1 // 2 // Copyright (C) 2014-2016 LunarG, Inc. 3 // Copyright (C) 2015-2018 Google, Inc. 4 // Copyright (C) 2017 ARM Limited. 5 // 6 // All rights reserved. 7 // 8 // Redistribution and use in source and binary forms, with or without 9 // modification, are permitted provided that the following conditions 10 // are met: 11 // 12 // Redistributions of source code must retain the above copyright 13 // notice, this list of conditions and the following disclaimer. 14 // 15 // Redistributions in binary form must reproduce the above 16 // copyright notice, this list of conditions and the following 17 // disclaimer in the documentation and/or other materials provided 18 // with the distribution. 19 // 20 // Neither the name of 3Dlabs Inc. Ltd. nor the names of its 21 // contributors may be used to endorse or promote products derived 22 // from this software without specific prior written permission. 23 // 24 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 25 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 26 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 27 // FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 28 // COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 29 // INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 30 // BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 31 // LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER 32 // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 33 // LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN 34 // ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 35 // POSSIBILITY OF SUCH DAMAGE. 36 37 // 38 // Visit the nodes in the glslang intermediate tree representation to 39 // translate them to SPIR-V. 40 // 41 42 #include "spirv.hpp" 43 #include "GlslangToSpv.h" 44 #include "SpvBuilder.h" 45 namespace spv { 46 #include "GLSL.std.450.h" 47 #include "GLSL.ext.KHR.h" 48 #include "GLSL.ext.EXT.h" 49 #ifdef AMD_EXTENSIONS 50 #include "GLSL.ext.AMD.h" 51 #endif 52 #ifdef NV_EXTENSIONS 53 #include "GLSL.ext.NV.h" 54 #endif 55 } 56 57 // Glslang includes 58 #include "../glslang/MachineIndependent/localintermediate.h" 59 #include "../glslang/MachineIndependent/SymbolTable.h" 60 #include "../glslang/Include/Common.h" 61 #include "../glslang/Include/revision.h" 62 63 #include <fstream> 64 #include <iomanip> 65 #include <list> 66 #include <map> 67 #include <stack> 68 #include <string> 69 #include <vector> 70 71 namespace { 72 73 namespace { 74 class SpecConstantOpModeGuard { 75 public: 76 SpecConstantOpModeGuard(spv::Builder* builder) 77 : builder_(builder) { 78 previous_flag_ = builder->isInSpecConstCodeGenMode(); 79 } 80 ~SpecConstantOpModeGuard() { 81 previous_flag_ ? builder_->setToSpecConstCodeGenMode() 82 : builder_->setToNormalCodeGenMode(); 83 } 84 void turnOnSpecConstantOpMode() { 85 builder_->setToSpecConstCodeGenMode(); 86 } 87 88 private: 89 spv::Builder* builder_; 90 bool previous_flag_; 91 }; 92 93 struct OpDecorations { 94 spv::Decoration precision; 95 spv::Decoration noContraction; 96 spv::Decoration nonUniform; 97 }; 98 99 } // namespace 100 101 // 102 // The main holder of information for translating glslang to SPIR-V. 103 // 104 // Derives from the AST walking base class. 105 // 106 class TGlslangToSpvTraverser : public glslang::TIntermTraverser { 107 public: 108 TGlslangToSpvTraverser(unsigned int spvVersion, const glslang::TIntermediate*, spv::SpvBuildLogger* logger, 109 glslang::SpvOptions& options); 110 virtual ~TGlslangToSpvTraverser() { } 111 112 bool visitAggregate(glslang::TVisit, glslang::TIntermAggregate*); 113 bool visitBinary(glslang::TVisit, glslang::TIntermBinary*); 114 void visitConstantUnion(glslang::TIntermConstantUnion*); 115 bool visitSelection(glslang::TVisit, glslang::TIntermSelection*); 116 bool visitSwitch(glslang::TVisit, glslang::TIntermSwitch*); 117 void visitSymbol(glslang::TIntermSymbol* symbol); 118 bool visitUnary(glslang::TVisit, glslang::TIntermUnary*); 119 bool visitLoop(glslang::TVisit, glslang::TIntermLoop*); 120 bool visitBranch(glslang::TVisit visit, glslang::TIntermBranch*); 121 122 void finishSpv(); 123 void dumpSpv(std::vector<unsigned int>& out); 124 125 protected: 126 TGlslangToSpvTraverser(TGlslangToSpvTraverser&); 127 TGlslangToSpvTraverser& operator=(TGlslangToSpvTraverser&); 128 129 spv::Decoration TranslateInterpolationDecoration(const glslang::TQualifier& qualifier); 130 spv::Decoration TranslateAuxiliaryStorageDecoration(const glslang::TQualifier& qualifier); 131 spv::Decoration TranslateNonUniformDecoration(const glslang::TQualifier& qualifier); 132 spv::Builder::AccessChain::CoherentFlags TranslateCoherent(const glslang::TType& type); 133 spv::MemoryAccessMask TranslateMemoryAccess(const spv::Builder::AccessChain::CoherentFlags &coherentFlags); 134 spv::ImageOperandsMask TranslateImageOperands(const spv::Builder::AccessChain::CoherentFlags &coherentFlags); 135 spv::Scope TranslateMemoryScope(const spv::Builder::AccessChain::CoherentFlags &coherentFlags); 136 spv::BuiltIn TranslateBuiltInDecoration(glslang::TBuiltInVariable, bool memberDeclaration); 137 spv::ImageFormat TranslateImageFormat(const glslang::TType& type); 138 spv::SelectionControlMask TranslateSelectionControl(const glslang::TIntermSelection&) const; 139 spv::SelectionControlMask TranslateSwitchControl(const glslang::TIntermSwitch&) const; 140 spv::LoopControlMask TranslateLoopControl(const glslang::TIntermLoop&, unsigned int& dependencyLength) const; 141 spv::StorageClass TranslateStorageClass(const glslang::TType&); 142 void addIndirectionIndexCapabilities(const glslang::TType& baseType, const glslang::TType& indexType); 143 spv::Id createSpvVariable(const glslang::TIntermSymbol*); 144 spv::Id getSampledType(const glslang::TSampler&); 145 spv::Id getInvertedSwizzleType(const glslang::TIntermTyped&); 146 spv::Id createInvertedSwizzle(spv::Decoration precision, const glslang::TIntermTyped&, spv::Id parentResult); 147 void convertSwizzle(const glslang::TIntermAggregate&, std::vector<unsigned>& swizzle); 148 spv::Id convertGlslangToSpvType(const glslang::TType& type, bool forwardReferenceOnly = false); 149 spv::Id convertGlslangToSpvType(const glslang::TType& type, glslang::TLayoutPacking, const glslang::TQualifier&, 150 bool lastBufferBlockMember, bool forwardReferenceOnly = false); 151 bool filterMember(const glslang::TType& member); 152 spv::Id convertGlslangStructToSpvType(const glslang::TType&, const glslang::TTypeList* glslangStruct, 153 glslang::TLayoutPacking, const glslang::TQualifier&); 154 void decorateStructType(const glslang::TType&, const glslang::TTypeList* glslangStruct, glslang::TLayoutPacking, 155 const glslang::TQualifier&, spv::Id); 156 spv::Id makeArraySizeId(const glslang::TArraySizes&, int dim); 157 spv::Id accessChainLoad(const glslang::TType& type); 158 void accessChainStore(const glslang::TType& type, spv::Id rvalue); 159 void multiTypeStore(const glslang::TType&, spv::Id rValue); 160 glslang::TLayoutPacking getExplicitLayout(const glslang::TType& type) const; 161 int getArrayStride(const glslang::TType& arrayType, glslang::TLayoutPacking, glslang::TLayoutMatrix); 162 int getMatrixStride(const glslang::TType& matrixType, glslang::TLayoutPacking, glslang::TLayoutMatrix); 163 void updateMemberOffset(const glslang::TType& structType, const glslang::TType& memberType, int& currentOffset, 164 int& nextOffset, glslang::TLayoutPacking, glslang::TLayoutMatrix); 165 void declareUseOfStructMember(const glslang::TTypeList& members, int glslangMember); 166 167 bool isShaderEntryPoint(const glslang::TIntermAggregate* node); 168 bool writableParam(glslang::TStorageQualifier) const; 169 bool originalParam(glslang::TStorageQualifier, const glslang::TType&, bool implicitThisParam); 170 void makeFunctions(const glslang::TIntermSequence&); 171 void makeGlobalInitializers(const glslang::TIntermSequence&); 172 void visitFunctions(const glslang::TIntermSequence&); 173 void handleFunctionEntry(const glslang::TIntermAggregate* node); 174 void translateArguments(const glslang::TIntermAggregate& node, std::vector<spv::Id>& arguments); 175 void translateArguments(glslang::TIntermUnary& node, std::vector<spv::Id>& arguments); 176 spv::Id createImageTextureFunctionCall(glslang::TIntermOperator* node); 177 spv::Id handleUserFunctionCall(const glslang::TIntermAggregate*); 178 179 spv::Id createBinaryOperation(glslang::TOperator op, OpDecorations&, spv::Id typeId, spv::Id left, spv::Id right, 180 glslang::TBasicType typeProxy, bool reduceComparison = true); 181 spv::Id createBinaryMatrixOperation(spv::Op, OpDecorations&, spv::Id typeId, spv::Id left, spv::Id right); 182 spv::Id createUnaryOperation(glslang::TOperator op, OpDecorations&, spv::Id typeId, spv::Id operand, 183 glslang::TBasicType typeProxy); 184 spv::Id createUnaryMatrixOperation(spv::Op op, OpDecorations&, spv::Id typeId, spv::Id operand, 185 glslang::TBasicType typeProxy); 186 spv::Id createConversion(glslang::TOperator op, OpDecorations&, spv::Id destTypeId, spv::Id operand, 187 glslang::TBasicType typeProxy); 188 spv::Id createIntWidthConversion(glslang::TOperator op, spv::Id operand, int vectorSize); 189 spv::Id makeSmearedConstant(spv::Id constant, int vectorSize); 190 spv::Id createAtomicOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy); 191 spv::Id createInvocationsOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy); 192 spv::Id CreateInvocationsVectorOperation(spv::Op op, spv::GroupOperation groupOperation, spv::Id typeId, std::vector<spv::Id>& operands); 193 spv::Id createSubgroupOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy); 194 spv::Id createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy); 195 spv::Id createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId); 196 spv::Id getSymbolId(const glslang::TIntermSymbol* node); 197 #ifdef NV_EXTENSIONS 198 void addMeshNVDecoration(spv::Id id, int member, const glslang::TQualifier & qualifier); 199 #endif 200 spv::Id createSpvConstant(const glslang::TIntermTyped&); 201 spv::Id createSpvConstantFromConstUnionArray(const glslang::TType& type, const glslang::TConstUnionArray&, int& nextConst, bool specConstant); 202 bool isTrivialLeaf(const glslang::TIntermTyped* node); 203 bool isTrivial(const glslang::TIntermTyped* node); 204 spv::Id createShortCircuit(glslang::TOperator, glslang::TIntermTyped& left, glslang::TIntermTyped& right); 205 #ifdef AMD_EXTENSIONS 206 spv::Id getExtBuiltins(const char* name); 207 #endif 208 void addPre13Extension(const char* ext) 209 { 210 if (builder.getSpvVersion() < glslang::EShTargetSpv_1_3) 211 builder.addExtension(ext); 212 } 213 214 unsigned int getBufferReferenceAlignment(const glslang::TType &type) const { 215 if (type.getBasicType() == glslang::EbtReference) { 216 return type.getReferentType()->getQualifier().hasBufferReferenceAlign() ? 217 (1u << type.getReferentType()->getQualifier().layoutBufferReferenceAlign) : 16u; 218 } else { 219 return 0; 220 } 221 } 222 223 glslang::SpvOptions& options; 224 spv::Function* shaderEntry; 225 spv::Function* currentFunction; 226 spv::Instruction* entryPoint; 227 int sequenceDepth; 228 229 spv::SpvBuildLogger* logger; 230 231 // There is a 1:1 mapping between a spv builder and a module; this is thread safe 232 spv::Builder builder; 233 bool inEntryPoint; 234 bool entryPointTerminated; 235 bool linkageOnly; // true when visiting the set of objects in the AST present only for establishing interface, whether or not they were statically used 236 std::set<spv::Id> iOSet; // all input/output variables from either static use or declaration of interface 237 const glslang::TIntermediate* glslangIntermediate; 238 spv::Id stdBuiltins; 239 std::unordered_map<const char*, spv::Id> extBuiltinMap; 240 241 std::unordered_map<int, spv::Id> symbolValues; 242 std::unordered_set<int> rValueParameters; // set of formal function parameters passed as rValues, rather than a pointer 243 std::unordered_map<std::string, spv::Function*> functionMap; 244 std::unordered_map<const glslang::TTypeList*, spv::Id> structMap[glslang::ElpCount][glslang::ElmCount]; 245 // for mapping glslang block indices to spv indices (e.g., due to hidden members): 246 std::unordered_map<const glslang::TTypeList*, std::vector<int> > memberRemapper; 247 std::stack<bool> breakForLoop; // false means break for switch 248 std::unordered_map<std::string, const glslang::TIntermSymbol*> counterOriginator; 249 // Map pointee types for EbtReference to their forward pointers 250 std::map<const glslang::TType *, spv::Id> forwardPointers; 251 }; 252 253 // 254 // Helper functions for translating glslang representations to SPIR-V enumerants. 255 // 256 257 // Translate glslang profile to SPIR-V source language. 258 spv::SourceLanguage TranslateSourceLanguage(glslang::EShSource source, EProfile profile) 259 { 260 switch (source) { 261 case glslang::EShSourceGlsl: 262 switch (profile) { 263 case ENoProfile: 264 case ECoreProfile: 265 case ECompatibilityProfile: 266 return spv::SourceLanguageGLSL; 267 case EEsProfile: 268 return spv::SourceLanguageESSL; 269 default: 270 return spv::SourceLanguageUnknown; 271 } 272 case glslang::EShSourceHlsl: 273 return spv::SourceLanguageHLSL; 274 default: 275 return spv::SourceLanguageUnknown; 276 } 277 } 278 279 // Translate glslang language (stage) to SPIR-V execution model. 280 spv::ExecutionModel TranslateExecutionModel(EShLanguage stage) 281 { 282 switch (stage) { 283 case EShLangVertex: return spv::ExecutionModelVertex; 284 case EShLangTessControl: return spv::ExecutionModelTessellationControl; 285 case EShLangTessEvaluation: return spv::ExecutionModelTessellationEvaluation; 286 case EShLangGeometry: return spv::ExecutionModelGeometry; 287 case EShLangFragment: return spv::ExecutionModelFragment; 288 case EShLangCompute: return spv::ExecutionModelGLCompute; 289 #ifdef NV_EXTENSIONS 290 case EShLangRayGenNV: return spv::ExecutionModelRayGenerationNV; 291 case EShLangIntersectNV: return spv::ExecutionModelIntersectionNV; 292 case EShLangAnyHitNV: return spv::ExecutionModelAnyHitNV; 293 case EShLangClosestHitNV: return spv::ExecutionModelClosestHitNV; 294 case EShLangMissNV: return spv::ExecutionModelMissNV; 295 case EShLangCallableNV: return spv::ExecutionModelCallableNV; 296 case EShLangTaskNV: return spv::ExecutionModelTaskNV; 297 case EShLangMeshNV: return spv::ExecutionModelMeshNV; 298 #endif 299 default: 300 assert(0); 301 return spv::ExecutionModelFragment; 302 } 303 } 304 305 // Translate glslang sampler type to SPIR-V dimensionality. 306 spv::Dim TranslateDimensionality(const glslang::TSampler& sampler) 307 { 308 switch (sampler.dim) { 309 case glslang::Esd1D: return spv::Dim1D; 310 case glslang::Esd2D: return spv::Dim2D; 311 case glslang::Esd3D: return spv::Dim3D; 312 case glslang::EsdCube: return spv::DimCube; 313 case glslang::EsdRect: return spv::DimRect; 314 case glslang::EsdBuffer: return spv::DimBuffer; 315 case glslang::EsdSubpass: return spv::DimSubpassData; 316 default: 317 assert(0); 318 return spv::Dim2D; 319 } 320 } 321 322 // Translate glslang precision to SPIR-V precision decorations. 323 spv::Decoration TranslatePrecisionDecoration(glslang::TPrecisionQualifier glslangPrecision) 324 { 325 switch (glslangPrecision) { 326 case glslang::EpqLow: return spv::DecorationRelaxedPrecision; 327 case glslang::EpqMedium: return spv::DecorationRelaxedPrecision; 328 default: 329 return spv::NoPrecision; 330 } 331 } 332 333 // Translate glslang type to SPIR-V precision decorations. 334 spv::Decoration TranslatePrecisionDecoration(const glslang::TType& type) 335 { 336 return TranslatePrecisionDecoration(type.getQualifier().precision); 337 } 338 339 // Translate glslang type to SPIR-V block decorations. 340 spv::Decoration TranslateBlockDecoration(const glslang::TType& type, bool useStorageBuffer) 341 { 342 if (type.getBasicType() == glslang::EbtBlock) { 343 switch (type.getQualifier().storage) { 344 case glslang::EvqUniform: return spv::DecorationBlock; 345 case glslang::EvqBuffer: return useStorageBuffer ? spv::DecorationBlock : spv::DecorationBufferBlock; 346 case glslang::EvqVaryingIn: return spv::DecorationBlock; 347 case glslang::EvqVaryingOut: return spv::DecorationBlock; 348 #ifdef NV_EXTENSIONS 349 case glslang::EvqPayloadNV: return spv::DecorationBlock; 350 case glslang::EvqPayloadInNV: return spv::DecorationBlock; 351 case glslang::EvqHitAttrNV: return spv::DecorationBlock; 352 case glslang::EvqCallableDataNV: return spv::DecorationBlock; 353 case glslang::EvqCallableDataInNV: return spv::DecorationBlock; 354 #endif 355 default: 356 assert(0); 357 break; 358 } 359 } 360 361 return spv::DecorationMax; 362 } 363 364 // Translate glslang type to SPIR-V memory decorations. 365 void TranslateMemoryDecoration(const glslang::TQualifier& qualifier, std::vector<spv::Decoration>& memory, bool useVulkanMemoryModel) 366 { 367 if (!useVulkanMemoryModel) { 368 if (qualifier.coherent) 369 memory.push_back(spv::DecorationCoherent); 370 if (qualifier.volatil) { 371 memory.push_back(spv::DecorationVolatile); 372 memory.push_back(spv::DecorationCoherent); 373 } 374 } 375 if (qualifier.restrict) 376 memory.push_back(spv::DecorationRestrict); 377 if (qualifier.readonly) 378 memory.push_back(spv::DecorationNonWritable); 379 if (qualifier.writeonly) 380 memory.push_back(spv::DecorationNonReadable); 381 } 382 383 // Translate glslang type to SPIR-V layout decorations. 384 spv::Decoration TranslateLayoutDecoration(const glslang::TType& type, glslang::TLayoutMatrix matrixLayout) 385 { 386 if (type.isMatrix()) { 387 switch (matrixLayout) { 388 case glslang::ElmRowMajor: 389 return spv::DecorationRowMajor; 390 case glslang::ElmColumnMajor: 391 return spv::DecorationColMajor; 392 default: 393 // opaque layouts don't need a majorness 394 return spv::DecorationMax; 395 } 396 } else { 397 switch (type.getBasicType()) { 398 default: 399 return spv::DecorationMax; 400 break; 401 case glslang::EbtBlock: 402 switch (type.getQualifier().storage) { 403 case glslang::EvqUniform: 404 case glslang::EvqBuffer: 405 switch (type.getQualifier().layoutPacking) { 406 case glslang::ElpShared: return spv::DecorationGLSLShared; 407 case glslang::ElpPacked: return spv::DecorationGLSLPacked; 408 default: 409 return spv::DecorationMax; 410 } 411 case glslang::EvqVaryingIn: 412 case glslang::EvqVaryingOut: 413 if (type.getQualifier().isTaskMemory()) { 414 switch (type.getQualifier().layoutPacking) { 415 case glslang::ElpShared: return spv::DecorationGLSLShared; 416 case glslang::ElpPacked: return spv::DecorationGLSLPacked; 417 default: break; 418 } 419 } else { 420 assert(type.getQualifier().layoutPacking == glslang::ElpNone); 421 } 422 return spv::DecorationMax; 423 #ifdef NV_EXTENSIONS 424 case glslang::EvqPayloadNV: 425 case glslang::EvqPayloadInNV: 426 case glslang::EvqHitAttrNV: 427 case glslang::EvqCallableDataNV: 428 case glslang::EvqCallableDataInNV: 429 return spv::DecorationMax; 430 #endif 431 default: 432 assert(0); 433 return spv::DecorationMax; 434 } 435 } 436 } 437 } 438 439 // Translate glslang type to SPIR-V interpolation decorations. 440 // Returns spv::DecorationMax when no decoration 441 // should be applied. 442 spv::Decoration TGlslangToSpvTraverser::TranslateInterpolationDecoration(const glslang::TQualifier& qualifier) 443 { 444 if (qualifier.smooth) 445 // Smooth decoration doesn't exist in SPIR-V 1.0 446 return spv::DecorationMax; 447 else if (qualifier.nopersp) 448 return spv::DecorationNoPerspective; 449 else if (qualifier.flat) 450 return spv::DecorationFlat; 451 #ifdef AMD_EXTENSIONS 452 else if (qualifier.explicitInterp) { 453 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); 454 return spv::DecorationExplicitInterpAMD; 455 } 456 #endif 457 else 458 return spv::DecorationMax; 459 } 460 461 // Translate glslang type to SPIR-V auxiliary storage decorations. 462 // Returns spv::DecorationMax when no decoration 463 // should be applied. 464 spv::Decoration TGlslangToSpvTraverser::TranslateAuxiliaryStorageDecoration(const glslang::TQualifier& qualifier) 465 { 466 if (qualifier.patch) 467 return spv::DecorationPatch; 468 else if (qualifier.centroid) 469 return spv::DecorationCentroid; 470 else if (qualifier.sample) { 471 builder.addCapability(spv::CapabilitySampleRateShading); 472 return spv::DecorationSample; 473 } else 474 return spv::DecorationMax; 475 } 476 477 // If glslang type is invariant, return SPIR-V invariant decoration. 478 spv::Decoration TranslateInvariantDecoration(const glslang::TQualifier& qualifier) 479 { 480 if (qualifier.invariant) 481 return spv::DecorationInvariant; 482 else 483 return spv::DecorationMax; 484 } 485 486 // If glslang type is noContraction, return SPIR-V NoContraction decoration. 487 spv::Decoration TranslateNoContractionDecoration(const glslang::TQualifier& qualifier) 488 { 489 if (qualifier.noContraction) 490 return spv::DecorationNoContraction; 491 else 492 return spv::DecorationMax; 493 } 494 495 // If glslang type is nonUniform, return SPIR-V NonUniform decoration. 496 spv::Decoration TGlslangToSpvTraverser::TranslateNonUniformDecoration(const glslang::TQualifier& qualifier) 497 { 498 if (qualifier.isNonUniform()) { 499 builder.addExtension("SPV_EXT_descriptor_indexing"); 500 builder.addCapability(spv::CapabilityShaderNonUniformEXT); 501 return spv::DecorationNonUniformEXT; 502 } else 503 return spv::DecorationMax; 504 } 505 506 spv::MemoryAccessMask TGlslangToSpvTraverser::TranslateMemoryAccess(const spv::Builder::AccessChain::CoherentFlags &coherentFlags) 507 { 508 if (!glslangIntermediate->usingVulkanMemoryModel() || coherentFlags.isImage) { 509 return spv::MemoryAccessMaskNone; 510 } 511 spv::MemoryAccessMask mask = spv::MemoryAccessMaskNone; 512 if (coherentFlags.volatil || 513 coherentFlags.coherent || 514 coherentFlags.devicecoherent || 515 coherentFlags.queuefamilycoherent || 516 coherentFlags.workgroupcoherent || 517 coherentFlags.subgroupcoherent) { 518 mask = mask | spv::MemoryAccessMakePointerAvailableKHRMask | 519 spv::MemoryAccessMakePointerVisibleKHRMask; 520 } 521 if (coherentFlags.nonprivate) { 522 mask = mask | spv::MemoryAccessNonPrivatePointerKHRMask; 523 } 524 if (coherentFlags.volatil) { 525 mask = mask | spv::MemoryAccessVolatileMask; 526 } 527 if (mask != spv::MemoryAccessMaskNone) { 528 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); 529 } 530 return mask; 531 } 532 533 spv::ImageOperandsMask TGlslangToSpvTraverser::TranslateImageOperands(const spv::Builder::AccessChain::CoherentFlags &coherentFlags) 534 { 535 if (!glslangIntermediate->usingVulkanMemoryModel()) { 536 return spv::ImageOperandsMaskNone; 537 } 538 spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone; 539 if (coherentFlags.volatil || 540 coherentFlags.coherent || 541 coherentFlags.devicecoherent || 542 coherentFlags.queuefamilycoherent || 543 coherentFlags.workgroupcoherent || 544 coherentFlags.subgroupcoherent) { 545 mask = mask | spv::ImageOperandsMakeTexelAvailableKHRMask | 546 spv::ImageOperandsMakeTexelVisibleKHRMask; 547 } 548 if (coherentFlags.nonprivate) { 549 mask = mask | spv::ImageOperandsNonPrivateTexelKHRMask; 550 } 551 if (coherentFlags.volatil) { 552 mask = mask | spv::ImageOperandsVolatileTexelKHRMask; 553 } 554 if (mask != spv::ImageOperandsMaskNone) { 555 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); 556 } 557 return mask; 558 } 559 560 spv::Builder::AccessChain::CoherentFlags TGlslangToSpvTraverser::TranslateCoherent(const glslang::TType& type) 561 { 562 spv::Builder::AccessChain::CoherentFlags flags; 563 flags.coherent = type.getQualifier().coherent; 564 flags.devicecoherent = type.getQualifier().devicecoherent; 565 flags.queuefamilycoherent = type.getQualifier().queuefamilycoherent; 566 // shared variables are implicitly workgroupcoherent in GLSL. 567 flags.workgroupcoherent = type.getQualifier().workgroupcoherent || 568 type.getQualifier().storage == glslang::EvqShared; 569 flags.subgroupcoherent = type.getQualifier().subgroupcoherent; 570 // *coherent variables are implicitly nonprivate in GLSL 571 flags.nonprivate = type.getQualifier().nonprivate || 572 flags.subgroupcoherent || 573 flags.workgroupcoherent || 574 flags.queuefamilycoherent || 575 flags.devicecoherent || 576 flags.coherent; 577 flags.volatil = type.getQualifier().volatil; 578 flags.isImage = type.getBasicType() == glslang::EbtSampler; 579 return flags; 580 } 581 582 spv::Scope TGlslangToSpvTraverser::TranslateMemoryScope(const spv::Builder::AccessChain::CoherentFlags &coherentFlags) 583 { 584 spv::Scope scope; 585 if (coherentFlags.coherent) { 586 // coherent defaults to Device scope in the old model, QueueFamilyKHR scope in the new model 587 scope = glslangIntermediate->usingVulkanMemoryModel() ? spv::ScopeQueueFamilyKHR : spv::ScopeDevice; 588 } else if (coherentFlags.devicecoherent) { 589 scope = spv::ScopeDevice; 590 } else if (coherentFlags.queuefamilycoherent) { 591 scope = spv::ScopeQueueFamilyKHR; 592 } else if (coherentFlags.workgroupcoherent) { 593 scope = spv::ScopeWorkgroup; 594 } else if (coherentFlags.subgroupcoherent) { 595 scope = spv::ScopeSubgroup; 596 } else { 597 scope = spv::ScopeMax; 598 } 599 if (glslangIntermediate->usingVulkanMemoryModel() && scope == spv::ScopeDevice) { 600 builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR); 601 } 602 return scope; 603 } 604 605 // Translate a glslang built-in variable to a SPIR-V built in decoration. Also generate 606 // associated capabilities when required. For some built-in variables, a capability 607 // is generated only when using the variable in an executable instruction, but not when 608 // just declaring a struct member variable with it. This is true for PointSize, 609 // ClipDistance, and CullDistance. 610 spv::BuiltIn TGlslangToSpvTraverser::TranslateBuiltInDecoration(glslang::TBuiltInVariable builtIn, bool memberDeclaration) 611 { 612 switch (builtIn) { 613 case glslang::EbvPointSize: 614 // Defer adding the capability until the built-in is actually used. 615 if (! memberDeclaration) { 616 switch (glslangIntermediate->getStage()) { 617 case EShLangGeometry: 618 builder.addCapability(spv::CapabilityGeometryPointSize); 619 break; 620 case EShLangTessControl: 621 case EShLangTessEvaluation: 622 builder.addCapability(spv::CapabilityTessellationPointSize); 623 break; 624 default: 625 break; 626 } 627 } 628 return spv::BuiltInPointSize; 629 630 // These *Distance capabilities logically belong here, but if the member is declared and 631 // then never used, consumers of SPIR-V prefer the capability not be declared. 632 // They are now generated when used, rather than here when declared. 633 // Potentially, the specification should be more clear what the minimum 634 // use needed is to trigger the capability. 635 // 636 case glslang::EbvClipDistance: 637 if (!memberDeclaration) 638 builder.addCapability(spv::CapabilityClipDistance); 639 return spv::BuiltInClipDistance; 640 641 case glslang::EbvCullDistance: 642 if (!memberDeclaration) 643 builder.addCapability(spv::CapabilityCullDistance); 644 return spv::BuiltInCullDistance; 645 646 case glslang::EbvViewportIndex: 647 builder.addCapability(spv::CapabilityMultiViewport); 648 if (glslangIntermediate->getStage() == EShLangVertex || 649 glslangIntermediate->getStage() == EShLangTessControl || 650 glslangIntermediate->getStage() == EShLangTessEvaluation) { 651 652 builder.addExtension(spv::E_SPV_EXT_shader_viewport_index_layer); 653 builder.addCapability(spv::CapabilityShaderViewportIndexLayerEXT); 654 } 655 return spv::BuiltInViewportIndex; 656 657 case glslang::EbvSampleId: 658 builder.addCapability(spv::CapabilitySampleRateShading); 659 return spv::BuiltInSampleId; 660 661 case glslang::EbvSamplePosition: 662 builder.addCapability(spv::CapabilitySampleRateShading); 663 return spv::BuiltInSamplePosition; 664 665 case glslang::EbvSampleMask: 666 return spv::BuiltInSampleMask; 667 668 case glslang::EbvLayer: 669 #ifdef NV_EXTENSIONS 670 if (glslangIntermediate->getStage() == EShLangMeshNV) { 671 return spv::BuiltInLayer; 672 } 673 #endif 674 builder.addCapability(spv::CapabilityGeometry); 675 if (glslangIntermediate->getStage() == EShLangVertex || 676 glslangIntermediate->getStage() == EShLangTessControl || 677 glslangIntermediate->getStage() == EShLangTessEvaluation) { 678 679 builder.addExtension(spv::E_SPV_EXT_shader_viewport_index_layer); 680 builder.addCapability(spv::CapabilityShaderViewportIndexLayerEXT); 681 } 682 return spv::BuiltInLayer; 683 684 case glslang::EbvPosition: return spv::BuiltInPosition; 685 case glslang::EbvVertexId: return spv::BuiltInVertexId; 686 case glslang::EbvInstanceId: return spv::BuiltInInstanceId; 687 case glslang::EbvVertexIndex: return spv::BuiltInVertexIndex; 688 case glslang::EbvInstanceIndex: return spv::BuiltInInstanceIndex; 689 690 case glslang::EbvBaseVertex: 691 addPre13Extension(spv::E_SPV_KHR_shader_draw_parameters); 692 builder.addCapability(spv::CapabilityDrawParameters); 693 return spv::BuiltInBaseVertex; 694 695 case glslang::EbvBaseInstance: 696 addPre13Extension(spv::E_SPV_KHR_shader_draw_parameters); 697 builder.addCapability(spv::CapabilityDrawParameters); 698 return spv::BuiltInBaseInstance; 699 700 case glslang::EbvDrawId: 701 addPre13Extension(spv::E_SPV_KHR_shader_draw_parameters); 702 builder.addCapability(spv::CapabilityDrawParameters); 703 return spv::BuiltInDrawIndex; 704 705 case glslang::EbvPrimitiveId: 706 if (glslangIntermediate->getStage() == EShLangFragment) 707 builder.addCapability(spv::CapabilityGeometry); 708 return spv::BuiltInPrimitiveId; 709 710 case glslang::EbvFragStencilRef: 711 builder.addExtension(spv::E_SPV_EXT_shader_stencil_export); 712 builder.addCapability(spv::CapabilityStencilExportEXT); 713 return spv::BuiltInFragStencilRefEXT; 714 715 case glslang::EbvInvocationId: return spv::BuiltInInvocationId; 716 case glslang::EbvTessLevelInner: return spv::BuiltInTessLevelInner; 717 case glslang::EbvTessLevelOuter: return spv::BuiltInTessLevelOuter; 718 case glslang::EbvTessCoord: return spv::BuiltInTessCoord; 719 case glslang::EbvPatchVertices: return spv::BuiltInPatchVertices; 720 case glslang::EbvFragCoord: return spv::BuiltInFragCoord; 721 case glslang::EbvPointCoord: return spv::BuiltInPointCoord; 722 case glslang::EbvFace: return spv::BuiltInFrontFacing; 723 case glslang::EbvFragDepth: return spv::BuiltInFragDepth; 724 case glslang::EbvHelperInvocation: return spv::BuiltInHelperInvocation; 725 case glslang::EbvNumWorkGroups: return spv::BuiltInNumWorkgroups; 726 case glslang::EbvWorkGroupSize: return spv::BuiltInWorkgroupSize; 727 case glslang::EbvWorkGroupId: return spv::BuiltInWorkgroupId; 728 case glslang::EbvLocalInvocationId: return spv::BuiltInLocalInvocationId; 729 case glslang::EbvLocalInvocationIndex: return spv::BuiltInLocalInvocationIndex; 730 case glslang::EbvGlobalInvocationId: return spv::BuiltInGlobalInvocationId; 731 732 case glslang::EbvSubGroupSize: 733 builder.addExtension(spv::E_SPV_KHR_shader_ballot); 734 builder.addCapability(spv::CapabilitySubgroupBallotKHR); 735 return spv::BuiltInSubgroupSize; 736 737 case glslang::EbvSubGroupInvocation: 738 builder.addExtension(spv::E_SPV_KHR_shader_ballot); 739 builder.addCapability(spv::CapabilitySubgroupBallotKHR); 740 return spv::BuiltInSubgroupLocalInvocationId; 741 742 case glslang::EbvSubGroupEqMask: 743 builder.addExtension(spv::E_SPV_KHR_shader_ballot); 744 builder.addCapability(spv::CapabilitySubgroupBallotKHR); 745 return spv::BuiltInSubgroupEqMaskKHR; 746 747 case glslang::EbvSubGroupGeMask: 748 builder.addExtension(spv::E_SPV_KHR_shader_ballot); 749 builder.addCapability(spv::CapabilitySubgroupBallotKHR); 750 return spv::BuiltInSubgroupGeMaskKHR; 751 752 case glslang::EbvSubGroupGtMask: 753 builder.addExtension(spv::E_SPV_KHR_shader_ballot); 754 builder.addCapability(spv::CapabilitySubgroupBallotKHR); 755 return spv::BuiltInSubgroupGtMaskKHR; 756 757 case glslang::EbvSubGroupLeMask: 758 builder.addExtension(spv::E_SPV_KHR_shader_ballot); 759 builder.addCapability(spv::CapabilitySubgroupBallotKHR); 760 return spv::BuiltInSubgroupLeMaskKHR; 761 762 case glslang::EbvSubGroupLtMask: 763 builder.addExtension(spv::E_SPV_KHR_shader_ballot); 764 builder.addCapability(spv::CapabilitySubgroupBallotKHR); 765 return spv::BuiltInSubgroupLtMaskKHR; 766 767 case glslang::EbvNumSubgroups: 768 builder.addCapability(spv::CapabilityGroupNonUniform); 769 return spv::BuiltInNumSubgroups; 770 771 case glslang::EbvSubgroupID: 772 builder.addCapability(spv::CapabilityGroupNonUniform); 773 return spv::BuiltInSubgroupId; 774 775 case glslang::EbvSubgroupSize2: 776 builder.addCapability(spv::CapabilityGroupNonUniform); 777 return spv::BuiltInSubgroupSize; 778 779 case glslang::EbvSubgroupInvocation2: 780 builder.addCapability(spv::CapabilityGroupNonUniform); 781 return spv::BuiltInSubgroupLocalInvocationId; 782 783 case glslang::EbvSubgroupEqMask2: 784 builder.addCapability(spv::CapabilityGroupNonUniform); 785 builder.addCapability(spv::CapabilityGroupNonUniformBallot); 786 return spv::BuiltInSubgroupEqMask; 787 788 case glslang::EbvSubgroupGeMask2: 789 builder.addCapability(spv::CapabilityGroupNonUniform); 790 builder.addCapability(spv::CapabilityGroupNonUniformBallot); 791 return spv::BuiltInSubgroupGeMask; 792 793 case glslang::EbvSubgroupGtMask2: 794 builder.addCapability(spv::CapabilityGroupNonUniform); 795 builder.addCapability(spv::CapabilityGroupNonUniformBallot); 796 return spv::BuiltInSubgroupGtMask; 797 798 case glslang::EbvSubgroupLeMask2: 799 builder.addCapability(spv::CapabilityGroupNonUniform); 800 builder.addCapability(spv::CapabilityGroupNonUniformBallot); 801 return spv::BuiltInSubgroupLeMask; 802 803 case glslang::EbvSubgroupLtMask2: 804 builder.addCapability(spv::CapabilityGroupNonUniform); 805 builder.addCapability(spv::CapabilityGroupNonUniformBallot); 806 return spv::BuiltInSubgroupLtMask; 807 #ifdef AMD_EXTENSIONS 808 case glslang::EbvBaryCoordNoPersp: 809 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); 810 return spv::BuiltInBaryCoordNoPerspAMD; 811 812 case glslang::EbvBaryCoordNoPerspCentroid: 813 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); 814 return spv::BuiltInBaryCoordNoPerspCentroidAMD; 815 816 case glslang::EbvBaryCoordNoPerspSample: 817 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); 818 return spv::BuiltInBaryCoordNoPerspSampleAMD; 819 820 case glslang::EbvBaryCoordSmooth: 821 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); 822 return spv::BuiltInBaryCoordSmoothAMD; 823 824 case glslang::EbvBaryCoordSmoothCentroid: 825 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); 826 return spv::BuiltInBaryCoordSmoothCentroidAMD; 827 828 case glslang::EbvBaryCoordSmoothSample: 829 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); 830 return spv::BuiltInBaryCoordSmoothSampleAMD; 831 832 case glslang::EbvBaryCoordPullModel: 833 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter); 834 return spv::BuiltInBaryCoordPullModelAMD; 835 #endif 836 837 case glslang::EbvDeviceIndex: 838 addPre13Extension(spv::E_SPV_KHR_device_group); 839 builder.addCapability(spv::CapabilityDeviceGroup); 840 return spv::BuiltInDeviceIndex; 841 842 case glslang::EbvViewIndex: 843 addPre13Extension(spv::E_SPV_KHR_multiview); 844 builder.addCapability(spv::CapabilityMultiView); 845 return spv::BuiltInViewIndex; 846 847 case glslang::EbvFragSizeEXT: 848 builder.addExtension(spv::E_SPV_EXT_fragment_invocation_density); 849 builder.addCapability(spv::CapabilityFragmentDensityEXT); 850 return spv::BuiltInFragSizeEXT; 851 852 case glslang::EbvFragInvocationCountEXT: 853 builder.addExtension(spv::E_SPV_EXT_fragment_invocation_density); 854 builder.addCapability(spv::CapabilityFragmentDensityEXT); 855 return spv::BuiltInFragInvocationCountEXT; 856 857 #ifdef NV_EXTENSIONS 858 case glslang::EbvViewportMaskNV: 859 if (!memberDeclaration) { 860 builder.addExtension(spv::E_SPV_NV_viewport_array2); 861 builder.addCapability(spv::CapabilityShaderViewportMaskNV); 862 } 863 return spv::BuiltInViewportMaskNV; 864 case glslang::EbvSecondaryPositionNV: 865 if (!memberDeclaration) { 866 builder.addExtension(spv::E_SPV_NV_stereo_view_rendering); 867 builder.addCapability(spv::CapabilityShaderStereoViewNV); 868 } 869 return spv::BuiltInSecondaryPositionNV; 870 case glslang::EbvSecondaryViewportMaskNV: 871 if (!memberDeclaration) { 872 builder.addExtension(spv::E_SPV_NV_stereo_view_rendering); 873 builder.addCapability(spv::CapabilityShaderStereoViewNV); 874 } 875 return spv::BuiltInSecondaryViewportMaskNV; 876 case glslang::EbvPositionPerViewNV: 877 if (!memberDeclaration) { 878 builder.addExtension(spv::E_SPV_NVX_multiview_per_view_attributes); 879 builder.addCapability(spv::CapabilityPerViewAttributesNV); 880 } 881 return spv::BuiltInPositionPerViewNV; 882 case glslang::EbvViewportMaskPerViewNV: 883 if (!memberDeclaration) { 884 builder.addExtension(spv::E_SPV_NVX_multiview_per_view_attributes); 885 builder.addCapability(spv::CapabilityPerViewAttributesNV); 886 } 887 return spv::BuiltInViewportMaskPerViewNV; 888 case glslang::EbvFragFullyCoveredNV: 889 builder.addExtension(spv::E_SPV_EXT_fragment_fully_covered); 890 builder.addCapability(spv::CapabilityFragmentFullyCoveredEXT); 891 return spv::BuiltInFullyCoveredEXT; 892 case glslang::EbvFragmentSizeNV: 893 builder.addExtension(spv::E_SPV_NV_shading_rate); 894 builder.addCapability(spv::CapabilityShadingRateNV); 895 return spv::BuiltInFragmentSizeNV; 896 case glslang::EbvInvocationsPerPixelNV: 897 builder.addExtension(spv::E_SPV_NV_shading_rate); 898 builder.addCapability(spv::CapabilityShadingRateNV); 899 return spv::BuiltInInvocationsPerPixelNV; 900 901 // raytracing 902 case glslang::EbvLaunchIdNV: 903 return spv::BuiltInLaunchIdNV; 904 case glslang::EbvLaunchSizeNV: 905 return spv::BuiltInLaunchSizeNV; 906 case glslang::EbvWorldRayOriginNV: 907 return spv::BuiltInWorldRayOriginNV; 908 case glslang::EbvWorldRayDirectionNV: 909 return spv::BuiltInWorldRayDirectionNV; 910 case glslang::EbvObjectRayOriginNV: 911 return spv::BuiltInObjectRayOriginNV; 912 case glslang::EbvObjectRayDirectionNV: 913 return spv::BuiltInObjectRayDirectionNV; 914 case glslang::EbvRayTminNV: 915 return spv::BuiltInRayTminNV; 916 case glslang::EbvRayTmaxNV: 917 return spv::BuiltInRayTmaxNV; 918 case glslang::EbvInstanceCustomIndexNV: 919 return spv::BuiltInInstanceCustomIndexNV; 920 case glslang::EbvHitTNV: 921 return spv::BuiltInHitTNV; 922 case glslang::EbvHitKindNV: 923 return spv::BuiltInHitKindNV; 924 case glslang::EbvObjectToWorldNV: 925 return spv::BuiltInObjectToWorldNV; 926 case glslang::EbvWorldToObjectNV: 927 return spv::BuiltInWorldToObjectNV; 928 case glslang::EbvIncomingRayFlagsNV: 929 return spv::BuiltInIncomingRayFlagsNV; 930 case glslang::EbvBaryCoordNV: 931 builder.addExtension(spv::E_SPV_NV_fragment_shader_barycentric); 932 builder.addCapability(spv::CapabilityFragmentBarycentricNV); 933 return spv::BuiltInBaryCoordNV; 934 case glslang::EbvBaryCoordNoPerspNV: 935 builder.addExtension(spv::E_SPV_NV_fragment_shader_barycentric); 936 builder.addCapability(spv::CapabilityFragmentBarycentricNV); 937 return spv::BuiltInBaryCoordNoPerspNV; 938 case glslang::EbvTaskCountNV: 939 return spv::BuiltInTaskCountNV; 940 case glslang::EbvPrimitiveCountNV: 941 return spv::BuiltInPrimitiveCountNV; 942 case glslang::EbvPrimitiveIndicesNV: 943 return spv::BuiltInPrimitiveIndicesNV; 944 case glslang::EbvClipDistancePerViewNV: 945 return spv::BuiltInClipDistancePerViewNV; 946 case glslang::EbvCullDistancePerViewNV: 947 return spv::BuiltInCullDistancePerViewNV; 948 case glslang::EbvLayerPerViewNV: 949 return spv::BuiltInLayerPerViewNV; 950 case glslang::EbvMeshViewCountNV: 951 return spv::BuiltInMeshViewCountNV; 952 case glslang::EbvMeshViewIndicesNV: 953 return spv::BuiltInMeshViewIndicesNV; 954 #endif 955 default: 956 return spv::BuiltInMax; 957 } 958 } 959 960 // Translate glslang image layout format to SPIR-V image format. 961 spv::ImageFormat TGlslangToSpvTraverser::TranslateImageFormat(const glslang::TType& type) 962 { 963 assert(type.getBasicType() == glslang::EbtSampler); 964 965 // Check for capabilities 966 switch (type.getQualifier().layoutFormat) { 967 case glslang::ElfRg32f: 968 case glslang::ElfRg16f: 969 case glslang::ElfR11fG11fB10f: 970 case glslang::ElfR16f: 971 case glslang::ElfRgba16: 972 case glslang::ElfRgb10A2: 973 case glslang::ElfRg16: 974 case glslang::ElfRg8: 975 case glslang::ElfR16: 976 case glslang::ElfR8: 977 case glslang::ElfRgba16Snorm: 978 case glslang::ElfRg16Snorm: 979 case glslang::ElfRg8Snorm: 980 case glslang::ElfR16Snorm: 981 case glslang::ElfR8Snorm: 982 983 case glslang::ElfRg32i: 984 case glslang::ElfRg16i: 985 case glslang::ElfRg8i: 986 case glslang::ElfR16i: 987 case glslang::ElfR8i: 988 989 case glslang::ElfRgb10a2ui: 990 case glslang::ElfRg32ui: 991 case glslang::ElfRg16ui: 992 case glslang::ElfRg8ui: 993 case glslang::ElfR16ui: 994 case glslang::ElfR8ui: 995 builder.addCapability(spv::CapabilityStorageImageExtendedFormats); 996 break; 997 998 default: 999 break; 1000 } 1001 1002 // do the translation 1003 switch (type.getQualifier().layoutFormat) { 1004 case glslang::ElfNone: return spv::ImageFormatUnknown; 1005 case glslang::ElfRgba32f: return spv::ImageFormatRgba32f; 1006 case glslang::ElfRgba16f: return spv::ImageFormatRgba16f; 1007 case glslang::ElfR32f: return spv::ImageFormatR32f; 1008 case glslang::ElfRgba8: return spv::ImageFormatRgba8; 1009 case glslang::ElfRgba8Snorm: return spv::ImageFormatRgba8Snorm; 1010 case glslang::ElfRg32f: return spv::ImageFormatRg32f; 1011 case glslang::ElfRg16f: return spv::ImageFormatRg16f; 1012 case glslang::ElfR11fG11fB10f: return spv::ImageFormatR11fG11fB10f; 1013 case glslang::ElfR16f: return spv::ImageFormatR16f; 1014 case glslang::ElfRgba16: return spv::ImageFormatRgba16; 1015 case glslang::ElfRgb10A2: return spv::ImageFormatRgb10A2; 1016 case glslang::ElfRg16: return spv::ImageFormatRg16; 1017 case glslang::ElfRg8: return spv::ImageFormatRg8; 1018 case glslang::ElfR16: return spv::ImageFormatR16; 1019 case glslang::ElfR8: return spv::ImageFormatR8; 1020 case glslang::ElfRgba16Snorm: return spv::ImageFormatRgba16Snorm; 1021 case glslang::ElfRg16Snorm: return spv::ImageFormatRg16Snorm; 1022 case glslang::ElfRg8Snorm: return spv::ImageFormatRg8Snorm; 1023 case glslang::ElfR16Snorm: return spv::ImageFormatR16Snorm; 1024 case glslang::ElfR8Snorm: return spv::ImageFormatR8Snorm; 1025 case glslang::ElfRgba32i: return spv::ImageFormatRgba32i; 1026 case glslang::ElfRgba16i: return spv::ImageFormatRgba16i; 1027 case glslang::ElfRgba8i: return spv::ImageFormatRgba8i; 1028 case glslang::ElfR32i: return spv::ImageFormatR32i; 1029 case glslang::ElfRg32i: return spv::ImageFormatRg32i; 1030 case glslang::ElfRg16i: return spv::ImageFormatRg16i; 1031 case glslang::ElfRg8i: return spv::ImageFormatRg8i; 1032 case glslang::ElfR16i: return spv::ImageFormatR16i; 1033 case glslang::ElfR8i: return spv::ImageFormatR8i; 1034 case glslang::ElfRgba32ui: return spv::ImageFormatRgba32ui; 1035 case glslang::ElfRgba16ui: return spv::ImageFormatRgba16ui; 1036 case glslang::ElfRgba8ui: return spv::ImageFormatRgba8ui; 1037 case glslang::ElfR32ui: return spv::ImageFormatR32ui; 1038 case glslang::ElfRg32ui: return spv::ImageFormatRg32ui; 1039 case glslang::ElfRg16ui: return spv::ImageFormatRg16ui; 1040 case glslang::ElfRgb10a2ui: return spv::ImageFormatRgb10a2ui; 1041 case glslang::ElfRg8ui: return spv::ImageFormatRg8ui; 1042 case glslang::ElfR16ui: return spv::ImageFormatR16ui; 1043 case glslang::ElfR8ui: return spv::ImageFormatR8ui; 1044 default: return spv::ImageFormatMax; 1045 } 1046 } 1047 1048 spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSelectionControl(const glslang::TIntermSelection& selectionNode) const 1049 { 1050 if (selectionNode.getFlatten()) 1051 return spv::SelectionControlFlattenMask; 1052 if (selectionNode.getDontFlatten()) 1053 return spv::SelectionControlDontFlattenMask; 1054 return spv::SelectionControlMaskNone; 1055 } 1056 1057 spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSwitchControl(const glslang::TIntermSwitch& switchNode) const 1058 { 1059 if (switchNode.getFlatten()) 1060 return spv::SelectionControlFlattenMask; 1061 if (switchNode.getDontFlatten()) 1062 return spv::SelectionControlDontFlattenMask; 1063 return spv::SelectionControlMaskNone; 1064 } 1065 1066 // return a non-0 dependency if the dependency argument must be set 1067 spv::LoopControlMask TGlslangToSpvTraverser::TranslateLoopControl(const glslang::TIntermLoop& loopNode, 1068 unsigned int& dependencyLength) const 1069 { 1070 spv::LoopControlMask control = spv::LoopControlMaskNone; 1071 1072 if (loopNode.getDontUnroll()) 1073 control = control | spv::LoopControlDontUnrollMask; 1074 if (loopNode.getUnroll()) 1075 control = control | spv::LoopControlUnrollMask; 1076 if (unsigned(loopNode.getLoopDependency()) == glslang::TIntermLoop::dependencyInfinite) 1077 control = control | spv::LoopControlDependencyInfiniteMask; 1078 else if (loopNode.getLoopDependency() > 0) { 1079 control = control | spv::LoopControlDependencyLengthMask; 1080 dependencyLength = loopNode.getLoopDependency(); 1081 } 1082 1083 return control; 1084 } 1085 1086 // Translate glslang type to SPIR-V storage class. 1087 spv::StorageClass TGlslangToSpvTraverser::TranslateStorageClass(const glslang::TType& type) 1088 { 1089 if (type.getQualifier().isPipeInput()) 1090 return spv::StorageClassInput; 1091 if (type.getQualifier().isPipeOutput()) 1092 return spv::StorageClassOutput; 1093 1094 if (glslangIntermediate->getSource() != glslang::EShSourceHlsl || 1095 type.getQualifier().storage == glslang::EvqUniform) { 1096 if (type.getBasicType() == glslang::EbtAtomicUint) 1097 return spv::StorageClassAtomicCounter; 1098 if (type.containsOpaque()) 1099 return spv::StorageClassUniformConstant; 1100 } 1101 1102 #ifdef NV_EXTENSIONS 1103 if (type.getQualifier().isUniformOrBuffer() && 1104 type.getQualifier().layoutShaderRecordNV) { 1105 return spv::StorageClassShaderRecordBufferNV; 1106 } 1107 #endif 1108 1109 if (glslangIntermediate->usingStorageBuffer() && type.getQualifier().storage == glslang::EvqBuffer) { 1110 addPre13Extension(spv::E_SPV_KHR_storage_buffer_storage_class); 1111 return spv::StorageClassStorageBuffer; 1112 } 1113 1114 if (type.getQualifier().isUniformOrBuffer()) { 1115 if (type.getQualifier().layoutPushConstant) 1116 return spv::StorageClassPushConstant; 1117 if (type.getBasicType() == glslang::EbtBlock) 1118 return spv::StorageClassUniform; 1119 return spv::StorageClassUniformConstant; 1120 } 1121 1122 switch (type.getQualifier().storage) { 1123 case glslang::EvqShared: return spv::StorageClassWorkgroup; 1124 case glslang::EvqGlobal: return spv::StorageClassPrivate; 1125 case glslang::EvqConstReadOnly: return spv::StorageClassFunction; 1126 case glslang::EvqTemporary: return spv::StorageClassFunction; 1127 #ifdef NV_EXTENSIONS 1128 case glslang::EvqPayloadNV: return spv::StorageClassRayPayloadNV; 1129 case glslang::EvqPayloadInNV: return spv::StorageClassIncomingRayPayloadNV; 1130 case glslang::EvqHitAttrNV: return spv::StorageClassHitAttributeNV; 1131 case glslang::EvqCallableDataNV: return spv::StorageClassCallableDataNV; 1132 case glslang::EvqCallableDataInNV: return spv::StorageClassIncomingCallableDataNV; 1133 #endif 1134 default: 1135 assert(0); 1136 break; 1137 } 1138 1139 return spv::StorageClassFunction; 1140 } 1141 1142 // Add capabilities pertaining to how an array is indexed. 1143 void TGlslangToSpvTraverser::addIndirectionIndexCapabilities(const glslang::TType& baseType, 1144 const glslang::TType& indexType) 1145 { 1146 if (indexType.getQualifier().isNonUniform()) { 1147 // deal with an asserted non-uniform index 1148 // SPV_EXT_descriptor_indexing already added in TranslateNonUniformDecoration 1149 if (baseType.getBasicType() == glslang::EbtSampler) { 1150 if (baseType.getQualifier().hasAttachment()) 1151 builder.addCapability(spv::CapabilityInputAttachmentArrayNonUniformIndexingEXT); 1152 else if (baseType.isImage() && baseType.getSampler().dim == glslang::EsdBuffer) 1153 builder.addCapability(spv::CapabilityStorageTexelBufferArrayNonUniformIndexingEXT); 1154 else if (baseType.isTexture() && baseType.getSampler().dim == glslang::EsdBuffer) 1155 builder.addCapability(spv::CapabilityUniformTexelBufferArrayNonUniformIndexingEXT); 1156 else if (baseType.isImage()) 1157 builder.addCapability(spv::CapabilityStorageImageArrayNonUniformIndexingEXT); 1158 else if (baseType.isTexture()) 1159 builder.addCapability(spv::CapabilitySampledImageArrayNonUniformIndexingEXT); 1160 } else if (baseType.getBasicType() == glslang::EbtBlock) { 1161 if (baseType.getQualifier().storage == glslang::EvqBuffer) 1162 builder.addCapability(spv::CapabilityStorageBufferArrayNonUniformIndexingEXT); 1163 else if (baseType.getQualifier().storage == glslang::EvqUniform) 1164 builder.addCapability(spv::CapabilityUniformBufferArrayNonUniformIndexingEXT); 1165 } 1166 } else { 1167 // assume a dynamically uniform index 1168 if (baseType.getBasicType() == glslang::EbtSampler) { 1169 if (baseType.getQualifier().hasAttachment()) { 1170 builder.addExtension("SPV_EXT_descriptor_indexing"); 1171 builder.addCapability(spv::CapabilityInputAttachmentArrayDynamicIndexingEXT); 1172 } else if (baseType.isImage() && baseType.getSampler().dim == glslang::EsdBuffer) { 1173 builder.addExtension("SPV_EXT_descriptor_indexing"); 1174 builder.addCapability(spv::CapabilityStorageTexelBufferArrayDynamicIndexingEXT); 1175 } else if (baseType.isTexture() && baseType.getSampler().dim == glslang::EsdBuffer) { 1176 builder.addExtension("SPV_EXT_descriptor_indexing"); 1177 builder.addCapability(spv::CapabilityUniformTexelBufferArrayDynamicIndexingEXT); 1178 } 1179 } 1180 } 1181 } 1182 1183 // Return whether or not the given type is something that should be tied to a 1184 // descriptor set. 1185 bool IsDescriptorResource(const glslang::TType& type) 1186 { 1187 // uniform and buffer blocks are included, unless it is a push_constant 1188 if (type.getBasicType() == glslang::EbtBlock) 1189 return type.getQualifier().isUniformOrBuffer() && 1190 #ifdef NV_EXTENSIONS 1191 ! type.getQualifier().layoutShaderRecordNV && 1192 #endif 1193 ! type.getQualifier().layoutPushConstant; 1194 1195 // non block... 1196 // basically samplerXXX/subpass/sampler/texture are all included 1197 // if they are the global-scope-class, not the function parameter 1198 // (or local, if they ever exist) class. 1199 if (type.getBasicType() == glslang::EbtSampler) 1200 return type.getQualifier().isUniformOrBuffer(); 1201 1202 // None of the above. 1203 return false; 1204 } 1205 1206 void InheritQualifiers(glslang::TQualifier& child, const glslang::TQualifier& parent) 1207 { 1208 if (child.layoutMatrix == glslang::ElmNone) 1209 child.layoutMatrix = parent.layoutMatrix; 1210 1211 if (parent.invariant) 1212 child.invariant = true; 1213 if (parent.nopersp) 1214 child.nopersp = true; 1215 #ifdef AMD_EXTENSIONS 1216 if (parent.explicitInterp) 1217 child.explicitInterp = true; 1218 #endif 1219 if (parent.flat) 1220 child.flat = true; 1221 if (parent.centroid) 1222 child.centroid = true; 1223 if (parent.patch) 1224 child.patch = true; 1225 if (parent.sample) 1226 child.sample = true; 1227 if (parent.coherent) 1228 child.coherent = true; 1229 if (parent.devicecoherent) 1230 child.devicecoherent = true; 1231 if (parent.queuefamilycoherent) 1232 child.queuefamilycoherent = true; 1233 if (parent.workgroupcoherent) 1234 child.workgroupcoherent = true; 1235 if (parent.subgroupcoherent) 1236 child.subgroupcoherent = true; 1237 if (parent.nonprivate) 1238 child.nonprivate = true; 1239 if (parent.volatil) 1240 child.volatil = true; 1241 if (parent.restrict) 1242 child.restrict = true; 1243 if (parent.readonly) 1244 child.readonly = true; 1245 if (parent.writeonly) 1246 child.writeonly = true; 1247 #ifdef NV_EXTENSIONS 1248 if (parent.perPrimitiveNV) 1249 child.perPrimitiveNV = true; 1250 if (parent.perViewNV) 1251 child.perViewNV = true; 1252 if (parent.perTaskNV) 1253 child.perTaskNV = true; 1254 #endif 1255 } 1256 1257 bool HasNonLayoutQualifiers(const glslang::TType& type, const glslang::TQualifier& qualifier) 1258 { 1259 // This should list qualifiers that simultaneous satisfy: 1260 // - struct members might inherit from a struct declaration 1261 // (note that non-block structs don't explicitly inherit, 1262 // only implicitly, meaning no decoration involved) 1263 // - affect decorations on the struct members 1264 // (note smooth does not, and expecting something like volatile 1265 // to effect the whole object) 1266 // - are not part of the offset/st430/etc or row/column-major layout 1267 return qualifier.invariant || (qualifier.hasLocation() && type.getBasicType() == glslang::EbtBlock); 1268 } 1269 1270 // 1271 // Implement the TGlslangToSpvTraverser class. 1272 // 1273 1274 TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const glslang::TIntermediate* glslangIntermediate, 1275 spv::SpvBuildLogger* buildLogger, glslang::SpvOptions& options) 1276 : TIntermTraverser(true, false, true), 1277 options(options), 1278 shaderEntry(nullptr), currentFunction(nullptr), 1279 sequenceDepth(0), logger(buildLogger), 1280 builder(spvVersion, (glslang::GetKhronosToolId() << 16) | glslang::GetSpirvGeneratorVersion(), logger), 1281 inEntryPoint(false), entryPointTerminated(false), linkageOnly(false), 1282 glslangIntermediate(glslangIntermediate) 1283 { 1284 spv::ExecutionModel executionModel = TranslateExecutionModel(glslangIntermediate->getStage()); 1285 1286 builder.clearAccessChain(); 1287 builder.setSource(TranslateSourceLanguage(glslangIntermediate->getSource(), glslangIntermediate->getProfile()), 1288 glslangIntermediate->getVersion()); 1289 1290 if (options.generateDebugInfo) { 1291 builder.setEmitOpLines(); 1292 builder.setSourceFile(glslangIntermediate->getSourceFile()); 1293 1294 // Set the source shader's text. If for SPV version 1.0, include 1295 // a preamble in comments stating the OpModuleProcessed instructions. 1296 // Otherwise, emit those as actual instructions. 1297 std::string text; 1298 const std::vector<std::string>& processes = glslangIntermediate->getProcesses(); 1299 for (int p = 0; p < (int)processes.size(); ++p) { 1300 if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_1) { 1301 text.append("// OpModuleProcessed "); 1302 text.append(processes[p]); 1303 text.append("\n"); 1304 } else 1305 builder.addModuleProcessed(processes[p]); 1306 } 1307 if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_1 && (int)processes.size() > 0) 1308 text.append("#line 1\n"); 1309 text.append(glslangIntermediate->getSourceText()); 1310 builder.setSourceText(text); 1311 // Pass name and text for all included files 1312 const std::map<std::string, std::string>& include_txt = glslangIntermediate->getIncludeText(); 1313 for (auto iItr = include_txt.begin(); iItr != include_txt.end(); ++iItr) 1314 builder.addInclude(iItr->first, iItr->second); 1315 } 1316 stdBuiltins = builder.import("GLSL.std.450"); 1317 1318 spv::AddressingModel addressingModel = spv::AddressingModelLogical; 1319 spv::MemoryModel memoryModel = spv::MemoryModelGLSL450; 1320 1321 if (glslangIntermediate->usingPhysicalStorageBuffer()) { 1322 addressingModel = spv::AddressingModelPhysicalStorageBuffer64EXT; 1323 builder.addExtension(spv::E_SPV_EXT_physical_storage_buffer); 1324 builder.addCapability(spv::CapabilityPhysicalStorageBufferAddressesEXT); 1325 }; 1326 if (glslangIntermediate->usingVulkanMemoryModel()) { 1327 memoryModel = spv::MemoryModelVulkanKHR; 1328 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); 1329 builder.addExtension(spv::E_SPV_KHR_vulkan_memory_model); 1330 } 1331 builder.setMemoryModel(addressingModel, memoryModel); 1332 1333 shaderEntry = builder.makeEntryPoint(glslangIntermediate->getEntryPointName().c_str()); 1334 entryPoint = builder.addEntryPoint(executionModel, shaderEntry, glslangIntermediate->getEntryPointName().c_str()); 1335 1336 // Add the source extensions 1337 const auto& sourceExtensions = glslangIntermediate->getRequestedExtensions(); 1338 for (auto it = sourceExtensions.begin(); it != sourceExtensions.end(); ++it) 1339 builder.addSourceExtension(it->c_str()); 1340 1341 // Add the top-level modes for this shader. 1342 1343 if (glslangIntermediate->getXfbMode()) { 1344 builder.addCapability(spv::CapabilityTransformFeedback); 1345 builder.addExecutionMode(shaderEntry, spv::ExecutionModeXfb); 1346 } 1347 1348 unsigned int mode; 1349 switch (glslangIntermediate->getStage()) { 1350 case EShLangVertex: 1351 builder.addCapability(spv::CapabilityShader); 1352 break; 1353 1354 case EShLangTessEvaluation: 1355 case EShLangTessControl: 1356 builder.addCapability(spv::CapabilityTessellation); 1357 1358 glslang::TLayoutGeometry primitive; 1359 1360 if (glslangIntermediate->getStage() == EShLangTessControl) { 1361 builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices()); 1362 primitive = glslangIntermediate->getOutputPrimitive(); 1363 } else { 1364 primitive = glslangIntermediate->getInputPrimitive(); 1365 } 1366 1367 switch (primitive) { 1368 case glslang::ElgTriangles: mode = spv::ExecutionModeTriangles; break; 1369 case glslang::ElgQuads: mode = spv::ExecutionModeQuads; break; 1370 case glslang::ElgIsolines: mode = spv::ExecutionModeIsolines; break; 1371 default: mode = spv::ExecutionModeMax; break; 1372 } 1373 if (mode != spv::ExecutionModeMax) 1374 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); 1375 1376 switch (glslangIntermediate->getVertexSpacing()) { 1377 case glslang::EvsEqual: mode = spv::ExecutionModeSpacingEqual; break; 1378 case glslang::EvsFractionalEven: mode = spv::ExecutionModeSpacingFractionalEven; break; 1379 case glslang::EvsFractionalOdd: mode = spv::ExecutionModeSpacingFractionalOdd; break; 1380 default: mode = spv::ExecutionModeMax; break; 1381 } 1382 if (mode != spv::ExecutionModeMax) 1383 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); 1384 1385 switch (glslangIntermediate->getVertexOrder()) { 1386 case glslang::EvoCw: mode = spv::ExecutionModeVertexOrderCw; break; 1387 case glslang::EvoCcw: mode = spv::ExecutionModeVertexOrderCcw; break; 1388 default: mode = spv::ExecutionModeMax; break; 1389 } 1390 if (mode != spv::ExecutionModeMax) 1391 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); 1392 1393 if (glslangIntermediate->getPointMode()) 1394 builder.addExecutionMode(shaderEntry, spv::ExecutionModePointMode); 1395 break; 1396 1397 case EShLangGeometry: 1398 builder.addCapability(spv::CapabilityGeometry); 1399 switch (glslangIntermediate->getInputPrimitive()) { 1400 case glslang::ElgPoints: mode = spv::ExecutionModeInputPoints; break; 1401 case glslang::ElgLines: mode = spv::ExecutionModeInputLines; break; 1402 case glslang::ElgLinesAdjacency: mode = spv::ExecutionModeInputLinesAdjacency; break; 1403 case glslang::ElgTriangles: mode = spv::ExecutionModeTriangles; break; 1404 case glslang::ElgTrianglesAdjacency: mode = spv::ExecutionModeInputTrianglesAdjacency; break; 1405 default: mode = spv::ExecutionModeMax; break; 1406 } 1407 if (mode != spv::ExecutionModeMax) 1408 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); 1409 1410 builder.addExecutionMode(shaderEntry, spv::ExecutionModeInvocations, glslangIntermediate->getInvocations()); 1411 1412 switch (glslangIntermediate->getOutputPrimitive()) { 1413 case glslang::ElgPoints: mode = spv::ExecutionModeOutputPoints; break; 1414 case glslang::ElgLineStrip: mode = spv::ExecutionModeOutputLineStrip; break; 1415 case glslang::ElgTriangleStrip: mode = spv::ExecutionModeOutputTriangleStrip; break; 1416 default: mode = spv::ExecutionModeMax; break; 1417 } 1418 if (mode != spv::ExecutionModeMax) 1419 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); 1420 builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices()); 1421 break; 1422 1423 case EShLangFragment: 1424 builder.addCapability(spv::CapabilityShader); 1425 if (glslangIntermediate->getPixelCenterInteger()) 1426 builder.addExecutionMode(shaderEntry, spv::ExecutionModePixelCenterInteger); 1427 1428 if (glslangIntermediate->getOriginUpperLeft()) 1429 builder.addExecutionMode(shaderEntry, spv::ExecutionModeOriginUpperLeft); 1430 else 1431 builder.addExecutionMode(shaderEntry, spv::ExecutionModeOriginLowerLeft); 1432 1433 if (glslangIntermediate->getEarlyFragmentTests()) 1434 builder.addExecutionMode(shaderEntry, spv::ExecutionModeEarlyFragmentTests); 1435 1436 if (glslangIntermediate->getPostDepthCoverage()) { 1437 builder.addCapability(spv::CapabilitySampleMaskPostDepthCoverage); 1438 builder.addExecutionMode(shaderEntry, spv::ExecutionModePostDepthCoverage); 1439 builder.addExtension(spv::E_SPV_KHR_post_depth_coverage); 1440 } 1441 1442 switch(glslangIntermediate->getDepth()) { 1443 case glslang::EldGreater: mode = spv::ExecutionModeDepthGreater; break; 1444 case glslang::EldLess: mode = spv::ExecutionModeDepthLess; break; 1445 default: mode = spv::ExecutionModeMax; break; 1446 } 1447 if (mode != spv::ExecutionModeMax) 1448 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); 1449 1450 if (glslangIntermediate->getDepth() != glslang::EldUnchanged && glslangIntermediate->isDepthReplacing()) 1451 builder.addExecutionMode(shaderEntry, spv::ExecutionModeDepthReplacing); 1452 break; 1453 1454 case EShLangCompute: 1455 builder.addCapability(spv::CapabilityShader); 1456 builder.addExecutionMode(shaderEntry, spv::ExecutionModeLocalSize, glslangIntermediate->getLocalSize(0), 1457 glslangIntermediate->getLocalSize(1), 1458 glslangIntermediate->getLocalSize(2)); 1459 #ifdef NV_EXTENSIONS 1460 if (glslangIntermediate->getLayoutDerivativeModeNone() == glslang::LayoutDerivativeGroupQuads) { 1461 builder.addCapability(spv::CapabilityComputeDerivativeGroupQuadsNV); 1462 builder.addExecutionMode(shaderEntry, spv::ExecutionModeDerivativeGroupQuadsNV); 1463 builder.addExtension(spv::E_SPV_NV_compute_shader_derivatives); 1464 } else if (glslangIntermediate->getLayoutDerivativeModeNone() == glslang::LayoutDerivativeGroupLinear) { 1465 builder.addCapability(spv::CapabilityComputeDerivativeGroupLinearNV); 1466 builder.addExecutionMode(shaderEntry, spv::ExecutionModeDerivativeGroupLinearNV); 1467 builder.addExtension(spv::E_SPV_NV_compute_shader_derivatives); 1468 } 1469 #endif 1470 break; 1471 1472 #ifdef NV_EXTENSIONS 1473 case EShLangRayGenNV: 1474 case EShLangIntersectNV: 1475 case EShLangAnyHitNV: 1476 case EShLangClosestHitNV: 1477 case EShLangMissNV: 1478 case EShLangCallableNV: 1479 builder.addCapability(spv::CapabilityRayTracingNV); 1480 builder.addExtension("SPV_NV_ray_tracing"); 1481 break; 1482 case EShLangTaskNV: 1483 case EShLangMeshNV: 1484 builder.addCapability(spv::CapabilityMeshShadingNV); 1485 builder.addExtension(spv::E_SPV_NV_mesh_shader); 1486 builder.addExecutionMode(shaderEntry, spv::ExecutionModeLocalSize, glslangIntermediate->getLocalSize(0), 1487 glslangIntermediate->getLocalSize(1), 1488 glslangIntermediate->getLocalSize(2)); 1489 if (glslangIntermediate->getStage() == EShLangMeshNV) { 1490 builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices()); 1491 builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputPrimitivesNV, glslangIntermediate->getPrimitives()); 1492 1493 switch (glslangIntermediate->getOutputPrimitive()) { 1494 case glslang::ElgPoints: mode = spv::ExecutionModeOutputPoints; break; 1495 case glslang::ElgLines: mode = spv::ExecutionModeOutputLinesNV; break; 1496 case glslang::ElgTriangles: mode = spv::ExecutionModeOutputTrianglesNV; break; 1497 default: mode = spv::ExecutionModeMax; break; 1498 } 1499 if (mode != spv::ExecutionModeMax) 1500 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode); 1501 } 1502 break; 1503 #endif 1504 1505 default: 1506 break; 1507 } 1508 } 1509 1510 // Finish creating SPV, after the traversal is complete. 1511 void TGlslangToSpvTraverser::finishSpv() 1512 { 1513 // Finish the entry point function 1514 if (! entryPointTerminated) { 1515 builder.setBuildPoint(shaderEntry->getLastBlock()); 1516 builder.leaveFunction(); 1517 } 1518 1519 // finish off the entry-point SPV instruction by adding the Input/Output <id> 1520 for (auto it = iOSet.cbegin(); it != iOSet.cend(); ++it) 1521 entryPoint->addIdOperand(*it); 1522 1523 // Add capabilities, extensions, remove unneeded decorations, etc., 1524 // based on the resulting SPIR-V. 1525 builder.postProcess(); 1526 } 1527 1528 // Write the SPV into 'out'. 1529 void TGlslangToSpvTraverser::dumpSpv(std::vector<unsigned int>& out) 1530 { 1531 builder.dump(out); 1532 } 1533 1534 // 1535 // Implement the traversal functions. 1536 // 1537 // Return true from interior nodes to have the external traversal 1538 // continue on to children. Return false if children were 1539 // already processed. 1540 // 1541 1542 // 1543 // Symbols can turn into 1544 // - uniform/input reads 1545 // - output writes 1546 // - complex lvalue base setups: foo.bar[3].... , where we see foo and start up an access chain 1547 // - something simple that degenerates into the last bullet 1548 // 1549 void TGlslangToSpvTraverser::visitSymbol(glslang::TIntermSymbol* symbol) 1550 { 1551 SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder); 1552 if (symbol->getType().getQualifier().isSpecConstant()) 1553 spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); 1554 1555 // getSymbolId() will set up all the IO decorations on the first call. 1556 // Formal function parameters were mapped during makeFunctions(). 1557 spv::Id id = getSymbolId(symbol); 1558 1559 // Include all "static use" and "linkage only" interface variables on the OpEntryPoint instruction 1560 if (builder.isPointer(id)) { 1561 spv::StorageClass sc = builder.getStorageClass(id); 1562 if (sc == spv::StorageClassInput || sc == spv::StorageClassOutput) { 1563 if (!symbol->getType().isStruct() || symbol->getType().getStruct()->size() > 0) 1564 iOSet.insert(id); 1565 } 1566 } 1567 1568 // Only process non-linkage-only nodes for generating actual static uses 1569 if (! linkageOnly || symbol->getQualifier().isSpecConstant()) { 1570 // Prepare to generate code for the access 1571 1572 // L-value chains will be computed left to right. We're on the symbol now, 1573 // which is the left-most part of the access chain, so now is "clear" time, 1574 // followed by setting the base. 1575 builder.clearAccessChain(); 1576 1577 // For now, we consider all user variables as being in memory, so they are pointers, 1578 // except for 1579 // A) R-Value arguments to a function, which are an intermediate object. 1580 // See comments in handleUserFunctionCall(). 1581 // B) Specialization constants (normal constants don't even come in as a variable), 1582 // These are also pure R-values. 1583 glslang::TQualifier qualifier = symbol->getQualifier(); 1584 if (qualifier.isSpecConstant() || rValueParameters.find(symbol->getId()) != rValueParameters.end()) 1585 builder.setAccessChainRValue(id); 1586 else 1587 builder.setAccessChainLValue(id); 1588 } 1589 1590 // Process linkage-only nodes for any special additional interface work. 1591 if (linkageOnly) { 1592 if (glslangIntermediate->getHlslFunctionality1()) { 1593 // Map implicit counter buffers to their originating buffers, which should have been 1594 // seen by now, given earlier pruning of unused counters, and preservation of order 1595 // of declaration. 1596 if (symbol->getType().getQualifier().isUniformOrBuffer()) { 1597 if (!glslangIntermediate->hasCounterBufferName(symbol->getName())) { 1598 // Save possible originating buffers for counter buffers, keyed by 1599 // making the potential counter-buffer name. 1600 std::string keyName = symbol->getName().c_str(); 1601 keyName = glslangIntermediate->addCounterBufferName(keyName); 1602 counterOriginator[keyName] = symbol; 1603 } else { 1604 // Handle a counter buffer, by finding the saved originating buffer. 1605 std::string keyName = symbol->getName().c_str(); 1606 auto it = counterOriginator.find(keyName); 1607 if (it != counterOriginator.end()) { 1608 id = getSymbolId(it->second); 1609 if (id != spv::NoResult) { 1610 spv::Id counterId = getSymbolId(symbol); 1611 if (counterId != spv::NoResult) { 1612 builder.addExtension("SPV_GOOGLE_hlsl_functionality1"); 1613 builder.addDecorationId(id, spv::DecorationHlslCounterBufferGOOGLE, counterId); 1614 } 1615 } 1616 } 1617 } 1618 } 1619 } 1620 } 1621 } 1622 1623 bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::TIntermBinary* node) 1624 { 1625 builder.setLine(node->getLoc().line, node->getLoc().getFilename()); 1626 1627 SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder); 1628 if (node->getType().getQualifier().isSpecConstant()) 1629 spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); 1630 1631 // First, handle special cases 1632 switch (node->getOp()) { 1633 case glslang::EOpAssign: 1634 case glslang::EOpAddAssign: 1635 case glslang::EOpSubAssign: 1636 case glslang::EOpMulAssign: 1637 case glslang::EOpVectorTimesMatrixAssign: 1638 case glslang::EOpVectorTimesScalarAssign: 1639 case glslang::EOpMatrixTimesScalarAssign: 1640 case glslang::EOpMatrixTimesMatrixAssign: 1641 case glslang::EOpDivAssign: 1642 case glslang::EOpModAssign: 1643 case glslang::EOpAndAssign: 1644 case glslang::EOpInclusiveOrAssign: 1645 case glslang::EOpExclusiveOrAssign: 1646 case glslang::EOpLeftShiftAssign: 1647 case glslang::EOpRightShiftAssign: 1648 // A bin-op assign "a += b" means the same thing as "a = a + b" 1649 // where a is evaluated before b. For a simple assignment, GLSL 1650 // says to evaluate the left before the right. So, always, left 1651 // node then right node. 1652 { 1653 // get the left l-value, save it away 1654 builder.clearAccessChain(); 1655 node->getLeft()->traverse(this); 1656 spv::Builder::AccessChain lValue = builder.getAccessChain(); 1657 1658 // evaluate the right 1659 builder.clearAccessChain(); 1660 node->getRight()->traverse(this); 1661 spv::Id rValue = accessChainLoad(node->getRight()->getType()); 1662 1663 if (node->getOp() != glslang::EOpAssign) { 1664 // the left is also an r-value 1665 builder.setAccessChain(lValue); 1666 spv::Id leftRValue = accessChainLoad(node->getLeft()->getType()); 1667 1668 // do the operation 1669 OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()), 1670 TranslateNoContractionDecoration(node->getType().getQualifier()), 1671 TranslateNonUniformDecoration(node->getType().getQualifier()) }; 1672 rValue = createBinaryOperation(node->getOp(), decorations, 1673 convertGlslangToSpvType(node->getType()), leftRValue, rValue, 1674 node->getType().getBasicType()); 1675 1676 // these all need their counterparts in createBinaryOperation() 1677 assert(rValue != spv::NoResult); 1678 } 1679 1680 // store the result 1681 builder.setAccessChain(lValue); 1682 multiTypeStore(node->getLeft()->getType(), rValue); 1683 1684 // assignments are expressions having an rValue after they are evaluated... 1685 builder.clearAccessChain(); 1686 builder.setAccessChainRValue(rValue); 1687 } 1688 return false; 1689 case glslang::EOpIndexDirect: 1690 case glslang::EOpIndexDirectStruct: 1691 { 1692 // Get the left part of the access chain. 1693 node->getLeft()->traverse(this); 1694 1695 // Add the next element in the chain 1696 1697 const int glslangIndex = node->getRight()->getAsConstantUnion()->getConstArray()[0].getIConst(); 1698 if (! node->getLeft()->getType().isArray() && 1699 node->getLeft()->getType().isVector() && 1700 node->getOp() == glslang::EOpIndexDirect) { 1701 // This is essentially a hard-coded vector swizzle of size 1, 1702 // so short circuit the access-chain stuff with a swizzle. 1703 std::vector<unsigned> swizzle; 1704 swizzle.push_back(glslangIndex); 1705 int dummySize; 1706 builder.accessChainPushSwizzle(swizzle, convertGlslangToSpvType(node->getLeft()->getType()), 1707 TranslateCoherent(node->getLeft()->getType()), 1708 glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize)); 1709 } else { 1710 1711 // Load through a block reference is performed with a dot operator that 1712 // is mapped to EOpIndexDirectStruct. When we get to the actual reference, 1713 // do a load and reset the access chain. 1714 if (node->getLeft()->getBasicType() == glslang::EbtReference && 1715 !node->getLeft()->getType().isArray() && 1716 node->getOp() == glslang::EOpIndexDirectStruct) 1717 { 1718 spv::Id left = accessChainLoad(node->getLeft()->getType()); 1719 builder.clearAccessChain(); 1720 builder.setAccessChainLValue(left); 1721 } 1722 1723 int spvIndex = glslangIndex; 1724 if (node->getLeft()->getBasicType() == glslang::EbtBlock && 1725 node->getOp() == glslang::EOpIndexDirectStruct) 1726 { 1727 // This may be, e.g., an anonymous block-member selection, which generally need 1728 // index remapping due to hidden members in anonymous blocks. 1729 std::vector<int>& remapper = memberRemapper[node->getLeft()->getType().getStruct()]; 1730 assert(remapper.size() > 0); 1731 spvIndex = remapper[glslangIndex]; 1732 } 1733 1734 // normal case for indexing array or structure or block 1735 builder.accessChainPush(builder.makeIntConstant(spvIndex), TranslateCoherent(node->getLeft()->getType()), getBufferReferenceAlignment(node->getLeft()->getType())); 1736 1737 // Add capabilities here for accessing PointSize and clip/cull distance. 1738 // We have deferred generation of associated capabilities until now. 1739 if (node->getLeft()->getType().isStruct() && ! node->getLeft()->getType().isArray()) 1740 declareUseOfStructMember(*(node->getLeft()->getType().getStruct()), glslangIndex); 1741 } 1742 } 1743 return false; 1744 case glslang::EOpIndexIndirect: 1745 { 1746 // Structure or array or vector indirection. 1747 // Will use native SPIR-V access-chain for struct and array indirection; 1748 // matrices are arrays of vectors, so will also work for a matrix. 1749 // Will use the access chain's 'component' for variable index into a vector. 1750 1751 // This adapter is building access chains left to right. 1752 // Set up the access chain to the left. 1753 node->getLeft()->traverse(this); 1754 1755 // save it so that computing the right side doesn't trash it 1756 spv::Builder::AccessChain partial = builder.getAccessChain(); 1757 1758 // compute the next index in the chain 1759 builder.clearAccessChain(); 1760 node->getRight()->traverse(this); 1761 spv::Id index = accessChainLoad(node->getRight()->getType()); 1762 1763 addIndirectionIndexCapabilities(node->getLeft()->getType(), node->getRight()->getType()); 1764 1765 // restore the saved access chain 1766 builder.setAccessChain(partial); 1767 1768 if (! node->getLeft()->getType().isArray() && node->getLeft()->getType().isVector()) { 1769 int dummySize; 1770 builder.accessChainPushComponent(index, convertGlslangToSpvType(node->getLeft()->getType()), 1771 TranslateCoherent(node->getLeft()->getType()), 1772 glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize)); 1773 } else 1774 builder.accessChainPush(index, TranslateCoherent(node->getLeft()->getType()), getBufferReferenceAlignment(node->getLeft()->getType())); 1775 } 1776 return false; 1777 case glslang::EOpVectorSwizzle: 1778 { 1779 node->getLeft()->traverse(this); 1780 std::vector<unsigned> swizzle; 1781 convertSwizzle(*node->getRight()->getAsAggregate(), swizzle); 1782 int dummySize; 1783 builder.accessChainPushSwizzle(swizzle, convertGlslangToSpvType(node->getLeft()->getType()), 1784 TranslateCoherent(node->getLeft()->getType()), 1785 glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize)); 1786 } 1787 return false; 1788 case glslang::EOpMatrixSwizzle: 1789 logger->missingFunctionality("matrix swizzle"); 1790 return true; 1791 case glslang::EOpLogicalOr: 1792 case glslang::EOpLogicalAnd: 1793 { 1794 1795 // These may require short circuiting, but can sometimes be done as straight 1796 // binary operations. The right operand must be short circuited if it has 1797 // side effects, and should probably be if it is complex. 1798 if (isTrivial(node->getRight()->getAsTyped())) 1799 break; // handle below as a normal binary operation 1800 // otherwise, we need to do dynamic short circuiting on the right operand 1801 spv::Id result = createShortCircuit(node->getOp(), *node->getLeft()->getAsTyped(), *node->getRight()->getAsTyped()); 1802 builder.clearAccessChain(); 1803 builder.setAccessChainRValue(result); 1804 } 1805 return false; 1806 default: 1807 break; 1808 } 1809 1810 // Assume generic binary op... 1811 1812 // get right operand 1813 builder.clearAccessChain(); 1814 node->getLeft()->traverse(this); 1815 spv::Id left = accessChainLoad(node->getLeft()->getType()); 1816 1817 // get left operand 1818 builder.clearAccessChain(); 1819 node->getRight()->traverse(this); 1820 spv::Id right = accessChainLoad(node->getRight()->getType()); 1821 1822 // get result 1823 OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()), 1824 TranslateNoContractionDecoration(node->getType().getQualifier()), 1825 TranslateNonUniformDecoration(node->getType().getQualifier()) }; 1826 spv::Id result = createBinaryOperation(node->getOp(), decorations, 1827 convertGlslangToSpvType(node->getType()), left, right, 1828 node->getLeft()->getType().getBasicType()); 1829 1830 builder.clearAccessChain(); 1831 if (! result) { 1832 logger->missingFunctionality("unknown glslang binary operation"); 1833 return true; // pick up a child as the place-holder result 1834 } else { 1835 builder.setAccessChainRValue(result); 1836 return false; 1837 } 1838 } 1839 1840 bool TGlslangToSpvTraverser::visitUnary(glslang::TVisit /* visit */, glslang::TIntermUnary* node) 1841 { 1842 builder.setLine(node->getLoc().line, node->getLoc().getFilename()); 1843 1844 SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder); 1845 if (node->getType().getQualifier().isSpecConstant()) 1846 spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); 1847 1848 spv::Id result = spv::NoResult; 1849 1850 // try texturing first 1851 result = createImageTextureFunctionCall(node); 1852 if (result != spv::NoResult) { 1853 builder.clearAccessChain(); 1854 builder.setAccessChainRValue(result); 1855 1856 return false; // done with this node 1857 } 1858 1859 // Non-texturing. 1860 1861 if (node->getOp() == glslang::EOpArrayLength) { 1862 // Quite special; won't want to evaluate the operand. 1863 1864 // Currently, the front-end does not allow .length() on an array until it is sized, 1865 // except for the last block membeor of an SSBO. 1866 // TODO: If this changes, link-time sized arrays might show up here, and need their 1867 // size extracted. 1868 1869 // Normal .length() would have been constant folded by the front-end. 1870 // So, this has to be block.lastMember.length(). 1871 // SPV wants "block" and member number as the operands, go get them. 1872 1873 glslang::TIntermTyped* block = node->getOperand()->getAsBinaryNode()->getLeft(); 1874 block->traverse(this); 1875 unsigned int member = node->getOperand()->getAsBinaryNode()->getRight()->getAsConstantUnion()->getConstArray()[0].getUConst(); 1876 spv::Id length = builder.createArrayLength(builder.accessChainGetLValue(), member); 1877 1878 // GLSL semantics say the result of .length() is an int, while SPIR-V says 1879 // signedness must be 0. So, convert from SPIR-V unsigned back to GLSL's 1880 // AST expectation of a signed result. 1881 if (glslangIntermediate->getSource() == glslang::EShSourceGlsl) 1882 length = builder.createUnaryOp(spv::OpBitcast, builder.makeIntType(32), length); 1883 1884 builder.clearAccessChain(); 1885 builder.setAccessChainRValue(length); 1886 1887 return false; 1888 } 1889 1890 // Start by evaluating the operand 1891 1892 // Does it need a swizzle inversion? If so, evaluation is inverted; 1893 // operate first on the swizzle base, then apply the swizzle. 1894 spv::Id invertedType = spv::NoType; 1895 auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ? invertedType : convertGlslangToSpvType(node->getType()); }; 1896 if (node->getOp() == glslang::EOpInterpolateAtCentroid) 1897 invertedType = getInvertedSwizzleType(*node->getOperand()); 1898 1899 builder.clearAccessChain(); 1900 if (invertedType != spv::NoType) 1901 node->getOperand()->getAsBinaryNode()->getLeft()->traverse(this); 1902 else 1903 node->getOperand()->traverse(this); 1904 1905 spv::Id operand = spv::NoResult; 1906 1907 if (node->getOp() == glslang::EOpAtomicCounterIncrement || 1908 node->getOp() == glslang::EOpAtomicCounterDecrement || 1909 node->getOp() == glslang::EOpAtomicCounter || 1910 node->getOp() == glslang::EOpInterpolateAtCentroid) 1911 operand = builder.accessChainGetLValue(); // Special case l-value operands 1912 else 1913 operand = accessChainLoad(node->getOperand()->getType()); 1914 1915 OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()), 1916 TranslateNoContractionDecoration(node->getType().getQualifier()), 1917 TranslateNonUniformDecoration(node->getType().getQualifier()) }; 1918 1919 // it could be a conversion 1920 if (! result) 1921 result = createConversion(node->getOp(), decorations, resultType(), operand, node->getOperand()->getBasicType()); 1922 1923 // if not, then possibly an operation 1924 if (! result) 1925 result = createUnaryOperation(node->getOp(), decorations, resultType(), operand, node->getOperand()->getBasicType()); 1926 1927 if (result) { 1928 if (invertedType) { 1929 result = createInvertedSwizzle(decorations.precision, *node->getOperand(), result); 1930 builder.addDecoration(result, decorations.nonUniform); 1931 } 1932 1933 builder.clearAccessChain(); 1934 builder.setAccessChainRValue(result); 1935 1936 return false; // done with this node 1937 } 1938 1939 // it must be a special case, check... 1940 switch (node->getOp()) { 1941 case glslang::EOpPostIncrement: 1942 case glslang::EOpPostDecrement: 1943 case glslang::EOpPreIncrement: 1944 case glslang::EOpPreDecrement: 1945 { 1946 // we need the integer value "1" or the floating point "1.0" to add/subtract 1947 spv::Id one = 0; 1948 if (node->getBasicType() == glslang::EbtFloat) 1949 one = builder.makeFloatConstant(1.0F); 1950 else if (node->getBasicType() == glslang::EbtDouble) 1951 one = builder.makeDoubleConstant(1.0); 1952 else if (node->getBasicType() == glslang::EbtFloat16) 1953 one = builder.makeFloat16Constant(1.0F); 1954 else if (node->getBasicType() == glslang::EbtInt8 || node->getBasicType() == glslang::EbtUint8) 1955 one = builder.makeInt8Constant(1); 1956 else if (node->getBasicType() == glslang::EbtInt16 || node->getBasicType() == glslang::EbtUint16) 1957 one = builder.makeInt16Constant(1); 1958 else if (node->getBasicType() == glslang::EbtInt64 || node->getBasicType() == glslang::EbtUint64) 1959 one = builder.makeInt64Constant(1); 1960 else 1961 one = builder.makeIntConstant(1); 1962 glslang::TOperator op; 1963 if (node->getOp() == glslang::EOpPreIncrement || 1964 node->getOp() == glslang::EOpPostIncrement) 1965 op = glslang::EOpAdd; 1966 else 1967 op = glslang::EOpSub; 1968 1969 spv::Id result = createBinaryOperation(op, decorations, 1970 convertGlslangToSpvType(node->getType()), operand, one, 1971 node->getType().getBasicType()); 1972 assert(result != spv::NoResult); 1973 1974 // The result of operation is always stored, but conditionally the 1975 // consumed result. The consumed result is always an r-value. 1976 builder.accessChainStore(result); 1977 builder.clearAccessChain(); 1978 if (node->getOp() == glslang::EOpPreIncrement || 1979 node->getOp() == glslang::EOpPreDecrement) 1980 builder.setAccessChainRValue(result); 1981 else 1982 builder.setAccessChainRValue(operand); 1983 } 1984 1985 return false; 1986 1987 case glslang::EOpEmitStreamVertex: 1988 builder.createNoResultOp(spv::OpEmitStreamVertex, operand); 1989 return false; 1990 case glslang::EOpEndStreamPrimitive: 1991 builder.createNoResultOp(spv::OpEndStreamPrimitive, operand); 1992 return false; 1993 1994 default: 1995 logger->missingFunctionality("unknown glslang unary"); 1996 return true; // pick up operand as placeholder result 1997 } 1998 } 1999 2000 bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TIntermAggregate* node) 2001 { 2002 SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder); 2003 if (node->getType().getQualifier().isSpecConstant()) 2004 spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); 2005 2006 spv::Id result = spv::NoResult; 2007 spv::Id invertedType = spv::NoType; // to use to override the natural type of the node 2008 auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ? invertedType : convertGlslangToSpvType(node->getType()); }; 2009 2010 // try texturing 2011 result = createImageTextureFunctionCall(node); 2012 if (result != spv::NoResult) { 2013 builder.clearAccessChain(); 2014 builder.setAccessChainRValue(result); 2015 2016 return false; 2017 } else if (node->getOp() == glslang::EOpImageStore || 2018 #ifdef AMD_EXTENSIONS 2019 node->getOp() == glslang::EOpImageStoreLod || 2020 #endif 2021 node->getOp() == glslang::EOpImageAtomicStore) { 2022 // "imageStore" is a special case, which has no result 2023 return false; 2024 } 2025 2026 glslang::TOperator binOp = glslang::EOpNull; 2027 bool reduceComparison = true; 2028 bool isMatrix = false; 2029 bool noReturnValue = false; 2030 bool atomic = false; 2031 2032 assert(node->getOp()); 2033 2034 spv::Decoration precision = TranslatePrecisionDecoration(node->getOperationPrecision()); 2035 2036 switch (node->getOp()) { 2037 case glslang::EOpSequence: 2038 { 2039 if (preVisit) 2040 ++sequenceDepth; 2041 else 2042 --sequenceDepth; 2043 2044 if (sequenceDepth == 1) { 2045 // If this is the parent node of all the functions, we want to see them 2046 // early, so all call points have actual SPIR-V functions to reference. 2047 // In all cases, still let the traverser visit the children for us. 2048 makeFunctions(node->getAsAggregate()->getSequence()); 2049 2050 // Also, we want all globals initializers to go into the beginning of the entry point, before 2051 // anything else gets there, so visit out of order, doing them all now. 2052 makeGlobalInitializers(node->getAsAggregate()->getSequence()); 2053 2054 // Initializers are done, don't want to visit again, but functions and link objects need to be processed, 2055 // so do them manually. 2056 visitFunctions(node->getAsAggregate()->getSequence()); 2057 2058 return false; 2059 } 2060 2061 return true; 2062 } 2063 case glslang::EOpLinkerObjects: 2064 { 2065 if (visit == glslang::EvPreVisit) 2066 linkageOnly = true; 2067 else 2068 linkageOnly = false; 2069 2070 return true; 2071 } 2072 case glslang::EOpComma: 2073 { 2074 // processing from left to right naturally leaves the right-most 2075 // lying around in the access chain 2076 glslang::TIntermSequence& glslangOperands = node->getSequence(); 2077 for (int i = 0; i < (int)glslangOperands.size(); ++i) 2078 glslangOperands[i]->traverse(this); 2079 2080 return false; 2081 } 2082 case glslang::EOpFunction: 2083 if (visit == glslang::EvPreVisit) { 2084 if (isShaderEntryPoint(node)) { 2085 inEntryPoint = true; 2086 builder.setBuildPoint(shaderEntry->getLastBlock()); 2087 currentFunction = shaderEntry; 2088 } else { 2089 handleFunctionEntry(node); 2090 } 2091 } else { 2092 if (inEntryPoint) 2093 entryPointTerminated = true; 2094 builder.leaveFunction(); 2095 inEntryPoint = false; 2096 } 2097 2098 return true; 2099 case glslang::EOpParameters: 2100 // Parameters will have been consumed by EOpFunction processing, but not 2101 // the body, so we still visited the function node's children, making this 2102 // child redundant. 2103 return false; 2104 case glslang::EOpFunctionCall: 2105 { 2106 builder.setLine(node->getLoc().line, node->getLoc().getFilename()); 2107 if (node->isUserDefined()) 2108 result = handleUserFunctionCall(node); 2109 // assert(result); // this can happen for bad shaders because the call graph completeness checking is not yet done 2110 if (result) { 2111 builder.clearAccessChain(); 2112 builder.setAccessChainRValue(result); 2113 } else 2114 logger->missingFunctionality("missing user function; linker needs to catch that"); 2115 2116 return false; 2117 } 2118 case glslang::EOpConstructMat2x2: 2119 case glslang::EOpConstructMat2x3: 2120 case glslang::EOpConstructMat2x4: 2121 case glslang::EOpConstructMat3x2: 2122 case glslang::EOpConstructMat3x3: 2123 case glslang::EOpConstructMat3x4: 2124 case glslang::EOpConstructMat4x2: 2125 case glslang::EOpConstructMat4x3: 2126 case glslang::EOpConstructMat4x4: 2127 case glslang::EOpConstructDMat2x2: 2128 case glslang::EOpConstructDMat2x3: 2129 case glslang::EOpConstructDMat2x4: 2130 case glslang::EOpConstructDMat3x2: 2131 case glslang::EOpConstructDMat3x3: 2132 case glslang::EOpConstructDMat3x4: 2133 case glslang::EOpConstructDMat4x2: 2134 case glslang::EOpConstructDMat4x3: 2135 case glslang::EOpConstructDMat4x4: 2136 case glslang::EOpConstructIMat2x2: 2137 case glslang::EOpConstructIMat2x3: 2138 case glslang::EOpConstructIMat2x4: 2139 case glslang::EOpConstructIMat3x2: 2140 case glslang::EOpConstructIMat3x3: 2141 case glslang::EOpConstructIMat3x4: 2142 case glslang::EOpConstructIMat4x2: 2143 case glslang::EOpConstructIMat4x3: 2144 case glslang::EOpConstructIMat4x4: 2145 case glslang::EOpConstructUMat2x2: 2146 case glslang::EOpConstructUMat2x3: 2147 case glslang::EOpConstructUMat2x4: 2148 case glslang::EOpConstructUMat3x2: 2149 case glslang::EOpConstructUMat3x3: 2150 case glslang::EOpConstructUMat3x4: 2151 case glslang::EOpConstructUMat4x2: 2152 case glslang::EOpConstructUMat4x3: 2153 case glslang::EOpConstructUMat4x4: 2154 case glslang::EOpConstructBMat2x2: 2155 case glslang::EOpConstructBMat2x3: 2156 case glslang::EOpConstructBMat2x4: 2157 case glslang::EOpConstructBMat3x2: 2158 case glslang::EOpConstructBMat3x3: 2159 case glslang::EOpConstructBMat3x4: 2160 case glslang::EOpConstructBMat4x2: 2161 case glslang::EOpConstructBMat4x3: 2162 case glslang::EOpConstructBMat4x4: 2163 case glslang::EOpConstructF16Mat2x2: 2164 case glslang::EOpConstructF16Mat2x3: 2165 case glslang::EOpConstructF16Mat2x4: 2166 case glslang::EOpConstructF16Mat3x2: 2167 case glslang::EOpConstructF16Mat3x3: 2168 case glslang::EOpConstructF16Mat3x4: 2169 case glslang::EOpConstructF16Mat4x2: 2170 case glslang::EOpConstructF16Mat4x3: 2171 case glslang::EOpConstructF16Mat4x4: 2172 isMatrix = true; 2173 // fall through 2174 case glslang::EOpConstructFloat: 2175 case glslang::EOpConstructVec2: 2176 case glslang::EOpConstructVec3: 2177 case glslang::EOpConstructVec4: 2178 case glslang::EOpConstructDouble: 2179 case glslang::EOpConstructDVec2: 2180 case glslang::EOpConstructDVec3: 2181 case glslang::EOpConstructDVec4: 2182 case glslang::EOpConstructFloat16: 2183 case glslang::EOpConstructF16Vec2: 2184 case glslang::EOpConstructF16Vec3: 2185 case glslang::EOpConstructF16Vec4: 2186 case glslang::EOpConstructBool: 2187 case glslang::EOpConstructBVec2: 2188 case glslang::EOpConstructBVec3: 2189 case glslang::EOpConstructBVec4: 2190 case glslang::EOpConstructInt8: 2191 case glslang::EOpConstructI8Vec2: 2192 case glslang::EOpConstructI8Vec3: 2193 case glslang::EOpConstructI8Vec4: 2194 case glslang::EOpConstructUint8: 2195 case glslang::EOpConstructU8Vec2: 2196 case glslang::EOpConstructU8Vec3: 2197 case glslang::EOpConstructU8Vec4: 2198 case glslang::EOpConstructInt16: 2199 case glslang::EOpConstructI16Vec2: 2200 case glslang::EOpConstructI16Vec3: 2201 case glslang::EOpConstructI16Vec4: 2202 case glslang::EOpConstructUint16: 2203 case glslang::EOpConstructU16Vec2: 2204 case glslang::EOpConstructU16Vec3: 2205 case glslang::EOpConstructU16Vec4: 2206 case glslang::EOpConstructInt: 2207 case glslang::EOpConstructIVec2: 2208 case glslang::EOpConstructIVec3: 2209 case glslang::EOpConstructIVec4: 2210 case glslang::EOpConstructUint: 2211 case glslang::EOpConstructUVec2: 2212 case glslang::EOpConstructUVec3: 2213 case glslang::EOpConstructUVec4: 2214 case glslang::EOpConstructInt64: 2215 case glslang::EOpConstructI64Vec2: 2216 case glslang::EOpConstructI64Vec3: 2217 case glslang::EOpConstructI64Vec4: 2218 case glslang::EOpConstructUint64: 2219 case glslang::EOpConstructU64Vec2: 2220 case glslang::EOpConstructU64Vec3: 2221 case glslang::EOpConstructU64Vec4: 2222 case glslang::EOpConstructStruct: 2223 case glslang::EOpConstructTextureSampler: 2224 case glslang::EOpConstructReference: 2225 { 2226 builder.setLine(node->getLoc().line, node->getLoc().getFilename()); 2227 std::vector<spv::Id> arguments; 2228 translateArguments(*node, arguments); 2229 spv::Id constructed; 2230 if (node->getOp() == glslang::EOpConstructTextureSampler) 2231 constructed = builder.createOp(spv::OpSampledImage, resultType(), arguments); 2232 else if (node->getOp() == glslang::EOpConstructStruct || node->getType().isArray()) { 2233 std::vector<spv::Id> constituents; 2234 for (int c = 0; c < (int)arguments.size(); ++c) 2235 constituents.push_back(arguments[c]); 2236 constructed = builder.createCompositeConstruct(resultType(), constituents); 2237 } else if (isMatrix) 2238 constructed = builder.createMatrixConstructor(precision, arguments, resultType()); 2239 else 2240 constructed = builder.createConstructor(precision, arguments, resultType()); 2241 2242 builder.clearAccessChain(); 2243 builder.setAccessChainRValue(constructed); 2244 2245 return false; 2246 } 2247 2248 // These six are component-wise compares with component-wise results. 2249 // Forward on to createBinaryOperation(), requesting a vector result. 2250 case glslang::EOpLessThan: 2251 case glslang::EOpGreaterThan: 2252 case glslang::EOpLessThanEqual: 2253 case glslang::EOpGreaterThanEqual: 2254 case glslang::EOpVectorEqual: 2255 case glslang::EOpVectorNotEqual: 2256 { 2257 // Map the operation to a binary 2258 binOp = node->getOp(); 2259 reduceComparison = false; 2260 switch (node->getOp()) { 2261 case glslang::EOpVectorEqual: binOp = glslang::EOpVectorEqual; break; 2262 case glslang::EOpVectorNotEqual: binOp = glslang::EOpVectorNotEqual; break; 2263 default: binOp = node->getOp(); break; 2264 } 2265 2266 break; 2267 } 2268 case glslang::EOpMul: 2269 // component-wise matrix multiply 2270 binOp = glslang::EOpMul; 2271 break; 2272 case glslang::EOpOuterProduct: 2273 // two vectors multiplied to make a matrix 2274 binOp = glslang::EOpOuterProduct; 2275 break; 2276 case glslang::EOpDot: 2277 { 2278 // for scalar dot product, use multiply 2279 glslang::TIntermSequence& glslangOperands = node->getSequence(); 2280 if (glslangOperands[0]->getAsTyped()->getVectorSize() == 1) 2281 binOp = glslang::EOpMul; 2282 break; 2283 } 2284 case glslang::EOpMod: 2285 // when an aggregate, this is the floating-point mod built-in function, 2286 // which can be emitted by the one in createBinaryOperation() 2287 binOp = glslang::EOpMod; 2288 break; 2289 case glslang::EOpEmitVertex: 2290 case glslang::EOpEndPrimitive: 2291 case glslang::EOpBarrier: 2292 case glslang::EOpMemoryBarrier: 2293 case glslang::EOpMemoryBarrierAtomicCounter: 2294 case glslang::EOpMemoryBarrierBuffer: 2295 case glslang::EOpMemoryBarrierImage: 2296 case glslang::EOpMemoryBarrierShared: 2297 case glslang::EOpGroupMemoryBarrier: 2298 case glslang::EOpDeviceMemoryBarrier: 2299 case glslang::EOpAllMemoryBarrierWithGroupSync: 2300 case glslang::EOpDeviceMemoryBarrierWithGroupSync: 2301 case glslang::EOpWorkgroupMemoryBarrier: 2302 case glslang::EOpWorkgroupMemoryBarrierWithGroupSync: 2303 case glslang::EOpSubgroupBarrier: 2304 case glslang::EOpSubgroupMemoryBarrier: 2305 case glslang::EOpSubgroupMemoryBarrierBuffer: 2306 case glslang::EOpSubgroupMemoryBarrierImage: 2307 case glslang::EOpSubgroupMemoryBarrierShared: 2308 noReturnValue = true; 2309 // These all have 0 operands and will naturally finish up in the code below for 0 operands 2310 break; 2311 2312 case glslang::EOpAtomicStore: 2313 noReturnValue = true; 2314 // fallthrough 2315 case glslang::EOpAtomicLoad: 2316 case glslang::EOpAtomicAdd: 2317 case glslang::EOpAtomicMin: 2318 case glslang::EOpAtomicMax: 2319 case glslang::EOpAtomicAnd: 2320 case glslang::EOpAtomicOr: 2321 case glslang::EOpAtomicXor: 2322 case glslang::EOpAtomicExchange: 2323 case glslang::EOpAtomicCompSwap: 2324 atomic = true; 2325 break; 2326 2327 case glslang::EOpAtomicCounterAdd: 2328 case glslang::EOpAtomicCounterSubtract: 2329 case glslang::EOpAtomicCounterMin: 2330 case glslang::EOpAtomicCounterMax: 2331 case glslang::EOpAtomicCounterAnd: 2332 case glslang::EOpAtomicCounterOr: 2333 case glslang::EOpAtomicCounterXor: 2334 case glslang::EOpAtomicCounterExchange: 2335 case glslang::EOpAtomicCounterCompSwap: 2336 builder.addExtension("SPV_KHR_shader_atomic_counter_ops"); 2337 builder.addCapability(spv::CapabilityAtomicStorageOps); 2338 atomic = true; 2339 break; 2340 2341 #ifdef NV_EXTENSIONS 2342 case glslang::EOpIgnoreIntersectionNV: 2343 case glslang::EOpTerminateRayNV: 2344 case glslang::EOpTraceNV: 2345 case glslang::EOpExecuteCallableNV: 2346 case glslang::EOpWritePackedPrimitiveIndices4x8NV: 2347 noReturnValue = true; 2348 break; 2349 #endif 2350 2351 default: 2352 break; 2353 } 2354 2355 // 2356 // See if it maps to a regular operation. 2357 // 2358 if (binOp != glslang::EOpNull) { 2359 glslang::TIntermTyped* left = node->getSequence()[0]->getAsTyped(); 2360 glslang::TIntermTyped* right = node->getSequence()[1]->getAsTyped(); 2361 assert(left && right); 2362 2363 builder.clearAccessChain(); 2364 left->traverse(this); 2365 spv::Id leftId = accessChainLoad(left->getType()); 2366 2367 builder.clearAccessChain(); 2368 right->traverse(this); 2369 spv::Id rightId = accessChainLoad(right->getType()); 2370 2371 builder.setLine(node->getLoc().line, node->getLoc().getFilename()); 2372 OpDecorations decorations = { precision, 2373 TranslateNoContractionDecoration(node->getType().getQualifier()), 2374 TranslateNonUniformDecoration(node->getType().getQualifier()) }; 2375 result = createBinaryOperation(binOp, decorations, 2376 resultType(), leftId, rightId, 2377 left->getType().getBasicType(), reduceComparison); 2378 2379 // code above should only make binOp that exists in createBinaryOperation 2380 assert(result != spv::NoResult); 2381 builder.clearAccessChain(); 2382 builder.setAccessChainRValue(result); 2383 2384 return false; 2385 } 2386 2387 // 2388 // Create the list of operands. 2389 // 2390 glslang::TIntermSequence& glslangOperands = node->getSequence(); 2391 std::vector<spv::Id> operands; 2392 for (int arg = 0; arg < (int)glslangOperands.size(); ++arg) { 2393 // special case l-value operands; there are just a few 2394 bool lvalue = false; 2395 switch (node->getOp()) { 2396 case glslang::EOpFrexp: 2397 case glslang::EOpModf: 2398 if (arg == 1) 2399 lvalue = true; 2400 break; 2401 case glslang::EOpInterpolateAtSample: 2402 case glslang::EOpInterpolateAtOffset: 2403 #ifdef AMD_EXTENSIONS 2404 case glslang::EOpInterpolateAtVertex: 2405 #endif 2406 if (arg == 0) { 2407 lvalue = true; 2408 2409 // Does it need a swizzle inversion? If so, evaluation is inverted; 2410 // operate first on the swizzle base, then apply the swizzle. 2411 if (glslangOperands[0]->getAsOperator() && 2412 glslangOperands[0]->getAsOperator()->getOp() == glslang::EOpVectorSwizzle) 2413 invertedType = convertGlslangToSpvType(glslangOperands[0]->getAsBinaryNode()->getLeft()->getType()); 2414 } 2415 break; 2416 case glslang::EOpAtomicAdd: 2417 case glslang::EOpAtomicMin: 2418 case glslang::EOpAtomicMax: 2419 case glslang::EOpAtomicAnd: 2420 case glslang::EOpAtomicOr: 2421 case glslang::EOpAtomicXor: 2422 case glslang::EOpAtomicExchange: 2423 case glslang::EOpAtomicCompSwap: 2424 case glslang::EOpAtomicLoad: 2425 case glslang::EOpAtomicStore: 2426 case glslang::EOpAtomicCounterAdd: 2427 case glslang::EOpAtomicCounterSubtract: 2428 case glslang::EOpAtomicCounterMin: 2429 case glslang::EOpAtomicCounterMax: 2430 case glslang::EOpAtomicCounterAnd: 2431 case glslang::EOpAtomicCounterOr: 2432 case glslang::EOpAtomicCounterXor: 2433 case glslang::EOpAtomicCounterExchange: 2434 case glslang::EOpAtomicCounterCompSwap: 2435 if (arg == 0) 2436 lvalue = true; 2437 break; 2438 case glslang::EOpAddCarry: 2439 case glslang::EOpSubBorrow: 2440 if (arg == 2) 2441 lvalue = true; 2442 break; 2443 case glslang::EOpUMulExtended: 2444 case glslang::EOpIMulExtended: 2445 if (arg >= 2) 2446 lvalue = true; 2447 break; 2448 default: 2449 break; 2450 } 2451 builder.clearAccessChain(); 2452 if (invertedType != spv::NoType && arg == 0) 2453 glslangOperands[0]->getAsBinaryNode()->getLeft()->traverse(this); 2454 else 2455 glslangOperands[arg]->traverse(this); 2456 if (lvalue) 2457 operands.push_back(builder.accessChainGetLValue()); 2458 else { 2459 builder.setLine(node->getLoc().line, node->getLoc().getFilename()); 2460 operands.push_back(accessChainLoad(glslangOperands[arg]->getAsTyped()->getType())); 2461 } 2462 } 2463 2464 builder.setLine(node->getLoc().line, node->getLoc().getFilename()); 2465 if (atomic) { 2466 // Handle all atomics 2467 result = createAtomicOperation(node->getOp(), precision, resultType(), operands, node->getBasicType()); 2468 } else { 2469 // Pass through to generic operations. 2470 switch (glslangOperands.size()) { 2471 case 0: 2472 result = createNoArgOperation(node->getOp(), precision, resultType()); 2473 break; 2474 case 1: 2475 { 2476 OpDecorations decorations = { precision, 2477 TranslateNoContractionDecoration(node->getType().getQualifier()), 2478 TranslateNonUniformDecoration(node->getType().getQualifier()) }; 2479 result = createUnaryOperation( 2480 node->getOp(), decorations, 2481 resultType(), operands.front(), 2482 glslangOperands[0]->getAsTyped()->getBasicType()); 2483 } 2484 break; 2485 default: 2486 result = createMiscOperation(node->getOp(), precision, resultType(), operands, node->getBasicType()); 2487 break; 2488 } 2489 if (invertedType) 2490 result = createInvertedSwizzle(precision, *glslangOperands[0]->getAsBinaryNode(), result); 2491 } 2492 2493 if (noReturnValue) 2494 return false; 2495 2496 if (! result) { 2497 logger->missingFunctionality("unknown glslang aggregate"); 2498 return true; // pick up a child as a placeholder operand 2499 } else { 2500 builder.clearAccessChain(); 2501 builder.setAccessChainRValue(result); 2502 return false; 2503 } 2504 } 2505 2506 // This path handles both if-then-else and ?: 2507 // The if-then-else has a node type of void, while 2508 // ?: has either a void or a non-void node type 2509 // 2510 // Leaving the result, when not void: 2511 // GLSL only has r-values as the result of a :?, but 2512 // if we have an l-value, that can be more efficient if it will 2513 // become the base of a complex r-value expression, because the 2514 // next layer copies r-values into memory to use the access-chain mechanism 2515 bool TGlslangToSpvTraverser::visitSelection(glslang::TVisit /* visit */, glslang::TIntermSelection* node) 2516 { 2517 // See if it simple and safe, or required, to execute both sides. 2518 // Crucially, side effects must be either semantically required or avoided, 2519 // and there are performance trade-offs. 2520 // Return true if required or a good idea (and safe) to execute both sides, 2521 // false otherwise. 2522 const auto bothSidesPolicy = [&]() -> bool { 2523 // do we have both sides? 2524 if (node->getTrueBlock() == nullptr || 2525 node->getFalseBlock() == nullptr) 2526 return false; 2527 2528 // required? (unless we write additional code to look for side effects 2529 // and make performance trade-offs if none are present) 2530 if (!node->getShortCircuit()) 2531 return true; 2532 2533 // if not required to execute both, decide based on performance/practicality... 2534 2535 // see if OpSelect can handle it 2536 if ((!node->getType().isScalar() && !node->getType().isVector()) || 2537 node->getBasicType() == glslang::EbtVoid) 2538 return false; 2539 2540 assert(node->getType() == node->getTrueBlock() ->getAsTyped()->getType() && 2541 node->getType() == node->getFalseBlock()->getAsTyped()->getType()); 2542 2543 // return true if a single operand to ? : is okay for OpSelect 2544 const auto operandOkay = [](glslang::TIntermTyped* node) { 2545 return node->getAsSymbolNode() || node->getType().getQualifier().isConstant(); 2546 }; 2547 2548 return operandOkay(node->getTrueBlock() ->getAsTyped()) && 2549 operandOkay(node->getFalseBlock()->getAsTyped()); 2550 }; 2551 2552 spv::Id result = spv::NoResult; // upcoming result selecting between trueValue and falseValue 2553 // emit the condition before doing anything with selection 2554 node->getCondition()->traverse(this); 2555 spv::Id condition = accessChainLoad(node->getCondition()->getType()); 2556 2557 // Find a way of executing both sides and selecting the right result. 2558 const auto executeBothSides = [&]() -> void { 2559 // execute both sides 2560 node->getTrueBlock()->traverse(this); 2561 spv::Id trueValue = accessChainLoad(node->getTrueBlock()->getAsTyped()->getType()); 2562 node->getFalseBlock()->traverse(this); 2563 spv::Id falseValue = accessChainLoad(node->getTrueBlock()->getAsTyped()->getType()); 2564 2565 builder.setLine(node->getLoc().line, node->getLoc().getFilename()); 2566 2567 // done if void 2568 if (node->getBasicType() == glslang::EbtVoid) 2569 return; 2570 2571 // emit code to select between trueValue and falseValue 2572 2573 // see if OpSelect can handle it 2574 if (node->getType().isScalar() || node->getType().isVector()) { 2575 // Emit OpSelect for this selection. 2576 2577 // smear condition to vector, if necessary (AST is always scalar) 2578 if (builder.isVector(trueValue)) 2579 condition = builder.smearScalar(spv::NoPrecision, condition, 2580 builder.makeVectorType(builder.makeBoolType(), 2581 builder.getNumComponents(trueValue))); 2582 2583 // OpSelect 2584 result = builder.createTriOp(spv::OpSelect, 2585 convertGlslangToSpvType(node->getType()), condition, 2586 trueValue, falseValue); 2587 2588 builder.clearAccessChain(); 2589 builder.setAccessChainRValue(result); 2590 } else { 2591 // We need control flow to select the result. 2592 // TODO: Once SPIR-V OpSelect allows arbitrary types, eliminate this path. 2593 result = builder.createVariable(spv::StorageClassFunction, convertGlslangToSpvType(node->getType())); 2594 2595 // Selection control: 2596 const spv::SelectionControlMask control = TranslateSelectionControl(*node); 2597 2598 // make an "if" based on the value created by the condition 2599 spv::Builder::If ifBuilder(condition, control, builder); 2600 2601 // emit the "then" statement 2602 builder.createStore(trueValue, result); 2603 ifBuilder.makeBeginElse(); 2604 // emit the "else" statement 2605 builder.createStore(falseValue, result); 2606 2607 // finish off the control flow 2608 ifBuilder.makeEndIf(); 2609 2610 builder.clearAccessChain(); 2611 builder.setAccessChainLValue(result); 2612 } 2613 }; 2614 2615 // Execute the one side needed, as per the condition 2616 const auto executeOneSide = [&]() { 2617 // Always emit control flow. 2618 if (node->getBasicType() != glslang::EbtVoid) 2619 result = builder.createVariable(spv::StorageClassFunction, convertGlslangToSpvType(node->getType())); 2620 2621 // Selection control: 2622 const spv::SelectionControlMask control = TranslateSelectionControl(*node); 2623 2624 // make an "if" based on the value created by the condition 2625 spv::Builder::If ifBuilder(condition, control, builder); 2626 2627 // emit the "then" statement 2628 if (node->getTrueBlock() != nullptr) { 2629 node->getTrueBlock()->traverse(this); 2630 if (result != spv::NoResult) 2631 builder.createStore(accessChainLoad(node->getTrueBlock()->getAsTyped()->getType()), result); 2632 } 2633 2634 if (node->getFalseBlock() != nullptr) { 2635 ifBuilder.makeBeginElse(); 2636 // emit the "else" statement 2637 node->getFalseBlock()->traverse(this); 2638 if (result != spv::NoResult) 2639 builder.createStore(accessChainLoad(node->getFalseBlock()->getAsTyped()->getType()), result); 2640 } 2641 2642 // finish off the control flow 2643 ifBuilder.makeEndIf(); 2644 2645 if (result != spv::NoResult) { 2646 builder.clearAccessChain(); 2647 builder.setAccessChainLValue(result); 2648 } 2649 }; 2650 2651 // Try for OpSelect (or a requirement to execute both sides) 2652 if (bothSidesPolicy()) { 2653 SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder); 2654 if (node->getType().getQualifier().isSpecConstant()) 2655 spec_constant_op_mode_setter.turnOnSpecConstantOpMode(); 2656 executeBothSides(); 2657 } else 2658 executeOneSide(); 2659 2660 return false; 2661 } 2662 2663 bool TGlslangToSpvTraverser::visitSwitch(glslang::TVisit /* visit */, glslang::TIntermSwitch* node) 2664 { 2665 // emit and get the condition before doing anything with switch 2666 node->getCondition()->traverse(this); 2667 spv::Id selector = accessChainLoad(node->getCondition()->getAsTyped()->getType()); 2668 2669 // Selection control: 2670 const spv::SelectionControlMask control = TranslateSwitchControl(*node); 2671 2672 // browse the children to sort out code segments 2673 int defaultSegment = -1; 2674 std::vector<TIntermNode*> codeSegments; 2675 glslang::TIntermSequence& sequence = node->getBody()->getSequence(); 2676 std::vector<int> caseValues; 2677 std::vector<int> valueIndexToSegment(sequence.size()); // note: probably not all are used, it is an overestimate 2678 for (glslang::TIntermSequence::iterator c = sequence.begin(); c != sequence.end(); ++c) { 2679 TIntermNode* child = *c; 2680 if (child->getAsBranchNode() && child->getAsBranchNode()->getFlowOp() == glslang::EOpDefault) 2681 defaultSegment = (int)codeSegments.size(); 2682 else if (child->getAsBranchNode() && child->getAsBranchNode()->getFlowOp() == glslang::EOpCase) { 2683 valueIndexToSegment[caseValues.size()] = (int)codeSegments.size(); 2684 caseValues.push_back(child->getAsBranchNode()->getExpression()->getAsConstantUnion()->getConstArray()[0].getIConst()); 2685 } else 2686 codeSegments.push_back(child); 2687 } 2688 2689 // handle the case where the last code segment is missing, due to no code 2690 // statements between the last case and the end of the switch statement 2691 if ((caseValues.size() && (int)codeSegments.size() == valueIndexToSegment[caseValues.size() - 1]) || 2692 (int)codeSegments.size() == defaultSegment) 2693 codeSegments.push_back(nullptr); 2694 2695 // make the switch statement 2696 std::vector<spv::Block*> segmentBlocks; // returned, as the blocks allocated in the call 2697 builder.makeSwitch(selector, control, (int)codeSegments.size(), caseValues, valueIndexToSegment, defaultSegment, segmentBlocks); 2698 2699 // emit all the code in the segments 2700 breakForLoop.push(false); 2701 for (unsigned int s = 0; s < codeSegments.size(); ++s) { 2702 builder.nextSwitchSegment(segmentBlocks, s); 2703 if (codeSegments[s]) 2704 codeSegments[s]->traverse(this); 2705 else 2706 builder.addSwitchBreak(); 2707 } 2708 breakForLoop.pop(); 2709 2710 builder.endSwitch(segmentBlocks); 2711 2712 return false; 2713 } 2714 2715 void TGlslangToSpvTraverser::visitConstantUnion(glslang::TIntermConstantUnion* node) 2716 { 2717 int nextConst = 0; 2718 spv::Id constant = createSpvConstantFromConstUnionArray(node->getType(), node->getConstArray(), nextConst, false); 2719 2720 builder.clearAccessChain(); 2721 builder.setAccessChainRValue(constant); 2722 } 2723 2724 bool TGlslangToSpvTraverser::visitLoop(glslang::TVisit /* visit */, glslang::TIntermLoop* node) 2725 { 2726 auto blocks = builder.makeNewLoop(); 2727 builder.createBranch(&blocks.head); 2728 2729 // Loop control: 2730 unsigned int dependencyLength = glslang::TIntermLoop::dependencyInfinite; 2731 const spv::LoopControlMask control = TranslateLoopControl(*node, dependencyLength); 2732 2733 // Spec requires back edges to target header blocks, and every header block 2734 // must dominate its merge block. Make a header block first to ensure these 2735 // conditions are met. By definition, it will contain OpLoopMerge, followed 2736 // by a block-ending branch. But we don't want to put any other body/test 2737 // instructions in it, since the body/test may have arbitrary instructions, 2738 // including merges of its own. 2739 builder.setLine(node->getLoc().line, node->getLoc().getFilename()); 2740 builder.setBuildPoint(&blocks.head); 2741 builder.createLoopMerge(&blocks.merge, &blocks.continue_target, control, dependencyLength); 2742 if (node->testFirst() && node->getTest()) { 2743 spv::Block& test = builder.makeNewBlock(); 2744 builder.createBranch(&test); 2745 2746 builder.setBuildPoint(&test); 2747 node->getTest()->traverse(this); 2748 spv::Id condition = accessChainLoad(node->getTest()->getType()); 2749 builder.createConditionalBranch(condition, &blocks.body, &blocks.merge); 2750 2751 builder.setBuildPoint(&blocks.body); 2752 breakForLoop.push(true); 2753 if (node->getBody()) 2754 node->getBody()->traverse(this); 2755 builder.createBranch(&blocks.continue_target); 2756 breakForLoop.pop(); 2757 2758 builder.setBuildPoint(&blocks.continue_target); 2759 if (node->getTerminal()) 2760 node->getTerminal()->traverse(this); 2761 builder.createBranch(&blocks.head); 2762 } else { 2763 builder.setLine(node->getLoc().line, node->getLoc().getFilename()); 2764 builder.createBranch(&blocks.body); 2765 2766 breakForLoop.push(true); 2767 builder.setBuildPoint(&blocks.body); 2768 if (node->getBody()) 2769 node->getBody()->traverse(this); 2770 builder.createBranch(&blocks.continue_target); 2771 breakForLoop.pop(); 2772 2773 builder.setBuildPoint(&blocks.continue_target); 2774 if (node->getTerminal()) 2775 node->getTerminal()->traverse(this); 2776 if (node->getTest()) { 2777 node->getTest()->traverse(this); 2778 spv::Id condition = 2779 accessChainLoad(node->getTest()->getType()); 2780 builder.createConditionalBranch(condition, &blocks.head, &blocks.merge); 2781 } else { 2782 // TODO: unless there was a break/return/discard instruction 2783 // somewhere in the body, this is an infinite loop, so we should 2784 // issue a warning. 2785 builder.createBranch(&blocks.head); 2786 } 2787 } 2788 builder.setBuildPoint(&blocks.merge); 2789 builder.closeLoop(); 2790 return false; 2791 } 2792 2793 bool TGlslangToSpvTraverser::visitBranch(glslang::TVisit /* visit */, glslang::TIntermBranch* node) 2794 { 2795 if (node->getExpression()) 2796 node->getExpression()->traverse(this); 2797 2798 builder.setLine(node->getLoc().line, node->getLoc().getFilename()); 2799 2800 switch (node->getFlowOp()) { 2801 case glslang::EOpKill: 2802 builder.makeDiscard(); 2803 break; 2804 case glslang::EOpBreak: 2805 if (breakForLoop.top()) 2806 builder.createLoopExit(); 2807 else 2808 builder.addSwitchBreak(); 2809 break; 2810 case glslang::EOpContinue: 2811 builder.createLoopContinue(); 2812 break; 2813 case glslang::EOpReturn: 2814 if (node->getExpression()) { 2815 const glslang::TType& glslangReturnType = node->getExpression()->getType(); 2816 spv::Id returnId = accessChainLoad(glslangReturnType); 2817 if (builder.getTypeId(returnId) != currentFunction->getReturnType()) { 2818 builder.clearAccessChain(); 2819 spv::Id copyId = builder.createVariable(spv::StorageClassFunction, currentFunction->getReturnType()); 2820 builder.setAccessChainLValue(copyId); 2821 multiTypeStore(glslangReturnType, returnId); 2822 returnId = builder.createLoad(copyId); 2823 } 2824 builder.makeReturn(false, returnId); 2825 } else 2826 builder.makeReturn(false); 2827 2828 builder.clearAccessChain(); 2829 break; 2830 2831 default: 2832 assert(0); 2833 break; 2834 } 2835 2836 return false; 2837 } 2838 2839 spv::Id TGlslangToSpvTraverser::createSpvVariable(const glslang::TIntermSymbol* node) 2840 { 2841 // First, steer off constants, which are not SPIR-V variables, but 2842 // can still have a mapping to a SPIR-V Id. 2843 // This includes specialization constants. 2844 if (node->getQualifier().isConstant()) { 2845 spv::Id result = createSpvConstant(*node); 2846 if (result != spv::NoResult) 2847 return result; 2848 } 2849 2850 // Now, handle actual variables 2851 spv::StorageClass storageClass = TranslateStorageClass(node->getType()); 2852 spv::Id spvType = convertGlslangToSpvType(node->getType()); 2853 2854 const bool contains16BitType = node->getType().containsBasicType(glslang::EbtFloat16) || 2855 node->getType().containsBasicType(glslang::EbtInt16) || 2856 node->getType().containsBasicType(glslang::EbtUint16); 2857 if (contains16BitType) { 2858 switch (storageClass) { 2859 case spv::StorageClassInput: 2860 case spv::StorageClassOutput: 2861 addPre13Extension(spv::E_SPV_KHR_16bit_storage); 2862 builder.addCapability(spv::CapabilityStorageInputOutput16); 2863 break; 2864 case spv::StorageClassPushConstant: 2865 addPre13Extension(spv::E_SPV_KHR_16bit_storage); 2866 builder.addCapability(spv::CapabilityStoragePushConstant16); 2867 break; 2868 case spv::StorageClassUniform: 2869 addPre13Extension(spv::E_SPV_KHR_16bit_storage); 2870 if (node->getType().getQualifier().storage == glslang::EvqBuffer) 2871 builder.addCapability(spv::CapabilityStorageUniformBufferBlock16); 2872 else 2873 builder.addCapability(spv::CapabilityStorageUniform16); 2874 break; 2875 case spv::StorageClassStorageBuffer: 2876 case spv::StorageClassPhysicalStorageBufferEXT: 2877 addPre13Extension(spv::E_SPV_KHR_16bit_storage); 2878 builder.addCapability(spv::CapabilityStorageUniformBufferBlock16); 2879 break; 2880 default: 2881 break; 2882 } 2883 } 2884 2885 const bool contains8BitType = node->getType().containsBasicType(glslang::EbtInt8) || 2886 node->getType().containsBasicType(glslang::EbtUint8); 2887 if (contains8BitType) { 2888 if (storageClass == spv::StorageClassPushConstant) { 2889 builder.addExtension(spv::E_SPV_KHR_8bit_storage); 2890 builder.addCapability(spv::CapabilityStoragePushConstant8); 2891 } else if (storageClass == spv::StorageClassUniform) { 2892 builder.addExtension(spv::E_SPV_KHR_8bit_storage); 2893 builder.addCapability(spv::CapabilityUniformAndStorageBuffer8BitAccess); 2894 } else if (storageClass == spv::StorageClassStorageBuffer) { 2895 builder.addExtension(spv::E_SPV_KHR_8bit_storage); 2896 builder.addCapability(spv::CapabilityStorageBuffer8BitAccess); 2897 } 2898 } 2899 2900 const char* name = node->getName().c_str(); 2901 if (glslang::IsAnonymous(name)) 2902 name = ""; 2903 2904 return builder.createVariable(storageClass, spvType, name); 2905 } 2906 2907 // Return type Id of the sampled type. 2908 spv::Id TGlslangToSpvTraverser::getSampledType(const glslang::TSampler& sampler) 2909 { 2910 switch (sampler.type) { 2911 case glslang::EbtFloat: return builder.makeFloatType(32); 2912 #ifdef AMD_EXTENSIONS 2913 case glslang::EbtFloat16: 2914 builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float_fetch); 2915 builder.addCapability(spv::CapabilityFloat16ImageAMD); 2916 return builder.makeFloatType(16); 2917 #endif 2918 case glslang::EbtInt: return builder.makeIntType(32); 2919 case glslang::EbtUint: return builder.makeUintType(32); 2920 default: 2921 assert(0); 2922 return builder.makeFloatType(32); 2923 } 2924 } 2925 2926 // If node is a swizzle operation, return the type that should be used if 2927 // the swizzle base is first consumed by another operation, before the swizzle 2928 // is applied. 2929 spv::Id TGlslangToSpvTraverser::getInvertedSwizzleType(const glslang::TIntermTyped& node) 2930 { 2931 if (node.getAsOperator() && 2932 node.getAsOperator()->getOp() == glslang::EOpVectorSwizzle) 2933 return convertGlslangToSpvType(node.getAsBinaryNode()->getLeft()->getType()); 2934 else 2935 return spv::NoType; 2936 } 2937 2938 // When inverting a swizzle with a parent op, this function 2939 // will apply the swizzle operation to a completed parent operation. 2940 spv::Id TGlslangToSpvTraverser::createInvertedSwizzle(spv::Decoration precision, const glslang::TIntermTyped& node, spv::Id parentResult) 2941 { 2942 std::vector<unsigned> swizzle; 2943 convertSwizzle(*node.getAsBinaryNode()->getRight()->getAsAggregate(), swizzle); 2944 return builder.createRvalueSwizzle(precision, convertGlslangToSpvType(node.getType()), parentResult, swizzle); 2945 } 2946 2947 // Convert a glslang AST swizzle node to a swizzle vector for building SPIR-V. 2948 void TGlslangToSpvTraverser::convertSwizzle(const glslang::TIntermAggregate& node, std::vector<unsigned>& swizzle) 2949 { 2950 const glslang::TIntermSequence& swizzleSequence = node.getSequence(); 2951 for (int i = 0; i < (int)swizzleSequence.size(); ++i) 2952 swizzle.push_back(swizzleSequence[i]->getAsConstantUnion()->getConstArray()[0].getIConst()); 2953 } 2954 2955 // Convert from a glslang type to an SPV type, by calling into a 2956 // recursive version of this function. This establishes the inherited 2957 // layout state rooted from the top-level type. 2958 spv::Id TGlslangToSpvTraverser::convertGlslangToSpvType(const glslang::TType& type, bool forwardReferenceOnly) 2959 { 2960 return convertGlslangToSpvType(type, getExplicitLayout(type), type.getQualifier(), false, forwardReferenceOnly); 2961 } 2962 2963 // Do full recursive conversion of an arbitrary glslang type to a SPIR-V Id. 2964 // explicitLayout can be kept the same throughout the hierarchical recursive walk. 2965 // Mutually recursive with convertGlslangStructToSpvType(). 2966 spv::Id TGlslangToSpvTraverser::convertGlslangToSpvType(const glslang::TType& type, 2967 glslang::TLayoutPacking explicitLayout, const glslang::TQualifier& qualifier, 2968 bool lastBufferBlockMember, bool forwardReferenceOnly) 2969 { 2970 spv::Id spvType = spv::NoResult; 2971 2972 switch (type.getBasicType()) { 2973 case glslang::EbtVoid: 2974 spvType = builder.makeVoidType(); 2975 assert (! type.isArray()); 2976 break; 2977 case glslang::EbtFloat: 2978 spvType = builder.makeFloatType(32); 2979 break; 2980 case glslang::EbtDouble: 2981 spvType = builder.makeFloatType(64); 2982 break; 2983 case glslang::EbtFloat16: 2984 spvType = builder.makeFloatType(16); 2985 break; 2986 case glslang::EbtBool: 2987 // "transparent" bool doesn't exist in SPIR-V. The GLSL convention is 2988 // a 32-bit int where non-0 means true. 2989 if (explicitLayout != glslang::ElpNone) 2990 spvType = builder.makeUintType(32); 2991 else 2992 spvType = builder.makeBoolType(); 2993 break; 2994 case glslang::EbtInt8: 2995 spvType = builder.makeIntType(8); 2996 break; 2997 case glslang::EbtUint8: 2998 spvType = builder.makeUintType(8); 2999 break; 3000 case glslang::EbtInt16: 3001 spvType = builder.makeIntType(16); 3002 break; 3003 case glslang::EbtUint16: 3004 spvType = builder.makeUintType(16); 3005 break; 3006 case glslang::EbtInt: 3007 spvType = builder.makeIntType(32); 3008 break; 3009 case glslang::EbtUint: 3010 spvType = builder.makeUintType(32); 3011 break; 3012 case glslang::EbtInt64: 3013 spvType = builder.makeIntType(64); 3014 break; 3015 case glslang::EbtUint64: 3016 spvType = builder.makeUintType(64); 3017 break; 3018 case glslang::EbtAtomicUint: 3019 builder.addCapability(spv::CapabilityAtomicStorage); 3020 spvType = builder.makeUintType(32); 3021 break; 3022 #ifdef NV_EXTENSIONS 3023 case glslang::EbtAccStructNV: 3024 spvType = builder.makeAccelerationStructureNVType(); 3025 break; 3026 #endif 3027 case glslang::EbtSampler: 3028 { 3029 const glslang::TSampler& sampler = type.getSampler(); 3030 if (sampler.sampler) { 3031 // pure sampler 3032 spvType = builder.makeSamplerType(); 3033 } else { 3034 // an image is present, make its type 3035 spvType = builder.makeImageType(getSampledType(sampler), TranslateDimensionality(sampler), sampler.shadow, sampler.arrayed, sampler.ms, 3036 sampler.image ? 2 : 1, TranslateImageFormat(type)); 3037 if (sampler.combined) { 3038 // already has both image and sampler, make the combined type 3039 spvType = builder.makeSampledImageType(spvType); 3040 } 3041 } 3042 } 3043 break; 3044 case glslang::EbtStruct: 3045 case glslang::EbtBlock: 3046 { 3047 // If we've seen this struct type, return it 3048 const glslang::TTypeList* glslangMembers = type.getStruct(); 3049 3050 // Try to share structs for different layouts, but not yet for other 3051 // kinds of qualification (primarily not yet including interpolant qualification). 3052 if (! HasNonLayoutQualifiers(type, qualifier)) 3053 spvType = structMap[explicitLayout][qualifier.layoutMatrix][glslangMembers]; 3054 if (spvType != spv::NoResult) 3055 break; 3056 3057 // else, we haven't seen it... 3058 if (type.getBasicType() == glslang::EbtBlock) 3059 memberRemapper[glslangMembers].resize(glslangMembers->size()); 3060 spvType = convertGlslangStructToSpvType(type, glslangMembers, explicitLayout, qualifier); 3061 } 3062 break; 3063 case glslang::EbtReference: 3064 { 3065 // Make the forward pointer, then recurse to convert the structure type, then 3066 // patch up the forward pointer with a real pointer type. 3067 if (forwardPointers.find(type.getReferentType()) == forwardPointers.end()) { 3068 spv::Id forwardId = builder.makeForwardPointer(spv::StorageClassPhysicalStorageBufferEXT); 3069 forwardPointers[type.getReferentType()] = forwardId; 3070 } 3071 spvType = forwardPointers[type.getReferentType()]; 3072 if (!forwardReferenceOnly) { 3073 spv::Id referentType = convertGlslangToSpvType(*type.getReferentType()); 3074 builder.makePointerFromForwardPointer(spv::StorageClassPhysicalStorageBufferEXT, 3075 forwardPointers[type.getReferentType()], 3076 referentType); 3077 } 3078 } 3079 break; 3080 default: 3081 assert(0); 3082 break; 3083 } 3084 3085 if (type.isMatrix()) 3086 spvType = builder.makeMatrixType(spvType, type.getMatrixCols(), type.getMatrixRows()); 3087 else { 3088 // If this variable has a vector element count greater than 1, create a SPIR-V vector 3089 if (type.getVectorSize() > 1) 3090 spvType = builder.makeVectorType(spvType, type.getVectorSize()); 3091 } 3092 3093 if (type.isArray()) { 3094 int stride = 0; // keep this 0 unless doing an explicit layout; 0 will mean no decoration, no stride 3095 3096 // Do all but the outer dimension 3097 if (type.getArraySizes()->getNumDims() > 1) { 3098 // We need to decorate array strides for types needing explicit layout, except blocks. 3099 if (explicitLayout != glslang::ElpNone && type.getBasicType() != glslang::EbtBlock) { 3100 // Use a dummy glslang type for querying internal strides of 3101 // arrays of arrays, but using just a one-dimensional array. 3102 glslang::TType simpleArrayType(type, 0); // deference type of the array 3103 while (simpleArrayType.getArraySizes()->getNumDims() > 1) 3104 simpleArrayType.getArraySizes()->dereference(); 3105 3106 // Will compute the higher-order strides here, rather than making a whole 3107 // pile of types and doing repetitive recursion on their contents. 3108 stride = getArrayStride(simpleArrayType, explicitLayout, qualifier.layoutMatrix); 3109 } 3110 3111 // make the arrays 3112 for (int dim = type.getArraySizes()->getNumDims() - 1; dim > 0; --dim) { 3113 spvType = builder.makeArrayType(spvType, makeArraySizeId(*type.getArraySizes(), dim), stride); 3114 if (stride > 0) 3115 builder.addDecoration(spvType, spv::DecorationArrayStride, stride); 3116 stride *= type.getArraySizes()->getDimSize(dim); 3117 } 3118 } else { 3119 // single-dimensional array, and don't yet have stride 3120 3121 // We need to decorate array strides for types needing explicit layout, except blocks. 3122 if (explicitLayout != glslang::ElpNone && type.getBasicType() != glslang::EbtBlock) 3123 stride = getArrayStride(type, explicitLayout, qualifier.layoutMatrix); 3124 } 3125 3126 // Do the outer dimension, which might not be known for a runtime-sized array. 3127 // (Unsized arrays that survive through linking will be runtime-sized arrays) 3128 if (type.isSizedArray()) 3129 spvType = builder.makeArrayType(spvType, makeArraySizeId(*type.getArraySizes(), 0), stride); 3130 else { 3131 if (!lastBufferBlockMember) { 3132 builder.addExtension("SPV_EXT_descriptor_indexing"); 3133 builder.addCapability(spv::CapabilityRuntimeDescriptorArrayEXT); 3134 } 3135 spvType = builder.makeRuntimeArray(spvType); 3136 } 3137 if (stride > 0) 3138 builder.addDecoration(spvType, spv::DecorationArrayStride, stride); 3139 } 3140 3141 return spvType; 3142 } 3143 3144 // TODO: this functionality should exist at a higher level, in creating the AST 3145 // 3146 // Identify interface members that don't have their required extension turned on. 3147 // 3148 bool TGlslangToSpvTraverser::filterMember(const glslang::TType& member) 3149 { 3150 #ifdef NV_EXTENSIONS 3151 auto& extensions = glslangIntermediate->getRequestedExtensions(); 3152 3153 if (member.getFieldName() == "gl_SecondaryViewportMaskNV" && 3154 extensions.find("GL_NV_stereo_view_rendering") == extensions.end()) 3155 return true; 3156 if (member.getFieldName() == "gl_SecondaryPositionNV" && 3157 extensions.find("GL_NV_stereo_view_rendering") == extensions.end()) 3158 return true; 3159 3160 if (glslangIntermediate->getStage() != EShLangMeshNV) { 3161 if (member.getFieldName() == "gl_ViewportMask" && 3162 extensions.find("GL_NV_viewport_array2") == extensions.end()) 3163 return true; 3164 if (member.getFieldName() == "gl_PositionPerViewNV" && 3165 extensions.find("GL_NVX_multiview_per_view_attributes") == extensions.end()) 3166 return true; 3167 if (member.getFieldName() == "gl_ViewportMaskPerViewNV" && 3168 extensions.find("GL_NVX_multiview_per_view_attributes") == extensions.end()) 3169 return true; 3170 } 3171 #endif 3172 3173 return false; 3174 }; 3175 3176 // Do full recursive conversion of a glslang structure (or block) type to a SPIR-V Id. 3177 // explicitLayout can be kept the same throughout the hierarchical recursive walk. 3178 // Mutually recursive with convertGlslangToSpvType(). 3179 spv::Id TGlslangToSpvTraverser::convertGlslangStructToSpvType(const glslang::TType& type, 3180 const glslang::TTypeList* glslangMembers, 3181 glslang::TLayoutPacking explicitLayout, 3182 const glslang::TQualifier& qualifier) 3183 { 3184 // Create a vector of struct types for SPIR-V to consume 3185 std::vector<spv::Id> spvMembers; 3186 int memberDelta = 0; // how much the member's index changes from glslang to SPIR-V, normally 0, except sometimes for blocks 3187 std::vector<std::pair<glslang::TType*, glslang::TQualifier> > deferredForwardPointers; 3188 for (int i = 0; i < (int)glslangMembers->size(); i++) { 3189 glslang::TType& glslangMember = *(*glslangMembers)[i].type; 3190 if (glslangMember.hiddenMember()) { 3191 ++memberDelta; 3192 if (type.getBasicType() == glslang::EbtBlock) 3193 memberRemapper[glslangMembers][i] = -1; 3194 } else { 3195 if (type.getBasicType() == glslang::EbtBlock) { 3196 memberRemapper[glslangMembers][i] = i - memberDelta; 3197 if (filterMember(glslangMember)) 3198 continue; 3199 } 3200 // modify just this child's view of the qualifier 3201 glslang::TQualifier memberQualifier = glslangMember.getQualifier(); 3202 InheritQualifiers(memberQualifier, qualifier); 3203 3204 // manually inherit location 3205 if (! memberQualifier.hasLocation() && qualifier.hasLocation()) 3206 memberQualifier.layoutLocation = qualifier.layoutLocation; 3207 3208 // recurse 3209 bool lastBufferBlockMember = qualifier.storage == glslang::EvqBuffer && 3210 i == (int)glslangMembers->size() - 1; 3211 3212 // Make forward pointers for any pointer members, and create a list of members to 3213 // convert to spirv types after creating the struct. 3214 if (glslangMember.getBasicType() == glslang::EbtReference) { 3215 if (forwardPointers.find(glslangMember.getReferentType()) == forwardPointers.end()) { 3216 deferredForwardPointers.push_back(std::make_pair(&glslangMember, memberQualifier)); 3217 } 3218 spvMembers.push_back( 3219 convertGlslangToSpvType(glslangMember, explicitLayout, memberQualifier, lastBufferBlockMember, true)); 3220 } else { 3221 spvMembers.push_back( 3222 convertGlslangToSpvType(glslangMember, explicitLayout, memberQualifier, lastBufferBlockMember, false)); 3223 } 3224 } 3225 } 3226 3227 // Make the SPIR-V type 3228 spv::Id spvType = builder.makeStructType(spvMembers, type.getTypeName().c_str()); 3229 if (! HasNonLayoutQualifiers(type, qualifier)) 3230 structMap[explicitLayout][qualifier.layoutMatrix][glslangMembers] = spvType; 3231 3232 // Decorate it 3233 decorateStructType(type, glslangMembers, explicitLayout, qualifier, spvType); 3234 3235 for (int i = 0; i < deferredForwardPointers.size(); ++i) { 3236 auto it = deferredForwardPointers[i]; 3237 convertGlslangToSpvType(*it.first, explicitLayout, it.second, false); 3238 } 3239 3240 return spvType; 3241 } 3242 3243 void TGlslangToSpvTraverser::decorateStructType(const glslang::TType& type, 3244 const glslang::TTypeList* glslangMembers, 3245 glslang::TLayoutPacking explicitLayout, 3246 const glslang::TQualifier& qualifier, 3247 spv::Id spvType) 3248 { 3249 // Name and decorate the non-hidden members 3250 int offset = -1; 3251 int locationOffset = 0; // for use within the members of this struct 3252 for (int i = 0; i < (int)glslangMembers->size(); i++) { 3253 glslang::TType& glslangMember = *(*glslangMembers)[i].type; 3254 int member = i; 3255 if (type.getBasicType() == glslang::EbtBlock) { 3256 member = memberRemapper[glslangMembers][i]; 3257 if (filterMember(glslangMember)) 3258 continue; 3259 } 3260 3261 // modify just this child's view of the qualifier 3262 glslang::TQualifier memberQualifier = glslangMember.getQualifier(); 3263 InheritQualifiers(memberQualifier, qualifier); 3264 3265 // using -1 above to indicate a hidden member 3266 if (member < 0) 3267 continue; 3268 3269 builder.addMemberName(spvType, member, glslangMember.getFieldName().c_str()); 3270 builder.addMemberDecoration(spvType, member, 3271 TranslateLayoutDecoration(glslangMember, memberQualifier.layoutMatrix)); 3272 builder.addMemberDecoration(spvType, member, TranslatePrecisionDecoration(glslangMember)); 3273 // Add interpolation and auxiliary storage decorations only to 3274 // top-level members of Input and Output storage classes 3275 if (type.getQualifier().storage == glslang::EvqVaryingIn || 3276 type.getQualifier().storage == glslang::EvqVaryingOut) { 3277 if (type.getBasicType() == glslang::EbtBlock || 3278 glslangIntermediate->getSource() == glslang::EShSourceHlsl) { 3279 builder.addMemberDecoration(spvType, member, TranslateInterpolationDecoration(memberQualifier)); 3280 builder.addMemberDecoration(spvType, member, TranslateAuxiliaryStorageDecoration(memberQualifier)); 3281 #ifdef NV_EXTENSIONS 3282 addMeshNVDecoration(spvType, member, memberQualifier); 3283 #endif 3284 } 3285 } 3286 builder.addMemberDecoration(spvType, member, TranslateInvariantDecoration(memberQualifier)); 3287 3288 if (type.getBasicType() == glslang::EbtBlock && 3289 qualifier.storage == glslang::EvqBuffer) { 3290 // Add memory decorations only to top-level members of shader storage block 3291 std::vector<spv::Decoration> memory; 3292 TranslateMemoryDecoration(memberQualifier, memory, glslangIntermediate->usingVulkanMemoryModel()); 3293 for (unsigned int i = 0; i < memory.size(); ++i) 3294 builder.addMemberDecoration(spvType, member, memory[i]); 3295 } 3296 3297 // Location assignment was already completed correctly by the front end, 3298 // just track whether a member needs to be decorated. 3299 // Ignore member locations if the container is an array, as that's 3300 // ill-specified and decisions have been made to not allow this. 3301 if (! type.isArray() && memberQualifier.hasLocation()) 3302 builder.addMemberDecoration(spvType, member, spv::DecorationLocation, memberQualifier.layoutLocation); 3303 3304 if (qualifier.hasLocation()) // track for upcoming inheritance 3305 locationOffset += glslangIntermediate->computeTypeLocationSize( 3306 glslangMember, glslangIntermediate->getStage()); 3307 3308 // component, XFB, others 3309 if (glslangMember.getQualifier().hasComponent()) 3310 builder.addMemberDecoration(spvType, member, spv::DecorationComponent, 3311 glslangMember.getQualifier().layoutComponent); 3312 if (glslangMember.getQualifier().hasXfbOffset()) 3313 builder.addMemberDecoration(spvType, member, spv::DecorationOffset, 3314 glslangMember.getQualifier().layoutXfbOffset); 3315 else if (explicitLayout != glslang::ElpNone) { 3316 // figure out what to do with offset, which is accumulating 3317 int nextOffset; 3318 updateMemberOffset(type, glslangMember, offset, nextOffset, explicitLayout, memberQualifier.layoutMatrix); 3319 if (offset >= 0) 3320 builder.addMemberDecoration(spvType, member, spv::DecorationOffset, offset); 3321 offset = nextOffset; 3322 } 3323 3324 if (glslangMember.isMatrix() && explicitLayout != glslang::ElpNone) 3325 builder.addMemberDecoration(spvType, member, spv::DecorationMatrixStride, 3326 getMatrixStride(glslangMember, explicitLayout, memberQualifier.layoutMatrix)); 3327 3328 // built-in variable decorations 3329 spv::BuiltIn builtIn = TranslateBuiltInDecoration(glslangMember.getQualifier().builtIn, true); 3330 if (builtIn != spv::BuiltInMax) 3331 builder.addMemberDecoration(spvType, member, spv::DecorationBuiltIn, (int)builtIn); 3332 3333 // nonuniform 3334 builder.addMemberDecoration(spvType, member, TranslateNonUniformDecoration(glslangMember.getQualifier())); 3335 3336 if (glslangIntermediate->getHlslFunctionality1() && memberQualifier.semanticName != nullptr) { 3337 builder.addExtension("SPV_GOOGLE_hlsl_functionality1"); 3338 builder.addMemberDecoration(spvType, member, (spv::Decoration)spv::DecorationHlslSemanticGOOGLE, 3339 memberQualifier.semanticName); 3340 } 3341 3342 #ifdef NV_EXTENSIONS 3343 if (builtIn == spv::BuiltInLayer) { 3344 // SPV_NV_viewport_array2 extension 3345 if (glslangMember.getQualifier().layoutViewportRelative){ 3346 builder.addMemberDecoration(spvType, member, (spv::Decoration)spv::DecorationViewportRelativeNV); 3347 builder.addCapability(spv::CapabilityShaderViewportMaskNV); 3348 builder.addExtension(spv::E_SPV_NV_viewport_array2); 3349 } 3350 if (glslangMember.getQualifier().layoutSecondaryViewportRelativeOffset != -2048){ 3351 builder.addMemberDecoration(spvType, member, 3352 (spv::Decoration)spv::DecorationSecondaryViewportRelativeNV, 3353 glslangMember.getQualifier().layoutSecondaryViewportRelativeOffset); 3354 builder.addCapability(spv::CapabilityShaderStereoViewNV); 3355 builder.addExtension(spv::E_SPV_NV_stereo_view_rendering); 3356 } 3357 } 3358 if (glslangMember.getQualifier().layoutPassthrough) { 3359 builder.addMemberDecoration(spvType, member, (spv::Decoration)spv::DecorationPassthroughNV); 3360 builder.addCapability(spv::CapabilityGeometryShaderPassthroughNV); 3361 builder.addExtension(spv::E_SPV_NV_geometry_shader_passthrough); 3362 } 3363 #endif 3364 } 3365 3366 // Decorate the structure 3367 builder.addDecoration(spvType, TranslateLayoutDecoration(type, qualifier.layoutMatrix)); 3368 builder.addDecoration(spvType, TranslateBlockDecoration(type, glslangIntermediate->usingStorageBuffer())); 3369 } 3370 3371 // Turn the expression forming the array size into an id. 3372 // This is not quite trivial, because of specialization constants. 3373 // Sometimes, a raw constant is turned into an Id, and sometimes 3374 // a specialization constant expression is. 3375 spv::Id TGlslangToSpvTraverser::makeArraySizeId(const glslang::TArraySizes& arraySizes, int dim) 3376 { 3377 // First, see if this is sized with a node, meaning a specialization constant: 3378 glslang::TIntermTyped* specNode = arraySizes.getDimNode(dim); 3379 if (specNode != nullptr) { 3380 builder.clearAccessChain(); 3381 specNode->traverse(this); 3382 return accessChainLoad(specNode->getAsTyped()->getType()); 3383 } 3384 3385 // Otherwise, need a compile-time (front end) size, get it: 3386 int size = arraySizes.getDimSize(dim); 3387 assert(size > 0); 3388 return builder.makeUintConstant(size); 3389 } 3390 3391 // Wrap the builder's accessChainLoad to: 3392 // - localize handling of RelaxedPrecision 3393 // - use the SPIR-V inferred type instead of another conversion of the glslang type 3394 // (avoids unnecessary work and possible type punning for structures) 3395 // - do conversion of concrete to abstract type 3396 spv::Id TGlslangToSpvTraverser::accessChainLoad(const glslang::TType& type) 3397 { 3398 spv::Id nominalTypeId = builder.accessChainGetInferredType(); 3399 3400 spv::Builder::AccessChain::CoherentFlags coherentFlags = builder.getAccessChain().coherentFlags; 3401 coherentFlags |= TranslateCoherent(type); 3402 3403 unsigned int alignment = builder.getAccessChain().alignment; 3404 alignment |= getBufferReferenceAlignment(type); 3405 3406 spv::Id loadedId = builder.accessChainLoad(TranslatePrecisionDecoration(type), 3407 TranslateNonUniformDecoration(type.getQualifier()), 3408 nominalTypeId, 3409 spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & ~spv::MemoryAccessMakePointerAvailableKHRMask), 3410 TranslateMemoryScope(coherentFlags), 3411 alignment); 3412 3413 // Need to convert to abstract types when necessary 3414 if (type.getBasicType() == glslang::EbtBool) { 3415 if (builder.isScalarType(nominalTypeId)) { 3416 // Conversion for bool 3417 spv::Id boolType = builder.makeBoolType(); 3418 if (nominalTypeId != boolType) 3419 loadedId = builder.createBinOp(spv::OpINotEqual, boolType, loadedId, builder.makeUintConstant(0)); 3420 } else if (builder.isVectorType(nominalTypeId)) { 3421 // Conversion for bvec 3422 int vecSize = builder.getNumTypeComponents(nominalTypeId); 3423 spv::Id bvecType = builder.makeVectorType(builder.makeBoolType(), vecSize); 3424 if (nominalTypeId != bvecType) 3425 loadedId = builder.createBinOp(spv::OpINotEqual, bvecType, loadedId, makeSmearedConstant(builder.makeUintConstant(0), vecSize)); 3426 } 3427 } 3428 3429 return loadedId; 3430 } 3431 3432 // Wrap the builder's accessChainStore to: 3433 // - do conversion of concrete to abstract type 3434 // 3435 // Implicitly uses the existing builder.accessChain as the storage target. 3436 void TGlslangToSpvTraverser::accessChainStore(const glslang::TType& type, spv::Id rvalue) 3437 { 3438 // Need to convert to abstract types when necessary 3439 if (type.getBasicType() == glslang::EbtBool) { 3440 spv::Id nominalTypeId = builder.accessChainGetInferredType(); 3441 3442 if (builder.isScalarType(nominalTypeId)) { 3443 // Conversion for bool 3444 spv::Id boolType = builder.makeBoolType(); 3445 if (nominalTypeId != boolType) { 3446 // keep these outside arguments, for determinant order-of-evaluation 3447 spv::Id one = builder.makeUintConstant(1); 3448 spv::Id zero = builder.makeUintConstant(0); 3449 rvalue = builder.createTriOp(spv::OpSelect, nominalTypeId, rvalue, one, zero); 3450 } else if (builder.getTypeId(rvalue) != boolType) 3451 rvalue = builder.createBinOp(spv::OpINotEqual, boolType, rvalue, builder.makeUintConstant(0)); 3452 } else if (builder.isVectorType(nominalTypeId)) { 3453 // Conversion for bvec 3454 int vecSize = builder.getNumTypeComponents(nominalTypeId); 3455 spv::Id bvecType = builder.makeVectorType(builder.makeBoolType(), vecSize); 3456 if (nominalTypeId != bvecType) { 3457 // keep these outside arguments, for determinant order-of-evaluation 3458 spv::Id one = makeSmearedConstant(builder.makeUintConstant(1), vecSize); 3459 spv::Id zero = makeSmearedConstant(builder.makeUintConstant(0), vecSize); 3460 rvalue = builder.createTriOp(spv::OpSelect, nominalTypeId, rvalue, one, zero); 3461 } else if (builder.getTypeId(rvalue) != bvecType) 3462 rvalue = builder.createBinOp(spv::OpINotEqual, bvecType, rvalue, 3463 makeSmearedConstant(builder.makeUintConstant(0), vecSize)); 3464 } 3465 } 3466 3467 spv::Builder::AccessChain::CoherentFlags coherentFlags = builder.getAccessChain().coherentFlags; 3468 coherentFlags |= TranslateCoherent(type); 3469 3470 unsigned int alignment = builder.getAccessChain().alignment; 3471 alignment |= getBufferReferenceAlignment(type); 3472 3473 builder.accessChainStore(rvalue, 3474 spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & ~spv::MemoryAccessMakePointerVisibleKHRMask), 3475 TranslateMemoryScope(coherentFlags), alignment); 3476 } 3477 3478 // For storing when types match at the glslang level, but not might match at the 3479 // SPIR-V level. 3480 // 3481 // This especially happens when a single glslang type expands to multiple 3482 // SPIR-V types, like a struct that is used in a member-undecorated way as well 3483 // as in a member-decorated way. 3484 // 3485 // NOTE: This function can handle any store request; if it's not special it 3486 // simplifies to a simple OpStore. 3487 // 3488 // Implicitly uses the existing builder.accessChain as the storage target. 3489 void TGlslangToSpvTraverser::multiTypeStore(const glslang::TType& type, spv::Id rValue) 3490 { 3491 // we only do the complex path here if it's an aggregate 3492 if (! type.isStruct() && ! type.isArray()) { 3493 accessChainStore(type, rValue); 3494 return; 3495 } 3496 3497 // and, it has to be a case of type aliasing 3498 spv::Id rType = builder.getTypeId(rValue); 3499 spv::Id lValue = builder.accessChainGetLValue(); 3500 spv::Id lType = builder.getContainedTypeId(builder.getTypeId(lValue)); 3501 if (lType == rType) { 3502 accessChainStore(type, rValue); 3503 return; 3504 } 3505 3506 // Recursively (as needed) copy an aggregate type to a different aggregate type, 3507 // where the two types were the same type in GLSL. This requires member 3508 // by member copy, recursively. 3509 3510 // If an array, copy element by element. 3511 if (type.isArray()) { 3512 glslang::TType glslangElementType(type, 0); 3513 spv::Id elementRType = builder.getContainedTypeId(rType); 3514 for (int index = 0; index < type.getOuterArraySize(); ++index) { 3515 // get the source member 3516 spv::Id elementRValue = builder.createCompositeExtract(rValue, elementRType, index); 3517 3518 // set up the target storage 3519 builder.clearAccessChain(); 3520 builder.setAccessChainLValue(lValue); 3521 builder.accessChainPush(builder.makeIntConstant(index), TranslateCoherent(type), getBufferReferenceAlignment(type)); 3522 3523 // store the member 3524 multiTypeStore(glslangElementType, elementRValue); 3525 } 3526 } else { 3527 assert(type.isStruct()); 3528 3529 // loop over structure members 3530 const glslang::TTypeList& members = *type.getStruct(); 3531 for (int m = 0; m < (int)members.size(); ++m) { 3532 const glslang::TType& glslangMemberType = *members[m].type; 3533 3534 // get the source member 3535 spv::Id memberRType = builder.getContainedTypeId(rType, m); 3536 spv::Id memberRValue = builder.createCompositeExtract(rValue, memberRType, m); 3537 3538 // set up the target storage 3539 builder.clearAccessChain(); 3540 builder.setAccessChainLValue(lValue); 3541 builder.accessChainPush(builder.makeIntConstant(m), TranslateCoherent(type), getBufferReferenceAlignment(type)); 3542 3543 // store the member 3544 multiTypeStore(glslangMemberType, memberRValue); 3545 } 3546 } 3547 } 3548 3549 // Decide whether or not this type should be 3550 // decorated with offsets and strides, and if so 3551 // whether std140 or std430 rules should be applied. 3552 glslang::TLayoutPacking TGlslangToSpvTraverser::getExplicitLayout(const glslang::TType& type) const 3553 { 3554 // has to be a block 3555 if (type.getBasicType() != glslang::EbtBlock) 3556 return glslang::ElpNone; 3557 3558 // has to be a uniform or buffer block or task in/out blocks 3559 if (type.getQualifier().storage != glslang::EvqUniform && 3560 type.getQualifier().storage != glslang::EvqBuffer && 3561 !type.getQualifier().isTaskMemory()) 3562 return glslang::ElpNone; 3563 3564 // return the layout to use 3565 switch (type.getQualifier().layoutPacking) { 3566 case glslang::ElpStd140: 3567 case glslang::ElpStd430: 3568 case glslang::ElpScalar: 3569 return type.getQualifier().layoutPacking; 3570 default: 3571 return glslang::ElpNone; 3572 } 3573 } 3574 3575 // Given an array type, returns the integer stride required for that array 3576 int TGlslangToSpvTraverser::getArrayStride(const glslang::TType& arrayType, glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout) 3577 { 3578 int size; 3579 int stride; 3580 glslangIntermediate->getMemberAlignment(arrayType, size, stride, explicitLayout, matrixLayout == glslang::ElmRowMajor); 3581 3582 return stride; 3583 } 3584 3585 // Given a matrix type, or array (of array) of matrixes type, returns the integer stride required for that matrix 3586 // when used as a member of an interface block 3587 int TGlslangToSpvTraverser::getMatrixStride(const glslang::TType& matrixType, glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout) 3588 { 3589 glslang::TType elementType; 3590 elementType.shallowCopy(matrixType); 3591 elementType.clearArraySizes(); 3592 3593 int size; 3594 int stride; 3595 glslangIntermediate->getMemberAlignment(elementType, size, stride, explicitLayout, matrixLayout == glslang::ElmRowMajor); 3596 3597 return stride; 3598 } 3599 3600 // Given a member type of a struct, realign the current offset for it, and compute 3601 // the next (not yet aligned) offset for the next member, which will get aligned 3602 // on the next call. 3603 // 'currentOffset' should be passed in already initialized, ready to modify, and reflecting 3604 // the migration of data from nextOffset -> currentOffset. It should be -1 on the first call. 3605 // -1 means a non-forced member offset (no decoration needed). 3606 void TGlslangToSpvTraverser::updateMemberOffset(const glslang::TType& structType, const glslang::TType& memberType, int& currentOffset, int& nextOffset, 3607 glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout) 3608 { 3609 // this will get a positive value when deemed necessary 3610 nextOffset = -1; 3611 3612 // override anything in currentOffset with user-set offset 3613 if (memberType.getQualifier().hasOffset()) 3614 currentOffset = memberType.getQualifier().layoutOffset; 3615 3616 // It could be that current linker usage in glslang updated all the layoutOffset, 3617 // in which case the following code does not matter. But, that's not quite right 3618 // once cross-compilation unit GLSL validation is done, as the original user 3619 // settings are needed in layoutOffset, and then the following will come into play. 3620 3621 if (explicitLayout == glslang::ElpNone) { 3622 if (! memberType.getQualifier().hasOffset()) 3623 currentOffset = -1; 3624 3625 return; 3626 } 3627 3628 // Getting this far means we need explicit offsets 3629 if (currentOffset < 0) 3630 currentOffset = 0; 3631 3632 // Now, currentOffset is valid (either 0, or from a previous nextOffset), 3633 // but possibly not yet correctly aligned. 3634 3635 int memberSize; 3636 int dummyStride; 3637 int memberAlignment = glslangIntermediate->getMemberAlignment(memberType, memberSize, dummyStride, explicitLayout, matrixLayout == glslang::ElmRowMajor); 3638 3639 // Adjust alignment for HLSL rules 3640 // TODO: make this consistent in early phases of code: 3641 // adjusting this late means inconsistencies with earlier code, which for reflection is an issue 3642 // Until reflection is brought in sync with these adjustments, don't apply to $Global, 3643 // which is the most likely to rely on reflection, and least likely to rely implicit layouts 3644 if (glslangIntermediate->usingHlslOffsets() && 3645 ! memberType.isArray() && memberType.isVector() && structType.getTypeName().compare("$Global") != 0) { 3646 int dummySize; 3647 int componentAlignment = glslangIntermediate->getBaseAlignmentScalar(memberType, dummySize); 3648 if (componentAlignment <= 4) 3649 memberAlignment = componentAlignment; 3650 } 3651 3652 // Bump up to member alignment 3653 glslang::RoundToPow2(currentOffset, memberAlignment); 3654 3655 // Bump up to vec4 if there is a bad straddle 3656 if (explicitLayout != glslang::ElpScalar && glslangIntermediate->improperStraddle(memberType, memberSize, currentOffset)) 3657 glslang::RoundToPow2(currentOffset, 16); 3658 3659 nextOffset = currentOffset + memberSize; 3660 } 3661 3662 void TGlslangToSpvTraverser::declareUseOfStructMember(const glslang::TTypeList& members, int glslangMember) 3663 { 3664 const glslang::TBuiltInVariable glslangBuiltIn = members[glslangMember].type->getQualifier().builtIn; 3665 switch (glslangBuiltIn) 3666 { 3667 case glslang::EbvClipDistance: 3668 case glslang::EbvCullDistance: 3669 case glslang::EbvPointSize: 3670 #ifdef NV_EXTENSIONS 3671 case glslang::EbvViewportMaskNV: 3672 case glslang::EbvSecondaryPositionNV: 3673 case glslang::EbvSecondaryViewportMaskNV: 3674 case glslang::EbvPositionPerViewNV: 3675 case glslang::EbvViewportMaskPerViewNV: 3676 case glslang::EbvTaskCountNV: 3677 case glslang::EbvPrimitiveCountNV: 3678 case glslang::EbvPrimitiveIndicesNV: 3679 case glslang::EbvClipDistancePerViewNV: 3680 case glslang::EbvCullDistancePerViewNV: 3681 case glslang::EbvLayerPerViewNV: 3682 case glslang::EbvMeshViewCountNV: 3683 case glslang::EbvMeshViewIndicesNV: 3684 #endif 3685 // Generate the associated capability. Delegate to TranslateBuiltInDecoration. 3686 // Alternately, we could just call this for any glslang built-in, since the 3687 // capability already guards against duplicates. 3688 TranslateBuiltInDecoration(glslangBuiltIn, false); 3689 break; 3690 default: 3691 // Capabilities were already generated when the struct was declared. 3692 break; 3693 } 3694 } 3695 3696 bool TGlslangToSpvTraverser::isShaderEntryPoint(const glslang::TIntermAggregate* node) 3697 { 3698 return node->getName().compare(glslangIntermediate->getEntryPointMangledName().c_str()) == 0; 3699 } 3700 3701 // Does parameter need a place to keep writes, separate from the original? 3702 // Assumes called after originalParam(), which filters out block/buffer/opaque-based 3703 // qualifiers such that we should have only in/out/inout/constreadonly here. 3704 bool TGlslangToSpvTraverser::writableParam(glslang::TStorageQualifier qualifier) const 3705 { 3706 assert(qualifier == glslang::EvqIn || 3707 qualifier == glslang::EvqOut || 3708 qualifier == glslang::EvqInOut || 3709 qualifier == glslang::EvqConstReadOnly); 3710 return qualifier != glslang::EvqConstReadOnly; 3711 } 3712 3713 // Is parameter pass-by-original? 3714 bool TGlslangToSpvTraverser::originalParam(glslang::TStorageQualifier qualifier, const glslang::TType& paramType, 3715 bool implicitThisParam) 3716 { 3717 if (implicitThisParam) // implicit this 3718 return true; 3719 if (glslangIntermediate->getSource() == glslang::EShSourceHlsl) 3720 return paramType.getBasicType() == glslang::EbtBlock; 3721 return paramType.containsOpaque() || // sampler, etc. 3722 (paramType.getBasicType() == glslang::EbtBlock && qualifier == glslang::EvqBuffer); // SSBO 3723 } 3724 3725 // Make all the functions, skeletally, without actually visiting their bodies. 3726 void TGlslangToSpvTraverser::makeFunctions(const glslang::TIntermSequence& glslFunctions) 3727 { 3728 const auto getParamDecorations = [&](std::vector<spv::Decoration>& decorations, const glslang::TType& type, bool useVulkanMemoryModel) { 3729 spv::Decoration paramPrecision = TranslatePrecisionDecoration(type); 3730 if (paramPrecision != spv::NoPrecision) 3731 decorations.push_back(paramPrecision); 3732 TranslateMemoryDecoration(type.getQualifier(), decorations, useVulkanMemoryModel); 3733 if (type.getBasicType() == glslang::EbtReference) { 3734 // Original and non-writable params pass the pointer directly and 3735 // use restrict/aliased, others are stored to a pointer in Function 3736 // memory and use RestrictPointer/AliasedPointer. 3737 if (originalParam(type.getQualifier().storage, type, false) || 3738 !writableParam(type.getQualifier().storage)) { 3739 decorations.push_back(type.getQualifier().restrict ? spv::DecorationRestrict : spv::DecorationAliased); 3740 } else { 3741 decorations.push_back(type.getQualifier().restrict ? spv::DecorationRestrictPointerEXT : spv::DecorationAliasedPointerEXT); 3742 } 3743 } 3744 }; 3745 3746 for (int f = 0; f < (int)glslFunctions.size(); ++f) { 3747 glslang::TIntermAggregate* glslFunction = glslFunctions[f]->getAsAggregate(); 3748 if (! glslFunction || glslFunction->getOp() != glslang::EOpFunction || isShaderEntryPoint(glslFunction)) 3749 continue; 3750 3751 // We're on a user function. Set up the basic interface for the function now, 3752 // so that it's available to call. Translating the body will happen later. 3753 // 3754 // Typically (except for a "const in" parameter), an address will be passed to the 3755 // function. What it is an address of varies: 3756 // 3757 // - "in" parameters not marked as "const" can be written to without modifying the calling 3758 // argument so that write needs to be to a copy, hence the address of a copy works. 3759 // 3760 // - "const in" parameters can just be the r-value, as no writes need occur. 3761 // 3762 // - "out" and "inout" arguments can't be done as pointers to the calling argument, because 3763 // GLSL has copy-in/copy-out semantics. They can be handled though with a pointer to a copy. 3764 3765 std::vector<spv::Id> paramTypes; 3766 std::vector<std::vector<spv::Decoration>> paramDecorations; // list of decorations per parameter 3767 glslang::TIntermSequence& parameters = glslFunction->getSequence()[0]->getAsAggregate()->getSequence(); 3768 3769 bool implicitThis = (int)parameters.size() > 0 && parameters[0]->getAsSymbolNode()->getName() == 3770 glslangIntermediate->implicitThisName; 3771 3772 paramDecorations.resize(parameters.size()); 3773 for (int p = 0; p < (int)parameters.size(); ++p) { 3774 const glslang::TType& paramType = parameters[p]->getAsTyped()->getType(); 3775 spv::Id typeId = convertGlslangToSpvType(paramType); 3776 if (originalParam(paramType.getQualifier().storage, paramType, implicitThis && p == 0)) 3777 typeId = builder.makePointer(TranslateStorageClass(paramType), typeId); 3778 else if (writableParam(paramType.getQualifier().storage)) 3779 typeId = builder.makePointer(spv::StorageClassFunction, typeId); 3780 else 3781 rValueParameters.insert(parameters[p]->getAsSymbolNode()->getId()); 3782 getParamDecorations(paramDecorations[p], paramType, glslangIntermediate->usingVulkanMemoryModel()); 3783 paramTypes.push_back(typeId); 3784 } 3785 3786 spv::Block* functionBlock; 3787 spv::Function *function = builder.makeFunctionEntry(TranslatePrecisionDecoration(glslFunction->getType()), 3788 convertGlslangToSpvType(glslFunction->getType()), 3789 glslFunction->getName().c_str(), paramTypes, 3790 paramDecorations, &functionBlock); 3791 if (implicitThis) 3792 function->setImplicitThis(); 3793 3794 // Track function to emit/call later 3795 functionMap[glslFunction->getName().c_str()] = function; 3796 3797 // Set the parameter id's 3798 for (int p = 0; p < (int)parameters.size(); ++p) { 3799 symbolValues[parameters[p]->getAsSymbolNode()->getId()] = function->getParamId(p); 3800 // give a name too 3801 builder.addName(function->getParamId(p), parameters[p]->getAsSymbolNode()->getName().c_str()); 3802 } 3803 } 3804 } 3805 3806 // Process all the initializers, while skipping the functions and link objects 3807 void TGlslangToSpvTraverser::makeGlobalInitializers(const glslang::TIntermSequence& initializers) 3808 { 3809 builder.setBuildPoint(shaderEntry->getLastBlock()); 3810 for (int i = 0; i < (int)initializers.size(); ++i) { 3811 glslang::TIntermAggregate* initializer = initializers[i]->getAsAggregate(); 3812 if (initializer && initializer->getOp() != glslang::EOpFunction && initializer->getOp() != glslang::EOpLinkerObjects) { 3813 3814 // We're on a top-level node that's not a function. Treat as an initializer, whose 3815 // code goes into the beginning of the entry point. 3816 initializer->traverse(this); 3817 } 3818 } 3819 } 3820 3821 // Process all the functions, while skipping initializers. 3822 void TGlslangToSpvTraverser::visitFunctions(const glslang::TIntermSequence& glslFunctions) 3823 { 3824 for (int f = 0; f < (int)glslFunctions.size(); ++f) { 3825 glslang::TIntermAggregate* node = glslFunctions[f]->getAsAggregate(); 3826 if (node && (node->getOp() == glslang::EOpFunction || node->getOp() == glslang::EOpLinkerObjects)) 3827 node->traverse(this); 3828 } 3829 } 3830 3831 void TGlslangToSpvTraverser::handleFunctionEntry(const glslang::TIntermAggregate* node) 3832 { 3833 // SPIR-V functions should already be in the functionMap from the prepass 3834 // that called makeFunctions(). 3835 currentFunction = functionMap[node->getName().c_str()]; 3836 spv::Block* functionBlock = currentFunction->getEntryBlock(); 3837 builder.setBuildPoint(functionBlock); 3838 } 3839 3840 void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate& node, std::vector<spv::Id>& arguments) 3841 { 3842 const glslang::TIntermSequence& glslangArguments = node.getSequence(); 3843 3844 glslang::TSampler sampler = {}; 3845 bool cubeCompare = false; 3846 #ifdef AMD_EXTENSIONS 3847 bool f16ShadowCompare = false; 3848 #endif 3849 if (node.isTexture() || node.isImage()) { 3850 sampler = glslangArguments[0]->getAsTyped()->getType().getSampler(); 3851 cubeCompare = sampler.dim == glslang::EsdCube && sampler.arrayed && sampler.shadow; 3852 #ifdef AMD_EXTENSIONS 3853 f16ShadowCompare = sampler.shadow && glslangArguments[1]->getAsTyped()->getType().getBasicType() == glslang::EbtFloat16; 3854 #endif 3855 } 3856 3857 for (int i = 0; i < (int)glslangArguments.size(); ++i) { 3858 builder.clearAccessChain(); 3859 glslangArguments[i]->traverse(this); 3860 3861 // Special case l-value operands 3862 bool lvalue = false; 3863 switch (node.getOp()) { 3864 case glslang::EOpImageAtomicAdd: 3865 case glslang::EOpImageAtomicMin: 3866 case glslang::EOpImageAtomicMax: 3867 case glslang::EOpImageAtomicAnd: 3868 case glslang::EOpImageAtomicOr: 3869 case glslang::EOpImageAtomicXor: 3870 case glslang::EOpImageAtomicExchange: 3871 case glslang::EOpImageAtomicCompSwap: 3872 case glslang::EOpImageAtomicLoad: 3873 case glslang::EOpImageAtomicStore: 3874 if (i == 0) 3875 lvalue = true; 3876 break; 3877 case glslang::EOpSparseImageLoad: 3878 if ((sampler.ms && i == 3) || (! sampler.ms && i == 2)) 3879 lvalue = true; 3880 break; 3881 #ifdef AMD_EXTENSIONS 3882 case glslang::EOpSparseTexture: 3883 if (((cubeCompare || f16ShadowCompare) && i == 3) || (! (cubeCompare || f16ShadowCompare) && i == 2)) 3884 lvalue = true; 3885 break; 3886 case glslang::EOpSparseTextureClamp: 3887 if (((cubeCompare || f16ShadowCompare) && i == 4) || (! (cubeCompare || f16ShadowCompare) && i == 3)) 3888 lvalue = true; 3889 break; 3890 case glslang::EOpSparseTextureLod: 3891 case glslang::EOpSparseTextureOffset: 3892 if ((f16ShadowCompare && i == 4) || (! f16ShadowCompare && i == 3)) 3893 lvalue = true; 3894 break; 3895 #else 3896 case glslang::EOpSparseTexture: 3897 if ((cubeCompare && i == 3) || (! cubeCompare && i == 2)) 3898 lvalue = true; 3899 break; 3900 case glslang::EOpSparseTextureClamp: 3901 if ((cubeCompare && i == 4) || (! cubeCompare && i == 3)) 3902 lvalue = true; 3903 break; 3904 case glslang::EOpSparseTextureLod: 3905 case glslang::EOpSparseTextureOffset: 3906 if (i == 3) 3907 lvalue = true; 3908 break; 3909 #endif 3910 case glslang::EOpSparseTextureFetch: 3911 if ((sampler.dim != glslang::EsdRect && i == 3) || (sampler.dim == glslang::EsdRect && i == 2)) 3912 lvalue = true; 3913 break; 3914 case glslang::EOpSparseTextureFetchOffset: 3915 if ((sampler.dim != glslang::EsdRect && i == 4) || (sampler.dim == glslang::EsdRect && i == 3)) 3916 lvalue = true; 3917 break; 3918 #ifdef AMD_EXTENSIONS 3919 case glslang::EOpSparseTextureLodOffset: 3920 case glslang::EOpSparseTextureGrad: 3921 case glslang::EOpSparseTextureOffsetClamp: 3922 if ((f16ShadowCompare && i == 5) || (! f16ShadowCompare && i == 4)) 3923 lvalue = true; 3924 break; 3925 case glslang::EOpSparseTextureGradOffset: 3926 case glslang::EOpSparseTextureGradClamp: 3927 if ((f16ShadowCompare && i == 6) || (! f16ShadowCompare && i == 5)) 3928 lvalue = true; 3929 break; 3930 case glslang::EOpSparseTextureGradOffsetClamp: 3931 if ((f16ShadowCompare && i == 7) || (! f16ShadowCompare && i == 6)) 3932 lvalue = true; 3933 break; 3934 #else 3935 case glslang::EOpSparseTextureLodOffset: 3936 case glslang::EOpSparseTextureGrad: 3937 case glslang::EOpSparseTextureOffsetClamp: 3938 if (i == 4) 3939 lvalue = true; 3940 break; 3941 case glslang::EOpSparseTextureGradOffset: 3942 case glslang::EOpSparseTextureGradClamp: 3943 if (i == 5) 3944 lvalue = true; 3945 break; 3946 case glslang::EOpSparseTextureGradOffsetClamp: 3947 if (i == 6) 3948 lvalue = true; 3949 break; 3950 #endif 3951 case glslang::EOpSparseTextureGather: 3952 if ((sampler.shadow && i == 3) || (! sampler.shadow && i == 2)) 3953 lvalue = true; 3954 break; 3955 case glslang::EOpSparseTextureGatherOffset: 3956 case glslang::EOpSparseTextureGatherOffsets: 3957 if ((sampler.shadow && i == 4) || (! sampler.shadow && i == 3)) 3958 lvalue = true; 3959 break; 3960 #ifdef AMD_EXTENSIONS 3961 case glslang::EOpSparseTextureGatherLod: 3962 if (i == 3) 3963 lvalue = true; 3964 break; 3965 case glslang::EOpSparseTextureGatherLodOffset: 3966 case glslang::EOpSparseTextureGatherLodOffsets: 3967 if (i == 4) 3968 lvalue = true; 3969 break; 3970 case glslang::EOpSparseImageLoadLod: 3971 if (i == 3) 3972 lvalue = true; 3973 break; 3974 #endif 3975 #ifdef NV_EXTENSIONS 3976 case glslang::EOpImageSampleFootprintNV: 3977 if (i == 4) 3978 lvalue = true; 3979 break; 3980 case glslang::EOpImageSampleFootprintClampNV: 3981 case glslang::EOpImageSampleFootprintLodNV: 3982 if (i == 5) 3983 lvalue = true; 3984 break; 3985 case glslang::EOpImageSampleFootprintGradNV: 3986 if (i == 6) 3987 lvalue = true; 3988 break; 3989 case glslang::EOpImageSampleFootprintGradClampNV: 3990 if (i == 7) 3991 lvalue = true; 3992 break; 3993 #endif 3994 default: 3995 break; 3996 } 3997 3998 if (lvalue) 3999 arguments.push_back(builder.accessChainGetLValue()); 4000 else 4001 arguments.push_back(accessChainLoad(glslangArguments[i]->getAsTyped()->getType())); 4002 } 4003 } 4004 4005 void TGlslangToSpvTraverser::translateArguments(glslang::TIntermUnary& node, std::vector<spv::Id>& arguments) 4006 { 4007 builder.clearAccessChain(); 4008 node.getOperand()->traverse(this); 4009 arguments.push_back(accessChainLoad(node.getOperand()->getType())); 4010 } 4011 4012 spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermOperator* node) 4013 { 4014 if (! node->isImage() && ! node->isTexture()) 4015 return spv::NoResult; 4016 4017 builder.setLine(node->getLoc().line, node->getLoc().getFilename()); 4018 4019 // Process a GLSL texturing op (will be SPV image) 4020 4021 const glslang::TType &imageType = node->getAsAggregate() ? node->getAsAggregate()->getSequence()[0]->getAsTyped()->getType() 4022 : node->getAsUnaryNode()->getOperand()->getAsTyped()->getType(); 4023 const glslang::TSampler sampler = imageType.getSampler(); 4024 #ifdef AMD_EXTENSIONS 4025 bool f16ShadowCompare = (sampler.shadow && node->getAsAggregate()) 4026 ? node->getAsAggregate()->getSequence()[1]->getAsTyped()->getType().getBasicType() == glslang::EbtFloat16 4027 : false; 4028 #endif 4029 4030 std::vector<spv::Id> arguments; 4031 if (node->getAsAggregate()) 4032 translateArguments(*node->getAsAggregate(), arguments); 4033 else 4034 translateArguments(*node->getAsUnaryNode(), arguments); 4035 spv::Decoration precision = TranslatePrecisionDecoration(node->getOperationPrecision()); 4036 4037 spv::Builder::TextureParameters params = { }; 4038 params.sampler = arguments[0]; 4039 4040 glslang::TCrackedTextureOp cracked; 4041 node->crackTexture(sampler, cracked); 4042 4043 const bool isUnsignedResult = node->getType().getBasicType() == glslang::EbtUint; 4044 4045 // Check for queries 4046 if (cracked.query) { 4047 // OpImageQueryLod works on a sampled image, for other queries the image has to be extracted first 4048 if (node->getOp() != glslang::EOpTextureQueryLod && builder.isSampledImage(params.sampler)) 4049 params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler); 4050 4051 switch (node->getOp()) { 4052 case glslang::EOpImageQuerySize: 4053 case glslang::EOpTextureQuerySize: 4054 if (arguments.size() > 1) { 4055 params.lod = arguments[1]; 4056 return builder.createTextureQueryCall(spv::OpImageQuerySizeLod, params, isUnsignedResult); 4057 } else 4058 return builder.createTextureQueryCall(spv::OpImageQuerySize, params, isUnsignedResult); 4059 case glslang::EOpImageQuerySamples: 4060 case glslang::EOpTextureQuerySamples: 4061 return builder.createTextureQueryCall(spv::OpImageQuerySamples, params, isUnsignedResult); 4062 case glslang::EOpTextureQueryLod: 4063 params.coords = arguments[1]; 4064 return builder.createTextureQueryCall(spv::OpImageQueryLod, params, isUnsignedResult); 4065 case glslang::EOpTextureQueryLevels: 4066 return builder.createTextureQueryCall(spv::OpImageQueryLevels, params, isUnsignedResult); 4067 case glslang::EOpSparseTexelsResident: 4068 return builder.createUnaryOp(spv::OpImageSparseTexelsResident, builder.makeBoolType(), arguments[0]); 4069 default: 4070 assert(0); 4071 break; 4072 } 4073 } 4074 4075 int components = node->getType().getVectorSize(); 4076 4077 if (node->getOp() == glslang::EOpTextureFetch) { 4078 // These must produce 4 components, per SPIR-V spec. We'll add a conversion constructor if needed. 4079 // This will only happen through the HLSL path for operator[], so we do not have to handle e.g. 4080 // the EOpTexture/Proj/Lod/etc family. It would be harmless to do so, but would need more logic 4081 // here around e.g. which ones return scalars or other types. 4082 components = 4; 4083 } 4084 4085 glslang::TType returnType(node->getType().getBasicType(), glslang::EvqTemporary, components); 4086 4087 auto resultType = [&returnType,this]{ return convertGlslangToSpvType(returnType); }; 4088 4089 // Check for image functions other than queries 4090 if (node->isImage()) { 4091 std::vector<spv::IdImmediate> operands; 4092 auto opIt = arguments.begin(); 4093 spv::IdImmediate image = { true, *(opIt++) }; 4094 operands.push_back(image); 4095 4096 // Handle subpass operations 4097 // TODO: GLSL should change to have the "MS" only on the type rather than the 4098 // built-in function. 4099 if (cracked.subpass) { 4100 // add on the (0,0) coordinate 4101 spv::Id zero = builder.makeIntConstant(0); 4102 std::vector<spv::Id> comps; 4103 comps.push_back(zero); 4104 comps.push_back(zero); 4105 spv::IdImmediate coord = { true, 4106 builder.makeCompositeConstant(builder.makeVectorType(builder.makeIntType(32), 2), comps) }; 4107 operands.push_back(coord); 4108 if (sampler.ms) { 4109 spv::IdImmediate imageOperands = { false, spv::ImageOperandsSampleMask }; 4110 operands.push_back(imageOperands); 4111 spv::IdImmediate imageOperand = { true, *(opIt++) }; 4112 operands.push_back(imageOperand); 4113 } 4114 spv::Id result = builder.createOp(spv::OpImageRead, resultType(), operands); 4115 builder.setPrecision(result, precision); 4116 return result; 4117 } 4118 4119 spv::IdImmediate coord = { true, *(opIt++) }; 4120 operands.push_back(coord); 4121 #ifdef AMD_EXTENSIONS 4122 if (node->getOp() == glslang::EOpImageLoad || node->getOp() == glslang::EOpImageLoadLod) { 4123 #else 4124 if (node->getOp() == glslang::EOpImageLoad) { 4125 #endif 4126 spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone; 4127 if (sampler.ms) { 4128 mask = mask | spv::ImageOperandsSampleMask; 4129 } 4130 #ifdef AMD_EXTENSIONS 4131 if (cracked.lod) { 4132 builder.addExtension(spv::E_SPV_AMD_shader_image_load_store_lod); 4133 builder.addCapability(spv::CapabilityImageReadWriteLodAMD); 4134 mask = mask | spv::ImageOperandsLodMask; 4135 } 4136 #endif 4137 mask = mask | TranslateImageOperands(TranslateCoherent(imageType)); 4138 mask = (spv::ImageOperandsMask)(mask & ~spv::ImageOperandsMakeTexelAvailableKHRMask); 4139 if (mask) { 4140 spv::IdImmediate imageOperands = { false, (unsigned int)mask }; 4141 operands.push_back(imageOperands); 4142 } 4143 if (mask & spv::ImageOperandsSampleMask) { 4144 spv::IdImmediate imageOperand = { true, *opIt++ }; 4145 operands.push_back(imageOperand); 4146 } 4147 #ifdef AMD_EXTENSIONS 4148 if (mask & spv::ImageOperandsLodMask) { 4149 spv::IdImmediate imageOperand = { true, *opIt++ }; 4150 operands.push_back(imageOperand); 4151 } 4152 #endif 4153 if (mask & spv::ImageOperandsMakeTexelVisibleKHRMask) { 4154 spv::IdImmediate imageOperand = { true, builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) }; 4155 operands.push_back(imageOperand); 4156 } 4157 4158 if (builder.getImageTypeFormat(builder.getImageType(operands.front().word)) == spv::ImageFormatUnknown) 4159 builder.addCapability(spv::CapabilityStorageImageReadWithoutFormat); 4160 4161 std::vector<spv::Id> result(1, builder.createOp(spv::OpImageRead, resultType(), operands)); 4162 builder.setPrecision(result[0], precision); 4163 4164 // If needed, add a conversion constructor to the proper size. 4165 if (components != node->getType().getVectorSize()) 4166 result[0] = builder.createConstructor(precision, result, convertGlslangToSpvType(node->getType())); 4167 4168 return result[0]; 4169 #ifdef AMD_EXTENSIONS 4170 } else if (node->getOp() == glslang::EOpImageStore || node->getOp() == glslang::EOpImageStoreLod) { 4171 #else 4172 } else if (node->getOp() == glslang::EOpImageStore) { 4173 #endif 4174 4175 // Push the texel value before the operands 4176 #ifdef AMD_EXTENSIONS 4177 if (sampler.ms || cracked.lod) { 4178 #else 4179 if (sampler.ms) { 4180 #endif 4181 spv::IdImmediate texel = { true, *(opIt + 1) }; 4182 operands.push_back(texel); 4183 } else { 4184 spv::IdImmediate texel = { true, *opIt }; 4185 operands.push_back(texel); 4186 } 4187 4188 spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone; 4189 if (sampler.ms) { 4190 mask = mask | spv::ImageOperandsSampleMask; 4191 } 4192 #ifdef AMD_EXTENSIONS 4193 if (cracked.lod) { 4194 builder.addExtension(spv::E_SPV_AMD_shader_image_load_store_lod); 4195 builder.addCapability(spv::CapabilityImageReadWriteLodAMD); 4196 mask = mask | spv::ImageOperandsLodMask; 4197 } 4198 #endif 4199 mask = mask | TranslateImageOperands(TranslateCoherent(imageType)); 4200 mask = (spv::ImageOperandsMask)(mask & ~spv::ImageOperandsMakeTexelVisibleKHRMask); 4201 if (mask) { 4202 spv::IdImmediate imageOperands = { false, (unsigned int)mask }; 4203 operands.push_back(imageOperands); 4204 } 4205 if (mask & spv::ImageOperandsSampleMask) { 4206 spv::IdImmediate imageOperand = { true, *opIt++ }; 4207 operands.push_back(imageOperand); 4208 } 4209 #ifdef AMD_EXTENSIONS 4210 if (mask & spv::ImageOperandsLodMask) { 4211 spv::IdImmediate imageOperand = { true, *opIt++ }; 4212 operands.push_back(imageOperand); 4213 } 4214 #endif 4215 if (mask & spv::ImageOperandsMakeTexelAvailableKHRMask) { 4216 spv::IdImmediate imageOperand = { true, builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) }; 4217 operands.push_back(imageOperand); 4218 } 4219 4220 builder.createNoResultOp(spv::OpImageWrite, operands); 4221 if (builder.getImageTypeFormat(builder.getImageType(operands.front().word)) == spv::ImageFormatUnknown) 4222 builder.addCapability(spv::CapabilityStorageImageWriteWithoutFormat); 4223 return spv::NoResult; 4224 #ifdef AMD_EXTENSIONS 4225 } else if (node->getOp() == glslang::EOpSparseImageLoad || node->getOp() == glslang::EOpSparseImageLoadLod) { 4226 #else 4227 } else if (node->getOp() == glslang::EOpSparseImageLoad) { 4228 #endif 4229 builder.addCapability(spv::CapabilitySparseResidency); 4230 if (builder.getImageTypeFormat(builder.getImageType(operands.front().word)) == spv::ImageFormatUnknown) 4231 builder.addCapability(spv::CapabilityStorageImageReadWithoutFormat); 4232 4233 spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone; 4234 if (sampler.ms) { 4235 mask = mask | spv::ImageOperandsSampleMask; 4236 } 4237 #ifdef AMD_EXTENSIONS 4238 if (cracked.lod) { 4239 builder.addExtension(spv::E_SPV_AMD_shader_image_load_store_lod); 4240 builder.addCapability(spv::CapabilityImageReadWriteLodAMD); 4241 4242 mask = mask | spv::ImageOperandsLodMask; 4243 } 4244 #endif 4245 mask = mask | TranslateImageOperands(TranslateCoherent(imageType)); 4246 mask = (spv::ImageOperandsMask)(mask & ~spv::ImageOperandsMakeTexelAvailableKHRMask); 4247 if (mask) { 4248 spv::IdImmediate imageOperands = { false, (unsigned int)mask }; 4249 operands.push_back(imageOperands); 4250 } 4251 if (mask & spv::ImageOperandsSampleMask) { 4252 spv::IdImmediate imageOperand = { true, *opIt++ }; 4253 operands.push_back(imageOperand); 4254 } 4255 #ifdef AMD_EXTENSIONS 4256 if (mask & spv::ImageOperandsLodMask) { 4257 spv::IdImmediate imageOperand = { true, *opIt++ }; 4258 operands.push_back(imageOperand); 4259 } 4260 #endif 4261 if (mask & spv::ImageOperandsMakeTexelVisibleKHRMask) { 4262 spv::IdImmediate imageOperand = { true, builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) }; 4263 operands.push_back(imageOperand); 4264 } 4265 4266 // Create the return type that was a special structure 4267 spv::Id texelOut = *opIt; 4268 spv::Id typeId0 = resultType(); 4269 spv::Id typeId1 = builder.getDerefTypeId(texelOut); 4270 spv::Id resultTypeId = builder.makeStructResultType(typeId0, typeId1); 4271 4272 spv::Id resultId = builder.createOp(spv::OpImageSparseRead, resultTypeId, operands); 4273 4274 // Decode the return type 4275 builder.createStore(builder.createCompositeExtract(resultId, typeId1, 1), texelOut); 4276 return builder.createCompositeExtract(resultId, typeId0, 0); 4277 } else { 4278 // Process image atomic operations 4279 4280 // GLSL "IMAGE_PARAMS" will involve in constructing an image texel pointer and this pointer, 4281 // as the first source operand, is required by SPIR-V atomic operations. 4282 // For non-MS, the sample value should be 0 4283 spv::IdImmediate sample = { true, sampler.ms ? *(opIt++) : builder.makeUintConstant(0) }; 4284 operands.push_back(sample); 4285 4286 spv::Id resultTypeId; 4287 // imageAtomicStore has a void return type so base the pointer type on 4288 // the type of the value operand. 4289 if (node->getOp() == glslang::EOpImageAtomicStore) { 4290 resultTypeId = builder.makePointer(spv::StorageClassImage, builder.getTypeId(operands[2].word)); 4291 } else { 4292 resultTypeId = builder.makePointer(spv::StorageClassImage, resultType()); 4293 } 4294 spv::Id pointer = builder.createOp(spv::OpImageTexelPointer, resultTypeId, operands); 4295 4296 std::vector<spv::Id> operands; 4297 operands.push_back(pointer); 4298 for (; opIt != arguments.end(); ++opIt) 4299 operands.push_back(*opIt); 4300 4301 return createAtomicOperation(node->getOp(), precision, resultType(), operands, node->getBasicType()); 4302 } 4303 } 4304 4305 #ifdef AMD_EXTENSIONS 4306 // Check for fragment mask functions other than queries 4307 if (cracked.fragMask) { 4308 assert(sampler.ms); 4309 4310 auto opIt = arguments.begin(); 4311 std::vector<spv::Id> operands; 4312 4313 // Extract the image if necessary 4314 if (builder.isSampledImage(params.sampler)) 4315 params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler); 4316 4317 operands.push_back(params.sampler); 4318 ++opIt; 4319 4320 if (sampler.isSubpass()) { 4321 // add on the (0,0) coordinate 4322 spv::Id zero = builder.makeIntConstant(0); 4323 std::vector<spv::Id> comps; 4324 comps.push_back(zero); 4325 comps.push_back(zero); 4326 operands.push_back(builder.makeCompositeConstant(builder.makeVectorType(builder.makeIntType(32), 2), comps)); 4327 } 4328 4329 for (; opIt != arguments.end(); ++opIt) 4330 operands.push_back(*opIt); 4331 4332 spv::Op fragMaskOp = spv::OpNop; 4333 if (node->getOp() == glslang::EOpFragmentMaskFetch) 4334 fragMaskOp = spv::OpFragmentMaskFetchAMD; 4335 else if (node->getOp() == glslang::EOpFragmentFetch) 4336 fragMaskOp = spv::OpFragmentFetchAMD; 4337 4338 builder.addExtension(spv::E_SPV_AMD_shader_fragment_mask); 4339 builder.addCapability(spv::CapabilityFragmentMaskAMD); 4340 return builder.createOp(fragMaskOp, resultType(), operands); 4341 } 4342 #endif 4343 4344 // Check for texture functions other than queries 4345 bool sparse = node->isSparseTexture(); 4346 #ifdef NV_EXTENSIONS 4347 bool imageFootprint = node->isImageFootprint(); 4348 #endif 4349 4350 bool cubeCompare = sampler.dim == glslang::EsdCube && sampler.arrayed && sampler.shadow; 4351 4352 // check for bias argument 4353 bool bias = false; 4354 #ifdef AMD_EXTENSIONS 4355 if (! cracked.lod && ! cracked.grad && ! cracked.fetch && ! cubeCompare) { 4356 #else 4357 if (! cracked.lod && ! cracked.gather && ! cracked.grad && ! cracked.fetch && ! cubeCompare) { 4358 #endif 4359 int nonBiasArgCount = 2; 4360 #ifdef AMD_EXTENSIONS 4361 if (cracked.gather) 4362 ++nonBiasArgCount; // comp argument should be present when bias argument is present 4363 4364 if (f16ShadowCompare) 4365 ++nonBiasArgCount; 4366 #endif 4367 if (cracked.offset) 4368 ++nonBiasArgCount; 4369 #ifdef AMD_EXTENSIONS 4370 else if (cracked.offsets) 4371 ++nonBiasArgCount; 4372 #endif 4373 if (cracked.grad) 4374 nonBiasArgCount += 2; 4375 if (cracked.lodClamp) 4376 ++nonBiasArgCount; 4377 if (sparse) 4378 ++nonBiasArgCount; 4379 #ifdef NV_EXTENSIONS 4380 if (imageFootprint) 4381 //Following three extra arguments 4382 // int granularity, bool coarse, out gl_TextureFootprint2DNV footprint 4383 nonBiasArgCount += 3; 4384 #endif 4385 if ((int)arguments.size() > nonBiasArgCount) 4386 bias = true; 4387 } 4388 4389 // See if the sampler param should really be just the SPV image part 4390 if (cracked.fetch) { 4391 // a fetch needs to have the image extracted first 4392 if (builder.isSampledImage(params.sampler)) 4393 params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler); 4394 } 4395 4396 #ifdef AMD_EXTENSIONS 4397 if (cracked.gather) { 4398 const auto& sourceExtensions = glslangIntermediate->getRequestedExtensions(); 4399 if (bias || cracked.lod || 4400 sourceExtensions.find(glslang::E_GL_AMD_texture_gather_bias_lod) != sourceExtensions.end()) { 4401 builder.addExtension(spv::E_SPV_AMD_texture_gather_bias_lod); 4402 builder.addCapability(spv::CapabilityImageGatherBiasLodAMD); 4403 } 4404 } 4405 #endif 4406 4407 // set the rest of the arguments 4408 4409 params.coords = arguments[1]; 4410 int extraArgs = 0; 4411 bool noImplicitLod = false; 4412 4413 // sort out where Dref is coming from 4414 #ifdef AMD_EXTENSIONS 4415 if (cubeCompare || f16ShadowCompare) { 4416 #else 4417 if (cubeCompare) { 4418 #endif 4419 params.Dref = arguments[2]; 4420 ++extraArgs; 4421 } else if (sampler.shadow && cracked.gather) { 4422 params.Dref = arguments[2]; 4423 ++extraArgs; 4424 } else if (sampler.shadow) { 4425 std::vector<spv::Id> indexes; 4426 int dRefComp; 4427 if (cracked.proj) 4428 dRefComp = 2; // "The resulting 3rd component of P in the shadow forms is used as Dref" 4429 else 4430 dRefComp = builder.getNumComponents(params.coords) - 1; 4431 indexes.push_back(dRefComp); 4432 params.Dref = builder.createCompositeExtract(params.coords, builder.getScalarTypeId(builder.getTypeId(params.coords)), indexes); 4433 } 4434 4435 // lod 4436 if (cracked.lod) { 4437 params.lod = arguments[2 + extraArgs]; 4438 ++extraArgs; 4439 } else if (glslangIntermediate->getStage() != EShLangFragment 4440 #ifdef NV_EXTENSIONS 4441 // NV_compute_shader_derivatives layout qualifiers allow for implicit LODs 4442 && !(glslangIntermediate->getStage() == EShLangCompute && 4443 (glslangIntermediate->getLayoutDerivativeModeNone() != glslang::LayoutDerivativeNone)) 4444 #endif 4445 ) { 4446 // we need to invent the default lod for an explicit lod instruction for a non-fragment stage 4447 noImplicitLod = true; 4448 } 4449 4450 // multisample 4451 if (sampler.ms) { 4452 params.sample = arguments[2 + extraArgs]; // For MS, "sample" should be specified 4453 ++extraArgs; 4454 } 4455 4456 // gradient 4457 if (cracked.grad) { 4458 params.gradX = arguments[2 + extraArgs]; 4459 params.gradY = arguments[3 + extraArgs]; 4460 extraArgs += 2; 4461 } 4462 4463 // offset and offsets 4464 if (cracked.offset) { 4465 params.offset = arguments[2 + extraArgs]; 4466 ++extraArgs; 4467 } else if (cracked.offsets) { 4468 params.offsets = arguments[2 + extraArgs]; 4469 ++extraArgs; 4470 } 4471 4472 // lod clamp 4473 if (cracked.lodClamp) { 4474 params.lodClamp = arguments[2 + extraArgs]; 4475 ++extraArgs; 4476 } 4477 // sparse 4478 if (sparse) { 4479 params.texelOut = arguments[2 + extraArgs]; 4480 ++extraArgs; 4481 } 4482 4483 // gather component 4484 if (cracked.gather && ! sampler.shadow) { 4485 // default component is 0, if missing, otherwise an argument 4486 if (2 + extraArgs < (int)arguments.size()) { 4487 params.component = arguments[2 + extraArgs]; 4488 ++extraArgs; 4489 } else 4490 params.component = builder.makeIntConstant(0); 4491 } 4492 #ifdef NV_EXTENSIONS 4493 spv::Id resultStruct = spv::NoResult; 4494 if (imageFootprint) { 4495 //Following three extra arguments 4496 // int granularity, bool coarse, out gl_TextureFootprint2DNV footprint 4497 params.granularity = arguments[2 + extraArgs]; 4498 params.coarse = arguments[3 + extraArgs]; 4499 resultStruct = arguments[4 + extraArgs]; 4500 extraArgs += 3; 4501 } 4502 #endif 4503 // bias 4504 if (bias) { 4505 params.bias = arguments[2 + extraArgs]; 4506 ++extraArgs; 4507 } 4508 4509 #ifdef NV_EXTENSIONS 4510 if (imageFootprint) { 4511 builder.addExtension(spv::E_SPV_NV_shader_image_footprint); 4512 builder.addCapability(spv::CapabilityImageFootprintNV); 4513 4514 4515 //resultStructType(OpenGL type) contains 5 elements: 4516 //struct gl_TextureFootprint2DNV { 4517 // uvec2 anchor; 4518 // uvec2 offset; 4519 // uvec2 mask; 4520 // uint lod; 4521 // uint granularity; 4522 //}; 4523 //or 4524 //struct gl_TextureFootprint3DNV { 4525 // uvec3 anchor; 4526 // uvec3 offset; 4527 // uvec2 mask; 4528 // uint lod; 4529 // uint granularity; 4530 //}; 4531 spv::Id resultStructType = builder.getContainedTypeId(builder.getTypeId(resultStruct)); 4532 assert(builder.isStructType(resultStructType)); 4533 4534 //resType (SPIR-V type) contains 6 elements: 4535 //Member 0 must be a Boolean type scalar(LOD), 4536 //Member 1 must be a vector of integer type, whose Signedness operand is 0(anchor), 4537 //Member 2 must be a vector of integer type, whose Signedness operand is 0(offset), 4538 //Member 3 must be a vector of integer type, whose Signedness operand is 0(mask), 4539 //Member 4 must be a scalar of integer type, whose Signedness operand is 0(lod), 4540 //Member 5 must be a scalar of integer type, whose Signedness operand is 0(granularity). 4541 std::vector<spv::Id> members; 4542 members.push_back(resultType()); 4543 for (int i = 0; i < 5; i++) { 4544 members.push_back(builder.getContainedTypeId(resultStructType, i)); 4545 } 4546 spv::Id resType = builder.makeStructType(members, "ResType"); 4547 4548 //call ImageFootprintNV 4549 spv::Id res = builder.createTextureCall(precision, resType, sparse, cracked.fetch, cracked.proj, cracked.gather, noImplicitLod, params); 4550 4551 //copy resType (SPIR-V type) to resultStructType(OpenGL type) 4552 for (int i = 0; i < 5; i++) { 4553 builder.clearAccessChain(); 4554 builder.setAccessChainLValue(resultStruct); 4555 4556 //Accessing to a struct we created, no coherent flag is set 4557 spv::Builder::AccessChain::CoherentFlags flags; 4558 flags.clear(); 4559 4560 builder.accessChainPush(builder.makeIntConstant(i), flags, 0); 4561 builder.accessChainStore(builder.createCompositeExtract(res, builder.getContainedTypeId(resType, i+1), i+1)); 4562 } 4563 return builder.createCompositeExtract(res, resultType(), 0); 4564 } 4565 #endif 4566 4567 // projective component (might not to move) 4568 // GLSL: "The texture coordinates consumed from P, not including the last component of P, 4569 // are divided by the last component of P." 4570 // SPIR-V: "... (u [, v] [, w], q)... It may be a vector larger than needed, but all 4571 // unused components will appear after all used components." 4572 if (cracked.proj) { 4573 int projSourceComp = builder.getNumComponents(params.coords) - 1; 4574 int projTargetComp; 4575 switch (sampler.dim) { 4576 case glslang::Esd1D: projTargetComp = 1; break; 4577 case glslang::Esd2D: projTargetComp = 2; break; 4578 case glslang::EsdRect: projTargetComp = 2; break; 4579 default: projTargetComp = projSourceComp; break; 4580 } 4581 // copy the projective coordinate if we have to 4582 if (projTargetComp != projSourceComp) { 4583 spv::Id projComp = builder.createCompositeExtract(params.coords, 4584 builder.getScalarTypeId(builder.getTypeId(params.coords)), 4585 projSourceComp); 4586 params.coords = builder.createCompositeInsert(projComp, params.coords, 4587 builder.getTypeId(params.coords), projTargetComp); 4588 } 4589 } 4590 4591 // nonprivate 4592 if (imageType.getQualifier().nonprivate) { 4593 params.nonprivate = true; 4594 } 4595 4596 // volatile 4597 if (imageType.getQualifier().volatil) { 4598 params.volatil = true; 4599 } 4600 4601 std::vector<spv::Id> result( 1, 4602 builder.createTextureCall(precision, resultType(), sparse, cracked.fetch, cracked.proj, cracked.gather, noImplicitLod, params) 4603 ); 4604 4605 if (components != node->getType().getVectorSize()) 4606 result[0] = builder.createConstructor(precision, result, convertGlslangToSpvType(node->getType())); 4607 4608 return result[0]; 4609 } 4610 4611 spv::Id TGlslangToSpvTraverser::handleUserFunctionCall(const glslang::TIntermAggregate* node) 4612 { 4613 // Grab the function's pointer from the previously created function 4614 spv::Function* function = functionMap[node->getName().c_str()]; 4615 if (! function) 4616 return 0; 4617 4618 const glslang::TIntermSequence& glslangArgs = node->getSequence(); 4619 const glslang::TQualifierList& qualifiers = node->getQualifierList(); 4620 4621 // See comments in makeFunctions() for details about the semantics for parameter passing. 4622 // 4623 // These imply we need a four step process: 4624 // 1. Evaluate the arguments 4625 // 2. Allocate and make copies of in, out, and inout arguments 4626 // 3. Make the call 4627 // 4. Copy back the results 4628 4629 // 1. Evaluate the arguments and their types 4630 std::vector<spv::Builder::AccessChain> lValues; 4631 std::vector<spv::Id> rValues; 4632 std::vector<const glslang::TType*> argTypes; 4633 for (int a = 0; a < (int)glslangArgs.size(); ++a) { 4634 argTypes.push_back(&glslangArgs[a]->getAsTyped()->getType()); 4635 // build l-value 4636 builder.clearAccessChain(); 4637 glslangArgs[a]->traverse(this); 4638 // keep outputs and pass-by-originals as l-values, evaluate others as r-values 4639 if (originalParam(qualifiers[a], *argTypes[a], function->hasImplicitThis() && a == 0) || 4640 writableParam(qualifiers[a])) { 4641 // save l-value 4642 lValues.push_back(builder.getAccessChain()); 4643 } else { 4644 // process r-value 4645 rValues.push_back(accessChainLoad(*argTypes.back())); 4646 } 4647 } 4648 4649 // 2. Allocate space for anything needing a copy, and if it's "in" or "inout" 4650 // copy the original into that space. 4651 // 4652 // Also, build up the list of actual arguments to pass in for the call 4653 int lValueCount = 0; 4654 int rValueCount = 0; 4655 std::vector<spv::Id> spvArgs; 4656 for (int a = 0; a < (int)glslangArgs.size(); ++a) { 4657 spv::Id arg; 4658 if (originalParam(qualifiers[a], *argTypes[a], function->hasImplicitThis() && a == 0)) { 4659 builder.setAccessChain(lValues[lValueCount]); 4660 arg = builder.accessChainGetLValue(); 4661 ++lValueCount; 4662 } else if (writableParam(qualifiers[a])) { 4663 // need space to hold the copy 4664 arg = builder.createVariable(spv::StorageClassFunction, builder.getContainedTypeId(function->getParamType(a)), "param"); 4665 if (qualifiers[a] == glslang::EvqIn || qualifiers[a] == glslang::EvqInOut) { 4666 // need to copy the input into output space 4667 builder.setAccessChain(lValues[lValueCount]); 4668 spv::Id copy = accessChainLoad(*argTypes[a]); 4669 builder.clearAccessChain(); 4670 builder.setAccessChainLValue(arg); 4671 multiTypeStore(*argTypes[a], copy); 4672 } 4673 ++lValueCount; 4674 } else { 4675 // process r-value, which involves a copy for a type mismatch 4676 if (function->getParamType(a) != convertGlslangToSpvType(*argTypes[a])) { 4677 spv::Id argCopy = builder.createVariable(spv::StorageClassFunction, function->getParamType(a), "arg"); 4678 builder.clearAccessChain(); 4679 builder.setAccessChainLValue(argCopy); 4680 multiTypeStore(*argTypes[a], rValues[rValueCount]); 4681 arg = builder.createLoad(argCopy); 4682 } else 4683 arg = rValues[rValueCount]; 4684 ++rValueCount; 4685 } 4686 spvArgs.push_back(arg); 4687 } 4688 4689 // 3. Make the call. 4690 spv::Id result = builder.createFunctionCall(function, spvArgs); 4691 builder.setPrecision(result, TranslatePrecisionDecoration(node->getType())); 4692 4693 // 4. Copy back out an "out" arguments. 4694 lValueCount = 0; 4695 for (int a = 0; a < (int)glslangArgs.size(); ++a) { 4696 if (originalParam(qualifiers[a], *argTypes[a], function->hasImplicitThis() && a == 0)) 4697 ++lValueCount; 4698 else if (writableParam(qualifiers[a])) { 4699 if (qualifiers[a] == glslang::EvqOut || qualifiers[a] == glslang::EvqInOut) { 4700 spv::Id copy = builder.createLoad(spvArgs[a]); 4701 builder.setAccessChain(lValues[lValueCount]); 4702 multiTypeStore(*argTypes[a], copy); 4703 } 4704 ++lValueCount; 4705 } 4706 } 4707 4708 return result; 4709 } 4710 4711 // Translate AST operation to SPV operation, already having SPV-based operands/types. 4712 spv::Id TGlslangToSpvTraverser::createBinaryOperation(glslang::TOperator op, OpDecorations& decorations, 4713 spv::Id typeId, spv::Id left, spv::Id right, 4714 glslang::TBasicType typeProxy, bool reduceComparison) 4715 { 4716 bool isUnsigned = isTypeUnsignedInt(typeProxy); 4717 bool isFloat = isTypeFloat(typeProxy); 4718 bool isBool = typeProxy == glslang::EbtBool; 4719 4720 spv::Op binOp = spv::OpNop; 4721 bool needMatchingVectors = true; // for non-matrix ops, would a scalar need to smear to match a vector? 4722 bool comparison = false; 4723 4724 switch (op) { 4725 case glslang::EOpAdd: 4726 case glslang::EOpAddAssign: 4727 if (isFloat) 4728 binOp = spv::OpFAdd; 4729 else 4730 binOp = spv::OpIAdd; 4731 break; 4732 case glslang::EOpSub: 4733 case glslang::EOpSubAssign: 4734 if (isFloat) 4735 binOp = spv::OpFSub; 4736 else 4737 binOp = spv::OpISub; 4738 break; 4739 case glslang::EOpMul: 4740 case glslang::EOpMulAssign: 4741 if (isFloat) 4742 binOp = spv::OpFMul; 4743 else 4744 binOp = spv::OpIMul; 4745 break; 4746 case glslang::EOpVectorTimesScalar: 4747 case glslang::EOpVectorTimesScalarAssign: 4748 if (isFloat && (builder.isVector(left) || builder.isVector(right))) { 4749 if (builder.isVector(right)) 4750 std::swap(left, right); 4751 assert(builder.isScalar(right)); 4752 needMatchingVectors = false; 4753 binOp = spv::OpVectorTimesScalar; 4754 } else if (isFloat) 4755 binOp = spv::OpFMul; 4756 else 4757 binOp = spv::OpIMul; 4758 break; 4759 case glslang::EOpVectorTimesMatrix: 4760 case glslang::EOpVectorTimesMatrixAssign: 4761 binOp = spv::OpVectorTimesMatrix; 4762 break; 4763 case glslang::EOpMatrixTimesVector: 4764 binOp = spv::OpMatrixTimesVector; 4765 break; 4766 case glslang::EOpMatrixTimesScalar: 4767 case glslang::EOpMatrixTimesScalarAssign: 4768 binOp = spv::OpMatrixTimesScalar; 4769 break; 4770 case glslang::EOpMatrixTimesMatrix: 4771 case glslang::EOpMatrixTimesMatrixAssign: 4772 binOp = spv::OpMatrixTimesMatrix; 4773 break; 4774 case glslang::EOpOuterProduct: 4775 binOp = spv::OpOuterProduct; 4776 needMatchingVectors = false; 4777 break; 4778 4779 case glslang::EOpDiv: 4780 case glslang::EOpDivAssign: 4781 if (isFloat) 4782 binOp = spv::OpFDiv; 4783 else if (isUnsigned) 4784 binOp = spv::OpUDiv; 4785 else 4786 binOp = spv::OpSDiv; 4787 break; 4788 case glslang::EOpMod: 4789 case glslang::EOpModAssign: 4790 if (isFloat) 4791 binOp = spv::OpFMod; 4792 else if (isUnsigned) 4793 binOp = spv::OpUMod; 4794 else 4795 binOp = spv::OpSMod; 4796 break; 4797 case glslang::EOpRightShift: 4798 case glslang::EOpRightShiftAssign: 4799 if (isUnsigned) 4800 binOp = spv::OpShiftRightLogical; 4801 else 4802 binOp = spv::OpShiftRightArithmetic; 4803 break; 4804 case glslang::EOpLeftShift: 4805 case glslang::EOpLeftShiftAssign: 4806 binOp = spv::OpShiftLeftLogical; 4807 break; 4808 case glslang::EOpAnd: 4809 case glslang::EOpAndAssign: 4810 binOp = spv::OpBitwiseAnd; 4811 break; 4812 case glslang::EOpLogicalAnd: 4813 needMatchingVectors = false; 4814 binOp = spv::OpLogicalAnd; 4815 break; 4816 case glslang::EOpInclusiveOr: 4817 case glslang::EOpInclusiveOrAssign: 4818 binOp = spv::OpBitwiseOr; 4819 break; 4820 case glslang::EOpLogicalOr: 4821 needMatchingVectors = false; 4822 binOp = spv::OpLogicalOr; 4823 break; 4824 case glslang::EOpExclusiveOr: 4825 case glslang::EOpExclusiveOrAssign: 4826 binOp = spv::OpBitwiseXor; 4827 break; 4828 case glslang::EOpLogicalXor: 4829 needMatchingVectors = false; 4830 binOp = spv::OpLogicalNotEqual; 4831 break; 4832 4833 case glslang::EOpLessThan: 4834 case glslang::EOpGreaterThan: 4835 case glslang::EOpLessThanEqual: 4836 case glslang::EOpGreaterThanEqual: 4837 case glslang::EOpEqual: 4838 case glslang::EOpNotEqual: 4839 case glslang::EOpVectorEqual: 4840 case glslang::EOpVectorNotEqual: 4841 comparison = true; 4842 break; 4843 default: 4844 break; 4845 } 4846 4847 // handle mapped binary operations (should be non-comparison) 4848 if (binOp != spv::OpNop) { 4849 assert(comparison == false); 4850 if (builder.isMatrix(left) || builder.isMatrix(right)) 4851 return createBinaryMatrixOperation(binOp, decorations, typeId, left, right); 4852 4853 // No matrix involved; make both operands be the same number of components, if needed 4854 if (needMatchingVectors) 4855 builder.promoteScalar(decorations.precision, left, right); 4856 4857 spv::Id result = builder.createBinOp(binOp, typeId, left, right); 4858 builder.addDecoration(result, decorations.noContraction); 4859 builder.addDecoration(result, decorations.nonUniform); 4860 return builder.setPrecision(result, decorations.precision); 4861 } 4862 4863 if (! comparison) 4864 return 0; 4865 4866 // Handle comparison instructions 4867 4868 if (reduceComparison && (op == glslang::EOpEqual || op == glslang::EOpNotEqual) 4869 && (builder.isVector(left) || builder.isMatrix(left) || builder.isAggregate(left))) { 4870 spv::Id result = builder.createCompositeCompare(decorations.precision, left, right, op == glslang::EOpEqual); 4871 builder.addDecoration(result, decorations.nonUniform); 4872 return result; 4873 } 4874 4875 switch (op) { 4876 case glslang::EOpLessThan: 4877 if (isFloat) 4878 binOp = spv::OpFOrdLessThan; 4879 else if (isUnsigned) 4880 binOp = spv::OpULessThan; 4881 else 4882 binOp = spv::OpSLessThan; 4883 break; 4884 case glslang::EOpGreaterThan: 4885 if (isFloat) 4886 binOp = spv::OpFOrdGreaterThan; 4887 else if (isUnsigned) 4888 binOp = spv::OpUGreaterThan; 4889 else 4890 binOp = spv::OpSGreaterThan; 4891 break; 4892 case glslang::EOpLessThanEqual: 4893 if (isFloat) 4894 binOp = spv::OpFOrdLessThanEqual; 4895 else if (isUnsigned) 4896 binOp = spv::OpULessThanEqual; 4897 else 4898 binOp = spv::OpSLessThanEqual; 4899 break; 4900 case glslang::EOpGreaterThanEqual: 4901 if (isFloat) 4902 binOp = spv::OpFOrdGreaterThanEqual; 4903 else if (isUnsigned) 4904 binOp = spv::OpUGreaterThanEqual; 4905 else 4906 binOp = spv::OpSGreaterThanEqual; 4907 break; 4908 case glslang::EOpEqual: 4909 case glslang::EOpVectorEqual: 4910 if (isFloat) 4911 binOp = spv::OpFOrdEqual; 4912 else if (isBool) 4913 binOp = spv::OpLogicalEqual; 4914 else 4915 binOp = spv::OpIEqual; 4916 break; 4917 case glslang::EOpNotEqual: 4918 case glslang::EOpVectorNotEqual: 4919 if (isFloat) 4920 binOp = spv::OpFOrdNotEqual; 4921 else if (isBool) 4922 binOp = spv::OpLogicalNotEqual; 4923 else 4924 binOp = spv::OpINotEqual; 4925 break; 4926 default: 4927 break; 4928 } 4929 4930 if (binOp != spv::OpNop) { 4931 spv::Id result = builder.createBinOp(binOp, typeId, left, right); 4932 builder.addDecoration(result, decorations.noContraction); 4933 builder.addDecoration(result, decorations.nonUniform); 4934 return builder.setPrecision(result, decorations.precision); 4935 } 4936 4937 return 0; 4938 } 4939 4940 // 4941 // Translate AST matrix operation to SPV operation, already having SPV-based operands/types. 4942 // These can be any of: 4943 // 4944 // matrix * scalar 4945 // scalar * matrix 4946 // matrix * matrix linear algebraic 4947 // matrix * vector 4948 // vector * matrix 4949 // matrix * matrix componentwise 4950 // matrix op matrix op in {+, -, /} 4951 // matrix op scalar op in {+, -, /} 4952 // scalar op matrix op in {+, -, /} 4953 // 4954 spv::Id TGlslangToSpvTraverser::createBinaryMatrixOperation(spv::Op op, OpDecorations& decorations, spv::Id typeId, 4955 spv::Id left, spv::Id right) 4956 { 4957 bool firstClass = true; 4958 4959 // First, handle first-class matrix operations (* and matrix/scalar) 4960 switch (op) { 4961 case spv::OpFDiv: 4962 if (builder.isMatrix(left) && builder.isScalar(right)) { 4963 // turn matrix / scalar into a multiply... 4964 spv::Id resultType = builder.getTypeId(right); 4965 right = builder.createBinOp(spv::OpFDiv, resultType, builder.makeFpConstant(resultType, 1.0), right); 4966 op = spv::OpMatrixTimesScalar; 4967 } else 4968 firstClass = false; 4969 break; 4970 case spv::OpMatrixTimesScalar: 4971 if (builder.isMatrix(right)) 4972 std::swap(left, right); 4973 assert(builder.isScalar(right)); 4974 break; 4975 case spv::OpVectorTimesMatrix: 4976 assert(builder.isVector(left)); 4977 assert(builder.isMatrix(right)); 4978 break; 4979 case spv::OpMatrixTimesVector: 4980 assert(builder.isMatrix(left)); 4981 assert(builder.isVector(right)); 4982 break; 4983 case spv::OpMatrixTimesMatrix: 4984 assert(builder.isMatrix(left)); 4985 assert(builder.isMatrix(right)); 4986 break; 4987 default: 4988 firstClass = false; 4989 break; 4990 } 4991 4992 if (firstClass) { 4993 spv::Id result = builder.createBinOp(op, typeId, left, right); 4994 builder.addDecoration(result, decorations.noContraction); 4995 builder.addDecoration(result, decorations.nonUniform); 4996 return builder.setPrecision(result, decorations.precision); 4997 } 4998 4999 // Handle component-wise +, -, *, %, and / for all combinations of type. 5000 // The result type of all of them is the same type as the (a) matrix operand. 5001 // The algorithm is to: 5002 // - break the matrix(es) into vectors 5003 // - smear any scalar to a vector 5004 // - do vector operations 5005 // - make a matrix out the vector results 5006 switch (op) { 5007 case spv::OpFAdd: 5008 case spv::OpFSub: 5009 case spv::OpFDiv: 5010 case spv::OpFMod: 5011 case spv::OpFMul: 5012 { 5013 // one time set up... 5014 bool leftMat = builder.isMatrix(left); 5015 bool rightMat = builder.isMatrix(right); 5016 unsigned int numCols = leftMat ? builder.getNumColumns(left) : builder.getNumColumns(right); 5017 int numRows = leftMat ? builder.getNumRows(left) : builder.getNumRows(right); 5018 spv::Id scalarType = builder.getScalarTypeId(typeId); 5019 spv::Id vecType = builder.makeVectorType(scalarType, numRows); 5020 std::vector<spv::Id> results; 5021 spv::Id smearVec = spv::NoResult; 5022 if (builder.isScalar(left)) 5023 smearVec = builder.smearScalar(decorations.precision, left, vecType); 5024 else if (builder.isScalar(right)) 5025 smearVec = builder.smearScalar(decorations.precision, right, vecType); 5026 5027 // do each vector op 5028 for (unsigned int c = 0; c < numCols; ++c) { 5029 std::vector<unsigned int> indexes; 5030 indexes.push_back(c); 5031 spv::Id leftVec = leftMat ? builder.createCompositeExtract( left, vecType, indexes) : smearVec; 5032 spv::Id rightVec = rightMat ? builder.createCompositeExtract(right, vecType, indexes) : smearVec; 5033 spv::Id result = builder.createBinOp(op, vecType, leftVec, rightVec); 5034 builder.addDecoration(result, decorations.noContraction); 5035 builder.addDecoration(result, decorations.nonUniform); 5036 results.push_back(builder.setPrecision(result, decorations.precision)); 5037 } 5038 5039 // put the pieces together 5040 spv::Id result = builder.setPrecision(builder.createCompositeConstruct(typeId, results), decorations.precision); 5041 builder.addDecoration(result, decorations.nonUniform); 5042 return result; 5043 } 5044 default: 5045 assert(0); 5046 return spv::NoResult; 5047 } 5048 } 5049 5050 spv::Id TGlslangToSpvTraverser::createUnaryOperation(glslang::TOperator op, OpDecorations& decorations, spv::Id typeId, 5051 spv::Id operand, glslang::TBasicType typeProxy) 5052 { 5053 spv::Op unaryOp = spv::OpNop; 5054 int extBuiltins = -1; 5055 int libCall = -1; 5056 bool isUnsigned = isTypeUnsignedInt(typeProxy); 5057 bool isFloat = isTypeFloat(typeProxy); 5058 5059 switch (op) { 5060 case glslang::EOpNegative: 5061 if (isFloat) { 5062 unaryOp = spv::OpFNegate; 5063 if (builder.isMatrixType(typeId)) 5064 return createUnaryMatrixOperation(unaryOp, decorations, typeId, operand, typeProxy); 5065 } else 5066 unaryOp = spv::OpSNegate; 5067 break; 5068 5069 case glslang::EOpLogicalNot: 5070 case glslang::EOpVectorLogicalNot: 5071 unaryOp = spv::OpLogicalNot; 5072 break; 5073 case glslang::EOpBitwiseNot: 5074 unaryOp = spv::OpNot; 5075 break; 5076 5077 case glslang::EOpDeterminant: 5078 libCall = spv::GLSLstd450Determinant; 5079 break; 5080 case glslang::EOpMatrixInverse: 5081 libCall = spv::GLSLstd450MatrixInverse; 5082 break; 5083 case glslang::EOpTranspose: 5084 unaryOp = spv::OpTranspose; 5085 break; 5086 5087 case glslang::EOpRadians: 5088 libCall = spv::GLSLstd450Radians; 5089 break; 5090 case glslang::EOpDegrees: 5091 libCall = spv::GLSLstd450Degrees; 5092 break; 5093 case glslang::EOpSin: 5094 libCall = spv::GLSLstd450Sin; 5095 break; 5096 case glslang::EOpCos: 5097 libCall = spv::GLSLstd450Cos; 5098 break; 5099 case glslang::EOpTan: 5100 libCall = spv::GLSLstd450Tan; 5101 break; 5102 case glslang::EOpAcos: 5103 libCall = spv::GLSLstd450Acos; 5104 break; 5105 case glslang::EOpAsin: 5106 libCall = spv::GLSLstd450Asin; 5107 break; 5108 case glslang::EOpAtan: 5109 libCall = spv::GLSLstd450Atan; 5110 break; 5111 5112 case glslang::EOpAcosh: 5113 libCall = spv::GLSLstd450Acosh; 5114 break; 5115 case glslang::EOpAsinh: 5116 libCall = spv::GLSLstd450Asinh; 5117 break; 5118 case glslang::EOpAtanh: 5119 libCall = spv::GLSLstd450Atanh; 5120 break; 5121 case glslang::EOpTanh: 5122 libCall = spv::GLSLstd450Tanh; 5123 break; 5124 case glslang::EOpCosh: 5125 libCall = spv::GLSLstd450Cosh; 5126 break; 5127 case glslang::EOpSinh: 5128 libCall = spv::GLSLstd450Sinh; 5129 break; 5130 5131 case glslang::EOpLength: 5132 libCall = spv::GLSLstd450Length; 5133 break; 5134 case glslang::EOpNormalize: 5135 libCall = spv::GLSLstd450Normalize; 5136 break; 5137 5138 case glslang::EOpExp: 5139 libCall = spv::GLSLstd450Exp; 5140 break; 5141 case glslang::EOpLog: 5142 libCall = spv::GLSLstd450Log; 5143 break; 5144 case glslang::EOpExp2: 5145 libCall = spv::GLSLstd450Exp2; 5146 break; 5147 case glslang::EOpLog2: 5148 libCall = spv::GLSLstd450Log2; 5149 break; 5150 case glslang::EOpSqrt: 5151 libCall = spv::GLSLstd450Sqrt; 5152 break; 5153 case glslang::EOpInverseSqrt: 5154 libCall = spv::GLSLstd450InverseSqrt; 5155 break; 5156 5157 case glslang::EOpFloor: 5158 libCall = spv::GLSLstd450Floor; 5159 break; 5160 case glslang::EOpTrunc: 5161 libCall = spv::GLSLstd450Trunc; 5162 break; 5163 case glslang::EOpRound: 5164 libCall = spv::GLSLstd450Round; 5165 break; 5166 case glslang::EOpRoundEven: 5167 libCall = spv::GLSLstd450RoundEven; 5168 break; 5169 case glslang::EOpCeil: 5170 libCall = spv::GLSLstd450Ceil; 5171 break; 5172 case glslang::EOpFract: 5173 libCall = spv::GLSLstd450Fract; 5174 break; 5175 5176 case glslang::EOpIsNan: 5177 unaryOp = spv::OpIsNan; 5178 break; 5179 case glslang::EOpIsInf: 5180 unaryOp = spv::OpIsInf; 5181 break; 5182 case glslang::EOpIsFinite: 5183 unaryOp = spv::OpIsFinite; 5184 break; 5185 5186 case glslang::EOpFloatBitsToInt: 5187 case glslang::EOpFloatBitsToUint: 5188 case glslang::EOpIntBitsToFloat: 5189 case glslang::EOpUintBitsToFloat: 5190 case glslang::EOpDoubleBitsToInt64: 5191 case glslang::EOpDoubleBitsToUint64: 5192 case glslang::EOpInt64BitsToDouble: 5193 case glslang::EOpUint64BitsToDouble: 5194 case glslang::EOpFloat16BitsToInt16: 5195 case glslang::EOpFloat16BitsToUint16: 5196 case glslang::EOpInt16BitsToFloat16: 5197 case glslang::EOpUint16BitsToFloat16: 5198 unaryOp = spv::OpBitcast; 5199 break; 5200 5201 case glslang::EOpPackSnorm2x16: 5202 libCall = spv::GLSLstd450PackSnorm2x16; 5203 break; 5204 case glslang::EOpUnpackSnorm2x16: 5205 libCall = spv::GLSLstd450UnpackSnorm2x16; 5206 break; 5207 case glslang::EOpPackUnorm2x16: 5208 libCall = spv::GLSLstd450PackUnorm2x16; 5209 break; 5210 case glslang::EOpUnpackUnorm2x16: 5211 libCall = spv::GLSLstd450UnpackUnorm2x16; 5212 break; 5213 case glslang::EOpPackHalf2x16: 5214 libCall = spv::GLSLstd450PackHalf2x16; 5215 break; 5216 case glslang::EOpUnpackHalf2x16: 5217 libCall = spv::GLSLstd450UnpackHalf2x16; 5218 break; 5219 case glslang::EOpPackSnorm4x8: 5220 libCall = spv::GLSLstd450PackSnorm4x8; 5221 break; 5222 case glslang::EOpUnpackSnorm4x8: 5223 libCall = spv::GLSLstd450UnpackSnorm4x8; 5224 break; 5225 case glslang::EOpPackUnorm4x8: 5226 libCall = spv::GLSLstd450PackUnorm4x8; 5227 break; 5228 case glslang::EOpUnpackUnorm4x8: 5229 libCall = spv::GLSLstd450UnpackUnorm4x8; 5230 break; 5231 case glslang::EOpPackDouble2x32: 5232 libCall = spv::GLSLstd450PackDouble2x32; 5233 break; 5234 case glslang::EOpUnpackDouble2x32: 5235 libCall = spv::GLSLstd450UnpackDouble2x32; 5236 break; 5237 5238 case glslang::EOpPackInt2x32: 5239 case glslang::EOpUnpackInt2x32: 5240 case glslang::EOpPackUint2x32: 5241 case glslang::EOpUnpackUint2x32: 5242 case glslang::EOpPack16: 5243 case glslang::EOpPack32: 5244 case glslang::EOpPack64: 5245 case glslang::EOpUnpack32: 5246 case glslang::EOpUnpack16: 5247 case glslang::EOpUnpack8: 5248 case glslang::EOpPackInt2x16: 5249 case glslang::EOpUnpackInt2x16: 5250 case glslang::EOpPackUint2x16: 5251 case glslang::EOpUnpackUint2x16: 5252 case glslang::EOpPackInt4x16: 5253 case glslang::EOpUnpackInt4x16: 5254 case glslang::EOpPackUint4x16: 5255 case glslang::EOpUnpackUint4x16: 5256 case glslang::EOpPackFloat2x16: 5257 case glslang::EOpUnpackFloat2x16: 5258 unaryOp = spv::OpBitcast; 5259 break; 5260 5261 case glslang::EOpDPdx: 5262 unaryOp = spv::OpDPdx; 5263 break; 5264 case glslang::EOpDPdy: 5265 unaryOp = spv::OpDPdy; 5266 break; 5267 case glslang::EOpFwidth: 5268 unaryOp = spv::OpFwidth; 5269 break; 5270 case glslang::EOpDPdxFine: 5271 unaryOp = spv::OpDPdxFine; 5272 break; 5273 case glslang::EOpDPdyFine: 5274 unaryOp = spv::OpDPdyFine; 5275 break; 5276 case glslang::EOpFwidthFine: 5277 unaryOp = spv::OpFwidthFine; 5278 break; 5279 case glslang::EOpDPdxCoarse: 5280 unaryOp = spv::OpDPdxCoarse; 5281 break; 5282 case glslang::EOpDPdyCoarse: 5283 unaryOp = spv::OpDPdyCoarse; 5284 break; 5285 case glslang::EOpFwidthCoarse: 5286 unaryOp = spv::OpFwidthCoarse; 5287 break; 5288 case glslang::EOpInterpolateAtCentroid: 5289 #ifdef AMD_EXTENSIONS 5290 if (typeProxy == glslang::EbtFloat16) 5291 builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float); 5292 #endif 5293 libCall = spv::GLSLstd450InterpolateAtCentroid; 5294 break; 5295 case glslang::EOpAny: 5296 unaryOp = spv::OpAny; 5297 break; 5298 case glslang::EOpAll: 5299 unaryOp = spv::OpAll; 5300 break; 5301 5302 case glslang::EOpAbs: 5303 if (isFloat) 5304 libCall = spv::GLSLstd450FAbs; 5305 else 5306 libCall = spv::GLSLstd450SAbs; 5307 break; 5308 case glslang::EOpSign: 5309 if (isFloat) 5310 libCall = spv::GLSLstd450FSign; 5311 else 5312 libCall = spv::GLSLstd450SSign; 5313 break; 5314 5315 case glslang::EOpAtomicCounterIncrement: 5316 case glslang::EOpAtomicCounterDecrement: 5317 case glslang::EOpAtomicCounter: 5318 { 5319 // Handle all of the atomics in one place, in createAtomicOperation() 5320 std::vector<spv::Id> operands; 5321 operands.push_back(operand); 5322 return createAtomicOperation(op, decorations.precision, typeId, operands, typeProxy); 5323 } 5324 5325 case glslang::EOpBitFieldReverse: 5326 unaryOp = spv::OpBitReverse; 5327 break; 5328 case glslang::EOpBitCount: 5329 unaryOp = spv::OpBitCount; 5330 break; 5331 case glslang::EOpFindLSB: 5332 libCall = spv::GLSLstd450FindILsb; 5333 break; 5334 case glslang::EOpFindMSB: 5335 if (isUnsigned) 5336 libCall = spv::GLSLstd450FindUMsb; 5337 else 5338 libCall = spv::GLSLstd450FindSMsb; 5339 break; 5340 5341 case glslang::EOpBallot: 5342 case glslang::EOpReadFirstInvocation: 5343 case glslang::EOpAnyInvocation: 5344 case glslang::EOpAllInvocations: 5345 case glslang::EOpAllInvocationsEqual: 5346 #ifdef AMD_EXTENSIONS 5347 case glslang::EOpMinInvocations: 5348 case glslang::EOpMaxInvocations: 5349 case glslang::EOpAddInvocations: 5350 case glslang::EOpMinInvocationsNonUniform: 5351 case glslang::EOpMaxInvocationsNonUniform: 5352 case glslang::EOpAddInvocationsNonUniform: 5353 case glslang::EOpMinInvocationsInclusiveScan: 5354 case glslang::EOpMaxInvocationsInclusiveScan: 5355 case glslang::EOpAddInvocationsInclusiveScan: 5356 case glslang::EOpMinInvocationsInclusiveScanNonUniform: 5357 case glslang::EOpMaxInvocationsInclusiveScanNonUniform: 5358 case glslang::EOpAddInvocationsInclusiveScanNonUniform: 5359 case glslang::EOpMinInvocationsExclusiveScan: 5360 case glslang::EOpMaxInvocationsExclusiveScan: 5361 case glslang::EOpAddInvocationsExclusiveScan: 5362 case glslang::EOpMinInvocationsExclusiveScanNonUniform: 5363 case glslang::EOpMaxInvocationsExclusiveScanNonUniform: 5364 case glslang::EOpAddInvocationsExclusiveScanNonUniform: 5365 #endif 5366 { 5367 std::vector<spv::Id> operands; 5368 operands.push_back(operand); 5369 return createInvocationsOperation(op, typeId, operands, typeProxy); 5370 } 5371 case glslang::EOpSubgroupAll: 5372 case glslang::EOpSubgroupAny: 5373 case glslang::EOpSubgroupAllEqual: 5374 case glslang::EOpSubgroupBroadcastFirst: 5375 case glslang::EOpSubgroupBallot: 5376 case glslang::EOpSubgroupInverseBallot: 5377 case glslang::EOpSubgroupBallotBitCount: 5378 case glslang::EOpSubgroupBallotInclusiveBitCount: 5379 case glslang::EOpSubgroupBallotExclusiveBitCount: 5380 case glslang::EOpSubgroupBallotFindLSB: 5381 case glslang::EOpSubgroupBallotFindMSB: 5382 case glslang::EOpSubgroupAdd: 5383 case glslang::EOpSubgroupMul: 5384 case glslang::EOpSubgroupMin: 5385 case glslang::EOpSubgroupMax: 5386 case glslang::EOpSubgroupAnd: 5387 case glslang::EOpSubgroupOr: 5388 case glslang::EOpSubgroupXor: 5389 case glslang::EOpSubgroupInclusiveAdd: 5390 case glslang::EOpSubgroupInclusiveMul: 5391 case glslang::EOpSubgroupInclusiveMin: 5392 case glslang::EOpSubgroupInclusiveMax: 5393 case glslang::EOpSubgroupInclusiveAnd: 5394 case glslang::EOpSubgroupInclusiveOr: 5395 case glslang::EOpSubgroupInclusiveXor: 5396 case glslang::EOpSubgroupExclusiveAdd: 5397 case glslang::EOpSubgroupExclusiveMul: 5398 case glslang::EOpSubgroupExclusiveMin: 5399 case glslang::EOpSubgroupExclusiveMax: 5400 case glslang::EOpSubgroupExclusiveAnd: 5401 case glslang::EOpSubgroupExclusiveOr: 5402 case glslang::EOpSubgroupExclusiveXor: 5403 case glslang::EOpSubgroupQuadSwapHorizontal: 5404 case glslang::EOpSubgroupQuadSwapVertical: 5405 case glslang::EOpSubgroupQuadSwapDiagonal: { 5406 std::vector<spv::Id> operands; 5407 operands.push_back(operand); 5408 return createSubgroupOperation(op, typeId, operands, typeProxy); 5409 } 5410 #ifdef AMD_EXTENSIONS 5411 case glslang::EOpMbcnt: 5412 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot); 5413 libCall = spv::MbcntAMD; 5414 break; 5415 5416 case glslang::EOpCubeFaceIndex: 5417 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_gcn_shader); 5418 libCall = spv::CubeFaceIndexAMD; 5419 break; 5420 5421 case glslang::EOpCubeFaceCoord: 5422 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_gcn_shader); 5423 libCall = spv::CubeFaceCoordAMD; 5424 break; 5425 #endif 5426 #ifdef NV_EXTENSIONS 5427 case glslang::EOpSubgroupPartition: 5428 unaryOp = spv::OpGroupNonUniformPartitionNV; 5429 break; 5430 #endif 5431 case glslang::EOpConstructReference: 5432 unaryOp = spv::OpBitcast; 5433 break; 5434 default: 5435 return 0; 5436 } 5437 5438 spv::Id id; 5439 if (libCall >= 0) { 5440 std::vector<spv::Id> args; 5441 args.push_back(operand); 5442 id = builder.createBuiltinCall(typeId, extBuiltins >= 0 ? extBuiltins : stdBuiltins, libCall, args); 5443 } else { 5444 id = builder.createUnaryOp(unaryOp, typeId, operand); 5445 } 5446 5447 builder.addDecoration(id, decorations.noContraction); 5448 builder.addDecoration(id, decorations.nonUniform); 5449 return builder.setPrecision(id, decorations.precision); 5450 } 5451 5452 // Create a unary operation on a matrix 5453 spv::Id TGlslangToSpvTraverser::createUnaryMatrixOperation(spv::Op op, OpDecorations& decorations, spv::Id typeId, 5454 spv::Id operand, glslang::TBasicType /* typeProxy */) 5455 { 5456 // Handle unary operations vector by vector. 5457 // The result type is the same type as the original type. 5458 // The algorithm is to: 5459 // - break the matrix into vectors 5460 // - apply the operation to each vector 5461 // - make a matrix out the vector results 5462 5463 // get the types sorted out 5464 int numCols = builder.getNumColumns(operand); 5465 int numRows = builder.getNumRows(operand); 5466 spv::Id srcVecType = builder.makeVectorType(builder.getScalarTypeId(builder.getTypeId(operand)), numRows); 5467 spv::Id destVecType = builder.makeVectorType(builder.getScalarTypeId(typeId), numRows); 5468 std::vector<spv::Id> results; 5469 5470 // do each vector op 5471 for (int c = 0; c < numCols; ++c) { 5472 std::vector<unsigned int> indexes; 5473 indexes.push_back(c); 5474 spv::Id srcVec = builder.createCompositeExtract(operand, srcVecType, indexes); 5475 spv::Id destVec = builder.createUnaryOp(op, destVecType, srcVec); 5476 builder.addDecoration(destVec, decorations.noContraction); 5477 builder.addDecoration(destVec, decorations.nonUniform); 5478 results.push_back(builder.setPrecision(destVec, decorations.precision)); 5479 } 5480 5481 // put the pieces together 5482 spv::Id result = builder.setPrecision(builder.createCompositeConstruct(typeId, results), decorations.precision); 5483 builder.addDecoration(result, decorations.nonUniform); 5484 return result; 5485 } 5486 5487 // For converting integers where both the bitwidth and the signedness could 5488 // change, but only do the width change here. The caller is still responsible 5489 // for the signedness conversion. 5490 spv::Id TGlslangToSpvTraverser::createIntWidthConversion(glslang::TOperator op, spv::Id operand, int vectorSize) 5491 { 5492 // Get the result type width, based on the type to convert to. 5493 int width = 32; 5494 switch(op) { 5495 case glslang::EOpConvInt16ToUint8: 5496 case glslang::EOpConvIntToUint8: 5497 case glslang::EOpConvInt64ToUint8: 5498 case glslang::EOpConvUint16ToInt8: 5499 case glslang::EOpConvUintToInt8: 5500 case glslang::EOpConvUint64ToInt8: 5501 width = 8; 5502 break; 5503 case glslang::EOpConvInt8ToUint16: 5504 case glslang::EOpConvIntToUint16: 5505 case glslang::EOpConvInt64ToUint16: 5506 case glslang::EOpConvUint8ToInt16: 5507 case glslang::EOpConvUintToInt16: 5508 case glslang::EOpConvUint64ToInt16: 5509 width = 16; 5510 break; 5511 case glslang::EOpConvInt8ToUint: 5512 case glslang::EOpConvInt16ToUint: 5513 case glslang::EOpConvInt64ToUint: 5514 case glslang::EOpConvUint8ToInt: 5515 case glslang::EOpConvUint16ToInt: 5516 case glslang::EOpConvUint64ToInt: 5517 width = 32; 5518 break; 5519 case glslang::EOpConvInt8ToUint64: 5520 case glslang::EOpConvInt16ToUint64: 5521 case glslang::EOpConvIntToUint64: 5522 case glslang::EOpConvUint8ToInt64: 5523 case glslang::EOpConvUint16ToInt64: 5524 case glslang::EOpConvUintToInt64: 5525 width = 64; 5526 break; 5527 5528 default: 5529 assert(false && "Default missing"); 5530 break; 5531 } 5532 5533 // Get the conversion operation and result type, 5534 // based on the target width, but the source type. 5535 spv::Id type = spv::NoType; 5536 spv::Op convOp = spv::OpNop; 5537 switch(op) { 5538 case glslang::EOpConvInt8ToUint16: 5539 case glslang::EOpConvInt8ToUint: 5540 case glslang::EOpConvInt8ToUint64: 5541 case glslang::EOpConvInt16ToUint8: 5542 case glslang::EOpConvInt16ToUint: 5543 case glslang::EOpConvInt16ToUint64: 5544 case glslang::EOpConvIntToUint8: 5545 case glslang::EOpConvIntToUint16: 5546 case glslang::EOpConvIntToUint64: 5547 case glslang::EOpConvInt64ToUint8: 5548 case glslang::EOpConvInt64ToUint16: 5549 case glslang::EOpConvInt64ToUint: 5550 convOp = spv::OpSConvert; 5551 type = builder.makeIntType(width); 5552 break; 5553 default: 5554 convOp = spv::OpUConvert; 5555 type = builder.makeUintType(width); 5556 break; 5557 } 5558 5559 if (vectorSize > 0) 5560 type = builder.makeVectorType(type, vectorSize); 5561 5562 return builder.createUnaryOp(convOp, type, operand); 5563 } 5564 5565 spv::Id TGlslangToSpvTraverser::createConversion(glslang::TOperator op, OpDecorations& decorations, spv::Id destType, 5566 spv::Id operand, glslang::TBasicType typeProxy) 5567 { 5568 spv::Op convOp = spv::OpNop; 5569 spv::Id zero = 0; 5570 spv::Id one = 0; 5571 5572 int vectorSize = builder.isVectorType(destType) ? builder.getNumTypeComponents(destType) : 0; 5573 5574 switch (op) { 5575 case glslang::EOpConvInt8ToBool: 5576 case glslang::EOpConvUint8ToBool: 5577 zero = builder.makeUint8Constant(0); 5578 zero = makeSmearedConstant(zero, vectorSize); 5579 return builder.createBinOp(spv::OpINotEqual, destType, operand, zero); 5580 case glslang::EOpConvInt16ToBool: 5581 case glslang::EOpConvUint16ToBool: 5582 zero = builder.makeUint16Constant(0); 5583 zero = makeSmearedConstant(zero, vectorSize); 5584 return builder.createBinOp(spv::OpINotEqual, destType, operand, zero); 5585 case glslang::EOpConvIntToBool: 5586 case glslang::EOpConvUintToBool: 5587 zero = builder.makeUintConstant(0); 5588 zero = makeSmearedConstant(zero, vectorSize); 5589 return builder.createBinOp(spv::OpINotEqual, destType, operand, zero); 5590 case glslang::EOpConvInt64ToBool: 5591 case glslang::EOpConvUint64ToBool: 5592 zero = builder.makeUint64Constant(0); 5593 zero = makeSmearedConstant(zero, vectorSize); 5594 return builder.createBinOp(spv::OpINotEqual, destType, operand, zero); 5595 5596 case glslang::EOpConvFloatToBool: 5597 zero = builder.makeFloatConstant(0.0F); 5598 zero = makeSmearedConstant(zero, vectorSize); 5599 return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero); 5600 5601 case glslang::EOpConvDoubleToBool: 5602 zero = builder.makeDoubleConstant(0.0); 5603 zero = makeSmearedConstant(zero, vectorSize); 5604 return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero); 5605 5606 case glslang::EOpConvFloat16ToBool: 5607 zero = builder.makeFloat16Constant(0.0F); 5608 zero = makeSmearedConstant(zero, vectorSize); 5609 return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero); 5610 5611 case glslang::EOpConvBoolToFloat: 5612 convOp = spv::OpSelect; 5613 zero = builder.makeFloatConstant(0.0F); 5614 one = builder.makeFloatConstant(1.0F); 5615 break; 5616 5617 case glslang::EOpConvBoolToDouble: 5618 convOp = spv::OpSelect; 5619 zero = builder.makeDoubleConstant(0.0); 5620 one = builder.makeDoubleConstant(1.0); 5621 break; 5622 5623 case glslang::EOpConvBoolToFloat16: 5624 convOp = spv::OpSelect; 5625 zero = builder.makeFloat16Constant(0.0F); 5626 one = builder.makeFloat16Constant(1.0F); 5627 break; 5628 5629 case glslang::EOpConvBoolToInt8: 5630 zero = builder.makeInt8Constant(0); 5631 one = builder.makeInt8Constant(1); 5632 convOp = spv::OpSelect; 5633 break; 5634 5635 case glslang::EOpConvBoolToUint8: 5636 zero = builder.makeUint8Constant(0); 5637 one = builder.makeUint8Constant(1); 5638 convOp = spv::OpSelect; 5639 break; 5640 5641 case glslang::EOpConvBoolToInt16: 5642 zero = builder.makeInt16Constant(0); 5643 one = builder.makeInt16Constant(1); 5644 convOp = spv::OpSelect; 5645 break; 5646 5647 case glslang::EOpConvBoolToUint16: 5648 zero = builder.makeUint16Constant(0); 5649 one = builder.makeUint16Constant(1); 5650 convOp = spv::OpSelect; 5651 break; 5652 5653 case glslang::EOpConvBoolToInt: 5654 case glslang::EOpConvBoolToInt64: 5655 if (op == glslang::EOpConvBoolToInt64) 5656 zero = builder.makeInt64Constant(0); 5657 else 5658 zero = builder.makeIntConstant(0); 5659 5660 if (op == glslang::EOpConvBoolToInt64) 5661 one = builder.makeInt64Constant(1); 5662 else 5663 one = builder.makeIntConstant(1); 5664 5665 convOp = spv::OpSelect; 5666 break; 5667 5668 case glslang::EOpConvBoolToUint: 5669 case glslang::EOpConvBoolToUint64: 5670 if (op == glslang::EOpConvBoolToUint64) 5671 zero = builder.makeUint64Constant(0); 5672 else 5673 zero = builder.makeUintConstant(0); 5674 5675 if (op == glslang::EOpConvBoolToUint64) 5676 one = builder.makeUint64Constant(1); 5677 else 5678 one = builder.makeUintConstant(1); 5679 5680 convOp = spv::OpSelect; 5681 break; 5682 5683 case glslang::EOpConvInt8ToFloat16: 5684 case glslang::EOpConvInt8ToFloat: 5685 case glslang::EOpConvInt8ToDouble: 5686 case glslang::EOpConvInt16ToFloat16: 5687 case glslang::EOpConvInt16ToFloat: 5688 case glslang::EOpConvInt16ToDouble: 5689 case glslang::EOpConvIntToFloat16: 5690 case glslang::EOpConvIntToFloat: 5691 case glslang::EOpConvIntToDouble: 5692 case glslang::EOpConvInt64ToFloat: 5693 case glslang::EOpConvInt64ToDouble: 5694 case glslang::EOpConvInt64ToFloat16: 5695 convOp = spv::OpConvertSToF; 5696 break; 5697 5698 case glslang::EOpConvUint8ToFloat16: 5699 case glslang::EOpConvUint8ToFloat: 5700 case glslang::EOpConvUint8ToDouble: 5701 case glslang::EOpConvUint16ToFloat16: 5702 case glslang::EOpConvUint16ToFloat: 5703 case glslang::EOpConvUint16ToDouble: 5704 case glslang::EOpConvUintToFloat16: 5705 case glslang::EOpConvUintToFloat: 5706 case glslang::EOpConvUintToDouble: 5707 case glslang::EOpConvUint64ToFloat: 5708 case glslang::EOpConvUint64ToDouble: 5709 case glslang::EOpConvUint64ToFloat16: 5710 convOp = spv::OpConvertUToF; 5711 break; 5712 5713 case glslang::EOpConvDoubleToFloat: 5714 case glslang::EOpConvFloatToDouble: 5715 case glslang::EOpConvDoubleToFloat16: 5716 case glslang::EOpConvFloat16ToDouble: 5717 case glslang::EOpConvFloatToFloat16: 5718 case glslang::EOpConvFloat16ToFloat: 5719 convOp = spv::OpFConvert; 5720 if (builder.isMatrixType(destType)) 5721 return createUnaryMatrixOperation(convOp, decorations, destType, operand, typeProxy); 5722 break; 5723 5724 case glslang::EOpConvFloat16ToInt8: 5725 case glslang::EOpConvFloatToInt8: 5726 case glslang::EOpConvDoubleToInt8: 5727 case glslang::EOpConvFloat16ToInt16: 5728 case glslang::EOpConvFloatToInt16: 5729 case glslang::EOpConvDoubleToInt16: 5730 case glslang::EOpConvFloat16ToInt: 5731 case glslang::EOpConvFloatToInt: 5732 case glslang::EOpConvDoubleToInt: 5733 case glslang::EOpConvFloat16ToInt64: 5734 case glslang::EOpConvFloatToInt64: 5735 case glslang::EOpConvDoubleToInt64: 5736 convOp = spv::OpConvertFToS; 5737 break; 5738 5739 case glslang::EOpConvUint8ToInt8: 5740 case glslang::EOpConvInt8ToUint8: 5741 case glslang::EOpConvUint16ToInt16: 5742 case glslang::EOpConvInt16ToUint16: 5743 case glslang::EOpConvUintToInt: 5744 case glslang::EOpConvIntToUint: 5745 case glslang::EOpConvUint64ToInt64: 5746 case glslang::EOpConvInt64ToUint64: 5747 if (builder.isInSpecConstCodeGenMode()) { 5748 // Build zero scalar or vector for OpIAdd. 5749 if(op == glslang::EOpConvUint8ToInt8 || op == glslang::EOpConvInt8ToUint8) { 5750 zero = builder.makeUint8Constant(0); 5751 } else if (op == glslang::EOpConvUint16ToInt16 || op == glslang::EOpConvInt16ToUint16) { 5752 zero = builder.makeUint16Constant(0); 5753 } else if (op == glslang::EOpConvUint64ToInt64 || op == glslang::EOpConvInt64ToUint64) { 5754 zero = builder.makeUint64Constant(0); 5755 } else { 5756 zero = builder.makeUintConstant(0); 5757 } 5758 zero = makeSmearedConstant(zero, vectorSize); 5759 // Use OpIAdd, instead of OpBitcast to do the conversion when 5760 // generating for OpSpecConstantOp instruction. 5761 return builder.createBinOp(spv::OpIAdd, destType, operand, zero); 5762 } 5763 // For normal run-time conversion instruction, use OpBitcast. 5764 convOp = spv::OpBitcast; 5765 break; 5766 5767 case glslang::EOpConvFloat16ToUint8: 5768 case glslang::EOpConvFloatToUint8: 5769 case glslang::EOpConvDoubleToUint8: 5770 case glslang::EOpConvFloat16ToUint16: 5771 case glslang::EOpConvFloatToUint16: 5772 case glslang::EOpConvDoubleToUint16: 5773 case glslang::EOpConvFloat16ToUint: 5774 case glslang::EOpConvFloatToUint: 5775 case glslang::EOpConvDoubleToUint: 5776 case glslang::EOpConvFloatToUint64: 5777 case glslang::EOpConvDoubleToUint64: 5778 case glslang::EOpConvFloat16ToUint64: 5779 convOp = spv::OpConvertFToU; 5780 break; 5781 5782 case glslang::EOpConvInt8ToInt16: 5783 case glslang::EOpConvInt8ToInt: 5784 case glslang::EOpConvInt8ToInt64: 5785 case glslang::EOpConvInt16ToInt8: 5786 case glslang::EOpConvInt16ToInt: 5787 case glslang::EOpConvInt16ToInt64: 5788 case glslang::EOpConvIntToInt8: 5789 case glslang::EOpConvIntToInt16: 5790 case glslang::EOpConvIntToInt64: 5791 case glslang::EOpConvInt64ToInt8: 5792 case glslang::EOpConvInt64ToInt16: 5793 case glslang::EOpConvInt64ToInt: 5794 convOp = spv::OpSConvert; 5795 break; 5796 5797 case glslang::EOpConvUint8ToUint16: 5798 case glslang::EOpConvUint8ToUint: 5799 case glslang::EOpConvUint8ToUint64: 5800 case glslang::EOpConvUint16ToUint8: 5801 case glslang::EOpConvUint16ToUint: 5802 case glslang::EOpConvUint16ToUint64: 5803 case glslang::EOpConvUintToUint8: 5804 case glslang::EOpConvUintToUint16: 5805 case glslang::EOpConvUintToUint64: 5806 case glslang::EOpConvUint64ToUint8: 5807 case glslang::EOpConvUint64ToUint16: 5808 case glslang::EOpConvUint64ToUint: 5809 convOp = spv::OpUConvert; 5810 break; 5811 5812 case glslang::EOpConvInt8ToUint16: 5813 case glslang::EOpConvInt8ToUint: 5814 case glslang::EOpConvInt8ToUint64: 5815 case glslang::EOpConvInt16ToUint8: 5816 case glslang::EOpConvInt16ToUint: 5817 case glslang::EOpConvInt16ToUint64: 5818 case glslang::EOpConvIntToUint8: 5819 case glslang::EOpConvIntToUint16: 5820 case glslang::EOpConvIntToUint64: 5821 case glslang::EOpConvInt64ToUint8: 5822 case glslang::EOpConvInt64ToUint16: 5823 case glslang::EOpConvInt64ToUint: 5824 case glslang::EOpConvUint8ToInt16: 5825 case glslang::EOpConvUint8ToInt: 5826 case glslang::EOpConvUint8ToInt64: 5827 case glslang::EOpConvUint16ToInt8: 5828 case glslang::EOpConvUint16ToInt: 5829 case glslang::EOpConvUint16ToInt64: 5830 case glslang::EOpConvUintToInt8: 5831 case glslang::EOpConvUintToInt16: 5832 case glslang::EOpConvUintToInt64: 5833 case glslang::EOpConvUint64ToInt8: 5834 case glslang::EOpConvUint64ToInt16: 5835 case glslang::EOpConvUint64ToInt: 5836 // OpSConvert/OpUConvert + OpBitCast 5837 operand = createIntWidthConversion(op, operand, vectorSize); 5838 5839 if (builder.isInSpecConstCodeGenMode()) { 5840 // Build zero scalar or vector for OpIAdd. 5841 switch(op) { 5842 case glslang::EOpConvInt16ToUint8: 5843 case glslang::EOpConvIntToUint8: 5844 case glslang::EOpConvInt64ToUint8: 5845 case glslang::EOpConvUint16ToInt8: 5846 case glslang::EOpConvUintToInt8: 5847 case glslang::EOpConvUint64ToInt8: 5848 zero = builder.makeUint8Constant(0); 5849 break; 5850 case glslang::EOpConvInt8ToUint16: 5851 case glslang::EOpConvIntToUint16: 5852 case glslang::EOpConvInt64ToUint16: 5853 case glslang::EOpConvUint8ToInt16: 5854 case glslang::EOpConvUintToInt16: 5855 case glslang::EOpConvUint64ToInt16: 5856 zero = builder.makeUint16Constant(0); 5857 break; 5858 case glslang::EOpConvInt8ToUint: 5859 case glslang::EOpConvInt16ToUint: 5860 case glslang::EOpConvInt64ToUint: 5861 case glslang::EOpConvUint8ToInt: 5862 case glslang::EOpConvUint16ToInt: 5863 case glslang::EOpConvUint64ToInt: 5864 zero = builder.makeUintConstant(0); 5865 break; 5866 case glslang::EOpConvInt8ToUint64: 5867 case glslang::EOpConvInt16ToUint64: 5868 case glslang::EOpConvIntToUint64: 5869 case glslang::EOpConvUint8ToInt64: 5870 case glslang::EOpConvUint16ToInt64: 5871 case glslang::EOpConvUintToInt64: 5872 zero = builder.makeUint64Constant(0); 5873 break; 5874 default: 5875 assert(false && "Default missing"); 5876 break; 5877 } 5878 zero = makeSmearedConstant(zero, vectorSize); 5879 // Use OpIAdd, instead of OpBitcast to do the conversion when 5880 // generating for OpSpecConstantOp instruction. 5881 return builder.createBinOp(spv::OpIAdd, destType, operand, zero); 5882 } 5883 // For normal run-time conversion instruction, use OpBitcast. 5884 convOp = spv::OpBitcast; 5885 break; 5886 case glslang::EOpConvUint64ToPtr: 5887 convOp = spv::OpConvertUToPtr; 5888 break; 5889 case glslang::EOpConvPtrToUint64: 5890 convOp = spv::OpConvertPtrToU; 5891 break; 5892 default: 5893 break; 5894 } 5895 5896 spv::Id result = 0; 5897 if (convOp == spv::OpNop) 5898 return result; 5899 5900 if (convOp == spv::OpSelect) { 5901 zero = makeSmearedConstant(zero, vectorSize); 5902 one = makeSmearedConstant(one, vectorSize); 5903 result = builder.createTriOp(convOp, destType, operand, one, zero); 5904 } else 5905 result = builder.createUnaryOp(convOp, destType, operand); 5906 5907 result = builder.setPrecision(result, decorations.precision); 5908 builder.addDecoration(result, decorations.nonUniform); 5909 return result; 5910 } 5911 5912 spv::Id TGlslangToSpvTraverser::makeSmearedConstant(spv::Id constant, int vectorSize) 5913 { 5914 if (vectorSize == 0) 5915 return constant; 5916 5917 spv::Id vectorTypeId = builder.makeVectorType(builder.getTypeId(constant), vectorSize); 5918 std::vector<spv::Id> components; 5919 for (int c = 0; c < vectorSize; ++c) 5920 components.push_back(constant); 5921 return builder.makeCompositeConstant(vectorTypeId, components); 5922 } 5923 5924 // For glslang ops that map to SPV atomic opCodes 5925 spv::Id TGlslangToSpvTraverser::createAtomicOperation(glslang::TOperator op, spv::Decoration /*precision*/, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy) 5926 { 5927 spv::Op opCode = spv::OpNop; 5928 5929 switch (op) { 5930 case glslang::EOpAtomicAdd: 5931 case glslang::EOpImageAtomicAdd: 5932 case glslang::EOpAtomicCounterAdd: 5933 opCode = spv::OpAtomicIAdd; 5934 break; 5935 case glslang::EOpAtomicCounterSubtract: 5936 opCode = spv::OpAtomicISub; 5937 break; 5938 case glslang::EOpAtomicMin: 5939 case glslang::EOpImageAtomicMin: 5940 case glslang::EOpAtomicCounterMin: 5941 opCode = (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) ? spv::OpAtomicUMin : spv::OpAtomicSMin; 5942 break; 5943 case glslang::EOpAtomicMax: 5944 case glslang::EOpImageAtomicMax: 5945 case glslang::EOpAtomicCounterMax: 5946 opCode = (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) ? spv::OpAtomicUMax : spv::OpAtomicSMax; 5947 break; 5948 case glslang::EOpAtomicAnd: 5949 case glslang::EOpImageAtomicAnd: 5950 case glslang::EOpAtomicCounterAnd: 5951 opCode = spv::OpAtomicAnd; 5952 break; 5953 case glslang::EOpAtomicOr: 5954 case glslang::EOpImageAtomicOr: 5955 case glslang::EOpAtomicCounterOr: 5956 opCode = spv::OpAtomicOr; 5957 break; 5958 case glslang::EOpAtomicXor: 5959 case glslang::EOpImageAtomicXor: 5960 case glslang::EOpAtomicCounterXor: 5961 opCode = spv::OpAtomicXor; 5962 break; 5963 case glslang::EOpAtomicExchange: 5964 case glslang::EOpImageAtomicExchange: 5965 case glslang::EOpAtomicCounterExchange: 5966 opCode = spv::OpAtomicExchange; 5967 break; 5968 case glslang::EOpAtomicCompSwap: 5969 case glslang::EOpImageAtomicCompSwap: 5970 case glslang::EOpAtomicCounterCompSwap: 5971 opCode = spv::OpAtomicCompareExchange; 5972 break; 5973 case glslang::EOpAtomicCounterIncrement: 5974 opCode = spv::OpAtomicIIncrement; 5975 break; 5976 case glslang::EOpAtomicCounterDecrement: 5977 opCode = spv::OpAtomicIDecrement; 5978 break; 5979 case glslang::EOpAtomicCounter: 5980 case glslang::EOpImageAtomicLoad: 5981 case glslang::EOpAtomicLoad: 5982 opCode = spv::OpAtomicLoad; 5983 break; 5984 case glslang::EOpAtomicStore: 5985 case glslang::EOpImageAtomicStore: 5986 opCode = spv::OpAtomicStore; 5987 break; 5988 default: 5989 assert(0); 5990 break; 5991 } 5992 5993 if (typeProxy == glslang::EbtInt64 || typeProxy == glslang::EbtUint64) 5994 builder.addCapability(spv::CapabilityInt64Atomics); 5995 5996 // Sort out the operands 5997 // - mapping from glslang -> SPV 5998 // - there are extra SPV operands that are optional in glslang 5999 // - compare-exchange swaps the value and comparator 6000 // - compare-exchange has an extra memory semantics 6001 // - EOpAtomicCounterDecrement needs a post decrement 6002 spv::Id pointerId = 0, compareId = 0, valueId = 0; 6003 // scope defaults to Device in the old model, QueueFamilyKHR in the new model 6004 spv::Id scopeId; 6005 if (glslangIntermediate->usingVulkanMemoryModel()) { 6006 scopeId = builder.makeUintConstant(spv::ScopeQueueFamilyKHR); 6007 } else { 6008 scopeId = builder.makeUintConstant(spv::ScopeDevice); 6009 } 6010 // semantics default to relaxed 6011 spv::Id semanticsId = builder.makeUintConstant(spv::MemorySemanticsMaskNone); 6012 spv::Id semanticsId2 = semanticsId; 6013 6014 pointerId = operands[0]; 6015 if (opCode == spv::OpAtomicIIncrement || opCode == spv::OpAtomicIDecrement) { 6016 // no additional operands 6017 } else if (opCode == spv::OpAtomicCompareExchange) { 6018 compareId = operands[1]; 6019 valueId = operands[2]; 6020 if (operands.size() > 3) { 6021 scopeId = operands[3]; 6022 semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[4]) | builder.getConstantScalar(operands[5])); 6023 semanticsId2 = builder.makeUintConstant(builder.getConstantScalar(operands[6]) | builder.getConstantScalar(operands[7])); 6024 } 6025 } else if (opCode == spv::OpAtomicLoad) { 6026 if (operands.size() > 1) { 6027 scopeId = operands[1]; 6028 semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[2]) | builder.getConstantScalar(operands[3])); 6029 } 6030 } else { 6031 // atomic store or RMW 6032 valueId = operands[1]; 6033 if (operands.size() > 2) { 6034 scopeId = operands[2]; 6035 semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[3]) | builder.getConstantScalar(operands[4])); 6036 } 6037 } 6038 6039 // Check for capabilities 6040 unsigned semanticsImmediate = builder.getConstantScalar(semanticsId) | builder.getConstantScalar(semanticsId2); 6041 if (semanticsImmediate & (spv::MemorySemanticsMakeAvailableKHRMask | spv::MemorySemanticsMakeVisibleKHRMask | spv::MemorySemanticsOutputMemoryKHRMask)) { 6042 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); 6043 } 6044 6045 if (glslangIntermediate->usingVulkanMemoryModel() && builder.getConstantScalar(scopeId) == spv::ScopeDevice) { 6046 builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR); 6047 } 6048 6049 std::vector<spv::Id> spvAtomicOperands; // hold the spv operands 6050 spvAtomicOperands.push_back(pointerId); 6051 spvAtomicOperands.push_back(scopeId); 6052 spvAtomicOperands.push_back(semanticsId); 6053 if (opCode == spv::OpAtomicCompareExchange) { 6054 spvAtomicOperands.push_back(semanticsId2); 6055 spvAtomicOperands.push_back(valueId); 6056 spvAtomicOperands.push_back(compareId); 6057 } else if (opCode != spv::OpAtomicLoad && opCode != spv::OpAtomicIIncrement && opCode != spv::OpAtomicIDecrement) { 6058 spvAtomicOperands.push_back(valueId); 6059 } 6060 6061 if (opCode == spv::OpAtomicStore) { 6062 builder.createNoResultOp(opCode, spvAtomicOperands); 6063 return 0; 6064 } else { 6065 spv::Id resultId = builder.createOp(opCode, typeId, spvAtomicOperands); 6066 6067 // GLSL and HLSL atomic-counter decrement return post-decrement value, 6068 // while SPIR-V returns pre-decrement value. Translate between these semantics. 6069 if (op == glslang::EOpAtomicCounterDecrement) 6070 resultId = builder.createBinOp(spv::OpISub, typeId, resultId, builder.makeIntConstant(1)); 6071 6072 return resultId; 6073 } 6074 } 6075 6076 // Create group invocation operations. 6077 spv::Id TGlslangToSpvTraverser::createInvocationsOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy) 6078 { 6079 #ifdef AMD_EXTENSIONS 6080 bool isUnsigned = isTypeUnsignedInt(typeProxy); 6081 bool isFloat = isTypeFloat(typeProxy); 6082 #endif 6083 6084 spv::Op opCode = spv::OpNop; 6085 std::vector<spv::IdImmediate> spvGroupOperands; 6086 spv::GroupOperation groupOperation = spv::GroupOperationMax; 6087 6088 if (op == glslang::EOpBallot || op == glslang::EOpReadFirstInvocation || 6089 op == glslang::EOpReadInvocation) { 6090 builder.addExtension(spv::E_SPV_KHR_shader_ballot); 6091 builder.addCapability(spv::CapabilitySubgroupBallotKHR); 6092 } else if (op == glslang::EOpAnyInvocation || 6093 op == glslang::EOpAllInvocations || 6094 op == glslang::EOpAllInvocationsEqual) { 6095 builder.addExtension(spv::E_SPV_KHR_subgroup_vote); 6096 builder.addCapability(spv::CapabilitySubgroupVoteKHR); 6097 } else { 6098 builder.addCapability(spv::CapabilityGroups); 6099 #ifdef AMD_EXTENSIONS 6100 if (op == glslang::EOpMinInvocationsNonUniform || 6101 op == glslang::EOpMaxInvocationsNonUniform || 6102 op == glslang::EOpAddInvocationsNonUniform || 6103 op == glslang::EOpMinInvocationsInclusiveScanNonUniform || 6104 op == glslang::EOpMaxInvocationsInclusiveScanNonUniform || 6105 op == glslang::EOpAddInvocationsInclusiveScanNonUniform || 6106 op == glslang::EOpMinInvocationsExclusiveScanNonUniform || 6107 op == glslang::EOpMaxInvocationsExclusiveScanNonUniform || 6108 op == glslang::EOpAddInvocationsExclusiveScanNonUniform) 6109 builder.addExtension(spv::E_SPV_AMD_shader_ballot); 6110 #endif 6111 6112 #ifdef AMD_EXTENSIONS 6113 switch (op) { 6114 case glslang::EOpMinInvocations: 6115 case glslang::EOpMaxInvocations: 6116 case glslang::EOpAddInvocations: 6117 case glslang::EOpMinInvocationsNonUniform: 6118 case glslang::EOpMaxInvocationsNonUniform: 6119 case glslang::EOpAddInvocationsNonUniform: 6120 groupOperation = spv::GroupOperationReduce; 6121 break; 6122 case glslang::EOpMinInvocationsInclusiveScan: 6123 case glslang::EOpMaxInvocationsInclusiveScan: 6124 case glslang::EOpAddInvocationsInclusiveScan: 6125 case glslang::EOpMinInvocationsInclusiveScanNonUniform: 6126 case glslang::EOpMaxInvocationsInclusiveScanNonUniform: 6127 case glslang::EOpAddInvocationsInclusiveScanNonUniform: 6128 groupOperation = spv::GroupOperationInclusiveScan; 6129 break; 6130 case glslang::EOpMinInvocationsExclusiveScan: 6131 case glslang::EOpMaxInvocationsExclusiveScan: 6132 case glslang::EOpAddInvocationsExclusiveScan: 6133 case glslang::EOpMinInvocationsExclusiveScanNonUniform: 6134 case glslang::EOpMaxInvocationsExclusiveScanNonUniform: 6135 case glslang::EOpAddInvocationsExclusiveScanNonUniform: 6136 groupOperation = spv::GroupOperationExclusiveScan; 6137 break; 6138 default: 6139 break; 6140 } 6141 spv::IdImmediate scope = { true, builder.makeUintConstant(spv::ScopeSubgroup) }; 6142 spvGroupOperands.push_back(scope); 6143 if (groupOperation != spv::GroupOperationMax) { 6144 spv::IdImmediate groupOp = { false, (unsigned)groupOperation }; 6145 spvGroupOperands.push_back(groupOp); 6146 } 6147 #endif 6148 } 6149 6150 for (auto opIt = operands.begin(); opIt != operands.end(); ++opIt) { 6151 spv::IdImmediate op = { true, *opIt }; 6152 spvGroupOperands.push_back(op); 6153 } 6154 6155 switch (op) { 6156 case glslang::EOpAnyInvocation: 6157 opCode = spv::OpSubgroupAnyKHR; 6158 break; 6159 case glslang::EOpAllInvocations: 6160 opCode = spv::OpSubgroupAllKHR; 6161 break; 6162 case glslang::EOpAllInvocationsEqual: 6163 opCode = spv::OpSubgroupAllEqualKHR; 6164 break; 6165 case glslang::EOpReadInvocation: 6166 opCode = spv::OpSubgroupReadInvocationKHR; 6167 if (builder.isVectorType(typeId)) 6168 return CreateInvocationsVectorOperation(opCode, groupOperation, typeId, operands); 6169 break; 6170 case glslang::EOpReadFirstInvocation: 6171 opCode = spv::OpSubgroupFirstInvocationKHR; 6172 break; 6173 case glslang::EOpBallot: 6174 { 6175 // NOTE: According to the spec, the result type of "OpSubgroupBallotKHR" must be a 4 component vector of 32 6176 // bit integer types. The GLSL built-in function "ballotARB()" assumes the maximum number of invocations in 6177 // a subgroup is 64. Thus, we have to convert uvec4.xy to uint64_t as follow: 6178 // 6179 // result = Bitcast(SubgroupBallotKHR(Predicate).xy) 6180 // 6181 spv::Id uintType = builder.makeUintType(32); 6182 spv::Id uvec4Type = builder.makeVectorType(uintType, 4); 6183 spv::Id result = builder.createOp(spv::OpSubgroupBallotKHR, uvec4Type, spvGroupOperands); 6184 6185 std::vector<spv::Id> components; 6186 components.push_back(builder.createCompositeExtract(result, uintType, 0)); 6187 components.push_back(builder.createCompositeExtract(result, uintType, 1)); 6188 6189 spv::Id uvec2Type = builder.makeVectorType(uintType, 2); 6190 return builder.createUnaryOp(spv::OpBitcast, typeId, 6191 builder.createCompositeConstruct(uvec2Type, components)); 6192 } 6193 6194 #ifdef AMD_EXTENSIONS 6195 case glslang::EOpMinInvocations: 6196 case glslang::EOpMaxInvocations: 6197 case glslang::EOpAddInvocations: 6198 case glslang::EOpMinInvocationsInclusiveScan: 6199 case glslang::EOpMaxInvocationsInclusiveScan: 6200 case glslang::EOpAddInvocationsInclusiveScan: 6201 case glslang::EOpMinInvocationsExclusiveScan: 6202 case glslang::EOpMaxInvocationsExclusiveScan: 6203 case glslang::EOpAddInvocationsExclusiveScan: 6204 if (op == glslang::EOpMinInvocations || 6205 op == glslang::EOpMinInvocationsInclusiveScan || 6206 op == glslang::EOpMinInvocationsExclusiveScan) { 6207 if (isFloat) 6208 opCode = spv::OpGroupFMin; 6209 else { 6210 if (isUnsigned) 6211 opCode = spv::OpGroupUMin; 6212 else 6213 opCode = spv::OpGroupSMin; 6214 } 6215 } else if (op == glslang::EOpMaxInvocations || 6216 op == glslang::EOpMaxInvocationsInclusiveScan || 6217 op == glslang::EOpMaxInvocationsExclusiveScan) { 6218 if (isFloat) 6219 opCode = spv::OpGroupFMax; 6220 else { 6221 if (isUnsigned) 6222 opCode = spv::OpGroupUMax; 6223 else 6224 opCode = spv::OpGroupSMax; 6225 } 6226 } else { 6227 if (isFloat) 6228 opCode = spv::OpGroupFAdd; 6229 else 6230 opCode = spv::OpGroupIAdd; 6231 } 6232 6233 if (builder.isVectorType(typeId)) 6234 return CreateInvocationsVectorOperation(opCode, groupOperation, typeId, operands); 6235 6236 break; 6237 case glslang::EOpMinInvocationsNonUniform: 6238 case glslang::EOpMaxInvocationsNonUniform: 6239 case glslang::EOpAddInvocationsNonUniform: 6240 case glslang::EOpMinInvocationsInclusiveScanNonUniform: 6241 case glslang::EOpMaxInvocationsInclusiveScanNonUniform: 6242 case glslang::EOpAddInvocationsInclusiveScanNonUniform: 6243 case glslang::EOpMinInvocationsExclusiveScanNonUniform: 6244 case glslang::EOpMaxInvocationsExclusiveScanNonUniform: 6245 case glslang::EOpAddInvocationsExclusiveScanNonUniform: 6246 if (op == glslang::EOpMinInvocationsNonUniform || 6247 op == glslang::EOpMinInvocationsInclusiveScanNonUniform || 6248 op == glslang::EOpMinInvocationsExclusiveScanNonUniform) { 6249 if (isFloat) 6250 opCode = spv::OpGroupFMinNonUniformAMD; 6251 else { 6252 if (isUnsigned) 6253 opCode = spv::OpGroupUMinNonUniformAMD; 6254 else 6255 opCode = spv::OpGroupSMinNonUniformAMD; 6256 } 6257 } 6258 else if (op == glslang::EOpMaxInvocationsNonUniform || 6259 op == glslang::EOpMaxInvocationsInclusiveScanNonUniform || 6260 op == glslang::EOpMaxInvocationsExclusiveScanNonUniform) { 6261 if (isFloat) 6262 opCode = spv::OpGroupFMaxNonUniformAMD; 6263 else { 6264 if (isUnsigned) 6265 opCode = spv::OpGroupUMaxNonUniformAMD; 6266 else 6267 opCode = spv::OpGroupSMaxNonUniformAMD; 6268 } 6269 } 6270 else { 6271 if (isFloat) 6272 opCode = spv::OpGroupFAddNonUniformAMD; 6273 else 6274 opCode = spv::OpGroupIAddNonUniformAMD; 6275 } 6276 6277 if (builder.isVectorType(typeId)) 6278 return CreateInvocationsVectorOperation(opCode, groupOperation, typeId, operands); 6279 6280 break; 6281 #endif 6282 default: 6283 logger->missingFunctionality("invocation operation"); 6284 return spv::NoResult; 6285 } 6286 6287 assert(opCode != spv::OpNop); 6288 return builder.createOp(opCode, typeId, spvGroupOperands); 6289 } 6290 6291 // Create group invocation operations on a vector 6292 spv::Id TGlslangToSpvTraverser::CreateInvocationsVectorOperation(spv::Op op, spv::GroupOperation groupOperation, 6293 spv::Id typeId, std::vector<spv::Id>& operands) 6294 { 6295 #ifdef AMD_EXTENSIONS 6296 assert(op == spv::OpGroupFMin || op == spv::OpGroupUMin || op == spv::OpGroupSMin || 6297 op == spv::OpGroupFMax || op == spv::OpGroupUMax || op == spv::OpGroupSMax || 6298 op == spv::OpGroupFAdd || op == spv::OpGroupIAdd || op == spv::OpGroupBroadcast || 6299 op == spv::OpSubgroupReadInvocationKHR || 6300 op == spv::OpGroupFMinNonUniformAMD || op == spv::OpGroupUMinNonUniformAMD || op == spv::OpGroupSMinNonUniformAMD || 6301 op == spv::OpGroupFMaxNonUniformAMD || op == spv::OpGroupUMaxNonUniformAMD || op == spv::OpGroupSMaxNonUniformAMD || 6302 op == spv::OpGroupFAddNonUniformAMD || op == spv::OpGroupIAddNonUniformAMD); 6303 #else 6304 assert(op == spv::OpGroupFMin || op == spv::OpGroupUMin || op == spv::OpGroupSMin || 6305 op == spv::OpGroupFMax || op == spv::OpGroupUMax || op == spv::OpGroupSMax || 6306 op == spv::OpGroupFAdd || op == spv::OpGroupIAdd || op == spv::OpGroupBroadcast || 6307 op == spv::OpSubgroupReadInvocationKHR); 6308 #endif 6309 6310 // Handle group invocation operations scalar by scalar. 6311 // The result type is the same type as the original type. 6312 // The algorithm is to: 6313 // - break the vector into scalars 6314 // - apply the operation to each scalar 6315 // - make a vector out the scalar results 6316 6317 // get the types sorted out 6318 int numComponents = builder.getNumComponents(operands[0]); 6319 spv::Id scalarType = builder.getScalarTypeId(builder.getTypeId(operands[0])); 6320 std::vector<spv::Id> results; 6321 6322 // do each scalar op 6323 for (int comp = 0; comp < numComponents; ++comp) { 6324 std::vector<unsigned int> indexes; 6325 indexes.push_back(comp); 6326 spv::IdImmediate scalar = { true, builder.createCompositeExtract(operands[0], scalarType, indexes) }; 6327 std::vector<spv::IdImmediate> spvGroupOperands; 6328 if (op == spv::OpSubgroupReadInvocationKHR) { 6329 spvGroupOperands.push_back(scalar); 6330 spv::IdImmediate operand = { true, operands[1] }; 6331 spvGroupOperands.push_back(operand); 6332 } else if (op == spv::OpGroupBroadcast) { 6333 spv::IdImmediate scope = { true, builder.makeUintConstant(spv::ScopeSubgroup) }; 6334 spvGroupOperands.push_back(scope); 6335 spvGroupOperands.push_back(scalar); 6336 spv::IdImmediate operand = { true, operands[1] }; 6337 spvGroupOperands.push_back(operand); 6338 } else { 6339 spv::IdImmediate scope = { true, builder.makeUintConstant(spv::ScopeSubgroup) }; 6340 spvGroupOperands.push_back(scope); 6341 spv::IdImmediate groupOp = { false, (unsigned)groupOperation }; 6342 spvGroupOperands.push_back(groupOp); 6343 spvGroupOperands.push_back(scalar); 6344 } 6345 6346 results.push_back(builder.createOp(op, scalarType, spvGroupOperands)); 6347 } 6348 6349 // put the pieces together 6350 return builder.createCompositeConstruct(typeId, results); 6351 } 6352 6353 // Create subgroup invocation operations. 6354 spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, spv::Id typeId, 6355 std::vector<spv::Id>& operands, glslang::TBasicType typeProxy) 6356 { 6357 // Add the required capabilities. 6358 switch (op) { 6359 case glslang::EOpSubgroupElect: 6360 builder.addCapability(spv::CapabilityGroupNonUniform); 6361 break; 6362 case glslang::EOpSubgroupAll: 6363 case glslang::EOpSubgroupAny: 6364 case glslang::EOpSubgroupAllEqual: 6365 builder.addCapability(spv::CapabilityGroupNonUniform); 6366 builder.addCapability(spv::CapabilityGroupNonUniformVote); 6367 break; 6368 case glslang::EOpSubgroupBroadcast: 6369 case glslang::EOpSubgroupBroadcastFirst: 6370 case glslang::EOpSubgroupBallot: 6371 case glslang::EOpSubgroupInverseBallot: 6372 case glslang::EOpSubgroupBallotBitExtract: 6373 case glslang::EOpSubgroupBallotBitCount: 6374 case glslang::EOpSubgroupBallotInclusiveBitCount: 6375 case glslang::EOpSubgroupBallotExclusiveBitCount: 6376 case glslang::EOpSubgroupBallotFindLSB: 6377 case glslang::EOpSubgroupBallotFindMSB: 6378 builder.addCapability(spv::CapabilityGroupNonUniform); 6379 builder.addCapability(spv::CapabilityGroupNonUniformBallot); 6380 break; 6381 case glslang::EOpSubgroupShuffle: 6382 case glslang::EOpSubgroupShuffleXor: 6383 builder.addCapability(spv::CapabilityGroupNonUniform); 6384 builder.addCapability(spv::CapabilityGroupNonUniformShuffle); 6385 break; 6386 case glslang::EOpSubgroupShuffleUp: 6387 case glslang::EOpSubgroupShuffleDown: 6388 builder.addCapability(spv::CapabilityGroupNonUniform); 6389 builder.addCapability(spv::CapabilityGroupNonUniformShuffleRelative); 6390 break; 6391 case glslang::EOpSubgroupAdd: 6392 case glslang::EOpSubgroupMul: 6393 case glslang::EOpSubgroupMin: 6394 case glslang::EOpSubgroupMax: 6395 case glslang::EOpSubgroupAnd: 6396 case glslang::EOpSubgroupOr: 6397 case glslang::EOpSubgroupXor: 6398 case glslang::EOpSubgroupInclusiveAdd: 6399 case glslang::EOpSubgroupInclusiveMul: 6400 case glslang::EOpSubgroupInclusiveMin: 6401 case glslang::EOpSubgroupInclusiveMax: 6402 case glslang::EOpSubgroupInclusiveAnd: 6403 case glslang::EOpSubgroupInclusiveOr: 6404 case glslang::EOpSubgroupInclusiveXor: 6405 case glslang::EOpSubgroupExclusiveAdd: 6406 case glslang::EOpSubgroupExclusiveMul: 6407 case glslang::EOpSubgroupExclusiveMin: 6408 case glslang::EOpSubgroupExclusiveMax: 6409 case glslang::EOpSubgroupExclusiveAnd: 6410 case glslang::EOpSubgroupExclusiveOr: 6411 case glslang::EOpSubgroupExclusiveXor: 6412 builder.addCapability(spv::CapabilityGroupNonUniform); 6413 builder.addCapability(spv::CapabilityGroupNonUniformArithmetic); 6414 break; 6415 case glslang::EOpSubgroupClusteredAdd: 6416 case glslang::EOpSubgroupClusteredMul: 6417 case glslang::EOpSubgroupClusteredMin: 6418 case glslang::EOpSubgroupClusteredMax: 6419 case glslang::EOpSubgroupClusteredAnd: 6420 case glslang::EOpSubgroupClusteredOr: 6421 case glslang::EOpSubgroupClusteredXor: 6422 builder.addCapability(spv::CapabilityGroupNonUniform); 6423 builder.addCapability(spv::CapabilityGroupNonUniformClustered); 6424 break; 6425 case glslang::EOpSubgroupQuadBroadcast: 6426 case glslang::EOpSubgroupQuadSwapHorizontal: 6427 case glslang::EOpSubgroupQuadSwapVertical: 6428 case glslang::EOpSubgroupQuadSwapDiagonal: 6429 builder.addCapability(spv::CapabilityGroupNonUniform); 6430 builder.addCapability(spv::CapabilityGroupNonUniformQuad); 6431 break; 6432 #ifdef NV_EXTENSIONS 6433 case glslang::EOpSubgroupPartitionedAdd: 6434 case glslang::EOpSubgroupPartitionedMul: 6435 case glslang::EOpSubgroupPartitionedMin: 6436 case glslang::EOpSubgroupPartitionedMax: 6437 case glslang::EOpSubgroupPartitionedAnd: 6438 case glslang::EOpSubgroupPartitionedOr: 6439 case glslang::EOpSubgroupPartitionedXor: 6440 case glslang::EOpSubgroupPartitionedInclusiveAdd: 6441 case glslang::EOpSubgroupPartitionedInclusiveMul: 6442 case glslang::EOpSubgroupPartitionedInclusiveMin: 6443 case glslang::EOpSubgroupPartitionedInclusiveMax: 6444 case glslang::EOpSubgroupPartitionedInclusiveAnd: 6445 case glslang::EOpSubgroupPartitionedInclusiveOr: 6446 case glslang::EOpSubgroupPartitionedInclusiveXor: 6447 case glslang::EOpSubgroupPartitionedExclusiveAdd: 6448 case glslang::EOpSubgroupPartitionedExclusiveMul: 6449 case glslang::EOpSubgroupPartitionedExclusiveMin: 6450 case glslang::EOpSubgroupPartitionedExclusiveMax: 6451 case glslang::EOpSubgroupPartitionedExclusiveAnd: 6452 case glslang::EOpSubgroupPartitionedExclusiveOr: 6453 case glslang::EOpSubgroupPartitionedExclusiveXor: 6454 builder.addExtension(spv::E_SPV_NV_shader_subgroup_partitioned); 6455 builder.addCapability(spv::CapabilityGroupNonUniformPartitionedNV); 6456 break; 6457 #endif 6458 default: assert(0 && "Unhandled subgroup operation!"); 6459 } 6460 6461 const bool isUnsigned = typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64; 6462 const bool isFloat = typeProxy == glslang::EbtFloat || typeProxy == glslang::EbtDouble; 6463 const bool isBool = typeProxy == glslang::EbtBool; 6464 6465 spv::Op opCode = spv::OpNop; 6466 6467 // Figure out which opcode to use. 6468 switch (op) { 6469 case glslang::EOpSubgroupElect: opCode = spv::OpGroupNonUniformElect; break; 6470 case glslang::EOpSubgroupAll: opCode = spv::OpGroupNonUniformAll; break; 6471 case glslang::EOpSubgroupAny: opCode = spv::OpGroupNonUniformAny; break; 6472 case glslang::EOpSubgroupAllEqual: opCode = spv::OpGroupNonUniformAllEqual; break; 6473 case glslang::EOpSubgroupBroadcast: opCode = spv::OpGroupNonUniformBroadcast; break; 6474 case glslang::EOpSubgroupBroadcastFirst: opCode = spv::OpGroupNonUniformBroadcastFirst; break; 6475 case glslang::EOpSubgroupBallot: opCode = spv::OpGroupNonUniformBallot; break; 6476 case glslang::EOpSubgroupInverseBallot: opCode = spv::OpGroupNonUniformInverseBallot; break; 6477 case glslang::EOpSubgroupBallotBitExtract: opCode = spv::OpGroupNonUniformBallotBitExtract; break; 6478 case glslang::EOpSubgroupBallotBitCount: 6479 case glslang::EOpSubgroupBallotInclusiveBitCount: 6480 case glslang::EOpSubgroupBallotExclusiveBitCount: opCode = spv::OpGroupNonUniformBallotBitCount; break; 6481 case glslang::EOpSubgroupBallotFindLSB: opCode = spv::OpGroupNonUniformBallotFindLSB; break; 6482 case glslang::EOpSubgroupBallotFindMSB: opCode = spv::OpGroupNonUniformBallotFindMSB; break; 6483 case glslang::EOpSubgroupShuffle: opCode = spv::OpGroupNonUniformShuffle; break; 6484 case glslang::EOpSubgroupShuffleXor: opCode = spv::OpGroupNonUniformShuffleXor; break; 6485 case glslang::EOpSubgroupShuffleUp: opCode = spv::OpGroupNonUniformShuffleUp; break; 6486 case glslang::EOpSubgroupShuffleDown: opCode = spv::OpGroupNonUniformShuffleDown; break; 6487 case glslang::EOpSubgroupAdd: 6488 case glslang::EOpSubgroupInclusiveAdd: 6489 case glslang::EOpSubgroupExclusiveAdd: 6490 case glslang::EOpSubgroupClusteredAdd: 6491 #ifdef NV_EXTENSIONS 6492 case glslang::EOpSubgroupPartitionedAdd: 6493 case glslang::EOpSubgroupPartitionedInclusiveAdd: 6494 case glslang::EOpSubgroupPartitionedExclusiveAdd: 6495 #endif 6496 if (isFloat) { 6497 opCode = spv::OpGroupNonUniformFAdd; 6498 } else { 6499 opCode = spv::OpGroupNonUniformIAdd; 6500 } 6501 break; 6502 case glslang::EOpSubgroupMul: 6503 case glslang::EOpSubgroupInclusiveMul: 6504 case glslang::EOpSubgroupExclusiveMul: 6505 case glslang::EOpSubgroupClusteredMul: 6506 #ifdef NV_EXTENSIONS 6507 case glslang::EOpSubgroupPartitionedMul: 6508 case glslang::EOpSubgroupPartitionedInclusiveMul: 6509 case glslang::EOpSubgroupPartitionedExclusiveMul: 6510 #endif 6511 if (isFloat) { 6512 opCode = spv::OpGroupNonUniformFMul; 6513 } else { 6514 opCode = spv::OpGroupNonUniformIMul; 6515 } 6516 break; 6517 case glslang::EOpSubgroupMin: 6518 case glslang::EOpSubgroupInclusiveMin: 6519 case glslang::EOpSubgroupExclusiveMin: 6520 case glslang::EOpSubgroupClusteredMin: 6521 #ifdef NV_EXTENSIONS 6522 case glslang::EOpSubgroupPartitionedMin: 6523 case glslang::EOpSubgroupPartitionedInclusiveMin: 6524 case glslang::EOpSubgroupPartitionedExclusiveMin: 6525 #endif 6526 if (isFloat) { 6527 opCode = spv::OpGroupNonUniformFMin; 6528 } else if (isUnsigned) { 6529 opCode = spv::OpGroupNonUniformUMin; 6530 } else { 6531 opCode = spv::OpGroupNonUniformSMin; 6532 } 6533 break; 6534 case glslang::EOpSubgroupMax: 6535 case glslang::EOpSubgroupInclusiveMax: 6536 case glslang::EOpSubgroupExclusiveMax: 6537 case glslang::EOpSubgroupClusteredMax: 6538 #ifdef NV_EXTENSIONS 6539 case glslang::EOpSubgroupPartitionedMax: 6540 case glslang::EOpSubgroupPartitionedInclusiveMax: 6541 case glslang::EOpSubgroupPartitionedExclusiveMax: 6542 #endif 6543 if (isFloat) { 6544 opCode = spv::OpGroupNonUniformFMax; 6545 } else if (isUnsigned) { 6546 opCode = spv::OpGroupNonUniformUMax; 6547 } else { 6548 opCode = spv::OpGroupNonUniformSMax; 6549 } 6550 break; 6551 case glslang::EOpSubgroupAnd: 6552 case glslang::EOpSubgroupInclusiveAnd: 6553 case glslang::EOpSubgroupExclusiveAnd: 6554 case glslang::EOpSubgroupClusteredAnd: 6555 #ifdef NV_EXTENSIONS 6556 case glslang::EOpSubgroupPartitionedAnd: 6557 case glslang::EOpSubgroupPartitionedInclusiveAnd: 6558 case glslang::EOpSubgroupPartitionedExclusiveAnd: 6559 #endif 6560 if (isBool) { 6561 opCode = spv::OpGroupNonUniformLogicalAnd; 6562 } else { 6563 opCode = spv::OpGroupNonUniformBitwiseAnd; 6564 } 6565 break; 6566 case glslang::EOpSubgroupOr: 6567 case glslang::EOpSubgroupInclusiveOr: 6568 case glslang::EOpSubgroupExclusiveOr: 6569 case glslang::EOpSubgroupClusteredOr: 6570 #ifdef NV_EXTENSIONS 6571 case glslang::EOpSubgroupPartitionedOr: 6572 case glslang::EOpSubgroupPartitionedInclusiveOr: 6573 case glslang::EOpSubgroupPartitionedExclusiveOr: 6574 #endif 6575 if (isBool) { 6576 opCode = spv::OpGroupNonUniformLogicalOr; 6577 } else { 6578 opCode = spv::OpGroupNonUniformBitwiseOr; 6579 } 6580 break; 6581 case glslang::EOpSubgroupXor: 6582 case glslang::EOpSubgroupInclusiveXor: 6583 case glslang::EOpSubgroupExclusiveXor: 6584 case glslang::EOpSubgroupClusteredXor: 6585 #ifdef NV_EXTENSIONS 6586 case glslang::EOpSubgroupPartitionedXor: 6587 case glslang::EOpSubgroupPartitionedInclusiveXor: 6588 case glslang::EOpSubgroupPartitionedExclusiveXor: 6589 #endif 6590 if (isBool) { 6591 opCode = spv::OpGroupNonUniformLogicalXor; 6592 } else { 6593 opCode = spv::OpGroupNonUniformBitwiseXor; 6594 } 6595 break; 6596 case glslang::EOpSubgroupQuadBroadcast: opCode = spv::OpGroupNonUniformQuadBroadcast; break; 6597 case glslang::EOpSubgroupQuadSwapHorizontal: 6598 case glslang::EOpSubgroupQuadSwapVertical: 6599 case glslang::EOpSubgroupQuadSwapDiagonal: opCode = spv::OpGroupNonUniformQuadSwap; break; 6600 default: assert(0 && "Unhandled subgroup operation!"); 6601 } 6602 6603 // get the right Group Operation 6604 spv::GroupOperation groupOperation = spv::GroupOperationMax; 6605 switch (op) { 6606 default: 6607 break; 6608 case glslang::EOpSubgroupBallotBitCount: 6609 case glslang::EOpSubgroupAdd: 6610 case glslang::EOpSubgroupMul: 6611 case glslang::EOpSubgroupMin: 6612 case glslang::EOpSubgroupMax: 6613 case glslang::EOpSubgroupAnd: 6614 case glslang::EOpSubgroupOr: 6615 case glslang::EOpSubgroupXor: 6616 groupOperation = spv::GroupOperationReduce; 6617 break; 6618 case glslang::EOpSubgroupBallotInclusiveBitCount: 6619 case glslang::EOpSubgroupInclusiveAdd: 6620 case glslang::EOpSubgroupInclusiveMul: 6621 case glslang::EOpSubgroupInclusiveMin: 6622 case glslang::EOpSubgroupInclusiveMax: 6623 case glslang::EOpSubgroupInclusiveAnd: 6624 case glslang::EOpSubgroupInclusiveOr: 6625 case glslang::EOpSubgroupInclusiveXor: 6626 groupOperation = spv::GroupOperationInclusiveScan; 6627 break; 6628 case glslang::EOpSubgroupBallotExclusiveBitCount: 6629 case glslang::EOpSubgroupExclusiveAdd: 6630 case glslang::EOpSubgroupExclusiveMul: 6631 case glslang::EOpSubgroupExclusiveMin: 6632 case glslang::EOpSubgroupExclusiveMax: 6633 case glslang::EOpSubgroupExclusiveAnd: 6634 case glslang::EOpSubgroupExclusiveOr: 6635 case glslang::EOpSubgroupExclusiveXor: 6636 groupOperation = spv::GroupOperationExclusiveScan; 6637 break; 6638 case glslang::EOpSubgroupClusteredAdd: 6639 case glslang::EOpSubgroupClusteredMul: 6640 case glslang::EOpSubgroupClusteredMin: 6641 case glslang::EOpSubgroupClusteredMax: 6642 case glslang::EOpSubgroupClusteredAnd: 6643 case glslang::EOpSubgroupClusteredOr: 6644 case glslang::EOpSubgroupClusteredXor: 6645 groupOperation = spv::GroupOperationClusteredReduce; 6646 break; 6647 #ifdef NV_EXTENSIONS 6648 case glslang::EOpSubgroupPartitionedAdd: 6649 case glslang::EOpSubgroupPartitionedMul: 6650 case glslang::EOpSubgroupPartitionedMin: 6651 case glslang::EOpSubgroupPartitionedMax: 6652 case glslang::EOpSubgroupPartitionedAnd: 6653 case glslang::EOpSubgroupPartitionedOr: 6654 case glslang::EOpSubgroupPartitionedXor: 6655 groupOperation = spv::GroupOperationPartitionedReduceNV; 6656 break; 6657 case glslang::EOpSubgroupPartitionedInclusiveAdd: 6658 case glslang::EOpSubgroupPartitionedInclusiveMul: 6659 case glslang::EOpSubgroupPartitionedInclusiveMin: 6660 case glslang::EOpSubgroupPartitionedInclusiveMax: 6661 case glslang::EOpSubgroupPartitionedInclusiveAnd: 6662 case glslang::EOpSubgroupPartitionedInclusiveOr: 6663 case glslang::EOpSubgroupPartitionedInclusiveXor: 6664 groupOperation = spv::GroupOperationPartitionedInclusiveScanNV; 6665 break; 6666 case glslang::EOpSubgroupPartitionedExclusiveAdd: 6667 case glslang::EOpSubgroupPartitionedExclusiveMul: 6668 case glslang::EOpSubgroupPartitionedExclusiveMin: 6669 case glslang::EOpSubgroupPartitionedExclusiveMax: 6670 case glslang::EOpSubgroupPartitionedExclusiveAnd: 6671 case glslang::EOpSubgroupPartitionedExclusiveOr: 6672 case glslang::EOpSubgroupPartitionedExclusiveXor: 6673 groupOperation = spv::GroupOperationPartitionedExclusiveScanNV; 6674 break; 6675 #endif 6676 } 6677 6678 // build the instruction 6679 std::vector<spv::IdImmediate> spvGroupOperands; 6680 6681 // Every operation begins with the Execution Scope operand. 6682 spv::IdImmediate executionScope = { true, builder.makeUintConstant(spv::ScopeSubgroup) }; 6683 spvGroupOperands.push_back(executionScope); 6684 6685 // Next, for all operations that use a Group Operation, push that as an operand. 6686 if (groupOperation != spv::GroupOperationMax) { 6687 spv::IdImmediate groupOperand = { false, (unsigned)groupOperation }; 6688 spvGroupOperands.push_back(groupOperand); 6689 } 6690 6691 // Push back the operands next. 6692 for (auto opIt = operands.cbegin(); opIt != operands.cend(); ++opIt) { 6693 spv::IdImmediate operand = { true, *opIt }; 6694 spvGroupOperands.push_back(operand); 6695 } 6696 6697 // Some opcodes have additional operands. 6698 spv::Id directionId = spv::NoResult; 6699 switch (op) { 6700 default: break; 6701 case glslang::EOpSubgroupQuadSwapHorizontal: directionId = builder.makeUintConstant(0); break; 6702 case glslang::EOpSubgroupQuadSwapVertical: directionId = builder.makeUintConstant(1); break; 6703 case glslang::EOpSubgroupQuadSwapDiagonal: directionId = builder.makeUintConstant(2); break; 6704 } 6705 if (directionId != spv::NoResult) { 6706 spv::IdImmediate direction = { true, directionId }; 6707 spvGroupOperands.push_back(direction); 6708 } 6709 6710 return builder.createOp(opCode, typeId, spvGroupOperands); 6711 } 6712 6713 spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy) 6714 { 6715 bool isUnsigned = isTypeUnsignedInt(typeProxy); 6716 bool isFloat = isTypeFloat(typeProxy); 6717 6718 spv::Op opCode = spv::OpNop; 6719 int extBuiltins = -1; 6720 int libCall = -1; 6721 size_t consumedOperands = operands.size(); 6722 spv::Id typeId0 = 0; 6723 if (consumedOperands > 0) 6724 typeId0 = builder.getTypeId(operands[0]); 6725 spv::Id typeId1 = 0; 6726 if (consumedOperands > 1) 6727 typeId1 = builder.getTypeId(operands[1]); 6728 spv::Id frexpIntType = 0; 6729 6730 switch (op) { 6731 case glslang::EOpMin: 6732 if (isFloat) 6733 libCall = spv::GLSLstd450FMin; 6734 else if (isUnsigned) 6735 libCall = spv::GLSLstd450UMin; 6736 else 6737 libCall = spv::GLSLstd450SMin; 6738 builder.promoteScalar(precision, operands.front(), operands.back()); 6739 break; 6740 case glslang::EOpModf: 6741 libCall = spv::GLSLstd450Modf; 6742 break; 6743 case glslang::EOpMax: 6744 if (isFloat) 6745 libCall = spv::GLSLstd450FMax; 6746 else if (isUnsigned) 6747 libCall = spv::GLSLstd450UMax; 6748 else 6749 libCall = spv::GLSLstd450SMax; 6750 builder.promoteScalar(precision, operands.front(), operands.back()); 6751 break; 6752 case glslang::EOpPow: 6753 libCall = spv::GLSLstd450Pow; 6754 break; 6755 case glslang::EOpDot: 6756 opCode = spv::OpDot; 6757 break; 6758 case glslang::EOpAtan: 6759 libCall = spv::GLSLstd450Atan2; 6760 break; 6761 6762 case glslang::EOpClamp: 6763 if (isFloat) 6764 libCall = spv::GLSLstd450FClamp; 6765 else if (isUnsigned) 6766 libCall = spv::GLSLstd450UClamp; 6767 else 6768 libCall = spv::GLSLstd450SClamp; 6769 builder.promoteScalar(precision, operands.front(), operands[1]); 6770 builder.promoteScalar(precision, operands.front(), operands[2]); 6771 break; 6772 case glslang::EOpMix: 6773 if (! builder.isBoolType(builder.getScalarTypeId(builder.getTypeId(operands.back())))) { 6774 assert(isFloat); 6775 libCall = spv::GLSLstd450FMix; 6776 } else { 6777 opCode = spv::OpSelect; 6778 std::swap(operands.front(), operands.back()); 6779 } 6780 builder.promoteScalar(precision, operands.front(), operands.back()); 6781 break; 6782 case glslang::EOpStep: 6783 libCall = spv::GLSLstd450Step; 6784 builder.promoteScalar(precision, operands.front(), operands.back()); 6785 break; 6786 case glslang::EOpSmoothStep: 6787 libCall = spv::GLSLstd450SmoothStep; 6788 builder.promoteScalar(precision, operands[0], operands[2]); 6789 builder.promoteScalar(precision, operands[1], operands[2]); 6790 break; 6791 6792 case glslang::EOpDistance: 6793 libCall = spv::GLSLstd450Distance; 6794 break; 6795 case glslang::EOpCross: 6796 libCall = spv::GLSLstd450Cross; 6797 break; 6798 case glslang::EOpFaceForward: 6799 libCall = spv::GLSLstd450FaceForward; 6800 break; 6801 case glslang::EOpReflect: 6802 libCall = spv::GLSLstd450Reflect; 6803 break; 6804 case glslang::EOpRefract: 6805 libCall = spv::GLSLstd450Refract; 6806 break; 6807 case glslang::EOpInterpolateAtSample: 6808 #ifdef AMD_EXTENSIONS 6809 if (typeProxy == glslang::EbtFloat16) 6810 builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float); 6811 #endif 6812 libCall = spv::GLSLstd450InterpolateAtSample; 6813 break; 6814 case glslang::EOpInterpolateAtOffset: 6815 #ifdef AMD_EXTENSIONS 6816 if (typeProxy == glslang::EbtFloat16) 6817 builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float); 6818 #endif 6819 libCall = spv::GLSLstd450InterpolateAtOffset; 6820 break; 6821 case glslang::EOpAddCarry: 6822 opCode = spv::OpIAddCarry; 6823 typeId = builder.makeStructResultType(typeId0, typeId0); 6824 consumedOperands = 2; 6825 break; 6826 case glslang::EOpSubBorrow: 6827 opCode = spv::OpISubBorrow; 6828 typeId = builder.makeStructResultType(typeId0, typeId0); 6829 consumedOperands = 2; 6830 break; 6831 case glslang::EOpUMulExtended: 6832 opCode = spv::OpUMulExtended; 6833 typeId = builder.makeStructResultType(typeId0, typeId0); 6834 consumedOperands = 2; 6835 break; 6836 case glslang::EOpIMulExtended: 6837 opCode = spv::OpSMulExtended; 6838 typeId = builder.makeStructResultType(typeId0, typeId0); 6839 consumedOperands = 2; 6840 break; 6841 case glslang::EOpBitfieldExtract: 6842 if (isUnsigned) 6843 opCode = spv::OpBitFieldUExtract; 6844 else 6845 opCode = spv::OpBitFieldSExtract; 6846 break; 6847 case glslang::EOpBitfieldInsert: 6848 opCode = spv::OpBitFieldInsert; 6849 break; 6850 6851 case glslang::EOpFma: 6852 libCall = spv::GLSLstd450Fma; 6853 break; 6854 case glslang::EOpFrexp: 6855 { 6856 libCall = spv::GLSLstd450FrexpStruct; 6857 assert(builder.isPointerType(typeId1)); 6858 typeId1 = builder.getContainedTypeId(typeId1); 6859 int width = builder.getScalarTypeWidth(typeId1); 6860 #ifdef AMD_EXTENSIONS 6861 if (width == 16) 6862 // Using 16-bit exp operand, enable extension SPV_AMD_gpu_shader_int16 6863 builder.addExtension(spv::E_SPV_AMD_gpu_shader_int16); 6864 #endif 6865 if (builder.getNumComponents(operands[0]) == 1) 6866 frexpIntType = builder.makeIntegerType(width, true); 6867 else 6868 frexpIntType = builder.makeVectorType(builder.makeIntegerType(width, true), builder.getNumComponents(operands[0])); 6869 typeId = builder.makeStructResultType(typeId0, frexpIntType); 6870 consumedOperands = 1; 6871 } 6872 break; 6873 case glslang::EOpLdexp: 6874 libCall = spv::GLSLstd450Ldexp; 6875 break; 6876 6877 case glslang::EOpReadInvocation: 6878 return createInvocationsOperation(op, typeId, operands, typeProxy); 6879 6880 case glslang::EOpSubgroupBroadcast: 6881 case glslang::EOpSubgroupBallotBitExtract: 6882 case glslang::EOpSubgroupShuffle: 6883 case glslang::EOpSubgroupShuffleXor: 6884 case glslang::EOpSubgroupShuffleUp: 6885 case glslang::EOpSubgroupShuffleDown: 6886 case glslang::EOpSubgroupClusteredAdd: 6887 case glslang::EOpSubgroupClusteredMul: 6888 case glslang::EOpSubgroupClusteredMin: 6889 case glslang::EOpSubgroupClusteredMax: 6890 case glslang::EOpSubgroupClusteredAnd: 6891 case glslang::EOpSubgroupClusteredOr: 6892 case glslang::EOpSubgroupClusteredXor: 6893 case glslang::EOpSubgroupQuadBroadcast: 6894 #ifdef NV_EXTENSIONS 6895 case glslang::EOpSubgroupPartitionedAdd: 6896 case glslang::EOpSubgroupPartitionedMul: 6897 case glslang::EOpSubgroupPartitionedMin: 6898 case glslang::EOpSubgroupPartitionedMax: 6899 case glslang::EOpSubgroupPartitionedAnd: 6900 case glslang::EOpSubgroupPartitionedOr: 6901 case glslang::EOpSubgroupPartitionedXor: 6902 case glslang::EOpSubgroupPartitionedInclusiveAdd: 6903 case glslang::EOpSubgroupPartitionedInclusiveMul: 6904 case glslang::EOpSubgroupPartitionedInclusiveMin: 6905 case glslang::EOpSubgroupPartitionedInclusiveMax: 6906 case glslang::EOpSubgroupPartitionedInclusiveAnd: 6907 case glslang::EOpSubgroupPartitionedInclusiveOr: 6908 case glslang::EOpSubgroupPartitionedInclusiveXor: 6909 case glslang::EOpSubgroupPartitionedExclusiveAdd: 6910 case glslang::EOpSubgroupPartitionedExclusiveMul: 6911 case glslang::EOpSubgroupPartitionedExclusiveMin: 6912 case glslang::EOpSubgroupPartitionedExclusiveMax: 6913 case glslang::EOpSubgroupPartitionedExclusiveAnd: 6914 case glslang::EOpSubgroupPartitionedExclusiveOr: 6915 case glslang::EOpSubgroupPartitionedExclusiveXor: 6916 #endif 6917 return createSubgroupOperation(op, typeId, operands, typeProxy); 6918 6919 #ifdef AMD_EXTENSIONS 6920 case glslang::EOpSwizzleInvocations: 6921 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot); 6922 libCall = spv::SwizzleInvocationsAMD; 6923 break; 6924 case glslang::EOpSwizzleInvocationsMasked: 6925 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot); 6926 libCall = spv::SwizzleInvocationsMaskedAMD; 6927 break; 6928 case glslang::EOpWriteInvocation: 6929 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot); 6930 libCall = spv::WriteInvocationAMD; 6931 break; 6932 6933 case glslang::EOpMin3: 6934 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax); 6935 if (isFloat) 6936 libCall = spv::FMin3AMD; 6937 else { 6938 if (isUnsigned) 6939 libCall = spv::UMin3AMD; 6940 else 6941 libCall = spv::SMin3AMD; 6942 } 6943 break; 6944 case glslang::EOpMax3: 6945 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax); 6946 if (isFloat) 6947 libCall = spv::FMax3AMD; 6948 else { 6949 if (isUnsigned) 6950 libCall = spv::UMax3AMD; 6951 else 6952 libCall = spv::SMax3AMD; 6953 } 6954 break; 6955 case glslang::EOpMid3: 6956 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax); 6957 if (isFloat) 6958 libCall = spv::FMid3AMD; 6959 else { 6960 if (isUnsigned) 6961 libCall = spv::UMid3AMD; 6962 else 6963 libCall = spv::SMid3AMD; 6964 } 6965 break; 6966 6967 case glslang::EOpInterpolateAtVertex: 6968 if (typeProxy == glslang::EbtFloat16) 6969 builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float); 6970 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_explicit_vertex_parameter); 6971 libCall = spv::InterpolateAtVertexAMD; 6972 break; 6973 #endif 6974 case glslang::EOpBarrier: 6975 { 6976 // This is for the extended controlBarrier function, with four operands. 6977 // The unextended barrier() goes through createNoArgOperation. 6978 assert(operands.size() == 4); 6979 unsigned int executionScope = builder.getConstantScalar(operands[0]); 6980 unsigned int memoryScope = builder.getConstantScalar(operands[1]); 6981 unsigned int semantics = builder.getConstantScalar(operands[2]) | builder.getConstantScalar(operands[3]); 6982 builder.createControlBarrier((spv::Scope)executionScope, (spv::Scope)memoryScope, (spv::MemorySemanticsMask)semantics); 6983 if (semantics & (spv::MemorySemanticsMakeAvailableKHRMask | spv::MemorySemanticsMakeVisibleKHRMask | spv::MemorySemanticsOutputMemoryKHRMask)) { 6984 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); 6985 } 6986 if (glslangIntermediate->usingVulkanMemoryModel() && (executionScope == spv::ScopeDevice || memoryScope == spv::ScopeDevice)) { 6987 builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR); 6988 } 6989 return 0; 6990 } 6991 break; 6992 case glslang::EOpMemoryBarrier: 6993 { 6994 // This is for the extended memoryBarrier function, with three operands. 6995 // The unextended memoryBarrier() goes through createNoArgOperation. 6996 assert(operands.size() == 3); 6997 unsigned int memoryScope = builder.getConstantScalar(operands[0]); 6998 unsigned int semantics = builder.getConstantScalar(operands[1]) | builder.getConstantScalar(operands[2]); 6999 builder.createMemoryBarrier((spv::Scope)memoryScope, (spv::MemorySemanticsMask)semantics); 7000 if (semantics & (spv::MemorySemanticsMakeAvailableKHRMask | spv::MemorySemanticsMakeVisibleKHRMask | spv::MemorySemanticsOutputMemoryKHRMask)) { 7001 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); 7002 } 7003 if (glslangIntermediate->usingVulkanMemoryModel() && memoryScope == spv::ScopeDevice) { 7004 builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR); 7005 } 7006 return 0; 7007 } 7008 break; 7009 7010 #ifdef NV_EXTENSIONS 7011 case glslang::EOpReportIntersectionNV: 7012 { 7013 typeId = builder.makeBoolType(); 7014 opCode = spv::OpReportIntersectionNV; 7015 } 7016 break; 7017 case glslang::EOpTraceNV: 7018 { 7019 builder.createNoResultOp(spv::OpTraceNV, operands); 7020 return 0; 7021 } 7022 break; 7023 case glslang::EOpExecuteCallableNV: 7024 { 7025 builder.createNoResultOp(spv::OpExecuteCallableNV, operands); 7026 return 0; 7027 } 7028 break; 7029 case glslang::EOpWritePackedPrimitiveIndices4x8NV: 7030 builder.createNoResultOp(spv::OpWritePackedPrimitiveIndices4x8NV, operands); 7031 return 0; 7032 #endif 7033 default: 7034 return 0; 7035 } 7036 7037 spv::Id id = 0; 7038 if (libCall >= 0) { 7039 // Use an extended instruction from the standard library. 7040 // Construct the call arguments, without modifying the original operands vector. 7041 // We might need the remaining arguments, e.g. in the EOpFrexp case. 7042 std::vector<spv::Id> callArguments(operands.begin(), operands.begin() + consumedOperands); 7043 id = builder.createBuiltinCall(typeId, extBuiltins >= 0 ? extBuiltins : stdBuiltins, libCall, callArguments); 7044 } else if (opCode == spv::OpDot && !isFloat) { 7045 // int dot(int, int) 7046 // NOTE: never called for scalar/vector1, this is turned into simple mul before this can be reached 7047 const int componentCount = builder.getNumComponents(operands[0]); 7048 spv::Id mulOp = builder.createBinOp(spv::OpIMul, builder.getTypeId(operands[0]), operands[0], operands[1]); 7049 builder.setPrecision(mulOp, precision); 7050 id = builder.createCompositeExtract(mulOp, typeId, 0); 7051 for (int i = 1; i < componentCount; ++i) { 7052 builder.setPrecision(id, precision); 7053 id = builder.createBinOp(spv::OpIAdd, typeId, id, builder.createCompositeExtract(operands[0], typeId, i)); 7054 } 7055 } else { 7056 switch (consumedOperands) { 7057 case 0: 7058 // should all be handled by visitAggregate and createNoArgOperation 7059 assert(0); 7060 return 0; 7061 case 1: 7062 // should all be handled by createUnaryOperation 7063 assert(0); 7064 return 0; 7065 case 2: 7066 id = builder.createBinOp(opCode, typeId, operands[0], operands[1]); 7067 break; 7068 default: 7069 // anything 3 or over doesn't have l-value operands, so all should be consumed 7070 assert(consumedOperands == operands.size()); 7071 id = builder.createOp(opCode, typeId, operands); 7072 break; 7073 } 7074 } 7075 7076 // Decode the return types that were structures 7077 switch (op) { 7078 case glslang::EOpAddCarry: 7079 case glslang::EOpSubBorrow: 7080 builder.createStore(builder.createCompositeExtract(id, typeId0, 1), operands[2]); 7081 id = builder.createCompositeExtract(id, typeId0, 0); 7082 break; 7083 case glslang::EOpUMulExtended: 7084 case glslang::EOpIMulExtended: 7085 builder.createStore(builder.createCompositeExtract(id, typeId0, 0), operands[3]); 7086 builder.createStore(builder.createCompositeExtract(id, typeId0, 1), operands[2]); 7087 break; 7088 case glslang::EOpFrexp: 7089 { 7090 assert(operands.size() == 2); 7091 if (builder.isFloatType(builder.getScalarTypeId(typeId1))) { 7092 // "exp" is floating-point type (from HLSL intrinsic) 7093 spv::Id member1 = builder.createCompositeExtract(id, frexpIntType, 1); 7094 member1 = builder.createUnaryOp(spv::OpConvertSToF, typeId1, member1); 7095 builder.createStore(member1, operands[1]); 7096 } else 7097 // "exp" is integer type (from GLSL built-in function) 7098 builder.createStore(builder.createCompositeExtract(id, frexpIntType, 1), operands[1]); 7099 id = builder.createCompositeExtract(id, typeId0, 0); 7100 } 7101 break; 7102 default: 7103 break; 7104 } 7105 7106 return builder.setPrecision(id, precision); 7107 } 7108 7109 // Intrinsics with no arguments (or no return value, and no precision). 7110 spv::Id TGlslangToSpvTraverser::createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId) 7111 { 7112 // GLSL memory barriers use queuefamily scope in new model, device scope in old model 7113 spv::Scope memoryBarrierScope = glslangIntermediate->usingVulkanMemoryModel() ? spv::ScopeQueueFamilyKHR : spv::ScopeDevice; 7114 7115 switch (op) { 7116 case glslang::EOpEmitVertex: 7117 builder.createNoResultOp(spv::OpEmitVertex); 7118 return 0; 7119 case glslang::EOpEndPrimitive: 7120 builder.createNoResultOp(spv::OpEndPrimitive); 7121 return 0; 7122 case glslang::EOpBarrier: 7123 if (glslangIntermediate->getStage() == EShLangTessControl) { 7124 if (glslangIntermediate->usingVulkanMemoryModel()) { 7125 builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup, 7126 spv::MemorySemanticsOutputMemoryKHRMask | 7127 spv::MemorySemanticsAcquireReleaseMask); 7128 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR); 7129 } else { 7130 builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeInvocation, spv::MemorySemanticsMaskNone); 7131 } 7132 } else { 7133 builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup, 7134 spv::MemorySemanticsWorkgroupMemoryMask | 7135 spv::MemorySemanticsAcquireReleaseMask); 7136 } 7137 return 0; 7138 case glslang::EOpMemoryBarrier: 7139 builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsAllMemory | 7140 spv::MemorySemanticsAcquireReleaseMask); 7141 return 0; 7142 case glslang::EOpMemoryBarrierAtomicCounter: 7143 builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsAtomicCounterMemoryMask | 7144 spv::MemorySemanticsAcquireReleaseMask); 7145 return 0; 7146 case glslang::EOpMemoryBarrierBuffer: 7147 builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsUniformMemoryMask | 7148 spv::MemorySemanticsAcquireReleaseMask); 7149 return 0; 7150 case glslang::EOpMemoryBarrierImage: 7151 builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsImageMemoryMask | 7152 spv::MemorySemanticsAcquireReleaseMask); 7153 return 0; 7154 case glslang::EOpMemoryBarrierShared: 7155 builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsWorkgroupMemoryMask | 7156 spv::MemorySemanticsAcquireReleaseMask); 7157 return 0; 7158 case glslang::EOpGroupMemoryBarrier: 7159 builder.createMemoryBarrier(spv::ScopeWorkgroup, spv::MemorySemanticsAllMemory | 7160 spv::MemorySemanticsAcquireReleaseMask); 7161 return 0; 7162 case glslang::EOpAllMemoryBarrierWithGroupSync: 7163 builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeDevice, 7164 spv::MemorySemanticsAllMemory | 7165 spv::MemorySemanticsAcquireReleaseMask); 7166 return 0; 7167 case glslang::EOpDeviceMemoryBarrier: 7168 builder.createMemoryBarrier(spv::ScopeDevice, spv::MemorySemanticsUniformMemoryMask | 7169 spv::MemorySemanticsImageMemoryMask | 7170 spv::MemorySemanticsAcquireReleaseMask); 7171 return 0; 7172 case glslang::EOpDeviceMemoryBarrierWithGroupSync: 7173 builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeDevice, spv::MemorySemanticsUniformMemoryMask | 7174 spv::MemorySemanticsImageMemoryMask | 7175 spv::MemorySemanticsAcquireReleaseMask); 7176 return 0; 7177 case glslang::EOpWorkgroupMemoryBarrier: 7178 builder.createMemoryBarrier(spv::ScopeWorkgroup, spv::MemorySemanticsWorkgroupMemoryMask | 7179 spv::MemorySemanticsAcquireReleaseMask); 7180 return 0; 7181 case glslang::EOpWorkgroupMemoryBarrierWithGroupSync: 7182 builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup, 7183 spv::MemorySemanticsWorkgroupMemoryMask | 7184 spv::MemorySemanticsAcquireReleaseMask); 7185 return 0; 7186 case glslang::EOpSubgroupBarrier: 7187 builder.createControlBarrier(spv::ScopeSubgroup, spv::ScopeSubgroup, spv::MemorySemanticsAllMemory | 7188 spv::MemorySemanticsAcquireReleaseMask); 7189 return spv::NoResult; 7190 case glslang::EOpSubgroupMemoryBarrier: 7191 builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsAllMemory | 7192 spv::MemorySemanticsAcquireReleaseMask); 7193 return spv::NoResult; 7194 case glslang::EOpSubgroupMemoryBarrierBuffer: 7195 builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsUniformMemoryMask | 7196 spv::MemorySemanticsAcquireReleaseMask); 7197 return spv::NoResult; 7198 case glslang::EOpSubgroupMemoryBarrierImage: 7199 builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsImageMemoryMask | 7200 spv::MemorySemanticsAcquireReleaseMask); 7201 return spv::NoResult; 7202 case glslang::EOpSubgroupMemoryBarrierShared: 7203 builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsWorkgroupMemoryMask | 7204 spv::MemorySemanticsAcquireReleaseMask); 7205 return spv::NoResult; 7206 case glslang::EOpSubgroupElect: { 7207 std::vector<spv::Id> operands; 7208 return createSubgroupOperation(op, typeId, operands, glslang::EbtVoid); 7209 } 7210 #ifdef AMD_EXTENSIONS 7211 case glslang::EOpTime: 7212 { 7213 std::vector<spv::Id> args; // Dummy arguments 7214 spv::Id id = builder.createBuiltinCall(typeId, getExtBuiltins(spv::E_SPV_AMD_gcn_shader), spv::TimeAMD, args); 7215 return builder.setPrecision(id, precision); 7216 } 7217 #endif 7218 #ifdef NV_EXTENSIONS 7219 case glslang::EOpIgnoreIntersectionNV: 7220 builder.createNoResultOp(spv::OpIgnoreIntersectionNV); 7221 return 0; 7222 case glslang::EOpTerminateRayNV: 7223 builder.createNoResultOp(spv::OpTerminateRayNV); 7224 return 0; 7225 #endif 7226 default: 7227 logger->missingFunctionality("unknown operation with no arguments"); 7228 return 0; 7229 } 7230 } 7231 7232 spv::Id TGlslangToSpvTraverser::getSymbolId(const glslang::TIntermSymbol* symbol) 7233 { 7234 auto iter = symbolValues.find(symbol->getId()); 7235 spv::Id id; 7236 if (symbolValues.end() != iter) { 7237 id = iter->second; 7238 return id; 7239 } 7240 7241 // it was not found, create it 7242 id = createSpvVariable(symbol); 7243 symbolValues[symbol->getId()] = id; 7244 7245 if (symbol->getBasicType() != glslang::EbtBlock) { 7246 builder.addDecoration(id, TranslatePrecisionDecoration(symbol->getType())); 7247 builder.addDecoration(id, TranslateInterpolationDecoration(symbol->getType().getQualifier())); 7248 builder.addDecoration(id, TranslateAuxiliaryStorageDecoration(symbol->getType().getQualifier())); 7249 #ifdef NV_EXTENSIONS 7250 addMeshNVDecoration(id, /*member*/ -1, symbol->getType().getQualifier()); 7251 #endif 7252 if (symbol->getType().getQualifier().hasSpecConstantId()) 7253 builder.addDecoration(id, spv::DecorationSpecId, symbol->getType().getQualifier().layoutSpecConstantId); 7254 if (symbol->getQualifier().hasIndex()) 7255 builder.addDecoration(id, spv::DecorationIndex, symbol->getQualifier().layoutIndex); 7256 if (symbol->getQualifier().hasComponent()) 7257 builder.addDecoration(id, spv::DecorationComponent, symbol->getQualifier().layoutComponent); 7258 // atomic counters use this: 7259 if (symbol->getQualifier().hasOffset()) 7260 builder.addDecoration(id, spv::DecorationOffset, symbol->getQualifier().layoutOffset); 7261 } 7262 7263 if (symbol->getQualifier().hasLocation()) 7264 builder.addDecoration(id, spv::DecorationLocation, symbol->getQualifier().layoutLocation); 7265 builder.addDecoration(id, TranslateInvariantDecoration(symbol->getType().getQualifier())); 7266 if (symbol->getQualifier().hasStream() && glslangIntermediate->isMultiStream()) { 7267 builder.addCapability(spv::CapabilityGeometryStreams); 7268 builder.addDecoration(id, spv::DecorationStream, symbol->getQualifier().layoutStream); 7269 } 7270 if (symbol->getQualifier().hasSet()) 7271 builder.addDecoration(id, spv::DecorationDescriptorSet, symbol->getQualifier().layoutSet); 7272 else if (IsDescriptorResource(symbol->getType())) { 7273 // default to 0 7274 builder.addDecoration(id, spv::DecorationDescriptorSet, 0); 7275 } 7276 if (symbol->getQualifier().hasBinding()) 7277 builder.addDecoration(id, spv::DecorationBinding, symbol->getQualifier().layoutBinding); 7278 else if (IsDescriptorResource(symbol->getType())) { 7279 // default to 0 7280 builder.addDecoration(id, spv::DecorationBinding, 0); 7281 } 7282 if (symbol->getQualifier().hasAttachment()) 7283 builder.addDecoration(id, spv::DecorationInputAttachmentIndex, symbol->getQualifier().layoutAttachment); 7284 if (glslangIntermediate->getXfbMode()) { 7285 builder.addCapability(spv::CapabilityTransformFeedback); 7286 if (symbol->getQualifier().hasXfbBuffer()) { 7287 builder.addDecoration(id, spv::DecorationXfbBuffer, symbol->getQualifier().layoutXfbBuffer); 7288 unsigned stride = glslangIntermediate->getXfbStride(symbol->getQualifier().layoutXfbBuffer); 7289 if (stride != glslang::TQualifier::layoutXfbStrideEnd) 7290 builder.addDecoration(id, spv::DecorationXfbStride, stride); 7291 } 7292 if (symbol->getQualifier().hasXfbOffset()) 7293 builder.addDecoration(id, spv::DecorationOffset, symbol->getQualifier().layoutXfbOffset); 7294 } 7295 7296 if (symbol->getType().isImage()) { 7297 std::vector<spv::Decoration> memory; 7298 TranslateMemoryDecoration(symbol->getType().getQualifier(), memory, glslangIntermediate->usingVulkanMemoryModel()); 7299 for (unsigned int i = 0; i < memory.size(); ++i) 7300 builder.addDecoration(id, memory[i]); 7301 } 7302 7303 // built-in variable decorations 7304 spv::BuiltIn builtIn = TranslateBuiltInDecoration(symbol->getQualifier().builtIn, false); 7305 if (builtIn != spv::BuiltInMax) 7306 builder.addDecoration(id, spv::DecorationBuiltIn, (int)builtIn); 7307 7308 // nonuniform 7309 builder.addDecoration(id, TranslateNonUniformDecoration(symbol->getType().getQualifier())); 7310 7311 #ifdef NV_EXTENSIONS 7312 if (builtIn == spv::BuiltInSampleMask) { 7313 spv::Decoration decoration; 7314 // GL_NV_sample_mask_override_coverage extension 7315 if (glslangIntermediate->getLayoutOverrideCoverage()) 7316 decoration = (spv::Decoration)spv::DecorationOverrideCoverageNV; 7317 else 7318 decoration = (spv::Decoration)spv::DecorationMax; 7319 builder.addDecoration(id, decoration); 7320 if (decoration != spv::DecorationMax) { 7321 builder.addExtension(spv::E_SPV_NV_sample_mask_override_coverage); 7322 } 7323 } 7324 else if (builtIn == spv::BuiltInLayer) { 7325 // SPV_NV_viewport_array2 extension 7326 if (symbol->getQualifier().layoutViewportRelative) { 7327 builder.addDecoration(id, (spv::Decoration)spv::DecorationViewportRelativeNV); 7328 builder.addCapability(spv::CapabilityShaderViewportMaskNV); 7329 builder.addExtension(spv::E_SPV_NV_viewport_array2); 7330 } 7331 if (symbol->getQualifier().layoutSecondaryViewportRelativeOffset != -2048) { 7332 builder.addDecoration(id, (spv::Decoration)spv::DecorationSecondaryViewportRelativeNV, 7333 symbol->getQualifier().layoutSecondaryViewportRelativeOffset); 7334 builder.addCapability(spv::CapabilityShaderStereoViewNV); 7335 builder.addExtension(spv::E_SPV_NV_stereo_view_rendering); 7336 } 7337 } 7338 7339 if (symbol->getQualifier().layoutPassthrough) { 7340 builder.addDecoration(id, spv::DecorationPassthroughNV); 7341 builder.addCapability(spv::CapabilityGeometryShaderPassthroughNV); 7342 builder.addExtension(spv::E_SPV_NV_geometry_shader_passthrough); 7343 } 7344 if (symbol->getQualifier().pervertexNV) { 7345 builder.addDecoration(id, spv::DecorationPerVertexNV); 7346 builder.addCapability(spv::CapabilityFragmentBarycentricNV); 7347 builder.addExtension(spv::E_SPV_NV_fragment_shader_barycentric); 7348 } 7349 #endif 7350 7351 if (glslangIntermediate->getHlslFunctionality1() && symbol->getType().getQualifier().semanticName != nullptr) { 7352 builder.addExtension("SPV_GOOGLE_hlsl_functionality1"); 7353 builder.addDecoration(id, (spv::Decoration)spv::DecorationHlslSemanticGOOGLE, 7354 symbol->getType().getQualifier().semanticName); 7355 } 7356 7357 if (symbol->getBasicType() == glslang::EbtReference) { 7358 builder.addDecoration(id, symbol->getType().getQualifier().restrict ? spv::DecorationRestrictPointerEXT : spv::DecorationAliasedPointerEXT); 7359 } 7360 7361 return id; 7362 } 7363 7364 #ifdef NV_EXTENSIONS 7365 // add per-primitive, per-view. per-task decorations to a struct member (member >= 0) or an object 7366 void TGlslangToSpvTraverser::addMeshNVDecoration(spv::Id id, int member, const glslang::TQualifier& qualifier) 7367 { 7368 if (member >= 0) { 7369 if (qualifier.perPrimitiveNV) { 7370 // Need to add capability/extension for fragment shader. 7371 // Mesh shader already adds this by default. 7372 if (glslangIntermediate->getStage() == EShLangFragment) { 7373 builder.addCapability(spv::CapabilityMeshShadingNV); 7374 builder.addExtension(spv::E_SPV_NV_mesh_shader); 7375 } 7376 builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerPrimitiveNV); 7377 } 7378 if (qualifier.perViewNV) 7379 builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerViewNV); 7380 if (qualifier.perTaskNV) 7381 builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerTaskNV); 7382 } else { 7383 if (qualifier.perPrimitiveNV) { 7384 // Need to add capability/extension for fragment shader. 7385 // Mesh shader already adds this by default. 7386 if (glslangIntermediate->getStage() == EShLangFragment) { 7387 builder.addCapability(spv::CapabilityMeshShadingNV); 7388 builder.addExtension(spv::E_SPV_NV_mesh_shader); 7389 } 7390 builder.addDecoration(id, spv::DecorationPerPrimitiveNV); 7391 } 7392 if (qualifier.perViewNV) 7393 builder.addDecoration(id, spv::DecorationPerViewNV); 7394 if (qualifier.perTaskNV) 7395 builder.addDecoration(id, spv::DecorationPerTaskNV); 7396 } 7397 } 7398 #endif 7399 7400 // Make a full tree of instructions to build a SPIR-V specialization constant, 7401 // or regular constant if possible. 7402 // 7403 // TBD: this is not yet done, nor verified to be the best design, it does do the leaf symbols though 7404 // 7405 // Recursively walk the nodes. The nodes form a tree whose leaves are 7406 // regular constants, which themselves are trees that createSpvConstant() 7407 // recursively walks. So, this function walks the "top" of the tree: 7408 // - emit specialization constant-building instructions for specConstant 7409 // - when running into a non-spec-constant, switch to createSpvConstant() 7410 spv::Id TGlslangToSpvTraverser::createSpvConstant(const glslang::TIntermTyped& node) 7411 { 7412 assert(node.getQualifier().isConstant()); 7413 7414 // Handle front-end constants first (non-specialization constants). 7415 if (! node.getQualifier().specConstant) { 7416 // hand off to the non-spec-constant path 7417 assert(node.getAsConstantUnion() != nullptr || node.getAsSymbolNode() != nullptr); 7418 int nextConst = 0; 7419 return createSpvConstantFromConstUnionArray(node.getType(), node.getAsConstantUnion() ? node.getAsConstantUnion()->getConstArray() : node.getAsSymbolNode()->getConstArray(), 7420 nextConst, false); 7421 } 7422 7423 // We now know we have a specialization constant to build 7424 7425 // gl_WorkGroupSize is a special case until the front-end handles hierarchical specialization constants, 7426 // even then, it's specialization ids are handled by special case syntax in GLSL: layout(local_size_x = ... 7427 if (node.getType().getQualifier().builtIn == glslang::EbvWorkGroupSize) { 7428 std::vector<spv::Id> dimConstId; 7429 for (int dim = 0; dim < 3; ++dim) { 7430 bool specConst = (glslangIntermediate->getLocalSizeSpecId(dim) != glslang::TQualifier::layoutNotSet); 7431 dimConstId.push_back(builder.makeUintConstant(glslangIntermediate->getLocalSize(dim), specConst)); 7432 if (specConst) { 7433 builder.addDecoration(dimConstId.back(), spv::DecorationSpecId, 7434 glslangIntermediate->getLocalSizeSpecId(dim)); 7435 } 7436 } 7437 return builder.makeCompositeConstant(builder.makeVectorType(builder.makeUintType(32), 3), dimConstId, true); 7438 } 7439 7440 // An AST node labelled as specialization constant should be a symbol node. 7441 // Its initializer should either be a sub tree with constant nodes, or a constant union array. 7442 if (auto* sn = node.getAsSymbolNode()) { 7443 spv::Id result; 7444 if (auto* sub_tree = sn->getConstSubtree()) { 7445 // Traverse the constant constructor sub tree like generating normal run-time instructions. 7446 // During the AST traversal, if the node is marked as 'specConstant', SpecConstantOpModeGuard 7447 // will set the builder into spec constant op instruction generating mode. 7448 sub_tree->traverse(this); 7449 result = accessChainLoad(sub_tree->getType()); 7450 } else if (auto* const_union_array = &sn->getConstArray()) { 7451 int nextConst = 0; 7452 result = createSpvConstantFromConstUnionArray(sn->getType(), *const_union_array, nextConst, true); 7453 } else { 7454 logger->missingFunctionality("Invalid initializer for spec onstant."); 7455 return spv::NoResult; 7456 } 7457 builder.addName(result, sn->getName().c_str()); 7458 return result; 7459 } 7460 7461 // Neither a front-end constant node, nor a specialization constant node with constant union array or 7462 // constant sub tree as initializer. 7463 logger->missingFunctionality("Neither a front-end constant nor a spec constant."); 7464 return spv::NoResult; 7465 } 7466 7467 // Use 'consts' as the flattened glslang source of scalar constants to recursively 7468 // build the aggregate SPIR-V constant. 7469 // 7470 // If there are not enough elements present in 'consts', 0 will be substituted; 7471 // an empty 'consts' can be used to create a fully zeroed SPIR-V constant. 7472 // 7473 spv::Id TGlslangToSpvTraverser::createSpvConstantFromConstUnionArray(const glslang::TType& glslangType, const glslang::TConstUnionArray& consts, int& nextConst, bool specConstant) 7474 { 7475 // vector of constants for SPIR-V 7476 std::vector<spv::Id> spvConsts; 7477 7478 // Type is used for struct and array constants 7479 spv::Id typeId = convertGlslangToSpvType(glslangType); 7480 7481 if (glslangType.isArray()) { 7482 glslang::TType elementType(glslangType, 0); 7483 for (int i = 0; i < glslangType.getOuterArraySize(); ++i) 7484 spvConsts.push_back(createSpvConstantFromConstUnionArray(elementType, consts, nextConst, false)); 7485 } else if (glslangType.isMatrix()) { 7486 glslang::TType vectorType(glslangType, 0); 7487 for (int col = 0; col < glslangType.getMatrixCols(); ++col) 7488 spvConsts.push_back(createSpvConstantFromConstUnionArray(vectorType, consts, nextConst, false)); 7489 } else if (glslangType.isStruct()) { 7490 glslang::TVector<glslang::TTypeLoc>::const_iterator iter; 7491 for (iter = glslangType.getStruct()->begin(); iter != glslangType.getStruct()->end(); ++iter) 7492 spvConsts.push_back(createSpvConstantFromConstUnionArray(*iter->type, consts, nextConst, false)); 7493 } else if (glslangType.getVectorSize() > 1) { 7494 for (unsigned int i = 0; i < (unsigned int)glslangType.getVectorSize(); ++i) { 7495 bool zero = nextConst >= consts.size(); 7496 switch (glslangType.getBasicType()) { 7497 case glslang::EbtInt8: 7498 spvConsts.push_back(builder.makeInt8Constant(zero ? 0 : consts[nextConst].getI8Const())); 7499 break; 7500 case glslang::EbtUint8: 7501 spvConsts.push_back(builder.makeUint8Constant(zero ? 0 : consts[nextConst].getU8Const())); 7502 break; 7503 case glslang::EbtInt16: 7504 spvConsts.push_back(builder.makeInt16Constant(zero ? 0 : consts[nextConst].getI16Const())); 7505 break; 7506 case glslang::EbtUint16: 7507 spvConsts.push_back(builder.makeUint16Constant(zero ? 0 : consts[nextConst].getU16Const())); 7508 break; 7509 case glslang::EbtInt: 7510 spvConsts.push_back(builder.makeIntConstant(zero ? 0 : consts[nextConst].getIConst())); 7511 break; 7512 case glslang::EbtUint: 7513 spvConsts.push_back(builder.makeUintConstant(zero ? 0 : consts[nextConst].getUConst())); 7514 break; 7515 case glslang::EbtInt64: 7516 spvConsts.push_back(builder.makeInt64Constant(zero ? 0 : consts[nextConst].getI64Const())); 7517 break; 7518 case glslang::EbtUint64: 7519 spvConsts.push_back(builder.makeUint64Constant(zero ? 0 : consts[nextConst].getU64Const())); 7520 break; 7521 case glslang::EbtFloat: 7522 spvConsts.push_back(builder.makeFloatConstant(zero ? 0.0F : (float)consts[nextConst].getDConst())); 7523 break; 7524 case glslang::EbtDouble: 7525 spvConsts.push_back(builder.makeDoubleConstant(zero ? 0.0 : consts[nextConst].getDConst())); 7526 break; 7527 case glslang::EbtFloat16: 7528 spvConsts.push_back(builder.makeFloat16Constant(zero ? 0.0F : (float)consts[nextConst].getDConst())); 7529 break; 7530 case glslang::EbtBool: 7531 spvConsts.push_back(builder.makeBoolConstant(zero ? false : consts[nextConst].getBConst())); 7532 break; 7533 default: 7534 assert(0); 7535 break; 7536 } 7537 ++nextConst; 7538 } 7539 } else { 7540 // we have a non-aggregate (scalar) constant 7541 bool zero = nextConst >= consts.size(); 7542 spv::Id scalar = 0; 7543 switch (glslangType.getBasicType()) { 7544 case glslang::EbtInt8: 7545 scalar = builder.makeInt8Constant(zero ? 0 : consts[nextConst].getI8Const(), specConstant); 7546 break; 7547 case glslang::EbtUint8: 7548 scalar = builder.makeUint8Constant(zero ? 0 : consts[nextConst].getU8Const(), specConstant); 7549 break; 7550 case glslang::EbtInt16: 7551 scalar = builder.makeInt16Constant(zero ? 0 : consts[nextConst].getI16Const(), specConstant); 7552 break; 7553 case glslang::EbtUint16: 7554 scalar = builder.makeUint16Constant(zero ? 0 : consts[nextConst].getU16Const(), specConstant); 7555 break; 7556 case glslang::EbtInt: 7557 scalar = builder.makeIntConstant(zero ? 0 : consts[nextConst].getIConst(), specConstant); 7558 break; 7559 case glslang::EbtUint: 7560 scalar = builder.makeUintConstant(zero ? 0 : consts[nextConst].getUConst(), specConstant); 7561 break; 7562 case glslang::EbtInt64: 7563 scalar = builder.makeInt64Constant(zero ? 0 : consts[nextConst].getI64Const(), specConstant); 7564 break; 7565 case glslang::EbtUint64: 7566 scalar = builder.makeUint64Constant(zero ? 0 : consts[nextConst].getU64Const(), specConstant); 7567 break; 7568 case glslang::EbtFloat: 7569 scalar = builder.makeFloatConstant(zero ? 0.0F : (float)consts[nextConst].getDConst(), specConstant); 7570 break; 7571 case glslang::EbtDouble: 7572 scalar = builder.makeDoubleConstant(zero ? 0.0 : consts[nextConst].getDConst(), specConstant); 7573 break; 7574 case glslang::EbtFloat16: 7575 scalar = builder.makeFloat16Constant(zero ? 0.0F : (float)consts[nextConst].getDConst(), specConstant); 7576 break; 7577 case glslang::EbtBool: 7578 scalar = builder.makeBoolConstant(zero ? false : consts[nextConst].getBConst(), specConstant); 7579 break; 7580 default: 7581 assert(0); 7582 break; 7583 } 7584 ++nextConst; 7585 return scalar; 7586 } 7587 7588 return builder.makeCompositeConstant(typeId, spvConsts); 7589 } 7590 7591 // Return true if the node is a constant or symbol whose reading has no 7592 // non-trivial observable cost or effect. 7593 bool TGlslangToSpvTraverser::isTrivialLeaf(const glslang::TIntermTyped* node) 7594 { 7595 // don't know what this is 7596 if (node == nullptr) 7597 return false; 7598 7599 // a constant is safe 7600 if (node->getAsConstantUnion() != nullptr) 7601 return true; 7602 7603 // not a symbol means non-trivial 7604 if (node->getAsSymbolNode() == nullptr) 7605 return false; 7606 7607 // a symbol, depends on what's being read 7608 switch (node->getType().getQualifier().storage) { 7609 case glslang::EvqTemporary: 7610 case glslang::EvqGlobal: 7611 case glslang::EvqIn: 7612 case glslang::EvqInOut: 7613 case glslang::EvqConst: 7614 case glslang::EvqConstReadOnly: 7615 case glslang::EvqUniform: 7616 return true; 7617 default: 7618 return false; 7619 } 7620 } 7621 7622 // A node is trivial if it is a single operation with no side effects. 7623 // HLSL (and/or vectors) are always trivial, as it does not short circuit. 7624 // Otherwise, error on the side of saying non-trivial. 7625 // Return true if trivial. 7626 bool TGlslangToSpvTraverser::isTrivial(const glslang::TIntermTyped* node) 7627 { 7628 if (node == nullptr) 7629 return false; 7630 7631 // count non scalars as trivial, as well as anything coming from HLSL 7632 if (! node->getType().isScalarOrVec1() || glslangIntermediate->getSource() == glslang::EShSourceHlsl) 7633 return true; 7634 7635 // symbols and constants are trivial 7636 if (isTrivialLeaf(node)) 7637 return true; 7638 7639 // otherwise, it needs to be a simple operation or one or two leaf nodes 7640 7641 // not a simple operation 7642 const glslang::TIntermBinary* binaryNode = node->getAsBinaryNode(); 7643 const glslang::TIntermUnary* unaryNode = node->getAsUnaryNode(); 7644 if (binaryNode == nullptr && unaryNode == nullptr) 7645 return false; 7646 7647 // not on leaf nodes 7648 if (binaryNode && (! isTrivialLeaf(binaryNode->getLeft()) || ! isTrivialLeaf(binaryNode->getRight()))) 7649 return false; 7650 7651 if (unaryNode && ! isTrivialLeaf(unaryNode->getOperand())) { 7652 return false; 7653 } 7654 7655 switch (node->getAsOperator()->getOp()) { 7656 case glslang::EOpLogicalNot: 7657 case glslang::EOpConvIntToBool: 7658 case glslang::EOpConvUintToBool: 7659 case glslang::EOpConvFloatToBool: 7660 case glslang::EOpConvDoubleToBool: 7661 case glslang::EOpEqual: 7662 case glslang::EOpNotEqual: 7663 case glslang::EOpLessThan: 7664 case glslang::EOpGreaterThan: 7665 case glslang::EOpLessThanEqual: 7666 case glslang::EOpGreaterThanEqual: 7667 case glslang::EOpIndexDirect: 7668 case glslang::EOpIndexDirectStruct: 7669 case glslang::EOpLogicalXor: 7670 case glslang::EOpAny: 7671 case glslang::EOpAll: 7672 return true; 7673 default: 7674 return false; 7675 } 7676 } 7677 7678 // Emit short-circuiting code, where 'right' is never evaluated unless 7679 // the left side is true (for &&) or false (for ||). 7680 spv::Id TGlslangToSpvTraverser::createShortCircuit(glslang::TOperator op, glslang::TIntermTyped& left, glslang::TIntermTyped& right) 7681 { 7682 spv::Id boolTypeId = builder.makeBoolType(); 7683 7684 // emit left operand 7685 builder.clearAccessChain(); 7686 left.traverse(this); 7687 spv::Id leftId = accessChainLoad(left.getType()); 7688 7689 // Operands to accumulate OpPhi operands 7690 std::vector<spv::Id> phiOperands; 7691 // accumulate left operand's phi information 7692 phiOperands.push_back(leftId); 7693 phiOperands.push_back(builder.getBuildPoint()->getId()); 7694 7695 // Make the two kinds of operation symmetric with a "!" 7696 // || => emit "if (! left) result = right" 7697 // && => emit "if ( left) result = right" 7698 // 7699 // TODO: this runtime "not" for || could be avoided by adding functionality 7700 // to 'builder' to have an "else" without an "then" 7701 if (op == glslang::EOpLogicalOr) 7702 leftId = builder.createUnaryOp(spv::OpLogicalNot, boolTypeId, leftId); 7703 7704 // make an "if" based on the left value 7705 spv::Builder::If ifBuilder(leftId, spv::SelectionControlMaskNone, builder); 7706 7707 // emit right operand as the "then" part of the "if" 7708 builder.clearAccessChain(); 7709 right.traverse(this); 7710 spv::Id rightId = accessChainLoad(right.getType()); 7711 7712 // accumulate left operand's phi information 7713 phiOperands.push_back(rightId); 7714 phiOperands.push_back(builder.getBuildPoint()->getId()); 7715 7716 // finish the "if" 7717 ifBuilder.makeEndIf(); 7718 7719 // phi together the two results 7720 return builder.createOp(spv::OpPhi, boolTypeId, phiOperands); 7721 } 7722 7723 #ifdef AMD_EXTENSIONS 7724 // Return type Id of the imported set of extended instructions corresponds to the name. 7725 // Import this set if it has not been imported yet. 7726 spv::Id TGlslangToSpvTraverser::getExtBuiltins(const char* name) 7727 { 7728 if (extBuiltinMap.find(name) != extBuiltinMap.end()) 7729 return extBuiltinMap[name]; 7730 else { 7731 builder.addExtension(name); 7732 spv::Id extBuiltins = builder.import(name); 7733 extBuiltinMap[name] = extBuiltins; 7734 return extBuiltins; 7735 } 7736 } 7737 #endif 7738 7739 }; // end anonymous namespace 7740 7741 namespace glslang { 7742 7743 void GetSpirvVersion(std::string& version) 7744 { 7745 const int bufSize = 100; 7746 char buf[bufSize]; 7747 snprintf(buf, bufSize, "0x%08x, Revision %d", spv::Version, spv::Revision); 7748 version = buf; 7749 } 7750 7751 // For low-order part of the generator's magic number. Bump up 7752 // when there is a change in the style (e.g., if SSA form changes, 7753 // or a different instruction sequence to do something gets used). 7754 int GetSpirvGeneratorVersion() 7755 { 7756 // return 1; // start 7757 // return 2; // EOpAtomicCounterDecrement gets a post decrement, to map between GLSL -> SPIR-V 7758 // return 3; // change/correct barrier-instruction operands, to match memory model group decisions 7759 // return 4; // some deeper access chains: for dynamic vector component, and local Boolean component 7760 // return 5; // make OpArrayLength result type be an int with signedness of 0 7761 // return 6; // revert version 5 change, which makes a different (new) kind of incorrect code, 7762 // versions 4 and 6 each generate OpArrayLength as it has long been done 7763 return 7; // GLSL volatile keyword maps to both SPIR-V decorations Volatile and Coherent 7764 } 7765 7766 // Write SPIR-V out to a binary file 7767 void OutputSpvBin(const std::vector<unsigned int>& spirv, const char* baseName) 7768 { 7769 std::ofstream out; 7770 out.open(baseName, std::ios::binary | std::ios::out); 7771 if (out.fail()) 7772 printf("ERROR: Failed to open file: %s\n", baseName); 7773 for (int i = 0; i < (int)spirv.size(); ++i) { 7774 unsigned int word = spirv[i]; 7775 out.write((const char*)&word, 4); 7776 } 7777 out.close(); 7778 } 7779 7780 // Write SPIR-V out to a text file with 32-bit hexadecimal words 7781 void OutputSpvHex(const std::vector<unsigned int>& spirv, const char* baseName, const char* varName) 7782 { 7783 std::ofstream out; 7784 out.open(baseName, std::ios::binary | std::ios::out); 7785 if (out.fail()) 7786 printf("ERROR: Failed to open file: %s\n", baseName); 7787 out << "\t// " << 7788 GetSpirvGeneratorVersion() << "." << GLSLANG_MINOR_VERSION << "." << GLSLANG_PATCH_LEVEL << 7789 std::endl; 7790 if (varName != nullptr) { 7791 out << "\t #pragma once" << std::endl; 7792 out << "const uint32_t " << varName << "[] = {" << std::endl; 7793 } 7794 const int WORDS_PER_LINE = 8; 7795 for (int i = 0; i < (int)spirv.size(); i += WORDS_PER_LINE) { 7796 out << "\t"; 7797 for (int j = 0; j < WORDS_PER_LINE && i + j < (int)spirv.size(); ++j) { 7798 const unsigned int word = spirv[i + j]; 7799 out << "0x" << std::hex << std::setw(8) << std::setfill('0') << word; 7800 if (i + j + 1 < (int)spirv.size()) { 7801 out << ","; 7802 } 7803 } 7804 out << std::endl; 7805 } 7806 if (varName != nullptr) { 7807 out << "};"; 7808 } 7809 out.close(); 7810 } 7811 7812 // 7813 // Set up the glslang traversal 7814 // 7815 void GlslangToSpv(const TIntermediate& intermediate, std::vector<unsigned int>& spirv, SpvOptions* options) 7816 { 7817 spv::SpvBuildLogger logger; 7818 GlslangToSpv(intermediate, spirv, &logger, options); 7819 } 7820 7821 void GlslangToSpv(const TIntermediate& intermediate, std::vector<unsigned int>& spirv, 7822 spv::SpvBuildLogger* logger, SpvOptions* options) 7823 { 7824 TIntermNode* root = intermediate.getTreeRoot(); 7825 7826 if (root == 0) 7827 return; 7828 7829 SpvOptions defaultOptions; 7830 if (options == nullptr) 7831 options = &defaultOptions; 7832 7833 GetThreadPoolAllocator().push(); 7834 7835 TGlslangToSpvTraverser it(intermediate.getSpv().spv, &intermediate, logger, *options); 7836 root->traverse(&it); 7837 it.finishSpv(); 7838 it.dumpSpv(spirv); 7839 7840 #if ENABLE_OPT 7841 // If from HLSL, run spirv-opt to "legalize" the SPIR-V for Vulkan 7842 // eg. forward and remove memory writes of opaque types. 7843 if ((intermediate.getSource() == EShSourceHlsl || options->optimizeSize) && !options->disableOptimizer) 7844 SpirvToolsLegalize(intermediate, spirv, logger, options); 7845 7846 if (options->validate) 7847 SpirvToolsValidate(intermediate, spirv, logger); 7848 7849 if (options->disassemble) 7850 SpirvToolsDisassemble(std::cout, spirv); 7851 7852 #endif 7853 7854 GetThreadPoolAllocator().pop(); 7855 } 7856 7857 }; // end namespace glslang 7858