Merge pull request #684 from cdavis5e/msl-builtin-vector-cast

MSL: Cast uses of builtin vectors to their declared SPIR-V type.
2018-09-11 19:59:58 +02:00 · 2018-09-11 19:59:58 +02:00 · 403011e973
--- a/reference/opt/shaders-msl/asm/comp/buffer-write-relative-addr.asm.comp
+++ b/reference/opt/shaders-msl/asm/comp/buffer-write-relative-addr.asm.comp
@ -18,8 +18,8 @@ uint2 spvTexelBufferCoord(uint tc)

 kernel void main0(constant cb5_struct& cb0_5 [[buffer(1)]], texture2d<uint, access::write> u0 [[texture(0)]], uint3 gl_LocalInvocationID [[thread_position_in_threadgroup]])
 {
-    uint _44 = as_type<uint>(as_type<float>(gl_LocalInvocationID.x << 4)) >> 2u;
-    uint4 _51 = as_type<uint4>(cb0_5._m0[uint(as_type<int>(as_type<float>(gl_LocalInvocationID.x))) + 1u]);
+    uint _44 = as_type<uint>(as_type<float>(int(gl_LocalInvocationID.x) << 4)) >> 2u;
+    uint4 _51 = as_type<uint4>(cb0_5._m0[uint(as_type<int>(as_type<float>(int(gl_LocalInvocationID.x)))) + 1u]);
    u0.write(_51.xxxx, spvTexelBufferCoord(_44));
    u0.write(_51.yyyy, spvTexelBufferCoord((_44 + 1u)));
    u0.write(_51.zzzz, spvTexelBufferCoord((_44 + 2u)));
--- a/reference/opt/shaders-msl/asm/comp/vector-builtin-type-cast-func.asm.comp
+++ b/reference/opt/shaders-msl/asm/comp/vector-builtin-type-cast-func.asm.comp
@ -0,0 +1,28 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+constant uint3 gl_WorkGroupSize = uint3(16u, 16u, 1u);
+
+struct cb1_struct
+{
+    float4 _m0[1];
+};
+
+kernel void main0(constant cb1_struct& cb0_1 [[buffer(0)]], texture2d<float, access::write> u0 [[texture(1)]], uint3 gl_LocalInvocationID [[thread_position_in_threadgroup]])
+{
+    int2 _46 = int2(u0.get_width(), u0.get_height()) >> int2(uint2(4u));
+    int _99;
+    _99 = 0;
+    for (; _99 < _46.y; _99++)
+    {
+        for (int _100 = 0; _100 < _46.x; )
+        {
+            u0.write(cb0_1._m0[0].xxxx, uint2(((_46 * int3(gl_LocalInvocationID).xy) + int2(_99, _100))));
+            _100++;
+            continue;
+        }
+    }
+}
+
--- a/reference/opt/shaders-msl/asm/comp/vector-builtin-type-cast.asm.comp
+++ b/reference/opt/shaders-msl/asm/comp/vector-builtin-type-cast.asm.comp
@ -0,0 +1,28 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+constant uint3 gl_WorkGroupSize = uint3(16u, 16u, 1u);
+
+struct cb1_struct
+{
+    float4 _m0[1];
+};
+
+kernel void main0(constant cb1_struct& cb0_1 [[buffer(0)]], texture2d<float, access::write> u0 [[texture(1)]], uint3 gl_LocalInvocationID [[thread_position_in_threadgroup]])
+{
+    int2 _40 = int2(u0.get_width(), u0.get_height()) >> int2(uint2(4u));
+    int _82;
+    _82 = 0;
+    for (; _82 < _40.y; _82++)
+    {
+        for (int _83 = 0; _83 < _40.x; )
+        {
+            u0.write(cb0_1._m0[0].xxxx, uint2(((_40 * int3(gl_LocalInvocationID).xy) + int2(_82, _83))));
+            _83++;
+            continue;
+        }
+    }
+}
+
--- a/reference/shaders-msl/asm/comp/buffer-write-relative-addr.asm.comp
+++ b/reference/shaders-msl/asm/comp/buffer-write-relative-addr.asm.comp
@ -19,8 +19,8 @@ uint2 spvTexelBufferCoord(uint tc)
 kernel void main0(constant cb5_struct& cb0_5 [[buffer(1)]], texture2d<uint, access::write> u0 [[texture(0)]], uint3 gl_LocalInvocationID [[thread_position_in_threadgroup]])
 {
    float4 r0;
-    r0.x = as_type<float>(gl_LocalInvocationID.x << 4);
-    r0.y = as_type<float>(gl_LocalInvocationID.x);
+    r0.x = as_type<float>(int(gl_LocalInvocationID.x) << 4);
+    r0.y = as_type<float>(int(gl_LocalInvocationID.x));
    uint _44 = as_type<uint>(r0.x) >> 2u;
    uint4 _51 = as_type<uint4>(cb0_5._m0[uint(as_type<int>(r0.y)) + 1u]);
    u0.write(_51.xxxx, spvTexelBufferCoord(_44));
--- a/reference/shaders-msl/asm/comp/vector-builtin-type-cast-func.asm.comp
+++ b/reference/shaders-msl/asm/comp/vector-builtin-type-cast-func.asm.comp
@ -0,0 +1,33 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+constant uint3 gl_WorkGroupSize = uint3(16u, 16u, 1u);
+
+struct cb1_struct
+{
+    float4 _m0[1];
+};
+
+int2 get_texcoord(thread const int2& base, thread const int2& index, thread uint3& gl_LocalInvocationID)
+{
+    return (base * int3(gl_LocalInvocationID).xy) + index;
+}
+
+kernel void main0(constant cb1_struct& cb0_1 [[buffer(0)]], texture2d<float, access::write> u0 [[texture(1)]], uint3 gl_LocalInvocationID [[thread_position_in_threadgroup]])
+{
+    int2 r0 = int2(int2(u0.get_width(), u0.get_height()) >> int2(uint2(4u)));
+    for (int i = 0; i < r0.y; i++)
+    {
+        for (int j = 0; j < r0.x; j++)
+        {
+            int2 param = r0;
+            int2 param_1 = int2(i, j);
+            u0.write(cb0_1._m0[0].xxxx, uint2(get_texcoord(param, param_1, gl_LocalInvocationID)));
+        }
+    }
+}
+
--- a/reference/shaders-msl/asm/comp/vector-builtin-type-cast.asm.comp
+++ b/reference/shaders-msl/asm/comp/vector-builtin-type-cast.asm.comp
@ -0,0 +1,24 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+constant uint3 gl_WorkGroupSize = uint3(16u, 16u, 1u);
+
+struct cb1_struct
+{
+    float4 _m0[1];
+};
+
+kernel void main0(constant cb1_struct& cb0_1 [[buffer(0)]], texture2d<float, access::write> u0 [[texture(1)]], uint3 gl_LocalInvocationID [[thread_position_in_threadgroup]])
+{
+    int2 r0 = int2(int2(u0.get_width(), u0.get_height()) >> int2(uint2(4u)));
+    for (int i = 0; i < r0.y; i++)
+    {
+        for (int j = 0; j < r0.x; j++)
+        {
+            u0.write(cb0_1._m0[0].xxxx, uint2(((r0 * int3(gl_LocalInvocationID).xy) + int2(i, j))));
+        }
+    }
+}
+
--- a/shaders-msl/asm/comp/vector-builtin-type-cast-func.asm.comp
+++ b/shaders-msl/asm/comp/vector-builtin-type-cast-func.asm.comp
@ -0,0 +1,147 @@
+; SPIR-V
+; Version: 1.3
+; Generator: Khronos Glslang Reference Front End; 6
+; Bound: 90
+; Schema: 0
+               OpCapability Shader
+               OpCapability ImageQuery
+               OpCapability StorageImageWriteWithoutFormat
+          %1 = OpExtInstImport "GLSL.std.450"
+               OpMemoryModel Logical GLSL450
+               OpEntryPoint GLCompute %main "main" %gl_LocalInvocationID
+               OpExecutionMode %main LocalSize 16 16 1
+               OpSource GLSL 450
+               OpName %main "main"
+               OpName %get_texcoord_vi2_vi2_ "get_texcoord(vi2;vi2;"
+               OpName %base "base"
+               OpName %index "index"
+               OpName %gl_LocalInvocationID "gl_LocalInvocationID"
+               OpName %r0 "r0"
+               OpName %u0 "u0"
+               OpName %i "i"
+               OpName %j "j"
+               OpName %param "param"
+               OpName %param_0 "param"
+               OpName %cb1_struct "cb1_struct"
+               OpMemberName %cb1_struct 0 "_m0"
+               OpName %cb0_1 "cb0_1"
+               OpDecorate %gl_LocalInvocationID BuiltIn LocalInvocationId
+               OpDecorate %u0 DescriptorSet 0
+               OpDecorate %u0 Binding 1
+               OpDecorate %u0 NonReadable
+               OpDecorate %_arr_v4float_uint_1 ArrayStride 16
+               OpMemberDecorate %cb1_struct 0 Offset 0
+               OpDecorate %cb1_struct Block
+               OpDecorate %cb0_1 DescriptorSet 0
+               OpDecorate %cb0_1 Binding 0
+               OpDecorate %gl_WorkGroupSize BuiltIn WorkgroupSize
+       %void = OpTypeVoid
+          %3 = OpTypeFunction %void
+        %int = OpTypeInt 32 1
+      %v2int = OpTypeVector %int 2
+%_ptr_Function_v2int = OpTypePointer Function %v2int
+          %9 = OpTypeFunction %v2int %_ptr_Function_v2int %_ptr_Function_v2int
+      %v3int = OpTypeVector %int 3
+%_ptr_Input_v3int = OpTypePointer Input %v3int
+%gl_LocalInvocationID = OpVariable %_ptr_Input_v3int Input
+       %uint = OpTypeInt 32 0
+     %v2uint = OpTypeVector %uint 2
+      %float = OpTypeFloat 32
+         %30 = OpTypeImage %float 2D 0 0 0 2 Unknown
+%_ptr_UniformConstant_30 = OpTypePointer UniformConstant %30
+         %u0 = OpVariable %_ptr_UniformConstant_30 UniformConstant
+     %uint_4 = OpConstant %uint 4
+%_ptr_Function_int = OpTypePointer Function %int
+      %int_0 = OpConstant %int 0
+     %uint_1 = OpConstant %uint 1
+       %bool = OpTypeBool
+     %uint_0 = OpConstant %uint 0
+    %v4float = OpTypeVector %float 4
+%_arr_v4float_uint_1 = OpTypeArray %v4float %uint_1
+ %cb1_struct = OpTypeStruct %_arr_v4float_uint_1
+%_ptr_Uniform_cb1_struct = OpTypePointer Uniform %cb1_struct
+      %cb0_1 = OpVariable %_ptr_Uniform_cb1_struct Uniform
+%_ptr_Uniform_v4float = OpTypePointer Uniform %v4float
+      %int_1 = OpConstant %int 1
+    %uint_16 = OpConstant %uint 16
+     %v3uint = OpTypeVector %uint 3
+%gl_WorkGroupSize = OpConstantComposite %v3uint %uint_16 %uint_16 %uint_1
+       %main = OpFunction %void None %3
+          %5 = OpLabel
+         %r0 = OpVariable %_ptr_Function_v2int Function
+          %i = OpVariable %_ptr_Function_int Function
+          %j = OpVariable %_ptr_Function_int Function
+      %param = OpVariable %_ptr_Function_v2int Function
+    %param_0 = OpVariable %_ptr_Function_v2int Function
+         %33 = OpLoad %30 %u0
+         %34 = OpImageQuerySize %v2int %33
+         %36 = OpCompositeConstruct %v2uint %uint_4 %uint_4
+         %37 = OpShiftRightArithmetic %v2int %34 %36
+         %38 = OpCompositeExtract %int %37 0
+         %39 = OpCompositeExtract %int %37 1
+         %40 = OpCompositeConstruct %v2int %38 %39
+               OpStore %r0 %40
+               OpStore %i %int_0
+               OpBranch %44
+         %44 = OpLabel
+               OpLoopMerge %46 %47 None
+               OpBranch %48
+         %48 = OpLabel
+         %49 = OpLoad %int %i
+         %51 = OpAccessChain %_ptr_Function_int %r0 %uint_1
+         %52 = OpLoad %int %51
+         %54 = OpSLessThan %bool %49 %52
+               OpBranchConditional %54 %45 %46
+         %45 = OpLabel
+               OpStore %j %int_0
+               OpBranch %56
+         %56 = OpLabel
+               OpLoopMerge %58 %59 None
+               OpBranch %60
+         %60 = OpLabel
+         %61 = OpLoad %int %j
+         %63 = OpAccessChain %_ptr_Function_int %r0 %uint_0
+         %64 = OpLoad %int %63
+         %65 = OpSLessThan %bool %61 %64
+               OpBranchConditional %65 %57 %58
+         %57 = OpLabel
+         %66 = OpLoad %30 %u0
+         %67 = OpLoad %int %i
+         %68 = OpLoad %int %j
+         %69 = OpCompositeConstruct %v2int %67 %68
+         %71 = OpLoad %v2int %r0
+               OpStore %param %71
+               OpStore %param_0 %69
+         %73 = OpFunctionCall %v2int %get_texcoord_vi2_vi2_ %param %param_0
+         %80 = OpAccessChain %_ptr_Uniform_v4float %cb0_1 %int_0 %int_0
+         %81 = OpLoad %v4float %80
+         %82 = OpVectorShuffle %v4float %81 %81 0 0 0 0
+               OpImageWrite %66 %73 %82
+               OpBranch %59
+         %59 = OpLabel
+         %83 = OpLoad %int %j
+         %85 = OpIAdd %int %83 %int_1
+               OpStore %j %85
+               OpBranch %56
+         %58 = OpLabel
+               OpBranch %47
+         %47 = OpLabel
+         %86 = OpLoad %int %i
+         %87 = OpIAdd %int %86 %int_1
+               OpStore %i %87
+               OpBranch %44
+         %46 = OpLabel
+               OpReturn
+               OpFunctionEnd
+%get_texcoord_vi2_vi2_ = OpFunction %v2int None %9
+       %base = OpFunctionParameter %_ptr_Function_v2int
+      %index = OpFunctionParameter %_ptr_Function_v2int
+         %13 = OpLabel
+         %14 = OpLoad %v2int %base
+         %20 = OpLoad %v3int %gl_LocalInvocationID
+         %21 = OpVectorShuffle %v2int %20 %20 0 1
+         %23 = OpIMul %v2int %14 %21
+         %24 = OpLoad %v2int %index
+         %25 = OpIAdd %v2int %23 %24
+               OpReturnValue %25
+               OpFunctionEnd
--- a/shaders-msl/asm/comp/vector-builtin-type-cast.asm.comp
+++ b/shaders-msl/asm/comp/vector-builtin-type-cast.asm.comp
@ -0,0 +1,128 @@
+; SPIR-V
+; Version: 1.3
+; Generator: Khronos Glslang Reference Front End; 6
+; Bound: 78
+; Schema: 0
+               OpCapability Shader
+               OpCapability ImageQuery
+               OpCapability StorageImageWriteWithoutFormat
+          %1 = OpExtInstImport "GLSL.std.450"
+               OpMemoryModel Logical GLSL450
+               OpEntryPoint GLCompute %main "main" %gl_LocalInvocationID
+               OpExecutionMode %main LocalSize 16 16 1
+               OpSource GLSL 450
+               OpName %main "main"
+               OpName %r0 "r0"
+               OpName %u0 "u0"
+               OpName %i "i"
+               OpName %j "j"
+               OpName %gl_LocalInvocationID "gl_LocalInvocationID"
+               OpName %cb1_struct "cb1_struct"
+               OpMemberName %cb1_struct 0 "_m0"
+               OpName %cb0_1 "cb0_1"
+               OpDecorate %u0 DescriptorSet 0
+               OpDecorate %u0 Binding 1
+               OpDecorate %u0 NonReadable
+               OpDecorate %gl_LocalInvocationID BuiltIn LocalInvocationId
+               OpDecorate %_arr_v4float_uint_1 ArrayStride 16
+               OpMemberDecorate %cb1_struct 0 Offset 0
+               OpDecorate %cb1_struct Block
+               OpDecorate %cb0_1 DescriptorSet 0
+               OpDecorate %cb0_1 Binding 0
+               OpDecorate %gl_WorkGroupSize BuiltIn WorkgroupSize
+       %void = OpTypeVoid
+          %3 = OpTypeFunction %void
+        %int = OpTypeInt 32 1
+      %v2int = OpTypeVector %int 2
+%_ptr_Function_v2int = OpTypePointer Function %v2int
+      %float = OpTypeFloat 32
+         %11 = OpTypeImage %float 2D 0 0 0 2 Unknown
+%_ptr_UniformConstant_11 = OpTypePointer UniformConstant %11
+         %u0 = OpVariable %_ptr_UniformConstant_11 UniformConstant
+       %uint = OpTypeInt 32 0
+     %uint_4 = OpConstant %uint 4
+     %v2uint = OpTypeVector %uint 2
+%_ptr_Function_int = OpTypePointer Function %int
+      %int_0 = OpConstant %int 0
+     %uint_1 = OpConstant %uint 1
+       %bool = OpTypeBool
+     %uint_0 = OpConstant %uint 0
+      %v3int = OpTypeVector %int 3
+%_ptr_Input_v3int = OpTypePointer Input %v3int
+%gl_LocalInvocationID = OpVariable %_ptr_Input_v3int Input
+    %v4float = OpTypeVector %float 4
+%_arr_v4float_uint_1 = OpTypeArray %v4float %uint_1
+ %cb1_struct = OpTypeStruct %_arr_v4float_uint_1
+%_ptr_Uniform_cb1_struct = OpTypePointer Uniform %cb1_struct
+      %cb0_1 = OpVariable %_ptr_Uniform_cb1_struct Uniform
+%_ptr_Uniform_v4float = OpTypePointer Uniform %v4float
+      %int_1 = OpConstant %int 1
+    %uint_16 = OpConstant %uint 16
+     %v3uint = OpTypeVector %uint 3
+%gl_WorkGroupSize = OpConstantComposite %v3uint %uint_16 %uint_16 %uint_1
+       %main = OpFunction %void None %3
+          %5 = OpLabel
+         %r0 = OpVariable %_ptr_Function_v2int Function
+          %i = OpVariable %_ptr_Function_int Function
+          %j = OpVariable %_ptr_Function_int Function
+         %14 = OpLoad %11 %u0
+         %15 = OpImageQuerySize %v2int %14
+         %19 = OpCompositeConstruct %v2uint %uint_4 %uint_4
+         %20 = OpShiftRightArithmetic %v2int %15 %19
+         %21 = OpCompositeExtract %int %20 0
+         %22 = OpCompositeExtract %int %20 1
+         %23 = OpCompositeConstruct %v2int %21 %22
+               OpStore %r0 %23
+               OpStore %i %int_0
+               OpBranch %27
+         %27 = OpLabel
+               OpLoopMerge %29 %30 None
+               OpBranch %31
+         %31 = OpLabel
+         %32 = OpLoad %int %i
+         %34 = OpAccessChain %_ptr_Function_int %r0 %uint_1
+         %35 = OpLoad %int %34
+         %37 = OpSLessThan %bool %32 %35
+               OpBranchConditional %37 %28 %29
+         %28 = OpLabel
+               OpStore %j %int_0
+               OpBranch %39
+         %39 = OpLabel
+               OpLoopMerge %41 %42 None
+               OpBranch %43
+         %43 = OpLabel
+         %44 = OpLoad %int %j
+         %46 = OpAccessChain %_ptr_Function_int %r0 %uint_0
+         %47 = OpLoad %int %46
+         %48 = OpSLessThan %bool %44 %47
+               OpBranchConditional %48 %40 %41
+         %40 = OpLabel
+         %49 = OpLoad %11 %u0
+         %50 = OpLoad %v2int %r0
+         %54 = OpLoad %v3int %gl_LocalInvocationID
+         %55 = OpVectorShuffle %v2int %54 %54 0 1
+         %57 = OpIMul %v2int %50 %55
+         %58 = OpLoad %int %i
+         %59 = OpLoad %int %j
+         %60 = OpCompositeConstruct %v2int %58 %59
+         %61 = OpIAdd %v2int %57 %60
+         %68 = OpAccessChain %_ptr_Uniform_v4float %cb0_1 %int_0 %int_0
+         %69 = OpLoad %v4float %68
+         %70 = OpVectorShuffle %v4float %69 %69 0 0 0 0
+               OpImageWrite %49 %61 %70
+               OpBranch %42
+         %42 = OpLabel
+         %71 = OpLoad %int %j
+         %73 = OpIAdd %int %71 %int_1
+               OpStore %j %73
+               OpBranch %39
+         %41 = OpLabel
+               OpBranch %30
+         %30 = OpLabel
+         %74 = OpLoad %int %i
+         %75 = OpIAdd %int %74 %int_1
+               OpStore %i %75
+               OpBranch %27
+         %29 = OpLabel
+               OpReturn
+               OpFunctionEnd
--- a/spirv_msl.cpp
+++ b/spirv_msl.cpp
@ -4545,11 +4545,38 @@ void CompilerMSL::remap_constexpr_sampler(uint32_t id, const MSLConstexprSampler
 	constexpr_samplers[id] = sampler;
 }

-// MSL always declares builtins with their SPIR-V type.
-void CompilerMSL::bitcast_from_builtin_load(uint32_t, std::string &, const SPIRType &)
+void CompilerMSL::bitcast_from_builtin_load(uint32_t source_id, std::string &expr, const SPIRType &expr_type)
 {
+	auto *var = maybe_get_backing_variable(source_id);
+	if (var)
+		source_id = var->self;
+
+	// Only interested in standalone builtin variables.
+	if (!has_decoration(source_id, DecorationBuiltIn))
+		return;
+
+	auto builtin = static_cast<BuiltIn>(get_decoration(source_id, DecorationBuiltIn));
+	auto expected_type = expr_type.basetype;
+	switch (builtin)
+	{
+	case BuiltInGlobalInvocationId:
+	case BuiltInLocalInvocationId:
+	case BuiltInWorkgroupId:
+	case BuiltInLocalInvocationIndex:
+	case BuiltInWorkgroupSize:
+	case BuiltInNumWorkgroups:
+		expected_type = SPIRType::UInt;
+		break;
+
+	default:
+		break;
+	}
+
+	if (expected_type != expr_type.basetype)
+		expr = bitcast_expression(expr_type, expected_type, expr);
 }

+// MSL always declares output builtins with the SPIR-V type.
 void CompilerMSL::bitcast_to_builtin_store(uint32_t, std::string &, const SPIRType &)
 {
 }