AuroraMiddleware/glslang
1
0
Fork 0
mirror of https://github.com/KhronosGroup/glslang synced 2024-11-09 12:00:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

GL_EXT_mesh_shader/SPV_EXT_mesh_shader implementation

Browse Source 
Added following updates to GL_EXT_mesh_shader implementation:

1. Added SPIRV and GLSL test cases
2. Added checks to ensure NV and EXT mesh shader builtins cannot be used interchangeably.
3. Updated the language name by removing the postfix "NV" to MeshShader and TaskShader.
4. Added checks for grammar checking to comply with the spec.

5. Added gl_NumWorkGroups builtin to Mesh shader
6. Fixed data type of gl_PrimitiveLineIndicesEXT and gl_PrimitiveTriangleIndicesEXT
7. Added new constants to the resources table
8. Updates to handle new storage qualifier "taskPayloadSharedEXT"
9. Updated test cases by replacing "taskEXT" with storage qualifier "taskPayloadSharedEXT"

Addressed  Review comments
1. Fixed instruction description used by glslang disassembly.
2. Updated OpEmitMeshTasksEXT as per spec update
3. Fixed implementation that errors out if there are more then one taskPayloadSharedEXT varjables.
4. Fixed miscellaneous error logs and removed unwanted code.

SPIRV 1.6 related build failure fixes
- Update SPIRV header to 1.6
- Fix conflict wiht SPIRV 1.6 change, where localSizeId is used for execution mode for mesh/task shaders

Enable SPIRV generated for EXT_mesh_shader to be version 1.4

GL_EXT_mesh_shader: Add checks for atomic support and corresponding test cases
This commit is contained in:
Pankaj Mistry 2021-08-25 13:41:32 -07:00
parent 69ae9e7460
commit 228c67228a
52 changed files with 8937 additions and 4185 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
SPIRV/CInterface/spirv_c_interface.cpp
View File

@ -71,10 +71,10 @@ static EShLanguage c_shader_stage(glslang_stage_t stage)
return EShLangMiss;
case GLSLANG_STAGE_CALLABLE_NV:
return EShLangCallable;
case GLSLANG_STAGE_TASK_NV:
return EShLangTaskNV;
case GLSLANG_STAGE_MESH_NV:
return EShLangMeshNV;
case GLSLANG_STAGE_TASK:
return EShLangTask;
case GLSLANG_STAGE_MESH:
return EShLangMesh;
default:
break;
}

1
SPIRV/GLSL.ext.EXT.h
View File

@ -39,5 +39,6 @@ static const char* const E_SPV_EXT_shader_atomic_float_add = "SPV_EXT_shader_ato
static const char* const E_SPV_EXT_shader_atomic_float16_add = "SPV_EXT_shader_atomic_float16_add";
static const char* const E_SPV_EXT_shader_atomic_float_min_max = "SPV_EXT_shader_atomic_float_min_max";
static const char* const E_SPV_EXT_shader_image_int64 = "SPV_EXT_shader_image_int64";
static const char* const E_SPV_EXT_mesh_shader = "SPV_EXT_mesh_shader";
#endif // #ifndef GLSLextEXT_H

83
SPIRV/GlslangToSpv.cpp
View File

@ -280,6 +280,9 @@ protected:
// Used later for generating OpTraceKHR/OpExecuteCallableKHR
std::unordered_map<unsigned int, glslang::TIntermSymbol *> locationToSymbol[2];
// Used by Task shader while generating opearnds for OpEmitMeshTasksEXT
spv::Id taskPayloadID;
};
//
@ -315,7 +318,7 @@ spv::SourceLanguage TranslateSourceLanguage(glslang::EShSource source, EProfile
}
// Translate glslang language (stage) to SPIR-V execution model.
spv::ExecutionModel TranslateExecutionModel(EShLanguage stage)
spv::ExecutionModel TranslateExecutionModel(EShLanguage stage, bool isMeshShaderEXT = false)
{
switch (stage) {
case EShLangVertex: return spv::ExecutionModelVertex;
@ -331,8 +334,8 @@ spv::ExecutionModel TranslateExecutionModel(EShLanguage stage)
case EShLangClosestHit: return spv::ExecutionModelClosestHitKHR;
case EShLangMiss: return spv::ExecutionModelMissKHR;
case EShLangCallable: return spv::ExecutionModelCallableKHR;
case EShLangTaskNV: return spv::ExecutionModelTaskNV;
case EShLangMeshNV: return spv::ExecutionModelMeshNV;
case EShLangTask: return (isMeshShaderEXT)? spv::ExecutionModelTaskEXT : spv::ExecutionModelTaskNV;
case EShLangMesh: return (isMeshShaderEXT)? spv::ExecutionModelMeshEXT: spv::ExecutionModelMeshNV;
#endif
default:
assert(0);
@ -763,7 +766,7 @@ spv::BuiltIn TGlslangToSpvTraverser::TranslateBuiltInDecoration(glslang::TBuiltI
return spv::BuiltInSampleMask;
case glslang::EbvLayer:
if (glslangIntermediate->getStage() == EShLangMeshNV) {
if (glslangIntermediate->getStage() == EShLangMesh) {
return spv::BuiltInLayer;
}
if (glslangIntermediate->getStage() == EShLangGeometry ||
@ -1078,6 +1081,16 @@ spv::BuiltIn TGlslangToSpvTraverser::TranslateBuiltInDecoration(glslang::TBuiltI
case glslang::EbvMeshViewIndicesNV:
return spv::BuiltInMeshViewIndicesNV;
// SPV_EXT_mesh_shader
case glslang::EbvPrimitivePointIndicesEXT:
return spv::BuiltInPrimitivePointIndicesEXT;
case glslang::EbvPrimitiveLineIndicesEXT:
return spv::BuiltInPrimitiveLineIndicesEXT;
case glslang::EbvPrimitiveTriangleIndicesEXT:
return spv::BuiltInPrimitiveTriangleIndicesEXT;
case glslang::EbvCullPrimitiveEXT:
return spv::BuiltInCullPrimitiveEXT;
// sm builtins
case glslang::EbvWarpsPerSM:
builder.addExtension(spv::E_SPV_NV_shader_sm_builtins);
@ -1321,6 +1334,7 @@ spv::StorageClass TGlslangToSpvTraverser::TranslateStorageClass(const glslang::T
case glslang::EvqHitAttr: return spv::StorageClassHitAttributeKHR;
case glslang::EvqCallableData: return spv::StorageClassCallableDataKHR;
case glslang::EvqCallableDataIn: return spv::StorageClassIncomingCallableDataKHR;
case glslang::EvqtaskPayloadSharedEXT : return spv::StorageClassTaskPayloadWorkgroupEXT;
case glslang::EvqSpirvStorageClass: return static_cast<spv::StorageClass>(type.getQualifier().spirvStorageClass);
#endif
default:
@ -1466,6 +1480,8 @@ void InheritQualifiers(glslang::TQualifier& child, const glslang::TQualifier& pa
child.perViewNV = true;
if (parent.perTaskNV)
child.perTaskNV = true;
if (parent.storage == glslang::EvqtaskPayloadSharedEXT)
child.storage = glslang::EvqtaskPayloadSharedEXT;
if (parent.patch)
child.patch = true;
if (parent.sample)
@ -1525,9 +1541,12 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion,
inEntryPoint(false), entryPointTerminated(false), linkageOnly(false),
glslangIntermediate(glslangIntermediate),
nanMinMaxClamp(glslangIntermediate->getNanMinMaxClamp()),
nonSemanticDebugPrintf(0)
nonSemanticDebugPrintf(0),
taskPayloadID(0)
{
spv::ExecutionModel executionModel = TranslateExecutionModel(glslangIntermediate->getStage());
bool isMeshShaderExt = (glslangIntermediate->getRequestedExtensions().find(glslang::E_GL_EXT_mesh_shader) !=
glslangIntermediate->getRequestedExtensions().end());
spv::ExecutionModel executionModel = TranslateExecutionModel(glslangIntermediate->getStage(), isMeshShaderExt);
builder.clearAccessChain();
builder.setSource(TranslateSourceLanguage(glslangIntermediate->getSource(), glslangIntermediate->getProfile()),
@ -1823,10 +1842,15 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion,
}
break;
}
case EShLangTaskNV:
case EShLangMeshNV:
builder.addCapability(spv::CapabilityMeshShadingNV);
builder.addExtension(spv::E_SPV_NV_mesh_shader);
case EShLangTask:
case EShLangMesh:
if(isMeshShaderExt) {
builder.addCapability(spv::CapabilityMeshShadingEXT);
builder.addExtension(spv::E_SPV_EXT_mesh_shader);
} else {
builder.addCapability(spv::CapabilityMeshShadingNV);
builder.addExtension(spv::E_SPV_NV_mesh_shader);
}
if (glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_6) {
std::vector<spv::Id> dimConstId;
for (int dim = 0; dim < 3; ++dim) {
@ -1843,7 +1867,7 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion,
glslangIntermediate->getLocalSize(1),
glslangIntermediate->getLocalSize(2));
}
if (glslangIntermediate->getStage() == EShLangMeshNV) {
if (glslangIntermediate->getStage() == EShLangMesh) {
builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices,
glslangIntermediate->getVertices());
builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputPrimitivesNV,
@ -1962,7 +1986,6 @@ void TGlslangToSpvTraverser::visitSymbol(glslang::TIntermSymbol* symbol)
if (symbol->getType().getQualifier().isSpecConstant())
spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
#ifdef ENABLE_HLSL
// Skip symbol handling if it is string-typed
if (symbol->getBasicType() == glslang::EbtString)
@ -1973,6 +1996,9 @@ void TGlslangToSpvTraverser::visitSymbol(glslang::TIntermSymbol* symbol)
// Formal function parameters were mapped during makeFunctions().
spv::Id id = getSymbolId(symbol);
if (symbol->getType().getQualifier().isTaskPayload())
taskPayloadID = id; // cache the taskPayloadID to be used it as operand for OpEmitMeshTasksEXT
if (builder.isPointer(id)) {
if (!symbol->getType().getQualifier().isParamInput() &&
!symbol->getType().getQualifier().isParamOutput()) {
@ -3111,6 +3137,8 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt
case glslang::EOpExecuteCallableNV:
case glslang::EOpExecuteCallableKHR:
case glslang::EOpWritePackedPrimitiveIndices4x8NV:
case glslang::EOpEmitMeshTasksEXT:
case glslang::EOpSetMeshOutputsEXT:
noReturnValue = true;
break;
case glslang::EOpRayQueryInitialize:
@ -4396,7 +4424,7 @@ bool TGlslangToSpvTraverser::filterMember(const glslang::TType& member)
extensions.find("GL_NV_stereo_view_rendering") == extensions.end())
return true;
if (glslangIntermediate->getStage() != EShLangMeshNV) {
if (glslangIntermediate->getStage() != EShLangMesh) {
if (member.getFieldName() == "gl_ViewportMask" &&
extensions.find("GL_NV_viewport_array2") == extensions.end())
return true;
@ -8527,6 +8555,14 @@ spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::
case glslang::EOpWritePackedPrimitiveIndices4x8NV:
builder.createNoResultOp(spv::OpWritePackedPrimitiveIndices4x8NV, operands);
return 0;
case glslang::EOpEmitMeshTasksEXT:
if(taskPayloadID)
operands.push_back(taskPayloadID);
builder.createNoResultOp(spv::OpEmitMeshTasksEXT, operands);
return 0;
case glslang::EOpSetMeshOutputsEXT:
builder.createNoResultOp(spv::OpSetMeshOutputsEXT, operands);
return 0;
case glslang::EOpCooperativeMatrixMulAdd:
opCode = spv::OpCooperativeMatrixMulAddNV;
break;
@ -9011,13 +9047,21 @@ spv::Id TGlslangToSpvTraverser::getSymbolId(const glslang::TIntermSymbol* symbol
// add per-primitive, per-view. per-task decorations to a struct member (member >= 0) or an object
void TGlslangToSpvTraverser::addMeshNVDecoration(spv::Id id, int member, const glslang::TQualifier& qualifier)
{
bool isMeshShaderExt = (glslangIntermediate->getRequestedExtensions().find(glslang::E_GL_EXT_mesh_shader) !=
glslangIntermediate->getRequestedExtensions().end());
if (member >= 0) {
if (qualifier.perPrimitiveNV) {
// Need to add capability/extension for fragment shader.
// Mesh shader already adds this by default.
if (glslangIntermediate->getStage() == EShLangFragment) {
builder.addCapability(spv::CapabilityMeshShadingNV);
builder.addExtension(spv::E_SPV_NV_mesh_shader);
if(isMeshShaderExt) {
builder.addCapability(spv::CapabilityMeshShadingEXT);
builder.addExtension(spv::E_SPV_EXT_mesh_shader);
} else {
builder.addCapability(spv::CapabilityMeshShadingNV);
builder.addExtension(spv::E_SPV_NV_mesh_shader);
}
}
builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerPrimitiveNV);
}
@ -9030,8 +9074,13 @@ void TGlslangToSpvTraverser::addMeshNVDecoration(spv::Id id, int member, const g
// Need to add capability/extension for fragment shader.
// Mesh shader already adds this by default.
if (glslangIntermediate->getStage() == EShLangFragment) {
builder.addCapability(spv::CapabilityMeshShadingNV);
builder.addExtension(spv::E_SPV_NV_mesh_shader);
if(isMeshShaderExt) {
builder.addCapability(spv::CapabilityMeshShadingEXT);
builder.addExtension(spv::E_SPV_EXT_mesh_shader);
} else {
builder.addCapability(spv::CapabilityMeshShadingNV);
builder.addExtension(spv::E_SPV_NV_mesh_shader);
}
}
builder.addDecoration(id, spv::DecorationPerPrimitiveNV);
}

22
SPIRV/doc.cpp
View File

@ -97,6 +97,8 @@ const char* ExecutionModelString(int model)
case 6: return "Kernel";
case ExecutionModelTaskNV: return "TaskNV";
case ExecutionModelMeshNV: return "MeshNV";
case ExecutionModelTaskEXT: return "TaskEXT";
case ExecutionModelMeshEXT: return "MeshEXT";
default: return "Bad";
@ -242,7 +244,7 @@ const char* StorageClassString(int StorageClass)
case StorageClassIncomingCallableDataKHR: return "IncomingCallableDataKHR";
case StorageClassPhysicalStorageBufferEXT: return "PhysicalStorageBufferEXT";
case StorageClassTaskPayloadWorkgroupEXT: return "TaskPayloadWorkgroupEXT";
default: return "Bad";
}
}
@ -433,6 +435,10 @@ const char* BuiltInString(int builtIn)
case BuiltInWarpIDNV: return "WarpIDNV";
case BuiltInSMIDNV: return "SMIDNV";
case BuiltInCurrentRayTimeNV: return "CurrentRayTimeNV";
case BuiltInPrimitivePointIndicesEXT: return "PrimitivePointIndicesEXT";
case BuiltInPrimitiveLineIndicesEXT: return "PrimitiveLineIndicesEXT";
case BuiltInPrimitiveTriangleIndicesEXT: return "PrimitiveTriangleIndicesEXT";
case BuiltInCullPrimitiveEXT: return "CullPrimitiveEXT";
default: return "Bad";
}
@ -940,6 +946,7 @@ const char* CapabilityString(int info)
case CapabilityFragmentBarycentricKHR: return "FragmentBarycentricKHR";
case CapabilityMeshShadingNV: return "MeshShadingNV";
case CapabilityImageFootprintNV: return "ImageFootprintNV";
case CapabilityMeshShadingEXT: return "MeshShadingEXT";
// case CapabilityShadingRateNV: return "ShadingRateNV"; // superseded by FragmentDensityEXT
case CapabilitySampleMaskOverrideCoverageNV: return "SampleMaskOverrideCoverageNV";
case CapabilityFragmentDensityEXT: return "FragmentDensityEXT";
@ -1407,6 +1414,8 @@ const char* OpcodeString(int op)
case OpGroupNonUniformPartitionNV: return "OpGroupNonUniformPartitionNV";
case OpImageSampleFootprintNV: return "OpImageSampleFootprintNV";
case OpWritePackedPrimitiveIndices4x8NV: return "OpWritePackedPrimitiveIndices4x8NV";
case OpEmitMeshTasksEXT: return "OpEmitMeshTasksEXT";
case OpSetMeshOutputsEXT: return "OpSetMeshOutputsEXT";
case OpTypeRayQueryKHR: return "OpTypeRayQueryKHR";
case OpRayQueryInitializeKHR: return "OpRayQueryInitializeKHR";
@ -2981,6 +2990,17 @@ void Parameterize()
InstructionDesc[OpWritePackedPrimitiveIndices4x8NV].operands.push(OperandId, "'Index Offset'");
InstructionDesc[OpWritePackedPrimitiveIndices4x8NV].operands.push(OperandId, "'Packed Indices'");
InstructionDesc[OpEmitMeshTasksEXT].operands.push(OperandId, "'groupCountX'");
InstructionDesc[OpEmitMeshTasksEXT].operands.push(OperandId, "'groupCountY'");
InstructionDesc[OpEmitMeshTasksEXT].operands.push(OperandId, "'groupCountZ'");
InstructionDesc[OpEmitMeshTasksEXT].operands.push(OperandId, "'Payload'");
InstructionDesc[OpEmitMeshTasksEXT].setResultAndType(false, false);
InstructionDesc[OpSetMeshOutputsEXT].operands.push(OperandId, "'vertexCount'");
InstructionDesc[OpSetMeshOutputsEXT].operands.push(OperandId, "'primitiveCount'");
InstructionDesc[OpSetMeshOutputsEXT].setResultAndType(false, false);
InstructionDesc[OpTypeCooperativeMatrixNV].operands.push(OperandId, "'Component Type'");
InstructionDesc[OpTypeCooperativeMatrixNV].operands.push(OperandId, "'Scope'");
InstructionDesc[OpTypeCooperativeMatrixNV].operands.push(OperandId, "'Rows'");

19
SPIRV/spirv.hpp
View File

@ -91,6 +91,8 @@ enum ExecutionModel {
ExecutionModelMissNV = 5317,
ExecutionModelCallableKHR = 5318,
ExecutionModelCallableNV = 5318,
ExecutionModelTaskEXT = 5364,
ExecutionModelMeshEXT = 5365,
ExecutionModelMax = 0x7fffffff,
};
@ -166,10 +168,13 @@ enum ExecutionMode {
ExecutionModeStencilRefUnchangedBackAMD = 5082,
ExecutionModeStencilRefGreaterBackAMD = 5083,
ExecutionModeStencilRefLessBackAMD = 5084,
ExecutionModeOutputLinesEXT = 5269,
ExecutionModeOutputLinesNV = 5269,
ExecutionModeOutputPrimitivesEXT = 5270,
ExecutionModeOutputPrimitivesNV = 5270,
ExecutionModeDerivativeGroupQuadsNV = 5289,
ExecutionModeDerivativeGroupLinearNV = 5290,
ExecutionModeOutputTrianglesEXT = 5298,
ExecutionModeOutputTrianglesNV = 5298,
ExecutionModePixelInterlockOrderedEXT = 5366,
ExecutionModePixelInterlockUnorderedEXT = 5367,
@ -218,6 +223,7 @@ enum StorageClass {
StorageClassShaderRecordBufferNV = 5343,
StorageClassPhysicalStorageBuffer = 5349,
StorageClassPhysicalStorageBufferEXT = 5349,
StorageClassTaskPayloadWorkgroupEXT = 5402,
StorageClassCodeSectionINTEL = 5605,
StorageClassDeviceOnlyINTEL = 5936,
StorageClassHostOnlyINTEL = 5937,
@ -500,6 +506,7 @@ enum Decoration {
DecorationPassthroughNV = 5250,
DecorationViewportRelativeNV = 5252,
DecorationSecondaryViewportRelativeNV = 5256,
DecorationPerPrimitiveEXT = 5271,
DecorationPerPrimitiveNV = 5271,
DecorationPerViewNV = 5272,
DecorationPerTaskNV = 5273,
@ -647,6 +654,10 @@ enum BuiltIn {
BuiltInFragmentSizeNV = 5292,
BuiltInFragInvocationCountEXT = 5293,
BuiltInInvocationsPerPixelNV = 5293,
BuiltInPrimitivePointIndicesEXT = 5294,
BuiltInPrimitiveLineIndicesEXT = 5295,
BuiltInPrimitiveTriangleIndicesEXT = 5296,
BuiltInCullPrimitiveEXT = 5299,
BuiltInLaunchIdKHR = 5319,
BuiltInLaunchIdNV = 5319,
BuiltInLaunchSizeKHR = 5320,
@ -983,7 +994,8 @@ enum Capability {
CapabilityFragmentFullyCoveredEXT = 5265,
CapabilityMeshShadingNV = 5266,
CapabilityImageFootprintNV = 5282,
CapabilityFragmentBarycentricKHR = 5284,
CapabilityMeshShadingEXT = 5283,
CapabilityFragmentBarycentricKHR = 5284,
CapabilityFragmentBarycentricNV = 5284,
CapabilityComputeDerivativeGroupQuadsNV = 5288,
CapabilityFragmentDensityEXT = 5291,
@ -1577,6 +1589,8 @@ enum Op {
OpFragmentFetchAMD = 5012,
OpReadClockKHR = 5056,
OpImageSampleFootprintNV = 5283,
OpEmitMeshTasksEXT = 5294,
OpSetMeshOutputsEXT = 5295,
OpGroupNonUniformPartitionNV = 5296,
OpWritePackedPrimitiveIndices4x8NV = 5299,
OpReportIntersectionKHR = 5334,
@ -2234,6 +2248,8 @@ inline void HasResultAndType(Op opcode, bool *hasResult, bool *hasResultType) {
case OpReadClockKHR: *hasResult = true; *hasResultType = true; break;
case OpImageSampleFootprintNV: *hasResult = true; *hasResultType = true; break;
case OpGroupNonUniformPartitionNV: *hasResult = true; *hasResultType = true; break;
case OpEmitMeshTasksEXT: *hasResult = false; *hasResultType = false; break;
case OpSetMeshOutputsEXT: *hasResult = false; *hasResultType = false; break;
case OpWritePackedPrimitiveIndices4x8NV: *hasResult = false; *hasResultType = false; break;
case OpReportIntersectionNV: *hasResult = true; *hasResultType = true; break;
case OpIgnoreIntersectionNV: *hasResult = false; *hasResultType = false; break;
@ -2515,4 +2531,3 @@ inline FragmentShadingRateMask operator|(FragmentShadingRateMask a, FragmentShad
} // end namespace spv
#endif // #ifndef spirv_HPP

36
StandAlone/ResourceLimits.cpp
View File

@ -134,6 +134,15 @@ const TBuiltInResource DefaultTBuiltInResource = {
/* .maxTaskWorkGroupSizeY_NV = */ 1,
/* .maxTaskWorkGroupSizeZ_NV = */ 1,
/* .maxMeshViewCountNV = */ 4,
/* .maxMeshOutputVerticesEXT = */ 256,
/* .maxMeshOutputPrimitivesEXT = */ 256,
/* .maxMeshWorkGroupSizeX_EXT = */ 128,
/* .maxMeshWorkGroupSizeY_EXT = */ 128,
/* .maxMeshWorkGroupSizeZ_EXT = */ 128,
/* .maxTaskWorkGroupSizeX_EXT = */ 128,
/* .maxTaskWorkGroupSizeY_EXT = */ 128,
/* .maxTaskWorkGroupSizeZ_EXT = */ 128,
/* .maxMeshViewCountEXT = */ 4,
/* .maxDualSourceDrawBuffersEXT = */ 1,
/* .limits = */ {
@ -244,6 +253,15 @@ std::string GetDefaultTBuiltInResourceString()
<< "MaxTaskWorkGroupSizeY_NV " << DefaultTBuiltInResource.maxTaskWorkGroupSizeY_NV << "\n"
<< "MaxTaskWorkGroupSizeZ_NV " << DefaultTBuiltInResource.maxTaskWorkGroupSizeZ_NV << "\n"
<< "MaxMeshViewCountNV " << DefaultTBuiltInResource.maxMeshViewCountNV << "\n"
<< "MaxMeshOutputVerticesEXT " << DefaultTBuiltInResource.maxMeshOutputVerticesEXT << "\n"
<< "MaxMeshOutputPrimitivesEXT " << DefaultTBuiltInResource.maxMeshOutputPrimitivesEXT << "\n"
<< "MaxMeshWorkGroupSizeX_EXT " << DefaultTBuiltInResource.maxMeshWorkGroupSizeX_EXT << "\n"
<< "MaxMeshWorkGroupSizeY_EXT " << DefaultTBuiltInResource.maxMeshWorkGroupSizeY_EXT << "\n"
<< "MaxMeshWorkGroupSizeZ_EXT " << DefaultTBuiltInResource.maxMeshWorkGroupSizeZ_EXT << "\n"
<< "MaxTaskWorkGroupSizeX_EXT " << DefaultTBuiltInResource.maxTaskWorkGroupSizeX_EXT << "\n"
<< "MaxTaskWorkGroupSizeY_EXT " << DefaultTBuiltInResource.maxTaskWorkGroupSizeY_EXT << "\n"
<< "MaxTaskWorkGroupSizeZ_EXT " << DefaultTBuiltInResource.maxTaskWorkGroupSizeZ_EXT << "\n"
<< "MaxMeshViewCountEXT " << DefaultTBuiltInResource.maxMeshViewCountEXT << "\n"
<< "MaxDualSourceDrawBuffersEXT " << DefaultTBuiltInResource.maxDualSourceDrawBuffersEXT << "\n"
<< "nonInductiveForLoops " << DefaultTBuiltInResource.limits.nonInductiveForLoops << "\n"
<< "whileLoops " << DefaultTBuiltInResource.limits.whileLoops << "\n"
@ -469,6 +487,24 @@ void DecodeResourceLimits(TBuiltInResource* resources, char* config)
resources->maxTaskWorkGroupSizeZ_NV = value;
else if (tokenStr == "MaxMeshViewCountNV")
resources->maxMeshViewCountNV = value;
else if (tokenStr == "MaxMeshOutputVerticesEXT")
resources->maxMeshOutputVerticesEXT = value;
else if (tokenStr == "MaxMeshOutputPrimitivesEXT")
resources->maxMeshOutputPrimitivesEXT = value;
else if (tokenStr == "MaxMeshWorkGroupSizeX_EXT")
resources->maxMeshWorkGroupSizeX_EXT = value;
else if (tokenStr == "MaxMeshWorkGroupSizeY_EXT")
resources->maxMeshWorkGroupSizeY_EXT = value;
else if (tokenStr == "MaxMeshWorkGroupSizeZ_EXT")
resources->maxMeshWorkGroupSizeZ_EXT = value;
else if (tokenStr == "MaxTaskWorkGroupSizeX_EXT")
resources->maxTaskWorkGroupSizeX_EXT = value;
else if (tokenStr == "MaxTaskWorkGroupSizeY_EXT")
resources->maxTaskWorkGroupSizeY_EXT = value;
else if (tokenStr == "MaxTaskWorkGroupSizeZ_EXT")
resources->maxTaskWorkGroupSizeZ_EXT = value;
else if (tokenStr == "MaxMeshViewCountEXT")
resources->maxMeshViewCountEXT = value;
else if (tokenStr == "nonInductiveForLoops")
resources->limits.nonInductiveForLoops = (value != 0);
else if (tokenStr == "whileLoops")

8
StandAlone/StandAlone.cpp
View File

@ -293,8 +293,8 @@ const char* GetBinaryName(EShLanguage stage)
case EShLangClosestHit: name = "rchit.spv"; break;
case EShLangMiss: name = "rmiss.spv"; break;
case EShLangCallable: name = "rcall.spv"; break;
case EShLangMeshNV: name = "mesh.spv"; break;
case EShLangTaskNV: name = "task.spv"; break;
case EShLangMesh : name = "mesh.spv"; break;
case EShLangTask : name = "task.spv"; break;
default: name = "unknown"; break;
}
} else
@ -1780,9 +1780,9 @@ EShLanguage FindLanguage(const std::string& name, bool parseStageName)
else if (stageName == "rcall")
return EShLangCallable;
else if (stageName == "mesh")
return EShLangMeshNV;
return EShLangMesh;
else if (stageName == "task")
return EShLangTaskNV;
return EShLangTask;
usage();
return EShLangVertex;

1039
Test/baseResults/glsl.460.subgroupEXT.mesh.out Normal file
View File

File diff suppressed because it is too large Load Diff

707
Test/baseResults/glsl.460.subgroupEXT.task.out Normal file
View File

@ -0,0 +1,707 @@
glsl.460.subgroupEXT.task
ERROR: 0:6: 'gl_SubgroupSize' : required extension not requested: GL_KHR_shader_subgroup_basic
ERROR: 0:7: 'gl_SubgroupInvocationID' : required extension not requested: GL_KHR_shader_subgroup_basic
ERROR: 0:8: 'subgroupBarrier' : required extension not requested: GL_KHR_shader_subgroup_basic
ERROR: 0:9: 'subgroupMemoryBarrier' : required extension not requested: GL_KHR_shader_subgroup_basic
ERROR: 0:10: 'subgroupMemoryBarrierBuffer' : required extension not requested: GL_KHR_shader_subgroup_basic
ERROR: 0:11: 'subgroupMemoryBarrierImage' : required extension not requested: GL_KHR_shader_subgroup_basic
ERROR: 0:12: 'subgroupElect' : required extension not requested: GL_KHR_shader_subgroup_basic
ERROR: 0:13: 'gl_NumSubgroups' : required extension not requested: GL_KHR_shader_subgroup_basic
ERROR: 0:14: 'gl_SubgroupID' : required extension not requested: GL_KHR_shader_subgroup_basic
ERROR: 0:15: 'subgroupMemoryBarrierShared' : required extension not requested: GL_KHR_shader_subgroup_basic
ERROR: 0:17: 'subgroupAll' : required extension not requested: GL_KHR_shader_subgroup_vote
ERROR: 0:18: 'subgroupAny' : required extension not requested: GL_KHR_shader_subgroup_vote
ERROR: 0:19: 'subgroupAllEqual' : required extension not requested: GL_KHR_shader_subgroup_vote
ERROR: 0:21: 'gl_SubgroupEqMask' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:22: 'gl_SubgroupGeMask' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:23: 'gl_SubgroupGtMask' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:24: 'gl_SubgroupLeMask' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:25: 'gl_SubgroupLtMask' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:26: 'subgroupBroadcast' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:27: 'subgroupBroadcastFirst' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:28: 'subgroupBallot' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:29: 'subgroupInverseBallot' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:30: 'subgroupBallotBitExtract' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:31: 'subgroupBallotBitCount' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:32: 'subgroupBallotInclusiveBitCount' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:33: 'subgroupBallotExclusiveBitCount' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:34: 'subgroupBallotFindLSB' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:35: 'subgroupBallotFindMSB' : required extension not requested: GL_KHR_shader_subgroup_ballot
ERROR: 0:37: 'subgroupShuffle' : required extension not requested: GL_KHR_shader_subgroup_shuffle
ERROR: 0:38: 'subgroupShuffleXor' : required extension not requested: GL_KHR_shader_subgroup_shuffle
ERROR: 0:39: 'subgroupShuffleUp' : required extension not requested: GL_KHR_shader_subgroup_shuffle_relative
ERROR: 0:40: 'subgroupShuffleDown' : required extension not requested: GL_KHR_shader_subgroup_shuffle_relative
ERROR: 0:42: 'subgroupAdd' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:43: 'subgroupMul' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:44: 'subgroupMin' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:45: 'subgroupMax' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:46: 'subgroupAnd' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:47: 'subgroupOr' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:48: 'subgroupXor' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:49: 'subgroupInclusiveAdd' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:50: 'subgroupInclusiveMul' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:51: 'subgroupInclusiveMin' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:52: 'subgroupInclusiveMax' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:53: 'subgroupInclusiveAnd' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:54: 'subgroupInclusiveOr' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:55: 'subgroupInclusiveXor' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:56: 'subgroupExclusiveAdd' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:57: 'subgroupExclusiveMul' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:58: 'subgroupExclusiveMin' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:59: 'subgroupExclusiveMax' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:60: 'subgroupExclusiveAnd' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:61: 'subgroupExclusiveOr' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:62: 'subgroupExclusiveXor' : required extension not requested: GL_KHR_shader_subgroup_arithmetic
ERROR: 0:64: 'subgroupClusteredAdd' : required extension not requested: GL_KHR_shader_subgroup_clustered
ERROR: 0:65: 'subgroupClusteredMul' : required extension not requested: GL_KHR_shader_subgroup_clustered
ERROR: 0:66: 'subgroupClusteredMin' : required extension not requested: GL_KHR_shader_subgroup_clustered
ERROR: 0:67: 'subgroupClusteredMax' : required extension not requested: GL_KHR_shader_subgroup_clustered
ERROR: 0:68: 'subgroupClusteredAnd' : required extension not requested: GL_KHR_shader_subgroup_clustered
ERROR: 0:69: 'subgroupClusteredOr' : required extension not requested: GL_KHR_shader_subgroup_clustered
ERROR: 0:70: 'subgroupClusteredXor' : required extension not requested: GL_KHR_shader_subgroup_clustered
ERROR: 0:72: 'subgroupQuadBroadcast' : required extension not requested: GL_KHR_shader_subgroup_quad
ERROR: 0:73: 'subgroupQuadSwapHorizontal' : required extension not requested: GL_KHR_shader_subgroup_quad
ERROR: 0:74: 'subgroupQuadSwapVertical' : required extension not requested: GL_KHR_shader_subgroup_quad
ERROR: 0:75: 'subgroupQuadSwapDiagonal' : required extension not requested: GL_KHR_shader_subgroup_quad
ERROR: 64 compilation errors. No code generated.
Shader version: 460
Requested GL_EXT_mesh_shader
Requested GL_KHR_shader_subgroup_arithmetic
Requested GL_KHR_shader_subgroup_ballot
Requested GL_KHR_shader_subgroup_basic
Requested GL_KHR_shader_subgroup_clustered
Requested GL_KHR_shader_subgroup_quad
Requested GL_KHR_shader_subgroup_shuffle
Requested GL_KHR_shader_subgroup_shuffle_relative
Requested GL_KHR_shader_subgroup_vote
local_size = (32, 1, 1)
ERROR: node is still EOpNull!
0:3 Function Definition: undeclared_errors(vf4; ( global 4-component vector of float)
0:3 Function Parameters:
0:3 'f4' ( in 4-component vector of float)
0:? Sequence
0:6 'gl_SubgroupSize' ( in uint SubgroupSize)
0:7 'gl_SubgroupInvocationID' ( in uint SubgroupInvocationID)
0:8 subgroupBarrier ( global void)
0:9 subgroupMemoryBarrier ( global void)
0:10 subgroupMemoryBarrierBuffer ( global void)
0:11 subgroupMemoryBarrierImage ( global void)
0:12 subgroupElect ( global bool)
0:13 'gl_NumSubgroups' ( in uint NumSubgroups)
0:14 'gl_SubgroupID' ( in uint SubgroupID)
0:15 subgroupMemoryBarrierShared ( global void)
0:17 subgroupAll ( global bool)
0:17 Constant:
0:17 true (const bool)
0:18 subgroupAny ( global bool)
0:18 Constant:
0:18 false (const bool)
0:19 subgroupAllEqual ( global bool)
0:19 'f4' ( in 4-component vector of float)
0:21 'gl_SubgroupEqMask' ( in 4-component vector of uint SubgroupEqMask)
0:22 'gl_SubgroupGeMask' ( in 4-component vector of uint SubgroupGeMask)
0:23 'gl_SubgroupGtMask' ( in 4-component vector of uint SubgroupGtMask)
0:24 'gl_SubgroupLeMask' ( in 4-component vector of uint SubgroupLeMask)
0:25 'gl_SubgroupLtMask' ( in 4-component vector of uint SubgroupLtMask)
0:26 subgroupBroadcast ( global 4-component vector of float)
0:26 'f4' ( in 4-component vector of float)
0:26 Constant:
0:26 0 (const uint)
0:27 subgroupBroadcastFirst ( global 4-component vector of float)
0:27 'f4' ( in 4-component vector of float)
0:28 Sequence
0:28 move second child to first child ( temp 4-component vector of uint)
0:28 'ballot' ( temp 4-component vector of uint)
0:28 subgroupBallot ( global 4-component vector of uint)
0:28 Constant:
0:28 false (const bool)
0:29 subgroupInverseBallot ( global bool)
0:29 Constant:
0:29 1 (const uint)
0:29 1 (const uint)
0:29 1 (const uint)
0:29 1 (const uint)
0:30 subgroupBallotBitExtract ( global bool)
0:30 'ballot' ( temp 4-component vector of uint)
0:30 Constant:
0:30 0 (const uint)
0:31 subgroupBallotBitCount ( global uint)
0:31 'ballot' ( temp 4-component vector of uint)
0:32 subgroupBallotInclusiveBitCount ( global uint)
0:32 'ballot' ( temp 4-component vector of uint)
0:33 subgroupBallotExclusiveBitCount ( global uint)
0:33 'ballot' ( temp 4-component vector of uint)
0:34 subgroupBallotFindLSB ( global uint)
0:34 'ballot' ( temp 4-component vector of uint)
0:35 subgroupBallotFindMSB ( global uint)
0:35 'ballot' ( temp 4-component vector of uint)
0:37 subgroupShuffle ( global 4-component vector of float)
0:37 'f4' ( in 4-component vector of float)
0:37 Constant:
0:37 0 (const uint)
0:38 subgroupShuffleXor ( global 4-component vector of float)
0:38 'f4' ( in 4-component vector of float)
0:38 Constant:
0:38 1 (const uint)
0:39 subgroupShuffleUp ( global 4-component vector of float)
0:39 'f4' ( in 4-component vector of float)
0:39 Constant:
0:39 1 (const uint)
0:40 subgroupShuffleDown ( global 4-component vector of float)
0:40 'f4' ( in 4-component vector of float)
0:40 Constant:
0:40 1 (const uint)
0:42 move second child to first child ( temp 4-component vector of float)
0:42 'result' ( temp 4-component vector of float)
0:42 subgroupAdd ( global 4-component vector of float)
0:42 'f4' ( in 4-component vector of float)
0:43 subgroupMul ( global 4-component vector of float)
0:43 'f4' ( in 4-component vector of float)
0:44 subgroupMin ( global 4-component vector of float)
0:44 'f4' ( in 4-component vector of float)
0:45 subgroupMax ( global 4-component vector of float)
0:45 'f4' ( in 4-component vector of float)
0:46 subgroupAnd ( global 4-component vector of uint)
0:46 'ballot' ( temp 4-component vector of uint)
0:47 subgroupOr ( global 4-component vector of uint)
0:47 'ballot' ( temp 4-component vector of uint)
0:48 subgroupXor ( global 4-component vector of uint)
0:48 'ballot' ( temp 4-component vector of uint)
0:49 subgroupInclusiveAdd ( global 4-component vector of float)
0:49 'f4' ( in 4-component vector of float)
0:50 subgroupInclusiveMul ( global 4-component vector of float)
0:50 'f4' ( in 4-component vector of float)
0:51 subgroupInclusiveMin ( global 4-component vector of float)
0:51 'f4' ( in 4-component vector of float)
0:52 subgroupInclusiveMax ( global 4-component vector of float)
0:52 'f4' ( in 4-component vector of float)
0:53 subgroupInclusiveAnd ( global 4-component vector of uint)
0:53 'ballot' ( temp 4-component vector of uint)
0:54 subgroupInclusiveOr ( global 4-component vector of uint)
0:54 'ballot' ( temp 4-component vector of uint)
0:55 subgroupInclusiveXor ( global 4-component vector of uint)
0:55 'ballot' ( temp 4-component vector of uint)
0:56 subgroupExclusiveAdd ( global 4-component vector of float)
0:56 'f4' ( in 4-component vector of float)
0:57 subgroupExclusiveMul ( global 4-component vector of float)
0:57 'f4' ( in 4-component vector of float)
0:58 subgroupExclusiveMin ( global 4-component vector of float)
0:58 'f4' ( in 4-component vector of float)
0:59 subgroupExclusiveMax ( global 4-component vector of float)
0:59 'f4' ( in 4-component vector of float)
0:60 subgroupExclusiveAnd ( global 4-component vector of uint)
0:60 'ballot' ( temp 4-component vector of uint)
0:61 subgroupExclusiveOr ( global 4-component vector of uint)
0:61 'ballot' ( temp 4-component vector of uint)
0:62 subgroupExclusiveXor ( global 4-component vector of uint)
0:62 'ballot' ( temp 4-component vector of uint)
0:64 subgroupClusteredAdd ( global 4-component vector of float)
0:64 'f4' ( in 4-component vector of float)
0:64 Constant:
0:64 2 (const uint)
0:65 subgroupClusteredMul ( global 4-component vector of float)
0:65 'f4' ( in 4-component vector of float)
0:65 Constant:
0:65 2 (const uint)
0:66 subgroupClusteredMin ( global 4-component vector of float)
0:66 'f4' ( in 4-component vector of float)
0:66 Constant:
0:66 2 (const uint)
0:67 subgroupClusteredMax ( global 4-component vector of float)
0:67 'f4' ( in 4-component vector of float)
0:67 Constant:
0:67 2 (const uint)
0:68 subgroupClusteredAnd ( global 4-component vector of uint)
0:68 'ballot' ( temp 4-component vector of uint)
0:68 Constant:
0:68 2 (const uint)
0:69 subgroupClusteredOr ( global 4-component vector of uint)
0:69 'ballot' ( temp 4-component vector of uint)
0:69 Constant:
0:69 2 (const uint)
0:70 subgroupClusteredXor ( global 4-component vector of uint)
0:70 'ballot' ( temp 4-component vector of uint)
0:70 Constant:
0:70 2 (const uint)
0:72 subgroupQuadBroadcast ( global 4-component vector of float)
0:72 'f4' ( in 4-component vector of float)
0:72 Constant:
0:72 0 (const uint)
0:73 subgroupQuadSwapHorizontal ( global 4-component vector of float)
0:73 'f4' ( in 4-component vector of float)
0:74 subgroupQuadSwapVertical ( global 4-component vector of float)
0:74 'f4' ( in 4-component vector of float)
0:75 subgroupQuadSwapDiagonal ( global 4-component vector of float)
0:75 'f4' ( in 4-component vector of float)
0:77 Branch: Return with expression
0:77 'result' ( temp 4-component vector of float)
0:102 Function Definition: main( ( global void)
0:102 Function Parameters:
0:104 Sequence
0:104 Sequence
0:104 move second child to first child ( temp uint)
0:104 'iid' ( temp uint)
0:104 direct index ( temp uint)
0:104 'gl_LocalInvocationID' ( in 3-component vector of uint LocalInvocationID)
0:104 Constant:
0:104 0 (const int)
0:105 Sequence
0:105 move second child to first child ( temp uint)
0:105 'gid' ( temp uint)
0:105 direct index ( temp uint)
0:105 'gl_WorkGroupID' ( in 3-component vector of uint WorkGroupID)
0:105 Constant:
0:105 0 (const int)
0:108 Sequence
0:108 Sequence
0:108 move second child to first child ( temp uint)
0:108 'i' ( temp uint)
0:108 Constant:
0:108 0 (const uint)
0:108 Loop with condition tested first
0:108 Loop Condition
0:108 Compare Less Than ( temp bool)
0:108 'i' ( temp uint)
0:108 Constant:
0:108 10 (const uint)
0:108 Loop Body
0:109 Sequence
0:109 move second child to first child ( temp 4-component vector of float)
0:109 indirect index ( temp 4-component vector of float)
0:109 'mem' ( shared 10-element array of 4-component vector of float)
0:109 'i' ( temp uint)
0:109 Construct vec4 ( temp 4-component vector of float)
0:109 Convert uint to float ( temp float)
0:109 add ( temp uint)
0:109 'i' ( temp uint)
0:109 uni_value: direct index for structure (layout( column_major shared) uniform uint)
0:109 'anon@0' (layout( column_major shared) uniform block{layout( column_major shared) uniform uint uni_value})
0:109 Constant:
0:109 0 (const uint)
0:108 Loop Terminal Expression
0:108 Pre-Increment ( temp uint)
0:108 'i' ( temp uint)
0:111 imageStore ( global void)
0:111 'uni_image' (layout( binding=0) writeonly uniform image2D)
0:111 Construct ivec2 ( temp 2-component vector of int)
0:111 Convert uint to int ( temp int)
0:111 'iid' ( temp uint)
0:111 indirect index ( temp 4-component vector of float)
0:111 'mem' ( shared 10-element array of 4-component vector of float)
0:111 'gid' ( temp uint)
0:112 imageStore ( global void)
0:112 'uni_image' (layout( binding=0) writeonly uniform image2D)
0:112 Construct ivec2 ( temp 2-component vector of int)
0:112 Convert uint to int ( temp int)
0:112 'iid' ( temp uint)
0:112 indirect index ( temp 4-component vector of float)
0:112 'mem' ( shared 10-element array of 4-component vector of float)
0:112 add ( temp uint)
0:112 'gid' ( temp uint)
0:112 Constant:
0:112 1 (const uint)
0:114 MemoryBarrierShared ( global void)
0:114 Barrier ( global void)
0:118 move second child to first child ( temp 2-component vector of float)
0:118 dummy: direct index for structure ( global 2-component vector of float)
0:118 'mytask' ( taskPayloadSharedEXT structure{ global 2-component vector of float dummy, global 3-element array of 2-component vector of float submesh})
0:118 Constant:
0:118 0 (const int)
0:118 Constant:
0:118 30.000000
0:118 31.000000
0:119 move second child to first child ( temp 2-component vector of float)
0:119 direct index ( temp 2-component vector of float)
0:119 submesh: direct index for structure ( global 3-element array of 2-component vector of float)
0:119 'mytask' ( taskPayloadSharedEXT structure{ global 2-component vector of float dummy, global 3-element array of 2-component vector of float submesh})
0:119 Constant:
0:119 1 (const int)
0:119 Constant:
0:119 0 (const int)
0:119 Constant:
0:119 32.000000
0:119 33.000000
0:120 move second child to first child ( temp 2-component vector of float)
0:120 direct index ( temp 2-component vector of float)
0:120 submesh: direct index for structure ( global 3-element array of 2-component vector of float)
0:120 'mytask' ( taskPayloadSharedEXT structure{ global 2-component vector of float dummy, global 3-element array of 2-component vector of float submesh})
0:120 Constant:
0:120 1 (const int)
0:120 Constant:
0:120 1 (const int)
0:120 Constant:
0:120 34.000000
0:120 35.000000
0:121 move second child to first child ( temp 2-component vector of float)
0:121 direct index ( temp 2-component vector of float)
0:121 submesh: direct index for structure ( global 3-element array of 2-component vector of float)
0:121 'mytask' ( taskPayloadSharedEXT structure{ global 2-component vector of float dummy, global 3-element array of 2-component vector of float submesh})
0:121 Constant:
0:121 1 (const int)
0:121 Constant:
0:121 2 (const int)
0:121 indirect index ( temp 2-component vector of float)
0:121 submesh: direct index for structure ( global 3-element array of 2-component vector of float)
0:121 'mytask' ( taskPayloadSharedEXT structure{ global 2-component vector of float dummy, global 3-element array of 2-component vector of float submesh})
0:121 Constant:
0:121 1 (const int)
0:121 mod ( temp uint)
0:121 'gid' ( temp uint)
0:121 Constant:
0:121 2 (const uint)
0:123 MemoryBarrierShared ( global void)
0:123 Barrier ( global void)
0:126 EmitMeshTasksEXT ( global void)
0:126 Constant:
0:126 3 (const uint)
0:126 Constant:
0:126 1 (const uint)
0:126 Constant:
0:126 1 (const uint)
0:130 Function Definition: basic_works( ( global void)
0:130 Function Parameters:
0:132 Sequence
0:132 'gl_SubgroupSize' ( in uint SubgroupSize)
0:133 'gl_SubgroupInvocationID' ( in uint SubgroupInvocationID)
0:134 subgroupBarrier ( global void)
0:135 subgroupMemoryBarrier ( global void)
0:136 subgroupMemoryBarrierBuffer ( global void)
0:137 subgroupMemoryBarrierImage ( global void)
0:138 subgroupElect ( global bool)
0:139 'gl_NumSubgroups' ( in uint NumSubgroups)
0:140 'gl_SubgroupID' ( in uint SubgroupID)
0:141 subgroupMemoryBarrierShared ( global void)
0:145 Function Definition: ballot_works(vf4; ( global void)
0:145 Function Parameters:
0:145 'f4' ( in 4-component vector of float)
0:146 Sequence
0:146 'gl_SubgroupEqMask' ( in 4-component vector of uint SubgroupEqMask)
0:147 'gl_SubgroupGeMask' ( in 4-component vector of uint SubgroupGeMask)
0:148 'gl_SubgroupGtMask' ( in 4-component vector of uint SubgroupGtMask)
0:149 'gl_SubgroupLeMask' ( in 4-component vector of uint SubgroupLeMask)
0:150 'gl_SubgroupLtMask' ( in 4-component vector of uint SubgroupLtMask)
0:151 subgroupBroadcast ( global 4-component vector of float)
0:151 'f4' ( in 4-component vector of float)
0:151 Constant:
0:151 0 (const uint)
0:152 subgroupBroadcastFirst ( global 4-component vector of float)
0:152 'f4' ( in 4-component vector of float)
0:153 Sequence
0:153 move second child to first child ( temp 4-component vector of uint)
0:153 'ballot' ( temp 4-component vector of uint)
0:153 subgroupBallot ( global 4-component vector of uint)
0:153 Constant:
0:153 false (const bool)
0:154 subgroupInverseBallot ( global bool)
0:154 Constant:
0:154 1 (const uint)
0:154 1 (const uint)
0:154 1 (const uint)
0:154 1 (const uint)
0:155 subgroupBallotBitExtract ( global bool)
0:155 'ballot' ( temp 4-component vector of uint)
0:155 Constant:
0:155 0 (const uint)
0:156 subgroupBallotBitCount ( global uint)
0:156 'ballot' ( temp 4-component vector of uint)
0:157 subgroupBallotInclusiveBitCount ( global uint)
0:157 'ballot' ( temp 4-component vector of uint)
0:158 subgroupBallotExclusiveBitCount ( global uint)
0:158 'ballot' ( temp 4-component vector of uint)
0:159 subgroupBallotFindLSB ( global uint)
0:159 'ballot' ( temp 4-component vector of uint)
0:160 subgroupBallotFindMSB ( global uint)
0:160 'ballot' ( temp 4-component vector of uint)
0:164 Function Definition: vote_works(vf4; ( global void)
0:164 Function Parameters:
0:164 'f4' ( in 4-component vector of float)
0:166 Sequence
0:166 subgroupAll ( global bool)
0:166 Constant:
0:166 true (const bool)
0:167 subgroupAny ( global bool)
0:167 Constant:
0:167 false (const bool)
0:168 subgroupAllEqual ( global bool)
0:168 'f4' ( in 4-component vector of float)
0:173 Function Definition: shuffle_works(vf4; ( global void)
0:173 Function Parameters:
0:173 'f4' ( in 4-component vector of float)
0:175 Sequence
0:175 subgroupShuffle ( global 4-component vector of float)
0:175 'f4' ( in 4-component vector of float)
0:175 Constant:
0:175 0 (const uint)
0:176 subgroupShuffleXor ( global 4-component vector of float)
0:176 'f4' ( in 4-component vector of float)
0:176 Constant:
0:176 1 (const uint)
0:177 subgroupShuffleUp ( global 4-component vector of float)
0:177 'f4' ( in 4-component vector of float)
0:177 Constant:
0:177 1 (const uint)
0:178 subgroupShuffleDown ( global 4-component vector of float)
0:178 'f4' ( in 4-component vector of float)
0:178 Constant:
0:178 1 (const uint)
0:182 Function Definition: arith_works(vf4; ( global void)
0:182 Function Parameters:
0:182 'f4' ( in 4-component vector of float)
0:? Sequence
0:185 subgroupAdd ( global 4-component vector of float)
0:185 'f4' ( in 4-component vector of float)
0:186 subgroupMul ( global 4-component vector of float)
0:186 'f4' ( in 4-component vector of float)
0:187 subgroupMin ( global 4-component vector of float)
0:187 'f4' ( in 4-component vector of float)
0:188 subgroupMax ( global 4-component vector of float)
0:188 'f4' ( in 4-component vector of float)
0:189 subgroupAnd ( global 4-component vector of uint)
0:189 'ballot' ( temp 4-component vector of uint)
0:190 subgroupOr ( global 4-component vector of uint)
0:190 'ballot' ( temp 4-component vector of uint)
0:191 subgroupXor ( global 4-component vector of uint)
0:191 'ballot' ( temp 4-component vector of uint)
0:192 subgroupInclusiveAdd ( global 4-component vector of float)
0:192 'f4' ( in 4-component vector of float)
0:193 subgroupInclusiveMul ( global 4-component vector of float)
0:193 'f4' ( in 4-component vector of float)
0:194 subgroupInclusiveMin ( global 4-component vector of float)
0:194 'f4' ( in 4-component vector of float)
0:195 subgroupInclusiveMax ( global 4-component vector of float)
0:195 'f4' ( in 4-component vector of float)
0:196 subgroupInclusiveAnd ( global 4-component vector of uint)
0:196 'ballot' ( temp 4-component vector of uint)
0:197 subgroupInclusiveOr ( global 4-component vector of uint)
0:197 'ballot' ( temp 4-component vector of uint)
0:198 subgroupInclusiveXor ( global 4-component vector of uint)
0:198 'ballot' ( temp 4-component vector of uint)
0:199 subgroupExclusiveAdd ( global 4-component vector of float)
0:199 'f4' ( in 4-component vector of float)
0:200 subgroupExclusiveMul ( global 4-component vector of float)
0:200 'f4' ( in 4-component vector of float)
0:201 subgroupExclusiveMin ( global 4-component vector of float)
0:201 'f4' ( in 4-component vector of float)
0:202 subgroupExclusiveMax ( global 4-component vector of float)
0:202 'f4' ( in 4-component vector of float)
0:203 subgroupExclusiveAnd ( global 4-component vector of uint)
0:203 'ballot' ( temp 4-component vector of uint)
0:204 subgroupExclusiveOr ( global 4-component vector of uint)
0:204 'ballot' ( temp 4-component vector of uint)
0:205 subgroupExclusiveXor ( global 4-component vector of uint)
0:205 'ballot' ( temp 4-component vector of uint)
0:209 Function Definition: clustered_works(vf4; ( global void)
0:209 Function Parameters:
0:209 'f4' ( in 4-component vector of float)
0:211 Sequence
0:211 Sequence
0:211 move second child to first child ( temp 4-component vector of uint)
0:211 'ballot' ( temp 4-component vector of uint)
0:211 Constant:
0:211 85 (const uint)
0:211 0 (const uint)
0:211 0 (const uint)
0:211 0 (const uint)
0:212 subgroupClusteredAdd ( global 4-component vector of float)
0:212 'f4' ( in 4-component vector of float)
0:212 Constant:
0:212 2 (const uint)
0:213 subgroupClusteredMul ( global 4-component vector of float)
0:213 'f4' ( in 4-component vector of float)
0:213 Constant:
0:213 2 (const uint)
0:214 subgroupClusteredMin ( global 4-component vector of float)
0:214 'f4' ( in 4-component vector of float)
0:214 Constant:
0:214 2 (const uint)
0:215 subgroupClusteredMax ( global 4-component vector of float)
0:215 'f4' ( in 4-component vector of float)
0:215 Constant:
0:215 2 (const uint)
0:216 subgroupClusteredAnd ( global 4-component vector of uint)
0:216 'ballot' ( temp 4-component vector of uint)
0:216 Constant:
0:216 2 (const uint)
0:217 subgroupClusteredOr ( global 4-component vector of uint)
0:217 'ballot' ( temp 4-component vector of uint)
0:217 Constant:
0:217 2 (const uint)
0:218 subgroupClusteredXor ( global 4-component vector of uint)
0:218 'ballot' ( temp 4-component vector of uint)
0:218 Constant:
0:218 2 (const uint)
0:222 Function Definition: quad_works(vf4; ( global void)
0:222 Function Parameters:
0:222 'f4' ( in 4-component vector of float)
0:224 Sequence
0:224 subgroupQuadBroadcast ( global 4-component vector of float)
0:224 'f4' ( in 4-component vector of float)
0:224 Constant:
0:224 0 (const uint)
0:225 subgroupQuadSwapHorizontal ( global 4-component vector of float)
0:225 'f4' ( in 4-component vector of float)
0:226 subgroupQuadSwapVertical ( global 4-component vector of float)
0:226 'f4' ( in 4-component vector of float)
0:227 subgroupQuadSwapDiagonal ( global 4-component vector of float)
0:227 'f4' ( in 4-component vector of float)
0:? Linker Objects
0:? 'gl_WorkGroupSize' ( const 3-component vector of uint WorkGroupSize)
0:? 32 (const uint)
0:? 1 (const uint)
0:? 1 (const uint)
0:? 'uni_image' (layout( binding=0) writeonly uniform image2D)
0:? 'anon@0' (layout( column_major shared) uniform block{layout( column_major shared) uniform uint uni_value})
0:? 'mem' ( shared 10-element array of 4-component vector of float)
0:? 'mytask' ( taskPayloadSharedEXT structure{ global 2-component vector of float dummy, global 3-element array of 2-component vector of float submesh})
Linked task stage:
Shader version: 460
Requested GL_EXT_mesh_shader
Requested GL_KHR_shader_subgroup_arithmetic
Requested GL_KHR_shader_subgroup_ballot
Requested GL_KHR_shader_subgroup_basic
Requested GL_KHR_shader_subgroup_clustered
Requested GL_KHR_shader_subgroup_quad
Requested GL_KHR_shader_subgroup_shuffle
Requested GL_KHR_shader_subgroup_shuffle_relative
Requested GL_KHR_shader_subgroup_vote
local_size = (32, 1, 1)
ERROR: node is still EOpNull!
0:102 Function Definition: main( ( global void)
0:102 Function Parameters:
0:104 Sequence
0:104 Sequence
0:104 move second child to first child ( temp uint)
0:104 'iid' ( temp uint)
0:104 direct index ( temp uint)
0:104 'gl_LocalInvocationID' ( in 3-component vector of uint LocalInvocationID)
0:104 Constant:
0:104 0 (const int)
0:105 Sequence
0:105 move second child to first child ( temp uint)
0:105 'gid' ( temp uint)
0:105 direct index ( temp uint)
0:105 'gl_WorkGroupID' ( in 3-component vector of uint WorkGroupID)
0:105 Constant:
0:105 0 (const int)
0:108 Sequence
0:108 Sequence
0:108 move second child to first child ( temp uint)
0:108 'i' ( temp uint)
0:108 Constant:
0:108 0 (const uint)
0:108 Loop with condition tested first
0:108 Loop Condition
0:108 Compare Less Than ( temp bool)
0:108 'i' ( temp uint)
0:108 Constant:
0:108 10 (const uint)
0:108 Loop Body
0:109 Sequence
0:109 move second child to first child ( temp 4-component vector of float)
0:109 indirect index ( temp 4-component vector of float)
0:109 'mem' ( shared 10-element array of 4-component vector of float)
0:109 'i' ( temp uint)
0:109 Construct vec4 ( temp 4-component vector of float)
0:109 Convert uint to float ( temp float)
0:109 add ( temp uint)
0:109 'i' ( temp uint)
0:109 uni_value: direct index for structure (layout( column_major shared) uniform uint)
0:109 'anon@0' (layout( column_major shared) uniform block{layout( column_major shared) uniform uint uni_value})
0:109 Constant:
0:109 0 (const uint)
0:108 Loop Terminal Expression
0:108 Pre-Increment ( temp uint)
0:108 'i' ( temp uint)
0:111 imageStore ( global void)
0:111 'uni_image' (layout( binding=0) writeonly uniform image2D)
0:111 Construct ivec2 ( temp 2-component vector of int)
0:111 Convert uint to int ( temp int)
0:111 'iid' ( temp uint)
0:111 indirect index ( temp 4-component vector of float)
0:111 'mem' ( shared 10-element array of 4-component vector of float)
0:111 'gid' ( temp uint)
0:112 imageStore ( global void)
0:112 'uni_image' (layout( binding=0) writeonly uniform image2D)
0:112 Construct ivec2 ( temp 2-component vector of int)
0:112 Convert uint to int ( temp int)
0:112 'iid' ( temp uint)
0:112 indirect index ( temp 4-component vector of float)
0:112 'mem' ( shared 10-element array of 4-component vector of float)
0:112 add ( temp uint)
0:112 'gid' ( temp uint)
0:112 Constant:
0:112 1 (const uint)
0:114 MemoryBarrierShared ( global void)
0:114 Barrier ( global void)
0:118 move second child to first child ( temp 2-component vector of float)
0:118 dummy: direct index for structure ( global 2-component vector of float)
0:118 'mytask' ( taskPayloadSharedEXT structure{ global 2-component vector of float dummy, global 3-element array of 2-component vector of float submesh})
0:118 Constant:
0:118 0 (const int)
0:118 Constant:
0:118 30.000000
0:118 31.000000
0:119 move second child to first child ( temp 2-component vector of float)
0:119 direct index ( temp 2-component vector of float)
0:119 submesh: direct index for structure ( global 3-element array of 2-component vector of float)
0:119 'mytask' ( taskPayloadSharedEXT structure{ global 2-component vector of float dummy, global 3-element array of 2-component vector of float submesh})
0:119 Constant:
0:119 1 (const int)
0:119 Constant:
0:119 0 (const int)
0:119 Constant:
0:119 32.000000
0:119 33.000000
0:120 move second child to first child ( temp 2-component vector of float)
0:120 direct index ( temp 2-component vector of float)
0:120 submesh: direct index for structure ( global 3-element array of 2-component vector of float)
0:120 'mytask' ( taskPayloadSharedEXT structure{ global 2-component vector of float dummy, global 3-element array of 2-component vector of float submesh})
0:120 Constant:
0:120 1 (const int)
0:120 Constant:
0:120 1 (const int)
0:120 Constant:
0:120 34.000000
0:120 35.000000
0:121 move second child to first child ( temp 2-component vector of float)
0:121 direct index ( temp 2-component vector of float)
0:121 submesh: direct index for structure ( global 3-element array of 2-component vector of float)
0:121 'mytask' ( taskPayloadSharedEXT structure{ global 2-component vector of float dummy, global 3-element array of 2-component vector of float submesh})
0:121 Constant:
0:121 1 (const int)
0:121 Constant:
0:121 2 (const int)
0:121 indirect index ( temp 2-component vector of float)
0:121 submesh: direct index for structure ( global 3-element array of 2-component vector of float)
0:121 'mytask' ( taskPayloadSharedEXT structure{ global 2-component vector of float dummy, global 3-element array of 2-component vector of float submesh})
0:121 Constant:
0:121 1 (const int)
0:121 mod ( temp uint)
0:121 'gid' ( temp uint)
0:121 Constant:
0:121 2 (const uint)
0:123 MemoryBarrierShared ( global void)
0:123 Barrier ( global void)
0:126 EmitMeshTasksEXT ( global void)
0:126 Constant:
0:126 3 (const uint)
0:126 Constant:
0:126 1 (const uint)
0:126 Constant:
0:126 1 (const uint)
0:? Linker Objects
0:? 'gl_WorkGroupSize' ( const 3-component vector of uint WorkGroupSize)
0:? 32 (const uint)
0:? 1 (const uint)
0:? 1 (const uint)
0:? 'uni_image' (layout( binding=0) writeonly uniform image2D)
0:? 'anon@0' (layout( column_major shared) uniform block{layout( column_major shared) uniform uint uni_value})
0:? 'mem' ( shared 10-element array of 4-component vector of float)
0:? 'mytask' ( taskPayloadSharedEXT structure{ global 2-component vector of float dummy, global 3-element array of 2-component vector of float submesh})

448
Test/baseResults/spv.460.subgroupEXT.mesh.out Normal file
View File

@ -0,0 +1,448 @@
spv.460.subgroupEXT.mesh
// Module Version 10400
// Generated by (magic number): 8000a
// Id's are bound by 280
Capability ClipDistance
Capability CullDistance
Capability GroupNonUniform
Capability GroupNonUniformVote
Capability GroupNonUniformArithmetic
Capability GroupNonUniformBallot
Capability GroupNonUniformShuffle
Capability GroupNonUniformShuffleRelative
Capability GroupNonUniformClustered
Capability GroupNonUniformQuad
Capability FragmentShadingRateKHR
Capability MeshShadingEXT
Extension "SPV_EXT_mesh_shader"
Extension "SPV_KHR_fragment_shading_rate"
1: ExtInstImport "GLSL.std.450"
MemoryModel Logical GLSL450
EntryPoint MeshEXT 4 "main" 35 41 57 109 148 162 163 168 169 172 173 174 175 176
ExecutionMode 4 LocalSize 32 1 1
ExecutionMode 4 OutputVertices 81
ExecutionMode 4 OutputPrimitivesNV 32
ExecutionMode 4 OutputTrianglesNV
Source GLSL 460
SourceExtension "GL_EXT_mesh_shader"
SourceExtension "GL_KHR_shader_subgroup_arithmetic"
SourceExtension "GL_KHR_shader_subgroup_ballot"
SourceExtension "GL_KHR_shader_subgroup_basic"
SourceExtension "GL_KHR_shader_subgroup_clustered"
SourceExtension "GL_KHR_shader_subgroup_quad"
SourceExtension "GL_KHR_shader_subgroup_shuffle"
SourceExtension "GL_KHR_shader_subgroup_shuffle_relative"
SourceExtension "GL_KHR_shader_subgroup_vote"
Name 4 "main"
Name 6 "basic_works("
Name 13 "ballot_works(vf4;"
Name 12 "f4"
Name 16 "vote_works(vf4;"
Name 15 "f4"
Name 19 "shuffle_works(vf4;"
Name 18 "f4"
Name 22 "arith_works(vf4;"
Name 21 "f4"
Name 25 "clustered_works(vf4;"
Name 24 "f4"
Name 28 "quad_works(vf4;"
Name 27 "f4"
Name 32 "iid"
Name 35 "gl_LocalInvocationID"
Name 40 "gid"
Name 41 "gl_WorkGroupID"
Name 44 "vertexCount"
Name 46 "primitiveCount"
Name 54 "gl_MeshPerVertexEXT"
MemberName 54(gl_MeshPerVertexEXT) 0 "gl_Position"
MemberName 54(gl_MeshPerVertexEXT) 1 "gl_PointSize"
MemberName 54(gl_MeshPerVertexEXT) 2 "gl_ClipDistance"
MemberName 54(gl_MeshPerVertexEXT) 3 "gl_CullDistance"
Name 57 "gl_MeshVerticesEXT"
Name 106 "gl_MeshPerPrimitiveEXT"
MemberName 106(gl_MeshPerPrimitiveEXT) 0 "gl_PrimitiveID"
MemberName 106(gl_MeshPerPrimitiveEXT) 1 "gl_Layer"
MemberName 106(gl_MeshPerPrimitiveEXT) 2 "gl_ViewportIndex"
MemberName 106(gl_MeshPerPrimitiveEXT) 3 "gl_CullPrimitiveEXT"
MemberName 106(gl_MeshPerPrimitiveEXT) 4 "gl_PrimitiveShadingRateEXT"
Name 109 "gl_MeshPrimitivesEXT"
Name 148 "gl_PrimitiveTriangleIndicesEXT"
Name 162 "gl_SubgroupSize"
Name 163 "gl_SubgroupInvocationID"
Name 168 "gl_NumSubgroups"
Name 169 "gl_SubgroupID"
Name 172 "gl_SubgroupEqMask"
Name 173 "gl_SubgroupGeMask"
Name 174 "gl_SubgroupGtMask"
Name 175 "gl_SubgroupLeMask"
Name 176 "gl_SubgroupLtMask"
Name 182 "ballot"
Name 219 "ballot"
Name 254 "ballot"
Decorate 35(gl_LocalInvocationID) BuiltIn LocalInvocationId
Decorate 41(gl_WorkGroupID) BuiltIn WorkgroupId
MemberDecorate 54(gl_MeshPerVertexEXT) 0 BuiltIn Position
MemberDecorate 54(gl_MeshPerVertexEXT) 1 BuiltIn PointSize
MemberDecorate 54(gl_MeshPerVertexEXT) 2 BuiltIn ClipDistance
MemberDecorate 54(gl_MeshPerVertexEXT) 3 BuiltIn CullDistance
Decorate 54(gl_MeshPerVertexEXT) Block
MemberDecorate 106(gl_MeshPerPrimitiveEXT) 0 PerPrimitiveNV
MemberDecorate 106(gl_MeshPerPrimitiveEXT) 0 BuiltIn PrimitiveId
MemberDecorate 106(gl_MeshPerPrimitiveEXT) 1 PerPrimitiveNV
MemberDecorate 106(gl_MeshPerPrimitiveEXT) 1 BuiltIn Layer
MemberDecorate 106(gl_MeshPerPrimitiveEXT) 2 PerPrimitiveNV
MemberDecorate 106(gl_MeshPerPrimitiveEXT) 2 BuiltIn ViewportIndex
MemberDecorate 106(gl_MeshPerPrimitiveEXT) 3 PerPrimitiveNV
MemberDecorate 106(gl_MeshPerPrimitiveEXT) 3 BuiltIn CullPrimitiveEXT
MemberDecorate 106(gl_MeshPerPrimitiveEXT) 4 PerPrimitiveNV
MemberDecorate 106(gl_MeshPerPrimitiveEXT) 4 BuiltIn PrimitiveShadingRateKHR
Decorate 106(gl_MeshPerPrimitiveEXT) Block
Decorate 148(gl_PrimitiveTriangleIndicesEXT) BuiltIn PrimitiveTriangleIndicesEXT
Decorate 162(gl_SubgroupSize) RelaxedPrecision
Decorate 162(gl_SubgroupSize) BuiltIn SubgroupSize
Decorate 163(gl_SubgroupInvocationID) RelaxedPrecision
Decorate 163(gl_SubgroupInvocationID) BuiltIn SubgroupLocalInvocationId
Decorate 168(gl_NumSubgroups) BuiltIn NumSubgroups
Decorate 169(gl_SubgroupID) BuiltIn SubgroupId
Decorate 172(gl_SubgroupEqMask) BuiltIn SubgroupEqMaskKHR
Decorate 173(gl_SubgroupGeMask) BuiltIn SubgroupGeMaskKHR
Decorate 174(gl_SubgroupGtMask) BuiltIn SubgroupGtMaskKHR
Decorate 175(gl_SubgroupLeMask) BuiltIn SubgroupLeMaskKHR
Decorate 176(gl_SubgroupLtMask) BuiltIn SubgroupLtMaskKHR
Decorate 279 BuiltIn WorkgroupSize
2: TypeVoid
3: TypeFunction 2
8: TypeFloat 32
9: TypeVector 8(float) 4
10: TypePointer Function 9(fvec4)
11: TypeFunction 2 10(ptr)
30: TypeInt 32 0
31: TypePointer Function 30(int)
33: TypeVector 30(int) 3
34: TypePointer Input 33(ivec3)
35(gl_LocalInvocationID): 34(ptr) Variable Input
36: 30(int) Constant 0
37: TypePointer Input 30(int)
41(gl_WorkGroupID): 34(ptr) Variable Input
45: 30(int) Constant 81
47: 30(int) Constant 32
50: 30(int) Constant 4
51: TypeArray 8(float) 50
52: 30(int) Constant 3
53: TypeArray 8(float) 52
54(gl_MeshPerVertexEXT): TypeStruct 9(fvec4) 8(float) 51 53
55: TypeArray 54(gl_MeshPerVertexEXT) 45
56: TypePointer Output 55
57(gl_MeshVerticesEXT): 56(ptr) Variable Output
59: TypeInt 32 1
60: 59(int) Constant 0
61: 8(float) Constant 1065353216
62: 9(fvec4) ConstantComposite 61 61 61 61
63: TypePointer Output 9(fvec4)
66: 59(int) Constant 1
67: 8(float) Constant 1073741824
68: TypePointer Output 8(float)
71: 59(int) Constant 2
72: 59(int) Constant 3
73: 8(float) Constant 1077936128
76: 8(float) Constant 1082130432
78: 30(int) Constant 1
79: 30(int) Constant 264
80: 30(int) Constant 2
105: TypeBool
106(gl_MeshPerPrimitiveEXT): TypeStruct 59(int) 59(int) 59(int) 105(bool) 59(int)
107: TypeArray 106(gl_MeshPerPrimitiveEXT) 47
108: TypePointer Output 107
109(gl_MeshPrimitivesEXT): 108(ptr) Variable Output
111: 59(int) Constant 6
112: TypePointer Output 59(int)
115: 59(int) Constant 7
118: 59(int) Constant 8
121: 105(bool) ConstantFalse
122: TypePointer Output 105(bool)
145: 30(int) Constant 96
146: TypeArray 33(ivec3) 145
147: TypePointer Output 146
148(gl_PrimitiveTriangleIndicesEXT): 147(ptr) Variable Output
149: 33(ivec3) ConstantComposite 78 78 78
150: TypePointer Output 33(ivec3)
154: 33(ivec3) ConstantComposite 80 80 80
162(gl_SubgroupSize): 37(ptr) Variable Input
163(gl_SubgroupInvocationID): 37(ptr) Variable Input
164: 30(int) Constant 3400
165: 30(int) Constant 72
166: 30(int) Constant 2056
168(gl_NumSubgroups): 37(ptr) Variable Input
169(gl_SubgroupID): 37(ptr) Variable Input
170: TypeVector 30(int) 4
171: TypePointer Input 170(ivec4)
172(gl_SubgroupEqMask): 171(ptr) Variable Input
173(gl_SubgroupGeMask): 171(ptr) Variable Input
174(gl_SubgroupGtMask): 171(ptr) Variable Input
175(gl_SubgroupLeMask): 171(ptr) Variable Input
176(gl_SubgroupLtMask): 171(ptr) Variable Input
181: TypePointer Function 170(ivec4)
184: 170(ivec4) ConstantComposite 78 78 78 78
198: 105(bool) ConstantTrue
255: 30(int) Constant 85
256: 170(ivec4) ConstantComposite 255 36 36 36
279: 33(ivec3) ConstantComposite 47 78 78
4(main): 2 Function None 3
5: Label
32(iid): 31(ptr) Variable Function
40(gid): 31(ptr) Variable Function
44(vertexCount): 31(ptr) Variable Function
46(primitiveCount): 31(ptr) Variable Function
38: 37(ptr) AccessChain 35(gl_LocalInvocationID) 36
39: 30(int) Load 38
Store 32(iid) 39
42: 37(ptr) AccessChain 41(gl_WorkGroupID) 36
43: 30(int) Load 42
Store 40(gid) 43
Store 44(vertexCount) 45
Store 46(primitiveCount) 47
48: 30(int) Load 44(vertexCount)
49: 30(int) Load 46(primitiveCount)
SetMeshOutputsEXT 48 49
58: 30(int) Load 32(iid)
64: 63(ptr) AccessChain 57(gl_MeshVerticesEXT) 58 60
Store 64 62
65: 30(int) Load 32(iid)
69: 68(ptr) AccessChain 57(gl_MeshVerticesEXT) 65 66
Store 69 67
70: 30(int) Load 32(iid)
74: 68(ptr) AccessChain 57(gl_MeshVerticesEXT) 70 71 72
Store 74 73
75: 30(int) Load 32(iid)
77: 68(ptr) AccessChain 57(gl_MeshVerticesEXT) 75 72 71
Store 77 76
MemoryBarrier 78 79
ControlBarrier 80 80 79
81: 30(int) Load 32(iid)
82: 30(int) IAdd 81 78
83: 30(int) Load 32(iid)
84: 63(ptr) AccessChain 57(gl_MeshVerticesEXT) 83 60
85: 9(fvec4) Load 84
86: 63(ptr) AccessChain 57(gl_MeshVerticesEXT) 82 60
Store 86 85
87: 30(int) Load 32(iid)
88: 30(int) IAdd 87 78
89: 30(int) Load 32(iid)
90: 68(ptr) AccessChain 57(gl_MeshVerticesEXT) 89 66
91: 8(float) Load 90
92: 68(ptr) AccessChain 57(gl_MeshVerticesEXT) 88 66
Store 92 91
93: 30(int) Load 32(iid)
94: 30(int) IAdd 93 78
95: 30(int) Load 32(iid)
96: 68(ptr) AccessChain 57(gl_MeshVerticesEXT) 95 71 72
97: 8(float) Load 96
98: 68(ptr) AccessChain 57(gl_MeshVerticesEXT) 94 71 72
Store 98 97
99: 30(int) Load 32(iid)
100: 30(int) IAdd 99 78
101: 30(int) Load 32(iid)
102: 68(ptr) AccessChain 57(gl_MeshVerticesEXT) 101 72 71
103: 8(float) Load 102
104: 68(ptr) AccessChain 57(gl_MeshVerticesEXT) 100 72 71
Store 104 103
MemoryBarrier 78 79
ControlBarrier 80 80 79
110: 30(int) Load 32(iid)
113: 112(ptr) AccessChain 109(gl_MeshPrimitivesEXT) 110 60
Store 113 111
114: 30(int) Load 32(iid)
116: 112(ptr) AccessChain 109(gl_MeshPrimitivesEXT) 114 66
Store 116 115
117: 30(int) Load 32(iid)
119: 112(ptr) AccessChain 109(gl_MeshPrimitivesEXT) 117 71
Store 119 118
120: 30(int) Load 32(iid)
123: 122(ptr) AccessChain 109(gl_MeshPrimitivesEXT) 120 72
Store 123 121
MemoryBarrier 78 79
ControlBarrier 80 80 79
124: 30(int) Load 32(iid)
125: 30(int) IAdd 124 78
126: 30(int) Load 32(iid)
127: 112(ptr) AccessChain 109(gl_MeshPrimitivesEXT) 126 60
128: 59(int) Load 127
129: 112(ptr) AccessChain 109(gl_MeshPrimitivesEXT) 125 60
Store 129 128
130: 30(int) Load 32(iid)
131: 30(int) IAdd 130 78
132: 30(int) Load 32(iid)
133: 112(ptr) AccessChain 109(gl_MeshPrimitivesEXT) 132 66
134: 59(int) Load 133
135: 112(ptr) AccessChain 109(gl_MeshPrimitivesEXT) 131 66
Store 135 134
136: 30(int) Load 32(iid)
137: 30(int) IAdd 136 78
138: 30(int) Load 32(iid)
139: 112(ptr) AccessChain 109(gl_MeshPrimitivesEXT) 138 71
140: 59(int) Load 139
141: 112(ptr) AccessChain 109(gl_MeshPrimitivesEXT) 137 71
Store 141 140
142: 30(int) Load 32(iid)
143: 30(int) IAdd 142 78
144: 122(ptr) AccessChain 109(gl_MeshPrimitivesEXT) 143 72
Store 144 121
MemoryBarrier 78 79
ControlBarrier 80 80 79
151: 150(ptr) AccessChain 148(gl_PrimitiveTriangleIndicesEXT) 60
Store 151 149
152: 30(int) Load 46(primitiveCount)
153: 30(int) ISub 152 78
155: 150(ptr) AccessChain 148(gl_PrimitiveTriangleIndicesEXT) 153
Store 155 154
156: 30(int) Load 40(gid)
157: 30(int) Load 40(gid)
158: 30(int) ISub 157 78
159: 150(ptr) AccessChain 148(gl_PrimitiveTriangleIndicesEXT) 158
160: 33(ivec3) Load 159
161: 150(ptr) AccessChain 148(gl_PrimitiveTriangleIndicesEXT) 156
Store 161 160
MemoryBarrier 78 79
ControlBarrier 80 80 79
Return
FunctionEnd
6(basic_works(): 2 Function None 3
7: Label
ControlBarrier 52 52 164
MemoryBarrier 52 164
MemoryBarrier 52 165
MemoryBarrier 52 166
167: 105(bool) GroupNonUniformElect 52
MemoryBarrier 52 79
Return
FunctionEnd
13(ballot_works(vf4;): 2 Function None 11
12(f4): 10(ptr) FunctionParameter
14: Label
182(ballot): 181(ptr) Variable Function
177: 9(fvec4) Load 12(f4)
178: 9(fvec4) GroupNonUniformBroadcast 52 177 36
179: 9(fvec4) Load 12(f4)
180: 9(fvec4) GroupNonUniformBroadcastFirst 52 179
183: 170(ivec4) GroupNonUniformBallot 52 121
Store 182(ballot) 183
185: 105(bool) GroupNonUniformInverseBallot 52 184
186: 170(ivec4) Load 182(ballot)
187: 105(bool) GroupNonUniformBallotBitExtract 52 186 36
188: 170(ivec4) Load 182(ballot)
189: 30(int) GroupNonUniformBallotBitCount 52 Reduce 188
190: 170(ivec4) Load 182(ballot)
191: 30(int) GroupNonUniformBallotBitCount 52 InclusiveScan 190
192: 170(ivec4) Load 182(ballot)
193: 30(int) GroupNonUniformBallotBitCount 52 ExclusiveScan 192
194: 170(ivec4) Load 182(ballot)
195: 30(int) GroupNonUniformBallotFindLSB 52 194
196: 170(ivec4) Load 182(ballot)
197: 30(int) GroupNonUniformBallotFindMSB 52 196
Return
FunctionEnd
16(vote_works(vf4;): 2 Function None 11
15(f4): 10(ptr) FunctionParameter
17: Label
199: 105(bool) GroupNonUniformAll 52 198
200: 105(bool) GroupNonUniformAny 52 121
201: 9(fvec4) Load 15(f4)
202: 105(bool) GroupNonUniformAllEqual 52 201
Return
FunctionEnd
19(shuffle_works(vf4;): 2 Function None 11
18(f4): 10(ptr) FunctionParameter
20: Label
203: 9(fvec4) Load 18(f4)
204: 9(fvec4) GroupNonUniformShuffle 52 203 36
205: 9(fvec4) Load 18(f4)
206: 9(fvec4) GroupNonUniformShuffleXor 52 205 78
207: 9(fvec4) Load 18(f4)
208: 9(fvec4) GroupNonUniformShuffleUp 52 207 78
209: 9(fvec4) Load 18(f4)
210: 9(fvec4) GroupNonUniformShuffleDown 52 209 78
Return
FunctionEnd
22(arith_works(vf4;): 2 Function None 11
21(f4): 10(ptr) FunctionParameter
23: Label
219(ballot): 181(ptr) Variable Function
211: 9(fvec4) Load 21(f4)
212: 9(fvec4) GroupNonUniformFAdd 52 Reduce 211
213: 9(fvec4) Load 21(f4)
214: 9(fvec4) GroupNonUniformFMul 52 Reduce 213
215: 9(fvec4) Load 21(f4)
216: 9(fvec4) GroupNonUniformFMin 52 Reduce 215
217: 9(fvec4) Load 21(f4)
218: 9(fvec4) GroupNonUniformFMax 52 Reduce 217
220: 170(ivec4) Load 219(ballot)
221: 170(ivec4) GroupNonUniformBitwiseAnd 52 Reduce 220
222: 170(ivec4) Load 219(ballot)
223: 170(ivec4) GroupNonUniformBitwiseOr 52 Reduce 222
224: 170(ivec4) Load 219(ballot)
225: 170(ivec4) GroupNonUniformBitwiseXor 52 Reduce 224
226: 9(fvec4) Load 21(f4)
227: 9(fvec4) GroupNonUniformFAdd 52 InclusiveScan 226
228: 9(fvec4) Load 21(f4)
229: 9(fvec4) GroupNonUniformFMul 52 InclusiveScan 228
230: 9(fvec4) Load 21(f4)
231: 9(fvec4) GroupNonUniformFMin 52 InclusiveScan 230
232: 9(fvec4) Load 21(f4)
233: 9(fvec4) GroupNonUniformFMax 52 InclusiveScan 232
234: 170(ivec4) Load 219(ballot)
235: 170(ivec4) GroupNonUniformBitwiseAnd 52 InclusiveScan 234
236: 170(ivec4) Load 219(ballot)
237: 170(ivec4) GroupNonUniformBitwiseOr 52 InclusiveScan 236
238: 170(ivec4) Load 219(ballot)
239: 170(ivec4) GroupNonUniformBitwiseXor 52 InclusiveScan 238
240: 9(fvec4) Load 21(f4)
241: 9(fvec4) GroupNonUniformFAdd 52 ExclusiveScan 240
242: 9(fvec4) Load 21(f4)
243: 9(fvec4) GroupNonUniformFMul 52 ExclusiveScan 242
244: 9(fvec4) Load 21(f4)
245: 9(fvec4) GroupNonUniformFMin 52 ExclusiveScan 244
246: 9(fvec4) Load 21(f4)
247: 9(fvec4) GroupNonUniformFMax 52 ExclusiveScan 246
248: 170(ivec4) Load 219(ballot)
249: 170(ivec4) GroupNonUniformBitwiseAnd 52 ExclusiveScan 248
250: 170(ivec4) Load 219(ballot)
251: 170(ivec4) GroupNonUniformBitwiseOr 52 ExclusiveScan 250
252: 170(ivec4) Load 219(ballot)
253: 170(ivec4) GroupNonUniformBitwiseXor 52 ExclusiveScan 252
Return
FunctionEnd
25(clustered_works(vf4;): 2 Function None 11
24(f4): 10(ptr) FunctionParameter
26: Label
254(ballot): 181(ptr) Variable Function
Store 254(ballot) 256
257: 9(fvec4) Load 24(f4)
258: 9(fvec4) GroupNonUniformFAdd 52 ClusteredReduce 257 80
259: 9(fvec4) Load 24(f4)
260: 9(fvec4) GroupNonUniformFMul 52 ClusteredReduce 259 80
261: 9(fvec4) Load 24(f4)
262: 9(fvec4) GroupNonUniformFMin 52 ClusteredReduce 261 80
263: 9(fvec4) Load 24(f4)
264: 9(fvec4) GroupNonUniformFMax 52 ClusteredReduce 263 80
265: 170(ivec4) Load 254(ballot)
266: 170(ivec4) GroupNonUniformBitwiseAnd 52 ClusteredReduce 265 80
267: 170(ivec4) Load 254(ballot)
268: 170(ivec4) GroupNonUniformBitwiseOr 52 ClusteredReduce 267 80
269: 170(ivec4) Load 254(ballot)
270: 170(ivec4) GroupNonUniformBitwiseXor 52 ClusteredReduce 269 80
Return
FunctionEnd
28(quad_works(vf4;): 2 Function None 11
27(f4): 10(ptr) FunctionParameter
29: Label
271: 9(fvec4) Load 27(f4)
272: 9(fvec4) GroupNonUniformQuadBroadcast 52 271 36
273: 9(fvec4) Load 27(f4)
274: 9(fvec4) GroupNonUniformQuadSwap 52 273 36
275: 9(fvec4) Load 27(f4)
276: 9(fvec4) GroupNonUniformQuadSwap 52 275 78
277: 9(fvec4) Load 27(f4)
278: 9(fvec4) GroupNonUniformQuadSwap 52 277 80
Return
FunctionEnd

376
Test/baseResults/spv.460.subgroupEXT.task.out Normal file
View File

@ -0,0 +1,376 @@
spv.460.subgroupEXT.task
// Module Version 10400
// Generated by (magic number): 8000a
// Id's are bound by 242
Capability StorageImageWriteWithoutFormat
Capability GroupNonUniform
Capability GroupNonUniformVote
Capability GroupNonUniformArithmetic
Capability GroupNonUniformBallot
Capability GroupNonUniformShuffle
Capability GroupNonUniformShuffleRelative
Capability GroupNonUniformClustered
Capability GroupNonUniformQuad
Capability MeshShadingEXT
Extension "SPV_EXT_mesh_shader"
1: ExtInstImport "GLSL.std.450"
MemoryModel Logical GLSL450
EntryPoint TaskEXT 4 "main" 35 41 56 61 77 102 122 123 128 129 132 133 134 135 136
ExecutionMode 4 LocalSize 32 1 1
Source GLSL 460
SourceExtension "GL_EXT_mesh_shader"
SourceExtension "GL_KHR_shader_subgroup_arithmetic"
SourceExtension "GL_KHR_shader_subgroup_ballot"
SourceExtension "GL_KHR_shader_subgroup_basic"
SourceExtension "GL_KHR_shader_subgroup_clustered"
SourceExtension "GL_KHR_shader_subgroup_quad"
SourceExtension "GL_KHR_shader_subgroup_shuffle"
SourceExtension "GL_KHR_shader_subgroup_shuffle_relative"
SourceExtension "GL_KHR_shader_subgroup_vote"
Name 4 "main"
Name 6 "basic_works("
Name 13 "ballot_works(vf4;"
Name 12 "f4"
Name 16 "vote_works(vf4;"
Name 15 "f4"
Name 19 "shuffle_works(vf4;"
Name 18 "f4"
Name 22 "arith_works(vf4;"
Name 21 "f4"
Name 25 "clustered_works(vf4;"
Name 24 "f4"
Name 28 "quad_works(vf4;"
Name 27 "f4"
Name 32 "iid"
Name 35 "gl_LocalInvocationID"
Name 40 "gid"
Name 41 "gl_WorkGroupID"
Name 44 "i"
Name 56 "mem"
Name 59 "block0"
MemberName 59(block0) 0 "uni_value"
Name 61 ""
Name 77 "uni_image"
Name 100 "Task"
MemberName 100(Task) 0 "dummy"
MemberName 100(Task) 1 "submesh"
Name 102 "mytask"
Name 122 "gl_SubgroupSize"
Name 123 "gl_SubgroupInvocationID"
Name 128 "gl_NumSubgroups"
Name 129 "gl_SubgroupID"
Name 132 "gl_SubgroupEqMask"
Name 133 "gl_SubgroupGeMask"
Name 134 "gl_SubgroupGtMask"
Name 135 "gl_SubgroupLeMask"
Name 136 "gl_SubgroupLtMask"
Name 142 "ballot"
Name 180 "ballot"
Name 215 "ballot"
Decorate 35(gl_LocalInvocationID) BuiltIn LocalInvocationId
Decorate 41(gl_WorkGroupID) BuiltIn WorkgroupId
MemberDecorate 59(block0) 0 Offset 0
Decorate 59(block0) Block
Decorate 61 DescriptorSet 0
Decorate 61 Binding 1
Decorate 77(uni_image) DescriptorSet 0
Decorate 77(uni_image) Binding 0
Decorate 77(uni_image) NonReadable
Decorate 122(gl_SubgroupSize) RelaxedPrecision
Decorate 122(gl_SubgroupSize) BuiltIn SubgroupSize
Decorate 123(gl_SubgroupInvocationID) RelaxedPrecision
Decorate 123(gl_SubgroupInvocationID) BuiltIn SubgroupLocalInvocationId
Decorate 128(gl_NumSubgroups) BuiltIn NumSubgroups
Decorate 129(gl_SubgroupID) BuiltIn SubgroupId
Decorate 132(gl_SubgroupEqMask) BuiltIn SubgroupEqMaskKHR
Decorate 133(gl_SubgroupGeMask) BuiltIn SubgroupGeMaskKHR
Decorate 134(gl_SubgroupGtMask) BuiltIn SubgroupGtMaskKHR
Decorate 135(gl_SubgroupLeMask) BuiltIn SubgroupLeMaskKHR
Decorate 136(gl_SubgroupLtMask) BuiltIn SubgroupLtMaskKHR
Decorate 241 BuiltIn WorkgroupSize
2: TypeVoid
3: TypeFunction 2
8: TypeFloat 32
9: TypeVector 8(float) 4
10: TypePointer Function 9(fvec4)
11: TypeFunction 2 10(ptr)
30: TypeInt 32 0
31: TypePointer Function 30(int)
33: TypeVector 30(int) 3
34: TypePointer Input 33(ivec3)
35(gl_LocalInvocationID): 34(ptr) Variable Input
36: 30(int) Constant 0
37: TypePointer Input 30(int)
41(gl_WorkGroupID): 34(ptr) Variable Input
51: 30(int) Constant 10
52: TypeBool
54: TypeArray 9(fvec4) 51
55: TypePointer Workgroup 54
56(mem): 55(ptr) Variable Workgroup
59(block0): TypeStruct 30(int)
60: TypePointer Uniform 59(block0)
61: 60(ptr) Variable Uniform
62: TypeInt 32 1
63: 62(int) Constant 0
64: TypePointer Uniform 30(int)
70: TypePointer Workgroup 9(fvec4)
73: 62(int) Constant 1
75: TypeImage 8(float) 2D nonsampled format:Unknown
76: TypePointer UniformConstant 75
77(uni_image): 76(ptr) Variable UniformConstant
81: TypeVector 62(int) 2
91: 30(int) Constant 1
95: 30(int) Constant 264
96: 30(int) Constant 2
97: TypeVector 8(float) 2
98: 30(int) Constant 3
99: TypeArray 97(fvec2) 98
100(Task): TypeStruct 97(fvec2) 99
101: TypePointer TaskPayloadWorkgroupEXT 100(Task)
102(mytask): 101(ptr) Variable TaskPayloadWorkgroupEXT
103: 8(float) Constant 1106247680
104: 8(float) Constant 1106771968
105: 97(fvec2) ConstantComposite 103 104
106: TypePointer TaskPayloadWorkgroupEXT 97(fvec2)
108: 8(float) Constant 1107296256
109: 8(float) Constant 1107558400
110: 97(fvec2) ConstantComposite 108 109
112: 8(float) Constant 1107820544
113: 8(float) Constant 1108082688
114: 97(fvec2) ConstantComposite 112 113
116: 62(int) Constant 2
122(gl_SubgroupSize): 37(ptr) Variable Input
123(gl_SubgroupInvocationID): 37(ptr) Variable Input
124: 30(int) Constant 3400
125: 30(int) Constant 72
126: 30(int) Constant 2056
128(gl_NumSubgroups): 37(ptr) Variable Input
129(gl_SubgroupID): 37(ptr) Variable Input
130: TypeVector 30(int) 4
131: TypePointer Input 130(ivec4)
132(gl_SubgroupEqMask): 131(ptr) Variable Input
133(gl_SubgroupGeMask): 131(ptr) Variable Input
134(gl_SubgroupGtMask): 131(ptr) Variable Input
135(gl_SubgroupLeMask): 131(ptr) Variable Input
136(gl_SubgroupLtMask): 131(ptr) Variable Input
141: TypePointer Function 130(ivec4)
143: 52(bool) ConstantFalse
145: 130(ivec4) ConstantComposite 91 91 91 91
159: 52(bool) ConstantTrue
216: 30(int) Constant 85
217: 130(ivec4) ConstantComposite 216 36 36 36
240: 30(int) Constant 32
241: 33(ivec3) ConstantComposite 240 91 91
4(main): 2 Function None 3
5: Label
32(iid): 31(ptr) Variable Function
40(gid): 31(ptr) Variable Function
44(i): 31(ptr) Variable Function
38: 37(ptr) AccessChain 35(gl_LocalInvocationID) 36
39: 30(int) Load 38
Store 32(iid) 39
42: 37(ptr) AccessChain 41(gl_WorkGroupID) 36
43: 30(int) Load 42
Store 40(gid) 43
Store 44(i) 36
Branch 45
45: Label
LoopMerge 47 48 None
Branch 49
49: Label
50: 30(int) Load 44(i)
53: 52(bool) ULessThan 50 51
BranchConditional 53 46 47
46: Label
57: 30(int) Load 44(i)
58: 30(int) Load 44(i)
65: 64(ptr) AccessChain 61 63
66: 30(int) Load 65
67: 30(int) IAdd 58 66
68: 8(float) ConvertUToF 67
69: 9(fvec4) CompositeConstruct 68 68 68 68
71: 70(ptr) AccessChain 56(mem) 57
Store 71 69
Branch 48
48: Label
72: 30(int) Load 44(i)
74: 30(int) IAdd 72 73
Store 44(i) 74
Branch 45
47: Label
78: 75 Load 77(uni_image)
79: 30(int) Load 32(iid)
80: 62(int) Bitcast 79
82: 81(ivec2) CompositeConstruct 80 80
83: 30(int) Load 40(gid)
84: 70(ptr) AccessChain 56(mem) 83
85: 9(fvec4) Load 84
ImageWrite 78 82 85
86: 75 Load 77(uni_image)
87: 30(int) Load 32(iid)
88: 62(int) Bitcast 87
89: 81(ivec2) CompositeConstruct 88 88
90: 30(int) Load 40(gid)
92: 30(int) IAdd 90 91
93: 70(ptr) AccessChain 56(mem) 92
94: 9(fvec4) Load 93
ImageWrite 86 89 94
MemoryBarrier 91 95
ControlBarrier 96 96 95
107: 106(ptr) AccessChain 102(mytask) 63
Store 107 105
111: 106(ptr) AccessChain 102(mytask) 73 63
Store 111 110
115: 106(ptr) AccessChain 102(mytask) 73 73
Store 115 114
117: 30(int) Load 40(gid)
118: 30(int) UMod 117 96
119: 106(ptr) AccessChain 102(mytask) 73 118
120: 97(fvec2) Load 119
121: 106(ptr) AccessChain 102(mytask) 73 116
Store 121 120
MemoryBarrier 91 95
ControlBarrier 96 96 95
EmitMeshTasksEXT 98 91 91 102(mytask)
Return
FunctionEnd
6(basic_works(): 2 Function None 3
7: Label
ControlBarrier 98 98 124
MemoryBarrier 98 124
MemoryBarrier 98 125
MemoryBarrier 98 126
127: 52(bool) GroupNonUniformElect 98
MemoryBarrier 98 95
Return
FunctionEnd
13(ballot_works(vf4;): 2 Function None 11
12(f4): 10(ptr) FunctionParameter
14: Label
142(ballot): 141(ptr) Variable Function
137: 9(fvec4) Load 12(f4)
138: 9(fvec4) GroupNonUniformBroadcast 98 137 36
139: 9(fvec4) Load 12(f4)
140: 9(fvec4) GroupNonUniformBroadcastFirst 98 139
144: 130(ivec4) GroupNonUniformBallot 98 143
Store 142(ballot) 144
146: 52(bool) GroupNonUniformInverseBallot 98 145
147: 130(ivec4) Load 142(ballot)
148: 52(bool) GroupNonUniformBallotBitExtract 98 147 36
149: 130(ivec4) Load 142(ballot)
150: 30(int) GroupNonUniformBallotBitCount 98 Reduce 149
151: 130(ivec4) Load 142(ballot)
152: 30(int) GroupNonUniformBallotBitCount 98 InclusiveScan 151
153: 130(ivec4) Load 142(ballot)
154: 30(int) GroupNonUniformBallotBitCount 98 ExclusiveScan 153
155: 130(ivec4) Load 142(ballot)
156: 30(int) GroupNonUniformBallotFindLSB 98 155
157: 130(ivec4) Load 142(ballot)
158: 30(int) GroupNonUniformBallotFindMSB 98 157
Return
FunctionEnd
16(vote_works(vf4;): 2 Function None 11
15(f4): 10(ptr) FunctionParameter
17: Label
160: 52(bool) GroupNonUniformAll 98 159
161: 52(bool) GroupNonUniformAny 98 143
162: 9(fvec4) Load 15(f4)
163: 52(bool) GroupNonUniformAllEqual 98 162
Return
FunctionEnd
19(shuffle_works(vf4;): 2 Function None 11
18(f4): 10(ptr) FunctionParameter
20: Label
164: 9(fvec4) Load 18(f4)
165: 9(fvec4) GroupNonUniformShuffle 98 164 36
166: 9(fvec4) Load 18(f4)
167: 9(fvec4) GroupNonUniformShuffleXor 98 166 91
168: 9(fvec4) Load 18(f4)
169: 9(fvec4) GroupNonUniformShuffleUp 98 168 91
170: 9(fvec4) Load 18(f4)
171: 9(fvec4) GroupNonUniformShuffleDown 98 170 91
Return
FunctionEnd
22(arith_works(vf4;): 2 Function None 11
21(f4): 10(ptr) FunctionParameter
23: Label
180(ballot): 141(ptr) Variable Function
172: 9(fvec4) Load 21(f4)
173: 9(fvec4) GroupNonUniformFAdd 98 Reduce 172
174: 9(fvec4) Load 21(f4)
175: 9(fvec4) GroupNonUniformFMul 98 Reduce 174
176: 9(fvec4) Load 21(f4)
177: 9(fvec4) GroupNonUniformFMin 98 Reduce 176
178: 9(fvec4) Load 21(f4)
179: 9(fvec4) GroupNonUniformFMax 98 Reduce 178
181: 130(ivec4) Load 180(ballot)
182: 130(ivec4) GroupNonUniformBitwiseAnd 98 Reduce 181
183: 130(ivec4) Load 180(ballot)
184: 130(ivec4) GroupNonUniformBitwiseOr 98 Reduce 183
185: 130(ivec4) Load 180(ballot)
186: 130(ivec4) GroupNonUniformBitwiseXor 98 Reduce 185
187: 9(fvec4) Load 21(f4)
188: 9(fvec4) GroupNonUniformFAdd 98 InclusiveScan 187
189: 9(fvec4) Load 21(f4)
190: 9(fvec4) GroupNonUniformFMul 98 InclusiveScan 189
191: 9(fvec4) Load 21(f4)
192: 9(fvec4) GroupNonUniformFMin 98 InclusiveScan 191
193: 9(fvec4) Load 21(f4)
194: 9(fvec4) GroupNonUniformFMax 98 InclusiveScan 193
195: 130(ivec4) Load 180(ballot)
196: 130(ivec4) GroupNonUniformBitwiseAnd 98 InclusiveScan 195
197: 130(ivec4) Load 180(ballot)
198: 130(ivec4) GroupNonUniformBitwiseOr 98 InclusiveScan 197
199: 130(ivec4) Load 180(ballot)
200: 130(ivec4) GroupNonUniformBitwiseXor 98 InclusiveScan 199
201: 9(fvec4) Load 21(f4)
202: 9(fvec4) GroupNonUniformFAdd 98 ExclusiveScan 201
203: 9(fvec4) Load 21(f4)
204: 9(fvec4) GroupNonUniformFMul 98 ExclusiveScan 203
205: 9(fvec4) Load 21(f4)
206: 9(fvec4) GroupNonUniformFMin 98 ExclusiveScan 205
207: 9(fvec4) Load 21(f4)
208: 9(fvec4) GroupNonUniformFMax 98 ExclusiveScan 207
209: 130(ivec4) Load 180(ballot)
210: 130(ivec4) GroupNonUniformBitwiseAnd 98 ExclusiveScan 209
211: 130(ivec4) Load 180(ballot)
212: 130(ivec4) GroupNonUniformBitwiseOr 98 ExclusiveScan 211
213: 130(ivec4) Load 180(ballot)
214: 130(ivec4) GroupNonUniformBitwiseXor 98 ExclusiveScan 213
Return
FunctionEnd
25(clustered_works(vf4;): 2 Function None 11
24(f4): 10(ptr) FunctionParameter
26: Label
215(ballot): 141(ptr) Variable Function
Store 215(ballot) 217
218: 9(fvec4) Load 24(f4)
219: 9(fvec4) GroupNonUniformFAdd 98 ClusteredReduce 218 96
220: 9(fvec4) Load 24(f4)
221: 9(fvec4) GroupNonUniformFMul 98 ClusteredReduce 220 96
222: 9(fvec4) Load 24(f4)
223: 9(fvec4) GroupNonUniformFMin 98 ClusteredReduce 222 96
224: 9(fvec4) Load 24(f4)
225: 9(fvec4) GroupNonUniformFMax 98 ClusteredReduce 224 96
226: 130(ivec4) Load 215(ballot)
227: 130(ivec4) GroupNonUniformBitwiseAnd 98 ClusteredReduce 226 96
228: 130(ivec4) Load 215(ballot)
229: 130(ivec4) GroupNonUniformBitwiseOr 98 ClusteredReduce 228 96
230: 130(ivec4) Load 215(ballot)
231: 130(ivec4) GroupNonUniformBitwiseXor 98 ClusteredReduce 230 96
Return
FunctionEnd
28(quad_works(vf4;): 2 Function None 11
27(f4): 10(ptr) FunctionParameter
29: Label
232: 9(fvec4) Load 27(f4)
233: 9(fvec4) GroupNonUniformQuadBroadcast 98 232 36
234: 9(fvec4) Load 27(f4)
235: 9(fvec4) GroupNonUniformQuadSwap 98 234 36
236: 9(fvec4) Load 27(f4)
237: 9(fvec4) GroupNonUniformQuadSwap 98 236 91
238: 9(fvec4) Load 27(f4)
239: 9(fvec4) GroupNonUniformQuadSwap 98 238 96
Return
FunctionEnd

7
Test/baseResults/spv.atomiAddEXT.error.mesh.out Normal file
View File

@ -0,0 +1,7 @@
spv.atomiAddEXT.error.mesh
ERROR: 0:21: 'assign' : l-value required "mytask" (can't modify variable with storage qualifier taskPayloadSharedEXT in mesh shaders)
ERROR: 0:21: 'out' : Non-L-value cannot be passed for 'out' or 'inout' parameters.
ERROR: 2 compilation errors. No code generated.
SPIR-V is not generated for failed compile or link

71
Test/baseResults/spv.atomiAddEXT.task.out Normal file
View File

@ -0,0 +1,71 @@
spv.atomiAddEXT.task
// Module Version 10400
// Generated by (magic number): 8000a
// Id's are bound by 34
Capability MeshShadingEXT
Extension "SPV_EXT_mesh_shader"
1: ExtInstImport "GLSL.std.450"
MemoryModel Logical GLSL450
EntryPoint TaskEXT 4 "main" 9 23 28
ExecutionMode 4 LocalSize 1 1 1
Source GLSL 460
SourceExtension "GL_EXT_mesh_shader"
Name 4 "main"
Name 7 "Buffer"
MemberName 7(Buffer) 0 "x"
Name 9 ""
Name 20 "structType"
MemberName 20(structType) 0 "y"
Name 21 "t2"
MemberName 21(t2) 0 "f"
Name 23 "t"
Name 26 "taskBlock"
MemberName 26(taskBlock) 0 "atom1"
Name 28 "mytask"
MemberDecorate 7(Buffer) 0 Coherent
MemberDecorate 7(Buffer) 0 Offset 0
Decorate 7(Buffer) Block
Decorate 9 DescriptorSet 0
Decorate 9 Binding 1
Decorate 19 ArrayStride 4
MemberDecorate 20(structType) 0 Offset 0
MemberDecorate 21(t2) 0 Offset 0
Decorate 21(t2) Block
Decorate 23(t) DescriptorSet 0
Decorate 23(t) Binding 0
Decorate 33 BuiltIn WorkgroupSize
2: TypeVoid
3: TypeFunction 2
6: TypeInt 32 1
7(Buffer): TypeStruct 6(int)
8: TypePointer StorageBuffer 7(Buffer)
9: 8(ptr) Variable StorageBuffer
10: 6(int) Constant 0
11: TypePointer StorageBuffer 6(int)
13: 6(int) Constant 1
14: TypeInt 32 0
15: 14(int) Constant 1
16: 14(int) Constant 0
18: 14(int) Constant 3
19: TypeArray 6(int) 18
20(structType): TypeStruct 19
21(t2): TypeStruct 20(structType)
22: TypePointer StorageBuffer 21(t2)
23(t): 22(ptr) Variable StorageBuffer
26(taskBlock): TypeStruct 6(int)
27: TypePointer TaskPayloadWorkgroupEXT 26(taskBlock)
28(mytask): 27(ptr) Variable TaskPayloadWorkgroupEXT
29: TypePointer TaskPayloadWorkgroupEXT 6(int)
32: TypeVector 14(int) 3
33: 32(ivec3) ConstantComposite 15 15 15
4(main): 2 Function None 3
5: Label
12: 11(ptr) AccessChain 9 10
17: 6(int) AtomicIAdd 12 15 16 13
24: 11(ptr) AccessChain 23(t) 10 10 13
25: 6(int) AtomicIAdd 24 15 16 13
30: 29(ptr) AccessChain 28(mytask) 10
31: 6(int) AtomicIAdd 30 15 16 13
Return
FunctionEnd

273
Test/baseResults/spv.ext.meshShaderBuiltins.mesh.out Normal file
View File

@ -0,0 +1,273 @@
spv.ext.meshShaderBuiltins.mesh
// Module Version 10400
// Generated by (magic number): 8000a
// Id's are bound by 159
Capability ClipDistance
Capability CullDistance
Capability FragmentShadingRateKHR
Capability DrawParameters
Capability MultiView
Capability MeshShadingEXT
Extension "SPV_EXT_mesh_shader"
Extension "SPV_KHR_fragment_shading_rate"
1: ExtInstImport "GLSL.std.450"
MemoryModel Logical GLSL450
EntryPoint MeshEXT 4 "main" 13 19 24 41 93 135 153 156
ExecutionMode 4 LocalSize 32 1 1
ExecutionMode 4 OutputVertices 81
ExecutionMode 4 OutputPrimitivesNV 32
ExecutionMode 4 OutputTrianglesNV
Source GLSL 460
SourceExtension "GL_ARB_shader_draw_parameters"
SourceExtension "GL_EXT_mesh_shader"
SourceExtension "GL_EXT_multiview"
Name 4 "main"
Name 6 "testAdditionalBuiltins("
Name 10 "iid"
Name 13 "gl_LocalInvocationID"
Name 18 "gid"
Name 19 "gl_WorkGroupID"
Name 23 "numWorkGrous"
Name 24 "gl_NumWorkGroups"
Name 26 "vertexCount"
Name 28 "primitiveCount"
Name 38 "gl_MeshPerVertexEXT"
MemberName 38(gl_MeshPerVertexEXT) 0 "gl_Position"
MemberName 38(gl_MeshPerVertexEXT) 1 "gl_PointSize"
MemberName 38(gl_MeshPerVertexEXT) 2 "gl_ClipDistance"
MemberName 38(gl_MeshPerVertexEXT) 3 "gl_CullDistance"
Name 41 "gl_MeshVerticesEXT"
Name 90 "gl_MeshPerPrimitiveEXT"
MemberName 90(gl_MeshPerPrimitiveEXT) 0 "gl_PrimitiveID"
MemberName 90(gl_MeshPerPrimitiveEXT) 1 "gl_Layer"
MemberName 90(gl_MeshPerPrimitiveEXT) 2 "gl_ViewportIndex"
MemberName 90(gl_MeshPerPrimitiveEXT) 3 "gl_CullPrimitiveEXT"
MemberName 90(gl_MeshPerPrimitiveEXT) 4 "gl_PrimitiveShadingRateEXT"
Name 93 "gl_MeshPrimitivesEXT"
Name 135 "gl_PrimitiveTriangleIndicesEXT"
Name 151 "id"
Name 153 "gl_DrawIDARB"
Name 155 "viewIdx"
Name 156 "gl_ViewIndex"
Decorate 13(gl_LocalInvocationID) BuiltIn LocalInvocationId
Decorate 19(gl_WorkGroupID) BuiltIn WorkgroupId
Decorate 24(gl_NumWorkGroups) BuiltIn NumWorkgroups
MemberDecorate 38(gl_MeshPerVertexEXT) 0 BuiltIn Position
MemberDecorate 38(gl_MeshPerVertexEXT) 1 BuiltIn PointSize
MemberDecorate 38(gl_MeshPerVertexEXT) 2 BuiltIn ClipDistance
MemberDecorate 38(gl_MeshPerVertexEXT) 3 BuiltIn CullDistance
Decorate 38(gl_MeshPerVertexEXT) Block
MemberDecorate 90(gl_MeshPerPrimitiveEXT) 0 PerPrimitiveNV
MemberDecorate 90(gl_MeshPerPrimitiveEXT) 0 BuiltIn PrimitiveId
MemberDecorate 90(gl_MeshPerPrimitiveEXT) 1 PerPrimitiveNV
MemberDecorate 90(gl_MeshPerPrimitiveEXT) 1 BuiltIn Layer
MemberDecorate 90(gl_MeshPerPrimitiveEXT) 2 PerPrimitiveNV
MemberDecorate 90(gl_MeshPerPrimitiveEXT) 2 BuiltIn ViewportIndex
MemberDecorate 90(gl_MeshPerPrimitiveEXT) 3 PerPrimitiveNV
MemberDecorate 90(gl_MeshPerPrimitiveEXT) 3 BuiltIn CullPrimitiveEXT
MemberDecorate 90(gl_MeshPerPrimitiveEXT) 4 PerPrimitiveNV
MemberDecorate 90(gl_MeshPerPrimitiveEXT) 4 BuiltIn PrimitiveShadingRateKHR
Decorate 90(gl_MeshPerPrimitiveEXT) Block
Decorate 135(gl_PrimitiveTriangleIndicesEXT) BuiltIn PrimitiveTriangleIndicesEXT
Decorate 153(gl_DrawIDARB) BuiltIn DrawIndex
Decorate 156(gl_ViewIndex) BuiltIn ViewIndex
Decorate 158 BuiltIn WorkgroupSize
2: TypeVoid
3: TypeFunction 2
8: TypeInt 32 0
9: TypePointer Function 8(int)
11: TypeVector 8(int) 3
12: TypePointer Input 11(ivec3)
13(gl_LocalInvocationID): 12(ptr) Variable Input
14: 8(int) Constant 0
15: TypePointer Input 8(int)
19(gl_WorkGroupID): 12(ptr) Variable Input
22: TypePointer Function 11(ivec3)
24(gl_NumWorkGroups): 12(ptr) Variable Input
27: 8(int) Constant 81
29: 8(int) Constant 32
32: TypeFloat 32
33: TypeVector 32(float) 4
34: 8(int) Constant 4
35: TypeArray 32(float) 34
36: 8(int) Constant 3
37: TypeArray 32(float) 36
38(gl_MeshPerVertexEXT): TypeStruct 33(fvec4) 32(float) 35 37
39: TypeArray 38(gl_MeshPerVertexEXT) 27
40: TypePointer Output 39
41(gl_MeshVerticesEXT): 40(ptr) Variable Output
43: TypeInt 32 1
44: 43(int) Constant 0
45: 32(float) Constant 1065353216
46: 33(fvec4) ConstantComposite 45 45 45 45
47: TypePointer Output 33(fvec4)
50: 43(int) Constant 1
51: 32(float) Constant 1073741824
52: TypePointer Output 32(float)
55: 43(int) Constant 2
56: 43(int) Constant 3
57: 32(float) Constant 1077936128
60: 32(float) Constant 1082130432
62: 8(int) Constant 1
63: 8(int) Constant 264
64: 8(int) Constant 2
89: TypeBool
90(gl_MeshPerPrimitiveEXT): TypeStruct 43(int) 43(int) 43(int) 89(bool) 43(int)
91: TypeArray 90(gl_MeshPerPrimitiveEXT) 29
92: TypePointer Output 91
93(gl_MeshPrimitivesEXT): 92(ptr) Variable Output
95: 43(int) Constant 6
96: TypePointer Output 43(int)
99: 43(int) Constant 7
102: 43(int) Constant 8
105: 89(bool) ConstantFalse
106: TypePointer Output 89(bool)
132: 8(int) Constant 96
133: TypeArray 11(ivec3) 132
134: TypePointer Output 133
135(gl_PrimitiveTriangleIndicesEXT): 134(ptr) Variable Output
136: 8(int) Constant 257
137: 11(ivec3) ConstantComposite 136 136 136
138: TypePointer Output 11(ivec3)
142: 11(ivec3) ConstantComposite 64 64 64
150: TypePointer Function 43(int)
152: TypePointer Input 43(int)
153(gl_DrawIDARB): 152(ptr) Variable Input
156(gl_ViewIndex): 152(ptr) Variable Input
158: 11(ivec3) ConstantComposite 29 62 62
4(main): 2 Function None 3
5: Label
10(iid): 9(ptr) Variable Function
18(gid): 9(ptr) Variable Function
23(numWorkGrous): 22(ptr) Variable Function
26(vertexCount): 9(ptr) Variable Function
28(primitiveCount): 9(ptr) Variable Function
16: 15(ptr) AccessChain 13(gl_LocalInvocationID) 14
17: 8(int) Load 16
Store 10(iid) 17
20: 15(ptr) AccessChain 19(gl_WorkGroupID) 14
21: 8(int) Load 20
Store 18(gid) 21
25: 11(ivec3) Load 24(gl_NumWorkGroups)
Store 23(numWorkGrous) 25
Store 26(vertexCount) 27
Store 28(primitiveCount) 29
30: 8(int) Load 26(vertexCount)
31: 8(int) Load 28(primitiveCount)
SetMeshOutputsEXT 30 31
42: 8(int) Load 10(iid)
48: 47(ptr) AccessChain 41(gl_MeshVerticesEXT) 42 44
Store 48 46
49: 8(int) Load 10(iid)
53: 52(ptr) AccessChain 41(gl_MeshVerticesEXT) 49 50
Store 53 51
54: 8(int) Load 10(iid)
58: 52(ptr) AccessChain 41(gl_MeshVerticesEXT) 54 55 56
Store 58 57
59: 8(int) Load 10(iid)
61: 52(ptr) AccessChain 41(gl_MeshVerticesEXT) 59 56 55
Store 61 60
MemoryBarrier 62 63
ControlBarrier 64 64 63
65: 8(int) Load 10(iid)
66: 8(int) IAdd 65 62
67: 8(int) Load 10(iid)
68: 47(ptr) AccessChain 41(gl_MeshVerticesEXT) 67 44
69: 33(fvec4) Load 68
70: 47(ptr) AccessChain 41(gl_MeshVerticesEXT) 66 44
Store 70 69
71: 8(int) Load 10(iid)
72: 8(int) IAdd 71 62
73: 8(int) Load 10(iid)
74: 52(ptr) AccessChain 41(gl_MeshVerticesEXT) 73 50
75: 32(float) Load 74
76: 52(ptr) AccessChain 41(gl_MeshVerticesEXT) 72 50
Store 76 75
77: 8(int) Load 10(iid)
78: 8(int) IAdd 77 62
79: 8(int) Load 10(iid)
80: 52(ptr) AccessChain 41(gl_MeshVerticesEXT) 79 55 56
81: 32(float) Load 80
82: 52(ptr) AccessChain 41(gl_MeshVerticesEXT) 78 55 56
Store 82 81
83: 8(int) Load 10(iid)
84: 8(int) IAdd 83 62
85: 8(int) Load 10(iid)
86: 52(ptr) AccessChain 41(gl_MeshVerticesEXT) 85 56 55
87: 32(float) Load 86
88: 52(ptr) AccessChain 41(gl_MeshVerticesEXT) 84 56 55
Store 88 87
MemoryBarrier 62 63
ControlBarrier 64 64 63
94: 8(int) Load 10(iid)
97: 96(ptr) AccessChain 93(gl_MeshPrimitivesEXT) 94 44
Store 97 95
98: 8(int) Load 10(iid)
100: 96(ptr) AccessChain 93(gl_MeshPrimitivesEXT) 98 50
Store 100 99
101: 8(int) Load 10(iid)
103: 96(ptr) AccessChain 93(gl_MeshPrimitivesEXT) 101 55
Store 103 102
104: 8(int) Load 10(iid)
107: 106(ptr) AccessChain 93(gl_MeshPrimitivesEXT) 104 56
Store 107 105
MemoryBarrier 62 63
ControlBarrier 64 64 63
108: 8(int) Load 10(iid)
109: 8(int) IAdd 108 62
110: 8(int) Load 10(iid)
111: 96(ptr) AccessChain 93(gl_MeshPrimitivesEXT) 110 44
112: 43(int) Load 111
113: 96(ptr) AccessChain 93(gl_MeshPrimitivesEXT) 109 44
Store 113 112
114: 8(int) Load 10(iid)
115: 8(int) IAdd 114 62
116: 8(int) Load 10(iid)
117: 96(ptr) AccessChain 93(gl_MeshPrimitivesEXT) 116 50
118: 43(int) Load 117
119: 96(ptr) AccessChain 93(gl_MeshPrimitivesEXT) 115 50
Store 119 118
120: 8(int) Load 10(iid)
121: 8(int) IAdd 120 62
122: 8(int) Load 10(iid)
123: 96(ptr) AccessChain 93(gl_MeshPrimitivesEXT) 122 55
124: 43(int) Load 123
125: 96(ptr) AccessChain 93(gl_MeshPrimitivesEXT) 121 55
Store 125 124
126: 8(int) Load 10(iid)
127: 8(int) IAdd 126 62
128: 8(int) Load 10(iid)
129: 106(ptr) AccessChain 93(gl_MeshPrimitivesEXT) 128 56
130: 89(bool) Load 129
131: 106(ptr) AccessChain 93(gl_MeshPrimitivesEXT) 127 56
Store 131 130
MemoryBarrier 62 63
ControlBarrier 64 64 63
139: 138(ptr) AccessChain 135(gl_PrimitiveTriangleIndicesEXT) 44
Store 139 137
140: 8(int) Load 28(primitiveCount)
141: 8(int) ISub 140 62
143: 138(ptr) AccessChain 135(gl_PrimitiveTriangleIndicesEXT) 141
Store 143 142
144: 8(int) Load 18(gid)
145: 8(int) Load 18(gid)
146: 8(int) ISub 145 62
147: 138(ptr) AccessChain 135(gl_PrimitiveTriangleIndicesEXT) 146
148: 11(ivec3) Load 147
149: 138(ptr) AccessChain 135(gl_PrimitiveTriangleIndicesEXT) 144
Store 149 148
MemoryBarrier 62 63
ControlBarrier 64 64 63
Return
FunctionEnd
6(testAdditionalBuiltins(): 2 Function None 3
7: Label
151(id): 150(ptr) Variable Function
155(viewIdx): 150(ptr) Variable Function
154: 43(int) Load 153(gl_DrawIDARB)
Store 151(id) 154
157: 43(int) Load 156(gl_ViewIndex)
Store 155(viewIdx) 157
Return
FunctionEnd

216
Test/baseResults/spv.ext.meshShaderRedeclBuiltins.mesh.out Normal file
View File

@ -0,0 +1,216 @@
spv.ext.meshShaderRedeclBuiltins.mesh
// Module Version 10400
// Generated by (magic number): 8000a
// Id's are bound by 128
Capability ClipDistance
Capability CullDistance
Capability FragmentShadingRateKHR
Capability MeshShadingEXT
Extension "SPV_EXT_mesh_shader"
Extension "SPV_KHR_fragment_shading_rate"
1: ExtInstImport "GLSL.std.450"
MemoryModel Logical GLSL450
EntryPoint MeshEXT 4 "main" 11 17 29 81 122
ExecutionMode 4 LocalSize 32 1 1
ExecutionMode 4 OutputVertices 81
ExecutionMode 4 OutputPrimitivesNV 32
ExecutionMode 4 OutputPoints
Source GLSL 460
SourceExtension "GL_EXT_mesh_shader"
Name 4 "main"
Name 8 "iid"
Name 11 "gl_LocalInvocationID"
Name 16 "gid"
Name 17 "gl_WorkGroupID"
Name 26 "gl_MeshPerVertexEXT"
MemberName 26(gl_MeshPerVertexEXT) 0 "gl_Position"
MemberName 26(gl_MeshPerVertexEXT) 1 "gl_PointSize"
MemberName 26(gl_MeshPerVertexEXT) 2 "gl_ClipDistance"
MemberName 26(gl_MeshPerVertexEXT) 3 "gl_CullDistance"
Name 29 "gl_MeshVerticesEXT"
Name 78 "gl_MeshPerPrimitiveEXT"
MemberName 78(gl_MeshPerPrimitiveEXT) 0 "gl_PrimitiveID"
MemberName 78(gl_MeshPerPrimitiveEXT) 1 "gl_Layer"
MemberName 78(gl_MeshPerPrimitiveEXT) 2 "gl_ViewportIndex"
MemberName 78(gl_MeshPerPrimitiveEXT) 3 "gl_CullPrimitiveEXT"
MemberName 78(gl_MeshPerPrimitiveEXT) 4 "gl_PrimitiveShadingRateEXT"
Name 81 "gl_MeshPrimitivesEXT"
Name 122 "gl_PrimitivePointIndicesEXT"
Decorate 11(gl_LocalInvocationID) BuiltIn LocalInvocationId
Decorate 17(gl_WorkGroupID) BuiltIn WorkgroupId
MemberDecorate 26(gl_MeshPerVertexEXT) 0 BuiltIn Position
MemberDecorate 26(gl_MeshPerVertexEXT) 1 BuiltIn PointSize
MemberDecorate 26(gl_MeshPerVertexEXT) 2 BuiltIn ClipDistance
MemberDecorate 26(gl_MeshPerVertexEXT) 3 BuiltIn CullDistance
Decorate 26(gl_MeshPerVertexEXT) Block
MemberDecorate 78(gl_MeshPerPrimitiveEXT) 0 PerPrimitiveNV
MemberDecorate 78(gl_MeshPerPrimitiveEXT) 0 BuiltIn PrimitiveId
MemberDecorate 78(gl_MeshPerPrimitiveEXT) 1 PerPrimitiveNV
MemberDecorate 78(gl_MeshPerPrimitiveEXT) 1 BuiltIn Layer
MemberDecorate 78(gl_MeshPerPrimitiveEXT) 2 PerPrimitiveNV
MemberDecorate 78(gl_MeshPerPrimitiveEXT) 2 BuiltIn ViewportIndex
MemberDecorate 78(gl_MeshPerPrimitiveEXT) 3 PerPrimitiveNV
MemberDecorate 78(gl_MeshPerPrimitiveEXT) 3 BuiltIn CullPrimitiveEXT
MemberDecorate 78(gl_MeshPerPrimitiveEXT) 4 PerPrimitiveNV
MemberDecorate 78(gl_MeshPerPrimitiveEXT) 4 BuiltIn PrimitiveShadingRateKHR
Decorate 78(gl_MeshPerPrimitiveEXT) Block
Decorate 122(gl_PrimitivePointIndicesEXT) BuiltIn PrimitivePointIndicesEXT
Decorate 127 BuiltIn WorkgroupSize
2: TypeVoid
3: TypeFunction 2
6: TypeInt 32 0
7: TypePointer Function 6(int)
9: TypeVector 6(int) 3
10: TypePointer Input 9(ivec3)
11(gl_LocalInvocationID): 10(ptr) Variable Input
12: 6(int) Constant 0
13: TypePointer Input 6(int)
17(gl_WorkGroupID): 10(ptr) Variable Input
20: 6(int) Constant 81
21: 6(int) Constant 32
22: TypeFloat 32
23: TypeVector 22(float) 4
24: 6(int) Constant 4
25: TypeArray 22(float) 24
26(gl_MeshPerVertexEXT): TypeStruct 23(fvec4) 22(float) 25 25
27: TypeArray 26(gl_MeshPerVertexEXT) 20
28: TypePointer Output 27
29(gl_MeshVerticesEXT): 28(ptr) Variable Output
31: TypeInt 32 1
32: 31(int) Constant 0
33: 22(float) Constant 1065353216
34: 23(fvec4) ConstantComposite 33 33 33 33
35: TypePointer Output 23(fvec4)
38: 31(int) Constant 1
39: 22(float) Constant 1073741824
40: TypePointer Output 22(float)
43: 31(int) Constant 2
44: 31(int) Constant 3
45: 22(float) Constant 1077936128
48: 22(float) Constant 1082130432
50: 6(int) Constant 1
51: 6(int) Constant 264
52: 6(int) Constant 2
77: TypeBool
78(gl_MeshPerPrimitiveEXT): TypeStruct 31(int) 31(int) 31(int) 77(bool) 31(int)
79: TypeArray 78(gl_MeshPerPrimitiveEXT) 21
80: TypePointer Output 79
81(gl_MeshPrimitivesEXT): 80(ptr) Variable Output
83: 31(int) Constant 6
84: TypePointer Output 31(int)
87: 31(int) Constant 7
90: 31(int) Constant 8
93: 77(bool) ConstantFalse
94: TypePointer Output 77(bool)
120: TypeArray 6(int) 21
121: TypePointer Output 120
122(gl_PrimitivePointIndicesEXT): 121(ptr) Variable Output
123: TypePointer Output 6(int)
125: 31(int) Constant 31
127: 9(ivec3) ConstantComposite 21 50 50
4(main): 2 Function None 3
5: Label
8(iid): 7(ptr) Variable Function
16(gid): 7(ptr) Variable Function
14: 13(ptr) AccessChain 11(gl_LocalInvocationID) 12
15: 6(int) Load 14
Store 8(iid) 15
18: 13(ptr) AccessChain 17(gl_WorkGroupID) 12
19: 6(int) Load 18
Store 16(gid) 19
SetMeshOutputsEXT 20 21
30: 6(int) Load 8(iid)
36: 35(ptr) AccessChain 29(gl_MeshVerticesEXT) 30 32
Store 36 34
37: 6(int) Load 8(iid)
41: 40(ptr) AccessChain 29(gl_MeshVerticesEXT) 37 38
Store 41 39
42: 6(int) Load 8(iid)
46: 40(ptr) AccessChain 29(gl_MeshVerticesEXT) 42 43 44
Store 46 45
47: 6(int) Load 8(iid)
49: 40(ptr) AccessChain 29(gl_MeshVerticesEXT) 47 44 43
Store 49 48
MemoryBarrier 50 51
ControlBarrier 52 52 51
53: 6(int) Load 8(iid)
54: 6(int) IAdd 53 50
55: 6(int) Load 8(iid)
56: 35(ptr) AccessChain 29(gl_MeshVerticesEXT) 55 32
57: 23(fvec4) Load 56
58: 35(ptr) AccessChain 29(gl_MeshVerticesEXT) 54 32
Store 58 57
59: 6(int) Load 8(iid)
60: 6(int) IAdd 59 50
61: 6(int) Load 8(iid)
62: 40(ptr) AccessChain 29(gl_MeshVerticesEXT) 61 38
63: 22(float) Load 62
64: 40(ptr) AccessChain 29(gl_MeshVerticesEXT) 60 38
Store 64 63
65: 6(int) Load 8(iid)
66: 6(int) IAdd 65 50
67: 6(int) Load 8(iid)
68: 40(ptr) AccessChain 29(gl_MeshVerticesEXT) 67 43 44
69: 22(float) Load 68
70: 40(ptr) AccessChain 29(gl_MeshVerticesEXT) 66 43 44
Store 70 69
71: 6(int) Load 8(iid)
72: 6(int) IAdd 71 50
73: 6(int) Load 8(iid)
74: 40(ptr) AccessChain 29(gl_MeshVerticesEXT) 73 44 43
75: 22(float) Load 74
76: 40(ptr) AccessChain 29(gl_MeshVerticesEXT) 72 44 43
Store 76 75
MemoryBarrier 50 51
ControlBarrier 52 52 51
82: 6(int) Load 8(iid)
85: 84(ptr) AccessChain 81(gl_MeshPrimitivesEXT) 82 32
Store 85 83
86: 6(int) Load 8(iid)
88: 84(ptr) AccessChain 81(gl_MeshPrimitivesEXT) 86 38
Store 88 87
89: 6(int) Load 8(iid)
91: 84(ptr) AccessChain 81(gl_MeshPrimitivesEXT) 89 43
Store 91 90
92: 6(int) Load 8(iid)
95: 94(ptr) AccessChain 81(gl_MeshPrimitivesEXT) 92 44
Store 95 93
MemoryBarrier 50 51
ControlBarrier 52 52 51
96: 6(int) Load 8(iid)
97: 6(int) IAdd 96 50
98: 6(int) Load 8(iid)
99: 84(ptr) AccessChain 81(gl_MeshPrimitivesEXT) 98 32
100: 31(int) Load 99
101: 84(ptr) AccessChain 81(gl_MeshPrimitivesEXT) 97 32
Store 101 100
102: 6(int) Load 8(iid)
103: 6(int) IAdd 102 50
104: 6(int) Load 8(iid)
105: 84(ptr) AccessChain 81(gl_MeshPrimitivesEXT) 104 38
106: 31(int) Load 105
107: 84(ptr) AccessChain 81(gl_MeshPrimitivesEXT) 103 38
Store 107 106
108: 6(int) Load 8(iid)
109: 6(int) IAdd 108 50
110: 6(int) Load 8(iid)
111: 84(ptr) AccessChain 81(gl_MeshPrimitivesEXT) 110 43
112: 31(int) Load 111
113: 84(ptr) AccessChain 81(gl_MeshPrimitivesEXT) 109 43
Store 113 112
114: 6(int) Load 8(iid)
115: 6(int) IAdd 114 50
116: 6(int) Load 8(iid)
117: 94(ptr) AccessChain 81(gl_MeshPrimitivesEXT) 116 44
118: 77(bool) Load 117
119: 94(ptr) AccessChain 81(gl_MeshPrimitivesEXT) 115 44
Store 119 118
MemoryBarrier 50 51
ControlBarrier 52 52 51
124: 123(ptr) AccessChain 122(gl_PrimitivePointIndicesEXT) 32
Store 124 50
126: 123(ptr) AccessChain 122(gl_PrimitivePointIndicesEXT) 125
Store 126 52
Return
FunctionEnd

102
Test/baseResults/spv.ext.meshShaderTaskMem.mesh.out Normal file
View File

@ -0,0 +1,102 @@
spv.ext.meshShaderTaskMem.mesh
// Module Version 10400
// Generated by (magic number): 8000a
// Id's are bound by 58
Capability MeshShadingEXT
Extension "SPV_EXT_mesh_shader"
1: ExtInstImport "GLSL.std.450"
MemoryModel Logical GLSL450
EntryPoint MeshEXT 4 "main" 11 22 30 38
ExecutionMode 4 LocalSize 32 1 1
ExecutionMode 4 OutputVertices 81
ExecutionMode 4 OutputPrimitivesNV 32
ExecutionMode 4 OutputTrianglesNV
Source GLSL 450
SourceExtension "GL_EXT_mesh_shader"
Name 4 "main"
Name 8 "iid"
Name 11 "gl_LocalInvocationID"
Name 18 "outBlock"
MemberName 18(outBlock) 0 "gid5"
MemberName 18(outBlock) 1 "gid6"
Name 22 "myblk"
Name 28 "taskBlock"
MemberName 28(taskBlock) 0 "gid1"
MemberName 28(taskBlock) 1 "gid2"
Name 30 "mytask"
Name 36 "bufferBlock"
MemberName 36(bufferBlock) 0 "gid3"
MemberName 36(bufferBlock) 1 "gid4"
Name 38 "mybuf"
Decorate 11(gl_LocalInvocationID) BuiltIn LocalInvocationId
Decorate 18(outBlock) Block
Decorate 22(myblk) Location 0
Decorate 35 ArrayStride 4
MemberDecorate 36(bufferBlock) 0 Offset 0
MemberDecorate 36(bufferBlock) 1 Offset 16
Decorate 36(bufferBlock) Block
Decorate 38(mybuf) DescriptorSet 0
Decorate 38(mybuf) Binding 0
Decorate 57 BuiltIn WorkgroupSize
2: TypeVoid
3: TypeFunction 2
6: TypeInt 32 0
7: TypePointer Function 6(int)
9: TypeVector 6(int) 3
10: TypePointer Input 9(ivec3)
11(gl_LocalInvocationID): 10(ptr) Variable Input
12: 6(int) Constant 0
13: TypePointer Input 6(int)
16: TypeFloat 32
17: TypeVector 16(float) 4
18(outBlock): TypeStruct 16(float) 17(fvec4)
19: 6(int) Constant 81
20: TypeArray 18(outBlock) 19
21: TypePointer Output 20
22(myblk): 21(ptr) Variable Output
24: TypeInt 32 1
25: 24(int) Constant 0
26: 6(int) Constant 2
27: TypeArray 16(float) 26
28(taskBlock): TypeStruct 27 17(fvec4)
29: TypePointer TaskPayloadWorkgroupEXT 28(taskBlock)
30(mytask): 29(ptr) Variable TaskPayloadWorkgroupEXT
31: 24(int) Constant 1
32: TypePointer TaskPayloadWorkgroupEXT 16(float)
35: TypeArray 16(float) 26
36(bufferBlock): TypeStruct 35 17(fvec4)
37: TypePointer StorageBuffer 36(bufferBlock)
38(mybuf): 37(ptr) Variable StorageBuffer
39: TypePointer StorageBuffer 16(float)
43: TypePointer Output 16(float)
46: TypePointer TaskPayloadWorkgroupEXT 17(fvec4)
49: TypePointer StorageBuffer 17(fvec4)
53: TypePointer Output 17(fvec4)
55: 6(int) Constant 32
56: 6(int) Constant 1
57: 9(ivec3) ConstantComposite 55 56 56
4(main): 2 Function None 3
5: Label
8(iid): 7(ptr) Variable Function
14: 13(ptr) AccessChain 11(gl_LocalInvocationID) 12
15: 6(int) Load 14
Store 8(iid) 15
23: 6(int) Load 8(iid)
33: 32(ptr) AccessChain 30(mytask) 25 31
34: 16(float) Load 33
40: 39(ptr) AccessChain 38(mybuf) 25 31
41: 16(float) Load 40
42: 16(float) FAdd 34 41
44: 43(ptr) AccessChain 22(myblk) 23 25
Store 44 42
45: 6(int) Load 8(iid)
47: 46(ptr) AccessChain 30(mytask) 31
48: 17(fvec4) Load 47
50: 49(ptr) AccessChain 38(mybuf) 31
51: 17(fvec4) Load 50
52: 17(fvec4) FAdd 48 51
54: 53(ptr) AccessChain 22(myblk) 45 31
Store 54 52
Return
FunctionEnd

208
Test/baseResults/spv.ext.meshShaderUserDefined.mesh.out Normal file
View File

@ -0,0 +1,208 @@
spv.ext.meshShaderUserDefined.mesh
// Module Version 10400
// Generated by (magic number): 8000a
// Id's are bound by 141
Capability MeshShadingEXT
Extension "SPV_EXT_mesh_shader"
1: ExtInstImport "GLSL.std.450"
MemoryModel Logical GLSL450
EntryPoint MeshEXT 4 "main" 11 17 34 104
ExecutionMode 4 LocalSize 32 1 1
ExecutionMode 4 OutputVertices 81
ExecutionMode 4 OutputPrimitivesNV 32
ExecutionMode 4 OutputTrianglesNV
Source GLSL 450
SourceExtension "GL_EXT_mesh_shader"
Name 4 "main"
Name 8 "iid"
Name 11 "gl_LocalInvocationID"
Name 16 "gid"
Name 17 "gl_WorkGroupID"
Name 30 "myblock"
MemberName 30(myblock) 0 "f"
MemberName 30(myblock) 1 "fArr"
MemberName 30(myblock) 2 "pos"
MemberName 30(myblock) 3 "posArr"
MemberName 30(myblock) 4 "m"
MemberName 30(myblock) 5 "mArr"
Name 34 "blk"
Name 100 "myblock2"
MemberName 100(myblock2) 0 "f"
MemberName 100(myblock2) 1 "pos"
MemberName 100(myblock2) 2 "m"
Name 104 "blk2"
Decorate 11(gl_LocalInvocationID) BuiltIn LocalInvocationId
Decorate 17(gl_WorkGroupID) BuiltIn WorkgroupId
MemberDecorate 30(myblock) 0 PerPrimitiveNV
MemberDecorate 30(myblock) 1 PerPrimitiveNV
MemberDecorate 30(myblock) 2 PerPrimitiveNV
MemberDecorate 30(myblock) 3 PerPrimitiveNV
MemberDecorate 30(myblock) 4 PerPrimitiveNV
MemberDecorate 30(myblock) 5 PerPrimitiveNV
Decorate 30(myblock) Block
Decorate 34(blk) Location 0
Decorate 100(myblock2) Block
Decorate 104(blk2) Location 20
Decorate 140 BuiltIn WorkgroupSize
2: TypeVoid
3: TypeFunction 2
6: TypeInt 32 0
7: TypePointer Function 6(int)
9: TypeVector 6(int) 3
10: TypePointer Input 9(ivec3)
11(gl_LocalInvocationID): 10(ptr) Variable Input
12: 6(int) Constant 0
13: TypePointer Input 6(int)
17(gl_WorkGroupID): 10(ptr) Variable Input
20: TypeFloat 32
21: 6(int) Constant 4
22: TypeArray 20(float) 21
23: TypeVector 20(float) 3
24: TypeVector 20(float) 4
25: TypeArray 24(fvec4) 21
26: TypeMatrix 24(fvec4) 4
27: TypeMatrix 23(fvec3) 3
28: 6(int) Constant 2
29: TypeArray 27 28
30(myblock): TypeStruct 20(float) 22 23(fvec3) 25 26 29
31: 6(int) Constant 32
32: TypeArray 30(myblock) 31
33: TypePointer Output 32
34(blk): 33(ptr) Variable Output
36: TypeInt 32 1
37: 36(int) Constant 0
38: 20(float) Constant 1093664768
39: TypePointer Output 20(float)
42: 6(int) Constant 1
44: 36(int) Constant 1
52: 36(int) Constant 2
53: 20(float) Constant 1096810496
54: 20(float) Constant 1097859072
55: 20(float) Constant 1095761920
56: 23(fvec3) ConstantComposite 53 54 55
57: TypePointer Output 23(fvec3)
63: 36(int) Constant 3
72: 6(int) Constant 3
77: 36(int) Constant 4
78: 20(float) Constant 1098907648
79: 24(fvec4) ConstantComposite 55 53 54 78
80: TypePointer Output 24(fvec4)
85: 36(int) Constant 5
94: 20(float) Constant 1099431936
95: 20(float) Constant 1099956224
96: 20(float) Constant 1100480512
97: 23(fvec3) ConstantComposite 94 95 96
99: 6(int) Constant 264
100(myblock2): TypeStruct 20(float) 24(fvec4) 26
101: 6(int) Constant 81
102: TypeArray 100(myblock2) 101
103: TypePointer Output 102
104(blk2): 103(ptr) Variable Output
110: 20(float) Constant 1101004800
114: 20(float) Constant 1101529088
115: 20(float) Constant 1102053376
116: 20(float) Constant 1102577664
117: 20(float) Constant 1103101952
118: 24(fvec4) ConstantComposite 114 115 116 117
130: 20(float) Constant 1105723392
140: 9(ivec3) ConstantComposite 31 42 42
4(main): 2 Function None 3
5: Label
8(iid): 7(ptr) Variable Function
16(gid): 7(ptr) Variable Function
14: 13(ptr) AccessChain 11(gl_LocalInvocationID) 12
15: 6(int) Load 14
Store 8(iid) 15
18: 13(ptr) AccessChain 17(gl_WorkGroupID) 12
19: 6(int) Load 18
Store 16(gid) 19
35: 6(int) Load 8(iid)
40: 39(ptr) AccessChain 34(blk) 35 37
Store 40 38
41: 6(int) Load 8(iid)
43: 6(int) IAdd 41 42
45: 6(int) Load 16(gid)
46: 6(int) Load 8(iid)
47: 39(ptr) AccessChain 34(blk) 46 37
48: 20(float) Load 47
49: 39(ptr) AccessChain 34(blk) 43 44 45
Store 49 48
50: 6(int) Load 8(iid)
51: 6(int) UDiv 50 28
58: 57(ptr) AccessChain 34(blk) 51 52
59: 23(fvec3) Load 58
60: 23(fvec3) VectorShuffle 59 56 5 3 4
Store 58 60
61: 6(int) Load 8(iid)
62: 6(int) IMul 61 28
64: 6(int) Load 8(iid)
65: 6(int) UDiv 64 28
66: 57(ptr) AccessChain 34(blk) 65 52
67: 23(fvec3) Load 66
68: 39(ptr) AccessChain 34(blk) 62 63 44 42
69: 20(float) CompositeExtract 67 0
Store 68 69
70: 39(ptr) AccessChain 34(blk) 62 63 44 28
71: 20(float) CompositeExtract 67 1
Store 70 71
73: 39(ptr) AccessChain 34(blk) 62 63 44 72
74: 20(float) CompositeExtract 67 2
Store 73 74
75: 6(int) Load 8(iid)
76: 6(int) UDiv 75 21
81: 80(ptr) AccessChain 34(blk) 76 77 52
82: 24(fvec4) Load 81
83: 24(fvec4) VectorShuffle 82 79 7 6 5 4
Store 81 83
84: 6(int) Load 8(iid)
86: 6(int) Load 8(iid)
87: 6(int) UDiv 86 21
88: 39(ptr) AccessChain 34(blk) 87 77 52 72
89: 20(float) Load 88
90: 39(ptr) AccessChain 34(blk) 84 85 37 44 42
Store 90 89
91: 6(int) Load 8(iid)
92: 6(int) IMul 91 21
93: 6(int) Load 16(gid)
98: 57(ptr) AccessChain 34(blk) 92 85 44 93
Store 98 97
MemoryBarrier 42 99
ControlBarrier 28 28 99
105: 6(int) Load 8(iid)
106: 6(int) Load 8(iid)
107: 6(int) ISub 106 42
108: 39(ptr) AccessChain 104(blk2) 107 37
109: 20(float) Load 108
111: 20(float) FAdd 109 110
112: 39(ptr) AccessChain 104(blk2) 105 37
Store 112 111
113: 6(int) Load 8(iid)
119: 80(ptr) AccessChain 104(blk2) 113 44
Store 119 118
120: 6(int) Load 8(iid)
121: 6(int) IAdd 120 42
122: 6(int) Load 16(gid)
123: 6(int) Load 8(iid)
124: 80(ptr) AccessChain 104(blk2) 123 44
125: 24(fvec4) Load 124
126: 80(ptr) AccessChain 104(blk2) 121 52 122
Store 126 125
127: 6(int) Load 8(iid)
128: 6(int) IAdd 127 42
129: 6(int) Load 16(gid)
131: 39(ptr) AccessChain 104(blk2) 128 52 129 28
Store 131 130
132: 6(int) Load 8(iid)
133: 6(int) IAdd 132 28
134: 6(int) Load 8(iid)
135: 6(int) IAdd 134 42
136: 6(int) Load 16(gid)
137: 80(ptr) AccessChain 104(blk2) 135 52 136
138: 24(fvec4) Load 137
139: 80(ptr) AccessChain 104(blk2) 133 52 63
Store 139 138
MemoryBarrier 42 99
ControlBarrier 28 28 99
Return
FunctionEnd

163
Test/baseResults/spv.ext.meshTaskShader.task.out Normal file
View File

@ -0,0 +1,163 @@
spv.ext.meshTaskShader.task
// Module Version 10400
// Generated by (magic number): 8000a
// Id's are bound by 102
Capability StorageImageWriteWithoutFormat
Capability MeshShadingEXT
Extension "SPV_EXT_mesh_shader"
1: ExtInstImport "GLSL.std.450"
MemoryModel Logical GLSL450
EntryPoint TaskEXT 4 "main" 11 17 34 39 55 80
ExecutionMode 4 LocalSize 32 1 1
Source GLSL 450
SourceExtension "GL_EXT_mesh_shader"
Name 4 "main"
Name 8 "iid"
Name 11 "gl_LocalInvocationID"
Name 16 "gid"
Name 17 "gl_WorkGroupID"
Name 20 "i"
Name 34 "mem"
Name 37 "block0"
MemberName 37(block0) 0 "uni_value"
Name 39 ""
Name 55 "uni_image"
Name 78 "Task"
MemberName 78(Task) 0 "dummy"
MemberName 78(Task) 1 "submesh"
Name 80 "mytask"
Decorate 11(gl_LocalInvocationID) BuiltIn LocalInvocationId
Decorate 17(gl_WorkGroupID) BuiltIn WorkgroupId
MemberDecorate 37(block0) 0 Offset 0
Decorate 37(block0) Block
Decorate 39 DescriptorSet 0
Decorate 39 Binding 1
Decorate 55(uni_image) DescriptorSet 0
Decorate 55(uni_image) Binding 0
Decorate 55(uni_image) NonReadable
Decorate 101 BuiltIn WorkgroupSize
2: TypeVoid
3: TypeFunction 2
6: TypeInt 32 0
7: TypePointer Function 6(int)
9: TypeVector 6(int) 3
10: TypePointer Input 9(ivec3)
11(gl_LocalInvocationID): 10(ptr) Variable Input
12: 6(int) Constant 0
13: TypePointer Input 6(int)
17(gl_WorkGroupID): 10(ptr) Variable Input
27: 6(int) Constant 10
28: TypeBool
30: TypeFloat 32
31: TypeVector 30(float) 4
32: TypeArray 31(fvec4) 27
33: TypePointer Workgroup 32
34(mem): 33(ptr) Variable Workgroup
37(block0): TypeStruct 6(int)
38: TypePointer Uniform 37(block0)
39: 38(ptr) Variable Uniform
40: TypeInt 32 1
41: 40(int) Constant 0
42: TypePointer Uniform 6(int)
48: TypePointer Workgroup 31(fvec4)
51: 40(int) Constant 1
53: TypeImage 30(float) 2D nonsampled format:Unknown
54: TypePointer UniformConstant 53
55(uni_image): 54(ptr) Variable UniformConstant
59: TypeVector 40(int) 2
69: 6(int) Constant 1
73: 6(int) Constant 264
74: 6(int) Constant 2
75: TypeVector 30(float) 2
76: 6(int) Constant 3
77: TypeArray 75(fvec2) 76
78(Task): TypeStruct 75(fvec2) 77
79: TypePointer TaskPayloadWorkgroupEXT 78(Task)
80(mytask): 79(ptr) Variable TaskPayloadWorkgroupEXT
81: 30(float) Constant 1106247680
82: 30(float) Constant 1106771968
83: 75(fvec2) ConstantComposite 81 82
84: TypePointer TaskPayloadWorkgroupEXT 75(fvec2)
86: 30(float) Constant 1107296256
87: 30(float) Constant 1107558400
88: 75(fvec2) ConstantComposite 86 87
90: 30(float) Constant 1107820544
91: 30(float) Constant 1108082688
92: 75(fvec2) ConstantComposite 90 91
94: 40(int) Constant 2
100: 6(int) Constant 32
101: 9(ivec3) ConstantComposite 100 69 69
4(main): 2 Function None 3
5: Label
8(iid): 7(ptr) Variable Function
16(gid): 7(ptr) Variable Function
20(i): 7(ptr) Variable Function
14: 13(ptr) AccessChain 11(gl_LocalInvocationID) 12
15: 6(int) Load 14
Store 8(iid) 15
18: 13(ptr) AccessChain 17(gl_WorkGroupID) 12
19: 6(int) Load 18
Store 16(gid) 19
Store 20(i) 12
Branch 21
21: Label
LoopMerge 23 24 None
Branch 25
25: Label
26: 6(int) Load 20(i)
29: 28(bool) ULessThan 26 27
BranchConditional 29 22 23
22: Label
35: 6(int) Load 20(i)
36: 6(int) Load 20(i)
43: 42(ptr) AccessChain 39 41
44: 6(int) Load 43
45: 6(int) IAdd 36 44
46: 30(float) ConvertUToF 45
47: 31(fvec4) CompositeConstruct 46 46 46 46
49: 48(ptr) AccessChain 34(mem) 35
Store 49 47
Branch 24
24: Label
50: 6(int) Load 20(i)
52: 6(int) IAdd 50 51
Store 20(i) 52
Branch 21
23: Label
56: 53 Load 55(uni_image)
57: 6(int) Load 8(iid)
58: 40(int) Bitcast 57
60: 59(ivec2) CompositeConstruct 58 58
61: 6(int) Load 16(gid)
62: 48(ptr) AccessChain 34(mem) 61
63: 31(fvec4) Load 62
ImageWrite 56 60 63
64: 53 Load 55(uni_image)
65: 6(int) Load 8(iid)
66: 40(int) Bitcast 65
67: 59(ivec2) CompositeConstruct 66 66
68: 6(int) Load 16(gid)
70: 6(int) IAdd 68 69
71: 48(ptr) AccessChain 34(mem) 70
72: 31(fvec4) Load 71
ImageWrite 64 67 72
MemoryBarrier 69 73
ControlBarrier 74 74 73
85: 84(ptr) AccessChain 80(mytask) 41
Store 85 83
89: 84(ptr) AccessChain 80(mytask) 51 41
Store 89 88
93: 84(ptr) AccessChain 80(mytask) 51 51
Store 93 92
95: 6(int) Load 16(gid)
96: 6(int) UMod 95 74
97: 84(ptr) AccessChain 80(mytask) 51 96
98: 75(fvec2) Load 97
99: 84(ptr) AccessChain 80(mytask) 51 94
Store 99 98
MemoryBarrier 69 73
ControlBarrier 74 74 73
EmitMeshTasksEXT 76 69 69 80(mytask)
Return
FunctionEnd

9
Test/baseResults/test.conf
View File

@ -90,6 +90,15 @@ MaxTaskWorkGroupSizeX_NV 32
MaxTaskWorkGroupSizeY_NV 1
MaxTaskWorkGroupSizeZ_NV 1
MaxMeshViewCountNV 4
MaxMeshOutputVerticesEXT 256
MaxMeshOutputPrimitivesEXT 256
MaxMeshWorkGroupSizeX_EXT 128
MaxMeshWorkGroupSizeY_EXT 128
MaxMeshWorkGroupSizeZ_EXT 128
MaxTaskWorkGroupSizeX_EXT 128
MaxTaskWorkGroupSizeY_EXT 128
MaxTaskWorkGroupSizeZ_EXT 128
MaxMeshViewCountEXT 4
MaxDualSourceDrawBuffersEXT 1
nonInductiveForLoops 1
whileLoops 1

164
Test/spv.460.subgroupEXT.mesh Normal file
View File

@ -0,0 +1,164 @@
#version 460
#define MAX_VER 81
#define MAX_PRIM 32
#define BARRIER() \
memoryBarrierShared(); \
barrier();
#extension GL_EXT_mesh_shader : enable
layout(local_size_x = 32, local_size_y=1, local_size_z=1) in;
layout(max_vertices=MAX_VER) out;
layout(max_primitives=MAX_PRIM) out;
layout(triangles) out;
// test use of builtins in mesh shaders:
void main()
{
uint iid = gl_LocalInvocationID.x;
uint gid = gl_WorkGroupID.x;
uint vertexCount = MAX_VER; // vertexCount <= max_vertices
uint primitiveCount = MAX_PRIM; // primitiveCount <= max_primtives
SetMeshOutputsEXT(vertexCount, primitiveCount);
gl_MeshVerticesEXT[iid].gl_Position = vec4(1.0);
gl_MeshVerticesEXT[iid].gl_PointSize = 2.0;
gl_MeshVerticesEXT[iid].gl_ClipDistance[3] = 3.0;
gl_MeshVerticesEXT[iid].gl_CullDistance[2] = 4.0;
BARRIER();
gl_MeshVerticesEXT[iid+1].gl_Position = gl_MeshVerticesEXT[iid].gl_Position;
gl_MeshVerticesEXT[iid+1].gl_PointSize = gl_MeshVerticesEXT[iid].gl_PointSize;
gl_MeshVerticesEXT[iid+1].gl_ClipDistance[3] = gl_MeshVerticesEXT[iid].gl_ClipDistance[3];
gl_MeshVerticesEXT[iid+1].gl_CullDistance[2] = gl_MeshVerticesEXT[iid].gl_CullDistance[2];
BARRIER();
gl_MeshPrimitivesEXT[iid].gl_PrimitiveID = 6;
gl_MeshPrimitivesEXT[iid].gl_Layer = 7;
gl_MeshPrimitivesEXT[iid].gl_ViewportIndex = 8;
gl_MeshPrimitivesEXT[iid].gl_CullPrimitiveEXT = false;
BARRIER();
gl_MeshPrimitivesEXT[iid+1].gl_PrimitiveID = gl_MeshPrimitivesEXT[iid].gl_PrimitiveID;
gl_MeshPrimitivesEXT[iid+1].gl_Layer = gl_MeshPrimitivesEXT[iid].gl_Layer;
gl_MeshPrimitivesEXT[iid+1].gl_ViewportIndex = gl_MeshPrimitivesEXT[iid].gl_ViewportIndex;
gl_MeshPrimitivesEXT[iid+1].gl_CullPrimitiveEXT = false;
BARRIER();
// check bound limits
gl_PrimitiveTriangleIndicesEXT[0] = uvec3(1, 1, 1); // range is between [0, vertexCount-1]
gl_PrimitiveTriangleIndicesEXT[primitiveCount - 1] = uvec3(2, 2, 2); // array size is primitiveCount*3 for triangle
gl_PrimitiveTriangleIndicesEXT[gid] = gl_PrimitiveTriangleIndicesEXT[gid-1];
BARRIER();
}
#extension GL_KHR_shader_subgroup_basic: enable
void basic_works (void)
{
gl_SubgroupSize;
gl_SubgroupInvocationID;
subgroupBarrier();
subgroupMemoryBarrier();
subgroupMemoryBarrierBuffer();
subgroupMemoryBarrierImage();
subgroupElect();
gl_NumSubgroups; // allowed in mesh
gl_SubgroupID; // allowed in mesh
subgroupMemoryBarrierShared(); // allowed in mesh
}
#extension GL_KHR_shader_subgroup_ballot: enable
void ballot_works(vec4 f4) {
gl_SubgroupEqMask;
gl_SubgroupGeMask;
gl_SubgroupGtMask;
gl_SubgroupLeMask;
gl_SubgroupLtMask;
subgroupBroadcast(f4, 0);
subgroupBroadcastFirst(f4);
uvec4 ballot = subgroupBallot(false);
subgroupInverseBallot(uvec4(0x1));
subgroupBallotBitExtract(ballot, 0);
subgroupBallotBitCount(ballot);
subgroupBallotInclusiveBitCount(ballot);
subgroupBallotExclusiveBitCount(ballot);
subgroupBallotFindLSB(ballot);
subgroupBallotFindMSB(ballot);
}
#extension GL_KHR_shader_subgroup_vote: enable
void vote_works(vec4 f4)
{
subgroupAll(true);
subgroupAny(false);
subgroupAllEqual(f4);
}
#extension GL_KHR_shader_subgroup_shuffle: enable
#extension GL_KHR_shader_subgroup_shuffle_relative: enable
void shuffle_works(vec4 f4)
{
subgroupShuffle(f4, 0);
subgroupShuffleXor(f4, 0x1);
subgroupShuffleUp(f4, 1);
subgroupShuffleDown(f4, 1);
}
#extension GL_KHR_shader_subgroup_arithmetic: enable
void arith_works(vec4 f4)
{
uvec4 ballot;
subgroupAdd(f4);
subgroupMul(f4);
subgroupMin(f4);
subgroupMax(f4);
subgroupAnd(ballot);
subgroupOr(ballot);
subgroupXor(ballot);
subgroupInclusiveAdd(f4);
subgroupInclusiveMul(f4);
subgroupInclusiveMin(f4);
subgroupInclusiveMax(f4);
subgroupInclusiveAnd(ballot);
subgroupInclusiveOr(ballot);
subgroupInclusiveXor(ballot);
subgroupExclusiveAdd(f4);
subgroupExclusiveMul(f4);
subgroupExclusiveMin(f4);
subgroupExclusiveMax(f4);
subgroupExclusiveAnd(ballot);
subgroupExclusiveOr(ballot);
subgroupExclusiveXor(ballot);
}
#extension GL_KHR_shader_subgroup_clustered: enable
void clustered_works(vec4 f4)
{
uvec4 ballot = uvec4(0x55,0,0,0);
subgroupClusteredAdd(f4, 2);
subgroupClusteredMul(f4, 2);
subgroupClusteredMin(f4, 2);
subgroupClusteredMax(f4, 2);
subgroupClusteredAnd(ballot, 2);
subgroupClusteredOr(ballot, 2);
subgroupClusteredXor(ballot, 2);
}
#extension GL_KHR_shader_subgroup_quad: enable
void quad_works(vec4 f4)
{
subgroupQuadBroadcast(f4, 0);
subgroupQuadSwapHorizontal(f4);
subgroupQuadSwapVertical(f4);
subgroupQuadSwapDiagonal(f4);
}

153
Test/spv.460.subgroupEXT.task Normal file
View File

@ -0,0 +1,153 @@
#version 460
#define BARRIER() \
memoryBarrierShared(); \
barrier();
#extension GL_EXT_mesh_shader : enable
layout(local_size_x = 32, local_size_y=1, local_size_z=1) in;
// test use of shared memory in task shaders:
layout(binding=0) writeonly uniform image2D uni_image;
uniform block0 {
uint uni_value;
};
shared vec4 mem[10];
// test use of task memory in task shaders:
struct Task {
vec2 dummy;
vec2 submesh[3];
};
taskPayloadSharedEXT Task mytask;
void main()
{
uint iid = gl_LocalInvocationID.x;
uint gid = gl_WorkGroupID.x;
// 1. shared memory load and stores
for (uint i = 0; i < 10; ++i) {
mem[i] = vec4(i + uni_value);
}
imageStore(uni_image, ivec2(iid), mem[gid]);
imageStore(uni_image, ivec2(iid), mem[gid+1]);
BARRIER();
// 2. task memory stores
mytask.dummy = vec2(30.0, 31.0);
mytask.submesh[0] = vec2(32.0, 33.0);
mytask.submesh[1] = vec2(34.0, 35.0);
mytask.submesh[2] = mytask.submesh[gid%2];
BARRIER();
// 3. emit task count under uniform control flow
EmitMeshTasksEXT(3U, 1U, 1U);
}
#extension GL_KHR_shader_subgroup_basic: enable
void basic_works (void)
{
gl_SubgroupSize;
gl_SubgroupInvocationID;
subgroupBarrier();
subgroupMemoryBarrier();
subgroupMemoryBarrierBuffer();
subgroupMemoryBarrierImage();
subgroupElect();
gl_NumSubgroups; // allowed in task
gl_SubgroupID; // allowed in task
subgroupMemoryBarrierShared(); // allowed in task
}
#extension GL_KHR_shader_subgroup_ballot: enable
void ballot_works(vec4 f4) {
gl_SubgroupEqMask;
gl_SubgroupGeMask;
gl_SubgroupGtMask;
gl_SubgroupLeMask;
gl_SubgroupLtMask;
subgroupBroadcast(f4, 0);
subgroupBroadcastFirst(f4);
uvec4 ballot = subgroupBallot(false);
subgroupInverseBallot(uvec4(0x1));
subgroupBallotBitExtract(ballot, 0);
subgroupBallotBitCount(ballot);
subgroupBallotInclusiveBitCount(ballot);
subgroupBallotExclusiveBitCount(ballot);
subgroupBallotFindLSB(ballot);
subgroupBallotFindMSB(ballot);
}
#extension GL_KHR_shader_subgroup_vote: enable
void vote_works(vec4 f4)
{
subgroupAll(true);
subgroupAny(false);
subgroupAllEqual(f4);
}
#extension GL_KHR_shader_subgroup_shuffle: enable
#extension GL_KHR_shader_subgroup_shuffle_relative: enable
void shuffle_works(vec4 f4)
{
subgroupShuffle(f4, 0);
subgroupShuffleXor(f4, 0x1);
subgroupShuffleUp(f4, 1);
subgroupShuffleDown(f4, 1);
}
#extension GL_KHR_shader_subgroup_arithmetic: enable
void arith_works(vec4 f4)
{
uvec4 ballot;
subgroupAdd(f4);
subgroupMul(f4);
subgroupMin(f4);
subgroupMax(f4);
subgroupAnd(ballot);
subgroupOr(ballot);
subgroupXor(ballot);
subgroupInclusiveAdd(f4);
subgroupInclusiveMul(f4);
subgroupInclusiveMin(f4);
subgroupInclusiveMax(f4);
subgroupInclusiveAnd(ballot);
subgroupInclusiveOr(ballot);
subgroupInclusiveXor(ballot);
subgroupExclusiveAdd(f4);
subgroupExclusiveMul(f4);
subgroupExclusiveMin(f4);
subgroupExclusiveMax(f4);
subgroupExclusiveAnd(ballot);
subgroupExclusiveOr(ballot);
subgroupExclusiveXor(ballot);
}
#extension GL_KHR_shader_subgroup_clustered: enable
void clustered_works(vec4 f4)
{
uvec4 ballot = uvec4(0x55,0,0,0);
subgroupClusteredAdd(f4, 2);
subgroupClusteredMul(f4, 2);
subgroupClusteredMin(f4, 2);
subgroupClusteredMax(f4, 2);
subgroupClusteredAnd(ballot, 2);
subgroupClusteredOr(ballot, 2);
subgroupClusteredXor(ballot, 2);
}
#extension GL_KHR_shader_subgroup_quad: enable
void quad_works(vec4 f4)
{
subgroupQuadBroadcast(f4, 0);
subgroupQuadSwapHorizontal(f4);
subgroupQuadSwapVertical(f4);
subgroupQuadSwapDiagonal(f4);
}

22
Test/spv.atomiAddEXT.error.mesh Normal file
View File

@ -0,0 +1,22 @@
#version 460
#extension GL_EXT_mesh_shader : enable
#define MAX_VER 81
#define MAX_PRIM 32
layout(local_size_x = 1) in;
layout(max_vertices=MAX_VER) out;
layout(max_primitives=MAX_PRIM) out;
layout(triangles) out;
// use of storage qualifier "taskPayloadSharedEXT" in mesh shaders:
struct taskBlock {
int atom1;
};
taskPayloadSharedEXT taskBlock mytask;
void main() {
atomicAdd(mytask.atom1, 1);
}

27
Test/spv.atomiAddEXT.task Normal file
View File

@ -0,0 +1,27 @@
#version 460
#extension GL_EXT_mesh_shader : enable
layout(local_size_x = 1) in;
struct structType{
int y[3];
};
layout(std430) buffer t2 {
structType f;
} t;
buffer coherent Buffer { int x; };
// use of storage qualifier "taskPayloadSharedEXT" in mesh shaders:
struct taskBlock {
int atom1;
};
taskPayloadSharedEXT taskBlock mytask;
void main() {
atomicAdd(x, 1);
atomicAdd(t.f.y[1], 1);
atomicAdd(mytask.atom1, 1);
}

74
Test/spv.ext.meshShaderBuiltins.mesh Normal file
View File

@ -0,0 +1,74 @@
#version 460
#define MAX_VER 81
#define MAX_PRIM 32
#define BARRIER() \
memoryBarrierShared(); \
barrier();
#extension GL_EXT_mesh_shader : enable
layout(local_size_x = 32, local_size_y=1, local_size_z=1) in;
layout(max_vertices=MAX_VER) out;
layout(max_primitives=MAX_PRIM) out;
layout(triangles) out;
// test use of builtins in mesh shaders:
void main()
{
uint iid = gl_LocalInvocationID.x;
uint gid = gl_WorkGroupID.x;
uvec3 numWorkGrous = gl_NumWorkGroups;
uint vertexCount = MAX_VER; // vertexCount <= max_vertices
uint primitiveCount = MAX_PRIM; // primitiveCount <= max_primtives
SetMeshOutputsEXT(vertexCount, primitiveCount);
gl_MeshVerticesEXT[iid].gl_Position = vec4(1.0);
gl_MeshVerticesEXT[iid].gl_PointSize = 2.0;
gl_MeshVerticesEXT[iid].gl_ClipDistance[3] = 3.0;
gl_MeshVerticesEXT[iid].gl_CullDistance[2] = 4.0;
BARRIER();
gl_MeshVerticesEXT[iid+1].gl_Position = gl_MeshVerticesEXT[iid].gl_Position;
gl_MeshVerticesEXT[iid+1].gl_PointSize = gl_MeshVerticesEXT[iid].gl_PointSize;
gl_MeshVerticesEXT[iid+1].gl_ClipDistance[3] = gl_MeshVerticesEXT[iid].gl_ClipDistance[3];
gl_MeshVerticesEXT[iid+1].gl_CullDistance[2] = gl_MeshVerticesEXT[iid].gl_CullDistance[2];
BARRIER();
gl_MeshPrimitivesEXT[iid].gl_PrimitiveID = 6;
gl_MeshPrimitivesEXT[iid].gl_Layer = 7;
gl_MeshPrimitivesEXT[iid].gl_ViewportIndex = 8;
gl_MeshPrimitivesEXT[iid].gl_CullPrimitiveEXT = false;
BARRIER();
gl_MeshPrimitivesEXT[iid+1].gl_PrimitiveID = gl_MeshPrimitivesEXT[iid].gl_PrimitiveID;
gl_MeshPrimitivesEXT[iid+1].gl_Layer = gl_MeshPrimitivesEXT[iid].gl_Layer;
gl_MeshPrimitivesEXT[iid+1].gl_ViewportIndex = gl_MeshPrimitivesEXT[iid].gl_ViewportIndex;
gl_MeshPrimitivesEXT[iid+1].gl_CullPrimitiveEXT = gl_MeshPrimitivesEXT[iid].gl_CullPrimitiveEXT;
BARRIER();
// check bound limits
gl_PrimitiveTriangleIndicesEXT[0] = uvec3(257); // should truncate 257 -> 1, range is between [0, vertexCount-1]
gl_PrimitiveTriangleIndicesEXT[primitiveCount - 1] = uvec3(2); // array size is primitiveCount*3 for triangle
gl_PrimitiveTriangleIndicesEXT[gid] = gl_PrimitiveTriangleIndicesEXT[gid-1];
BARRIER();
}
// test use of builtins enabled by other extensions
#extension GL_ARB_shader_draw_parameters : enable
#extension GL_EXT_multiview : enable
void testAdditionalBuiltins()
{
int id = gl_DrawIDARB; // GL_ARB_shader_draw_parameters
int viewIdx = gl_ViewIndex; // GL_EXT_multiview
}

75
Test/spv.ext.meshShaderRedeclBuiltins.mesh Normal file
View File

@ -0,0 +1,75 @@
#version 460
#define MAX_VER 81
#define MAX_PRIM 32
#define BARRIER() \
memoryBarrierShared(); \
barrier();
#extension GL_EXT_mesh_shader : enable
layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
layout(max_vertices=MAX_VER) out;
layout(max_primitives=MAX_PRIM) out;
layout(points) out;
// test use of redeclared single-view builtins in mesh shaders:
out gl_MeshPerVertexEXT {
vec4 gl_Position;
float gl_PointSize;
float gl_ClipDistance[4];
float gl_CullDistance[4];
} gl_MeshVerticesEXT[MAX_VER]; // explicitly sized to MAX_VER
perprimitiveEXT out gl_MeshPerPrimitiveEXT {
int gl_PrimitiveID;
int gl_Layer;
int gl_ViewportIndex;
bool gl_CullPrimitiveEXT;
int gl_PrimitiveShadingRateEXT;
} gl_MeshPrimitivesEXT[]; // implicitly sized to MAX_PRIM
out uint gl_PrimitivePointIndicesEXT[MAX_PRIM]; // explicitly sized to MAX_PRIM
void main()
{
uint iid = gl_LocalInvocationID.x;
uint gid = gl_WorkGroupID.x;
SetMeshOutputsEXT(MAX_VER, MAX_PRIM);
gl_MeshVerticesEXT[iid].gl_Position = vec4(1.0);
gl_MeshVerticesEXT[iid].gl_PointSize = 2.0;
gl_MeshVerticesEXT[iid].gl_ClipDistance[3] = 3.0;
gl_MeshVerticesEXT[iid].gl_CullDistance[2] = 4.0;
BARRIER();
gl_MeshVerticesEXT[iid+1].gl_Position = gl_MeshVerticesEXT[iid].gl_Position;
gl_MeshVerticesEXT[iid+1].gl_PointSize = gl_MeshVerticesEXT[iid].gl_PointSize;
gl_MeshVerticesEXT[iid+1].gl_ClipDistance[3] = gl_MeshVerticesEXT[iid].gl_ClipDistance[3];
gl_MeshVerticesEXT[iid+1].gl_CullDistance[2] = gl_MeshVerticesEXT[iid].gl_CullDistance[2];
BARRIER();
gl_MeshPrimitivesEXT[iid].gl_PrimitiveID = 6;
gl_MeshPrimitivesEXT[iid].gl_Layer = 7;
gl_MeshPrimitivesEXT[iid].gl_ViewportIndex = 8;
gl_MeshPrimitivesEXT[iid].gl_CullPrimitiveEXT = false;
BARRIER();
gl_MeshPrimitivesEXT[iid+1].gl_PrimitiveID = gl_MeshPrimitivesEXT[iid].gl_PrimitiveID;
gl_MeshPrimitivesEXT[iid+1].gl_Layer = gl_MeshPrimitivesEXT[iid].gl_Layer;
gl_MeshPrimitivesEXT[iid+1].gl_ViewportIndex = gl_MeshPrimitivesEXT[iid].gl_ViewportIndex;
gl_MeshPrimitivesEXT[iid+1].gl_CullPrimitiveEXT = gl_MeshPrimitivesEXT[iid].gl_CullPrimitiveEXT;
BARRIER();
// check bound limits
gl_PrimitivePointIndicesEXT[0] = 1;
gl_PrimitivePointIndicesEXT[MAX_PRIM - 1] = 2;
}

41
Test/spv.ext.meshShaderTaskMem.mesh Normal file
View File

@ -0,0 +1,41 @@
#version 450
#define MAX_VER 81
#define MAX_PRIM 32
#define BARRIER() \
memoryBarrierShared(); \
barrier();
#extension GL_EXT_mesh_shader : enable
layout(local_size_x = 32) in;
layout(max_vertices=MAX_VER) out;
layout(max_primitives=MAX_PRIM) out;
layout(triangles) out;
// use of storage qualifier "taskPayloadSharedEXT" in mesh shaders:
struct taskBlock {
float gid1[2];
vec4 gid2;
};
taskPayloadSharedEXT taskBlock mytask;
buffer bufferBlock {
float gid3[2];
vec4 gid4;
} mybuf;
layout(location=0) out outBlock {
float gid5;
vec4 gid6;
} myblk[];
void main()
{
uint iid = gl_LocalInvocationID.x;
myblk[iid].gid5 = mytask.gid1[1] + mybuf.gid3[1];
myblk[iid].gid6 = mytask.gid2 + mybuf.gid4;
}

59
Test/spv.ext.meshShaderUserDefined.mesh Normal file
View File

@ -0,0 +1,59 @@
#version 450
#define MAX_VER 81
#define MAX_PRIM 32
#define BARRIER() \
memoryBarrierShared(); \
barrier();
#extension GL_EXT_mesh_shader : enable
layout(local_size_x = 32) in;
layout(max_vertices=MAX_VER) out;
layout(max_primitives=MAX_PRIM) out;
layout(triangles) out;
// test use of user defined interface out blocks:
// per-primitive block
perprimitiveEXT layout(location=0) out myblock {
float f;
float fArr[4];
vec3 pos;
vec4 posArr[4];
mat4 m;
mat3 mArr[2];
} blk[];
// per-vertex block
layout(location=20) out myblock2 {
float f;
vec4 pos;
mat4 m;
} blk2[];
void main()
{
uint iid = gl_LocalInvocationID.x;
uint gid = gl_WorkGroupID.x;
blk[iid].f = 11.0;
blk[iid+1].fArr[gid] = blk[iid].f;
blk[iid/2].pos.yzx = vec3(14.0, 15.0, 13.0);
blk[iid*2].posArr[1].yzw = blk[iid/2].pos;
blk[iid/4].m[2].wzyx = vec4(13.0, 14.0, 15.0, 16.0);
blk[iid].mArr[0][1][1] = blk[iid/4].m[2].w;
blk[iid*4].mArr[1][gid] = vec3(17.0, 18.0, 19.0);
BARRIER();
blk2[iid].f = blk2[iid-1].f + 20.0;
blk2[iid].pos = vec4(21.0, 22.0, 23.0, 24.0);
blk2[iid+1].m[gid] = blk2[iid].pos;
blk2[iid+1].m[gid][2] = 29.0;
blk2[iid+2].m[3] = blk2[iid+1].m[gid];
BARRIER();
}

51
Test/spv.ext.meshTaskShader.task Normal file
View File

@ -0,0 +1,51 @@
#version 450
#define BARRIER() \
memoryBarrierShared(); \
barrier();
#extension GL_EXT_mesh_shader : enable
layout(local_size_x = 32) in;
// test use of shared memory in task shaders:
layout(binding=0) writeonly uniform image2D uni_image;
uniform block0 {
uint uni_value;
};
shared vec4 mem[10];
// use of storage qualifier "taskPayloadSharedEXT" in task shaders
struct Task {
vec2 dummy;
vec2 submesh[3];
};
taskPayloadSharedEXT Task mytask;
void main()
{
uint iid = gl_LocalInvocationID.x;
uint gid = gl_WorkGroupID.x;
// 1. shared memory load and stores
for (uint i = 0; i < 10; ++i) {
mem[i] = vec4(i + uni_value);
}
imageStore(uni_image, ivec2(iid), mem[gid]);
imageStore(uni_image, ivec2(iid), mem[gid+1]);
BARRIER();
// 2. task memory stores
mytask.dummy = vec2(30.0, 31.0);
mytask.submesh[0] = vec2(32.0, 33.0);
mytask.submesh[1] = vec2(34.0, 35.0);
mytask.submesh[2] = mytask.submesh[gid%2];
BARRIER();
// 3. emit task count under uniform control flow
EmitMeshTasksEXT(3U, 1U, 1U);
}

8
glslang/CInterface/glslang_c_interface.cpp
View File

@ -192,10 +192,10 @@ static EShLanguage c_shader_stage(glslang_stage_t stage)
return EShLangMiss;
case GLSLANG_STAGE_CALLABLE_NV:
return EShLangCallable;
case GLSLANG_STAGE_TASK_NV:
return EShLangTaskNV;
case GLSLANG_STAGE_MESH_NV:
return EShLangMeshNV;
case GLSLANG_STAGE_TASK:
return EShLangTask;
case GLSLANG_STAGE_MESH:
return EShLangMesh;
default:
break;
}

13
glslang/Include/BaseTypes.h
View File

@ -105,6 +105,8 @@ enum TStorageQualifier {
EvqCallableData,
EvqCallableDataIn,
EvqtaskPayloadSharedEXT,
// parameters
EvqIn, // also, for 'in' in the grammar before we know if it's a pipeline input or an 'in' parameter
EvqOut, // also, for 'out' in the grammar before we know if it's a pipeline output or an 'out' parameter
@ -287,6 +289,11 @@ enum TBuiltInVariable {
EbvLayerPerViewNV,
EbvMeshViewCountNV,
EbvMeshViewIndicesNV,
//GL_EXT_mesh_shader
EbvPrimitivePointIndicesEXT,
EbvPrimitiveLineIndicesEXT,
EbvPrimitiveTriangleIndicesEXT,
EbvCullPrimitiveEXT,
// sm builtins
EbvWarpsPerSM,
@ -360,6 +367,7 @@ __inline const char* GetStorageQualifierString(TStorageQualifier q)
case EvqHitAttr: return "hitAttributeNV"; break;
case EvqCallableData: return "callableDataNV"; break;
case EvqCallableDataIn: return "callableDataInNV"; break;
case EvqtaskPayloadSharedEXT: return "taskPayloadSharedEXT"; break;
default: return "unknown qualifier";
}
}
@ -496,6 +504,11 @@ __inline const char* GetBuiltInVariableString(TBuiltInVariable v)
case EbvLayerPerViewNV: return "LayerPerViewNV";
case EbvMeshViewCountNV: return "MeshViewCountNV";
case EbvMeshViewIndicesNV: return "MeshViewIndicesNV";
// GL_EXT_mesh_shader
case EbvPrimitivePointIndicesEXT: return "PrimitivePointIndicesEXT";
case EbvPrimitiveLineIndicesEXT: return "PrimitiveLineIndicesEXT";
case EbvPrimitiveTriangleIndicesEXT: return "PrimitiveTriangleIndicesEXT";
case EbvCullPrimitiveEXT: return "CullPrimitiveEXT";
case EbvWarpsPerSM: return "WarpsPerSMNV";
case EbvSMCount: return "SMCountNV";

9
glslang/Include/ResourceLimits.h
View File

@ -142,6 +142,15 @@ struct TBuiltInResource {
int maxTaskWorkGroupSizeY_NV;
int maxTaskWorkGroupSizeZ_NV;
int maxMeshViewCountNV;
int maxMeshOutputVerticesEXT;
int maxMeshOutputPrimitivesEXT;
int maxMeshWorkGroupSizeX_EXT;
int maxMeshWorkGroupSizeY_EXT;
int maxMeshWorkGroupSizeZ_EXT;
int maxTaskWorkGroupSizeX_EXT;
int maxTaskWorkGroupSizeY_EXT;
int maxTaskWorkGroupSizeZ_EXT;
int maxMeshViewCountEXT;
int maxDualSourceDrawBuffersEXT;
TLimits limits;

5
glslang/Include/Types.h
View File

@ -856,6 +856,7 @@ public:
bool isPerPrimitive() const { return perPrimitiveNV; }
bool isPerView() const { return perViewNV; }
bool isTaskMemory() const { return perTaskNV; }
bool isTaskPayload() const { return storage == EvqtaskPayloadSharedEXT; }
bool isAnyPayload() const {
return storage == EvqPayload || storage == EvqPayloadIn;
}
@ -874,8 +875,8 @@ public:
case EShLangTessEvaluation:
return ! patch && isPipeInput();
case EShLangFragment:
return (pervertexNV || pervertexEXT) && isPipeInput();
case EShLangMeshNV:
return pervertexNV && isPipeInput();
case EShLangMesh:
return ! perTaskNV && isPipeOutput();
default:

9
glslang/Include/glslang_c_interface.h
View File

@ -148,6 +148,15 @@ typedef struct glslang_resource_s {
int max_task_work_group_size_y_nv;
int max_task_work_group_size_z_nv;
int max_mesh_view_count_nv;
int max_mesh_output_vertices_ext;
int max_mesh_output_primitives_ext;
int max_mesh_work_group_size_x_ext;
int max_mesh_work_group_size_y_ext;
int max_mesh_work_group_size_z_ext;
int max_task_work_group_size_x_ext;
int max_task_work_group_size_y_ext;
int max_task_work_group_size_z_ext;
int max_mesh_view_count_ext;
int maxDualSourceDrawBuffersEXT;
glslang_limits_t limits;

12
glslang/Include/glslang_c_shader_types.h
View File

@ -49,8 +49,10 @@ typedef enum {
GLSLANG_STAGE_CLOSESTHIT_NV,
GLSLANG_STAGE_MISS_NV,
GLSLANG_STAGE_CALLABLE_NV,
GLSLANG_STAGE_TASK_NV,
GLSLANG_STAGE_MESH_NV,
GLSLANG_STAGE_TASK,
GLSLANG_STAGE_TASK_NV = GLSLANG_STAGE_TASK,
GLSLANG_STAGE_MESH,
GLSLANG_STAGE_MESH_NV = GLSLANG_STAGE_MESH,
LAST_ELEMENT_MARKER(GLSLANG_STAGE_COUNT),
} glslang_stage_t; // would be better as stage, but this is ancient now
@ -68,8 +70,10 @@ typedef enum {
GLSLANG_STAGE_CLOSESTHIT_NV_MASK = (1 << GLSLANG_STAGE_CLOSESTHIT_NV),
GLSLANG_STAGE_MISS_NV_MASK = (1 << GLSLANG_STAGE_MISS_NV),
GLSLANG_STAGE_CALLABLE_NV_MASK = (1 << GLSLANG_STAGE_CALLABLE_NV),
GLSLANG_STAGE_TASK_NV_MASK = (1 << GLSLANG_STAGE_TASK_NV),
GLSLANG_STAGE_MESH_NV_MASK = (1 << GLSLANG_STAGE_MESH_NV),
GLSLANG_STAGE_TASK_MASK = (1 << GLSLANG_STAGE_TASK),
GLSLANG_STAGE_TASK_NV_MASK = GLSLANG_STAGE_TASK_MASK,
GLSLANG_STAGE_MESH_MASK = (1 << GLSLANG_STAGE_MESH),
GLSLANG_STAGE_MESH_NV_MASK = GLSLANG_STAGE_MESH_MASK,
LAST_ELEMENT_MARKER(GLSLANG_STAGE_MASK_COUNT),
} glslang_stage_mask_t;

2
glslang/Include/intermediate.h
View File

@ -934,6 +934,8 @@ enum TOperator {
EOpExecuteCallableNV,
EOpExecuteCallableKHR,
EOpWritePackedPrimitiveIndices4x8NV,
EOpEmitMeshTasksEXT,
EOpSetMeshOutputsEXT,
//
// GL_EXT_ray_query operations

211
glslang/MachineIndependent/Initialize.cpp
View File

@ -2268,11 +2268,11 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
"\n"
);
stageBuiltins[EShLangMeshNV].append(
stageBuiltins[EShLangMesh].append(
"void subgroupMemoryBarrierShared();"
"\n"
);
stageBuiltins[EShLangTaskNV].append(
stageBuiltins[EShLangTask].append(
"void subgroupMemoryBarrierShared();"
"\n"
);
@ -4298,10 +4298,10 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
"void barrier();"
);
if ((profile != EEsProfile && version >= 450) || (profile == EEsProfile && version >= 320)) {
stageBuiltins[EShLangMeshNV].append(
stageBuiltins[EShLangMesh].append(
"void barrier();"
);
stageBuiltins[EShLangTaskNV].append(
stageBuiltins[EShLangTask].append(
"void barrier();"
);
}
@ -4326,11 +4326,11 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
commonBuiltins.append("void memoryBarrierImage();");
}
if ((profile != EEsProfile && version >= 450) || (profile == EEsProfile && version >= 320)) {
stageBuiltins[EShLangMeshNV].append(
stageBuiltins[EShLangMesh].append(
"void memoryBarrierShared();"
"void groupMemoryBarrier();"
);
stageBuiltins[EShLangTaskNV].append(
stageBuiltins[EShLangTask].append(
"void memoryBarrierShared();"
"void groupMemoryBarrier();"
);
@ -4655,10 +4655,21 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
// Builtins for GL_NV_mesh_shader
if ((profile != EEsProfile && version >= 450) || (profile == EEsProfile && version >= 320)) {
stageBuiltins[EShLangMeshNV].append(
stageBuiltins[EShLangMesh].append(
"void writePackedPrimitiveIndices4x8NV(uint, uint);"
"\n");
}
// Builtins for GL_EXT_mesh_shader
if ((profile != EEsProfile && version >= 450) || (profile == EEsProfile && version >= 320)) {
// Builtins for GL_EXT_mesh_shader
stageBuiltins[EShLangTask].append(
"void EmitMeshTasksEXT(uint, uint, uint);"
"\n");
stageBuiltins[EShLangMesh].append(
"void SetMeshOutputsEXT(uint, uint);"
"\n");
}
#endif // !GLSLANG_ANGLE
#endif // !GLSLANG_WEB
@ -4855,7 +4866,7 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
if ((profile != EEsProfile && version >= 450) || (profile == EEsProfile && version >= 320)) {
// per-vertex attributes
stageBuiltins[EShLangMeshNV].append(
stageBuiltins[EShLangMesh].append(
"out gl_MeshPerVertexNV {"
"vec4 gl_Position;"
"float gl_PointSize;"
@ -4868,7 +4879,7 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
);
// per-primitive attributes
stageBuiltins[EShLangMeshNV].append(
stageBuiltins[EShLangMesh].append(
"perprimitiveNV out gl_MeshPerPrimitiveNV {"
"int gl_PrimitiveID;"
"int gl_Layer;"
@ -4879,7 +4890,7 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
"} gl_MeshPrimitivesNV[];"
);
stageBuiltins[EShLangMeshNV].append(
stageBuiltins[EShLangMesh].append(
"out uint gl_PrimitiveCountNV;"
"out uint gl_PrimitiveIndicesNV[];"
@ -4893,10 +4904,38 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
"in highp uvec3 gl_GlobalInvocationID;"
"in highp uint gl_LocalInvocationIndex;"
"\n");
stageBuiltins[EShLangTaskNV].append(
// GL_EXT_mesh_shader
stageBuiltins[EShLangMesh].append(
"out uint gl_PrimitivePointIndicesEXT[];"
"out uvec2 gl_PrimitiveLineIndicesEXT[];"
"out uvec3 gl_PrimitiveTriangleIndicesEXT[];"
"in highp uvec3 gl_NumWorkGroups;"
"\n");
// per-vertex attributes
stageBuiltins[EShLangMesh].append(
"out gl_MeshPerVertexEXT {"
"vec4 gl_Position;"
"float gl_PointSize;"
"float gl_ClipDistance[];"
"float gl_CullDistance[];"
"} gl_MeshVerticesEXT[];"
);
// per-primitive attributes
stageBuiltins[EShLangMesh].append(
"perprimitiveEXT out gl_MeshPerPrimitiveEXT {"
"int gl_PrimitiveID;"
"int gl_Layer;"
"int gl_ViewportIndex;"
"bool gl_CullPrimitiveEXT;"
"int gl_PrimitiveShadingRateEXT;"
"} gl_MeshPrimitivesEXT[];"
);
stageBuiltins[EShLangTask].append(
"out uint gl_TaskCountNV;"
"const highp uvec3 gl_WorkGroupSize = uvec3(1,1,1);"
@ -4909,27 +4948,28 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
"in uint gl_MeshViewCountNV;"
"in uint gl_MeshViewIndicesNV[4];"
"in highp uvec3 gl_NumWorkGroups;"
"\n");
}
if (profile != EEsProfile && version >= 450) {
stageBuiltins[EShLangMeshNV].append(
stageBuiltins[EShLangMesh].append(
"in highp int gl_DeviceIndex;" // GL_EXT_device_group
"in int gl_DrawIDARB;" // GL_ARB_shader_draw_parameters
"in int gl_ViewIndex;" // GL_EXT_multiview
"\n");
stageBuiltins[EShLangTaskNV].append(
stageBuiltins[EShLangTask].append(
"in highp int gl_DeviceIndex;" // GL_EXT_device_group
"in int gl_DrawIDARB;" // GL_ARB_shader_draw_parameters
"\n");
if (version >= 460) {
stageBuiltins[EShLangMeshNV].append(
stageBuiltins[EShLangMesh].append(
"in int gl_DrawID;"
"\n");
stageBuiltins[EShLangTaskNV].append(
stageBuiltins[EShLangTask].append(
"in int gl_DrawID;"
"\n");
}
@ -5704,8 +5744,8 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
stageBuiltins[EShLangGeometry] .append(ballotDecls);
stageBuiltins[EShLangCompute] .append(ballotDecls);
stageBuiltins[EShLangFragment] .append(fragmentBallotDecls);
stageBuiltins[EShLangMeshNV] .append(ballotDecls);
stageBuiltins[EShLangTaskNV] .append(ballotDecls);
stageBuiltins[EShLangMesh] .append(ballotDecls);
stageBuiltins[EShLangTask] .append(ballotDecls);
stageBuiltins[EShLangRayGen] .append(rtBallotDecls);
stageBuiltins[EShLangIntersect] .append(rtBallotDecls);
// No volatile qualifier on these builtins in any-hit
@ -5773,10 +5813,10 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
stageBuiltins[EShLangCompute] .append(subgroupDecls);
stageBuiltins[EShLangCompute] .append(computeSubgroupDecls);
stageBuiltins[EShLangFragment] .append(fragmentSubgroupDecls);
stageBuiltins[EShLangMeshNV] .append(subgroupDecls);
stageBuiltins[EShLangMeshNV] .append(computeSubgroupDecls);
stageBuiltins[EShLangTaskNV] .append(subgroupDecls);
stageBuiltins[EShLangTaskNV] .append(computeSubgroupDecls);
stageBuiltins[EShLangMesh] .append(subgroupDecls);
stageBuiltins[EShLangMesh] .append(computeSubgroupDecls);
stageBuiltins[EShLangTask] .append(subgroupDecls);
stageBuiltins[EShLangTask] .append(computeSubgroupDecls);
stageBuiltins[EShLangRayGen] .append(rtSubgroupDecls);
stageBuiltins[EShLangIntersect] .append(rtSubgroupDecls);
// No volatile qualifier on these builtins in any-hit
@ -8883,7 +8923,7 @@ void TBuiltIns::identifyBuiltIns(int version, EProfile profile, const SpvVersion
}
break;
case EShLangMeshNV:
case EShLangMesh:
if ((profile != EEsProfile && version >= 450) || (profile == EEsProfile && version >= 320)) {
// per-vertex builtins
symbolTable.setVariableExtensions("gl_MeshVerticesNV", "gl_Position", 1, &E_GL_NV_mesh_shader);
@ -8927,12 +8967,19 @@ void TBuiltIns::identifyBuiltIns(int version, EProfile profile, const SpvVersion
symbolTable.setVariableExtensions("gl_PrimitiveIndicesNV", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_MeshViewCountNV", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_MeshViewIndicesNV", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_WorkGroupSize", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_WorkGroupID", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_LocalInvocationID", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_GlobalInvocationID", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_LocalInvocationIndex", 1, &E_GL_NV_mesh_shader);
if (profile != EEsProfile) {
symbolTable.setVariableExtensions("gl_WorkGroupSize", Num_AEP_mesh_shader, AEP_mesh_shader);
symbolTable.setVariableExtensions("gl_WorkGroupID", Num_AEP_mesh_shader, AEP_mesh_shader);
symbolTable.setVariableExtensions("gl_LocalInvocationID", Num_AEP_mesh_shader, AEP_mesh_shader);
symbolTable.setVariableExtensions("gl_GlobalInvocationID", Num_AEP_mesh_shader, AEP_mesh_shader);
symbolTable.setVariableExtensions("gl_LocalInvocationIndex", Num_AEP_mesh_shader, AEP_mesh_shader);
} else {
symbolTable.setVariableExtensions("gl_WorkGroupSize", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_WorkGroupID", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_LocalInvocationID", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_GlobalInvocationID", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_LocalInvocationIndex", 1, &E_GL_NV_mesh_shader);
}
BuiltInVariable("gl_PrimitiveCountNV", EbvPrimitiveCountNV, symbolTable);
BuiltInVariable("gl_PrimitiveIndicesNV", EbvPrimitiveIndicesNV, symbolTable);
BuiltInVariable("gl_MeshViewCountNV", EbvMeshViewCountNV, symbolTable);
@ -8950,12 +8997,54 @@ void TBuiltIns::identifyBuiltIns(int version, EProfile profile, const SpvVersion
symbolTable.setVariableExtensions("gl_MaxMeshViewCountNV", 1, &E_GL_NV_mesh_shader);
// builtin functions
symbolTable.setFunctionExtensions("barrier", 1, &E_GL_NV_mesh_shader);
symbolTable.setFunctionExtensions("memoryBarrierShared", 1, &E_GL_NV_mesh_shader);
symbolTable.setFunctionExtensions("groupMemoryBarrier", 1, &E_GL_NV_mesh_shader);
if (profile != EEsProfile) {
symbolTable.setFunctionExtensions("barrier", Num_AEP_mesh_shader, AEP_mesh_shader);
symbolTable.setFunctionExtensions("memoryBarrierShared", Num_AEP_mesh_shader, AEP_mesh_shader);
symbolTable.setFunctionExtensions("groupMemoryBarrier", Num_AEP_mesh_shader, AEP_mesh_shader);
} else {
symbolTable.setFunctionExtensions("barrier", 1, &E_GL_NV_mesh_shader);
symbolTable.setFunctionExtensions("memoryBarrierShared", 1, &E_GL_NV_mesh_shader);
symbolTable.setFunctionExtensions("groupMemoryBarrier", 1, &E_GL_NV_mesh_shader);
}
symbolTable.setFunctionExtensions("writePackedPrimitiveIndices4x8NV", 1, &E_GL_NV_mesh_shader);
}
if (profile != EEsProfile && version >= 450) {
// GL_EXT_Mesh_shader
symbolTable.setVariableExtensions("gl_PrimitivePointIndicesEXT", 1, &E_GL_EXT_mesh_shader);
symbolTable.setVariableExtensions("gl_PrimitiveLineIndicesEXT", 1, &E_GL_EXT_mesh_shader);
symbolTable.setVariableExtensions("gl_PrimitiveTriangleIndicesEXT", 1, &E_GL_EXT_mesh_shader);
symbolTable.setVariableExtensions("gl_NumWorkGroups", 1, &E_GL_EXT_mesh_shader);
BuiltInVariable("gl_PrimitivePointIndicesEXT", EbvPrimitivePointIndicesEXT, symbolTable);
BuiltInVariable("gl_PrimitiveLineIndicesEXT", EbvPrimitiveLineIndicesEXT, symbolTable);
BuiltInVariable("gl_PrimitiveTriangleIndicesEXT", EbvPrimitiveTriangleIndicesEXT, symbolTable);
BuiltInVariable("gl_NumWorkGroups", EbvNumWorkGroups, symbolTable);
symbolTable.setVariableExtensions("gl_MeshVerticesEXT", "gl_Position", 1, &E_GL_EXT_mesh_shader);
symbolTable.setVariableExtensions("gl_MeshVerticesEXT", "gl_PointSize", 1, &E_GL_EXT_mesh_shader);
symbolTable.setVariableExtensions("gl_MeshVerticesEXT", "gl_ClipDistance", 1, &E_GL_EXT_mesh_shader);
symbolTable.setVariableExtensions("gl_MeshVerticesEXT", "gl_CullDistance", 1, &E_GL_EXT_mesh_shader);
BuiltInVariable("gl_MeshVerticesEXT", "gl_Position", EbvPosition, symbolTable);
BuiltInVariable("gl_MeshVerticesEXT", "gl_PointSize", EbvPointSize, symbolTable);
BuiltInVariable("gl_MeshVerticesEXT", "gl_ClipDistance", EbvClipDistance, symbolTable);
BuiltInVariable("gl_MeshVerticesEXT", "gl_CullDistance", EbvCullDistance, symbolTable);
symbolTable.setVariableExtensions("gl_MeshPrimitivesEXT", "gl_PrimitiveID", 1, &E_GL_EXT_mesh_shader);
symbolTable.setVariableExtensions("gl_MeshPrimitivesEXT", "gl_Layer", 1, &E_GL_EXT_mesh_shader);
symbolTable.setVariableExtensions("gl_MeshPrimitivesEXT", "gl_ViewportIndex", 1, &E_GL_EXT_mesh_shader);
symbolTable.setVariableExtensions("gl_MeshPrimitivesEXT", "gl_CullPrimitiveEXT", 1, &E_GL_EXT_mesh_shader);
symbolTable.setVariableExtensions("gl_MeshPrimitivesEXT", "gl_PrimitiveShadingRateEXT", 1, &E_GL_EXT_mesh_shader);
BuiltInVariable("gl_MeshPrimitivesEXT", "gl_PrimitiveID", EbvPrimitiveId, symbolTable);
BuiltInVariable("gl_MeshPrimitivesEXT", "gl_Layer", EbvLayer, symbolTable);
BuiltInVariable("gl_MeshPrimitivesEXT", "gl_ViewportIndex", EbvViewportIndex, symbolTable);
BuiltInVariable("gl_MeshPrimitivesEXT", "gl_CullPrimitiveEXT", EbvCullPrimitiveEXT, symbolTable);
BuiltInVariable("gl_MeshPrimitivesEXT", "gl_PrimitiveShadingRateEXT", EbvPrimitiveShadingRateKHR, symbolTable);
symbolTable.setFunctionExtensions("SetMeshOutputsEXT", 1, &E_GL_EXT_mesh_shader);
// GL_EXT_device_group
symbolTable.setVariableExtensions("gl_DeviceIndex", 1, &E_GL_EXT_device_group);
BuiltInVariable("gl_DeviceIndex", EbvDeviceIndex, symbolTable);
@ -8966,6 +9055,9 @@ void TBuiltIns::identifyBuiltIns(int version, EProfile profile, const SpvVersion
if (version >= 460) {
BuiltInVariable("gl_DrawID", EbvDrawId, symbolTable);
}
// GL_EXT_multiview
BuiltInVariable("gl_ViewIndex", EbvViewIndex, symbolTable);
symbolTable.setVariableExtensions("gl_ViewIndex", 1, &E_GL_EXT_multiview);
// GL_ARB_shader_ballot
symbolTable.setVariableExtensions("gl_SubGroupSizeARB", 1, &E_GL_ARB_shader_ballot);
@ -9035,16 +9127,24 @@ void TBuiltIns::identifyBuiltIns(int version, EProfile profile, const SpvVersion
}
break;
case EShLangTaskNV:
case EShLangTask:
if ((profile != EEsProfile && version >= 450) || (profile == EEsProfile && version >= 320)) {
symbolTable.setVariableExtensions("gl_TaskCountNV", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_WorkGroupSize", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_WorkGroupID", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_LocalInvocationID", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_GlobalInvocationID", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_LocalInvocationIndex", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_MeshViewCountNV", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_MeshViewIndicesNV", 1, &E_GL_NV_mesh_shader);
if (profile != EEsProfile) {
symbolTable.setVariableExtensions("gl_WorkGroupSize", Num_AEP_mesh_shader, AEP_mesh_shader);
symbolTable.setVariableExtensions("gl_WorkGroupID", Num_AEP_mesh_shader, AEP_mesh_shader);
symbolTable.setVariableExtensions("gl_LocalInvocationID", Num_AEP_mesh_shader, AEP_mesh_shader);
symbolTable.setVariableExtensions("gl_GlobalInvocationID", Num_AEP_mesh_shader, AEP_mesh_shader);
symbolTable.setVariableExtensions("gl_LocalInvocationIndex", Num_AEP_mesh_shader, AEP_mesh_shader);
} else {
symbolTable.setVariableExtensions("gl_WorkGroupSize", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_WorkGroupID", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_LocalInvocationID", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_GlobalInvocationID", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_LocalInvocationIndex", 1, &E_GL_NV_mesh_shader);
}
BuiltInVariable("gl_TaskCountNV", EbvTaskCountNV, symbolTable);
BuiltInVariable("gl_WorkGroupSize", EbvWorkGroupSize, symbolTable);
@ -9058,12 +9158,23 @@ void TBuiltIns::identifyBuiltIns(int version, EProfile profile, const SpvVersion
symbolTable.setVariableExtensions("gl_MaxTaskWorkGroupSizeNV", 1, &E_GL_NV_mesh_shader);
symbolTable.setVariableExtensions("gl_MaxMeshViewCountNV", 1, &E_GL_NV_mesh_shader);
symbolTable.setFunctionExtensions("barrier", 1, &E_GL_NV_mesh_shader);
symbolTable.setFunctionExtensions("memoryBarrierShared", 1, &E_GL_NV_mesh_shader);
symbolTable.setFunctionExtensions("groupMemoryBarrier", 1, &E_GL_NV_mesh_shader);
if (profile != EEsProfile) {
symbolTable.setFunctionExtensions("barrier", Num_AEP_mesh_shader, AEP_mesh_shader);
symbolTable.setFunctionExtensions("memoryBarrierShared", Num_AEP_mesh_shader, AEP_mesh_shader);
symbolTable.setFunctionExtensions("groupMemoryBarrier", Num_AEP_mesh_shader, AEP_mesh_shader);
} else {
symbolTable.setFunctionExtensions("barrier", 1, &E_GL_NV_mesh_shader);
symbolTable.setFunctionExtensions("memoryBarrierShared", 1, &E_GL_NV_mesh_shader);
symbolTable.setFunctionExtensions("groupMemoryBarrier", 1, &E_GL_NV_mesh_shader);
}
}
if (profile != EEsProfile && version >= 450) {
// GL_EXT_mesh_shader
symbolTable.setFunctionExtensions("EmitMeshTasksEXT", 1, &E_GL_EXT_mesh_shader);
symbolTable.setVariableExtensions("gl_NumWorkGroups", 1, &E_GL_EXT_mesh_shader);
BuiltInVariable("gl_NumWorkGroups", EbvNumWorkGroups, symbolTable);
// GL_EXT_device_group
symbolTable.setVariableExtensions("gl_DeviceIndex", 1, &E_GL_EXT_device_group);
BuiltInVariable("gl_DeviceIndex", EbvDeviceIndex, symbolTable);
@ -9687,17 +9798,27 @@ void TBuiltIns::identifyBuiltIns(int version, EProfile profile, const SpvVersion
symbolTable.relateToOperator("executeCallableEXT", EOpExecuteCallableKHR);
}
break;
case EShLangMeshNV:
case EShLangMesh:
if ((profile != EEsProfile && version >= 450) || (profile == EEsProfile && version >= 320)) {
symbolTable.relateToOperator("writePackedPrimitiveIndices4x8NV", EOpWritePackedPrimitiveIndices4x8NV);
symbolTable.relateToOperator("memoryBarrierShared", EOpMemoryBarrierShared);
symbolTable.relateToOperator("groupMemoryBarrier", EOpGroupMemoryBarrier);
symbolTable.relateToOperator("subgroupMemoryBarrierShared", EOpSubgroupMemoryBarrierShared);
}
// fall through
case EShLangTaskNV:
if (profile != EEsProfile && version >= 450) {
symbolTable.relateToOperator("SetMeshOutputsEXT", EOpSetMeshOutputsEXT);
}
break;
case EShLangTask:
if ((profile != EEsProfile && version >= 450) || (profile == EEsProfile && version >= 320)) {
symbolTable.relateToOperator("memoryBarrierShared", EOpMemoryBarrierShared);
symbolTable.relateToOperator("groupMemoryBarrier", EOpGroupMemoryBarrier);
symbolTable.relateToOperator("subgroupMemoryBarrierShared", EOpSubgroupMemoryBarrierShared);
}
if (profile != EEsProfile && version >= 450) {
symbolTable.relateToOperator("EmitMeshTasksEXT", EOpEmitMeshTasksEXT);
}
break;
default:

152
glslang/MachineIndependent/ParseHelper.cpp
View File

@ -502,6 +502,16 @@ TIntermTyped* TParseContext::handleVariable(const TSourceLoc& loc, TSymbol* symb
error(loc, "cannot be used (maybe an instance name is needed)", string->c_str(), "");
variable = nullptr;
}
if (language == EShLangMesh && variable) {
TLayoutGeometry primitiveType = intermediate.getOutputPrimitive();
if ((variable->getMangledName() == "gl_PrimitiveTriangleIndicesEXT" && primitiveType != ElgTriangles) ||
(variable->getMangledName() == "gl_PrimitiveLineIndicesEXT" && primitiveType != ElgLines) ||
(variable->getMangledName() == "gl_PrimitivePointIndicesEXT" && primitiveType != ElgPoints)) {
error(loc, "cannot be used (ouput primitive type mismatch)", string->c_str(), "");
variable = nullptr;
}
}
} else {
if (symbol)
error(loc, "variable name expected", string->c_str(), "");
@ -717,7 +727,7 @@ bool TParseContext::isIoResizeArray(const TType& type) const
! type.getQualifier().patch) ||
(language == EShLangFragment && type.getQualifier().storage == EvqVaryingIn &&
(type.getQualifier().pervertexNV || type.getQualifier().pervertexEXT)) ||
(language == EShLangMeshNV && type.getQualifier().storage == EvqVaryingOut &&
(language == EShLangMesh && type.getQualifier().storage == EvqVaryingOut &&
!type.getQualifier().perTaskNV));
}
@ -794,7 +804,7 @@ void TParseContext::checkIoArraysConsistency(const TSourceLoc &loc, bool tailOnl
// As I/O array sizes don't change, fetch requiredSize only once,
// except for mesh shaders which could have different I/O array sizes based on type qualifiers.
if (firstIteration || (language == EShLangMeshNV)) {
if (firstIteration || (language == EShLangMesh)) {
requiredSize = getIoArrayImplicitSize(type.getQualifier(), &featureString);
if (requiredSize == 0)
break;
@ -823,10 +833,11 @@ int TParseContext::getIoArrayImplicitSize(const TQualifier &qualifier, TString *
// Number of vertices for Fragment shader is always three.
expectedSize = 3;
str = "vertices";
} else if (language == EShLangMeshNV) {
} else if (language == EShLangMesh) {
unsigned int maxPrimitives =
intermediate.getPrimitives() != TQualifier::layoutNotSet ? intermediate.getPrimitives() : 0;
if (qualifier.builtIn == EbvPrimitiveIndicesNV) {
if (qualifier.builtIn == EbvPrimitiveIndicesNV || qualifier.builtIn == EbvPrimitiveTriangleIndicesEXT ||
qualifier.builtIn == EbvPrimitiveLineIndicesEXT || qualifier.builtIn == EbvPrimitivePointIndicesEXT) {
expectedSize = maxPrimitives * TQualifier::mapGeometryToSize(intermediate.getOutputPrimitive());
str = "max_primitives*";
str += TQualifier::getGeometryString(intermediate.getOutputPrimitive());
@ -858,7 +869,7 @@ void TParseContext::checkIoArrayConsistency(const TSourceLoc& loc, int requiredS
if (type.getOuterArraySize() > requiredSize)
error(loc, " cannot be greater than 3 for pervertexEXT", feature, name.c_str());
}
else if (language == EShLangMeshNV)
else if (language == EShLangMesh)
error(loc, "inconsistent output array size of", feature, name.c_str());
else
assert(0);
@ -2462,7 +2473,7 @@ void TParseContext::builtInOpCheck(const TSourceLoc& loc, const TFunction& fnCan
const TIntermTyped* base = TIntermediate::findLValueBase(arg0, true , true);
const TType* refType = (base->getType().isReference()) ? base->getType().getReferentType() : nullptr;
const TQualifier& qualifier = (refType != nullptr) ? refType->getQualifier() : base->getType().getQualifier();
if (qualifier.storage != EvqShared && qualifier.storage != EvqBuffer)
if (qualifier.storage != EvqShared && qualifier.storage != EvqBuffer && qualifier.storage != EvqtaskPayloadSharedEXT)
error(loc,"Atomic memory function can only be used for shader storage block member or shared variable.",
fnCandidate.getName().c_str(), "");
@ -2995,6 +3006,10 @@ bool TParseContext::lValueErrorCheck(const TSourceLoc& loc, const char* op, TInt
message = "can't modify EvqFragStencil if using early_fragment_tests";
break;
case EvqtaskPayloadSharedEXT:
if (language == EShLangMesh)
message = "can't modify variable with storage qualifier taskPayloadSharedEXT in mesh shaders";
break;
default:
break;
}
@ -3831,6 +3846,9 @@ void TParseContext::globalQualifierTypeCheck(const TSourceLoc& loc, const TQuali
if (qualifier.isPatch() && qualifier.isInterpolation())
error(loc, "cannot use interpolation qualifiers with patch", "patch", "");
if (qualifier.isTaskPayload() && publicType.basicType == EbtBlock)
error(loc, "taskPayloadSharedEXT variables should not be declared as interface blocks", "taskPayloadSharedEXT", "");
if (qualifier.isTaskMemory() && publicType.basicType != EbtBlock)
error(loc, "taskNV variables can be declared only as blocks", "taskNV", "");
@ -4346,10 +4364,10 @@ void TParseContext::arraySizesCheck(const TSourceLoc& loc, const TQualifier& qua
extensionsTurnedOn(Num_AEP_tessellation_shader, AEP_tessellation_shader))
return;
break;
case EShLangMeshNV:
case EShLangMesh:
if (qualifier.storage == EvqVaryingOut)
if ((isEsProfile() && version >= 320) ||
extensionTurnedOn(E_GL_NV_mesh_shader))
extensionsTurnedOn(Num_AEP_mesh_shader, AEP_mesh_shader))
return;
break;
default:
@ -4633,6 +4651,9 @@ TSymbol* TParseContext::redeclareBuiltinVariable(const TSourceLoc& loc, const TS
identifier == "gl_SampleMask" ||
identifier == "gl_Layer" ||
identifier == "gl_PrimitiveIndicesNV" ||
identifier == "gl_PrimitivePointIndicesEXT" ||
identifier == "gl_PrimitiveLineIndicesEXT" ||
identifier == "gl_PrimitiveTriangleIndicesEXT" ||
identifier == "gl_TexCoord") {
// Find the existing symbol, if any.
@ -4771,7 +4792,8 @@ void TParseContext::redeclareBuiltinBlock(const TSourceLoc& loc, TTypeList& newT
profileRequires(loc, ~EEsProfile, 410, E_GL_ARB_separate_shader_objects, feature);
if (blockName != "gl_PerVertex" && blockName != "gl_PerFragment" &&
blockName != "gl_MeshPerVertexNV" && blockName != "gl_MeshPerPrimitiveNV") {
blockName != "gl_MeshPerVertexNV" && blockName != "gl_MeshPerPrimitiveNV" &&
blockName != "gl_MeshPerVertexEXT" && blockName != "gl_MeshPerPrimitiveEXT") {
error(loc, "cannot redeclare block: ", "block declaration", blockName.c_str());
return;
}
@ -5340,11 +5362,11 @@ void TParseContext::finish()
if (!isEsProfile() && version < 430)
requireExtensions(getCurrentLoc(), 1, &E_GL_ARB_compute_shader, "compute shaders");
break;
case EShLangTaskNV:
requireExtensions(getCurrentLoc(), 1, &E_GL_NV_mesh_shader, "task shaders");
case EShLangTask:
requireExtensions(getCurrentLoc(), Num_AEP_mesh_shader, AEP_mesh_shader, "task shaders");
break;
case EShLangMeshNV:
requireExtensions(getCurrentLoc(), 1, &E_GL_NV_mesh_shader, "mesh shaders");
case EShLangMesh:
requireExtensions(getCurrentLoc(), Num_AEP_mesh_shader, AEP_mesh_shader, "mesh shaders");
break;
default:
break;
@ -5454,12 +5476,12 @@ void TParseContext::setLayoutQualifier(const TSourceLoc& loc, TPublicType& publi
intermediate.setUsePhysicalStorageBuffer();
return;
}
if (language == EShLangGeometry || language == EShLangTessEvaluation || language == EShLangMeshNV) {
if (language == EShLangGeometry || language == EShLangTessEvaluation || language == EShLangMesh) {
if (id == TQualifier::getGeometryString(ElgTriangles)) {
publicType.shaderQualifiers.geometry = ElgTriangles;
return;
}
if (language == EShLangGeometry || language == EShLangMeshNV) {
if (language == EShLangGeometry || language == EShLangMesh) {
if (id == TQualifier::getGeometryString(ElgPoints)) {
publicType.shaderQualifiers.geometry = ElgPoints;
return;
@ -5947,9 +5969,9 @@ void TParseContext::setLayoutQualifier(const TSourceLoc& loc, TPublicType& publi
}
break;
case EShLangMeshNV:
case EShLangMesh:
if (id == "max_vertices") {
requireExtensions(loc, 1, &E_GL_NV_mesh_shader, "max_vertices");
requireExtensions(loc, Num_AEP_mesh_shader, AEP_mesh_shader, "max_vertices");
publicType.shaderQualifiers.vertices = value;
if (value > resources.maxMeshOutputVerticesNV)
error(loc, "too large, must be less than gl_MaxMeshOutputVerticesNV", "max_vertices", "");
@ -5958,7 +5980,7 @@ void TParseContext::setLayoutQualifier(const TSourceLoc& loc, TPublicType& publi
return;
}
if (id == "max_primitives") {
requireExtensions(loc, 1, &E_GL_NV_mesh_shader, "max_primitives");
requireExtensions(loc, Num_AEP_mesh_shader, AEP_mesh_shader, "max_primitives");
publicType.shaderQualifiers.primitives = value;
if (value > resources.maxMeshOutputPrimitivesNV)
error(loc, "too large, must be less than gl_MaxMeshOutputPrimitivesNV", "max_primitives", "");
@ -5968,14 +5990,14 @@ void TParseContext::setLayoutQualifier(const TSourceLoc& loc, TPublicType& publi
}
// Fall through
case EShLangTaskNV:
case EShLangTask:
// Fall through
#endif
case EShLangCompute:
if (id.compare(0, 11, "local_size_") == 0) {
#ifndef GLSLANG_WEB
if (language == EShLangMeshNV || language == EShLangTaskNV) {
requireExtensions(loc, 1, &E_GL_NV_mesh_shader, "gl_WorkGroupSize");
if (language == EShLangMesh || language == EShLangTask) {
requireExtensions(loc, Num_AEP_mesh_shader, AEP_mesh_shader, "gl_WorkGroupSize");
} else {
profileRequires(loc, EEsProfile, 310, 0, "gl_WorkGroupSize");
profileRequires(loc, ~EEsProfile, 430, E_GL_ARB_compute_shader, "gl_WorkGroupSize");
@ -6251,6 +6273,9 @@ void TParseContext::layoutTypeCheck(const TSourceLoc& loc, const TType& type)
if (type.getBasicType() == EbtBlock)
error(loc, "cannot apply to uniform or buffer block", "location", "");
break;
case EvqtaskPayloadSharedEXT:
error(loc, "cannot apply to taskPayloadSharedEXT", "location", "");
break;
#ifndef GLSLANG_WEB
case EvqPayload:
case EvqPayloadIn:
@ -6612,7 +6637,7 @@ void TParseContext::checkNoShaderLayouts(const TSourceLoc& loc, const TShaderQua
error(loc, message, "local_size id", "");
}
if (shaderQualifiers.vertices != TQualifier::layoutNotSet) {
if (language == EShLangGeometry || language == EShLangMeshNV)
if (language == EShLangGeometry || language == EShLangMesh)
error(loc, message, "max_vertices", "");
else if (language == EShLangTessControl)
error(loc, message, "vertices", "");
@ -6624,7 +6649,7 @@ void TParseContext::checkNoShaderLayouts(const TSourceLoc& loc, const TShaderQua
if (shaderQualifiers.postDepthCoverage)
error(loc, message, "post_depth_coverage", "");
if (shaderQualifiers.primitives != TQualifier::layoutNotSet) {
if (language == EShLangMeshNV)
if (language == EShLangMesh)
error(loc, message, "max_primitives", "");
else
assert(0);
@ -7270,6 +7295,8 @@ TIntermNode* TParseContext::declareVariable(const TSourceLoc& loc, TString& iden
requireInt8Arithmetic(loc, "qualifier", "(u)int8 types can only be in uniform block or buffer storage");
}
if (type.getQualifier().storage == EvqtaskPayloadSharedEXT)
intermediate.addTaskPayloadEXTCount();
if (type.getQualifier().storage == EvqShared && type.containsCoopMat())
error(loc, "qualifier", "Cooperative matrix types must not be used in shared memory", "");
@ -8236,6 +8263,8 @@ void TParseContext::declareBlock(const TSourceLoc& loc, TTypeList& typeList, con
memberQualifier.perViewNV = currentBlockQualifier.perViewNV;
if (currentBlockQualifier.perTaskNV)
memberQualifier.perTaskNV = currentBlockQualifier.perTaskNV;
if (currentBlockQualifier.storage == EvqtaskPayloadSharedEXT)
memberQualifier.storage = EvqtaskPayloadSharedEXT;
if (memberQualifier.storage == EvqSpirvStorageClass)
error(memberLoc, "member cannot have a spirv_storage_class qualifier", memberType.getFieldName().c_str(), "");
if (memberQualifier.hasSprivDecorate() && !memberQualifier.getSpirvDecorate().decorateIds.empty())
@ -8536,23 +8565,23 @@ void TParseContext::blockStageIoCheck(const TSourceLoc& loc, const TQualifier& q
// It is a compile-time error to have an input block in a vertex shader or an output block in a fragment shader
// "Compute shaders do not permit user-defined input variables..."
requireStage(loc, (EShLanguageMask)(EShLangTessControlMask|EShLangTessEvaluationMask|EShLangGeometryMask|
EShLangFragmentMask|EShLangMeshNVMask), "input block");
EShLangFragmentMask|EShLangMeshMask), "input block");
if (language == EShLangFragment) {
profileRequires(loc, EEsProfile, 320, Num_AEP_shader_io_blocks, AEP_shader_io_blocks, "fragment input block");
} else if (language == EShLangMeshNV && ! qualifier.isTaskMemory()) {
} else if (language == EShLangMesh && ! qualifier.isTaskMemory()) {
error(loc, "input blocks cannot be used in a mesh shader", "out", "");
}
break;
case EvqVaryingOut:
profileRequires(loc, ~EEsProfile, 150, E_GL_ARB_separate_shader_objects, "output block");
requireStage(loc, (EShLanguageMask)(EShLangVertexMask|EShLangTessControlMask|EShLangTessEvaluationMask|
EShLangGeometryMask|EShLangMeshNVMask|EShLangTaskNVMask), "output block");
EShLangGeometryMask|EShLangMeshMask|EShLangTaskMask), "output block");
// ES 310 can have a block before shader_io is turned on, so skip this test for built-ins
if (language == EShLangVertex && ! parsingBuiltins) {
profileRequires(loc, EEsProfile, 320, Num_AEP_shader_io_blocks, AEP_shader_io_blocks, "vertex output block");
} else if (language == EShLangMeshNV && qualifier.isTaskMemory()) {
} else if (language == EShLangMesh && qualifier.isTaskMemory()) {
error(loc, "can only use on input blocks in mesh shader", "taskNV", "");
} else if (language == EShLangTaskNV && ! qualifier.isTaskMemory()) {
} else if (language == EShLangTask && ! qualifier.isTaskMemory()) {
error(loc, "output blocks cannot be used in a task shader", "out", "");
}
break;
@ -8966,7 +8995,7 @@ void TParseContext::updateStandaloneQualifierDefaults(const TSourceLoc& loc, con
{
#ifndef GLSLANG_WEB
if (publicType.shaderQualifiers.vertices != TQualifier::layoutNotSet) {
assert(language == EShLangTessControl || language == EShLangGeometry || language == EShLangMeshNV);
assert(language == EShLangTessControl || language == EShLangGeometry || language == EShLangMesh);
const char* id = (language == EShLangTessControl) ? "vertices" : "max_vertices";
if (publicType.qualifier.storage != EvqVaryingOut)
@ -8978,7 +9007,7 @@ void TParseContext::updateStandaloneQualifierDefaults(const TSourceLoc& loc, con
checkIoArraysConsistency(loc);
}
if (publicType.shaderQualifiers.primitives != TQualifier::layoutNotSet) {
assert(language == EShLangMeshNV);
assert(language == EShLangMesh);
const char* id = "max_primitives";
if (publicType.qualifier.storage != EvqVaryingOut)
@ -9002,7 +9031,7 @@ void TParseContext::updateStandaloneQualifierDefaults(const TSourceLoc& loc, con
case ElgTrianglesAdjacency:
case ElgQuads:
case ElgIsolines:
if (language == EShLangMeshNV) {
if (language == EShLangMesh) {
error(loc, "cannot apply to input", TQualifier::getGeometryString(publicType.shaderQualifiers.geometry), "");
break;
}
@ -9019,7 +9048,7 @@ void TParseContext::updateStandaloneQualifierDefaults(const TSourceLoc& loc, con
switch (publicType.shaderQualifiers.geometry) {
case ElgLines:
case ElgTriangles:
if (language != EShLangMeshNV) {
if (language != EShLangMesh) {
error(loc, "cannot apply to 'out'", TQualifier::getGeometryString(publicType.shaderQualifiers.geometry), "");
break;
}
@ -9075,24 +9104,56 @@ void TParseContext::updateStandaloneQualifierDefaults(const TSourceLoc& loc, con
error(loc, "too large; see gl_MaxComputeWorkGroupSize", "local_size", "");
}
#ifndef GLSLANG_WEB
else if (language == EShLangMeshNV) {
else if (language == EShLangMesh) {
switch (i) {
case 0: max = resources.maxMeshWorkGroupSizeX_NV; break;
case 1: max = resources.maxMeshWorkGroupSizeY_NV; break;
case 2: max = resources.maxMeshWorkGroupSizeZ_NV; break;
case 0:
max = extensionTurnedOn(E_GL_EXT_mesh_shader) ?
resources.maxMeshWorkGroupSizeX_EXT :
resources.maxMeshWorkGroupSizeX_NV;
break;
case 1:
max = extensionTurnedOn(E_GL_EXT_mesh_shader) ?
resources.maxMeshWorkGroupSizeY_EXT :
resources.maxMeshWorkGroupSizeY_NV ;
break;
case 2:
max = extensionTurnedOn(E_GL_EXT_mesh_shader) ?
resources.maxMeshWorkGroupSizeZ_EXT :
resources.maxMeshWorkGroupSizeZ_NV ;
break;
default: break;
}
if (intermediate.getLocalSize(i) > (unsigned int)max)
error(loc, "too large; see gl_MaxMeshWorkGroupSizeNV", "local_size", "");
} else if (language == EShLangTaskNV) {
if (intermediate.getLocalSize(i) > (unsigned int)max) {
TString maxsErrtring = "too large, see ";
maxsErrtring.append(extensionTurnedOn(E_GL_EXT_mesh_shader) ?
"gl_MaxMeshWorkGroupSizeEXT" : "gl_MaxMeshWorkGroupSizeNV");
error(loc, maxsErrtring.c_str(), "local_size", "");
}
} else if (language == EShLangTask) {
switch (i) {
case 0: max = resources.maxTaskWorkGroupSizeX_NV; break;
case 1: max = resources.maxTaskWorkGroupSizeY_NV; break;
case 2: max = resources.maxTaskWorkGroupSizeZ_NV; break;
case 0:
max = extensionTurnedOn(E_GL_EXT_mesh_shader) ?
resources.maxTaskWorkGroupSizeX_EXT :
resources.maxTaskWorkGroupSizeX_NV;
break;
case 1:
max = extensionTurnedOn(E_GL_EXT_mesh_shader) ?
resources.maxTaskWorkGroupSizeY_EXT:
resources.maxTaskWorkGroupSizeY_NV;
break;
case 2:
max = extensionTurnedOn(E_GL_EXT_mesh_shader) ?
resources.maxTaskWorkGroupSizeZ_EXT:
resources.maxTaskWorkGroupSizeZ_NV;
break;
default: break;
}
if (intermediate.getLocalSize(i) > (unsigned int)max)
error(loc, "too large; see gl_MaxTaskWorkGroupSizeNV", "local_size", "");
if (intermediate.getLocalSize(i) > (unsigned int)max) {
TString maxsErrtring = "too large, see ";
maxsErrtring.append(extensionTurnedOn(E_GL_EXT_mesh_shader) ?
"gl_MaxTaskWorkGroupSizeEXT" : "gl_MaxTaskWorkGroupSizeNV");
error(loc, maxsErrtring.c_str(), "local_size", "");
}
}
#endif
else {
@ -9181,7 +9242,7 @@ void TParseContext::updateStandaloneQualifierDefaults(const TSourceLoc& loc, con
error(loc, "can only apply to 'in'", "derivative_group_linearNV", "");
}
// Check mesh out array sizes, once all the necessary out qualifiers are defined.
if ((language == EShLangMeshNV) &&
if ((language == EShLangMesh) &&
(intermediate.getVertices() != TQualifier::layoutNotSet) &&
(intermediate.getPrimitives() != TQualifier::layoutNotSet) &&
(intermediate.getOutputPrimitive() != ElgNone))
@ -9396,4 +9457,3 @@ const TTypeList* TParseContext::recordStructCopy(TStructRecord& record, const TT
}
} // end namespace glslang

12
glslang/MachineIndependent/Scan.cpp
View File

@ -758,6 +758,8 @@ void TScanContext::fillInKeywordMap()
(*KeywordMap)["perprimitiveNV"] = PERPRIMITIVENV;
(*KeywordMap)["perviewNV"] = PERVIEWNV;
(*KeywordMap)["taskNV"] = PERTASKNV;
(*KeywordMap)["perprimitiveEXT"] = PERPRIMITIVEEXT;
(*KeywordMap)["taskPayloadSharedEXT"] = TASKPAYLOADWORKGROUPEXT;
(*KeywordMap)["fcoopmatNV"] = FCOOPMATNV;
(*KeywordMap)["icoopmatNV"] = ICOOPMATNV;
@ -1740,12 +1742,18 @@ int TScanContext::tokenizeIdentifier()
case PERPRIMITIVENV:
case PERVIEWNV:
case PERTASKNV:
if ((!parseContext.isEsProfile() && parseContext.version >= 450) ||
(parseContext.isEsProfile() && parseContext.version >= 320) ||
if (parseContext.symbolTable.atBuiltInLevel() ||
parseContext.extensionTurnedOn(E_GL_NV_mesh_shader))
return keyword;
return identifierOrType();
case PERPRIMITIVEEXT:
case TASKPAYLOADWORKGROUPEXT:
if (parseContext.symbolTable.atBuiltInLevel() ||
parseContext.extensionTurnedOn(E_GL_EXT_mesh_shader))
return keyword;
return identifierOrType();
case FCOOPMATNV:
afterType = true;
if (parseContext.symbolTable.atBuiltInLevel() ||

8
glslang/MachineIndependent/ShaderLang.cpp
View File

@ -391,13 +391,13 @@ bool InitializeSymbolTables(TInfoSink& infoSink, TSymbolTable** commonTable, TS
// check for mesh
if ((profile != EEsProfile && version >= 450) ||
(profile == EEsProfile && version >= 320))
InitializeStageSymbolTable(*builtInParseables, version, profile, spvVersion, EShLangMeshNV, source,
InitializeStageSymbolTable(*builtInParseables, version, profile, spvVersion, EShLangMesh, source,
infoSink, commonTable, symbolTables);
// check for task
if ((profile != EEsProfile && version >= 450) ||
(profile == EEsProfile && version >= 320))
InitializeStageSymbolTable(*builtInParseables, version, profile, spvVersion, EShLangTaskNV, source,
InitializeStageSymbolTable(*builtInParseables, version, profile, spvVersion, EShLangTask, source,
infoSink, commonTable, symbolTables);
#endif // !GLSLANG_ANGLE
#endif // !GLSLANG_WEB
@ -650,8 +650,8 @@ bool DeduceVersionProfile(TInfoSink& infoSink, EShLanguage stage, bool versionNo
version = 460;
}
break;
case EShLangMeshNV:
case EShLangTaskNV:
case EShLangMesh:
case EShLangTask:
if ((profile == EEsProfile && version < 320) ||
(profile != EEsProfile && version < 450)) {
correct = false;

27
glslang/MachineIndependent/Versions.cpp
View File

@ -166,7 +166,8 @@ void TParseVersions::initializeExtensionBehavior()
} extensionData;
const extensionData exts[] = { {E_GL_EXT_ray_tracing, EShTargetSpv_1_4},
{E_GL_NV_ray_tracing_motion_blur, EShTargetSpv_1_4}
{E_GL_NV_ray_tracing_motion_blur, EShTargetSpv_1_4},
{E_GL_EXT_mesh_shader, EShTargetSpv_1_4}
};
for (size_t ii = 0; ii < sizeof(exts) / sizeof(exts[0]); ii++) {
@ -345,6 +346,7 @@ void TParseVersions::initializeExtensionBehavior()
extensionBehavior[E_GL_EXT_terminate_invocation] = EBhDisable;
extensionBehavior[E_GL_EXT_shared_memory_block] = EBhDisable;
extensionBehavior[E_GL_EXT_spirv_intrinsics] = EBhDisable;
extensionBehavior[E_GL_EXT_mesh_shader] = EBhDisable;
// OVR extensions
extensionBehavior[E_GL_OVR_multiview] = EBhDisable;
@ -511,6 +513,7 @@ void TParseVersions::getPreamble(std::string& preamble)
"#define GL_EXT_ray_flags_primitive_culling 1\n"
"#define GL_EXT_ray_cull_mask 1\n"
"#define GL_EXT_spirv_intrinsics 1\n"
"#define GL_EXT_mesh_shader 1\n"
"#define GL_AMD_shader_ballot 1\n"
"#define GL_AMD_shader_trinary_minmax 1\n"
@ -641,8 +644,8 @@ void TParseVersions::getPreamble(std::string& preamble)
case EShLangClosestHit: preamble += "#define GL_CLOSEST_HIT_SHADER_EXT 1 \n"; break;
case EShLangMiss: preamble += "#define GL_MISS_SHADER_EXT 1 \n"; break;
case EShLangCallable: preamble += "#define GL_CALLABLE_SHADER_EXT 1 \n"; break;
case EShLangTaskNV: preamble += "#define GL_TASK_SHADER_NV 1 \n"; break;
case EShLangMeshNV: preamble += "#define GL_MESH_SHADER_NV 1 \n"; break;
case EShLangTask: preamble += "#define GL_TASK_SHADER_NV 1 \n"; break;
case EShLangMesh: preamble += "#define GL_MESH_SHADER_NV 1 \n"; break;
default: break;
}
}
@ -668,8 +671,8 @@ const char* StageName(EShLanguage stage)
case EShLangClosestHit: return "closest-hit";
case EShLangMiss: return "miss";
case EShLangCallable: return "callable";
case EShLangMeshNV: return "mesh";
case EShLangTaskNV: return "task";
case EShLangMesh: return "mesh";
case EShLangTask: return "task";
#endif
default: return "unknown stage";
}
@ -1060,10 +1063,22 @@ void TParseVersions::checkExtensionStage(const TSourceLoc& loc, const char * con
{
// GL_NV_mesh_shader extension is only allowed in task/mesh shaders
if (strcmp(extension, "GL_NV_mesh_shader") == 0) {
requireStage(loc, (EShLanguageMask)(EShLangTaskNVMask | EShLangMeshNVMask | EShLangFragmentMask),
requireStage(loc, (EShLanguageMask)(EShLangTaskMask | EShLangMeshMask | EShLangFragmentMask),
"#extension GL_NV_mesh_shader");
profileRequires(loc, ECoreProfile, 450, 0, "#extension GL_NV_mesh_shader");
profileRequires(loc, EEsProfile, 320, 0, "#extension GL_NV_mesh_shader");
if (extensionTurnedOn(E_GL_EXT_mesh_shader)) {
error(loc, "GL_EXT_mesh_shader is already turned on, and not allowed with", "#extension", extension);
}
}
else if (strcmp(extension, "GL_EXT_mesh_shader") == 0) {
requireStage(loc, (EShLanguageMask)(EShLangTaskMask | EShLangMeshMask | EShLangFragmentMask),
"#extension GL_EXT_mesh_shader");
profileRequires(loc, ECoreProfile, 450, 0, "#extension GL_EXT_mesh_shader");
profileRequires(loc, EEsProfile, 320, 0, "#extension GL_EXT_mesh_shader");
if (extensionTurnedOn(E_GL_NV_mesh_shader)) {
error(loc, "GL_NV_mesh_shader is already turned on, and not allowed with", "#extension", extension);
}
}
}

4
glslang/MachineIndependent/Versions.h
View File

@ -211,6 +211,7 @@ const char* const E_GL_EXT_shared_memory_block = "GL_EXT_shared_mem
const char* const E_GL_EXT_subgroup_uniform_control_flow = "GL_EXT_subgroup_uniform_control_flow";
const char* const E_GL_EXT_spirv_intrinsics = "GL_EXT_spirv_intrinsics";
const char* const E_GL_EXT_fragment_shader_barycentric = "GL_EXT_fragment_shader_barycentric";
const char* const E_GL_EXT_mesh_shader = "GL_EXT_mesh_shader";
// Arrays of extensions for the above viewportEXTs duplications
@ -349,6 +350,9 @@ const int Num_AEP_texture_buffer = sizeof(AEP_texture_buffer)/sizeof(AEP_texture
const char* const AEP_texture_cube_map_array[] = { E_GL_EXT_texture_cube_map_array, E_GL_OES_texture_cube_map_array };
const int Num_AEP_texture_cube_map_array = sizeof(AEP_texture_cube_map_array)/sizeof(AEP_texture_cube_map_array[0]);
const char* const AEP_mesh_shader[] = { E_GL_NV_mesh_shader, E_GL_EXT_mesh_shader };
const int Num_AEP_mesh_shader = sizeof(AEP_mesh_shader)/sizeof(AEP_mesh_shader[0]);
} // end namespace glslang
#endif // _VERSIONS_INCLUDED_

27
glslang/MachineIndependent/glslang.m4
View File

@ -315,7 +315,7 @@ GLSLANG_WEB_EXCLUDE_ON
%token <lex> PATCH SAMPLE NONUNIFORM
%token <lex> COHERENT VOLATILE RESTRICT READONLY WRITEONLY DEVICECOHERENT QUEUEFAMILYCOHERENT WORKGROUPCOHERENT
%token <lex> SUBGROUPCOHERENT NONPRIVATE SHADERCALLCOHERENT
%token <lex> NOPERSPECTIVE EXPLICITINTERPAMD PERVERTEXEXT PERVERTEXNV PERPRIMITIVENV PERVIEWNV PERTASKNV
%token <lex> NOPERSPECTIVE EXPLICITINTERPAMD PERVERTEXEXT PERVERTEXNV PERPRIMITIVENV PERVIEWNV PERTASKNV PERPRIMITIVEEXT TASKPAYLOADWORKGROUPEXT
%token <lex> PRECISE
GLSLANG_WEB_EXCLUDE_OFF
@ -1301,24 +1301,34 @@ GLSLANG_WEB_EXCLUDE_ON
| PERPRIMITIVENV {
// No need for profile version or extension check. Shader stage already checks both.
parseContext.globalCheck($1.loc, "perprimitiveNV");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangFragmentMask | EShLangMeshNVMask), "perprimitiveNV");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangFragmentMask | EShLangMeshMask), "perprimitiveNV");
// Fragment shader stage doesn't check for extension. So we explicitly add below extension check.
if (parseContext.language == EShLangFragment)
parseContext.requireExtensions($1.loc, 1, &E_GL_NV_mesh_shader, "perprimitiveNV");
$$.init($1.loc);
$$.qualifier.perPrimitiveNV = true;
}
| PERPRIMITIVEEXT {
// No need for profile version or extension check. Shader stage already checks both.
parseContext.globalCheck($1.loc, "perprimitiveEXT");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangFragmentMask | EShLangMeshMask), "perprimitiveEXT");
// Fragment shader stage doesn't check for extension. So we explicitly add below extension check.
if (parseContext.language == EShLangFragment)
parseContext.requireExtensions($1.loc, 1, &E_GL_EXT_mesh_shader, "perprimitiveEXT");
$$.init($1.loc);
$$.qualifier.perPrimitiveNV = true;
}
| PERVIEWNV {
// No need for profile version or extension check. Shader stage already checks both.
parseContext.globalCheck($1.loc, "perviewNV");
parseContext.requireStage($1.loc, EShLangMeshNV, "perviewNV");
parseContext.requireStage($1.loc, EShLangMesh, "perviewNV");
$$.init($1.loc);
$$.qualifier.perViewNV = true;
}
| PERTASKNV {
// No need for profile version or extension check. Shader stage already checks both.
parseContext.globalCheck($1.loc, "taskNV");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangTaskNVMask | EShLangMeshNVMask), "taskNV");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangTaskMask | EShLangMeshMask), "taskNV");
$$.init($1.loc);
$$.qualifier.perTaskNV = true;
}
@ -1469,7 +1479,7 @@ storage_qualifier
parseContext.globalCheck($1.loc, "shared");
parseContext.profileRequires($1.loc, ECoreProfile | ECompatibilityProfile, 430, E_GL_ARB_compute_shader, "shared");
parseContext.profileRequires($1.loc, EEsProfile, 310, 0, "shared");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangComputeMask | EShLangMeshNVMask | EShLangTaskNVMask), "shared");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangComputeMask | EShLangMeshMask | EShLangTaskMask), "shared");
$$.init($1.loc);
$$.qualifier.storage = EvqShared;
}
@ -1656,6 +1666,13 @@ GLSLANG_WEB_EXCLUDE_ON
parseContext.unimplemented($1.loc, "subroutine");
$$.init($1.loc);
}
| TASKPAYLOADWORKGROUPEXT {
// No need for profile version or extension check. Shader stage already checks both.
parseContext.globalCheck($1.loc, "taskPayloadSharedEXT");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangTaskMask | EShLangMeshMask), "taskPayloadSharedEXT ");
$$.init($1.loc);
$$.qualifier.storage = EvqtaskPayloadSharedEXT;
}
GLSLANG_WEB_EXCLUDE_OFF
;

27
glslang/MachineIndependent/glslang.y
View File

@ -315,7 +315,7 @@ extern int yylex(YYSTYPE*, TParseContext&);
%token <lex> PATCH SAMPLE NONUNIFORM
%token <lex> COHERENT VOLATILE RESTRICT READONLY WRITEONLY DEVICECOHERENT QUEUEFAMILYCOHERENT WORKGROUPCOHERENT
%token <lex> SUBGROUPCOHERENT NONPRIVATE SHADERCALLCOHERENT
%token <lex> NOPERSPECTIVE EXPLICITINTERPAMD PERVERTEXEXT PERVERTEXNV PERPRIMITIVENV PERVIEWNV PERTASKNV
%token <lex> NOPERSPECTIVE EXPLICITINTERPAMD PERVERTEXEXT PERVERTEXNV PERPRIMITIVENV PERVIEWNV PERTASKNV PERPRIMITIVEEXT TASKPAYLOADWORKGROUPEXT
%token <lex> PRECISE
@ -1301,24 +1301,34 @@ interpolation_qualifier
| PERPRIMITIVENV {
// No need for profile version or extension check. Shader stage already checks both.
parseContext.globalCheck($1.loc, "perprimitiveNV");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangFragmentMask | EShLangMeshNVMask), "perprimitiveNV");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangFragmentMask | EShLangMeshMask), "perprimitiveNV");
// Fragment shader stage doesn't check for extension. So we explicitly add below extension check.
if (parseContext.language == EShLangFragment)
parseContext.requireExtensions($1.loc, 1, &E_GL_NV_mesh_shader, "perprimitiveNV");
$$.init($1.loc);
$$.qualifier.perPrimitiveNV = true;
}
| PERPRIMITIVEEXT {
// No need for profile version or extension check. Shader stage already checks both.
parseContext.globalCheck($1.loc, "perprimitiveEXT");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangFragmentMask | EShLangMeshMask), "perprimitiveEXT");
// Fragment shader stage doesn't check for extension. So we explicitly add below extension check.
if (parseContext.language == EShLangFragment)
parseContext.requireExtensions($1.loc, 1, &E_GL_EXT_mesh_shader, "perprimitiveEXT");
$$.init($1.loc);
$$.qualifier.perPrimitiveNV = true;
}
| PERVIEWNV {
// No need for profile version or extension check. Shader stage already checks both.
parseContext.globalCheck($1.loc, "perviewNV");
parseContext.requireStage($1.loc, EShLangMeshNV, "perviewNV");
parseContext.requireStage($1.loc, EShLangMesh, "perviewNV");
$$.init($1.loc);
$$.qualifier.perViewNV = true;
}
| PERTASKNV {
// No need for profile version or extension check. Shader stage already checks both.
parseContext.globalCheck($1.loc, "taskNV");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangTaskNVMask | EShLangMeshNVMask), "taskNV");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangTaskMask | EShLangMeshMask), "taskNV");
$$.init($1.loc);
$$.qualifier.perTaskNV = true;
}
@ -1469,7 +1479,7 @@ storage_qualifier
parseContext.globalCheck($1.loc, "shared");
parseContext.profileRequires($1.loc, ECoreProfile | ECompatibilityProfile, 430, E_GL_ARB_compute_shader, "shared");
parseContext.profileRequires($1.loc, EEsProfile, 310, 0, "shared");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangComputeMask | EShLangMeshNVMask | EShLangTaskNVMask), "shared");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangComputeMask | EShLangMeshMask | EShLangTaskMask), "shared");
$$.init($1.loc);
$$.qualifier.storage = EvqShared;
}
@ -1656,6 +1666,13 @@ storage_qualifier
parseContext.unimplemented($1.loc, "subroutine");
$$.init($1.loc);
}
| TASKPAYLOADWORKGROUPEXT {
// No need for profile version or extension check. Shader stage already checks both.
parseContext.globalCheck($1.loc, "taskPayloadSharedEXT");
parseContext.requireStage($1.loc, (EShLanguageMask)(EShLangTaskMask | EShLangMeshMask), "taskPayloadSharedEXT ");
$$.init($1.loc);
$$.qualifier.storage = EvqtaskPayloadSharedEXT;
}
;

7999
glslang/MachineIndependent/glslang_tab.cpp
View File

File diff suppressed because it is too large Load Diff

6
glslang/MachineIndependent/glslang_tab.cpp.h
View File

@ -506,7 +506,9 @@ extern int yydebug;
PERPRIMITIVENV = 707, /* PERPRIMITIVENV */
PERVIEWNV = 708, /* PERVIEWNV */
PERTASKNV = 709, /* PERTASKNV */
PRECISE = 710 /* PRECISE */
PERPRIMITIVEEXT = 710, /* PERPRIMITIVEEXT */
TASKPAYLOADWORKGROUPEXT = 711, /* TASKPAYLOADWORKGROUPEXT */
PRECISE = 712 /* PRECISE */
};
typedef enum yytokentype yytoken_kind_t;
#endif
@ -554,7 +556,7 @@ union YYSTYPE
glslang::TArraySizes* typeParameters;
} interm;
#line 558 "MachineIndependent/glslang_tab.cpp.h"
#line 560 "MachineIndependent/glslang_tab.cpp.h"
};
typedef union YYSTYPE YYSTYPE;

6
glslang/MachineIndependent/intermOut.cpp
View File

@ -1068,6 +1068,8 @@ bool TOutputTraverser::visitAggregate(TVisit /* visit */, TIntermAggregate* node
case EOpExecuteCallableNV: out.debug << "executeCallableNV"; break;
case EOpExecuteCallableKHR: out.debug << "executeCallableKHR"; break;
case EOpWritePackedPrimitiveIndices4x8NV: out.debug << "writePackedPrimitiveIndices4x8NV"; break;
case EOpEmitMeshTasksEXT: out.debug << "EmitMeshTasksEXT"; break;
case EOpSetMeshOutputsEXT: out.debug << "SetMeshOutputsEXT"; break;
case EOpRayQueryInitialize: out.debug << "rayQueryInitializeEXT"; break;
case EOpRayQueryTerminate: out.debug << "rayQueryTerminateEXT"; break;
@ -1522,12 +1524,12 @@ void TIntermediate::output(TInfoSink& infoSink, bool tree)
infoSink.debug << "interlock ordering = " << TQualifier::getInterlockOrderingString(interlockOrdering) << "\n";
break;
case EShLangMeshNV:
case EShLangMesh:
infoSink.debug << "max_vertices = " << vertices << "\n";
infoSink.debug << "max_primitives = " << primitives << "\n";
infoSink.debug << "output primitive = " << TQualifier::getGeometryString(outputPrimitive) << "\n";
// Fall through
case EShLangTaskNV:
case EShLangTask:
// Fall through
case EShLangCompute:
infoSink.debug << "local_size = (" << localSize[0] << ", " << localSize[1] << ", " << localSize[2] << ")\n";

12
glslang/MachineIndependent/linkValidate.cpp
View File

@ -212,7 +212,7 @@ void TIntermediate::mergeModes(TInfoSink& infoSink, TIntermediate& unit)
if (vertices == TQualifier::layoutNotSet)
vertices = unit.vertices;
else if (unit.vertices != TQualifier::layoutNotSet && vertices != unit.vertices) {
if (language == EShLangGeometry || language == EShLangMeshNV)
if (language == EShLangGeometry || language == EShLangMesh)
error(infoSink, "Contradictory layout max_vertices values");
else if (language == EShLangTessControl)
error(infoSink, "Contradictory layout vertices values");
@ -222,7 +222,7 @@ void TIntermediate::mergeModes(TInfoSink& infoSink, TIntermediate& unit)
if (primitives == TQualifier::layoutNotSet)
primitives = unit.primitives;
else if (primitives != unit.primitives) {
if (language == EShLangMeshNV)
if (language == EShLangMesh)
error(infoSink, "Contradictory layout max_primitives values");
else
assert(0);
@ -1312,7 +1312,7 @@ void TIntermediate::finalCheck(TInfoSink& infoSink, bool keepUncalled)
if (numShaderRecordBlocks > 1)
error(infoSink, "Only one shaderRecordNV buffer block is allowed per stage");
break;
case EShLangMeshNV:
case EShLangMesh:
// NV_mesh_shader doesn't allow use of both single-view and per-view builtins.
if (inIoAccessed("gl_Position") && inIoAccessed("gl_PositionPerViewNV"))
error(infoSink, "Can only use one of gl_Position or gl_PositionPerViewNV");
@ -1331,9 +1331,11 @@ void TIntermediate::finalCheck(TInfoSink& infoSink, bool keepUncalled)
if (primitives == TQualifier::layoutNotSet)
error(infoSink, "At least one shader must specify a layout(max_primitives = value)");
// fall through
case EShLangTaskNV:
case EShLangTask:
if (numTaskNVBlocks > 1)
error(infoSink, "Only one taskNV interface block is allowed per shader");
if (numTaskEXTPayloads > 1)
error(infoSink, "Only single variable of type taskPayloadSharedEXT is allowed per shader");
sharedBlockCheck(infoSink);
break;
default:
@ -2338,7 +2340,7 @@ bool TIntermediate::isIoResizeArray(const TType& type, EShLanguage language) {
(language == EShLangTessEvaluation && type.getQualifier().storage == EvqVaryingIn) ||
(language == EShLangFragment && type.getQualifier().storage == EvqVaryingIn &&
(type.getQualifier().pervertexNV || type.getQualifier().pervertexEXT)) ||
(language == EShLangMeshNV && type.getQualifier().storage == EvqVaryingOut &&
(language == EShLangMesh && type.getQualifier().storage == EvqVaryingOut &&
!type.getQualifier().perTaskNV));
}
#endif // not GLSLANG_WEB

4
glslang/MachineIndependent/localintermediate.h
View File

@ -322,6 +322,7 @@ public:
primitives(TQualifier::layoutNotSet),
numTaskNVBlocks(0),
layoutPrimitiveCulling(false),
numTaskEXTPayloads(0),
autoMapBindings(false),
autoMapLocations(false),
flattenUniformArrays(false),
@ -639,6 +640,7 @@ public:
int getNumPushConstants() const { return 0; }
void addShaderRecordCount() { }
void addTaskNVCount() { }
void addTaskPayloadEXTCount() { }
void setUseVulkanMemoryModel() { }
bool usingVulkanMemoryModel() const { return false; }
bool usingPhysicalStorageBuffer() const { return false; }
@ -756,6 +758,7 @@ public:
int getNumPushConstants() const { return numPushConstants; }
void addShaderRecordCount() { ++numShaderRecordBlocks; }
void addTaskNVCount() { ++numTaskNVBlocks; }
void addTaskPayloadEXTCount() { ++numTaskEXTPayloads; }
bool setInvocations(int i)
{
@ -1160,6 +1163,7 @@ protected:
int primitives;
int numTaskNVBlocks;
bool layoutPrimitiveCulling;
int numTaskEXTPayloads;
// Base shift values
std::array<unsigned int, EResCount> shiftBinding;

12
glslang/Public/ShaderLang.h
View File

@ -108,8 +108,10 @@ typedef enum {
EShLangMissNV = EShLangMiss,
EShLangCallable,
EShLangCallableNV = EShLangCallable,
EShLangTaskNV,
EShLangMeshNV,
EShLangTask,
EShLangTaskNV = EShLangTask,
EShLangMesh,
EShLangMeshNV = EShLangMesh,
LAST_ELEMENT_MARKER(EShLangCount),
} EShLanguage; // would be better as stage, but this is ancient now
@ -132,8 +134,10 @@ typedef enum : unsigned {
EShLangMissNVMask = EShLangMissMask,
EShLangCallableMask = (1 << EShLangCallable),
EShLangCallableNVMask = EShLangCallableMask,
EShLangTaskNVMask = (1 << EShLangTaskNV),
EShLangMeshNVMask = (1 << EShLangMeshNV),
EShLangTaskMask = (1 << EShLangTask),
EShLangTaskNVMask = EShLangTaskMask,
EShLangMeshMask = (1 << EShLangMesh),
EShLangMeshNVMask = EShLangMeshMask,
LAST_ELEMENT_MARKER(EShLanguageMaskCount),
} EShLanguageMask;

11
gtests/Spv.FromFile.cpp
View File

@ -630,6 +630,17 @@ INSTANTIATE_TEST_SUITE_P(
"spv.WorkgroupMemoryExplicitLayout.std140.comp",
"spv.WorkgroupMemoryExplicitLayout.std430.comp",
"spv.WorkgroupMemoryExplicitLayout.scalar.comp",
// SPV_EXT_mesh_shader
"spv.ext.meshShaderBuiltins.mesh",
"spv.ext.meshShaderRedeclBuiltins.mesh",
"spv.ext.meshShaderTaskMem.mesh",
"spv.ext.meshShaderUserDefined.mesh",
"spv.ext.meshTaskShader.task",
"spv.atomiAddEXT.error.mesh",
"spv.atomiAddEXT.task",
"spv.460.subgroupEXT.task",
"spv.460.subgroupEXT.mesh",
})),
FileNameAsCustomTestSuffix
);

4
gtests/TestFixture.cpp
View File

@ -73,9 +73,9 @@ EShLanguage GetShaderStage(const std::string& stage)
} else if (stage == "rcall") {
return EShLangCallable;
} else if (stage == "task") {
return EShLangTaskNV;
return EShLangTask;
} else if (stage == "mesh") {
return EShLangMeshNV;
return EShLangMesh;
} else {
assert(0 && "Unknown shader stage");
return EShLangCount;

4
known_good.json
View File

@ -5,14 +5,14 @@
"site" : "github",
"subrepo" : "KhronosGroup/SPIRV-Tools",
"subdir" : "External/spirv-tools",
"commit" : "5e61ea2098220059e89523f1f47b0bcd8c33b89a"
"commit" : "4c456f7da67c5437a6fb7d4d20d78e2a5ae2acf2"
},
{
"name" : "spirv-tools/external/spirv-headers",
"site" : "github",
"subrepo" : "KhronosGroup/SPIRV-Headers",
"subdir" : "External/spirv-tools/external/spirv-headers",
"commit" : "b2a156e1c0434bc8c99aaebba1c7be98be7ac580"
"commit" : "87d5b782bec60822aa878941e6b13c0a9a954c9b"
}
]
}