Merge pull request #1566 from KhronosGroup/subgroup-table-fix

GLSL: Fix KHR subgroup extension table for subgroups.

7 files changed, 122 insertions, 23 deletions

diff --git a/reference/shaders-hlsl-no-opt/frag/nonuniform-constructor.sm51.fxconly.frag b/reference/shaders-hlsl-no-opt/frag/nonuniform-constructor.sm51.nonuniformresource.frag
index ca9a116f..ca9a116f 100644
--- a/reference/shaders-hlsl-no-opt/frag/nonuniform-constructor.sm51.fxconly.frag
+++ b/reference/shaders-hlsl-no-opt/frag/nonuniform-constructor.sm51.nonuniformresource.frag
diff --git a/reference/shaders-no-opt/asm/frag/subgroup-ballot-only.vk.asm.frag b/reference/shaders-no-opt/asm/frag/subgroup-ballot-only.vk.asm.frag
new file mode 100644
index 00000000..8a918c03
--- /dev/null
+++ b/reference/shaders-no-opt/asm/frag/subgroup-ballot-only.vk.asm.frag
@@ -0,0 +1,32 @@
+#version 450
+
+#if defined(GL_KHR_shader_subgroup_ballot)
+#extension GL_KHR_shader_subgroup_ballot : require
+#elif defined(GL_NV_shader_thread_group)
+#extension GL_NV_shader_thread_group : require
+#elif defined(GL_ARB_shader_ballot) && defined(GL_ARB_shader_int64)
+#extension GL_ARB_shader_int64 : enable
+#extension GL_ARB_shader_ballot : require
+#else
+#error No extensions available to emulate requested subgroup feature.
+#endif
+
+layout(location = 0) flat in uint INDEX;
+layout(location = 0) out uvec4 SV_Target;
+
+#if defined(GL_KHR_shader_subgroup_ballot)
+#elif defined(GL_NV_shader_thread_group)
+uvec4 subgroupBallot(bool v) { return uvec4(ballotThreadNV(v), 0u, 0u, 0u); }
+#elif defined(GL_ARB_shader_ballot)
+uvec4 subgroupBallot(bool v) { return uvec4(unpackUint2x32(ballotARB(v)), 0u, 0u); }
+#endif
+
+void main()
+{
+    uvec4 _21 = subgroupBallot(INDEX < 100u);
+    SV_Target.x = _21.x;
+    SV_Target.y = _21.y;
+    SV_Target.z = _21.z;
+    SV_Target.w = _21.w;
+}
+
diff --git a/reference/shaders-no-opt/asm/frag/subgroup-ballot-only.vk.asm.frag.vk b/reference/shaders-no-opt/asm/frag/subgroup-ballot-only.vk.asm.frag.vk
new file mode 100644
index 00000000..ed5933f3
--- /dev/null
+++ b/reference/shaders-no-opt/asm/frag/subgroup-ballot-only.vk.asm.frag.vk
@@ -0,0 +1,15 @@
+#version 450
+#extension GL_KHR_shader_subgroup_ballot : require
+
+layout(location = 0) flat in uint INDEX;
+layout(location = 0) out uvec4 SV_Target;
+
+void main()
+{
+    uvec4 _21 = subgroupBallot(INDEX < 100u);
+    SV_Target.x = _21.x;
+    SV_Target.y = _21.y;
+    SV_Target.z = _21.z;
+    SV_Target.w = _21.w;
+}
+
diff --git a/shaders-hlsl-no-opt/frag/nonuniform-constructor.sm51.fxconly.frag b/shaders-hlsl-no-opt/frag/nonuniform-constructor.sm51.nonuniformresource.frag
index 452aa953..452aa953 100644
--- a/shaders-hlsl-no-opt/frag/nonuniform-constructor.sm51.fxconly.frag
+++ b/shaders-hlsl-no-opt/frag/nonuniform-constructor.sm51.nonuniformresource.frag
diff --git a/shaders-no-opt/asm/frag/subgroup-ballot-only.vk.asm.frag b/shaders-no-opt/asm/frag/subgroup-ballot-only.vk.asm.frag
new file mode 100644
index 00000000..39f4d066
--- /dev/null
+++ b/shaders-no-opt/asm/frag/subgroup-ballot-only.vk.asm.frag
@@ -0,0 +1,52 @@
+; SPIR-V
+; Version: 1.3
+; Generator: Unknown(30017); 21022
+; Bound: 31
+; Schema: 0
+               OpCapability Shader
+               OpCapability GroupNonUniformBallot
+               OpMemoryModel Logical GLSL450
+               OpEntryPoint Fragment %main "main" %INDEX %SV_Target
+               OpExecutionMode %main OriginUpperLeft
+               OpName %main "main"
+               OpName %INDEX "INDEX"
+               OpName %SV_Target "SV_Target"
+               OpDecorate %INDEX Flat
+               OpDecorate %INDEX Location 0
+               OpDecorate %SV_Target Location 0
+       %void = OpTypeVoid
+          %2 = OpTypeFunction %void
+       %uint = OpTypeInt 32 0
+%_ptr_Input_uint = OpTypePointer Input %uint
+      %INDEX = OpVariable %_ptr_Input_uint Input
+     %v4uint = OpTypeVector %uint 4
+%_ptr_Output_v4uint = OpTypePointer Output %v4uint
+  %SV_Target = OpVariable %_ptr_Output_v4uint Output
+       %bool = OpTypeBool
+   %uint_100 = OpConstant %uint 100
+     %uint_3 = OpConstant %uint 3
+%_ptr_Output_uint = OpTypePointer Output %uint
+     %uint_0 = OpConstant %uint 0
+     %uint_1 = OpConstant %uint 1
+     %uint_2 = OpConstant %uint 2
+       %main = OpFunction %void None %2
+          %4 = OpLabel
+               OpBranch %29
+         %29 = OpLabel
+         %11 = OpLoad %uint %INDEX
+         %13 = OpULessThan %bool %11 %uint_100
+         %15 = OpGroupNonUniformBallot %v4uint %uint_3 %13
+         %17 = OpCompositeExtract %uint %15 0
+         %18 = OpCompositeExtract %uint %15 1
+         %19 = OpCompositeExtract %uint %15 2
+         %20 = OpCompositeExtract %uint %15 3
+         %22 = OpAccessChain %_ptr_Output_uint %SV_Target %uint_0
+               OpStore %22 %17
+         %24 = OpAccessChain %_ptr_Output_uint %SV_Target %uint_1
+               OpStore %24 %18
+         %26 = OpAccessChain %_ptr_Output_uint %SV_Target %uint_2
+               OpStore %26 %19
+         %28 = OpAccessChain %_ptr_Output_uint %SV_Target %uint_3
+               OpStore %28 %20
+               OpReturn
+               OpFunctionEnd
diff --git a/spirv_glsl.cpp b/spirv_glsl.cpp
index 12d72dc6..ca68e87f 100644
--- a/spirv_glsl.cpp
+++ b/spirv_glsl.cpp
@@ -3569,9 +3569,9 @@ void CompilerGLSL::emit_extension_workarounds(spv::ExecutionModel model)
 			statement("");
 		}
 
-		if (shader_subgroup_supporter.is_feature_requested(Supp::SubgroupBrodcast_First))
+		if (shader_subgroup_supporter.is_feature_requested(Supp::SubgroupBroadcast_First))
 		{
-			auto exts = Supp::get_candidates_for_feature(Supp::SubgroupBrodcast_First, result);
+			auto exts = Supp::get_candidates_for_feature(Supp::SubgroupBroadcast_First, result);
 
 			for (auto &e : exts)
 			{
@@ -7380,7 +7380,7 @@ void CompilerGLSL::emit_subgroup_op(const Instruction &i)
 
 	case OpGroupNonUniformBroadcast:
 	case OpGroupNonUniformBroadcastFirst:
-		request_subgroup_feature(ShaderSubgroupSupportHelper::SubgroupBrodcast_First);
+		request_subgroup_feature(ShaderSubgroupSupportHelper::SubgroupBroadcast_First);
 		break;
 
 	case OpGroupNonUniformShuffle:
@@ -15150,7 +15150,7 @@ CompilerGLSL::ShaderSubgroupSupportHelper::FeatureVector CompilerGLSL::ShaderSub
 	switch (feature)
 	{
 	case SubgroupAllEqualT:
-		return { SubgroupBrodcast_First, SubgroupAll_Any_AllEqualBool };
+		return { SubgroupBroadcast_First, SubgroupAll_Any_AllEqualBool };
 	case SubgroupElect:
 		return { SubgroupBallotFindLSB_MSB, SubgroupBallot, SubgroupInvocationID };
 	case SubgroupInverseBallot_InclBitCount_ExclBitCout:
@@ -15185,8 +15185,8 @@ CompilerGLSL::ShaderSubgroupSupportHelper::Candidate CompilerGLSL::ShaderSubgrou
 	static const Candidate extensions[FeatureCount] = {
 		KHR_shader_subgroup_ballot, KHR_shader_subgroup_basic,  KHR_shader_subgroup_basic,  KHR_shader_subgroup_basic,
 		KHR_shader_subgroup_basic,  KHR_shader_subgroup_ballot, KHR_shader_subgroup_ballot, KHR_shader_subgroup_vote,
-		KHR_shader_subgroup_vote,   KHR_shader_subgroup_basic,  KHR_shader_subgroup_ballot, KHR_shader_subgroup_basic,
-		KHR_shader_subgroup_basic,  KHR_shader_subgroup_ballot, KHR_shader_subgroup_ballot, KHR_shader_subgroup_ballot
+		KHR_shader_subgroup_vote,   KHR_shader_subgroup_basic,  KHR_shader_subgroup_basic, KHR_shader_subgroup_basic,
+		KHR_shader_subgroup_ballot, KHR_shader_subgroup_ballot, KHR_shader_subgroup_ballot, KHR_shader_subgroup_ballot
 	};
 
 	return extensions[feature];
@@ -15260,7 +15260,7 @@ CompilerGLSL::ShaderSubgroupSupportHelper::CandidateVector CompilerGLSL::ShaderS
 		return { KHR_shader_subgroup_basic, NV_shader_thread_group };
 	case NumSubgroups:
 		return { KHR_shader_subgroup_basic, NV_shader_thread_group };
-	case SubgroupBrodcast_First:
+	case SubgroupBroadcast_First:
 		return { KHR_shader_subgroup_ballot, NV_shader_thread_shuffle, ARB_shader_ballot };
 	case SubgroupBallotFindLSB_MSB:
 		return { KHR_shader_subgroup_ballot, NV_shader_thread_group };
diff --git a/spirv_glsl.hpp b/spirv_glsl.hpp
index add549aa..c382db67 100644
--- a/spirv_glsl.hpp
+++ b/spirv_glsl.hpp
@@ -275,22 +275,22 @@ protected:
 
 		enum Feature
 		{
-			SubgroupMask,
-			SubgroupSize,
-			SubgroupInvocationID,
-			SubgroupID,
-			NumSubgroups,
-			SubgroupBrodcast_First,
-			SubgroupBallotFindLSB_MSB,
-			SubgroupAll_Any_AllEqualBool,
-			SubgroupAllEqualT,
-			SubgroupElect,
-			SubgroupBarrier,
-			SubgroupMemBarrier,
-			SubgroupBallot,
-			SubgroupInverseBallot_InclBitCount_ExclBitCout,
-			SubgroupBallotBitExtract,
-			SubgroupBallotBitCount,
+			SubgroupMask = 0,
+			SubgroupSize = 1,
+			SubgroupInvocationID = 2,
+			SubgroupID = 3,
+			NumSubgroups = 4,
+			SubgroupBroadcast_First = 5,
+			SubgroupBallotFindLSB_MSB = 6,
+			SubgroupAll_Any_AllEqualBool = 7,
+			SubgroupAllEqualT = 8,
+			SubgroupElect = 9,
+			SubgroupBarrier = 10,
+			SubgroupMemBarrier = 11,
+			SubgroupBallot = 12,
+			SubgroupInverseBallot_InclBitCount_ExclBitCout = 13,
+			SubgroupBallotBitExtract = 14,
+			SubgroupBallotBitCount = 15,
 
 			FeatureCount
 		};