HLSL: Handle case where PerVertex block is not used.

3 files changed, 239 insertions, 15 deletions

diff --git a/reference/shaders-hlsl-no-opt/asm/mesh/mesh-shader-plain-builtin-outputs.spv14.asm.vk.nocompat.mesh b/reference/shaders-hlsl-no-opt/asm/mesh/mesh-shader-plain-builtin-outputs.spv14.asm.vk.nocompat.mesh
new file mode 100644
index 00000000..2f4e548d
--- /dev/null
+++ b/reference/shaders-hlsl-no-opt/asm/mesh/mesh-shader-plain-builtin-outputs.spv14.asm.vk.nocompat.mesh
@@ -0,0 +1,63 @@
+struct _12
+{
+    float _m0;
+};
+
+static uint gl_LocalInvocationIndex;
+struct SPIRV_Cross_Input
+{
+    uint gl_LocalInvocationIndex : SV_GroupIndex;
+};
+
+struct gl_MeshPerVertexEXT
+{
+    float4 B : TEXCOORD1;
+    float4 gl_Position : SV_Position;
+};
+
+struct gl_MeshPerPrimitiveEXT
+{
+    float4 C : TEXCOORD3;
+    uint gl_PrimitiveID : SV_PrimitiveID;
+    uint gl_Layer : SV_RenderTargetArrayIndex;
+    bool gl_CullPrimitiveEXT : SV_CullPrimitive;
+};
+
+groupshared float _9[64];
+
+void mesh_main(inout gl_MeshPerVertexEXT gl_MeshVerticesEXT[24], _12 _11, inout uint3 gl_PrimitiveTriangleIndicesEXT[8], inout gl_MeshPerPrimitiveEXT gl_MeshPrimitivesEXT[8])
+{
+    _9[gl_LocalInvocationIndex] = float(gl_LocalInvocationIndex);
+    GroupMemoryBarrierWithGroupSync();
+    SetMeshOutputCounts(24u, 8u);
+    gl_Position[gl_LocalInvocationIndex].x = _9[gl_LocalInvocationIndex];
+    gl_Position[gl_LocalInvocationIndex].y = _9[gl_LocalInvocationIndex];
+    gl_Position[gl_LocalInvocationIndex].z = _9[gl_LocalInvocationIndex];
+    gl_Position[gl_LocalInvocationIndex].w = _9[gl_LocalInvocationIndex];
+    float _63 = _11._m0 + _9[gl_LocalInvocationIndex ^ 1u];
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].B.x = _63;
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].B.y = _63;
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].B.z = _63;
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].B.w = _63;
+    if (gl_LocalInvocationIndex < 8u)
+    {
+        uint _71 = gl_LocalInvocationIndex * 3u;
+        gl_PrimitiveTriangleIndicesEXT[gl_LocalInvocationIndex] = uint3(_71, _71 + 1u, _71 + 2u);
+        gl_CullPrimitiveEXT[gl_LocalInvocationIndex] = (gl_LocalInvocationIndex & 1u) != 0u;
+        gl_PrimitiveID[gl_LocalInvocationIndex] = int(gl_LocalInvocationIndex);
+        gl_Layer[gl_LocalInvocationIndex] = int(gl_LocalInvocationIndex);
+        uint _81 = gl_LocalInvocationIndex ^ 2u;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].C.x = _9[_81];
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].C.y = _9[_81];
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].C.z = _9[_81];
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].C.w = _9[_81];
+    }
+}
+
+[outputtopology("triangle")]
+[numthreads(2, 3, 4)]
+void main(SPIRV_Cross_Input stage_input, out vertices gl_MeshPerVertexEXT gl_MeshVerticesEXT[24], in payload _12 _11, out indices uint3 gl_PrimitiveTriangleIndicesEXT[8], out primitives gl_MeshPerPrimitiveEXT gl_MeshPrimitivesEXT[8])
+{
+    gl_LocalInvocationIndex = stage_input.gl_LocalInvocationIndex;
+    mesh_main(gl_MeshVerticesEXT, _11, gl_PrimitiveTriangleIndicesEXT, gl_MeshPrimitivesEXT);
+}
diff --git a/shaders-hlsl-no-opt/asm/mesh/mesh-shader-plain-builtin-outputs.spv14.asm.vk.nocompat.mesh b/shaders-hlsl-no-opt/asm/mesh/mesh-shader-plain-builtin-outputs.spv14.asm.vk.nocompat.mesh
new file mode 100644
index 00000000..7b38001d
--- /dev/null
+++ b/shaders-hlsl-no-opt/asm/mesh/mesh-shader-plain-builtin-outputs.spv14.asm.vk.nocompat.mesh
@@ -0,0 +1,150 @@
+; 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 %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/spirv_hlsl.cpp b/spirv_hlsl.cpp
index fe3ed058..86dcbd1c 100644
--- a/spirv_hlsl.cpp
+++ b/spirv_hlsl.cpp
@@ -2309,6 +2309,7 @@ void CompilerHLSL::analyze_meshlet_writes()
 	uint32_t id_per_vertex = 0;
 	uint32_t id_per_primitive = 0;
 	bool need_per_primitive = false;
+	bool need_per_vertex = false;
 
 	ir.for_each_typed_id<SPIRVariable>([&](uint32_t, SPIRVariable &var) {
 		auto &type = this->get<SPIRType>(var.basetype);
@@ -2331,13 +2332,16 @@ void CompilerHLSL::analyze_meshlet_writes()
 
 			if (flags.get(DecorationPerPrimitiveEXT))
 				need_per_primitive = true;
+			else
+				need_per_vertex = true;
 		}
 	});
 
 	// If we have per-primitive outputs, and no per-primitive builtins,
-	// empty version of gl_MeshPerPrimitiveEXT will be emitted
-	if (id_per_primitive == 0 && need_per_primitive)
-	{
+	// empty version of gl_MeshPerPrimitiveEXT will be emitted.
+	// If we don't use block IO for vertex output, we'll also need to synthesize the PerVertex block.
+
+	const auto generate_block = [&](const char *block_name, const char *instance_name, bool per_primitive) -> uint32_t {
 		auto &execution = get_entry_point();
 
 		uint32_t op_type = ir.increase_bound_by(4);
@@ -2345,32 +2349,39 @@ void CompilerHLSL::analyze_meshlet_writes()
 		uint32_t op_ptr = op_type + 2;
 		uint32_t op_var = op_type + 3;
 
-		auto& type = set<SPIRType>(op_type);
+		auto &type = set<SPIRType>(op_type);
 		type.basetype = SPIRType::Struct;
-		set_name(op_type, "gl_MeshPerPrimitiveEXT");
+		set_name(op_type, block_name);
 		set_decoration(op_type, DecorationBlock);
-		set_decoration(op_type, DecorationPerPrimitiveEXT);
+		if (per_primitive)
+			set_decoration(op_type, DecorationPerPrimitiveEXT);
 
-		auto& arr = set<SPIRType>(op_arr, type);
+		auto &arr = set<SPIRType>(op_arr, type);
 		arr.parent_type = type.self;
-		arr.array.push_back(execution.output_primitives);
+		arr.array.push_back(per_primitive ? execution.output_primitives : execution.output_vertices);
 		arr.array_size_literal.push_back(true);
 
-		auto& ptr = set<SPIRType>(op_ptr, arr);
+		auto &ptr = set<SPIRType>(op_ptr, arr);
 		ptr.parent_type = arr.self;
 		ptr.pointer = true;
 		ptr.pointer_depth++;
 		ptr.storage = StorageClassOutput;
 		set_decoration(op_ptr, DecorationBlock);
-		set_name(op_ptr, "gl_MeshPerPrimitiveEXT");
+		set_name(op_ptr, block_name);
 
-		auto& var = set<SPIRVariable>(op_var, op_ptr, StorageClassOutput);
-		set_decoration(op_var, DecorationPerPrimitiveEXT);
-		set_name(op_var, "gl_MeshPrimitivesEXT");
+		auto &var = set<SPIRVariable>(op_var, op_ptr, StorageClassOutput);
+		if (per_primitive)
+			set_decoration(op_var, DecorationPerPrimitiveEXT);
+		set_name(op_var, instance_name);
 		execution.interface_variables.push_back(var.self);
 
-		id_per_primitive = op_var;
-	}
+		return op_var;
+	};
+
+	if (id_per_vertex == 0 && need_per_vertex)
+		id_per_vertex = generate_block("gl_MeshPerVertexEXT", "gl_MeshVerticesEXT", false);
+	if (id_per_primitive == 0 && need_per_primitive)
+		id_per_primitive = generate_block("gl_MeshPerPrimitiveEXT", "gl_MeshPrimitivesEXT", true);
 
 	unordered_set<uint32_t> processed_func_ids;
 	analyze_meshlet_writes(ir.default_entry_point, id_per_vertex, id_per_primitive, processed_func_ids);