        <register type="NOA" address="0x00009888" value="0x02535000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x43900063" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extra Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extra"
+       hw_config_guid="efc497ac-884e-4ee4-a4a8-15fba22aaf21"
+       chipset="KBLGT3"
+       symbol_name="ComputeExtra"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active including Ext Math"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing including Extended Math processing"
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ  C 5 READ C 6 READ FADD C 7 READ FADD C 2 READ FADD C 3 READ FADD C 4 READ FADD 8 FMUL FADD 100 FMUL $EuCoresTotalCount FDIV $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active_adjusted"
+             units="percent"
+             symbol_name="Fpu1ActiveAdjusted"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x121203E0" />
+        <register type="NOA" address="0x00009888" value="0x123203E0" />
+        <register type="NOA" address="0x00009888" value="0x125203E0" />
+        <register type="NOA" address="0x00009888" value="0x129203E0" />
+        <register type="NOA" address="0x00009888" value="0x12B203E0" />
+        <register type="NOA" address="0x00009888" value="0x12D203E0" />
+        <register type="NOA" address="0x00009888" value="0x024EC000" />
+        <register type="NOA" address="0x00009888" value="0x044EC000" />
+        <register type="NOA" address="0x00009888" value="0x064EC000" />
+        <register type="NOA" address="0x00009888" value="0x022F4000" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0042" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F006D" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x042D8000" />
+        <register type="NOA" address="0x00009888" value="0x06104000" />
+        <register type="NOA" address="0x00009888" value="0x06114000" />
+        <register type="NOA" address="0x00009888" value="0x06120033" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x06131000" />
+        <register type="NOA" address="0x00009888" value="0x04308000" />
+        <register type="NOA" address="0x00009888" value="0x04318000" />
+        <register type="NOA" address="0x00009888" value="0x04321980" />
+        <register type="NOA" address="0x00009888" value="0x00320000" />
+        <register type="NOA" address="0x00009888" value="0x04334000" />
+        <register type="NOA" address="0x00009888" value="0x04504000" />
+        <register type="NOA" address="0x00009888" value="0x04514000" />
+        <register type="NOA" address="0x00009888" value="0x04520033" />
+        <register type="NOA" address="0x00009888" value="0x00520000" />
+        <register type="NOA" address="0x00009888" value="0x04531000" />
+        <register type="NOA" address="0x00009888" value="0x00AF8000" />
+        <register type="NOA" address="0x00009888" value="0x0ACC0001" />
+        <register type="NOA" address="0x00009888" value="0x008D8000" />
+        <register type="NOA" address="0x00009888" value="0x028DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C8FB000" />
+        <register type="NOA" address="0x00009888" value="0x0E8F0001" />
+        <register type="NOA" address="0x00009888" value="0x06AC8000" />
+        <register type="NOA" address="0x00009888" value="0x02AD4000" />
+        <register type="NOA" address="0x00009888" value="0x02908000" />
+        <register type="NOA" address="0x00009888" value="0x02918000" />
+        <register type="NOA" address="0x00009888" value="0x02921980" />
+        <register type="NOA" address="0x00009888" value="0x00920000" />
+        <register type="NOA" address="0x00009888" value="0x02934000" />
+        <register type="NOA" address="0x00009888" value="0x02B04000" />
+        <register type="NOA" address="0x00009888" value="0x02B14000" />
+        <register type="NOA" address="0x00009888" value="0x02B20033" />
+        <register type="NOA" address="0x00009888" value="0x00B20000" />
+        <register type="NOA" address="0x00009888" value="0x02B31000" />
+        <register type="NOA" address="0x00009888" value="0x00D08000" />
+        <register type="NOA" address="0x00009888" value="0x00D18000" />
+        <register type="NOA" address="0x00009888" value="0x00D21980" />
+        <register type="NOA" address="0x00009888" value="0x00D34000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x43900002" />
+        <register type="NOA" address="0x00009888" value="0x53900420" />
+        <register type="NOA" address="0x00009888" value="0x459000A1" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+  </set>
+
+  <set name="Media Vme Pipe Gen9"
+       mdapi_supported_apis="MEDIA IO BB"
+       underscore_name="vme_pipe"
+       hw_config_guid="bfd9764d-2c5b-4c16-bfc1-89de3ca10917"
+       chipset="KBLGT3"
+       symbol_name="VMEPipe"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL  GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="8 A 10 READ FMUL $EuThreadsCount FDIV $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VME Busy"
+             description="The percentage of time in which VME (IME or CRE) was actively processing data."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ B 3 READ FADD 2 FDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vme_busy"
+             units="percent"
+             symbol_name="VMEBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Overview System Batch Tier2"
+             mdapi_group="VME Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x141A5800" />
+        <register type="NOA" address="0x00009888" value="0x161A00C0" />
+        <register type="NOA" address="0x00009888" value="0x12180240" />
+        <register type="NOA" address="0x00009888" value="0x14180002" />
+        <register type="NOA" address="0x00009888" value="0x149A5800" />
+        <register type="NOA" address="0x00009888" value="0x169A00C0" />
+        <register type="NOA" address="0x00009888" value="0x12980240" />
+        <register type="NOA" address="0x00009888" value="0x14980002" />
+        <register type="NOA" address="0x00009888" value="0x1A4E3FC0" />
+        <register type="NOA" address="0x00009888" value="0x002F1000" />
+        <register type="NOA" address="0x00009888" value="0x022F8000" />
+        <register type="NOA" address="0x00009888" value="0x042F3000" />
+        <register type="NOA" address="0x00009888" value="0x004C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C9500" />
+        <register type="NOA" address="0x00009888" value="0x0C4C002A" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5500" />
+        <register type="NOA" address="0x00009888" value="0x100F0015" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162CAA00" />
+        <register type="NOA" address="0x00009888" value="0x182C000A" />
+        <register type="NOA" address="0x00009888" value="0x04193000" />
+        <register type="NOA" address="0x00009888" value="0x081A28C1" />
+        <register type="NOA" address="0x00009888" value="0x001A0000" />
+        <register type="NOA" address="0x00009888" value="0x00133000" />
+        <register type="NOA" address="0x00009888" value="0x0613C000" />
+        <register type="NOA" address="0x00009888" value="0x0813F000" />
+        <register type="NOA" address="0x00009888" value="0x00172000" />
+        <register type="NOA" address="0x00009888" value="0x06178000" />
+        <register type="NOA" address="0x00009888" value="0x0817A000" />
+        <register type="NOA" address="0x00009888" value="0x00180037" />
+        <register type="NOA" address="0x00009888" value="0x06180940" />
+        <register type="NOA" address="0x00009888" value="0x08180000" />
+        <register type="NOA" address="0x00009888" value="0x02180000" />
+        <register type="NOA" address="0x00009888" value="0x04183000" />
+        <register type="NOA" address="0x00009888" value="0x04AFC000" />
+        <register type="NOA" address="0x00009888" value="0x06AF3000" />
+        <register type="NOA" address="0x00009888" value="0x0ACC4000" />
+        <register type="NOA" address="0x00009888" value="0x0CCC0015" />
+        <register type="NOA" address="0x00009888" value="0x0A8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E8F4000" />
+        <register type="NOA" address="0x00009888" value="0x108F0015" />
+        <register type="NOA" address="0x00009888" value="0x16ACA000" />
+        <register type="NOA" address="0x00009888" value="0x18AC000A" />
+        <register type="NOA" address="0x00009888" value="0x06993000" />
+        <register type="NOA" address="0x00009888" value="0x0C9A28C1" />
+        <register type="NOA" address="0x00009888" value="0x009A0000" />
+        <register type="NOA" address="0x00009888" value="0x0A93F000" />
+        <register type="NOA" address="0x00009888" value="0x0C93F000" />
+        <register type="NOA" address="0x00009888" value="0x0A97A000" />
+        <register type="NOA" address="0x00009888" value="0x0C97A000" />
+        <register type="NOA" address="0x00009888" value="0x0A980977" />
+        <register type="NOA" address="0x00009888" value="0x08980000" />
+        <register type="NOA" address="0x00009888" value="0x04980000" />
+        <register type="NOA" address="0x00009888" value="0x06983000" />
+        <register type="NOA" address="0x00009888" value="0x119000FF" />
+        <register type="NOA" address="0x00009888" value="0x51900040" />
+        <register type="NOA" address="0x00009888" value="0x41900020" />
+        <register type="NOA" address="0x00009888" value="0x55900004" />
+        <register type="NOA" address="0x00009888" value="0x45900400" />
+        <register type="NOA" address="0x00009888" value="0x479008A5" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900002" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00100030" />
+        <register type="OA" address="0x00002774" value="0x0000FFF9" />
+        <register type="OA" address="0x00002778" value="0x00000002" />
+        <register type="OA" address="0x0000277C" value="0x0000FFFC" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000FFF3" />
+        <register type="OA" address="0x00002788" value="0x00100180" />
+        <register type="OA" address="0x0000278C" value="0x0000FFCF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00000002" />
+        <register type="OA" address="0x0000279C" value="0x0000FF3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00008003" />
+    </register_config>
+  </set>
+
+  <set name="MDAPI testing set Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="test_oa"
+       hw_config_guid="f1792f32-6db2-4b50-b4b2-557128f1688d"
+       chipset="KBLGT3"
+       symbol_name="TestOa"
+       >
+    <counter name="TestCounter7"
+             description="HW test counter 7. Factor: 0.666"
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="counter7"
+             units="events"
+             symbol_name="Counter7"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TestCounter8"
+             description="HW test counter 8. Should be equal to 1."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="counter8"
+             units="events"
+             symbol_name="Counter8"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter4"
+             description="HW test counter 4. Factor: 0.333"
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="counter4"
+             units="events"
+             symbol_name="Counter4"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter5"
+             description="HW test counter 5. Factor: 0.333"
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="counter5"
+             units="events"
+             symbol_name="Counter5"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter6"
+             description="HW test counter 6. Factor: 0.166"
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="counter6"
+             units="events"
+             symbol_name="Counter6"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter3"
+             description="HW test counter 3. Factor: 0.5"
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="counter3"
+             units="events"
+             symbol_name="Counter3"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter0"
+             description="HW test counter 0. Factor: 0.0"
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="counter0"
+             units="events"
+             symbol_name="Counter0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter1"
+             description="HW test counter 1. Factor: 1.0"
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="counter1"
+             units="events"
+             symbol_name="Counter1"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter2"
+             description="HW test counter 2. Factor: 1.0"
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="counter2"
+             units="events"
+             symbol_name="Counter2"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810000" />
+        <register type="NOA" address="0x00009888" value="0x07810013" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930040" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x00000004" />
+        <register type="OA" address="0x00002774" value="0x00000000" />
+        <register type="OA" address="0x00002778" value="0x00000003" />
+        <register type="OA" address="0x0000277C" value="0x00000000" />
+        <register type="OA" address="0x00002780" value="0x00000007" />
+        <register type="OA" address="0x00002784" value="0x00000000" />
+        <register type="OA" address="0x00002788" value="0x00100002" />
+        <register type="OA" address="0x0000278C" value="0x0000FFF7" />
+        <register type="OA" address="0x00002790" value="0x00100002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00100082" />
+        <register type="OA" address="0x0000279C" value="0x0000FFEF" />
+        <register type="OA" address="0x000027A0" value="0x001000C2" />
+        <register type="OA" address="0x000027A4" value="0x0000FFE7" />
+        <register type="OA" address="0x000027A8" value="0x00100001" />
+        <register type="OA" address="0x000027AC" value="0x0000FFE7" />
+    </register_config>
+  </set>
+
+</metrics>
diff --git a/src/mesa/drivers/dri/i965/brw_performance_query.c b/src/mesa/drivers/dri/i965/brw_performance_query.c
index 90a3d588bdf..efb0b9c3fab 100644
--- a/src/mesa/drivers/dri/i965/brw_performance_query.c
+++ b/src/mesa/drivers/dri/i965/brw_performance_query.c
@@ -78,6 +78,8 @@
 #include "brw_oa_sklgt3.h"
 #include "brw_oa_sklgt4.h"
 #include "brw_oa_bxt.h"
+#include "brw_oa_kblgt2.h"
+#include "brw_oa_kblgt3.h"
 #include "intel_batchbuffer.h"
 
 #define FILE_DEBUG_FLAG DEBUG_PERFMON
@@ -2009,6 +2011,12 @@ get_register_queries_function(const struct gen_device_info *devinfo)
       if (devinfo->gt == 4)
          return brw_oa_register_queries_sklgt4;
    }
+   if (devinfo->is_kabylake) {
+      if (devinfo->gt == 2)
+         return brw_oa_register_queries_kblgt2;
+      if (devinfo->gt == 3)
+         return brw_oa_register_queries_kblgt3;
+   }
    return NULL;
 }