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 #include "GLSL.ext.AMD.h"
50 #include "GLSL.ext.NV.h"
51 }
52
53 // Glslang includes
54 #include "../glslang/MachineIndependent/localintermediate.h"
55 #include "../glslang/MachineIndependent/SymbolTable.h"
56 #include "../glslang/Include/Common.h"
57 #include "../glslang/Include/revision.h"
58
59 #include <fstream>
60 #include <iomanip>
61 #include <list>
62 #include <map>
63 #include <stack>
64 #include <string>
65 #include <vector>
66
67 namespace {
68
69 namespace {
70 class SpecConstantOpModeGuard {
71 public:
SpecConstantOpModeGuard(spv::Builder * builder)72 SpecConstantOpModeGuard(spv::Builder* builder)
73 : builder_(builder) {
74 previous_flag_ = builder->isInSpecConstCodeGenMode();
75 }
~SpecConstantOpModeGuard()76 ~SpecConstantOpModeGuard() {
77 previous_flag_ ? builder_->setToSpecConstCodeGenMode()
78 : builder_->setToNormalCodeGenMode();
79 }
turnOnSpecConstantOpMode()80 void turnOnSpecConstantOpMode() {
81 builder_->setToSpecConstCodeGenMode();
82 }
83
84 private:
85 spv::Builder* builder_;
86 bool previous_flag_;
87 };
88
89 struct OpDecorations {
90 public:
OpDecorations__anonf2472eb70111::__anonf2472eb70211::OpDecorations91 OpDecorations(spv::Decoration precision, spv::Decoration noContraction, spv::Decoration nonUniform) :
92 precision(precision)
93 #ifndef GLSLANG_WEB
94 ,
95 noContraction(noContraction),
96 nonUniform(nonUniform)
97 #endif
98 { }
99
100 spv::Decoration precision;
101
102 #ifdef GLSLANG_WEB
addNoContraction__anonf2472eb70111::__anonf2472eb70211::OpDecorations103 void addNoContraction(spv::Builder&, spv::Id) const { }
addNonUniform__anonf2472eb70111::__anonf2472eb70211::OpDecorations104 void addNonUniform(spv::Builder&, spv::Id) const { }
105 #else
addNoContraction__anonf2472eb70111::__anonf2472eb70211::OpDecorations106 void addNoContraction(spv::Builder& builder, spv::Id t) { builder.addDecoration(t, noContraction); }
addNonUniform__anonf2472eb70111::__anonf2472eb70211::OpDecorations107 void addNonUniform(spv::Builder& builder, spv::Id t) { builder.addDecoration(t, nonUniform); }
108 protected:
109 spv::Decoration noContraction;
110 spv::Decoration nonUniform;
111 #endif
112
113 };
114
115 } // namespace
116
117 //
118 // The main holder of information for translating glslang to SPIR-V.
119 //
120 // Derives from the AST walking base class.
121 //
122 class TGlslangToSpvTraverser : public glslang::TIntermTraverser {
123 public:
124 TGlslangToSpvTraverser(unsigned int spvVersion, const glslang::TIntermediate*, spv::SpvBuildLogger* logger,
125 glslang::SpvOptions& options);
~TGlslangToSpvTraverser()126 virtual ~TGlslangToSpvTraverser() { }
127
128 bool visitAggregate(glslang::TVisit, glslang::TIntermAggregate*);
129 bool visitBinary(glslang::TVisit, glslang::TIntermBinary*);
130 void visitConstantUnion(glslang::TIntermConstantUnion*);
131 bool visitSelection(glslang::TVisit, glslang::TIntermSelection*);
132 bool visitSwitch(glslang::TVisit, glslang::TIntermSwitch*);
133 void visitSymbol(glslang::TIntermSymbol* symbol);
134 bool visitUnary(glslang::TVisit, glslang::TIntermUnary*);
135 bool visitLoop(glslang::TVisit, glslang::TIntermLoop*);
136 bool visitBranch(glslang::TVisit visit, glslang::TIntermBranch*);
137
138 void finishSpv();
139 void dumpSpv(std::vector<unsigned int>& out);
140
141 protected:
142 TGlslangToSpvTraverser(TGlslangToSpvTraverser&);
143 TGlslangToSpvTraverser& operator=(TGlslangToSpvTraverser&);
144
145 spv::Decoration TranslateInterpolationDecoration(const glslang::TQualifier& qualifier);
146 spv::Decoration TranslateAuxiliaryStorageDecoration(const glslang::TQualifier& qualifier);
147 spv::Decoration TranslateNonUniformDecoration(const glslang::TQualifier& qualifier);
148 spv::Builder::AccessChain::CoherentFlags TranslateCoherent(const glslang::TType& type);
149 spv::MemoryAccessMask TranslateMemoryAccess(const spv::Builder::AccessChain::CoherentFlags &coherentFlags);
150 spv::ImageOperandsMask TranslateImageOperands(const spv::Builder::AccessChain::CoherentFlags &coherentFlags);
151 spv::Scope TranslateMemoryScope(const spv::Builder::AccessChain::CoherentFlags &coherentFlags);
152 spv::BuiltIn TranslateBuiltInDecoration(glslang::TBuiltInVariable, bool memberDeclaration);
153 spv::ImageFormat TranslateImageFormat(const glslang::TType& type);
154 spv::SelectionControlMask TranslateSelectionControl(const glslang::TIntermSelection&) const;
155 spv::SelectionControlMask TranslateSwitchControl(const glslang::TIntermSwitch&) const;
156 spv::LoopControlMask TranslateLoopControl(const glslang::TIntermLoop&, std::vector<unsigned int>& operands) const;
157 spv::StorageClass TranslateStorageClass(const glslang::TType&);
158 void addIndirectionIndexCapabilities(const glslang::TType& baseType, const glslang::TType& indexType);
159 spv::Id createSpvVariable(const glslang::TIntermSymbol*, spv::Id forcedType);
160 spv::Id getSampledType(const glslang::TSampler&);
161 spv::Id getInvertedSwizzleType(const glslang::TIntermTyped&);
162 spv::Id createInvertedSwizzle(spv::Decoration precision, const glslang::TIntermTyped&, spv::Id parentResult);
163 void convertSwizzle(const glslang::TIntermAggregate&, std::vector<unsigned>& swizzle);
164 spv::Id convertGlslangToSpvType(const glslang::TType& type, bool forwardReferenceOnly = false);
165 spv::Id convertGlslangToSpvType(const glslang::TType& type, glslang::TLayoutPacking, const glslang::TQualifier&,
166 bool lastBufferBlockMember, bool forwardReferenceOnly = false);
167 bool filterMember(const glslang::TType& member);
168 spv::Id convertGlslangStructToSpvType(const glslang::TType&, const glslang::TTypeList* glslangStruct,
169 glslang::TLayoutPacking, const glslang::TQualifier&);
170 void decorateStructType(const glslang::TType&, const glslang::TTypeList* glslangStruct, glslang::TLayoutPacking,
171 const glslang::TQualifier&, spv::Id);
172 spv::Id makeArraySizeId(const glslang::TArraySizes&, int dim);
173 spv::Id accessChainLoad(const glslang::TType& type);
174 void accessChainStore(const glslang::TType& type, spv::Id rvalue);
175 void multiTypeStore(const glslang::TType&, spv::Id rValue);
176 glslang::TLayoutPacking getExplicitLayout(const glslang::TType& type) const;
177 int getArrayStride(const glslang::TType& arrayType, glslang::TLayoutPacking, glslang::TLayoutMatrix);
178 int getMatrixStride(const glslang::TType& matrixType, glslang::TLayoutPacking, glslang::TLayoutMatrix);
179 void updateMemberOffset(const glslang::TType& structType, const glslang::TType& memberType, int& currentOffset,
180 int& nextOffset, glslang::TLayoutPacking, glslang::TLayoutMatrix);
181 void declareUseOfStructMember(const glslang::TTypeList& members, int glslangMember);
182
183 bool isShaderEntryPoint(const glslang::TIntermAggregate* node);
184 bool writableParam(glslang::TStorageQualifier) const;
185 bool originalParam(glslang::TStorageQualifier, const glslang::TType&, bool implicitThisParam);
186 void makeFunctions(const glslang::TIntermSequence&);
187 void makeGlobalInitializers(const glslang::TIntermSequence&);
188 void visitFunctions(const glslang::TIntermSequence&);
189 void handleFunctionEntry(const glslang::TIntermAggregate* node);
190 void translateArguments(const glslang::TIntermAggregate& node, std::vector<spv::Id>& arguments, spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags);
191 void translateArguments(glslang::TIntermUnary& node, std::vector<spv::Id>& arguments);
192 spv::Id createImageTextureFunctionCall(glslang::TIntermOperator* node);
193 spv::Id handleUserFunctionCall(const glslang::TIntermAggregate*);
194
195 spv::Id createBinaryOperation(glslang::TOperator op, OpDecorations&, spv::Id typeId, spv::Id left, spv::Id right,
196 glslang::TBasicType typeProxy, bool reduceComparison = true);
197 spv::Id createBinaryMatrixOperation(spv::Op, OpDecorations&, spv::Id typeId, spv::Id left, spv::Id right);
198 spv::Id createUnaryOperation(glslang::TOperator op, OpDecorations&, spv::Id typeId, spv::Id operand,
199 glslang::TBasicType typeProxy, const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags);
200 spv::Id createUnaryMatrixOperation(spv::Op op, OpDecorations&, spv::Id typeId, spv::Id operand,
201 glslang::TBasicType typeProxy);
202 spv::Id createConversion(glslang::TOperator op, OpDecorations&, spv::Id destTypeId, spv::Id operand,
203 glslang::TBasicType typeProxy);
204 spv::Id createIntWidthConversion(glslang::TOperator op, spv::Id operand, int vectorSize);
205 spv::Id makeSmearedConstant(spv::Id constant, int vectorSize);
206 spv::Id createAtomicOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy, const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags);
207 spv::Id createInvocationsOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy);
208 spv::Id CreateInvocationsVectorOperation(spv::Op op, spv::GroupOperation groupOperation, spv::Id typeId, std::vector<spv::Id>& operands);
209 spv::Id createSubgroupOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy);
210 spv::Id createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy);
211 spv::Id createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId);
212 spv::Id getSymbolId(const glslang::TIntermSymbol* node);
213 void addMeshNVDecoration(spv::Id id, int member, const glslang::TQualifier & qualifier);
214 spv::Id createSpvConstant(const glslang::TIntermTyped&);
215 spv::Id createSpvConstantFromConstUnionArray(const glslang::TType& type, const glslang::TConstUnionArray&, int& nextConst, bool specConstant);
216 bool isTrivialLeaf(const glslang::TIntermTyped* node);
217 bool isTrivial(const glslang::TIntermTyped* node);
218 spv::Id createShortCircuit(glslang::TOperator, glslang::TIntermTyped& left, glslang::TIntermTyped& right);
219 spv::Id getExtBuiltins(const char* name);
220 std::pair<spv::Id, spv::Id> getForcedType(spv::BuiltIn, const glslang::TType&);
221 spv::Id translateForcedType(spv::Id object);
222 spv::Id createCompositeConstruct(spv::Id typeId, std::vector<spv::Id> constituents);
223
224 glslang::SpvOptions& options;
225 spv::Function* shaderEntry;
226 spv::Function* currentFunction;
227 spv::Instruction* entryPoint;
228 int sequenceDepth;
229
230 spv::SpvBuildLogger* logger;
231
232 // There is a 1:1 mapping between a spv builder and a module; this is thread safe
233 spv::Builder builder;
234 bool inEntryPoint;
235 bool entryPointTerminated;
236 bool linkageOnly; // true when visiting the set of objects in the AST present only for establishing interface, whether or not they were statically used
237 std::set<spv::Id> iOSet; // all input/output variables from either static use or declaration of interface
238 const glslang::TIntermediate* glslangIntermediate;
239 bool nanMinMaxClamp; // true if use NMin/NMax/NClamp instead of FMin/FMax/FClamp
240 spv::Id stdBuiltins;
241 std::unordered_map<const char*, spv::Id> extBuiltinMap;
242
243 std::unordered_map<int, spv::Id> symbolValues;
244 std::unordered_set<int> rValueParameters; // set of formal function parameters passed as rValues, rather than a pointer
245 std::unordered_map<std::string, spv::Function*> functionMap;
246 std::unordered_map<const glslang::TTypeList*, spv::Id> structMap[glslang::ElpCount][glslang::ElmCount];
247 // for mapping glslang block indices to spv indices (e.g., due to hidden members):
248 std::unordered_map<const glslang::TTypeList*, std::vector<int> > memberRemapper;
249 std::stack<bool> breakForLoop; // false means break for switch
250 std::unordered_map<std::string, const glslang::TIntermSymbol*> counterOriginator;
251 // Map pointee types for EbtReference to their forward pointers
252 std::map<const glslang::TType *, spv::Id> forwardPointers;
253 // Type forcing, for when SPIR-V wants a different type than the AST,
254 // requiring local translation to and from SPIR-V type on every access.
255 // Maps <builtin-variable-id -> AST-required-type-id>
256 std::unordered_map<spv::Id, spv::Id> forceType;
257 };
258
259 //
260 // Helper functions for translating glslang representations to SPIR-V enumerants.
261 //
262
263 // Translate glslang profile to SPIR-V source language.
TranslateSourceLanguage(glslang::EShSource source,EProfile profile)264 spv::SourceLanguage TranslateSourceLanguage(glslang::EShSource source, EProfile profile)
265 {
266 #ifdef GLSLANG_WEB
267 return spv::SourceLanguageESSL;
268 #endif
269
270 switch (source) {
271 case glslang::EShSourceGlsl:
272 switch (profile) {
273 case ENoProfile:
274 case ECoreProfile:
275 case ECompatibilityProfile:
276 return spv::SourceLanguageGLSL;
277 case EEsProfile:
278 return spv::SourceLanguageESSL;
279 default:
280 return spv::SourceLanguageUnknown;
281 }
282 case glslang::EShSourceHlsl:
283 return spv::SourceLanguageHLSL;
284 default:
285 return spv::SourceLanguageUnknown;
286 }
287 }
288
289 // Translate glslang language (stage) to SPIR-V execution model.
TranslateExecutionModel(EShLanguage stage)290 spv::ExecutionModel TranslateExecutionModel(EShLanguage stage)
291 {
292 switch (stage) {
293 case EShLangVertex: return spv::ExecutionModelVertex;
294 case EShLangFragment: return spv::ExecutionModelFragment;
295 case EShLangCompute: return spv::ExecutionModelGLCompute;
296 #ifndef GLSLANG_WEB
297 case EShLangTessControl: return spv::ExecutionModelTessellationControl;
298 case EShLangTessEvaluation: return spv::ExecutionModelTessellationEvaluation;
299 case EShLangGeometry: return spv::ExecutionModelGeometry;
300 case EShLangRayGenNV: return spv::ExecutionModelRayGenerationNV;
301 case EShLangIntersectNV: return spv::ExecutionModelIntersectionNV;
302 case EShLangAnyHitNV: return spv::ExecutionModelAnyHitNV;
303 case EShLangClosestHitNV: return spv::ExecutionModelClosestHitNV;
304 case EShLangMissNV: return spv::ExecutionModelMissNV;
305 case EShLangCallableNV: return spv::ExecutionModelCallableNV;
306 case EShLangTaskNV: return spv::ExecutionModelTaskNV;
307 case EShLangMeshNV: return spv::ExecutionModelMeshNV;
308 #endif
309 default:
310 assert(0);
311 return spv::ExecutionModelFragment;
312 }
313 }
314
315 // Translate glslang sampler type to SPIR-V dimensionality.
TranslateDimensionality(const glslang::TSampler & sampler)316 spv::Dim TranslateDimensionality(const glslang::TSampler& sampler)
317 {
318 switch (sampler.dim) {
319 case glslang::Esd1D: return spv::Dim1D;
320 case glslang::Esd2D: return spv::Dim2D;
321 case glslang::Esd3D: return spv::Dim3D;
322 case glslang::EsdCube: return spv::DimCube;
323 case glslang::EsdRect: return spv::DimRect;
324 case glslang::EsdBuffer: return spv::DimBuffer;
325 case glslang::EsdSubpass: return spv::DimSubpassData;
326 default:
327 assert(0);
328 return spv::Dim2D;
329 }
330 }
331
332 // Translate glslang precision to SPIR-V precision decorations.
TranslatePrecisionDecoration(glslang::TPrecisionQualifier glslangPrecision)333 spv::Decoration TranslatePrecisionDecoration(glslang::TPrecisionQualifier glslangPrecision)
334 {
335 switch (glslangPrecision) {
336 case glslang::EpqLow: return spv::DecorationRelaxedPrecision;
337 case glslang::EpqMedium: return spv::DecorationRelaxedPrecision;
338 default:
339 return spv::NoPrecision;
340 }
341 }
342
343 // Translate glslang type to SPIR-V precision decorations.
TranslatePrecisionDecoration(const glslang::TType & type)344 spv::Decoration TranslatePrecisionDecoration(const glslang::TType& type)
345 {
346 return TranslatePrecisionDecoration(type.getQualifier().precision);
347 }
348
349 // Translate glslang type to SPIR-V block decorations.
TranslateBlockDecoration(const glslang::TType & type,bool useStorageBuffer)350 spv::Decoration TranslateBlockDecoration(const glslang::TType& type, bool useStorageBuffer)
351 {
352 if (type.getBasicType() == glslang::EbtBlock) {
353 switch (type.getQualifier().storage) {
354 case glslang::EvqUniform: return spv::DecorationBlock;
355 case glslang::EvqBuffer: return useStorageBuffer ? spv::DecorationBlock : spv::DecorationBufferBlock;
356 case glslang::EvqVaryingIn: return spv::DecorationBlock;
357 case glslang::EvqVaryingOut: return spv::DecorationBlock;
358 #ifndef GLSLANG_WEB
359 case glslang::EvqPayloadNV: return spv::DecorationBlock;
360 case glslang::EvqPayloadInNV: return spv::DecorationBlock;
361 case glslang::EvqHitAttrNV: return spv::DecorationBlock;
362 case glslang::EvqCallableDataNV: return spv::DecorationBlock;
363 case glslang::EvqCallableDataInNV: return spv::DecorationBlock;
364 #endif
365 default:
366 assert(0);
367 break;
368 }
369 }
370
371 return spv::DecorationMax;
372 }
373
374 // Translate glslang type to SPIR-V memory decorations.
TranslateMemoryDecoration(const glslang::TQualifier & qualifier,std::vector<spv::Decoration> & memory,bool useVulkanMemoryModel)375 void TranslateMemoryDecoration(const glslang::TQualifier& qualifier, std::vector<spv::Decoration>& memory, bool useVulkanMemoryModel)
376 {
377 if (!useVulkanMemoryModel) {
378 if (qualifier.isCoherent())
379 memory.push_back(spv::DecorationCoherent);
380 if (qualifier.isVolatile()) {
381 memory.push_back(spv::DecorationVolatile);
382 memory.push_back(spv::DecorationCoherent);
383 }
384 }
385 if (qualifier.isRestrict())
386 memory.push_back(spv::DecorationRestrict);
387 if (qualifier.isReadOnly())
388 memory.push_back(spv::DecorationNonWritable);
389 if (qualifier.isWriteOnly())
390 memory.push_back(spv::DecorationNonReadable);
391 }
392
393 // Translate glslang type to SPIR-V layout decorations.
TranslateLayoutDecoration(const glslang::TType & type,glslang::TLayoutMatrix matrixLayout)394 spv::Decoration TranslateLayoutDecoration(const glslang::TType& type, glslang::TLayoutMatrix matrixLayout)
395 {
396 if (type.isMatrix()) {
397 switch (matrixLayout) {
398 case glslang::ElmRowMajor:
399 return spv::DecorationRowMajor;
400 case glslang::ElmColumnMajor:
401 return spv::DecorationColMajor;
402 default:
403 // opaque layouts don't need a majorness
404 return spv::DecorationMax;
405 }
406 } else {
407 switch (type.getBasicType()) {
408 default:
409 return spv::DecorationMax;
410 break;
411 case glslang::EbtBlock:
412 switch (type.getQualifier().storage) {
413 case glslang::EvqUniform:
414 case glslang::EvqBuffer:
415 switch (type.getQualifier().layoutPacking) {
416 case glslang::ElpShared: return spv::DecorationGLSLShared;
417 case glslang::ElpPacked: return spv::DecorationGLSLPacked;
418 default:
419 return spv::DecorationMax;
420 }
421 case glslang::EvqVaryingIn:
422 case glslang::EvqVaryingOut:
423 if (type.getQualifier().isTaskMemory()) {
424 switch (type.getQualifier().layoutPacking) {
425 case glslang::ElpShared: return spv::DecorationGLSLShared;
426 case glslang::ElpPacked: return spv::DecorationGLSLPacked;
427 default: break;
428 }
429 } else {
430 assert(type.getQualifier().layoutPacking == glslang::ElpNone);
431 }
432 return spv::DecorationMax;
433 #ifndef GLSLANG_WEB
434 case glslang::EvqPayloadNV:
435 case glslang::EvqPayloadInNV:
436 case glslang::EvqHitAttrNV:
437 case glslang::EvqCallableDataNV:
438 case glslang::EvqCallableDataInNV:
439 return spv::DecorationMax;
440 #endif
441 default:
442 assert(0);
443 return spv::DecorationMax;
444 }
445 }
446 }
447 }
448
449 // Translate glslang type to SPIR-V interpolation decorations.
450 // Returns spv::DecorationMax when no decoration
451 // should be applied.
TranslateInterpolationDecoration(const glslang::TQualifier & qualifier)452 spv::Decoration TGlslangToSpvTraverser::TranslateInterpolationDecoration(const glslang::TQualifier& qualifier)
453 {
454 if (qualifier.smooth)
455 // Smooth decoration doesn't exist in SPIR-V 1.0
456 return spv::DecorationMax;
457 else if (qualifier.isNonPerspective())
458 return spv::DecorationNoPerspective;
459 else if (qualifier.flat)
460 return spv::DecorationFlat;
461 else if (qualifier.isExplicitInterpolation()) {
462 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
463 return spv::DecorationExplicitInterpAMD;
464 }
465 else
466 return spv::DecorationMax;
467 }
468
469 // Translate glslang type to SPIR-V auxiliary storage decorations.
470 // Returns spv::DecorationMax when no decoration
471 // should be applied.
TranslateAuxiliaryStorageDecoration(const glslang::TQualifier & qualifier)472 spv::Decoration TGlslangToSpvTraverser::TranslateAuxiliaryStorageDecoration(const glslang::TQualifier& qualifier)
473 {
474 if (qualifier.centroid)
475 return spv::DecorationCentroid;
476 #ifndef GLSLANG_WEB
477 else if (qualifier.patch)
478 return spv::DecorationPatch;
479 else if (qualifier.sample) {
480 builder.addCapability(spv::CapabilitySampleRateShading);
481 return spv::DecorationSample;
482 }
483 #endif
484
485 return spv::DecorationMax;
486 }
487
488 // If glslang type is invariant, return SPIR-V invariant decoration.
TranslateInvariantDecoration(const glslang::TQualifier & qualifier)489 spv::Decoration TranslateInvariantDecoration(const glslang::TQualifier& qualifier)
490 {
491 if (qualifier.invariant)
492 return spv::DecorationInvariant;
493 else
494 return spv::DecorationMax;
495 }
496
497 // If glslang type is noContraction, return SPIR-V NoContraction decoration.
TranslateNoContractionDecoration(const glslang::TQualifier & qualifier)498 spv::Decoration TranslateNoContractionDecoration(const glslang::TQualifier& qualifier)
499 {
500 #ifndef GLSLANG_WEB
501 if (qualifier.isNoContraction())
502 return spv::DecorationNoContraction;
503 else
504 #endif
505 return spv::DecorationMax;
506 }
507
508 // If glslang type is nonUniform, return SPIR-V NonUniform decoration.
TranslateNonUniformDecoration(const glslang::TQualifier & qualifier)509 spv::Decoration TGlslangToSpvTraverser::TranslateNonUniformDecoration(const glslang::TQualifier& qualifier)
510 {
511 #ifndef GLSLANG_WEB
512 if (qualifier.isNonUniform()) {
513 builder.addIncorporatedExtension("SPV_EXT_descriptor_indexing", spv::Spv_1_5);
514 builder.addCapability(spv::CapabilityShaderNonUniformEXT);
515 return spv::DecorationNonUniformEXT;
516 } else
517 #endif
518 return spv::DecorationMax;
519 }
520
TranslateMemoryAccess(const spv::Builder::AccessChain::CoherentFlags & coherentFlags)521 spv::MemoryAccessMask TGlslangToSpvTraverser::TranslateMemoryAccess(
522 const spv::Builder::AccessChain::CoherentFlags &coherentFlags)
523 {
524 spv::MemoryAccessMask mask = spv::MemoryAccessMaskNone;
525
526 #ifndef GLSLANG_WEB
527 if (!glslangIntermediate->usingVulkanMemoryModel() || coherentFlags.isImage)
528 return mask;
529
530 if (coherentFlags.volatil ||
531 coherentFlags.coherent ||
532 coherentFlags.devicecoherent ||
533 coherentFlags.queuefamilycoherent ||
534 coherentFlags.workgroupcoherent ||
535 coherentFlags.subgroupcoherent) {
536 mask = mask | spv::MemoryAccessMakePointerAvailableKHRMask |
537 spv::MemoryAccessMakePointerVisibleKHRMask;
538 }
539 if (coherentFlags.nonprivate) {
540 mask = mask | spv::MemoryAccessNonPrivatePointerKHRMask;
541 }
542 if (coherentFlags.volatil) {
543 mask = mask | spv::MemoryAccessVolatileMask;
544 }
545 if (mask != spv::MemoryAccessMaskNone) {
546 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
547 }
548 #endif
549
550 return mask;
551 }
552
TranslateImageOperands(const spv::Builder::AccessChain::CoherentFlags & coherentFlags)553 spv::ImageOperandsMask TGlslangToSpvTraverser::TranslateImageOperands(
554 const spv::Builder::AccessChain::CoherentFlags &coherentFlags)
555 {
556 spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone;
557
558 #ifndef GLSLANG_WEB
559 if (!glslangIntermediate->usingVulkanMemoryModel())
560 return mask;
561
562 if (coherentFlags.volatil ||
563 coherentFlags.coherent ||
564 coherentFlags.devicecoherent ||
565 coherentFlags.queuefamilycoherent ||
566 coherentFlags.workgroupcoherent ||
567 coherentFlags.subgroupcoherent) {
568 mask = mask | spv::ImageOperandsMakeTexelAvailableKHRMask |
569 spv::ImageOperandsMakeTexelVisibleKHRMask;
570 }
571 if (coherentFlags.nonprivate) {
572 mask = mask | spv::ImageOperandsNonPrivateTexelKHRMask;
573 }
574 if (coherentFlags.volatil) {
575 mask = mask | spv::ImageOperandsVolatileTexelKHRMask;
576 }
577 if (mask != spv::ImageOperandsMaskNone) {
578 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
579 }
580 #endif
581
582 return mask;
583 }
584
TranslateCoherent(const glslang::TType & type)585 spv::Builder::AccessChain::CoherentFlags TGlslangToSpvTraverser::TranslateCoherent(const glslang::TType& type)
586 {
587 spv::Builder::AccessChain::CoherentFlags flags = {};
588 #ifndef GLSLANG_WEB
589 flags.coherent = type.getQualifier().coherent;
590 flags.devicecoherent = type.getQualifier().devicecoherent;
591 flags.queuefamilycoherent = type.getQualifier().queuefamilycoherent;
592 // shared variables are implicitly workgroupcoherent in GLSL.
593 flags.workgroupcoherent = type.getQualifier().workgroupcoherent ||
594 type.getQualifier().storage == glslang::EvqShared;
595 flags.subgroupcoherent = type.getQualifier().subgroupcoherent;
596 flags.volatil = type.getQualifier().volatil;
597 // *coherent variables are implicitly nonprivate in GLSL
598 flags.nonprivate = type.getQualifier().nonprivate ||
599 flags.subgroupcoherent ||
600 flags.workgroupcoherent ||
601 flags.queuefamilycoherent ||
602 flags.devicecoherent ||
603 flags.coherent ||
604 flags.volatil;
605 flags.isImage = type.getBasicType() == glslang::EbtSampler;
606 #endif
607 return flags;
608 }
609
TranslateMemoryScope(const spv::Builder::AccessChain::CoherentFlags & coherentFlags)610 spv::Scope TGlslangToSpvTraverser::TranslateMemoryScope(
611 const spv::Builder::AccessChain::CoherentFlags &coherentFlags)
612 {
613 spv::Scope scope = spv::ScopeMax;
614
615 #ifndef GLSLANG_WEB
616 if (coherentFlags.volatil || coherentFlags.coherent) {
617 // coherent defaults to Device scope in the old model, QueueFamilyKHR scope in the new model
618 scope = glslangIntermediate->usingVulkanMemoryModel() ? spv::ScopeQueueFamilyKHR : spv::ScopeDevice;
619 } else if (coherentFlags.devicecoherent) {
620 scope = spv::ScopeDevice;
621 } else if (coherentFlags.queuefamilycoherent) {
622 scope = spv::ScopeQueueFamilyKHR;
623 } else if (coherentFlags.workgroupcoherent) {
624 scope = spv::ScopeWorkgroup;
625 } else if (coherentFlags.subgroupcoherent) {
626 scope = spv::ScopeSubgroup;
627 }
628 if (glslangIntermediate->usingVulkanMemoryModel() && scope == spv::ScopeDevice) {
629 builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR);
630 }
631 #endif
632
633 return scope;
634 }
635
636 // Translate a glslang built-in variable to a SPIR-V built in decoration. Also generate
637 // associated capabilities when required. For some built-in variables, a capability
638 // is generated only when using the variable in an executable instruction, but not when
639 // just declaring a struct member variable with it. This is true for PointSize,
640 // ClipDistance, and CullDistance.
TranslateBuiltInDecoration(glslang::TBuiltInVariable builtIn,bool memberDeclaration)641 spv::BuiltIn TGlslangToSpvTraverser::TranslateBuiltInDecoration(glslang::TBuiltInVariable builtIn, bool memberDeclaration)
642 {
643 switch (builtIn) {
644 case glslang::EbvPointSize:
645 #ifndef GLSLANG_WEB
646 // Defer adding the capability until the built-in is actually used.
647 if (! memberDeclaration) {
648 switch (glslangIntermediate->getStage()) {
649 case EShLangGeometry:
650 builder.addCapability(spv::CapabilityGeometryPointSize);
651 break;
652 case EShLangTessControl:
653 case EShLangTessEvaluation:
654 builder.addCapability(spv::CapabilityTessellationPointSize);
655 break;
656 default:
657 break;
658 }
659 }
660 #endif
661 return spv::BuiltInPointSize;
662
663 case glslang::EbvPosition: return spv::BuiltInPosition;
664 case glslang::EbvVertexId: return spv::BuiltInVertexId;
665 case glslang::EbvInstanceId: return spv::BuiltInInstanceId;
666 case glslang::EbvVertexIndex: return spv::BuiltInVertexIndex;
667 case glslang::EbvInstanceIndex: return spv::BuiltInInstanceIndex;
668
669 case glslang::EbvFragCoord: return spv::BuiltInFragCoord;
670 case glslang::EbvPointCoord: return spv::BuiltInPointCoord;
671 case glslang::EbvFace: return spv::BuiltInFrontFacing;
672 case glslang::EbvFragDepth: return spv::BuiltInFragDepth;
673
674 case glslang::EbvNumWorkGroups: return spv::BuiltInNumWorkgroups;
675 case glslang::EbvWorkGroupSize: return spv::BuiltInWorkgroupSize;
676 case glslang::EbvWorkGroupId: return spv::BuiltInWorkgroupId;
677 case glslang::EbvLocalInvocationId: return spv::BuiltInLocalInvocationId;
678 case glslang::EbvLocalInvocationIndex: return spv::BuiltInLocalInvocationIndex;
679 case glslang::EbvGlobalInvocationId: return spv::BuiltInGlobalInvocationId;
680
681 #ifndef GLSLANG_WEB
682 // These *Distance capabilities logically belong here, but if the member is declared and
683 // then never used, consumers of SPIR-V prefer the capability not be declared.
684 // They are now generated when used, rather than here when declared.
685 // Potentially, the specification should be more clear what the minimum
686 // use needed is to trigger the capability.
687 //
688 case glslang::EbvClipDistance:
689 if (!memberDeclaration)
690 builder.addCapability(spv::CapabilityClipDistance);
691 return spv::BuiltInClipDistance;
692
693 case glslang::EbvCullDistance:
694 if (!memberDeclaration)
695 builder.addCapability(spv::CapabilityCullDistance);
696 return spv::BuiltInCullDistance;
697
698 case glslang::EbvViewportIndex:
699 builder.addCapability(spv::CapabilityMultiViewport);
700 if (glslangIntermediate->getStage() == EShLangVertex ||
701 glslangIntermediate->getStage() == EShLangTessControl ||
702 glslangIntermediate->getStage() == EShLangTessEvaluation) {
703
704 builder.addIncorporatedExtension(spv::E_SPV_EXT_shader_viewport_index_layer, spv::Spv_1_5);
705 builder.addCapability(spv::CapabilityShaderViewportIndexLayerEXT);
706 }
707 return spv::BuiltInViewportIndex;
708
709 case glslang::EbvSampleId:
710 builder.addCapability(spv::CapabilitySampleRateShading);
711 return spv::BuiltInSampleId;
712
713 case glslang::EbvSamplePosition:
714 builder.addCapability(spv::CapabilitySampleRateShading);
715 return spv::BuiltInSamplePosition;
716
717 case glslang::EbvSampleMask:
718 return spv::BuiltInSampleMask;
719
720 case glslang::EbvLayer:
721 if (glslangIntermediate->getStage() == EShLangMeshNV) {
722 return spv::BuiltInLayer;
723 }
724 builder.addCapability(spv::CapabilityGeometry);
725 if (glslangIntermediate->getStage() == EShLangVertex ||
726 glslangIntermediate->getStage() == EShLangTessControl ||
727 glslangIntermediate->getStage() == EShLangTessEvaluation) {
728
729 builder.addIncorporatedExtension(spv::E_SPV_EXT_shader_viewport_index_layer, spv::Spv_1_5);
730 builder.addCapability(spv::CapabilityShaderViewportIndexLayerEXT);
731 }
732 return spv::BuiltInLayer;
733
734 case glslang::EbvBaseVertex:
735 builder.addIncorporatedExtension(spv::E_SPV_KHR_shader_draw_parameters, spv::Spv_1_3);
736 builder.addCapability(spv::CapabilityDrawParameters);
737 return spv::BuiltInBaseVertex;
738
739 case glslang::EbvBaseInstance:
740 builder.addIncorporatedExtension(spv::E_SPV_KHR_shader_draw_parameters, spv::Spv_1_3);
741 builder.addCapability(spv::CapabilityDrawParameters);
742 return spv::BuiltInBaseInstance;
743
744 case glslang::EbvDrawId:
745 builder.addIncorporatedExtension(spv::E_SPV_KHR_shader_draw_parameters, spv::Spv_1_3);
746 builder.addCapability(spv::CapabilityDrawParameters);
747 return spv::BuiltInDrawIndex;
748
749 case glslang::EbvPrimitiveId:
750 if (glslangIntermediate->getStage() == EShLangFragment)
751 builder.addCapability(spv::CapabilityGeometry);
752 return spv::BuiltInPrimitiveId;
753
754 case glslang::EbvFragStencilRef:
755 builder.addExtension(spv::E_SPV_EXT_shader_stencil_export);
756 builder.addCapability(spv::CapabilityStencilExportEXT);
757 return spv::BuiltInFragStencilRefEXT;
758
759 case glslang::EbvInvocationId: return spv::BuiltInInvocationId;
760 case glslang::EbvTessLevelInner: return spv::BuiltInTessLevelInner;
761 case glslang::EbvTessLevelOuter: return spv::BuiltInTessLevelOuter;
762 case glslang::EbvTessCoord: return spv::BuiltInTessCoord;
763 case glslang::EbvPatchVertices: return spv::BuiltInPatchVertices;
764 case glslang::EbvHelperInvocation: return spv::BuiltInHelperInvocation;
765
766 case glslang::EbvSubGroupSize:
767 builder.addExtension(spv::E_SPV_KHR_shader_ballot);
768 builder.addCapability(spv::CapabilitySubgroupBallotKHR);
769 return spv::BuiltInSubgroupSize;
770
771 case glslang::EbvSubGroupInvocation:
772 builder.addExtension(spv::E_SPV_KHR_shader_ballot);
773 builder.addCapability(spv::CapabilitySubgroupBallotKHR);
774 return spv::BuiltInSubgroupLocalInvocationId;
775
776 case glslang::EbvSubGroupEqMask:
777 builder.addExtension(spv::E_SPV_KHR_shader_ballot);
778 builder.addCapability(spv::CapabilitySubgroupBallotKHR);
779 return spv::BuiltInSubgroupEqMask;
780
781 case glslang::EbvSubGroupGeMask:
782 builder.addExtension(spv::E_SPV_KHR_shader_ballot);
783 builder.addCapability(spv::CapabilitySubgroupBallotKHR);
784 return spv::BuiltInSubgroupGeMask;
785
786 case glslang::EbvSubGroupGtMask:
787 builder.addExtension(spv::E_SPV_KHR_shader_ballot);
788 builder.addCapability(spv::CapabilitySubgroupBallotKHR);
789 return spv::BuiltInSubgroupGtMask;
790
791 case glslang::EbvSubGroupLeMask:
792 builder.addExtension(spv::E_SPV_KHR_shader_ballot);
793 builder.addCapability(spv::CapabilitySubgroupBallotKHR);
794 return spv::BuiltInSubgroupLeMask;
795
796 case glslang::EbvSubGroupLtMask:
797 builder.addExtension(spv::E_SPV_KHR_shader_ballot);
798 builder.addCapability(spv::CapabilitySubgroupBallotKHR);
799 return spv::BuiltInSubgroupLtMask;
800
801 case glslang::EbvNumSubgroups:
802 builder.addCapability(spv::CapabilityGroupNonUniform);
803 return spv::BuiltInNumSubgroups;
804
805 case glslang::EbvSubgroupID:
806 builder.addCapability(spv::CapabilityGroupNonUniform);
807 return spv::BuiltInSubgroupId;
808
809 case glslang::EbvSubgroupSize2:
810 builder.addCapability(spv::CapabilityGroupNonUniform);
811 return spv::BuiltInSubgroupSize;
812
813 case glslang::EbvSubgroupInvocation2:
814 builder.addCapability(spv::CapabilityGroupNonUniform);
815 return spv::BuiltInSubgroupLocalInvocationId;
816
817 case glslang::EbvSubgroupEqMask2:
818 builder.addCapability(spv::CapabilityGroupNonUniform);
819 builder.addCapability(spv::CapabilityGroupNonUniformBallot);
820 return spv::BuiltInSubgroupEqMask;
821
822 case glslang::EbvSubgroupGeMask2:
823 builder.addCapability(spv::CapabilityGroupNonUniform);
824 builder.addCapability(spv::CapabilityGroupNonUniformBallot);
825 return spv::BuiltInSubgroupGeMask;
826
827 case glslang::EbvSubgroupGtMask2:
828 builder.addCapability(spv::CapabilityGroupNonUniform);
829 builder.addCapability(spv::CapabilityGroupNonUniformBallot);
830 return spv::BuiltInSubgroupGtMask;
831
832 case glslang::EbvSubgroupLeMask2:
833 builder.addCapability(spv::CapabilityGroupNonUniform);
834 builder.addCapability(spv::CapabilityGroupNonUniformBallot);
835 return spv::BuiltInSubgroupLeMask;
836
837 case glslang::EbvSubgroupLtMask2:
838 builder.addCapability(spv::CapabilityGroupNonUniform);
839 builder.addCapability(spv::CapabilityGroupNonUniformBallot);
840 return spv::BuiltInSubgroupLtMask;
841
842 case glslang::EbvBaryCoordNoPersp:
843 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
844 return spv::BuiltInBaryCoordNoPerspAMD;
845
846 case glslang::EbvBaryCoordNoPerspCentroid:
847 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
848 return spv::BuiltInBaryCoordNoPerspCentroidAMD;
849
850 case glslang::EbvBaryCoordNoPerspSample:
851 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
852 return spv::BuiltInBaryCoordNoPerspSampleAMD;
853
854 case glslang::EbvBaryCoordSmooth:
855 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
856 return spv::BuiltInBaryCoordSmoothAMD;
857
858 case glslang::EbvBaryCoordSmoothCentroid:
859 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
860 return spv::BuiltInBaryCoordSmoothCentroidAMD;
861
862 case glslang::EbvBaryCoordSmoothSample:
863 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
864 return spv::BuiltInBaryCoordSmoothSampleAMD;
865
866 case glslang::EbvBaryCoordPullModel:
867 builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
868 return spv::BuiltInBaryCoordPullModelAMD;
869
870 case glslang::EbvDeviceIndex:
871 builder.addIncorporatedExtension(spv::E_SPV_KHR_device_group, spv::Spv_1_3);
872 builder.addCapability(spv::CapabilityDeviceGroup);
873 return spv::BuiltInDeviceIndex;
874
875 case glslang::EbvViewIndex:
876 builder.addIncorporatedExtension(spv::E_SPV_KHR_multiview, spv::Spv_1_3);
877 builder.addCapability(spv::CapabilityMultiView);
878 return spv::BuiltInViewIndex;
879
880 case glslang::EbvFragSizeEXT:
881 builder.addExtension(spv::E_SPV_EXT_fragment_invocation_density);
882 builder.addCapability(spv::CapabilityFragmentDensityEXT);
883 return spv::BuiltInFragSizeEXT;
884
885 case glslang::EbvFragInvocationCountEXT:
886 builder.addExtension(spv::E_SPV_EXT_fragment_invocation_density);
887 builder.addCapability(spv::CapabilityFragmentDensityEXT);
888 return spv::BuiltInFragInvocationCountEXT;
889
890 case glslang::EbvViewportMaskNV:
891 if (!memberDeclaration) {
892 builder.addExtension(spv::E_SPV_NV_viewport_array2);
893 builder.addCapability(spv::CapabilityShaderViewportMaskNV);
894 }
895 return spv::BuiltInViewportMaskNV;
896 case glslang::EbvSecondaryPositionNV:
897 if (!memberDeclaration) {
898 builder.addExtension(spv::E_SPV_NV_stereo_view_rendering);
899 builder.addCapability(spv::CapabilityShaderStereoViewNV);
900 }
901 return spv::BuiltInSecondaryPositionNV;
902 case glslang::EbvSecondaryViewportMaskNV:
903 if (!memberDeclaration) {
904 builder.addExtension(spv::E_SPV_NV_stereo_view_rendering);
905 builder.addCapability(spv::CapabilityShaderStereoViewNV);
906 }
907 return spv::BuiltInSecondaryViewportMaskNV;
908 case glslang::EbvPositionPerViewNV:
909 if (!memberDeclaration) {
910 builder.addExtension(spv::E_SPV_NVX_multiview_per_view_attributes);
911 builder.addCapability(spv::CapabilityPerViewAttributesNV);
912 }
913 return spv::BuiltInPositionPerViewNV;
914 case glslang::EbvViewportMaskPerViewNV:
915 if (!memberDeclaration) {
916 builder.addExtension(spv::E_SPV_NVX_multiview_per_view_attributes);
917 builder.addCapability(spv::CapabilityPerViewAttributesNV);
918 }
919 return spv::BuiltInViewportMaskPerViewNV;
920 case glslang::EbvFragFullyCoveredNV:
921 builder.addExtension(spv::E_SPV_EXT_fragment_fully_covered);
922 builder.addCapability(spv::CapabilityFragmentFullyCoveredEXT);
923 return spv::BuiltInFullyCoveredEXT;
924 case glslang::EbvFragmentSizeNV:
925 builder.addExtension(spv::E_SPV_NV_shading_rate);
926 builder.addCapability(spv::CapabilityShadingRateNV);
927 return spv::BuiltInFragmentSizeNV;
928 case glslang::EbvInvocationsPerPixelNV:
929 builder.addExtension(spv::E_SPV_NV_shading_rate);
930 builder.addCapability(spv::CapabilityShadingRateNV);
931 return spv::BuiltInInvocationsPerPixelNV;
932
933 // ray tracing
934 case glslang::EbvLaunchIdNV:
935 return spv::BuiltInLaunchIdNV;
936 case glslang::EbvLaunchSizeNV:
937 return spv::BuiltInLaunchSizeNV;
938 case glslang::EbvWorldRayOriginNV:
939 return spv::BuiltInWorldRayOriginNV;
940 case glslang::EbvWorldRayDirectionNV:
941 return spv::BuiltInWorldRayDirectionNV;
942 case glslang::EbvObjectRayOriginNV:
943 return spv::BuiltInObjectRayOriginNV;
944 case glslang::EbvObjectRayDirectionNV:
945 return spv::BuiltInObjectRayDirectionNV;
946 case glslang::EbvRayTminNV:
947 return spv::BuiltInRayTminNV;
948 case glslang::EbvRayTmaxNV:
949 return spv::BuiltInRayTmaxNV;
950 case glslang::EbvInstanceCustomIndexNV:
951 return spv::BuiltInInstanceCustomIndexNV;
952 case glslang::EbvHitTNV:
953 return spv::BuiltInHitTNV;
954 case glslang::EbvHitKindNV:
955 return spv::BuiltInHitKindNV;
956 case glslang::EbvObjectToWorldNV:
957 return spv::BuiltInObjectToWorldNV;
958 case glslang::EbvWorldToObjectNV:
959 return spv::BuiltInWorldToObjectNV;
960 case glslang::EbvIncomingRayFlagsNV:
961 return spv::BuiltInIncomingRayFlagsNV;
962
963 // barycentrics
964 case glslang::EbvBaryCoordNV:
965 builder.addExtension(spv::E_SPV_NV_fragment_shader_barycentric);
966 builder.addCapability(spv::CapabilityFragmentBarycentricNV);
967 return spv::BuiltInBaryCoordNV;
968 case glslang::EbvBaryCoordNoPerspNV:
969 builder.addExtension(spv::E_SPV_NV_fragment_shader_barycentric);
970 builder.addCapability(spv::CapabilityFragmentBarycentricNV);
971 return spv::BuiltInBaryCoordNoPerspNV;
972
973 // mesh shaders
974 case glslang::EbvTaskCountNV:
975 return spv::BuiltInTaskCountNV;
976 case glslang::EbvPrimitiveCountNV:
977 return spv::BuiltInPrimitiveCountNV;
978 case glslang::EbvPrimitiveIndicesNV:
979 return spv::BuiltInPrimitiveIndicesNV;
980 case glslang::EbvClipDistancePerViewNV:
981 return spv::BuiltInClipDistancePerViewNV;
982 case glslang::EbvCullDistancePerViewNV:
983 return spv::BuiltInCullDistancePerViewNV;
984 case glslang::EbvLayerPerViewNV:
985 return spv::BuiltInLayerPerViewNV;
986 case glslang::EbvMeshViewCountNV:
987 return spv::BuiltInMeshViewCountNV;
988 case glslang::EbvMeshViewIndicesNV:
989 return spv::BuiltInMeshViewIndicesNV;
990
991 // sm builtins
992 case glslang::EbvWarpsPerSM:
993 builder.addExtension(spv::E_SPV_NV_shader_sm_builtins);
994 builder.addCapability(spv::CapabilityShaderSMBuiltinsNV);
995 return spv::BuiltInWarpsPerSMNV;
996 case glslang::EbvSMCount:
997 builder.addExtension(spv::E_SPV_NV_shader_sm_builtins);
998 builder.addCapability(spv::CapabilityShaderSMBuiltinsNV);
999 return spv::BuiltInSMCountNV;
1000 case glslang::EbvWarpID:
1001 builder.addExtension(spv::E_SPV_NV_shader_sm_builtins);
1002 builder.addCapability(spv::CapabilityShaderSMBuiltinsNV);
1003 return spv::BuiltInWarpIDNV;
1004 case glslang::EbvSMID:
1005 builder.addExtension(spv::E_SPV_NV_shader_sm_builtins);
1006 builder.addCapability(spv::CapabilityShaderSMBuiltinsNV);
1007 return spv::BuiltInSMIDNV;
1008 #endif
1009
1010 default:
1011 return spv::BuiltInMax;
1012 }
1013 }
1014
1015 // Translate glslang image layout format to SPIR-V image format.
TranslateImageFormat(const glslang::TType & type)1016 spv::ImageFormat TGlslangToSpvTraverser::TranslateImageFormat(const glslang::TType& type)
1017 {
1018 assert(type.getBasicType() == glslang::EbtSampler);
1019
1020 #ifdef GLSLANG_WEB
1021 return spv::ImageFormatUnknown;
1022 #endif
1023
1024 // Check for capabilities
1025 switch (type.getQualifier().getFormat()) {
1026 case glslang::ElfRg32f:
1027 case glslang::ElfRg16f:
1028 case glslang::ElfR11fG11fB10f:
1029 case glslang::ElfR16f:
1030 case glslang::ElfRgba16:
1031 case glslang::ElfRgb10A2:
1032 case glslang::ElfRg16:
1033 case glslang::ElfRg8:
1034 case glslang::ElfR16:
1035 case glslang::ElfR8:
1036 case glslang::ElfRgba16Snorm:
1037 case glslang::ElfRg16Snorm:
1038 case glslang::ElfRg8Snorm:
1039 case glslang::ElfR16Snorm:
1040 case glslang::ElfR8Snorm:
1041
1042 case glslang::ElfRg32i:
1043 case glslang::ElfRg16i:
1044 case glslang::ElfRg8i:
1045 case glslang::ElfR16i:
1046 case glslang::ElfR8i:
1047
1048 case glslang::ElfRgb10a2ui:
1049 case glslang::ElfRg32ui:
1050 case glslang::ElfRg16ui:
1051 case glslang::ElfRg8ui:
1052 case glslang::ElfR16ui:
1053 case glslang::ElfR8ui:
1054 builder.addCapability(spv::CapabilityStorageImageExtendedFormats);
1055 break;
1056
1057 default:
1058 break;
1059 }
1060
1061 // do the translation
1062 switch (type.getQualifier().getFormat()) {
1063 case glslang::ElfNone: return spv::ImageFormatUnknown;
1064 case glslang::ElfRgba32f: return spv::ImageFormatRgba32f;
1065 case glslang::ElfRgba16f: return spv::ImageFormatRgba16f;
1066 case glslang::ElfR32f: return spv::ImageFormatR32f;
1067 case glslang::ElfRgba8: return spv::ImageFormatRgba8;
1068 case glslang::ElfRgba8Snorm: return spv::ImageFormatRgba8Snorm;
1069 case glslang::ElfRg32f: return spv::ImageFormatRg32f;
1070 case glslang::ElfRg16f: return spv::ImageFormatRg16f;
1071 case glslang::ElfR11fG11fB10f: return spv::ImageFormatR11fG11fB10f;
1072 case glslang::ElfR16f: return spv::ImageFormatR16f;
1073 case glslang::ElfRgba16: return spv::ImageFormatRgba16;
1074 case glslang::ElfRgb10A2: return spv::ImageFormatRgb10A2;
1075 case glslang::ElfRg16: return spv::ImageFormatRg16;
1076 case glslang::ElfRg8: return spv::ImageFormatRg8;
1077 case glslang::ElfR16: return spv::ImageFormatR16;
1078 case glslang::ElfR8: return spv::ImageFormatR8;
1079 case glslang::ElfRgba16Snorm: return spv::ImageFormatRgba16Snorm;
1080 case glslang::ElfRg16Snorm: return spv::ImageFormatRg16Snorm;
1081 case glslang::ElfRg8Snorm: return spv::ImageFormatRg8Snorm;
1082 case glslang::ElfR16Snorm: return spv::ImageFormatR16Snorm;
1083 case glslang::ElfR8Snorm: return spv::ImageFormatR8Snorm;
1084 case glslang::ElfRgba32i: return spv::ImageFormatRgba32i;
1085 case glslang::ElfRgba16i: return spv::ImageFormatRgba16i;
1086 case glslang::ElfRgba8i: return spv::ImageFormatRgba8i;
1087 case glslang::ElfR32i: return spv::ImageFormatR32i;
1088 case glslang::ElfRg32i: return spv::ImageFormatRg32i;
1089 case glslang::ElfRg16i: return spv::ImageFormatRg16i;
1090 case glslang::ElfRg8i: return spv::ImageFormatRg8i;
1091 case glslang::ElfR16i: return spv::ImageFormatR16i;
1092 case glslang::ElfR8i: return spv::ImageFormatR8i;
1093 case glslang::ElfRgba32ui: return spv::ImageFormatRgba32ui;
1094 case glslang::ElfRgba16ui: return spv::ImageFormatRgba16ui;
1095 case glslang::ElfRgba8ui: return spv::ImageFormatRgba8ui;
1096 case glslang::ElfR32ui: return spv::ImageFormatR32ui;
1097 case glslang::ElfRg32ui: return spv::ImageFormatRg32ui;
1098 case glslang::ElfRg16ui: return spv::ImageFormatRg16ui;
1099 case glslang::ElfRgb10a2ui: return spv::ImageFormatRgb10a2ui;
1100 case glslang::ElfRg8ui: return spv::ImageFormatRg8ui;
1101 case glslang::ElfR16ui: return spv::ImageFormatR16ui;
1102 case glslang::ElfR8ui: return spv::ImageFormatR8ui;
1103 default: return spv::ImageFormatMax;
1104 }
1105 }
1106
TranslateSelectionControl(const glslang::TIntermSelection & selectionNode) const1107 spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSelectionControl(const glslang::TIntermSelection& selectionNode) const
1108 {
1109 if (selectionNode.getFlatten())
1110 return spv::SelectionControlFlattenMask;
1111 if (selectionNode.getDontFlatten())
1112 return spv::SelectionControlDontFlattenMask;
1113 return spv::SelectionControlMaskNone;
1114 }
1115
TranslateSwitchControl(const glslang::TIntermSwitch & switchNode) const1116 spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSwitchControl(const glslang::TIntermSwitch& switchNode) const
1117 {
1118 if (switchNode.getFlatten())
1119 return spv::SelectionControlFlattenMask;
1120 if (switchNode.getDontFlatten())
1121 return spv::SelectionControlDontFlattenMask;
1122 return spv::SelectionControlMaskNone;
1123 }
1124
1125 // return a non-0 dependency if the dependency argument must be set
TranslateLoopControl(const glslang::TIntermLoop & loopNode,std::vector<unsigned int> & operands) const1126 spv::LoopControlMask TGlslangToSpvTraverser::TranslateLoopControl(const glslang::TIntermLoop& loopNode,
1127 std::vector<unsigned int>& operands) const
1128 {
1129 spv::LoopControlMask control = spv::LoopControlMaskNone;
1130
1131 if (loopNode.getDontUnroll())
1132 control = control | spv::LoopControlDontUnrollMask;
1133 if (loopNode.getUnroll())
1134 control = control | spv::LoopControlUnrollMask;
1135 if (unsigned(loopNode.getLoopDependency()) == glslang::TIntermLoop::dependencyInfinite)
1136 control = control | spv::LoopControlDependencyInfiniteMask;
1137 else if (loopNode.getLoopDependency() > 0) {
1138 control = control | spv::LoopControlDependencyLengthMask;
1139 operands.push_back((unsigned int)loopNode.getLoopDependency());
1140 }
1141 if (glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4) {
1142 if (loopNode.getMinIterations() > 0) {
1143 control = control | spv::LoopControlMinIterationsMask;
1144 operands.push_back(loopNode.getMinIterations());
1145 }
1146 if (loopNode.getMaxIterations() < glslang::TIntermLoop::iterationsInfinite) {
1147 control = control | spv::LoopControlMaxIterationsMask;
1148 operands.push_back(loopNode.getMaxIterations());
1149 }
1150 if (loopNode.getIterationMultiple() > 1) {
1151 control = control | spv::LoopControlIterationMultipleMask;
1152 operands.push_back(loopNode.getIterationMultiple());
1153 }
1154 if (loopNode.getPeelCount() > 0) {
1155 control = control | spv::LoopControlPeelCountMask;
1156 operands.push_back(loopNode.getPeelCount());
1157 }
1158 if (loopNode.getPartialCount() > 0) {
1159 control = control | spv::LoopControlPartialCountMask;
1160 operands.push_back(loopNode.getPartialCount());
1161 }
1162 }
1163
1164 return control;
1165 }
1166
1167 // Translate glslang type to SPIR-V storage class.
TranslateStorageClass(const glslang::TType & type)1168 spv::StorageClass TGlslangToSpvTraverser::TranslateStorageClass(const glslang::TType& type)
1169 {
1170 if (type.getQualifier().isPipeInput())
1171 return spv::StorageClassInput;
1172 if (type.getQualifier().isPipeOutput())
1173 return spv::StorageClassOutput;
1174
1175 if (glslangIntermediate->getSource() != glslang::EShSourceHlsl ||
1176 type.getQualifier().storage == glslang::EvqUniform) {
1177 if (type.isAtomic())
1178 return spv::StorageClassAtomicCounter;
1179 if (type.containsOpaque())
1180 return spv::StorageClassUniformConstant;
1181 }
1182
1183 if (type.getQualifier().isUniformOrBuffer() &&
1184 type.getQualifier().isShaderRecordNV()) {
1185 return spv::StorageClassShaderRecordBufferNV;
1186 }
1187
1188 if (glslangIntermediate->usingStorageBuffer() && type.getQualifier().storage == glslang::EvqBuffer) {
1189 builder.addIncorporatedExtension(spv::E_SPV_KHR_storage_buffer_storage_class, spv::Spv_1_3);
1190 return spv::StorageClassStorageBuffer;
1191 }
1192
1193 if (type.getQualifier().isUniformOrBuffer()) {
1194 if (type.getQualifier().isPushConstant())
1195 return spv::StorageClassPushConstant;
1196 if (type.getBasicType() == glslang::EbtBlock)
1197 return spv::StorageClassUniform;
1198 return spv::StorageClassUniformConstant;
1199 }
1200
1201 switch (type.getQualifier().storage) {
1202 case glslang::EvqGlobal: return spv::StorageClassPrivate;
1203 case glslang::EvqConstReadOnly: return spv::StorageClassFunction;
1204 case glslang::EvqTemporary: return spv::StorageClassFunction;
1205 case glslang::EvqShared: return spv::StorageClassWorkgroup;
1206 #ifndef GLSLANG_WEB
1207 case glslang::EvqPayloadNV: return spv::StorageClassRayPayloadNV;
1208 case glslang::EvqPayloadInNV: return spv::StorageClassIncomingRayPayloadNV;
1209 case glslang::EvqHitAttrNV: return spv::StorageClassHitAttributeNV;
1210 case glslang::EvqCallableDataNV: return spv::StorageClassCallableDataNV;
1211 case glslang::EvqCallableDataInNV: return spv::StorageClassIncomingCallableDataNV;
1212 #endif
1213 default:
1214 assert(0);
1215 break;
1216 }
1217
1218 return spv::StorageClassFunction;
1219 }
1220
1221 // Add capabilities pertaining to how an array is indexed.
addIndirectionIndexCapabilities(const glslang::TType & baseType,const glslang::TType & indexType)1222 void TGlslangToSpvTraverser::addIndirectionIndexCapabilities(const glslang::TType& baseType,
1223 const glslang::TType& indexType)
1224 {
1225 #ifndef GLSLANG_WEB
1226 if (indexType.getQualifier().isNonUniform()) {
1227 // deal with an asserted non-uniform index
1228 // SPV_EXT_descriptor_indexing already added in TranslateNonUniformDecoration
1229 if (baseType.getBasicType() == glslang::EbtSampler) {
1230 if (baseType.getQualifier().hasAttachment())
1231 builder.addCapability(spv::CapabilityInputAttachmentArrayNonUniformIndexingEXT);
1232 else if (baseType.isImage() && baseType.getSampler().isBuffer())
1233 builder.addCapability(spv::CapabilityStorageTexelBufferArrayNonUniformIndexingEXT);
1234 else if (baseType.isTexture() && baseType.getSampler().isBuffer())
1235 builder.addCapability(spv::CapabilityUniformTexelBufferArrayNonUniformIndexingEXT);
1236 else if (baseType.isImage())
1237 builder.addCapability(spv::CapabilityStorageImageArrayNonUniformIndexingEXT);
1238 else if (baseType.isTexture())
1239 builder.addCapability(spv::CapabilitySampledImageArrayNonUniformIndexingEXT);
1240 } else if (baseType.getBasicType() == glslang::EbtBlock) {
1241 if (baseType.getQualifier().storage == glslang::EvqBuffer)
1242 builder.addCapability(spv::CapabilityStorageBufferArrayNonUniformIndexingEXT);
1243 else if (baseType.getQualifier().storage == glslang::EvqUniform)
1244 builder.addCapability(spv::CapabilityUniformBufferArrayNonUniformIndexingEXT);
1245 }
1246 } else {
1247 // assume a dynamically uniform index
1248 if (baseType.getBasicType() == glslang::EbtSampler) {
1249 if (baseType.getQualifier().hasAttachment()) {
1250 builder.addIncorporatedExtension("SPV_EXT_descriptor_indexing", spv::Spv_1_5);
1251 builder.addCapability(spv::CapabilityInputAttachmentArrayDynamicIndexingEXT);
1252 } else if (baseType.isImage() && baseType.getSampler().isBuffer()) {
1253 builder.addIncorporatedExtension("SPV_EXT_descriptor_indexing", spv::Spv_1_5);
1254 builder.addCapability(spv::CapabilityStorageTexelBufferArrayDynamicIndexingEXT);
1255 } else if (baseType.isTexture() && baseType.getSampler().isBuffer()) {
1256 builder.addIncorporatedExtension("SPV_EXT_descriptor_indexing", spv::Spv_1_5);
1257 builder.addCapability(spv::CapabilityUniformTexelBufferArrayDynamicIndexingEXT);
1258 }
1259 }
1260 }
1261 #endif
1262 }
1263
1264 // Return whether or not the given type is something that should be tied to a
1265 // descriptor set.
IsDescriptorResource(const glslang::TType & type)1266 bool IsDescriptorResource(const glslang::TType& type)
1267 {
1268 // uniform and buffer blocks are included, unless it is a push_constant
1269 if (type.getBasicType() == glslang::EbtBlock)
1270 return type.getQualifier().isUniformOrBuffer() &&
1271 ! type.getQualifier().isShaderRecordNV() &&
1272 ! type.getQualifier().isPushConstant();
1273
1274 // non block...
1275 // basically samplerXXX/subpass/sampler/texture are all included
1276 // if they are the global-scope-class, not the function parameter
1277 // (or local, if they ever exist) class.
1278 if (type.getBasicType() == glslang::EbtSampler)
1279 return type.getQualifier().isUniformOrBuffer();
1280
1281 // None of the above.
1282 return false;
1283 }
1284
InheritQualifiers(glslang::TQualifier & child,const glslang::TQualifier & parent)1285 void InheritQualifiers(glslang::TQualifier& child, const glslang::TQualifier& parent)
1286 {
1287 if (child.layoutMatrix == glslang::ElmNone)
1288 child.layoutMatrix = parent.layoutMatrix;
1289
1290 if (parent.invariant)
1291 child.invariant = true;
1292 if (parent.flat)
1293 child.flat = true;
1294 if (parent.centroid)
1295 child.centroid = true;
1296 #ifndef GLSLANG_WEB
1297 if (parent.nopersp)
1298 child.nopersp = true;
1299 if (parent.explicitInterp)
1300 child.explicitInterp = true;
1301 if (parent.perPrimitiveNV)
1302 child.perPrimitiveNV = true;
1303 if (parent.perViewNV)
1304 child.perViewNV = true;
1305 if (parent.perTaskNV)
1306 child.perTaskNV = true;
1307 if (parent.patch)
1308 child.patch = true;
1309 if (parent.sample)
1310 child.sample = true;
1311 if (parent.coherent)
1312 child.coherent = true;
1313 if (parent.devicecoherent)
1314 child.devicecoherent = true;
1315 if (parent.queuefamilycoherent)
1316 child.queuefamilycoherent = true;
1317 if (parent.workgroupcoherent)
1318 child.workgroupcoherent = true;
1319 if (parent.subgroupcoherent)
1320 child.subgroupcoherent = true;
1321 if (parent.nonprivate)
1322 child.nonprivate = true;
1323 if (parent.volatil)
1324 child.volatil = true;
1325 if (parent.restrict)
1326 child.restrict = true;
1327 if (parent.readonly)
1328 child.readonly = true;
1329 if (parent.writeonly)
1330 child.writeonly = true;
1331 #endif
1332 }
1333
HasNonLayoutQualifiers(const glslang::TType & type,const glslang::TQualifier & qualifier)1334 bool HasNonLayoutQualifiers(const glslang::TType& type, const glslang::TQualifier& qualifier)
1335 {
1336 // This should list qualifiers that simultaneous satisfy:
1337 // - struct members might inherit from a struct declaration
1338 // (note that non-block structs don't explicitly inherit,
1339 // only implicitly, meaning no decoration involved)
1340 // - affect decorations on the struct members
1341 // (note smooth does not, and expecting something like volatile
1342 // to effect the whole object)
1343 // - are not part of the offset/st430/etc or row/column-major layout
1344 return qualifier.invariant || (qualifier.hasLocation() && type.getBasicType() == glslang::EbtBlock);
1345 }
1346
1347 //
1348 // Implement the TGlslangToSpvTraverser class.
1349 //
1350
TGlslangToSpvTraverser(unsigned int spvVersion,const glslang::TIntermediate * glslangIntermediate,spv::SpvBuildLogger * buildLogger,glslang::SpvOptions & options)1351 TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const glslang::TIntermediate* glslangIntermediate,
1352 spv::SpvBuildLogger* buildLogger, glslang::SpvOptions& options)
1353 : TIntermTraverser(true, false, true),
1354 options(options),
1355 shaderEntry(nullptr), currentFunction(nullptr),
1356 sequenceDepth(0), logger(buildLogger),
1357 builder(spvVersion, (glslang::GetKhronosToolId() << 16) | glslang::GetSpirvGeneratorVersion(), logger),
1358 inEntryPoint(false), entryPointTerminated(false), linkageOnly(false),
1359 glslangIntermediate(glslangIntermediate),
1360 nanMinMaxClamp(glslangIntermediate->getNanMinMaxClamp())
1361 {
1362 spv::ExecutionModel executionModel = TranslateExecutionModel(glslangIntermediate->getStage());
1363
1364 builder.clearAccessChain();
1365 builder.setSource(TranslateSourceLanguage(glslangIntermediate->getSource(), glslangIntermediate->getProfile()),
1366 glslangIntermediate->getVersion());
1367
1368 if (options.generateDebugInfo) {
1369 builder.setEmitOpLines();
1370 builder.setSourceFile(glslangIntermediate->getSourceFile());
1371
1372 // Set the source shader's text. If for SPV version 1.0, include
1373 // a preamble in comments stating the OpModuleProcessed instructions.
1374 // Otherwise, emit those as actual instructions.
1375 std::string text;
1376 const std::vector<std::string>& processes = glslangIntermediate->getProcesses();
1377 for (int p = 0; p < (int)processes.size(); ++p) {
1378 if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_1) {
1379 text.append("// OpModuleProcessed ");
1380 text.append(processes[p]);
1381 text.append("\n");
1382 } else
1383 builder.addModuleProcessed(processes[p]);
1384 }
1385 if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_1 && (int)processes.size() > 0)
1386 text.append("#line 1\n");
1387 text.append(glslangIntermediate->getSourceText());
1388 builder.setSourceText(text);
1389 // Pass name and text for all included files
1390 const std::map<std::string, std::string>& include_txt = glslangIntermediate->getIncludeText();
1391 for (auto iItr = include_txt.begin(); iItr != include_txt.end(); ++iItr)
1392 builder.addInclude(iItr->first, iItr->second);
1393 }
1394 stdBuiltins = builder.import("GLSL.std.450");
1395
1396 spv::AddressingModel addressingModel = spv::AddressingModelLogical;
1397 spv::MemoryModel memoryModel = spv::MemoryModelGLSL450;
1398
1399 if (glslangIntermediate->usingPhysicalStorageBuffer()) {
1400 addressingModel = spv::AddressingModelPhysicalStorageBuffer64EXT;
1401 builder.addIncorporatedExtension(spv::E_SPV_EXT_physical_storage_buffer, spv::Spv_1_5);
1402 builder.addCapability(spv::CapabilityPhysicalStorageBufferAddressesEXT);
1403 };
1404 if (glslangIntermediate->usingVulkanMemoryModel()) {
1405 memoryModel = spv::MemoryModelVulkanKHR;
1406 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
1407 builder.addIncorporatedExtension(spv::E_SPV_KHR_vulkan_memory_model, spv::Spv_1_5);
1408 }
1409 builder.setMemoryModel(addressingModel, memoryModel);
1410
1411 if (glslangIntermediate->usingVariablePointers()) {
1412 builder.addCapability(spv::CapabilityVariablePointers);
1413 }
1414
1415 shaderEntry = builder.makeEntryPoint(glslangIntermediate->getEntryPointName().c_str());
1416 entryPoint = builder.addEntryPoint(executionModel, shaderEntry, glslangIntermediate->getEntryPointName().c_str());
1417
1418 // Add the source extensions
1419 const auto& sourceExtensions = glslangIntermediate->getRequestedExtensions();
1420 for (auto it = sourceExtensions.begin(); it != sourceExtensions.end(); ++it)
1421 builder.addSourceExtension(it->c_str());
1422
1423 // Add the top-level modes for this shader.
1424
1425 if (glslangIntermediate->getXfbMode()) {
1426 builder.addCapability(spv::CapabilityTransformFeedback);
1427 builder.addExecutionMode(shaderEntry, spv::ExecutionModeXfb);
1428 }
1429
1430 unsigned int mode;
1431 switch (glslangIntermediate->getStage()) {
1432 case EShLangVertex:
1433 builder.addCapability(spv::CapabilityShader);
1434 break;
1435
1436 case EShLangFragment:
1437 builder.addCapability(spv::CapabilityShader);
1438 if (glslangIntermediate->getPixelCenterInteger())
1439 builder.addExecutionMode(shaderEntry, spv::ExecutionModePixelCenterInteger);
1440
1441 if (glslangIntermediate->getOriginUpperLeft())
1442 builder.addExecutionMode(shaderEntry, spv::ExecutionModeOriginUpperLeft);
1443 else
1444 builder.addExecutionMode(shaderEntry, spv::ExecutionModeOriginLowerLeft);
1445
1446 if (glslangIntermediate->getEarlyFragmentTests())
1447 builder.addExecutionMode(shaderEntry, spv::ExecutionModeEarlyFragmentTests);
1448
1449 if (glslangIntermediate->getPostDepthCoverage()) {
1450 builder.addCapability(spv::CapabilitySampleMaskPostDepthCoverage);
1451 builder.addExecutionMode(shaderEntry, spv::ExecutionModePostDepthCoverage);
1452 builder.addExtension(spv::E_SPV_KHR_post_depth_coverage);
1453 }
1454
1455 if (glslangIntermediate->getDepth() != glslang::EldUnchanged && glslangIntermediate->isDepthReplacing())
1456 builder.addExecutionMode(shaderEntry, spv::ExecutionModeDepthReplacing);
1457
1458 #ifndef GLSLANG_WEB
1459 switch(glslangIntermediate->getDepth()) {
1460 case glslang::EldGreater: mode = spv::ExecutionModeDepthGreater; break;
1461 case glslang::EldLess: mode = spv::ExecutionModeDepthLess; break;
1462 default: mode = spv::ExecutionModeMax; break;
1463 }
1464 if (mode != spv::ExecutionModeMax)
1465 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
1466 switch (glslangIntermediate->getInterlockOrdering()) {
1467 case glslang::EioPixelInterlockOrdered: mode = spv::ExecutionModePixelInterlockOrderedEXT;
1468 break;
1469 case glslang::EioPixelInterlockUnordered: mode = spv::ExecutionModePixelInterlockUnorderedEXT;
1470 break;
1471 case glslang::EioSampleInterlockOrdered: mode = spv::ExecutionModeSampleInterlockOrderedEXT;
1472 break;
1473 case glslang::EioSampleInterlockUnordered: mode = spv::ExecutionModeSampleInterlockUnorderedEXT;
1474 break;
1475 case glslang::EioShadingRateInterlockOrdered: mode = spv::ExecutionModeShadingRateInterlockOrderedEXT;
1476 break;
1477 case glslang::EioShadingRateInterlockUnordered: mode = spv::ExecutionModeShadingRateInterlockUnorderedEXT;
1478 break;
1479 default: mode = spv::ExecutionModeMax;
1480 break;
1481 }
1482 if (mode != spv::ExecutionModeMax) {
1483 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
1484 if (mode == spv::ExecutionModeShadingRateInterlockOrderedEXT ||
1485 mode == spv::ExecutionModeShadingRateInterlockUnorderedEXT) {
1486 builder.addCapability(spv::CapabilityFragmentShaderShadingRateInterlockEXT);
1487 } else if (mode == spv::ExecutionModePixelInterlockOrderedEXT ||
1488 mode == spv::ExecutionModePixelInterlockUnorderedEXT) {
1489 builder.addCapability(spv::CapabilityFragmentShaderPixelInterlockEXT);
1490 } else {
1491 builder.addCapability(spv::CapabilityFragmentShaderSampleInterlockEXT);
1492 }
1493 builder.addExtension(spv::E_SPV_EXT_fragment_shader_interlock);
1494 }
1495 #endif
1496 break;
1497
1498 case EShLangCompute:
1499 builder.addCapability(spv::CapabilityShader);
1500 builder.addExecutionMode(shaderEntry, spv::ExecutionModeLocalSize, glslangIntermediate->getLocalSize(0),
1501 glslangIntermediate->getLocalSize(1),
1502 glslangIntermediate->getLocalSize(2));
1503 if (glslangIntermediate->getLayoutDerivativeModeNone() == glslang::LayoutDerivativeGroupQuads) {
1504 builder.addCapability(spv::CapabilityComputeDerivativeGroupQuadsNV);
1505 builder.addExecutionMode(shaderEntry, spv::ExecutionModeDerivativeGroupQuadsNV);
1506 builder.addExtension(spv::E_SPV_NV_compute_shader_derivatives);
1507 } else if (glslangIntermediate->getLayoutDerivativeModeNone() == glslang::LayoutDerivativeGroupLinear) {
1508 builder.addCapability(spv::CapabilityComputeDerivativeGroupLinearNV);
1509 builder.addExecutionMode(shaderEntry, spv::ExecutionModeDerivativeGroupLinearNV);
1510 builder.addExtension(spv::E_SPV_NV_compute_shader_derivatives);
1511 }
1512 break;
1513 #ifndef GLSLANG_WEB
1514 case EShLangTessEvaluation:
1515 case EShLangTessControl:
1516 builder.addCapability(spv::CapabilityTessellation);
1517
1518 glslang::TLayoutGeometry primitive;
1519
1520 if (glslangIntermediate->getStage() == EShLangTessControl) {
1521 builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices());
1522 primitive = glslangIntermediate->getOutputPrimitive();
1523 } else {
1524 primitive = glslangIntermediate->getInputPrimitive();
1525 }
1526
1527 switch (primitive) {
1528 case glslang::ElgTriangles: mode = spv::ExecutionModeTriangles; break;
1529 case glslang::ElgQuads: mode = spv::ExecutionModeQuads; break;
1530 case glslang::ElgIsolines: mode = spv::ExecutionModeIsolines; break;
1531 default: mode = spv::ExecutionModeMax; break;
1532 }
1533 if (mode != spv::ExecutionModeMax)
1534 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
1535
1536 switch (glslangIntermediate->getVertexSpacing()) {
1537 case glslang::EvsEqual: mode = spv::ExecutionModeSpacingEqual; break;
1538 case glslang::EvsFractionalEven: mode = spv::ExecutionModeSpacingFractionalEven; break;
1539 case glslang::EvsFractionalOdd: mode = spv::ExecutionModeSpacingFractionalOdd; break;
1540 default: mode = spv::ExecutionModeMax; break;
1541 }
1542 if (mode != spv::ExecutionModeMax)
1543 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
1544
1545 switch (glslangIntermediate->getVertexOrder()) {
1546 case glslang::EvoCw: mode = spv::ExecutionModeVertexOrderCw; break;
1547 case glslang::EvoCcw: mode = spv::ExecutionModeVertexOrderCcw; break;
1548 default: mode = spv::ExecutionModeMax; break;
1549 }
1550 if (mode != spv::ExecutionModeMax)
1551 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
1552
1553 if (glslangIntermediate->getPointMode())
1554 builder.addExecutionMode(shaderEntry, spv::ExecutionModePointMode);
1555 break;
1556
1557 case EShLangGeometry:
1558 builder.addCapability(spv::CapabilityGeometry);
1559 switch (glslangIntermediate->getInputPrimitive()) {
1560 case glslang::ElgPoints: mode = spv::ExecutionModeInputPoints; break;
1561 case glslang::ElgLines: mode = spv::ExecutionModeInputLines; break;
1562 case glslang::ElgLinesAdjacency: mode = spv::ExecutionModeInputLinesAdjacency; break;
1563 case glslang::ElgTriangles: mode = spv::ExecutionModeTriangles; break;
1564 case glslang::ElgTrianglesAdjacency: mode = spv::ExecutionModeInputTrianglesAdjacency; break;
1565 default: mode = spv::ExecutionModeMax; break;
1566 }
1567 if (mode != spv::ExecutionModeMax)
1568 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
1569
1570 builder.addExecutionMode(shaderEntry, spv::ExecutionModeInvocations, glslangIntermediate->getInvocations());
1571
1572 switch (glslangIntermediate->getOutputPrimitive()) {
1573 case glslang::ElgPoints: mode = spv::ExecutionModeOutputPoints; break;
1574 case glslang::ElgLineStrip: mode = spv::ExecutionModeOutputLineStrip; break;
1575 case glslang::ElgTriangleStrip: mode = spv::ExecutionModeOutputTriangleStrip; break;
1576 default: mode = spv::ExecutionModeMax; break;
1577 }
1578 if (mode != spv::ExecutionModeMax)
1579 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
1580 builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices());
1581 break;
1582
1583 case EShLangRayGenNV:
1584 case EShLangIntersectNV:
1585 case EShLangAnyHitNV:
1586 case EShLangClosestHitNV:
1587 case EShLangMissNV:
1588 case EShLangCallableNV:
1589 builder.addCapability(spv::CapabilityRayTracingNV);
1590 builder.addExtension("SPV_NV_ray_tracing");
1591 break;
1592 case EShLangTaskNV:
1593 case EShLangMeshNV:
1594 builder.addCapability(spv::CapabilityMeshShadingNV);
1595 builder.addExtension(spv::E_SPV_NV_mesh_shader);
1596 builder.addExecutionMode(shaderEntry, spv::ExecutionModeLocalSize, glslangIntermediate->getLocalSize(0),
1597 glslangIntermediate->getLocalSize(1),
1598 glslangIntermediate->getLocalSize(2));
1599 if (glslangIntermediate->getStage() == EShLangMeshNV) {
1600 builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices());
1601 builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputPrimitivesNV, glslangIntermediate->getPrimitives());
1602
1603 switch (glslangIntermediate->getOutputPrimitive()) {
1604 case glslang::ElgPoints: mode = spv::ExecutionModeOutputPoints; break;
1605 case glslang::ElgLines: mode = spv::ExecutionModeOutputLinesNV; break;
1606 case glslang::ElgTriangles: mode = spv::ExecutionModeOutputTrianglesNV; break;
1607 default: mode = spv::ExecutionModeMax; break;
1608 }
1609 if (mode != spv::ExecutionModeMax)
1610 builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
1611 }
1612 break;
1613 #endif
1614
1615 default:
1616 break;
1617 }
1618 }
1619
1620 // Finish creating SPV, after the traversal is complete.
finishSpv()1621 void TGlslangToSpvTraverser::finishSpv()
1622 {
1623 // Finish the entry point function
1624 if (! entryPointTerminated) {
1625 builder.setBuildPoint(shaderEntry->getLastBlock());
1626 builder.leaveFunction();
1627 }
1628
1629 // finish off the entry-point SPV instruction by adding the Input/Output <id>
1630 for (auto it = iOSet.cbegin(); it != iOSet.cend(); ++it)
1631 entryPoint->addIdOperand(*it);
1632
1633 // Add capabilities, extensions, remove unneeded decorations, etc.,
1634 // based on the resulting SPIR-V.
1635 // Note: WebGPU code generation must have the opportunity to aggressively
1636 // prune unreachable merge blocks and continue targets.
1637 builder.postProcess();
1638 }
1639
1640 // Write the SPV into 'out'.
dumpSpv(std::vector<unsigned int> & out)1641 void TGlslangToSpvTraverser::dumpSpv(std::vector<unsigned int>& out)
1642 {
1643 builder.dump(out);
1644 }
1645
1646 //
1647 // Implement the traversal functions.
1648 //
1649 // Return true from interior nodes to have the external traversal
1650 // continue on to children. Return false if children were
1651 // already processed.
1652 //
1653
1654 //
1655 // Symbols can turn into
1656 // - uniform/input reads
1657 // - output writes
1658 // - complex lvalue base setups: foo.bar[3].... , where we see foo and start up an access chain
1659 // - something simple that degenerates into the last bullet
1660 //
visitSymbol(glslang::TIntermSymbol * symbol)1661 void TGlslangToSpvTraverser::visitSymbol(glslang::TIntermSymbol* symbol)
1662 {
1663 SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder);
1664 if (symbol->getType().getQualifier().isSpecConstant())
1665 spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
1666
1667 // getSymbolId() will set up all the IO decorations on the first call.
1668 // Formal function parameters were mapped during makeFunctions().
1669 spv::Id id = getSymbolId(symbol);
1670
1671 if (builder.isPointer(id)) {
1672 // Include all "static use" and "linkage only" interface variables on the OpEntryPoint instruction
1673 // Consider adding to the OpEntryPoint interface list.
1674 // Only looking at structures if they have at least one member.
1675 if (!symbol->getType().isStruct() || symbol->getType().getStruct()->size() > 0) {
1676 spv::StorageClass sc = builder.getStorageClass(id);
1677 // Before SPIR-V 1.4, we only want to include Input and Output.
1678 // Starting with SPIR-V 1.4, we want all globals.
1679 if ((glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4 && sc != spv::StorageClassFunction) ||
1680 (sc == spv::StorageClassInput || sc == spv::StorageClassOutput)) {
1681 iOSet.insert(id);
1682 }
1683 }
1684
1685 // If the SPIR-V type is required to be different than the AST type,
1686 // translate now from the SPIR-V type to the AST type, for the consuming
1687 // operation.
1688 // Note this turns it from an l-value to an r-value.
1689 // Currently, all symbols needing this are inputs; avoid the map lookup when non-input.
1690 if (symbol->getType().getQualifier().storage == glslang::EvqVaryingIn)
1691 id = translateForcedType(id);
1692 }
1693
1694 // Only process non-linkage-only nodes for generating actual static uses
1695 if (! linkageOnly || symbol->getQualifier().isSpecConstant()) {
1696 // Prepare to generate code for the access
1697
1698 // L-value chains will be computed left to right. We're on the symbol now,
1699 // which is the left-most part of the access chain, so now is "clear" time,
1700 // followed by setting the base.
1701 builder.clearAccessChain();
1702
1703 // For now, we consider all user variables as being in memory, so they are pointers,
1704 // except for
1705 // A) R-Value arguments to a function, which are an intermediate object.
1706 // See comments in handleUserFunctionCall().
1707 // B) Specialization constants (normal constants don't even come in as a variable),
1708 // These are also pure R-values.
1709 // C) R-Values from type translation, see above call to translateForcedType()
1710 glslang::TQualifier qualifier = symbol->getQualifier();
1711 if (qualifier.isSpecConstant() || rValueParameters.find(symbol->getId()) != rValueParameters.end() ||
1712 !builder.isPointerType(builder.getTypeId(id)))
1713 builder.setAccessChainRValue(id);
1714 else
1715 builder.setAccessChainLValue(id);
1716 }
1717
1718 #ifdef ENABLE_HLSL
1719 // Process linkage-only nodes for any special additional interface work.
1720 if (linkageOnly) {
1721 if (glslangIntermediate->getHlslFunctionality1()) {
1722 // Map implicit counter buffers to their originating buffers, which should have been
1723 // seen by now, given earlier pruning of unused counters, and preservation of order
1724 // of declaration.
1725 if (symbol->getType().getQualifier().isUniformOrBuffer()) {
1726 if (!glslangIntermediate->hasCounterBufferName(symbol->getName())) {
1727 // Save possible originating buffers for counter buffers, keyed by
1728 // making the potential counter-buffer name.
1729 std::string keyName = symbol->getName().c_str();
1730 keyName = glslangIntermediate->addCounterBufferName(keyName);
1731 counterOriginator[keyName] = symbol;
1732 } else {
1733 // Handle a counter buffer, by finding the saved originating buffer.
1734 std::string keyName = symbol->getName().c_str();
1735 auto it = counterOriginator.find(keyName);
1736 if (it != counterOriginator.end()) {
1737 id = getSymbolId(it->second);
1738 if (id != spv::NoResult) {
1739 spv::Id counterId = getSymbolId(symbol);
1740 if (counterId != spv::NoResult) {
1741 builder.addExtension("SPV_GOOGLE_hlsl_functionality1");
1742 builder.addDecorationId(id, spv::DecorationHlslCounterBufferGOOGLE, counterId);
1743 }
1744 }
1745 }
1746 }
1747 }
1748 }
1749 }
1750 #endif
1751 }
1752
visitBinary(glslang::TVisit,glslang::TIntermBinary * node)1753 bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::TIntermBinary* node)
1754 {
1755 builder.setLine(node->getLoc().line, node->getLoc().getFilename());
1756
1757 SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder);
1758 if (node->getType().getQualifier().isSpecConstant())
1759 spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
1760
1761 // First, handle special cases
1762 switch (node->getOp()) {
1763 case glslang::EOpAssign:
1764 case glslang::EOpAddAssign:
1765 case glslang::EOpSubAssign:
1766 case glslang::EOpMulAssign:
1767 case glslang::EOpVectorTimesMatrixAssign:
1768 case glslang::EOpVectorTimesScalarAssign:
1769 case glslang::EOpMatrixTimesScalarAssign:
1770 case glslang::EOpMatrixTimesMatrixAssign:
1771 case glslang::EOpDivAssign:
1772 case glslang::EOpModAssign:
1773 case glslang::EOpAndAssign:
1774 case glslang::EOpInclusiveOrAssign:
1775 case glslang::EOpExclusiveOrAssign:
1776 case glslang::EOpLeftShiftAssign:
1777 case glslang::EOpRightShiftAssign:
1778 // A bin-op assign "a += b" means the same thing as "a = a + b"
1779 // where a is evaluated before b. For a simple assignment, GLSL
1780 // says to evaluate the left before the right. So, always, left
1781 // node then right node.
1782 {
1783 // get the left l-value, save it away
1784 builder.clearAccessChain();
1785 node->getLeft()->traverse(this);
1786 spv::Builder::AccessChain lValue = builder.getAccessChain();
1787
1788 // evaluate the right
1789 builder.clearAccessChain();
1790 node->getRight()->traverse(this);
1791 spv::Id rValue = accessChainLoad(node->getRight()->getType());
1792
1793 if (node->getOp() != glslang::EOpAssign) {
1794 // the left is also an r-value
1795 builder.setAccessChain(lValue);
1796 spv::Id leftRValue = accessChainLoad(node->getLeft()->getType());
1797
1798 // do the operation
1799 OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()),
1800 TranslateNoContractionDecoration(node->getType().getQualifier()),
1801 TranslateNonUniformDecoration(node->getType().getQualifier()) };
1802 rValue = createBinaryOperation(node->getOp(), decorations,
1803 convertGlslangToSpvType(node->getType()), leftRValue, rValue,
1804 node->getType().getBasicType());
1805
1806 // these all need their counterparts in createBinaryOperation()
1807 assert(rValue != spv::NoResult);
1808 }
1809
1810 // store the result
1811 builder.setAccessChain(lValue);
1812 multiTypeStore(node->getLeft()->getType(), rValue);
1813
1814 // assignments are expressions having an rValue after they are evaluated...
1815 builder.clearAccessChain();
1816 builder.setAccessChainRValue(rValue);
1817 }
1818 return false;
1819 case glslang::EOpIndexDirect:
1820 case glslang::EOpIndexDirectStruct:
1821 {
1822 // Structure, array, matrix, or vector indirection with statically known index.
1823 // Get the left part of the access chain.
1824 node->getLeft()->traverse(this);
1825
1826 // Add the next element in the chain
1827
1828 const int glslangIndex = node->getRight()->getAsConstantUnion()->getConstArray()[0].getIConst();
1829 if (! node->getLeft()->getType().isArray() &&
1830 node->getLeft()->getType().isVector() &&
1831 node->getOp() == glslang::EOpIndexDirect) {
1832 // This is essentially a hard-coded vector swizzle of size 1,
1833 // so short circuit the access-chain stuff with a swizzle.
1834 std::vector<unsigned> swizzle;
1835 swizzle.push_back(glslangIndex);
1836 int dummySize;
1837 builder.accessChainPushSwizzle(swizzle, convertGlslangToSpvType(node->getLeft()->getType()),
1838 TranslateCoherent(node->getLeft()->getType()),
1839 glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize));
1840 } else {
1841
1842 // Load through a block reference is performed with a dot operator that
1843 // is mapped to EOpIndexDirectStruct. When we get to the actual reference,
1844 // do a load and reset the access chain.
1845 if (node->getLeft()->isReference() &&
1846 !node->getLeft()->getType().isArray() &&
1847 node->getOp() == glslang::EOpIndexDirectStruct)
1848 {
1849 spv::Id left = accessChainLoad(node->getLeft()->getType());
1850 builder.clearAccessChain();
1851 builder.setAccessChainLValue(left);
1852 }
1853
1854 int spvIndex = glslangIndex;
1855 if (node->getLeft()->getBasicType() == glslang::EbtBlock &&
1856 node->getOp() == glslang::EOpIndexDirectStruct)
1857 {
1858 // This may be, e.g., an anonymous block-member selection, which generally need
1859 // index remapping due to hidden members in anonymous blocks.
1860 std::vector<int>& remapper = memberRemapper[node->getLeft()->getType().getStruct()];
1861 assert(remapper.size() > 0);
1862 spvIndex = remapper[glslangIndex];
1863 }
1864
1865 // normal case for indexing array or structure or block
1866 builder.accessChainPush(builder.makeIntConstant(spvIndex), TranslateCoherent(node->getLeft()->getType()), node->getLeft()->getType().getBufferReferenceAlignment());
1867
1868 // Add capabilities here for accessing PointSize and clip/cull distance.
1869 // We have deferred generation of associated capabilities until now.
1870 if (node->getLeft()->getType().isStruct() && ! node->getLeft()->getType().isArray())
1871 declareUseOfStructMember(*(node->getLeft()->getType().getStruct()), glslangIndex);
1872 }
1873 }
1874 return false;
1875 case glslang::EOpIndexIndirect:
1876 {
1877 // Array, matrix, or vector indirection with variable index.
1878 // Will use native SPIR-V access-chain for and array indirection;
1879 // matrices are arrays of vectors, so will also work for a matrix.
1880 // Will use the access chain's 'component' for variable index into a vector.
1881
1882 // This adapter is building access chains left to right.
1883 // Set up the access chain to the left.
1884 node->getLeft()->traverse(this);
1885
1886 // save it so that computing the right side doesn't trash it
1887 spv::Builder::AccessChain partial = builder.getAccessChain();
1888
1889 // compute the next index in the chain
1890 builder.clearAccessChain();
1891 node->getRight()->traverse(this);
1892 spv::Id index = accessChainLoad(node->getRight()->getType());
1893
1894 addIndirectionIndexCapabilities(node->getLeft()->getType(), node->getRight()->getType());
1895
1896 // restore the saved access chain
1897 builder.setAccessChain(partial);
1898
1899 if (! node->getLeft()->getType().isArray() && node->getLeft()->getType().isVector()) {
1900 int dummySize;
1901 builder.accessChainPushComponent(index, convertGlslangToSpvType(node->getLeft()->getType()),
1902 TranslateCoherent(node->getLeft()->getType()),
1903 glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize));
1904 } else
1905 builder.accessChainPush(index, TranslateCoherent(node->getLeft()->getType()), node->getLeft()->getType().getBufferReferenceAlignment());
1906 }
1907 return false;
1908 case glslang::EOpVectorSwizzle:
1909 {
1910 node->getLeft()->traverse(this);
1911 std::vector<unsigned> swizzle;
1912 convertSwizzle(*node->getRight()->getAsAggregate(), swizzle);
1913 int dummySize;
1914 builder.accessChainPushSwizzle(swizzle, convertGlslangToSpvType(node->getLeft()->getType()),
1915 TranslateCoherent(node->getLeft()->getType()),
1916 glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize));
1917 }
1918 return false;
1919 case glslang::EOpMatrixSwizzle:
1920 logger->missingFunctionality("matrix swizzle");
1921 return true;
1922 case glslang::EOpLogicalOr:
1923 case glslang::EOpLogicalAnd:
1924 {
1925
1926 // These may require short circuiting, but can sometimes be done as straight
1927 // binary operations. The right operand must be short circuited if it has
1928 // side effects, and should probably be if it is complex.
1929 if (isTrivial(node->getRight()->getAsTyped()))
1930 break; // handle below as a normal binary operation
1931 // otherwise, we need to do dynamic short circuiting on the right operand
1932 spv::Id result = createShortCircuit(node->getOp(), *node->getLeft()->getAsTyped(), *node->getRight()->getAsTyped());
1933 builder.clearAccessChain();
1934 builder.setAccessChainRValue(result);
1935 }
1936 return false;
1937 default:
1938 break;
1939 }
1940
1941 // Assume generic binary op...
1942
1943 // get right operand
1944 builder.clearAccessChain();
1945 node->getLeft()->traverse(this);
1946 spv::Id left = accessChainLoad(node->getLeft()->getType());
1947
1948 // get left operand
1949 builder.clearAccessChain();
1950 node->getRight()->traverse(this);
1951 spv::Id right = accessChainLoad(node->getRight()->getType());
1952
1953 // get result
1954 OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()),
1955 TranslateNoContractionDecoration(node->getType().getQualifier()),
1956 TranslateNonUniformDecoration(node->getType().getQualifier()) };
1957 spv::Id result = createBinaryOperation(node->getOp(), decorations,
1958 convertGlslangToSpvType(node->getType()), left, right,
1959 node->getLeft()->getType().getBasicType());
1960
1961 builder.clearAccessChain();
1962 if (! result) {
1963 logger->missingFunctionality("unknown glslang binary operation");
1964 return true; // pick up a child as the place-holder result
1965 } else {
1966 builder.setAccessChainRValue(result);
1967 return false;
1968 }
1969 }
1970
1971 // Figure out what, if any, type changes are needed when accessing a specific built-in.
1972 // Returns <the type SPIR-V requires for declarion, the type to translate to on use>.
1973 // Also see comment for 'forceType', regarding tracking SPIR-V-required types.
getForcedType(spv::BuiltIn builtIn,const glslang::TType & glslangType)1974 std::pair<spv::Id, spv::Id> TGlslangToSpvTraverser::getForcedType(spv::BuiltIn builtIn,
1975 const glslang::TType& glslangType)
1976 {
1977 switch(builtIn)
1978 {
1979 case spv::BuiltInSubgroupEqMask:
1980 case spv::BuiltInSubgroupGeMask:
1981 case spv::BuiltInSubgroupGtMask:
1982 case spv::BuiltInSubgroupLeMask:
1983 case spv::BuiltInSubgroupLtMask: {
1984 // these require changing a 64-bit scaler -> a vector of 32-bit components
1985 if (glslangType.isVector())
1986 break;
1987 std::pair<spv::Id, spv::Id> ret(builder.makeVectorType(builder.makeUintType(32), 4),
1988 builder.makeUintType(64));
1989 return ret;
1990 }
1991 default:
1992 break;
1993 }
1994
1995 std::pair<spv::Id, spv::Id> ret(spv::NoType, spv::NoType);
1996 return ret;
1997 }
1998
1999 // For an object previously identified (see getForcedType() and forceType)
2000 // as needing type translations, do the translation needed for a load, turning
2001 // an L-value into in R-value.
translateForcedType(spv::Id object)2002 spv::Id TGlslangToSpvTraverser::translateForcedType(spv::Id object)
2003 {
2004 const auto forceIt = forceType.find(object);
2005 if (forceIt == forceType.end())
2006 return object;
2007
2008 spv::Id desiredTypeId = forceIt->second;
2009 spv::Id objectTypeId = builder.getTypeId(object);
2010 assert(builder.isPointerType(objectTypeId));
2011 objectTypeId = builder.getContainedTypeId(objectTypeId);
2012 if (builder.isVectorType(objectTypeId) &&
2013 builder.getScalarTypeWidth(builder.getContainedTypeId(objectTypeId)) == 32) {
2014 if (builder.getScalarTypeWidth(desiredTypeId) == 64) {
2015 // handle 32-bit v.xy* -> 64-bit
2016 builder.clearAccessChain();
2017 builder.setAccessChainLValue(object);
2018 object = builder.accessChainLoad(spv::NoPrecision, spv::DecorationMax, objectTypeId);
2019 std::vector<spv::Id> components;
2020 components.push_back(builder.createCompositeExtract(object, builder.getContainedTypeId(objectTypeId), 0));
2021 components.push_back(builder.createCompositeExtract(object, builder.getContainedTypeId(objectTypeId), 1));
2022
2023 spv::Id vecType = builder.makeVectorType(builder.getContainedTypeId(objectTypeId), 2);
2024 return builder.createUnaryOp(spv::OpBitcast, desiredTypeId,
2025 builder.createCompositeConstruct(vecType, components));
2026 } else {
2027 logger->missingFunctionality("forcing 32-bit vector type to non 64-bit scalar");
2028 }
2029 } else {
2030 logger->missingFunctionality("forcing non 32-bit vector type");
2031 }
2032
2033 return object;
2034 }
2035
visitUnary(glslang::TVisit,glslang::TIntermUnary * node)2036 bool TGlslangToSpvTraverser::visitUnary(glslang::TVisit /* visit */, glslang::TIntermUnary* node)
2037 {
2038 builder.setLine(node->getLoc().line, node->getLoc().getFilename());
2039
2040 SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder);
2041 if (node->getType().getQualifier().isSpecConstant())
2042 spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
2043
2044 spv::Id result = spv::NoResult;
2045
2046 // try texturing first
2047 result = createImageTextureFunctionCall(node);
2048 if (result != spv::NoResult) {
2049 builder.clearAccessChain();
2050 builder.setAccessChainRValue(result);
2051
2052 return false; // done with this node
2053 }
2054
2055 // Non-texturing.
2056
2057 if (node->getOp() == glslang::EOpArrayLength) {
2058 // Quite special; won't want to evaluate the operand.
2059
2060 // Currently, the front-end does not allow .length() on an array until it is sized,
2061 // except for the last block membeor of an SSBO.
2062 // TODO: If this changes, link-time sized arrays might show up here, and need their
2063 // size extracted.
2064
2065 // Normal .length() would have been constant folded by the front-end.
2066 // So, this has to be block.lastMember.length().
2067 // SPV wants "block" and member number as the operands, go get them.
2068
2069 spv::Id length;
2070 if (node->getOperand()->getType().isCoopMat()) {
2071 spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
2072
2073 spv::Id typeId = convertGlslangToSpvType(node->getOperand()->getType());
2074 assert(builder.isCooperativeMatrixType(typeId));
2075
2076 length = builder.createCooperativeMatrixLength(typeId);
2077 } else {
2078 glslang::TIntermTyped* block = node->getOperand()->getAsBinaryNode()->getLeft();
2079 block->traverse(this);
2080 unsigned int member = node->getOperand()->getAsBinaryNode()->getRight()->getAsConstantUnion()->getConstArray()[0].getUConst();
2081 length = builder.createArrayLength(builder.accessChainGetLValue(), member);
2082 }
2083
2084 // GLSL semantics say the result of .length() is an int, while SPIR-V says
2085 // signedness must be 0. So, convert from SPIR-V unsigned back to GLSL's
2086 // AST expectation of a signed result.
2087 if (glslangIntermediate->getSource() == glslang::EShSourceGlsl) {
2088 if (builder.isInSpecConstCodeGenMode()) {
2089 length = builder.createBinOp(spv::OpIAdd, builder.makeIntType(32), length, builder.makeIntConstant(0));
2090 } else {
2091 length = builder.createUnaryOp(spv::OpBitcast, builder.makeIntType(32), length);
2092 }
2093 }
2094
2095 builder.clearAccessChain();
2096 builder.setAccessChainRValue(length);
2097
2098 return false;
2099 }
2100
2101 // Start by evaluating the operand
2102
2103 // Does it need a swizzle inversion? If so, evaluation is inverted;
2104 // operate first on the swizzle base, then apply the swizzle.
2105 spv::Id invertedType = spv::NoType;
2106 auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ? invertedType : convertGlslangToSpvType(node->getType()); };
2107 if (node->getOp() == glslang::EOpInterpolateAtCentroid)
2108 invertedType = getInvertedSwizzleType(*node->getOperand());
2109
2110 builder.clearAccessChain();
2111 TIntermNode *operandNode;
2112 if (invertedType != spv::NoType)
2113 operandNode = node->getOperand()->getAsBinaryNode()->getLeft();
2114 else
2115 operandNode = node->getOperand();
2116
2117 operandNode->traverse(this);
2118
2119 spv::Id operand = spv::NoResult;
2120
2121 spv::Builder::AccessChain::CoherentFlags lvalueCoherentFlags;
2122
2123 #ifndef GLSLANG_WEB
2124 if (node->getOp() == glslang::EOpAtomicCounterIncrement ||
2125 node->getOp() == glslang::EOpAtomicCounterDecrement ||
2126 node->getOp() == glslang::EOpAtomicCounter ||
2127 node->getOp() == glslang::EOpInterpolateAtCentroid) {
2128 operand = builder.accessChainGetLValue(); // Special case l-value operands
2129 lvalueCoherentFlags = builder.getAccessChain().coherentFlags;
2130 lvalueCoherentFlags |= TranslateCoherent(operandNode->getAsTyped()->getType());
2131 } else
2132 #endif
2133 {
2134 operand = accessChainLoad(node->getOperand()->getType());
2135 }
2136
2137 OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()),
2138 TranslateNoContractionDecoration(node->getType().getQualifier()),
2139 TranslateNonUniformDecoration(node->getType().getQualifier()) };
2140
2141 // it could be a conversion
2142 if (! result)
2143 result = createConversion(node->getOp(), decorations, resultType(), operand, node->getOperand()->getBasicType());
2144
2145 // if not, then possibly an operation
2146 if (! result)
2147 result = createUnaryOperation(node->getOp(), decorations, resultType(), operand, node->getOperand()->getBasicType(), lvalueCoherentFlags);
2148
2149 if (result) {
2150 if (invertedType) {
2151 result = createInvertedSwizzle(decorations.precision, *node->getOperand(), result);
2152 decorations.addNonUniform(builder, result);
2153 }
2154
2155 builder.clearAccessChain();
2156 builder.setAccessChainRValue(result);
2157
2158 return false; // done with this node
2159 }
2160
2161 // it must be a special case, check...
2162 switch (node->getOp()) {
2163 case glslang::EOpPostIncrement:
2164 case glslang::EOpPostDecrement:
2165 case glslang::EOpPreIncrement:
2166 case glslang::EOpPreDecrement:
2167 {
2168 // we need the integer value "1" or the floating point "1.0" to add/subtract
2169 spv::Id one = 0;
2170 if (node->getBasicType() == glslang::EbtFloat)
2171 one = builder.makeFloatConstant(1.0F);
2172 #ifndef GLSLANG_WEB
2173 else if (node->getBasicType() == glslang::EbtDouble)
2174 one = builder.makeDoubleConstant(1.0);
2175 else if (node->getBasicType() == glslang::EbtFloat16)
2176 one = builder.makeFloat16Constant(1.0F);
2177 else if (node->getBasicType() == glslang::EbtInt8 || node->getBasicType() == glslang::EbtUint8)
2178 one = builder.makeInt8Constant(1);
2179 else if (node->getBasicType() == glslang::EbtInt16 || node->getBasicType() == glslang::EbtUint16)
2180 one = builder.makeInt16Constant(1);
2181 else if (node->getBasicType() == glslang::EbtInt64 || node->getBasicType() == glslang::EbtUint64)
2182 one = builder.makeInt64Constant(1);
2183 #endif
2184 else
2185 one = builder.makeIntConstant(1);
2186 glslang::TOperator op;
2187 if (node->getOp() == glslang::EOpPreIncrement ||
2188 node->getOp() == glslang::EOpPostIncrement)
2189 op = glslang::EOpAdd;
2190 else
2191 op = glslang::EOpSub;
2192
2193 spv::Id result = createBinaryOperation(op, decorations,
2194 convertGlslangToSpvType(node->getType()), operand, one,
2195 node->getType().getBasicType());
2196 assert(result != spv::NoResult);
2197
2198 // The result of operation is always stored, but conditionally the
2199 // consumed result. The consumed result is always an r-value.
2200 builder.accessChainStore(result);
2201 builder.clearAccessChain();
2202 if (node->getOp() == glslang::EOpPreIncrement ||
2203 node->getOp() == glslang::EOpPreDecrement)
2204 builder.setAccessChainRValue(result);
2205 else
2206 builder.setAccessChainRValue(operand);
2207 }
2208
2209 return false;
2210
2211 #ifndef GLSLANG_WEB
2212 case glslang::EOpEmitStreamVertex:
2213 builder.createNoResultOp(spv::OpEmitStreamVertex, operand);
2214 return false;
2215 case glslang::EOpEndStreamPrimitive:
2216 builder.createNoResultOp(spv::OpEndStreamPrimitive, operand);
2217 return false;
2218 #endif
2219
2220 default:
2221 logger->missingFunctionality("unknown glslang unary");
2222 return true; // pick up operand as placeholder result
2223 }
2224 }
2225
2226 // Construct a composite object, recursively copying members if their types don't match
createCompositeConstruct(spv::Id resultTypeId,std::vector<spv::Id> constituents)2227 spv::Id TGlslangToSpvTraverser::createCompositeConstruct(spv::Id resultTypeId, std::vector<spv::Id> constituents)
2228 {
2229 for (int c = 0; c < (int)constituents.size(); ++c) {
2230 spv::Id& constituent = constituents[c];
2231 spv::Id lType = builder.getContainedTypeId(resultTypeId, c);
2232 spv::Id rType = builder.getTypeId(constituent);
2233 if (lType != rType) {
2234 if (glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4) {
2235 constituent = builder.createUnaryOp(spv::OpCopyLogical, lType, constituent);
2236 } else if (builder.isStructType(rType)) {
2237 std::vector<spv::Id> rTypeConstituents;
2238 int numrTypeConstituents = builder.getNumTypeConstituents(rType);
2239 for (int i = 0; i < numrTypeConstituents; ++i) {
2240 rTypeConstituents.push_back(builder.createCompositeExtract(constituent, builder.getContainedTypeId(rType, i), i));
2241 }
2242 constituents[c] = createCompositeConstruct(lType, rTypeConstituents);
2243 } else {
2244 assert(builder.isArrayType(rType));
2245 std::vector<spv::Id> rTypeConstituents;
2246 int numrTypeConstituents = builder.getNumTypeConstituents(rType);
2247
2248 spv::Id elementRType = builder.getContainedTypeId(rType);
2249 for (int i = 0; i < numrTypeConstituents; ++i) {
2250 rTypeConstituents.push_back(builder.createCompositeExtract(constituent, elementRType, i));
2251 }
2252 constituents[c] = createCompositeConstruct(lType, rTypeConstituents);
2253 }
2254 }
2255 }
2256 return builder.createCompositeConstruct(resultTypeId, constituents);
2257 }
2258
visitAggregate(glslang::TVisit visit,glslang::TIntermAggregate * node)2259 bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TIntermAggregate* node)
2260 {
2261 SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder);
2262 if (node->getType().getQualifier().isSpecConstant())
2263 spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
2264
2265 spv::Id result = spv::NoResult;
2266 spv::Id invertedType = spv::NoType; // to use to override the natural type of the node
2267 auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ? invertedType : convertGlslangToSpvType(node->getType()); };
2268
2269 // try texturing
2270 result = createImageTextureFunctionCall(node);
2271 if (result != spv::NoResult) {
2272 builder.clearAccessChain();
2273 builder.setAccessChainRValue(result);
2274
2275 return false;
2276 }
2277 #ifndef GLSLANG_WEB
2278 else if (node->getOp() == glslang::EOpImageStore ||
2279 node->getOp() == glslang::EOpImageStoreLod ||
2280 node->getOp() == glslang::EOpImageAtomicStore) {
2281 // "imageStore" is a special case, which has no result
2282 return false;
2283 }
2284 #endif
2285
2286 glslang::TOperator binOp = glslang::EOpNull;
2287 bool reduceComparison = true;
2288 bool isMatrix = false;
2289 bool noReturnValue = false;
2290 bool atomic = false;
2291
2292 spv::Builder::AccessChain::CoherentFlags lvalueCoherentFlags;
2293
2294 assert(node->getOp());
2295
2296 spv::Decoration precision = TranslatePrecisionDecoration(node->getOperationPrecision());
2297
2298 switch (node->getOp()) {
2299 case glslang::EOpSequence:
2300 {
2301 if (preVisit)
2302 ++sequenceDepth;
2303 else
2304 --sequenceDepth;
2305
2306 if (sequenceDepth == 1) {
2307 // If this is the parent node of all the functions, we want to see them
2308 // early, so all call points have actual SPIR-V functions to reference.
2309 // In all cases, still let the traverser visit the children for us.
2310 makeFunctions(node->getAsAggregate()->getSequence());
2311
2312 // Also, we want all globals initializers to go into the beginning of the entry point, before
2313 // anything else gets there, so visit out of order, doing them all now.
2314 makeGlobalInitializers(node->getAsAggregate()->getSequence());
2315
2316 // Initializers are done, don't want to visit again, but functions and link objects need to be processed,
2317 // so do them manually.
2318 visitFunctions(node->getAsAggregate()->getSequence());
2319
2320 return false;
2321 }
2322
2323 return true;
2324 }
2325 case glslang::EOpLinkerObjects:
2326 {
2327 if (visit == glslang::EvPreVisit)
2328 linkageOnly = true;
2329 else
2330 linkageOnly = false;
2331
2332 return true;
2333 }
2334 case glslang::EOpComma:
2335 {
2336 // processing from left to right naturally leaves the right-most
2337 // lying around in the access chain
2338 glslang::TIntermSequence& glslangOperands = node->getSequence();
2339 for (int i = 0; i < (int)glslangOperands.size(); ++i)
2340 glslangOperands[i]->traverse(this);
2341
2342 return false;
2343 }
2344 case glslang::EOpFunction:
2345 if (visit == glslang::EvPreVisit) {
2346 if (isShaderEntryPoint(node)) {
2347 inEntryPoint = true;
2348 builder.setBuildPoint(shaderEntry->getLastBlock());
2349 currentFunction = shaderEntry;
2350 } else {
2351 handleFunctionEntry(node);
2352 }
2353 } else {
2354 if (inEntryPoint)
2355 entryPointTerminated = true;
2356 builder.leaveFunction();
2357 inEntryPoint = false;
2358 }
2359
2360 return true;
2361 case glslang::EOpParameters:
2362 // Parameters will have been consumed by EOpFunction processing, but not
2363 // the body, so we still visited the function node's children, making this
2364 // child redundant.
2365 return false;
2366 case glslang::EOpFunctionCall:
2367 {
2368 builder.setLine(node->getLoc().line, node->getLoc().getFilename());
2369 if (node->isUserDefined())
2370 result = handleUserFunctionCall(node);
2371 // assert(result); // this can happen for bad shaders because the call graph completeness checking is not yet done
2372 if (result) {
2373 builder.clearAccessChain();
2374 builder.setAccessChainRValue(result);
2375 } else
2376 logger->missingFunctionality("missing user function; linker needs to catch that");
2377
2378 return false;
2379 }
2380 case glslang::EOpConstructMat2x2:
2381 case glslang::EOpConstructMat2x3:
2382 case glslang::EOpConstructMat2x4:
2383 case glslang::EOpConstructMat3x2:
2384 case glslang::EOpConstructMat3x3:
2385 case glslang::EOpConstructMat3x4:
2386 case glslang::EOpConstructMat4x2:
2387 case glslang::EOpConstructMat4x3:
2388 case glslang::EOpConstructMat4x4:
2389 case glslang::EOpConstructDMat2x2:
2390 case glslang::EOpConstructDMat2x3:
2391 case glslang::EOpConstructDMat2x4:
2392 case glslang::EOpConstructDMat3x2:
2393 case glslang::EOpConstructDMat3x3:
2394 case glslang::EOpConstructDMat3x4:
2395 case glslang::EOpConstructDMat4x2:
2396 case glslang::EOpConstructDMat4x3:
2397 case glslang::EOpConstructDMat4x4:
2398 case glslang::EOpConstructIMat2x2:
2399 case glslang::EOpConstructIMat2x3:
2400 case glslang::EOpConstructIMat2x4:
2401 case glslang::EOpConstructIMat3x2:
2402 case glslang::EOpConstructIMat3x3:
2403 case glslang::EOpConstructIMat3x4:
2404 case glslang::EOpConstructIMat4x2:
2405 case glslang::EOpConstructIMat4x3:
2406 case glslang::EOpConstructIMat4x4:
2407 case glslang::EOpConstructUMat2x2:
2408 case glslang::EOpConstructUMat2x3:
2409 case glslang::EOpConstructUMat2x4:
2410 case glslang::EOpConstructUMat3x2:
2411 case glslang::EOpConstructUMat3x3:
2412 case glslang::EOpConstructUMat3x4:
2413 case glslang::EOpConstructUMat4x2:
2414 case glslang::EOpConstructUMat4x3:
2415 case glslang::EOpConstructUMat4x4:
2416 case glslang::EOpConstructBMat2x2:
2417 case glslang::EOpConstructBMat2x3:
2418 case glslang::EOpConstructBMat2x4:
2419 case glslang::EOpConstructBMat3x2:
2420 case glslang::EOpConstructBMat3x3:
2421 case glslang::EOpConstructBMat3x4:
2422 case glslang::EOpConstructBMat4x2:
2423 case glslang::EOpConstructBMat4x3:
2424 case glslang::EOpConstructBMat4x4:
2425 case glslang::EOpConstructF16Mat2x2:
2426 case glslang::EOpConstructF16Mat2x3:
2427 case glslang::EOpConstructF16Mat2x4:
2428 case glslang::EOpConstructF16Mat3x2:
2429 case glslang::EOpConstructF16Mat3x3:
2430 case glslang::EOpConstructF16Mat3x4:
2431 case glslang::EOpConstructF16Mat4x2:
2432 case glslang::EOpConstructF16Mat4x3:
2433 case glslang::EOpConstructF16Mat4x4:
2434 isMatrix = true;
2435 // fall through
2436 case glslang::EOpConstructFloat:
2437 case glslang::EOpConstructVec2:
2438 case glslang::EOpConstructVec3:
2439 case glslang::EOpConstructVec4:
2440 case glslang::EOpConstructDouble:
2441 case glslang::EOpConstructDVec2:
2442 case glslang::EOpConstructDVec3:
2443 case glslang::EOpConstructDVec4:
2444 case glslang::EOpConstructFloat16:
2445 case glslang::EOpConstructF16Vec2:
2446 case glslang::EOpConstructF16Vec3:
2447 case glslang::EOpConstructF16Vec4:
2448 case glslang::EOpConstructBool:
2449 case glslang::EOpConstructBVec2:
2450 case glslang::EOpConstructBVec3:
2451 case glslang::EOpConstructBVec4:
2452 case glslang::EOpConstructInt8:
2453 case glslang::EOpConstructI8Vec2:
2454 case glslang::EOpConstructI8Vec3:
2455 case glslang::EOpConstructI8Vec4:
2456 case glslang::EOpConstructUint8:
2457 case glslang::EOpConstructU8Vec2:
2458 case glslang::EOpConstructU8Vec3:
2459 case glslang::EOpConstructU8Vec4:
2460 case glslang::EOpConstructInt16:
2461 case glslang::EOpConstructI16Vec2:
2462 case glslang::EOpConstructI16Vec3:
2463 case glslang::EOpConstructI16Vec4:
2464 case glslang::EOpConstructUint16:
2465 case glslang::EOpConstructU16Vec2:
2466 case glslang::EOpConstructU16Vec3:
2467 case glslang::EOpConstructU16Vec4:
2468 case glslang::EOpConstructInt:
2469 case glslang::EOpConstructIVec2:
2470 case glslang::EOpConstructIVec3:
2471 case glslang::EOpConstructIVec4:
2472 case glslang::EOpConstructUint:
2473 case glslang::EOpConstructUVec2:
2474 case glslang::EOpConstructUVec3:
2475 case glslang::EOpConstructUVec4:
2476 case glslang::EOpConstructInt64:
2477 case glslang::EOpConstructI64Vec2:
2478 case glslang::EOpConstructI64Vec3:
2479 case glslang::EOpConstructI64Vec4:
2480 case glslang::EOpConstructUint64:
2481 case glslang::EOpConstructU64Vec2:
2482 case glslang::EOpConstructU64Vec3:
2483 case glslang::EOpConstructU64Vec4:
2484 case glslang::EOpConstructStruct:
2485 case glslang::EOpConstructTextureSampler:
2486 case glslang::EOpConstructReference:
2487 case glslang::EOpConstructCooperativeMatrix:
2488 {
2489 builder.setLine(node->getLoc().line, node->getLoc().getFilename());
2490 std::vector<spv::Id> arguments;
2491 translateArguments(*node, arguments, lvalueCoherentFlags);
2492 spv::Id constructed;
2493 if (node->getOp() == glslang::EOpConstructTextureSampler)
2494 constructed = builder.createOp(spv::OpSampledImage, resultType(), arguments);
2495 else if (node->getOp() == glslang::EOpConstructStruct ||
2496 node->getOp() == glslang::EOpConstructCooperativeMatrix ||
2497 node->getType().isArray()) {
2498 std::vector<spv::Id> constituents;
2499 for (int c = 0; c < (int)arguments.size(); ++c)
2500 constituents.push_back(arguments[c]);
2501 constructed = createCompositeConstruct(resultType(), constituents);
2502 } else if (isMatrix)
2503 constructed = builder.createMatrixConstructor(precision, arguments, resultType());
2504 else
2505 constructed = builder.createConstructor(precision, arguments, resultType());
2506
2507 builder.clearAccessChain();
2508 builder.setAccessChainRValue(constructed);
2509
2510 return false;
2511 }
2512
2513 // These six are component-wise compares with component-wise results.
2514 // Forward on to createBinaryOperation(), requesting a vector result.
2515 case glslang::EOpLessThan:
2516 case glslang::EOpGreaterThan:
2517 case glslang::EOpLessThanEqual:
2518 case glslang::EOpGreaterThanEqual:
2519 case glslang::EOpVectorEqual:
2520 case glslang::EOpVectorNotEqual:
2521 {
2522 // Map the operation to a binary
2523 binOp = node->getOp();
2524 reduceComparison = false;
2525 switch (node->getOp()) {
2526 case glslang::EOpVectorEqual: binOp = glslang::EOpVectorEqual; break;
2527 case glslang::EOpVectorNotEqual: binOp = glslang::EOpVectorNotEqual; break;
2528 default: binOp = node->getOp(); break;
2529 }
2530
2531 break;
2532 }
2533 case glslang::EOpMul:
2534 // component-wise matrix multiply
2535 binOp = glslang::EOpMul;
2536 break;
2537 case glslang::EOpOuterProduct:
2538 // two vectors multiplied to make a matrix
2539 binOp = glslang::EOpOuterProduct;
2540 break;
2541 case glslang::EOpDot:
2542 {
2543 // for scalar dot product, use multiply
2544 glslang::TIntermSequence& glslangOperands = node->getSequence();
2545 if (glslangOperands[0]->getAsTyped()->getVectorSize() == 1)
2546 binOp = glslang::EOpMul;
2547 break;
2548 }
2549 case glslang::EOpMod:
2550 // when an aggregate, this is the floating-point mod built-in function,
2551 // which can be emitted by the one in createBinaryOperation()
2552 binOp = glslang::EOpMod;
2553 break;
2554
2555 case glslang::EOpEmitVertex:
2556 case glslang::EOpEndPrimitive:
2557 case glslang::EOpBarrier:
2558 case glslang::EOpMemoryBarrier:
2559 case glslang::EOpMemoryBarrierAtomicCounter:
2560 case glslang::EOpMemoryBarrierBuffer:
2561 case glslang::EOpMemoryBarrierImage:
2562 case glslang::EOpMemoryBarrierShared:
2563 case glslang::EOpGroupMemoryBarrier:
2564 case glslang::EOpDeviceMemoryBarrier:
2565 case glslang::EOpAllMemoryBarrierWithGroupSync:
2566 case glslang::EOpDeviceMemoryBarrierWithGroupSync:
2567 case glslang::EOpWorkgroupMemoryBarrier:
2568 case glslang::EOpWorkgroupMemoryBarrierWithGroupSync:
2569 case glslang::EOpSubgroupBarrier:
2570 case glslang::EOpSubgroupMemoryBarrier:
2571 case glslang::EOpSubgroupMemoryBarrierBuffer:
2572 case glslang::EOpSubgroupMemoryBarrierImage:
2573 case glslang::EOpSubgroupMemoryBarrierShared:
2574 noReturnValue = true;
2575 // These all have 0 operands and will naturally finish up in the code below for 0 operands
2576 break;
2577
2578 case glslang::EOpAtomicAdd:
2579 case glslang::EOpAtomicMin:
2580 case glslang::EOpAtomicMax:
2581 case glslang::EOpAtomicAnd:
2582 case glslang::EOpAtomicOr:
2583 case glslang::EOpAtomicXor:
2584 case glslang::EOpAtomicExchange:
2585 case glslang::EOpAtomicCompSwap:
2586 atomic = true;
2587 break;
2588
2589 #ifndef GLSLANG_WEB
2590 case glslang::EOpAtomicStore:
2591 noReturnValue = true;
2592 // fallthrough
2593 case glslang::EOpAtomicLoad:
2594 atomic = true;
2595 break;
2596
2597 case glslang::EOpAtomicCounterAdd:
2598 case glslang::EOpAtomicCounterSubtract:
2599 case glslang::EOpAtomicCounterMin:
2600 case glslang::EOpAtomicCounterMax:
2601 case glslang::EOpAtomicCounterAnd:
2602 case glslang::EOpAtomicCounterOr:
2603 case glslang::EOpAtomicCounterXor:
2604 case glslang::EOpAtomicCounterExchange:
2605 case glslang::EOpAtomicCounterCompSwap:
2606 builder.addExtension("SPV_KHR_shader_atomic_counter_ops");
2607 builder.addCapability(spv::CapabilityAtomicStorageOps);
2608 atomic = true;
2609 break;
2610
2611 case glslang::EOpIgnoreIntersectionNV:
2612 case glslang::EOpTerminateRayNV:
2613 case glslang::EOpTraceNV:
2614 case glslang::EOpExecuteCallableNV:
2615 case glslang::EOpWritePackedPrimitiveIndices4x8NV:
2616 noReturnValue = true;
2617 break;
2618 case glslang::EOpCooperativeMatrixLoad:
2619 case glslang::EOpCooperativeMatrixStore:
2620 noReturnValue = true;
2621 break;
2622 case glslang::EOpBeginInvocationInterlock:
2623 case glslang::EOpEndInvocationInterlock:
2624 builder.addExtension(spv::E_SPV_EXT_fragment_shader_interlock);
2625 noReturnValue = true;
2626 break;
2627 #endif
2628
2629 default:
2630 break;
2631 }
2632
2633 //
2634 // See if it maps to a regular operation.
2635 //
2636 if (binOp != glslang::EOpNull) {
2637 glslang::TIntermTyped* left = node->getSequence()[0]->getAsTyped();
2638 glslang::TIntermTyped* right = node->getSequence()[1]->getAsTyped();
2639 assert(left && right);
2640
2641 builder.clearAccessChain();
2642 left->traverse(this);
2643 spv::Id leftId = accessChainLoad(left->getType());
2644
2645 builder.clearAccessChain();
2646 right->traverse(this);
2647 spv::Id rightId = accessChainLoad(right->getType());
2648
2649 builder.setLine(node->getLoc().line, node->getLoc().getFilename());
2650 OpDecorations decorations = { precision,
2651 TranslateNoContractionDecoration(node->getType().getQualifier()),
2652 TranslateNonUniformDecoration(node->getType().getQualifier()) };
2653 result = createBinaryOperation(binOp, decorations,
2654 resultType(), leftId, rightId,
2655 left->getType().getBasicType(), reduceComparison);
2656
2657 // code above should only make binOp that exists in createBinaryOperation
2658 assert(result != spv::NoResult);
2659 builder.clearAccessChain();
2660 builder.setAccessChainRValue(result);
2661
2662 return false;
2663 }
2664
2665 //
2666 // Create the list of operands.
2667 //
2668 glslang::TIntermSequence& glslangOperands = node->getSequence();
2669 std::vector<spv::Id> operands;
2670 std::vector<spv::IdImmediate> memoryAccessOperands;
2671 for (int arg = 0; arg < (int)glslangOperands.size(); ++arg) {
2672 // special case l-value operands; there are just a few
2673 bool lvalue = false;
2674 switch (node->getOp()) {
2675 case glslang::EOpModf:
2676 if (arg == 1)
2677 lvalue = true;
2678 break;
2679
2680 case glslang::EOpAtomicAdd:
2681 case glslang::EOpAtomicMin:
2682 case glslang::EOpAtomicMax:
2683 case glslang::EOpAtomicAnd:
2684 case glslang::EOpAtomicOr:
2685 case glslang::EOpAtomicXor:
2686 case glslang::EOpAtomicExchange:
2687 case glslang::EOpAtomicCompSwap:
2688 if (arg == 0)
2689 lvalue = true;
2690 break;
2691
2692 #ifndef GLSLANG_WEB
2693 case glslang::EOpFrexp:
2694 if (arg == 1)
2695 lvalue = true;
2696 break;
2697 case glslang::EOpInterpolateAtSample:
2698 case glslang::EOpInterpolateAtOffset:
2699 case glslang::EOpInterpolateAtVertex:
2700 if (arg == 0) {
2701 lvalue = true;
2702
2703 // Does it need a swizzle inversion? If so, evaluation is inverted;
2704 // operate first on the swizzle base, then apply the swizzle.
2705 if (glslangOperands[0]->getAsOperator() &&
2706 glslangOperands[0]->getAsOperator()->getOp() == glslang::EOpVectorSwizzle)
2707 invertedType = convertGlslangToSpvType(glslangOperands[0]->getAsBinaryNode()->getLeft()->getType());
2708 }
2709 break;
2710 case glslang::EOpAtomicLoad:
2711 case glslang::EOpAtomicStore:
2712 case glslang::EOpAtomicCounterAdd:
2713 case glslang::EOpAtomicCounterSubtract:
2714 case glslang::EOpAtomicCounterMin:
2715 case glslang::EOpAtomicCounterMax:
2716 case glslang::EOpAtomicCounterAnd:
2717 case glslang::EOpAtomicCounterOr:
2718 case glslang::EOpAtomicCounterXor:
2719 case glslang::EOpAtomicCounterExchange:
2720 case glslang::EOpAtomicCounterCompSwap:
2721 if (arg == 0)
2722 lvalue = true;
2723 break;
2724 case glslang::EOpAddCarry:
2725 case glslang::EOpSubBorrow:
2726 if (arg == 2)
2727 lvalue = true;
2728 break;
2729 case glslang::EOpUMulExtended:
2730 case glslang::EOpIMulExtended:
2731 if (arg >= 2)
2732 lvalue = true;
2733 break;
2734 case glslang::EOpCooperativeMatrixLoad:
2735 if (arg == 0 || arg == 1)
2736 lvalue = true;
2737 break;
2738 case glslang::EOpCooperativeMatrixStore:
2739 if (arg == 1)
2740 lvalue = true;
2741 break;
2742 #endif
2743 default:
2744 break;
2745 }
2746 builder.clearAccessChain();
2747 if (invertedType != spv::NoType && arg == 0)
2748 glslangOperands[0]->getAsBinaryNode()->getLeft()->traverse(this);
2749 else
2750 glslangOperands[arg]->traverse(this);
2751
2752 #ifndef GLSLANG_WEB
2753 if (node->getOp() == glslang::EOpCooperativeMatrixLoad ||
2754 node->getOp() == glslang::EOpCooperativeMatrixStore) {
2755
2756 if (arg == 1) {
2757 // fold "element" parameter into the access chain
2758 spv::Builder::AccessChain save = builder.getAccessChain();
2759 builder.clearAccessChain();
2760 glslangOperands[2]->traverse(this);
2761
2762 spv::Id elementId = accessChainLoad(glslangOperands[2]->getAsTyped()->getType());
2763
2764 builder.setAccessChain(save);
2765
2766 // Point to the first element of the array.
2767 builder.accessChainPush(elementId, TranslateCoherent(glslangOperands[arg]->getAsTyped()->getType()),
2768 glslangOperands[arg]->getAsTyped()->getType().getBufferReferenceAlignment());
2769
2770 spv::Builder::AccessChain::CoherentFlags coherentFlags = builder.getAccessChain().coherentFlags;
2771 unsigned int alignment = builder.getAccessChain().alignment;
2772
2773 int memoryAccess = TranslateMemoryAccess(coherentFlags);
2774 if (node->getOp() == glslang::EOpCooperativeMatrixLoad)
2775 memoryAccess &= ~spv::MemoryAccessMakePointerAvailableKHRMask;
2776 if (node->getOp() == glslang::EOpCooperativeMatrixStore)
2777 memoryAccess &= ~spv::MemoryAccessMakePointerVisibleKHRMask;
2778 if (builder.getStorageClass(builder.getAccessChain().base) == spv::StorageClassPhysicalStorageBufferEXT) {
2779 memoryAccess = (spv::MemoryAccessMask)(memoryAccess | spv::MemoryAccessAlignedMask);
2780 }
2781
2782 memoryAccessOperands.push_back(spv::IdImmediate(false, memoryAccess));
2783
2784 if (memoryAccess & spv::MemoryAccessAlignedMask) {
2785 memoryAccessOperands.push_back(spv::IdImmediate(false, alignment));
2786 }
2787
2788 if (memoryAccess & (spv::MemoryAccessMakePointerAvailableKHRMask | spv::MemoryAccessMakePointerVisibleKHRMask)) {
2789 memoryAccessOperands.push_back(spv::IdImmediate(true, builder.makeUintConstant(TranslateMemoryScope(coherentFlags))));
2790 }
2791 } else if (arg == 2) {
2792 continue;
2793 }
2794 }
2795 #endif
2796
2797 if (lvalue) {
2798 operands.push_back(builder.accessChainGetLValue());
2799 lvalueCoherentFlags = builder.getAccessChain().coherentFlags;
2800 lvalueCoherentFlags |= TranslateCoherent(glslangOperands[arg]->getAsTyped()->getType());
2801 } else {
2802 builder.setLine(node->getLoc().line, node->getLoc().getFilename());
2803 operands.push_back(accessChainLoad(glslangOperands[arg]->getAsTyped()->getType()));
2804 }
2805 }
2806
2807 builder.setLine(node->getLoc().line, node->getLoc().getFilename());
2808 #ifndef GLSLANG_WEB
2809 if (node->getOp() == glslang::EOpCooperativeMatrixLoad) {
2810 std::vector<spv::IdImmediate> idImmOps;
2811
2812 idImmOps.push_back(spv::IdImmediate(true, operands[1])); // buf
2813 idImmOps.push_back(spv::IdImmediate(true, operands[2])); // stride
2814 idImmOps.push_back(spv::IdImmediate(true, operands[3])); // colMajor
2815 idImmOps.insert(idImmOps.end(), memoryAccessOperands.begin(), memoryAccessOperands.end());
2816 // get the pointee type
2817 spv::Id typeId = builder.getContainedTypeId(builder.getTypeId(operands[0]));
2818 assert(builder.isCooperativeMatrixType(typeId));
2819 // do the op
2820 spv::Id result = builder.createOp(spv::OpCooperativeMatrixLoadNV, typeId, idImmOps);
2821 // store the result to the pointer (out param 'm')
2822 builder.createStore(result, operands[0]);
2823 result = 0;
2824 } else if (node->getOp() == glslang::EOpCooperativeMatrixStore) {
2825 std::vector<spv::IdImmediate> idImmOps;
2826
2827 idImmOps.push_back(spv::IdImmediate(true, operands[1])); // buf
2828 idImmOps.push_back(spv::IdImmediate(true, operands[0])); // object
2829 idImmOps.push_back(spv::IdImmediate(true, operands[2])); // stride
2830 idImmOps.push_back(spv::IdImmediate(true, operands[3])); // colMajor
2831 idImmOps.insert(idImmOps.end(), memoryAccessOperands.begin(), memoryAccessOperands.end());
2832
2833 builder.createNoResultOp(spv::OpCooperativeMatrixStoreNV, idImmOps);
2834 result = 0;
2835 } else
2836 #endif
2837 if (atomic) {
2838 // Handle all atomics
2839 result = createAtomicOperation(node->getOp(), precision, resultType(), operands, node->getBasicType(), lvalueCoherentFlags);
2840 } else {
2841 // Pass through to generic operations.
2842 switch (glslangOperands.size()) {
2843 case 0:
2844 result = createNoArgOperation(node->getOp(), precision, resultType());
2845 break;
2846 case 1:
2847 {
2848 OpDecorations decorations = { precision,
2849 TranslateNoContractionDecoration(node->getType().getQualifier()),
2850 TranslateNonUniformDecoration(node->getType().getQualifier()) };
2851 result = createUnaryOperation(
2852 node->getOp(), decorations,
2853 resultType(), operands.front(),
2854 glslangOperands[0]->getAsTyped()->getBasicType(), lvalueCoherentFlags);
2855 }
2856 break;
2857 default:
2858 result = createMiscOperation(node->getOp(), precision, resultType(), operands, node->getBasicType());
2859 break;
2860 }
2861 if (invertedType)
2862 result = createInvertedSwizzle(precision, *glslangOperands[0]->getAsBinaryNode(), result);
2863 }
2864
2865 if (noReturnValue)
2866 return false;
2867
2868 if (! result) {
2869 logger->missingFunctionality("unknown glslang aggregate");
2870 return true; // pick up a child as a placeholder operand
2871 } else {
2872 builder.clearAccessChain();
2873 builder.setAccessChainRValue(result);
2874 return false;
2875 }
2876 }
2877
2878 // This path handles both if-then-else and ?:
2879 // The if-then-else has a node type of void, while
2880 // ?: has either a void or a non-void node type
2881 //
2882 // Leaving the result, when not void:
2883 // GLSL only has r-values as the result of a :?, but
2884 // if we have an l-value, that can be more efficient if it will
2885 // become the base of a complex r-value expression, because the
2886 // next layer copies r-values into memory to use the access-chain mechanism
visitSelection(glslang::TVisit,glslang::TIntermSelection * node)2887 bool TGlslangToSpvTraverser::visitSelection(glslang::TVisit /* visit */, glslang::TIntermSelection* node)
2888 {
2889 // see if OpSelect can handle it
2890 const auto isOpSelectable = [&]() {
2891 if (node->getBasicType() == glslang::EbtVoid)
2892 return false;
2893 // OpSelect can do all other types starting with SPV 1.4
2894 if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_4) {
2895 // pre-1.4, only scalars and vectors can be handled
2896 if ((!node->getType().isScalar() && !node->getType().isVector()))
2897 return false;
2898 }
2899 return true;
2900 };
2901
2902 // See if it simple and safe, or required, to execute both sides.
2903 // Crucially, side effects must be either semantically required or avoided,
2904 // and there are performance trade-offs.
2905 // Return true if required or a good idea (and safe) to execute both sides,
2906 // false otherwise.
2907 const auto bothSidesPolicy = [&]() -> bool {
2908 // do we have both sides?
2909 if (node->getTrueBlock() == nullptr ||
2910 node->getFalseBlock() == nullptr)
2911 return false;
2912
2913 // required? (unless we write additional code to look for side effects
2914 // and make performance trade-offs if none are present)
2915 if (!node->getShortCircuit())
2916 return true;
2917
2918 // if not required to execute both, decide based on performance/practicality...
2919
2920 if (!isOpSelectable())
2921 return false;
2922
2923 assert(node->getType() == node->getTrueBlock() ->getAsTyped()->getType() &&
2924 node->getType() == node->getFalseBlock()->getAsTyped()->getType());
2925
2926 // return true if a single operand to ? : is okay for OpSelect
2927 const auto operandOkay = [](glslang::TIntermTyped* node) {
2928 return node->getAsSymbolNode() || node->getType().getQualifier().isConstant();
2929 };
2930
2931 return operandOkay(node->getTrueBlock() ->getAsTyped()) &&
2932 operandOkay(node->getFalseBlock()->getAsTyped());
2933 };
2934
2935 spv::Id result = spv::NoResult; // upcoming result selecting between trueValue and falseValue
2936 // emit the condition before doing anything with selection
2937 node->getCondition()->traverse(this);
2938 spv::Id condition = accessChainLoad(node->getCondition()->getType());
2939
2940 // Find a way of executing both sides and selecting the right result.
2941 const auto executeBothSides = [&]() -> void {
2942 // execute both sides
2943 node->getTrueBlock()->traverse(this);
2944 spv::Id trueValue = accessChainLoad(node->getTrueBlock()->getAsTyped()->getType());
2945 node->getFalseBlock()->traverse(this);
2946 spv::Id falseValue = accessChainLoad(node->getTrueBlock()->getAsTyped()->getType());
2947
2948 builder.setLine(node->getLoc().line, node->getLoc().getFilename());
2949
2950 // done if void
2951 if (node->getBasicType() == glslang::EbtVoid)
2952 return;
2953
2954 // emit code to select between trueValue and falseValue
2955
2956 // see if OpSelect can handle it
2957 if (isOpSelectable()) {
2958 // Emit OpSelect for this selection.
2959
2960 // smear condition to vector, if necessary (AST is always scalar)
2961 // Before 1.4, smear like for mix(), starting with 1.4, keep it scalar
2962 if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_4 && builder.isVector(trueValue)) {
2963 condition = builder.smearScalar(spv::NoPrecision, condition,
2964 builder.makeVectorType(builder.makeBoolType(),
2965 builder.getNumComponents(trueValue)));
2966 }
2967
2968 // OpSelect
2969 result = builder.createTriOp(spv::OpSelect,
2970 convertGlslangToSpvType(node->getType()), condition,
2971 trueValue, falseValue);
2972
2973 builder.clearAccessChain();
2974 builder.setAccessChainRValue(result);
2975 } else {
2976 // We need control flow to select the result.
2977 // TODO: Once SPIR-V OpSelect allows arbitrary types, eliminate this path.
2978 result = builder.createVariable(spv::StorageClassFunction, convertGlslangToSpvType(node->getType()));
2979
2980 // Selection control:
2981 const spv::SelectionControlMask control = TranslateSelectionControl(*node);
2982
2983 // make an "if" based on the value created by the condition
2984 spv::Builder::If ifBuilder(condition, control, builder);
2985
2986 // emit the "then" statement
2987 builder.createStore(trueValue, result);
2988 ifBuilder.makeBeginElse();
2989 // emit the "else" statement
2990 builder.createStore(falseValue, result);
2991
2992 // finish off the control flow
2993 ifBuilder.makeEndIf();
2994
2995 builder.clearAccessChain();
2996 builder.setAccessChainLValue(result);
2997 }
2998 };
2999
3000 // Execute the one side needed, as per the condition
3001 const auto executeOneSide = [&]() {
3002 // Always emit control flow.
3003 if (node->getBasicType() != glslang::EbtVoid)
3004 result = builder.createVariable(spv::StorageClassFunction, convertGlslangToSpvType(node->getType()));
3005
3006 // Selection control:
3007 const spv::SelectionControlMask control = TranslateSelectionControl(*node);
3008
3009 // make an "if" based on the value created by the condition
3010 spv::Builder::If ifBuilder(condition, control, builder);
3011
3012 // emit the "then" statement
3013 if (node->getTrueBlock() != nullptr) {
3014 node->getTrueBlock()->traverse(this);
3015 if (result != spv::NoResult)
3016 builder.createStore(accessChainLoad(node->getTrueBlock()->getAsTyped()->getType()), result);
3017 }
3018
3019 if (node->getFalseBlock() != nullptr) {
3020 ifBuilder.makeBeginElse();
3021 // emit the "else" statement
3022 node->getFalseBlock()->traverse(this);
3023 if (result != spv::NoResult)
3024 builder.createStore(accessChainLoad(node->getFalseBlock()->getAsTyped()->getType()), result);
3025 }
3026
3027 // finish off the control flow
3028 ifBuilder.makeEndIf();
3029
3030 if (result != spv::NoResult) {
3031 builder.clearAccessChain();
3032 builder.setAccessChainLValue(result);
3033 }
3034 };
3035
3036 // Try for OpSelect (or a requirement to execute both sides)
3037 if (bothSidesPolicy()) {
3038 SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder);
3039 if (node->getType().getQualifier().isSpecConstant())
3040 spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
3041 executeBothSides();
3042 } else
3043 executeOneSide();
3044
3045 return false;
3046 }
3047
visitSwitch(glslang::TVisit,glslang::TIntermSwitch * node)3048 bool TGlslangToSpvTraverser::visitSwitch(glslang::TVisit /* visit */, glslang::TIntermSwitch* node)
3049 {
3050 // emit and get the condition before doing anything with switch
3051 node->getCondition()->traverse(this);
3052 spv::Id selector = accessChainLoad(node->getCondition()->getAsTyped()->getType());
3053
3054 // Selection control:
3055 const spv::SelectionControlMask control = TranslateSwitchControl(*node);
3056
3057 // browse the children to sort out code segments
3058 int defaultSegment = -1;
3059 std::vector<TIntermNode*> codeSegments;
3060 glslang::TIntermSequence& sequence = node->getBody()->getSequence();
3061 std::vector<int> caseValues;
3062 std::vector<int> valueIndexToSegment(sequence.size()); // note: probably not all are used, it is an overestimate
3063 for (glslang::TIntermSequence::iterator c = sequence.begin(); c != sequence.end(); ++c) {
3064 TIntermNode* child = *c;
3065 if (child->getAsBranchNode() && child->getAsBranchNode()->getFlowOp() == glslang::EOpDefault)
3066 defaultSegment = (int)codeSegments.size();
3067 else if (child->getAsBranchNode() && child->getAsBranchNode()->getFlowOp() == glslang::EOpCase) {
3068 valueIndexToSegment[caseValues.size()] = (int)codeSegments.size();
3069 caseValues.push_back(child->getAsBranchNode()->getExpression()->getAsConstantUnion()->getConstArray()[0].getIConst());
3070 } else
3071 codeSegments.push_back(child);
3072 }
3073
3074 // handle the case where the last code segment is missing, due to no code
3075 // statements between the last case and the end of the switch statement
3076 if ((caseValues.size() && (int)codeSegments.size() == valueIndexToSegment[caseValues.size() - 1]) ||
3077 (int)codeSegments.size() == defaultSegment)
3078 codeSegments.push_back(nullptr);
3079
3080 // make the switch statement
3081 std::vector<spv::Block*> segmentBlocks; // returned, as the blocks allocated in the call
3082 builder.makeSwitch(selector, control, (int)codeSegments.size(), caseValues, valueIndexToSegment, defaultSegment, segmentBlocks);
3083
3084 // emit all the code in the segments
3085 breakForLoop.push(false);
3086 for (unsigned int s = 0; s < codeSegments.size(); ++s) {
3087 builder.nextSwitchSegment(segmentBlocks, s);
3088 if (codeSegments[s])
3089 codeSegments[s]->traverse(this);
3090 else
3091 builder.addSwitchBreak();
3092 }
3093 breakForLoop.pop();
3094
3095 builder.endSwitch(segmentBlocks);
3096
3097 return false;
3098 }
3099
visitConstantUnion(glslang::TIntermConstantUnion * node)3100 void TGlslangToSpvTraverser::visitConstantUnion(glslang::TIntermConstantUnion* node)
3101 {
3102 int nextConst = 0;
3103 spv::Id constant = createSpvConstantFromConstUnionArray(node->getType(), node->getConstArray(), nextConst, false);
3104
3105 builder.clearAccessChain();
3106 builder.setAccessChainRValue(constant);
3107 }
3108
visitLoop(glslang::TVisit,glslang::TIntermLoop * node)3109 bool TGlslangToSpvTraverser::visitLoop(glslang::TVisit /* visit */, glslang::TIntermLoop* node)
3110 {
3111 auto blocks = builder.makeNewLoop();
3112 builder.createBranch(&blocks.head);
3113
3114 // Loop control:
3115 std::vector<unsigned int> operands;
3116 const spv::LoopControlMask control = TranslateLoopControl(*node, operands);
3117
3118 // Spec requires back edges to target header blocks, and every header block
3119 // must dominate its merge block. Make a header block first to ensure these
3120 // conditions are met. By definition, it will contain OpLoopMerge, followed
3121 // by a block-ending branch. But we don't want to put any other body/test
3122 // instructions in it, since the body/test may have arbitrary instructions,
3123 // including merges of its own.
3124 builder.setLine(node->getLoc().line, node->getLoc().getFilename());
3125 builder.setBuildPoint(&blocks.head);
3126 builder.createLoopMerge(&blocks.merge, &blocks.continue_target, control, operands);
3127 if (node->testFirst() && node->getTest()) {
3128 spv::Block& test = builder.makeNewBlock();
3129 builder.createBranch(&test);
3130
3131 builder.setBuildPoint(&test);
3132 node->getTest()->traverse(this);
3133 spv::Id condition = accessChainLoad(node->getTest()->getType());
3134 builder.createConditionalBranch(condition, &blocks.body, &blocks.merge);
3135
3136 builder.setBuildPoint(&blocks.body);
3137 breakForLoop.push(true);
3138 if (node->getBody())
3139 node->getBody()->traverse(this);
3140 builder.createBranch(&blocks.continue_target);
3141 breakForLoop.pop();
3142
3143 builder.setBuildPoint(&blocks.continue_target);
3144 if (node->getTerminal())
3145 node->getTerminal()->traverse(this);
3146 builder.createBranch(&blocks.head);
3147 } else {
3148 builder.setLine(node->getLoc().line, node->getLoc().getFilename());
3149 builder.createBranch(&blocks.body);
3150
3151 breakForLoop.push(true);
3152 builder.setBuildPoint(&blocks.body);
3153 if (node->getBody())
3154 node->getBody()->traverse(this);
3155 builder.createBranch(&blocks.continue_target);
3156 breakForLoop.pop();
3157
3158 builder.setBuildPoint(&blocks.continue_target);
3159 if (node->getTerminal())
3160 node->getTerminal()->traverse(this);
3161 if (node->getTest()) {
3162 node->getTest()->traverse(this);
3163 spv::Id condition =
3164 accessChainLoad(node->getTest()->getType());
3165 builder.createConditionalBranch(condition, &blocks.head, &blocks.merge);
3166 } else {
3167 // TODO: unless there was a break/return/discard instruction
3168 // somewhere in the body, this is an infinite loop, so we should
3169 // issue a warning.
3170 builder.createBranch(&blocks.head);
3171 }
3172 }
3173 builder.setBuildPoint(&blocks.merge);
3174 builder.closeLoop();
3175 return false;
3176 }
3177
visitBranch(glslang::TVisit,glslang::TIntermBranch * node)3178 bool TGlslangToSpvTraverser::visitBranch(glslang::TVisit /* visit */, glslang::TIntermBranch* node)
3179 {
3180 if (node->getExpression())
3181 node->getExpression()->traverse(this);
3182
3183 builder.setLine(node->getLoc().line, node->getLoc().getFilename());
3184
3185 switch (node->getFlowOp()) {
3186 case glslang::EOpKill:
3187 builder.makeDiscard();
3188 break;
3189 case glslang::EOpBreak:
3190 if (breakForLoop.top())
3191 builder.createLoopExit();
3192 else
3193 builder.addSwitchBreak();
3194 break;
3195 case glslang::EOpContinue:
3196 builder.createLoopContinue();
3197 break;
3198 case glslang::EOpReturn:
3199 if (node->getExpression()) {
3200 const glslang::TType& glslangReturnType = node->getExpression()->getType();
3201 spv::Id returnId = accessChainLoad(glslangReturnType);
3202 if (builder.getTypeId(returnId) != currentFunction->getReturnType()) {
3203 builder.clearAccessChain();
3204 spv::Id copyId = builder.createVariable(spv::StorageClassFunction, currentFunction->getReturnType());
3205 builder.setAccessChainLValue(copyId);
3206 multiTypeStore(glslangReturnType, returnId);
3207 returnId = builder.createLoad(copyId);
3208 }
3209 builder.makeReturn(false, returnId);
3210 } else
3211 builder.makeReturn(false);
3212
3213 builder.clearAccessChain();
3214 break;
3215
3216 #ifndef GLSLANG_WEB
3217 case glslang::EOpDemote:
3218 builder.createNoResultOp(spv::OpDemoteToHelperInvocationEXT);
3219 builder.addExtension(spv::E_SPV_EXT_demote_to_helper_invocation);
3220 builder.addCapability(spv::CapabilityDemoteToHelperInvocationEXT);
3221 break;
3222 #endif
3223
3224 default:
3225 assert(0);
3226 break;
3227 }
3228
3229 return false;
3230 }
3231
createSpvVariable(const glslang::TIntermSymbol * node,spv::Id forcedType)3232 spv::Id TGlslangToSpvTraverser::createSpvVariable(const glslang::TIntermSymbol* node, spv::Id forcedType)
3233 {
3234 // First, steer off constants, which are not SPIR-V variables, but
3235 // can still have a mapping to a SPIR-V Id.
3236 // This includes specialization constants.
3237 if (node->getQualifier().isConstant()) {
3238 spv::Id result = createSpvConstant(*node);
3239 if (result != spv::NoResult)
3240 return result;
3241 }
3242
3243 // Now, handle actual variables
3244 spv::StorageClass storageClass = TranslateStorageClass(node->getType());
3245 spv::Id spvType = forcedType == spv::NoType ? convertGlslangToSpvType(node->getType())
3246 : forcedType;
3247
3248 const bool contains16BitType = node->getType().contains16BitFloat() ||
3249 node->getType().contains16BitInt();
3250 if (contains16BitType) {
3251 switch (storageClass) {
3252 case spv::StorageClassInput:
3253 case spv::StorageClassOutput:
3254 builder.addIncorporatedExtension(spv::E_SPV_KHR_16bit_storage, spv::Spv_1_3);
3255 builder.addCapability(spv::CapabilityStorageInputOutput16);
3256 break;
3257 case spv::StorageClassUniform:
3258 builder.addIncorporatedExtension(spv::E_SPV_KHR_16bit_storage, spv::Spv_1_3);
3259 if (node->getType().getQualifier().storage == glslang::EvqBuffer)
3260 builder.addCapability(spv::CapabilityStorageUniformBufferBlock16);
3261 else
3262 builder.addCapability(spv::CapabilityStorageUniform16);
3263 break;
3264 #ifndef GLSLANG_WEB
3265 case spv::StorageClassPushConstant:
3266 builder.addIncorporatedExtension(spv::E_SPV_KHR_16bit_storage, spv::Spv_1_3);
3267 builder.addCapability(spv::CapabilityStoragePushConstant16);
3268 break;
3269 case spv::StorageClassStorageBuffer:
3270 case spv::StorageClassPhysicalStorageBufferEXT:
3271 builder.addIncorporatedExtension(spv::E_SPV_KHR_16bit_storage, spv::Spv_1_3);
3272 builder.addCapability(spv::CapabilityStorageUniformBufferBlock16);
3273 break;
3274 #endif
3275 default:
3276 if (node->getType().contains16BitFloat())
3277 builder.addCapability(spv::CapabilityFloat16);
3278 if (node->getType().contains16BitInt())
3279 builder.addCapability(spv::CapabilityInt16);
3280 break;
3281 }
3282 }
3283
3284 if (node->getType().contains8BitInt()) {
3285 if (storageClass == spv::StorageClassPushConstant) {
3286 builder.addIncorporatedExtension(spv::E_SPV_KHR_8bit_storage, spv::Spv_1_5);
3287 builder.addCapability(spv::CapabilityStoragePushConstant8);
3288 } else if (storageClass == spv::StorageClassUniform) {
3289 builder.addIncorporatedExtension(spv::E_SPV_KHR_8bit_storage, spv::Spv_1_5);
3290 builder.addCapability(spv::CapabilityUniformAndStorageBuffer8BitAccess);
3291 } else if (storageClass == spv::StorageClassStorageBuffer) {
3292 builder.addIncorporatedExtension(spv::E_SPV_KHR_8bit_storage, spv::Spv_1_5);
3293 builder.addCapability(spv::CapabilityStorageBuffer8BitAccess);
3294 } else {
3295 builder.addCapability(spv::CapabilityInt8);
3296 }
3297 }
3298
3299 const char* name = node->getName().c_str();
3300 if (glslang::IsAnonymous(name))
3301 name = "";
3302
3303 return builder.createVariable(storageClass, spvType, name);
3304 }
3305
3306 // Return type Id of the sampled type.
getSampledType(const glslang::TSampler & sampler)3307 spv::Id TGlslangToSpvTraverser::getSampledType(const glslang::TSampler& sampler)
3308 {
3309 switch (sampler.type) {
3310 case glslang::EbtInt: return builder.makeIntType(32);
3311 case glslang::EbtUint: return builder.makeUintType(32);
3312 case glslang::EbtFloat: return builder.makeFloatType(32);
3313 #ifndef GLSLANG_WEB
3314 case glslang::EbtFloat16:
3315 builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float_fetch);
3316 builder.addCapability(spv::CapabilityFloat16ImageAMD);
3317 return builder.makeFloatType(16);
3318 #endif
3319 default:
3320 assert(0);
3321 return builder.makeFloatType(32);
3322 }
3323 }
3324
3325 // If node is a swizzle operation, return the type that should be used if
3326 // the swizzle base is first consumed by another operation, before the swizzle
3327 // is applied.
getInvertedSwizzleType(const glslang::TIntermTyped & node)3328 spv::Id TGlslangToSpvTraverser::getInvertedSwizzleType(const glslang::TIntermTyped& node)
3329 {
3330 if (node.getAsOperator() &&
3331 node.getAsOperator()->getOp() == glslang::EOpVectorSwizzle)
3332 return convertGlslangToSpvType(node.getAsBinaryNode()->getLeft()->getType());
3333 else
3334 return spv::NoType;
3335 }
3336
3337 // When inverting a swizzle with a parent op, this function
3338 // will apply the swizzle operation to a completed parent operation.
createInvertedSwizzle(spv::Decoration precision,const glslang::TIntermTyped & node,spv::Id parentResult)3339 spv::Id TGlslangToSpvTraverser::createInvertedSwizzle(spv::Decoration precision, const glslang::TIntermTyped& node, spv::Id parentResult)
3340 {
3341 std::vector<unsigned> swizzle;
3342 convertSwizzle(*node.getAsBinaryNode()->getRight()->getAsAggregate(), swizzle);
3343 return builder.createRvalueSwizzle(precision, convertGlslangToSpvType(node.getType()), parentResult, swizzle);
3344 }
3345
3346 // Convert a glslang AST swizzle node to a swizzle vector for building SPIR-V.
convertSwizzle(const glslang::TIntermAggregate & node,std::vector<unsigned> & swizzle)3347 void TGlslangToSpvTraverser::convertSwizzle(const glslang::TIntermAggregate& node, std::vector<unsigned>& swizzle)
3348 {
3349 const glslang::TIntermSequence& swizzleSequence = node.getSequence();
3350 for (int i = 0; i < (int)swizzleSequence.size(); ++i)
3351 swizzle.push_back(swizzleSequence[i]->getAsConstantUnion()->getConstArray()[0].getIConst());
3352 }
3353
3354 // Convert from a glslang type to an SPV type, by calling into a
3355 // recursive version of this function. This establishes the inherited
3356 // layout state rooted from the top-level type.
convertGlslangToSpvType(const glslang::TType & type,bool forwardReferenceOnly)3357 spv::Id TGlslangToSpvTraverser::convertGlslangToSpvType(const glslang::TType& type, bool forwardReferenceOnly)
3358 {
3359 return convertGlslangToSpvType(type, getExplicitLayout(type), type.getQualifier(), false, forwardReferenceOnly);
3360 }
3361
3362 // Do full recursive conversion of an arbitrary glslang type to a SPIR-V Id.
3363 // explicitLayout can be kept the same throughout the hierarchical recursive walk.
3364 // Mutually recursive with convertGlslangStructToSpvType().
convertGlslangToSpvType(const glslang::TType & type,glslang::TLayoutPacking explicitLayout,const glslang::TQualifier & qualifier,bool lastBufferBlockMember,bool forwardReferenceOnly)3365 spv::Id TGlslangToSpvTraverser::convertGlslangToSpvType(const glslang::TType& type,
3366 glslang::TLayoutPacking explicitLayout, const glslang::TQualifier& qualifier,
3367 bool lastBufferBlockMember, bool forwardReferenceOnly)
3368 {
3369 spv::Id spvType = spv::NoResult;
3370
3371 switch (type.getBasicType()) {
3372 case glslang::EbtVoid:
3373 spvType = builder.makeVoidType();
3374 assert (! type.isArray());
3375 break;
3376 case glslang::EbtBool:
3377 // "transparent" bool doesn't exist in SPIR-V. The GLSL convention is
3378 // a 32-bit int where non-0 means true.
3379 if (explicitLayout != glslang::ElpNone)
3380 spvType = builder.makeUintType(32);
3381 else
3382 spvType = builder.makeBoolType();
3383 break;
3384 case glslang::EbtInt:
3385 spvType = builder.makeIntType(32);
3386 break;
3387 case glslang::EbtUint:
3388 spvType = builder.makeUintType(32);
3389 break;
3390 case glslang::EbtFloat:
3391 spvType = builder.makeFloatType(32);
3392 break;
3393 #ifndef GLSLANG_WEB
3394 case glslang::EbtDouble:
3395 spvType = builder.makeFloatType(64);
3396 break;
3397 case glslang::EbtFloat16:
3398 spvType = builder.makeFloatType(16);
3399 break;
3400 case glslang::EbtInt8:
3401 spvType = builder.makeIntType(8);
3402 break;
3403 case glslang::EbtUint8:
3404 spvType = builder.makeUintType(8);
3405 break;
3406 case glslang::EbtInt16:
3407 spvType = builder.makeIntType(16);
3408 break;
3409 case glslang::EbtUint16:
3410 spvType = builder.makeUintType(16);
3411 break;
3412 case glslang::EbtInt64:
3413 spvType = builder.makeIntType(64);
3414 break;
3415 case glslang::EbtUint64:
3416 spvType = builder.makeUintType(64);
3417 break;
3418 case glslang::EbtAtomicUint:
3419 builder.addCapability(spv::CapabilityAtomicStorage);
3420 spvType = builder.makeUintType(32);
3421 break;
3422 case glslang::EbtAccStructNV:
3423 spvType = builder.makeAccelerationStructureNVType();
3424 break;
3425 case glslang::EbtReference:
3426 {
3427 // Make the forward pointer, then recurse to convert the structure type, then
3428 // patch up the forward pointer with a real pointer type.
3429 if (forwardPointers.find(type.getReferentType()) == forwardPointers.end()) {
3430 spv::Id forwardId = builder.makeForwardPointer(spv::StorageClassPhysicalStorageBufferEXT);
3431 forwardPointers[type.getReferentType()] = forwardId;
3432 }
3433 spvType = forwardPointers[type.getReferentType()];
3434 if (!forwardReferenceOnly) {
3435 spv::Id referentType = convertGlslangToSpvType(*type.getReferentType());
3436 builder.makePointerFromForwardPointer(spv::StorageClassPhysicalStorageBufferEXT,
3437 forwardPointers[type.getReferentType()],
3438 referentType);
3439 }
3440 }
3441 break;
3442 #endif
3443 case glslang::EbtSampler:
3444 {
3445 const glslang::TSampler& sampler = type.getSampler();
3446 if (sampler.isPureSampler()) {
3447 spvType = builder.makeSamplerType();
3448 } else {
3449 // an image is present, make its type
3450 spvType = builder.makeImageType(getSampledType(sampler), TranslateDimensionality(sampler),
3451 sampler.isShadow(), sampler.isArrayed(), sampler.isMultiSample(),
3452 sampler.isImageClass() ? 2 : 1, TranslateImageFormat(type));
3453 if (sampler.isCombined()) {
3454 // already has both image and sampler, make the combined type
3455 spvType = builder.makeSampledImageType(spvType);
3456 }
3457 }
3458 }
3459 break;
3460 case glslang::EbtStruct:
3461 case glslang::EbtBlock:
3462 {
3463 // If we've seen this struct type, return it
3464 const glslang::TTypeList* glslangMembers = type.getStruct();
3465
3466 // Try to share structs for different layouts, but not yet for other
3467 // kinds of qualification (primarily not yet including interpolant qualification).
3468 if (! HasNonLayoutQualifiers(type, qualifier))
3469 spvType = structMap[explicitLayout][qualifier.layoutMatrix][glslangMembers];
3470 if (spvType != spv::NoResult)
3471 break;
3472
3473 // else, we haven't seen it...
3474 if (type.getBasicType() == glslang::EbtBlock)
3475 memberRemapper[glslangMembers].resize(glslangMembers->size());
3476 spvType = convertGlslangStructToSpvType(type, glslangMembers, explicitLayout, qualifier);
3477 }
3478 break;
3479 default:
3480 assert(0);
3481 break;
3482 }
3483
3484 if (type.isMatrix())
3485 spvType = builder.makeMatrixType(spvType, type.getMatrixCols(), type.getMatrixRows());
3486 else {
3487 // If this variable has a vector element count greater than 1, create a SPIR-V vector
3488 if (type.getVectorSize() > 1)
3489 spvType = builder.makeVectorType(spvType, type.getVectorSize());
3490 }
3491
3492 if (type.isCoopMat()) {
3493 builder.addCapability(spv::CapabilityCooperativeMatrixNV);
3494 builder.addExtension(spv::E_SPV_NV_cooperative_matrix);
3495 if (type.getBasicType() == glslang::EbtFloat16)
3496 builder.addCapability(spv::CapabilityFloat16);
3497 if (type.getBasicType() == glslang::EbtUint8 ||
3498 type.getBasicType() == glslang::EbtInt8) {
3499 builder.addCapability(spv::CapabilityInt8);
3500 }
3501
3502 spv::Id scope = makeArraySizeId(*type.getTypeParameters(), 1);
3503 spv::Id rows = makeArraySizeId(*type.getTypeParameters(), 2);
3504 spv::Id cols = makeArraySizeId(*type.getTypeParameters(), 3);
3505
3506 spvType = builder.makeCooperativeMatrixType(spvType, scope, rows, cols);
3507 }
3508
3509 if (type.isArray()) {
3510 int stride = 0; // keep this 0 unless doing an explicit layout; 0 will mean no decoration, no stride
3511
3512 // Do all but the outer dimension
3513 if (type.getArraySizes()->getNumDims() > 1) {
3514 // We need to decorate array strides for types needing explicit layout, except blocks.
3515 if (explicitLayout != glslang::ElpNone && type.getBasicType() != glslang::EbtBlock) {
3516 // Use a dummy glslang type for querying internal strides of
3517 // arrays of arrays, but using just a one-dimensional array.
3518 glslang::TType simpleArrayType(type, 0); // deference type of the array
3519 while (simpleArrayType.getArraySizes()->getNumDims() > 1)
3520 simpleArrayType.getArraySizes()->dereference();
3521
3522 // Will compute the higher-order strides here, rather than making a whole
3523 // pile of types and doing repetitive recursion on their contents.
3524 stride = getArrayStride(simpleArrayType, explicitLayout, qualifier.layoutMatrix);
3525 }
3526
3527 // make the arrays
3528 for (int dim = type.getArraySizes()->getNumDims() - 1; dim > 0; --dim) {
3529 spvType = builder.makeArrayType(spvType, makeArraySizeId(*type.getArraySizes(), dim), stride);
3530 if (stride > 0)
3531 builder.addDecoration(spvType, spv::DecorationArrayStride, stride);
3532 stride *= type.getArraySizes()->getDimSize(dim);
3533 }
3534 } else {
3535 // single-dimensional array, and don't yet have stride
3536
3537 // We need to decorate array strides for types needing explicit layout, except blocks.
3538 if (explicitLayout != glslang::ElpNone && type.getBasicType() != glslang::EbtBlock)
3539 stride = getArrayStride(type, explicitLayout, qualifier.layoutMatrix);
3540 }
3541
3542 // Do the outer dimension, which might not be known for a runtime-sized array.
3543 // (Unsized arrays that survive through linking will be runtime-sized arrays)
3544 if (type.isSizedArray())
3545 spvType = builder.makeArrayType(spvType, makeArraySizeId(*type.getArraySizes(), 0), stride);
3546 else {
3547 #ifndef GLSLANG_WEB
3548 if (!lastBufferBlockMember) {
3549 builder.addIncorporatedExtension("SPV_EXT_descriptor_indexing", spv::Spv_1_5);
3550 builder.addCapability(spv::CapabilityRuntimeDescriptorArrayEXT);
3551 }
3552 #endif
3553 spvType = builder.makeRuntimeArray(spvType);
3554 }
3555 if (stride > 0)
3556 builder.addDecoration(spvType, spv::DecorationArrayStride, stride);
3557 }
3558
3559 return spvType;
3560 }
3561
3562 // TODO: this functionality should exist at a higher level, in creating the AST
3563 //
3564 // Identify interface members that don't have their required extension turned on.
3565 //
filterMember(const glslang::TType & member)3566 bool TGlslangToSpvTraverser::filterMember(const glslang::TType& member)
3567 {
3568 #ifndef GLSLANG_WEB
3569 auto& extensions = glslangIntermediate->getRequestedExtensions();
3570
3571 if (member.getFieldName() == "gl_SecondaryViewportMaskNV" &&
3572 extensions.find("GL_NV_stereo_view_rendering") == extensions.end())
3573 return true;
3574 if (member.getFieldName() == "gl_SecondaryPositionNV" &&
3575 extensions.find("GL_NV_stereo_view_rendering") == extensions.end())
3576 return true;
3577
3578 if (glslangIntermediate->getStage() != EShLangMeshNV) {
3579 if (member.getFieldName() == "gl_ViewportMask" &&
3580 extensions.find("GL_NV_viewport_array2") == extensions.end())
3581 return true;
3582 if (member.getFieldName() == "gl_PositionPerViewNV" &&
3583 extensions.find("GL_NVX_multiview_per_view_attributes") == extensions.end())
3584 return true;
3585 if (member.getFieldName() == "gl_ViewportMaskPerViewNV" &&
3586 extensions.find("GL_NVX_multiview_per_view_attributes") == extensions.end())
3587 return true;
3588 }
3589 #endif
3590
3591 return false;
3592 };
3593
3594 // Do full recursive conversion of a glslang structure (or block) type to a SPIR-V Id.
3595 // explicitLayout can be kept the same throughout the hierarchical recursive walk.
3596 // Mutually recursive with convertGlslangToSpvType().
convertGlslangStructToSpvType(const glslang::TType & type,const glslang::TTypeList * glslangMembers,glslang::TLayoutPacking explicitLayout,const glslang::TQualifier & qualifier)3597 spv::Id TGlslangToSpvTraverser::convertGlslangStructToSpvType(const glslang::TType& type,
3598 const glslang::TTypeList* glslangMembers,
3599 glslang::TLayoutPacking explicitLayout,
3600 const glslang::TQualifier& qualifier)
3601 {
3602 // Create a vector of struct types for SPIR-V to consume
3603 std::vector<spv::Id> spvMembers;
3604 int memberDelta = 0; // how much the member's index changes from glslang to SPIR-V, normally 0, except sometimes for blocks
3605 std::vector<std::pair<glslang::TType*, glslang::TQualifier> > deferredForwardPointers;
3606 for (int i = 0; i < (int)glslangMembers->size(); i++) {
3607 glslang::TType& glslangMember = *(*glslangMembers)[i].type;
3608 if (glslangMember.hiddenMember()) {
3609 ++memberDelta;
3610 if (type.getBasicType() == glslang::EbtBlock)
3611 memberRemapper[glslangMembers][i] = -1;
3612 } else {
3613 if (type.getBasicType() == glslang::EbtBlock) {
3614 if (filterMember(glslangMember)) {
3615 memberDelta++;
3616 memberRemapper[glslangMembers][i] = -1;
3617 continue;
3618 }
3619 memberRemapper[glslangMembers][i] = i - memberDelta;
3620 }
3621 // modify just this child's view of the qualifier
3622 glslang::TQualifier memberQualifier = glslangMember.getQualifier();
3623 InheritQualifiers(memberQualifier, qualifier);
3624
3625 // manually inherit location
3626 if (! memberQualifier.hasLocation() && qualifier.hasLocation())
3627 memberQualifier.layoutLocation = qualifier.layoutLocation;
3628
3629 // recurse
3630 bool lastBufferBlockMember = qualifier.storage == glslang::EvqBuffer &&
3631 i == (int)glslangMembers->size() - 1;
3632
3633 // Make forward pointers for any pointer members, and create a list of members to
3634 // convert to spirv types after creating the struct.
3635 if (glslangMember.isReference()) {
3636 if (forwardPointers.find(glslangMember.getReferentType()) == forwardPointers.end()) {
3637 deferredForwardPointers.push_back(std::make_pair(&glslangMember, memberQualifier));
3638 }
3639 spvMembers.push_back(
3640 convertGlslangToSpvType(glslangMember, explicitLayout, memberQualifier, lastBufferBlockMember, true));
3641 } else {
3642 spvMembers.push_back(
3643 convertGlslangToSpvType(glslangMember, explicitLayout, memberQualifier, lastBufferBlockMember, false));
3644 }
3645 }
3646 }
3647
3648 // Make the SPIR-V type
3649 spv::Id spvType = builder.makeStructType(spvMembers, type.getTypeName().c_str());
3650 if (! HasNonLayoutQualifiers(type, qualifier))
3651 structMap[explicitLayout][qualifier.layoutMatrix][glslangMembers] = spvType;
3652
3653 // Decorate it
3654 decorateStructType(type, glslangMembers, explicitLayout, qualifier, spvType);
3655
3656 for (int i = 0; i < (int)deferredForwardPointers.size(); ++i) {
3657 auto it = deferredForwardPointers[i];
3658 convertGlslangToSpvType(*it.first, explicitLayout, it.second, false);
3659 }
3660
3661 return spvType;
3662 }
3663
decorateStructType(const glslang::TType & type,const glslang::TTypeList * glslangMembers,glslang::TLayoutPacking explicitLayout,const glslang::TQualifier & qualifier,spv::Id spvType)3664 void TGlslangToSpvTraverser::decorateStructType(const glslang::TType& type,
3665 const glslang::TTypeList* glslangMembers,
3666 glslang::TLayoutPacking explicitLayout,
3667 const glslang::TQualifier& qualifier,
3668 spv::Id spvType)
3669 {
3670 // Name and decorate the non-hidden members
3671 int offset = -1;
3672 int locationOffset = 0; // for use within the members of this struct
3673 for (int i = 0; i < (int)glslangMembers->size(); i++) {
3674 glslang::TType& glslangMember = *(*glslangMembers)[i].type;
3675 int member = i;
3676 if (type.getBasicType() == glslang::EbtBlock) {
3677 member = memberRemapper[glslangMembers][i];
3678 if (filterMember(glslangMember))
3679 continue;
3680 }
3681
3682 // modify just this child's view of the qualifier
3683 glslang::TQualifier memberQualifier = glslangMember.getQualifier();
3684 InheritQualifiers(memberQualifier, qualifier);
3685
3686 // using -1 above to indicate a hidden member
3687 if (member < 0)
3688 continue;
3689
3690 builder.addMemberName(spvType, member, glslangMember.getFieldName().c_str());
3691 builder.addMemberDecoration(spvType, member,
3692 TranslateLayoutDecoration(glslangMember, memberQualifier.layoutMatrix));
3693 builder.addMemberDecoration(spvType, member, TranslatePrecisionDecoration(glslangMember));
3694 // Add interpolation and auxiliary storage decorations only to
3695 // top-level members of Input and Output storage classes
3696 if (type.getQualifier().storage == glslang::EvqVaryingIn ||
3697 type.getQualifier().storage == glslang::EvqVaryingOut) {
3698 if (type.getBasicType() == glslang::EbtBlock ||
3699 glslangIntermediate->getSource() == glslang::EShSourceHlsl) {
3700 builder.addMemberDecoration(spvType, member, TranslateInterpolationDecoration(memberQualifier));
3701 builder.addMemberDecoration(spvType, member, TranslateAuxiliaryStorageDecoration(memberQualifier));
3702 #ifndef GLSLANG_WEB
3703 addMeshNVDecoration(spvType, member, memberQualifier);
3704 #endif
3705 }
3706 }
3707 builder.addMemberDecoration(spvType, member, TranslateInvariantDecoration(memberQualifier));
3708
3709 #ifndef GLSLANG_WEB
3710 if (type.getBasicType() == glslang::EbtBlock &&
3711 qualifier.storage == glslang::EvqBuffer) {
3712 // Add memory decorations only to top-level members of shader storage block
3713 std::vector<spv::Decoration> memory;
3714 TranslateMemoryDecoration(memberQualifier, memory, glslangIntermediate->usingVulkanMemoryModel());
3715 for (unsigned int i = 0; i < memory.size(); ++i)
3716 builder.addMemberDecoration(spvType, member, memory[i]);
3717 }
3718
3719 #endif
3720
3721 // Location assignment was already completed correctly by the front end,
3722 // just track whether a member needs to be decorated.
3723 // Ignore member locations if the container is an array, as that's
3724 // ill-specified and decisions have been made to not allow this.
3725 if (! type.isArray() && memberQualifier.hasLocation())
3726 builder.addMemberDecoration(spvType, member, spv::DecorationLocation, memberQualifier.layoutLocation);
3727
3728 if (qualifier.hasLocation()) // track for upcoming inheritance
3729 locationOffset += glslangIntermediate->computeTypeLocationSize(
3730 glslangMember, glslangIntermediate->getStage());
3731
3732 // component, XFB, others
3733 if (glslangMember.getQualifier().hasComponent())
3734 builder.addMemberDecoration(spvType, member, spv::DecorationComponent,
3735 glslangMember.getQualifier().layoutComponent);
3736 if (glslangMember.getQualifier().hasXfbOffset())
3737 builder.addMemberDecoration(spvType, member, spv::DecorationOffset,
3738 glslangMember.getQualifier().layoutXfbOffset);
3739 else if (explicitLayout != glslang::ElpNone) {
3740 // figure out what to do with offset, which is accumulating
3741 int nextOffset;
3742 updateMemberOffset(type, glslangMember, offset, nextOffset, explicitLayout, memberQualifier.layoutMatrix);
3743 if (offset >= 0)
3744 builder.addMemberDecoration(spvType, member, spv::DecorationOffset, offset);
3745 offset = nextOffset;
3746 }
3747
3748 if (glslangMember.isMatrix() && explicitLayout != glslang::ElpNone)
3749 builder.addMemberDecoration(spvType, member, spv::DecorationMatrixStride,
3750 getMatrixStride(glslangMember, explicitLayout, memberQualifier.layoutMatrix));
3751
3752 // built-in variable decorations
3753 spv::BuiltIn builtIn = TranslateBuiltInDecoration(glslangMember.getQualifier().builtIn, true);
3754 if (builtIn != spv::BuiltInMax)
3755 builder.addMemberDecoration(spvType, member, spv::DecorationBuiltIn, (int)builtIn);
3756
3757 #ifndef GLSLANG_WEB
3758 // nonuniform
3759 builder.addMemberDecoration(spvType, member, TranslateNonUniformDecoration(glslangMember.getQualifier()));
3760
3761 if (glslangIntermediate->getHlslFunctionality1() && memberQualifier.semanticName != nullptr) {
3762 builder.addExtension("SPV_GOOGLE_hlsl_functionality1");
3763 builder.addMemberDecoration(spvType, member, (spv::Decoration)spv::DecorationHlslSemanticGOOGLE,
3764 memberQualifier.semanticName);
3765 }
3766
3767 if (builtIn == spv::BuiltInLayer) {
3768 // SPV_NV_viewport_array2 extension
3769 if (glslangMember.getQualifier().layoutViewportRelative){
3770 builder.addMemberDecoration(spvType, member, (spv::Decoration)spv::DecorationViewportRelativeNV);
3771 builder.addCapability(spv::CapabilityShaderViewportMaskNV);
3772 builder.addExtension(spv::E_SPV_NV_viewport_array2);
3773 }
3774 if (glslangMember.getQualifier().layoutSecondaryViewportRelativeOffset != -2048){
3775 builder.addMemberDecoration(spvType, member,
3776 (spv::Decoration)spv::DecorationSecondaryViewportRelativeNV,
3777 glslangMember.getQualifier().layoutSecondaryViewportRelativeOffset);
3778 builder.addCapability(spv::CapabilityShaderStereoViewNV);
3779 builder.addExtension(spv::E_SPV_NV_stereo_view_rendering);
3780 }
3781 }
3782 if (glslangMember.getQualifier().layoutPassthrough) {
3783 builder.addMemberDecoration(spvType, member, (spv::Decoration)spv::DecorationPassthroughNV);
3784 builder.addCapability(spv::CapabilityGeometryShaderPassthroughNV);
3785 builder.addExtension(spv::E_SPV_NV_geometry_shader_passthrough);
3786 }
3787 #endif
3788 }
3789
3790 // Decorate the structure
3791 builder.addDecoration(spvType, TranslateLayoutDecoration(type, qualifier.layoutMatrix));
3792 builder.addDecoration(spvType, TranslateBlockDecoration(type, glslangIntermediate->usingStorageBuffer()));
3793 }
3794
3795 // Turn the expression forming the array size into an id.
3796 // This is not quite trivial, because of specialization constants.
3797 // Sometimes, a raw constant is turned into an Id, and sometimes
3798 // a specialization constant expression is.
makeArraySizeId(const glslang::TArraySizes & arraySizes,int dim)3799 spv::Id TGlslangToSpvTraverser::makeArraySizeId(const glslang::TArraySizes& arraySizes, int dim)
3800 {
3801 // First, see if this is sized with a node, meaning a specialization constant:
3802 glslang::TIntermTyped* specNode = arraySizes.getDimNode(dim);
3803 if (specNode != nullptr) {
3804 builder.clearAccessChain();
3805 specNode->traverse(this);
3806 return accessChainLoad(specNode->getAsTyped()->getType());
3807 }
3808
3809 // Otherwise, need a compile-time (front end) size, get it:
3810 int size = arraySizes.getDimSize(dim);
3811 assert(size > 0);
3812 return builder.makeUintConstant(size);
3813 }
3814
3815 // Wrap the builder's accessChainLoad to:
3816 // - localize handling of RelaxedPrecision
3817 // - use the SPIR-V inferred type instead of another conversion of the glslang type
3818 // (avoids unnecessary work and possible type punning for structures)
3819 // - do conversion of concrete to abstract type
accessChainLoad(const glslang::TType & type)3820 spv::Id TGlslangToSpvTraverser::accessChainLoad(const glslang::TType& type)
3821 {
3822 spv::Id nominalTypeId = builder.accessChainGetInferredType();
3823
3824 spv::Builder::AccessChain::CoherentFlags coherentFlags = builder.getAccessChain().coherentFlags;
3825 coherentFlags |= TranslateCoherent(type);
3826
3827 unsigned int alignment = builder.getAccessChain().alignment;
3828 alignment |= type.getBufferReferenceAlignment();
3829
3830 spv::Id loadedId = builder.accessChainLoad(TranslatePrecisionDecoration(type),
3831 TranslateNonUniformDecoration(type.getQualifier()),
3832 nominalTypeId,
3833 spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & ~spv::MemoryAccessMakePointerAvailableKHRMask),
3834 TranslateMemoryScope(coherentFlags),
3835 alignment);
3836
3837 // Need to convert to abstract types when necessary
3838 if (type.getBasicType() == glslang::EbtBool) {
3839 if (builder.isScalarType(nominalTypeId)) {
3840 // Conversion for bool
3841 spv::Id boolType = builder.makeBoolType();
3842 if (nominalTypeId != boolType)
3843 loadedId = builder.createBinOp(spv::OpINotEqual, boolType, loadedId, builder.makeUintConstant(0));
3844 } else if (builder.isVectorType(nominalTypeId)) {
3845 // Conversion for bvec
3846 int vecSize = builder.getNumTypeComponents(nominalTypeId);
3847 spv::Id bvecType = builder.makeVectorType(builder.makeBoolType(), vecSize);
3848 if (nominalTypeId != bvecType)
3849 loadedId = builder.createBinOp(spv::OpINotEqual, bvecType, loadedId, makeSmearedConstant(builder.makeUintConstant(0), vecSize));
3850 }
3851 }
3852
3853 return loadedId;
3854 }
3855
3856 // Wrap the builder's accessChainStore to:
3857 // - do conversion of concrete to abstract type
3858 //
3859 // Implicitly uses the existing builder.accessChain as the storage target.
accessChainStore(const glslang::TType & type,spv::Id rvalue)3860 void TGlslangToSpvTraverser::accessChainStore(const glslang::TType& type, spv::Id rvalue)
3861 {
3862 // Need to convert to abstract types when necessary
3863 if (type.getBasicType() == glslang::EbtBool) {
3864 spv::Id nominalTypeId = builder.accessChainGetInferredType();
3865
3866 if (builder.isScalarType(nominalTypeId)) {
3867 // Conversion for bool
3868 spv::Id boolType = builder.makeBoolType();
3869 if (nominalTypeId != boolType) {
3870 // keep these outside arguments, for determinant order-of-evaluation
3871 spv::Id one = builder.makeUintConstant(1);
3872 spv::Id zero = builder.makeUintConstant(0);
3873 rvalue = builder.createTriOp(spv::OpSelect, nominalTypeId, rvalue, one, zero);
3874 } else if (builder.getTypeId(rvalue) != boolType)
3875 rvalue = builder.createBinOp(spv::OpINotEqual, boolType, rvalue, builder.makeUintConstant(0));
3876 } else if (builder.isVectorType(nominalTypeId)) {
3877 // Conversion for bvec
3878 int vecSize = builder.getNumTypeComponents(nominalTypeId);
3879 spv::Id bvecType = builder.makeVectorType(builder.makeBoolType(), vecSize);
3880 if (nominalTypeId != bvecType) {
3881 // keep these outside arguments, for determinant order-of-evaluation
3882 spv::Id one = makeSmearedConstant(builder.makeUintConstant(1), vecSize);
3883 spv::Id zero = makeSmearedConstant(builder.makeUintConstant(0), vecSize);
3884 rvalue = builder.createTriOp(spv::OpSelect, nominalTypeId, rvalue, one, zero);
3885 } else if (builder.getTypeId(rvalue) != bvecType)
3886 rvalue = builder.createBinOp(spv::OpINotEqual, bvecType, rvalue,
3887 makeSmearedConstant(builder.makeUintConstant(0), vecSize));
3888 }
3889 }
3890
3891 spv::Builder::AccessChain::CoherentFlags coherentFlags = builder.getAccessChain().coherentFlags;
3892 coherentFlags |= TranslateCoherent(type);
3893
3894 unsigned int alignment = builder.getAccessChain().alignment;
3895 alignment |= type.getBufferReferenceAlignment();
3896
3897 builder.accessChainStore(rvalue,
3898 spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & ~spv::MemoryAccessMakePointerVisibleKHRMask),
3899 TranslateMemoryScope(coherentFlags), alignment);
3900 }
3901
3902 // For storing when types match at the glslang level, but not might match at the
3903 // SPIR-V level.
3904 //
3905 // This especially happens when a single glslang type expands to multiple
3906 // SPIR-V types, like a struct that is used in a member-undecorated way as well
3907 // as in a member-decorated way.
3908 //
3909 // NOTE: This function can handle any store request; if it's not special it
3910 // simplifies to a simple OpStore.
3911 //
3912 // Implicitly uses the existing builder.accessChain as the storage target.
multiTypeStore(const glslang::TType & type,spv::Id rValue)3913 void TGlslangToSpvTraverser::multiTypeStore(const glslang::TType& type, spv::Id rValue)
3914 {
3915 // we only do the complex path here if it's an aggregate
3916 if (! type.isStruct() && ! type.isArray()) {
3917 accessChainStore(type, rValue);
3918 return;
3919 }
3920
3921 // and, it has to be a case of type aliasing
3922 spv::Id rType = builder.getTypeId(rValue);
3923 spv::Id lValue = builder.accessChainGetLValue();
3924 spv::Id lType = builder.getContainedTypeId(builder.getTypeId(lValue));
3925 if (lType == rType) {
3926 accessChainStore(type, rValue);
3927 return;
3928 }
3929
3930 // Recursively (as needed) copy an aggregate type to a different aggregate type,
3931 // where the two types were the same type in GLSL. This requires member
3932 // by member copy, recursively.
3933
3934 // SPIR-V 1.4 added an instruction to do help do this.
3935 if (glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4) {
3936 // However, bool in uniform space is changed to int, so
3937 // OpCopyLogical does not work for that.
3938 // TODO: It would be more robust to do a full recursive verification of the types satisfying SPIR-V rules.
3939 bool rBool = builder.containsType(builder.getTypeId(rValue), spv::OpTypeBool, 0);
3940 bool lBool = builder.containsType(lType, spv::OpTypeBool, 0);
3941 if (lBool == rBool) {
3942 spv::Id logicalCopy = builder.createUnaryOp(spv::OpCopyLogical, lType, rValue);
3943 accessChainStore(type, logicalCopy);
3944 return;
3945 }
3946 }
3947
3948 // If an array, copy element by element.
3949 if (type.isArray()) {
3950 glslang::TType glslangElementType(type, 0);
3951 spv::Id elementRType = builder.getContainedTypeId(rType);
3952 for (int index = 0; index < type.getOuterArraySize(); ++index) {
3953 // get the source member
3954 spv::Id elementRValue = builder.createCompositeExtract(rValue, elementRType, index);
3955
3956 // set up the target storage
3957 builder.clearAccessChain();
3958 builder.setAccessChainLValue(lValue);
3959 builder.accessChainPush(builder.makeIntConstant(index), TranslateCoherent(type), type.getBufferReferenceAlignment());
3960
3961 // store the member
3962 multiTypeStore(glslangElementType, elementRValue);
3963 }
3964 } else {
3965 assert(type.isStruct());
3966
3967 // loop over structure members
3968 const glslang::TTypeList& members = *type.getStruct();
3969 for (int m = 0; m < (int)members.size(); ++m) {
3970 const glslang::TType& glslangMemberType = *members[m].type;
3971
3972 // get the source member
3973 spv::Id memberRType = builder.getContainedTypeId(rType, m);
3974 spv::Id memberRValue = builder.createCompositeExtract(rValue, memberRType, m);
3975
3976 // set up the target storage
3977 builder.clearAccessChain();
3978 builder.setAccessChainLValue(lValue);
3979 builder.accessChainPush(builder.makeIntConstant(m), TranslateCoherent(type), type.getBufferReferenceAlignment());
3980
3981 // store the member
3982 multiTypeStore(glslangMemberType, memberRValue);
3983 }
3984 }
3985 }
3986
3987 // Decide whether or not this type should be
3988 // decorated with offsets and strides, and if so
3989 // whether std140 or std430 rules should be applied.
getExplicitLayout(const glslang::TType & type) const3990 glslang::TLayoutPacking TGlslangToSpvTraverser::getExplicitLayout(const glslang::TType& type) const
3991 {
3992 // has to be a block
3993 if (type.getBasicType() != glslang::EbtBlock)
3994 return glslang::ElpNone;
3995
3996 // has to be a uniform or buffer block or task in/out blocks
3997 if (type.getQualifier().storage != glslang::EvqUniform &&
3998 type.getQualifier().storage != glslang::EvqBuffer &&
3999 !type.getQualifier().isTaskMemory())
4000 return glslang::ElpNone;
4001
4002 // return the layout to use
4003 switch (type.getQualifier().layoutPacking) {
4004 case glslang::ElpStd140:
4005 case glslang::ElpStd430:
4006 case glslang::ElpScalar:
4007 return type.getQualifier().layoutPacking;
4008 default:
4009 return glslang::ElpNone;
4010 }
4011 }
4012
4013 // Given an array type, returns the integer stride required for that array
getArrayStride(const glslang::TType & arrayType,glslang::TLayoutPacking explicitLayout,glslang::TLayoutMatrix matrixLayout)4014 int TGlslangToSpvTraverser::getArrayStride(const glslang::TType& arrayType, glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout)
4015 {
4016 int size;
4017 int stride;
4018 glslangIntermediate->getMemberAlignment(arrayType, size, stride, explicitLayout, matrixLayout == glslang::ElmRowMajor);
4019
4020 return stride;
4021 }
4022
4023 // Given a matrix type, or array (of array) of matrixes type, returns the integer stride required for that matrix
4024 // when used as a member of an interface block
getMatrixStride(const glslang::TType & matrixType,glslang::TLayoutPacking explicitLayout,glslang::TLayoutMatrix matrixLayout)4025 int TGlslangToSpvTraverser::getMatrixStride(const glslang::TType& matrixType, glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout)
4026 {
4027 glslang::TType elementType;
4028 elementType.shallowCopy(matrixType);
4029 elementType.clearArraySizes();
4030
4031 int size;
4032 int stride;
4033 glslangIntermediate->getMemberAlignment(elementType, size, stride, explicitLayout, matrixLayout == glslang::ElmRowMajor);
4034
4035 return stride;
4036 }
4037
4038 // Given a member type of a struct, realign the current offset for it, and compute
4039 // the next (not yet aligned) offset for the next member, which will get aligned
4040 // on the next call.
4041 // 'currentOffset' should be passed in already initialized, ready to modify, and reflecting
4042 // the migration of data from nextOffset -> currentOffset. It should be -1 on the first call.
4043 // -1 means a non-forced member offset (no decoration needed).
updateMemberOffset(const glslang::TType & structType,const glslang::TType & memberType,int & currentOffset,int & nextOffset,glslang::TLayoutPacking explicitLayout,glslang::TLayoutMatrix matrixLayout)4044 void TGlslangToSpvTraverser::updateMemberOffset(const glslang::TType& structType, const glslang::TType& memberType, int& currentOffset, int& nextOffset,
4045 glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout)
4046 {
4047 // this will get a positive value when deemed necessary
4048 nextOffset = -1;
4049
4050 // override anything in currentOffset with user-set offset
4051 if (memberType.getQualifier().hasOffset())
4052 currentOffset = memberType.getQualifier().layoutOffset;
4053
4054 // It could be that current linker usage in glslang updated all the layoutOffset,
4055 // in which case the following code does not matter. But, that's not quite right
4056 // once cross-compilation unit GLSL validation is done, as the original user
4057 // settings are needed in layoutOffset, and then the following will come into play.
4058
4059 if (explicitLayout == glslang::ElpNone) {
4060 if (! memberType.getQualifier().hasOffset())
4061 currentOffset = -1;
4062
4063 return;
4064 }
4065
4066 // Getting this far means we need explicit offsets
4067 if (currentOffset < 0)
4068 currentOffset = 0;
4069
4070 // Now, currentOffset is valid (either 0, or from a previous nextOffset),
4071 // but possibly not yet correctly aligned.
4072
4073 int memberSize;
4074 int dummyStride;
4075 int memberAlignment = glslangIntermediate->getMemberAlignment(memberType, memberSize, dummyStride, explicitLayout, matrixLayout == glslang::ElmRowMajor);
4076
4077 // Adjust alignment for HLSL rules
4078 // TODO: make this consistent in early phases of code:
4079 // adjusting this late means inconsistencies with earlier code, which for reflection is an issue
4080 // Until reflection is brought in sync with these adjustments, don't apply to $Global,
4081 // which is the most likely to rely on reflection, and least likely to rely implicit layouts
4082 if (glslangIntermediate->usingHlslOffsets() &&
4083 ! memberType.isArray() && memberType.isVector() && structType.getTypeName().compare("$Global") != 0) {
4084 int dummySize;
4085 int componentAlignment = glslangIntermediate->getBaseAlignmentScalar(memberType, dummySize);
4086 if (componentAlignment <= 4)
4087 memberAlignment = componentAlignment;
4088 }
4089
4090 // Bump up to member alignment
4091 glslang::RoundToPow2(currentOffset, memberAlignment);
4092
4093 // Bump up to vec4 if there is a bad straddle
4094 if (explicitLayout != glslang::ElpScalar && glslangIntermediate->improperStraddle(memberType, memberSize, currentOffset))
4095 glslang::RoundToPow2(currentOffset, 16);
4096
4097 nextOffset = currentOffset + memberSize;
4098 }
4099
declareUseOfStructMember(const glslang::TTypeList & members,int glslangMember)4100 void TGlslangToSpvTraverser::declareUseOfStructMember(const glslang::TTypeList& members, int glslangMember)
4101 {
4102 const glslang::TBuiltInVariable glslangBuiltIn = members[glslangMember].type->getQualifier().builtIn;
4103 switch (glslangBuiltIn)
4104 {
4105 case glslang::EbvPointSize:
4106 #ifndef GLSLANG_WEB
4107 case glslang::EbvClipDistance:
4108 case glslang::EbvCullDistance:
4109 case glslang::EbvViewportMaskNV:
4110 case glslang::EbvSecondaryPositionNV:
4111 case glslang::EbvSecondaryViewportMaskNV:
4112 case glslang::EbvPositionPerViewNV:
4113 case glslang::EbvViewportMaskPerViewNV:
4114 case glslang::EbvTaskCountNV:
4115 case glslang::EbvPrimitiveCountNV:
4116 case glslang::EbvPrimitiveIndicesNV:
4117 case glslang::EbvClipDistancePerViewNV:
4118 case glslang::EbvCullDistancePerViewNV:
4119 case glslang::EbvLayerPerViewNV:
4120 case glslang::EbvMeshViewCountNV:
4121 case glslang::EbvMeshViewIndicesNV:
4122 #endif
4123 // Generate the associated capability. Delegate to TranslateBuiltInDecoration.
4124 // Alternately, we could just call this for any glslang built-in, since the
4125 // capability already guards against duplicates.
4126 TranslateBuiltInDecoration(glslangBuiltIn, false);
4127 break;
4128 default:
4129 // Capabilities were already generated when the struct was declared.
4130 break;
4131 }
4132 }
4133
isShaderEntryPoint(const glslang::TIntermAggregate * node)4134 bool TGlslangToSpvTraverser::isShaderEntryPoint(const glslang::TIntermAggregate* node)
4135 {
4136 return node->getName().compare(glslangIntermediate->getEntryPointMangledName().c_str()) == 0;
4137 }
4138
4139 // Does parameter need a place to keep writes, separate from the original?
4140 // Assumes called after originalParam(), which filters out block/buffer/opaque-based
4141 // qualifiers such that we should have only in/out/inout/constreadonly here.
writableParam(glslang::TStorageQualifier qualifier) const4142 bool TGlslangToSpvTraverser::writableParam(glslang::TStorageQualifier qualifier) const
4143 {
4144 assert(qualifier == glslang::EvqIn ||
4145 qualifier == glslang::EvqOut ||
4146 qualifier == glslang::EvqInOut ||
4147 qualifier == glslang::EvqConstReadOnly);
4148 return qualifier != glslang::EvqConstReadOnly;
4149 }
4150
4151 // Is parameter pass-by-original?
originalParam(glslang::TStorageQualifier qualifier,const glslang::TType & paramType,bool implicitThisParam)4152 bool TGlslangToSpvTraverser::originalParam(glslang::TStorageQualifier qualifier, const glslang::TType& paramType,
4153 bool implicitThisParam)
4154 {
4155 if (implicitThisParam) // implicit this
4156 return true;
4157 if (glslangIntermediate->getSource() == glslang::EShSourceHlsl)
4158 return paramType.getBasicType() == glslang::EbtBlock;
4159 return paramType.containsOpaque() || // sampler, etc.
4160 (paramType.getBasicType() == glslang::EbtBlock && qualifier == glslang::EvqBuffer); // SSBO
4161 }
4162
4163 // Make all the functions, skeletally, without actually visiting their bodies.
makeFunctions(const glslang::TIntermSequence & glslFunctions)4164 void TGlslangToSpvTraverser::makeFunctions(const glslang::TIntermSequence& glslFunctions)
4165 {
4166 const auto getParamDecorations = [&](std::vector<spv::Decoration>& decorations, const glslang::TType& type, bool useVulkanMemoryModel) {
4167 spv::Decoration paramPrecision = TranslatePrecisionDecoration(type);
4168 if (paramPrecision != spv::NoPrecision)
4169 decorations.push_back(paramPrecision);
4170 TranslateMemoryDecoration(type.getQualifier(), decorations, useVulkanMemoryModel);
4171 if (type.isReference()) {
4172 // Original and non-writable params pass the pointer directly and
4173 // use restrict/aliased, others are stored to a pointer in Function
4174 // memory and use RestrictPointer/AliasedPointer.
4175 if (originalParam(type.getQualifier().storage, type, false) ||
4176 !writableParam(type.getQualifier().storage)) {
4177 decorations.push_back(type.getQualifier().isRestrict() ? spv::DecorationRestrict :
4178 spv::DecorationAliased);
4179 } else {
4180 decorations.push_back(type.getQualifier().isRestrict() ? spv::DecorationRestrictPointerEXT :
4181 spv::DecorationAliasedPointerEXT);
4182 }
4183 }
4184 };
4185
4186 for (int f = 0; f < (int)glslFunctions.size(); ++f) {
4187 glslang::TIntermAggregate* glslFunction = glslFunctions[f]->getAsAggregate();
4188 if (! glslFunction || glslFunction->getOp() != glslang::EOpFunction || isShaderEntryPoint(glslFunction))
4189 continue;
4190
4191 // We're on a user function. Set up the basic interface for the function now,
4192 // so that it's available to call. Translating the body will happen later.
4193 //
4194 // Typically (except for a "const in" parameter), an address will be passed to the
4195 // function. What it is an address of varies:
4196 //
4197 // - "in" parameters not marked as "const" can be written to without modifying the calling
4198 // argument so that write needs to be to a copy, hence the address of a copy works.
4199 //
4200 // - "const in" parameters can just be the r-value, as no writes need occur.
4201 //
4202 // - "out" and "inout" arguments can't be done as pointers to the calling argument, because
4203 // GLSL has copy-in/copy-out semantics. They can be handled though with a pointer to a copy.
4204
4205 std::vector<spv::Id> paramTypes;
4206 std::vector<std::vector<spv::Decoration>> paramDecorations; // list of decorations per parameter
4207 glslang::TIntermSequence& parameters = glslFunction->getSequence()[0]->getAsAggregate()->getSequence();
4208
4209 #ifdef ENABLE_HLSL
4210 bool implicitThis = (int)parameters.size() > 0 && parameters[0]->getAsSymbolNode()->getName() ==
4211 glslangIntermediate->implicitThisName;
4212 #else
4213 bool implicitThis = false;
4214 #endif
4215
4216 paramDecorations.resize(parameters.size());
4217 for (int p = 0; p < (int)parameters.size(); ++p) {
4218 const glslang::TType& paramType = parameters[p]->getAsTyped()->getType();
4219 spv::Id typeId = convertGlslangToSpvType(paramType);
4220 if (originalParam(paramType.getQualifier().storage, paramType, implicitThis && p == 0))
4221 typeId = builder.makePointer(TranslateStorageClass(paramType), typeId);
4222 else if (writableParam(paramType.getQualifier().storage))
4223 typeId = builder.makePointer(spv::StorageClassFunction, typeId);
4224 else
4225 rValueParameters.insert(parameters[p]->getAsSymbolNode()->getId());
4226 getParamDecorations(paramDecorations[p], paramType, glslangIntermediate->usingVulkanMemoryModel());
4227 paramTypes.push_back(typeId);
4228 }
4229
4230 spv::Block* functionBlock;
4231 spv::Function *function = builder.makeFunctionEntry(TranslatePrecisionDecoration(glslFunction->getType()),
4232 convertGlslangToSpvType(glslFunction->getType()),
4233 glslFunction->getName().c_str(), paramTypes,
4234 paramDecorations, &functionBlock);
4235 if (implicitThis)
4236 function->setImplicitThis();
4237
4238 // Track function to emit/call later
4239 functionMap[glslFunction->getName().c_str()] = function;
4240
4241 // Set the parameter id's
4242 for (int p = 0; p < (int)parameters.size(); ++p) {
4243 symbolValues[parameters[p]->getAsSymbolNode()->getId()] = function->getParamId(p);
4244 // give a name too
4245 builder.addName(function->getParamId(p), parameters[p]->getAsSymbolNode()->getName().c_str());
4246
4247 const glslang::TType& paramType = parameters[p]->getAsTyped()->getType();
4248 if (paramType.contains8BitInt())
4249 builder.addCapability(spv::CapabilityInt8);
4250 if (paramType.contains16BitInt())
4251 builder.addCapability(spv::CapabilityInt16);
4252 if (paramType.contains16BitFloat())
4253 builder.addCapability(spv::CapabilityFloat16);
4254 }
4255 }
4256 }
4257
4258 // Process all the initializers, while skipping the functions and link objects
makeGlobalInitializers(const glslang::TIntermSequence & initializers)4259 void TGlslangToSpvTraverser::makeGlobalInitializers(const glslang::TIntermSequence& initializers)
4260 {
4261 builder.setBuildPoint(shaderEntry->getLastBlock());
4262 for (int i = 0; i < (int)initializers.size(); ++i) {
4263 glslang::TIntermAggregate* initializer = initializers[i]->getAsAggregate();
4264 if (initializer && initializer->getOp() != glslang::EOpFunction && initializer->getOp() != glslang::EOpLinkerObjects) {
4265
4266 // We're on a top-level node that's not a function. Treat as an initializer, whose
4267 // code goes into the beginning of the entry point.
4268 initializer->traverse(this);
4269 }
4270 }
4271 }
4272
4273 // Process all the functions, while skipping initializers.
visitFunctions(const glslang::TIntermSequence & glslFunctions)4274 void TGlslangToSpvTraverser::visitFunctions(const glslang::TIntermSequence& glslFunctions)
4275 {
4276 for (int f = 0; f < (int)glslFunctions.size(); ++f) {
4277 glslang::TIntermAggregate* node = glslFunctions[f]->getAsAggregate();
4278 if (node && (node->getOp() == glslang::EOpFunction || node->getOp() == glslang::EOpLinkerObjects))
4279 node->traverse(this);
4280 }
4281 }
4282
handleFunctionEntry(const glslang::TIntermAggregate * node)4283 void TGlslangToSpvTraverser::handleFunctionEntry(const glslang::TIntermAggregate* node)
4284 {
4285 // SPIR-V functions should already be in the functionMap from the prepass
4286 // that called makeFunctions().
4287 currentFunction = functionMap[node->getName().c_str()];
4288 spv::Block* functionBlock = currentFunction->getEntryBlock();
4289 builder.setBuildPoint(functionBlock);
4290 }
4291
translateArguments(const glslang::TIntermAggregate & node,std::vector<spv::Id> & arguments,spv::Builder::AccessChain::CoherentFlags & lvalueCoherentFlags)4292 void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate& node, std::vector<spv::Id>& arguments, spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags)
4293 {
4294 const glslang::TIntermSequence& glslangArguments = node.getSequence();
4295
4296 glslang::TSampler sampler = {};
4297 bool cubeCompare = false;
4298 #ifndef GLSLANG_WEB
4299 bool f16ShadowCompare = false;
4300 #endif
4301 if (node.isTexture() || node.isImage()) {
4302 sampler = glslangArguments[0]->getAsTyped()->getType().getSampler();
4303 cubeCompare = sampler.dim == glslang::EsdCube && sampler.arrayed && sampler.shadow;
4304 #ifndef GLSLANG_WEB
4305 f16ShadowCompare = sampler.shadow && glslangArguments[1]->getAsTyped()->getType().getBasicType() == glslang::EbtFloat16;
4306 #endif
4307 }
4308
4309 for (int i = 0; i < (int)glslangArguments.size(); ++i) {
4310 builder.clearAccessChain();
4311 glslangArguments[i]->traverse(this);
4312
4313 #ifndef GLSLANG_WEB
4314 // Special case l-value operands
4315 bool lvalue = false;
4316 switch (node.getOp()) {
4317 case glslang::EOpImageAtomicAdd:
4318 case glslang::EOpImageAtomicMin:
4319 case glslang::EOpImageAtomicMax:
4320 case glslang::EOpImageAtomicAnd:
4321 case glslang::EOpImageAtomicOr:
4322 case glslang::EOpImageAtomicXor:
4323 case glslang::EOpImageAtomicExchange:
4324 case glslang::EOpImageAtomicCompSwap:
4325 case glslang::EOpImageAtomicLoad:
4326 case glslang::EOpImageAtomicStore:
4327 if (i == 0)
4328 lvalue = true;
4329 break;
4330 case glslang::EOpSparseImageLoad:
4331 if ((sampler.ms && i == 3) || (! sampler.ms && i == 2))
4332 lvalue = true;
4333 break;
4334 case glslang::EOpSparseTexture:
4335 if (((cubeCompare || f16ShadowCompare) && i == 3) || (! (cubeCompare || f16ShadowCompare) && i == 2))
4336 lvalue = true;
4337 break;
4338 case glslang::EOpSparseTextureClamp:
4339 if (((cubeCompare || f16ShadowCompare) && i == 4) || (! (cubeCompare || f16ShadowCompare) && i == 3))
4340 lvalue = true;
4341 break;
4342 case glslang::EOpSparseTextureLod:
4343 case glslang::EOpSparseTextureOffset:
4344 if ((f16ShadowCompare && i == 4) || (! f16ShadowCompare && i == 3))
4345 lvalue = true;
4346 break;
4347 case glslang::EOpSparseTextureFetch:
4348 if ((sampler.dim != glslang::EsdRect && i == 3) || (sampler.dim == glslang::EsdRect && i == 2))
4349 lvalue = true;
4350 break;
4351 case glslang::EOpSparseTextureFetchOffset:
4352 if ((sampler.dim != glslang::EsdRect && i == 4) || (sampler.dim == glslang::EsdRect && i == 3))
4353 lvalue = true;
4354 break;
4355 case glslang::EOpSparseTextureLodOffset:
4356 case glslang::EOpSparseTextureGrad:
4357 case glslang::EOpSparseTextureOffsetClamp:
4358 if ((f16ShadowCompare && i == 5) || (! f16ShadowCompare && i == 4))
4359 lvalue = true;
4360 break;
4361 case glslang::EOpSparseTextureGradOffset:
4362 case glslang::EOpSparseTextureGradClamp:
4363 if ((f16ShadowCompare && i == 6) || (! f16ShadowCompare && i == 5))
4364 lvalue = true;
4365 break;
4366 case glslang::EOpSparseTextureGradOffsetClamp:
4367 if ((f16ShadowCompare && i == 7) || (! f16ShadowCompare && i == 6))
4368 lvalue = true;
4369 break;
4370 case glslang::EOpSparseTextureGather:
4371 if ((sampler.shadow && i == 3) || (! sampler.shadow && i == 2))
4372 lvalue = true;
4373 break;
4374 case glslang::EOpSparseTextureGatherOffset:
4375 case glslang::EOpSparseTextureGatherOffsets:
4376 if ((sampler.shadow && i == 4) || (! sampler.shadow && i == 3))
4377 lvalue = true;
4378 break;
4379 case glslang::EOpSparseTextureGatherLod:
4380 if (i == 3)
4381 lvalue = true;
4382 break;
4383 case glslang::EOpSparseTextureGatherLodOffset:
4384 case glslang::EOpSparseTextureGatherLodOffsets:
4385 if (i == 4)
4386 lvalue = true;
4387 break;
4388 case glslang::EOpSparseImageLoadLod:
4389 if (i == 3)
4390 lvalue = true;
4391 break;
4392 case glslang::EOpImageSampleFootprintNV:
4393 if (i == 4)
4394 lvalue = true;
4395 break;
4396 case glslang::EOpImageSampleFootprintClampNV:
4397 case glslang::EOpImageSampleFootprintLodNV:
4398 if (i == 5)
4399 lvalue = true;
4400 break;
4401 case glslang::EOpImageSampleFootprintGradNV:
4402 if (i == 6)
4403 lvalue = true;
4404 break;
4405 case glslang::EOpImageSampleFootprintGradClampNV:
4406 if (i == 7)
4407 lvalue = true;
4408 break;
4409 default:
4410 break;
4411 }
4412
4413 if (lvalue) {
4414 arguments.push_back(builder.accessChainGetLValue());
4415 lvalueCoherentFlags = builder.getAccessChain().coherentFlags;
4416 lvalueCoherentFlags |= TranslateCoherent(glslangArguments[i]->getAsTyped()->getType());
4417 } else
4418 #endif
4419 arguments.push_back(accessChainLoad(glslangArguments[i]->getAsTyped()->getType()));
4420 }
4421 }
4422
translateArguments(glslang::TIntermUnary & node,std::vector<spv::Id> & arguments)4423 void TGlslangToSpvTraverser::translateArguments(glslang::TIntermUnary& node, std::vector<spv::Id>& arguments)
4424 {
4425 builder.clearAccessChain();
4426 node.getOperand()->traverse(this);
4427 arguments.push_back(accessChainLoad(node.getOperand()->getType()));
4428 }
4429
createImageTextureFunctionCall(glslang::TIntermOperator * node)4430 spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermOperator* node)
4431 {
4432 if (! node->isImage() && ! node->isTexture())
4433 return spv::NoResult;
4434
4435 builder.setLine(node->getLoc().line, node->getLoc().getFilename());
4436
4437 // Process a GLSL texturing op (will be SPV image)
4438
4439 const glslang::TType &imageType = node->getAsAggregate()
4440 ? node->getAsAggregate()->getSequence()[0]->getAsTyped()->getType()
4441 : node->getAsUnaryNode()->getOperand()->getAsTyped()->getType();
4442 const glslang::TSampler sampler = imageType.getSampler();
4443 #ifdef GLSLANG_WEB
4444 const bool f16ShadowCompare = false;
4445 #else
4446 bool f16ShadowCompare = (sampler.shadow && node->getAsAggregate())
4447 ? node->getAsAggregate()->getSequence()[1]->getAsTyped()->getType().getBasicType() == glslang::EbtFloat16
4448 : false;
4449 #endif
4450
4451 const auto signExtensionMask = [&]() {
4452 if (builder.getSpvVersion() >= spv::Spv_1_4) {
4453 if (sampler.type == glslang::EbtUint)
4454 return spv::ImageOperandsZeroExtendMask;
4455 else if (sampler.type == glslang::EbtInt)
4456 return spv::ImageOperandsSignExtendMask;
4457 }
4458 return spv::ImageOperandsMaskNone;
4459 };
4460
4461 spv::Builder::AccessChain::CoherentFlags lvalueCoherentFlags;
4462
4463 std::vector<spv::Id> arguments;
4464 if (node->getAsAggregate())
4465 translateArguments(*node->getAsAggregate(), arguments, lvalueCoherentFlags);
4466 else
4467 translateArguments(*node->getAsUnaryNode(), arguments);
4468 spv::Decoration precision = TranslatePrecisionDecoration(node->getOperationPrecision());
4469
4470 spv::Builder::TextureParameters params = { };
4471 params.sampler = arguments[0];
4472
4473 glslang::TCrackedTextureOp cracked;
4474 node->crackTexture(sampler, cracked);
4475
4476 const bool isUnsignedResult = node->getType().getBasicType() == glslang::EbtUint;
4477
4478 // Check for queries
4479 if (cracked.query) {
4480 // OpImageQueryLod works on a sampled image, for other queries the image has to be extracted first
4481 if (node->getOp() != glslang::EOpTextureQueryLod && builder.isSampledImage(params.sampler))
4482 params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler);
4483
4484 switch (node->getOp()) {
4485 case glslang::EOpImageQuerySize:
4486 case glslang::EOpTextureQuerySize:
4487 if (arguments.size() > 1) {
4488 params.lod = arguments[1];
4489 return builder.createTextureQueryCall(spv::OpImageQuerySizeLod, params, isUnsignedResult);
4490 } else
4491 return builder.createTextureQueryCall(spv::OpImageQuerySize, params, isUnsignedResult);
4492 #ifndef GLSLANG_WEB
4493 case glslang::EOpImageQuerySamples:
4494 case glslang::EOpTextureQuerySamples:
4495 return builder.createTextureQueryCall(spv::OpImageQuerySamples, params, isUnsignedResult);
4496 case glslang::EOpTextureQueryLod:
4497 params.coords = arguments[1];
4498 return builder.createTextureQueryCall(spv::OpImageQueryLod, params, isUnsignedResult);
4499 case glslang::EOpTextureQueryLevels:
4500 return builder.createTextureQueryCall(spv::OpImageQueryLevels, params, isUnsignedResult);
4501 case glslang::EOpSparseTexelsResident:
4502 return builder.createUnaryOp(spv::OpImageSparseTexelsResident, builder.makeBoolType(), arguments[0]);
4503 #endif
4504 default:
4505 assert(0);
4506 break;
4507 }
4508 }
4509
4510 int components = node->getType().getVectorSize();
4511
4512 if (node->getOp() == glslang::EOpTextureFetch) {
4513 // These must produce 4 components, per SPIR-V spec. We'll add a conversion constructor if needed.
4514 // This will only happen through the HLSL path for operator[], so we do not have to handle e.g.
4515 // the EOpTexture/Proj/Lod/etc family. It would be harmless to do so, but would need more logic
4516 // here around e.g. which ones return scalars or other types.
4517 components = 4;
4518 }
4519
4520 glslang::TType returnType(node->getType().getBasicType(), glslang::EvqTemporary, components);
4521
4522 auto resultType = [&returnType,this]{ return convertGlslangToSpvType(returnType); };
4523
4524 // Check for image functions other than queries
4525 if (node->isImage()) {
4526 std::vector<spv::IdImmediate> operands;
4527 auto opIt = arguments.begin();
4528 spv::IdImmediate image = { true, *(opIt++) };
4529 operands.push_back(image);
4530
4531 // Handle subpass operations
4532 // TODO: GLSL should change to have the "MS" only on the type rather than the
4533 // built-in function.
4534 if (cracked.subpass) {
4535 // add on the (0,0) coordinate
4536 spv::Id zero = builder.makeIntConstant(0);
4537 std::vector<spv::Id> comps;
4538 comps.push_back(zero);
4539 comps.push_back(zero);
4540 spv::IdImmediate coord = { true,
4541 builder.makeCompositeConstant(builder.makeVectorType(builder.makeIntType(32), 2), comps) };
4542 operands.push_back(coord);
4543 spv::IdImmediate imageOperands = { false, spv::ImageOperandsMaskNone };
4544 imageOperands.word = imageOperands.word | signExtensionMask();
4545 if (sampler.isMultiSample()) {
4546 imageOperands.word = imageOperands.word | spv::ImageOperandsSampleMask;
4547 }
4548 if (imageOperands.word != spv::ImageOperandsMaskNone) {
4549 operands.push_back(imageOperands);
4550 if (sampler.isMultiSample()) {
4551 spv::IdImmediate imageOperand = { true, *(opIt++) };
4552 operands.push_back(imageOperand);
4553 }
4554 }
4555 spv::Id result = builder.createOp(spv::OpImageRead, resultType(), operands);
4556 builder.setPrecision(result, precision);
4557 return result;
4558 }
4559
4560 spv::IdImmediate coord = { true, *(opIt++) };
4561 operands.push_back(coord);
4562 if (node->getOp() == glslang::EOpImageLoad || node->getOp() == glslang::EOpImageLoadLod) {
4563 spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone;
4564 if (sampler.isMultiSample()) {
4565 mask = mask | spv::ImageOperandsSampleMask;
4566 }
4567 if (cracked.lod) {
4568 builder.addExtension(spv::E_SPV_AMD_shader_image_load_store_lod);
4569 builder.addCapability(spv::CapabilityImageReadWriteLodAMD);
4570 mask = mask | spv::ImageOperandsLodMask;
4571 }
4572 mask = mask | TranslateImageOperands(TranslateCoherent(imageType));
4573 mask = (spv::ImageOperandsMask)(mask & ~spv::ImageOperandsMakeTexelAvailableKHRMask);
4574 mask = mask | signExtensionMask();
4575 if (mask != spv::ImageOperandsMaskNone) {
4576 spv::IdImmediate imageOperands = { false, (unsigned int)mask };
4577 operands.push_back(imageOperands);
4578 }
4579 if (mask & spv::ImageOperandsSampleMask) {
4580 spv::IdImmediate imageOperand = { true, *opIt++ };
4581 operands.push_back(imageOperand);
4582 }
4583 if (mask & spv::ImageOperandsLodMask) {
4584 spv::IdImmediate imageOperand = { true, *opIt++ };
4585 operands.push_back(imageOperand);
4586 }
4587 if (mask & spv::ImageOperandsMakeTexelVisibleKHRMask) {
4588 spv::IdImmediate imageOperand = { true,
4589 builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) };
4590 operands.push_back(imageOperand);
4591 }
4592
4593 if (builder.getImageTypeFormat(builder.getImageType(operands.front().word)) == spv::ImageFormatUnknown)
4594 builder.addCapability(spv::CapabilityStorageImageReadWithoutFormat);
4595
4596 std::vector<spv::Id> result(1, builder.createOp(spv::OpImageRead, resultType(), operands));
4597 builder.setPrecision(result[0], precision);
4598
4599 // If needed, add a conversion constructor to the proper size.
4600 if (components != node->getType().getVectorSize())
4601 result[0] = builder.createConstructor(precision, result, convertGlslangToSpvType(node->getType()));
4602
4603 return result[0];
4604 } else if (node->getOp() == glslang::EOpImageStore || node->getOp() == glslang::EOpImageStoreLod) {
4605
4606 // Push the texel value before the operands
4607 if (sampler.isMultiSample() || cracked.lod) {
4608 spv::IdImmediate texel = { true, *(opIt + 1) };
4609 operands.push_back(texel);
4610 } else {
4611 spv::IdImmediate texel = { true, *opIt };
4612 operands.push_back(texel);
4613 }
4614
4615 spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone;
4616 if (sampler.isMultiSample()) {
4617 mask = mask | spv::ImageOperandsSampleMask;
4618 }
4619 if (cracked.lod) {
4620 builder.addExtension(spv::E_SPV_AMD_shader_image_load_store_lod);
4621 builder.addCapability(spv::CapabilityImageReadWriteLodAMD);
4622 mask = mask | spv::ImageOperandsLodMask;
4623 }
4624 mask = mask | TranslateImageOperands(TranslateCoherent(imageType));
4625 mask = (spv::ImageOperandsMask)(mask & ~spv::ImageOperandsMakeTexelVisibleKHRMask);
4626 mask = mask | signExtensionMask();
4627 if (mask != spv::ImageOperandsMaskNone) {
4628 spv::IdImmediate imageOperands = { false, (unsigned int)mask };
4629 operands.push_back(imageOperands);
4630 }
4631 if (mask & spv::ImageOperandsSampleMask) {
4632 spv::IdImmediate imageOperand = { true, *opIt++ };
4633 operands.push_back(imageOperand);
4634 }
4635 if (mask & spv::ImageOperandsLodMask) {
4636 spv::IdImmediate imageOperand = { true, *opIt++ };
4637 operands.push_back(imageOperand);
4638 }
4639 if (mask & spv::ImageOperandsMakeTexelAvailableKHRMask) {
4640 spv::IdImmediate imageOperand = { true,
4641 builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) };
4642 operands.push_back(imageOperand);
4643 }
4644
4645 builder.createNoResultOp(spv::OpImageWrite, operands);
4646 if (builder.getImageTypeFormat(builder.getImageType(operands.front().word)) == spv::ImageFormatUnknown)
4647 builder.addCapability(spv::CapabilityStorageImageWriteWithoutFormat);
4648 return spv::NoResult;
4649 } else if (node->getOp() == glslang::EOpSparseImageLoad ||
4650 node->getOp() == glslang::EOpSparseImageLoadLod) {
4651 builder.addCapability(spv::CapabilitySparseResidency);
4652 if (builder.getImageTypeFormat(builder.getImageType(operands.front().word)) == spv::ImageFormatUnknown)
4653 builder.addCapability(spv::CapabilityStorageImageReadWithoutFormat);
4654
4655 spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone;
4656 if (sampler.isMultiSample()) {
4657 mask = mask | spv::ImageOperandsSampleMask;
4658 }
4659 if (cracked.lod) {
4660 builder.addExtension(spv::E_SPV_AMD_shader_image_load_store_lod);
4661 builder.addCapability(spv::CapabilityImageReadWriteLodAMD);
4662
4663 mask = mask | spv::ImageOperandsLodMask;
4664 }
4665 mask = mask | TranslateImageOperands(TranslateCoherent(imageType));
4666 mask = (spv::ImageOperandsMask)(mask & ~spv::ImageOperandsMakeTexelAvailableKHRMask);
4667 mask = mask | signExtensionMask();
4668 if (mask != spv::ImageOperandsMaskNone) {
4669 spv::IdImmediate imageOperands = { false, (unsigned int)mask };
4670 operands.push_back(imageOperands);
4671 }
4672 if (mask & spv::ImageOperandsSampleMask) {
4673 spv::IdImmediate imageOperand = { true, *opIt++ };
4674 operands.push_back(imageOperand);
4675 }
4676 if (mask & spv::ImageOperandsLodMask) {
4677 spv::IdImmediate imageOperand = { true, *opIt++ };
4678 operands.push_back(imageOperand);
4679 }
4680 if (mask & spv::ImageOperandsMakeTexelVisibleKHRMask) {
4681 spv::IdImmediate imageOperand = { true, builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) };
4682 operands.push_back(imageOperand);
4683 }
4684
4685 // Create the return type that was a special structure
4686 spv::Id texelOut = *opIt;
4687 spv::Id typeId0 = resultType();
4688 spv::Id typeId1 = builder.getDerefTypeId(texelOut);
4689 spv::Id resultTypeId = builder.makeStructResultType(typeId0, typeId1);
4690
4691 spv::Id resultId = builder.createOp(spv::OpImageSparseRead, resultTypeId, operands);
4692
4693 // Decode the return type
4694 builder.createStore(builder.createCompositeExtract(resultId, typeId1, 1), texelOut);
4695 return builder.createCompositeExtract(resultId, typeId0, 0);
4696 } else {
4697 // Process image atomic operations
4698
4699 // GLSL "IMAGE_PARAMS" will involve in constructing an image texel pointer and this pointer,
4700 // as the first source operand, is required by SPIR-V atomic operations.
4701 // For non-MS, the sample value should be 0
4702 spv::IdImmediate sample = { true, sampler.isMultiSample() ? *(opIt++) : builder.makeUintConstant(0) };
4703 operands.push_back(sample);
4704
4705 spv::Id resultTypeId;
4706 // imageAtomicStore has a void return type so base the pointer type on
4707 // the type of the value operand.
4708 if (node->getOp() == glslang::EOpImageAtomicStore) {
4709 resultTypeId = builder.makePointer(spv::StorageClassImage, builder.getTypeId(operands[2].word));
4710 } else {
4711 resultTypeId = builder.makePointer(spv::StorageClassImage, resultType());
4712 }
4713 spv::Id pointer = builder.createOp(spv::OpImageTexelPointer, resultTypeId, operands);
4714
4715 std::vector<spv::Id> operands;
4716 operands.push_back(pointer);
4717 for (; opIt != arguments.end(); ++opIt)
4718 operands.push_back(*opIt);
4719
4720 return createAtomicOperation(node->getOp(), precision, resultType(), operands, node->getBasicType(), lvalueCoherentFlags);
4721 }
4722 }
4723
4724 #ifndef GLSLANG_WEB
4725 // Check for fragment mask functions other than queries
4726 if (cracked.fragMask) {
4727 assert(sampler.ms);
4728
4729 auto opIt = arguments.begin();
4730 std::vector<spv::Id> operands;
4731
4732 // Extract the image if necessary
4733 if (builder.isSampledImage(params.sampler))
4734 params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler);
4735
4736 operands.push_back(params.sampler);
4737 ++opIt;
4738
4739 if (sampler.isSubpass()) {
4740 // add on the (0,0) coordinate
4741 spv::Id zero = builder.makeIntConstant(0);
4742 std::vector<spv::Id> comps;
4743 comps.push_back(zero);
4744 comps.push_back(zero);
4745 operands.push_back(builder.makeCompositeConstant(builder.makeVectorType(builder.makeIntType(32), 2), comps));
4746 }
4747
4748 for (; opIt != arguments.end(); ++opIt)
4749 operands.push_back(*opIt);
4750
4751 spv::Op fragMaskOp = spv::OpNop;
4752 if (node->getOp() == glslang::EOpFragmentMaskFetch)
4753 fragMaskOp = spv::OpFragmentMaskFetchAMD;
4754 else if (node->getOp() == glslang::EOpFragmentFetch)
4755 fragMaskOp = spv::OpFragmentFetchAMD;
4756
4757 builder.addExtension(spv::E_SPV_AMD_shader_fragment_mask);
4758 builder.addCapability(spv::CapabilityFragmentMaskAMD);
4759 return builder.createOp(fragMaskOp, resultType(), operands);
4760 }
4761 #endif
4762
4763 // Check for texture functions other than queries
4764 bool sparse = node->isSparseTexture();
4765 bool imageFootprint = node->isImageFootprint();
4766 bool cubeCompare = sampler.dim == glslang::EsdCube && sampler.isArrayed() && sampler.isShadow();
4767
4768 // check for bias argument
4769 bool bias = false;
4770 if (! cracked.lod && ! cracked.grad && ! cracked.fetch && ! cubeCompare) {
4771 int nonBiasArgCount = 2;
4772 if (cracked.gather)
4773 ++nonBiasArgCount; // comp argument should be present when bias argument is present
4774
4775 if (f16ShadowCompare)
4776 ++nonBiasArgCount;
4777 if (cracked.offset)
4778 ++nonBiasArgCount;
4779 else if (cracked.offsets)
4780 ++nonBiasArgCount;
4781 if (cracked.grad)
4782 nonBiasArgCount += 2;
4783 if (cracked.lodClamp)
4784 ++nonBiasArgCount;
4785 if (sparse)
4786 ++nonBiasArgCount;
4787 if (imageFootprint)
4788 //Following three extra arguments
4789 // int granularity, bool coarse, out gl_TextureFootprint2DNV footprint
4790 nonBiasArgCount += 3;
4791 if ((int)arguments.size() > nonBiasArgCount)
4792 bias = true;
4793 }
4794
4795 // See if the sampler param should really be just the SPV image part
4796 if (cracked.fetch) {
4797 // a fetch needs to have the image extracted first
4798 if (builder.isSampledImage(params.sampler))
4799 params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler);
4800 }
4801
4802 #ifndef GLSLANG_WEB
4803 if (cracked.gather) {
4804 const auto& sourceExtensions = glslangIntermediate->getRequestedExtensions();
4805 if (bias || cracked.lod ||
4806 sourceExtensions.find(glslang::E_GL_AMD_texture_gather_bias_lod) != sourceExtensions.end()) {
4807 builder.addExtension(spv::E_SPV_AMD_texture_gather_bias_lod);
4808 builder.addCapability(spv::CapabilityImageGatherBiasLodAMD);
4809 }
4810 }
4811 #endif
4812
4813 // set the rest of the arguments
4814
4815 params.coords = arguments[1];
4816 int extraArgs = 0;
4817 bool noImplicitLod = false;
4818
4819 // sort out where Dref is coming from
4820 if (cubeCompare || f16ShadowCompare) {
4821 params.Dref = arguments[2];
4822 ++extraArgs;
4823 } else if (sampler.shadow && cracked.gather) {
4824 params.Dref = arguments[2];
4825 ++extraArgs;
4826 } else if (sampler.shadow) {
4827 std::vector<spv::Id> indexes;
4828 int dRefComp;
4829 if (cracked.proj)
4830 dRefComp = 2; // "The resulting 3rd component of P in the shadow forms is used as Dref"
4831 else
4832 dRefComp = builder.getNumComponents(params.coords) - 1;
4833 indexes.push_back(dRefComp);
4834 params.Dref = builder.createCompositeExtract(params.coords, builder.getScalarTypeId(builder.getTypeId(params.coords)), indexes);
4835 }
4836
4837 // lod
4838 if (cracked.lod) {
4839 params.lod = arguments[2 + extraArgs];
4840 ++extraArgs;
4841 } else if (glslangIntermediate->getStage() != EShLangFragment &&
4842 !(glslangIntermediate->getStage() == EShLangCompute &&
4843 glslangIntermediate->hasLayoutDerivativeModeNone())) {
4844 // we need to invent the default lod for an explicit lod instruction for a non-fragment stage
4845 noImplicitLod = true;
4846 }
4847
4848 // multisample
4849 if (sampler.isMultiSample()) {
4850 params.sample = arguments[2 + extraArgs]; // For MS, "sample" should be specified
4851 ++extraArgs;
4852 }
4853
4854 // gradient
4855 if (cracked.grad) {
4856 params.gradX = arguments[2 + extraArgs];
4857 params.gradY = arguments[3 + extraArgs];
4858 extraArgs += 2;
4859 }
4860
4861 // offset and offsets
4862 if (cracked.offset) {
4863 params.offset = arguments[2 + extraArgs];
4864 ++extraArgs;
4865 } else if (cracked.offsets) {
4866 params.offsets = arguments[2 + extraArgs];
4867 ++extraArgs;
4868 }
4869
4870 #ifndef GLSLANG_WEB
4871 // lod clamp
4872 if (cracked.lodClamp) {
4873 params.lodClamp = arguments[2 + extraArgs];
4874 ++extraArgs;
4875 }
4876 // sparse
4877 if (sparse) {
4878 params.texelOut = arguments[2 + extraArgs];
4879 ++extraArgs;
4880 }
4881 // gather component
4882 if (cracked.gather && ! sampler.shadow) {
4883 // default component is 0, if missing, otherwise an argument
4884 if (2 + extraArgs < (int)arguments.size()) {
4885 params.component = arguments[2 + extraArgs];
4886 ++extraArgs;
4887 } else
4888 params.component = builder.makeIntConstant(0);
4889 }
4890 spv::Id resultStruct = spv::NoResult;
4891 if (imageFootprint) {
4892 //Following three extra arguments
4893 // int granularity, bool coarse, out gl_TextureFootprint2DNV footprint
4894 params.granularity = arguments[2 + extraArgs];
4895 params.coarse = arguments[3 + extraArgs];
4896 resultStruct = arguments[4 + extraArgs];
4897 extraArgs += 3;
4898 }
4899 #endif
4900 // bias
4901 if (bias) {
4902 params.bias = arguments[2 + extraArgs];
4903 ++extraArgs;
4904 }
4905
4906 #ifndef GLSLANG_WEB
4907 if (imageFootprint) {
4908 builder.addExtension(spv::E_SPV_NV_shader_image_footprint);
4909 builder.addCapability(spv::CapabilityImageFootprintNV);
4910
4911
4912 //resultStructType(OpenGL type) contains 5 elements:
4913 //struct gl_TextureFootprint2DNV {
4914 // uvec2 anchor;
4915 // uvec2 offset;
4916 // uvec2 mask;
4917 // uint lod;
4918 // uint granularity;
4919 //};
4920 //or
4921 //struct gl_TextureFootprint3DNV {
4922 // uvec3 anchor;
4923 // uvec3 offset;
4924 // uvec2 mask;
4925 // uint lod;
4926 // uint granularity;
4927 //};
4928 spv::Id resultStructType = builder.getContainedTypeId(builder.getTypeId(resultStruct));
4929 assert(builder.isStructType(resultStructType));
4930
4931 //resType (SPIR-V type) contains 6 elements:
4932 //Member 0 must be a Boolean type scalar(LOD),
4933 //Member 1 must be a vector of integer type, whose Signedness operand is 0(anchor),
4934 //Member 2 must be a vector of integer type, whose Signedness operand is 0(offset),
4935 //Member 3 must be a vector of integer type, whose Signedness operand is 0(mask),
4936 //Member 4 must be a scalar of integer type, whose Signedness operand is 0(lod),
4937 //Member 5 must be a scalar of integer type, whose Signedness operand is 0(granularity).
4938 std::vector<spv::Id> members;
4939 members.push_back(resultType());
4940 for (int i = 0; i < 5; i++) {
4941 members.push_back(builder.getContainedTypeId(resultStructType, i));
4942 }
4943 spv::Id resType = builder.makeStructType(members, "ResType");
4944
4945 //call ImageFootprintNV
4946 spv::Id res = builder.createTextureCall(precision, resType, sparse, cracked.fetch, cracked.proj,
4947 cracked.gather, noImplicitLod, params, signExtensionMask());
4948
4949 //copy resType (SPIR-V type) to resultStructType(OpenGL type)
4950 for (int i = 0; i < 5; i++) {
4951 builder.clearAccessChain();
4952 builder.setAccessChainLValue(resultStruct);
4953
4954 //Accessing to a struct we created, no coherent flag is set
4955 spv::Builder::AccessChain::CoherentFlags flags;
4956 flags.clear();
4957
4958 builder.accessChainPush(builder.makeIntConstant(i), flags, 0);
4959 builder.accessChainStore(builder.createCompositeExtract(res, builder.getContainedTypeId(resType, i+1), i+1));
4960 }
4961 return builder.createCompositeExtract(res, resultType(), 0);
4962 }
4963 #endif
4964
4965 // projective component (might not to move)
4966 // GLSL: "The texture coordinates consumed from P, not including the last component of P,
4967 // are divided by the last component of P."
4968 // SPIR-V: "... (u [, v] [, w], q)... It may be a vector larger than needed, but all
4969 // unused components will appear after all used components."
4970 if (cracked.proj) {
4971 int projSourceComp = builder.getNumComponents(params.coords) - 1;
4972 int projTargetComp;
4973 switch (sampler.dim) {
4974 case glslang::Esd1D: projTargetComp = 1; break;
4975 case glslang::Esd2D: projTargetComp = 2; break;
4976 case glslang::EsdRect: projTargetComp = 2; break;
4977 default: projTargetComp = projSourceComp; break;
4978 }
4979 // copy the projective coordinate if we have to
4980 if (projTargetComp != projSourceComp) {
4981 spv::Id projComp = builder.createCompositeExtract(params.coords,
4982 builder.getScalarTypeId(builder.getTypeId(params.coords)),
4983 projSourceComp);
4984 params.coords = builder.createCompositeInsert(projComp, params.coords,
4985 builder.getTypeId(params.coords), projTargetComp);
4986 }
4987 }
4988
4989 #ifndef GLSLANG_WEB
4990 // nonprivate
4991 if (imageType.getQualifier().nonprivate) {
4992 params.nonprivate = true;
4993 }
4994
4995 // volatile
4996 if (imageType.getQualifier().volatil) {
4997 params.volatil = true;
4998 }
4999 #endif
5000
5001 std::vector<spv::Id> result( 1,
5002 builder.createTextureCall(precision, resultType(), sparse, cracked.fetch, cracked.proj, cracked.gather,
5003 noImplicitLod, params, signExtensionMask())
5004 );
5005
5006 if (components != node->getType().getVectorSize())
5007 result[0] = builder.createConstructor(precision, result, convertGlslangToSpvType(node->getType()));
5008
5009 return result[0];
5010 }
5011
handleUserFunctionCall(const glslang::TIntermAggregate * node)5012 spv::Id TGlslangToSpvTraverser::handleUserFunctionCall(const glslang::TIntermAggregate* node)
5013 {
5014 // Grab the function's pointer from the previously created function
5015 spv::Function* function = functionMap[node->getName().c_str()];
5016 if (! function)
5017 return 0;
5018
5019 const glslang::TIntermSequence& glslangArgs = node->getSequence();
5020 const glslang::TQualifierList& qualifiers = node->getQualifierList();
5021
5022 // See comments in makeFunctions() for details about the semantics for parameter passing.
5023 //
5024 // These imply we need a four step process:
5025 // 1. Evaluate the arguments
5026 // 2. Allocate and make copies of in, out, and inout arguments
5027 // 3. Make the call
5028 // 4. Copy back the results
5029
5030 // 1. Evaluate the arguments and their types
5031 std::vector<spv::Builder::AccessChain> lValues;
5032 std::vector<spv::Id> rValues;
5033 std::vector<const glslang::TType*> argTypes;
5034 for (int a = 0; a < (int)glslangArgs.size(); ++a) {
5035 argTypes.push_back(&glslangArgs[a]->getAsTyped()->getType());
5036 // build l-value
5037 builder.clearAccessChain();
5038 glslangArgs[a]->traverse(this);
5039 // keep outputs and pass-by-originals as l-values, evaluate others as r-values
5040 if (originalParam(qualifiers[a], *argTypes[a], function->hasImplicitThis() && a == 0) ||
5041 writableParam(qualifiers[a])) {
5042 // save l-value
5043 lValues.push_back(builder.getAccessChain());
5044 } else {
5045 // process r-value
5046 rValues.push_back(accessChainLoad(*argTypes.back()));
5047 }
5048 }
5049
5050 // 2. Allocate space for anything needing a copy, and if it's "in" or "inout"
5051 // copy the original into that space.
5052 //
5053 // Also, build up the list of actual arguments to pass in for the call
5054 int lValueCount = 0;
5055 int rValueCount = 0;
5056 std::vector<spv::Id> spvArgs;
5057 for (int a = 0; a < (int)glslangArgs.size(); ++a) {
5058 spv::Id arg;
5059 if (originalParam(qualifiers[a], *argTypes[a], function->hasImplicitThis() && a == 0)) {
5060 builder.setAccessChain(lValues[lValueCount]);
5061 arg = builder.accessChainGetLValue();
5062 ++lValueCount;
5063 } else if (writableParam(qualifiers[a])) {
5064 // need space to hold the copy
5065 arg = builder.createVariable(spv::StorageClassFunction, builder.getContainedTypeId(function->getParamType(a)), "param");
5066 if (qualifiers[a] == glslang::EvqIn || qualifiers[a] == glslang::EvqInOut) {
5067 // need to copy the input into output space
5068 builder.setAccessChain(lValues[lValueCount]);
5069 spv::Id copy = accessChainLoad(*argTypes[a]);
5070 builder.clearAccessChain();
5071 builder.setAccessChainLValue(arg);
5072 multiTypeStore(*argTypes[a], copy);
5073 }
5074 ++lValueCount;
5075 } else {
5076 // process r-value, which involves a copy for a type mismatch
5077 if (function->getParamType(a) != convertGlslangToSpvType(*argTypes[a])) {
5078 spv::Id argCopy = builder.createVariable(spv::StorageClassFunction, function->getParamType(a), "arg");
5079 builder.clearAccessChain();
5080 builder.setAccessChainLValue(argCopy);
5081 multiTypeStore(*argTypes[a], rValues[rValueCount]);
5082 arg = builder.createLoad(argCopy);
5083 } else
5084 arg = rValues[rValueCount];
5085 ++rValueCount;
5086 }
5087 spvArgs.push_back(arg);
5088 }
5089
5090 // 3. Make the call.
5091 spv::Id result = builder.createFunctionCall(function, spvArgs);
5092 builder.setPrecision(result, TranslatePrecisionDecoration(node->getType()));
5093
5094 // 4. Copy back out an "out" arguments.
5095 lValueCount = 0;
5096 for (int a = 0; a < (int)glslangArgs.size(); ++a) {
5097 if (originalParam(qualifiers[a], *argTypes[a], function->hasImplicitThis() && a == 0))
5098 ++lValueCount;
5099 else if (writableParam(qualifiers[a])) {
5100 if (qualifiers[a] == glslang::EvqOut || qualifiers[a] == glslang::EvqInOut) {
5101 spv::Id copy = builder.createLoad(spvArgs[a]);
5102 builder.setAccessChain(lValues[lValueCount]);
5103 multiTypeStore(*argTypes[a], copy);
5104 }
5105 ++lValueCount;
5106 }
5107 }
5108
5109 return result;
5110 }
5111
5112 // Translate AST operation to SPV operation, already having SPV-based operands/types.
createBinaryOperation(glslang::TOperator op,OpDecorations & decorations,spv::Id typeId,spv::Id left,spv::Id right,glslang::TBasicType typeProxy,bool reduceComparison)5113 spv::Id TGlslangToSpvTraverser::createBinaryOperation(glslang::TOperator op, OpDecorations& decorations,
5114 spv::Id typeId, spv::Id left, spv::Id right,
5115 glslang::TBasicType typeProxy, bool reduceComparison)
5116 {
5117 bool isUnsigned = isTypeUnsignedInt(typeProxy);
5118 bool isFloat = isTypeFloat(typeProxy);
5119 bool isBool = typeProxy == glslang::EbtBool;
5120
5121 spv::Op binOp = spv::OpNop;
5122 bool needMatchingVectors = true; // for non-matrix ops, would a scalar need to smear to match a vector?
5123 bool comparison = false;
5124
5125 switch (op) {
5126 case glslang::EOpAdd:
5127 case glslang::EOpAddAssign:
5128 if (isFloat)
5129 binOp = spv::OpFAdd;
5130 else
5131 binOp = spv::OpIAdd;
5132 break;
5133 case glslang::EOpSub:
5134 case glslang::EOpSubAssign:
5135 if (isFloat)
5136 binOp = spv::OpFSub;
5137 else
5138 binOp = spv::OpISub;
5139 break;
5140 case glslang::EOpMul:
5141 case glslang::EOpMulAssign:
5142 if (isFloat)
5143 binOp = spv::OpFMul;
5144 else
5145 binOp = spv::OpIMul;
5146 break;
5147 case glslang::EOpVectorTimesScalar:
5148 case glslang::EOpVectorTimesScalarAssign:
5149 if (isFloat && (builder.isVector(left) || builder.isVector(right))) {
5150 if (builder.isVector(right))
5151 std::swap(left, right);
5152 assert(builder.isScalar(right));
5153 needMatchingVectors = false;
5154 binOp = spv::OpVectorTimesScalar;
5155 } else if (isFloat)
5156 binOp = spv::OpFMul;
5157 else
5158 binOp = spv::OpIMul;
5159 break;
5160 case glslang::EOpVectorTimesMatrix:
5161 case glslang::EOpVectorTimesMatrixAssign:
5162 binOp = spv::OpVectorTimesMatrix;
5163 break;
5164 case glslang::EOpMatrixTimesVector:
5165 binOp = spv::OpMatrixTimesVector;
5166 break;
5167 case glslang::EOpMatrixTimesScalar:
5168 case glslang::EOpMatrixTimesScalarAssign:
5169 binOp = spv::OpMatrixTimesScalar;
5170 break;
5171 case glslang::EOpMatrixTimesMatrix:
5172 case glslang::EOpMatrixTimesMatrixAssign:
5173 binOp = spv::OpMatrixTimesMatrix;
5174 break;
5175 case glslang::EOpOuterProduct:
5176 binOp = spv::OpOuterProduct;
5177 needMatchingVectors = false;
5178 break;
5179
5180 case glslang::EOpDiv:
5181 case glslang::EOpDivAssign:
5182 if (isFloat)
5183 binOp = spv::OpFDiv;
5184 else if (isUnsigned)
5185 binOp = spv::OpUDiv;
5186 else
5187 binOp = spv::OpSDiv;
5188 break;
5189 case glslang::EOpMod:
5190 case glslang::EOpModAssign:
5191 if (isFloat)
5192 binOp = spv::OpFMod;
5193 else if (isUnsigned)
5194 binOp = spv::OpUMod;
5195 else
5196 binOp = spv::OpSMod;
5197 break;
5198 case glslang::EOpRightShift:
5199 case glslang::EOpRightShiftAssign:
5200 if (isUnsigned)
5201 binOp = spv::OpShiftRightLogical;
5202 else
5203 binOp = spv::OpShiftRightArithmetic;
5204 break;
5205 case glslang::EOpLeftShift:
5206 case glslang::EOpLeftShiftAssign:
5207 binOp = spv::OpShiftLeftLogical;
5208 break;
5209 case glslang::EOpAnd:
5210 case glslang::EOpAndAssign:
5211 binOp = spv::OpBitwiseAnd;
5212 break;
5213 case glslang::EOpLogicalAnd:
5214 needMatchingVectors = false;
5215 binOp = spv::OpLogicalAnd;
5216 break;
5217 case glslang::EOpInclusiveOr:
5218 case glslang::EOpInclusiveOrAssign:
5219 binOp = spv::OpBitwiseOr;
5220 break;
5221 case glslang::EOpLogicalOr:
5222 needMatchingVectors = false;
5223 binOp = spv::OpLogicalOr;
5224 break;
5225 case glslang::EOpExclusiveOr:
5226 case glslang::EOpExclusiveOrAssign:
5227 binOp = spv::OpBitwiseXor;
5228 break;
5229 case glslang::EOpLogicalXor:
5230 needMatchingVectors = false;
5231 binOp = spv::OpLogicalNotEqual;
5232 break;
5233
5234 case glslang::EOpLessThan:
5235 case glslang::EOpGreaterThan:
5236 case glslang::EOpLessThanEqual:
5237 case glslang::EOpGreaterThanEqual:
5238 case glslang::EOpEqual:
5239 case glslang::EOpNotEqual:
5240 case glslang::EOpVectorEqual:
5241 case glslang::EOpVectorNotEqual:
5242 comparison = true;
5243 break;
5244 default:
5245 break;
5246 }
5247
5248 // handle mapped binary operations (should be non-comparison)
5249 if (binOp != spv::OpNop) {
5250 assert(comparison == false);
5251 if (builder.isMatrix(left) || builder.isMatrix(right) ||
5252 builder.isCooperativeMatrix(left) || builder.isCooperativeMatrix(right))
5253 return createBinaryMatrixOperation(binOp, decorations, typeId, left, right);
5254
5255 // No matrix involved; make both operands be the same number of components, if needed
5256 if (needMatchingVectors)
5257 builder.promoteScalar(decorations.precision, left, right);
5258
5259 spv::Id result = builder.createBinOp(binOp, typeId, left, right);
5260 decorations.addNoContraction(builder, result);
5261 decorations.addNonUniform(builder, result);
5262 return builder.setPrecision(result, decorations.precision);
5263 }
5264
5265 if (! comparison)
5266 return 0;
5267
5268 // Handle comparison instructions
5269
5270 if (reduceComparison && (op == glslang::EOpEqual || op == glslang::EOpNotEqual)
5271 && (builder.isVector(left) || builder.isMatrix(left) || builder.isAggregate(left))) {
5272 spv::Id result = builder.createCompositeCompare(decorations.precision, left, right, op == glslang::EOpEqual);
5273 decorations.addNonUniform(builder, result);
5274 return result;
5275 }
5276
5277 switch (op) {
5278 case glslang::EOpLessThan:
5279 if (isFloat)
5280 binOp = spv::OpFOrdLessThan;
5281 else if (isUnsigned)
5282 binOp = spv::OpULessThan;
5283 else
5284 binOp = spv::OpSLessThan;
5285 break;
5286 case glslang::EOpGreaterThan:
5287 if (isFloat)
5288 binOp = spv::OpFOrdGreaterThan;
5289 else if (isUnsigned)
5290 binOp = spv::OpUGreaterThan;
5291 else
5292 binOp = spv::OpSGreaterThan;
5293 break;
5294 case glslang::EOpLessThanEqual:
5295 if (isFloat)
5296 binOp = spv::OpFOrdLessThanEqual;
5297 else if (isUnsigned)
5298 binOp = spv::OpULessThanEqual;
5299 else
5300 binOp = spv::OpSLessThanEqual;
5301 break;
5302 case glslang::EOpGreaterThanEqual:
5303 if (isFloat)
5304 binOp = spv::OpFOrdGreaterThanEqual;
5305 else if (isUnsigned)
5306 binOp = spv::OpUGreaterThanEqual;
5307 else
5308 binOp = spv::OpSGreaterThanEqual;
5309 break;
5310 case glslang::EOpEqual:
5311 case glslang::EOpVectorEqual:
5312 if (isFloat)
5313 binOp = spv::OpFOrdEqual;
5314 else if (isBool)
5315 binOp = spv::OpLogicalEqual;
5316 else
5317 binOp = spv::OpIEqual;
5318 break;
5319 case glslang::EOpNotEqual:
5320 case glslang::EOpVectorNotEqual:
5321 if (isFloat)
5322 binOp = spv::OpFOrdNotEqual;
5323 else if (isBool)
5324 binOp = spv::OpLogicalNotEqual;
5325 else
5326 binOp = spv::OpINotEqual;
5327 break;
5328 default:
5329 break;
5330 }
5331
5332 if (binOp != spv::OpNop) {
5333 spv::Id result = builder.createBinOp(binOp, typeId, left, right);
5334 decorations.addNoContraction(builder, result);
5335 decorations.addNonUniform(builder, result);
5336 return builder.setPrecision(result, decorations.precision);
5337 }
5338
5339 return 0;
5340 }
5341
5342 //
5343 // Translate AST matrix operation to SPV operation, already having SPV-based operands/types.
5344 // These can be any of:
5345 //
5346 // matrix * scalar
5347 // scalar * matrix
5348 // matrix * matrix linear algebraic
5349 // matrix * vector
5350 // vector * matrix
5351 // matrix * matrix componentwise
5352 // matrix op matrix op in {+, -, /}
5353 // matrix op scalar op in {+, -, /}
5354 // scalar op matrix op in {+, -, /}
5355 //
createBinaryMatrixOperation(spv::Op op,OpDecorations & decorations,spv::Id typeId,spv::Id left,spv::Id right)5356 spv::Id TGlslangToSpvTraverser::createBinaryMatrixOperation(spv::Op op, OpDecorations& decorations, spv::Id typeId,
5357 spv::Id left, spv::Id right)
5358 {
5359 bool firstClass = true;
5360
5361 // First, handle first-class matrix operations (* and matrix/scalar)
5362 switch (op) {
5363 case spv::OpFDiv:
5364 if (builder.isMatrix(left) && builder.isScalar(right)) {
5365 // turn matrix / scalar into a multiply...
5366 spv::Id resultType = builder.getTypeId(right);
5367 right = builder.createBinOp(spv::OpFDiv, resultType, builder.makeFpConstant(resultType, 1.0), right);
5368 op = spv::OpMatrixTimesScalar;
5369 } else
5370 firstClass = false;
5371 break;
5372 case spv::OpMatrixTimesScalar:
5373 if (builder.isMatrix(right) || builder.isCooperativeMatrix(right))
5374 std::swap(left, right);
5375 assert(builder.isScalar(right));
5376 break;
5377 case spv::OpVectorTimesMatrix:
5378 assert(builder.isVector(left));
5379 assert(builder.isMatrix(right));
5380 break;
5381 case spv::OpMatrixTimesVector:
5382 assert(builder.isMatrix(left));
5383 assert(builder.isVector(right));
5384 break;
5385 case spv::OpMatrixTimesMatrix:
5386 assert(builder.isMatrix(left));
5387 assert(builder.isMatrix(right));
5388 break;
5389 default:
5390 firstClass = false;
5391 break;
5392 }
5393
5394 if (builder.isCooperativeMatrix(left) || builder.isCooperativeMatrix(right))
5395 firstClass = true;
5396
5397 if (firstClass) {
5398 spv::Id result = builder.createBinOp(op, typeId, left, right);
5399 decorations.addNoContraction(builder, result);
5400 decorations.addNonUniform(builder, result);
5401 return builder.setPrecision(result, decorations.precision);
5402 }
5403
5404 // Handle component-wise +, -, *, %, and / for all combinations of type.
5405 // The result type of all of them is the same type as the (a) matrix operand.
5406 // The algorithm is to:
5407 // - break the matrix(es) into vectors
5408 // - smear any scalar to a vector
5409 // - do vector operations
5410 // - make a matrix out the vector results
5411 switch (op) {
5412 case spv::OpFAdd:
5413 case spv::OpFSub:
5414 case spv::OpFDiv:
5415 case spv::OpFMod:
5416 case spv::OpFMul:
5417 {
5418 // one time set up...
5419 bool leftMat = builder.isMatrix(left);
5420 bool rightMat = builder.isMatrix(right);
5421 unsigned int numCols = leftMat ? builder.getNumColumns(left) : builder.getNumColumns(right);
5422 int numRows = leftMat ? builder.getNumRows(left) : builder.getNumRows(right);
5423 spv::Id scalarType = builder.getScalarTypeId(typeId);
5424 spv::Id vecType = builder.makeVectorType(scalarType, numRows);
5425 std::vector<spv::Id> results;
5426 spv::Id smearVec = spv::NoResult;
5427 if (builder.isScalar(left))
5428 smearVec = builder.smearScalar(decorations.precision, left, vecType);
5429 else if (builder.isScalar(right))
5430 smearVec = builder.smearScalar(decorations.precision, right, vecType);
5431
5432 // do each vector op
5433 for (unsigned int c = 0; c < numCols; ++c) {
5434 std::vector<unsigned int> indexes;
5435 indexes.push_back(c);
5436 spv::Id leftVec = leftMat ? builder.createCompositeExtract( left, vecType, indexes) : smearVec;
5437 spv::Id rightVec = rightMat ? builder.createCompositeExtract(right, vecType, indexes) : smearVec;
5438 spv::Id result = builder.createBinOp(op, vecType, leftVec, rightVec);
5439 decorations.addNoContraction(builder, result);
5440 decorations.addNonUniform(builder, result);
5441 results.push_back(builder.setPrecision(result, decorations.precision));
5442 }
5443
5444 // put the pieces together
5445 spv::Id result = builder.setPrecision(builder.createCompositeConstruct(typeId, results), decorations.precision);
5446 decorations.addNonUniform(builder, result);
5447 return result;
5448 }
5449 default:
5450 assert(0);
5451 return spv::NoResult;
5452 }
5453 }
5454
createUnaryOperation(glslang::TOperator op,OpDecorations & decorations,spv::Id typeId,spv::Id operand,glslang::TBasicType typeProxy,const spv::Builder::AccessChain::CoherentFlags & lvalueCoherentFlags)5455 spv::Id TGlslangToSpvTraverser::createUnaryOperation(glslang::TOperator op, OpDecorations& decorations, spv::Id typeId,
5456 spv::Id operand, glslang::TBasicType typeProxy, const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags)
5457 {
5458 spv::Op unaryOp = spv::OpNop;
5459 int extBuiltins = -1;
5460 int libCall = -1;
5461 bool isUnsigned = isTypeUnsignedInt(typeProxy);
5462 bool isFloat = isTypeFloat(typeProxy);
5463
5464 switch (op) {
5465 case glslang::EOpNegative:
5466 if (isFloat) {
5467 unaryOp = spv::OpFNegate;
5468 if (builder.isMatrixType(typeId))
5469 return createUnaryMatrixOperation(unaryOp, decorations, typeId, operand, typeProxy);
5470 } else
5471 unaryOp = spv::OpSNegate;
5472 break;
5473
5474 case glslang::EOpLogicalNot:
5475 case glslang::EOpVectorLogicalNot:
5476 unaryOp = spv::OpLogicalNot;
5477 break;
5478 case glslang::EOpBitwiseNot:
5479 unaryOp = spv::OpNot;
5480 break;
5481
5482 case glslang::EOpDeterminant:
5483 libCall = spv::GLSLstd450Determinant;
5484 break;
5485 case glslang::EOpMatrixInverse:
5486 libCall = spv::GLSLstd450MatrixInverse;
5487 break;
5488 case glslang::EOpTranspose:
5489 unaryOp = spv::OpTranspose;
5490 break;
5491
5492 case glslang::EOpRadians:
5493 libCall = spv::GLSLstd450Radians;
5494 break;
5495 case glslang::EOpDegrees:
5496 libCall = spv::GLSLstd450Degrees;
5497 break;
5498 case glslang::EOpSin:
5499 libCall = spv::GLSLstd450Sin;
5500 break;
5501 case glslang::EOpCos:
5502 libCall = spv::GLSLstd450Cos;
5503 break;
5504 case glslang::EOpTan:
5505 libCall = spv::GLSLstd450Tan;
5506 break;
5507 case glslang::EOpAcos:
5508 libCall = spv::GLSLstd450Acos;
5509 break;
5510 case glslang::EOpAsin:
5511 libCall = spv::GLSLstd450Asin;
5512 break;
5513 case glslang::EOpAtan:
5514 libCall = spv::GLSLstd450Atan;
5515 break;
5516
5517 case glslang::EOpAcosh:
5518 libCall = spv::GLSLstd450Acosh;
5519 break;
5520 case glslang::EOpAsinh:
5521 libCall = spv::GLSLstd450Asinh;
5522 break;
5523 case glslang::EOpAtanh:
5524 libCall = spv::GLSLstd450Atanh;
5525 break;
5526 case glslang::EOpTanh:
5527 libCall = spv::GLSLstd450Tanh;
5528 break;
5529 case glslang::EOpCosh:
5530 libCall = spv::GLSLstd450Cosh;
5531 break;
5532 case glslang::EOpSinh:
5533 libCall = spv::GLSLstd450Sinh;
5534 break;
5535
5536 case glslang::EOpLength:
5537 libCall = spv::GLSLstd450Length;
5538 break;
5539 case glslang::EOpNormalize:
5540 libCall = spv::GLSLstd450Normalize;
5541 break;
5542
5543 case glslang::EOpExp:
5544 libCall = spv::GLSLstd450Exp;
5545 break;
5546 case glslang::EOpLog:
5547 libCall = spv::GLSLstd450Log;
5548 break;
5549 case glslang::EOpExp2:
5550 libCall = spv::GLSLstd450Exp2;
5551 break;
5552 case glslang::EOpLog2:
5553 libCall = spv::GLSLstd450Log2;
5554 break;
5555 case glslang::EOpSqrt:
5556 libCall = spv::GLSLstd450Sqrt;
5557 break;
5558 case glslang::EOpInverseSqrt:
5559 libCall = spv::GLSLstd450InverseSqrt;
5560 break;
5561
5562 case glslang::EOpFloor:
5563 libCall = spv::GLSLstd450Floor;
5564 break;
5565 case glslang::EOpTrunc:
5566 libCall = spv::GLSLstd450Trunc;
5567 break;
5568 case glslang::EOpRound:
5569 libCall = spv::GLSLstd450Round;
5570 break;
5571 case glslang::EOpRoundEven:
5572 libCall = spv::GLSLstd450RoundEven;
5573 break;
5574 case glslang::EOpCeil:
5575 libCall = spv::GLSLstd450Ceil;
5576 break;
5577 case glslang::EOpFract:
5578 libCall = spv::GLSLstd450Fract;
5579 break;
5580
5581 case glslang::EOpIsNan:
5582 unaryOp = spv::OpIsNan;
5583 break;
5584 case glslang::EOpIsInf:
5585 unaryOp = spv::OpIsInf;
5586 break;
5587 case glslang::EOpIsFinite:
5588 unaryOp = spv::OpIsFinite;
5589 break;
5590
5591 case glslang::EOpFloatBitsToInt:
5592 case glslang::EOpFloatBitsToUint:
5593 case glslang::EOpIntBitsToFloat:
5594 case glslang::EOpUintBitsToFloat:
5595 case glslang::EOpDoubleBitsToInt64:
5596 case glslang::EOpDoubleBitsToUint64:
5597 case glslang::EOpInt64BitsToDouble:
5598 case glslang::EOpUint64BitsToDouble:
5599 case glslang::EOpFloat16BitsToInt16:
5600 case glslang::EOpFloat16BitsToUint16:
5601 case glslang::EOpInt16BitsToFloat16:
5602 case glslang::EOpUint16BitsToFloat16:
5603 unaryOp = spv::OpBitcast;
5604 break;
5605
5606 case glslang::EOpPackSnorm2x16:
5607 libCall = spv::GLSLstd450PackSnorm2x16;
5608 break;
5609 case glslang::EOpUnpackSnorm2x16:
5610 libCall = spv::GLSLstd450UnpackSnorm2x16;
5611 break;
5612 case glslang::EOpPackUnorm2x16:
5613 libCall = spv::GLSLstd450PackUnorm2x16;
5614 break;
5615 case glslang::EOpUnpackUnorm2x16:
5616 libCall = spv::GLSLstd450UnpackUnorm2x16;
5617 break;
5618 case glslang::EOpPackHalf2x16:
5619 libCall = spv::GLSLstd450PackHalf2x16;
5620 break;
5621 case glslang::EOpUnpackHalf2x16:
5622 libCall = spv::GLSLstd450UnpackHalf2x16;
5623 break;
5624 #ifndef GLSLANG_WEB
5625 case glslang::EOpPackSnorm4x8:
5626 libCall = spv::GLSLstd450PackSnorm4x8;
5627 break;
5628 case glslang::EOpUnpackSnorm4x8:
5629 libCall = spv::GLSLstd450UnpackSnorm4x8;
5630 break;
5631 case glslang::EOpPackUnorm4x8:
5632 libCall = spv::GLSLstd450PackUnorm4x8;
5633 break;
5634 case glslang::EOpUnpackUnorm4x8:
5635 libCall = spv::GLSLstd450UnpackUnorm4x8;
5636 break;
5637 case glslang::EOpPackDouble2x32:
5638 libCall = spv::GLSLstd450PackDouble2x32;
5639 break;
5640 case glslang::EOpUnpackDouble2x32:
5641 libCall = spv::GLSLstd450UnpackDouble2x32;
5642 break;
5643 #endif
5644
5645 case glslang::EOpPackInt2x32:
5646 case glslang::EOpUnpackInt2x32:
5647 case glslang::EOpPackUint2x32:
5648 case glslang::EOpUnpackUint2x32:
5649 case glslang::EOpPack16:
5650 case glslang::EOpPack32:
5651 case glslang::EOpPack64:
5652 case glslang::EOpUnpack32:
5653 case glslang::EOpUnpack16:
5654 case glslang::EOpUnpack8:
5655 case glslang::EOpPackInt2x16:
5656 case glslang::EOpUnpackInt2x16:
5657 case glslang::EOpPackUint2x16:
5658 case glslang::EOpUnpackUint2x16:
5659 case glslang::EOpPackInt4x16:
5660 case glslang::EOpUnpackInt4x16:
5661 case glslang::EOpPackUint4x16:
5662 case glslang::EOpUnpackUint4x16:
5663 case glslang::EOpPackFloat2x16:
5664 case glslang::EOpUnpackFloat2x16:
5665 unaryOp = spv::OpBitcast;
5666 break;
5667
5668 case glslang::EOpDPdx:
5669 unaryOp = spv::OpDPdx;
5670 break;
5671 case glslang::EOpDPdy:
5672 unaryOp = spv::OpDPdy;
5673 break;
5674 case glslang::EOpFwidth:
5675 unaryOp = spv::OpFwidth;
5676 break;
5677
5678 case glslang::EOpAny:
5679 unaryOp = spv::OpAny;
5680 break;
5681 case glslang::EOpAll:
5682 unaryOp = spv::OpAll;
5683 break;
5684
5685 case glslang::EOpAbs:
5686 if (isFloat)
5687 libCall = spv::GLSLstd450FAbs;
5688 else
5689 libCall = spv::GLSLstd450SAbs;
5690 break;
5691 case glslang::EOpSign:
5692 if (isFloat)
5693 libCall = spv::GLSLstd450FSign;
5694 else
5695 libCall = spv::GLSLstd450SSign;
5696 break;
5697
5698 #ifndef GLSLANG_WEB
5699 case glslang::EOpDPdxFine:
5700 unaryOp = spv::OpDPdxFine;
5701 break;
5702 case glslang::EOpDPdyFine:
5703 unaryOp = spv::OpDPdyFine;
5704 break;
5705 case glslang::EOpFwidthFine:
5706 unaryOp = spv::OpFwidthFine;
5707 break;
5708 case glslang::EOpDPdxCoarse:
5709 unaryOp = spv::OpDPdxCoarse;
5710 break;
5711 case glslang::EOpDPdyCoarse:
5712 unaryOp = spv::OpDPdyCoarse;
5713 break;
5714 case glslang::EOpFwidthCoarse:
5715 unaryOp = spv::OpFwidthCoarse;
5716 break;
5717 case glslang::EOpInterpolateAtCentroid:
5718 if (typeProxy == glslang::EbtFloat16)
5719 builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float);
5720 libCall = spv::GLSLstd450InterpolateAtCentroid;
5721 break;
5722 case glslang::EOpAtomicCounterIncrement:
5723 case glslang::EOpAtomicCounterDecrement:
5724 case glslang::EOpAtomicCounter:
5725 {
5726 // Handle all of the atomics in one place, in createAtomicOperation()
5727 std::vector<spv::Id> operands;
5728 operands.push_back(operand);
5729 return createAtomicOperation(op, decorations.precision, typeId, operands, typeProxy, lvalueCoherentFlags);
5730 }
5731
5732 case glslang::EOpBitFieldReverse:
5733 unaryOp = spv::OpBitReverse;
5734 break;
5735 case glslang::EOpBitCount:
5736 unaryOp = spv::OpBitCount;
5737 break;
5738 case glslang::EOpFindLSB:
5739 libCall = spv::GLSLstd450FindILsb;
5740 break;
5741 case glslang::EOpFindMSB:
5742 if (isUnsigned)
5743 libCall = spv::GLSLstd450FindUMsb;
5744 else
5745 libCall = spv::GLSLstd450FindSMsb;
5746 break;
5747
5748 case glslang::EOpBallot:
5749 case glslang::EOpReadFirstInvocation:
5750 case glslang::EOpAnyInvocation:
5751 case glslang::EOpAllInvocations:
5752 case glslang::EOpAllInvocationsEqual:
5753 case glslang::EOpMinInvocations:
5754 case glslang::EOpMaxInvocations:
5755 case glslang::EOpAddInvocations:
5756 case glslang::EOpMinInvocationsNonUniform:
5757 case glslang::EOpMaxInvocationsNonUniform:
5758 case glslang::EOpAddInvocationsNonUniform:
5759 case glslang::EOpMinInvocationsInclusiveScan:
5760 case glslang::EOpMaxInvocationsInclusiveScan:
5761 case glslang::EOpAddInvocationsInclusiveScan:
5762 case glslang::EOpMinInvocationsInclusiveScanNonUniform:
5763 case glslang::EOpMaxInvocationsInclusiveScanNonUniform:
5764 case glslang::EOpAddInvocationsInclusiveScanNonUniform:
5765 case glslang::EOpMinInvocationsExclusiveScan:
5766 case glslang::EOpMaxInvocationsExclusiveScan:
5767 case glslang::EOpAddInvocationsExclusiveScan:
5768 case glslang::EOpMinInvocationsExclusiveScanNonUniform:
5769 case glslang::EOpMaxInvocationsExclusiveScanNonUniform:
5770 case glslang::EOpAddInvocationsExclusiveScanNonUniform:
5771 {
5772 std::vector<spv::Id> operands;
5773 operands.push_back(operand);
5774 return createInvocationsOperation(op, typeId, operands, typeProxy);
5775 }
5776 case glslang::EOpSubgroupAll:
5777 case glslang::EOpSubgroupAny:
5778 case glslang::EOpSubgroupAllEqual:
5779 case glslang::EOpSubgroupBroadcastFirst:
5780 case glslang::EOpSubgroupBallot:
5781 case glslang::EOpSubgroupInverseBallot:
5782 case glslang::EOpSubgroupBallotBitCount:
5783 case glslang::EOpSubgroupBallotInclusiveBitCount:
5784 case glslang::EOpSubgroupBallotExclusiveBitCount:
5785 case glslang::EOpSubgroupBallotFindLSB:
5786 case glslang::EOpSubgroupBallotFindMSB:
5787 case glslang::EOpSubgroupAdd:
5788 case glslang::EOpSubgroupMul:
5789 case glslang::EOpSubgroupMin:
5790 case glslang::EOpSubgroupMax:
5791 case glslang::EOpSubgroupAnd:
5792 case glslang::EOpSubgroupOr:
5793 case glslang::EOpSubgroupXor:
5794 case glslang::EOpSubgroupInclusiveAdd:
5795 case glslang::EOpSubgroupInclusiveMul:
5796 case glslang::EOpSubgroupInclusiveMin:
5797 case glslang::EOpSubgroupInclusiveMax:
5798 case glslang::EOpSubgroupInclusiveAnd:
5799 case glslang::EOpSubgroupInclusiveOr:
5800 case glslang::EOpSubgroupInclusiveXor:
5801 case glslang::EOpSubgroupExclusiveAdd:
5802 case glslang::EOpSubgroupExclusiveMul:
5803 case glslang::EOpSubgroupExclusiveMin:
5804 case glslang::EOpSubgroupExclusiveMax:
5805 case glslang::EOpSubgroupExclusiveAnd:
5806 case glslang::EOpSubgroupExclusiveOr:
5807 case glslang::EOpSubgroupExclusiveXor:
5808 case glslang::EOpSubgroupQuadSwapHorizontal:
5809 case glslang::EOpSubgroupQuadSwapVertical:
5810 case glslang::EOpSubgroupQuadSwapDiagonal: {
5811 std::vector<spv::Id> operands;
5812 operands.push_back(operand);
5813 return createSubgroupOperation(op, typeId, operands, typeProxy);
5814 }
5815 case glslang::EOpMbcnt:
5816 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot);
5817 libCall = spv::MbcntAMD;
5818 break;
5819
5820 case glslang::EOpCubeFaceIndex:
5821 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_gcn_shader);
5822 libCall = spv::CubeFaceIndexAMD;
5823 break;
5824
5825 case glslang::EOpCubeFaceCoord:
5826 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_gcn_shader);
5827 libCall = spv::CubeFaceCoordAMD;
5828 break;
5829 case glslang::EOpSubgroupPartition:
5830 unaryOp = spv::OpGroupNonUniformPartitionNV;
5831 break;
5832 case glslang::EOpConstructReference:
5833 unaryOp = spv::OpBitcast;
5834 break;
5835 #endif
5836
5837 case glslang::EOpCopyObject:
5838 unaryOp = spv::OpCopyObject;
5839 break;
5840
5841 default:
5842 return 0;
5843 }
5844
5845 spv::Id id;
5846 if (libCall >= 0) {
5847 std::vector<spv::Id> args;
5848 args.push_back(operand);
5849 id = builder.createBuiltinCall(typeId, extBuiltins >= 0 ? extBuiltins : stdBuiltins, libCall, args);
5850 } else {
5851 id = builder.createUnaryOp(unaryOp, typeId, operand);
5852 }
5853
5854 decorations.addNoContraction(builder, id);
5855 decorations.addNonUniform(builder, id);
5856 return builder.setPrecision(id, decorations.precision);
5857 }
5858
5859 // Create a unary operation on a matrix
createUnaryMatrixOperation(spv::Op op,OpDecorations & decorations,spv::Id typeId,spv::Id operand,glslang::TBasicType)5860 spv::Id TGlslangToSpvTraverser::createUnaryMatrixOperation(spv::Op op, OpDecorations& decorations, spv::Id typeId,
5861 spv::Id operand, glslang::TBasicType /* typeProxy */)
5862 {
5863 // Handle unary operations vector by vector.
5864 // The result type is the same type as the original type.
5865 // The algorithm is to:
5866 // - break the matrix into vectors
5867 // - apply the operation to each vector
5868 // - make a matrix out the vector results
5869
5870 // get the types sorted out
5871 int numCols = builder.getNumColumns(operand);
5872 int numRows = builder.getNumRows(operand);
5873 spv::Id srcVecType = builder.makeVectorType(builder.getScalarTypeId(builder.getTypeId(operand)), numRows);
5874 spv::Id destVecType = builder.makeVectorType(builder.getScalarTypeId(typeId), numRows);
5875 std::vector<spv::Id> results;
5876
5877 // do each vector op
5878 for (int c = 0; c < numCols; ++c) {
5879 std::vector<unsigned int> indexes;
5880 indexes.push_back(c);
5881 spv::Id srcVec = builder.createCompositeExtract(operand, srcVecType, indexes);
5882 spv::Id destVec = builder.createUnaryOp(op, destVecType, srcVec);
5883 decorations.addNoContraction(builder, destVec);
5884 decorations.addNonUniform(builder, destVec);
5885 results.push_back(builder.setPrecision(destVec, decorations.precision));
5886 }
5887
5888 // put the pieces together
5889 spv::Id result = builder.setPrecision(builder.createCompositeConstruct(typeId, results), decorations.precision);
5890 decorations.addNonUniform(builder, result);
5891 return result;
5892 }
5893
5894 // For converting integers where both the bitwidth and the signedness could
5895 // change, but only do the width change here. The caller is still responsible
5896 // for the signedness conversion.
createIntWidthConversion(glslang::TOperator op,spv::Id operand,int vectorSize)5897 spv::Id TGlslangToSpvTraverser::createIntWidthConversion(glslang::TOperator op, spv::Id operand, int vectorSize)
5898 {
5899 // Get the result type width, based on the type to convert to.
5900 int width = 32;
5901 switch(op) {
5902 case glslang::EOpConvInt16ToUint8:
5903 case glslang::EOpConvIntToUint8:
5904 case glslang::EOpConvInt64ToUint8:
5905 case glslang::EOpConvUint16ToInt8:
5906 case glslang::EOpConvUintToInt8:
5907 case glslang::EOpConvUint64ToInt8:
5908 width = 8;
5909 break;
5910 case glslang::EOpConvInt8ToUint16:
5911 case glslang::EOpConvIntToUint16:
5912 case glslang::EOpConvInt64ToUint16:
5913 case glslang::EOpConvUint8ToInt16:
5914 case glslang::EOpConvUintToInt16:
5915 case glslang::EOpConvUint64ToInt16:
5916 width = 16;
5917 break;
5918 case glslang::EOpConvInt8ToUint:
5919 case glslang::EOpConvInt16ToUint:
5920 case glslang::EOpConvInt64ToUint:
5921 case glslang::EOpConvUint8ToInt:
5922 case glslang::EOpConvUint16ToInt:
5923 case glslang::EOpConvUint64ToInt:
5924 width = 32;
5925 break;
5926 case glslang::EOpConvInt8ToUint64:
5927 case glslang::EOpConvInt16ToUint64:
5928 case glslang::EOpConvIntToUint64:
5929 case glslang::EOpConvUint8ToInt64:
5930 case glslang::EOpConvUint16ToInt64:
5931 case glslang::EOpConvUintToInt64:
5932 width = 64;
5933 break;
5934
5935 default:
5936 assert(false && "Default missing");
5937 break;
5938 }
5939
5940 // Get the conversion operation and result type,
5941 // based on the target width, but the source type.
5942 spv::Id type = spv::NoType;
5943 spv::Op convOp = spv::OpNop;
5944 switch(op) {
5945 case glslang::EOpConvInt8ToUint16:
5946 case glslang::EOpConvInt8ToUint:
5947 case glslang::EOpConvInt8ToUint64:
5948 case glslang::EOpConvInt16ToUint8:
5949 case glslang::EOpConvInt16ToUint:
5950 case glslang::EOpConvInt16ToUint64:
5951 case glslang::EOpConvIntToUint8:
5952 case glslang::EOpConvIntToUint16:
5953 case glslang::EOpConvIntToUint64:
5954 case glslang::EOpConvInt64ToUint8:
5955 case glslang::EOpConvInt64ToUint16:
5956 case glslang::EOpConvInt64ToUint:
5957 convOp = spv::OpSConvert;
5958 type = builder.makeIntType(width);
5959 break;
5960 default:
5961 convOp = spv::OpUConvert;
5962 type = builder.makeUintType(width);
5963 break;
5964 }
5965
5966 if (vectorSize > 0)
5967 type = builder.makeVectorType(type, vectorSize);
5968
5969 return builder.createUnaryOp(convOp, type, operand);
5970 }
5971
createConversion(glslang::TOperator op,OpDecorations & decorations,spv::Id destType,spv::Id operand,glslang::TBasicType typeProxy)5972 spv::Id TGlslangToSpvTraverser::createConversion(glslang::TOperator op, OpDecorations& decorations, spv::Id destType,
5973 spv::Id operand, glslang::TBasicType typeProxy)
5974 {
5975 spv::Op convOp = spv::OpNop;
5976 spv::Id zero = 0;
5977 spv::Id one = 0;
5978
5979 int vectorSize = builder.isVectorType(destType) ? builder.getNumTypeComponents(destType) : 0;
5980
5981 switch (op) {
5982 case glslang::EOpConvIntToBool:
5983 case glslang::EOpConvUintToBool:
5984 zero = builder.makeUintConstant(0);
5985 zero = makeSmearedConstant(zero, vectorSize);
5986 return builder.createBinOp(spv::OpINotEqual, destType, operand, zero);
5987 case glslang::EOpConvFloatToBool:
5988 zero = builder.makeFloatConstant(0.0F);
5989 zero = makeSmearedConstant(zero, vectorSize);
5990 return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero);
5991 case glslang::EOpConvBoolToFloat:
5992 convOp = spv::OpSelect;
5993 zero = builder.makeFloatConstant(0.0F);
5994 one = builder.makeFloatConstant(1.0F);
5995 break;
5996
5997 case glslang::EOpConvBoolToInt:
5998 case glslang::EOpConvBoolToInt64:
5999 #ifndef GLSLANG_WEB
6000 if (op == glslang::EOpConvBoolToInt64) {
6001 zero = builder.makeInt64Constant(0);
6002 one = builder.makeInt64Constant(1);
6003 } else
6004 #endif
6005 {
6006 zero = builder.makeIntConstant(0);
6007 one = builder.makeIntConstant(1);
6008 }
6009
6010 convOp = spv::OpSelect;
6011 break;
6012
6013 case glslang::EOpConvBoolToUint:
6014 case glslang::EOpConvBoolToUint64:
6015 #ifndef GLSLANG_WEB
6016 if (op == glslang::EOpConvBoolToUint64) {
6017 zero = builder.makeUint64Constant(0);
6018 one = builder.makeUint64Constant(1);
6019 } else
6020 #endif
6021 {
6022 zero = builder.makeUintConstant(0);
6023 one = builder.makeUintConstant(1);
6024 }
6025
6026 convOp = spv::OpSelect;
6027 break;
6028
6029 case glslang::EOpConvInt8ToFloat16:
6030 case glslang::EOpConvInt8ToFloat:
6031 case glslang::EOpConvInt8ToDouble:
6032 case glslang::EOpConvInt16ToFloat16:
6033 case glslang::EOpConvInt16ToFloat:
6034 case glslang::EOpConvInt16ToDouble:
6035 case glslang::EOpConvIntToFloat16:
6036 case glslang::EOpConvIntToFloat:
6037 case glslang::EOpConvIntToDouble:
6038 case glslang::EOpConvInt64ToFloat:
6039 case glslang::EOpConvInt64ToDouble:
6040 case glslang::EOpConvInt64ToFloat16:
6041 convOp = spv::OpConvertSToF;
6042 break;
6043
6044 case glslang::EOpConvUint8ToFloat16:
6045 case glslang::EOpConvUint8ToFloat:
6046 case glslang::EOpConvUint8ToDouble:
6047 case glslang::EOpConvUint16ToFloat16:
6048 case glslang::EOpConvUint16ToFloat:
6049 case glslang::EOpConvUint16ToDouble:
6050 case glslang::EOpConvUintToFloat16:
6051 case glslang::EOpConvUintToFloat:
6052 case glslang::EOpConvUintToDouble:
6053 case glslang::EOpConvUint64ToFloat:
6054 case glslang::EOpConvUint64ToDouble:
6055 case glslang::EOpConvUint64ToFloat16:
6056 convOp = spv::OpConvertUToF;
6057 break;
6058
6059 case glslang::EOpConvFloat16ToInt8:
6060 case glslang::EOpConvFloatToInt8:
6061 case glslang::EOpConvDoubleToInt8:
6062 case glslang::EOpConvFloat16ToInt16:
6063 case glslang::EOpConvFloatToInt16:
6064 case glslang::EOpConvDoubleToInt16:
6065 case glslang::EOpConvFloat16ToInt:
6066 case glslang::EOpConvFloatToInt:
6067 case glslang::EOpConvDoubleToInt:
6068 case glslang::EOpConvFloat16ToInt64:
6069 case glslang::EOpConvFloatToInt64:
6070 case glslang::EOpConvDoubleToInt64:
6071 convOp = spv::OpConvertFToS;
6072 break;
6073
6074 case glslang::EOpConvUint8ToInt8:
6075 case glslang::EOpConvInt8ToUint8:
6076 case glslang::EOpConvUint16ToInt16:
6077 case glslang::EOpConvInt16ToUint16:
6078 case glslang::EOpConvUintToInt:
6079 case glslang::EOpConvIntToUint:
6080 case glslang::EOpConvUint64ToInt64:
6081 case glslang::EOpConvInt64ToUint64:
6082 if (builder.isInSpecConstCodeGenMode()) {
6083 // Build zero scalar or vector for OpIAdd.
6084 #ifndef GLSLANG_WEB
6085 if(op == glslang::EOpConvUint8ToInt8 || op == glslang::EOpConvInt8ToUint8) {
6086 zero = builder.makeUint8Constant(0);
6087 } else if (op == glslang::EOpConvUint16ToInt16 || op == glslang::EOpConvInt16ToUint16) {
6088 zero = builder.makeUint16Constant(0);
6089 } else if (op == glslang::EOpConvUint64ToInt64 || op == glslang::EOpConvInt64ToUint64) {
6090 zero = builder.makeUint64Constant(0);
6091 } else
6092 #endif
6093 {
6094 zero = builder.makeUintConstant(0);
6095 }
6096 zero = makeSmearedConstant(zero, vectorSize);
6097 // Use OpIAdd, instead of OpBitcast to do the conversion when
6098 // generating for OpSpecConstantOp instruction.
6099 return builder.createBinOp(spv::OpIAdd, destType, operand, zero);
6100 }
6101 // For normal run-time conversion instruction, use OpBitcast.
6102 convOp = spv::OpBitcast;
6103 break;
6104
6105 case glslang::EOpConvFloat16ToUint8:
6106 case glslang::EOpConvFloatToUint8:
6107 case glslang::EOpConvDoubleToUint8:
6108 case glslang::EOpConvFloat16ToUint16:
6109 case glslang::EOpConvFloatToUint16:
6110 case glslang::EOpConvDoubleToUint16:
6111 case glslang::EOpConvFloat16ToUint:
6112 case glslang::EOpConvFloatToUint:
6113 case glslang::EOpConvDoubleToUint:
6114 case glslang::EOpConvFloatToUint64:
6115 case glslang::EOpConvDoubleToUint64:
6116 case glslang::EOpConvFloat16ToUint64:
6117 convOp = spv::OpConvertFToU;
6118 break;
6119
6120 #ifndef GLSLANG_WEB
6121 case glslang::EOpConvInt8ToBool:
6122 case glslang::EOpConvUint8ToBool:
6123 zero = builder.makeUint8Constant(0);
6124 zero = makeSmearedConstant(zero, vectorSize);
6125 return builder.createBinOp(spv::OpINotEqual, destType, operand, zero);
6126 case glslang::EOpConvInt16ToBool:
6127 case glslang::EOpConvUint16ToBool:
6128 zero = builder.makeUint16Constant(0);
6129 zero = makeSmearedConstant(zero, vectorSize);
6130 return builder.createBinOp(spv::OpINotEqual, destType, operand, zero);
6131 case glslang::EOpConvInt64ToBool:
6132 case glslang::EOpConvUint64ToBool:
6133 zero = builder.makeUint64Constant(0);
6134 zero = makeSmearedConstant(zero, vectorSize);
6135 return builder.createBinOp(spv::OpINotEqual, destType, operand, zero);
6136 case glslang::EOpConvDoubleToBool:
6137 zero = builder.makeDoubleConstant(0.0);
6138 zero = makeSmearedConstant(zero, vectorSize);
6139 return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero);
6140 case glslang::EOpConvFloat16ToBool:
6141 zero = builder.makeFloat16Constant(0.0F);
6142 zero = makeSmearedConstant(zero, vectorSize);
6143 return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero);
6144 case glslang::EOpConvBoolToDouble:
6145 convOp = spv::OpSelect;
6146 zero = builder.makeDoubleConstant(0.0);
6147 one = builder.makeDoubleConstant(1.0);
6148 break;
6149 case glslang::EOpConvBoolToFloat16:
6150 convOp = spv::OpSelect;
6151 zero = builder.makeFloat16Constant(0.0F);
6152 one = builder.makeFloat16Constant(1.0F);
6153 break;
6154 case glslang::EOpConvBoolToInt8:
6155 zero = builder.makeInt8Constant(0);
6156 one = builder.makeInt8Constant(1);
6157 convOp = spv::OpSelect;
6158 break;
6159 case glslang::EOpConvBoolToUint8:
6160 zero = builder.makeUint8Constant(0);
6161 one = builder.makeUint8Constant(1);
6162 convOp = spv::OpSelect;
6163 break;
6164 case glslang::EOpConvBoolToInt16:
6165 zero = builder.makeInt16Constant(0);
6166 one = builder.makeInt16Constant(1);
6167 convOp = spv::OpSelect;
6168 break;
6169 case glslang::EOpConvBoolToUint16:
6170 zero = builder.makeUint16Constant(0);
6171 one = builder.makeUint16Constant(1);
6172 convOp = spv::OpSelect;
6173 break;
6174 case glslang::EOpConvDoubleToFloat:
6175 case glslang::EOpConvFloatToDouble:
6176 case glslang::EOpConvDoubleToFloat16:
6177 case glslang::EOpConvFloat16ToDouble:
6178 case glslang::EOpConvFloatToFloat16:
6179 case glslang::EOpConvFloat16ToFloat:
6180 convOp = spv::OpFConvert;
6181 if (builder.isMatrixType(destType))
6182 return createUnaryMatrixOperation(convOp, decorations, destType, operand, typeProxy);
6183 break;
6184
6185 case glslang::EOpConvInt8ToInt16:
6186 case glslang::EOpConvInt8ToInt:
6187 case glslang::EOpConvInt8ToInt64:
6188 case glslang::EOpConvInt16ToInt8:
6189 case glslang::EOpConvInt16ToInt:
6190 case glslang::EOpConvInt16ToInt64:
6191 case glslang::EOpConvIntToInt8:
6192 case glslang::EOpConvIntToInt16:
6193 case glslang::EOpConvIntToInt64:
6194 case glslang::EOpConvInt64ToInt8:
6195 case glslang::EOpConvInt64ToInt16:
6196 case glslang::EOpConvInt64ToInt:
6197 convOp = spv::OpSConvert;
6198 break;
6199
6200 case glslang::EOpConvUint8ToUint16:
6201 case glslang::EOpConvUint8ToUint:
6202 case glslang::EOpConvUint8ToUint64:
6203 case glslang::EOpConvUint16ToUint8:
6204 case glslang::EOpConvUint16ToUint:
6205 case glslang::EOpConvUint16ToUint64:
6206 case glslang::EOpConvUintToUint8:
6207 case glslang::EOpConvUintToUint16:
6208 case glslang::EOpConvUintToUint64:
6209 case glslang::EOpConvUint64ToUint8:
6210 case glslang::EOpConvUint64ToUint16:
6211 case glslang::EOpConvUint64ToUint:
6212 convOp = spv::OpUConvert;
6213 break;
6214
6215 case glslang::EOpConvInt8ToUint16:
6216 case glslang::EOpConvInt8ToUint:
6217 case glslang::EOpConvInt8ToUint64:
6218 case glslang::EOpConvInt16ToUint8:
6219 case glslang::EOpConvInt16ToUint:
6220 case glslang::EOpConvInt16ToUint64:
6221 case glslang::EOpConvIntToUint8:
6222 case glslang::EOpConvIntToUint16:
6223 case glslang::EOpConvIntToUint64:
6224 case glslang::EOpConvInt64ToUint8:
6225 case glslang::EOpConvInt64ToUint16:
6226 case glslang::EOpConvInt64ToUint:
6227 case glslang::EOpConvUint8ToInt16:
6228 case glslang::EOpConvUint8ToInt:
6229 case glslang::EOpConvUint8ToInt64:
6230 case glslang::EOpConvUint16ToInt8:
6231 case glslang::EOpConvUint16ToInt:
6232 case glslang::EOpConvUint16ToInt64:
6233 case glslang::EOpConvUintToInt8:
6234 case glslang::EOpConvUintToInt16:
6235 case glslang::EOpConvUintToInt64:
6236 case glslang::EOpConvUint64ToInt8:
6237 case glslang::EOpConvUint64ToInt16:
6238 case glslang::EOpConvUint64ToInt:
6239 // OpSConvert/OpUConvert + OpBitCast
6240 operand = createIntWidthConversion(op, operand, vectorSize);
6241
6242 if (builder.isInSpecConstCodeGenMode()) {
6243 // Build zero scalar or vector for OpIAdd.
6244 switch(op) {
6245 case glslang::EOpConvInt16ToUint8:
6246 case glslang::EOpConvIntToUint8:
6247 case glslang::EOpConvInt64ToUint8:
6248 case glslang::EOpConvUint16ToInt8:
6249 case glslang::EOpConvUintToInt8:
6250 case glslang::EOpConvUint64ToInt8:
6251 zero = builder.makeUint8Constant(0);
6252 break;
6253 case glslang::EOpConvInt8ToUint16:
6254 case glslang::EOpConvIntToUint16:
6255 case glslang::EOpConvInt64ToUint16:
6256 case glslang::EOpConvUint8ToInt16:
6257 case glslang::EOpConvUintToInt16:
6258 case glslang::EOpConvUint64ToInt16:
6259 zero = builder.makeUint16Constant(0);
6260 break;
6261 case glslang::EOpConvInt8ToUint:
6262 case glslang::EOpConvInt16ToUint:
6263 case glslang::EOpConvInt64ToUint:
6264 case glslang::EOpConvUint8ToInt:
6265 case glslang::EOpConvUint16ToInt:
6266 case glslang::EOpConvUint64ToInt:
6267 zero = builder.makeUintConstant(0);
6268 break;
6269 case glslang::EOpConvInt8ToUint64:
6270 case glslang::EOpConvInt16ToUint64:
6271 case glslang::EOpConvIntToUint64:
6272 case glslang::EOpConvUint8ToInt64:
6273 case glslang::EOpConvUint16ToInt64:
6274 case glslang::EOpConvUintToInt64:
6275 zero = builder.makeUint64Constant(0);
6276 break;
6277 default:
6278 assert(false && "Default missing");
6279 break;
6280 }
6281 zero = makeSmearedConstant(zero, vectorSize);
6282 // Use OpIAdd, instead of OpBitcast to do the conversion when
6283 // generating for OpSpecConstantOp instruction.
6284 return builder.createBinOp(spv::OpIAdd, destType, operand, zero);
6285 }
6286 // For normal run-time conversion instruction, use OpBitcast.
6287 convOp = spv::OpBitcast;
6288 break;
6289 case glslang::EOpConvUint64ToPtr:
6290 convOp = spv::OpConvertUToPtr;
6291 break;
6292 case glslang::EOpConvPtrToUint64:
6293 convOp = spv::OpConvertPtrToU;
6294 break;
6295 case glslang::EOpConvPtrToUvec2:
6296 case glslang::EOpConvUvec2ToPtr:
6297 if (builder.isVector(operand))
6298 builder.promoteIncorporatedExtension(spv::E_SPV_EXT_physical_storage_buffer,
6299 spv::E_SPV_KHR_physical_storage_buffer, spv::Spv_1_5);
6300 convOp = spv::OpBitcast;
6301 break;
6302 #endif
6303
6304 default:
6305 break;
6306 }
6307
6308 spv::Id result = 0;
6309 if (convOp == spv::OpNop)
6310 return result;
6311
6312 if (convOp == spv::OpSelect) {
6313 zero = makeSmearedConstant(zero, vectorSize);
6314 one = makeSmearedConstant(one, vectorSize);
6315 result = builder.createTriOp(convOp, destType, operand, one, zero);
6316 } else
6317 result = builder.createUnaryOp(convOp, destType, operand);
6318
6319 result = builder.setPrecision(result, decorations.precision);
6320 decorations.addNonUniform(builder, result);
6321 return result;
6322 }
6323
makeSmearedConstant(spv::Id constant,int vectorSize)6324 spv::Id TGlslangToSpvTraverser::makeSmearedConstant(spv::Id constant, int vectorSize)
6325 {
6326 if (vectorSize == 0)
6327 return constant;
6328
6329 spv::Id vectorTypeId = builder.makeVectorType(builder.getTypeId(constant), vectorSize);
6330 std::vector<spv::Id> components;
6331 for (int c = 0; c < vectorSize; ++c)
6332 components.push_back(constant);
6333 return builder.makeCompositeConstant(vectorTypeId, components);
6334 }
6335
6336 // For glslang ops that map to SPV atomic opCodes
createAtomicOperation(glslang::TOperator op,spv::Decoration,spv::Id typeId,std::vector<spv::Id> & operands,glslang::TBasicType typeProxy,const spv::Builder::AccessChain::CoherentFlags & lvalueCoherentFlags)6337 spv::Id TGlslangToSpvTraverser::createAtomicOperation(glslang::TOperator op, spv::Decoration /*precision*/, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy, const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags)
6338 {
6339 spv::Op opCode = spv::OpNop;
6340
6341 switch (op) {
6342 case glslang::EOpAtomicAdd:
6343 case glslang::EOpImageAtomicAdd:
6344 case glslang::EOpAtomicCounterAdd:
6345 opCode = spv::OpAtomicIAdd;
6346 break;
6347 case glslang::EOpAtomicCounterSubtract:
6348 opCode = spv::OpAtomicISub;
6349 break;
6350 case glslang::EOpAtomicMin:
6351 case glslang::EOpImageAtomicMin:
6352 case glslang::EOpAtomicCounterMin:
6353 opCode = (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) ? spv::OpAtomicUMin : spv::OpAtomicSMin;
6354 break;
6355 case glslang::EOpAtomicMax:
6356 case glslang::EOpImageAtomicMax:
6357 case glslang::EOpAtomicCounterMax:
6358 opCode = (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) ? spv::OpAtomicUMax : spv::OpAtomicSMax;
6359 break;
6360 case glslang::EOpAtomicAnd:
6361 case glslang::EOpImageAtomicAnd:
6362 case glslang::EOpAtomicCounterAnd:
6363 opCode = spv::OpAtomicAnd;
6364 break;
6365 case glslang::EOpAtomicOr:
6366 case glslang::EOpImageAtomicOr:
6367 case glslang::EOpAtomicCounterOr:
6368 opCode = spv::OpAtomicOr;
6369 break;
6370 case glslang::EOpAtomicXor:
6371 case glslang::EOpImageAtomicXor:
6372 case glslang::EOpAtomicCounterXor:
6373 opCode = spv::OpAtomicXor;
6374 break;
6375 case glslang::EOpAtomicExchange:
6376 case glslang::EOpImageAtomicExchange:
6377 case glslang::EOpAtomicCounterExchange:
6378 opCode = spv::OpAtomicExchange;
6379 break;
6380 case glslang::EOpAtomicCompSwap:
6381 case glslang::EOpImageAtomicCompSwap:
6382 case glslang::EOpAtomicCounterCompSwap:
6383 opCode = spv::OpAtomicCompareExchange;
6384 break;
6385 case glslang::EOpAtomicCounterIncrement:
6386 opCode = spv::OpAtomicIIncrement;
6387 break;
6388 case glslang::EOpAtomicCounterDecrement:
6389 opCode = spv::OpAtomicIDecrement;
6390 break;
6391 case glslang::EOpAtomicCounter:
6392 case glslang::EOpImageAtomicLoad:
6393 case glslang::EOpAtomicLoad:
6394 opCode = spv::OpAtomicLoad;
6395 break;
6396 case glslang::EOpAtomicStore:
6397 case glslang::EOpImageAtomicStore:
6398 opCode = spv::OpAtomicStore;
6399 break;
6400 default:
6401 assert(0);
6402 break;
6403 }
6404
6405 if (typeProxy == glslang::EbtInt64 || typeProxy == glslang::EbtUint64)
6406 builder.addCapability(spv::CapabilityInt64Atomics);
6407
6408 // Sort out the operands
6409 // - mapping from glslang -> SPV
6410 // - there are extra SPV operands that are optional in glslang
6411 // - compare-exchange swaps the value and comparator
6412 // - compare-exchange has an extra memory semantics
6413 // - EOpAtomicCounterDecrement needs a post decrement
6414 spv::Id pointerId = 0, compareId = 0, valueId = 0;
6415 // scope defaults to Device in the old model, QueueFamilyKHR in the new model
6416 spv::Id scopeId;
6417 if (glslangIntermediate->usingVulkanMemoryModel()) {
6418 scopeId = builder.makeUintConstant(spv::ScopeQueueFamilyKHR);
6419 } else {
6420 scopeId = builder.makeUintConstant(spv::ScopeDevice);
6421 }
6422 // semantics default to relaxed
6423 spv::Id semanticsId = builder.makeUintConstant(lvalueCoherentFlags.isVolatile() && glslangIntermediate->usingVulkanMemoryModel() ?
6424 spv::MemorySemanticsVolatileMask :
6425 spv::MemorySemanticsMaskNone);
6426 spv::Id semanticsId2 = semanticsId;
6427
6428 pointerId = operands[0];
6429 if (opCode == spv::OpAtomicIIncrement || opCode == spv::OpAtomicIDecrement) {
6430 // no additional operands
6431 } else if (opCode == spv::OpAtomicCompareExchange) {
6432 compareId = operands[1];
6433 valueId = operands[2];
6434 if (operands.size() > 3) {
6435 scopeId = operands[3];
6436 semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[4]) | builder.getConstantScalar(operands[5]));
6437 semanticsId2 = builder.makeUintConstant(builder.getConstantScalar(operands[6]) | builder.getConstantScalar(operands[7]));
6438 }
6439 } else if (opCode == spv::OpAtomicLoad) {
6440 if (operands.size() > 1) {
6441 scopeId = operands[1];
6442 semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[2]) | builder.getConstantScalar(operands[3]));
6443 }
6444 } else {
6445 // atomic store or RMW
6446 valueId = operands[1];
6447 if (operands.size() > 2) {
6448 scopeId = operands[2];
6449 semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[3]) | builder.getConstantScalar(operands[4]));
6450 }
6451 }
6452
6453 // Check for capabilities
6454 unsigned semanticsImmediate = builder.getConstantScalar(semanticsId) | builder.getConstantScalar(semanticsId2);
6455 if (semanticsImmediate & (spv::MemorySemanticsMakeAvailableKHRMask |
6456 spv::MemorySemanticsMakeVisibleKHRMask |
6457 spv::MemorySemanticsOutputMemoryKHRMask |
6458 spv::MemorySemanticsVolatileMask)) {
6459 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
6460 }
6461
6462 if (glslangIntermediate->usingVulkanMemoryModel() && builder.getConstantScalar(scopeId) == spv::ScopeDevice) {
6463 builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR);
6464 }
6465
6466 std::vector<spv::Id> spvAtomicOperands; // hold the spv operands
6467 spvAtomicOperands.push_back(pointerId);
6468 spvAtomicOperands.push_back(scopeId);
6469 spvAtomicOperands.push_back(semanticsId);
6470 if (opCode == spv::OpAtomicCompareExchange) {
6471 spvAtomicOperands.push_back(semanticsId2);
6472 spvAtomicOperands.push_back(valueId);
6473 spvAtomicOperands.push_back(compareId);
6474 } else if (opCode != spv::OpAtomicLoad && opCode != spv::OpAtomicIIncrement && opCode != spv::OpAtomicIDecrement) {
6475 spvAtomicOperands.push_back(valueId);
6476 }
6477
6478 if (opCode == spv::OpAtomicStore) {
6479 builder.createNoResultOp(opCode, spvAtomicOperands);
6480 return 0;
6481 } else {
6482 spv::Id resultId = builder.createOp(opCode, typeId, spvAtomicOperands);
6483
6484 // GLSL and HLSL atomic-counter decrement return post-decrement value,
6485 // while SPIR-V returns pre-decrement value. Translate between these semantics.
6486 if (op == glslang::EOpAtomicCounterDecrement)
6487 resultId = builder.createBinOp(spv::OpISub, typeId, resultId, builder.makeIntConstant(1));
6488
6489 return resultId;
6490 }
6491 }
6492
6493 // Create group invocation operations.
createInvocationsOperation(glslang::TOperator op,spv::Id typeId,std::vector<spv::Id> & operands,glslang::TBasicType typeProxy)6494 spv::Id TGlslangToSpvTraverser::createInvocationsOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy)
6495 {
6496 bool isUnsigned = isTypeUnsignedInt(typeProxy);
6497 bool isFloat = isTypeFloat(typeProxy);
6498
6499 spv::Op opCode = spv::OpNop;
6500 std::vector<spv::IdImmediate> spvGroupOperands;
6501 spv::GroupOperation groupOperation = spv::GroupOperationMax;
6502
6503 if (op == glslang::EOpBallot || op == glslang::EOpReadFirstInvocation ||
6504 op == glslang::EOpReadInvocation) {
6505 builder.addExtension(spv::E_SPV_KHR_shader_ballot);
6506 builder.addCapability(spv::CapabilitySubgroupBallotKHR);
6507 } else if (op == glslang::EOpAnyInvocation ||
6508 op == glslang::EOpAllInvocations ||
6509 op == glslang::EOpAllInvocationsEqual) {
6510 builder.addExtension(spv::E_SPV_KHR_subgroup_vote);
6511 builder.addCapability(spv::CapabilitySubgroupVoteKHR);
6512 } else {
6513 builder.addCapability(spv::CapabilityGroups);
6514 if (op == glslang::EOpMinInvocationsNonUniform ||
6515 op == glslang::EOpMaxInvocationsNonUniform ||
6516 op == glslang::EOpAddInvocationsNonUniform ||
6517 op == glslang::EOpMinInvocationsInclusiveScanNonUniform ||
6518 op == glslang::EOpMaxInvocationsInclusiveScanNonUniform ||
6519 op == glslang::EOpAddInvocationsInclusiveScanNonUniform ||
6520 op == glslang::EOpMinInvocationsExclusiveScanNonUniform ||
6521 op == glslang::EOpMaxInvocationsExclusiveScanNonUniform ||
6522 op == glslang::EOpAddInvocationsExclusiveScanNonUniform)
6523 builder.addExtension(spv::E_SPV_AMD_shader_ballot);
6524
6525 switch (op) {
6526 case glslang::EOpMinInvocations:
6527 case glslang::EOpMaxInvocations:
6528 case glslang::EOpAddInvocations:
6529 case glslang::EOpMinInvocationsNonUniform:
6530 case glslang::EOpMaxInvocationsNonUniform:
6531 case glslang::EOpAddInvocationsNonUniform:
6532 groupOperation = spv::GroupOperationReduce;
6533 break;
6534 case glslang::EOpMinInvocationsInclusiveScan:
6535 case glslang::EOpMaxInvocationsInclusiveScan:
6536 case glslang::EOpAddInvocationsInclusiveScan:
6537 case glslang::EOpMinInvocationsInclusiveScanNonUniform:
6538 case glslang::EOpMaxInvocationsInclusiveScanNonUniform:
6539 case glslang::EOpAddInvocationsInclusiveScanNonUniform:
6540 groupOperation = spv::GroupOperationInclusiveScan;
6541 break;
6542 case glslang::EOpMinInvocationsExclusiveScan:
6543 case glslang::EOpMaxInvocationsExclusiveScan:
6544 case glslang::EOpAddInvocationsExclusiveScan:
6545 case glslang::EOpMinInvocationsExclusiveScanNonUniform:
6546 case glslang::EOpMaxInvocationsExclusiveScanNonUniform:
6547 case glslang::EOpAddInvocationsExclusiveScanNonUniform:
6548 groupOperation = spv::GroupOperationExclusiveScan;
6549 break;
6550 default:
6551 break;
6552 }
6553 spv::IdImmediate scope = { true, builder.makeUintConstant(spv::ScopeSubgroup) };
6554 spvGroupOperands.push_back(scope);
6555 if (groupOperation != spv::GroupOperationMax) {
6556 spv::IdImmediate groupOp = { false, (unsigned)groupOperation };
6557 spvGroupOperands.push_back(groupOp);
6558 }
6559 }
6560
6561 for (auto opIt = operands.begin(); opIt != operands.end(); ++opIt) {
6562 spv::IdImmediate op = { true, *opIt };
6563 spvGroupOperands.push_back(op);
6564 }
6565
6566 switch (op) {
6567 case glslang::EOpAnyInvocation:
6568 opCode = spv::OpSubgroupAnyKHR;
6569 break;
6570 case glslang::EOpAllInvocations:
6571 opCode = spv::OpSubgroupAllKHR;
6572 break;
6573 case glslang::EOpAllInvocationsEqual:
6574 opCode = spv::OpSubgroupAllEqualKHR;
6575 break;
6576 case glslang::EOpReadInvocation:
6577 opCode = spv::OpSubgroupReadInvocationKHR;
6578 if (builder.isVectorType(typeId))
6579 return CreateInvocationsVectorOperation(opCode, groupOperation, typeId, operands);
6580 break;
6581 case glslang::EOpReadFirstInvocation:
6582 opCode = spv::OpSubgroupFirstInvocationKHR;
6583 break;
6584 case glslang::EOpBallot:
6585 {
6586 // NOTE: According to the spec, the result type of "OpSubgroupBallotKHR" must be a 4 component vector of 32
6587 // bit integer types. The GLSL built-in function "ballotARB()" assumes the maximum number of invocations in
6588 // a subgroup is 64. Thus, we have to convert uvec4.xy to uint64_t as follow:
6589 //
6590 // result = Bitcast(SubgroupBallotKHR(Predicate).xy)
6591 //
6592 spv::Id uintType = builder.makeUintType(32);
6593 spv::Id uvec4Type = builder.makeVectorType(uintType, 4);
6594 spv::Id result = builder.createOp(spv::OpSubgroupBallotKHR, uvec4Type, spvGroupOperands);
6595
6596 std::vector<spv::Id> components;
6597 components.push_back(builder.createCompositeExtract(result, uintType, 0));
6598 components.push_back(builder.createCompositeExtract(result, uintType, 1));
6599
6600 spv::Id uvec2Type = builder.makeVectorType(uintType, 2);
6601 return builder.createUnaryOp(spv::OpBitcast, typeId,
6602 builder.createCompositeConstruct(uvec2Type, components));
6603 }
6604
6605 case glslang::EOpMinInvocations:
6606 case glslang::EOpMaxInvocations:
6607 case glslang::EOpAddInvocations:
6608 case glslang::EOpMinInvocationsInclusiveScan:
6609 case glslang::EOpMaxInvocationsInclusiveScan:
6610 case glslang::EOpAddInvocationsInclusiveScan:
6611 case glslang::EOpMinInvocationsExclusiveScan:
6612 case glslang::EOpMaxInvocationsExclusiveScan:
6613 case glslang::EOpAddInvocationsExclusiveScan:
6614 if (op == glslang::EOpMinInvocations ||
6615 op == glslang::EOpMinInvocationsInclusiveScan ||
6616 op == glslang::EOpMinInvocationsExclusiveScan) {
6617 if (isFloat)
6618 opCode = spv::OpGroupFMin;
6619 else {
6620 if (isUnsigned)
6621 opCode = spv::OpGroupUMin;
6622 else
6623 opCode = spv::OpGroupSMin;
6624 }
6625 } else if (op == glslang::EOpMaxInvocations ||
6626 op == glslang::EOpMaxInvocationsInclusiveScan ||
6627 op == glslang::EOpMaxInvocationsExclusiveScan) {
6628 if (isFloat)
6629 opCode = spv::OpGroupFMax;
6630 else {
6631 if (isUnsigned)
6632 opCode = spv::OpGroupUMax;
6633 else
6634 opCode = spv::OpGroupSMax;
6635 }
6636 } else {
6637 if (isFloat)
6638 opCode = spv::OpGroupFAdd;
6639 else
6640 opCode = spv::OpGroupIAdd;
6641 }
6642
6643 if (builder.isVectorType(typeId))
6644 return CreateInvocationsVectorOperation(opCode, groupOperation, typeId, operands);
6645
6646 break;
6647 case glslang::EOpMinInvocationsNonUniform:
6648 case glslang::EOpMaxInvocationsNonUniform:
6649 case glslang::EOpAddInvocationsNonUniform:
6650 case glslang::EOpMinInvocationsInclusiveScanNonUniform:
6651 case glslang::EOpMaxInvocationsInclusiveScanNonUniform:
6652 case glslang::EOpAddInvocationsInclusiveScanNonUniform:
6653 case glslang::EOpMinInvocationsExclusiveScanNonUniform:
6654 case glslang::EOpMaxInvocationsExclusiveScanNonUniform:
6655 case glslang::EOpAddInvocationsExclusiveScanNonUniform:
6656 if (op == glslang::EOpMinInvocationsNonUniform ||
6657 op == glslang::EOpMinInvocationsInclusiveScanNonUniform ||
6658 op == glslang::EOpMinInvocationsExclusiveScanNonUniform) {
6659 if (isFloat)
6660 opCode = spv::OpGroupFMinNonUniformAMD;
6661 else {
6662 if (isUnsigned)
6663 opCode = spv::OpGroupUMinNonUniformAMD;
6664 else
6665 opCode = spv::OpGroupSMinNonUniformAMD;
6666 }
6667 }
6668 else if (op == glslang::EOpMaxInvocationsNonUniform ||
6669 op == glslang::EOpMaxInvocationsInclusiveScanNonUniform ||
6670 op == glslang::EOpMaxInvocationsExclusiveScanNonUniform) {
6671 if (isFloat)
6672 opCode = spv::OpGroupFMaxNonUniformAMD;
6673 else {
6674 if (isUnsigned)
6675 opCode = spv::OpGroupUMaxNonUniformAMD;
6676 else
6677 opCode = spv::OpGroupSMaxNonUniformAMD;
6678 }
6679 }
6680 else {
6681 if (isFloat)
6682 opCode = spv::OpGroupFAddNonUniformAMD;
6683 else
6684 opCode = spv::OpGroupIAddNonUniformAMD;
6685 }
6686
6687 if (builder.isVectorType(typeId))
6688 return CreateInvocationsVectorOperation(opCode, groupOperation, typeId, operands);
6689
6690 break;
6691 default:
6692 logger->missingFunctionality("invocation operation");
6693 return spv::NoResult;
6694 }
6695
6696 assert(opCode != spv::OpNop);
6697 return builder.createOp(opCode, typeId, spvGroupOperands);
6698 }
6699
6700 // Create group invocation operations on a vector
CreateInvocationsVectorOperation(spv::Op op,spv::GroupOperation groupOperation,spv::Id typeId,std::vector<spv::Id> & operands)6701 spv::Id TGlslangToSpvTraverser::CreateInvocationsVectorOperation(spv::Op op, spv::GroupOperation groupOperation,
6702 spv::Id typeId, std::vector<spv::Id>& operands)
6703 {
6704 assert(op == spv::OpGroupFMin || op == spv::OpGroupUMin || op == spv::OpGroupSMin ||
6705 op == spv::OpGroupFMax || op == spv::OpGroupUMax || op == spv::OpGroupSMax ||
6706 op == spv::OpGroupFAdd || op == spv::OpGroupIAdd || op == spv::OpGroupBroadcast ||
6707 op == spv::OpSubgroupReadInvocationKHR ||
6708 op == spv::OpGroupFMinNonUniformAMD || op == spv::OpGroupUMinNonUniformAMD || op == spv::OpGroupSMinNonUniformAMD ||
6709 op == spv::OpGroupFMaxNonUniformAMD || op == spv::OpGroupUMaxNonUniformAMD || op == spv::OpGroupSMaxNonUniformAMD ||
6710 op == spv::OpGroupFAddNonUniformAMD || op == spv::OpGroupIAddNonUniformAMD);
6711
6712 // Handle group invocation operations scalar by scalar.
6713 // The result type is the same type as the original type.
6714 // The algorithm is to:
6715 // - break the vector into scalars
6716 // - apply the operation to each scalar
6717 // - make a vector out the scalar results
6718
6719 // get the types sorted out
6720 int numComponents = builder.getNumComponents(operands[0]);
6721 spv::Id scalarType = builder.getScalarTypeId(builder.getTypeId(operands[0]));
6722 std::vector<spv::Id> results;
6723
6724 // do each scalar op
6725 for (int comp = 0; comp < numComponents; ++comp) {
6726 std::vector<unsigned int> indexes;
6727 indexes.push_back(comp);
6728 spv::IdImmediate scalar = { true, builder.createCompositeExtract(operands[0], scalarType, indexes) };
6729 std::vector<spv::IdImmediate> spvGroupOperands;
6730 if (op == spv::OpSubgroupReadInvocationKHR) {
6731 spvGroupOperands.push_back(scalar);
6732 spv::IdImmediate operand = { true, operands[1] };
6733 spvGroupOperands.push_back(operand);
6734 } else if (op == spv::OpGroupBroadcast) {
6735 spv::IdImmediate scope = { true, builder.makeUintConstant(spv::ScopeSubgroup) };
6736 spvGroupOperands.push_back(scope);
6737 spvGroupOperands.push_back(scalar);
6738 spv::IdImmediate operand = { true, operands[1] };
6739 spvGroupOperands.push_back(operand);
6740 } else {
6741 spv::IdImmediate scope = { true, builder.makeUintConstant(spv::ScopeSubgroup) };
6742 spvGroupOperands.push_back(scope);
6743 spv::IdImmediate groupOp = { false, (unsigned)groupOperation };
6744 spvGroupOperands.push_back(groupOp);
6745 spvGroupOperands.push_back(scalar);
6746 }
6747
6748 results.push_back(builder.createOp(op, scalarType, spvGroupOperands));
6749 }
6750
6751 // put the pieces together
6752 return builder.createCompositeConstruct(typeId, results);
6753 }
6754
6755 // Create subgroup invocation operations.
createSubgroupOperation(glslang::TOperator op,spv::Id typeId,std::vector<spv::Id> & operands,glslang::TBasicType typeProxy)6756 spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, spv::Id typeId,
6757 std::vector<spv::Id>& operands, glslang::TBasicType typeProxy)
6758 {
6759 // Add the required capabilities.
6760 switch (op) {
6761 case glslang::EOpSubgroupElect:
6762 builder.addCapability(spv::CapabilityGroupNonUniform);
6763 break;
6764 case glslang::EOpSubgroupAll:
6765 case glslang::EOpSubgroupAny:
6766 case glslang::EOpSubgroupAllEqual:
6767 builder.addCapability(spv::CapabilityGroupNonUniform);
6768 builder.addCapability(spv::CapabilityGroupNonUniformVote);
6769 break;
6770 case glslang::EOpSubgroupBroadcast:
6771 case glslang::EOpSubgroupBroadcastFirst:
6772 case glslang::EOpSubgroupBallot:
6773 case glslang::EOpSubgroupInverseBallot:
6774 case glslang::EOpSubgroupBallotBitExtract:
6775 case glslang::EOpSubgroupBallotBitCount:
6776 case glslang::EOpSubgroupBallotInclusiveBitCount:
6777 case glslang::EOpSubgroupBallotExclusiveBitCount:
6778 case glslang::EOpSubgroupBallotFindLSB:
6779 case glslang::EOpSubgroupBallotFindMSB:
6780 builder.addCapability(spv::CapabilityGroupNonUniform);
6781 builder.addCapability(spv::CapabilityGroupNonUniformBallot);
6782 break;
6783 case glslang::EOpSubgroupShuffle:
6784 case glslang::EOpSubgroupShuffleXor:
6785 builder.addCapability(spv::CapabilityGroupNonUniform);
6786 builder.addCapability(spv::CapabilityGroupNonUniformShuffle);
6787 break;
6788 case glslang::EOpSubgroupShuffleUp:
6789 case glslang::EOpSubgroupShuffleDown:
6790 builder.addCapability(spv::CapabilityGroupNonUniform);
6791 builder.addCapability(spv::CapabilityGroupNonUniformShuffleRelative);
6792 break;
6793 case glslang::EOpSubgroupAdd:
6794 case glslang::EOpSubgroupMul:
6795 case glslang::EOpSubgroupMin:
6796 case glslang::EOpSubgroupMax:
6797 case glslang::EOpSubgroupAnd:
6798 case glslang::EOpSubgroupOr:
6799 case glslang::EOpSubgroupXor:
6800 case glslang::EOpSubgroupInclusiveAdd:
6801 case glslang::EOpSubgroupInclusiveMul:
6802 case glslang::EOpSubgroupInclusiveMin:
6803 case glslang::EOpSubgroupInclusiveMax:
6804 case glslang::EOpSubgroupInclusiveAnd:
6805 case glslang::EOpSubgroupInclusiveOr:
6806 case glslang::EOpSubgroupInclusiveXor:
6807 case glslang::EOpSubgroupExclusiveAdd:
6808 case glslang::EOpSubgroupExclusiveMul:
6809 case glslang::EOpSubgroupExclusiveMin:
6810 case glslang::EOpSubgroupExclusiveMax:
6811 case glslang::EOpSubgroupExclusiveAnd:
6812 case glslang::EOpSubgroupExclusiveOr:
6813 case glslang::EOpSubgroupExclusiveXor:
6814 builder.addCapability(spv::CapabilityGroupNonUniform);
6815 builder.addCapability(spv::CapabilityGroupNonUniformArithmetic);
6816 break;
6817 case glslang::EOpSubgroupClusteredAdd:
6818 case glslang::EOpSubgroupClusteredMul:
6819 case glslang::EOpSubgroupClusteredMin:
6820 case glslang::EOpSubgroupClusteredMax:
6821 case glslang::EOpSubgroupClusteredAnd:
6822 case glslang::EOpSubgroupClusteredOr:
6823 case glslang::EOpSubgroupClusteredXor:
6824 builder.addCapability(spv::CapabilityGroupNonUniform);
6825 builder.addCapability(spv::CapabilityGroupNonUniformClustered);
6826 break;
6827 case glslang::EOpSubgroupQuadBroadcast:
6828 case glslang::EOpSubgroupQuadSwapHorizontal:
6829 case glslang::EOpSubgroupQuadSwapVertical:
6830 case glslang::EOpSubgroupQuadSwapDiagonal:
6831 builder.addCapability(spv::CapabilityGroupNonUniform);
6832 builder.addCapability(spv::CapabilityGroupNonUniformQuad);
6833 break;
6834 case glslang::EOpSubgroupPartitionedAdd:
6835 case glslang::EOpSubgroupPartitionedMul:
6836 case glslang::EOpSubgroupPartitionedMin:
6837 case glslang::EOpSubgroupPartitionedMax:
6838 case glslang::EOpSubgroupPartitionedAnd:
6839 case glslang::EOpSubgroupPartitionedOr:
6840 case glslang::EOpSubgroupPartitionedXor:
6841 case glslang::EOpSubgroupPartitionedInclusiveAdd:
6842 case glslang::EOpSubgroupPartitionedInclusiveMul:
6843 case glslang::EOpSubgroupPartitionedInclusiveMin:
6844 case glslang::EOpSubgroupPartitionedInclusiveMax:
6845 case glslang::EOpSubgroupPartitionedInclusiveAnd:
6846 case glslang::EOpSubgroupPartitionedInclusiveOr:
6847 case glslang::EOpSubgroupPartitionedInclusiveXor:
6848 case glslang::EOpSubgroupPartitionedExclusiveAdd:
6849 case glslang::EOpSubgroupPartitionedExclusiveMul:
6850 case glslang::EOpSubgroupPartitionedExclusiveMin:
6851 case glslang::EOpSubgroupPartitionedExclusiveMax:
6852 case glslang::EOpSubgroupPartitionedExclusiveAnd:
6853 case glslang::EOpSubgroupPartitionedExclusiveOr:
6854 case glslang::EOpSubgroupPartitionedExclusiveXor:
6855 builder.addExtension(spv::E_SPV_NV_shader_subgroup_partitioned);
6856 builder.addCapability(spv::CapabilityGroupNonUniformPartitionedNV);
6857 break;
6858 default: assert(0 && "Unhandled subgroup operation!");
6859 }
6860
6861
6862 const bool isUnsigned = isTypeUnsignedInt(typeProxy);
6863 const bool isFloat = isTypeFloat(typeProxy);
6864 const bool isBool = typeProxy == glslang::EbtBool;
6865
6866 spv::Op opCode = spv::OpNop;
6867
6868 // Figure out which opcode to use.
6869 switch (op) {
6870 case glslang::EOpSubgroupElect: opCode = spv::OpGroupNonUniformElect; break;
6871 case glslang::EOpSubgroupAll: opCode = spv::OpGroupNonUniformAll; break;
6872 case glslang::EOpSubgroupAny: opCode = spv::OpGroupNonUniformAny; break;
6873 case glslang::EOpSubgroupAllEqual: opCode = spv::OpGroupNonUniformAllEqual; break;
6874 case glslang::EOpSubgroupBroadcast: opCode = spv::OpGroupNonUniformBroadcast; break;
6875 case glslang::EOpSubgroupBroadcastFirst: opCode = spv::OpGroupNonUniformBroadcastFirst; break;
6876 case glslang::EOpSubgroupBallot: opCode = spv::OpGroupNonUniformBallot; break;
6877 case glslang::EOpSubgroupInverseBallot: opCode = spv::OpGroupNonUniformInverseBallot; break;
6878 case glslang::EOpSubgroupBallotBitExtract: opCode = spv::OpGroupNonUniformBallotBitExtract; break;
6879 case glslang::EOpSubgroupBallotBitCount:
6880 case glslang::EOpSubgroupBallotInclusiveBitCount:
6881 case glslang::EOpSubgroupBallotExclusiveBitCount: opCode = spv::OpGroupNonUniformBallotBitCount; break;
6882 case glslang::EOpSubgroupBallotFindLSB: opCode = spv::OpGroupNonUniformBallotFindLSB; break;
6883 case glslang::EOpSubgroupBallotFindMSB: opCode = spv::OpGroupNonUniformBallotFindMSB; break;
6884 case glslang::EOpSubgroupShuffle: opCode = spv::OpGroupNonUniformShuffle; break;
6885 case glslang::EOpSubgroupShuffleXor: opCode = spv::OpGroupNonUniformShuffleXor; break;
6886 case glslang::EOpSubgroupShuffleUp: opCode = spv::OpGroupNonUniformShuffleUp; break;
6887 case glslang::EOpSubgroupShuffleDown: opCode = spv::OpGroupNonUniformShuffleDown; break;
6888 case glslang::EOpSubgroupAdd:
6889 case glslang::EOpSubgroupInclusiveAdd:
6890 case glslang::EOpSubgroupExclusiveAdd:
6891 case glslang::EOpSubgroupClusteredAdd:
6892 case glslang::EOpSubgroupPartitionedAdd:
6893 case glslang::EOpSubgroupPartitionedInclusiveAdd:
6894 case glslang::EOpSubgroupPartitionedExclusiveAdd:
6895 if (isFloat) {
6896 opCode = spv::OpGroupNonUniformFAdd;
6897 } else {
6898 opCode = spv::OpGroupNonUniformIAdd;
6899 }
6900 break;
6901 case glslang::EOpSubgroupMul:
6902 case glslang::EOpSubgroupInclusiveMul:
6903 case glslang::EOpSubgroupExclusiveMul:
6904 case glslang::EOpSubgroupClusteredMul:
6905 case glslang::EOpSubgroupPartitionedMul:
6906 case glslang::EOpSubgroupPartitionedInclusiveMul:
6907 case glslang::EOpSubgroupPartitionedExclusiveMul:
6908 if (isFloat) {
6909 opCode = spv::OpGroupNonUniformFMul;
6910 } else {
6911 opCode = spv::OpGroupNonUniformIMul;
6912 }
6913 break;
6914 case glslang::EOpSubgroupMin:
6915 case glslang::EOpSubgroupInclusiveMin:
6916 case glslang::EOpSubgroupExclusiveMin:
6917 case glslang::EOpSubgroupClusteredMin:
6918 case glslang::EOpSubgroupPartitionedMin:
6919 case glslang::EOpSubgroupPartitionedInclusiveMin:
6920 case glslang::EOpSubgroupPartitionedExclusiveMin:
6921 if (isFloat) {
6922 opCode = spv::OpGroupNonUniformFMin;
6923 } else if (isUnsigned) {
6924 opCode = spv::OpGroupNonUniformUMin;
6925 } else {
6926 opCode = spv::OpGroupNonUniformSMin;
6927 }
6928 break;
6929 case glslang::EOpSubgroupMax:
6930 case glslang::EOpSubgroupInclusiveMax:
6931 case glslang::EOpSubgroupExclusiveMax:
6932 case glslang::EOpSubgroupClusteredMax:
6933 case glslang::EOpSubgroupPartitionedMax:
6934 case glslang::EOpSubgroupPartitionedInclusiveMax:
6935 case glslang::EOpSubgroupPartitionedExclusiveMax:
6936 if (isFloat) {
6937 opCode = spv::OpGroupNonUniformFMax;
6938 } else if (isUnsigned) {
6939 opCode = spv::OpGroupNonUniformUMax;
6940 } else {
6941 opCode = spv::OpGroupNonUniformSMax;
6942 }
6943 break;
6944 case glslang::EOpSubgroupAnd:
6945 case glslang::EOpSubgroupInclusiveAnd:
6946 case glslang::EOpSubgroupExclusiveAnd:
6947 case glslang::EOpSubgroupClusteredAnd:
6948 case glslang::EOpSubgroupPartitionedAnd:
6949 case glslang::EOpSubgroupPartitionedInclusiveAnd:
6950 case glslang::EOpSubgroupPartitionedExclusiveAnd:
6951 if (isBool) {
6952 opCode = spv::OpGroupNonUniformLogicalAnd;
6953 } else {
6954 opCode = spv::OpGroupNonUniformBitwiseAnd;
6955 }
6956 break;
6957 case glslang::EOpSubgroupOr:
6958 case glslang::EOpSubgroupInclusiveOr:
6959 case glslang::EOpSubgroupExclusiveOr:
6960 case glslang::EOpSubgroupClusteredOr:
6961 case glslang::EOpSubgroupPartitionedOr:
6962 case glslang::EOpSubgroupPartitionedInclusiveOr:
6963 case glslang::EOpSubgroupPartitionedExclusiveOr:
6964 if (isBool) {
6965 opCode = spv::OpGroupNonUniformLogicalOr;
6966 } else {
6967 opCode = spv::OpGroupNonUniformBitwiseOr;
6968 }
6969 break;
6970 case glslang::EOpSubgroupXor:
6971 case glslang::EOpSubgroupInclusiveXor:
6972 case glslang::EOpSubgroupExclusiveXor:
6973 case glslang::EOpSubgroupClusteredXor:
6974 case glslang::EOpSubgroupPartitionedXor:
6975 case glslang::EOpSubgroupPartitionedInclusiveXor:
6976 case glslang::EOpSubgroupPartitionedExclusiveXor:
6977 if (isBool) {
6978 opCode = spv::OpGroupNonUniformLogicalXor;
6979 } else {
6980 opCode = spv::OpGroupNonUniformBitwiseXor;
6981 }
6982 break;
6983 case glslang::EOpSubgroupQuadBroadcast: opCode = spv::OpGroupNonUniformQuadBroadcast; break;
6984 case glslang::EOpSubgroupQuadSwapHorizontal:
6985 case glslang::EOpSubgroupQuadSwapVertical:
6986 case glslang::EOpSubgroupQuadSwapDiagonal: opCode = spv::OpGroupNonUniformQuadSwap; break;
6987 default: assert(0 && "Unhandled subgroup operation!");
6988 }
6989
6990 // get the right Group Operation
6991 spv::GroupOperation groupOperation = spv::GroupOperationMax;
6992 switch (op) {
6993 default:
6994 break;
6995 case glslang::EOpSubgroupBallotBitCount:
6996 case glslang::EOpSubgroupAdd:
6997 case glslang::EOpSubgroupMul:
6998 case glslang::EOpSubgroupMin:
6999 case glslang::EOpSubgroupMax:
7000 case glslang::EOpSubgroupAnd:
7001 case glslang::EOpSubgroupOr:
7002 case glslang::EOpSubgroupXor:
7003 groupOperation = spv::GroupOperationReduce;
7004 break;
7005 case glslang::EOpSubgroupBallotInclusiveBitCount:
7006 case glslang::EOpSubgroupInclusiveAdd:
7007 case glslang::EOpSubgroupInclusiveMul:
7008 case glslang::EOpSubgroupInclusiveMin:
7009 case glslang::EOpSubgroupInclusiveMax:
7010 case glslang::EOpSubgroupInclusiveAnd:
7011 case glslang::EOpSubgroupInclusiveOr:
7012 case glslang::EOpSubgroupInclusiveXor:
7013 groupOperation = spv::GroupOperationInclusiveScan;
7014 break;
7015 case glslang::EOpSubgroupBallotExclusiveBitCount:
7016 case glslang::EOpSubgroupExclusiveAdd:
7017 case glslang::EOpSubgroupExclusiveMul:
7018 case glslang::EOpSubgroupExclusiveMin:
7019 case glslang::EOpSubgroupExclusiveMax:
7020 case glslang::EOpSubgroupExclusiveAnd:
7021 case glslang::EOpSubgroupExclusiveOr:
7022 case glslang::EOpSubgroupExclusiveXor:
7023 groupOperation = spv::GroupOperationExclusiveScan;
7024 break;
7025 case glslang::EOpSubgroupClusteredAdd:
7026 case glslang::EOpSubgroupClusteredMul:
7027 case glslang::EOpSubgroupClusteredMin:
7028 case glslang::EOpSubgroupClusteredMax:
7029 case glslang::EOpSubgroupClusteredAnd:
7030 case glslang::EOpSubgroupClusteredOr:
7031 case glslang::EOpSubgroupClusteredXor:
7032 groupOperation = spv::GroupOperationClusteredReduce;
7033 break;
7034 case glslang::EOpSubgroupPartitionedAdd:
7035 case glslang::EOpSubgroupPartitionedMul:
7036 case glslang::EOpSubgroupPartitionedMin:
7037 case glslang::EOpSubgroupPartitionedMax:
7038 case glslang::EOpSubgroupPartitionedAnd:
7039 case glslang::EOpSubgroupPartitionedOr:
7040 case glslang::EOpSubgroupPartitionedXor:
7041 groupOperation = spv::GroupOperationPartitionedReduceNV;
7042 break;
7043 case glslang::EOpSubgroupPartitionedInclusiveAdd:
7044 case glslang::EOpSubgroupPartitionedInclusiveMul:
7045 case glslang::EOpSubgroupPartitionedInclusiveMin:
7046 case glslang::EOpSubgroupPartitionedInclusiveMax:
7047 case glslang::EOpSubgroupPartitionedInclusiveAnd:
7048 case glslang::EOpSubgroupPartitionedInclusiveOr:
7049 case glslang::EOpSubgroupPartitionedInclusiveXor:
7050 groupOperation = spv::GroupOperationPartitionedInclusiveScanNV;
7051 break;
7052 case glslang::EOpSubgroupPartitionedExclusiveAdd:
7053 case glslang::EOpSubgroupPartitionedExclusiveMul:
7054 case glslang::EOpSubgroupPartitionedExclusiveMin:
7055 case glslang::EOpSubgroupPartitionedExclusiveMax:
7056 case glslang::EOpSubgroupPartitionedExclusiveAnd:
7057 case glslang::EOpSubgroupPartitionedExclusiveOr:
7058 case glslang::EOpSubgroupPartitionedExclusiveXor:
7059 groupOperation = spv::GroupOperationPartitionedExclusiveScanNV;
7060 break;
7061 }
7062
7063 // build the instruction
7064 std::vector<spv::IdImmediate> spvGroupOperands;
7065
7066 // Every operation begins with the Execution Scope operand.
7067 spv::IdImmediate executionScope = { true, builder.makeUintConstant(spv::ScopeSubgroup) };
7068 spvGroupOperands.push_back(executionScope);
7069
7070 // Next, for all operations that use a Group Operation, push that as an operand.
7071 if (groupOperation != spv::GroupOperationMax) {
7072 spv::IdImmediate groupOperand = { false, (unsigned)groupOperation };
7073 spvGroupOperands.push_back(groupOperand);
7074 }
7075
7076 // Push back the operands next.
7077 for (auto opIt = operands.cbegin(); opIt != operands.cend(); ++opIt) {
7078 spv::IdImmediate operand = { true, *opIt };
7079 spvGroupOperands.push_back(operand);
7080 }
7081
7082 // Some opcodes have additional operands.
7083 spv::Id directionId = spv::NoResult;
7084 switch (op) {
7085 default: break;
7086 case glslang::EOpSubgroupQuadSwapHorizontal: directionId = builder.makeUintConstant(0); break;
7087 case glslang::EOpSubgroupQuadSwapVertical: directionId = builder.makeUintConstant(1); break;
7088 case glslang::EOpSubgroupQuadSwapDiagonal: directionId = builder.makeUintConstant(2); break;
7089 }
7090 if (directionId != spv::NoResult) {
7091 spv::IdImmediate direction = { true, directionId };
7092 spvGroupOperands.push_back(direction);
7093 }
7094
7095 return builder.createOp(opCode, typeId, spvGroupOperands);
7096 }
7097
createMiscOperation(glslang::TOperator op,spv::Decoration precision,spv::Id typeId,std::vector<spv::Id> & operands,glslang::TBasicType typeProxy)7098 spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy)
7099 {
7100 bool isUnsigned = isTypeUnsignedInt(typeProxy);
7101 bool isFloat = isTypeFloat(typeProxy);
7102
7103 spv::Op opCode = spv::OpNop;
7104 int extBuiltins = -1;
7105 int libCall = -1;
7106 size_t consumedOperands = operands.size();
7107 spv::Id typeId0 = 0;
7108 if (consumedOperands > 0)
7109 typeId0 = builder.getTypeId(operands[0]);
7110 spv::Id typeId1 = 0;
7111 if (consumedOperands > 1)
7112 typeId1 = builder.getTypeId(operands[1]);
7113 spv::Id frexpIntType = 0;
7114
7115 switch (op) {
7116 case glslang::EOpMin:
7117 if (isFloat)
7118 libCall = nanMinMaxClamp ? spv::GLSLstd450NMin : spv::GLSLstd450FMin;
7119 else if (isUnsigned)
7120 libCall = spv::GLSLstd450UMin;
7121 else
7122 libCall = spv::GLSLstd450SMin;
7123 builder.promoteScalar(precision, operands.front(), operands.back());
7124 break;
7125 case glslang::EOpModf:
7126 libCall = spv::GLSLstd450Modf;
7127 break;
7128 case glslang::EOpMax:
7129 if (isFloat)
7130 libCall = nanMinMaxClamp ? spv::GLSLstd450NMax : spv::GLSLstd450FMax;
7131 else if (isUnsigned)
7132 libCall = spv::GLSLstd450UMax;
7133 else
7134 libCall = spv::GLSLstd450SMax;
7135 builder.promoteScalar(precision, operands.front(), operands.back());
7136 break;
7137 case glslang::EOpPow:
7138 libCall = spv::GLSLstd450Pow;
7139 break;
7140 case glslang::EOpDot:
7141 opCode = spv::OpDot;
7142 break;
7143 case glslang::EOpAtan:
7144 libCall = spv::GLSLstd450Atan2;
7145 break;
7146
7147 case glslang::EOpClamp:
7148 if (isFloat)
7149 libCall = nanMinMaxClamp ? spv::GLSLstd450NClamp : spv::GLSLstd450FClamp;
7150 else if (isUnsigned)
7151 libCall = spv::GLSLstd450UClamp;
7152 else
7153 libCall = spv::GLSLstd450SClamp;
7154 builder.promoteScalar(precision, operands.front(), operands[1]);
7155 builder.promoteScalar(precision, operands.front(), operands[2]);
7156 break;
7157 case glslang::EOpMix:
7158 if (! builder.isBoolType(builder.getScalarTypeId(builder.getTypeId(operands.back())))) {
7159 assert(isFloat);
7160 libCall = spv::GLSLstd450FMix;
7161 } else {
7162 opCode = spv::OpSelect;
7163 std::swap(operands.front(), operands.back());
7164 }
7165 builder.promoteScalar(precision, operands.front(), operands.back());
7166 break;
7167 case glslang::EOpStep:
7168 libCall = spv::GLSLstd450Step;
7169 builder.promoteScalar(precision, operands.front(), operands.back());
7170 break;
7171 case glslang::EOpSmoothStep:
7172 libCall = spv::GLSLstd450SmoothStep;
7173 builder.promoteScalar(precision, operands[0], operands[2]);
7174 builder.promoteScalar(precision, operands[1], operands[2]);
7175 break;
7176
7177 case glslang::EOpDistance:
7178 libCall = spv::GLSLstd450Distance;
7179 break;
7180 case glslang::EOpCross:
7181 libCall = spv::GLSLstd450Cross;
7182 break;
7183 case glslang::EOpFaceForward:
7184 libCall = spv::GLSLstd450FaceForward;
7185 break;
7186 case glslang::EOpReflect:
7187 libCall = spv::GLSLstd450Reflect;
7188 break;
7189 case glslang::EOpRefract:
7190 libCall = spv::GLSLstd450Refract;
7191 break;
7192 case glslang::EOpBarrier:
7193 {
7194 // This is for the extended controlBarrier function, with four operands.
7195 // The unextended barrier() goes through createNoArgOperation.
7196 assert(operands.size() == 4);
7197 unsigned int executionScope = builder.getConstantScalar(operands[0]);
7198 unsigned int memoryScope = builder.getConstantScalar(operands[1]);
7199 unsigned int semantics = builder.getConstantScalar(operands[2]) | builder.getConstantScalar(operands[3]);
7200 builder.createControlBarrier((spv::Scope)executionScope, (spv::Scope)memoryScope, (spv::MemorySemanticsMask)semantics);
7201 if (semantics & (spv::MemorySemanticsMakeAvailableKHRMask |
7202 spv::MemorySemanticsMakeVisibleKHRMask |
7203 spv::MemorySemanticsOutputMemoryKHRMask |
7204 spv::MemorySemanticsVolatileMask)) {
7205 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
7206 }
7207 if (glslangIntermediate->usingVulkanMemoryModel() && (executionScope == spv::ScopeDevice || memoryScope == spv::ScopeDevice)) {
7208 builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR);
7209 }
7210 return 0;
7211 }
7212 break;
7213 case glslang::EOpMemoryBarrier:
7214 {
7215 // This is for the extended memoryBarrier function, with three operands.
7216 // The unextended memoryBarrier() goes through createNoArgOperation.
7217 assert(operands.size() == 3);
7218 unsigned int memoryScope = builder.getConstantScalar(operands[0]);
7219 unsigned int semantics = builder.getConstantScalar(operands[1]) | builder.getConstantScalar(operands[2]);
7220 builder.createMemoryBarrier((spv::Scope)memoryScope, (spv::MemorySemanticsMask)semantics);
7221 if (semantics & (spv::MemorySemanticsMakeAvailableKHRMask |
7222 spv::MemorySemanticsMakeVisibleKHRMask |
7223 spv::MemorySemanticsOutputMemoryKHRMask |
7224 spv::MemorySemanticsVolatileMask)) {
7225 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
7226 }
7227 if (glslangIntermediate->usingVulkanMemoryModel() && memoryScope == spv::ScopeDevice) {
7228 builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR);
7229 }
7230 return 0;
7231 }
7232 break;
7233
7234 #ifndef GLSLANG_WEB
7235 case glslang::EOpInterpolateAtSample:
7236 if (typeProxy == glslang::EbtFloat16)
7237 builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float);
7238 libCall = spv::GLSLstd450InterpolateAtSample;
7239 break;
7240 case glslang::EOpInterpolateAtOffset:
7241 if (typeProxy == glslang::EbtFloat16)
7242 builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float);
7243 libCall = spv::GLSLstd450InterpolateAtOffset;
7244 break;
7245 case glslang::EOpAddCarry:
7246 opCode = spv::OpIAddCarry;
7247 typeId = builder.makeStructResultType(typeId0, typeId0);
7248 consumedOperands = 2;
7249 break;
7250 case glslang::EOpSubBorrow:
7251 opCode = spv::OpISubBorrow;
7252 typeId = builder.makeStructResultType(typeId0, typeId0);
7253 consumedOperands = 2;
7254 break;
7255 case glslang::EOpUMulExtended:
7256 opCode = spv::OpUMulExtended;
7257 typeId = builder.makeStructResultType(typeId0, typeId0);
7258 consumedOperands = 2;
7259 break;
7260 case glslang::EOpIMulExtended:
7261 opCode = spv::OpSMulExtended;
7262 typeId = builder.makeStructResultType(typeId0, typeId0);
7263 consumedOperands = 2;
7264 break;
7265 case glslang::EOpBitfieldExtract:
7266 if (isUnsigned)
7267 opCode = spv::OpBitFieldUExtract;
7268 else
7269 opCode = spv::OpBitFieldSExtract;
7270 break;
7271 case glslang::EOpBitfieldInsert:
7272 opCode = spv::OpBitFieldInsert;
7273 break;
7274
7275 case glslang::EOpFma:
7276 libCall = spv::GLSLstd450Fma;
7277 break;
7278 case glslang::EOpFrexp:
7279 {
7280 libCall = spv::GLSLstd450FrexpStruct;
7281 assert(builder.isPointerType(typeId1));
7282 typeId1 = builder.getContainedTypeId(typeId1);
7283 int width = builder.getScalarTypeWidth(typeId1);
7284 if (width == 16)
7285 // Using 16-bit exp operand, enable extension SPV_AMD_gpu_shader_int16
7286 builder.addExtension(spv::E_SPV_AMD_gpu_shader_int16);
7287 if (builder.getNumComponents(operands[0]) == 1)
7288 frexpIntType = builder.makeIntegerType(width, true);
7289 else
7290 frexpIntType = builder.makeVectorType(builder.makeIntegerType(width, true), builder.getNumComponents(operands[0]));
7291 typeId = builder.makeStructResultType(typeId0, frexpIntType);
7292 consumedOperands = 1;
7293 }
7294 break;
7295 case glslang::EOpLdexp:
7296 libCall = spv::GLSLstd450Ldexp;
7297 break;
7298
7299 case glslang::EOpReadInvocation:
7300 return createInvocationsOperation(op, typeId, operands, typeProxy);
7301
7302 case glslang::EOpSubgroupBroadcast:
7303 case glslang::EOpSubgroupBallotBitExtract:
7304 case glslang::EOpSubgroupShuffle:
7305 case glslang::EOpSubgroupShuffleXor:
7306 case glslang::EOpSubgroupShuffleUp:
7307 case glslang::EOpSubgroupShuffleDown:
7308 case glslang::EOpSubgroupClusteredAdd:
7309 case glslang::EOpSubgroupClusteredMul:
7310 case glslang::EOpSubgroupClusteredMin:
7311 case glslang::EOpSubgroupClusteredMax:
7312 case glslang::EOpSubgroupClusteredAnd:
7313 case glslang::EOpSubgroupClusteredOr:
7314 case glslang::EOpSubgroupClusteredXor:
7315 case glslang::EOpSubgroupQuadBroadcast:
7316 case glslang::EOpSubgroupPartitionedAdd:
7317 case glslang::EOpSubgroupPartitionedMul:
7318 case glslang::EOpSubgroupPartitionedMin:
7319 case glslang::EOpSubgroupPartitionedMax:
7320 case glslang::EOpSubgroupPartitionedAnd:
7321 case glslang::EOpSubgroupPartitionedOr:
7322 case glslang::EOpSubgroupPartitionedXor:
7323 case glslang::EOpSubgroupPartitionedInclusiveAdd:
7324 case glslang::EOpSubgroupPartitionedInclusiveMul:
7325 case glslang::EOpSubgroupPartitionedInclusiveMin:
7326 case glslang::EOpSubgroupPartitionedInclusiveMax:
7327 case glslang::EOpSubgroupPartitionedInclusiveAnd:
7328 case glslang::EOpSubgroupPartitionedInclusiveOr:
7329 case glslang::EOpSubgroupPartitionedInclusiveXor:
7330 case glslang::EOpSubgroupPartitionedExclusiveAdd:
7331 case glslang::EOpSubgroupPartitionedExclusiveMul:
7332 case glslang::EOpSubgroupPartitionedExclusiveMin:
7333 case glslang::EOpSubgroupPartitionedExclusiveMax:
7334 case glslang::EOpSubgroupPartitionedExclusiveAnd:
7335 case glslang::EOpSubgroupPartitionedExclusiveOr:
7336 case glslang::EOpSubgroupPartitionedExclusiveXor:
7337 return createSubgroupOperation(op, typeId, operands, typeProxy);
7338
7339 case glslang::EOpSwizzleInvocations:
7340 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot);
7341 libCall = spv::SwizzleInvocationsAMD;
7342 break;
7343 case glslang::EOpSwizzleInvocationsMasked:
7344 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot);
7345 libCall = spv::SwizzleInvocationsMaskedAMD;
7346 break;
7347 case glslang::EOpWriteInvocation:
7348 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot);
7349 libCall = spv::WriteInvocationAMD;
7350 break;
7351
7352 case glslang::EOpMin3:
7353 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax);
7354 if (isFloat)
7355 libCall = spv::FMin3AMD;
7356 else {
7357 if (isUnsigned)
7358 libCall = spv::UMin3AMD;
7359 else
7360 libCall = spv::SMin3AMD;
7361 }
7362 break;
7363 case glslang::EOpMax3:
7364 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax);
7365 if (isFloat)
7366 libCall = spv::FMax3AMD;
7367 else {
7368 if (isUnsigned)
7369 libCall = spv::UMax3AMD;
7370 else
7371 libCall = spv::SMax3AMD;
7372 }
7373 break;
7374 case glslang::EOpMid3:
7375 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax);
7376 if (isFloat)
7377 libCall = spv::FMid3AMD;
7378 else {
7379 if (isUnsigned)
7380 libCall = spv::UMid3AMD;
7381 else
7382 libCall = spv::SMid3AMD;
7383 }
7384 break;
7385
7386 case glslang::EOpInterpolateAtVertex:
7387 if (typeProxy == glslang::EbtFloat16)
7388 builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float);
7389 extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
7390 libCall = spv::InterpolateAtVertexAMD;
7391 break;
7392
7393 case glslang::EOpReportIntersectionNV:
7394 {
7395 typeId = builder.makeBoolType();
7396 opCode = spv::OpReportIntersectionNV;
7397 }
7398 break;
7399 case glslang::EOpTraceNV:
7400 {
7401 builder.createNoResultOp(spv::OpTraceNV, operands);
7402 return 0;
7403 }
7404 break;
7405 case glslang::EOpExecuteCallableNV:
7406 {
7407 builder.createNoResultOp(spv::OpExecuteCallableNV, operands);
7408 return 0;
7409 }
7410 break;
7411 case glslang::EOpWritePackedPrimitiveIndices4x8NV:
7412 builder.createNoResultOp(spv::OpWritePackedPrimitiveIndices4x8NV, operands);
7413 return 0;
7414 case glslang::EOpCooperativeMatrixMulAdd:
7415 opCode = spv::OpCooperativeMatrixMulAddNV;
7416 break;
7417 #endif // GLSLANG_WEB
7418 default:
7419 return 0;
7420 }
7421
7422 spv::Id id = 0;
7423 if (libCall >= 0) {
7424 // Use an extended instruction from the standard library.
7425 // Construct the call arguments, without modifying the original operands vector.
7426 // We might need the remaining arguments, e.g. in the EOpFrexp case.
7427 std::vector<spv::Id> callArguments(operands.begin(), operands.begin() + consumedOperands);
7428 id = builder.createBuiltinCall(typeId, extBuiltins >= 0 ? extBuiltins : stdBuiltins, libCall, callArguments);
7429 } else if (opCode == spv::OpDot && !isFloat) {
7430 // int dot(int, int)
7431 // NOTE: never called for scalar/vector1, this is turned into simple mul before this can be reached
7432 const int componentCount = builder.getNumComponents(operands[0]);
7433 spv::Id mulOp = builder.createBinOp(spv::OpIMul, builder.getTypeId(operands[0]), operands[0], operands[1]);
7434 builder.setPrecision(mulOp, precision);
7435 id = builder.createCompositeExtract(mulOp, typeId, 0);
7436 for (int i = 1; i < componentCount; ++i) {
7437 builder.setPrecision(id, precision);
7438 id = builder.createBinOp(spv::OpIAdd, typeId, id, builder.createCompositeExtract(operands[0], typeId, i));
7439 }
7440 } else {
7441 switch (consumedOperands) {
7442 case 0:
7443 // should all be handled by visitAggregate and createNoArgOperation
7444 assert(0);
7445 return 0;
7446 case 1:
7447 // should all be handled by createUnaryOperation
7448 assert(0);
7449 return 0;
7450 case 2:
7451 id = builder.createBinOp(opCode, typeId, operands[0], operands[1]);
7452 break;
7453 default:
7454 // anything 3 or over doesn't have l-value operands, so all should be consumed
7455 assert(consumedOperands == operands.size());
7456 id = builder.createOp(opCode, typeId, operands);
7457 break;
7458 }
7459 }
7460
7461 #ifndef GLSLANG_WEB
7462 // Decode the return types that were structures
7463 switch (op) {
7464 case glslang::EOpAddCarry:
7465 case glslang::EOpSubBorrow:
7466 builder.createStore(builder.createCompositeExtract(id, typeId0, 1), operands[2]);
7467 id = builder.createCompositeExtract(id, typeId0, 0);
7468 break;
7469 case glslang::EOpUMulExtended:
7470 case glslang::EOpIMulExtended:
7471 builder.createStore(builder.createCompositeExtract(id, typeId0, 0), operands[3]);
7472 builder.createStore(builder.createCompositeExtract(id, typeId0, 1), operands[2]);
7473 break;
7474 case glslang::EOpFrexp:
7475 {
7476 assert(operands.size() == 2);
7477 if (builder.isFloatType(builder.getScalarTypeId(typeId1))) {
7478 // "exp" is floating-point type (from HLSL intrinsic)
7479 spv::Id member1 = builder.createCompositeExtract(id, frexpIntType, 1);
7480 member1 = builder.createUnaryOp(spv::OpConvertSToF, typeId1, member1);
7481 builder.createStore(member1, operands[1]);
7482 } else
7483 // "exp" is integer type (from GLSL built-in function)
7484 builder.createStore(builder.createCompositeExtract(id, frexpIntType, 1), operands[1]);
7485 id = builder.createCompositeExtract(id, typeId0, 0);
7486 }
7487 break;
7488 default:
7489 break;
7490 }
7491 #endif
7492
7493 return builder.setPrecision(id, precision);
7494 }
7495
7496 // Intrinsics with no arguments (or no return value, and no precision).
createNoArgOperation(glslang::TOperator op,spv::Decoration precision,spv::Id typeId)7497 spv::Id TGlslangToSpvTraverser::createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId)
7498 {
7499 // GLSL memory barriers use queuefamily scope in new model, device scope in old model
7500 spv::Scope memoryBarrierScope = glslangIntermediate->usingVulkanMemoryModel() ? spv::ScopeQueueFamilyKHR : spv::ScopeDevice;
7501
7502 switch (op) {
7503 case glslang::EOpBarrier:
7504 if (glslangIntermediate->getStage() == EShLangTessControl) {
7505 if (glslangIntermediate->usingVulkanMemoryModel()) {
7506 builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup,
7507 spv::MemorySemanticsOutputMemoryKHRMask |
7508 spv::MemorySemanticsAcquireReleaseMask);
7509 builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
7510 } else {
7511 builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeInvocation, spv::MemorySemanticsMaskNone);
7512 }
7513 } else {
7514 builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup,
7515 spv::MemorySemanticsWorkgroupMemoryMask |
7516 spv::MemorySemanticsAcquireReleaseMask);
7517 }
7518 return 0;
7519 case glslang::EOpMemoryBarrier:
7520 builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsAllMemory |
7521 spv::MemorySemanticsAcquireReleaseMask);
7522 return 0;
7523 case glslang::EOpMemoryBarrierBuffer:
7524 builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsUniformMemoryMask |
7525 spv::MemorySemanticsAcquireReleaseMask);
7526 return 0;
7527 case glslang::EOpMemoryBarrierShared:
7528 builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsWorkgroupMemoryMask |
7529 spv::MemorySemanticsAcquireReleaseMask);
7530 return 0;
7531 case glslang::EOpGroupMemoryBarrier:
7532 builder.createMemoryBarrier(spv::ScopeWorkgroup, spv::MemorySemanticsAllMemory |
7533 spv::MemorySemanticsAcquireReleaseMask);
7534 return 0;
7535 #ifndef GLSLANG_WEB
7536 case glslang::EOpMemoryBarrierAtomicCounter:
7537 builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsAtomicCounterMemoryMask |
7538 spv::MemorySemanticsAcquireReleaseMask);
7539 return 0;
7540 case glslang::EOpMemoryBarrierImage:
7541 builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsImageMemoryMask |
7542 spv::MemorySemanticsAcquireReleaseMask);
7543 return 0;
7544 case glslang::EOpAllMemoryBarrierWithGroupSync:
7545 builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeDevice,
7546 spv::MemorySemanticsAllMemory |
7547 spv::MemorySemanticsAcquireReleaseMask);
7548 return 0;
7549 case glslang::EOpDeviceMemoryBarrier:
7550 builder.createMemoryBarrier(spv::ScopeDevice, spv::MemorySemanticsUniformMemoryMask |
7551 spv::MemorySemanticsImageMemoryMask |
7552 spv::MemorySemanticsAcquireReleaseMask);
7553 return 0;
7554 case glslang::EOpDeviceMemoryBarrierWithGroupSync:
7555 builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeDevice, spv::MemorySemanticsUniformMemoryMask |
7556 spv::MemorySemanticsImageMemoryMask |
7557 spv::MemorySemanticsAcquireReleaseMask);
7558 return 0;
7559 case glslang::EOpWorkgroupMemoryBarrier:
7560 builder.createMemoryBarrier(spv::ScopeWorkgroup, spv::MemorySemanticsWorkgroupMemoryMask |
7561 spv::MemorySemanticsAcquireReleaseMask);
7562 return 0;
7563 case glslang::EOpWorkgroupMemoryBarrierWithGroupSync:
7564 builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup,
7565 spv::MemorySemanticsWorkgroupMemoryMask |
7566 spv::MemorySemanticsAcquireReleaseMask);
7567 return 0;
7568 case glslang::EOpSubgroupBarrier:
7569 builder.createControlBarrier(spv::ScopeSubgroup, spv::ScopeSubgroup, spv::MemorySemanticsAllMemory |
7570 spv::MemorySemanticsAcquireReleaseMask);
7571 return spv::NoResult;
7572 case glslang::EOpSubgroupMemoryBarrier:
7573 builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsAllMemory |
7574 spv::MemorySemanticsAcquireReleaseMask);
7575 return spv::NoResult;
7576 case glslang::EOpSubgroupMemoryBarrierBuffer:
7577 builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsUniformMemoryMask |
7578 spv::MemorySemanticsAcquireReleaseMask);
7579 return spv::NoResult;
7580 case glslang::EOpSubgroupMemoryBarrierImage:
7581 builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsImageMemoryMask |
7582 spv::MemorySemanticsAcquireReleaseMask);
7583 return spv::NoResult;
7584 case glslang::EOpSubgroupMemoryBarrierShared:
7585 builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsWorkgroupMemoryMask |
7586 spv::MemorySemanticsAcquireReleaseMask);
7587 return spv::NoResult;
7588
7589 case glslang::EOpEmitVertex:
7590 builder.createNoResultOp(spv::OpEmitVertex);
7591 return 0;
7592 case glslang::EOpEndPrimitive:
7593 builder.createNoResultOp(spv::OpEndPrimitive);
7594 return 0;
7595
7596 case glslang::EOpSubgroupElect: {
7597 std::vector<spv::Id> operands;
7598 return createSubgroupOperation(op, typeId, operands, glslang::EbtVoid);
7599 }
7600 case glslang::EOpTime:
7601 {
7602 std::vector<spv::Id> args; // Dummy arguments
7603 spv::Id id = builder.createBuiltinCall(typeId, getExtBuiltins(spv::E_SPV_AMD_gcn_shader), spv::TimeAMD, args);
7604 return builder.setPrecision(id, precision);
7605 }
7606 case glslang::EOpIgnoreIntersectionNV:
7607 builder.createNoResultOp(spv::OpIgnoreIntersectionNV);
7608 return 0;
7609 case glslang::EOpTerminateRayNV:
7610 builder.createNoResultOp(spv::OpTerminateRayNV);
7611 return 0;
7612
7613 case glslang::EOpBeginInvocationInterlock:
7614 builder.createNoResultOp(spv::OpBeginInvocationInterlockEXT);
7615 return 0;
7616 case glslang::EOpEndInvocationInterlock:
7617 builder.createNoResultOp(spv::OpEndInvocationInterlockEXT);
7618 return 0;
7619
7620 case glslang::EOpIsHelperInvocation:
7621 {
7622 std::vector<spv::Id> args; // Dummy arguments
7623 builder.addExtension(spv::E_SPV_EXT_demote_to_helper_invocation);
7624 builder.addCapability(spv::CapabilityDemoteToHelperInvocationEXT);
7625 return builder.createOp(spv::OpIsHelperInvocationEXT, typeId, args);
7626 }
7627
7628 case glslang::EOpReadClockSubgroupKHR: {
7629 std::vector<spv::Id> args;
7630 args.push_back(builder.makeUintConstant(spv::ScopeSubgroup));
7631 builder.addExtension(spv::E_SPV_KHR_shader_clock);
7632 builder.addCapability(spv::CapabilityShaderClockKHR);
7633 return builder.createOp(spv::OpReadClockKHR, typeId, args);
7634 }
7635
7636 case glslang::EOpReadClockDeviceKHR: {
7637 std::vector<spv::Id> args;
7638 args.push_back(builder.makeUintConstant(spv::ScopeDevice));
7639 builder.addExtension(spv::E_SPV_KHR_shader_clock);
7640 builder.addCapability(spv::CapabilityShaderClockKHR);
7641 return builder.createOp(spv::OpReadClockKHR, typeId, args);
7642 }
7643 #endif
7644 default:
7645 break;
7646 }
7647
7648 logger->missingFunctionality("unknown operation with no arguments");
7649
7650 return 0;
7651 }
7652
getSymbolId(const glslang::TIntermSymbol * symbol)7653 spv::Id TGlslangToSpvTraverser::getSymbolId(const glslang::TIntermSymbol* symbol)
7654 {
7655 auto iter = symbolValues.find(symbol->getId());
7656 spv::Id id;
7657 if (symbolValues.end() != iter) {
7658 id = iter->second;
7659 return id;
7660 }
7661
7662 // it was not found, create it
7663 spv::BuiltIn builtIn = TranslateBuiltInDecoration(symbol->getQualifier().builtIn, false);
7664 auto forcedType = getForcedType(builtIn, symbol->getType());
7665 id = createSpvVariable(symbol, forcedType.first);
7666 symbolValues[symbol->getId()] = id;
7667 if (forcedType.second != spv::NoType)
7668 forceType[id] = forcedType.second;
7669
7670 if (symbol->getBasicType() != glslang::EbtBlock) {
7671 builder.addDecoration(id, TranslatePrecisionDecoration(symbol->getType()));
7672 builder.addDecoration(id, TranslateInterpolationDecoration(symbol->getType().getQualifier()));
7673 builder.addDecoration(id, TranslateAuxiliaryStorageDecoration(symbol->getType().getQualifier()));
7674 #ifndef GLSLANG_WEB
7675 addMeshNVDecoration(id, /*member*/ -1, symbol->getType().getQualifier());
7676 if (symbol->getQualifier().hasComponent())
7677 builder.addDecoration(id, spv::DecorationComponent, symbol->getQualifier().layoutComponent);
7678 if (symbol->getQualifier().hasIndex())
7679 builder.addDecoration(id, spv::DecorationIndex, symbol->getQualifier().layoutIndex);
7680 #endif
7681 if (symbol->getType().getQualifier().hasSpecConstantId())
7682 builder.addDecoration(id, spv::DecorationSpecId, symbol->getType().getQualifier().layoutSpecConstantId);
7683 // atomic counters use this:
7684 if (symbol->getQualifier().hasOffset())
7685 builder.addDecoration(id, spv::DecorationOffset, symbol->getQualifier().layoutOffset);
7686 }
7687
7688 if (symbol->getQualifier().hasLocation())
7689 builder.addDecoration(id, spv::DecorationLocation, symbol->getQualifier().layoutLocation);
7690 builder.addDecoration(id, TranslateInvariantDecoration(symbol->getType().getQualifier()));
7691 if (symbol->getQualifier().hasStream() && glslangIntermediate->isMultiStream()) {
7692 builder.addCapability(spv::CapabilityGeometryStreams);
7693 builder.addDecoration(id, spv::DecorationStream, symbol->getQualifier().layoutStream);
7694 }
7695 if (symbol->getQualifier().hasSet())
7696 builder.addDecoration(id, spv::DecorationDescriptorSet, symbol->getQualifier().layoutSet);
7697 else if (IsDescriptorResource(symbol->getType())) {
7698 // default to 0
7699 builder.addDecoration(id, spv::DecorationDescriptorSet, 0);
7700 }
7701 if (symbol->getQualifier().hasBinding())
7702 builder.addDecoration(id, spv::DecorationBinding, symbol->getQualifier().layoutBinding);
7703 else if (IsDescriptorResource(symbol->getType())) {
7704 // default to 0
7705 builder.addDecoration(id, spv::DecorationBinding, 0);
7706 }
7707 if (symbol->getQualifier().hasAttachment())
7708 builder.addDecoration(id, spv::DecorationInputAttachmentIndex, symbol->getQualifier().layoutAttachment);
7709 if (glslangIntermediate->getXfbMode()) {
7710 builder.addCapability(spv::CapabilityTransformFeedback);
7711 if (symbol->getQualifier().hasXfbBuffer()) {
7712 builder.addDecoration(id, spv::DecorationXfbBuffer, symbol->getQualifier().layoutXfbBuffer);
7713 unsigned stride = glslangIntermediate->getXfbStride(symbol->getQualifier().layoutXfbBuffer);
7714 if (stride != glslang::TQualifier::layoutXfbStrideEnd)
7715 builder.addDecoration(id, spv::DecorationXfbStride, stride);
7716 }
7717 if (symbol->getQualifier().hasXfbOffset())
7718 builder.addDecoration(id, spv::DecorationOffset, symbol->getQualifier().layoutXfbOffset);
7719 }
7720
7721 // add built-in variable decoration
7722 if (builtIn != spv::BuiltInMax) {
7723 builder.addDecoration(id, spv::DecorationBuiltIn, (int)builtIn);
7724 }
7725
7726 #ifndef GLSLANG_WEB
7727 if (symbol->getType().isImage()) {
7728 std::vector<spv::Decoration> memory;
7729 TranslateMemoryDecoration(symbol->getType().getQualifier(), memory, glslangIntermediate->usingVulkanMemoryModel());
7730 for (unsigned int i = 0; i < memory.size(); ++i)
7731 builder.addDecoration(id, memory[i]);
7732 }
7733
7734 // nonuniform
7735 builder.addDecoration(id, TranslateNonUniformDecoration(symbol->getType().getQualifier()));
7736
7737 if (builtIn == spv::BuiltInSampleMask) {
7738 spv::Decoration decoration;
7739 // GL_NV_sample_mask_override_coverage extension
7740 if (glslangIntermediate->getLayoutOverrideCoverage())
7741 decoration = (spv::Decoration)spv::DecorationOverrideCoverageNV;
7742 else
7743 decoration = (spv::Decoration)spv::DecorationMax;
7744 builder.addDecoration(id, decoration);
7745 if (decoration != spv::DecorationMax) {
7746 builder.addCapability(spv::CapabilitySampleMaskOverrideCoverageNV);
7747 builder.addExtension(spv::E_SPV_NV_sample_mask_override_coverage);
7748 }
7749 }
7750 else if (builtIn == spv::BuiltInLayer) {
7751 // SPV_NV_viewport_array2 extension
7752 if (symbol->getQualifier().layoutViewportRelative) {
7753 builder.addDecoration(id, (spv::Decoration)spv::DecorationViewportRelativeNV);
7754 builder.addCapability(spv::CapabilityShaderViewportMaskNV);
7755 builder.addExtension(spv::E_SPV_NV_viewport_array2);
7756 }
7757 if (symbol->getQualifier().layoutSecondaryViewportRelativeOffset != -2048) {
7758 builder.addDecoration(id, (spv::Decoration)spv::DecorationSecondaryViewportRelativeNV,
7759 symbol->getQualifier().layoutSecondaryViewportRelativeOffset);
7760 builder.addCapability(spv::CapabilityShaderStereoViewNV);
7761 builder.addExtension(spv::E_SPV_NV_stereo_view_rendering);
7762 }
7763 }
7764
7765 if (symbol->getQualifier().layoutPassthrough) {
7766 builder.addDecoration(id, spv::DecorationPassthroughNV);
7767 builder.addCapability(spv::CapabilityGeometryShaderPassthroughNV);
7768 builder.addExtension(spv::E_SPV_NV_geometry_shader_passthrough);
7769 }
7770 if (symbol->getQualifier().pervertexNV) {
7771 builder.addDecoration(id, spv::DecorationPerVertexNV);
7772 builder.addCapability(spv::CapabilityFragmentBarycentricNV);
7773 builder.addExtension(spv::E_SPV_NV_fragment_shader_barycentric);
7774 }
7775
7776 if (glslangIntermediate->getHlslFunctionality1() && symbol->getType().getQualifier().semanticName != nullptr) {
7777 builder.addExtension("SPV_GOOGLE_hlsl_functionality1");
7778 builder.addDecoration(id, (spv::Decoration)spv::DecorationHlslSemanticGOOGLE,
7779 symbol->getType().getQualifier().semanticName);
7780 }
7781
7782 if (symbol->isReference()) {
7783 builder.addDecoration(id, symbol->getType().getQualifier().restrict ? spv::DecorationRestrictPointerEXT : spv::DecorationAliasedPointerEXT);
7784 }
7785 #endif
7786
7787 return id;
7788 }
7789
7790 #ifndef GLSLANG_WEB
7791 // add per-primitive, per-view. per-task decorations to a struct member (member >= 0) or an object
addMeshNVDecoration(spv::Id id,int member,const glslang::TQualifier & qualifier)7792 void TGlslangToSpvTraverser::addMeshNVDecoration(spv::Id id, int member, const glslang::TQualifier& qualifier)
7793 {
7794 if (member >= 0) {
7795 if (qualifier.perPrimitiveNV) {
7796 // Need to add capability/extension for fragment shader.
7797 // Mesh shader already adds this by default.
7798 if (glslangIntermediate->getStage() == EShLangFragment) {
7799 builder.addCapability(spv::CapabilityMeshShadingNV);
7800 builder.addExtension(spv::E_SPV_NV_mesh_shader);
7801 }
7802 builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerPrimitiveNV);
7803 }
7804 if (qualifier.perViewNV)
7805 builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerViewNV);
7806 if (qualifier.perTaskNV)
7807 builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerTaskNV);
7808 } else {
7809 if (qualifier.perPrimitiveNV) {
7810 // Need to add capability/extension for fragment shader.
7811 // Mesh shader already adds this by default.
7812 if (glslangIntermediate->getStage() == EShLangFragment) {
7813 builder.addCapability(spv::CapabilityMeshShadingNV);
7814 builder.addExtension(spv::E_SPV_NV_mesh_shader);
7815 }
7816 builder.addDecoration(id, spv::DecorationPerPrimitiveNV);
7817 }
7818 if (qualifier.perViewNV)
7819 builder.addDecoration(id, spv::DecorationPerViewNV);
7820 if (qualifier.perTaskNV)
7821 builder.addDecoration(id, spv::DecorationPerTaskNV);
7822 }
7823 }
7824 #endif
7825
7826 // Make a full tree of instructions to build a SPIR-V specialization constant,
7827 // or regular constant if possible.
7828 //
7829 // TBD: this is not yet done, nor verified to be the best design, it does do the leaf symbols though
7830 //
7831 // Recursively walk the nodes. The nodes form a tree whose leaves are
7832 // regular constants, which themselves are trees that createSpvConstant()
7833 // recursively walks. So, this function walks the "top" of the tree:
7834 // - emit specialization constant-building instructions for specConstant
7835 // - when running into a non-spec-constant, switch to createSpvConstant()
createSpvConstant(const glslang::TIntermTyped & node)7836 spv::Id TGlslangToSpvTraverser::createSpvConstant(const glslang::TIntermTyped& node)
7837 {
7838 assert(node.getQualifier().isConstant());
7839
7840 // Handle front-end constants first (non-specialization constants).
7841 if (! node.getQualifier().specConstant) {
7842 // hand off to the non-spec-constant path
7843 assert(node.getAsConstantUnion() != nullptr || node.getAsSymbolNode() != nullptr);
7844 int nextConst = 0;
7845 return createSpvConstantFromConstUnionArray(node.getType(), node.getAsConstantUnion() ? node.getAsConstantUnion()->getConstArray() : node.getAsSymbolNode()->getConstArray(),
7846 nextConst, false);
7847 }
7848
7849 // We now know we have a specialization constant to build
7850
7851 // gl_WorkGroupSize is a special case until the front-end handles hierarchical specialization constants,
7852 // even then, it's specialization ids are handled by special case syntax in GLSL: layout(local_size_x = ...
7853 if (node.getType().getQualifier().builtIn == glslang::EbvWorkGroupSize) {
7854 std::vector<spv::Id> dimConstId;
7855 for (int dim = 0; dim < 3; ++dim) {
7856 bool specConst = (glslangIntermediate->getLocalSizeSpecId(dim) != glslang::TQualifier::layoutNotSet);
7857 dimConstId.push_back(builder.makeUintConstant(glslangIntermediate->getLocalSize(dim), specConst));
7858 if (specConst) {
7859 builder.addDecoration(dimConstId.back(), spv::DecorationSpecId,
7860 glslangIntermediate->getLocalSizeSpecId(dim));
7861 }
7862 }
7863 return builder.makeCompositeConstant(builder.makeVectorType(builder.makeUintType(32), 3), dimConstId, true);
7864 }
7865
7866 // An AST node labelled as specialization constant should be a symbol node.
7867 // Its initializer should either be a sub tree with constant nodes, or a constant union array.
7868 if (auto* sn = node.getAsSymbolNode()) {
7869 spv::Id result;
7870 if (auto* sub_tree = sn->getConstSubtree()) {
7871 // Traverse the constant constructor sub tree like generating normal run-time instructions.
7872 // During the AST traversal, if the node is marked as 'specConstant', SpecConstantOpModeGuard
7873 // will set the builder into spec constant op instruction generating mode.
7874 sub_tree->traverse(this);
7875 result = accessChainLoad(sub_tree->getType());
7876 } else if (auto* const_union_array = &sn->getConstArray()) {
7877 int nextConst = 0;
7878 result = createSpvConstantFromConstUnionArray(sn->getType(), *const_union_array, nextConst, true);
7879 } else {
7880 logger->missingFunctionality("Invalid initializer for spec onstant.");
7881 return spv::NoResult;
7882 }
7883 builder.addName(result, sn->getName().c_str());
7884 return result;
7885 }
7886
7887 // Neither a front-end constant node, nor a specialization constant node with constant union array or
7888 // constant sub tree as initializer.
7889 logger->missingFunctionality("Neither a front-end constant nor a spec constant.");
7890 return spv::NoResult;
7891 }
7892
7893 // Use 'consts' as the flattened glslang source of scalar constants to recursively
7894 // build the aggregate SPIR-V constant.
7895 //
7896 // If there are not enough elements present in 'consts', 0 will be substituted;
7897 // an empty 'consts' can be used to create a fully zeroed SPIR-V constant.
7898 //
createSpvConstantFromConstUnionArray(const glslang::TType & glslangType,const glslang::TConstUnionArray & consts,int & nextConst,bool specConstant)7899 spv::Id TGlslangToSpvTraverser::createSpvConstantFromConstUnionArray(const glslang::TType& glslangType, const glslang::TConstUnionArray& consts, int& nextConst, bool specConstant)
7900 {
7901 // vector of constants for SPIR-V
7902 std::vector<spv::Id> spvConsts;
7903
7904 // Type is used for struct and array constants
7905 spv::Id typeId = convertGlslangToSpvType(glslangType);
7906
7907 if (glslangType.isArray()) {
7908 glslang::TType elementType(glslangType, 0);
7909 for (int i = 0; i < glslangType.getOuterArraySize(); ++i)
7910 spvConsts.push_back(createSpvConstantFromConstUnionArray(elementType, consts, nextConst, false));
7911 } else if (glslangType.isMatrix()) {
7912 glslang::TType vectorType(glslangType, 0);
7913 for (int col = 0; col < glslangType.getMatrixCols(); ++col)
7914 spvConsts.push_back(createSpvConstantFromConstUnionArray(vectorType, consts, nextConst, false));
7915 } else if (glslangType.isCoopMat()) {
7916 glslang::TType componentType(glslangType.getBasicType());
7917 spvConsts.push_back(createSpvConstantFromConstUnionArray(componentType, consts, nextConst, false));
7918 } else if (glslangType.isStruct()) {
7919 glslang::TVector<glslang::TTypeLoc>::const_iterator iter;
7920 for (iter = glslangType.getStruct()->begin(); iter != glslangType.getStruct()->end(); ++iter)
7921 spvConsts.push_back(createSpvConstantFromConstUnionArray(*iter->type, consts, nextConst, false));
7922 } else if (glslangType.getVectorSize() > 1) {
7923 for (unsigned int i = 0; i < (unsigned int)glslangType.getVectorSize(); ++i) {
7924 bool zero = nextConst >= consts.size();
7925 switch (glslangType.getBasicType()) {
7926 case glslang::EbtInt:
7927 spvConsts.push_back(builder.makeIntConstant(zero ? 0 : consts[nextConst].getIConst()));
7928 break;
7929 case glslang::EbtUint:
7930 spvConsts.push_back(builder.makeUintConstant(zero ? 0 : consts[nextConst].getUConst()));
7931 break;
7932 case glslang::EbtFloat:
7933 spvConsts.push_back(builder.makeFloatConstant(zero ? 0.0F : (float)consts[nextConst].getDConst()));
7934 break;
7935 case glslang::EbtBool:
7936 spvConsts.push_back(builder.makeBoolConstant(zero ? false : consts[nextConst].getBConst()));
7937 break;
7938 #ifndef GLSLANG_WEB
7939 case glslang::EbtInt8:
7940 spvConsts.push_back(builder.makeInt8Constant(zero ? 0 : consts[nextConst].getI8Const()));
7941 break;
7942 case glslang::EbtUint8:
7943 spvConsts.push_back(builder.makeUint8Constant(zero ? 0 : consts[nextConst].getU8Const()));
7944 break;
7945 case glslang::EbtInt16:
7946 spvConsts.push_back(builder.makeInt16Constant(zero ? 0 : consts[nextConst].getI16Const()));
7947 break;
7948 case glslang::EbtUint16:
7949 spvConsts.push_back(builder.makeUint16Constant(zero ? 0 : consts[nextConst].getU16Const()));
7950 break;
7951 case glslang::EbtInt64:
7952 spvConsts.push_back(builder.makeInt64Constant(zero ? 0 : consts[nextConst].getI64Const()));
7953 break;
7954 case glslang::EbtUint64:
7955 spvConsts.push_back(builder.makeUint64Constant(zero ? 0 : consts[nextConst].getU64Const()));
7956 break;
7957 case glslang::EbtDouble:
7958 spvConsts.push_back(builder.makeDoubleConstant(zero ? 0.0 : consts[nextConst].getDConst()));
7959 break;
7960 case glslang::EbtFloat16:
7961 spvConsts.push_back(builder.makeFloat16Constant(zero ? 0.0F : (float)consts[nextConst].getDConst()));
7962 break;
7963 #endif
7964 default:
7965 assert(0);
7966 break;
7967 }
7968 ++nextConst;
7969 }
7970 } else {
7971 // we have a non-aggregate (scalar) constant
7972 bool zero = nextConst >= consts.size();
7973 spv::Id scalar = 0;
7974 switch (glslangType.getBasicType()) {
7975 case glslang::EbtInt:
7976 scalar = builder.makeIntConstant(zero ? 0 : consts[nextConst].getIConst(), specConstant);
7977 break;
7978 case glslang::EbtUint:
7979 scalar = builder.makeUintConstant(zero ? 0 : consts[nextConst].getUConst(), specConstant);
7980 break;
7981 case glslang::EbtFloat:
7982 scalar = builder.makeFloatConstant(zero ? 0.0F : (float)consts[nextConst].getDConst(), specConstant);
7983 break;
7984 case glslang::EbtBool:
7985 scalar = builder.makeBoolConstant(zero ? false : consts[nextConst].getBConst(), specConstant);
7986 break;
7987 #ifndef GLSLANG_WEB
7988 case glslang::EbtInt8:
7989 scalar = builder.makeInt8Constant(zero ? 0 : consts[nextConst].getI8Const(), specConstant);
7990 break;
7991 case glslang::EbtUint8:
7992 scalar = builder.makeUint8Constant(zero ? 0 : consts[nextConst].getU8Const(), specConstant);
7993 break;
7994 case glslang::EbtInt16:
7995 scalar = builder.makeInt16Constant(zero ? 0 : consts[nextConst].getI16Const(), specConstant);
7996 break;
7997 case glslang::EbtUint16:
7998 scalar = builder.makeUint16Constant(zero ? 0 : consts[nextConst].getU16Const(), specConstant);
7999 break;
8000 case glslang::EbtInt64:
8001 scalar = builder.makeInt64Constant(zero ? 0 : consts[nextConst].getI64Const(), specConstant);
8002 break;
8003 case glslang::EbtUint64:
8004 scalar = builder.makeUint64Constant(zero ? 0 : consts[nextConst].getU64Const(), specConstant);
8005 break;
8006 case glslang::EbtDouble:
8007 scalar = builder.makeDoubleConstant(zero ? 0.0 : consts[nextConst].getDConst(), specConstant);
8008 break;
8009 case glslang::EbtFloat16:
8010 scalar = builder.makeFloat16Constant(zero ? 0.0F : (float)consts[nextConst].getDConst(), specConstant);
8011 break;
8012 case glslang::EbtReference:
8013 scalar = builder.makeUint64Constant(zero ? 0 : consts[nextConst].getU64Const(), specConstant);
8014 scalar = builder.createUnaryOp(spv::OpBitcast, typeId, scalar);
8015 break;
8016 #endif
8017 default:
8018 assert(0);
8019 break;
8020 }
8021 ++nextConst;
8022 return scalar;
8023 }
8024
8025 return builder.makeCompositeConstant(typeId, spvConsts);
8026 }
8027
8028 // Return true if the node is a constant or symbol whose reading has no
8029 // non-trivial observable cost or effect.
isTrivialLeaf(const glslang::TIntermTyped * node)8030 bool TGlslangToSpvTraverser::isTrivialLeaf(const glslang::TIntermTyped* node)
8031 {
8032 // don't know what this is
8033 if (node == nullptr)
8034 return false;
8035
8036 // a constant is safe
8037 if (node->getAsConstantUnion() != nullptr)
8038 return true;
8039
8040 // not a symbol means non-trivial
8041 if (node->getAsSymbolNode() == nullptr)
8042 return false;
8043
8044 // a symbol, depends on what's being read
8045 switch (node->getType().getQualifier().storage) {
8046 case glslang::EvqTemporary:
8047 case glslang::EvqGlobal:
8048 case glslang::EvqIn:
8049 case glslang::EvqInOut:
8050 case glslang::EvqConst:
8051 case glslang::EvqConstReadOnly:
8052 case glslang::EvqUniform:
8053 return true;
8054 default:
8055 return false;
8056 }
8057 }
8058
8059 // A node is trivial if it is a single operation with no side effects.
8060 // HLSL (and/or vectors) are always trivial, as it does not short circuit.
8061 // Otherwise, error on the side of saying non-trivial.
8062 // Return true if trivial.
isTrivial(const glslang::TIntermTyped * node)8063 bool TGlslangToSpvTraverser::isTrivial(const glslang::TIntermTyped* node)
8064 {
8065 if (node == nullptr)
8066 return false;
8067
8068 // count non scalars as trivial, as well as anything coming from HLSL
8069 if (! node->getType().isScalarOrVec1() || glslangIntermediate->getSource() == glslang::EShSourceHlsl)
8070 return true;
8071
8072 // symbols and constants are trivial
8073 if (isTrivialLeaf(node))
8074 return true;
8075
8076 // otherwise, it needs to be a simple operation or one or two leaf nodes
8077
8078 // not a simple operation
8079 const glslang::TIntermBinary* binaryNode = node->getAsBinaryNode();
8080 const glslang::TIntermUnary* unaryNode = node->getAsUnaryNode();
8081 if (binaryNode == nullptr && unaryNode == nullptr)
8082 return false;
8083
8084 // not on leaf nodes
8085 if (binaryNode && (! isTrivialLeaf(binaryNode->getLeft()) || ! isTrivialLeaf(binaryNode->getRight())))
8086 return false;
8087
8088 if (unaryNode && ! isTrivialLeaf(unaryNode->getOperand())) {
8089 return false;
8090 }
8091
8092 switch (node->getAsOperator()->getOp()) {
8093 case glslang::EOpLogicalNot:
8094 case glslang::EOpConvIntToBool:
8095 case glslang::EOpConvUintToBool:
8096 case glslang::EOpConvFloatToBool:
8097 case glslang::EOpConvDoubleToBool:
8098 case glslang::EOpEqual:
8099 case glslang::EOpNotEqual:
8100 case glslang::EOpLessThan:
8101 case glslang::EOpGreaterThan:
8102 case glslang::EOpLessThanEqual:
8103 case glslang::EOpGreaterThanEqual:
8104 case glslang::EOpIndexDirect:
8105 case glslang::EOpIndexDirectStruct:
8106 case glslang::EOpLogicalXor:
8107 case glslang::EOpAny:
8108 case glslang::EOpAll:
8109 return true;
8110 default:
8111 return false;
8112 }
8113 }
8114
8115 // Emit short-circuiting code, where 'right' is never evaluated unless
8116 // the left side is true (for &&) or false (for ||).
createShortCircuit(glslang::TOperator op,glslang::TIntermTyped & left,glslang::TIntermTyped & right)8117 spv::Id TGlslangToSpvTraverser::createShortCircuit(glslang::TOperator op, glslang::TIntermTyped& left, glslang::TIntermTyped& right)
8118 {
8119 spv::Id boolTypeId = builder.makeBoolType();
8120
8121 // emit left operand
8122 builder.clearAccessChain();
8123 left.traverse(this);
8124 spv::Id leftId = accessChainLoad(left.getType());
8125
8126 // Operands to accumulate OpPhi operands
8127 std::vector<spv::Id> phiOperands;
8128 // accumulate left operand's phi information
8129 phiOperands.push_back(leftId);
8130 phiOperands.push_back(builder.getBuildPoint()->getId());
8131
8132 // Make the two kinds of operation symmetric with a "!"
8133 // || => emit "if (! left) result = right"
8134 // && => emit "if ( left) result = right"
8135 //
8136 // TODO: this runtime "not" for || could be avoided by adding functionality
8137 // to 'builder' to have an "else" without an "then"
8138 if (op == glslang::EOpLogicalOr)
8139 leftId = builder.createUnaryOp(spv::OpLogicalNot, boolTypeId, leftId);
8140
8141 // make an "if" based on the left value
8142 spv::Builder::If ifBuilder(leftId, spv::SelectionControlMaskNone, builder);
8143
8144 // emit right operand as the "then" part of the "if"
8145 builder.clearAccessChain();
8146 right.traverse(this);
8147 spv::Id rightId = accessChainLoad(right.getType());
8148
8149 // accumulate left operand's phi information
8150 phiOperands.push_back(rightId);
8151 phiOperands.push_back(builder.getBuildPoint()->getId());
8152
8153 // finish the "if"
8154 ifBuilder.makeEndIf();
8155
8156 // phi together the two results
8157 return builder.createOp(spv::OpPhi, boolTypeId, phiOperands);
8158 }
8159
8160 #ifndef GLSLANG_WEB
8161 // Return type Id of the imported set of extended instructions corresponds to the name.
8162 // Import this set if it has not been imported yet.
getExtBuiltins(const char * name)8163 spv::Id TGlslangToSpvTraverser::getExtBuiltins(const char* name)
8164 {
8165 if (extBuiltinMap.find(name) != extBuiltinMap.end())
8166 return extBuiltinMap[name];
8167 else {
8168 builder.addExtension(name);
8169 spv::Id extBuiltins = builder.import(name);
8170 extBuiltinMap[name] = extBuiltins;
8171 return extBuiltins;
8172 }
8173 }
8174 #endif
8175
8176 }; // end anonymous namespace
8177
8178 namespace glslang {
8179
GetSpirvVersion(std::string & version)8180 void GetSpirvVersion(std::string& version)
8181 {
8182 const int bufSize = 100;
8183 char buf[bufSize];
8184 snprintf(buf, bufSize, "0x%08x, Revision %d", spv::Version, spv::Revision);
8185 version = buf;
8186 }
8187
8188 // For low-order part of the generator's magic number. Bump up
8189 // when there is a change in the style (e.g., if SSA form changes,
8190 // or a different instruction sequence to do something gets used).
GetSpirvGeneratorVersion()8191 int GetSpirvGeneratorVersion()
8192 {
8193 // return 1; // start
8194 // return 2; // EOpAtomicCounterDecrement gets a post decrement, to map between GLSL -> SPIR-V
8195 // return 3; // change/correct barrier-instruction operands, to match memory model group decisions
8196 // return 4; // some deeper access chains: for dynamic vector component, and local Boolean component
8197 // return 5; // make OpArrayLength result type be an int with signedness of 0
8198 // return 6; // revert version 5 change, which makes a different (new) kind of incorrect code,
8199 // versions 4 and 6 each generate OpArrayLength as it has long been done
8200 // return 7; // GLSL volatile keyword maps to both SPIR-V decorations Volatile and Coherent
8201 return 8; // switch to new dead block eliminator; use OpUnreachable
8202 }
8203
8204 // Write SPIR-V out to a binary file
OutputSpvBin(const std::vector<unsigned int> & spirv,const char * baseName)8205 void OutputSpvBin(const std::vector<unsigned int>& spirv, const char* baseName)
8206 {
8207 std::ofstream out;
8208 out.open(baseName, std::ios::binary | std::ios::out);
8209 if (out.fail())
8210 printf("ERROR: Failed to open file: %s\n", baseName);
8211 for (int i = 0; i < (int)spirv.size(); ++i) {
8212 unsigned int word = spirv[i];
8213 out.write((const char*)&word, 4);
8214 }
8215 out.close();
8216 }
8217
8218 // Write SPIR-V out to a text file with 32-bit hexadecimal words
OutputSpvHex(const std::vector<unsigned int> & spirv,const char * baseName,const char * varName)8219 void OutputSpvHex(const std::vector<unsigned int>& spirv, const char* baseName, const char* varName)
8220 {
8221 #ifndef GLSLANG_WEB
8222 std::ofstream out;
8223 out.open(baseName, std::ios::binary | std::ios::out);
8224 if (out.fail())
8225 printf("ERROR: Failed to open file: %s\n", baseName);
8226 out << "\t// " <<
8227 GetSpirvGeneratorVersion() << "." << GLSLANG_MINOR_VERSION << "." << GLSLANG_PATCH_LEVEL <<
8228 std::endl;
8229 if (varName != nullptr) {
8230 out << "\t #pragma once" << std::endl;
8231 out << "const uint32_t " << varName << "[] = {" << std::endl;
8232 }
8233 const int WORDS_PER_LINE = 8;
8234 for (int i = 0; i < (int)spirv.size(); i += WORDS_PER_LINE) {
8235 out << "\t";
8236 for (int j = 0; j < WORDS_PER_LINE && i + j < (int)spirv.size(); ++j) {
8237 const unsigned int word = spirv[i + j];
8238 out << "0x" << std::hex << std::setw(8) << std::setfill('0') << word;
8239 if (i + j + 1 < (int)spirv.size()) {
8240 out << ",";
8241 }
8242 }
8243 out << std::endl;
8244 }
8245 if (varName != nullptr) {
8246 out << "};";
8247 }
8248 out.close();
8249 #endif
8250 }
8251
8252 //
8253 // Set up the glslang traversal
8254 //
GlslangToSpv(const TIntermediate & intermediate,std::vector<unsigned int> & spirv,SpvOptions * options)8255 void GlslangToSpv(const TIntermediate& intermediate, std::vector<unsigned int>& spirv, SpvOptions* options)
8256 {
8257 spv::SpvBuildLogger logger;
8258 GlslangToSpv(intermediate, spirv, &logger, options);
8259 }
8260
GlslangToSpv(const TIntermediate & intermediate,std::vector<unsigned int> & spirv,spv::SpvBuildLogger * logger,SpvOptions * options)8261 void GlslangToSpv(const TIntermediate& intermediate, std::vector<unsigned int>& spirv,
8262 spv::SpvBuildLogger* logger, SpvOptions* options)
8263 {
8264 TIntermNode* root = intermediate.getTreeRoot();
8265
8266 if (root == 0)
8267 return;
8268
8269 SpvOptions defaultOptions;
8270 if (options == nullptr)
8271 options = &defaultOptions;
8272
8273 GetThreadPoolAllocator().push();
8274
8275 TGlslangToSpvTraverser it(intermediate.getSpv().spv, &intermediate, logger, *options);
8276 root->traverse(&it);
8277 it.finishSpv();
8278 it.dumpSpv(spirv);
8279
8280 #if ENABLE_OPT
8281 // If from HLSL, run spirv-opt to "legalize" the SPIR-V for Vulkan
8282 // eg. forward and remove memory writes of opaque types.
8283 bool prelegalization = intermediate.getSource() == EShSourceHlsl;
8284 if ((intermediate.getSource() == EShSourceHlsl || options->optimizeSize) && !options->disableOptimizer) {
8285 SpirvToolsLegalize(intermediate, spirv, logger, options);
8286 prelegalization = false;
8287 }
8288
8289 if (options->validate)
8290 SpirvToolsValidate(intermediate, spirv, logger, prelegalization);
8291
8292 if (options->disassemble)
8293 SpirvToolsDisassemble(std::cout, spirv);
8294
8295 #endif
8296
8297 GetThreadPoolAllocator().pop();
8298 }
8299
8300 }; // end namespace glslang
8301