nir: Add a load_global_constant intrinsic

[mesa.git] / src / compiler / nir / nir_intrinsics.py

diff --git a/src/compiler/nir/nir_intrinsics.py b/src/compiler/nir/nir_intrinsics.py
index 80c61650f78894159e192eff6615181949c2a9d7..2ef14a56984fada90bf5a07b22704ac6abff0efa 100644 (file)
--- a/src/compiler/nir/nir_intrinsics.py
+++ b/src/compiler/nir/nir_intrinsics.py
@@ -134,6 +134,7 @@ MEMORY_MODES = "NIR_INTRINSIC_MEMORY_MODES"
 MEMORY_SCOPE = "NIR_INTRINSIC_MEMORY_SCOPE"
 # Scope of a control barrier
 EXECUTION_SCOPE = "NIR_INTRINSIC_EXECUTION_SCOPE"
+IO_SEMANTICS = "NIR_INTRINSIC_IO_SEMANTICS"
 
 #
 # Possible flags:
@@ -385,16 +386,16 @@ atomic3("atomic_counter_comp_swap")
 # argument with the value to be written, and image atomic operations take
 # either one or two additional scalar arguments with the same meaning as in
 # the ARB_shader_image_load_store specification.
-def image(name, src_comp=[], **kwargs):
+def image(name, src_comp=[], extra_indices=[], **kwargs):
     intrinsic("image_deref_" + name, src_comp=[1] + src_comp,
-              indices=[ACCESS], **kwargs)
+              indices=[ACCESS] + extra_indices, **kwargs)
     intrinsic("image_" + name, src_comp=[1] + src_comp,
-              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS], **kwargs)
+              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS] + extra_indices, **kwargs)
     intrinsic("bindless_image_" + name, src_comp=[1] + src_comp,
-              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS], **kwargs)
+              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS] + extra_indices, **kwargs)
 
-image("load", src_comp=[4, 1, 1], dest_comp=0, flags=[CAN_ELIMINATE])
-image("store", src_comp=[4, 1, 0, 1])
+image("load", src_comp=[4, 1, 1], extra_indices=[TYPE], dest_comp=0, flags=[CAN_ELIMINATE])
+image("store", src_comp=[4, 1, 0, 1], extra_indices=[TYPE])
 image("atomic_add",  src_comp=[4, 1, 1], dest_comp=1)
 image("atomic_imin",  src_comp=[4, 1, 1], dest_comp=1)
 image("atomic_umin",  src_comp=[4, 1, 1], dest_comp=1)
@@ -410,6 +411,9 @@ image("size",    dest_comp=0, src_comp=[1], flags=[CAN_ELIMINATE, CAN_REORDER])
 image("samples", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 image("atomic_inc_wrap",  src_comp=[4, 1, 1], dest_comp=1)
 image("atomic_dec_wrap",  src_comp=[4, 1, 1], dest_comp=1)
+# CL-specific format queries
+image("format", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
+image("order", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 
 # Vulkan descriptor set intrinsics
 #
@@ -705,7 +709,7 @@ intrinsic("load_size_ir3", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 #    float result = iid.x + iid.y * bary.y + iid.z * bary.x
 
 intrinsic("load_fs_input_interp_deltas", src_comp=[1], dest_comp=3,
-          indices=[BASE, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
+          indices=[BASE, COMPONENT, IO_SEMANTICS], flags=[CAN_ELIMINATE, CAN_REORDER])
 
 # Load operations pull data from some piece of GPU memory.  All load
 # operations operate in terms of offsets into some piece of theoretical
@@ -735,23 +739,25 @@ def load(name, src_comp, indices=[], flags=[]):
 load("uniform", [1], [BASE, RANGE, TYPE], [CAN_ELIMINATE, CAN_REORDER])
 # src[] = { buffer_index, offset }.
 load("ubo", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE, CAN_REORDER])
+# src[] = { buffer_index, offset in vec4 units }
+load("ubo_vec4", [-1, 1], [ACCESS, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
 # src[] = { offset }.
-load("input", [1], [BASE, COMPONENT, TYPE], [CAN_ELIMINATE, CAN_REORDER])
+load("input", [1], [BASE, COMPONENT, TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
 # src[] = { vertex_id, offset }.
-load("input_vertex", [1, 1], [BASE, COMPONENT, TYPE], [CAN_ELIMINATE, CAN_REORDER])
+load("input_vertex", [1, 1], [BASE, COMPONENT, TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
 # src[] = { vertex, offset }.
-load("per_vertex_input", [1, 1], [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
+load("per_vertex_input", [1, 1], [BASE, COMPONENT, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
 # src[] = { barycoord, offset }.
-load("interpolated_input", [2, 1], [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
+load("interpolated_input", [2, 1], [BASE, COMPONENT, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
 
 # src[] = { buffer_index, offset }.
 load("ssbo", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 # src[] = { buffer_index }
 load("ssbo_address", [1], [], [CAN_ELIMINATE, CAN_REORDER])
 # src[] = { offset }.
-load("output", [1], [BASE, COMPONENT], flags=[CAN_ELIMINATE])
+load("output", [1], [BASE, COMPONENT, IO_SEMANTICS], flags=[CAN_ELIMINATE])
 # src[] = { vertex, offset }.
-load("per_vertex_output", [1, 1], [BASE, COMPONENT], [CAN_ELIMINATE])
+load("per_vertex_output", [1, 1], [BASE, COMPONENT, IO_SEMANTICS], [CAN_ELIMINATE])
 # src[] = { offset }.
 load("shared", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 # src[] = { offset }.
@@ -762,6 +768,9 @@ load("constant", [1], [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET],
 # src[] = { address }.
 load("global", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 # src[] = { address }.
+load("global_constant", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET],
+     [CAN_ELIMINATE, CAN_REORDER])
+# src[] = { address }.
 load("kernel_input", [1], [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE, CAN_REORDER])
 # src[] = { offset }.
 load("scratch", [1], [ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
@@ -775,9 +784,9 @@ def store(name, srcs, indices=[], flags=[]):
     intrinsic("store_" + name, [0] + srcs, indices=indices, flags=flags)
 
 # src[] = { value, offset }.
-store("output", [1], [BASE, WRMASK, COMPONENT, TYPE])
+store("output", [1], [BASE, WRMASK, COMPONENT, TYPE, IO_SEMANTICS])
 # src[] = { value, vertex, offset }.
-store("per_vertex_output", [1, 1], [BASE, WRMASK, COMPONENT])
+store("per_vertex_output", [1, 1], [BASE, WRMASK, COMPONENT, IO_SEMANTICS])
 # src[] = { value, block_index, offset }
 store("ssbo", [-1, 1], [WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 # src[] = { value, offset }.
@@ -813,12 +822,6 @@ intrinsic("ssbo_atomic_xor_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1, i
 intrinsic("ssbo_atomic_exchange_ir3",   src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
 intrinsic("ssbo_atomic_comp_swap_ir3",  src_comp=[1, 1, 1, 1, 1], dest_comp=1, indices=[ACCESS])
 
-# IR3-specific instruction for UBO loads using the ldc instruction. The second
-# source is the indirect offset, in units of four dwords. The base is a
-# component offset, in dword units.
-intrinsic("load_ubo_ir3", src_comp=[1, 1], bit_sizes=[32], dest_comp=0, indices=[BASE],
-          flags=[CAN_REORDER, CAN_ELIMINATE])
-
 # System values for freedreno geometry shaders.
 system_value("vs_primitive_stride_ir3", 1)
 system_value("vs_vertex_stride_ir3", 1)