Skip to content

G
glslang
Commit 2abe9a4d by Jeff Bolz
Options

Implement GL_NV_shader_subgroup_partitioned

parent 6b1ede41
Showing with 211 additions and 9 deletions
SPIRV/GLSL.ext.NV.h
......@@ -51,4 +51,7 @@ const char* const E_SPV_NV_stereo_view_rendering = "SPV_NV_stereo_view_rendering
//SPV_NVX_multiview_per_view_attributes
const char* const E_SPV_NVX_multiview_per_view_attributes = "SPV_NVX_multiview_per_view_attributes";
//SPV_NV_shader_subgroup_partitioned
const char* const E_SPV_NV_shader_subgroup_partitioned = "SPV_NV_shader_subgroup_partitioned";
#endif // #ifndef GLSLextNV_H
\ No newline at end of file
SPIRV/GlslangToSpv.cpp 100644 → 100755
......@@ -4666,7 +4666,13 @@ spv::Id TGlslangToSpvTraverser::createUnaryOperation(glslang::TOperator op, OpDe
libCall = spv::CubeFaceCoordAMD;
break;
#endif
#ifdef NV_EXTENSIONS
case glslang::EOpSubgroupPartition:
builder.addExtension(spv::E_SPV_NV_shader_subgroup_partitioned);
builder.addCapability(spv::CapabilityGroupNonUniformPartitionedNV);
unaryOp = spv::OpGroupNonUniformPartitionNV;
break;
#endif
default:
return 0;
}
......@@ -5629,6 +5635,32 @@ spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, s
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformQuad);
break;
#ifdef NV_EXTENSIONS
case glslang::EOpSubgroupPartitionedAdd:
case glslang::EOpSubgroupPartitionedMul:
case glslang::EOpSubgroupPartitionedMin:
case glslang::EOpSubgroupPartitionedMax:
case glslang::EOpSubgroupPartitionedAnd:
case glslang::EOpSubgroupPartitionedOr:
case glslang::EOpSubgroupPartitionedXor:
case glslang::EOpSubgroupPartitionedInclusiveAdd:
case glslang::EOpSubgroupPartitionedInclusiveMul:
case glslang::EOpSubgroupPartitionedInclusiveMin:
case glslang::EOpSubgroupPartitionedInclusiveMax:
case glslang::EOpSubgroupPartitionedInclusiveAnd:
case glslang::EOpSubgroupPartitionedInclusiveOr:
case glslang::EOpSubgroupPartitionedInclusiveXor:
case glslang::EOpSubgroupPartitionedExclusiveAdd:
case glslang::EOpSubgroupPartitionedExclusiveMul:
case glslang::EOpSubgroupPartitionedExclusiveMin:
case glslang::EOpSubgroupPartitionedExclusiveMax:
case glslang::EOpSubgroupPartitionedExclusiveAnd:
case glslang::EOpSubgroupPartitionedExclusiveOr:
case glslang::EOpSubgroupPartitionedExclusiveXor:
builder.addExtension(spv::E_SPV_NV_shader_subgroup_partitioned);
builder.addCapability(spv::CapabilityGroupNonUniformPartitionedNV);
break;
#endif
default: assert(0 && "Unhandled subgroup operation!");
}
......@@ -5662,6 +5694,11 @@ spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, s
case glslang::EOpSubgroupInclusiveAdd:
case glslang::EOpSubgroupExclusiveAdd:
case glslang::EOpSubgroupClusteredAdd:
#ifdef NV_EXTENSIONS
case glslang::EOpSubgroupPartitionedAdd:
case glslang::EOpSubgroupPartitionedInclusiveAdd:
case glslang::EOpSubgroupPartitionedExclusiveAdd:
#endif
if (isFloat) {
opCode = spv::OpGroupNonUniformFAdd;
} else {
......@@ -5672,6 +5709,11 @@ spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, s
case glslang::EOpSubgroupInclusiveMul:
case glslang::EOpSubgroupExclusiveMul:
case glslang::EOpSubgroupClusteredMul:
#ifdef NV_EXTENSIONS
case glslang::EOpSubgroupPartitionedMul:
case glslang::EOpSubgroupPartitionedInclusiveMul:
case glslang::EOpSubgroupPartitionedExclusiveMul:
#endif
if (isFloat) {
opCode = spv::OpGroupNonUniformFMul;
} else {
......@@ -5682,6 +5724,11 @@ spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, s
case glslang::EOpSubgroupInclusiveMin:
case glslang::EOpSubgroupExclusiveMin:
case glslang::EOpSubgroupClusteredMin:
#ifdef NV_EXTENSIONS
case glslang::EOpSubgroupPartitionedMin:
case glslang::EOpSubgroupPartitionedInclusiveMin:
case glslang::EOpSubgroupPartitionedExclusiveMin:
#endif
if (isFloat) {
opCode = spv::OpGroupNonUniformFMin;
} else if (isUnsigned) {
......@@ -5694,6 +5741,11 @@ spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, s
case glslang::EOpSubgroupInclusiveMax:
case glslang::EOpSubgroupExclusiveMax:
case glslang::EOpSubgroupClusteredMax:
#ifdef NV_EXTENSIONS
case glslang::EOpSubgroupPartitionedMax:
case glslang::EOpSubgroupPartitionedInclusiveMax:
case glslang::EOpSubgroupPartitionedExclusiveMax:
#endif
if (isFloat) {
opCode = spv::OpGroupNonUniformFMax;
} else if (isUnsigned) {
......@@ -5706,6 +5758,11 @@ spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, s
case glslang::EOpSubgroupInclusiveAnd:
case glslang::EOpSubgroupExclusiveAnd:
case glslang::EOpSubgroupClusteredAnd:
#ifdef NV_EXTENSIONS
case glslang::EOpSubgroupPartitionedAnd:
case glslang::EOpSubgroupPartitionedInclusiveAnd:
case glslang::EOpSubgroupPartitionedExclusiveAnd:
#endif
if (isBool) {
opCode = spv::OpGroupNonUniformLogicalAnd;
} else {
......@@ -5716,6 +5773,11 @@ spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, s
case glslang::EOpSubgroupInclusiveOr:
case glslang::EOpSubgroupExclusiveOr:
case glslang::EOpSubgroupClusteredOr:
#ifdef NV_EXTENSIONS
case glslang::EOpSubgroupPartitionedOr:
case glslang::EOpSubgroupPartitionedInclusiveOr:
case glslang::EOpSubgroupPartitionedExclusiveOr:
#endif
if (isBool) {
opCode = spv::OpGroupNonUniformLogicalOr;
} else {
......@@ -5726,6 +5788,11 @@ spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, s
case glslang::EOpSubgroupInclusiveXor:
case glslang::EOpSubgroupExclusiveXor:
case glslang::EOpSubgroupClusteredXor:
#ifdef NV_EXTENSIONS
case glslang::EOpSubgroupPartitionedXor:
case glslang::EOpSubgroupPartitionedInclusiveXor:
case glslang::EOpSubgroupPartitionedExclusiveXor:
#endif
if (isBool) {
opCode = spv::OpGroupNonUniformLogicalXor;
} else {
......@@ -5786,6 +5853,35 @@ spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, s
case glslang::EOpSubgroupClusteredXor:
spvGroupOperands.push_back(spv::GroupOperationClusteredReduce);
break;
#ifdef NV_EXTENSIONS
case glslang::EOpSubgroupPartitionedAdd:
case glslang::EOpSubgroupPartitionedMul:
case glslang::EOpSubgroupPartitionedMin:
case glslang::EOpSubgroupPartitionedMax:
case glslang::EOpSubgroupPartitionedAnd:
case glslang::EOpSubgroupPartitionedOr:
case glslang::EOpSubgroupPartitionedXor:
spvGroupOperands.push_back(spv::GroupOperationPartitionedReduceNV);
break;
case glslang::EOpSubgroupPartitionedInclusiveAdd:
case glslang::EOpSubgroupPartitionedInclusiveMul:
case glslang::EOpSubgroupPartitionedInclusiveMin:
case glslang::EOpSubgroupPartitionedInclusiveMax:
case glslang::EOpSubgroupPartitionedInclusiveAnd:
case glslang::EOpSubgroupPartitionedInclusiveOr:
case glslang::EOpSubgroupPartitionedInclusiveXor:
spvGroupOperands.push_back(spv::GroupOperationPartitionedInclusiveScanNV);
break;
case glslang::EOpSubgroupPartitionedExclusiveAdd:
case glslang::EOpSubgroupPartitionedExclusiveMul:
case glslang::EOpSubgroupPartitionedExclusiveMin:
case glslang::EOpSubgroupPartitionedExclusiveMax:
case glslang::EOpSubgroupPartitionedExclusiveAnd:
case glslang::EOpSubgroupPartitionedExclusiveOr:
case glslang::EOpSubgroupPartitionedExclusiveXor:
spvGroupOperands.push_back(spv::GroupOperationPartitionedExclusiveScanNV);
break;
#endif
}
// Push back the operands next.
......@@ -5974,6 +6070,29 @@ spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::
case glslang::EOpSubgroupClusteredOr:
case glslang::EOpSubgroupClusteredXor:
case glslang::EOpSubgroupQuadBroadcast:
#ifdef NV_EXTENSIONS
case glslang::EOpSubgroupPartitionedAdd:
case glslang::EOpSubgroupPartitionedMul:
case glslang::EOpSubgroupPartitionedMin:
case glslang::EOpSubgroupPartitionedMax:
case glslang::EOpSubgroupPartitionedAnd:
case glslang::EOpSubgroupPartitionedOr:
case glslang::EOpSubgroupPartitionedXor:
case glslang::EOpSubgroupPartitionedInclusiveAdd:
case glslang::EOpSubgroupPartitionedInclusiveMul:
case glslang::EOpSubgroupPartitionedInclusiveMin:
case glslang::EOpSubgroupPartitionedInclusiveMax:
case glslang::EOpSubgroupPartitionedInclusiveAnd:
case glslang::EOpSubgroupPartitionedInclusiveOr:
case glslang::EOpSubgroupPartitionedInclusiveXor:
case glslang::EOpSubgroupPartitionedExclusiveAdd:
case glslang::EOpSubgroupPartitionedExclusiveMul:
case glslang::EOpSubgroupPartitionedExclusiveMin:
case glslang::EOpSubgroupPartitionedExclusiveMax:
case glslang::EOpSubgroupPartitionedExclusiveAnd:
case glslang::EOpSubgroupPartitionedExclusiveOr:
case glslang::EOpSubgroupPartitionedExclusiveXor:
#endif
return createSubgroupOperation(op, typeId, operands, typeProxy);
#ifdef AMD_EXTENSIONS
......
SPIRV/doc.cpp 100644 → 100755
......@@ -728,17 +728,22 @@ const char* ScopeString(int mem)
}
}
const int GroupOperationCeiling = 4;
const int GroupOperationCeiling = 9;
const char* GroupOperationString(int gop)
{
switch (gop)
{
case 0: return "Reduce";
case 1: return "InclusiveScan";
case 2: return "ExclusiveScan";
case 3: return "ClusteredReduce";
case GroupOperationReduce: return "Reduce";
case GroupOperationInclusiveScan: return "InclusiveScan";
case GroupOperationExclusiveScan: return "ExclusiveScan";
case GroupOperationClusteredReduce: return "ClusteredReduce";
#ifdef NV_EXTENSIONS
case GroupOperationPartitionedReduceNV: return "PartitionedReduceNV";
case GroupOperationPartitionedInclusiveScanNV: return "PartitionedInclusiveScanNV";
case GroupOperationPartitionedExclusiveScanNV: return "PartitionedExclusiveScanNV";
#endif
case GroupOperationCeiling:
default: return "Bad";
......@@ -876,6 +881,7 @@ const char* CapabilityString(int info)
case 5255: return "ShaderViewportMaskNV";
case 5259: return "ShaderStereoViewNV";
case 5260: return "PerViewAttributesNV";
case 5297: return "GroupNonUniformPartitionedNV";
#endif
case 5265: return "FragmentFullyCoveredEXT";
......@@ -1272,6 +1278,9 @@ const char* OpcodeString(int op)
case OpDecorateStringGOOGLE: return "OpDecorateStringGOOGLE";
case OpMemberDecorateStringGOOGLE: return "OpMemberDecorateStringGOOGLE";
#ifdef NV_EXTENSIONS
case 5296: return "OpGroupNonUniformPartitionNV";
#endif
case OpcodeCeiling:
default:
return "Bad";
......@@ -3138,6 +3147,11 @@ void Parameterize()
InstructionDesc[OpFragmentFetchAMD].operands.push(OperandId, "'Coordinate'");
InstructionDesc[OpFragmentFetchAMD].operands.push(OperandId, "'Fragment Index'");
#endif
#ifdef NV_EXTENSIONS
InstructionDesc[OpGroupNonUniformPartitionNV].capabilities.push_back(CapabilityGroupNonUniformPartitionedNV);
InstructionDesc[OpGroupNonUniformPartitionNV].operands.push(OperandId, "X");
#endif
}
}; // end spv namespace
SPIRV/spirv.hpp
......@@ -572,6 +572,9 @@ enum GroupOperation {
GroupOperationInclusiveScan = 1,
GroupOperationExclusiveScan = 2,
GroupOperationClusteredReduce = 3,
GroupOperationPartitionedReduceNV = 6,
GroupOperationPartitionedInclusiveScanNV = 7,
GroupOperationPartitionedExclusiveScanNV = 8,
GroupOperationMax = 0x7fffffff,
};
......@@ -688,6 +691,7 @@ enum Capability {
CapabilityShaderStereoViewNV = 5259,
CapabilityPerViewAttributesNV = 5260,
CapabilityFragmentFullyCoveredEXT = 5265,
CapabilityGroupNonUniformPartitionedNV = 5297,
CapabilitySubgroupShuffleINTEL = 5568,
CapabilitySubgroupBufferBlockIOINTEL = 5569,
CapabilitySubgroupImageBlockIOINTEL = 5570,
......@@ -1051,6 +1055,7 @@ enum Op {
OpGroupSMaxNonUniformAMD = 5007,
OpFragmentMaskFetchAMD = 5011,
OpFragmentFetchAMD = 5012,
OpGroupNonUniformPartitionNV = 5296,
OpSubgroupShuffleINTEL = 5571,
OpSubgroupShuffleDownINTEL = 5572,
OpSubgroupShuffleUpINTEL = 5573,
......
Test/baseResults/spv.specConstant.vert.out
......@@ -11,7 +11,7 @@ spv.specConstant.vert
Source GLSL 400
Name 4 "main"
Name 9 "arraySize"
Name 14 "foo(vf4[s2148];"
Name 14 "foo(vf4[s2468];"
Name 13 "p"
Name 17 "builtin_spec_constant("
Name 20 "color"
......@@ -102,10 +102,10 @@ spv.specConstant.vert
Store 20(color) 46
48: 10 Load 22(ucol)
Store 47(param) 48
49: 2 FunctionCall 14(foo(vf4[s2148];) 47(param)
49: 2 FunctionCall 14(foo(vf4[s2468];) 47(param)
Return
FunctionEnd
14(foo(vf4[s2148];): 2 Function None 12
14(foo(vf4[s2468];): 2 Function None 12
13(p): 11(ptr) FunctionParameter
15: Label
54: 24(ptr) AccessChain 53(dupUcol) 23
......
Test/baseResults/spv.subgroupPartitioned.comp.out 0 → 100755
This source diff could not be displayed because it is too large. You can view the blob instead.
glslang/Include/intermediate.h 100644 → 100755
......@@ -527,6 +527,32 @@ enum TOperator {
EOpSubgroupQuadSwapHorizontal,
EOpSubgroupQuadSwapVertical,
EOpSubgroupQuadSwapDiagonal,
#ifdef NV_EXTENSIONS
EOpSubgroupPartition,
EOpSubgroupPartitionedAdd,
EOpSubgroupPartitionedMul,
EOpSubgroupPartitionedMin,
EOpSubgroupPartitionedMax,
EOpSubgroupPartitionedAnd,
EOpSubgroupPartitionedOr,
EOpSubgroupPartitionedXor,
EOpSubgroupPartitionedInclusiveAdd,
EOpSubgroupPartitionedInclusiveMul,
EOpSubgroupPartitionedInclusiveMin,
EOpSubgroupPartitionedInclusiveMax,
EOpSubgroupPartitionedInclusiveAnd,
EOpSubgroupPartitionedInclusiveOr,
EOpSubgroupPartitionedInclusiveXor,
EOpSubgroupPartitionedExclusiveAdd,
EOpSubgroupPartitionedExclusiveMul,
EOpSubgroupPartitionedExclusiveMin,
EOpSubgroupPartitionedExclusiveMax,
EOpSubgroupPartitionedExclusiveAnd,
EOpSubgroupPartitionedExclusiveOr,
EOpSubgroupPartitionedExclusiveXor,
#endif
EOpSubgroupGuardStop,
#ifdef AMD_EXTENSIONS
......
glslang/MachineIndependent/Versions.cpp 100644 → 100755
......@@ -226,6 +226,7 @@ void TParseVersions::initializeExtensionBehavior()
extensionBehavior[E_GL_NV_shader_atomic_int64] = EBhDisable;
extensionBehavior[E_GL_NV_conservative_raster_underestimation] = EBhDisable;
extensionBehavior[E_GL_NV_shader_noperspective_interpolation] = EBhDisable;
extensionBehavior[E_GL_NV_shader_subgroup_partitioned] = EBhDisable;
#endif
// AEP
......@@ -389,6 +390,7 @@ void TParseVersions::getPreamble(std::string& preamble)
"#define GL_NV_viewport_array2 1\n"
"#define GL_NV_shader_atomic_int64 1\n"
"#define GL_NV_conservative_raster_underestimation 1\n"
"#define GL_NV_shader_subgroup_partitioned 1\n"
#endif
"#define GL_KHX_shader_explicit_arithmetic_types 1\n"
"#define GL_KHX_shader_explicit_arithmetic_types_int8 1\n"
......@@ -750,6 +752,10 @@ void TParseVersions::updateExtensionBehavior(int line, const char* extension, co
updateExtensionBehavior(line, "GL_KHR_shader_subgroup_basic", behaviorString);
else if (strcmp(extension, "GL_KHR_shader_subgroup_quad") == 0)
updateExtensionBehavior(line, "GL_KHR_shader_subgroup_basic", behaviorString);
#ifdef NV_EXTENSIONS
else if (strcmp(extension, "GL_NV_shader_subgroup_partitioned") == 0)
updateExtensionBehavior(line, "GL_KHR_shader_subgroup_basic", behaviorString);
#endif
}
void TParseVersions::updateExtensionBehavior(const char* extension, TExtensionBehavior behavior)
......
glslang/MachineIndependent/Versions.h 100644 → 100755
......@@ -197,6 +197,7 @@ const char* const E_GL_NVX_multiview_per_view_attributes = "GL_NVX_multiv
const char* const E_GL_NV_shader_atomic_int64 = "GL_NV_shader_atomic_int64";
const char* const E_GL_NV_conservative_raster_underestimation = "GL_NV_conservative_raster_underestimation";
const char* const E_GL_NV_shader_noperspective_interpolation = "GL_NV_shader_noperspective_interpolation";
const char* const E_GL_NV_shader_subgroup_partitioned = "GL_NV_shader_subgroup_partitioned";
// Arrays of extensions for the above viewportEXTs duplications
......
glslang/MachineIndependent/intermOut.cpp
......@@ -599,6 +599,31 @@ bool TOutputTraverser::visitUnary(TVisit /* visit */, TIntermUnary* node)
case EOpSubgroupQuadSwapVertical: out.debug << "subgroupQuadSwapVertical"; break;
case EOpSubgroupQuadSwapDiagonal: out.debug << "subgroupQuadSwapDiagonal"; break;
#ifdef NV_EXTENSIONS
case EOpSubgroupPartition: out.debug << "subgroupPartitionNV"; break;
case EOpSubgroupPartitionedAdd: out.debug << "subgroupPartitionedAddNV"; break;
case EOpSubgroupPartitionedMul: out.debug << "subgroupPartitionedMulNV"; break;
case EOpSubgroupPartitionedMin: out.debug << "subgroupPartitionedMinNV"; break;
case EOpSubgroupPartitionedMax: out.debug << "subgroupPartitionedMaxNV"; break;
case EOpSubgroupPartitionedAnd: out.debug << "subgroupPartitionedAndNV"; break;
case EOpSubgroupPartitionedOr: out.debug << "subgroupPartitionedOrNV"; break;
case EOpSubgroupPartitionedXor: out.debug << "subgroupPartitionedXorNV"; break;
case EOpSubgroupPartitionedInclusiveAdd: out.debug << "subgroupPartitionedInclusiveAddNV"; break;
case EOpSubgroupPartitionedInclusiveMul: out.debug << "subgroupPartitionedInclusiveMulNV"; break;
case EOpSubgroupPartitionedInclusiveMin: out.debug << "subgroupPartitionedInclusiveMinNV"; break;
case EOpSubgroupPartitionedInclusiveMax: out.debug << "subgroupPartitionedInclusiveMaxNV"; break;
case EOpSubgroupPartitionedInclusiveAnd: out.debug << "subgroupPartitionedInclusiveAndNV"; break;
case EOpSubgroupPartitionedInclusiveOr: out.debug << "subgroupPartitionedInclusiveOrNV"; break;
case EOpSubgroupPartitionedInclusiveXor: out.debug << "subgroupPartitionedInclusiveXorNV"; break;
case EOpSubgroupPartitionedExclusiveAdd: out.debug << "subgroupPartitionedExclusiveAddNV"; break;
case EOpSubgroupPartitionedExclusiveMul: out.debug << "subgroupPartitionedExclusiveMulNV"; break;
case EOpSubgroupPartitionedExclusiveMin: out.debug << "subgroupPartitionedExclusiveMinNV"; break;
case EOpSubgroupPartitionedExclusiveMax: out.debug << "subgroupPartitionedExclusiveMaxNV"; break;
case EOpSubgroupPartitionedExclusiveAnd: out.debug << "subgroupPartitionedExclusiveAndNV"; break;
case EOpSubgroupPartitionedExclusiveOr: out.debug << "subgroupPartitionedExclusiveOrNV"; break;
case EOpSubgroupPartitionedExclusiveXor: out.debug << "subgroupPartitionedExclusiveXorNV"; break;
#endif
case EOpClip: out.debug << "clip"; break;
case EOpIsFinite: out.debug << "isfinite"; break;
case EOpLog10: out.debug << "log10"; break;
......
gtests/Spv.FromFile.cpp 100644 → 100755
......@@ -338,6 +338,7 @@ INSTANTIATE_TEST_CASE_P(
"spv.precise.tese",
"spv.precise.tesc",
"spv.vulkan100.subgroupArithmetic.comp",
"spv.vulkan100.subgroupPartitioned.comp",
"spv.xfb.vert",
"spv.xfb2.vert",
"spv.xfb3.vert",
......@@ -368,6 +369,7 @@ INSTANTIATE_TEST_CASE_P(
"spv.subgroupBallot.comp",
"spv.subgroupClustered.comp",
"spv.subgroupClusteredNeg.comp",
"spv.subgroupPartitioned.comp",
"spv.subgroupShuffle.comp",
"spv.subgroupShuffleRelative.comp",
"spv.subgroupQuad.comp",
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment