Merge pull request #2187 from KhronosGroup/mesh-invariant-position

GLSL: Emit inline invariant position for mesh shaders.
KhronosGroup · Aug 17, 2023 · 633dc30 · 633dc30
2 parents a3f0268 + 79d6393
commit 633dc30
Show file tree

Hide file tree

Showing 9 changed files with 253 additions and 13 deletions.
diff --git a/reference/opt/shaders/asm/vert/invariant-block.sso.asm.vert b/reference/opt/shaders/asm/vert/invariant-block.sso.asm.vert
@@ -2,14 +2,12 @@
 
 out gl_PerVertex
 {
-    vec4 gl_Position;
+    invariant vec4 gl_Position;
     float gl_PointSize;
     float gl_ClipDistance[1];
     float gl_CullDistance[1];
 };
 
-invariant gl_Position;
-
 void main()
 {
     gl_Position = vec4(1.0);

diff --git a/reference/opt/shaders/asm/vert/invariant.sso.asm.vert b/reference/opt/shaders/asm/vert/invariant.sso.asm.vert
@@ -2,11 +2,9 @@
 
 out gl_PerVertex
 {
-    vec4 gl_Position;
+    invariant vec4 gl_Position;
 };
 
-invariant gl_Position;
-
 void main()
 {
     gl_Position = vec4(1.0);

diff --git a/reference/shaders-no-opt/asm/mesh/invariant-position-dx-style.vk.nocompat.spv14.asm.mesh.vk b/reference/shaders-no-opt/asm/mesh/invariant-position-dx-style.vk.nocompat.spv14.asm.mesh.vk
@@ -0,0 +1,49 @@
+#version 450
+#extension GL_EXT_mesh_shader : require
+layout(local_size_x = 2, local_size_y = 3, local_size_z = 4) in;
+layout(max_vertices = 24, max_primitives = 8, triangles) out;
+
+out gl_MeshPerVertexEXT
+{
+    invariant vec4 gl_Position;
+} gl_MeshVerticesEXT[];
+
+struct _12
+{
+    float _m0;
+};
+
+layout(location = 1) out vec4 B[24];
+layout(location = 3) perprimitiveEXT out vec4 C[8];
+shared float _32[64];
+taskPayloadSharedEXT _12 _38;
+
+void main()
+{
+    _32[gl_LocalInvocationIndex] = float(gl_LocalInvocationIndex);
+    barrier();
+    SetMeshOutputsEXT(24u, 8u);
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_Position.x = _32[gl_LocalInvocationIndex];
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_Position.y = _32[gl_LocalInvocationIndex];
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_Position.z = _32[gl_LocalInvocationIndex];
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_Position.w = _32[gl_LocalInvocationIndex];
+    float _59 = _38._m0 + _32[gl_LocalInvocationIndex ^ 1u];
+    B[gl_LocalInvocationIndex].x = _59;
+    B[gl_LocalInvocationIndex].y = _59;
+    B[gl_LocalInvocationIndex].z = _59;
+    B[gl_LocalInvocationIndex].w = _59;
+    if (gl_LocalInvocationIndex < 8u)
+    {
+        uint _65 = gl_LocalInvocationIndex * 3u;
+        gl_PrimitiveTriangleIndicesEXT[gl_LocalInvocationIndex] = uvec3(_65, _65 + 1u, _65 + 2u);
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_CullPrimitiveEXT = (gl_LocalInvocationIndex & 1u) != 0u;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_PrimitiveID = int(gl_LocalInvocationIndex);
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_Layer = int(gl_LocalInvocationIndex);
+        uint _78 = gl_LocalInvocationIndex ^ 2u;
+        C[gl_LocalInvocationIndex].x = _32[_78];
+        C[gl_LocalInvocationIndex].y = _32[_78];
+        C[gl_LocalInvocationIndex].z = _32[_78];
+        C[gl_LocalInvocationIndex].w = _32[_78];
+    }
+}
+
diff --git a/reference/shaders-no-opt/mesh/invariant-position-mesh.spv14.nocompat.vk.mesh.vk b/reference/shaders-no-opt/mesh/invariant-position-mesh.spv14.nocompat.vk.mesh.vk
@@ -0,0 +1,19 @@
+#version 450
+#extension GL_EXT_mesh_shader : require
+layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+layout(max_vertices = 3, max_primitives = 1, triangles) out;
+
+out gl_MeshPerVertexEXT
+{
+    invariant vec4 gl_Position;
+} gl_MeshVerticesEXT[];
+
+void main()
+{
+    SetMeshOutputsEXT(3u, 1u);
+    gl_MeshVerticesEXT[0].gl_Position = vec4(1.0);
+    gl_MeshVerticesEXT[1].gl_Position = vec4(1.0);
+    gl_MeshVerticesEXT[2].gl_Position = vec4(1.0);
+    gl_PrimitiveTriangleIndicesEXT[0] = uvec3(0u, 1u, 2u);
+}
+
diff --git a/reference/shaders/asm/vert/invariant-block.sso.asm.vert b/reference/shaders/asm/vert/invariant-block.sso.asm.vert
@@ -2,14 +2,12 @@
 
 out gl_PerVertex
 {
-    vec4 gl_Position;
+    invariant vec4 gl_Position;
     float gl_PointSize;
     float gl_ClipDistance[1];
     float gl_CullDistance[1];
 };
 
-invariant gl_Position;
-
 void main()
 {
     gl_Position = vec4(1.0);

diff --git a/reference/shaders/asm/vert/invariant.sso.asm.vert b/reference/shaders/asm/vert/invariant.sso.asm.vert
@@ -2,11 +2,9 @@
 
 out gl_PerVertex
 {
-    vec4 gl_Position;
+    invariant vec4 gl_Position;
 };
 
-invariant gl_Position;
-
 vec4 _main()
 {
     return vec4(1.0);

diff --git a/shaders-no-opt/asm/mesh/invariant-position-dx-style.vk.nocompat.spv14.asm.mesh b/shaders-no-opt/asm/mesh/invariant-position-dx-style.vk.nocompat.spv14.asm.mesh
@@ -0,0 +1,151 @@
+; SPIR-V
+; Version: 1.4
+; Generator: Unknown(30017); 21022
+; Bound: 89
+; Schema: 0
+               OpCapability Shader
+               OpCapability Geometry
+               OpCapability ShaderViewportIndexLayerEXT
+               OpCapability MeshShadingEXT
+               OpExtension "SPV_EXT_mesh_shader"
+               OpExtension "SPV_EXT_shader_viewport_index_layer"
+               OpMemoryModel Logical GLSL450
+               OpEntryPoint MeshEXT %main "main" %SV_Position %B %SV_CullPrimitive %SV_RenderTargetArrayIndex %SV_PrimitiveID %C %indices %32 %gl_LocalInvocationIndex %38
+               OpExecutionMode %main OutputVertices 24
+               OpExecutionMode %main OutputPrimitivesNV 8
+               OpExecutionMode %main OutputTrianglesNV
+               OpExecutionMode %main LocalSize 2 3 4
+               OpName %main "main"
+               OpName %SV_Position "SV_Position"
+               OpName %B "B"
+               OpName %SV_CullPrimitive "SV_CullPrimitive"
+               OpName %SV_RenderTargetArrayIndex "SV_RenderTargetArrayIndex"
+               OpName %SV_PrimitiveID "SV_PrimitiveID"
+               OpName %C "C"
+               OpName %indices "indices"
+               OpName %_ ""
+               OpDecorate %SV_Position BuiltIn Position
+               OpDecorate %SV_Position Invariant
+               OpDecorate %B Location 1
+               OpDecorate %SV_CullPrimitive BuiltIn CullPrimitiveEXT
+               OpDecorate %SV_CullPrimitive PerPrimitiveNV
+               OpDecorate %SV_RenderTargetArrayIndex BuiltIn Layer
+               OpDecorate %SV_RenderTargetArrayIndex PerPrimitiveNV
+               OpDecorate %SV_PrimitiveID BuiltIn PrimitiveId
+               OpDecorate %SV_PrimitiveID PerPrimitiveNV
+               OpDecorate %C Location 3
+               OpDecorate %C PerPrimitiveNV
+               OpDecorate %indices BuiltIn PrimitiveTriangleIndicesEXT
+               OpDecorate %gl_LocalInvocationIndex BuiltIn LocalInvocationIndex
+       %void = OpTypeVoid
+          %2 = OpTypeFunction %void
+      %float = OpTypeFloat 32
+    %v4float = OpTypeVector %float 4
+       %uint = OpTypeInt 32 0
+    %uint_24 = OpConstant %uint 24
+%_arr_v4float_uint_24 = OpTypeArray %v4float %uint_24
+%_ptr_Output__arr_v4float_uint_24 = OpTypePointer Output %_arr_v4float_uint_24
+%SV_Position = OpVariable %_ptr_Output__arr_v4float_uint_24 Output
+          %B = OpVariable %_ptr_Output__arr_v4float_uint_24 Output
+       %bool = OpTypeBool
+     %uint_8 = OpConstant %uint 8
+%_arr_bool_uint_8 = OpTypeArray %bool %uint_8
+%_ptr_Output__arr_bool_uint_8 = OpTypePointer Output %_arr_bool_uint_8
+%SV_CullPrimitive = OpVariable %_ptr_Output__arr_bool_uint_8 Output
+%_arr_uint_uint_8 = OpTypeArray %uint %uint_8
+%_ptr_Output__arr_uint_uint_8 = OpTypePointer Output %_arr_uint_uint_8
+%SV_RenderTargetArrayIndex = OpVariable %_ptr_Output__arr_uint_uint_8 Output
+%SV_PrimitiveID = OpVariable %_ptr_Output__arr_uint_uint_8 Output
+%_arr_v4float_uint_8 = OpTypeArray %v4float %uint_8
+%_ptr_Output__arr_v4float_uint_8 = OpTypePointer Output %_arr_v4float_uint_8
+          %C = OpVariable %_ptr_Output__arr_v4float_uint_8 Output
+     %v3uint = OpTypeVector %uint 3
+%_arr_v3uint_uint_8 = OpTypeArray %v3uint %uint_8
+%_ptr_Output__arr_v3uint_uint_8 = OpTypePointer Output %_arr_v3uint_uint_8
+    %indices = OpVariable %_ptr_Output__arr_v3uint_uint_8 Output
+    %uint_64 = OpConstant %uint 64
+%_arr_float_uint_64 = OpTypeArray %float %uint_64
+%_ptr_Workgroup__arr_float_uint_64 = OpTypePointer Workgroup %_arr_float_uint_64
+         %32 = OpVariable %_ptr_Workgroup__arr_float_uint_64 Workgroup
+%_ptr_Input_uint = OpTypePointer Input %uint
+%gl_LocalInvocationIndex = OpVariable %_ptr_Input_uint Input
+          %_ = OpTypeStruct %float
+%_ptr_TaskPayloadWorkgroupEXT__ = OpTypePointer TaskPayloadWorkgroupEXT %_
+         %38 = OpVariable %_ptr_TaskPayloadWorkgroupEXT__ TaskPayloadWorkgroupEXT
+%_ptr_Workgroup_float = OpTypePointer Workgroup %float
+     %uint_2 = OpConstant %uint 2
+   %uint_264 = OpConstant %uint 264
+%_ptr_Output_float = OpTypePointer Output %float
+     %uint_0 = OpConstant %uint 0
+     %uint_1 = OpConstant %uint 1
+     %uint_3 = OpConstant %uint 3
+%_ptr_TaskPayloadWorkgroupEXT_float = OpTypePointer TaskPayloadWorkgroupEXT %float
+%_ptr_Output_v3uint = OpTypePointer Output %v3uint
+%_ptr_Output_bool = OpTypePointer Output %bool
+%_ptr_Output_uint = OpTypePointer Output %uint
+       %main = OpFunction %void None %2
+          %4 = OpLabel
+               OpBranch %85
+         %85 = OpLabel
+         %35 = OpLoad %uint %gl_LocalInvocationIndex
+         %39 = OpConvertUToF %float %35
+         %41 = OpAccessChain %_ptr_Workgroup_float %32 %35
+               OpStore %41 %39
+               OpControlBarrier %uint_2 %uint_2 %uint_264
+               OpSetMeshOutputsEXT %uint_24 %uint_8
+         %44 = OpLoad %float %41
+         %46 = OpAccessChain %_ptr_Output_float %SV_Position %35 %uint_0
+               OpStore %46 %44
+         %48 = OpAccessChain %_ptr_Output_float %SV_Position %35 %uint_1
+               OpStore %48 %44
+         %50 = OpAccessChain %_ptr_Output_float %SV_Position %35 %uint_2
+               OpStore %50 %44
+         %51 = OpAccessChain %_ptr_Output_float %SV_Position %35 %uint_3
+               OpStore %51 %44
+         %53 = OpBitwiseXor %uint %35 %uint_1
+         %54 = OpAccessChain %_ptr_Workgroup_float %32 %53
+         %55 = OpLoad %float %54
+         %57 = OpInBoundsAccessChain %_ptr_TaskPayloadWorkgroupEXT_float %38 %uint_0
+         %58 = OpLoad %float %57
+         %59 = OpFAdd %float %58 %55
+         %60 = OpAccessChain %_ptr_Output_float %B %35 %uint_0
+               OpStore %60 %59
+         %61 = OpAccessChain %_ptr_Output_float %B %35 %uint_1
+               OpStore %61 %59
+         %62 = OpAccessChain %_ptr_Output_float %B %35 %uint_2
+               OpStore %62 %59
+         %63 = OpAccessChain %_ptr_Output_float %B %35 %uint_3
+               OpStore %63 %59
+         %64 = OpULessThan %bool %35 %uint_8
+               OpSelectionMerge %87 None
+               OpBranchConditional %64 %86 %87
+         %86 = OpLabel
+         %65 = OpIMul %uint %35 %uint_3
+         %66 = OpIAdd %uint %65 %uint_1
+         %67 = OpIAdd %uint %65 %uint_2
+         %68 = OpCompositeConstruct %v3uint %65 %66 %67
+         %70 = OpAccessChain %_ptr_Output_v3uint %indices %35
+               OpStore %70 %68
+         %71 = OpBitwiseAnd %uint %35 %uint_1
+         %72 = OpINotEqual %bool %71 %uint_0
+         %74 = OpAccessChain %_ptr_Output_bool %SV_CullPrimitive %35
+               OpStore %74 %72
+         %76 = OpAccessChain %_ptr_Output_uint %SV_PrimitiveID %35
+               OpStore %76 %35
+         %77 = OpAccessChain %_ptr_Output_uint %SV_RenderTargetArrayIndex %35
+               OpStore %77 %35
+         %78 = OpBitwiseXor %uint %35 %uint_2
+         %79 = OpAccessChain %_ptr_Workgroup_float %32 %78
+         %80 = OpLoad %float %79
+         %81 = OpAccessChain %_ptr_Output_float %C %35 %uint_0
+               OpStore %81 %80
+         %82 = OpAccessChain %_ptr_Output_float %C %35 %uint_1
+               OpStore %82 %80
+         %83 = OpAccessChain %_ptr_Output_float %C %35 %uint_2
+               OpStore %83 %80
+         %84 = OpAccessChain %_ptr_Output_float %C %35 %uint_3
+               OpStore %84 %80
+               OpBranch %87
+         %87 = OpLabel
+               OpReturn
+               OpFunctionEnd
diff --git a/shaders-no-opt/mesh/invariant-position-mesh.spv14.nocompat.vk.mesh b/shaders-no-opt/mesh/invariant-position-mesh.spv14.nocompat.vk.mesh
@@ -0,0 +1,18 @@
+#version 450
+#extension GL_EXT_mesh_shader : require
+layout(max_vertices = 3, max_primitives = 1, triangles) out;
+layout(local_size_x = 1) in;
+
+out gl_MeshPerVertexEXT
+{
+	invariant vec4 gl_Position;
+} gl_MeshVerticesEXT[3];
+
+void main()
+{
+	SetMeshOutputsEXT(3, 1);
+	gl_MeshVerticesEXT[0].gl_Position = vec4(1.0);
+	gl_MeshVerticesEXT[1].gl_Position = vec4(1.0);
+	gl_MeshVerticesEXT[2].gl_Position = vec4(1.0);
+	gl_PrimitiveTriangleIndicesEXT[0] = uvec3(0, 1, 2);
+}
diff --git a/spirv_glsl.cpp b/spirv_glsl.cpp
@@ -3162,6 +3162,10 @@ bool CompilerGLSL::should_force_emit_builtin_block(StorageClass storage)
 		should_force = true;
 	}
 
+	// Either glslang bug or oversight, but global invariant position does not work in mesh shaders.
+	if (get_execution_model() == ExecutionModelMeshEXT && position_invariant)
+		should_force = true;
+
 	return should_force;
 }
 
@@ -3410,6 +3414,8 @@ void CompilerGLSL::emit_declared_builtin_block(StorageClass storage, ExecutionMo
 		auto itr = builtin_xfb_offsets.find(BuiltInPosition);
 		if (itr != end(builtin_xfb_offsets))
 			statement("layout(xfb_offset = ", itr->second, ") vec4 gl_Position;");
+		else if (position_invariant)
+			statement("invariant vec4 gl_Position;");
 		else
 			statement("vec4 gl_Position;");
 	}
@@ -3506,6 +3512,8 @@ void CompilerGLSL::emit_resources()
 		break;
 	}
 
+	bool global_invariant_position = position_invariant && (options.es || options.version >= 120);
+
 	// Emit custom gl_PerVertex for SSO compatibility.
 	if (options.separate_shader_objects && !options.es && execution.model != ExecutionModelFragment)
 	{
@@ -3516,11 +3524,13 @@ void CompilerGLSL::emit_resources()
 		case ExecutionModelTessellationEvaluation:
 			emit_declared_builtin_block(StorageClassInput, execution.model);
 			emit_declared_builtin_block(StorageClassOutput, execution.model);
+			global_invariant_position = false;
 			break;
 
 		case ExecutionModelVertex:
 		case ExecutionModelMeshEXT:
 			emit_declared_builtin_block(StorageClassOutput, execution.model);
+			global_invariant_position = false;
 			break;
 
 		default:
@@ -3530,6 +3540,7 @@ void CompilerGLSL::emit_resources()
 	else if (should_force_emit_builtin_block(StorageClassOutput))
 	{
 		emit_declared_builtin_block(StorageClassOutput, execution.model);
+		global_invariant_position = false;
 	}
 	else if (execution.geometry_passthrough)
 	{
@@ -3550,7 +3561,7 @@ void CompilerGLSL::emit_resources()
 			statement("");
 	}
 
-	if (position_invariant && (options.es || options.version >= 120))
+	if (global_invariant_position)
 	{
 		statement("invariant gl_Position;");
 		statement("");