diff options
Diffstat (limited to 'thirdparty/glslang/SPIRV')
20 files changed, 2728 insertions, 540 deletions
diff --git a/thirdparty/glslang/SPIRV/CInterface/spirv_c_interface.cpp b/thirdparty/glslang/SPIRV/CInterface/spirv_c_interface.cpp new file mode 100644 index 0000000000..a0790f48f1 --- /dev/null +++ b/thirdparty/glslang/SPIRV/CInterface/spirv_c_interface.cpp @@ -0,0 +1,110 @@ +/** + This code is based on the glslang_c_interface implementation by Viktor Latypov +**/ + +/** +BSD 2-Clause License + +Copyright (c) 2019, Viktor Latypov +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +1. Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + +2. Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE +FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR +SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, +OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +**/ + +#include "glslang/Include/glslang_c_interface.h" + +#include "SPIRV/GlslangToSpv.h" +#include "SPIRV/Logger.h" +#include "SPIRV/SpvTools.h" + +typedef struct glslang_program_s { + glslang::TProgram* program; + std::vector<unsigned int> spirv; + std::string loggerMessages; +} glslang_program_t; + +static EShLanguage c_shader_stage(glslang_stage_t stage) +{ + switch (stage) { + case GLSLANG_STAGE_VERTEX: + return EShLangVertex; + case GLSLANG_STAGE_TESSCONTROL: + return EShLangTessControl; + case GLSLANG_STAGE_TESSEVALUATION: + return EShLangTessEvaluation; + case GLSLANG_STAGE_GEOMETRY: + return EShLangGeometry; + case GLSLANG_STAGE_FRAGMENT: + return EShLangFragment; + case GLSLANG_STAGE_COMPUTE: + return EShLangCompute; + case GLSLANG_STAGE_RAYGEN_NV: + return EShLangRayGen; + case GLSLANG_STAGE_INTERSECT_NV: + return EShLangIntersect; + case GLSLANG_STAGE_ANYHIT_NV: + return EShLangAnyHit; + case GLSLANG_STAGE_CLOSESTHIT_NV: + return EShLangClosestHit; + case GLSLANG_STAGE_MISS_NV: + return EShLangMiss; + case GLSLANG_STAGE_CALLABLE_NV: + return EShLangCallable; + case GLSLANG_STAGE_TASK_NV: + return EShLangTaskNV; + case GLSLANG_STAGE_MESH_NV: + return EShLangMeshNV; + default: + break; + } + return EShLangCount; +} + +GLSLANG_EXPORT void glslang_program_SPIRV_generate(glslang_program_t* program, glslang_stage_t stage) +{ + spv::SpvBuildLogger logger; + glslang::SpvOptions spvOptions; + spvOptions.validate = true; + + const glslang::TIntermediate* intermediate = program->program->getIntermediate(c_shader_stage(stage)); + + glslang::GlslangToSpv(*intermediate, program->spirv, &logger, &spvOptions); + + program->loggerMessages = logger.getAllMessages(); +} + +GLSLANG_EXPORT size_t glslang_program_SPIRV_get_size(glslang_program_t* program) { return program->spirv.size(); } + +GLSLANG_EXPORT void glslang_program_SPIRV_get(glslang_program_t* program, unsigned int* out) +{ + memcpy(out, program->spirv.data(), program->spirv.size() * sizeof(unsigned int)); +} + +GLSLANG_EXPORT unsigned int* glslang_program_SPIRV_get_ptr(glslang_program_t* program) +{ + return program->spirv.data(); +} + +GLSLANG_EXPORT const char* glslang_program_SPIRV_get_messages(glslang_program_t* program) +{ + return program->loggerMessages.empty() ? nullptr : program->loggerMessages.c_str(); +} diff --git a/thirdparty/glslang/SPIRV/GLSL.ext.EXT.h b/thirdparty/glslang/SPIRV/GLSL.ext.EXT.h index 40164b6187..f48f1304d6 100644 --- a/thirdparty/glslang/SPIRV/GLSL.ext.EXT.h +++ b/thirdparty/glslang/SPIRV/GLSL.ext.EXT.h @@ -35,5 +35,9 @@ static const char* const E_SPV_EXT_shader_viewport_index_layer = "SPV_EXT_shade static const char* const E_SPV_EXT_fragment_fully_covered = "SPV_EXT_fragment_fully_covered"; static const char* const E_SPV_EXT_fragment_invocation_density = "SPV_EXT_fragment_invocation_density"; static const char* const E_SPV_EXT_demote_to_helper_invocation = "SPV_EXT_demote_to_helper_invocation"; +static const char* const E_SPV_EXT_shader_atomic_float_add = "SPV_EXT_shader_atomic_float_add"; +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"; #endif // #ifndef GLSLextEXT_H diff --git a/thirdparty/glslang/SPIRV/GLSL.ext.KHR.h b/thirdparty/glslang/SPIRV/GLSL.ext.KHR.h index e58e836a8d..5eb3e94482 100644 --- a/thirdparty/glslang/SPIRV/GLSL.ext.KHR.h +++ b/thirdparty/glslang/SPIRV/GLSL.ext.KHR.h @@ -1,5 +1,6 @@ /* -** Copyright (c) 2014-2016 The Khronos Group Inc. +** Copyright (c) 2014-2020 The Khronos Group Inc. +** Modifications Copyright (C) 2020 Advanced Micro Devices, Inc. All rights reserved. ** ** Permission is hereby granted, free of charge, to any person obtaining a copy ** of this software and/or associated documentation files (the "Materials"), @@ -44,5 +45,12 @@ static const char* const E_SPV_EXT_physical_storage_buffer = "SPV_EXT_physi static const char* const E_SPV_KHR_physical_storage_buffer = "SPV_KHR_physical_storage_buffer"; static const char* const E_SPV_EXT_fragment_shader_interlock = "SPV_EXT_fragment_shader_interlock"; static const char* const E_SPV_KHR_shader_clock = "SPV_KHR_shader_clock"; +static const char* const E_SPV_KHR_non_semantic_info = "SPV_KHR_non_semantic_info"; +static const char* const E_SPV_KHR_ray_tracing = "SPV_KHR_ray_tracing"; +static const char* const E_SPV_KHR_ray_query = "SPV_KHR_ray_query"; +static const char* const E_SPV_KHR_fragment_shading_rate = "SPV_KHR_fragment_shading_rate"; +static const char* const E_SPV_KHR_terminate_invocation = "SPV_KHR_terminate_invocation"; +static const char* const E_SPV_KHR_workgroup_memory_explicit_layout = "SPV_KHR_workgroup_memory_explicit_layout"; +static const char* const E_SPV_KHR_subgroup_uniform_control_flow = "SPV_KHR_subgroup_uniform_control_flow"; #endif // #ifndef GLSLextKHR_H diff --git a/thirdparty/glslang/SPIRV/GLSL.ext.NV.h b/thirdparty/glslang/SPIRV/GLSL.ext.NV.h index 50146da104..93c98bf626 100644 --- a/thirdparty/glslang/SPIRV/GLSL.ext.NV.h +++ b/thirdparty/glslang/SPIRV/GLSL.ext.NV.h @@ -69,6 +69,9 @@ const char* const E_SPV_NV_mesh_shader = "SPV_NV_mesh_shader"; //SPV_NV_raytracing const char* const E_SPV_NV_ray_tracing = "SPV_NV_ray_tracing"; +//SPV_NV_ray_tracing_motion_blur +const char* const E_SPV_NV_ray_tracing_motion_blur = "SPV_NV_ray_tracing_motion_blur"; + //SPV_NV_shading_rate const char* const E_SPV_NV_shading_rate = "SPV_NV_shading_rate"; diff --git a/thirdparty/glslang/SPIRV/GlslangToSpv.cpp b/thirdparty/glslang/SPIRV/GlslangToSpv.cpp index 0c8a87e3ce..c323bcdb09 100755..100644 --- a/thirdparty/glslang/SPIRV/GlslangToSpv.cpp +++ b/thirdparty/glslang/SPIRV/GlslangToSpv.cpp @@ -1,7 +1,8 @@ // // Copyright (C) 2014-2016 LunarG, Inc. -// Copyright (C) 2015-2018 Google, Inc. +// Copyright (C) 2015-2020 Google, Inc. // Copyright (C) 2017 ARM Limited. +// Modifications Copyright (C) 2020 Advanced Micro Devices, Inc. All rights reserved. // // All rights reserved. // @@ -48,13 +49,16 @@ namespace spv { #include "GLSL.ext.EXT.h" #include "GLSL.ext.AMD.h" #include "GLSL.ext.NV.h" + #include "NonSemanticDebugPrintf.h" } // Glslang includes #include "../glslang/MachineIndependent/localintermediate.h" #include "../glslang/MachineIndependent/SymbolTable.h" #include "../glslang/Include/Common.h" -#include "../glslang/Include/revision.h" + +// Build-time generated includes +#include "glslang/build_info.h" #include <fstream> #include <iomanip> @@ -145,6 +149,7 @@ protected: spv::Decoration TranslateInterpolationDecoration(const glslang::TQualifier& qualifier); spv::Decoration TranslateAuxiliaryStorageDecoration(const glslang::TQualifier& qualifier); spv::Decoration TranslateNonUniformDecoration(const glslang::TQualifier& qualifier); + spv::Decoration TranslateNonUniformDecoration(const spv::Builder::AccessChain::CoherentFlags& coherentFlags); spv::Builder::AccessChain::CoherentFlags TranslateCoherent(const glslang::TType& type); spv::MemoryAccessMask TranslateMemoryAccess(const spv::Builder::AccessChain::CoherentFlags &coherentFlags); spv::ImageOperandsMask TranslateImageOperands(const spv::Builder::AccessChain::CoherentFlags &coherentFlags); @@ -155,6 +160,7 @@ protected: spv::SelectionControlMask TranslateSwitchControl(const glslang::TIntermSwitch&) const; spv::LoopControlMask TranslateLoopControl(const glslang::TIntermLoop&, std::vector<unsigned int>& operands) const; spv::StorageClass TranslateStorageClass(const glslang::TType&); + void TranslateLiterals(const glslang::TVector<const glslang::TIntermConstantUnion*>&, std::vector<unsigned>&) const; void addIndirectionIndexCapabilities(const glslang::TType& baseType, const glslang::TType& indexType); spv::Id createSpvVariable(const glslang::TIntermSymbol*, spv::Id forcedType); spv::Id getSampledType(const glslang::TSampler&); @@ -173,6 +179,7 @@ protected: spv::Id accessChainLoad(const glslang::TType& type); void accessChainStore(const glslang::TType& type, spv::Id rvalue); void multiTypeStore(const glslang::TType&, spv::Id rValue); + spv::Id convertLoadedBoolInUniformToUint(const glslang::TType& type, spv::Id nominalTypeId, spv::Id loadedId); glslang::TLayoutPacking getExplicitLayout(const glslang::TType& type) const; int getArrayStride(const glslang::TType& arrayType, glslang::TLayoutPacking, glslang::TLayoutMatrix); int getMatrixStride(const glslang::TType& matrixType, glslang::TLayoutPacking, glslang::TLayoutMatrix); @@ -185,9 +192,11 @@ protected: bool originalParam(glslang::TStorageQualifier, const glslang::TType&, bool implicitThisParam); void makeFunctions(const glslang::TIntermSequence&); void makeGlobalInitializers(const glslang::TIntermSequence&); + void collectRayTracingLinkerObjects(); void visitFunctions(const glslang::TIntermSequence&); void handleFunctionEntry(const glslang::TIntermAggregate* node); - void translateArguments(const glslang::TIntermAggregate& node, std::vector<spv::Id>& arguments, spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags); + void translateArguments(const glslang::TIntermAggregate& node, std::vector<spv::Id>& arguments, + spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags); void translateArguments(glslang::TIntermUnary& node, std::vector<spv::Id>& arguments); spv::Id createImageTextureFunctionCall(glslang::TIntermOperator* node); spv::Id handleUserFunctionCall(const glslang::TIntermAggregate*); @@ -196,28 +205,36 @@ protected: glslang::TBasicType typeProxy, bool reduceComparison = true); spv::Id createBinaryMatrixOperation(spv::Op, OpDecorations&, spv::Id typeId, spv::Id left, spv::Id right); spv::Id createUnaryOperation(glslang::TOperator op, OpDecorations&, spv::Id typeId, spv::Id operand, - glslang::TBasicType typeProxy, const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags); + glslang::TBasicType typeProxy, + const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags); spv::Id createUnaryMatrixOperation(spv::Op op, OpDecorations&, spv::Id typeId, spv::Id operand, glslang::TBasicType typeProxy); spv::Id createConversion(glslang::TOperator op, OpDecorations&, spv::Id destTypeId, spv::Id operand, glslang::TBasicType typeProxy); spv::Id createIntWidthConversion(glslang::TOperator op, spv::Id operand, int vectorSize); spv::Id makeSmearedConstant(spv::Id constant, int vectorSize); - spv::Id createAtomicOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy, const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags); - spv::Id createInvocationsOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy); - spv::Id CreateInvocationsVectorOperation(spv::Op op, spv::GroupOperation groupOperation, spv::Id typeId, std::vector<spv::Id>& operands); - spv::Id createSubgroupOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy); - spv::Id createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy); + spv::Id createAtomicOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, + std::vector<spv::Id>& operands, glslang::TBasicType typeProxy, + const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags); + spv::Id createInvocationsOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, + glslang::TBasicType typeProxy); + spv::Id CreateInvocationsVectorOperation(spv::Op op, spv::GroupOperation groupOperation, + spv::Id typeId, std::vector<spv::Id>& operands); + spv::Id createSubgroupOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, + glslang::TBasicType typeProxy); + spv::Id createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, + std::vector<spv::Id>& operands, glslang::TBasicType typeProxy); spv::Id createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId); spv::Id getSymbolId(const glslang::TIntermSymbol* node); void addMeshNVDecoration(spv::Id id, int member, const glslang::TQualifier & qualifier); spv::Id createSpvConstant(const glslang::TIntermTyped&); - spv::Id createSpvConstantFromConstUnionArray(const glslang::TType& type, const glslang::TConstUnionArray&, int& nextConst, bool specConstant); + spv::Id createSpvConstantFromConstUnionArray(const glslang::TType& type, const glslang::TConstUnionArray&, + int& nextConst, bool specConstant); bool isTrivialLeaf(const glslang::TIntermTyped* node); bool isTrivial(const glslang::TIntermTyped* node); spv::Id createShortCircuit(glslang::TOperator, glslang::TIntermTyped& left, glslang::TIntermTyped& right); spv::Id getExtBuiltins(const char* name); - std::pair<spv::Id, spv::Id> getForcedType(spv::BuiltIn, const glslang::TType&); + std::pair<spv::Id, spv::Id> getForcedType(glslang::TBuiltInVariable builtIn, const glslang::TType&); spv::Id translateForcedType(spv::Id object); spv::Id createCompositeConstruct(spv::Id typeId, std::vector<spv::Id> constituents); @@ -233,21 +250,24 @@ protected: spv::Builder builder; bool inEntryPoint; bool entryPointTerminated; - bool linkageOnly; // true when visiting the set of objects in the AST present only for establishing interface, whether or not they were statically used + bool linkageOnly; // true when visiting the set of objects in the AST present only for + // establishing interface, whether or not they were statically used std::set<spv::Id> iOSet; // all input/output variables from either static use or declaration of interface const glslang::TIntermediate* glslangIntermediate; bool nanMinMaxClamp; // true if use NMin/NMax/NClamp instead of FMin/FMax/FClamp spv::Id stdBuiltins; - std::unordered_map<const char*, spv::Id> extBuiltinMap; + spv::Id nonSemanticDebugPrintf; + std::unordered_map<std::string, spv::Id> extBuiltinMap; - std::unordered_map<int, spv::Id> symbolValues; - std::unordered_set<int> rValueParameters; // set of formal function parameters passed as rValues, rather than a pointer + std::unordered_map<long long, spv::Id> symbolValues; + std::unordered_set<long long> rValueParameters; // set of formal function parameters passed as rValues, + // rather than a pointer std::unordered_map<std::string, spv::Function*> functionMap; std::unordered_map<const glslang::TTypeList*, spv::Id> structMap[glslang::ElpCount][glslang::ElmCount]; // for mapping glslang block indices to spv indices (e.g., due to hidden members): - std::unordered_map<int, std::vector<int>> memberRemapper; + std::unordered_map<long long, std::vector<int>> memberRemapper; // for mapping glslang symbol struct to symbol Id - std::unordered_map<const glslang::TTypeList*, int> glslangTypeToIdMap; + std::unordered_map<const glslang::TTypeList*, long long> glslangTypeToIdMap; std::stack<bool> breakForLoop; // false means break for switch std::unordered_map<std::string, const glslang::TIntermSymbol*> counterOriginator; // Map pointee types for EbtReference to their forward pointers @@ -256,6 +276,9 @@ protected: // requiring local translation to and from SPIR-V type on every access. // Maps <builtin-variable-id -> AST-required-type-id> std::unordered_map<spv::Id, spv::Id> forceType; + + // Used later for generating OpTraceKHR/OpExecuteCallableKHR + std::unordered_map<unsigned int, glslang::TIntermSymbol *> locationToSymbol[2]; }; // @@ -267,6 +290,8 @@ spv::SourceLanguage TranslateSourceLanguage(glslang::EShSource source, EProfile { #ifdef GLSLANG_WEB return spv::SourceLanguageESSL; +#elif defined(GLSLANG_ANGLE) + return spv::SourceLanguageGLSL; #endif switch (source) { @@ -299,12 +324,12 @@ spv::ExecutionModel TranslateExecutionModel(EShLanguage stage) case EShLangTessControl: return spv::ExecutionModelTessellationControl; case EShLangTessEvaluation: return spv::ExecutionModelTessellationEvaluation; case EShLangGeometry: return spv::ExecutionModelGeometry; - case EShLangRayGenNV: return spv::ExecutionModelRayGenerationNV; - case EShLangIntersectNV: return spv::ExecutionModelIntersectionNV; - case EShLangAnyHitNV: return spv::ExecutionModelAnyHitNV; - case EShLangClosestHitNV: return spv::ExecutionModelClosestHitNV; - case EShLangMissNV: return spv::ExecutionModelMissNV; - case EShLangCallableNV: return spv::ExecutionModelCallableNV; + case EShLangRayGen: return spv::ExecutionModelRayGenerationKHR; + case EShLangIntersect: return spv::ExecutionModelIntersectionKHR; + case EShLangAnyHit: return spv::ExecutionModelAnyHitKHR; + 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; #endif @@ -357,12 +382,13 @@ spv::Decoration TranslateBlockDecoration(const glslang::TType& type, bool useSto case glslang::EvqBuffer: return useStorageBuffer ? spv::DecorationBlock : spv::DecorationBufferBlock; case glslang::EvqVaryingIn: return spv::DecorationBlock; case glslang::EvqVaryingOut: return spv::DecorationBlock; + case glslang::EvqShared: return spv::DecorationBlock; #ifndef GLSLANG_WEB - case glslang::EvqPayloadNV: return spv::DecorationBlock; - case glslang::EvqPayloadInNV: return spv::DecorationBlock; - case glslang::EvqHitAttrNV: return spv::DecorationBlock; - case glslang::EvqCallableDataNV: return spv::DecorationBlock; - case glslang::EvqCallableDataInNV: return spv::DecorationBlock; + case glslang::EvqPayload: return spv::DecorationBlock; + case glslang::EvqPayloadIn: return spv::DecorationBlock; + case glslang::EvqHitAttr: return spv::DecorationBlock; + case glslang::EvqCallableData: return spv::DecorationBlock; + case glslang::EvqCallableDataIn: return spv::DecorationBlock; #endif default: assert(0); @@ -374,7 +400,8 @@ spv::Decoration TranslateBlockDecoration(const glslang::TType& type, bool useSto } // Translate glslang type to SPIR-V memory decorations. -void TranslateMemoryDecoration(const glslang::TQualifier& qualifier, std::vector<spv::Decoration>& memory, bool useVulkanMemoryModel) +void TranslateMemoryDecoration(const glslang::TQualifier& qualifier, std::vector<spv::Decoration>& memory, + bool useVulkanMemoryModel) { if (!useVulkanMemoryModel) { if (qualifier.isCoherent()) @@ -412,6 +439,7 @@ spv::Decoration TranslateLayoutDecoration(const glslang::TType& type, glslang::T break; case glslang::EbtBlock: switch (type.getQualifier().storage) { + case glslang::EvqShared: case glslang::EvqUniform: case glslang::EvqBuffer: switch (type.getQualifier().layoutPacking) { @@ -433,11 +461,11 @@ spv::Decoration TranslateLayoutDecoration(const glslang::TType& type, glslang::T } return spv::DecorationMax; #ifndef GLSLANG_WEB - case glslang::EvqPayloadNV: - case glslang::EvqPayloadInNV: - case glslang::EvqHitAttrNV: - case glslang::EvqCallableDataNV: - case glslang::EvqCallableDataInNV: + case glslang::EvqPayload: + case glslang::EvqPayloadIn: + case glslang::EvqHitAttr: + case glslang::EvqCallableData: + case glslang::EvqCallableDataIn: return spv::DecorationMax; #endif default: @@ -520,6 +548,20 @@ spv::Decoration TGlslangToSpvTraverser::TranslateNonUniformDecoration(const glsl return spv::DecorationMax; } +// If lvalue flags contains nonUniform, return SPIR-V NonUniform decoration. +spv::Decoration TGlslangToSpvTraverser::TranslateNonUniformDecoration( + const spv::Builder::AccessChain::CoherentFlags& coherentFlags) +{ +#ifndef GLSLANG_WEB + if (coherentFlags.isNonUniform()) { + builder.addIncorporatedExtension("SPV_EXT_descriptor_indexing", spv::Spv_1_5); + builder.addCapability(spv::CapabilityShaderNonUniformEXT); + return spv::DecorationNonUniformEXT; + } else +#endif + return spv::DecorationMax; +} + spv::MemoryAccessMask TGlslangToSpvTraverser::TranslateMemoryAccess( const spv::Builder::AccessChain::CoherentFlags &coherentFlags) { @@ -529,15 +571,11 @@ spv::MemoryAccessMask TGlslangToSpvTraverser::TranslateMemoryAccess( if (!glslangIntermediate->usingVulkanMemoryModel() || coherentFlags.isImage) return mask; - if (coherentFlags.volatil || - coherentFlags.coherent || - coherentFlags.devicecoherent || - coherentFlags.queuefamilycoherent || - coherentFlags.workgroupcoherent || - coherentFlags.subgroupcoherent) { + if (coherentFlags.isVolatile() || coherentFlags.anyCoherent()) { mask = mask | spv::MemoryAccessMakePointerAvailableKHRMask | spv::MemoryAccessMakePointerVisibleKHRMask; } + if (coherentFlags.nonprivate) { mask = mask | spv::MemoryAccessNonPrivatePointerKHRMask; } @@ -562,11 +600,7 @@ spv::ImageOperandsMask TGlslangToSpvTraverser::TranslateImageOperands( return mask; if (coherentFlags.volatil || - coherentFlags.coherent || - coherentFlags.devicecoherent || - coherentFlags.queuefamilycoherent || - coherentFlags.workgroupcoherent || - coherentFlags.subgroupcoherent) { + coherentFlags.anyCoherent()) { mask = mask | spv::ImageOperandsMakeTexelAvailableKHRMask | spv::ImageOperandsMakeTexelVisibleKHRMask; } @@ -595,17 +629,15 @@ spv::Builder::AccessChain::CoherentFlags TGlslangToSpvTraverser::TranslateCohere flags.workgroupcoherent = type.getQualifier().workgroupcoherent || type.getQualifier().storage == glslang::EvqShared; flags.subgroupcoherent = type.getQualifier().subgroupcoherent; + flags.shadercallcoherent = type.getQualifier().shadercallcoherent; flags.volatil = type.getQualifier().volatil; // *coherent variables are implicitly nonprivate in GLSL flags.nonprivate = type.getQualifier().nonprivate || - flags.subgroupcoherent || - flags.workgroupcoherent || - flags.queuefamilycoherent || - flags.devicecoherent || - flags.coherent || + flags.anyCoherent() || flags.volatil; flags.isImage = type.getBasicType() == glslang::EbtSampler; #endif + flags.nonUniform = type.getQualifier().nonUniform; return flags; } @@ -626,6 +658,8 @@ spv::Scope TGlslangToSpvTraverser::TranslateMemoryScope( scope = spv::ScopeWorkgroup; } else if (coherentFlags.subgroupcoherent) { scope = spv::ScopeSubgroup; + } else if (coherentFlags.shadercallcoherent) { + scope = spv::ScopeShaderCallKHR; } if (glslangIntermediate->usingVulkanMemoryModel() && scope == spv::ScopeDevice) { builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR); @@ -640,7 +674,8 @@ spv::Scope TGlslangToSpvTraverser::TranslateMemoryScope( // is generated only when using the variable in an executable instruction, but not when // just declaring a struct member variable with it. This is true for PointSize, // ClipDistance, and CullDistance. -spv::BuiltIn TGlslangToSpvTraverser::TranslateBuiltInDecoration(glslang::TBuiltInVariable builtIn, bool memberDeclaration) +spv::BuiltIn TGlslangToSpvTraverser::TranslateBuiltInDecoration(glslang::TBuiltInVariable builtIn, + bool memberDeclaration) { switch (builtIn) { case glslang::EbvPointSize: @@ -698,13 +733,20 @@ spv::BuiltIn TGlslangToSpvTraverser::TranslateBuiltInDecoration(glslang::TBuiltI return spv::BuiltInCullDistance; case glslang::EbvViewportIndex: - builder.addCapability(spv::CapabilityMultiViewport); + if (glslangIntermediate->getStage() == EShLangGeometry || + glslangIntermediate->getStage() == EShLangFragment) { + builder.addCapability(spv::CapabilityMultiViewport); + } if (glslangIntermediate->getStage() == EShLangVertex || glslangIntermediate->getStage() == EShLangTessControl || glslangIntermediate->getStage() == EShLangTessEvaluation) { - builder.addIncorporatedExtension(spv::E_SPV_EXT_shader_viewport_index_layer, spv::Spv_1_5); - builder.addCapability(spv::CapabilityShaderViewportIndexLayerEXT); + if (builder.getSpvVersion() < spv::Spv_1_5) { + builder.addIncorporatedExtension(spv::E_SPV_EXT_shader_viewport_index_layer, spv::Spv_1_5); + builder.addCapability(spv::CapabilityShaderViewportIndexLayerEXT); + } + else + builder.addCapability(spv::CapabilityShaderViewportIndex); } return spv::BuiltInViewportIndex; @@ -723,13 +765,19 @@ spv::BuiltIn TGlslangToSpvTraverser::TranslateBuiltInDecoration(glslang::TBuiltI if (glslangIntermediate->getStage() == EShLangMeshNV) { return spv::BuiltInLayer; } - builder.addCapability(spv::CapabilityGeometry); + if (glslangIntermediate->getStage() == EShLangGeometry || + glslangIntermediate->getStage() == EShLangFragment) { + builder.addCapability(spv::CapabilityGeometry); + } if (glslangIntermediate->getStage() == EShLangVertex || glslangIntermediate->getStage() == EShLangTessControl || glslangIntermediate->getStage() == EShLangTessEvaluation) { - builder.addIncorporatedExtension(spv::E_SPV_EXT_shader_viewport_index_layer, spv::Spv_1_5); - builder.addCapability(spv::CapabilityShaderViewportIndexLayerEXT); + if (builder.getSpvVersion() < spv::Spv_1_5) { + builder.addIncorporatedExtension(spv::E_SPV_EXT_shader_viewport_index_layer, spv::Spv_1_5); + builder.addCapability(spv::CapabilityShaderViewportIndexLayerEXT); + } else + builder.addCapability(spv::CapabilityShaderLayer); } return spv::BuiltInLayer; @@ -758,6 +806,16 @@ spv::BuiltIn TGlslangToSpvTraverser::TranslateBuiltInDecoration(glslang::TBuiltI builder.addCapability(spv::CapabilityStencilExportEXT); return spv::BuiltInFragStencilRefEXT; + case glslang::EbvShadingRateKHR: + builder.addExtension(spv::E_SPV_KHR_fragment_shading_rate); + builder.addCapability(spv::CapabilityFragmentShadingRateKHR); + return spv::BuiltInShadingRateKHR; + + case glslang::EbvPrimitiveShadingRateKHR: + builder.addExtension(spv::E_SPV_KHR_fragment_shading_rate); + builder.addCapability(spv::CapabilityFragmentShadingRateKHR); + return spv::BuiltInPrimitiveShadingRateKHR; + case glslang::EbvInvocationId: return spv::BuiltInInvocationId; case glslang::EbvTessLevelInner: return spv::BuiltInTessLevelInner; case glslang::EbvTessLevelOuter: return spv::BuiltInTessLevelOuter; @@ -933,34 +991,52 @@ spv::BuiltIn TGlslangToSpvTraverser::TranslateBuiltInDecoration(glslang::TBuiltI return spv::BuiltInInvocationsPerPixelNV; // ray tracing - case glslang::EbvLaunchIdNV: - return spv::BuiltInLaunchIdNV; - case glslang::EbvLaunchSizeNV: - return spv::BuiltInLaunchSizeNV; - case glslang::EbvWorldRayOriginNV: - return spv::BuiltInWorldRayOriginNV; - case glslang::EbvWorldRayDirectionNV: - return spv::BuiltInWorldRayDirectionNV; - case glslang::EbvObjectRayOriginNV: - return spv::BuiltInObjectRayOriginNV; - case glslang::EbvObjectRayDirectionNV: - return spv::BuiltInObjectRayDirectionNV; - case glslang::EbvRayTminNV: - return spv::BuiltInRayTminNV; - case glslang::EbvRayTmaxNV: - return spv::BuiltInRayTmaxNV; - case glslang::EbvInstanceCustomIndexNV: - return spv::BuiltInInstanceCustomIndexNV; - case glslang::EbvHitTNV: - return spv::BuiltInHitTNV; - case glslang::EbvHitKindNV: - return spv::BuiltInHitKindNV; - case glslang::EbvObjectToWorldNV: - return spv::BuiltInObjectToWorldNV; - case glslang::EbvWorldToObjectNV: - return spv::BuiltInWorldToObjectNV; - case glslang::EbvIncomingRayFlagsNV: - return spv::BuiltInIncomingRayFlagsNV; + case glslang::EbvLaunchId: + return spv::BuiltInLaunchIdKHR; + case glslang::EbvLaunchSize: + return spv::BuiltInLaunchSizeKHR; + case glslang::EbvWorldRayOrigin: + return spv::BuiltInWorldRayOriginKHR; + case glslang::EbvWorldRayDirection: + return spv::BuiltInWorldRayDirectionKHR; + case glslang::EbvObjectRayOrigin: + return spv::BuiltInObjectRayOriginKHR; + case glslang::EbvObjectRayDirection: + return spv::BuiltInObjectRayDirectionKHR; + case glslang::EbvRayTmin: + return spv::BuiltInRayTminKHR; + case glslang::EbvRayTmax: + return spv::BuiltInRayTmaxKHR; + case glslang::EbvInstanceCustomIndex: + return spv::BuiltInInstanceCustomIndexKHR; + case glslang::EbvHitT: + { + // this is a GLSL alias of RayTmax + // in SPV_NV_ray_tracing it has a dedicated builtin + // but in SPV_KHR_ray_tracing it gets mapped to RayTmax + auto& extensions = glslangIntermediate->getRequestedExtensions(); + if (extensions.find("GL_NV_ray_tracing") != extensions.end()) { + return spv::BuiltInHitTNV; + } else { + return spv::BuiltInRayTmaxKHR; + } + } + case glslang::EbvHitKind: + return spv::BuiltInHitKindKHR; + case glslang::EbvObjectToWorld: + case glslang::EbvObjectToWorld3x4: + return spv::BuiltInObjectToWorldKHR; + case glslang::EbvWorldToObject: + case glslang::EbvWorldToObject3x4: + return spv::BuiltInWorldToObjectKHR; + case glslang::EbvIncomingRayFlags: + return spv::BuiltInIncomingRayFlagsKHR; + case glslang::EbvGeometryIndex: + return spv::BuiltInRayGeometryIndexKHR; + case glslang::EbvCurrentRayTimeNV: + builder.addExtension(spv::E_SPV_NV_ray_tracing_motion_blur); + builder.addCapability(spv::CapabilityRayTracingMotionBlurNV); + return spv::BuiltInCurrentRayTimeNV; // barycentrics case glslang::EbvBaryCoordNV: @@ -1056,6 +1132,10 @@ spv::ImageFormat TGlslangToSpvTraverser::TranslateImageFormat(const glslang::TTy builder.addCapability(spv::CapabilityStorageImageExtendedFormats); break; + case glslang::ElfR64ui: + case glslang::ElfR64i: + builder.addExtension(spv::E_SPV_EXT_shader_image_int64); + builder.addCapability(spv::CapabilityInt64ImageEXT); default: break; } @@ -1102,11 +1182,14 @@ spv::ImageFormat TGlslangToSpvTraverser::TranslateImageFormat(const glslang::TTy case glslang::ElfRg8ui: return spv::ImageFormatRg8ui; case glslang::ElfR16ui: return spv::ImageFormatR16ui; case glslang::ElfR8ui: return spv::ImageFormatR8ui; + case glslang::ElfR64ui: return spv::ImageFormatR64ui; + case glslang::ElfR64i: return spv::ImageFormatR64i; default: return spv::ImageFormatMax; } } -spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSelectionControl(const glslang::TIntermSelection& selectionNode) const +spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSelectionControl( + const glslang::TIntermSelection& selectionNode) const { if (selectionNode.getFlatten()) return spv::SelectionControlFlattenMask; @@ -1115,7 +1198,8 @@ spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSelectionControl(cons return spv::SelectionControlMaskNone; } -spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSwitchControl(const glslang::TIntermSwitch& switchNode) const +spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSwitchControl(const glslang::TIntermSwitch& switchNode) + const { if (switchNode.getFlatten()) return spv::SelectionControlFlattenMask; @@ -1169,6 +1253,12 @@ spv::LoopControlMask TGlslangToSpvTraverser::TranslateLoopControl(const glslang: // Translate glslang type to SPIR-V storage class. spv::StorageClass TGlslangToSpvTraverser::TranslateStorageClass(const glslang::TType& type) { + if (type.getBasicType() == glslang::EbtRayQuery) + return spv::StorageClassPrivate; +#ifndef GLSLANG_WEB + if (type.getQualifier().isSpirvByReference()) + return spv::StorageClassFunction; +#endif if (type.getQualifier().isPipeInput()) return spv::StorageClassInput; if (type.getQualifier().isPipeOutput()) @@ -1183,8 +1273,8 @@ spv::StorageClass TGlslangToSpvTraverser::TranslateStorageClass(const glslang::T } if (type.getQualifier().isUniformOrBuffer() && - type.getQualifier().isShaderRecordNV()) { - return spv::StorageClassShaderRecordBufferNV; + type.getQualifier().isShaderRecord()) { + return spv::StorageClassShaderRecordBufferKHR; } if (glslangIntermediate->usingStorageBuffer() && type.getQualifier().storage == glslang::EvqBuffer) { @@ -1200,17 +1290,24 @@ spv::StorageClass TGlslangToSpvTraverser::TranslateStorageClass(const glslang::T return spv::StorageClassUniformConstant; } + if (type.getQualifier().storage == glslang::EvqShared && type.getBasicType() == glslang::EbtBlock) { + builder.addExtension(spv::E_SPV_KHR_workgroup_memory_explicit_layout); + builder.addCapability(spv::CapabilityWorkgroupMemoryExplicitLayoutKHR); + return spv::StorageClassWorkgroup; + } + switch (type.getQualifier().storage) { case glslang::EvqGlobal: return spv::StorageClassPrivate; case glslang::EvqConstReadOnly: return spv::StorageClassFunction; case glslang::EvqTemporary: return spv::StorageClassFunction; case glslang::EvqShared: return spv::StorageClassWorkgroup; #ifndef GLSLANG_WEB - case glslang::EvqPayloadNV: return spv::StorageClassRayPayloadNV; - case glslang::EvqPayloadInNV: return spv::StorageClassIncomingRayPayloadNV; - case glslang::EvqHitAttrNV: return spv::StorageClassHitAttributeNV; - case glslang::EvqCallableDataNV: return spv::StorageClassCallableDataNV; - case glslang::EvqCallableDataInNV: return spv::StorageClassIncomingCallableDataNV; + case glslang::EvqPayload: return spv::StorageClassRayPayloadKHR; + case glslang::EvqPayloadIn: return spv::StorageClassIncomingRayPayloadKHR; + case glslang::EvqHitAttr: return spv::StorageClassHitAttributeKHR; + case glslang::EvqCallableData: return spv::StorageClassCallableDataKHR; + case glslang::EvqCallableDataIn: return spv::StorageClassIncomingCallableDataKHR; + case glslang::EvqSpirvStorageClass: return static_cast<spv::StorageClass>(type.getQualifier().spirvStorageClass); #endif default: assert(0); @@ -1220,6 +1317,52 @@ spv::StorageClass TGlslangToSpvTraverser::TranslateStorageClass(const glslang::T return spv::StorageClassFunction; } +// Translate glslang constants to SPIR-V literals +void TGlslangToSpvTraverser::TranslateLiterals(const glslang::TVector<const glslang::TIntermConstantUnion*>& constants, + std::vector<unsigned>& literals) const +{ + for (auto constant : constants) { + if (constant->getBasicType() == glslang::EbtFloat) { + float floatValue = static_cast<float>(constant->getConstArray()[0].getDConst()); + unsigned literal = *reinterpret_cast<unsigned*>(&floatValue); + literals.push_back(literal); + } else if (constant->getBasicType() == glslang::EbtInt) { + unsigned literal = constant->getConstArray()[0].getIConst(); + literals.push_back(literal); + } else if (constant->getBasicType() == glslang::EbtUint) { + unsigned literal = constant->getConstArray()[0].getUConst(); + literals.push_back(literal); + } else if (constant->getBasicType() == glslang::EbtBool) { + unsigned literal = constant->getConstArray()[0].getBConst(); + literals.push_back(literal); + } else if (constant->getBasicType() == glslang::EbtString) { + auto str = constant->getConstArray()[0].getSConst()->c_str(); + unsigned literal = 0; + char* literalPtr = reinterpret_cast<char*>(&literal); + unsigned charCount = 0; + char ch = 0; + do { + ch = *(str++); + *(literalPtr++) = ch; + ++charCount; + if (charCount == 4) { + literals.push_back(literal); + literalPtr = reinterpret_cast<char*>(&literal); + charCount = 0; + } + } while (ch != 0); + + // Partial literal is padded with 0 + if (charCount > 0) { + for (; charCount < 4; ++charCount) + *(literalPtr++) = 0; + literals.push_back(literal); + } + } else + assert(0); // Unexpected type + } +} + // Add capabilities pertaining to how an array is indexed. void TGlslangToSpvTraverser::addIndirectionIndexCapabilities(const glslang::TType& baseType, const glslang::TType& indexType) @@ -1270,14 +1413,15 @@ bool IsDescriptorResource(const glslang::TType& type) // uniform and buffer blocks are included, unless it is a push_constant if (type.getBasicType() == glslang::EbtBlock) return type.getQualifier().isUniformOrBuffer() && - ! type.getQualifier().isShaderRecordNV() && + ! type.getQualifier().isShaderRecord() && ! type.getQualifier().isPushConstant(); // non block... // basically samplerXXX/subpass/sampler/texture are all included // if they are the global-scope-class, not the function parameter // (or local, if they ever exist) class. - if (type.getBasicType() == glslang::EbtSampler) + if (type.getBasicType() == glslang::EbtSampler || + type.getBasicType() == glslang::EbtAccStruct) return type.getQualifier().isUniformOrBuffer(); // None of the above. @@ -1320,6 +1464,8 @@ void InheritQualifiers(glslang::TQualifier& child, const glslang::TQualifier& pa child.workgroupcoherent = true; if (parent.subgroupcoherent) child.subgroupcoherent = true; + if (parent.shadercallcoherent) + child.shadercallcoherent = true; if (parent.nonprivate) child.nonprivate = true; if (parent.volatil) @@ -1331,6 +1477,8 @@ void InheritQualifiers(glslang::TQualifier& child, const glslang::TQualifier& pa if (parent.writeonly) child.writeonly = true; #endif + if (parent.nonUniform) + child.nonUniform = true; } bool HasNonLayoutQualifiers(const glslang::TType& type, const glslang::TQualifier& qualifier) @@ -1350,16 +1498,18 @@ bool HasNonLayoutQualifiers(const glslang::TType& type, const glslang::TQualifie // Implement the TGlslangToSpvTraverser class. // -TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const glslang::TIntermediate* glslangIntermediate, - spv::SpvBuildLogger* buildLogger, glslang::SpvOptions& options) - : TIntermTraverser(true, false, true), - options(options), - shaderEntry(nullptr), currentFunction(nullptr), - sequenceDepth(0), logger(buildLogger), - builder(spvVersion, (glslang::GetKhronosToolId() << 16) | glslang::GetSpirvGeneratorVersion(), logger), - inEntryPoint(false), entryPointTerminated(false), linkageOnly(false), - glslangIntermediate(glslangIntermediate), - nanMinMaxClamp(glslangIntermediate->getNanMinMaxClamp()) +TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, + const glslang::TIntermediate* glslangIntermediate, + spv::SpvBuildLogger* buildLogger, glslang::SpvOptions& options) : + TIntermTraverser(true, false, true), + options(options), + shaderEntry(nullptr), currentFunction(nullptr), + sequenceDepth(0), logger(buildLogger), + builder(spvVersion, (glslang::GetKhronosToolId() << 16) | glslang::GetSpirvGeneratorVersion(), logger), + inEntryPoint(false), entryPointTerminated(false), linkageOnly(false), + glslangIntermediate(glslangIntermediate), + nanMinMaxClamp(glslangIntermediate->getNanMinMaxClamp()), + nonSemanticDebugPrintf(0) { spv::ExecutionModel executionModel = TranslateExecutionModel(glslangIntermediate->getStage()); @@ -1400,9 +1550,9 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const gl if (glslangIntermediate->usingPhysicalStorageBuffer()) { addressingModel = spv::AddressingModelPhysicalStorageBuffer64EXT; - builder.addIncorporatedExtension(spv::E_SPV_EXT_physical_storage_buffer, spv::Spv_1_5); + builder.addIncorporatedExtension(spv::E_SPV_KHR_physical_storage_buffer, spv::Spv_1_5); builder.addCapability(spv::CapabilityPhysicalStorageBufferAddressesEXT); - }; + } if (glslangIntermediate->usingVulkanMemoryModel()) { memoryModel = spv::MemoryModelVulkanKHR; builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); @@ -1429,6 +1579,17 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const gl builder.addExecutionMode(shaderEntry, spv::ExecutionModeXfb); } + if (glslangIntermediate->getLayoutPrimitiveCulling()) { + builder.addCapability(spv::CapabilityRayTraversalPrimitiveCullingKHR); + } + +#ifndef GLSLANG_WEB + if (glslangIntermediate->getSubgroupUniformControlFlow()) { + builder.addExtension(spv::E_SPV_KHR_subgroup_uniform_control_flow); + builder.addExecutionMode(shaderEntry, spv::ExecutionModeSubgroupUniformControlFlowKHR); + } +#endif + unsigned int mode; switch (glslangIntermediate->getStage()) { case EShLangVertex: @@ -1454,14 +1615,16 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const gl builder.addExtension(spv::E_SPV_KHR_post_depth_coverage); } - if (glslangIntermediate->getDepth() != glslang::EldUnchanged && glslangIntermediate->isDepthReplacing()) + if (glslangIntermediate->isDepthReplacing()) builder.addExecutionMode(shaderEntry, spv::ExecutionModeDepthReplacing); #ifndef GLSLANG_WEB + switch(glslangIntermediate->getDepth()) { - case glslang::EldGreater: mode = spv::ExecutionModeDepthGreater; break; - case glslang::EldLess: mode = spv::ExecutionModeDepthLess; break; - default: mode = spv::ExecutionModeMax; break; + case glslang::EldGreater: mode = spv::ExecutionModeDepthGreater; break; + case glslang::EldLess: mode = spv::ExecutionModeDepthLess; break; + case glslang::EldUnchanged: mode = spv::ExecutionModeDepthUnchanged; break; + default: mode = spv::ExecutionModeMax; break; } if (mode != spv::ExecutionModeMax) builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); @@ -1495,7 +1658,7 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const gl builder.addExtension(spv::E_SPV_EXT_fragment_shader_interlock); } #endif - break; + break; case EShLangCompute: builder.addCapability(spv::CapabilityShader); @@ -1520,7 +1683,8 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const gl glslang::TLayoutGeometry primitive; if (glslangIntermediate->getStage() == EShLangTessControl) { - builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices()); + builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, + glslangIntermediate->getVertices()); primitive = glslangIntermediate->getOutputPrimitive(); } else { primitive = glslangIntermediate->getInputPrimitive(); @@ -1582,15 +1746,24 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const gl builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices()); break; - case EShLangRayGenNV: - case EShLangIntersectNV: - case EShLangAnyHitNV: - case EShLangClosestHitNV: - case EShLangMissNV: - case EShLangCallableNV: - builder.addCapability(spv::CapabilityRayTracingNV); - builder.addExtension("SPV_NV_ray_tracing"); + case EShLangRayGen: + case EShLangIntersect: + case EShLangAnyHit: + case EShLangClosestHit: + case EShLangMiss: + case EShLangCallable: + { + auto& extensions = glslangIntermediate->getRequestedExtensions(); + if (extensions.find("GL_NV_ray_tracing") == extensions.end()) { + builder.addCapability(spv::CapabilityRayTracingKHR); + builder.addExtension("SPV_KHR_ray_tracing"); + } + else { + builder.addCapability(spv::CapabilityRayTracingNV); + builder.addExtension("SPV_NV_ray_tracing"); + } break; + } case EShLangTaskNV: case EShLangMeshNV: builder.addCapability(spv::CapabilityMeshShadingNV); @@ -1599,8 +1772,10 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const gl glslangIntermediate->getLocalSize(1), glslangIntermediate->getLocalSize(2)); if (glslangIntermediate->getStage() == EShLangMeshNV) { - builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices()); - builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputPrimitivesNV, glslangIntermediate->getPrimitives()); + builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, + glslangIntermediate->getVertices()); + builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputPrimitivesNV, + glslangIntermediate->getPrimitives()); switch (glslangIntermediate->getOutputPrimitive()) { case glslang::ElgPoints: mode = spv::ExecutionModeOutputPoints; break; @@ -1617,6 +1792,53 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const gl default: break; } + +#ifndef GLSLANG_WEB + // + // Add SPIR-V requirements (GL_EXT_spirv_intrinsics) + // + if (glslangIntermediate->hasSpirvRequirement()) { + const glslang::TSpirvRequirement& spirvRequirement = glslangIntermediate->getSpirvRequirement(); + + // Add SPIR-V extension requirement + for (auto& extension : spirvRequirement.extensions) + builder.addExtension(extension.c_str()); + + // Add SPIR-V capability requirement + for (auto capability : spirvRequirement.capabilities) + builder.addCapability(static_cast<spv::Capability>(capability)); + } + + // + // Add SPIR-V execution mode qualifiers (GL_EXT_spirv_intrinsics) + // + if (glslangIntermediate->hasSpirvExecutionMode()) { + const glslang::TSpirvExecutionMode spirvExecutionMode = glslangIntermediate->getSpirvExecutionMode(); + + // Add spirv_execution_mode + for (auto& mode : spirvExecutionMode.modes) { + if (!mode.second.empty()) { + std::vector<unsigned> literals; + TranslateLiterals(mode.second, literals); + builder.addExecutionMode(shaderEntry, static_cast<spv::ExecutionMode>(mode.first), literals); + } else + builder.addExecutionMode(shaderEntry, static_cast<spv::ExecutionMode>(mode.first)); + } + + // Add spirv_execution_mode_id + for (auto& modeId : spirvExecutionMode.modeIds) { + std::vector<spv::Id> operandIds; + assert(!modeId.second.empty()); + for (auto extraOperand : modeId.second) { + int nextConst = 0; + spv::Id operandId = createSpvConstantFromConstUnionArray( + extraOperand->getType(), extraOperand->getConstArray(), nextConst, false); + operandIds.push_back(operandId); + } + builder.addExecutionModeId(shaderEntry, static_cast<spv::ExecutionMode>(modeId.first), operandIds); + } + } +#endif } // Finish creating SPV, after the traversal is complete. @@ -1669,25 +1891,35 @@ 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) + return; +#endif + // getSymbolId() will set up all the IO decorations on the first call. // Formal function parameters were mapped during makeFunctions(). spv::Id id = getSymbolId(symbol); if (builder.isPointer(id)) { - // Include all "static use" and "linkage only" interface variables on the OpEntryPoint instruction - // Consider adding to the OpEntryPoint interface list. - // Only looking at structures if they have at least one member. - if (!symbol->getType().isStruct() || symbol->getType().getStruct()->size() > 0) { - spv::StorageClass sc = builder.getStorageClass(id); - // Before SPIR-V 1.4, we only want to include Input and Output. - // Starting with SPIR-V 1.4, we want all globals. - if ((glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4 && sc != spv::StorageClassFunction) || - (sc == spv::StorageClassInput || sc == spv::StorageClassOutput)) { - iOSet.insert(id); + if (!symbol->getType().getQualifier().isParamInput() && + !symbol->getType().getQualifier().isParamOutput()) { + // Include all "static use" and "linkage only" interface variables on the OpEntryPoint instruction + // Consider adding to the OpEntryPoint interface list. + // Only looking at structures if they have at least one member. + if (!symbol->getType().isStruct() || symbol->getType().getStruct()->size() > 0) { + spv::StorageClass sc = builder.getStorageClass(id); + // Before SPIR-V 1.4, we only want to include Input and Output. + // Starting with SPIR-V 1.4, we want all globals. + if ((glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4 && builder.isGlobalStorage(id)) || + (sc == spv::StorageClassInput || sc == spv::StorageClassOutput)) { + iOSet.insert(id); + } } } - // If the SPIR-V type is required to be different than the AST type, + // If the SPIR-V type is required to be different than the AST type + // (for ex SubgroupMasks or 3x4 ObjectToWorld/WorldToObject matrices), // translate now from the SPIR-V type to the AST type, for the consuming // operation. // Note this turns it from an l-value to an r-value. @@ -1807,9 +2039,11 @@ bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::T spv::Id leftRValue = accessChainLoad(node->getLeft()->getType()); // do the operation + spv::Builder::AccessChain::CoherentFlags coherentFlags = TranslateCoherent(node->getLeft()->getType()); + coherentFlags |= TranslateCoherent(node->getRight()->getType()); OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()), TranslateNoContractionDecoration(node->getType().getQualifier()), - TranslateNonUniformDecoration(node->getType().getQualifier()) }; + TranslateNonUniformDecoration(coherentFlags) }; rValue = createBinaryOperation(node->getOp(), decorations, convertGlslangToSpvType(node->getType()), leftRValue, rValue, node->getType().getBasicType()); @@ -1840,14 +2074,18 @@ bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::T if (! node->getLeft()->getType().isArray() && node->getLeft()->getType().isVector() && node->getOp() == glslang::EOpIndexDirect) { + // Swizzle is uniform so propagate uniform into access chain + spv::Builder::AccessChain::CoherentFlags coherentFlags = TranslateCoherent(node->getLeft()->getType()); + coherentFlags.nonUniform = 0; // This is essentially a hard-coded vector swizzle of size 1, // so short circuit the access-chain stuff with a swizzle. std::vector<unsigned> swizzle; swizzle.push_back(glslangIndex); int dummySize; builder.accessChainPushSwizzle(swizzle, convertGlslangToSpvType(node->getLeft()->getType()), - TranslateCoherent(node->getLeft()->getType()), - glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize)); + coherentFlags, + glslangIntermediate->getBaseAlignmentScalar( + node->getLeft()->getType(), dummySize)); } else { // Load through a block reference is performed with a dot operator that @@ -1868,7 +2106,7 @@ bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::T { // This may be, e.g., an anonymous block-member selection, which generally need // index remapping due to hidden members in anonymous blocks. - int glslangId = glslangTypeToIdMap[node->getLeft()->getType().getStruct()]; + long long glslangId = glslangTypeToIdMap[node->getLeft()->getType().getStruct()]; if (memberRemapper.find(glslangId) != memberRemapper.end()) { std::vector<int>& remapper = memberRemapper[glslangId]; assert(remapper.size() > 0); @@ -1876,8 +2114,15 @@ bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::T } } + // Struct reference propagates uniform lvalue + spv::Builder::AccessChain::CoherentFlags coherentFlags = + TranslateCoherent(node->getLeft()->getType()); + coherentFlags.nonUniform = 0; + // normal case for indexing array or structure or block - builder.accessChainPush(builder.makeIntConstant(spvIndex), TranslateCoherent(node->getLeft()->getType()), node->getLeft()->getType().getBufferReferenceAlignment()); + builder.accessChainPush(builder.makeIntConstant(spvIndex), + coherentFlags, + node->getLeft()->getType().getBufferReferenceAlignment()); // Add capabilities here for accessing PointSize and clip/cull distance. // We have deferred generation of associated capabilities until now. @@ -1910,13 +2155,20 @@ bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::T // restore the saved access chain builder.setAccessChain(partial); + // Only if index is nonUniform should we propagate nonUniform into access chain + spv::Builder::AccessChain::CoherentFlags index_flags = TranslateCoherent(node->getRight()->getType()); + spv::Builder::AccessChain::CoherentFlags coherent_flags = TranslateCoherent(node->getLeft()->getType()); + coherent_flags.nonUniform = index_flags.nonUniform; + if (! node->getLeft()->getType().isArray() && node->getLeft()->getType().isVector()) { int dummySize; - builder.accessChainPushComponent(index, convertGlslangToSpvType(node->getLeft()->getType()), - TranslateCoherent(node->getLeft()->getType()), - glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize)); + builder.accessChainPushComponent( + index, convertGlslangToSpvType(node->getLeft()->getType()), coherent_flags, + glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), + dummySize)); } else - builder.accessChainPush(index, TranslateCoherent(node->getLeft()->getType()), node->getLeft()->getType().getBufferReferenceAlignment()); + builder.accessChainPush(index, coherent_flags, + node->getLeft()->getType().getBufferReferenceAlignment()); } return false; case glslang::EOpVectorSwizzle: @@ -1927,7 +2179,8 @@ bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::T int dummySize; builder.accessChainPushSwizzle(swizzle, convertGlslangToSpvType(node->getLeft()->getType()), TranslateCoherent(node->getLeft()->getType()), - glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize)); + glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), + dummySize)); } return false; case glslang::EOpMatrixSwizzle: @@ -1943,7 +2196,8 @@ bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::T if (isTrivial(node->getRight()->getAsTyped())) break; // handle below as a normal binary operation // otherwise, we need to do dynamic short circuiting on the right operand - spv::Id result = createShortCircuit(node->getOp(), *node->getLeft()->getAsTyped(), *node->getRight()->getAsTyped()); + spv::Id result = createShortCircuit(node->getOp(), *node->getLeft()->getAsTyped(), + *node->getRight()->getAsTyped()); builder.clearAccessChain(); builder.setAccessChainRValue(result); } @@ -1982,24 +2236,77 @@ bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::T } } +spv::Id TGlslangToSpvTraverser::convertLoadedBoolInUniformToUint(const glslang::TType& type, + spv::Id nominalTypeId, + spv::Id loadedId) +{ + if (builder.isScalarType(nominalTypeId)) { + // Conversion for bool + spv::Id boolType = builder.makeBoolType(); + if (nominalTypeId != boolType) + return builder.createBinOp(spv::OpINotEqual, boolType, loadedId, builder.makeUintConstant(0)); + } else if (builder.isVectorType(nominalTypeId)) { + // Conversion for bvec + int vecSize = builder.getNumTypeComponents(nominalTypeId); + spv::Id bvecType = builder.makeVectorType(builder.makeBoolType(), vecSize); + if (nominalTypeId != bvecType) + loadedId = builder.createBinOp(spv::OpINotEqual, bvecType, loadedId, + makeSmearedConstant(builder.makeUintConstant(0), vecSize)); + } else if (builder.isArrayType(nominalTypeId)) { + // Conversion for bool array + spv::Id boolArrayTypeId = convertGlslangToSpvType(type); + if (nominalTypeId != boolArrayTypeId) + { + // Use OpCopyLogical from SPIR-V 1.4 if available. + if (glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4) + return builder.createUnaryOp(spv::OpCopyLogical, boolArrayTypeId, loadedId); + + glslang::TType glslangElementType(type, 0); + spv::Id elementNominalTypeId = builder.getContainedTypeId(nominalTypeId); + std::vector<spv::Id> constituents; + for (int index = 0; index < type.getOuterArraySize(); ++index) { + // get the element + spv::Id elementValue = builder.createCompositeExtract(loadedId, elementNominalTypeId, index); + + // recursively convert it + spv::Id elementConvertedValue = convertLoadedBoolInUniformToUint(glslangElementType, elementNominalTypeId, elementValue); + constituents.push_back(elementConvertedValue); + } + return builder.createCompositeConstruct(boolArrayTypeId, constituents); + } + } + + return loadedId; +} + // Figure out what, if any, type changes are needed when accessing a specific built-in. // Returns <the type SPIR-V requires for declarion, the type to translate to on use>. // Also see comment for 'forceType', regarding tracking SPIR-V-required types. -std::pair<spv::Id, spv::Id> TGlslangToSpvTraverser::getForcedType(spv::BuiltIn builtIn, +std::pair<spv::Id, spv::Id> TGlslangToSpvTraverser::getForcedType(glslang::TBuiltInVariable glslangBuiltIn, const glslang::TType& glslangType) { - switch(builtIn) + switch(glslangBuiltIn) { - case spv::BuiltInSubgroupEqMask: - case spv::BuiltInSubgroupGeMask: - case spv::BuiltInSubgroupGtMask: - case spv::BuiltInSubgroupLeMask: - case spv::BuiltInSubgroupLtMask: { + case glslang::EbvSubGroupEqMask: + case glslang::EbvSubGroupGeMask: + case glslang::EbvSubGroupGtMask: + case glslang::EbvSubGroupLeMask: + case glslang::EbvSubGroupLtMask: { // these require changing a 64-bit scaler -> a vector of 32-bit components if (glslangType.isVector()) break; - std::pair<spv::Id, spv::Id> ret(builder.makeVectorType(builder.makeUintType(32), 4), - builder.makeUintType(64)); + spv::Id ivec4_type = builder.makeVectorType(builder.makeUintType(32), 4); + spv::Id uint64_type = builder.makeUintType(64); + std::pair<spv::Id, spv::Id> ret(ivec4_type, uint64_type); + return ret; + } + // There are no SPIR-V builtins defined for these and map onto original non-transposed + // builtins. During visitBinary we insert a transpose + case glslang::EbvWorldToObject3x4: + case glslang::EbvObjectToWorld3x4: { + spv::Id mat43 = builder.makeMatrixType(builder.makeFloatType(32), 4, 3); + spv::Id mat34 = builder.makeMatrixType(builder.makeFloatType(32), 3, 4); + std::pair<spv::Id, spv::Id> ret(mat43, mat34); return ret; } default: @@ -2029,7 +2336,7 @@ spv::Id TGlslangToSpvTraverser::translateForcedType(spv::Id object) // handle 32-bit v.xy* -> 64-bit builder.clearAccessChain(); builder.setAccessChainLValue(object); - object = builder.accessChainLoad(spv::NoPrecision, spv::DecorationMax, objectTypeId); + object = builder.accessChainLoad(spv::NoPrecision, spv::DecorationMax, spv::DecorationMax, objectTypeId); std::vector<spv::Id> components; components.push_back(builder.createCompositeExtract(object, builder.getContainedTypeId(objectTypeId), 0)); components.push_back(builder.createCompositeExtract(object, builder.getContainedTypeId(objectTypeId), 1)); @@ -2040,7 +2347,15 @@ spv::Id TGlslangToSpvTraverser::translateForcedType(spv::Id object) } else { logger->missingFunctionality("forcing 32-bit vector type to non 64-bit scalar"); } - } else { + } else if (builder.isMatrixType(objectTypeId)) { + // There are no SPIR-V builtins defined for 3x4 variants of ObjectToWorld/WorldToObject + // and we insert a transpose after loading the original non-transposed builtins + builder.clearAccessChain(); + builder.setAccessChainLValue(object); + object = builder.accessChainLoad(spv::NoPrecision, spv::DecorationMax, spv::DecorationMax, objectTypeId); + return builder.createUnaryOp(spv::OpTranspose, desiredTypeId, object); + + } else { logger->missingFunctionality("forcing non 32-bit vector type"); } @@ -2091,7 +2406,8 @@ bool TGlslangToSpvTraverser::visitUnary(glslang::TVisit /* visit */, glslang::TI } else { glslang::TIntermTyped* block = node->getOperand()->getAsBinaryNode()->getLeft(); block->traverse(this); - unsigned int member = node->getOperand()->getAsBinaryNode()->getRight()->getAsConstantUnion()->getConstArray()[0].getUConst(); + unsigned int member = node->getOperand()->getAsBinaryNode()->getRight()->getAsConstantUnion() + ->getConstArray()[0].getUConst(); length = builder.createArrayLength(builder.accessChainGetLValue(), member); } @@ -2117,7 +2433,8 @@ bool TGlslangToSpvTraverser::visitUnary(glslang::TVisit /* visit */, glslang::TI // Does it need a swizzle inversion? If so, evaluation is inverted; // operate first on the swizzle base, then apply the swizzle. spv::Id invertedType = spv::NoType; - auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ? invertedType : convertGlslangToSpvType(node->getType()); }; + auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ? + invertedType : convertGlslangToSpvType(node->getType()); }; if (node->getOp() == glslang::EOpInterpolateAtCentroid) invertedType = getInvertedSwizzleType(*node->getOperand()); @@ -2138,10 +2455,23 @@ bool TGlslangToSpvTraverser::visitUnary(glslang::TVisit /* visit */, glslang::TI if (node->getOp() == glslang::EOpAtomicCounterIncrement || node->getOp() == glslang::EOpAtomicCounterDecrement || node->getOp() == glslang::EOpAtomicCounter || - node->getOp() == glslang::EOpInterpolateAtCentroid) { + (node->getOp() == glslang::EOpInterpolateAtCentroid && + glslangIntermediate->getSource() != glslang::EShSourceHlsl) || + node->getOp() == glslang::EOpRayQueryProceed || + node->getOp() == glslang::EOpRayQueryGetRayTMin || + node->getOp() == glslang::EOpRayQueryGetRayFlags || + node->getOp() == glslang::EOpRayQueryGetWorldRayOrigin || + node->getOp() == glslang::EOpRayQueryGetWorldRayDirection || + node->getOp() == glslang::EOpRayQueryGetIntersectionCandidateAABBOpaque || + node->getOp() == glslang::EOpRayQueryTerminate || + node->getOp() == glslang::EOpRayQueryConfirmIntersection || + (node->getOp() == glslang::EOpSpirvInst && operandNode->getAsTyped()->getQualifier().isSpirvByReference())) { operand = builder.accessChainGetLValue(); // Special case l-value operands lvalueCoherentFlags = builder.getAccessChain().coherentFlags; lvalueCoherentFlags |= TranslateCoherent(operandNode->getAsTyped()->getType()); + } else if (operandNode->getAsTyped()->getQualifier().isSpirvLiteral()) { + // Will be translated to a literal value, make a placeholder here + operand = spv::NoResult; } else #endif { @@ -2154,11 +2484,45 @@ bool TGlslangToSpvTraverser::visitUnary(glslang::TVisit /* visit */, glslang::TI // it could be a conversion if (! result) - result = createConversion(node->getOp(), decorations, resultType(), operand, node->getOperand()->getBasicType()); + result = createConversion(node->getOp(), decorations, resultType(), operand, + node->getOperand()->getBasicType()); // if not, then possibly an operation if (! result) - result = createUnaryOperation(node->getOp(), decorations, resultType(), operand, node->getOperand()->getBasicType(), lvalueCoherentFlags); + result = createUnaryOperation(node->getOp(), decorations, resultType(), operand, + node->getOperand()->getBasicType(), lvalueCoherentFlags); + +#ifndef GLSLANG_WEB + // it could be attached to a SPIR-V intruction + if (!result) { + if (node->getOp() == glslang::EOpSpirvInst) { + const auto& spirvInst = node->getSpirvInstruction(); + if (spirvInst.set == "") { + spv::IdImmediate idImmOp = {true, operand}; + if (operandNode->getAsTyped()->getQualifier().isSpirvLiteral()) { + // Translate the constant to a literal value + std::vector<unsigned> literals; + glslang::TVector<const glslang::TIntermConstantUnion*> constants; + constants.push_back(operandNode->getAsConstantUnion()); + TranslateLiterals(constants, literals); + idImmOp = {false, literals[0]}; + } + + if (node->getBasicType() == glslang::EbtVoid) + builder.createNoResultOp(static_cast<spv::Op>(spirvInst.id), {idImmOp}); + else + result = builder.createOp(static_cast<spv::Op>(spirvInst.id), resultType(), {idImmOp}); + } else { + result = builder.createBuiltinCall( + resultType(), spirvInst.set == "GLSL.std.450" ? stdBuiltins : getExtBuiltins(spirvInst.set.c_str()), + spirvInst.id, {operand}); + } + + if (node->getBasicType() == glslang::EbtVoid) + return false; // done with this node + } + } +#endif if (result) { if (invertedType) { @@ -2211,7 +2575,8 @@ bool TGlslangToSpvTraverser::visitUnary(glslang::TVisit /* visit */, glslang::TI // The result of operation is always stored, but conditionally the // consumed result. The consumed result is always an r-value. - builder.accessChainStore(result); + builder.accessChainStore(result, + TranslateNonUniformDecoration(builder.getAccessChain().coherentFlags)); builder.clearAccessChain(); if (node->getOp() == glslang::EOpPreIncrement || node->getOp() == glslang::EOpPreDecrement) @@ -2229,6 +2594,12 @@ bool TGlslangToSpvTraverser::visitUnary(glslang::TVisit /* visit */, glslang::TI case glslang::EOpEndStreamPrimitive: builder.createNoResultOp(spv::OpEndStreamPrimitive, operand); return false; + case glslang::EOpRayQueryTerminate: + builder.createNoResultOp(spv::OpRayQueryTerminateKHR, operand); + return false; + case glslang::EOpRayQueryConfirmIntersection: + builder.createNoResultOp(spv::OpRayQueryConfirmIntersectionKHR, operand); + return false; #endif default: @@ -2251,7 +2622,8 @@ spv::Id TGlslangToSpvTraverser::createCompositeConstruct(spv::Id resultTypeId, s std::vector<spv::Id> rTypeConstituents; int numrTypeConstituents = builder.getNumTypeConstituents(rType); for (int i = 0; i < numrTypeConstituents; ++i) { - rTypeConstituents.push_back(builder.createCompositeExtract(constituent, builder.getContainedTypeId(rType, i), i)); + rTypeConstituents.push_back(builder.createCompositeExtract(constituent, + builder.getContainedTypeId(rType, i), i)); } constituents[c] = createCompositeConstruct(lType, rTypeConstituents); } else { @@ -2277,8 +2649,14 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); spv::Id result = spv::NoResult; - spv::Id invertedType = spv::NoType; // to use to override the natural type of the node - auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ? invertedType : convertGlslangToSpvType(node->getType()); }; + spv::Id invertedType = spv::NoType; // to use to override the natural type of the node + std::vector<spv::Builder::AccessChain> complexLvalues; // for holding swizzling l-values too complex for + // SPIR-V, for an out parameter + std::vector<spv::Id> temporaryLvalues; // temporaries to pass, as proxies for complexLValues + + auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ? + invertedType : + convertGlslangToSpvType(node->getType()); }; // try texturing result = createImageTextureFunctionCall(node); @@ -2327,6 +2705,10 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt // anything else gets there, so visit out of order, doing them all now. makeGlobalInitializers(node->getAsAggregate()->getSequence()); + //Pre process linker objects for ray tracing stages + if (glslangIntermediate->isRayTracingStage()) + collectRayTracingLinkerObjects(); + // Initializers are done, don't want to visit again, but functions and link objects need to be processed, // so do them manually. visitFunctions(node->getAsAggregate()->getSequence()); @@ -2382,7 +2764,6 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt builder.setLine(node->getLoc().line, node->getLoc().getFilename()); if (node->isUserDefined()) result = handleUserFunctionCall(node); - // assert(result); // this can happen for bad shaders because the call graph completeness checking is not yet done if (result) { builder.clearAccessChain(); builder.setAccessChainRValue(result); @@ -2518,6 +2899,10 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt else constructed = builder.createConstructor(precision, arguments, resultType()); + if (node->getType().getQualifier().isNonUniform()) { + builder.addDecoration(constructed, spv::DecorationNonUniformEXT); + } + builder.clearAccessChain(); builder.setAccessChainRValue(constructed); @@ -2590,6 +2975,7 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt break; case glslang::EOpAtomicAdd: + case glslang::EOpAtomicSubtract: case glslang::EOpAtomicMin: case glslang::EOpAtomicMax: case glslang::EOpAtomicAnd: @@ -2636,10 +3022,43 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt case glslang::EOpIgnoreIntersectionNV: case glslang::EOpTerminateRayNV: case glslang::EOpTraceNV: + case glslang::EOpTraceRayMotionNV: + case glslang::EOpTraceKHR: case glslang::EOpExecuteCallableNV: + case glslang::EOpExecuteCallableKHR: case glslang::EOpWritePackedPrimitiveIndices4x8NV: noReturnValue = true; break; + case glslang::EOpRayQueryInitialize: + case glslang::EOpRayQueryTerminate: + case glslang::EOpRayQueryGenerateIntersection: + case glslang::EOpRayQueryConfirmIntersection: + builder.addExtension("SPV_KHR_ray_query"); + builder.addCapability(spv::CapabilityRayQueryKHR); + noReturnValue = true; + break; + case glslang::EOpRayQueryProceed: + case glslang::EOpRayQueryGetIntersectionType: + case glslang::EOpRayQueryGetRayTMin: + case glslang::EOpRayQueryGetRayFlags: + case glslang::EOpRayQueryGetIntersectionT: + case glslang::EOpRayQueryGetIntersectionInstanceCustomIndex: + case glslang::EOpRayQueryGetIntersectionInstanceId: + case glslang::EOpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffset: + case glslang::EOpRayQueryGetIntersectionGeometryIndex: + case glslang::EOpRayQueryGetIntersectionPrimitiveIndex: + case glslang::EOpRayQueryGetIntersectionBarycentrics: + case glslang::EOpRayQueryGetIntersectionFrontFace: + case glslang::EOpRayQueryGetIntersectionCandidateAABBOpaque: + case glslang::EOpRayQueryGetIntersectionObjectRayDirection: + case glslang::EOpRayQueryGetIntersectionObjectRayOrigin: + case glslang::EOpRayQueryGetWorldRayDirection: + case glslang::EOpRayQueryGetWorldRayOrigin: + case glslang::EOpRayQueryGetIntersectionObjectToWorld: + case glslang::EOpRayQueryGetIntersectionWorldToObject: + builder.addExtension("SPV_KHR_ray_query"); + builder.addCapability(spv::CapabilityRayQueryKHR); + break; case glslang::EOpCooperativeMatrixLoad: case glslang::EOpCooperativeMatrixStore: noReturnValue = true; @@ -2651,6 +3070,10 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt break; #endif + case glslang::EOpDebugPrintf: + noReturnValue = true; + break; + default: break; } @@ -2702,7 +3125,30 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt lvalue = true; break; + case glslang::EOpRayQueryInitialize: + case glslang::EOpRayQueryTerminate: + case glslang::EOpRayQueryConfirmIntersection: + case glslang::EOpRayQueryProceed: + case glslang::EOpRayQueryGenerateIntersection: + case glslang::EOpRayQueryGetIntersectionType: + case glslang::EOpRayQueryGetIntersectionT: + case glslang::EOpRayQueryGetIntersectionInstanceCustomIndex: + case glslang::EOpRayQueryGetIntersectionInstanceId: + case glslang::EOpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffset: + case glslang::EOpRayQueryGetIntersectionGeometryIndex: + case glslang::EOpRayQueryGetIntersectionPrimitiveIndex: + case glslang::EOpRayQueryGetIntersectionBarycentrics: + case glslang::EOpRayQueryGetIntersectionFrontFace: + case glslang::EOpRayQueryGetIntersectionObjectRayDirection: + case glslang::EOpRayQueryGetIntersectionObjectRayOrigin: + case glslang::EOpRayQueryGetIntersectionObjectToWorld: + case glslang::EOpRayQueryGetIntersectionWorldToObject: + if (arg == 0) + lvalue = true; + break; + case glslang::EOpAtomicAdd: + case glslang::EOpAtomicSubtract: case glslang::EOpAtomicMin: case glslang::EOpAtomicMax: case glslang::EOpAtomicAnd: @@ -2723,13 +3169,24 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt case glslang::EOpInterpolateAtOffset: case glslang::EOpInterpolateAtVertex: if (arg == 0) { - lvalue = true; + // If GLSL, use the address of the interpolant argument. + // If HLSL, use an internal version of OpInterolates that takes + // the rvalue of the interpolant. A fixup pass in spirv-opt + // legalization will remove the OpLoad and convert to an lvalue. + // Had to do this because legalization will only propagate a + // builtin into an rvalue. + lvalue = glslangIntermediate->getSource() != glslang::EShSourceHlsl; // Does it need a swizzle inversion? If so, evaluation is inverted; // operate first on the swizzle base, then apply the swizzle. + // That is, we transform + // + // interpolate(v.zy) -> interpolate(v).zy + // if (glslangOperands[0]->getAsOperator() && glslangOperands[0]->getAsOperator()->getOp() == glslang::EOpVectorSwizzle) - invertedType = convertGlslangToSpvType(glslangOperands[0]->getAsBinaryNode()->getLeft()->getType()); + invertedType = convertGlslangToSpvType( + glslangOperands[0]->getAsBinaryNode()->getLeft()->getType()); } break; case glslang::EOpAtomicLoad: @@ -2764,6 +3221,10 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt if (arg == 1) lvalue = true; break; + case glslang::EOpSpirvInst: + if (glslangOperands[arg]->getAsTyped()->getQualifier().isSpirvByReference()) + lvalue = true; + break; #endif default: break; @@ -2789,8 +3250,9 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt builder.setAccessChain(save); // Point to the first element of the array. - builder.accessChainPush(elementId, TranslateCoherent(glslangOperands[arg]->getAsTyped()->getType()), - glslangOperands[arg]->getAsTyped()->getType().getBufferReferenceAlignment()); + builder.accessChainPush(elementId, + TranslateCoherent(glslangOperands[arg]->getAsTyped()->getType()), + glslangOperands[arg]->getAsTyped()->getType().getBufferReferenceAlignment()); spv::Builder::AccessChain::CoherentFlags coherentFlags = builder.getAccessChain().coherentFlags; unsigned int alignment = builder.getAccessChain().alignment; @@ -2800,7 +3262,8 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt memoryAccess &= ~spv::MemoryAccessMakePointerAvailableKHRMask; if (node->getOp() == glslang::EOpCooperativeMatrixStore) memoryAccess &= ~spv::MemoryAccessMakePointerVisibleKHRMask; - if (builder.getStorageClass(builder.getAccessChain().base) == spv::StorageClassPhysicalStorageBufferEXT) { + if (builder.getStorageClass(builder.getAccessChain().base) == + spv::StorageClassPhysicalStorageBufferEXT) { memoryAccess = (spv::MemoryAccessMask)(memoryAccess | spv::MemoryAccessAlignedMask); } @@ -2810,8 +3273,10 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt memoryAccessOperands.push_back(spv::IdImmediate(false, alignment)); } - if (memoryAccess & (spv::MemoryAccessMakePointerAvailableKHRMask | spv::MemoryAccessMakePointerVisibleKHRMask)) { - memoryAccessOperands.push_back(spv::IdImmediate(true, builder.makeUintConstant(TranslateMemoryScope(coherentFlags)))); + if (memoryAccess & + (spv::MemoryAccessMakePointerAvailableKHRMask | spv::MemoryAccessMakePointerVisibleKHRMask)) { + memoryAccessOperands.push_back(spv::IdImmediate(true, + builder.makeUintConstant(TranslateMemoryScope(coherentFlags)))); } } else if (arg == 2) { continue; @@ -2819,13 +3284,62 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt } #endif + // for l-values, pass the address, for r-values, pass the value if (lvalue) { - operands.push_back(builder.accessChainGetLValue()); + if (invertedType == spv::NoType && !builder.isSpvLvalue()) { + // SPIR-V cannot represent an l-value containing a swizzle that doesn't + // reduce to a simple access chain. So, we need a temporary vector to + // receive the result, and must later swizzle that into the original + // l-value. + complexLvalues.push_back(builder.getAccessChain()); + temporaryLvalues.push_back(builder.createVariable( + spv::NoPrecision, spv::StorageClassFunction, + builder.accessChainGetInferredType(), "swizzleTemp")); + operands.push_back(temporaryLvalues.back()); + } else { + operands.push_back(builder.accessChainGetLValue()); + } lvalueCoherentFlags = builder.getAccessChain().coherentFlags; lvalueCoherentFlags |= TranslateCoherent(glslangOperands[arg]->getAsTyped()->getType()); } else { builder.setLine(node->getLoc().line, node->getLoc().getFilename()); - operands.push_back(accessChainLoad(glslangOperands[arg]->getAsTyped()->getType())); + glslang::TOperator glslangOp = node->getOp(); + if (arg == 1 && + (glslangOp == glslang::EOpRayQueryGetIntersectionType || + glslangOp == glslang::EOpRayQueryGetIntersectionT || + glslangOp == glslang::EOpRayQueryGetIntersectionInstanceCustomIndex || + glslangOp == glslang::EOpRayQueryGetIntersectionInstanceId || + glslangOp == glslang::EOpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffset || + glslangOp == glslang::EOpRayQueryGetIntersectionGeometryIndex || + glslangOp == glslang::EOpRayQueryGetIntersectionPrimitiveIndex || + glslangOp == glslang::EOpRayQueryGetIntersectionBarycentrics || + glslangOp == glslang::EOpRayQueryGetIntersectionFrontFace || + glslangOp == glslang::EOpRayQueryGetIntersectionObjectRayDirection || + glslangOp == glslang::EOpRayQueryGetIntersectionObjectRayOrigin || + glslangOp == glslang::EOpRayQueryGetIntersectionObjectToWorld || + glslangOp == glslang::EOpRayQueryGetIntersectionWorldToObject + )) { + bool cond = glslangOperands[arg]->getAsConstantUnion()->getConstArray()[0].getBConst(); + operands.push_back(builder.makeIntConstant(cond ? 1 : 0)); + } else if ((arg == 10 && glslangOp == glslang::EOpTraceKHR) || + (arg == 11 && glslangOp == glslang::EOpTraceRayMotionNV) || + (arg == 1 && glslangOp == glslang::EOpExecuteCallableKHR)) { + const int opdNum = glslangOp == glslang::EOpTraceKHR ? 10 : (glslangOp == glslang::EOpTraceRayMotionNV ? 11 : 1); + const int set = glslangOp == glslang::EOpExecuteCallableKHR ? 1 : 0; + + const int location = glslangOperands[opdNum]->getAsConstantUnion()->getConstArray()[0].getUConst(); + auto itNode = locationToSymbol[set].find(location); + visitSymbol(itNode->second); + spv::Id symId = getSymbolId(itNode->second); + operands.push_back(symId); +#ifndef GLSLANG_WEB + } else if (glslangOperands[arg]->getAsTyped()->getQualifier().isSpirvLiteral()) { + // Will be translated to a literal value, make a placeholder here + operands.push_back(spv::NoResult); +#endif + } else { + operands.push_back(accessChainLoad(glslangOperands[arg]->getAsTyped()->getType())); + } } } @@ -2861,7 +3375,44 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt #endif if (atomic) { // Handle all atomics - result = createAtomicOperation(node->getOp(), precision, resultType(), operands, node->getBasicType(), lvalueCoherentFlags); + glslang::TBasicType typeProxy = (node->getOp() == glslang::EOpAtomicStore) + ? node->getSequence()[0]->getAsTyped()->getBasicType() : node->getBasicType(); + result = createAtomicOperation(node->getOp(), precision, resultType(), operands, typeProxy, + lvalueCoherentFlags); +#ifndef GLSLANG_WEB + } else if (node->getOp() == glslang::EOpSpirvInst) { + const auto& spirvInst = node->getSpirvInstruction(); + if (spirvInst.set == "") { + std::vector<spv::IdImmediate> idImmOps; + for (int i = 0; i < glslangOperands.size(); ++i) { + if (glslangOperands[i]->getAsTyped()->getQualifier().isSpirvLiteral()) { + // Translate the constant to a literal value + std::vector<unsigned> literals; + glslang::TVector<const glslang::TIntermConstantUnion*> constants; + constants.push_back(glslangOperands[i]->getAsConstantUnion()); + TranslateLiterals(constants, literals); + idImmOps.push_back({false, literals[0]}); + } else + idImmOps.push_back({true, operands[i]}); + } + + if (node->getBasicType() == glslang::EbtVoid) + builder.createNoResultOp(static_cast<spv::Op>(spirvInst.id), idImmOps); + else + result = builder.createOp(static_cast<spv::Op>(spirvInst.id), resultType(), idImmOps); + } else { + result = builder.createBuiltinCall( + resultType(), spirvInst.set == "GLSL.std.450" ? stdBuiltins : getExtBuiltins(spirvInst.set.c_str()), + spirvInst.id, operands); + } + noReturnValue = node->getBasicType() == glslang::EbtVoid; +#endif + } else if (node->getOp() == glslang::EOpDebugPrintf) { + if (!nonSemanticDebugPrintf) { + nonSemanticDebugPrintf = builder.import("NonSemantic.DebugPrintf"); + } + result = builder.createBuiltinCall(builder.makeVoidType(), nonSemanticDebugPrintf, spv::NonSemanticDebugPrintfDebugPrintf, operands); + builder.addExtension(spv::E_SPV_KHR_non_semantic_info); } else { // Pass through to generic operations. switch (glslangOperands.size()) { @@ -2883,8 +3434,15 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt result = createMiscOperation(node->getOp(), precision, resultType(), operands, node->getBasicType()); break; } - if (invertedType) + + if (invertedType != spv::NoResult) result = createInvertedSwizzle(precision, *glslangOperands[0]->getAsBinaryNode(), result); + + for (unsigned int i = 0; i < temporaryLvalues.size(); ++i) { + builder.setAccessChain(complexLvalues[i]); + builder.accessChainStore(builder.createLoad(temporaryLvalues[i], spv::NoPrecision), + TranslateNonUniformDecoration(complexLvalues[i].coherentFlags)); + } } if (noReturnValue) @@ -3000,7 +3558,8 @@ bool TGlslangToSpvTraverser::visitSelection(glslang::TVisit /* visit */, glslang } else { // We need control flow to select the result. // TODO: Once SPIR-V OpSelect allows arbitrary types, eliminate this path. - result = builder.createVariable(spv::StorageClassFunction, convertGlslangToSpvType(node->getType())); + result = builder.createVariable(TranslatePrecisionDecoration(node->getType()), + spv::StorageClassFunction, convertGlslangToSpvType(node->getType())); // Selection control: const spv::SelectionControlMask control = TranslateSelectionControl(*node); @@ -3025,8 +3584,10 @@ bool TGlslangToSpvTraverser::visitSelection(glslang::TVisit /* visit */, glslang // Execute the one side needed, as per the condition const auto executeOneSide = [&]() { // Always emit control flow. - if (node->getBasicType() != glslang::EbtVoid) - result = builder.createVariable(spv::StorageClassFunction, convertGlslangToSpvType(node->getType())); + if (node->getBasicType() != glslang::EbtVoid) { + result = builder.createVariable(TranslatePrecisionDecoration(node->getType()), spv::StorageClassFunction, + convertGlslangToSpvType(node->getType())); + } // Selection control: const spv::SelectionControlMask control = TranslateSelectionControl(*node); @@ -3091,7 +3652,8 @@ bool TGlslangToSpvTraverser::visitSwitch(glslang::TVisit /* visit */, glslang::T defaultSegment = (int)codeSegments.size(); else if (child->getAsBranchNode() && child->getAsBranchNode()->getFlowOp() == glslang::EOpCase) { valueIndexToSegment[caseValues.size()] = (int)codeSegments.size(); - caseValues.push_back(child->getAsBranchNode()->getExpression()->getAsConstantUnion()->getConstArray()[0].getIConst()); + caseValues.push_back(child->getAsBranchNode()->getExpression()->getAsConstantUnion() + ->getConstArray()[0].getIConst()); } else codeSegments.push_back(child); } @@ -3104,7 +3666,8 @@ bool TGlslangToSpvTraverser::visitSwitch(glslang::TVisit /* visit */, glslang::T // make the switch statement std::vector<spv::Block*> segmentBlocks; // returned, as the blocks allocated in the call - builder.makeSwitch(selector, control, (int)codeSegments.size(), caseValues, valueIndexToSegment, defaultSegment, segmentBlocks); + builder.makeSwitch(selector, control, (int)codeSegments.size(), caseValues, valueIndexToSegment, defaultSegment, + segmentBlocks); // emit all the code in the segments breakForLoop.push(false); @@ -3124,6 +3687,11 @@ bool TGlslangToSpvTraverser::visitSwitch(glslang::TVisit /* visit */, glslang::T void TGlslangToSpvTraverser::visitConstantUnion(glslang::TIntermConstantUnion* node) { +#ifndef GLSLANG_WEB + if (node->getQualifier().isSpirvLiteral()) + return; // Translated to a literal value, skip further processing +#endif + int nextConst = 0; spv::Id constant = createSpvConstantFromConstUnionArray(node->getType(), node->getConstArray(), nextConst, false); @@ -3209,7 +3777,11 @@ bool TGlslangToSpvTraverser::visitBranch(glslang::TVisit /* visit */, glslang::T switch (node->getFlowOp()) { case glslang::EOpKill: - builder.makeDiscard(); + builder.makeStatementTerminator(spv::OpKill, "post-discard"); + break; + case glslang::EOpTerminateInvocation: + builder.addExtension(spv::E_SPV_KHR_terminate_invocation); + builder.makeStatementTerminator(spv::OpTerminateInvocation, "post-terminate-invocation"); break; case glslang::EOpBreak: if (breakForLoop.top()) @@ -3221,15 +3793,17 @@ bool TGlslangToSpvTraverser::visitBranch(glslang::TVisit /* visit */, glslang::T builder.createLoopContinue(); break; case glslang::EOpReturn: - if (node->getExpression()) { + if (node->getExpression() != nullptr) { const glslang::TType& glslangReturnType = node->getExpression()->getType(); spv::Id returnId = accessChainLoad(glslangReturnType); - if (builder.getTypeId(returnId) != currentFunction->getReturnType()) { + if (builder.getTypeId(returnId) != currentFunction->getReturnType() || + TranslatePrecisionDecoration(glslangReturnType) != currentFunction->getReturnPrecision()) { builder.clearAccessChain(); - spv::Id copyId = builder.createVariable(spv::StorageClassFunction, currentFunction->getReturnType()); + spv::Id copyId = builder.createVariable(currentFunction->getReturnPrecision(), + spv::StorageClassFunction, currentFunction->getReturnType()); builder.setAccessChainLValue(copyId); multiTypeStore(glslangReturnType, returnId); - returnId = builder.createLoad(copyId); + returnId = builder.createLoad(copyId, currentFunction->getReturnPrecision()); } builder.makeReturn(false, returnId); } else @@ -3244,6 +3818,12 @@ bool TGlslangToSpvTraverser::visitBranch(glslang::TVisit /* visit */, glslang::T builder.addExtension(spv::E_SPV_EXT_demote_to_helper_invocation); builder.addCapability(spv::CapabilityDemoteToHelperInvocationEXT); break; + case glslang::EOpTerminateRayKHR: + builder.makeStatementTerminator(spv::OpTerminateRayKHR, "post-terminateRayKHR"); + break; + case glslang::EOpIgnoreIntersectionKHR: + builder.makeStatementTerminator(spv::OpIgnoreIntersectionKHR, "post-ignoreIntersectionKHR"); + break; #endif default: @@ -3298,6 +3878,11 @@ spv::Id TGlslangToSpvTraverser::createSpvVariable(const glslang::TIntermSymbol* break; #endif default: + if (storageClass == spv::StorageClassWorkgroup && + node->getType().getBasicType() == glslang::EbtBlock) { + builder.addCapability(spv::CapabilityWorkgroupMemoryExplicitLayout16BitAccessKHR); + break; + } if (node->getType().contains16BitFloat()) builder.addCapability(spv::CapabilityFloat16); if (node->getType().contains16BitInt()) @@ -3316,6 +3901,9 @@ spv::Id TGlslangToSpvTraverser::createSpvVariable(const glslang::TIntermSymbol* } else if (storageClass == spv::StorageClassStorageBuffer) { builder.addIncorporatedExtension(spv::E_SPV_KHR_8bit_storage, spv::Spv_1_5); builder.addCapability(spv::CapabilityStorageBuffer8BitAccess); + } else if (storageClass == spv::StorageClassWorkgroup && + node->getType().getBasicType() == glslang::EbtBlock) { + builder.addCapability(spv::CapabilityWorkgroupMemoryExplicitLayout8BitAccessKHR); } else { builder.addCapability(spv::CapabilityInt8); } @@ -3325,7 +3913,19 @@ spv::Id TGlslangToSpvTraverser::createSpvVariable(const glslang::TIntermSymbol* if (glslang::IsAnonymous(name)) name = ""; - return builder.createVariable(storageClass, spvType, name); + spv::Id initializer = spv::NoResult; + + if (node->getType().getQualifier().storage == glslang::EvqUniform && !node->getConstArray().empty()) { + int nextConst = 0; + initializer = createSpvConstantFromConstUnionArray(node->getType(), + node->getConstArray(), + nextConst, + false /* specConst */); + } else if (node->getType().getQualifier().isNullInit()) { + initializer = builder.makeNullConstant(spvType); + } + + return builder.createVariable(spv::NoPrecision, storageClass, spvType, name, initializer); } // Return type Id of the sampled type. @@ -3340,6 +3940,12 @@ spv::Id TGlslangToSpvTraverser::getSampledType(const glslang::TSampler& sampler) builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float_fetch); builder.addCapability(spv::CapabilityFloat16ImageAMD); return builder.makeFloatType(16); + case glslang::EbtInt64: return builder.makeIntType(64); + builder.addExtension(spv::E_SPV_EXT_shader_image_int64); + builder.addCapability(spv::CapabilityFloat16ImageAMD); + case glslang::EbtUint64: return builder.makeUintType(64); + builder.addExtension(spv::E_SPV_EXT_shader_image_int64); + builder.addCapability(spv::CapabilityFloat16ImageAMD); #endif default: assert(0); @@ -3361,7 +3967,8 @@ spv::Id TGlslangToSpvTraverser::getInvertedSwizzleType(const glslang::TIntermTyp // When inverting a swizzle with a parent op, this function // will apply the swizzle operation to a completed parent operation. -spv::Id TGlslangToSpvTraverser::createInvertedSwizzle(spv::Decoration precision, const glslang::TIntermTyped& node, spv::Id parentResult) +spv::Id TGlslangToSpvTraverser::createInvertedSwizzle(spv::Decoration precision, const glslang::TIntermTyped& node, + spv::Id parentResult) { std::vector<unsigned> swizzle; convertSwizzle(*node.getAsBinaryNode()->getRight()->getAsAggregate(), swizzle); @@ -3444,8 +4051,37 @@ spv::Id TGlslangToSpvTraverser::convertGlslangToSpvType(const glslang::TType& ty builder.addCapability(spv::CapabilityAtomicStorage); spvType = builder.makeUintType(32); break; - case glslang::EbtAccStructNV: - spvType = builder.makeAccelerationStructureNVType(); + case glslang::EbtAccStruct: + switch (glslangIntermediate->getStage()) { + case EShLangRayGen: + case EShLangIntersect: + case EShLangAnyHit: + case EShLangClosestHit: + case EShLangMiss: + case EShLangCallable: + // these all should have the RayTracingNV/KHR capability already + break; + default: + { + auto& extensions = glslangIntermediate->getRequestedExtensions(); + if (extensions.find("GL_EXT_ray_query") != extensions.end()) { + builder.addExtension(spv::E_SPV_KHR_ray_query); + builder.addCapability(spv::CapabilityRayQueryKHR); + } + } + break; + } + spvType = builder.makeAccelerationStructureType(); + break; + case glslang::EbtRayQuery: + { + auto& extensions = glslangIntermediate->getRequestedExtensions(); + if (extensions.find("GL_EXT_ray_query") != extensions.end()) { + builder.addExtension(spv::E_SPV_KHR_ray_query); + builder.addCapability(spv::CapabilityRayQueryKHR); + } + spvType = builder.makeRayQueryType(); + } break; case glslang::EbtReference: { @@ -3501,6 +4137,80 @@ spv::Id TGlslangToSpvTraverser::convertGlslangToSpvType(const glslang::TType& ty spvType = convertGlslangStructToSpvType(type, glslangMembers, explicitLayout, qualifier); } break; + case glslang::EbtString: + // no type used for OpString + return 0; +#ifndef GLSLANG_WEB + case glslang::EbtSpirvType: { + // GL_EXT_spirv_intrinsics + const auto& spirvType = type.getSpirvType(); + const auto& spirvInst = spirvType.spirvInst; + + std::vector<spv::Id> operands; + for (const auto& typeParam : spirvType.typeParams) { + if (typeParam.isConstant) { + // Constant expression + if (typeParam.constant->isLiteral()) { + if (typeParam.constant->getBasicType() == glslang::EbtFloat) { + float floatValue = static_cast<float>(typeParam.constant->getConstArray()[0].getDConst()); + unsigned literal = *reinterpret_cast<unsigned*>(&floatValue); + operands.push_back(literal); + } else if (typeParam.constant->getBasicType() == glslang::EbtInt) { + unsigned literal = typeParam.constant->getConstArray()[0].getIConst(); + operands.push_back(literal); + } else if (typeParam.constant->getBasicType() == glslang::EbtUint) { + unsigned literal = typeParam.constant->getConstArray()[0].getUConst(); + operands.push_back(literal); + } else if (typeParam.constant->getBasicType() == glslang::EbtBool) { + unsigned literal = typeParam.constant->getConstArray()[0].getBConst(); + operands.push_back(literal); + } else if (typeParam.constant->getBasicType() == glslang::EbtString) { + auto str = typeParam.constant->getConstArray()[0].getSConst()->c_str(); + unsigned literal = 0; + char* literalPtr = reinterpret_cast<char*>(&literal); + unsigned charCount = 0; + char ch = 0; + do { + ch = *(str++); + *(literalPtr++) = ch; + ++charCount; + if (charCount == 4) { + operands.push_back(literal); + literalPtr = reinterpret_cast<char*>(&literal); + charCount = 0; + } + } while (ch != 0); + + // Partial literal is padded with 0 + if (charCount > 0) { + for (; charCount < 4; ++charCount) + *(literalPtr++) = 0; + operands.push_back(literal); + } + } else + assert(0); // Unexpected type + } else { + int nextConst = 0; + spv::Id constant = createSpvConstantFromConstUnionArray( + typeParam.constant->getType(), typeParam.constant->getConstArray(), nextConst, false); + operands.push_back(constant); + } + } else { + // Type specifier + spv::Id typeId = convertGlslangToSpvType(*typeParam.type); + operands.push_back(typeId); + } + } + + if (spirvInst.set == "") + spvType = builder.createOp(static_cast<spv::Op>(spirvInst.id), spv::NoType, operands); + else { + spvType = builder.createBuiltinCall( + spv::NoType, getExtBuiltins(spirvInst.set.c_str()), spirvInst.id, operands); + } + break; + } +#endif default: assert(0); break; @@ -3626,7 +4336,8 @@ spv::Id TGlslangToSpvTraverser::convertGlslangStructToSpvType(const glslang::TTy { // Create a vector of struct types for SPIR-V to consume std::vector<spv::Id> spvMembers; - int memberDelta = 0; // how much the member's index changes from glslang to SPIR-V, normally 0, except sometimes for blocks + int memberDelta = 0; // how much the member's index changes from glslang to SPIR-V, normally 0, + // except sometimes for blocks std::vector<std::pair<glslang::TType*, glslang::TQualifier> > deferredForwardPointers; for (int i = 0; i < (int)glslangMembers->size(); i++) { glslang::TType& glslangMember = *(*glslangMembers)[i].type; @@ -3662,10 +4373,12 @@ spv::Id TGlslangToSpvTraverser::convertGlslangStructToSpvType(const glslang::TTy deferredForwardPointers.push_back(std::make_pair(&glslangMember, memberQualifier)); } spvMembers.push_back( - convertGlslangToSpvType(glslangMember, explicitLayout, memberQualifier, lastBufferBlockMember, true)); + convertGlslangToSpvType(glslangMember, explicitLayout, memberQualifier, lastBufferBlockMember, + true)); } else { spvMembers.push_back( - convertGlslangToSpvType(glslangMember, explicitLayout, memberQualifier, lastBufferBlockMember, false)); + convertGlslangToSpvType(glslangMember, explicitLayout, memberQualifier, lastBufferBlockMember, + false)); } } } @@ -3694,7 +4407,8 @@ void TGlslangToSpvTraverser::decorateStructType(const glslang::TType& type, { // Name and decorate the non-hidden members int offset = -1; - int locationOffset = 0; // for use within the members of this struct + bool memberLocationInvalid = type.isArrayOfArrays() || + (type.isArray() && (type.getQualifier().isArrayedIo(glslangIntermediate->getStage()) == false)); for (int i = 0; i < (int)glslangMembers->size(); i++) { glslang::TType& glslangMember = *(*glslangMembers)[i].type; int member = i; @@ -3747,13 +4461,9 @@ void TGlslangToSpvTraverser::decorateStructType(const glslang::TType& type, // just track whether a member needs to be decorated. // Ignore member locations if the container is an array, as that's // ill-specified and decisions have been made to not allow this. - if (! type.isArray() && memberQualifier.hasLocation()) + if (!memberLocationInvalid && memberQualifier.hasLocation()) builder.addMemberDecoration(spvType, member, spv::DecorationLocation, memberQualifier.layoutLocation); - if (qualifier.hasLocation()) // track for upcoming inheritance - locationOffset += glslangIntermediate->computeTypeLocationSize( - glslangMember, glslangIntermediate->getStage()); - // component, XFB, others if (glslangMember.getQualifier().hasComponent()) builder.addMemberDecoration(spvType, member, spv::DecorationComponent, @@ -3809,6 +4519,38 @@ void TGlslangToSpvTraverser::decorateStructType(const glslang::TType& type, builder.addCapability(spv::CapabilityGeometryShaderPassthroughNV); builder.addExtension(spv::E_SPV_NV_geometry_shader_passthrough); } + + // + // Add SPIR-V decorations for members (GL_EXT_spirv_intrinsics) + // + if (glslangMember.getQualifier().hasSprivDecorate()) { + const glslang::TSpirvDecorate& spirvDecorate = glslangMember.getQualifier().getSpirvDecorate(); + + // Add spirv_decorate + for (auto& decorate : spirvDecorate.decorates) { + if (!decorate.second.empty()) { + std::vector<unsigned> literals; + TranslateLiterals(decorate.second, literals); + builder.addMemberDecoration(spvType, member, static_cast<spv::Decoration>(decorate.first), literals); + } + else + builder.addMemberDecoration(spvType, member, static_cast<spv::Decoration>(decorate.first)); + } + + // spirv_decorate_id not applied to members + assert(spirvDecorate.decorateIds.empty()); + + // Add spirv_decorate_string + for (auto& decorateString : spirvDecorate.decorateStrings) { + std::vector<const char*> strings; + assert(!decorateString.second.empty()); + for (auto extraOperand : decorateString.second) { + const char* string = extraOperand->getConstArray()[0].getSConst()->c_str(); + strings.push_back(string); + } + builder.addDecoration(spvType, static_cast<spv::Decoration>(decorateString.first), strings); + } + } #endif } @@ -3827,6 +4569,8 @@ spv::Id TGlslangToSpvTraverser::makeArraySizeId(const glslang::TArraySizes& arra glslang::TIntermTyped* specNode = arraySizes.getDimNode(dim); if (specNode != nullptr) { builder.clearAccessChain(); + SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder); + spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); specNode->traverse(this); return accessChainLoad(specNode->getAsTyped()->getType()); } @@ -3853,26 +4597,16 @@ spv::Id TGlslangToSpvTraverser::accessChainLoad(const glslang::TType& type) alignment |= type.getBufferReferenceAlignment(); spv::Id loadedId = builder.accessChainLoad(TranslatePrecisionDecoration(type), - TranslateNonUniformDecoration(type.getQualifier()), - nominalTypeId, - spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & ~spv::MemoryAccessMakePointerAvailableKHRMask), - TranslateMemoryScope(coherentFlags), - alignment); + TranslateNonUniformDecoration(builder.getAccessChain().coherentFlags), + TranslateNonUniformDecoration(type.getQualifier()), + nominalTypeId, + spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & ~spv::MemoryAccessMakePointerAvailableKHRMask), + TranslateMemoryScope(coherentFlags), + alignment); // Need to convert to abstract types when necessary if (type.getBasicType() == glslang::EbtBool) { - if (builder.isScalarType(nominalTypeId)) { - // Conversion for bool - spv::Id boolType = builder.makeBoolType(); - if (nominalTypeId != boolType) - loadedId = builder.createBinOp(spv::OpINotEqual, boolType, loadedId, builder.makeUintConstant(0)); - } else if (builder.isVectorType(nominalTypeId)) { - // Conversion for bvec - int vecSize = builder.getNumTypeComponents(nominalTypeId); - spv::Id bvecType = builder.makeVectorType(builder.makeBoolType(), vecSize); - if (nominalTypeId != bvecType) - loadedId = builder.createBinOp(spv::OpINotEqual, bvecType, loadedId, makeSmearedConstant(builder.makeUintConstant(0), vecSize)); - } + loadedId = convertLoadedBoolInUniformToUint(type, nominalTypeId, loadedId); } return loadedId; @@ -3919,8 +4653,9 @@ void TGlslangToSpvTraverser::accessChainStore(const glslang::TType& type, spv::I unsigned int alignment = builder.getAccessChain().alignment; alignment |= type.getBufferReferenceAlignment(); - builder.accessChainStore(rvalue, - spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & ~spv::MemoryAccessMakePointerVisibleKHRMask), + builder.accessChainStore(rvalue, TranslateNonUniformDecoration(builder.getAccessChain().coherentFlags), + spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & + ~spv::MemoryAccessMakePointerVisibleKHRMask), TranslateMemoryScope(coherentFlags), alignment); } @@ -3981,7 +4716,8 @@ void TGlslangToSpvTraverser::multiTypeStore(const glslang::TType& type, spv::Id // set up the target storage builder.clearAccessChain(); builder.setAccessChainLValue(lValue); - builder.accessChainPush(builder.makeIntConstant(index), TranslateCoherent(type), type.getBufferReferenceAlignment()); + builder.accessChainPush(builder.makeIntConstant(index), TranslateCoherent(type), + type.getBufferReferenceAlignment()); // store the member multiTypeStore(glslangElementType, elementRValue); @@ -4001,7 +4737,8 @@ void TGlslangToSpvTraverser::multiTypeStore(const glslang::TType& type, spv::Id // set up the target storage builder.clearAccessChain(); builder.setAccessChainLValue(lValue); - builder.accessChainPush(builder.makeIntConstant(m), TranslateCoherent(type), type.getBufferReferenceAlignment()); + builder.accessChainPush(builder.makeIntConstant(m), TranslateCoherent(type), + type.getBufferReferenceAlignment()); // store the member multiTypeStore(glslangMemberType, memberRValue); @@ -4021,6 +4758,7 @@ glslang::TLayoutPacking TGlslangToSpvTraverser::getExplicitLayout(const glslang: // has to be a uniform or buffer block or task in/out blocks if (type.getQualifier().storage != glslang::EvqUniform && type.getQualifier().storage != glslang::EvqBuffer && + type.getQualifier().storage != glslang::EvqShared && !type.getQualifier().isTaskMemory()) return glslang::ElpNone; @@ -4036,18 +4774,21 @@ glslang::TLayoutPacking TGlslangToSpvTraverser::getExplicitLayout(const glslang: } // Given an array type, returns the integer stride required for that array -int TGlslangToSpvTraverser::getArrayStride(const glslang::TType& arrayType, glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout) +int TGlslangToSpvTraverser::getArrayStride(const glslang::TType& arrayType, glslang::TLayoutPacking explicitLayout, + glslang::TLayoutMatrix matrixLayout) { int size; int stride; - glslangIntermediate->getMemberAlignment(arrayType, size, stride, explicitLayout, matrixLayout == glslang::ElmRowMajor); + glslangIntermediate->getMemberAlignment(arrayType, size, stride, explicitLayout, + matrixLayout == glslang::ElmRowMajor); return stride; } // Given a matrix type, or array (of array) of matrixes type, returns the integer stride required for that matrix // when used as a member of an interface block -int TGlslangToSpvTraverser::getMatrixStride(const glslang::TType& matrixType, glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout) +int TGlslangToSpvTraverser::getMatrixStride(const glslang::TType& matrixType, glslang::TLayoutPacking explicitLayout, + glslang::TLayoutMatrix matrixLayout) { glslang::TType elementType; elementType.shallowCopy(matrixType); @@ -4055,7 +4796,8 @@ int TGlslangToSpvTraverser::getMatrixStride(const glslang::TType& matrixType, gl int size; int stride; - glslangIntermediate->getMemberAlignment(elementType, size, stride, explicitLayout, matrixLayout == glslang::ElmRowMajor); + glslangIntermediate->getMemberAlignment(elementType, size, stride, explicitLayout, + matrixLayout == glslang::ElmRowMajor); return stride; } @@ -4066,8 +4808,8 @@ int TGlslangToSpvTraverser::getMatrixStride(const glslang::TType& matrixType, gl // 'currentOffset' should be passed in already initialized, ready to modify, and reflecting // the migration of data from nextOffset -> currentOffset. It should be -1 on the first call. // -1 means a non-forced member offset (no decoration needed). -void TGlslangToSpvTraverser::updateMemberOffset(const glslang::TType& structType, const glslang::TType& memberType, int& currentOffset, int& nextOffset, - glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout) +void TGlslangToSpvTraverser::updateMemberOffset(const glslang::TType& structType, const glslang::TType& memberType, + int& currentOffset, int& nextOffset, glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout) { // this will get a positive value when deemed necessary nextOffset = -1; @@ -4097,7 +4839,8 @@ void TGlslangToSpvTraverser::updateMemberOffset(const glslang::TType& structType int memberSize; int dummyStride; - int memberAlignment = glslangIntermediate->getMemberAlignment(memberType, memberSize, dummyStride, explicitLayout, matrixLayout == glslang::ElmRowMajor); + int memberAlignment = glslangIntermediate->getMemberAlignment(memberType, memberSize, dummyStride, explicitLayout, + matrixLayout == glslang::ElmRowMajor); // Adjust alignment for HLSL rules // TODO: make this consistent in early phases of code: @@ -4116,7 +4859,8 @@ void TGlslangToSpvTraverser::updateMemberOffset(const glslang::TType& structType glslang::RoundToPow2(currentOffset, memberAlignment); // Bump up to vec4 if there is a bad straddle - if (explicitLayout != glslang::ElpScalar && glslangIntermediate->improperStraddle(memberType, memberSize, currentOffset)) + if (explicitLayout != glslang::ElpScalar && glslangIntermediate->improperStraddle(memberType, memberSize, + currentOffset)) glslang::RoundToPow2(currentOffset, 16); nextOffset = currentOffset + memberSize; @@ -4169,8 +4913,10 @@ bool TGlslangToSpvTraverser::writableParam(glslang::TStorageQualifier qualifier) assert(qualifier == glslang::EvqIn || qualifier == glslang::EvqOut || qualifier == glslang::EvqInOut || + qualifier == glslang::EvqUniform || qualifier == glslang::EvqConstReadOnly); - return qualifier != glslang::EvqConstReadOnly; + return qualifier != glslang::EvqConstReadOnly && + qualifier != glslang::EvqUniform; } // Is parameter pass-by-original? @@ -4182,13 +4928,17 @@ bool TGlslangToSpvTraverser::originalParam(glslang::TStorageQualifier qualifier, if (glslangIntermediate->getSource() == glslang::EShSourceHlsl) return paramType.getBasicType() == glslang::EbtBlock; return paramType.containsOpaque() || // sampler, etc. +#ifndef GLSLANG_WEB + paramType.getQualifier().isSpirvByReference() || // spirv_by_reference +#endif (paramType.getBasicType() == glslang::EbtBlock && qualifier == glslang::EvqBuffer); // SSBO } // Make all the functions, skeletally, without actually visiting their bodies. void TGlslangToSpvTraverser::makeFunctions(const glslang::TIntermSequence& glslFunctions) { - const auto getParamDecorations = [&](std::vector<spv::Decoration>& decorations, const glslang::TType& type, bool useVulkanMemoryModel) { + const auto getParamDecorations = [&](std::vector<spv::Decoration>& decorations, const glslang::TType& type, + bool useVulkanMemoryModel) { spv::Decoration paramPrecision = TranslatePrecisionDecoration(type); if (paramPrecision != spv::NoPrecision) decorations.push_back(paramPrecision); @@ -4286,7 +5036,8 @@ void TGlslangToSpvTraverser::makeGlobalInitializers(const glslang::TIntermSequen builder.setBuildPoint(shaderEntry->getLastBlock()); for (int i = 0; i < (int)initializers.size(); ++i) { glslang::TIntermAggregate* initializer = initializers[i]->getAsAggregate(); - if (initializer && initializer->getOp() != glslang::EOpFunction && initializer->getOp() != glslang::EOpLinkerObjects) { + if (initializer && initializer->getOp() != glslang::EOpFunction && initializer->getOp() != + glslang::EOpLinkerObjects) { // We're on a top-level node that's not a function. Treat as an initializer, whose // code goes into the beginning of the entry point. @@ -4294,7 +5045,39 @@ void TGlslangToSpvTraverser::makeGlobalInitializers(const glslang::TIntermSequen } } } +// Walk over all linker objects to create a map for payload and callable data linker objects +// and their location to be used during codegen for OpTraceKHR and OpExecuteCallableKHR +// This is done here since it is possible that these linker objects are not be referenced in the AST +void TGlslangToSpvTraverser::collectRayTracingLinkerObjects() +{ + glslang::TIntermAggregate* linkerObjects = glslangIntermediate->findLinkerObjects(); + for (auto& objSeq : linkerObjects->getSequence()) { + auto objNode = objSeq->getAsSymbolNode(); + if (objNode != nullptr) { + if (objNode->getQualifier().hasLocation()) { + unsigned int location = objNode->getQualifier().layoutLocation; + auto st = objNode->getQualifier().storage; + int set; + switch (st) + { + case glslang::EvqPayload: + case glslang::EvqPayloadIn: + set = 0; + break; + case glslang::EvqCallableData: + case glslang::EvqCallableDataIn: + set = 1; + break; + default: + set = -1; + } + if (set != -1) + locationToSymbol[set].insert(std::make_pair(location, objNode)); + } + } + } +} // Process all the functions, while skipping initializers. void TGlslangToSpvTraverser::visitFunctions(const glslang::TIntermSequence& glslFunctions) { @@ -4314,7 +5097,8 @@ void TGlslangToSpvTraverser::handleFunctionEntry(const glslang::TIntermAggregate builder.setBuildPoint(functionBlock); } -void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate& node, std::vector<spv::Id>& arguments, spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags) +void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate& node, std::vector<spv::Id>& arguments, + spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags) { const glslang::TIntermSequence& glslangArguments = node.getSequence(); @@ -4327,7 +5111,8 @@ void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate& sampler = glslangArguments[0]->getAsTyped()->getType().getSampler(); cubeCompare = sampler.dim == glslang::EsdCube && sampler.arrayed && sampler.shadow; #ifndef GLSLANG_WEB - f16ShadowCompare = sampler.shadow && glslangArguments[1]->getAsTyped()->getType().getBasicType() == glslang::EbtFloat16; + f16ShadowCompare = sampler.shadow && + glslangArguments[1]->getAsTyped()->getType().getBasicType() == glslang::EbtFloat16; #endif } @@ -4436,8 +5221,10 @@ void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate& } if (lvalue) { - arguments.push_back(builder.accessChainGetLValue()); + spv::Id lvalue_id = builder.accessChainGetLValue(); + arguments.push_back(lvalue_id); lvalueCoherentFlags = builder.getAccessChain().coherentFlags; + builder.addDecoration(lvalue_id, TranslateNonUniformDecoration(lvalueCoherentFlags)); lvalueCoherentFlags |= TranslateCoherent(glslangArguments[i]->getAsTyped()->getType()); } else #endif @@ -4490,7 +5277,7 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO translateArguments(*node->getAsAggregate(), arguments, lvalueCoherentFlags); else translateArguments(*node->getAsUnaryNode(), arguments); - spv::Decoration precision = TranslatePrecisionDecoration(node->getOperationPrecision()); + spv::Decoration precision = TranslatePrecisionDecoration(node->getType()); spv::Builder::TextureParameters params = { }; params.sampler = arguments[0]; @@ -4500,12 +5287,15 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO const bool isUnsignedResult = node->getType().getBasicType() == glslang::EbtUint; + if (builder.isSampledImage(params.sampler) && + ((cracked.query && node->getOp() != glslang::EOpTextureQueryLod) || cracked.fragMask || cracked.fetch)) { + params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler); + if (imageType.getQualifier().isNonUniform()) { + builder.addDecoration(params.sampler, spv::DecorationNonUniformEXT); + } + } // Check for queries if (cracked.query) { - // OpImageQueryLod works on a sampled image, for other queries the image has to be extracted first - if (node->getOp() != glslang::EOpTextureQueryLod && builder.isSampledImage(params.sampler)) - params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler); - switch (node->getOp()) { case glslang::EOpImageQuerySize: case glslang::EOpTextureQuerySize: @@ -4534,7 +5324,10 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO int components = node->getType().getVectorSize(); - if (node->getOp() == glslang::EOpTextureFetch) { + if (node->getOp() == glslang::EOpImageLoad || + node->getOp() == glslang::EOpImageLoadLod || + node->getOp() == glslang::EOpTextureFetch || + node->getOp() == glslang::EOpTextureFetchOffset) { // These must produce 4 components, per SPIR-V spec. We'll add a conversion constructor if needed. // This will only happen through the HLSL path for operator[], so we do not have to handle e.g. // the EOpTexture/Proj/Lod/etc family. It would be harmless to do so, but would need more logic @@ -4703,7 +5496,8 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO operands.push_back(imageOperand); } if (mask & spv::ImageOperandsMakeTexelVisibleKHRMask) { - spv::IdImmediate imageOperand = { true, builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) }; + spv::IdImmediate imageOperand = { true, builder.makeUintConstant(TranslateMemoryScope( + TranslateCoherent(imageType))) }; operands.push_back(imageOperand); } @@ -4731,18 +5525,22 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO // imageAtomicStore has a void return type so base the pointer type on // the type of the value operand. if (node->getOp() == glslang::EOpImageAtomicStore) { - resultTypeId = builder.makePointer(spv::StorageClassImage, builder.getTypeId(operands[2].word)); + resultTypeId = builder.makePointer(spv::StorageClassImage, builder.getTypeId(*opIt)); } else { resultTypeId = builder.makePointer(spv::StorageClassImage, resultType()); } spv::Id pointer = builder.createOp(spv::OpImageTexelPointer, resultTypeId, operands); + if (imageType.getQualifier().nonUniform) { + builder.addDecoration(pointer, spv::DecorationNonUniformEXT); + } std::vector<spv::Id> operands; operands.push_back(pointer); for (; opIt != arguments.end(); ++opIt) operands.push_back(*opIt); - return createAtomicOperation(node->getOp(), precision, resultType(), operands, node->getBasicType(), lvalueCoherentFlags); + return createAtomicOperation(node->getOp(), precision, resultType(), operands, node->getBasicType(), + lvalueCoherentFlags); } } @@ -4754,10 +5552,6 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO auto opIt = arguments.begin(); std::vector<spv::Id> operands; - // Extract the image if necessary - if (builder.isSampledImage(params.sampler)) - params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler); - operands.push_back(params.sampler); ++opIt; @@ -4767,7 +5561,8 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO std::vector<spv::Id> comps; comps.push_back(zero); comps.push_back(zero); - operands.push_back(builder.makeCompositeConstant(builder.makeVectorType(builder.makeIntType(32), 2), comps)); + operands.push_back(builder.makeCompositeConstant( + builder.makeVectorType(builder.makeIntType(32), 2), comps)); } for (; opIt != arguments.end(); ++opIt) @@ -4817,13 +5612,6 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO bias = true; } - // See if the sampler param should really be just the SPV image part - if (cracked.fetch) { - // a fetch needs to have the image extracted first - if (builder.isSampledImage(params.sampler)) - params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler); - } - #ifndef GLSLANG_WEB if (cracked.gather) { const auto& sourceExtensions = glslangIntermediate->getRequestedExtensions(); @@ -4856,7 +5644,8 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO else dRefComp = builder.getNumComponents(params.coords) - 1; indexes.push_back(dRefComp); - params.Dref = builder.createCompositeExtract(params.coords, builder.getScalarTypeId(builder.getTypeId(params.coords)), indexes); + params.Dref = builder.createCompositeExtract(params.coords, + builder.getScalarTypeId(builder.getTypeId(params.coords)), indexes); } // lod @@ -4981,7 +5770,8 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO flags.clear(); builder.accessChainPush(builder.makeIntConstant(i), flags, 0); - builder.accessChainStore(builder.createCompositeExtract(res, builder.getContainedTypeId(resType, i+1), i+1)); + builder.accessChainStore(builder.createCompositeExtract(res, builder.getContainedTypeId(resType, i+1), + i+1), TranslateNonUniformDecoration(imageType.getQualifier())); } return builder.createCompositeExtract(res, resultType(), 0); } @@ -5004,10 +5794,9 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO // copy the projective coordinate if we have to if (projTargetComp != projSourceComp) { spv::Id projComp = builder.createCompositeExtract(params.coords, - builder.getScalarTypeId(builder.getTypeId(params.coords)), - projSourceComp); + builder.getScalarTypeId(builder.getTypeId(params.coords)), projSourceComp); params.coords = builder.createCompositeInsert(projComp, params.coords, - builder.getTypeId(params.coords), projTargetComp); + builder.getTypeId(params.coords), projTargetComp); } } @@ -5087,7 +5876,8 @@ spv::Id TGlslangToSpvTraverser::handleUserFunctionCall(const glslang::TIntermAgg ++lValueCount; } else if (writableParam(qualifiers[a])) { // need space to hold the copy - arg = builder.createVariable(spv::StorageClassFunction, builder.getContainedTypeId(function->getParamType(a)), "param"); + arg = builder.createVariable(function->getParamPrecision(a), spv::StorageClassFunction, + builder.getContainedTypeId(function->getParamType(a)), "param"); if (qualifiers[a] == glslang::EvqIn || qualifiers[a] == glslang::EvqInOut) { // need to copy the input into output space builder.setAccessChain(lValues[lValueCount]); @@ -5099,12 +5889,14 @@ spv::Id TGlslangToSpvTraverser::handleUserFunctionCall(const glslang::TIntermAgg ++lValueCount; } else { // process r-value, which involves a copy for a type mismatch - if (function->getParamType(a) != convertGlslangToSpvType(*argTypes[a])) { - spv::Id argCopy = builder.createVariable(spv::StorageClassFunction, function->getParamType(a), "arg"); + if (function->getParamType(a) != builder.getTypeId(rValues[rValueCount]) || + TranslatePrecisionDecoration(*argTypes[a]) != function->getParamPrecision(a)) + { + spv::Id argCopy = builder.createVariable(function->getParamPrecision(a), spv::StorageClassFunction, function->getParamType(a), "arg"); builder.clearAccessChain(); builder.setAccessChainLValue(argCopy); multiTypeStore(*argTypes[a], rValues[rValueCount]); - arg = builder.createLoad(argCopy); + arg = builder.createLoad(argCopy, function->getParamPrecision(a)); } else arg = rValues[rValueCount]; ++rValueCount; @@ -5115,6 +5907,7 @@ spv::Id TGlslangToSpvTraverser::handleUserFunctionCall(const glslang::TIntermAgg // 3. Make the call. spv::Id result = builder.createFunctionCall(function, spvArgs); builder.setPrecision(result, TranslatePrecisionDecoration(node->getType())); + builder.addDecoration(result, TranslateNonUniformDecoration(node->getType().getQualifier())); // 4. Copy back out an "out" arguments. lValueCount = 0; @@ -5123,7 +5916,8 @@ spv::Id TGlslangToSpvTraverser::handleUserFunctionCall(const glslang::TIntermAgg ++lValueCount; else if (writableParam(qualifiers[a])) { if (qualifiers[a] == glslang::EvqOut || qualifiers[a] == glslang::EvqInOut) { - spv::Id copy = builder.createLoad(spvArgs[a]); + spv::Id copy = builder.createLoad(spvArgs[a], spv::NoPrecision); + builder.addDecoration(copy, TranslateNonUniformDecoration(argTypes[a]->getQualifier())); builder.setAccessChain(lValues[lValueCount]); multiTypeStore(*argTypes[a], copy); } @@ -5368,7 +6162,7 @@ spv::Id TGlslangToSpvTraverser::createBinaryOperation(glslang::TOperator op, OpD case glslang::EOpNotEqual: case glslang::EOpVectorNotEqual: if (isFloat) - binOp = spv::OpFOrdNotEqual; + binOp = spv::OpFUnordNotEqual; else if (isBool) binOp = spv::OpLogicalNotEqual; else @@ -5502,7 +6296,7 @@ spv::Id TGlslangToSpvTraverser::createBinaryMatrixOperation(spv::Op op, OpDecora } spv::Id TGlslangToSpvTraverser::createUnaryOperation(glslang::TOperator op, OpDecorations& decorations, spv::Id typeId, - spv::Id operand, glslang::TBasicType typeProxy, const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags) + spv::Id operand, glslang::TBasicType typeProxy, const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags) { spv::Op unaryOp = spv::OpNop; int extBuiltins = -1; @@ -5763,6 +6557,24 @@ spv::Id TGlslangToSpvTraverser::createUnaryOperation(glslang::TOperator op, OpDe case glslang::EOpFwidthCoarse: unaryOp = spv::OpFwidthCoarse; break; + case glslang::EOpRayQueryProceed: + unaryOp = spv::OpRayQueryProceedKHR; + break; + case glslang::EOpRayQueryGetRayTMin: + unaryOp = spv::OpRayQueryGetRayTMinKHR; + break; + case glslang::EOpRayQueryGetRayFlags: + unaryOp = spv::OpRayQueryGetRayFlagsKHR; + break; + case glslang::EOpRayQueryGetWorldRayOrigin: + unaryOp = spv::OpRayQueryGetWorldRayOriginKHR; + break; + case glslang::EOpRayQueryGetWorldRayDirection: + unaryOp = spv::OpRayQueryGetWorldRayDirectionKHR; + break; + case glslang::EOpRayQueryGetIntersectionCandidateAABBOpaque: + unaryOp = spv::OpRayQueryGetIntersectionCandidateAABBOpaqueKHR; + break; case glslang::EOpInterpolateAtCentroid: if (typeProxy == glslang::EbtFloat16) builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float); @@ -5893,6 +6705,11 @@ spv::Id TGlslangToSpvTraverser::createUnaryOperation(glslang::TOperator op, OpDe case glslang::EOpConstructReference: unaryOp = spv::OpBitcast; break; + + case glslang::EOpConvUint64ToAccStruct: + case glslang::EOpConvUvec2ToAccStruct: + unaryOp = spv::OpConvertUToAccelerationStructureKHR; + break; #endif case glslang::EOpCopyObject: @@ -6048,7 +6865,7 @@ spv::Id TGlslangToSpvTraverser::createConversion(glslang::TOperator op, OpDecora case glslang::EOpConvFloatToBool: zero = builder.makeFloatConstant(0.0F); zero = makeSmearedConstant(zero, vectorSize); - return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero); + return builder.createBinOp(spv::OpFUnordNotEqual, destType, operand, zero); case glslang::EOpConvBoolToFloat: convOp = spv::OpSelect; zero = builder.makeFloatConstant(0.0F); @@ -6197,11 +7014,11 @@ spv::Id TGlslangToSpvTraverser::createConversion(glslang::TOperator op, OpDecora case glslang::EOpConvDoubleToBool: zero = builder.makeDoubleConstant(0.0); zero = makeSmearedConstant(zero, vectorSize); - return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero); + return builder.createBinOp(spv::OpFUnordNotEqual, destType, operand, zero); case glslang::EOpConvFloat16ToBool: zero = builder.makeFloat16Constant(0.0F); zero = makeSmearedConstant(zero, vectorSize); - return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero); + return builder.createBinOp(spv::OpFUnordNotEqual, destType, operand, zero); case glslang::EOpConvBoolToDouble: convOp = spv::OpSelect; zero = builder.makeDoubleConstant(0.0); @@ -6355,9 +7172,6 @@ spv::Id TGlslangToSpvTraverser::createConversion(glslang::TOperator op, OpDecora break; case glslang::EOpConvPtrToUvec2: case glslang::EOpConvUvec2ToPtr: - if (builder.isVector(operand)) - builder.promoteIncorporatedExtension(spv::E_SPV_EXT_physical_storage_buffer, - spv::E_SPV_KHR_physical_storage_buffer, spv::Spv_1_5); convOp = spv::OpBitcast; break; #endif @@ -6395,7 +7209,9 @@ spv::Id TGlslangToSpvTraverser::makeSmearedConstant(spv::Id constant, int vector } // For glslang ops that map to SPV atomic opCodes -spv::Id TGlslangToSpvTraverser::createAtomicOperation(glslang::TOperator op, spv::Decoration /*precision*/, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy, const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags) +spv::Id TGlslangToSpvTraverser::createAtomicOperation(glslang::TOperator op, spv::Decoration /*precision*/, + spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy, + const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags) { spv::Op opCode = spv::OpNop; @@ -6404,19 +7220,58 @@ spv::Id TGlslangToSpvTraverser::createAtomicOperation(glslang::TOperator op, spv case glslang::EOpImageAtomicAdd: case glslang::EOpAtomicCounterAdd: opCode = spv::OpAtomicIAdd; + if (typeProxy == glslang::EbtFloat16 || typeProxy == glslang::EbtFloat || typeProxy == glslang::EbtDouble) { + opCode = spv::OpAtomicFAddEXT; + builder.addExtension(spv::E_SPV_EXT_shader_atomic_float_add); + if (typeProxy == glslang::EbtFloat16) { + builder.addExtension(spv::E_SPV_EXT_shader_atomic_float16_add); + builder.addCapability(spv::CapabilityAtomicFloat16AddEXT); + } else if (typeProxy == glslang::EbtFloat) { + builder.addCapability(spv::CapabilityAtomicFloat32AddEXT); + } else { + builder.addCapability(spv::CapabilityAtomicFloat64AddEXT); + } + } break; + case glslang::EOpAtomicSubtract: case glslang::EOpAtomicCounterSubtract: opCode = spv::OpAtomicISub; break; case glslang::EOpAtomicMin: case glslang::EOpImageAtomicMin: case glslang::EOpAtomicCounterMin: - opCode = (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) ? spv::OpAtomicUMin : spv::OpAtomicSMin; + if (typeProxy == glslang::EbtFloat16 || typeProxy == glslang::EbtFloat || typeProxy == glslang::EbtDouble) { + opCode = spv::OpAtomicFMinEXT; + builder.addExtension(spv::E_SPV_EXT_shader_atomic_float_min_max); + if (typeProxy == glslang::EbtFloat16) + builder.addCapability(spv::CapabilityAtomicFloat16MinMaxEXT); + else if (typeProxy == glslang::EbtFloat) + builder.addCapability(spv::CapabilityAtomicFloat32MinMaxEXT); + else + builder.addCapability(spv::CapabilityAtomicFloat64MinMaxEXT); + } else if (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) { + opCode = spv::OpAtomicUMin; + } else { + opCode = spv::OpAtomicSMin; + } break; case glslang::EOpAtomicMax: case glslang::EOpImageAtomicMax: case glslang::EOpAtomicCounterMax: - opCode = (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) ? spv::OpAtomicUMax : spv::OpAtomicSMax; + if (typeProxy == glslang::EbtFloat16 || typeProxy == glslang::EbtFloat || typeProxy == glslang::EbtDouble) { + opCode = spv::OpAtomicFMaxEXT; + builder.addExtension(spv::E_SPV_EXT_shader_atomic_float_min_max); + if (typeProxy == glslang::EbtFloat16) + builder.addCapability(spv::CapabilityAtomicFloat16MinMaxEXT); + else if (typeProxy == glslang::EbtFloat) + builder.addCapability(spv::CapabilityAtomicFloat32MinMaxEXT); + else + builder.addCapability(spv::CapabilityAtomicFloat64MinMaxEXT); + } else if (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) { + opCode = spv::OpAtomicUMax; + } else { + opCode = spv::OpAtomicSMax; + } break; case glslang::EOpAtomicAnd: case glslang::EOpImageAtomicAnd: @@ -6481,7 +7336,8 @@ spv::Id TGlslangToSpvTraverser::createAtomicOperation(glslang::TOperator op, spv scopeId = builder.makeUintConstant(spv::ScopeDevice); } // semantics default to relaxed - spv::Id semanticsId = builder.makeUintConstant(lvalueCoherentFlags.isVolatile() && glslangIntermediate->usingVulkanMemoryModel() ? + spv::Id semanticsId = builder.makeUintConstant(lvalueCoherentFlags.isVolatile() && + glslangIntermediate->usingVulkanMemoryModel() ? spv::MemorySemanticsVolatileMask : spv::MemorySemanticsMaskNone); spv::Id semanticsId2 = semanticsId; @@ -6494,20 +7350,24 @@ spv::Id TGlslangToSpvTraverser::createAtomicOperation(glslang::TOperator op, spv valueId = operands[2]; if (operands.size() > 3) { scopeId = operands[3]; - semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[4]) | builder.getConstantScalar(operands[5])); - semanticsId2 = builder.makeUintConstant(builder.getConstantScalar(operands[6]) | builder.getConstantScalar(operands[7])); + semanticsId = builder.makeUintConstant( + builder.getConstantScalar(operands[4]) | builder.getConstantScalar(operands[5])); + semanticsId2 = builder.makeUintConstant( + builder.getConstantScalar(operands[6]) | builder.getConstantScalar(operands[7])); } } else if (opCode == spv::OpAtomicLoad) { if (operands.size() > 1) { scopeId = operands[1]; - semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[2]) | builder.getConstantScalar(operands[3])); + semanticsId = builder.makeUintConstant( + builder.getConstantScalar(operands[2]) | builder.getConstantScalar(operands[3])); } } else { // atomic store or RMW valueId = operands[1]; if (operands.size() > 2) { scopeId = operands[2]; - semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[3]) | builder.getConstantScalar(operands[4])); + semanticsId = builder.makeUintConstant + (builder.getConstantScalar(operands[3]) | builder.getConstantScalar(operands[4])); } } @@ -6520,6 +7380,10 @@ spv::Id TGlslangToSpvTraverser::createAtomicOperation(glslang::TOperator op, spv builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); } + if (builder.getConstantScalar(scopeId) == spv::ScopeQueueFamily) { + builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); + } + if (glslangIntermediate->usingVulkanMemoryModel() && builder.getConstantScalar(scopeId) == spv::ScopeDevice) { builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR); } @@ -6552,7 +7416,8 @@ spv::Id TGlslangToSpvTraverser::createAtomicOperation(glslang::TOperator op, spv } // Create group invocation operations. -spv::Id TGlslangToSpvTraverser::createInvocationsOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy) +spv::Id TGlslangToSpvTraverser::createInvocationsOperation(glslang::TOperator op, spv::Id typeId, + std::vector<spv::Id>& operands, glslang::TBasicType typeProxy) { bool isUnsigned = isTypeUnsignedInt(typeProxy); bool isFloat = isTypeFloat(typeProxy); @@ -6766,8 +7631,10 @@ spv::Id TGlslangToSpvTraverser::CreateInvocationsVectorOperation(spv::Op op, spv op == spv::OpGroupFMax || op == spv::OpGroupUMax || op == spv::OpGroupSMax || op == spv::OpGroupFAdd || op == spv::OpGroupIAdd || op == spv::OpGroupBroadcast || op == spv::OpSubgroupReadInvocationKHR || - op == spv::OpGroupFMinNonUniformAMD || op == spv::OpGroupUMinNonUniformAMD || op == spv::OpGroupSMinNonUniformAMD || - op == spv::OpGroupFMaxNonUniformAMD || op == spv::OpGroupUMaxNonUniformAMD || op == spv::OpGroupSMaxNonUniformAMD || + op == spv::OpGroupFMinNonUniformAMD || op == spv::OpGroupUMinNonUniformAMD || + op == spv::OpGroupSMinNonUniformAMD || + op == spv::OpGroupFMaxNonUniformAMD || op == spv::OpGroupUMaxNonUniformAMD || + op == spv::OpGroupSMaxNonUniformAMD || op == spv::OpGroupFAddNonUniformAMD || op == spv::OpGroupIAddNonUniformAMD); // Handle group invocation operations scalar by scalar. @@ -7156,7 +8023,8 @@ spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, s return builder.createOp(opCode, typeId, spvGroupOperands); } -spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy) +spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::Decoration precision, + spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy) { bool isUnsigned = isTypeUnsignedInt(typeProxy); bool isFloat = isTypeFloat(typeProxy); @@ -7258,14 +8126,16 @@ spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv:: unsigned int executionScope = builder.getConstantScalar(operands[0]); unsigned int memoryScope = builder.getConstantScalar(operands[1]); unsigned int semantics = builder.getConstantScalar(operands[2]) | builder.getConstantScalar(operands[3]); - builder.createControlBarrier((spv::Scope)executionScope, (spv::Scope)memoryScope, (spv::MemorySemanticsMask)semantics); + builder.createControlBarrier((spv::Scope)executionScope, (spv::Scope)memoryScope, + (spv::MemorySemanticsMask)semantics); if (semantics & (spv::MemorySemanticsMakeAvailableKHRMask | spv::MemorySemanticsMakeVisibleKHRMask | spv::MemorySemanticsOutputMemoryKHRMask | spv::MemorySemanticsVolatileMask)) { builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); } - if (glslangIntermediate->usingVulkanMemoryModel() && (executionScope == spv::ScopeDevice || memoryScope == spv::ScopeDevice)) { + if (glslangIntermediate->usingVulkanMemoryModel() && (executionScope == spv::ScopeDevice || + memoryScope == spv::ScopeDevice)) { builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR); } return 0; @@ -7348,7 +8218,8 @@ spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv:: if (builder.getNumComponents(operands[0]) == 1) frexpIntType = builder.makeIntegerType(width, true); else - frexpIntType = builder.makeVectorType(builder.makeIntegerType(width, true), builder.getNumComponents(operands[0])); + frexpIntType = builder.makeVectorType(builder.makeIntegerType(width, true), + builder.getNumComponents(operands[0])); typeId = builder.makeStructResultType(typeId0, frexpIntType); consumedOperands = 1; } @@ -7451,24 +8322,116 @@ spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv:: libCall = spv::InterpolateAtVertexAMD; break; - case glslang::EOpReportIntersectionNV: - { + case glslang::EOpReportIntersection: typeId = builder.makeBoolType(); - opCode = spv::OpReportIntersectionNV; - } - break; + opCode = spv::OpReportIntersectionKHR; + break; case glslang::EOpTraceNV: - { builder.createNoResultOp(spv::OpTraceNV, operands); return 0; - } - break; + case glslang::EOpTraceRayMotionNV: + builder.addExtension(spv::E_SPV_NV_ray_tracing_motion_blur); + builder.addCapability(spv::CapabilityRayTracingMotionBlurNV); + builder.createNoResultOp(spv::OpTraceRayMotionNV, operands); + return 0; + case glslang::EOpTraceKHR: + builder.createNoResultOp(spv::OpTraceRayKHR, operands); + return 0; case glslang::EOpExecuteCallableNV: - { builder.createNoResultOp(spv::OpExecuteCallableNV, operands); return 0; - } - break; + case glslang::EOpExecuteCallableKHR: + builder.createNoResultOp(spv::OpExecuteCallableKHR, operands); + return 0; + + case glslang::EOpRayQueryInitialize: + builder.createNoResultOp(spv::OpRayQueryInitializeKHR, operands); + return 0; + case glslang::EOpRayQueryTerminate: + builder.createNoResultOp(spv::OpRayQueryTerminateKHR, operands); + return 0; + case glslang::EOpRayQueryGenerateIntersection: + builder.createNoResultOp(spv::OpRayQueryGenerateIntersectionKHR, operands); + return 0; + case glslang::EOpRayQueryConfirmIntersection: + builder.createNoResultOp(spv::OpRayQueryConfirmIntersectionKHR, operands); + return 0; + case glslang::EOpRayQueryProceed: + typeId = builder.makeBoolType(); + opCode = spv::OpRayQueryProceedKHR; + break; + case glslang::EOpRayQueryGetIntersectionType: + typeId = builder.makeUintType(32); + opCode = spv::OpRayQueryGetIntersectionTypeKHR; + break; + case glslang::EOpRayQueryGetRayTMin: + typeId = builder.makeFloatType(32); + opCode = spv::OpRayQueryGetRayTMinKHR; + break; + case glslang::EOpRayQueryGetRayFlags: + typeId = builder.makeIntType(32); + opCode = spv::OpRayQueryGetRayFlagsKHR; + break; + case glslang::EOpRayQueryGetIntersectionT: + typeId = builder.makeFloatType(32); + opCode = spv::OpRayQueryGetIntersectionTKHR; + break; + case glslang::EOpRayQueryGetIntersectionInstanceCustomIndex: + typeId = builder.makeIntType(32); + opCode = spv::OpRayQueryGetIntersectionInstanceCustomIndexKHR; + break; + case glslang::EOpRayQueryGetIntersectionInstanceId: + typeId = builder.makeIntType(32); + opCode = spv::OpRayQueryGetIntersectionInstanceIdKHR; + break; + case glslang::EOpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffset: + typeId = builder.makeUintType(32); + opCode = spv::OpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR; + break; + case glslang::EOpRayQueryGetIntersectionGeometryIndex: + typeId = builder.makeIntType(32); + opCode = spv::OpRayQueryGetIntersectionGeometryIndexKHR; + break; + case glslang::EOpRayQueryGetIntersectionPrimitiveIndex: + typeId = builder.makeIntType(32); + opCode = spv::OpRayQueryGetIntersectionPrimitiveIndexKHR; + break; + case glslang::EOpRayQueryGetIntersectionBarycentrics: + typeId = builder.makeVectorType(builder.makeFloatType(32), 2); + opCode = spv::OpRayQueryGetIntersectionBarycentricsKHR; + break; + case glslang::EOpRayQueryGetIntersectionFrontFace: + typeId = builder.makeBoolType(); + opCode = spv::OpRayQueryGetIntersectionFrontFaceKHR; + break; + case glslang::EOpRayQueryGetIntersectionCandidateAABBOpaque: + typeId = builder.makeBoolType(); + opCode = spv::OpRayQueryGetIntersectionCandidateAABBOpaqueKHR; + break; + case glslang::EOpRayQueryGetIntersectionObjectRayDirection: + typeId = builder.makeVectorType(builder.makeFloatType(32), 3); + opCode = spv::OpRayQueryGetIntersectionObjectRayDirectionKHR; + break; + case glslang::EOpRayQueryGetIntersectionObjectRayOrigin: + typeId = builder.makeVectorType(builder.makeFloatType(32), 3); + opCode = spv::OpRayQueryGetIntersectionObjectRayOriginKHR; + break; + case glslang::EOpRayQueryGetWorldRayDirection: + typeId = builder.makeVectorType(builder.makeFloatType(32), 3); + opCode = spv::OpRayQueryGetWorldRayDirectionKHR; + break; + case glslang::EOpRayQueryGetWorldRayOrigin: + typeId = builder.makeVectorType(builder.makeFloatType(32), 3); + opCode = spv::OpRayQueryGetWorldRayOriginKHR; + break; + case glslang::EOpRayQueryGetIntersectionObjectToWorld: + typeId = builder.makeMatrixType(builder.makeFloatType(32), 4, 3); + opCode = spv::OpRayQueryGetIntersectionObjectToWorldKHR; + break; + case glslang::EOpRayQueryGetIntersectionWorldToObject: + typeId = builder.makeMatrixType(builder.makeFloatType(32), 4, 3); + opCode = spv::OpRayQueryGetIntersectionWorldToObjectKHR; + break; case glslang::EOpWritePackedPrimitiveIndices4x8NV: builder.createNoResultOp(spv::OpWritePackedPrimitiveIndices4x8NV, operands); return 0; @@ -7558,7 +8521,8 @@ spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv:: spv::Id TGlslangToSpvTraverser::createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId) { // GLSL memory barriers use queuefamily scope in new model, device scope in old model - spv::Scope memoryBarrierScope = glslangIntermediate->usingVulkanMemoryModel() ? spv::ScopeQueueFamilyKHR : spv::ScopeDevice; + spv::Scope memoryBarrierScope = glslangIntermediate->usingVulkanMemoryModel() ? + spv::ScopeQueueFamilyKHR : spv::ScopeDevice; switch (op) { case glslang::EOpBarrier: @@ -7670,7 +8634,18 @@ spv::Id TGlslangToSpvTraverser::createNoArgOperation(glslang::TOperator op, spv: case glslang::EOpTerminateRayNV: builder.createNoResultOp(spv::OpTerminateRayNV); return 0; - + case glslang::EOpRayQueryInitialize: + builder.createNoResultOp(spv::OpRayQueryInitializeKHR); + return 0; + case glslang::EOpRayQueryTerminate: + builder.createNoResultOp(spv::OpRayQueryTerminateKHR); + return 0; + case glslang::EOpRayQueryGenerateIntersection: + builder.createNoResultOp(spv::OpRayQueryGenerateIntersectionKHR); + return 0; + case glslang::EOpRayQueryConfirmIntersection: + builder.createNoResultOp(spv::OpRayQueryConfirmIntersectionKHR); + return 0; case glslang::EOpBeginInvocationInterlock: builder.createNoResultOp(spv::OpBeginInvocationInterlockEXT); return 0; @@ -7722,7 +8697,7 @@ spv::Id TGlslangToSpvTraverser::getSymbolId(const glslang::TIntermSymbol* symbol // it was not found, create it spv::BuiltIn builtIn = TranslateBuiltInDecoration(symbol->getQualifier().builtIn, false); - auto forcedType = getForcedType(builtIn, symbol->getType()); + auto forcedType = getForcedType(symbol->getQualifier().builtIn, symbol->getType()); id = createSpvVariable(symbol, forcedType.first); symbolValues[symbol->getId()] = id; if (forcedType.second != spv::NoType) @@ -7785,16 +8760,15 @@ spv::Id TGlslangToSpvTraverser::getSymbolId(const glslang::TIntermSymbol* symbol } #ifndef GLSLANG_WEB - if (symbol->getType().isImage()) { + // Subgroup builtins which have input storage class are volatile for ray tracing stages. + if (symbol->getType().isImage() || symbol->getQualifier().isPipeInput()) { std::vector<spv::Decoration> memory; - TranslateMemoryDecoration(symbol->getType().getQualifier(), memory, glslangIntermediate->usingVulkanMemoryModel()); + TranslateMemoryDecoration(symbol->getType().getQualifier(), memory, + glslangIntermediate->usingVulkanMemoryModel()); for (unsigned int i = 0; i < memory.size(); ++i) builder.addDecoration(id, memory[i]); } - // nonuniform - builder.addDecoration(id, TranslateNonUniformDecoration(symbol->getType().getQualifier())); - if (builtIn == spv::BuiltInSampleMask) { spv::Decoration decoration; // GL_NV_sample_mask_override_coverage extension @@ -7841,7 +8815,50 @@ spv::Id TGlslangToSpvTraverser::getSymbolId(const glslang::TIntermSymbol* symbol } if (symbol->isReference()) { - builder.addDecoration(id, symbol->getType().getQualifier().restrict ? spv::DecorationRestrictPointerEXT : spv::DecorationAliasedPointerEXT); + builder.addDecoration(id, symbol->getType().getQualifier().restrict ? + spv::DecorationRestrictPointerEXT : spv::DecorationAliasedPointerEXT); + } + + // + // Add SPIR-V decorations for structure (GL_EXT_spirv_intrinsics) + // + if (symbol->getType().getQualifier().hasSprivDecorate()) { + const glslang::TSpirvDecorate& spirvDecorate = symbol->getType().getQualifier().getSpirvDecorate(); + + // Add spirv_decorate + for (auto& decorate : spirvDecorate.decorates) { + if (!decorate.second.empty()) { + std::vector<unsigned> literals; + TranslateLiterals(decorate.second, literals); + builder.addDecoration(id, static_cast<spv::Decoration>(decorate.first), literals); + } + else + builder.addDecoration(id, static_cast<spv::Decoration>(decorate.first)); + } + + // Add spirv_decorate_id + for (auto& decorateId : spirvDecorate.decorateIds) { + std::vector<spv::Id> operandIds; + assert(!decorateId.second.empty()); + for (auto extraOperand : decorateId.second) { + int nextConst = 0; + spv::Id operandId = createSpvConstantFromConstUnionArray( + extraOperand->getType(), extraOperand->getConstArray(), nextConst, false); + operandIds.push_back(operandId); + } + builder.addDecoration(id, static_cast<spv::Decoration>(decorateId.first), operandIds); + } + + // Add spirv_decorate_string + for (auto& decorateString : spirvDecorate.decorateStrings) { + std::vector<const char*> strings; + assert(!decorateString.second.empty()); + for (auto extraOperand : decorateString.second) { + const char* string = extraOperand->getConstArray()[0].getSConst()->c_str(); + strings.push_back(string); + } + builder.addDecoration(id, static_cast<spv::Decoration>(decorateString.first), strings); + } } #endif @@ -7903,12 +8920,25 @@ spv::Id TGlslangToSpvTraverser::createSpvConstant(const glslang::TIntermTyped& n // hand off to the non-spec-constant path assert(node.getAsConstantUnion() != nullptr || node.getAsSymbolNode() != nullptr); int nextConst = 0; - return createSpvConstantFromConstUnionArray(node.getType(), node.getAsConstantUnion() ? node.getAsConstantUnion()->getConstArray() : node.getAsSymbolNode()->getConstArray(), - nextConst, false); + return createSpvConstantFromConstUnionArray(node.getType(), node.getAsConstantUnion() ? + node.getAsConstantUnion()->getConstArray() : node.getAsSymbolNode()->getConstArray(), + nextConst, false); } // We now know we have a specialization constant to build + // Extra capabilities may be needed. + if (node.getType().contains8BitInt()) + builder.addCapability(spv::CapabilityInt8); + if (node.getType().contains16BitFloat()) + builder.addCapability(spv::CapabilityFloat16); + if (node.getType().contains16BitInt()) + builder.addCapability(spv::CapabilityInt16); + if (node.getType().contains64BitInt()) + builder.addCapability(spv::CapabilityInt64); + if (node.getType().containsDouble()) + builder.addCapability(spv::CapabilityFloat64); + // gl_WorkGroupSize is a special case until the front-end handles hierarchical specialization constants, // even then, it's specialization ids are handled by special case syntax in GLSL: layout(local_size_x = ... if (node.getType().getQualifier().builtIn == glslang::EbvWorkGroupSize) { @@ -7957,7 +8987,8 @@ spv::Id TGlslangToSpvTraverser::createSpvConstant(const glslang::TIntermTyped& n // If there are not enough elements present in 'consts', 0 will be substituted; // an empty 'consts' can be used to create a fully zeroed SPIR-V constant. // -spv::Id TGlslangToSpvTraverser::createSpvConstantFromConstUnionArray(const glslang::TType& glslangType, const glslang::TConstUnionArray& consts, int& nextConst, bool specConstant) +spv::Id TGlslangToSpvTraverser::createSpvConstantFromConstUnionArray(const glslang::TType& glslangType, + const glslang::TConstUnionArray& consts, int& nextConst, bool specConstant) { // vector of constants for SPIR-V std::vector<spv::Id> spvConsts; @@ -8075,6 +9106,9 @@ spv::Id TGlslangToSpvTraverser::createSpvConstantFromConstUnionArray(const glsla scalar = builder.createUnaryOp(spv::OpBitcast, typeId, scalar); break; #endif + case glslang::EbtString: + scalar = builder.getStringId(consts[nextConst].getSConst()->c_str()); + break; default: assert(0); break; @@ -8175,7 +9209,8 @@ bool TGlslangToSpvTraverser::isTrivial(const glslang::TIntermTyped* node) // Emit short-circuiting code, where 'right' is never evaluated unless // the left side is true (for &&) or false (for ||). -spv::Id TGlslangToSpvTraverser::createShortCircuit(glslang::TOperator op, glslang::TIntermTyped& left, glslang::TIntermTyped& right) +spv::Id TGlslangToSpvTraverser::createShortCircuit(glslang::TOperator op, glslang::TIntermTyped& left, + glslang::TIntermTyped& right) { spv::Id boolTypeId = builder.makeBoolType(); @@ -8259,7 +9294,9 @@ int GetSpirvGeneratorVersion() // return 6; // revert version 5 change, which makes a different (new) kind of incorrect code, // versions 4 and 6 each generate OpArrayLength as it has long been done // return 7; // GLSL volatile keyword maps to both SPIR-V decorations Volatile and Coherent - return 8; // switch to new dead block eliminator; use OpUnreachable + // return 8; // switch to new dead block eliminator; use OpUnreachable + // return 9; // don't include opaque function parameters in OpEntryPoint global's operand list + return 10; // Generate OpFUnordNotEqual for != comparisons } // Write SPIR-V out to a binary file @@ -8279,14 +9316,15 @@ void OutputSpvBin(const std::vector<unsigned int>& spirv, const char* baseName) // Write SPIR-V out to a text file with 32-bit hexadecimal words void OutputSpvHex(const std::vector<unsigned int>& spirv, const char* baseName, const char* varName) { -#ifndef GLSLANG_WEB +#if !defined(GLSLANG_WEB) && !defined(GLSLANG_ANGLE) std::ofstream out; out.open(baseName, std::ios::binary | std::ios::out); if (out.fail()) printf("ERROR: Failed to open file: %s\n", baseName); - out << "\t// " << - GetSpirvGeneratorVersion() << "." << GLSLANG_MINOR_VERSION << "." << GLSLANG_PATCH_LEVEL << - std::endl; + out << "\t// " << + GetSpirvGeneratorVersion() << + GLSLANG_VERSION_MAJOR << "." << GLSLANG_VERSION_MINOR << "." << GLSLANG_VERSION_PATCH << + GLSLANG_VERSION_FLAVOR << std::endl; if (varName != nullptr) { out << "\t #pragma once" << std::endl; out << "const uint32_t " << varName << "[] = {" << std::endl; @@ -8305,6 +9343,7 @@ void OutputSpvHex(const std::vector<unsigned int>& spirv, const char* baseName, } if (varName != nullptr) { out << "};"; + out << std::endl; } out.close(); #endif @@ -8342,10 +9381,14 @@ void GlslangToSpv(const TIntermediate& intermediate, std::vector<unsigned int>& // If from HLSL, run spirv-opt to "legalize" the SPIR-V for Vulkan // eg. forward and remove memory writes of opaque types. bool prelegalization = intermediate.getSource() == EShSourceHlsl; - if ((intermediate.getSource() == EShSourceHlsl || options->optimizeSize) && !options->disableOptimizer) { - SpirvToolsLegalize(intermediate, spirv, logger, options); + if ((prelegalization || options->optimizeSize) && !options->disableOptimizer) { + SpirvToolsTransform(intermediate, spirv, logger, options); prelegalization = false; } + else if (options->stripDebugInfo) { + // Strip debug info even if optimization is disabled. + SpirvToolsStripDebugInfo(intermediate, spirv, logger); + } if (options->validate) SpirvToolsValidate(intermediate, spirv, logger, prelegalization); diff --git a/thirdparty/glslang/SPIRV/GlslangToSpv.h b/thirdparty/glslang/SPIRV/GlslangToSpv.h index 3907be43b7..3907be43b7 100755..100644 --- a/thirdparty/glslang/SPIRV/GlslangToSpv.h +++ b/thirdparty/glslang/SPIRV/GlslangToSpv.h diff --git a/thirdparty/glslang/SPIRV/Logger.cpp b/thirdparty/glslang/SPIRV/Logger.cpp index 7ea0c6342b..cdc8469c44 100644 --- a/thirdparty/glslang/SPIRV/Logger.cpp +++ b/thirdparty/glslang/SPIRV/Logger.cpp @@ -69,4 +69,4 @@ std::string SpvBuildLogger::getAllMessages() const { } // end spv namespace -#endif
\ No newline at end of file +#endif diff --git a/thirdparty/glslang/SPIRV/NonSemanticDebugPrintf.h b/thirdparty/glslang/SPIRV/NonSemanticDebugPrintf.h new file mode 100644 index 0000000000..83796d75e5 --- /dev/null +++ b/thirdparty/glslang/SPIRV/NonSemanticDebugPrintf.h @@ -0,0 +1,50 @@ +// Copyright (c) 2020 The Khronos Group Inc. +// +// Permission is hereby granted, free of charge, to any person obtaining a +// copy of this software and/or associated documentation files (the +// "Materials"), to deal in the Materials without restriction, including +// without limitation the rights to use, copy, modify, merge, publish, +// distribute, sublicense, and/or sell copies of the Materials, and to +// permit persons to whom the Materials are furnished to do so, subject to +// the following conditions: +// +// The above copyright notice and this permission notice shall be included +// in all copies or substantial portions of the Materials. +// +// MODIFICATIONS TO THIS FILE MAY MEAN IT NO LONGER ACCURATELY REFLECTS +// KHRONOS STANDARDS. THE UNMODIFIED, NORMATIVE VERSIONS OF KHRONOS +// SPECIFICATIONS AND HEADER INFORMATION ARE LOCATED AT +// https://www.khronos.org/registry/ +// +// THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, +// EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF +// MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. +// IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY +// CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, +// TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE +// MATERIALS OR THE USE OR OTHER DEALINGS IN THE MATERIALS. +// + +#ifndef SPIRV_UNIFIED1_NonSemanticDebugPrintf_H_ +#define SPIRV_UNIFIED1_NonSemanticDebugPrintf_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +enum { + NonSemanticDebugPrintfRevision = 1, + NonSemanticDebugPrintfRevision_BitWidthPadding = 0x7fffffff +}; + +enum NonSemanticDebugPrintfInstructions { + NonSemanticDebugPrintfDebugPrintf = 1, + NonSemanticDebugPrintfInstructionsMax = 0x7fffffff +}; + + +#ifdef __cplusplus +} +#endif + +#endif // SPIRV_UNIFIED1_NonSemanticDebugPrintf_H_ diff --git a/thirdparty/glslang/SPIRV/SPVRemapper.cpp b/thirdparty/glslang/SPIRV/SPVRemapper.cpp index fd0bb8950c..56d6d5d4a5 100644 --- a/thirdparty/glslang/SPIRV/SPVRemapper.cpp +++ b/thirdparty/glslang/SPIRV/SPVRemapper.cpp @@ -544,6 +544,9 @@ namespace spv { // Extended instructions: currently, assume everything is an ID. // TODO: add whatever data we need for exceptions to that if (opCode == spv::OpExtInst) { + + idFn(asId(word)); // Instruction set is an ID that also needs to be mapped + word += 2; // instruction set, and instruction from set numOperands -= 2; @@ -625,6 +628,9 @@ namespace spv { break; } + case spv::OperandVariableLiteralStrings: + return nextInst; + // Execution mode might have extra literal operands. Skip them. case spv::OperandExecutionMode: return nextInst; @@ -827,7 +833,15 @@ namespace spv { [&](spv::Id& id) { if (thisOpCode != spv::OpNop) { ++idCounter; - const std::uint32_t hashval = opCounter[thisOpCode] * thisOpCode * 50047 + idCounter + fnId * 117; + const std::uint32_t hashval = + // Explicitly cast operands to unsigned int to avoid integer + // promotion to signed int followed by integer overflow, + // which would result in undefined behavior. + static_cast<unsigned int>(opCounter[thisOpCode]) + * thisOpCode + * 50047 + + idCounter + + static_cast<unsigned int>(fnId) * 117; if (isOldIdUnmapped(id)) localId(id, nextUnusedId(hashval % softTypeIdLimit + firstMappedID)); diff --git a/thirdparty/glslang/SPIRV/SpvBuilder.cpp b/thirdparty/glslang/SPIRV/SpvBuilder.cpp index bd208952e0..e83306ebcb 100644 --- a/thirdparty/glslang/SPIRV/SpvBuilder.cpp +++ b/thirdparty/glslang/SPIRV/SpvBuilder.cpp @@ -1,6 +1,7 @@ // // Copyright (C) 2014-2015 LunarG, Inc. // Copyright (C) 2015-2018 Google, Inc. +// Modifications Copyright (C) 2020 Advanced Micro Devices, Inc. All rights reserved. // // All rights reserved. // @@ -496,7 +497,8 @@ Id Builder::makeFunctionType(Id returnType, const std::vector<Id>& paramTypes) return type->getResultId(); } -Id Builder::makeImageType(Id sampledType, Dim dim, bool depth, bool arrayed, bool ms, unsigned sampled, ImageFormat format) +Id Builder::makeImageType(Id sampledType, Dim dim, bool depth, bool arrayed, bool ms, unsigned sampled, + ImageFormat format) { assert(sampled == 1 || sampled == 2); @@ -601,16 +603,31 @@ Id Builder::makeSampledImageType(Id imageType) } #ifndef GLSLANG_WEB -Id Builder::makeAccelerationStructureNVType() +Id Builder::makeAccelerationStructureType() { Instruction *type; - if (groupedTypes[OpTypeAccelerationStructureNV].size() == 0) { - type = new Instruction(getUniqueId(), NoType, OpTypeAccelerationStructureNV); - groupedTypes[OpTypeAccelerationStructureNV].push_back(type); + if (groupedTypes[OpTypeAccelerationStructureKHR].size() == 0) { + type = new Instruction(getUniqueId(), NoType, OpTypeAccelerationStructureKHR); + groupedTypes[OpTypeAccelerationStructureKHR].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); module.mapInstruction(type); } else { - type = groupedTypes[OpTypeAccelerationStructureNV].back(); + type = groupedTypes[OpTypeAccelerationStructureKHR].back(); + } + + return type->getResultId(); +} + +Id Builder::makeRayQueryType() +{ + Instruction *type; + if (groupedTypes[OpTypeRayQueryKHR].size() == 0) { + type = new Instruction(getUniqueId(), NoType, OpTypeRayQueryKHR); + groupedTypes[OpTypeRayQueryKHR].push_back(type); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(type)); + module.mapInstruction(type); + } else { + type = groupedTypes[OpTypeRayQueryKHR].back(); } return type->getResultId(); @@ -726,6 +743,26 @@ Id Builder::getContainedTypeId(Id typeId, int member) const } } +// Figure out the final resulting type of the access chain. +Id Builder::getResultingAccessChainType() const +{ + assert(accessChain.base != NoResult); + Id typeId = getTypeId(accessChain.base); + + assert(isPointerType(typeId)); + typeId = getContainedTypeId(typeId); + + for (int i = 0; i < (int)accessChain.indexChain.size(); ++i) { + if (isStructType(typeId)) { + assert(isConstantScalar(accessChain.indexChain[i])); + typeId = getContainedTypeId(typeId, getConstantScalar(accessChain.indexChain[i])); + } else + typeId = getContainedTypeId(typeId, accessChain.indexChain[i]); + } + + return typeId; +} + // Return the immediately contained type of a given composite type. Id Builder::getContainedTypeId(Id typeId) const { @@ -852,6 +889,30 @@ bool Builder::isSpecConstantOpCode(Op opcode) const } } +Id Builder::makeNullConstant(Id typeId) +{ + Instruction* constant; + + // See if we already made it. + Id existing = NoResult; + for (int i = 0; i < (int)nullConstants.size(); ++i) { + constant = nullConstants[i]; + if (constant->getTypeId() == typeId) + existing = constant->getResultId(); + } + + if (existing != NoResult) + return existing; + + // Make it + Instruction* c = new Instruction(getUniqueId(), typeId, OpConstantNull); + constantsTypesGlobals.push_back(std::unique_ptr<Instruction>(c)); + nullConstants.push_back(c); + module.mapInstruction(c); + + return c->getResultId(); +} + Id Builder::makeBoolConstant(bool b, bool specConstant) { Id typeId = makeBoolType(); @@ -1166,6 +1227,28 @@ void Builder::addExecutionMode(Function* entryPoint, ExecutionMode mode, int val executionModes.push_back(std::unique_ptr<Instruction>(instr)); } +void Builder::addExecutionMode(Function* entryPoint, ExecutionMode mode, const std::vector<unsigned>& literals) +{ + Instruction* instr = new Instruction(OpExecutionMode); + instr->addIdOperand(entryPoint->getId()); + instr->addImmediateOperand(mode); + for (auto literal : literals) + instr->addImmediateOperand(literal); + + executionModes.push_back(std::unique_ptr<Instruction>(instr)); +} + +void Builder::addExecutionModeId(Function* entryPoint, ExecutionMode mode, const std::vector<Id>& operandIds) +{ + Instruction* instr = new Instruction(OpExecutionModeId); + instr->addIdOperand(entryPoint->getId()); + instr->addImmediateOperand(mode); + for (auto operandId : operandIds) + instr->addIdOperand(operandId); + + executionModes.push_back(std::unique_ptr<Instruction>(instr)); +} + void Builder::addName(Id id, const char* string) { Instruction* name = new Instruction(OpName); @@ -1204,7 +1287,7 @@ void Builder::addDecoration(Id id, Decoration decoration, const char* s) if (decoration == spv::DecorationMax) return; - Instruction* dec = new Instruction(OpDecorateStringGOOGLE); + Instruction* dec = new Instruction(OpDecorateString); dec->addIdOperand(id); dec->addImmediateOperand(decoration); dec->addStringOperand(s); @@ -1212,6 +1295,34 @@ void Builder::addDecoration(Id id, Decoration decoration, const char* s) decorations.push_back(std::unique_ptr<Instruction>(dec)); } +void Builder::addDecoration(Id id, Decoration decoration, const std::vector<unsigned>& literals) +{ + if (decoration == spv::DecorationMax) + return; + + Instruction* dec = new Instruction(OpDecorate); + dec->addIdOperand(id); + dec->addImmediateOperand(decoration); + for (auto literal : literals) + dec->addImmediateOperand(literal); + + decorations.push_back(std::unique_ptr<Instruction>(dec)); +} + +void Builder::addDecoration(Id id, Decoration decoration, const std::vector<const char*>& strings) +{ + if (decoration == spv::DecorationMax) + return; + + Instruction* dec = new Instruction(OpDecorateString); + dec->addIdOperand(id); + dec->addImmediateOperand(decoration); + for (auto string : strings) + dec->addStringOperand(string); + + decorations.push_back(std::unique_ptr<Instruction>(dec)); +} + void Builder::addDecorationId(Id id, Decoration decoration, Id idDecoration) { if (decoration == spv::DecorationMax) @@ -1225,6 +1336,21 @@ void Builder::addDecorationId(Id id, Decoration decoration, Id idDecoration) decorations.push_back(std::unique_ptr<Instruction>(dec)); } +void Builder::addDecorationId(Id id, Decoration decoration, const std::vector<Id>& operandIds) +{ + if(decoration == spv::DecorationMax) + return; + + Instruction* dec = new Instruction(OpDecorateId); + dec->addIdOperand(id); + dec->addImmediateOperand(decoration); + + for (auto operandId : operandIds) + dec->addIdOperand(operandId); + + decorations.push_back(std::unique_ptr<Instruction>(dec)); +} + void Builder::addMemberDecoration(Id id, unsigned int member, Decoration decoration, int num) { if (decoration == spv::DecorationMax) @@ -1254,6 +1380,36 @@ void Builder::addMemberDecoration(Id id, unsigned int member, Decoration decorat decorations.push_back(std::unique_ptr<Instruction>(dec)); } +void Builder::addMemberDecoration(Id id, unsigned int member, Decoration decoration, const std::vector<unsigned>& literals) +{ + if (decoration == spv::DecorationMax) + return; + + Instruction* dec = new Instruction(OpMemberDecorate); + dec->addIdOperand(id); + dec->addImmediateOperand(member); + dec->addImmediateOperand(decoration); + for (auto literal : literals) + dec->addImmediateOperand(literal); + + decorations.push_back(std::unique_ptr<Instruction>(dec)); +} + +void Builder::addMemberDecoration(Id id, unsigned int member, Decoration decoration, const std::vector<const char*>& strings) +{ + if (decoration == spv::DecorationMax) + return; + + Instruction* dec = new Instruction(OpMemberDecorateString); + dec->addIdOperand(id); + dec->addImmediateOperand(member); + dec->addImmediateOperand(decoration); + for (auto string : strings) + dec->addStringOperand(string); + + decorations.push_back(std::unique_ptr<Instruction>(dec)); +} + // Comments in header Function* Builder::makeEntryPoint(const char* entryPoint) { @@ -1270,7 +1426,8 @@ Function* Builder::makeEntryPoint(const char* entryPoint) // Comments in header Function* Builder::makeFunctionEntry(Decoration precision, Id returnType, const char* name, - const std::vector<Id>& paramTypes, const std::vector<std::vector<Decoration>>& decorations, Block **entry) + const std::vector<Id>& paramTypes, + const std::vector<std::vector<Decoration>>& decorations, Block **entry) { // Make the function and initial instructions in it Id typeId = makeFunctionType(returnType, paramTypes); @@ -1279,9 +1436,12 @@ Function* Builder::makeFunctionEntry(Decoration precision, Id returnType, const // Set up the precisions setPrecision(function->getId(), precision); + function->setReturnPrecision(precision); for (unsigned p = 0; p < (unsigned)decorations.size(); ++p) { - for (int d = 0; d < (int)decorations[p].size(); ++d) + for (int d = 0; d < (int)decorations[p].size(); ++d) { addDecoration(firstParamId + p, decorations[p][d]); + function->addParamPrecision(p, decorations[p][d]); + } } // CFG @@ -1331,14 +1491,14 @@ void Builder::leaveFunction() } // Comments in header -void Builder::makeDiscard() +void Builder::makeStatementTerminator(spv::Op opcode, const char *name) { - buildPoint->addInstruction(std::unique_ptr<Instruction>(new Instruction(OpKill))); - createAndSetNoPredecessorBlock("post-discard"); + buildPoint->addInstruction(std::unique_ptr<Instruction>(new Instruction(opcode))); + createAndSetNoPredecessorBlock(name); } // Comments in header -Id Builder::createVariable(StorageClass storageClass, Id type, const char* name, Id initializer) +Id Builder::createVariable(Decoration precision, StorageClass storageClass, Id type, const char* name, Id initializer) { Id pointerType = makePointer(storageClass, type); Instruction* inst = new Instruction(getUniqueId(), pointerType, OpVariable); @@ -1360,6 +1520,7 @@ Id Builder::createVariable(StorageClass storageClass, Id type, const char* name, if (name) addName(inst->getResultId(), name); + setPrecision(inst->getResultId(), precision); return inst->getResultId(); } @@ -1373,7 +1534,8 @@ Id Builder::createUndefined(Id type) } // av/vis/nonprivate are unnecessary and illegal for some storage classes. -spv::MemoryAccessMask Builder::sanitizeMemoryAccessForStorageClass(spv::MemoryAccessMask memoryAccess, StorageClass sc) const +spv::MemoryAccessMask Builder::sanitizeMemoryAccessForStorageClass(spv::MemoryAccessMask memoryAccess, StorageClass sc) + const { switch (sc) { case spv::StorageClassUniform: @@ -1392,7 +1554,8 @@ spv::MemoryAccessMask Builder::sanitizeMemoryAccessForStorageClass(spv::MemoryAc } // Comments in header -void Builder::createStore(Id rValue, Id lValue, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) +void Builder::createStore(Id rValue, Id lValue, spv::MemoryAccessMask memoryAccess, spv::Scope scope, + unsigned int alignment) { Instruction* store = new Instruction(OpStore); store->addIdOperand(lValue); @@ -1414,7 +1577,8 @@ void Builder::createStore(Id rValue, Id lValue, spv::MemoryAccessMask memoryAcce } // Comments in header -Id Builder::createLoad(Id lValue, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) +Id Builder::createLoad(Id lValue, spv::Decoration precision, spv::MemoryAccessMask memoryAccess, + spv::Scope scope, unsigned int alignment) { Instruction* load = new Instruction(getUniqueId(), getDerefTypeId(lValue), OpLoad); load->addIdOperand(lValue); @@ -1432,6 +1596,7 @@ Id Builder::createLoad(Id lValue, spv::MemoryAccessMask memoryAccess, spv::Scope } buildPoint->addInstruction(std::unique_ptr<Instruction>(load)); + setPrecision(load->getResultId(), precision); return load->getResultId(); } @@ -1440,16 +1605,7 @@ Id Builder::createLoad(Id lValue, spv::MemoryAccessMask memoryAccess, spv::Scope Id Builder::createAccessChain(StorageClass storageClass, Id base, const std::vector<Id>& offsets) { // Figure out the final resulting type. - spv::Id typeId = getTypeId(base); - assert(isPointerType(typeId) && offsets.size() > 0); - typeId = getContainedTypeId(typeId); - for (int i = 0; i < (int)offsets.size(); ++i) { - if (isStructType(typeId)) { - assert(isConstantScalar(offsets[i])); - typeId = getContainedTypeId(typeId, getConstantScalar(offsets[i])); - } else - typeId = getContainedTypeId(typeId, offsets[i]); - } + Id typeId = getResultingAccessChainType(); typeId = makePointer(storageClass, typeId); // Make the instruction @@ -1495,7 +1651,8 @@ Id Builder::createCompositeExtract(Id composite, Id typeId, unsigned index) // Generate code for spec constants if in spec constant operation // generation mode. if (generatingOpCodeForSpecConst) { - return createSpecConstantOp(OpCompositeExtract, typeId, std::vector<Id>(1, composite), std::vector<Id>(1, index)); + return createSpecConstantOp(OpCompositeExtract, typeId, std::vector<Id>(1, composite), + std::vector<Id>(1, index)); } Instruction* extract = new Instruction(getUniqueId(), typeId, OpCompositeExtract); extract->addIdOperand(composite); @@ -1697,7 +1854,8 @@ Id Builder::createOp(Op opCode, Id typeId, const std::vector<IdImmediate>& opera return op->getResultId(); } -Id Builder::createSpecConstantOp(Op opCode, Id typeId, const std::vector<Id>& operands, const std::vector<unsigned>& literals) +Id Builder::createSpecConstantOp(Op opCode, Id typeId, const std::vector<Id>& operands, + const std::vector<unsigned>& literals) { Instruction* op = new Instruction(getUniqueId(), typeId, OpSpecConstantOp); op->addImmediateOperand((unsigned) opCode); @@ -2144,7 +2302,7 @@ Id Builder::createCompositeCompare(Decoration precision, Id value1, Id value2, b Op op; switch (getMostBasicTypeClass(valueType)) { case OpTypeFloat: - op = equal ? OpFOrdEqual : OpFOrdNotEqual; + op = equal ? OpFOrdEqual : OpFUnordNotEqual; break; case OpTypeInt: default: @@ -2187,7 +2345,8 @@ Id Builder::createCompositeCompare(Decoration precision, Id value1, Id value2, b if (constituent == 0) resultId = subResultId; else - resultId = setPrecision(createBinOp(equal ? OpLogicalAnd : OpLogicalOr, boolType, resultId, subResultId), precision); + resultId = setPrecision(createBinOp(equal ? OpLogicalAnd : OpLogicalOr, boolType, resultId, subResultId), + precision); } return resultId; @@ -2196,7 +2355,8 @@ Id Builder::createCompositeCompare(Decoration precision, Id value1, Id value2, b // OpCompositeConstruct Id Builder::createCompositeConstruct(Id typeId, const std::vector<Id>& constituents) { - assert(isAggregateType(typeId) || (getNumTypeConstituents(typeId) > 1 && getNumTypeConstituents(typeId) == (int)constituents.size())); + assert(isAggregateType(typeId) || (getNumTypeConstituents(typeId) > 1 && + getNumTypeConstituents(typeId) == (int)constituents.size())); if (generatingOpCodeForSpecConst) { // Sometime, even in spec-constant-op mode, the constant composite to be @@ -2394,7 +2554,7 @@ Id Builder::createMatrixConstructor(Decoration precision, const std::vector<Id>& int row = 0; int col = 0; - for (int arg = 0; arg < (int)sources.size(); ++arg) { + for (int arg = 0; arg < (int)sources.size() && col < numCols; ++arg) { Id argComp = sources[arg]; for (int comp = 0; comp < getNumComponents(sources[arg]); ++comp) { if (getNumComponents(sources[arg]) > 1) { @@ -2406,6 +2566,10 @@ Id Builder::createMatrixConstructor(Decoration precision, const std::vector<Id>& row = 0; col++; } + if (col == numCols) { + // If more components are provided than fit the matrix, discard the rest. + break; + } } } } @@ -2609,7 +2773,8 @@ void Builder::clearAccessChain() } // Comments in header -void Builder::accessChainPushSwizzle(std::vector<unsigned>& swizzle, Id preSwizzleBaseType, AccessChain::CoherentFlags coherentFlags, unsigned int alignment) +void Builder::accessChainPushSwizzle(std::vector<unsigned>& swizzle, Id preSwizzleBaseType, + AccessChain::CoherentFlags coherentFlags, unsigned int alignment) { accessChain.coherentFlags |= coherentFlags; accessChain.alignment |= alignment; @@ -2635,35 +2800,70 @@ void Builder::accessChainPushSwizzle(std::vector<unsigned>& swizzle, Id preSwizz } // Comments in header -void Builder::accessChainStore(Id rvalue, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) +void Builder::accessChainStore(Id rvalue, Decoration nonUniform, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) { assert(accessChain.isRValue == false); transferAccessChainSwizzle(true); - Id base = collapseAccessChain(); - Id source = rvalue; - // dynamic component should be gone - assert(accessChain.component == NoResult); + // If a swizzle exists and is not full and is not dynamic, then the swizzle will be broken into individual stores. + if (accessChain.swizzle.size() > 0 && + getNumTypeComponents(getResultingAccessChainType()) != (int)accessChain.swizzle.size() && + accessChain.component == NoResult) { + for (unsigned int i = 0; i < accessChain.swizzle.size(); ++i) { + accessChain.indexChain.push_back(makeUintConstant(accessChain.swizzle[i])); + accessChain.instr = NoResult; - // If swizzle still exists, it is out-of-order or not full, we must load the target vector, - // extract and insert elements to perform writeMask and/or swizzle. - if (accessChain.swizzle.size() > 0) { - Id tempBaseId = createLoad(base); - source = createLvalueSwizzle(getTypeId(tempBaseId), tempBaseId, source, accessChain.swizzle); - } + Id base = collapseAccessChain(); + addDecoration(base, nonUniform); + + accessChain.indexChain.pop_back(); + accessChain.instr = NoResult; - // take LSB of alignment - alignment = alignment & ~(alignment & (alignment-1)); - if (getStorageClass(base) == StorageClassPhysicalStorageBufferEXT) { - memoryAccess = (spv::MemoryAccessMask)(memoryAccess | spv::MemoryAccessAlignedMask); + // dynamic component should be gone + assert(accessChain.component == NoResult); + + Id source = createCompositeExtract(rvalue, getContainedTypeId(getTypeId(rvalue)), i); + + // take LSB of alignment + alignment = alignment & ~(alignment & (alignment-1)); + if (getStorageClass(base) == StorageClassPhysicalStorageBufferEXT) { + memoryAccess = (spv::MemoryAccessMask)(memoryAccess | spv::MemoryAccessAlignedMask); + } + + createStore(source, base, memoryAccess, scope, alignment); + } } + else { + Id base = collapseAccessChain(); + addDecoration(base, nonUniform); + + Id source = rvalue; + + // dynamic component should be gone + assert(accessChain.component == NoResult); - createStore(source, base, memoryAccess, scope, alignment); + // If swizzle still exists, it may be out-of-order, we must load the target vector, + // extract and insert elements to perform writeMask and/or swizzle. + if (accessChain.swizzle.size() > 0) { + Id tempBaseId = createLoad(base, spv::NoPrecision); + source = createLvalueSwizzle(getTypeId(tempBaseId), tempBaseId, source, accessChain.swizzle); + } + + // take LSB of alignment + alignment = alignment & ~(alignment & (alignment-1)); + if (getStorageClass(base) == StorageClassPhysicalStorageBufferEXT) { + memoryAccess = (spv::MemoryAccessMask)(memoryAccess | spv::MemoryAccessAlignedMask); + } + + createStore(source, base, memoryAccess, scope, alignment); + } } // Comments in header -Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resultType, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) +Id Builder::accessChainLoad(Decoration precision, Decoration l_nonUniform, + Decoration r_nonUniform, Id resultType, spv::MemoryAccessMask memoryAccess, + spv::Scope scope, unsigned int alignment) { Id id; @@ -2687,17 +2887,19 @@ Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resu if (constant) { id = createCompositeExtract(accessChain.base, swizzleBase, indexes); + setPrecision(id, precision); } else { Id lValue = NoResult; - if (spvVersion >= Spv_1_4) { + if (spvVersion >= Spv_1_4 && isValidInitializer(accessChain.base)) { // make a new function variable for this r-value, using an initializer, // and mark it as NonWritable so that downstream it can be detected as a lookup // table - lValue = createVariable(StorageClassFunction, getTypeId(accessChain.base), "indexable", - accessChain.base); + lValue = createVariable(NoPrecision, StorageClassFunction, getTypeId(accessChain.base), + "indexable", accessChain.base); addDecoration(lValue, DecorationNonWritable); } else { - lValue = createVariable(StorageClassFunction, getTypeId(accessChain.base), "indexable"); + lValue = createVariable(NoPrecision, StorageClassFunction, getTypeId(accessChain.base), + "indexable"); // store into it createStore(accessChain.base, lValue); } @@ -2706,9 +2908,8 @@ Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resu accessChain.isRValue = false; // load through the access chain - id = createLoad(collapseAccessChain()); + id = createLoad(collapseAccessChain(), precision); } - setPrecision(id, precision); } else id = accessChain.base; // no precision, it was set when this was defined } else { @@ -2721,9 +2922,14 @@ Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resu } // load through the access chain - id = createLoad(collapseAccessChain(), memoryAccess, scope, alignment); - setPrecision(id, precision); - addDecoration(id, nonUniform); + id = collapseAccessChain(); + // Apply nonuniform both to the access chain and the loaded value. + // Buffer accesses need the access chain decorated, and this is where + // loaded image types get decorated. TODO: This should maybe move to + // createImageTextureFunctionCall. + addDecoration(id, l_nonUniform); + id = createLoad(id, precision, memoryAccess, scope, alignment); + addDecoration(id, r_nonUniform); } // Done, unless there are swizzles to do @@ -2744,7 +2950,7 @@ Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resu if (accessChain.component != NoResult) id = setPrecision(createVectorExtractDynamic(id, resultType, accessChain.component), precision); - addDecoration(id, nonUniform); + addDecoration(id, r_nonUniform); return id; } @@ -3075,7 +3281,8 @@ void Builder::dumpSourceInstructions(std::vector<unsigned int>& out) const dumpSourceInstructions(iItr->first, *iItr->second, out); } -void Builder::dumpInstructions(std::vector<unsigned int>& out, const std::vector<std::unique_ptr<Instruction> >& instructions) const +void Builder::dumpInstructions(std::vector<unsigned int>& out, + const std::vector<std::unique_ptr<Instruction> >& instructions) const { for (int i = 0; i < (int)instructions.size(); ++i) { instructions[i]->dump(out); diff --git a/thirdparty/glslang/SPIRV/SpvBuilder.h b/thirdparty/glslang/SPIRV/SpvBuilder.h index 31fee975fc..251b9ee823 100644 --- a/thirdparty/glslang/SPIRV/SpvBuilder.h +++ b/thirdparty/glslang/SPIRV/SpvBuilder.h @@ -1,7 +1,8 @@ // // Copyright (C) 2014-2015 LunarG, Inc. -// Copyright (C) 2015-2018 Google, Inc. +// Copyright (C) 2015-2020 Google, Inc. // Copyright (C) 2017 ARM Limited. +// Modifications Copyright (C) 2020 Advanced Micro Devices, Inc. All rights reserved. // // All rights reserved. // @@ -94,6 +95,7 @@ public: const char* file_c_str = str.c_str(); fileString->addStringOperand(file_c_str); strings.push_back(std::unique_ptr<Instruction>(fileString)); + module.mapInstruction(fileString); stringIds[file_c_str] = strId; return strId; } @@ -181,7 +183,9 @@ public: Id makeCooperativeMatrixType(Id component, Id scope, Id rows, Id cols); // accelerationStructureNV type - Id makeAccelerationStructureNVType(); + Id makeAccelerationStructureType(); + // rayQueryEXT type + Id makeRayQueryType(); // For querying about types. Id getTypeId(Id resultId) const { return module.getTypeId(resultId); } @@ -196,7 +200,9 @@ public: Id getContainedTypeId(Id typeId) const; Id getContainedTypeId(Id typeId, int) const; StorageClass getTypeStorageClass(Id typeId) const { return module.getStorageClass(typeId); } - ImageFormat getImageTypeFormat(Id typeId) const { return (ImageFormat)module.getInstruction(typeId)->getImmediateOperand(6); } + ImageFormat getImageTypeFormat(Id typeId) const + { return (ImageFormat)module.getInstruction(typeId)->getImmediateOperand(6); } + Id getResultingAccessChainType() const; bool isPointer(Id resultId) const { return isPointerType(getTypeId(resultId)); } bool isScalar(Id resultId) const { return isScalarType(getTypeId(resultId)); } @@ -206,12 +212,17 @@ public: bool isAggregate(Id resultId) const { return isAggregateType(getTypeId(resultId)); } bool isSampledImage(Id resultId) const { return isSampledImageType(getTypeId(resultId)); } - bool isBoolType(Id typeId) { return groupedTypes[OpTypeBool].size() > 0 && typeId == groupedTypes[OpTypeBool].back()->getResultId(); } - bool isIntType(Id typeId) const { return getTypeClass(typeId) == OpTypeInt && module.getInstruction(typeId)->getImmediateOperand(1) != 0; } - bool isUintType(Id typeId) const { return getTypeClass(typeId) == OpTypeInt && module.getInstruction(typeId)->getImmediateOperand(1) == 0; } + bool isBoolType(Id typeId) + { return groupedTypes[OpTypeBool].size() > 0 && typeId == groupedTypes[OpTypeBool].back()->getResultId(); } + bool isIntType(Id typeId) const + { return getTypeClass(typeId) == OpTypeInt && module.getInstruction(typeId)->getImmediateOperand(1) != 0; } + bool isUintType(Id typeId) const + { return getTypeClass(typeId) == OpTypeInt && module.getInstruction(typeId)->getImmediateOperand(1) == 0; } bool isFloatType(Id typeId) const { return getTypeClass(typeId) == OpTypeFloat; } bool isPointerType(Id typeId) const { return getTypeClass(typeId) == OpTypePointer; } - bool isScalarType(Id typeId) const { return getTypeClass(typeId) == OpTypeFloat || getTypeClass(typeId) == OpTypeInt || getTypeClass(typeId) == OpTypeBool; } + bool isScalarType(Id typeId) const + { return getTypeClass(typeId) == OpTypeFloat || getTypeClass(typeId) == OpTypeInt || + getTypeClass(typeId) == OpTypeBool; } bool isVectorType(Id typeId) const { return getTypeClass(typeId) == OpTypeVector; } bool isMatrixType(Id typeId) const { return getTypeClass(typeId) == OpTypeMatrix; } bool isStructType(Id typeId) const { return getTypeClass(typeId) == OpTypeStruct; } @@ -221,7 +232,8 @@ public: #else bool isCooperativeMatrixType(Id typeId)const { return getTypeClass(typeId) == OpTypeCooperativeMatrixNV; } #endif - bool isAggregateType(Id typeId) const { return isArrayType(typeId) || isStructType(typeId) || isCooperativeMatrixType(typeId); } + bool isAggregateType(Id typeId) const + { return isArrayType(typeId) || isStructType(typeId) || isCooperativeMatrixType(typeId); } bool isImageType(Id typeId) const { return getTypeClass(typeId) == OpTypeImage; } bool isSamplerType(Id typeId) const { return getTypeClass(typeId) == OpTypeSampler; } bool isSampledImageType(Id typeId) const { return getTypeClass(typeId) == OpTypeSampledImage; } @@ -233,9 +245,17 @@ public: bool isConstant(Id resultId) const { return isConstantOpCode(getOpCode(resultId)); } bool isConstantScalar(Id resultId) const { return getOpCode(resultId) == OpConstant; } bool isSpecConstant(Id resultId) const { return isSpecConstantOpCode(getOpCode(resultId)); } - unsigned int getConstantScalar(Id resultId) const { return module.getInstruction(resultId)->getImmediateOperand(0); } + unsigned int getConstantScalar(Id resultId) const + { return module.getInstruction(resultId)->getImmediateOperand(0); } StorageClass getStorageClass(Id resultId) const { return getTypeStorageClass(getTypeId(resultId)); } + bool isVariableOpCode(Op opcode) const { return opcode == OpVariable; } + bool isVariable(Id resultId) const { return isVariableOpCode(getOpCode(resultId)); } + bool isGlobalStorage(Id resultId) const { return getStorageClass(resultId) != StorageClassFunction; } + bool isGlobalVariable(Id resultId) const { return isVariable(resultId) && isGlobalStorage(resultId); } + // See if a resultId is valid for use as an initializer. + bool isValidInitializer(Id resultId) const { return isConstant(resultId) || isGlobalVariable(resultId); } + int getScalarTypeWidth(Id typeId) const { Id scalarTypeId = getScalarTypeId(typeId); @@ -274,15 +294,24 @@ public: } // For making new constants (will return old constant if the requested one was already made). + Id makeNullConstant(Id typeId); Id makeBoolConstant(bool b, bool specConstant = false); - Id makeInt8Constant(int i, bool specConstant = false) { return makeIntConstant(makeIntType(8), (unsigned)i, specConstant); } - Id makeUint8Constant(unsigned u, bool specConstant = false) { return makeIntConstant(makeUintType(8), u, specConstant); } - Id makeInt16Constant(int i, bool specConstant = false) { return makeIntConstant(makeIntType(16), (unsigned)i, specConstant); } - Id makeUint16Constant(unsigned u, bool specConstant = false) { return makeIntConstant(makeUintType(16), u, specConstant); } - Id makeIntConstant(int i, bool specConstant = false) { return makeIntConstant(makeIntType(32), (unsigned)i, specConstant); } - Id makeUintConstant(unsigned u, bool specConstant = false) { return makeIntConstant(makeUintType(32), u, specConstant); } - Id makeInt64Constant(long long i, bool specConstant = false) { return makeInt64Constant(makeIntType(64), (unsigned long long)i, specConstant); } - Id makeUint64Constant(unsigned long long u, bool specConstant = false) { return makeInt64Constant(makeUintType(64), u, specConstant); } + Id makeInt8Constant(int i, bool specConstant = false) + { return makeIntConstant(makeIntType(8), (unsigned)i, specConstant); } + Id makeUint8Constant(unsigned u, bool specConstant = false) + { return makeIntConstant(makeUintType(8), u, specConstant); } + Id makeInt16Constant(int i, bool specConstant = false) + { return makeIntConstant(makeIntType(16), (unsigned)i, specConstant); } + Id makeUint16Constant(unsigned u, bool specConstant = false) + { return makeIntConstant(makeUintType(16), u, specConstant); } + Id makeIntConstant(int i, bool specConstant = false) + { return makeIntConstant(makeIntType(32), (unsigned)i, specConstant); } + Id makeUintConstant(unsigned u, bool specConstant = false) + { return makeIntConstant(makeUintType(32), u, specConstant); } + Id makeInt64Constant(long long i, bool specConstant = false) + { return makeInt64Constant(makeIntType(64), (unsigned long long)i, specConstant); } + Id makeUint64Constant(unsigned long long u, bool specConstant = false) + { return makeInt64Constant(makeUintType(64), u, specConstant); } Id makeFloatConstant(float f, bool specConstant = false); Id makeDoubleConstant(double d, bool specConstant = false); Id makeFloat16Constant(float f16, bool specConstant = false); @@ -294,13 +323,20 @@ public: // Methods for adding information outside the CFG. Instruction* addEntryPoint(ExecutionModel, Function*, const char* name); void addExecutionMode(Function*, ExecutionMode mode, int value1 = -1, int value2 = -1, int value3 = -1); + void addExecutionMode(Function*, ExecutionMode mode, const std::vector<unsigned>& literals); + void addExecutionModeId(Function*, ExecutionMode mode, const std::vector<Id>& operandIds); void addName(Id, const char* name); void addMemberName(Id, int member, const char* name); void addDecoration(Id, Decoration, int num = -1); void addDecoration(Id, Decoration, const char*); + void addDecoration(Id, Decoration, const std::vector<unsigned>& literals); + void addDecoration(Id, Decoration, const std::vector<const char*>& strings); void addDecorationId(Id id, Decoration, Id idDecoration); + void addDecorationId(Id id, Decoration, const std::vector<Id>& operandIds); void addMemberDecoration(Id, unsigned int member, Decoration, int num = -1); void addMemberDecoration(Id, unsigned int member, Decoration, const char*); + void addMemberDecoration(Id, unsigned int member, Decoration, const std::vector<unsigned>& literals); + void addMemberDecoration(Id, unsigned int member, Decoration, const std::vector<const char*>& strings); // At the end of what block do the next create*() instructions go? void setBuildPoint(Block* bp) { buildPoint = bp; } @@ -313,8 +349,8 @@ public: // Make a shader-style function, and create its entry block if entry is non-zero. // Return the function, pass back the entry. // The returned pointer is only valid for the lifetime of this builder. - Function* makeFunctionEntry(Decoration precision, Id returnType, const char* name, const std::vector<Id>& paramTypes, - const std::vector<std::vector<Decoration>>& precisions, Block **entry = 0); + Function* makeFunctionEntry(Decoration precision, Id returnType, const char* name, + const std::vector<Id>& paramTypes, const std::vector<std::vector<Decoration>>& precisions, Block **entry = 0); // Create a return. An 'implicit' return is one not appearing in the source // code. In the case of an implicit return, no post-return block is inserted. @@ -323,20 +359,25 @@ public: // Generate all the code needed to finish up a function. void leaveFunction(); - // Create a discard. - void makeDiscard(); + // Create block terminator instruction for certain statements like + // discard, terminate-invocation, terminateRayEXT, or ignoreIntersectionEXT + void makeStatementTerminator(spv::Op opcode, const char *name); // Create a global or function local or IO variable. - Id createVariable(StorageClass, Id type, const char* name = 0, Id initializer = NoResult); + Id createVariable(Decoration precision, StorageClass, Id type, const char* name = nullptr, + Id initializer = NoResult); // Create an intermediate with an undefined value. Id createUndefined(Id type); // Store into an Id and return the l-value - void createStore(Id rValue, Id lValue, spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0); + void createStore(Id rValue, Id lValue, spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, + spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0); // Load from an Id and return it - Id createLoad(Id lValue, spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0); + Id createLoad(Id lValue, spv::Decoration precision, + spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, + spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0); // Create an OpAccessChain instruction Id createAccessChain(StorageClass, Id base, const std::vector<Id>& offsets); @@ -495,7 +536,7 @@ public: // recursion stack can hold the memory for it. // void makeSwitch(Id condition, unsigned int control, int numSegments, const std::vector<int>& caseValues, - const std::vector<int>& valueToSegment, int defaultSegment, std::vector<Block*>& segmentBB); // return argument + const std::vector<int>& valueToSegment, int defaultSegment, std::vector<Block*>& segmentBB); // Add a branch to the innermost switch's merge block. void addSwitchBreak(); @@ -512,7 +553,7 @@ public: Block &head, &body, &merge, &continue_target; private: LoopBlocks(); - LoopBlocks& operator=(const LoopBlocks&); + LoopBlocks& operator=(const LoopBlocks&) = delete; }; // Start a new loop and prepare the builder to generate code for it. Until @@ -569,10 +610,13 @@ public: std::vector<Id> indexChain; Id instr; // cache the instruction that generates this access chain std::vector<unsigned> swizzle; // each std::vector element selects the next GLSL component number - Id component; // a dynamic component index, can coexist with a swizzle, done after the swizzle, NoResult if not present - Id preSwizzleBaseType; // dereferenced type, before swizzle or component is applied; NoType unless a swizzle or component is present + Id component; // a dynamic component index, can coexist with a swizzle, + // done after the swizzle, NoResult if not present + Id preSwizzleBaseType; // dereferenced type, before swizzle or component is applied; + // NoType unless a swizzle or component is present bool isRValue; // true if 'base' is an r-value, otherwise, base is an l-value - unsigned int alignment; // bitwise OR of alignment values passed in. Accumulates worst alignment. Only tracks base and (optional) component selection alignment. + unsigned int alignment; // bitwise OR of alignment values passed in. Accumulates worst alignment. + // Only tracks base and (optional) component selection alignment. // Accumulate whether anything in the chain of structures has coherent decorations. struct CoherentFlags { @@ -583,15 +627,22 @@ public: CoherentFlags operator |=(const CoherentFlags &other) { return *this; } #else bool isVolatile() const { return volatil; } + bool isNonUniform() const { return nonUniform; } + bool anyCoherent() const { + return coherent || devicecoherent || queuefamilycoherent || workgroupcoherent || + subgroupcoherent || shadercallcoherent; + } unsigned coherent : 1; unsigned devicecoherent : 1; unsigned queuefamilycoherent : 1; unsigned workgroupcoherent : 1; unsigned subgroupcoherent : 1; + unsigned shadercallcoherent : 1; unsigned nonprivate : 1; unsigned volatil : 1; unsigned isImage : 1; + unsigned nonUniform : 1; void clear() { coherent = 0; @@ -599,9 +650,11 @@ public: queuefamilycoherent = 0; workgroupcoherent = 0; subgroupcoherent = 0; + shadercallcoherent = 0; nonprivate = 0; volatil = 0; isImage = 0; + nonUniform = 0; } CoherentFlags operator |=(const CoherentFlags &other) { @@ -610,9 +663,11 @@ public: queuefamilycoherent |= other.queuefamilycoherent; workgroupcoherent |= other.workgroupcoherent; subgroupcoherent |= other.subgroupcoherent; + shadercallcoherent |= other.shadercallcoherent; nonprivate |= other.nonprivate; volatil |= other.volatil; isImage |= other.isImage; + nonUniform |= other.nonUniform; return *this; } #endif @@ -655,11 +710,13 @@ public: } // push new swizzle onto the end of any existing swizzle, merging into a single swizzle - void accessChainPushSwizzle(std::vector<unsigned>& swizzle, Id preSwizzleBaseType, AccessChain::CoherentFlags coherentFlags, unsigned int alignment); + void accessChainPushSwizzle(std::vector<unsigned>& swizzle, Id preSwizzleBaseType, + AccessChain::CoherentFlags coherentFlags, unsigned int alignment); // push a dynamic component selection onto the access chain, only applicable with a // non-trivial swizzle or no swizzle - void accessChainPushComponent(Id component, Id preSwizzleBaseType, AccessChain::CoherentFlags coherentFlags, unsigned int alignment) + void accessChainPushComponent(Id component, Id preSwizzleBaseType, AccessChain::CoherentFlags coherentFlags, + unsigned int alignment) { if (accessChain.swizzle.size() != 1) { accessChain.component = component; @@ -671,10 +728,19 @@ public: } // use accessChain and swizzle to store value - void accessChainStore(Id rvalue, spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0); + void accessChainStore(Id rvalue, Decoration nonUniform, + spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, + spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0); // use accessChain and swizzle to load an r-value - Id accessChainLoad(Decoration precision, Decoration nonUniform, Id ResultType, spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0); + Id accessChainLoad(Decoration precision, Decoration l_nonUniform, Decoration r_nonUniform, Id ResultType, + spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, spv::Scope scope = spv::ScopeMax, + unsigned int alignment = 0); + + // Return whether or not the access chain can be represented in SPIR-V + // as an l-value. + // E.g., a[3].yx cannot be, while a[3].y and a[3].y[x] can be. + bool isSpvLvalue() const { return accessChain.swizzle.size() <= 1; } // get the direct pointer for an l-value Id accessChainGetLValue(); @@ -703,7 +769,8 @@ public: void createBranch(Block* block); void createConditionalBranch(Id condition, Block* thenBlock, Block* elseBlock); - void createLoopMerge(Block* mergeBlock, Block* continueBlock, unsigned int control, const std::vector<unsigned int>& operands); + void createLoopMerge(Block* mergeBlock, Block* continueBlock, unsigned int control, + const std::vector<unsigned int>& operands); // Sets to generate opcode for specialization constants. void setToSpecConstCodeGenMode() { generatingOpCodeForSpecConst = true; } @@ -729,7 +796,8 @@ public: void dumpSourceInstructions(const spv::Id fileId, const std::string& text, std::vector<unsigned int>&) const; void dumpInstructions(std::vector<unsigned int>&, const std::vector<std::unique_ptr<Instruction> >&) const; void dumpModuleProcesses(std::vector<unsigned int>&) const; - spv::MemoryAccessMask sanitizeMemoryAccessForStorageClass(spv::MemoryAccessMask memoryAccess, StorageClass sc) const; + spv::MemoryAccessMask sanitizeMemoryAccessForStorageClass(spv::MemoryAccessMask memoryAccess, StorageClass sc) + const; unsigned int spvVersion; // the version of SPIR-V to emit in the header SourceLanguage source; @@ -764,10 +832,16 @@ public: std::vector<std::unique_ptr<Instruction> > externals; std::vector<std::unique_ptr<Function> > functions; - // not output, internally used for quick & dirty canonical (unique) creation - std::unordered_map<unsigned int, std::vector<Instruction*>> groupedConstants; // map type opcodes to constant inst. - std::unordered_map<unsigned int, std::vector<Instruction*>> groupedStructConstants; // map struct-id to constant instructions - std::unordered_map<unsigned int, std::vector<Instruction*>> groupedTypes; // map type opcodes to type instructions + // not output, internally used for quick & dirty canonical (unique) creation + + // map type opcodes to constant inst. + std::unordered_map<unsigned int, std::vector<Instruction*>> groupedConstants; + // map struct-id to constant instructions + std::unordered_map<unsigned int, std::vector<Instruction*>> groupedStructConstants; + // map type opcodes to type instructions + std::unordered_map<unsigned int, std::vector<Instruction*>> groupedTypes; + // list of OpConstantNull instructions + std::vector<Instruction*> nullConstants; // stack of switches std::stack<Block*> switchMerges; diff --git a/thirdparty/glslang/SPIRV/SpvPostProcess.cpp b/thirdparty/glslang/SPIRV/SpvPostProcess.cpp index d40174d172..23d7b5a461 100644 --- a/thirdparty/glslang/SPIRV/SpvPostProcess.cpp +++ b/thirdparty/glslang/SPIRV/SpvPostProcess.cpp @@ -436,6 +436,38 @@ void Builder::postProcessFeatures() { } } } + + // If any Vulkan memory model-specific functionality is used, update the + // OpMemoryModel to match. + if (capabilities.find(spv::CapabilityVulkanMemoryModelKHR) != capabilities.end()) { + memoryModel = spv::MemoryModelVulkanKHR; + addIncorporatedExtension(spv::E_SPV_KHR_vulkan_memory_model, spv::Spv_1_5); + } + + // Add Aliased decoration if there's more than one Workgroup Block variable. + if (capabilities.find(spv::CapabilityWorkgroupMemoryExplicitLayoutKHR) != capabilities.end()) { + assert(entryPoints.size() == 1); + auto &ep = entryPoints[0]; + + std::vector<Id> workgroup_variables; + for (int i = 0; i < (int)ep->getNumOperands(); i++) { + if (!ep->isIdOperand(i)) + continue; + + const Id id = ep->getIdOperand(i); + const Instruction *instr = module.getInstruction(id); + if (instr->getOpCode() != spv::OpVariable) + continue; + + if (instr->getImmediateOperand(0) == spv::StorageClassWorkgroup) + workgroup_variables.push_back(id); + } + + if (workgroup_variables.size() > 1) { + for (size_t i = 0; i < workgroup_variables.size(); i++) + addDecoration(workgroup_variables[i], spv::DecorationAliased); + } + } } #endif diff --git a/thirdparty/glslang/SPIRV/SpvTools.cpp b/thirdparty/glslang/SPIRV/SpvTools.cpp index 97bd4e7427..8acf9b139a 100644 --- a/thirdparty/glslang/SPIRV/SpvTools.cpp +++ b/thirdparty/glslang/SPIRV/SpvTools.cpp @@ -1,6 +1,6 @@ // // Copyright (C) 2014-2016 LunarG, Inc. -// Copyright (C) 2018 Google, Inc. +// Copyright (C) 2018-2020 Google, Inc. // // All rights reserved. // @@ -44,7 +44,6 @@ #include "SpvTools.h" #include "spirv-tools/optimizer.hpp" -#include "spirv-tools/libspirv.h" namespace glslang { @@ -80,12 +79,52 @@ spv_target_env MapToSpirvToolsEnv(const SpvVersion& spvVersion, spv::SpvBuildLog return spv_target_env::SPV_ENV_UNIVERSAL_1_0; } +// Callback passed to spvtools::Optimizer::SetMessageConsumer +void OptimizerMesssageConsumer(spv_message_level_t level, const char *source, + const spv_position_t &position, const char *message) +{ + auto &out = std::cerr; + switch (level) + { + case SPV_MSG_FATAL: + case SPV_MSG_INTERNAL_ERROR: + case SPV_MSG_ERROR: + out << "error: "; + break; + case SPV_MSG_WARNING: + out << "warning: "; + break; + case SPV_MSG_INFO: + case SPV_MSG_DEBUG: + out << "info: "; + break; + default: + break; + } + if (source) + { + out << source << ":"; + } + out << position.line << ":" << position.column << ":" << position.index << ":"; + if (message) + { + out << " " << message; + } + out << std::endl; +} -// Use the SPIRV-Tools disassembler to print SPIR-V. +// Use the SPIRV-Tools disassembler to print SPIR-V using a SPV_ENV_UNIVERSAL_1_3 environment. void SpirvToolsDisassemble(std::ostream& out, const std::vector<unsigned int>& spirv) { + SpirvToolsDisassemble(out, spirv, spv_target_env::SPV_ENV_UNIVERSAL_1_3); +} + +// Use the SPIRV-Tools disassembler to print SPIR-V with a provided SPIR-V environment. +void SpirvToolsDisassemble(std::ostream& out, const std::vector<unsigned int>& spirv, + spv_target_env requested_context) +{ // disassemble - spv_context context = spvContextCreate(SPV_ENV_UNIVERSAL_1_3); + spv_context context = spvContextCreate(requested_context); spv_text text; spv_diagnostic diagnostic = nullptr; spvBinaryToText(context, spirv.data(), spirv.size(), @@ -114,6 +153,8 @@ void SpirvToolsValidate(const glslang::TIntermediate& intermediate, std::vector< spv_validator_options options = spvValidatorOptionsCreate(); spvValidatorOptionsSetRelaxBlockLayout(options, intermediate.usingHlslOffsets()); spvValidatorOptionsSetBeforeHlslLegalization(options, prelegalization); + spvValidatorOptionsSetScalarBlockLayout(options, intermediate.usingScalarBlockLayout()); + spvValidatorOptionsSetWorkgroupScalarBlockLayout(options, intermediate.usingScalarBlockLayout()); spvValidateWithOptions(context, options, &binary, &diagnostic); // report @@ -128,52 +169,21 @@ void SpirvToolsValidate(const glslang::TIntermediate& intermediate, std::vector< spvContextDestroy(context); } -// Apply the SPIRV-Tools optimizer to generated SPIR-V, for the purpose of -// legalizing HLSL SPIR-V. -void SpirvToolsLegalize(const glslang::TIntermediate&, std::vector<unsigned int>& spirv, - spv::SpvBuildLogger*, const SpvOptions* options) +// Apply the SPIRV-Tools optimizer to generated SPIR-V. HLSL SPIR-V is legalized in the process. +void SpirvToolsTransform(const glslang::TIntermediate& intermediate, std::vector<unsigned int>& spirv, + spv::SpvBuildLogger* logger, const SpvOptions* options) { - spv_target_env target_env = SPV_ENV_UNIVERSAL_1_2; + spv_target_env target_env = MapToSpirvToolsEnv(intermediate.getSpv(), logger); spvtools::Optimizer optimizer(target_env); - optimizer.SetMessageConsumer( - [](spv_message_level_t level, const char *source, const spv_position_t &position, const char *message) { - auto &out = std::cerr; - switch (level) - { - case SPV_MSG_FATAL: - case SPV_MSG_INTERNAL_ERROR: - case SPV_MSG_ERROR: - out << "error: "; - break; - case SPV_MSG_WARNING: - out << "warning: "; - break; - case SPV_MSG_INFO: - case SPV_MSG_DEBUG: - out << "info: "; - break; - default: - break; - } - if (source) - { - out << source << ":"; - } - out << position.line << ":" << position.column << ":" << position.index << ":"; - if (message) - { - out << " " << message; - } - out << std::endl; - }); + optimizer.SetMessageConsumer(OptimizerMesssageConsumer); // If debug (specifically source line info) is being generated, propagate // line information into all SPIR-V instructions. This avoids loss of // information when instructions are deleted or moved. Later, remove // redundant information to minimize final SPRIR-V size. - if (options->generateDebugInfo) { - optimizer.RegisterPass(spvtools::CreatePropagateLineInfoPass()); + if (options->stripDebugInfo) { + optimizer.RegisterPass(spvtools::CreateStripDebugInfoPass()); } optimizer.RegisterPass(spvtools::CreateWrapOpKillPass()); optimizer.RegisterPass(spvtools::CreateDeadBranchElimPass()); @@ -197,17 +207,35 @@ void SpirvToolsLegalize(const glslang::TIntermediate&, std::vector<unsigned int> optimizer.RegisterPass(spvtools::CreateAggressiveDCEPass()); optimizer.RegisterPass(spvtools::CreateVectorDCEPass()); optimizer.RegisterPass(spvtools::CreateDeadInsertElimPass()); + optimizer.RegisterPass(spvtools::CreateInterpolateFixupPass()); if (options->optimizeSize) { optimizer.RegisterPass(spvtools::CreateRedundancyEliminationPass()); } optimizer.RegisterPass(spvtools::CreateAggressiveDCEPass()); optimizer.RegisterPass(spvtools::CreateCFGCleanupPass()); - if (options->generateDebugInfo) { - optimizer.RegisterPass(spvtools::CreateRedundantLineInfoElimPass()); - } spvtools::OptimizerOptions spvOptOptions; - spvOptOptions.set_run_validator(false); // The validator may run as a seperate step later on + optimizer.SetTargetEnv(MapToSpirvToolsEnv(intermediate.getSpv(), logger)); + spvOptOptions.set_run_validator(false); // The validator may run as a separate step later on + optimizer.Run(spirv.data(), spirv.size(), &spirv, spvOptOptions); +} + +// Apply the SPIRV-Tools optimizer to strip debug info from SPIR-V. This is implicitly done by +// SpirvToolsTransform if spvOptions->stripDebugInfo is set, but can be called separately if +// optimization is disabled. +void SpirvToolsStripDebugInfo(const glslang::TIntermediate& intermediate, + std::vector<unsigned int>& spirv, spv::SpvBuildLogger* logger) +{ + spv_target_env target_env = MapToSpirvToolsEnv(intermediate.getSpv(), logger); + + spvtools::Optimizer optimizer(target_env); + optimizer.SetMessageConsumer(OptimizerMesssageConsumer); + + optimizer.RegisterPass(spvtools::CreateStripDebugInfoPass()); + + spvtools::OptimizerOptions spvOptOptions; + optimizer.SetTargetEnv(MapToSpirvToolsEnv(intermediate.getSpv(), logger)); + spvOptOptions.set_run_validator(false); // The validator may run as a separate step later on optimizer.Run(spirv.data(), spirv.size(), &spirv, spvOptOptions); } diff --git a/thirdparty/glslang/SPIRV/SpvTools.h b/thirdparty/glslang/SPIRV/SpvTools.h index 59c914da0b..3fb3cbacd3 100644 --- a/thirdparty/glslang/SPIRV/SpvTools.h +++ b/thirdparty/glslang/SPIRV/SpvTools.h @@ -41,9 +41,10 @@ #ifndef GLSLANG_SPV_TOOLS_H #define GLSLANG_SPV_TOOLS_H -#ifdef ENABLE_OPT +#if ENABLE_OPT #include <vector> #include <ostream> +#include "spirv-tools/libspirv.h" #endif #include "glslang/MachineIndependent/localintermediate.h" @@ -52,28 +53,38 @@ namespace glslang { struct SpvOptions { - SpvOptions() : generateDebugInfo(false), disableOptimizer(true), + SpvOptions() : generateDebugInfo(false), stripDebugInfo(false), disableOptimizer(true), optimizeSize(false), disassemble(false), validate(false) { } bool generateDebugInfo; + bool stripDebugInfo; bool disableOptimizer; bool optimizeSize; bool disassemble; bool validate; }; -#ifdef ENABLE_OPT +#if ENABLE_OPT -// Use the SPIRV-Tools disassembler to print SPIR-V. +// Use the SPIRV-Tools disassembler to print SPIR-V using a SPV_ENV_UNIVERSAL_1_3 environment. void SpirvToolsDisassemble(std::ostream& out, const std::vector<unsigned int>& spirv); +// Use the SPIRV-Tools disassembler to print SPIR-V with a provided SPIR-V environment. +void SpirvToolsDisassemble(std::ostream& out, const std::vector<unsigned int>& spirv, + spv_target_env requested_context); + // Apply the SPIRV-Tools validator to generated SPIR-V. void SpirvToolsValidate(const glslang::TIntermediate& intermediate, std::vector<unsigned int>& spirv, spv::SpvBuildLogger*, bool prelegalization); -// Apply the SPIRV-Tools optimizer to generated SPIR-V, for the purpose of -// legalizing HLSL SPIR-V. -void SpirvToolsLegalize(const glslang::TIntermediate& intermediate, std::vector<unsigned int>& spirv, - spv::SpvBuildLogger*, const SpvOptions*); +// Apply the SPIRV-Tools optimizer to generated SPIR-V. HLSL SPIR-V is legalized in the process. +void SpirvToolsTransform(const glslang::TIntermediate& intermediate, std::vector<unsigned int>& spirv, + spv::SpvBuildLogger*, const SpvOptions*); + +// Apply the SPIRV-Tools optimizer to strip debug info from SPIR-V. This is implicitly done by +// SpirvToolsTransform if spvOptions->stripDebugInfo is set, but can be called separately if +// optimization is disabled. +void SpirvToolsStripDebugInfo(const glslang::TIntermediate& intermediate, + std::vector<unsigned int>& spirv, spv::SpvBuildLogger*); #endif diff --git a/thirdparty/glslang/SPIRV/disassemble.cpp b/thirdparty/glslang/SPIRV/disassemble.cpp index 930e799493..73c988c5b3 100644 --- a/thirdparty/glslang/SPIRV/disassemble.cpp +++ b/thirdparty/glslang/SPIRV/disassemble.cpp @@ -46,7 +46,6 @@ #include "disassemble.h" #include "doc.h" -#include "SpvTools.h" namespace spv { extern "C" { @@ -75,6 +74,7 @@ enum ExtInstSet { GLSLextAMDInst, GLSLextNVInst, OpenCLExtInst, + NonSemanticDebugPrintfExtInst, }; // Container class for a single instance of a SPIR-V stream, with methods for disassembly. @@ -480,8 +480,12 @@ void SpirvStream::disassembleInstruction(Id resultId, Id /*typeId*/, Op opCode, if (opCode == OpExtInst) { ExtInstSet extInstSet = GLSL450Inst; const char* name = idDescriptor[stream[word - 2]].c_str(); - if (0 == memcmp("OpenCL", name, 6)) { + if (strcmp("OpenCL.std", name) == 0) { extInstSet = OpenCLExtInst; + } else if (strcmp("OpenCL.DebugInfo.100", name) == 0) { + extInstSet = OpenCLExtInst; + } else if (strcmp("NonSemantic.DebugPrintf", name) == 0) { + extInstSet = NonSemanticDebugPrintfExtInst; } else if (strcmp(spv::E_SPV_AMD_shader_ballot, name) == 0 || strcmp(spv::E_SPV_AMD_shader_trinary_minmax, name) == 0 || strcmp(spv::E_SPV_AMD_shader_explicit_vertex_parameter, name) == 0 || @@ -505,6 +509,8 @@ void SpirvStream::disassembleInstruction(Id resultId, Id /*typeId*/, Op opCode, } else if (extInstSet == GLSLextNVInst) { out << "(" << GLSLextNVGetDebugNames(name, entrypoint) << ")"; + } else if (extInstSet == NonSemanticDebugPrintfExtInst) { + out << "(DebugPrintf)"; } } break; @@ -512,6 +518,10 @@ void SpirvStream::disassembleInstruction(Id resultId, Id /*typeId*/, Op opCode, case OperandLiteralString: numOperands -= disassembleString(); break; + case OperandVariableLiteralStrings: + while (numOperands > 0) + numOperands -= disassembleString(); + return; case OperandMemoryAccess: outputMask(OperandMemoryAccess, stream[word++]); --numOperands; diff --git a/thirdparty/glslang/SPIRV/doc.cpp b/thirdparty/glslang/SPIRV/doc.cpp index bee5c79729..dbdf7077a6 100644 --- a/thirdparty/glslang/SPIRV/doc.cpp +++ b/thirdparty/glslang/SPIRV/doc.cpp @@ -1,5 +1,6 @@ // // Copyright (C) 2014-2015 LunarG, Inc. +// Modifications Copyright (C) 2020 Advanced Micro Devices, Inc. All rights reserved. // // All rights reserved. // @@ -99,12 +100,12 @@ const char* ExecutionModelString(int model) default: return "Bad"; - case ExecutionModelRayGenerationNV: return "RayGenerationNV"; - case ExecutionModelIntersectionNV: return "IntersectionNV"; - case ExecutionModelAnyHitNV: return "AnyHitNV"; - case ExecutionModelClosestHitNV: return "ClosestHitNV"; - case ExecutionModelMissNV: return "MissNV"; - case ExecutionModelCallableNV: return "CallableNV"; + case ExecutionModelRayGenerationKHR: return "RayGenerationKHR"; + case ExecutionModelIntersectionKHR: return "IntersectionKHR"; + case ExecutionModelAnyHitKHR: return "AnyHitKHR"; + case ExecutionModelClosestHitKHR: return "ClosestHitKHR"; + case ExecutionModelMissKHR: return "MissKHR"; + case ExecutionModelCallableKHR: return "CallableKHR"; } } @@ -133,7 +134,7 @@ const char* MemoryString(int mem) } } -const int ExecutionModeCeiling = 33; +const int ExecutionModeCeiling = 40; const char* ExecutionModeString(int mode) { @@ -172,7 +173,22 @@ const char* ExecutionModeString(int mode) case 31: return "ContractionOff"; case 32: return "Bad"; - case 4446: return "PostDepthCoverage"; + case ExecutionModeInitializer: return "Initializer"; + case ExecutionModeFinalizer: return "Finalizer"; + case ExecutionModeSubgroupSize: return "SubgroupSize"; + case ExecutionModeSubgroupsPerWorkgroup: return "SubgroupsPerWorkgroup"; + case ExecutionModeSubgroupsPerWorkgroupId: return "SubgroupsPerWorkgroupId"; + case ExecutionModeLocalSizeId: return "LocalSizeId"; + case ExecutionModeLocalSizeHintId: return "LocalSizeHintId"; + + case ExecutionModePostDepthCoverage: return "PostDepthCoverage"; + case ExecutionModeDenormPreserve: return "DenormPreserve"; + case ExecutionModeDenormFlushToZero: return "DenormFlushToZero"; + case ExecutionModeSignedZeroInfNanPreserve: return "SignedZeroInfNanPreserve"; + case ExecutionModeRoundingModeRTE: return "RoundingModeRTE"; + case ExecutionModeRoundingModeRTZ: return "RoundingModeRTZ"; + case ExecutionModeStencilRefReplacingEXT: return "StencilRefReplacingEXT"; + case ExecutionModeSubgroupUniformControlFlowKHR: return "SubgroupUniformControlFlow"; case ExecutionModeOutputLinesNV: return "OutputLinesNV"; case ExecutionModeOutputPrimitivesNV: return "OutputPrimitivesNV"; @@ -187,6 +203,11 @@ const char* ExecutionModeString(int mode) case ExecutionModeShadingRateInterlockOrderedEXT: return "ShadingRateInterlockOrderedEXT"; case ExecutionModeShadingRateInterlockUnorderedEXT: return "ShadingRateInterlockUnorderedEXT"; + case ExecutionModeMaxWorkgroupSizeINTEL: return "MaxWorkgroupSizeINTEL"; + case ExecutionModeMaxWorkDimINTEL: return "MaxWorkDimINTEL"; + case ExecutionModeNoGlobalOffsetINTEL: return "NoGlobalOffsetINTEL"; + case ExecutionModeNumSIMDWorkitemsINTEL: return "NumSIMDWorkitemsINTEL"; + case ExecutionModeCeiling: default: return "Bad"; } @@ -209,12 +230,12 @@ const char* StorageClassString(int StorageClass) case 11: return "Image"; case 12: return "StorageBuffer"; - case StorageClassRayPayloadNV: return "RayPayloadNV"; - case StorageClassHitAttributeNV: return "HitAttributeNV"; - case StorageClassIncomingRayPayloadNV: return "IncomingRayPayloadNV"; - case StorageClassShaderRecordBufferNV: return "ShaderRecordBufferNV"; - case StorageClassCallableDataNV: return "CallableDataNV"; - case StorageClassIncomingCallableDataNV: return "IncomingCallableDataNV"; + case StorageClassRayPayloadKHR: return "RayPayloadKHR"; + case StorageClassHitAttributeKHR: return "HitAttributeKHR"; + case StorageClassIncomingRayPayloadKHR: return "IncomingRayPayloadKHR"; + case StorageClassShaderRecordBufferKHR: return "ShaderRecordBufferKHR"; + case StorageClassCallableDataKHR: return "CallableDataKHR"; + case StorageClassIncomingCallableDataKHR: return "IncomingCallableDataKHR"; case StorageClassPhysicalStorageBufferEXT: return "PhysicalStorageBufferEXT"; @@ -352,6 +373,8 @@ const char* BuiltInString(int builtIn) case 4424: return "BaseVertex"; case 4425: return "BaseInstance"; case 4426: return "DrawIndex"; + case 4432: return "PrimitiveShadingRateKHR"; + case 4444: return "ShadingRateKHR"; case 5014: return "FragStencilRefEXT"; case 4992: return "BaryCoordNoPerspAMD"; @@ -361,32 +384,33 @@ const char* BuiltInString(int builtIn) case 4996: return "BaryCoordSmoothCentroidAMD"; case 4997: return "BaryCoordSmoothSampleAMD"; case 4998: return "BaryCoordPullModelAMD"; - case BuiltInLaunchIdNV: return "LaunchIdNV"; - case BuiltInLaunchSizeNV: return "LaunchSizeNV"; - case BuiltInWorldRayOriginNV: return "WorldRayOriginNV"; - case BuiltInWorldRayDirectionNV: return "WorldRayDirectionNV"; - case BuiltInObjectRayOriginNV: return "ObjectRayOriginNV"; - case BuiltInObjectRayDirectionNV: return "ObjectRayDirectionNV"; - case BuiltInRayTminNV: return "RayTminNV"; - case BuiltInRayTmaxNV: return "RayTmaxNV"; - case BuiltInInstanceCustomIndexNV: return "InstanceCustomIndexNV"; - case BuiltInObjectToWorldNV: return "ObjectToWorldNV"; - case BuiltInWorldToObjectNV: return "WorldToObjectNV"; - case BuiltInHitTNV: return "HitTNV"; - case BuiltInHitKindNV: return "HitKindNV"; - case BuiltInIncomingRayFlagsNV: return "IncomingRayFlagsNV"; - case BuiltInViewportMaskNV: return "ViewportMaskNV"; - case BuiltInSecondaryPositionNV: return "SecondaryPositionNV"; - case BuiltInSecondaryViewportMaskNV: return "SecondaryViewportMaskNV"; - case BuiltInPositionPerViewNV: return "PositionPerViewNV"; - case BuiltInViewportMaskPerViewNV: return "ViewportMaskPerViewNV"; + case BuiltInLaunchIdKHR: return "LaunchIdKHR"; + case BuiltInLaunchSizeKHR: return "LaunchSizeKHR"; + case BuiltInWorldRayOriginKHR: return "WorldRayOriginKHR"; + case BuiltInWorldRayDirectionKHR: return "WorldRayDirectionKHR"; + case BuiltInObjectRayOriginKHR: return "ObjectRayOriginKHR"; + case BuiltInObjectRayDirectionKHR: return "ObjectRayDirectionKHR"; + case BuiltInRayTminKHR: return "RayTminKHR"; + case BuiltInRayTmaxKHR: return "RayTmaxKHR"; + case BuiltInInstanceCustomIndexKHR: return "InstanceCustomIndexKHR"; + case BuiltInRayGeometryIndexKHR: return "RayGeometryIndexKHR"; + case BuiltInObjectToWorldKHR: return "ObjectToWorldKHR"; + case BuiltInWorldToObjectKHR: return "WorldToObjectKHR"; + case BuiltInHitTNV: return "HitTNV"; + case BuiltInHitKindKHR: return "HitKindKHR"; + case BuiltInIncomingRayFlagsKHR: return "IncomingRayFlagsKHR"; + case BuiltInViewportMaskNV: return "ViewportMaskNV"; + case BuiltInSecondaryPositionNV: return "SecondaryPositionNV"; + case BuiltInSecondaryViewportMaskNV: return "SecondaryViewportMaskNV"; + case BuiltInPositionPerViewNV: return "PositionPerViewNV"; + case BuiltInViewportMaskPerViewNV: return "ViewportMaskPerViewNV"; // case BuiltInFragmentSizeNV: return "FragmentSizeNV"; // superseded by BuiltInFragSizeEXT // case BuiltInInvocationsPerPixelNV: return "InvocationsPerPixelNV"; // superseded by BuiltInFragInvocationCountEXT - case BuiltInBaryCoordNV: return "BaryCoordNV"; - case BuiltInBaryCoordNoPerspNV: return "BaryCoordNoPerspNV"; + case BuiltInBaryCoordNV: return "BaryCoordNV"; + case BuiltInBaryCoordNoPerspNV: return "BaryCoordNoPerspNV"; - case BuiltInFragSizeEXT: return "FragSizeEXT"; - case BuiltInFragInvocationCountEXT: return "FragInvocationCountEXT"; + case BuiltInFragSizeEXT: return "FragSizeEXT"; + case BuiltInFragInvocationCountEXT: return "FragInvocationCountEXT"; case 5264: return "FullyCoveredEXT"; @@ -402,6 +426,7 @@ const char* BuiltInString(int builtIn) case BuiltInSMCountNV: return "SMCountNV"; case BuiltInWarpIDNV: return "WarpIDNV"; case BuiltInSMIDNV: return "SMIDNV"; + case BuiltInCurrentRayTimeNV: return "CurrentRayTimeNV"; default: return "Bad"; } @@ -500,6 +525,8 @@ const char* ImageFormatString(int format) case 37: return "Rg8ui"; case 38: return "R16ui"; case 39: return "R8ui"; + case 40: return "R64ui"; + case 41: return "R64i"; default: return "Bad"; @@ -890,6 +917,11 @@ const char* CapabilityString(int info) case CapabilityPerViewAttributesNV: return "PerViewAttributesNV"; case CapabilityGroupNonUniformPartitionedNV: return "GroupNonUniformPartitionedNV"; case CapabilityRayTracingNV: return "RayTracingNV"; + case CapabilityRayTracingMotionBlurNV: return "RayTracingMotionBlurNV"; + case CapabilityRayTracingKHR: return "RayTracingKHR"; + case CapabilityRayQueryKHR: return "RayQueryKHR"; + case CapabilityRayTracingProvisionalKHR: return "RayTracingProvisionalKHR"; + case CapabilityRayTraversalPrimitiveCullingKHR: return "RayTraversalPrimitiveCullingKHR"; case CapabilityComputeDerivativeGroupQuadsNV: return "ComputeDerivativeGroupQuadsNV"; case CapabilityComputeDerivativeGroupLinearNV: return "ComputeDerivativeGroupLinearNV"; case CapabilityFragmentBarycentricNV: return "FragmentBarycentricNV"; @@ -928,11 +960,25 @@ const char* CapabilityString(int info) case CapabilityFragmentShaderPixelInterlockEXT: return "CapabilityFragmentShaderPixelInterlockEXT"; case CapabilityFragmentShaderShadingRateInterlockEXT: return "CapabilityFragmentShaderShadingRateInterlockEXT"; + case CapabilityFragmentShadingRateKHR: return "FragmentShadingRateKHR"; + case CapabilityDemoteToHelperInvocationEXT: return "DemoteToHelperInvocationEXT"; case CapabilityShaderClockKHR: return "ShaderClockKHR"; + case CapabilityInt64ImageEXT: return "Int64ImageEXT"; case CapabilityIntegerFunctions2INTEL: return "CapabilityIntegerFunctions2INTEL"; + case CapabilityAtomicFloat16AddEXT: return "AtomicFloat16AddEXT"; + case CapabilityAtomicFloat32AddEXT: return "AtomicFloat32AddEXT"; + case CapabilityAtomicFloat64AddEXT: return "AtomicFloat64AddEXT"; + case CapabilityAtomicFloat16MinMaxEXT: return "AtomicFloat16MinMaxEXT"; + case CapabilityAtomicFloat32MinMaxEXT: return "AtomicFloat32MinMaxEXT"; + case CapabilityAtomicFloat64MinMaxEXT: return "AtomicFloat64MinMaxEXT"; + + case CapabilityWorkgroupMemoryExplicitLayoutKHR: return "CapabilityWorkgroupMemoryExplicitLayoutKHR"; + case CapabilityWorkgroupMemoryExplicitLayout8BitAccessKHR: return "CapabilityWorkgroupMemoryExplicitLayout8BitAccessKHR"; + case CapabilityWorkgroupMemoryExplicitLayout16BitAccessKHR: return "CapabilityWorkgroupMemoryExplicitLayout16BitAccessKHR"; + default: return "Bad"; } } @@ -1264,6 +1310,7 @@ const char* OpcodeString(int op) case 320: return "OpImageSparseRead"; case OpModuleProcessed: return "OpModuleProcessed"; + case OpExecutionModeId: return "OpExecutionModeId"; case OpDecorateId: return "OpDecorateId"; case 333: return "OpGroupNonUniformElect"; @@ -1301,6 +1348,8 @@ const char* OpcodeString(int op) case 365: return "OpGroupNonUniformQuadBroadcast"; case 366: return "OpGroupNonUniformQuadSwap"; + case OpTerminateInvocation: return "OpTerminateInvocation"; + case 4421: return "OpSubgroupBallotKHR"; case 4422: return "OpSubgroupFirstInvocationKHR"; case 4428: return "OpSubgroupAllKHR"; @@ -1308,6 +1357,10 @@ const char* OpcodeString(int op) case 4430: return "OpSubgroupAllEqualKHR"; case 4432: return "OpSubgroupReadInvocationKHR"; + case OpAtomicFAddEXT: return "OpAtomicFAddEXT"; + case OpAtomicFMinEXT: return "OpAtomicFMinEXT"; + case OpAtomicFMaxEXT: return "OpAtomicFMaxEXT"; + case 5000: return "OpGroupIAddNonUniformAMD"; case 5001: return "OpGroupFAddNonUniformAMD"; case 5002: return "OpGroupFMinNonUniformAMD"; @@ -1325,16 +1378,48 @@ const char* OpcodeString(int op) case OpDecorateStringGOOGLE: return "OpDecorateStringGOOGLE"; case OpMemberDecorateStringGOOGLE: return "OpMemberDecorateStringGOOGLE"; + case OpReportIntersectionKHR: return "OpReportIntersectionKHR"; + case OpIgnoreIntersectionNV: return "OpIgnoreIntersectionNV"; + case OpIgnoreIntersectionKHR: return "OpIgnoreIntersectionKHR"; + case OpTerminateRayNV: return "OpTerminateRayNV"; + case OpTerminateRayKHR: return "OpTerminateRayKHR"; + case OpTraceNV: return "OpTraceNV"; + case OpTraceRayMotionNV: return "OpTraceRayMotionNV"; + case OpTraceRayKHR: return "OpTraceRayKHR"; + case OpTypeAccelerationStructureKHR: return "OpTypeAccelerationStructureKHR"; + case OpExecuteCallableNV: return "OpExecuteCallableNV"; + case OpExecuteCallableKHR: return "OpExecuteCallableKHR"; + case OpConvertUToAccelerationStructureKHR: return "OpConvertUToAccelerationStructureKHR"; + case OpGroupNonUniformPartitionNV: return "OpGroupNonUniformPartitionNV"; - case OpReportIntersectionNV: return "OpReportIntersectionNV"; - case OpIgnoreIntersectionNV: return "OpIgnoreIntersectionNV"; - case OpTerminateRayNV: return "OpTerminateRayNV"; - case OpTraceNV: return "OpTraceNV"; - case OpTypeAccelerationStructureNV: return "OpTypeAccelerationStructureNV"; - case OpExecuteCallableNV: return "OpExecuteCallableNV"; case OpImageSampleFootprintNV: return "OpImageSampleFootprintNV"; case OpWritePackedPrimitiveIndices4x8NV: return "OpWritePackedPrimitiveIndices4x8NV"; + case OpTypeRayQueryKHR: return "OpTypeRayQueryKHR"; + case OpRayQueryInitializeKHR: return "OpRayQueryInitializeKHR"; + case OpRayQueryTerminateKHR: return "OpRayQueryTerminateKHR"; + case OpRayQueryGenerateIntersectionKHR: return "OpRayQueryGenerateIntersectionKHR"; + case OpRayQueryConfirmIntersectionKHR: return "OpRayQueryConfirmIntersectionKHR"; + case OpRayQueryProceedKHR: return "OpRayQueryProceedKHR"; + case OpRayQueryGetIntersectionTypeKHR: return "OpRayQueryGetIntersectionTypeKHR"; + case OpRayQueryGetRayTMinKHR: return "OpRayQueryGetRayTMinKHR"; + case OpRayQueryGetRayFlagsKHR: return "OpRayQueryGetRayFlagsKHR"; + case OpRayQueryGetIntersectionTKHR: return "OpRayQueryGetIntersectionTKHR"; + case OpRayQueryGetIntersectionInstanceCustomIndexKHR: return "OpRayQueryGetIntersectionInstanceCustomIndexKHR"; + case OpRayQueryGetIntersectionInstanceIdKHR: return "OpRayQueryGetIntersectionInstanceIdKHR"; + case OpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR: return "OpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR"; + case OpRayQueryGetIntersectionGeometryIndexKHR: return "OpRayQueryGetIntersectionGeometryIndexKHR"; + case OpRayQueryGetIntersectionPrimitiveIndexKHR: return "OpRayQueryGetIntersectionPrimitiveIndexKHR"; + case OpRayQueryGetIntersectionBarycentricsKHR: return "OpRayQueryGetIntersectionBarycentricsKHR"; + case OpRayQueryGetIntersectionFrontFaceKHR: return "OpRayQueryGetIntersectionFrontFaceKHR"; + case OpRayQueryGetIntersectionCandidateAABBOpaqueKHR: return "OpRayQueryGetIntersectionCandidateAABBOpaqueKHR"; + case OpRayQueryGetIntersectionObjectRayDirectionKHR: return "OpRayQueryGetIntersectionObjectRayDirectionKHR"; + case OpRayQueryGetIntersectionObjectRayOriginKHR: return "OpRayQueryGetIntersectionObjectRayOriginKHR"; + case OpRayQueryGetWorldRayDirectionKHR: return "OpRayQueryGetWorldRayDirectionKHR"; + case OpRayQueryGetWorldRayOriginKHR: return "OpRayQueryGetWorldRayOriginKHR"; + case OpRayQueryGetIntersectionObjectToWorldKHR: return "OpRayQueryGetIntersectionObjectToWorldKHR"; + case OpRayQueryGetIntersectionWorldToObjectKHR: return "OpRayQueryGetIntersectionWorldToObjectKHR"; + case OpTypeCooperativeMatrixNV: return "OpTypeCooperativeMatrixNV"; case OpCooperativeMatrixLoadNV: return "OpCooperativeMatrixLoadNV"; case OpCooperativeMatrixStoreNV: return "OpCooperativeMatrixStoreNV"; @@ -1388,6 +1473,7 @@ void Parameterize() InstructionDesc[OpMemoryModel].setResultAndType(false, false); InstructionDesc[OpEntryPoint].setResultAndType(false, false); InstructionDesc[OpExecutionMode].setResultAndType(false, false); + InstructionDesc[OpExecutionModeId].setResultAndType(false, false); InstructionDesc[OpTypeVoid].setResultAndType(true, false); InstructionDesc[OpTypeBool].setResultAndType(true, false); InstructionDesc[OpTypeInt].setResultAndType(true, false); @@ -1441,6 +1527,7 @@ void Parameterize() InstructionDesc[OpBranchConditional].setResultAndType(false, false); InstructionDesc[OpSwitch].setResultAndType(false, false); InstructionDesc[OpKill].setResultAndType(false, false); + InstructionDesc[OpTerminateInvocation].setResultAndType(false, false); InstructionDesc[OpReturn].setResultAndType(false, false); InstructionDesc[OpReturnValue].setResultAndType(false, false); InstructionDesc[OpUnreachable].setResultAndType(false, false); @@ -1574,6 +1661,10 @@ void Parameterize() InstructionDesc[OpExecutionMode].operands.push(OperandExecutionMode, "'Mode'"); InstructionDesc[OpExecutionMode].operands.push(OperandOptionalLiteral, "See <<Execution_Mode,Execution Mode>>"); + InstructionDesc[OpExecutionModeId].operands.push(OperandId, "'Entry Point'"); + InstructionDesc[OpExecutionModeId].operands.push(OperandExecutionMode, "'Mode'"); + InstructionDesc[OpExecutionModeId].operands.push(OperandVariableIds, "See <<Execution_Mode,Execution Mode>>"); + InstructionDesc[OpTypeInt].operands.push(OperandLiteralNumber, "'Width'"); InstructionDesc[OpTypeInt].operands.push(OperandLiteralNumber, "'Signedness'"); @@ -1667,7 +1758,7 @@ void Parameterize() InstructionDesc[OpDecorateStringGOOGLE].operands.push(OperandId, "'Target'"); InstructionDesc[OpDecorateStringGOOGLE].operands.push(OperandDecoration, ""); - InstructionDesc[OpDecorateStringGOOGLE].operands.push(OperandLiteralString, "'Literal String'"); + InstructionDesc[OpDecorateStringGOOGLE].operands.push(OperandVariableLiteralStrings, "'Literal Strings'"); InstructionDesc[OpMemberDecorate].operands.push(OperandId, "'Structure Type'"); InstructionDesc[OpMemberDecorate].operands.push(OperandLiteralNumber, "'Member'"); @@ -1677,7 +1768,7 @@ void Parameterize() InstructionDesc[OpMemberDecorateStringGOOGLE].operands.push(OperandId, "'Structure Type'"); InstructionDesc[OpMemberDecorateStringGOOGLE].operands.push(OperandLiteralNumber, "'Member'"); InstructionDesc[OpMemberDecorateStringGOOGLE].operands.push(OperandDecoration, ""); - InstructionDesc[OpMemberDecorateStringGOOGLE].operands.push(OperandLiteralString, "'Literal String'"); + InstructionDesc[OpMemberDecorateStringGOOGLE].operands.push(OperandVariableLiteralStrings, "'Literal Strings'"); InstructionDesc[OpGroupDecorate].operands.push(OperandId, "'Decoration Group'"); InstructionDesc[OpGroupDecorate].operands.push(OperandVariableIds, "'Targets'"); @@ -2230,6 +2321,11 @@ void Parameterize() InstructionDesc[OpAtomicIAdd].operands.push(OperandMemorySemantics, "'Semantics'"); InstructionDesc[OpAtomicIAdd].operands.push(OperandId, "'Value'"); + InstructionDesc[OpAtomicFAddEXT].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicFAddEXT].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicFAddEXT].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicFAddEXT].operands.push(OperandId, "'Value'"); + InstructionDesc[OpAtomicISub].operands.push(OperandId, "'Pointer'"); InstructionDesc[OpAtomicISub].operands.push(OperandScope, "'Scope'"); InstructionDesc[OpAtomicISub].operands.push(OperandMemorySemantics, "'Semantics'"); @@ -2255,6 +2351,16 @@ void Parameterize() InstructionDesc[OpAtomicSMax].operands.push(OperandMemorySemantics, "'Semantics'"); InstructionDesc[OpAtomicSMax].operands.push(OperandId, "'Value'"); + InstructionDesc[OpAtomicFMinEXT].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicFMinEXT].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicFMinEXT].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicFMinEXT].operands.push(OperandId, "'Value'"); + + InstructionDesc[OpAtomicFMaxEXT].operands.push(OperandId, "'Pointer'"); + InstructionDesc[OpAtomicFMaxEXT].operands.push(OperandScope, "'Scope'"); + InstructionDesc[OpAtomicFMaxEXT].operands.push(OperandMemorySemantics, "'Semantics'"); + InstructionDesc[OpAtomicFMaxEXT].operands.push(OperandId, "'Value'"); + InstructionDesc[OpAtomicAnd].operands.push(OperandId, "'Pointer'"); InstructionDesc[OpAtomicAnd].operands.push(OperandScope, "'Scope'"); InstructionDesc[OpAtomicAnd].operands.push(OperandMemorySemantics, "'Semantics'"); @@ -2633,7 +2739,7 @@ void Parameterize() InstructionDesc[OpGroupNonUniformQuadSwap].operands.push(OperandScope, "'Execution'"); InstructionDesc[OpGroupNonUniformQuadSwap].operands.push(OperandId, "X"); - InstructionDesc[OpGroupNonUniformQuadSwap].operands.push(OperandLiteralNumber, "'Direction'"); + InstructionDesc[OpGroupNonUniformQuadSwap].operands.push(OperandId, "'Direction'"); InstructionDesc[OpSubgroupBallotKHR].operands.push(OperandId, "'Predicate'"); @@ -2694,9 +2800,9 @@ void Parameterize() InstructionDesc[OpGroupNonUniformPartitionNV].operands.push(OperandId, "X"); - InstructionDesc[OpTypeAccelerationStructureNV].setResultAndType(true, false); + InstructionDesc[OpTypeAccelerationStructureKHR].setResultAndType(true, false); - InstructionDesc[OpTraceNV].operands.push(OperandId, "'NV Acceleration Structure'"); + InstructionDesc[OpTraceNV].operands.push(OperandId, "'Acceleration Structure'"); InstructionDesc[OpTraceNV].operands.push(OperandId, "'Ray Flags'"); InstructionDesc[OpTraceNV].operands.push(OperandId, "'Cull Mask'"); InstructionDesc[OpTraceNV].operands.push(OperandId, "'SBT Record Offset'"); @@ -2709,17 +2815,149 @@ void Parameterize() InstructionDesc[OpTraceNV].operands.push(OperandId, "'Payload'"); InstructionDesc[OpTraceNV].setResultAndType(false, false); - InstructionDesc[OpReportIntersectionNV].operands.push(OperandId, "'Hit Parameter'"); - InstructionDesc[OpReportIntersectionNV].operands.push(OperandId, "'Hit Kind'"); + InstructionDesc[OpTraceRayMotionNV].operands.push(OperandId, "'Acceleration Structure'"); + InstructionDesc[OpTraceRayMotionNV].operands.push(OperandId, "'Ray Flags'"); + InstructionDesc[OpTraceRayMotionNV].operands.push(OperandId, "'Cull Mask'"); + InstructionDesc[OpTraceRayMotionNV].operands.push(OperandId, "'SBT Record Offset'"); + InstructionDesc[OpTraceRayMotionNV].operands.push(OperandId, "'SBT Record Stride'"); + InstructionDesc[OpTraceRayMotionNV].operands.push(OperandId, "'Miss Index'"); + InstructionDesc[OpTraceRayMotionNV].operands.push(OperandId, "'Ray Origin'"); + InstructionDesc[OpTraceRayMotionNV].operands.push(OperandId, "'TMin'"); + InstructionDesc[OpTraceRayMotionNV].operands.push(OperandId, "'Ray Direction'"); + InstructionDesc[OpTraceRayMotionNV].operands.push(OperandId, "'TMax'"); + InstructionDesc[OpTraceRayMotionNV].operands.push(OperandId, "'Time'"); + InstructionDesc[OpTraceRayMotionNV].operands.push(OperandId, "'Payload'"); + InstructionDesc[OpTraceRayMotionNV].setResultAndType(false, false); + + InstructionDesc[OpTraceRayKHR].operands.push(OperandId, "'Acceleration Structure'"); + InstructionDesc[OpTraceRayKHR].operands.push(OperandId, "'Ray Flags'"); + InstructionDesc[OpTraceRayKHR].operands.push(OperandId, "'Cull Mask'"); + InstructionDesc[OpTraceRayKHR].operands.push(OperandId, "'SBT Record Offset'"); + InstructionDesc[OpTraceRayKHR].operands.push(OperandId, "'SBT Record Stride'"); + InstructionDesc[OpTraceRayKHR].operands.push(OperandId, "'Miss Index'"); + InstructionDesc[OpTraceRayKHR].operands.push(OperandId, "'Ray Origin'"); + InstructionDesc[OpTraceRayKHR].operands.push(OperandId, "'TMin'"); + InstructionDesc[OpTraceRayKHR].operands.push(OperandId, "'Ray Direction'"); + InstructionDesc[OpTraceRayKHR].operands.push(OperandId, "'TMax'"); + InstructionDesc[OpTraceRayKHR].operands.push(OperandId, "'Payload'"); + InstructionDesc[OpTraceRayKHR].setResultAndType(false, false); + + InstructionDesc[OpReportIntersectionKHR].operands.push(OperandId, "'Hit Parameter'"); + InstructionDesc[OpReportIntersectionKHR].operands.push(OperandId, "'Hit Kind'"); InstructionDesc[OpIgnoreIntersectionNV].setResultAndType(false, false); + InstructionDesc[OpIgnoreIntersectionKHR].setResultAndType(false, false); + InstructionDesc[OpTerminateRayNV].setResultAndType(false, false); + + InstructionDesc[OpTerminateRayKHR].setResultAndType(false, false); InstructionDesc[OpExecuteCallableNV].operands.push(OperandId, "SBT Record Index"); InstructionDesc[OpExecuteCallableNV].operands.push(OperandId, "CallableData ID"); InstructionDesc[OpExecuteCallableNV].setResultAndType(false, false); + InstructionDesc[OpExecuteCallableKHR].operands.push(OperandId, "SBT Record Index"); + InstructionDesc[OpExecuteCallableKHR].operands.push(OperandId, "CallableData"); + InstructionDesc[OpExecuteCallableKHR].setResultAndType(false, false); + + InstructionDesc[OpConvertUToAccelerationStructureKHR].operands.push(OperandId, "Value"); + InstructionDesc[OpConvertUToAccelerationStructureKHR].setResultAndType(true, true); + + // Ray Query + InstructionDesc[OpTypeAccelerationStructureKHR].setResultAndType(true, false); + InstructionDesc[OpTypeRayQueryKHR].setResultAndType(true, false); + + InstructionDesc[OpRayQueryInitializeKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryInitializeKHR].operands.push(OperandId, "'AccelerationS'"); + InstructionDesc[OpRayQueryInitializeKHR].operands.push(OperandId, "'RayFlags'"); + InstructionDesc[OpRayQueryInitializeKHR].operands.push(OperandId, "'CullMask'"); + InstructionDesc[OpRayQueryInitializeKHR].operands.push(OperandId, "'Origin'"); + InstructionDesc[OpRayQueryInitializeKHR].operands.push(OperandId, "'Tmin'"); + InstructionDesc[OpRayQueryInitializeKHR].operands.push(OperandId, "'Direction'"); + InstructionDesc[OpRayQueryInitializeKHR].operands.push(OperandId, "'Tmax'"); + InstructionDesc[OpRayQueryInitializeKHR].setResultAndType(false, false); + + InstructionDesc[OpRayQueryTerminateKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryTerminateKHR].setResultAndType(false, false); + + InstructionDesc[OpRayQueryGenerateIntersectionKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGenerateIntersectionKHR].operands.push(OperandId, "'THit'"); + InstructionDesc[OpRayQueryGenerateIntersectionKHR].setResultAndType(false, false); + + InstructionDesc[OpRayQueryConfirmIntersectionKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryConfirmIntersectionKHR].setResultAndType(false, false); + + InstructionDesc[OpRayQueryProceedKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryProceedKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionTypeKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionTypeKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionTypeKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetRayTMinKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetRayTMinKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetRayFlagsKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetRayFlagsKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionTKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionTKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionTKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionInstanceCustomIndexKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionInstanceCustomIndexKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionInstanceCustomIndexKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionInstanceIdKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionInstanceIdKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionInstanceIdKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionGeometryIndexKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionGeometryIndexKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionGeometryIndexKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionPrimitiveIndexKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionPrimitiveIndexKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionPrimitiveIndexKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionBarycentricsKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionBarycentricsKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionBarycentricsKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionFrontFaceKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionFrontFaceKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionFrontFaceKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionCandidateAABBOpaqueKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionCandidateAABBOpaqueKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionObjectRayDirectionKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionObjectRayDirectionKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionObjectRayDirectionKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionObjectRayOriginKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionObjectRayOriginKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionObjectRayOriginKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetWorldRayDirectionKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetWorldRayDirectionKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetWorldRayOriginKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetWorldRayOriginKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionObjectToWorldKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionObjectToWorldKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionObjectToWorldKHR].setResultAndType(true, true); + + InstructionDesc[OpRayQueryGetIntersectionWorldToObjectKHR].operands.push(OperandId, "'RayQuery'"); + InstructionDesc[OpRayQueryGetIntersectionWorldToObjectKHR].operands.push(OperandId, "'Committed'"); + InstructionDesc[OpRayQueryGetIntersectionWorldToObjectKHR].setResultAndType(true, true); + InstructionDesc[OpImageSampleFootprintNV].operands.push(OperandId, "'Sampled Image'"); InstructionDesc[OpImageSampleFootprintNV].operands.push(OperandId, "'Coordinate'"); InstructionDesc[OpImageSampleFootprintNV].operands.push(OperandId, "'Granularity'"); diff --git a/thirdparty/glslang/SPIRV/doc.h b/thirdparty/glslang/SPIRV/doc.h index 293256a2c6..2a0b28c6b3 100644 --- a/thirdparty/glslang/SPIRV/doc.h +++ b/thirdparty/glslang/SPIRV/doc.h @@ -125,6 +125,7 @@ enum OperandClass { OperandVariableLiteralId, OperandLiteralNumber, OperandLiteralString, + OperandVariableLiteralStrings, OperandSource, OperandExecutionModel, OperandAddressing, diff --git a/thirdparty/glslang/SPIRV/hex_float.h b/thirdparty/glslang/SPIRV/hex_float.h index 905b21a45a..8be8e9f7e3 100644 --- a/thirdparty/glslang/SPIRV/hex_float.h +++ b/thirdparty/glslang/SPIRV/hex_float.h @@ -784,8 +784,8 @@ inline std::istream& ParseNormalFloat(std::istream& is, bool negate_value, if (val.isInfinity()) { // Fail the parse. Emulate standard behaviour by setting the value to // the closest normal value, and set the fail bit on the stream. - value.set_value((value.isNegative() | negate_value) ? T::lowest() - : T::max()); + value.set_value((value.isNegative() || negate_value) ? T::lowest() + : T::max()); is.setstate(std::ios_base::failbit); } return is; diff --git a/thirdparty/glslang/SPIRV/spirv.hpp b/thirdparty/glslang/SPIRV/spirv.hpp index 1e96f7b4a9..e0fe24980d 100644 --- a/thirdparty/glslang/SPIRV/spirv.hpp +++ b/thirdparty/glslang/SPIRV/spirv.hpp @@ -1,4 +1,4 @@ -// Copyright (c) 2014-2019 The Khronos Group Inc. +// Copyright (c) 2014-2020 The Khronos Group Inc. // // Permission is hereby granted, free of charge, to any person obtaining a copy // of this software and/or associated documentation files (the "Materials"), @@ -49,12 +49,12 @@ namespace spv { typedef unsigned int Id; -#define SPV_VERSION 0x10400 -#define SPV_REVISION 1 +#define SPV_VERSION 0x10500 +#define SPV_REVISION 4 static const unsigned int MagicNumber = 0x07230203; -static const unsigned int Version = 0x00010400; -static const unsigned int Revision = 1; +static const unsigned int Version = 0x00010500; +static const unsigned int Revision = 4; static const unsigned int OpCodeMask = 0xffff; static const unsigned int WordCountShift = 16; @@ -78,11 +78,17 @@ enum ExecutionModel { ExecutionModelKernel = 6, ExecutionModelTaskNV = 5267, ExecutionModelMeshNV = 5268, + ExecutionModelRayGenerationKHR = 5313, ExecutionModelRayGenerationNV = 5313, + ExecutionModelIntersectionKHR = 5314, ExecutionModelIntersectionNV = 5314, + ExecutionModelAnyHitKHR = 5315, ExecutionModelAnyHitNV = 5315, + ExecutionModelClosestHitKHR = 5316, ExecutionModelClosestHitNV = 5316, + ExecutionModelMissKHR = 5317, ExecutionModelMissNV = 5317, + ExecutionModelCallableKHR = 5318, ExecutionModelCallableNV = 5318, ExecutionModelMax = 0x7fffffff, }; @@ -144,6 +150,7 @@ enum ExecutionMode { ExecutionModeSubgroupsPerWorkgroupId = 37, ExecutionModeLocalSizeId = 38, ExecutionModeLocalSizeHintId = 39, + ExecutionModeSubgroupUniformControlFlowKHR = 4421, ExecutionModePostDepthCoverage = 4446, ExecutionModeDenormPreserve = 4459, ExecutionModeDenormFlushToZero = 4460, @@ -162,6 +169,16 @@ enum ExecutionMode { ExecutionModeSampleInterlockUnorderedEXT = 5369, ExecutionModeShadingRateInterlockOrderedEXT = 5370, ExecutionModeShadingRateInterlockUnorderedEXT = 5371, + ExecutionModeSharedLocalMemorySizeINTEL = 5618, + ExecutionModeRoundingModeRTPINTEL = 5620, + ExecutionModeRoundingModeRTNINTEL = 5621, + ExecutionModeFloatingPointModeALTINTEL = 5622, + ExecutionModeFloatingPointModeIEEEINTEL = 5623, + ExecutionModeMaxWorkgroupSizeINTEL = 5893, + ExecutionModeMaxWorkDimINTEL = 5894, + ExecutionModeNoGlobalOffsetINTEL = 5895, + ExecutionModeNumSIMDWorkitemsINTEL = 5896, + ExecutionModeSchedulerTargetFmaxMhzINTEL = 5903, ExecutionModeMax = 0x7fffffff, }; @@ -179,14 +196,23 @@ enum StorageClass { StorageClassAtomicCounter = 10, StorageClassImage = 11, StorageClassStorageBuffer = 12, + StorageClassCallableDataKHR = 5328, StorageClassCallableDataNV = 5328, + StorageClassIncomingCallableDataKHR = 5329, StorageClassIncomingCallableDataNV = 5329, + StorageClassRayPayloadKHR = 5338, StorageClassRayPayloadNV = 5338, + StorageClassHitAttributeKHR = 5339, StorageClassHitAttributeNV = 5339, + StorageClassIncomingRayPayloadKHR = 5342, StorageClassIncomingRayPayloadNV = 5342, + StorageClassShaderRecordBufferKHR = 5343, StorageClassShaderRecordBufferNV = 5343, StorageClassPhysicalStorageBuffer = 5349, StorageClassPhysicalStorageBufferEXT = 5349, + StorageClassCodeSectionINTEL = 5605, + StorageClassDeviceOnlyINTEL = 5936, + StorageClassHostOnlyINTEL = 5937, StorageClassMax = 0x7fffffff, }; @@ -257,6 +283,8 @@ enum ImageFormat { ImageFormatRg8ui = 37, ImageFormatR16ui = 38, ImageFormatR8ui = 39, + ImageFormatR64ui = 40, + ImageFormatR64i = 41, ImageFormatMax = 0x7fffffff, }; @@ -355,6 +383,8 @@ enum FPFastMathModeShift { FPFastMathModeNSZShift = 2, FPFastMathModeAllowRecipShift = 3, FPFastMathModeFastShift = 4, + FPFastMathModeAllowContractFastINTELShift = 16, + FPFastMathModeAllowReassocINTELShift = 17, FPFastMathModeMax = 0x7fffffff, }; @@ -365,6 +395,8 @@ enum FPFastMathModeMask { FPFastMathModeNSZMask = 0x00000004, FPFastMathModeAllowRecipMask = 0x00000008, FPFastMathModeFastMask = 0x00000010, + FPFastMathModeAllowContractFastINTELMask = 0x00010000, + FPFastMathModeAllowReassocINTELMask = 0x00020000, }; enum FPRoundingMode { @@ -378,6 +410,7 @@ enum FPRoundingMode { enum LinkageType { LinkageTypeExport = 0, LinkageTypeImport = 1, + LinkageTypeLinkOnceODR = 2, LinkageTypeMax = 0x7fffffff, }; @@ -465,11 +498,45 @@ enum Decoration { DecorationRestrictPointerEXT = 5355, DecorationAliasedPointer = 5356, DecorationAliasedPointerEXT = 5356, + DecorationSIMTCallINTEL = 5599, + DecorationReferencedIndirectlyINTEL = 5602, + DecorationClobberINTEL = 5607, + DecorationSideEffectsINTEL = 5608, + DecorationVectorComputeVariableINTEL = 5624, + DecorationFuncParamIOKindINTEL = 5625, + DecorationVectorComputeFunctionINTEL = 5626, + DecorationStackCallINTEL = 5627, + DecorationGlobalVariableOffsetINTEL = 5628, DecorationCounterBuffer = 5634, DecorationHlslCounterBufferGOOGLE = 5634, DecorationHlslSemanticGOOGLE = 5635, DecorationUserSemantic = 5635, DecorationUserTypeGOOGLE = 5636, + DecorationFunctionRoundingModeINTEL = 5822, + DecorationFunctionDenormModeINTEL = 5823, + DecorationRegisterINTEL = 5825, + DecorationMemoryINTEL = 5826, + DecorationNumbanksINTEL = 5827, + DecorationBankwidthINTEL = 5828, + DecorationMaxPrivateCopiesINTEL = 5829, + DecorationSinglepumpINTEL = 5830, + DecorationDoublepumpINTEL = 5831, + DecorationMaxReplicatesINTEL = 5832, + DecorationSimpleDualPortINTEL = 5833, + DecorationMergeINTEL = 5834, + DecorationBankBitsINTEL = 5835, + DecorationForcePow2DepthINTEL = 5836, + DecorationBurstCoalesceINTEL = 5899, + DecorationCacheSizeINTEL = 5900, + DecorationDontStaticallyCoalesceINTEL = 5901, + DecorationPrefetchINTEL = 5902, + DecorationStallEnableINTEL = 5905, + DecorationFuseLoopsInFunctionINTEL = 5907, + DecorationBufferLocationINTEL = 5921, + DecorationIOPipeStorageINTEL = 5944, + DecorationFunctionFloatingPointModeINTEL = 6080, + DecorationSingleElementVectorINTEL = 6085, + DecorationVectorComputeCallableFunctionINTEL = 6087, DecorationMax = 0x7fffffff, }; @@ -528,8 +595,10 @@ enum BuiltIn { BuiltInBaseVertex = 4424, BuiltInBaseInstance = 4425, BuiltInDrawIndex = 4426, + BuiltInPrimitiveShadingRateKHR = 4432, BuiltInDeviceIndex = 4438, BuiltInViewIndex = 4440, + BuiltInShadingRateKHR = 4444, BuiltInBaryCoordNoPerspAMD = 4992, BuiltInBaryCoordNoPerspCentroidAMD = 4993, BuiltInBaryCoordNoPerspSampleAMD = 4994, @@ -558,20 +627,35 @@ enum BuiltIn { BuiltInFragmentSizeNV = 5292, BuiltInFragInvocationCountEXT = 5293, BuiltInInvocationsPerPixelNV = 5293, + BuiltInLaunchIdKHR = 5319, BuiltInLaunchIdNV = 5319, + BuiltInLaunchSizeKHR = 5320, BuiltInLaunchSizeNV = 5320, + BuiltInWorldRayOriginKHR = 5321, BuiltInWorldRayOriginNV = 5321, + BuiltInWorldRayDirectionKHR = 5322, BuiltInWorldRayDirectionNV = 5322, + BuiltInObjectRayOriginKHR = 5323, BuiltInObjectRayOriginNV = 5323, + BuiltInObjectRayDirectionKHR = 5324, BuiltInObjectRayDirectionNV = 5324, + BuiltInRayTminKHR = 5325, BuiltInRayTminNV = 5325, + BuiltInRayTmaxKHR = 5326, BuiltInRayTmaxNV = 5326, + BuiltInInstanceCustomIndexKHR = 5327, BuiltInInstanceCustomIndexNV = 5327, + BuiltInObjectToWorldKHR = 5330, BuiltInObjectToWorldNV = 5330, + BuiltInWorldToObjectKHR = 5331, BuiltInWorldToObjectNV = 5331, BuiltInHitTNV = 5332, + BuiltInHitKindKHR = 5333, BuiltInHitKindNV = 5333, + BuiltInCurrentRayTimeNV = 5334, + BuiltInIncomingRayFlagsKHR = 5351, BuiltInIncomingRayFlagsNV = 5351, + BuiltInRayGeometryIndexKHR = 5352, BuiltInWarpsPerSMNV = 5374, BuiltInSMCountNV = 5375, BuiltInWarpIDNV = 5376, @@ -601,6 +685,14 @@ enum LoopControlShift { LoopControlIterationMultipleShift = 6, LoopControlPeelCountShift = 7, LoopControlPartialCountShift = 8, + LoopControlInitiationIntervalINTELShift = 16, + LoopControlMaxConcurrencyINTELShift = 17, + LoopControlDependencyArrayINTELShift = 18, + LoopControlPipelineEnableINTELShift = 19, + LoopControlLoopCoalesceINTELShift = 20, + LoopControlMaxInterleavingINTELShift = 21, + LoopControlSpeculatedIterationsINTELShift = 22, + LoopControlNoFusionINTELShift = 23, LoopControlMax = 0x7fffffff, }; @@ -615,6 +707,14 @@ enum LoopControlMask { LoopControlIterationMultipleMask = 0x00000040, LoopControlPeelCountMask = 0x00000080, LoopControlPartialCountMask = 0x00000100, + LoopControlInitiationIntervalINTELMask = 0x00010000, + LoopControlMaxConcurrencyINTELMask = 0x00020000, + LoopControlDependencyArrayINTELMask = 0x00040000, + LoopControlPipelineEnableINTELMask = 0x00080000, + LoopControlLoopCoalesceINTELMask = 0x00100000, + LoopControlMaxInterleavingINTELMask = 0x00200000, + LoopControlSpeculatedIterationsINTELMask = 0x00400000, + LoopControlNoFusionINTELMask = 0x00800000, }; enum FunctionControlShift { @@ -709,6 +809,7 @@ enum Scope { ScopeInvocation = 4, ScopeQueueFamily = 5, ScopeQueueFamilyKHR = 5, + ScopeShaderCallKHR = 6, ScopeMax = 0x7fffffff, }; @@ -810,8 +911,12 @@ enum Capability { CapabilityGroupNonUniformQuad = 68, CapabilityShaderLayer = 69, CapabilityShaderViewportIndex = 70, + CapabilityFragmentShadingRateKHR = 4422, CapabilitySubgroupBallotKHR = 4423, CapabilityDrawParameters = 4427, + CapabilityWorkgroupMemoryExplicitLayoutKHR = 4428, + CapabilityWorkgroupMemoryExplicitLayout8BitAccessKHR = 4429, + CapabilityWorkgroupMemoryExplicitLayout16BitAccessKHR = 4430, CapabilitySubgroupVoteKHR = 4431, CapabilityStorageBuffer16BitAccess = 4433, CapabilityStorageUniformBufferBlock16 = 4433, @@ -833,11 +938,16 @@ enum Capability { CapabilitySignedZeroInfNanPreserve = 4466, CapabilityRoundingModeRTE = 4467, CapabilityRoundingModeRTZ = 4468, + CapabilityRayQueryProvisionalKHR = 4471, + CapabilityRayQueryKHR = 4472, + CapabilityRayTraversalPrimitiveCullingKHR = 4478, + CapabilityRayTracingKHR = 4479, CapabilityFloat16ImageAMD = 5008, CapabilityImageGatherBiasLodAMD = 5009, CapabilityFragmentMaskAMD = 5010, CapabilityStencilExportEXT = 5013, CapabilityImageReadWriteLodAMD = 5015, + CapabilityInt64ImageEXT = 5016, CapabilityShaderClockKHR = 5055, CapabilitySampleMaskOverrideCoverageNV = 5249, CapabilityGeometryShaderPassthroughNV = 5251, @@ -879,6 +989,7 @@ enum Capability { CapabilityStorageTexelBufferArrayNonUniformIndexing = 5312, CapabilityStorageTexelBufferArrayNonUniformIndexingEXT = 5312, CapabilityRayTracingNV = 5340, + CapabilityRayTracingMotionBlurNV = 5341, CapabilityVulkanMemoryModel = 5345, CapabilityVulkanMemoryModelKHR = 5345, CapabilityVulkanMemoryModelDeviceScope = 5346, @@ -886,6 +997,7 @@ enum Capability { CapabilityPhysicalStorageBufferAddresses = 5347, CapabilityPhysicalStorageBufferAddressesEXT = 5347, CapabilityComputeDerivativeGroupLinearNV = 5350, + CapabilityRayTracingProvisionalKHR = 5353, CapabilityCooperativeMatrixNV = 5357, CapabilityFragmentShaderSampleInterlockEXT = 5363, CapabilityFragmentShaderShadingRateInterlockEXT = 5372, @@ -896,13 +1008,120 @@ enum Capability { CapabilitySubgroupBufferBlockIOINTEL = 5569, CapabilitySubgroupImageBlockIOINTEL = 5570, CapabilitySubgroupImageMediaBlockIOINTEL = 5579, + CapabilityRoundToInfinityINTEL = 5582, + CapabilityFloatingPointModeINTEL = 5583, CapabilityIntegerFunctions2INTEL = 5584, + CapabilityFunctionPointersINTEL = 5603, + CapabilityIndirectReferencesINTEL = 5604, + CapabilityAsmINTEL = 5606, + CapabilityAtomicFloat32MinMaxEXT = 5612, + CapabilityAtomicFloat64MinMaxEXT = 5613, + CapabilityAtomicFloat16MinMaxEXT = 5616, + CapabilityVectorComputeINTEL = 5617, + CapabilityVectorAnyINTEL = 5619, + CapabilityExpectAssumeKHR = 5629, CapabilitySubgroupAvcMotionEstimationINTEL = 5696, CapabilitySubgroupAvcMotionEstimationIntraINTEL = 5697, CapabilitySubgroupAvcMotionEstimationChromaINTEL = 5698, + CapabilityVariableLengthArrayINTEL = 5817, + CapabilityFunctionFloatControlINTEL = 5821, + CapabilityFPGAMemoryAttributesINTEL = 5824, + CapabilityFPFastMathModeINTEL = 5837, + CapabilityArbitraryPrecisionIntegersINTEL = 5844, + CapabilityUnstructuredLoopControlsINTEL = 5886, + CapabilityFPGALoopControlsINTEL = 5888, + CapabilityKernelAttributesINTEL = 5892, + CapabilityFPGAKernelAttributesINTEL = 5897, + CapabilityFPGAMemoryAccessesINTEL = 5898, + CapabilityFPGAClusterAttributesINTEL = 5904, + CapabilityLoopFuseINTEL = 5906, + CapabilityFPGABufferLocationINTEL = 5920, + CapabilityUSMStorageClassesINTEL = 5935, + CapabilityIOPipesINTEL = 5943, + CapabilityBlockingPipesINTEL = 5945, + CapabilityFPGARegINTEL = 5948, + CapabilityAtomicFloat32AddEXT = 6033, + CapabilityAtomicFloat64AddEXT = 6034, + CapabilityLongConstantCompositeINTEL = 6089, + CapabilityAtomicFloat16AddEXT = 6095, CapabilityMax = 0x7fffffff, }; +enum RayFlagsShift { + RayFlagsOpaqueKHRShift = 0, + RayFlagsNoOpaqueKHRShift = 1, + RayFlagsTerminateOnFirstHitKHRShift = 2, + RayFlagsSkipClosestHitShaderKHRShift = 3, + RayFlagsCullBackFacingTrianglesKHRShift = 4, + RayFlagsCullFrontFacingTrianglesKHRShift = 5, + RayFlagsCullOpaqueKHRShift = 6, + RayFlagsCullNoOpaqueKHRShift = 7, + RayFlagsSkipTrianglesKHRShift = 8, + RayFlagsSkipAABBsKHRShift = 9, + RayFlagsMax = 0x7fffffff, +}; + +enum RayFlagsMask { + RayFlagsMaskNone = 0, + RayFlagsOpaqueKHRMask = 0x00000001, + RayFlagsNoOpaqueKHRMask = 0x00000002, + RayFlagsTerminateOnFirstHitKHRMask = 0x00000004, + RayFlagsSkipClosestHitShaderKHRMask = 0x00000008, + RayFlagsCullBackFacingTrianglesKHRMask = 0x00000010, + RayFlagsCullFrontFacingTrianglesKHRMask = 0x00000020, + RayFlagsCullOpaqueKHRMask = 0x00000040, + RayFlagsCullNoOpaqueKHRMask = 0x00000080, + RayFlagsSkipTrianglesKHRMask = 0x00000100, + RayFlagsSkipAABBsKHRMask = 0x00000200, +}; + +enum RayQueryIntersection { + RayQueryIntersectionRayQueryCandidateIntersectionKHR = 0, + RayQueryIntersectionRayQueryCommittedIntersectionKHR = 1, + RayQueryIntersectionMax = 0x7fffffff, +}; + +enum RayQueryCommittedIntersectionType { + RayQueryCommittedIntersectionTypeRayQueryCommittedIntersectionNoneKHR = 0, + RayQueryCommittedIntersectionTypeRayQueryCommittedIntersectionTriangleKHR = 1, + RayQueryCommittedIntersectionTypeRayQueryCommittedIntersectionGeneratedKHR = 2, + RayQueryCommittedIntersectionTypeMax = 0x7fffffff, +}; + +enum RayQueryCandidateIntersectionType { + RayQueryCandidateIntersectionTypeRayQueryCandidateIntersectionTriangleKHR = 0, + RayQueryCandidateIntersectionTypeRayQueryCandidateIntersectionAABBKHR = 1, + RayQueryCandidateIntersectionTypeMax = 0x7fffffff, +}; + +enum FragmentShadingRateShift { + FragmentShadingRateVertical2PixelsShift = 0, + FragmentShadingRateVertical4PixelsShift = 1, + FragmentShadingRateHorizontal2PixelsShift = 2, + FragmentShadingRateHorizontal4PixelsShift = 3, + FragmentShadingRateMax = 0x7fffffff, +}; + +enum FragmentShadingRateMask { + FragmentShadingRateMaskNone = 0, + FragmentShadingRateVertical2PixelsMask = 0x00000001, + FragmentShadingRateVertical4PixelsMask = 0x00000002, + FragmentShadingRateHorizontal2PixelsMask = 0x00000004, + FragmentShadingRateHorizontal4PixelsMask = 0x00000008, +}; + +enum FPDenormMode { + FPDenormModePreserve = 0, + FPDenormModeFlushToZero = 1, + FPDenormModeMax = 0x7fffffff, +}; + +enum FPOperationMode { + FPOperationModeIEEE = 0, + FPOperationModeALT = 1, + FPOperationModeMax = 0x7fffffff, +}; + enum Op { OpNop = 0, OpUndef = 1, @@ -1248,12 +1467,25 @@ enum Op { OpPtrEqual = 401, OpPtrNotEqual = 402, OpPtrDiff = 403, + OpTerminateInvocation = 4416, OpSubgroupBallotKHR = 4421, OpSubgroupFirstInvocationKHR = 4422, OpSubgroupAllKHR = 4428, OpSubgroupAnyKHR = 4429, OpSubgroupAllEqualKHR = 4430, OpSubgroupReadInvocationKHR = 4432, + OpTraceRayKHR = 4445, + OpExecuteCallableKHR = 4446, + OpConvertUToAccelerationStructureKHR = 4447, + OpIgnoreIntersectionKHR = 4448, + OpTerminateRayKHR = 4449, + OpTypeRayQueryKHR = 4472, + OpRayQueryInitializeKHR = 4473, + OpRayQueryTerminateKHR = 4474, + OpRayQueryGenerateIntersectionKHR = 4475, + OpRayQueryConfirmIntersectionKHR = 4476, + OpRayQueryProceedKHR = 4477, + OpRayQueryGetIntersectionTypeKHR = 4479, OpGroupIAddNonUniformAMD = 5000, OpGroupFAddNonUniformAMD = 5001, OpGroupFMinNonUniformAMD = 5002, @@ -1268,10 +1500,14 @@ enum Op { OpImageSampleFootprintNV = 5283, OpGroupNonUniformPartitionNV = 5296, OpWritePackedPrimitiveIndices4x8NV = 5299, + OpReportIntersectionKHR = 5334, OpReportIntersectionNV = 5334, OpIgnoreIntersectionNV = 5335, OpTerminateRayNV = 5336, OpTraceNV = 5337, + OpTraceMotionNV = 5338, + OpTraceRayMotionNV = 5339, + OpTypeAccelerationStructureKHR = 5341, OpTypeAccelerationStructureNV = 5341, OpExecuteCallableNV = 5344, OpTypeCooperativeMatrixNV = 5358, @@ -1307,6 +1543,15 @@ enum Op { OpUSubSatINTEL = 5596, OpIMul32x16INTEL = 5597, OpUMul32x16INTEL = 5598, + OpConstFunctionPointerINTEL = 5600, + OpFunctionPointerCallINTEL = 5601, + OpAsmTargetINTEL = 5609, + OpAsmINTEL = 5610, + OpAsmCallINTEL = 5611, + OpAtomicFMinEXT = 5614, + OpAtomicFMaxEXT = 5615, + OpAssumeTrueKHR = 5630, + OpExpectKHR = 5631, OpDecorateString = 5632, OpDecorateStringGOOGLE = 5632, OpMemberDecorateString = 5633, @@ -1429,6 +1674,37 @@ enum Op { OpSubgroupAvcSicGetPackedSkcLumaCountThresholdINTEL = 5814, OpSubgroupAvcSicGetPackedSkcLumaSumThresholdINTEL = 5815, OpSubgroupAvcSicGetInterRawSadsINTEL = 5816, + OpVariableLengthArrayINTEL = 5818, + OpSaveMemoryINTEL = 5819, + OpRestoreMemoryINTEL = 5820, + OpLoopControlINTEL = 5887, + OpPtrCastToCrossWorkgroupINTEL = 5934, + OpCrossWorkgroupCastToPtrINTEL = 5938, + OpReadPipeBlockingINTEL = 5946, + OpWritePipeBlockingINTEL = 5947, + OpFPGARegINTEL = 5949, + OpRayQueryGetRayTMinKHR = 6016, + OpRayQueryGetRayFlagsKHR = 6017, + OpRayQueryGetIntersectionTKHR = 6018, + OpRayQueryGetIntersectionInstanceCustomIndexKHR = 6019, + OpRayQueryGetIntersectionInstanceIdKHR = 6020, + OpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR = 6021, + OpRayQueryGetIntersectionGeometryIndexKHR = 6022, + OpRayQueryGetIntersectionPrimitiveIndexKHR = 6023, + OpRayQueryGetIntersectionBarycentricsKHR = 6024, + OpRayQueryGetIntersectionFrontFaceKHR = 6025, + OpRayQueryGetIntersectionCandidateAABBOpaqueKHR = 6026, + OpRayQueryGetIntersectionObjectRayDirectionKHR = 6027, + OpRayQueryGetIntersectionObjectRayOriginKHR = 6028, + OpRayQueryGetWorldRayDirectionKHR = 6029, + OpRayQueryGetWorldRayOriginKHR = 6030, + OpRayQueryGetIntersectionObjectToWorldKHR = 6031, + OpRayQueryGetIntersectionWorldToObjectKHR = 6032, + OpAtomicFAddEXT = 6035, + OpTypeBufferSurfaceINTEL = 6086, + OpTypeStructContinuedINTEL = 6090, + OpConstantCompositeContinuedINTEL = 6091, + OpSpecConstantCompositeContinuedINTEL = 6092, OpMax = 0x7fffffff, }; @@ -1781,12 +2057,25 @@ inline void HasResultAndType(Op opcode, bool *hasResult, bool *hasResultType) { case OpPtrEqual: *hasResult = true; *hasResultType = true; break; case OpPtrNotEqual: *hasResult = true; *hasResultType = true; break; case OpPtrDiff: *hasResult = true; *hasResultType = true; break; + case OpTerminateInvocation: *hasResult = false; *hasResultType = false; break; case OpSubgroupBallotKHR: *hasResult = true; *hasResultType = true; break; case OpSubgroupFirstInvocationKHR: *hasResult = true; *hasResultType = true; break; case OpSubgroupAllKHR: *hasResult = true; *hasResultType = true; break; case OpSubgroupAnyKHR: *hasResult = true; *hasResultType = true; break; case OpSubgroupAllEqualKHR: *hasResult = true; *hasResultType = true; break; case OpSubgroupReadInvocationKHR: *hasResult = true; *hasResultType = true; break; + case OpTraceRayKHR: *hasResult = false; *hasResultType = false; break; + case OpExecuteCallableKHR: *hasResult = false; *hasResultType = false; break; + case OpConvertUToAccelerationStructureKHR: *hasResult = true; *hasResultType = true; break; + case OpIgnoreIntersectionKHR: *hasResult = false; *hasResultType = false; break; + case OpTerminateRayKHR: *hasResult = false; *hasResultType = false; break; + case OpTypeRayQueryKHR: *hasResult = true; *hasResultType = false; break; + case OpRayQueryInitializeKHR: *hasResult = false; *hasResultType = false; break; + case OpRayQueryTerminateKHR: *hasResult = false; *hasResultType = false; break; + case OpRayQueryGenerateIntersectionKHR: *hasResult = false; *hasResultType = false; break; + case OpRayQueryConfirmIntersectionKHR: *hasResult = false; *hasResultType = false; break; + case OpRayQueryProceedKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionTypeKHR: *hasResult = true; *hasResultType = true; break; case OpGroupIAddNonUniformAMD: *hasResult = true; *hasResultType = true; break; case OpGroupFAddNonUniformAMD: *hasResult = true; *hasResultType = true; break; case OpGroupFMinNonUniformAMD: *hasResult = true; *hasResultType = true; break; @@ -1805,6 +2094,8 @@ inline void HasResultAndType(Op opcode, bool *hasResult, bool *hasResultType) { case OpIgnoreIntersectionNV: *hasResult = false; *hasResultType = false; break; case OpTerminateRayNV: *hasResult = false; *hasResultType = false; break; case OpTraceNV: *hasResult = false; *hasResultType = false; break; + case OpTraceMotionNV: *hasResult = false; *hasResultType = false; break; + case OpTraceRayMotionNV: *hasResult = false; *hasResultType = false; break; case OpTypeAccelerationStructureNV: *hasResult = true; *hasResultType = false; break; case OpExecuteCallableNV: *hasResult = false; *hasResultType = false; break; case OpTypeCooperativeMatrixNV: *hasResult = true; *hasResultType = false; break; @@ -1840,6 +2131,15 @@ inline void HasResultAndType(Op opcode, bool *hasResult, bool *hasResultType) { case OpUSubSatINTEL: *hasResult = true; *hasResultType = true; break; case OpIMul32x16INTEL: *hasResult = true; *hasResultType = true; break; case OpUMul32x16INTEL: *hasResult = true; *hasResultType = true; break; + case OpConstFunctionPointerINTEL: *hasResult = true; *hasResultType = true; break; + case OpFunctionPointerCallINTEL: *hasResult = true; *hasResultType = true; break; + case OpAsmTargetINTEL: *hasResult = true; *hasResultType = true; break; + case OpAsmINTEL: *hasResult = true; *hasResultType = true; break; + case OpAsmCallINTEL: *hasResult = true; *hasResultType = true; break; + case OpAtomicFMinEXT: *hasResult = true; *hasResultType = true; break; + case OpAtomicFMaxEXT: *hasResult = true; *hasResultType = true; break; + case OpAssumeTrueKHR: *hasResult = false; *hasResultType = false; break; + case OpExpectKHR: *hasResult = true; *hasResultType = true; break; case OpDecorateString: *hasResult = false; *hasResultType = false; break; case OpMemberDecorateString: *hasResult = false; *hasResultType = false; break; case OpVmeImageINTEL: *hasResult = true; *hasResultType = true; break; @@ -1960,6 +2260,37 @@ inline void HasResultAndType(Op opcode, bool *hasResult, bool *hasResultType) { case OpSubgroupAvcSicGetPackedSkcLumaCountThresholdINTEL: *hasResult = true; *hasResultType = true; break; case OpSubgroupAvcSicGetPackedSkcLumaSumThresholdINTEL: *hasResult = true; *hasResultType = true; break; case OpSubgroupAvcSicGetInterRawSadsINTEL: *hasResult = true; *hasResultType = true; break; + case OpVariableLengthArrayINTEL: *hasResult = true; *hasResultType = true; break; + case OpSaveMemoryINTEL: *hasResult = true; *hasResultType = true; break; + case OpRestoreMemoryINTEL: *hasResult = false; *hasResultType = false; break; + case OpLoopControlINTEL: *hasResult = false; *hasResultType = false; break; + case OpPtrCastToCrossWorkgroupINTEL: *hasResult = true; *hasResultType = true; break; + case OpCrossWorkgroupCastToPtrINTEL: *hasResult = true; *hasResultType = true; break; + case OpReadPipeBlockingINTEL: *hasResult = true; *hasResultType = true; break; + case OpWritePipeBlockingINTEL: *hasResult = true; *hasResultType = true; break; + case OpFPGARegINTEL: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetRayTMinKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetRayFlagsKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionTKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionInstanceCustomIndexKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionInstanceIdKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionGeometryIndexKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionPrimitiveIndexKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionBarycentricsKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionFrontFaceKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionCandidateAABBOpaqueKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionObjectRayDirectionKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionObjectRayOriginKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetWorldRayDirectionKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetWorldRayOriginKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionObjectToWorldKHR: *hasResult = true; *hasResultType = true; break; + case OpRayQueryGetIntersectionWorldToObjectKHR: *hasResult = true; *hasResultType = true; break; + case OpAtomicFAddEXT: *hasResult = true; *hasResultType = true; break; + case OpTypeBufferSurfaceINTEL: *hasResult = true; *hasResultType = false; break; + case OpTypeStructContinuedINTEL: *hasResult = false; *hasResultType = false; break; + case OpConstantCompositeContinuedINTEL: *hasResult = false; *hasResultType = false; break; + case OpSpecConstantCompositeContinuedINTEL: *hasResult = false; *hasResultType = false; break; } } #endif /* SPV_ENABLE_UTILITY_CODE */ @@ -1974,6 +2305,8 @@ inline FunctionControlMask operator|(FunctionControlMask a, FunctionControlMask inline MemorySemanticsMask operator|(MemorySemanticsMask a, MemorySemanticsMask b) { return MemorySemanticsMask(unsigned(a) | unsigned(b)); } inline MemoryAccessMask operator|(MemoryAccessMask a, MemoryAccessMask b) { return MemoryAccessMask(unsigned(a) | unsigned(b)); } inline KernelProfilingInfoMask operator|(KernelProfilingInfoMask a, KernelProfilingInfoMask b) { return KernelProfilingInfoMask(unsigned(a) | unsigned(b)); } +inline RayFlagsMask operator|(RayFlagsMask a, RayFlagsMask b) { return RayFlagsMask(unsigned(a) | unsigned(b)); } +inline FragmentShadingRateMask operator|(FragmentShadingRateMask a, FragmentShadingRateMask b) { return FragmentShadingRateMask(unsigned(a) | unsigned(b)); } } // end namespace spv diff --git a/thirdparty/glslang/SPIRV/spvIR.h b/thirdparty/glslang/SPIRV/spvIR.h index cf6a71159a..486e80d000 100755..100644 --- a/thirdparty/glslang/SPIRV/spvIR.h +++ b/thirdparty/glslang/SPIRV/spvIR.h @@ -55,6 +55,7 @@ #include <iostream> #include <memory> #include <vector> +#include <set> namespace spv { @@ -235,8 +236,7 @@ public: assert(instructions.size() > 0); instructions.resize(1); successors.clear(); - Instruction* unreachable = new Instruction(OpUnreachable); - addInstruction(std::unique_ptr<Instruction>(unreachable)); + addInstruction(std::unique_ptr<Instruction>(new Instruction(OpUnreachable))); } // Change this block into a canonical dead continue target branching to the // given header ID. Delete instructions as necessary. A canonical dead continue @@ -263,6 +263,7 @@ public: case OpBranchConditional: case OpSwitch: case OpKill: + case OpTerminateInvocation: case OpReturn: case OpReturnValue: case OpUnreachable: @@ -352,10 +353,28 @@ public: const std::vector<Block*>& getBlocks() const { return blocks; } void addLocalVariable(std::unique_ptr<Instruction> inst); Id getReturnType() const { return functionInstruction.getTypeId(); } + void setReturnPrecision(Decoration precision) + { + if (precision == DecorationRelaxedPrecision) + reducedPrecisionReturn = true; + } + Decoration getReturnPrecision() const + { return reducedPrecisionReturn ? DecorationRelaxedPrecision : NoPrecision; } void setImplicitThis() { implicitThis = true; } bool hasImplicitThis() const { return implicitThis; } + void addParamPrecision(unsigned param, Decoration precision) + { + if (precision == DecorationRelaxedPrecision) + reducedPrecisionParams.insert(param); + } + Decoration getParamPrecision(unsigned param) const + { + return reducedPrecisionParams.find(param) != reducedPrecisionParams.end() ? + DecorationRelaxedPrecision : NoPrecision; + } + void dump(std::vector<unsigned int>& out) const { // OpFunction @@ -380,6 +399,8 @@ protected: std::vector<Instruction*> parameterInstructions; std::vector<Block*> blocks; bool implicitThis; // true if this is a member function expecting to be passed a 'this' as the first argument + bool reducedPrecisionReturn; + std::set<int> reducedPrecisionParams; // list of parameter indexes that need a relaxed precision arg }; // @@ -440,7 +461,8 @@ protected: // - the OpFunction instruction // - all the OpFunctionParameter instructions __inline Function::Function(Id id, Id resultType, Id functionType, Id firstParamId, Module& parent) - : parent(parent), functionInstruction(id, resultType, OpFunction), implicitThis(false) + : parent(parent), functionInstruction(id, resultType, OpFunction), implicitThis(false), + reducedPrecisionReturn(false) { // OpFunction functionInstruction.addImmediateOperand(FunctionControlMaskNone); |