From: Robert Bragg <robert@sixbynine.org>
Date: Fri, 5 Jun 2015 12:44:56 +0000 (+0100)
Subject: i965: Add more Haswell OA metrics sets
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=a678b79ef4e97cad7cc488dd55f6983d9861e823;p=mesa.git

i965: Add more Haswell OA metrics sets

This extends the brw_oa_hsw.xml to expose these additional queries:

- Compute Metrics Basic Gen7.5
- Compute Metrics Extended Gen7.5
- Memory Reads Distribution Gen7.5
- Memory Writes Distribution Gen7.5
- Metric set Sampler Balance

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

diff --git a/src/mesa/drivers/dri/i965/brw_oa_hsw.xml b/src/mesa/drivers/dri/i965/brw_oa_hsw.xml
index 49476712633..739ea0e377d 100644
--- a/src/mesa/drivers/dri/i965/brw_oa_hsw.xml
+++ b/src/mesa/drivers/dri/i965/brw_oa_hsw.xml
@@ -1,5 +1,5 @@
 <?xml version="1.0"?>
-<metrics version="1486657040" merge_md5="">
+<metrics version="1486658299" merge_md5="">
   <set name="Render Metrics Basic Gen7.5"
        mdapi_supported_apis="OGL OCL IO BB"
        underscore_name="render_basic"
@@ -995,4 +995,3406 @@
     </register_config>
   </set>
 
+  <set name="Compute Metrics Basic Gen7.5"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_basic"
+       hw_config_guid="39ad14bc-2380-45c4-91eb-fbcb3aa7ae7b"
+       chipset="HSW"
+       symbol_name="ComputeBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="C 2 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OGL4 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 0 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="TES EU Stall"
+             description="The percentage of time in which evaluation shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_eu_stall"
+             units="percent"
+             symbol_name="DsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Bytes Written"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 6 READ B 7 READ 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="Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Alpha Test Fails"
+             description="The total number of pixels dropped on post-FS alpha test."
+             data_type="uint64"
+             equation="A 37 READ"
+             underscore_name="alpha_test_fails"
+             units="pixels"
+             symbol_name="AlphaTestFails"
+             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="TCS Threads Dispatched"
+             description="The total number of control shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 10 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/Control Shader"
+             />
+    <counter name="TCS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which control shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 7 READ $HsThreads UDIV"
+             underscore_name="hs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="HsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Control 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 25 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="GS EU Stall"
+             description="The percentage of time in which geometry shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 23 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_eu_stall"
+             units="percent"
+             symbol_name="GsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Geometry 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 5 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             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/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 30 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             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/Fragment Shader"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 20 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="CS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which compute shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 17 READ $CsThreads UDIV"
+             underscore_name="cs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="CsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 0 READ B 1 READ 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="Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which geometry shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 23 READ $GsThreads UDIV"
+             underscore_name="gs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="GsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Typed Atomics"
+             description="The total number of typed atomics."
+             data_type="uint64"
+             equation="C 0 READ C 1 READ UADD $EuSlicesTotalCount UMUL"
+             underscore_name="typed_atomics"
+             units="messages"
+             symbol_name="TypedAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 6 READ C 7 READ UADD $EuSlicesTotalCount 64 UMUL 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 Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which fragment shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 28 READ $PsThreads UDIV"
+             underscore_name="ps_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="PsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="TES EU Active"
+             description="The percentage of time in which evaluation shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_eu_active"
+             units="percent"
+             symbol_name="DsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which vertex shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 3 READ $VsThreads UDIV"
+             underscore_name="vs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="VsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="TCS EU Stall"
+             description="The percentage of time in which control shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_eu_stall"
+             units="percent"
+             symbol_name="HsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has being processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 41 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="FS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which fragment shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 27 READ $PsThreads UDIV"
+             underscore_name="ps_eu_active_per_thread"
+             units="cycles"
+             symbol_name="PsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="TES Threads Dispatched"
+             description="The total number of evaluation shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 15 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/Evaluation Shader"
+             />
+    <counter name="TES AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which evaluation shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 12 READ $DsThreads UDIV"
+             underscore_name="ds_eu_active_per_thread"
+             units="cycles"
+             symbol_name="DsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS EU Active"
+             description="The percentage of time in which geometry shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 22 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_eu_active"
+             units="percent"
+             symbol_name="GsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS EU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 27 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_active"
+             units="percent"
+             symbol_name="PsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TES AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which evaluation shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 13 READ $DsThreads UDIV"
+             underscore_name="ds_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="DsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Evaluation Shader"
+             />
+    <counter name="GS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which geometry shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 22 READ $GsThreads UDIV"
+             underscore_name="gs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="GsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Writes"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 2 READ B 3 READ UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_written"
+             units="bytes"
+             symbol_name="UntypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of byten written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ C 5 READ UADD $EuSlicesTotalCount 64 UMUL 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 Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TCS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which control shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 8 READ $HsThreads UDIV"
+             underscore_name="hs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="HsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Control Shader"
+             />
+    <counter name="TCS EU Active"
+             description="The percentage of time in which control shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_eu_active"
+             units="percent"
+             symbol_name="HsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             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 36 READ"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Late Depth Test Fails"
+             description="The total number of pixels dropped on post-FS depth test."
+             data_type="uint64"
+             equation="A 39 READ $SamplesKilledInPs USUB"
+             underscore_name="post_ps_depth_test_fails"
+             units="pixels"
+             symbol_name="PostPsDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="FS EU Stall"
+             description="The percentage of time in which fragment shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 28 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_stall"
+             units="percent"
+             symbol_name="PsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Fragment 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 1 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="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 33 READ"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 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 EU Active"
+             description="The percentage of time in which vertex shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 2 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_eu_active"
+             units="percent"
+             symbol_name="VsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS EU Active"
+             description="The percentage of time in which compute shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cs_eu_active"
+             units="percent"
+             symbol_name="CsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which compute shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 18 READ $CsThreads UDIV"
+             underscore_name="cs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="CsEuStallPerThread"
+             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="Late Stencil Test Fails"
+             description="The total number of pixels dropped on post-FS stencil test."
+             data_type="uint64"
+             equation="A 38 READ"
+             underscore_name="post_ps_stencil_test_fails"
+             units="pixels"
+             symbol_name="PostPsStencilTestFails"
+             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="VS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which vertex shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 2 READ $VsThreads UDIV"
+             underscore_name="vs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="VsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS EU Stall"
+             description="The percentage of time in which compute shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cs_eu_stall"
+             units="percent"
+             symbol_name="CsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 4 READ B 5 READ 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="Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 40 READ"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="VS EU Stall"
+             description="The percentage of time in which vertex shaders were stalled on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 3 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_eu_stall"
+             units="percent"
+             symbol_name="VsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config id="0">
+        <register type="NOA" address="0x000253A4" value="0x00000000" />
+        <register type="NOA" address="0x0002681C" value="0x01F00800" />
+        <register type="NOA" address="0x00026820" value="0x00001000" />
+        <register type="NOA" address="0x0002781C" value="0x01F00800" />
+        <register type="NOA" address="0x00026520" value="0x00000007" />
+        <register type="NOA" address="0x000265A0" value="0x00000007" />
+        <register type="NOA" address="0x00025380" value="0x00000010" />
+        <register type="NOA" address="0x0002538C" value="0x00300000" />
+        <register type="NOA" address="0x00025384" value="0xAA8AAAAA" />
+        <register type="NOA" address="0x00025404" value="0xFFFFFFFF" />
+        <register type="NOA" address="0x00026800" value="0x00004202" />
+        <register type="NOA" address="0x00026808" value="0x00605817" />
+        <register type="NOA" address="0x0002680C" value="0x10001005" />
+        <register type="NOA" address="0x00026804" value="0x00000000" />
+        <register type="NOA" address="0x00027800" value="0x00000102" />
+        <register type="NOA" address="0x00027808" value="0x0C0701E0" />
+        <register type="NOA" address="0x0002780C" value="0x000200A0" />
+        <register type="NOA" address="0x00027804" value="0x00000000" />
+        <register type="NOA" address="0x00026484" value="0x44000000" />
+        <register type="NOA" address="0x00026704" value="0x44000000" />
+        <register type="NOA" address="0x00026500" value="0x00000006" />
+        <register type="NOA" address="0x00026510" value="0x00000001" />
+        <register type="NOA" address="0x00026504" value="0x88000000" />
+        <register type="NOA" address="0x00026580" value="0x00000006" />
+        <register type="NOA" address="0x00026590" value="0x00000020" />
+        <register type="NOA" address="0x00026584" value="0x00000000" />
+        <register type="NOA" address="0x00026104" value="0x55822222" />
+        <register type="NOA" address="0x00026184" value="0xAA866666" />
+        <register type="NOA" address="0x00025420" value="0x08320C83" />
+        <register type="NOA" address="0x00025424" value="0x06820C83" />
+        <register type="NOA" address="0x0002541C" value="0x00000000" />
+        <register type="NOA" address="0x00025428" value="0x00000C03" />
+    </register_config>
+    <register_config id="1">
+        <register type="OA" address="0x00002710" value="0x00000000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002714" value="0x00800000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002718" value="0xAAAAAAAA" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x0000271C" value="0xAAAAAAAA" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002720" value="0x00000000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002724" value="0x00800000" /> <!--OASTARTTRIG6 -->
+        <register type="OA" address="0x00002728" value="0xAAAAAAAA" /> <!--OASTARTTRIG7 -->
+        <register type="OA" address="0x0000272C" value="0xAAAAAAAA" /> <!--OASTARTTRIG8 -->
+        <register type="OA" address="0x00002740" value="0x00000000" /> <!--OAREPORTTRIG1 -->
+        <register type="OA" address="0x00002744" value="0x00000000" /> <!--OAREPORTTRIG2 -->
+        <register type="OA" address="0x00002748" value="0x00000000" /> <!--OAREPORTTRIG3 -->
+        <register type="OA" address="0x0000274C" value="0x00000000" /> <!--OAREPORTTRIG4 -->
+        <register type="OA" address="0x00002750" value="0x00000000" /> <!--OAREPORTTRIG5 -->
+        <register type="OA" address="0x00002754" value="0x00000000" /> <!--OAREPORTTRIG6 -->
+        <register type="OA" address="0x00002758" value="0x00000000" /> <!--OAREPORTTRIG7 -->
+        <register type="OA" address="0x0000275C" value="0x00000000" /> <!--OAREPORTTRIG8 -->
+        <register type="OA" address="0x0000236C" value="0x00000000" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extended Gen7.5"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extended"
+       hw_config_guid="3865be28-6982-49fe-9494-e4d1b4795413"
+       chipset="HSW"
+       symbol_name="ComputeExtended"
+       >
+    <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="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="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 $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="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="Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <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="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="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="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="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="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="Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gpu_clocks"
+             units="cycles"
+             symbol_name="GpuClocks"
+             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="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 20 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="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="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="Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <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="Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUrbAtomics0"
+             description="The subslice 0 EU URB Atomics subslice 0."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="eu_urb_atomics0"
+             units="messages"
+             symbol_name="EuUrbAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="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="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 $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="Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TypedReadsPerCacheLine"
+             description="The ratio of EU typed read requests to L3 cache line reads."
+             data_type="float"
+             equation="$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="Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <register_config id="0">
+        <register type="NOA" address="0x0002681C" value="0x3EB00800" />
+        <register type="NOA" address="0x00026820" value="0x00900000" />
+        <register type="NOA" address="0x00025384" value="0x02AAAAAA" />
+        <register type="NOA" address="0x00025404" value="0x03FFFFFF" />
+        <register type="NOA" address="0x00026800" value="0x00142284" />
+        <register type="NOA" address="0x00026808" value="0x0E629062" />
+        <register type="NOA" address="0x0002680C" value="0x3F6F55CB" />
+        <register type="NOA" address="0x00026810" value="0x00000014" />
+        <register type="NOA" address="0x00026804" value="0x00000000" />
+        <register type="NOA" address="0x00026104" value="0x02AAAAAA" />
+        <register type="NOA" address="0x00026184" value="0x02AAAAAA" />
+        <register type="NOA" address="0x00025420" value="0x00000000" />
+        <register type="NOA" address="0x00025424" value="0x00000000" />
+        <register type="NOA" address="0x0002541C" value="0x00000000" />
+        <register type="NOA" address="0x00025428" value="0x00000000" />
+    </register_config>
+    <register_config id="1">
+        <register type="OA" address="0x00002724" value="0xF0800000" /> <!--OASTARTTRIG6 -->
+        <register type="OA" address="0x00002720" value="0x00000000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002714" value="0xF0800000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002710" value="0x00000000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002770" value="0x0007FE2A" /> <!--OACEC0_0 -->
+        <register type="OA" address="0x00002774" value="0x0000FF00" /> <!--OACEC0_1 -->
+        <register type="OA" address="0x00002778" value="0x0007FE6A" /> <!--OACEC1_0 -->
+        <register type="OA" address="0x0000277C" value="0x0000FF00" /> <!--OACEC1_1 -->
+        <register type="OA" address="0x00002780" value="0x0007FE92" /> <!--OACEC2_0 -->
+        <register type="OA" address="0x00002784" value="0x0000FF00" /> <!--OACEC2_1 -->
+        <register type="OA" address="0x00002788" value="0x0007FEA2" /> <!--OACEC3_0 -->
+        <register type="OA" address="0x0000278C" value="0x0000FF00" /> <!--OACEC3_1 -->
+        <register type="OA" address="0x00002790" value="0x0007FE32" /> <!--OACEC4_0 -->
+        <register type="OA" address="0x00002794" value="0x0000FF00" /> <!--OACEC4_1 -->
+        <register type="OA" address="0x00002798" value="0x0007FE9A" /> <!--OACEC5_0 -->
+        <register type="OA" address="0x0000279C" value="0x0000FF00" /> <!--OACEC5_1 -->
+        <register type="OA" address="0x000027A0" value="0x0007FF23" /> <!--OACEC6_0 -->
+        <register type="OA" address="0x000027A4" value="0x0000FF00" /> <!--OACEC6_1 -->
+        <register type="OA" address="0x000027A8" value="0x0007FFF3" /> <!--OACEC7_0 -->
+        <register type="OA" address="0x000027AC" value="0x0000FFFE" /> <!--OACEC7_1 -->
+    </register_config>
+  </set>
+
+  <set name="Memory Reads Distribution Gen7.5"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_reads"
+       hw_config_guid="bb5ed49b-2497-4095-94f6-26ba294db88a"
+       chipset="HSW"
+       symbol_name="MemoryReads"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="C 7 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 0 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="TES EU Stall"
+             description="The percentage of time in which evaluation shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_eu_stall"
+             units="percent"
+             symbol_name="DsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Alpha Test Fails"
+             description="The total number of pixels dropped on post-FS alpha test."
+             data_type="uint64"
+             equation="A 37 READ"
+             underscore_name="alpha_test_fails"
+             units="pixels"
+             symbol_name="AlphaTestFails"
+             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="TES Threads Dispatched"
+             description="The total number of evaluation shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 15 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/Evaluation Shader"
+             />
+    <counter name="TES AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which evaluation shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 12 READ $DsThreads UDIV"
+             underscore_name="ds_eu_active_per_thread"
+             units="cycles"
+             symbol_name="DsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation 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 25 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="GS EU Stall"
+             description="The percentage of time in which geometry shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 23 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_eu_stall"
+             units="percent"
+             symbol_name="GsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Geometry 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 5 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="LLC GPU Read Accesses"
+             description="The total number of LLC cache lookups for reads done from the GPU."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="llc_read_accesses"
+             units="messages"
+             symbol_name="LlcReadAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="LLC"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 30 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="GtiMemoryReads"
+             description="The total number of GTI memory reads (64B each)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_memory_reads"
+             units="messages"
+             symbol_name="GtiMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 20 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="CS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which compute shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 17 READ $CsThreads UDIV"
+             underscore_name="cs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="CsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiRczMemoryReads"
+             description="The total number of GTI memory reads from Render Depth Cache."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_reads"
+             units="messages"
+             symbol_name="GtiRczMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="GS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which geometry shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 23 READ $GsThreads UDIV"
+             underscore_name="gs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="GsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <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="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=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="FS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which fragment shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 28 READ $PsThreads UDIV"
+             underscore_name="ps_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="PsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="TES EU Active"
+             description="The percentage of time in which evaluation shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_eu_active"
+             units="percent"
+             symbol_name="DsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which vertex shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 3 READ $VsThreads UDIV"
+             underscore_name="vs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="VsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="TCS EU Stall"
+             description="The percentage of time in which control shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_eu_stall"
+             units="percent"
+             symbol_name="HsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has being processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 41 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="FS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which fragment shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 27 READ $PsThreads UDIV"
+             underscore_name="ps_eu_active_per_thread"
+             units="cycles"
+             symbol_name="PsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Fragment Shader"
+             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="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 35 READ"
+             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="GS EU Active"
+             description="The percentage of time in which geometry shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 22 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_eu_active"
+             units="percent"
+             symbol_name="GsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS EU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 27 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_active"
+             units="percent"
+             symbol_name="PsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TES AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which evaluation shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 13 READ $DsThreads UDIV"
+             underscore_name="ds_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="DsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Evaluation Shader"
+             />
+    <counter name="GS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which geometry shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 22 READ $GsThreads UDIV"
+             underscore_name="gs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="GsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiHiDepthMemoryReads"
+             description="The total number of GTI memory reads from Hierarchical Depth Cache."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hi_depth_memory_reads"
+             units="messages"
+             symbol_name="GtiHiDepthMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="TCS Threads Dispatched"
+             description="The total number of control shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 10 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/Control Shader"
+             />
+    <counter name="TCS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which control shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 8 READ $HsThreads UDIV"
+             underscore_name="hs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="HsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Control Shader"
+             />
+    <counter name="TCS EU Active"
+             description="The percentage of time in which control shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_eu_active"
+             units="percent"
+             symbol_name="HsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             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 36 READ"
+             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="Late Depth Test Fails"
+             description="The total number of pixels dropped on post-FS depth test."
+             data_type="uint64"
+             equation="A 39 READ $SamplesKilledInPs USUB"
+             underscore_name="post_ps_depth_test_fails"
+             units="pixels"
+             symbol_name="PostPsDepthTestFails"
+             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="TCS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which control shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 7 READ $HsThreads UDIV"
+             underscore_name="hs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="HsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS EU Stall"
+             description="The percentage of time in which fragment shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 28 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_stall"
+             units="percent"
+             symbol_name="PsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Fragment 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 1 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="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=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 33 READ"
+             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 EU Active"
+             description="The percentage of time in which vertex shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 2 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_eu_active"
+             units="percent"
+             symbol_name="VsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS EU Active"
+             description="The percentage of time in which compute shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cs_eu_active"
+             units="percent"
+             symbol_name="CsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which compute shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 18 READ $CsThreads UDIV"
+             underscore_name="cs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="CsEuStallPerThread"
+             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="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=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Vertex Fetch"
+             />
+    <counter name="Late Stencil Test Fails"
+             description="The total number of pixels dropped on post-FS stencil test."
+             data_type="uint64"
+             equation="A 38 READ"
+             underscore_name="post_ps_stencil_test_fails"
+             units="pixels"
+             symbol_name="PostPsStencilTestFails"
+             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="GtiStcMemoryReads"
+             description="The total number of GTI memory reads from Stencil Cache."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_reads"
+             units="messages"
+             symbol_name="GtiStcMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="VS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which vertex shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 2 READ $VsThreads UDIV"
+             underscore_name="vs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="VsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <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="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Resource Streamer"
+             />
+    <counter name="CS EU Stall"
+             description="The percentage of time in which compute shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cs_eu_stall"
+             units="percent"
+             symbol_name="CsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 40 READ"
+             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="GtiL3Reads"
+             description="The total number of GTI memory reads from L3 (L3 misses)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_reads"
+             units="messages"
+             symbol_name="GtiL3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="VS EU Stall"
+             description="The percentage of time in which vertex shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 3 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_eu_stall"
+             units="percent"
+             symbol_name="VsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config id="0">
+        <register type="NOA" address="0x000253A4" value="0x34300000" />
+        <register type="NOA" address="0x00025440" value="0x2D800000" />
+        <register type="NOA" address="0x00025444" value="0x00000008" />
+        <register type="NOA" address="0x00025128" value="0x0E600000" />
+        <register type="NOA" address="0x00025380" value="0x00000450" />
+        <register type="NOA" address="0x00025390" value="0x00052C43" />
+        <register type="NOA" address="0x00025384" value="0x00000000" />
+        <register type="NOA" address="0x00025400" value="0x00006144" />
+        <register type="NOA" address="0x00025408" value="0x0A418820" />
+        <register type="NOA" address="0x0002540C" value="0x000820E6" />
+        <register type="NOA" address="0x00025404" value="0xFF500000" />
+        <register type="NOA" address="0x00025100" value="0x000005D6" />
+        <register type="NOA" address="0x0002510C" value="0x0EF00000" />
+        <register type="NOA" address="0x00025104" value="0x00000000" />
+        <register type="NOA" address="0x00025420" value="0x02108421" />
+        <register type="NOA" address="0x00025424" value="0x00008421" />
+        <register type="NOA" address="0x0002541C" value="0x00000000" />
+        <register type="NOA" address="0x00025428" value="0x00000000" />
+    </register_config>
+    <register_config id="1">
+        <register type="OA" address="0x00002724" value="0xF0800000" /> <!--OASTARTTRIG6 -->
+        <register type="OA" address="0x00002720" value="0x00000000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002714" value="0xF0800000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002710" value="0x00000000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x0000274C" value="0x76543298" /> <!--OAREPORTTRIG4 -->
+        <register type="OA" address="0x00002748" value="0x98989898" /> <!--OAREPORTTRIG3 -->
+        <register type="OA" address="0x00002744" value="0x000000E4" /> <!--OAREPORTTRIG2 -->
+        <register type="OA" address="0x00002740" value="0x00000000" /> <!--OAREPORTTRIG1 -->
+        <register type="OA" address="0x0000275C" value="0x98A98A98" /> <!--OAREPORTTRIG8 -->
+        <register type="OA" address="0x00002758" value="0x88888888" /> <!--OAREPORTTRIG7 -->
+        <register type="OA" address="0x00002754" value="0x000C5500" /> <!--OAREPORTTRIG6 -->
+        <register type="OA" address="0x00002750" value="0x00000000" /> <!--OAREPORTTRIG5 -->
+        <register type="OA" address="0x00002770" value="0x0007F81A" /> <!--OACEC0_0 -->
+        <register type="OA" address="0x00002774" value="0x0000FC00" /> <!--OACEC0_1 -->
+        <register type="OA" address="0x00002778" value="0x0007F82A" /> <!--OACEC1_0 -->
+        <register type="OA" address="0x0000277C" value="0x0000FC00" /> <!--OACEC1_1 -->
+        <register type="OA" address="0x00002780" value="0x0007F872" /> <!--OACEC2_0 -->
+        <register type="OA" address="0x00002784" value="0x0000FC00" /> <!--OACEC2_1 -->
+        <register type="OA" address="0x00002788" value="0x0007F8BA" /> <!--OACEC3_0 -->
+        <register type="OA" address="0x0000278C" value="0x0000FC00" /> <!--OACEC3_1 -->
+        <register type="OA" address="0x00002790" value="0x0007F87A" /> <!--OACEC4_0 -->
+        <register type="OA" address="0x00002794" value="0x0000FC00" /> <!--OACEC4_1 -->
+        <register type="OA" address="0x00002798" value="0x0007F8EA" /> <!--OACEC5_0 -->
+        <register type="OA" address="0x0000279C" value="0x0000FC00" /> <!--OACEC5_1 -->
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" /> <!--OACEC6_0 -->
+        <register type="OA" address="0x000027A4" value="0x0000FC00" /> <!--OACEC6_1 -->
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" /> <!--OACEC7_0 -->
+        <register type="OA" address="0x000027AC" value="0x0000FC00" /> <!--OACEC7_1 -->
+    </register_config>
+  </set>
+
+  <set name="Memory Writes Distribution Gen7.5"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_writes"
+       hw_config_guid="3358d639-9b5f-45ab-976d-9b08cbfc6240"
+       chipset="HSW"
+       symbol_name="MemoryWrites"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="C 7 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 0 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="TES EU Stall"
+             description="The percentage of time in which evaluation shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_eu_stall"
+             units="percent"
+             symbol_name="DsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Alpha Test Fails"
+             description="The total number of pixels dropped on post-FS alpha test."
+             data_type="uint64"
+             equation="A 37 READ"
+             underscore_name="alpha_test_fails"
+             units="pixels"
+             symbol_name="AlphaTestFails"
+             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="TES Threads Dispatched"
+             description="The total number of evaluation shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 15 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/Evaluation Shader"
+             />
+    <counter name="TES AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which evaluation shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 12 READ $DsThreads UDIV"
+             underscore_name="ds_eu_active_per_thread"
+             units="cycles"
+             symbol_name="DsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation 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 25 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="GS EU Stall"
+             description="The percentage of time in which geometry shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 23 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_eu_stall"
+             units="percent"
+             symbol_name="GsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Geometry 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 5 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 30 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=""
+             mdapi_usage_flags="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 20 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="CS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which compute shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 17 READ $CsThreads UDIV"
+             underscore_name="cs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="CsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <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="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="GS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which geometry shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 23 READ $GsThreads UDIV"
+             underscore_name="gs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="GsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiL3Writes"
+             description="The total number of GTI memory writes from L3 (L3 invalidations)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_writes"
+             units="messages"
+             symbol_name="GtiL3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <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=""
+             mdapi_usage_flags="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="FS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which fragment shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 28 READ $PsThreads UDIV"
+             underscore_name="ps_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="PsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <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=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="TES EU Active"
+             description="The percentage of time in which evaluation shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_eu_active"
+             units="percent"
+             symbol_name="DsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which vertex shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 3 READ $VsThreads UDIV"
+             underscore_name="vs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="VsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="TCS EU Stall"
+             description="The percentage of time in which control shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_eu_stall"
+             units="percent"
+             symbol_name="HsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has being processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 41 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="FS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which fragment shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 27 READ $PsThreads UDIV"
+             underscore_name="ps_eu_active_per_thread"
+             units="cycles"
+             symbol_name="PsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 35 READ"
+             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="LLC GPU Write Accesses"
+             description="The total number of LLC cache lookups for write done from the GPU (32B writes)."
+             data_type="uint64"
+             equation="C 6 READ 2 UMUL"
+             underscore_name="llc_wr_accesses"
+             units="messages"
+             symbol_name="LlcWrAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="LLC"
+             />
+    <counter name="GS EU Active"
+             description="The percentage of time in which geometry shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 22 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_eu_active"
+             units="percent"
+             symbol_name="GsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS EU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 27 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_active"
+             units="percent"
+             symbol_name="PsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Fragment Shader"
+             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=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="TES AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which evaluation shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 13 READ $DsThreads UDIV"
+             underscore_name="ds_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="DsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Evaluation Shader"
+             />
+    <counter name="GS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which geometry shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 22 READ $GsThreads UDIV"
+             underscore_name="gs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="GsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TCS Threads Dispatched"
+             description="The total number of control shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 10 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/Control Shader"
+             />
+    <counter name="TCS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which control shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 8 READ $HsThreads UDIV"
+             underscore_name="hs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="HsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Control Shader"
+             />
+    <counter name="TCS EU Active"
+             description="The percentage of time in which control shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_eu_active"
+             units="percent"
+             symbol_name="HsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             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 36 READ"
+             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="Late Depth Test Fails"
+             description="The total number of pixels dropped on post-FS depth test."
+             data_type="uint64"
+             equation="A 39 READ $SamplesKilledInPs USUB"
+             underscore_name="post_ps_depth_test_fails"
+             units="pixels"
+             symbol_name="PostPsDepthTestFails"
+             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="TCS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which control shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 7 READ $HsThreads UDIV"
+             underscore_name="hs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="HsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS EU Stall"
+             description="The percentage of time in which fragment shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 28 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_stall"
+             units="percent"
+             symbol_name="PsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Fragment 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 1 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="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 33 READ"
+             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 EU Active"
+             description="The percentage of time in which vertex shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 2 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_eu_active"
+             units="percent"
+             symbol_name="VsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS EU Active"
+             description="The percentage of time in which compute shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cs_eu_active"
+             units="percent"
+             symbol_name="CsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which compute shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 18 READ $CsThreads UDIV"
+             underscore_name="cs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="CsEuStallPerThread"
+             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="Late Stencil Test Fails"
+             description="The total number of pixels dropped on post-FS stencil test."
+             data_type="uint64"
+             equation="A 38 READ"
+             underscore_name="post_ps_stencil_test_fails"
+             units="pixels"
+             symbol_name="PostPsStencilTestFails"
+             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="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=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Stream Output"
+             />
+    <counter name="VS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which vertex shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 2 READ $VsThreads UDIV"
+             underscore_name="vs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="VsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <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=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="CS EU Stall"
+             description="The percentage of time in which compute shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cs_eu_stall"
+             units="percent"
+             symbol_name="CsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 40 READ"
+             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="GtiMemoryWrites"
+             description="The total number of GTI memory writes (64B each)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_memory_writes"
+             units="messages"
+             symbol_name="GtiMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="VS EU Stall"
+             description="The percentage of time in which vertex shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 3 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_eu_stall"
+             units="percent"
+             symbol_name="VsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config id="0">
+        <register type="NOA" address="0x000253A4" value="0x34300000" />
+        <register type="NOA" address="0x00025440" value="0x01500000" />
+        <register type="NOA" address="0x00025444" value="0x00000120" />
+        <register type="NOA" address="0x00025128" value="0x0C200000" />
+        <register type="NOA" address="0x00025380" value="0x00000450" />
+        <register type="NOA" address="0x00025390" value="0x00052C43" />
+        <register type="NOA" address="0x00025384" value="0x00000000" />
+        <register type="NOA" address="0x00025400" value="0x00007184" />
+        <register type="NOA" address="0x00025408" value="0x0A418820" />
+        <register type="NOA" address="0x0002540C" value="0x000820E6" />
+        <register type="NOA" address="0x00025404" value="0xFF500000" />
+        <register type="NOA" address="0x00025100" value="0x000005D6" />
+        <register type="NOA" address="0x0002510C" value="0x1E700000" />
+        <register type="NOA" address="0x00025104" value="0x00000000" />
+        <register type="NOA" address="0x00025420" value="0x02108421" />
+        <register type="NOA" address="0x00025424" value="0x00008421" />
+        <register type="NOA" address="0x0002541C" value="0x00000000" />
+        <register type="NOA" address="0x00025428" value="0x00000000" />
+    </register_config>
+    <register_config id="1">
+        <register type="OA" address="0x00002724" value="0xF0800000" /> <!--OASTARTTRIG6 -->
+        <register type="OA" address="0x00002720" value="0x00000000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002714" value="0xF0800000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002710" value="0x00000000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x0000274C" value="0x76543298" /> <!--OAREPORTTRIG4 -->
+        <register type="OA" address="0x00002748" value="0x98989898" /> <!--OAREPORTTRIG3 -->
+        <register type="OA" address="0x00002744" value="0x000000E4" /> <!--OAREPORTTRIG2 -->
+        <register type="OA" address="0x00002740" value="0x00000000" /> <!--OAREPORTTRIG1 -->
+        <register type="OA" address="0x0000275C" value="0xBABABABA" /> <!--OAREPORTTRIG8 -->
+        <register type="OA" address="0x00002758" value="0x88888888" /> <!--OAREPORTTRIG7 -->
+        <register type="OA" address="0x00002754" value="0x000C5500" /> <!--OAREPORTTRIG6 -->
+        <register type="OA" address="0x00002750" value="0x00000000" /> <!--OAREPORTTRIG5 -->
+        <register type="OA" address="0x00002770" value="0x0007F81A" /> <!--OACEC0_0 -->
+        <register type="OA" address="0x00002774" value="0x0000FC00" /> <!--OACEC0_1 -->
+        <register type="OA" address="0x00002778" value="0x0007F82A" /> <!--OACEC1_0 -->
+        <register type="OA" address="0x0000277C" value="0x0000FC00" /> <!--OACEC1_1 -->
+        <register type="OA" address="0x00002780" value="0x0007F822" /> <!--OACEC2_0 -->
+        <register type="OA" address="0x00002784" value="0x0000FC00" /> <!--OACEC2_1 -->
+        <register type="OA" address="0x00002788" value="0x0007F8BA" /> <!--OACEC3_0 -->
+        <register type="OA" address="0x0000278C" value="0x0000FC00" /> <!--OACEC3_1 -->
+        <register type="OA" address="0x00002790" value="0x0007F87A" /> <!--OACEC4_0 -->
+        <register type="OA" address="0x00002794" value="0x0000FC00" /> <!--OACEC4_1 -->
+        <register type="OA" address="0x00002798" value="0x0007F8EA" /> <!--OACEC5_0 -->
+        <register type="OA" address="0x0000279C" value="0x0000FC00" /> <!--OACEC5_1 -->
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" /> <!--OACEC6_0 -->
+        <register type="OA" address="0x000027A4" value="0x0000FC00" /> <!--OACEC6_1 -->
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" /> <!--OACEC7_0 -->
+        <register type="OA" address="0x000027AC" value="0x0000FC00" /> <!--OACEC7_1 -->
+    </register_config>
+  </set>
+
+  <set name="Metric set SamplerBalance"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="sampler_balance"
+       hw_config_guid="bc274488-b4b6-40c7-90da-b77d7ad16189"
+       chipset="HSW"
+       symbol_name="SamplerBalance"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="B 7 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 0 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="TES EU Stall"
+             description="The percentage of time in which evaluation shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_eu_stall"
+             units="percent"
+             symbol_name="DsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler L2 cache misses (ss3)"
+             description="Number of sampler L2 cache misses (ss3)"
+             data_type="uint64"
+             equation="C 1 READ C 0 READ UADD"
+             underscore_name="sampler3_l2_cache_misses"
+             units="messages"
+             symbol_name="Sampler3L2CacheMisses"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="Alpha Test Fails"
+             description="The total number of pixels dropped on post-FS alpha test."
+             data_type="uint64"
+             equation="A 37 READ"
+             underscore_name="alpha_test_fails"
+             units="pixels"
+             symbol_name="AlphaTestFails"
+             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="TES Threads Dispatched"
+             description="The total number of evaluation shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 15 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/Evaluation Shader"
+             />
+    <counter name="TES AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which evaluation shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 12 READ $DsThreads UDIV"
+             underscore_name="ds_eu_active_per_thread"
+             units="cycles"
+             symbol_name="DsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation 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 25 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="GS EU Stall"
+             description="The percentage of time in which geometry shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 23 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_eu_stall"
+             units="percent"
+             symbol_name="GsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS EU Active"
+             description="The percentage of time in which compute shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cs_eu_active"
+             units="percent"
+             symbol_name="CsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS EU Active"
+             description="The percentage of time in which vertex shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 2 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_eu_active"
+             units="percent"
+             symbol_name="VsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TCS EU Active"
+             description="The percentage of time in which control shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_eu_active"
+             units="percent"
+             symbol_name="HsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TES EU Active"
+             description="The percentage of time in which evaluation shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_eu_active"
+             units="percent"
+             symbol_name="DsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS EU Active"
+             description="The percentage of time in which geometry shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 22 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_eu_active"
+             units="percent"
+             symbol_name="GsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS EU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 27 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_active"
+             units="percent"
+             symbol_name="PsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS EU Stall"
+             description="The percentage of time in which compute shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cs_eu_stall"
+             units="percent"
+             symbol_name="CsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Compute 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 1 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="VS EU Stall"
+             description="The percentage of time in which vertex shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 3 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_eu_stall"
+             units="percent"
+             symbol_name="VsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TCS EU Stall"
+             description="The percentage of time in which control shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_eu_stall"
+             units="percent"
+             symbol_name="HsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS EU Stall"
+             description="The percentage of time in which fragment shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 28 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_stall"
+             units="percent"
+             symbol_name="PsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Fragment Shader"
+             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="CS Duration"
+             description="Total Compute Shader GPU duration."
+             data_type="uint64"
+             equation="A 17 READ  A 0 READ UMUL  A 2 READ  A 7 READ UADD  A 12 READ UADD  A 17 READ UADD  A 22 READ UADD  A 27 READ UADD UDIV  A 18 READ  A 1 READ UMUL  A 3 READ  A 8 READ UADD  A 13 READ UADD  A 18 READ UADD  A 23 READ UADD  A 28 READ UADD UDIV UADD $GpuTime UMUL $GpuCoreClocks $EuCoresTotalCount UMUL 1000 UMUL UDIV"
+             underscore_name="cs_duration"
+             units="us"
+             symbol_name="CsDuration"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 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 30 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="TES Duration"
+             description="Total Evaluation Shader GPU duration."
+             data_type="uint64"
+             equation="A 12 READ  A 0 READ UMUL  A 2 READ  A 7 READ UADD  A 12 READ UADD  A 17 READ UADD  A 22 READ UADD  A 27 READ UADD UDIV  A 13 READ  A 1 READ UMUL  A 3 READ  A 8 READ UADD  A 13 READ UADD  A 18 READ UADD  A 23 READ UADD  A 28 READ UADD UDIV UADD $GpuTime UMUL $GpuCoreClocks $EuCoresTotalCount UMUL 1000 UMUL UDIV"
+             underscore_name="ds_duration"
+             units="us"
+             symbol_name="DsDuration"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Evaluation Shader"
+             />
+    <counter name="GS Duration"
+             description="Total Geometry Shader GPU duration."
+             data_type="uint64"
+             equation="A 22 READ  A 0 READ UMUL  A 2 READ  A 7 READ UADD  A 12 READ UADD  A 17 READ UADD  A 22 READ UADD  A 27 READ UADD UDIV  A 23 READ  A 1 READ UMUL  A 3 READ  A 8 READ UADD  A 13 READ UADD  A 18 READ UADD  A 23 READ UADD  A 28 READ UADD UDIV UADD $GpuTime UMUL $GpuCoreClocks $EuCoresTotalCount UMUL 1000 UMUL UDIV"
+             underscore_name="gs_duration"
+             units="us"
+             symbol_name="GsDuration"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 20 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="CS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which compute shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 17 READ $CsThreads UDIV"
+             underscore_name="cs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="CsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler L2 cache misses"
+             description="Number of sampler L2 cache misses"
+             data_type="uint64"
+             equation="C 7 READ C 6 READ UADD C 5 READ UADD C 4 READ UADD C 3 READ UADD C 2 READ UADD C 1 READ UADD C 0 READ UADD"
+             underscore_name="sampler_l2_cache_misses"
+             units="messages"
+             symbol_name="SamplerL2CacheMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler"
+             />
+    <counter name="GS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which geometry shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 23 READ $GsThreads UDIV"
+             underscore_name="gs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="GsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Sampler L2 cache misses (ss2)"
+             description="Number of sampler L2 cache misses (ss2)"
+             data_type="uint64"
+             equation="C 3 READ C 2 READ UADD"
+             underscore_name="sampler2_l2_cache_misses"
+             units="messages"
+             symbol_name="Sampler2L2CacheMisses"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <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 Idle"
+             description="The percentage of time in which the Execution Units were idle."
+             data_type="float"
+             max_equation="100"
+             equation="100 $EuActive $EuStall FADD FSUB"
+             underscore_name="eu_idle"
+             units="percent"
+             symbol_name="EuIdle"
+             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="FS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which fragment shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 28 READ $PsThreads UDIV"
+             underscore_name="ps_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="PsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="VS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which vertex shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 3 READ $VsThreads UDIV"
+             underscore_name="vs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="VsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has being processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 41 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="FS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which fragment shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 27 READ $PsThreads UDIV"
+             underscore_name="ps_eu_active_per_thread"
+             units="cycles"
+             symbol_name="PsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 35 READ"
+             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="TCS Duration"
+             description="Total Control Shader GPU duration."
+             data_type="uint64"
+             equation="A 7 READ  A 0 READ UMUL  A 2 READ  A 7 READ UADD  A 12 READ UADD  A 17 READ UADD  A 22 READ UADD  A 27 READ UADD UDIV  A 8 READ  A 1 READ UMUL  A 3 READ  A 8 READ UADD  A 13 READ UADD  A 18 READ UADD  A 23 READ UADD  A 28 READ UADD UDIV UADD $GpuTime UMUL $GpuCoreClocks $EuCoresTotalCount UMUL 1000 UMUL UDIV"
+             underscore_name="hs_duration"
+             units="us"
+             symbol_name="HsDuration"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Control Shader"
+             />
+    <counter name="TES AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which evaluation shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 13 READ $DsThreads UDIV"
+             underscore_name="ds_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="DsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Evaluation Shader"
+             />
+    <counter name="GS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which geometry shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 22 READ $GsThreads UDIV"
+             underscore_name="gs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="GsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TCS Threads Dispatched"
+             description="The total number of control shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 10 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/Control Shader"
+             />
+    <counter name="TCS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which control shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 8 READ $HsThreads UDIV"
+             underscore_name="hs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="HsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Control Shader"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 36 READ"
+             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="Late Depth Test Fails"
+             description="The total number of pixels dropped on post-FS depth test."
+             data_type="uint64"
+             equation="A 39 READ $SamplesKilledInPs USUB"
+             underscore_name="post_ps_depth_test_fails"
+             units="pixels"
+             symbol_name="PostPsDepthTestFails"
+             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="TCS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which control shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 7 READ $HsThreads UDIV"
+             underscore_name="hs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="HsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS Duration"
+             description="Total Fragment Shader GPU duration."
+             data_type="uint64"
+             equation="A 27 READ  A 0 READ UMUL  A 2 READ  A 7 READ UADD  A 12 READ UADD  A 17 READ UADD  A 22 READ UADD  A 27 READ UADD UDIV  A 28 READ  A 1 READ UMUL  A 3 READ  A 8 READ UADD  A 13 READ UADD  A 18 READ UADD  A 23 READ UADD  A 28 READ UADD UDIV UADD $GpuTime UMUL $GpuCoreClocks $EuCoresTotalCount UMUL 1000 UMUL UDIV"
+             underscore_name="ps_duration"
+             units="us"
+             symbol_name="PsDuration"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Sampler L2 cache misses (ss1)"
+             description="Number of sampler L2 cache misses (ss1)"
+             data_type="uint64"
+             equation="C 5 READ C 4 READ UADD"
+             underscore_name="sampler1_l2_cache_misses"
+             units="messages"
+             symbol_name="Sampler1L2CacheMisses"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 33 READ"
+             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="CS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which compute shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 18 READ $CsThreads UDIV"
+             underscore_name="cs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="CsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Late Stencil Test Fails"
+             description="The total number of pixels dropped on post-FS stencil test."
+             data_type="uint64"
+             equation="A 38 READ"
+             underscore_name="post_ps_stencil_test_fails"
+             units="pixels"
+             symbol_name="PostPsStencilTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Sampler L2 cache misses (ss0)"
+             description="Number of sampler L2 cache misses (ss0)"
+             data_type="uint64"
+             equation="C 7 READ C 6 READ UADD"
+             underscore_name="sampler0_l2_cache_misses"
+             units="messages"
+             symbol_name="Sampler0L2CacheMisses"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="VS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which vertex shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 2 READ $VsThreads UDIV"
+             underscore_name="vs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="VsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Duration"
+             description="Total Vertex Shader GPU duration."
+             data_type="uint64"
+             equation="A 2 READ  A 0 READ UMUL  A 2 READ  A 7 READ UADD  A 12 READ UADD  A 17 READ UADD  A 22 READ UADD  A 27 READ UADD UDIV  A 3 READ  A 1 READ UMUL  A 3 READ  A 8 READ UADD  A 13 READ UADD  A 18 READ UADD  A 23 READ UADD  A 28 READ UADD UDIV UADD $GpuTime UMUL $GpuCoreClocks $EuCoresTotalCount UMUL 1000 UMUL UDIV"
+             underscore_name="vs_duration"
+             units="us"
+             symbol_name="VsDuration"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 40 READ"
+             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"
+             />
+    <register_config id="0">
+        <register type="NOA" address="0x0002EB9C" value="0x01906400" />
+        <register type="NOA" address="0x0002FB9C" value="0x01906400" />
+        <register type="NOA" address="0x000253A4" value="0x00000000" />
+        <register type="NOA" address="0x00026B9C" value="0x01906400" />
+        <register type="NOA" address="0x00027B9C" value="0x01906400" />
+        <register type="NOA" address="0x00027104" value="0x00A00000" />
+        <register type="NOA" address="0x00027184" value="0x00A50000" />
+        <register type="NOA" address="0x0002E804" value="0x00500000" />
+        <register type="NOA" address="0x0002E984" value="0x00500000" />
+        <register type="NOA" address="0x0002EB04" value="0x00500000" />
+        <register type="NOA" address="0x0002EB80" value="0x00000084" />
+        <register type="NOA" address="0x0002EB8C" value="0x14200000" />
+        <register type="NOA" address="0x0002EB84" value="0x00000000" />
+        <register type="NOA" address="0x0002F804" value="0x00050000" />
+        <register type="NOA" address="0x0002F984" value="0x00050000" />
+        <register type="NOA" address="0x0002FB04" value="0x00050000" />
+        <register type="NOA" address="0x0002FB80" value="0x00000084" />
+        <register type="NOA" address="0x0002FB8C" value="0x00050800" />
+        <register type="NOA" address="0x0002FB84" value="0x00000000" />
+        <register type="NOA" address="0x00025380" value="0x00000010" />
+        <register type="NOA" address="0x0002538C" value="0x000000C0" />
+        <register type="NOA" address="0x00025384" value="0xAA550000" />
+        <register type="NOA" address="0x00025404" value="0xFFFFC000" />
+        <register type="NOA" address="0x00026804" value="0x50000000" />
+        <register type="NOA" address="0x00026984" value="0x50000000" />
+        <register type="NOA" address="0x00026B04" value="0x50000000" />
+        <register type="NOA" address="0x00026B80" value="0x00000084" />
+        <register type="NOA" address="0x00026B90" value="0x00050800" />
+        <register type="NOA" address="0x00026B84" value="0x00000000" />
+        <register type="NOA" address="0x00027804" value="0x05000000" />
+        <register type="NOA" address="0x00027984" value="0x05000000" />
+        <register type="NOA" address="0x00027B04" value="0x05000000" />
+        <register type="NOA" address="0x00027B80" value="0x00000084" />
+        <register type="NOA" address="0x00027B90" value="0x00000142" />
+        <register type="NOA" address="0x00027B84" value="0x00000000" />
+        <register type="NOA" address="0x00026104" value="0xA0000000" />
+        <register type="NOA" address="0x00026184" value="0xA5000000" />
+        <register type="NOA" address="0x00025424" value="0x00008620" />
+        <register type="NOA" address="0x0002541C" value="0x00000000" />
+        <register type="NOA" address="0x00025428" value="0x0004A54A" />
+    </register_config>
+    <register_config id="1">
+        <register type="OA" address="0x00002740" value="0x00000000" /> <!--OAREPORTTRIG1 -->
+        <register type="OA" address="0x00002744" value="0x00800000" /> <!--OAREPORTTRIG2 -->
+        <register type="OA" address="0x00002710" value="0x00000000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002714" value="0x00800000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002720" value="0x00000000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002724" value="0x00800000" /> <!--OASTARTTRIG6 -->
+    </register_config>
+  </set>
+
 </metrics>