AuroraMiddleware/glslang
1
0
Fork 0
mirror of https://github.com/KhronosGroup/glslang synced 2024-11-08 19:40:06 +00:00
Code Issues Packages Projects Releases Wiki Activity
b9197c812e
glslang/SPIRV/GlslangToSpv.cpp

8279 lines
346 KiB
C++
Raw Normal View History

Final round for line endings.
2015-06-26 22:58:36 +00:00
//
Non-functional: White space after "//", mostly for copyrights.
2017-01-06 19:34:14 +00:00
// Copyright (C) 2014-2016 LunarG, Inc.
Licensing. Fixes #958. Add licenes file and update copyrights.
2018-12-14 17:47:35 +00:00
// Copyright (C) 2015-2018 Google, Inc.
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
// Copyright (C) 2017 ARM Limited.
Final round for line endings.
2015-06-26 22:58:36 +00:00
//
Non-functional: White space after "//", mostly for copyrights.
2017-01-06 19:34:14 +00:00
// All rights reserved.
Final round for line endings.
2015-06-26 22:58:36 +00:00
//
Non-functional: White space after "//", mostly for copyrights.
2017-01-06 19:34:14 +00:00
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions
// are met:
Final round for line endings.
2015-06-26 22:58:36 +00:00
//
// Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// Redistributions in binary form must reproduce the above
// copyright notice, this list of conditions and the following
// disclaimer in the documentation and/or other materials provided
// with the distribution.
//
// Neither the name of 3Dlabs Inc. Ltd. nor the names of its
// contributors may be used to endorse or promote products derived
// from this software without specific prior written permission.
//
Non-functional: White space after "//", mostly for copyrights.
2017-01-06 19:34:14 +00:00
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
// POSSIBILITY OF SUCH DAMAGE.
Final round for line endings.
2015-06-26 22:58:36 +00:00
//
// Visit the nodes in the glslang intermediate tree representation to
// translate them to SPIR-V.
//
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
#include "spirv.hpp"
Final round for line endings.
2015-06-26 22:58:36 +00:00
#include "GlslangToSpv.h"
#include "SpvBuilder.h"
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
namespace spv {
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
#include "GLSL.std.450.h"
#include "GLSL.ext.KHR.h"
Add implementation of SPV_EXT_fragment_fully_covered This implementation uses the GLSL extension GL_NV_conservative_raster_underestimation to generate the new SPIR-V FullyCoveredEXT built in.
2017-12-13 20:07:22 +00:00
#include "GLSL.ext.EXT.h"
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
#include "GLSL.ext.AMD.h"
Add support for SPV_NV_sample_mask_override_coverage
2016-12-20 00:29:34 +00:00
#include "GLSL.ext.NV.h"
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
// Glslang includes
include by relative path, so root doesn't need to be in the search path
2015-07-08 13:11:59 +00:00
#include "../glslang/MachineIndependent/localintermediate.h"
#include "../glslang/MachineIndependent/SymbolTable.h"
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
#include "../glslang/Include/Common.h"
Added -x option to save SPIR-V as 32-bit hexadecimal numbers to a text file.
2016-05-27 17:55:53 +00:00
#include "../glslang/Include/revision.h"
Final round for line endings.
2015-06-26 22:58:36 +00:00
Create a new logger class for all messages w.r.t. SPIR-V build.
2016-05-04 19:55:59 +00:00
#include <fstream>
Added -x option to save SPIR-V as 32-bit hexadecimal numbers to a text file.
2016-05-27 17:55:53 +00:00
#include <iomanip>
Final round for line endings.
2015-06-26 22:58:36 +00:00
#include <list>
Create a new logger class for all messages w.r.t. SPIR-V build.
2016-05-04 19:55:59 +00:00
#include <map>
Final round for line endings.
2015-06-26 22:58:36 +00:00
#include <stack>
Create a new logger class for all messages w.r.t. SPIR-V build.
2016-05-04 19:55:59 +00:00
#include <string>
#include <vector>
Final round for line endings.
2015-06-26 22:58:36 +00:00
namespace {
fix the wrong generated code when a non-constant array is declared with its size derived from spec constant operations
2016-04-01 14:35:16 +00:00
namespace {
class SpecConstantOpModeGuard {
public:
SpecConstantOpModeGuard(spv::Builder* builder)
: builder_(builder) {
previous_flag_ = builder->isInSpecConstCodeGenMode();
}
~SpecConstantOpModeGuard() {
previous_flag_ ? builder_->setToSpecConstCodeGenMode()
: builder_->setToNormalCodeGenMode();
}
Turn on SpecConstantOpMode based on node qualifier Move SpecConstantOpModeGuard from makeSpvConstantFromConstSubTree() to visitbinary() and visitunary(). Checking if the visiting node is a spec constants, if so, turn on the SpecConstantOpMode, otherwise, stay in the normal mode.
2016-04-04 02:20:42 +00:00
void turnOnSpecConstantOpMode() {
builder_->setToSpecConstCodeGenMode();
}
fix the wrong generated code when a non-constant array is declared with its size derived from spec constant operations
2016-04-01 14:35:16 +00:00
private:
spv::Builder* builder_;
bool previous_flag_;
};
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
struct OpDecorations {
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
public:
OpDecorations(spv::Decoration precision, spv::Decoration noContraction, spv::Decoration nonUniform) :
precision(precision)
#ifndef GLSLANG_WEB
,
noContraction(noContraction),
nonUniform(nonUniform)
#endif
{ }
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Decoration precision;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifdef GLSLANG_WEB
void addNoContraction(spv::Builder&, spv::Id) const { };
void addNonUniform(spv::Builder&, spv::Id) const { };
#else
void addNoContraction(spv::Builder& builder, spv::Id t) { builder.addDecoration(t, noContraction); };
void addNonUniform(spv::Builder& builder, spv::Id t) { builder.addDecoration(t, nonUniform); };
protected:
spv::Decoration noContraction;
spv::Decoration nonUniform;
#endif
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
};
} // namespace
fix the wrong generated code when a non-constant array is declared with its size derived from spec constant operations
2016-04-01 14:35:16 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
//
// The main holder of information for translating glslang to SPIR-V.
//
// Derives from the AST walking base class.
//
class TGlslangToSpvTraverser : public glslang::TIntermTraverser {
public:
SPV Version: Emit the requested SPIR-V version, not the header version. Fixes #1236.
2018-02-01 01:35:56 +00:00
TGlslangToSpvTraverser(unsigned int spvVersion, const glslang::TIntermediate*, spv::SpvBuildLogger* logger,
glslang::SpvOptions& options);
Always correctly terminate main. Issue #588, PR #600.
2016-11-26 20:23:20 +00:00
virtual ~TGlslangToSpvTraverser() { }
Final round for line endings.
2015-06-26 22:58:36 +00:00
bool visitAggregate(glslang::TVisit, glslang::TIntermAggregate*);
bool visitBinary(glslang::TVisit, glslang::TIntermBinary*);
void visitConstantUnion(glslang::TIntermConstantUnion*);
bool visitSelection(glslang::TVisit, glslang::TIntermSelection*);
bool visitSwitch(glslang::TVisit, glslang::TIntermSwitch*);
void visitSymbol(glslang::TIntermSymbol* symbol);
bool visitUnary(glslang::TVisit, glslang::TIntermUnary*);
bool visitLoop(glslang::TVisit, glslang::TIntermLoop*);
bool visitBranch(glslang::TVisit visit, glslang::TIntermBranch*);
Always correctly terminate main. Issue #588, PR #600.
2016-11-26 20:23:20 +00:00
void finishSpv();
SPV: Make OpEntryPoint I/O list include all static uses. Before, it was only including explicit interface, sufficient for IO-Block-declared oriented interface, but not sufficient for all modes GLSL might be used with SPIR-V. Two things are accomplished now: 1) each id will appear exactly once 2) the OpEntryPoint list will union static use with declarations
2015-12-21 00:37:07 +00:00
void dumpSpv(std::vector<unsigned int>& out);
Final round for line endings.
2015-06-26 22:58:36 +00:00
protected:
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
TGlslangToSpvTraverser(TGlslangToSpvTraverser&);
TGlslangToSpvTraverser& operator=(TGlslangToSpvTraverser&);
SPV: Add missing extension declarations for certain AMD extensions.
2016-10-14 09:41:45 +00:00
spv::Decoration TranslateInterpolationDecoration(const glslang::TQualifier& qualifier);
SPV: Fix an issue of interpolation decoration. GLSL interpolation qualifiers and auxiliary storage qualifiers are not mutually exclusive. So when they are translated to SPIR-V decorations, two independent utility methods should be employed to do this job.
2016-05-21 01:40:44 +00:00
spv::Decoration TranslateAuxiliaryStorageDecoration(const glslang::TQualifier& qualifier);
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
spv::Decoration TranslateNonUniformDecoration(const glslang::TQualifier& qualifier);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spv::Builder::AccessChain::CoherentFlags TranslateCoherent(const glslang::TType& type);
spv::MemoryAccessMask TranslateMemoryAccess(const spv::Builder::AccessChain::CoherentFlags &coherentFlags);
spv::ImageOperandsMask TranslateImageOperands(const spv::Builder::AccessChain::CoherentFlags &coherentFlags);
spv::Scope TranslateMemoryScope(const spv::Builder::AccessChain::CoherentFlags &coherentFlags);
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
spv::BuiltIn TranslateBuiltInDecoration(glslang::TBuiltInVariable, bool memberDeclaration);
SPV Capabilities: StorageImageExtendedFormats, StorageImageReadWithoutFormat, and StorageImageWriteWithoutFormat.
2016-02-15 18:57:00 +00:00
spv::ImageFormat TranslateImageFormat(const glslang::TType& type);
HLSL: Refactor attribute implementation. - make it sharable with GLSL - correct the case insensitivity - remove the map; queries are not needed, all entries need processing - make it easier to build bottom up (will help GLSL parsing) - support semantic checking and reporting - allow front-end dependent semantics and attribute name mapping
2018-01-30 18:01:39 +00:00
spv::SelectionControlMask TranslateSelectionControl(const glslang::TIntermSelection&) const;
spv::SelectionControlMask TranslateSwitchControl(const glslang::TIntermSwitch&) const;
SPV 1.4: Implement the 5 new loop controls.
2019-01-12 10:31:41 +00:00
spv::LoopControlMask TranslateLoopControl(const glslang::TIntermLoop&, std::vector<unsigned int>& operands) const;
SPV: Emit extension for using StorageClassStorageBuffer.
2017-05-05 11:09:58 +00:00
spv::StorageClass TranslateStorageClass(const glslang::TType&);
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
void addIndirectionIndexCapabilities(const glslang::TType& baseType, const glslang::TType& indexType);
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
spv::Id createSpvVariable(const glslang::TIntermSymbol*, spv::Id forcedType);
Final round for line endings.
2015-06-26 22:58:36 +00:00
spv::Id getSampledType(const glslang::TSampler&);
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
spv::Id getInvertedSwizzleType(const glslang::TIntermTyped&);
spv::Id createInvertedSwizzle(spv::Decoration precision, const glslang::TIntermTyped&, spv::Id parentResult);
void convertSwizzle(const glslang::TIntermAggregate&, std::vector<unsigned>& swizzle);
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
spv::Id convertGlslangToSpvType(const glslang::TType& type, bool forwardReferenceOnly = false);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id convertGlslangToSpvType(const glslang::TType& type, glslang::TLayoutPacking, const glslang::TQualifier&,
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
bool lastBufferBlockMember, bool forwardReferenceOnly = false);
Fix #790: Don't emit NV-specific interface members unless enabled by extension.
2017-03-25 00:38:16 +00:00
bool filterMember(const glslang::TType& member);
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
spv::Id convertGlslangStructToSpvType(const glslang::TType&, const glslang::TTypeList* glslangStruct,
glslang::TLayoutPacking, const glslang::TQualifier&);
void decorateStructType(const glslang::TType&, const glslang::TTypeList* glslangStruct, glslang::TLayoutPacking,
const glslang::TQualifier&, spv::Id);
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
spv::Id makeArraySizeId(const glslang::TArraySizes&, int dim);
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
spv::Id accessChainLoad(const glslang::TType& type);
SPV: Fix an issue caused by bool-to-uint32 conversion. This issue is related with the commit 103bef9d74d768f0690ed53f52681baead384d1e.
2016-02-23 09:51:09 +00:00
void accessChainStore(const glslang::TType& type, spv::Id rvalue);
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
void multiTypeStore(const glslang::TType&, spv::Id rValue);
SPV: Nested layouts: Recursively send around the top-level std140 or std430 packing. This also distinguishes between the same struct included in multiple parent packings.
2015-12-19 20:57:10 +00:00
glslang::TLayoutPacking getExplicitLayout(const glslang::TType& type) const;
SPV: recursively propagate row/col majorness through nested structures. This includes doing structure uniqueness modulo majorness, for shared nested structures.
2015-12-20 18:29:16 +00:00
int getArrayStride(const glslang::TType& arrayType, glslang::TLayoutPacking, glslang::TLayoutMatrix);
int getMatrixStride(const glslang::TType& matrixType, glslang::TLayoutPacking, glslang::TLayoutMatrix);
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
void updateMemberOffset(const glslang::TType& structType, const glslang::TType& memberType, int& currentOffset,
int& nextOffset, glslang::TLayoutPacking, glslang::TLayoutMatrix);
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
void declareUseOfStructMember(const glslang::TTypeList& members, int glslangMember);
Final round for line endings.
2015-06-26 22:58:36 +00:00
Non-functional: Sweep through the stack for consistent with "main" and entry point. Partially addresses issue #513.
2016-09-19 22:01:41 +00:00
bool isShaderEntryPoint(const glslang::TIntermAggregate* node);
Fix #944: Convert argument type to match formal parameter type.
2018-05-04 17:43:03 +00:00
bool writableParam(glslang::TStorageQualifier) const;
SPV: rationalize parameter handling for "original" and "writable" parameters.
2017-09-10 21:21:05 +00:00
bool originalParam(glslang::TStorageQualifier, const glslang::TType&, bool implicitThisParam);
Final round for line endings.
2015-06-26 22:58:36 +00:00
void makeFunctions(const glslang::TIntermSequence&);
void makeGlobalInitializers(const glslang::TIntermSequence&);
void visitFunctions(const glslang::TIntermSequence&);
void handleFunctionEntry(const glslang::TIntermAggregate* node);
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
void translateArguments(const glslang::TIntermAggregate& node, std::vector<spv::Id>& arguments, spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags);
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
void translateArguments(glslang::TIntermUnary& node, std::vector<spv::Id>& arguments);
spv::Id createImageTextureFunctionCall(glslang::TIntermOperator* node);
Final round for line endings.
2015-06-26 22:58:36 +00:00
spv::Id handleUserFunctionCall(const glslang::TIntermAggregate*);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id createBinaryOperation(glslang::TOperator op, OpDecorations&, spv::Id typeId, spv::Id left, spv::Id right,
glslang::TBasicType typeProxy, bool reduceComparison = true);
spv::Id createBinaryMatrixOperation(spv::Op, OpDecorations&, spv::Id typeId, spv::Id left, spv::Id right);
spv::Id createUnaryOperation(glslang::TOperator op, OpDecorations&, spv::Id typeId, spv::Id operand,
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
glslang::TBasicType typeProxy, const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id createUnaryMatrixOperation(spv::Op op, OpDecorations&, spv::Id typeId, spv::Id operand,
glslang::TBasicType typeProxy);
spv::Id createConversion(glslang::TOperator op, OpDecorations&, spv::Id destTypeId, spv::Id operand,
glslang::TBasicType typeProxy);
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
spv::Id createIntWidthConversion(glslang::TOperator op, spv::Id operand, int vectorSize);
Final round for line endings.
2015-06-26 22:58:36 +00:00
spv::Id makeSmearedConstant(spv::Id constant, int vectorSize);
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
spv::Id createAtomicOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy, const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags);
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
spv::Id createInvocationsOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy);
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
spv::Id CreateInvocationsVectorOperation(spv::Op op, spv::GroupOperation groupOperation, spv::Id typeId, std::vector<spv::Id>& operands);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
spv::Id createSubgroupOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy);
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
spv::Id createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy);
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
spv::Id createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId);
Final round for line endings.
2015-06-26 22:58:36 +00:00
spv::Id getSymbolId(const glslang::TIntermSymbol* node);
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
void addMeshNVDecoration(spv::Id id, int member, const glslang::TQualifier & qualifier);
Support specialization composite constants Fix issue #163, support creation and reference of composite type specialization constants. e.g.: ``` layout(constant_id = 200) const float myfloat = 1.25; layout(constant_id = 201) const int myint = 14; struct structtype { float f; int i; }; const structtype outer_struct_var = {myfloat, myint}; void main(){} ``` generated code (use glslangValidator): ``` // Module Version 10000 // Generated by (magic number): 80001 // Id's are bound by 12 Capability Shader 1: ExtInstImport "GLSL.std.450" MemoryModel Logical GLSL450 EntryPoint Vertex 4 "main" Source GLSL 450 Name 4 "main" Name 10 "structtype" MemberName 10(structtype) 0 "f" MemberName 10(structtype) 1 "i" Decorate 7 SpecId 200 Decorate 9 SpecId 201 2: TypeVoid 3: TypeFunction 2 6: TypeFloat 32 7: 6(float) SpecConstant 1067450368 8: TypeInt 32 1 9: 8(int) SpecConstant 14 10(structtype): TypeStruct 6(float) 8(int) 11:10(structtype) SpecConstantComposite 7 9 4(main): 2 Function None 3 5: Label Return FunctionEnd ``` Rname two function names to match their functionalities. 1) Rename `GlslangToSpvTraverser::createSpvSpecConstant()` to `createSpvConstant()`; 2) Rename `GlslangToSpvTraverser::createSpvConstant()` to `createSpvConstantFromConstUnionArray()` Add function `GlslangToSpvTraverser::createSpvConstantFromSubTree()` to handle constant creation from sub trees (e.g.: specialization constants). Related PR: #208
2016-03-21 13:51:37 +00:00
spv::Id createSpvConstant(const glslang::TIntermTyped&);
spv::Id createSpvConstantFromConstUnionArray(const glslang::TType& type, const glslang::TConstUnionArray&, int& nextConst, bool specConstant);
SPV: Implement short circuiting of && and || when emitting SPIR-V.
2015-10-15 19:29:11 +00:00
bool isTrivialLeaf(const glslang::TIntermTyped* node);
bool isTrivial(const glslang::TIntermTyped* node);
spv::Id createShortCircuit(glslang::TOperator, glslang::TIntermTyped& left, glslang::TIntermTyped& right);
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
spv::Id getExtBuiltins(const char* name);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
void addPre13Extension(const char* ext)
{
if (builder.getSpvVersion() < glslang::EShTargetSpv_1_3)
builder.addExtension(ext);
}
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
std::pair<spv::Id, spv::Id> getForcedType(spv::BuiltIn, const glslang::TType&);
spv::Id translateForcedType(spv::Id object);
Handle SPIR-V type mismatch when constructing a composite
2019-06-25 18:31:10 +00:00
spv::Id createCompositeConstruct(spv::Id typeId, std::vector<spv::Id> constituents);
Final round for line endings.
2015-06-26 22:58:36 +00:00
SPV: Add OpSource shader source code and file name.
2017-05-31 23:11:16 +00:00
glslang::SpvOptions& options;
Final round for line endings.
2015-06-26 22:58:36 +00:00
spv::Function* shaderEntry;
HLSL: Do structure conversion for return type struct-punning on non-entry-point functions.
2016-10-06 18:59:51 +00:00
spv::Function* currentFunction;
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
spv::Instruction* entryPoint;
Final round for line endings.
2015-06-26 22:58:36 +00:00
int sequenceDepth;
Create a new logger class for all messages w.r.t. SPIR-V build.
2016-05-04 19:55:59 +00:00
spv::SpvBuildLogger* logger;
Avoid printing to stdout directly in library functions. Previously GlslangToSpv() reported missing/TBD functionalities by directly writing to stdout using printf. That could cause problems to callers of GlslangToSpv(). This patch cleans up the error reporting logic in GlslangToSpv(), TGlslangToSpvTraverser, and spv::Builder a little bit to use ostringstream. Also fixed the usage of GlslangToSpv() in GTest fixtures to capture warnings/errors reported when translating AST to SPIR-V.
2016-05-02 22:11:54 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
// There is a 1:1 mapping between a spv builder and a module; this is thread safe
spv::Builder builder;
Non-functional: Rename some entry-point variables to entryPoint, not main.
2016-11-26 20:31:47 +00:00
bool inEntryPoint;
bool entryPointTerminated;
SPV: Make OpEntryPoint I/O list include all static uses. Before, it was only including explicit interface, sufficient for IO-Block-declared oriented interface, but not sufficient for all modes GLSL might be used with SPIR-V. Two things are accomplished now: 1) each id will appear exactly once 2) the OpEntryPoint list will union static use with declarations
2015-12-21 00:37:07 +00:00
bool linkageOnly; // true when visiting the set of objects in the AST present only for establishing interface, whether or not they were statically used
SPV: Move to deterministic results for OpEntryPoint, for test stability across platforms.
2015-12-21 18:45:34 +00:00
std::set<spv::Id> iOSet; // all input/output variables from either static use or declaration of interface
Final round for line endings.
2015-06-26 22:58:36 +00:00
const glslang::TIntermediate* glslangIntermediate;
SPV: Add a switch for favoring non-NaN operands in min, max, and clamp.
2019-06-18 05:33:09 +00:00
bool nanMinMaxClamp; // true if use NMin/NMax/NClamp instead of FMin/FMax/FClamp
Final round for line endings.
2015-06-26 22:58:36 +00:00
spv::Id stdBuiltins;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
std::unordered_map<const char*, spv::Id> extBuiltinMap;
Final round for line endings.
2015-06-26 22:58:36 +00:00
Performance: Simple upgrade/cleanup of stl c++11 containers (10% perf. increase).
2015-07-19 04:34:27 +00:00
std::unordered_map<int, spv::Id> symbolValues;
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
std::unordered_set<int> rValueParameters; // set of formal function parameters passed as rValues, rather than a pointer
Performance: Simple upgrade/cleanup of stl c++11 containers (10% perf. increase).
2015-07-19 04:34:27 +00:00
std::unordered_map<std::string, spv::Function*> functionMap;
SPV: recursively propagate row/col majorness through nested structures. This includes doing structure uniqueness modulo majorness, for shared nested structures.
2015-12-20 18:29:16 +00:00
std::unordered_map<const glslang::TTypeList*, spv::Id> structMap[glslang::ElpCount][glslang::ElmCount];
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
// for mapping glslang block indices to spv indices (e.g., due to hidden members):
std::unordered_map<const glslang::TTypeList*, std::vector<int> > memberRemapper;
Final round for line endings.
2015-06-26 22:58:36 +00:00
std::stack<bool> breakForLoop; // false means break for switch
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
std::unordered_map<std::string, const glslang::TIntermSymbol*> counterOriginator;
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
// Map pointee types for EbtReference to their forward pointers
std::map<const glslang::TType *, spv::Id> forwardPointers;
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
// Type forcing, for when SPIR-V wants a different type than the AST,
// requiring local translation to and from SPIR-V type on every access.
// Maps <builtin-variable-id -> AST-required-type-id>
std::unordered_map<spv::Id, spv::Id> forceType;
Final round for line endings.
2015-06-26 22:58:36 +00:00
};
//
// Helper functions for translating glslang representations to SPIR-V enumerants.
//
// Translate glslang profile to SPIR-V source language.
Support multiple source languages, adding HLSL as an option.
2016-03-13 01:34:36 +00:00
spv::SourceLanguage TranslateSourceLanguage(glslang::EShSource source, EProfile profile)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#ifdef GLSLANG_WEB
return spv::SourceLanguageESSL;
#endif
Support multiple source languages, adding HLSL as an option.
2016-03-13 01:34:36 +00:00
switch (source) {
case glslang::EShSourceGlsl:
switch (profile) {
case ENoProfile:
case ECoreProfile:
case ECompatibilityProfile:
return spv::SourceLanguageGLSL;
case EEsProfile:
return spv::SourceLanguageESSL;
default:
return spv::SourceLanguageUnknown;
}
case glslang::EShSourceHlsl:
HLSL: Emit the OpSource HLSL instruction for HLSL, using new headers.
2017-04-07 21:33:08 +00:00
return spv::SourceLanguageHLSL;
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
return spv::SourceLanguageUnknown;
}
}
// Translate glslang language (stage) to SPIR-V execution model.
spv::ExecutionModel TranslateExecutionModel(EShLanguage stage)
{
switch (stage) {
case EShLangVertex: return spv::ExecutionModelVertex;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
case EShLangFragment: return spv::ExecutionModelFragment;
#ifndef GLSLANG_WEB
case EShLangCompute: return spv::ExecutionModelGLCompute;
Final round for line endings.
2015-06-26 22:58:36 +00:00
case EShLangTessControl: return spv::ExecutionModelTessellationControl;
case EShLangTessEvaluation: return spv::ExecutionModelTessellationEvaluation;
case EShLangGeometry: return spv::ExecutionModelGeometry;
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
case EShLangRayGenNV: return spv::ExecutionModelRayGenerationNV;
case EShLangIntersectNV: return spv::ExecutionModelIntersectionNV;
case EShLangAnyHitNV: return spv::ExecutionModelAnyHitNV;
case EShLangClosestHitNV: return spv::ExecutionModelClosestHitNV;
case EShLangMissNV: return spv::ExecutionModelMissNV;
case EShLangCallableNV: return spv::ExecutionModelCallableNV;
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
case EShLangTaskNV: return spv::ExecutionModelTaskNV;
case EShLangMeshNV: return spv::ExecutionModelMeshNV;
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(0);
Final round for line endings.
2015-06-26 22:58:36 +00:00
return spv::ExecutionModelFragment;
}
}
// Translate glslang sampler type to SPIR-V dimensionality.
spv::Dim TranslateDimensionality(const glslang::TSampler& sampler)
{
switch (sampler.dim) {
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
case glslang::Esd1D: return spv::Dim1D;
case glslang::Esd2D: return spv::Dim2D;
case glslang::Esd3D: return spv::Dim3D;
case glslang::EsdCube: return spv::DimCube;
case glslang::EsdRect: return spv::DimRect;
case glslang::EsdBuffer: return spv::DimBuffer;
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
case glslang::EsdSubpass: return spv::DimSubpassData;
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(0);
Final round for line endings.
2015-06-26 22:58:36 +00:00
return spv::Dim2D;
}
}
Front-end: Implement 2nd task of issue #400; precision of result and operation. From the ES spec + Bugzilla 15931 and GL_KHR_vulkan_glsl: - Update precision qualifiers for all built-in function prototypes. - Implement the new algorithm used to distinguish built-in function operation precisions from result precisions. Also add tracking of separate result and operation precisions, and use that in generating SPIR-V. (SPIR-V cares about precision of operation, while the front-end cares about precision of result, for propagation.)
2016-08-02 01:44:00 +00:00
// Translate glslang precision to SPIR-V precision decorations.
spv::Decoration TranslatePrecisionDecoration(glslang::TPrecisionQualifier glslangPrecision)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
Front-end: Implement 2nd task of issue #400; precision of result and operation. From the ES spec + Bugzilla 15931 and GL_KHR_vulkan_glsl: - Update precision qualifiers for all built-in function prototypes. - Implement the new algorithm used to distinguish built-in function operation precisions from result precisions. Also add tracking of separate result and operation precisions, and use that in generating SPIR-V. (SPIR-V cares about precision of operation, while the front-end cares about precision of result, for propagation.)
2016-08-02 01:44:00 +00:00
switch (glslangPrecision) {
Constant folding: Correct result type of non-square matrix folding. This also made the function easier to read by identifying left and right operands more clearly.
2015-12-15 01:21:19 +00:00
case glslang::EpqLow: return spv::DecorationRelaxedPrecision;
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
case glslang::EpqMedium: return spv::DecorationRelaxedPrecision;
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
return spv::NoPrecision;
}
}
Front-end: Implement 2nd task of issue #400; precision of result and operation. From the ES spec + Bugzilla 15931 and GL_KHR_vulkan_glsl: - Update precision qualifiers for all built-in function prototypes. - Implement the new algorithm used to distinguish built-in function operation precisions from result precisions. Also add tracking of separate result and operation precisions, and use that in generating SPIR-V. (SPIR-V cares about precision of operation, while the front-end cares about precision of result, for propagation.)
2016-08-02 01:44:00 +00:00
// Translate glslang type to SPIR-V precision decorations.
spv::Decoration TranslatePrecisionDecoration(const glslang::TType& type)
{
return TranslatePrecisionDecoration(type.getQualifier().precision);
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
// Translate glslang type to SPIR-V block decorations.
SPV: Support test #pragma for generating the StorageBuffer storage class. Longer term, this storage class should be generated based on the mode of compilation.
2017-04-20 00:34:49 +00:00
spv::Decoration TranslateBlockDecoration(const glslang::TType& type, bool useStorageBuffer)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
if (type.getBasicType() == glslang::EbtBlock) {
switch (type.getQualifier().storage) {
case glslang::EvqUniform: return spv::DecorationBlock;
SPV: Support test #pragma for generating the StorageBuffer storage class. Longer term, this storage class should be generated based on the mode of compilation.
2017-04-20 00:34:49 +00:00
case glslang::EvqBuffer: return useStorageBuffer ? spv::DecorationBlock : spv::DecorationBufferBlock;
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EvqVaryingIn: return spv::DecorationBlock;
case glslang::EvqVaryingOut: return spv::DecorationBlock;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EvqPayloadNV: return spv::DecorationBlock;
case glslang::EvqPayloadInNV: return spv::DecorationBlock;
case glslang::EvqHitAttrNV: return spv::DecorationBlock;
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
case glslang::EvqCallableDataNV: return spv::DecorationBlock;
case glslang::EvqCallableDataInNV: return spv::DecorationBlock;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(0);
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
}
}
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
return spv::DecorationMax;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
SPV: Add support for memory qualifiers.
2016-02-21 12:59:01 +00:00
// Translate glslang type to SPIR-V memory decorations.
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
void TranslateMemoryDecoration(const glslang::TQualifier& qualifier, std::vector<spv::Decoration>& memory, bool useVulkanMemoryModel)
SPV: Add support for memory qualifiers.
2016-02-21 12:59:01 +00:00
{
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
if (!useVulkanMemoryModel) {
if (qualifier.coherent)
memory.push_back(spv::DecorationCoherent);
if (qualifier.volatil) {
memory.push_back(spv::DecorationVolatile);
memory.push_back(spv::DecorationCoherent);
}
Fix #1395: GLSL volatile maps to SPIR-V Volatile and Coherent. The major version number was bumped in the previous commit to support this.
2018-06-04 21:36:03 +00:00
}
SPV: Add support for memory qualifiers.
2016-02-21 12:59:01 +00:00
if (qualifier.restrict)
memory.push_back(spv::DecorationRestrict);
if (qualifier.readonly)
memory.push_back(spv::DecorationNonWritable);
if (qualifier.writeonly)
memory.push_back(spv::DecorationNonReadable);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
SPV: Add support for memory qualifiers.
2016-02-21 12:59:01 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
// Translate glslang type to SPIR-V layout decorations.
SPV: recursively propagate row/col majorness through nested structures. This includes doing structure uniqueness modulo majorness, for shared nested structures.
2015-12-20 18:29:16 +00:00
spv::Decoration TranslateLayoutDecoration(const glslang::TType& type, glslang::TLayoutMatrix matrixLayout)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
if (type.isMatrix()) {
SPV: recursively propagate row/col majorness through nested structures. This includes doing structure uniqueness modulo majorness, for shared nested structures.
2015-12-20 18:29:16 +00:00
switch (matrixLayout) {
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::ElmRowMajor:
return spv::DecorationRowMajor;
SPV: recursively propagate row/col majorness through nested structures. This includes doing structure uniqueness modulo majorness, for shared nested structures.
2015-12-20 18:29:16 +00:00
case glslang::ElmColumnMajor:
Final round for line endings.
2015-06-26 22:58:36 +00:00
return spv::DecorationColMajor;
SPV: recursively propagate row/col majorness through nested structures. This includes doing structure uniqueness modulo majorness, for shared nested structures.
2015-12-20 18:29:16 +00:00
default:
// opaque layouts don't need a majorness
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
return spv::DecorationMax;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
} else {
switch (type.getBasicType()) {
default:
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
return spv::DecorationMax;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EbtBlock:
switch (type.getQualifier().storage) {
case glslang::EvqUniform:
case glslang::EvqBuffer:
switch (type.getQualifier().layoutPacking) {
case glslang::ElpShared: return spv::DecorationGLSLShared;
case glslang::ElpPacked: return spv::DecorationGLSLPacked;
default:
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
return spv::DecorationMax;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
case glslang::EvqVaryingIn:
case glslang::EvqVaryingOut:
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
if (type.getQualifier().isTaskMemory()) {
switch (type.getQualifier().layoutPacking) {
case glslang::ElpShared: return spv::DecorationGLSLShared;
case glslang::ElpPacked: return spv::DecorationGLSLPacked;
default: break;
}
} else {
assert(type.getQualifier().layoutPacking == glslang::ElpNone);
}
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
return spv::DecorationMax;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EvqPayloadNV:
case glslang::EvqPayloadInNV:
case glslang::EvqHitAttrNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
case glslang::EvqCallableDataNV:
case glslang::EvqCallableDataInNV:
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
return spv::DecorationMax;
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(0);
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
return spv::DecorationMax;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
}
}
}
// Translate glslang type to SPIR-V interpolation decorations.
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
// Returns spv::DecorationMax when no decoration
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
// should be applied.
SPV: Add missing extension declarations for certain AMD extensions.
2016-10-14 09:41:45 +00:00
spv::Decoration TGlslangToSpvTraverser::TranslateInterpolationDecoration(const glslang::TQualifier& qualifier)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
SPV: Fix an issue of interpolation decoration. GLSL interpolation qualifiers and auxiliary storage qualifiers are not mutually exclusive. So when they are translated to SPIR-V decorations, two independent utility methods should be employed to do this job.
2016-05-21 01:40:44 +00:00
if (qualifier.smooth)
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
// Smooth decoration doesn't exist in SPIR-V 1.0
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
return spv::DecorationMax;
Web: Remove/rationalize a set of *_EXTENSIONS, using GLSLANG_WEB. Focus was on the front end (not SPIR-V), minus the grammar. Reduces #ifdef count by around 320 and makes the web build 270K smaller, which is about 90% the target size. The grammar and scanner will be another step, as will the SPIR-V backend. This makes heavy use of methods #ifdef'd to return false as a global way of turning off code, relying on C++ DCE to do the rest.
2019-08-01 09:28:08 +00:00
else if (qualifier.isNonPerspective())
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
return spv::DecorationNoPerspective;
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
else if (qualifier.flat)
Final round for line endings.
2015-06-26 22:58:36 +00:00
return spv::DecorationFlat;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
else if (qualifier.isExplicitInterpolation()) {
SPV: Add missing extension declarations for certain AMD extensions.
2016-10-14 09:41:45 +00:00
builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
return spv::DecorationExplicitInterpAMD;
SPV: Add missing extension declarations for certain AMD extensions.
2016-10-14 09:41:45 +00:00
}
SPV: Fix an issue of interpolation decoration. GLSL interpolation qualifiers and auxiliary storage qualifiers are not mutually exclusive. So when they are translated to SPIR-V decorations, two independent utility methods should be employed to do this job.
2016-05-21 01:40:44 +00:00
else
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
return spv::DecorationMax;
SPV: Fix an issue of interpolation decoration. GLSL interpolation qualifiers and auxiliary storage qualifiers are not mutually exclusive. So when they are translated to SPIR-V decorations, two independent utility methods should be employed to do this job.
2016-05-21 01:40:44 +00:00
}
// Translate glslang type to SPIR-V auxiliary storage decorations.
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
// Returns spv::DecorationMax when no decoration
SPV: Fix an issue of interpolation decoration. GLSL interpolation qualifiers and auxiliary storage qualifiers are not mutually exclusive. So when they are translated to SPIR-V decorations, two independent utility methods should be employed to do this job.
2016-05-21 01:40:44 +00:00
// should be applied.
spv::Decoration TGlslangToSpvTraverser::TranslateAuxiliaryStorageDecoration(const glslang::TQualifier& qualifier)
{
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
if (qualifier.centroid)
Final round for line endings.
2015-06-26 22:58:36 +00:00
return spv::DecorationCentroid;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
else if (qualifier.patch)
return spv::DecorationPatch;
SPV Capabilities: SampleRateShading, SparseResidency, MinLod, and ImageQuery.
2016-02-15 18:09:46 +00:00
else if (qualifier.sample) {
builder.addCapability(spv::CapabilitySampleRateShading);
Final round for line endings.
2015-06-26 22:58:36 +00:00
return spv::DecorationSample;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
}
#endif
return spv::DecorationMax;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
// If glslang type is invariant, return SPIR-V invariant decoration.
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
spv::Decoration TranslateInvariantDecoration(const glslang::TQualifier& qualifier)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
if (qualifier.invariant)
Final round for line endings.
2015-06-26 22:58:36 +00:00
return spv::DecorationInvariant;
else
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
return spv::DecorationMax;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
Precise and noContraction propagation Reimplement the whole workflow to make that: precise'ness of struct members won't spread to other non-precise members of the same struct instance. Approach: 1. Build the map from symbols to their defining nodes. And for each object node (StructIndex, DirectIndex, Symbol nodes, etc), generates an accesschain path. Different AST nodes that indicating a same object should have the same accesschain path. 2. Along the building phase in step 1, collect the initial set of 'precise' (AST qualifier: 'noContraction') objects' accesschain paths. 3. Start with the initial set of 'precise' accesschain paths, use it as a worklist, do as the following steps until the worklist is empty: 1) Pop an accesschain path from worklist. 2) Get the symbol part from the accesschain path. 3) Find the defining nodes of that symbol. 4) For each defining node, check whether it is defining a 'precise' object, or its assignee has nested 'precise' object. Get the incremental path from assignee to its nested 'precise' object (if any). 5) Traverse the right side of the defining node, obtain the accesschain paths of the corresponding involved 'precise' objects. Update the worklist with those new objects' accesschain paths. Label involved operations with 'noContraction'. In each step, whenever we find the parent object of an nested object is 'precise' (has 'noContraction' qualifier), we let the nested object inherit the 'precise'ness from its parent object.
2016-05-04 21:34:38 +00:00
// If glslang type is noContraction, return SPIR-V NoContraction decoration.
spv::Decoration TranslateNoContractionDecoration(const glslang::TQualifier& qualifier)
{
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
if (qualifier.isNoContraction())
Precise and noContraction propagation Reimplement the whole workflow to make that: precise'ness of struct members won't spread to other non-precise members of the same struct instance. Approach: 1. Build the map from symbols to their defining nodes. And for each object node (StructIndex, DirectIndex, Symbol nodes, etc), generates an accesschain path. Different AST nodes that indicating a same object should have the same accesschain path. 2. Along the building phase in step 1, collect the initial set of 'precise' (AST qualifier: 'noContraction') objects' accesschain paths. 3. Start with the initial set of 'precise' accesschain paths, use it as a worklist, do as the following steps until the worklist is empty: 1) Pop an accesschain path from worklist. 2) Get the symbol part from the accesschain path. 3) Find the defining nodes of that symbol. 4) For each defining node, check whether it is defining a 'precise' object, or its assignee has nested 'precise' object. Get the incremental path from assignee to its nested 'precise' object (if any). 5) Traverse the right side of the defining node, obtain the accesschain paths of the corresponding involved 'precise' objects. Update the worklist with those new objects' accesschain paths. Label involved operations with 'noContraction'. In each step, whenever we find the parent object of an nested object is 'precise' (has 'noContraction' qualifier), we let the nested object inherit the 'precise'ness from its parent object.
2016-05-04 21:34:38 +00:00
return spv::DecorationNoContraction;
else
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
return spv::DecorationMax;
Precise and noContraction propagation Reimplement the whole workflow to make that: precise'ness of struct members won't spread to other non-precise members of the same struct instance. Approach: 1. Build the map from symbols to their defining nodes. And for each object node (StructIndex, DirectIndex, Symbol nodes, etc), generates an accesschain path. Different AST nodes that indicating a same object should have the same accesschain path. 2. Along the building phase in step 1, collect the initial set of 'precise' (AST qualifier: 'noContraction') objects' accesschain paths. 3. Start with the initial set of 'precise' accesschain paths, use it as a worklist, do as the following steps until the worklist is empty: 1) Pop an accesschain path from worklist. 2) Get the symbol part from the accesschain path. 3) Find the defining nodes of that symbol. 4) For each defining node, check whether it is defining a 'precise' object, or its assignee has nested 'precise' object. Get the incremental path from assignee to its nested 'precise' object (if any). 5) Traverse the right side of the defining node, obtain the accesschain paths of the corresponding involved 'precise' objects. Update the worklist with those new objects' accesschain paths. Label involved operations with 'noContraction'. In each step, whenever we find the parent object of an nested object is 'precise' (has 'noContraction' qualifier), we let the nested object inherit the 'precise'ness from its parent object.
2016-05-04 21:34:38 +00:00
}
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
// If glslang type is nonUniform, return SPIR-V NonUniform decoration.
spv::Decoration TGlslangToSpvTraverser::TranslateNonUniformDecoration(const glslang::TQualifier& qualifier)
{
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
if (qualifier.isNonUniform()) {
builder.addExtension("SPV_EXT_descriptor_indexing");
builder.addCapability(spv::CapabilityShaderNonUniformEXT);
return spv::DecorationNonUniformEXT;
} else
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
return spv::DecorationMax;
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
spv::MemoryAccessMask TGlslangToSpvTraverser::TranslateMemoryAccess(
const spv::Builder::AccessChain::CoherentFlags &coherentFlags)
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
{
spv::MemoryAccessMask mask = spv::MemoryAccessMaskNone;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
if (!glslangIntermediate->usingVulkanMemoryModel() || coherentFlags.isImage)
return mask;
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
if (coherentFlags.volatil ||
coherentFlags.coherent ||
coherentFlags.devicecoherent ||
coherentFlags.queuefamilycoherent ||
coherentFlags.workgroupcoherent ||
coherentFlags.subgroupcoherent) {
mask = mask | spv::MemoryAccessMakePointerAvailableKHRMask |
spv::MemoryAccessMakePointerVisibleKHRMask;
}
if (coherentFlags.nonprivate) {
mask = mask | spv::MemoryAccessNonPrivatePointerKHRMask;
}
if (coherentFlags.volatil) {
mask = mask | spv::MemoryAccessVolatileMask;
}
if (mask != spv::MemoryAccessMaskNone) {
builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
return mask;
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
spv::ImageOperandsMask TGlslangToSpvTraverser::TranslateImageOperands(
const spv::Builder::AccessChain::CoherentFlags &coherentFlags)
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
{
spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
if (!glslangIntermediate->usingVulkanMemoryModel())
return mask;
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
if (coherentFlags.volatil ||
coherentFlags.coherent ||
coherentFlags.devicecoherent ||
coherentFlags.queuefamilycoherent ||
coherentFlags.workgroupcoherent ||
coherentFlags.subgroupcoherent) {
mask = mask | spv::ImageOperandsMakeTexelAvailableKHRMask |
spv::ImageOperandsMakeTexelVisibleKHRMask;
}
if (coherentFlags.nonprivate) {
mask = mask | spv::ImageOperandsNonPrivateTexelKHRMask;
}
if (coherentFlags.volatil) {
mask = mask | spv::ImageOperandsVolatileTexelKHRMask;
}
if (mask != spv::ImageOperandsMaskNone) {
builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
return mask;
}
spv::Builder::AccessChain::CoherentFlags TGlslangToSpvTraverser::TranslateCoherent(const glslang::TType& type)
{
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
spv::Builder::AccessChain::CoherentFlags flags = {};
#ifndef GLSLANG_WEB
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
flags.coherent = type.getQualifier().coherent;
flags.devicecoherent = type.getQualifier().devicecoherent;
flags.queuefamilycoherent = type.getQualifier().queuefamilycoherent;
// shared variables are implicitly workgroupcoherent in GLSL.
flags.workgroupcoherent = type.getQualifier().workgroupcoherent ||
type.getQualifier().storage == glslang::EvqShared;
flags.subgroupcoherent = type.getQualifier().subgroupcoherent;
Fix interactions between 'volatile' and the Vulkan memory model Last year we changed 'volatile' to also act as 'coherent', but when I resolved the memory model changes against that change I missed handling volatile in a couple places that we check for coherent. There was also a place in post-processing that acted as if the volatile memory access flag has a literal number associated with it, when it doesn't.
2019-03-05 20:40:07 +00:00
flags.volatil = type.getQualifier().volatil;
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
// *coherent variables are implicitly nonprivate in GLSL
flags.nonprivate = type.getQualifier().nonprivate ||
Fix flags.nonprivate to be true for workgroup memory, which is implicitly workgroupcoherent/nonprivate
2018-10-16 03:46:48 +00:00
flags.subgroupcoherent ||
flags.workgroupcoherent ||
flags.queuefamilycoherent ||
flags.devicecoherent ||
Fix interactions between 'volatile' and the Vulkan memory model Last year we changed 'volatile' to also act as 'coherent', but when I resolved the memory model changes against that change I missed handling volatile in a couple places that we check for coherent. There was also a place in post-processing that acted as if the volatile memory access flag has a literal number associated with it, when it doesn't.
2019-03-05 20:40:07 +00:00
flags.coherent ||
flags.volatil;
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
flags.isImage = type.getBasicType() == glslang::EbtSampler;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
return flags;
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
spv::Scope TGlslangToSpvTraverser::TranslateMemoryScope(
const spv::Builder::AccessChain::CoherentFlags &coherentFlags)
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
{
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
spv::Scope scope = spv::ScopeMax;
#ifndef GLSLANG_WEB
Fix interactions between 'volatile' and the Vulkan memory model Last year we changed 'volatile' to also act as 'coherent', but when I resolved the memory model changes against that change I missed handling volatile in a couple places that we check for coherent. There was also a place in post-processing that acted as if the volatile memory access flag has a literal number associated with it, when it doesn't.
2019-03-05 20:40:07 +00:00
if (coherentFlags.volatil || coherentFlags.coherent) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
// coherent defaults to Device scope in the old model, QueueFamilyKHR scope in the new model
scope = glslangIntermediate->usingVulkanMemoryModel() ? spv::ScopeQueueFamilyKHR : spv::ScopeDevice;
} else if (coherentFlags.devicecoherent) {
scope = spv::ScopeDevice;
} else if (coherentFlags.queuefamilycoherent) {
scope = spv::ScopeQueueFamilyKHR;
} else if (coherentFlags.workgroupcoherent) {
scope = spv::ScopeWorkgroup;
} else if (coherentFlags.subgroupcoherent) {
scope = spv::ScopeSubgroup;
}
if (glslangIntermediate->usingVulkanMemoryModel() && scope == spv::ScopeDevice) {
builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR);
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
return scope;
}
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
// Translate a glslang built-in variable to a SPIR-V built in decoration. Also generate
// associated capabilities when required. For some built-in variables, a capability
// is generated only when using the variable in an executable instruction, but not when
// just declaring a struct member variable with it. This is true for PointSize,
// ClipDistance, and CullDistance.
spv::BuiltIn TGlslangToSpvTraverser::TranslateBuiltInDecoration(glslang::TBuiltInVariable builtIn, bool memberDeclaration)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
switch (builtIn) {
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
case glslang::EbvPointSize:
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#ifndef GLSLANG_WEB
SPV: Fix issue #376; declare Geometry capability for fragment use of gl_Layer.
2016-07-08 20:05:15 +00:00
// Defer adding the capability until the built-in is actually used.
if (! memberDeclaration) {
switch (glslangIntermediate->getStage()) {
case EShLangGeometry:
builder.addCapability(spv::CapabilityGeometryPointSize);
break;
case EShLangTessControl:
case EShLangTessEvaluation:
builder.addCapability(spv::CapabilityTessellationPointSize);
break;
default:
break;
}
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
}
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#endif
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
return spv::BuiltInPointSize;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
case glslang::EbvPosition: return spv::BuiltInPosition;
case glslang::EbvVertexId: return spv::BuiltInVertexId;
case glslang::EbvInstanceId: return spv::BuiltInInstanceId;
case glslang::EbvVertexIndex: return spv::BuiltInVertexIndex;
case glslang::EbvInstanceIndex: return spv::BuiltInInstanceIndex;
case glslang::EbvFragCoord: return spv::BuiltInFragCoord;
case glslang::EbvPointCoord: return spv::BuiltInPointCoord;
case glslang::EbvFace: return spv::BuiltInFrontFacing;
case glslang::EbvFragDepth: return spv::BuiltInFragDepth;
#ifndef GLSLANG_WEB
SPV: Don't add clip/cull distance capabilities unless used. These capabalities were added on declaration of the members, but that is considered too aggressive, as those members are automatically declared in some shaders that don't use them. Now, actual access is needed to make the capabalities be declared.
2016-05-17 01:22:05 +00:00
// These *Distance capabilities logically belong here, but if the member is declared and
// then never used, consumers of SPIR-V prefer the capability not be declared.
// They are now generated when used, rather than here when declared.
// Potentially, the specification should be more clear what the minimum
// use needed is to trigger the capability.
//
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
case glslang::EbvClipDistance:
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
if (!memberDeclaration)
SPV: Unexpected declarations of capabilities from NV extensions.
2017-02-22 08:44:48 +00:00
builder.addCapability(spv::CapabilityClipDistance);
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
return spv::BuiltInClipDistance;
case glslang::EbvCullDistance:
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
if (!memberDeclaration)
SPV: Unexpected declarations of capabilities from NV extensions.
2017-02-22 08:44:48 +00:00
builder.addCapability(spv::CapabilityCullDistance);
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
return spv::BuiltInCullDistance;
case glslang::EbvViewportIndex:
GLSL: Make gl_Layer and gl_ViewportIndex always be outside blocks. There was some ambiguity/contradiction in this behavior, and Khronos decided glslang should always have these outside blocks, rather than have stage/vendor/target variations.
2017-09-13 19:22:50 +00:00
builder.addCapability(spv::CapabilityMultiViewport);
if (glslangIntermediate->getStage() == EShLangVertex ||
glslangIntermediate->getStage() == EShLangTessControl ||
glslangIntermediate->getStage() == EShLangTessEvaluation) {
SPV: Fix unexpected declarations of capability and extension
2017-03-16 15:02:39 +00:00
GLSL: Make gl_Layer and gl_ViewportIndex always be outside blocks. There was some ambiguity/contradiction in this behavior, and Khronos decided glslang should always have these outside blocks, rather than have stage/vendor/target variations.
2017-09-13 19:22:50 +00:00
builder.addExtension(spv::E_SPV_EXT_shader_viewport_index_layer);
builder.addCapability(spv::CapabilityShaderViewportIndexLayerEXT);
SPV: Fix unexpected declarations of capability and extension
2017-03-16 15:02:39 +00:00
}
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
return spv::BuiltInViewportIndex;
SPV Capabilities: SampleRateShading, SparseResidency, MinLod, and ImageQuery.
2016-02-15 18:09:46 +00:00
case glslang::EbvSampleId:
builder.addCapability(spv::CapabilitySampleRateShading);
return spv::BuiltInSampleId;
case glslang::EbvSamplePosition:
builder.addCapability(spv::CapabilitySampleRateShading);
return spv::BuiltInSamplePosition;
case glslang::EbvSampleMask:
return spv::BuiltInSampleMask;
SPV: Fix issue #376; declare Geometry capability for fragment use of gl_Layer.
2016-07-08 20:05:15 +00:00
case glslang::EbvLayer:
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
if (glslangIntermediate->getStage() == EShLangMeshNV) {
return spv::BuiltInLayer;
}
GLSL: Make gl_Layer and gl_ViewportIndex always be outside blocks. There was some ambiguity/contradiction in this behavior, and Khronos decided glslang should always have these outside blocks, rather than have stage/vendor/target variations.
2017-09-13 19:22:50 +00:00
builder.addCapability(spv::CapabilityGeometry);
if (glslangIntermediate->getStage() == EShLangVertex ||
glslangIntermediate->getStage() == EShLangTessControl ||
glslangIntermediate->getStage() == EShLangTessEvaluation) {
SPV: Fix unexpected declarations of capability and extension
2017-03-16 15:02:39 +00:00
GLSL: Make gl_Layer and gl_ViewportIndex always be outside blocks. There was some ambiguity/contradiction in this behavior, and Khronos decided glslang should always have these outside blocks, rather than have stage/vendor/target variations.
2017-09-13 19:22:50 +00:00
builder.addExtension(spv::E_SPV_EXT_shader_viewport_index_layer);
builder.addCapability(spv::CapabilityShaderViewportIndexLayerEXT);
SPV: Fix unexpected declarations of capability and extension
2017-03-16 15:02:39 +00:00
}
SPV: Fix issue #376; declare Geometry capability for fragment use of gl_Layer.
2016-07-08 20:05:15 +00:00
return spv::BuiltInLayer;
GL_ARB_shader_draw_parameters: From @amdrexu, implement extension.
2015-10-14 20:10:30 +00:00
case glslang::EbvBaseVertex:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
addPre13Extension(spv::E_SPV_KHR_shader_draw_parameters);
SPV: Implement extension SPV_KHR_shader_draw_parameters.
2016-07-22 10:15:31 +00:00
builder.addCapability(spv::CapabilityDrawParameters);
return spv::BuiltInBaseVertex;
GL_ARB_shader_draw_parameters: From @amdrexu, implement extension.
2015-10-14 20:10:30 +00:00
case glslang::EbvBaseInstance:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
addPre13Extension(spv::E_SPV_KHR_shader_draw_parameters);
SPV: Implement extension SPV_KHR_shader_draw_parameters.
2016-07-22 10:15:31 +00:00
builder.addCapability(spv::CapabilityDrawParameters);
return spv::BuiltInBaseInstance;
SPV: PrimitiveId in frag shader will emit Geometry capability Using PrimitiveId in a fragment shader requires declaring an OpCapability with either Geometry or Tessellation.
2016-09-26 14:49:09 +00:00
GL_ARB_shader_draw_parameters: From @amdrexu, implement extension.
2015-10-14 20:10:30 +00:00
case glslang::EbvDrawId:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
addPre13Extension(spv::E_SPV_KHR_shader_draw_parameters);
SPV: Implement extension SPV_KHR_shader_draw_parameters.
2016-07-22 10:15:31 +00:00
builder.addCapability(spv::CapabilityDrawParameters);
return spv::BuiltInDrawIndex;
SPV: PrimitiveId in frag shader will emit Geometry capability Using PrimitiveId in a fragment shader requires declaring an OpCapability with either Geometry or Tessellation.
2016-09-26 14:49:09 +00:00
case glslang::EbvPrimitiveId:
if (glslangIntermediate->getStage() == EShLangFragment)
builder.addCapability(spv::CapabilityGeometry);
return spv::BuiltInPrimitiveId;
Implement extension GL_ARB_shader_stencil_export
2017-06-29 09:46:34 +00:00
case glslang::EbvFragStencilRef:
SPV: Implement extension SPV_EXT_shader_stencil_export
2017-08-23 15:24:42 +00:00
builder.addExtension(spv::E_SPV_EXT_shader_stencil_export);
builder.addCapability(spv::CapabilityStencilExportEXT);
return spv::BuiltInFragStencilRefEXT;
Implement extension GL_ARB_shader_stencil_export
2017-06-29 09:46:34 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EbvInvocationId: return spv::BuiltInInvocationId;
case glslang::EbvTessLevelInner: return spv::BuiltInTessLevelInner;
case glslang::EbvTessLevelOuter: return spv::BuiltInTessLevelOuter;
case glslang::EbvTessCoord: return spv::BuiltInTessCoord;
case glslang::EbvPatchVertices: return spv::BuiltInPatchVertices;
case glslang::EbvHelperInvocation: return spv::BuiltInHelperInvocation;
case glslang::EbvNumWorkGroups: return spv::BuiltInNumWorkgroups;
case glslang::EbvWorkGroupSize: return spv::BuiltInWorkgroupSize;
case glslang::EbvWorkGroupId: return spv::BuiltInWorkgroupId;
case glslang::EbvLocalInvocationId: return spv::BuiltInLocalInvocationId;
case glslang::EbvLocalInvocationIndex: return spv::BuiltInLocalInvocationIndex;
case glslang::EbvGlobalInvocationId: return spv::BuiltInGlobalInvocationId;
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
Implement the extension GL_ARB_shader_ballot Add new built-in variables and functions to the parser (SPIR-V tokens are missing).
2016-04-14 08:53:07 +00:00
case glslang::EbvSubGroupSize:
SPV: Still have to specify SPIR-V extension for gl_SubGroupSizeARB and gl_SubGroupInvocationARB.
2016-09-23 14:13:43 +00:00
builder.addExtension(spv::E_SPV_KHR_shader_ballot);
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
builder.addCapability(spv::CapabilitySubgroupBallotKHR);
return spv::BuiltInSubgroupSize;
Implement the extension GL_ARB_shader_ballot Add new built-in variables and functions to the parser (SPIR-V tokens are missing).
2016-04-14 08:53:07 +00:00
case glslang::EbvSubGroupInvocation:
SPV: Still have to specify SPIR-V extension for gl_SubGroupSizeARB and gl_SubGroupInvocationARB.
2016-09-23 14:13:43 +00:00
builder.addExtension(spv::E_SPV_KHR_shader_ballot);
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
builder.addCapability(spv::CapabilitySubgroupBallotKHR);
return spv::BuiltInSubgroupLocalInvocationId;
Implement the extension GL_ARB_shader_ballot Add new built-in variables and functions to the parser (SPIR-V tokens are missing).
2016-04-14 08:53:07 +00:00
case glslang::EbvSubGroupEqMask:
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
builder.addExtension(spv::E_SPV_KHR_shader_ballot);
builder.addCapability(spv::CapabilitySubgroupBallotKHR);
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
return spv::BuiltInSubgroupEqMask;
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
Implement the extension GL_ARB_shader_ballot Add new built-in variables and functions to the parser (SPIR-V tokens are missing).
2016-04-14 08:53:07 +00:00
case glslang::EbvSubGroupGeMask:
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
builder.addExtension(spv::E_SPV_KHR_shader_ballot);
builder.addCapability(spv::CapabilitySubgroupBallotKHR);
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
return spv::BuiltInSubgroupGeMask;
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
Implement the extension GL_ARB_shader_ballot Add new built-in variables and functions to the parser (SPIR-V tokens are missing).
2016-04-14 08:53:07 +00:00
case glslang::EbvSubGroupGtMask:
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
builder.addExtension(spv::E_SPV_KHR_shader_ballot);
builder.addCapability(spv::CapabilitySubgroupBallotKHR);
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
return spv::BuiltInSubgroupGtMask;
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
Implement the extension GL_ARB_shader_ballot Add new built-in variables and functions to the parser (SPIR-V tokens are missing).
2016-04-14 08:53:07 +00:00
case glslang::EbvSubGroupLeMask:
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
builder.addExtension(spv::E_SPV_KHR_shader_ballot);
builder.addCapability(spv::CapabilitySubgroupBallotKHR);
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
return spv::BuiltInSubgroupLeMask;
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
Implement the extension GL_ARB_shader_ballot Add new built-in variables and functions to the parser (SPIR-V tokens are missing).
2016-04-14 08:53:07 +00:00
case glslang::EbvSubGroupLtMask:
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
builder.addExtension(spv::E_SPV_KHR_shader_ballot);
builder.addCapability(spv::CapabilitySubgroupBallotKHR);
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
return spv::BuiltInSubgroupLtMask;
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EbvNumSubgroups:
builder.addCapability(spv::CapabilityGroupNonUniform);
return spv::BuiltInNumSubgroups;
case glslang::EbvSubgroupID:
builder.addCapability(spv::CapabilityGroupNonUniform);
return spv::BuiltInSubgroupId;
case glslang::EbvSubgroupSize2:
builder.addCapability(spv::CapabilityGroupNonUniform);
return spv::BuiltInSubgroupSize;
case glslang::EbvSubgroupInvocation2:
builder.addCapability(spv::CapabilityGroupNonUniform);
return spv::BuiltInSubgroupLocalInvocationId;
case glslang::EbvSubgroupEqMask2:
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformBallot);
return spv::BuiltInSubgroupEqMask;
case glslang::EbvSubgroupGeMask2:
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformBallot);
return spv::BuiltInSubgroupGeMask;
case glslang::EbvSubgroupGtMask2:
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformBallot);
return spv::BuiltInSubgroupGtMask;
case glslang::EbvSubgroupLeMask2:
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformBallot);
return spv::BuiltInSubgroupLeMask;
case glslang::EbvSubgroupLtMask2:
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformBallot);
return spv::BuiltInSubgroupLtMask;
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
SPV: Add missing extension declarations for certain AMD extensions.
2016-10-14 09:41:45 +00:00
case glslang::EbvBaryCoordNoPersp:
builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
return spv::BuiltInBaryCoordNoPerspAMD;
case glslang::EbvBaryCoordNoPerspCentroid:
builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
return spv::BuiltInBaryCoordNoPerspCentroidAMD;
case glslang::EbvBaryCoordNoPerspSample:
builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
return spv::BuiltInBaryCoordNoPerspSampleAMD;
case glslang::EbvBaryCoordSmooth:
builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
return spv::BuiltInBaryCoordSmoothAMD;
case glslang::EbvBaryCoordSmoothCentroid:
builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
return spv::BuiltInBaryCoordSmoothCentroidAMD;
case glslang::EbvBaryCoordSmoothSample:
builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
return spv::BuiltInBaryCoordSmoothSampleAMD;
case glslang::EbvBaryCoordPullModel:
builder.addExtension(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
return spv::BuiltInBaryCoordPullModelAMD;
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
SPV: Implement new extensions GL_KHX_device_group and GL_KHX_multiview. These correspond to SPV_KHR_device_group and SPV_KHR_multiview. Also, bring tests up to date with Khronos internals, and some misc. related changes.
2017-02-27 08:20:51 +00:00
case glslang::EbvDeviceIndex:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
addPre13Extension(spv::E_SPV_KHR_device_group);
SPV: Implement new extensions GL_KHX_device_group and GL_KHX_multiview. These correspond to SPV_KHR_device_group and SPV_KHR_multiview. Also, bring tests up to date with Khronos internals, and some misc. related changes.
2017-02-27 08:20:51 +00:00
builder.addCapability(spv::CapabilityDeviceGroup);
SPV: Update SPV header to official Rev. 10 of 1.0, and fix the consequences of doing so.
2017-02-27 08:50:28 +00:00
return spv::BuiltInDeviceIndex;
SPV: Implement new extensions GL_KHX_device_group and GL_KHX_multiview. These correspond to SPV_KHR_device_group and SPV_KHR_multiview. Also, bring tests up to date with Khronos internals, and some misc. related changes.
2017-02-27 08:20:51 +00:00
case glslang::EbvViewIndex:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
addPre13Extension(spv::E_SPV_KHR_multiview);
SPV: Implement new extensions GL_KHX_device_group and GL_KHX_multiview. These correspond to SPV_KHR_device_group and SPV_KHR_multiview. Also, bring tests up to date with Khronos internals, and some misc. related changes.
2017-02-27 08:20:51 +00:00
builder.addCapability(spv::CapabilityMultiView);
SPV: Update SPV header to official Rev. 10 of 1.0, and fix the consequences of doing so.
2017-02-27 08:50:28 +00:00
return spv::BuiltInViewIndex;
SPV: Implement new extensions GL_KHX_device_group and GL_KHX_multiview. These correspond to SPV_KHR_device_group and SPV_KHR_multiview. Also, bring tests up to date with Khronos internals, and some misc. related changes.
2017-02-27 08:20:51 +00:00
Implement GL_EXT_fragment_invocation_density including SPV generation using SPV_EXT_fragment_invocation_density. This is an alias of the functionality in SPV_NV_shading_rate, and thus in some cases we can only have one set of the tokens present (switch statements), so we have picked the EXT version. This required updating the expected test results for SPV_NV_shading_rate. Also updated the known-good for spirv-headers so that the validator in spirv-tools knows about the new extension.
2018-11-26 15:01:58 +00:00
case glslang::EbvFragSizeEXT:
builder.addExtension(spv::E_SPV_EXT_fragment_invocation_density);
builder.addCapability(spv::CapabilityFragmentDensityEXT);
return spv::BuiltInFragSizeEXT;
case glslang::EbvFragInvocationCountEXT:
builder.addExtension(spv::E_SPV_EXT_fragment_invocation_density);
builder.addCapability(spv::CapabilityFragmentDensityEXT);
return spv::BuiltInFragInvocationCountEXT;
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
case glslang::EbvViewportMaskNV:
SPV: Fix unexpected declarations of capability and extension
2017-03-16 15:02:39 +00:00
if (!memberDeclaration) {
builder.addExtension(spv::E_SPV_NV_viewport_array2);
builder.addCapability(spv::CapabilityShaderViewportMaskNV);
}
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
return spv::BuiltInViewportMaskNV;
case glslang::EbvSecondaryPositionNV:
SPV: Unexpected declarations of capabilities from NV extensions.
2017-02-22 08:44:48 +00:00
if (!memberDeclaration) {
builder.addExtension(spv::E_SPV_NV_stereo_view_rendering);
builder.addCapability(spv::CapabilityShaderStereoViewNV);
}
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
return spv::BuiltInSecondaryPositionNV;
case glslang::EbvSecondaryViewportMaskNV:
SPV: Unexpected declarations of capabilities from NV extensions.
2017-02-22 08:44:48 +00:00
if (!memberDeclaration) {
builder.addExtension(spv::E_SPV_NV_stereo_view_rendering);
builder.addCapability(spv::CapabilityShaderStereoViewNV);
}
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
return spv::BuiltInSecondaryViewportMaskNV;
Implement NVX_multiview_per_view_attributes
2017-02-14 22:52:34 +00:00
case glslang::EbvPositionPerViewNV:
SPV: Unexpected declarations of capabilities from NV extensions.
2017-02-22 08:44:48 +00:00
if (!memberDeclaration) {
builder.addExtension(spv::E_SPV_NVX_multiview_per_view_attributes);
builder.addCapability(spv::CapabilityPerViewAttributesNV);
}
Implement NVX_multiview_per_view_attributes
2017-02-14 22:52:34 +00:00
return spv::BuiltInPositionPerViewNV;
case glslang::EbvViewportMaskPerViewNV:
SPV: Unexpected declarations of capabilities from NV extensions.
2017-02-22 08:44:48 +00:00
if (!memberDeclaration) {
builder.addExtension(spv::E_SPV_NVX_multiview_per_view_attributes);
builder.addCapability(spv::CapabilityPerViewAttributesNV);
}
Implement NVX_multiview_per_view_attributes
2017-02-14 22:52:34 +00:00
return spv::BuiltInViewportMaskPerViewNV;
Add implementation of SPV_EXT_fragment_fully_covered This implementation uses the GLSL extension GL_NV_conservative_raster_underestimation to generate the new SPIR-V FullyCoveredEXT built in.
2017-12-13 20:07:22 +00:00
case glslang::EbvFragFullyCoveredNV:
builder.addExtension(spv::E_SPV_EXT_fragment_fully_covered);
builder.addCapability(spv::CapabilityFragmentFullyCoveredEXT);
return spv::BuiltInFullyCoveredEXT;
Add-support-for-SPV_NV_shading_rate
2018-09-19 18:43:21 +00:00
case glslang::EbvFragmentSizeNV:
builder.addExtension(spv::E_SPV_NV_shading_rate);
builder.addCapability(spv::CapabilityShadingRateNV);
return spv::BuiltInFragmentSizeNV;
case glslang::EbvInvocationsPerPixelNV:
builder.addExtension(spv::E_SPV_NV_shading_rate);
builder.addCapability(spv::CapabilityShadingRateNV);
return spv::BuiltInInvocationsPerPixelNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
Fix subgroup support for ray tracing Closes #1735 GlslangToSpv.cpp - minor formatting cleanup BaseTypes.h - minor formatting cleanup - add subgroup builtins to GetBuiltInVariableString (was resulting in "unknown built-in variable" messages in test output) Initialize.cpp - better naming and re-use of strings for subgroup builtin variable declarations - define subgroup builtin variables in ray-tracing shaders intermOut.cpp - add handling of the EOpSubgroupParition* variables (was resulting in "ERROR: Bad aggregation op" messages in test output) Update test results.
2019-05-27 20:46:31 +00:00
// ray tracing
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvLaunchIdNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInLaunchIdNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvLaunchSizeNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInLaunchSizeNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvWorldRayOriginNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInWorldRayOriginNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvWorldRayDirectionNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInWorldRayDirectionNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvObjectRayOriginNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInObjectRayOriginNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvObjectRayDirectionNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInObjectRayDirectionNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvRayTminNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInRayTminNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvRayTmaxNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInRayTmaxNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvInstanceCustomIndexNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInInstanceCustomIndexNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvHitTNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInHitTNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvHitKindNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInHitKindNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvObjectToWorldNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInObjectToWorldNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbvWorldToObjectNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
return spv::BuiltInWorldToObjectNV;
case glslang::EbvIncomingRayFlagsNV:
return spv::BuiltInIncomingRayFlagsNV;
Fix subgroup support for ray tracing Closes #1735 GlslangToSpv.cpp - minor formatting cleanup BaseTypes.h - minor formatting cleanup - add subgroup builtins to GetBuiltInVariableString (was resulting in "unknown built-in variable" messages in test output) Initialize.cpp - better naming and re-use of strings for subgroup builtin variable declarations - define subgroup builtin variables in ray-tracing shaders intermOut.cpp - add handling of the EOpSubgroupParition* variables (was resulting in "ERROR: Bad aggregation op" messages in test output) Update test results.
2019-05-27 20:46:31 +00:00
// barycentrics
Add-support-for-SPV_NV_fragment_shader_barycentric
2018-09-19 18:39:56 +00:00
case glslang::EbvBaryCoordNV:
builder.addExtension(spv::E_SPV_NV_fragment_shader_barycentric);
builder.addCapability(spv::CapabilityFragmentBarycentricNV);
return spv::BuiltInBaryCoordNV;
case glslang::EbvBaryCoordNoPerspNV:
builder.addExtension(spv::E_SPV_NV_fragment_shader_barycentric);
builder.addCapability(spv::CapabilityFragmentBarycentricNV);
return spv::BuiltInBaryCoordNoPerspNV;
Fix subgroup support for ray tracing Closes #1735 GlslangToSpv.cpp - minor formatting cleanup BaseTypes.h - minor formatting cleanup - add subgroup builtins to GetBuiltInVariableString (was resulting in "unknown built-in variable" messages in test output) Initialize.cpp - better naming and re-use of strings for subgroup builtin variable declarations - define subgroup builtin variables in ray-tracing shaders intermOut.cpp - add handling of the EOpSubgroupParition* variables (was resulting in "ERROR: Bad aggregation op" messages in test output) Update test results.
2019-05-27 20:46:31 +00:00
// mesh shaders
case glslang::EbvTaskCountNV:
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
return spv::BuiltInTaskCountNV;
Fix subgroup support for ray tracing Closes #1735 GlslangToSpv.cpp - minor formatting cleanup BaseTypes.h - minor formatting cleanup - add subgroup builtins to GetBuiltInVariableString (was resulting in "unknown built-in variable" messages in test output) Initialize.cpp - better naming and re-use of strings for subgroup builtin variable declarations - define subgroup builtin variables in ray-tracing shaders intermOut.cpp - add handling of the EOpSubgroupParition* variables (was resulting in "ERROR: Bad aggregation op" messages in test output) Update test results.
2019-05-27 20:46:31 +00:00
case glslang::EbvPrimitiveCountNV:
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
return spv::BuiltInPrimitiveCountNV;
Fix subgroup support for ray tracing Closes #1735 GlslangToSpv.cpp - minor formatting cleanup BaseTypes.h - minor formatting cleanup - add subgroup builtins to GetBuiltInVariableString (was resulting in "unknown built-in variable" messages in test output) Initialize.cpp - better naming and re-use of strings for subgroup builtin variable declarations - define subgroup builtin variables in ray-tracing shaders intermOut.cpp - add handling of the EOpSubgroupParition* variables (was resulting in "ERROR: Bad aggregation op" messages in test output) Update test results.
2019-05-27 20:46:31 +00:00
case glslang::EbvPrimitiveIndicesNV:
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
return spv::BuiltInPrimitiveIndicesNV;
Fix subgroup support for ray tracing Closes #1735 GlslangToSpv.cpp - minor formatting cleanup BaseTypes.h - minor formatting cleanup - add subgroup builtins to GetBuiltInVariableString (was resulting in "unknown built-in variable" messages in test output) Initialize.cpp - better naming and re-use of strings for subgroup builtin variable declarations - define subgroup builtin variables in ray-tracing shaders intermOut.cpp - add handling of the EOpSubgroupParition* variables (was resulting in "ERROR: Bad aggregation op" messages in test output) Update test results.
2019-05-27 20:46:31 +00:00
case glslang::EbvClipDistancePerViewNV:
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
return spv::BuiltInClipDistancePerViewNV;
Fix subgroup support for ray tracing Closes #1735 GlslangToSpv.cpp - minor formatting cleanup BaseTypes.h - minor formatting cleanup - add subgroup builtins to GetBuiltInVariableString (was resulting in "unknown built-in variable" messages in test output) Initialize.cpp - better naming and re-use of strings for subgroup builtin variable declarations - define subgroup builtin variables in ray-tracing shaders intermOut.cpp - add handling of the EOpSubgroupParition* variables (was resulting in "ERROR: Bad aggregation op" messages in test output) Update test results.
2019-05-27 20:46:31 +00:00
case glslang::EbvCullDistancePerViewNV:
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
return spv::BuiltInCullDistancePerViewNV;
Fix subgroup support for ray tracing Closes #1735 GlslangToSpv.cpp - minor formatting cleanup BaseTypes.h - minor formatting cleanup - add subgroup builtins to GetBuiltInVariableString (was resulting in "unknown built-in variable" messages in test output) Initialize.cpp - better naming and re-use of strings for subgroup builtin variable declarations - define subgroup builtin variables in ray-tracing shaders intermOut.cpp - add handling of the EOpSubgroupParition* variables (was resulting in "ERROR: Bad aggregation op" messages in test output) Update test results.
2019-05-27 20:46:31 +00:00
case glslang::EbvLayerPerViewNV:
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
return spv::BuiltInLayerPerViewNV;
Fix subgroup support for ray tracing Closes #1735 GlslangToSpv.cpp - minor formatting cleanup BaseTypes.h - minor formatting cleanup - add subgroup builtins to GetBuiltInVariableString (was resulting in "unknown built-in variable" messages in test output) Initialize.cpp - better naming and re-use of strings for subgroup builtin variable declarations - define subgroup builtin variables in ray-tracing shaders intermOut.cpp - add handling of the EOpSubgroupParition* variables (was resulting in "ERROR: Bad aggregation op" messages in test output) Update test results.
2019-05-27 20:46:31 +00:00
case glslang::EbvMeshViewCountNV:
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
return spv::BuiltInMeshViewCountNV;
Fix subgroup support for ray tracing Closes #1735 GlslangToSpv.cpp - minor formatting cleanup BaseTypes.h - minor formatting cleanup - add subgroup builtins to GetBuiltInVariableString (was resulting in "unknown built-in variable" messages in test output) Initialize.cpp - better naming and re-use of strings for subgroup builtin variable declarations - define subgroup builtin variables in ray-tracing shaders intermOut.cpp - add handling of the EOpSubgroupParition* variables (was resulting in "ERROR: Bad aggregation op" messages in test output) Update test results.
2019-05-27 20:46:31 +00:00
case glslang::EbvMeshViewIndicesNV:
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
return spv::BuiltInMeshViewIndicesNV;
Add support for GL_NV_shader_sm_builtins Including spirv and AST tests Also increase size of TBuiltInVariable bitfields since we've now exceeded 127 and add a static_assert to make this easier to find next time it happens!
2019-06-04 12:43:32 +00:00
// sm builtins
case glslang::EbvWarpsPerSM:
builder.addExtension(spv::E_SPV_NV_shader_sm_builtins);
builder.addCapability(spv::CapabilityShaderSMBuiltinsNV);
return spv::BuiltInWarpsPerSMNV;
case glslang::EbvSMCount:
builder.addExtension(spv::E_SPV_NV_shader_sm_builtins);
builder.addCapability(spv::CapabilityShaderSMBuiltinsNV);
return spv::BuiltInSMCountNV;
case glslang::EbvWarpID:
builder.addExtension(spv::E_SPV_NV_shader_sm_builtins);
builder.addCapability(spv::CapabilityShaderSMBuiltinsNV);
return spv::BuiltInWarpIDNV;
case glslang::EbvSMID:
builder.addExtension(spv::E_SPV_NV_shader_sm_builtins);
builder.addCapability(spv::CapabilityShaderSMBuiltinsNV);
return spv::BuiltInSMIDNV;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#endif
SPV: Unexpected declarations of capabilities from NV extensions.
2017-02-22 08:44:48 +00:00
default:
return spv::BuiltInMax;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
}
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
// Translate glslang image layout format to SPIR-V image format.
SPV Capabilities: StorageImageExtendedFormats, StorageImageReadWithoutFormat, and StorageImageWriteWithoutFormat.
2016-02-15 18:57:00 +00:00
spv::ImageFormat TGlslangToSpvTraverser::TranslateImageFormat(const glslang::TType& type)
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
{
assert(type.getBasicType() == glslang::EbtSampler);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifdef GLSLANG_WEB
return spv::ImageFormatUnknown;
#endif
SPV Capabilities: StorageImageExtendedFormats, StorageImageReadWithoutFormat, and StorageImageWriteWithoutFormat.
2016-02-15 18:57:00 +00:00
// Check for capabilities
Web: Remove/rationalize a set of *_EXTENSIONS, using GLSLANG_WEB. Focus was on the front end (not SPIR-V), minus the grammar. Reduces #ifdef count by around 320 and makes the web build 270K smaller, which is about 90% the target size. The grammar and scanner will be another step, as will the SPIR-V backend. This makes heavy use of methods #ifdef'd to return false as a global way of turning off code, relying on C++ DCE to do the rest.
2019-08-01 09:28:08 +00:00
switch (type.getQualifier().getFormat()) {
SPV Capabilities: StorageImageExtendedFormats, StorageImageReadWithoutFormat, and StorageImageWriteWithoutFormat.
2016-02-15 18:57:00 +00:00
case glslang::ElfRg32f:
case glslang::ElfRg16f:
case glslang::ElfR11fG11fB10f:
case glslang::ElfR16f:
case glslang::ElfRgba16:
case glslang::ElfRgb10A2:
case glslang::ElfRg16:
case glslang::ElfRg8:
case glslang::ElfR16:
case glslang::ElfR8:
case glslang::ElfRgba16Snorm:
case glslang::ElfRg16Snorm:
case glslang::ElfRg8Snorm:
case glslang::ElfR16Snorm:
case glslang::ElfR8Snorm:
case glslang::ElfRg32i:
case glslang::ElfRg16i:
case glslang::ElfRg8i:
case glslang::ElfR16i:
case glslang::ElfR8i:
case glslang::ElfRgb10a2ui:
case glslang::ElfRg32ui:
case glslang::ElfRg16ui:
case glslang::ElfRg8ui:
case glslang::ElfR16ui:
case glslang::ElfR8ui:
builder.addCapability(spv::CapabilityStorageImageExtendedFormats);
break;
default:
break;
}
// do the translation
Web: Remove/rationalize a set of *_EXTENSIONS, using GLSLANG_WEB. Focus was on the front end (not SPIR-V), minus the grammar. Reduces #ifdef count by around 320 and makes the web build 270K smaller, which is about 90% the target size. The grammar and scanner will be another step, as will the SPIR-V backend. This makes heavy use of methods #ifdef'd to return false as a global way of turning off code, relying on C++ DCE to do the rest.
2019-08-01 09:28:08 +00:00
switch (type.getQualifier().getFormat()) {
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
case glslang::ElfNone: return spv::ImageFormatUnknown;
case glslang::ElfRgba32f: return spv::ImageFormatRgba32f;
case glslang::ElfRgba16f: return spv::ImageFormatRgba16f;
case glslang::ElfR32f: return spv::ImageFormatR32f;
case glslang::ElfRgba8: return spv::ImageFormatRgba8;
case glslang::ElfRgba8Snorm: return spv::ImageFormatRgba8Snorm;
case glslang::ElfRg32f: return spv::ImageFormatRg32f;
case glslang::ElfRg16f: return spv::ImageFormatRg16f;
case glslang::ElfR11fG11fB10f: return spv::ImageFormatR11fG11fB10f;
case glslang::ElfR16f: return spv::ImageFormatR16f;
case glslang::ElfRgba16: return spv::ImageFormatRgba16;
case glslang::ElfRgb10A2: return spv::ImageFormatRgb10A2;
case glslang::ElfRg16: return spv::ImageFormatRg16;
case glslang::ElfRg8: return spv::ImageFormatRg8;
case glslang::ElfR16: return spv::ImageFormatR16;
case glslang::ElfR8: return spv::ImageFormatR8;
case glslang::ElfRgba16Snorm: return spv::ImageFormatRgba16Snorm;
case glslang::ElfRg16Snorm: return spv::ImageFormatRg16Snorm;
case glslang::ElfRg8Snorm: return spv::ImageFormatRg8Snorm;
case glslang::ElfR16Snorm: return spv::ImageFormatR16Snorm;
case glslang::ElfR8Snorm: return spv::ImageFormatR8Snorm;
case glslang::ElfRgba32i: return spv::ImageFormatRgba32i;
case glslang::ElfRgba16i: return spv::ImageFormatRgba16i;
case glslang::ElfRgba8i: return spv::ImageFormatRgba8i;
case glslang::ElfR32i: return spv::ImageFormatR32i;
case glslang::ElfRg32i: return spv::ImageFormatRg32i;
case glslang::ElfRg16i: return spv::ImageFormatRg16i;
case glslang::ElfRg8i: return spv::ImageFormatRg8i;
case glslang::ElfR16i: return spv::ImageFormatR16i;
case glslang::ElfR8i: return spv::ImageFormatR8i;
case glslang::ElfRgba32ui: return spv::ImageFormatRgba32ui;
case glslang::ElfRgba16ui: return spv::ImageFormatRgba16ui;
case glslang::ElfRgba8ui: return spv::ImageFormatRgba8ui;
case glslang::ElfR32ui: return spv::ImageFormatR32ui;
case glslang::ElfRg32ui: return spv::ImageFormatRg32ui;
case glslang::ElfRg16ui: return spv::ImageFormatRg16ui;
case glslang::ElfRgb10a2ui: return spv::ImageFormatRgb10a2ui;
case glslang::ElfRg8ui: return spv::ImageFormatRg8ui;
case glslang::ElfR16ui: return spv::ImageFormatR16ui;
case glslang::ElfR8ui: return spv::ImageFormatR8ui;
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
default: return spv::ImageFormatMax;
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
}
}
HLSL: Refactor attribute implementation. - make it sharable with GLSL - correct the case insensitivity - remove the map; queries are not needed, all entries need processing - make it easier to build bottom up (will help GLSL parsing) - support semantic checking and reporting - allow front-end dependent semantics and attribute name mapping
2018-01-30 18:01:39 +00:00
spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSelectionControl(const glslang::TIntermSelection& selectionNode) const
HLSL: Translate directive [flatten] and [branch] to SPV control mask.
2017-07-04 15:23:40 +00:00
{
HLSL: Refactor attribute implementation. - make it sharable with GLSL - correct the case insensitivity - remove the map; queries are not needed, all entries need processing - make it easier to build bottom up (will help GLSL parsing) - support semantic checking and reporting - allow front-end dependent semantics and attribute name mapping
2018-01-30 18:01:39 +00:00
if (selectionNode.getFlatten())
return spv::SelectionControlFlattenMask;
if (selectionNode.getDontFlatten())
return spv::SelectionControlDontFlattenMask;
return spv::SelectionControlMaskNone;
HLSL: Translate directive [flatten] and [branch] to SPV control mask.
2017-07-04 15:23:40 +00:00
}
HLSL: Refactor attribute implementation. - make it sharable with GLSL - correct the case insensitivity - remove the map; queries are not needed, all entries need processing - make it easier to build bottom up (will help GLSL parsing) - support semantic checking and reporting - allow front-end dependent semantics and attribute name mapping
2018-01-30 18:01:39 +00:00
spv::SelectionControlMask TGlslangToSpvTraverser::TranslateSwitchControl(const glslang::TIntermSwitch& switchNode) const
HLSL: implement [unroll] and [loop] attributes This adds infrastructure suitable for any front end to create SPIR-V loop control flags. The only current front end doing so is HLSL. [unroll] turns into spv::LoopControlUnrollMask [loop] turns into spv::LoopControlDontUnrollMask no specification means spv::LoopControlMaskNone
2017-05-03 02:14:50 +00:00
{
HLSL: Refactor attribute implementation. - make it sharable with GLSL - correct the case insensitivity - remove the map; queries are not needed, all entries need processing - make it easier to build bottom up (will help GLSL parsing) - support semantic checking and reporting - allow front-end dependent semantics and attribute name mapping
2018-01-30 18:01:39 +00:00
if (switchNode.getFlatten())
return spv::SelectionControlFlattenMask;
if (switchNode.getDontFlatten())
return spv::SelectionControlDontFlattenMask;
return spv::SelectionControlMaskNone;
}
GLSL: Implement GL_EXT_control_flow_attributes. See https://github.com/KhronosGroup/GLSL/pull/11.
2018-01-31 15:11:18 +00:00
// return a non-0 dependency if the dependency argument must be set
spv::LoopControlMask TGlslangToSpvTraverser::TranslateLoopControl(const glslang::TIntermLoop& loopNode,
SPV 1.4: Implement the 5 new loop controls.
2019-01-12 10:31:41 +00:00
std::vector<unsigned int>& operands) const
HLSL: Refactor attribute implementation. - make it sharable with GLSL - correct the case insensitivity - remove the map; queries are not needed, all entries need processing - make it easier to build bottom up (will help GLSL parsing) - support semantic checking and reporting - allow front-end dependent semantics and attribute name mapping
2018-01-30 18:01:39 +00:00
{
spv::LoopControlMask control = spv::LoopControlMaskNone;
if (loopNode.getDontUnroll())
control = control | spv::LoopControlDontUnrollMask;
if (loopNode.getUnroll())
control = control | spv::LoopControlUnrollMask;
HLSL: Add conversions for image ops during SPV construction HLSL allows image and texture types to be templatized on sub-vec4 types, or even structures. This was mostly handled already during creation of sampling operations. However, for operator[] which can generate image loads, this wasn't happening. It also isn't very easy to do at that point in time, because operator[] does not know where the results it produces will end up. They may be an lvalue or an rvalue, and there's a post-process to convert loads to stores. They may end up in atomic ops. To bypass that difficulty, GlslangToSpv now looks for this case and adds the appropriate conversion. LIMITATION: this only works for cases for which a simple conversion opcode suffices. That is to say, it will not work if the type is templatized on a struct.
2018-02-18 18:40:01 +00:00
if (unsigned(loopNode.getLoopDependency()) == glslang::TIntermLoop::dependencyInfinite)
GLSL: Implement GL_EXT_control_flow_attributes. See https://github.com/KhronosGroup/GLSL/pull/11.
2018-01-31 15:11:18 +00:00
control = control | spv::LoopControlDependencyInfiniteMask;
else if (loopNode.getLoopDependency() > 0) {
control = control | spv::LoopControlDependencyLengthMask;
SPV 1.4: Implement the 5 new loop controls.
2019-01-12 10:31:41 +00:00
operands.push_back((unsigned int)loopNode.getLoopDependency());
}
if (glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4) {
if (loopNode.getMinIterations() > 0) {
control = control | spv::LoopControlMinIterationsMask;
operands.push_back(loopNode.getMinIterations());
}
if (loopNode.getMaxIterations() < glslang::TIntermLoop::iterationsInfinite) {
control = control | spv::LoopControlMaxIterationsMask;
operands.push_back(loopNode.getMaxIterations());
}
if (loopNode.getIterationMultiple() > 1) {
control = control | spv::LoopControlIterationMultipleMask;
operands.push_back(loopNode.getIterationMultiple());
}
if (loopNode.getPeelCount() > 0) {
control = control | spv::LoopControlPeelCountMask;
operands.push_back(loopNode.getPeelCount());
}
if (loopNode.getPartialCount() > 0) {
control = control | spv::LoopControlPartialCountMask;
operands.push_back(loopNode.getPartialCount());
}
GLSL: Implement GL_EXT_control_flow_attributes. See https://github.com/KhronosGroup/GLSL/pull/11.
2018-01-31 15:11:18 +00:00
}
HLSL: Refactor attribute implementation. - make it sharable with GLSL - correct the case insensitivity - remove the map; queries are not needed, all entries need processing - make it easier to build bottom up (will help GLSL parsing) - support semantic checking and reporting - allow front-end dependent semantics and attribute name mapping
2018-01-30 18:01:39 +00:00
return control;
HLSL: implement [unroll] and [loop] attributes This adds infrastructure suitable for any front end to create SPIR-V loop control flags. The only current front end doing so is HLSL. [unroll] turns into spv::LoopControlUnrollMask [loop] turns into spv::LoopControlDontUnrollMask no specification means spv::LoopControlMaskNone
2017-05-03 02:14:50 +00:00
}
SPV: Emit extension for using StorageClassStorageBuffer.
2017-05-05 11:09:58 +00:00
// Translate glslang type to SPIR-V storage class.
spv::StorageClass TGlslangToSpvTraverser::TranslateStorageClass(const glslang::TType& type)
{
if (type.getQualifier().isPipeInput())
return spv::StorageClassInput;
HLSL: Pass opaques by local copy, instead of by interface original. Also, remove assumption that if something is opaque that it must be in the UniformConstant storage class. This allows function declarations to know all parameters will be in the Function storage class.
2017-09-08 08:38:07 +00:00
if (type.getQualifier().isPipeOutput())
SPV: Emit extension for using StorageClassStorageBuffer.
2017-05-05 11:09:58 +00:00
return spv::StorageClassOutput;
HLSL: Pass opaques by local copy, instead of by interface original. Also, remove assumption that if something is opaque that it must be in the UniformConstant storage class. This allows function declarations to know all parameters will be in the Function storage class.
2017-09-08 08:38:07 +00:00
if (glslangIntermediate->getSource() != glslang::EShSourceHlsl ||
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
type.getQualifier().storage == glslang::EvqUniform) {
#ifndef GLSLANG_WEB
HLSL: Pass opaques by local copy, instead of by interface original. Also, remove assumption that if something is opaque that it must be in the UniformConstant storage class. This allows function declarations to know all parameters will be in the Function storage class.
2017-09-08 08:38:07 +00:00
if (type.getBasicType() == glslang::EbtAtomicUint)
return spv::StorageClassAtomicCounter;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#endif
HLSL: Pass opaques by local copy, instead of by interface original. Also, remove assumption that if something is opaque that it must be in the UniformConstant storage class. This allows function declarations to know all parameters will be in the Function storage class.
2017-09-08 08:38:07 +00:00
if (type.containsOpaque())
return spv::StorageClassUniformConstant;
}
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Fix bug where shaderRecordNV buffers were using StorageBuffer storage class
2018-12-15 02:59:53 +00:00
if (type.getQualifier().isUniformOrBuffer() &&
type.getQualifier().layoutShaderRecordNV) {
return spv::StorageClassShaderRecordBufferNV;
}
#endif
HLSL: Pass opaques by local copy, instead of by interface original. Also, remove assumption that if something is opaque that it must be in the UniformConstant storage class. This allows function declarations to know all parameters will be in the Function storage class.
2017-09-08 08:38:07 +00:00
if (glslangIntermediate->usingStorageBuffer() && type.getQualifier().storage == glslang::EvqBuffer) {
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
addPre13Extension(spv::E_SPV_KHR_storage_buffer_storage_class);
SPV: Emit extension for using StorageClassStorageBuffer.
2017-05-05 11:09:58 +00:00
return spv::StorageClassStorageBuffer;
HLSL: Pass opaques by local copy, instead of by interface original. Also, remove assumption that if something is opaque that it must be in the UniformConstant storage class. This allows function declarations to know all parameters will be in the Function storage class.
2017-09-08 08:38:07 +00:00
}
if (type.getQualifier().isUniformOrBuffer()) {
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Web: Remove/rationalize a set of *_EXTENSIONS, using GLSLANG_WEB. Focus was on the front end (not SPIR-V), minus the grammar. Reduces #ifdef count by around 320 and makes the web build 270K smaller, which is about 90% the target size. The grammar and scanner will be another step, as will the SPIR-V backend. This makes heavy use of methods #ifdef'd to return false as a global way of turning off code, relying on C++ DCE to do the rest.
2019-08-01 09:28:08 +00:00
if (type.getQualifier().isPushConstant())
SPV: Emit extension for using StorageClassStorageBuffer.
2017-05-05 11:09:58 +00:00
return spv::StorageClassPushConstant;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#endif
SPV: Emit extension for using StorageClassStorageBuffer.
2017-05-05 11:09:58 +00:00
if (type.getBasicType() == glslang::EbtBlock)
return spv::StorageClassUniform;
HLSL: Pass opaques by local copy, instead of by interface original. Also, remove assumption that if something is opaque that it must be in the UniformConstant storage class. This allows function declarations to know all parameters will be in the Function storage class.
2017-09-08 08:38:07 +00:00
return spv::StorageClassUniformConstant;
}
switch (type.getQualifier().storage) {
case glslang::EvqGlobal: return spv::StorageClassPrivate;
case glslang::EvqConstReadOnly: return spv::StorageClassFunction;
case glslang::EvqTemporary: return spv::StorageClassFunction;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
case glslang::EvqShared: return spv::StorageClassWorkgroup;
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
case glslang::EvqPayloadNV: return spv::StorageClassRayPayloadNV;
case glslang::EvqPayloadInNV: return spv::StorageClassIncomingRayPayloadNV;
case glslang::EvqHitAttrNV: return spv::StorageClassHitAttributeNV;
case glslang::EvqCallableDataNV: return spv::StorageClassCallableDataNV;
case glslang::EvqCallableDataInNV: return spv::StorageClassIncomingCallableDataNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
#endif
HLSL: Pass opaques by local copy, instead of by interface original. Also, remove assumption that if something is opaque that it must be in the UniformConstant storage class. This allows function declarations to know all parameters will be in the Function storage class.
2017-09-08 08:38:07 +00:00
default:
assert(0);
break;
SPV: Emit extension for using StorageClassStorageBuffer.
2017-05-05 11:09:58 +00:00
}
HLSL: Pass opaques by local copy, instead of by interface original. Also, remove assumption that if something is opaque that it must be in the UniformConstant storage class. This allows function declarations to know all parameters will be in the Function storage class.
2017-09-08 08:38:07 +00:00
return spv::StorageClassFunction;
SPV: Emit extension for using StorageClassStorageBuffer.
2017-05-05 11:09:58 +00:00
}
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
// Add capabilities pertaining to how an array is indexed.
void TGlslangToSpvTraverser::addIndirectionIndexCapabilities(const glslang::TType& baseType,
const glslang::TType& indexType)
{
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
if (indexType.getQualifier().isNonUniform()) {
// deal with an asserted non-uniform index
Add missing OpExtension SPV_EXT_descriptor_indexing for dynamic indexing capabilities
2018-07-12 21:51:18 +00:00
// SPV_EXT_descriptor_indexing already added in TranslateNonUniformDecoration
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
if (baseType.getBasicType() == glslang::EbtSampler) {
if (baseType.getQualifier().hasAttachment())
builder.addCapability(spv::CapabilityInputAttachmentArrayNonUniformIndexingEXT);
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
else if (baseType.isImage() && baseType.getSampler().isBuffer())
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
builder.addCapability(spv::CapabilityStorageTexelBufferArrayNonUniformIndexingEXT);
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
else if (baseType.isTexture() && baseType.getSampler().isBuffer())
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
builder.addCapability(spv::CapabilityUniformTexelBufferArrayNonUniformIndexingEXT);
else if (baseType.isImage())
builder.addCapability(spv::CapabilityStorageImageArrayNonUniformIndexingEXT);
else if (baseType.isTexture())
builder.addCapability(spv::CapabilitySampledImageArrayNonUniformIndexingEXT);
} else if (baseType.getBasicType() == glslang::EbtBlock) {
if (baseType.getQualifier().storage == glslang::EvqBuffer)
builder.addCapability(spv::CapabilityStorageBufferArrayNonUniformIndexingEXT);
else if (baseType.getQualifier().storage == glslang::EvqUniform)
builder.addCapability(spv::CapabilityUniformBufferArrayNonUniformIndexingEXT);
}
} else {
// assume a dynamically uniform index
if (baseType.getBasicType() == glslang::EbtSampler) {
Add missing OpExtension SPV_EXT_descriptor_indexing for dynamic indexing capabilities
2018-07-12 21:51:18 +00:00
if (baseType.getQualifier().hasAttachment()) {
builder.addExtension("SPV_EXT_descriptor_indexing");
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
builder.addCapability(spv::CapabilityInputAttachmentArrayDynamicIndexingEXT);
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
} else if (baseType.isImage() && baseType.getSampler().isBuffer()) {
Add missing OpExtension SPV_EXT_descriptor_indexing for dynamic indexing capabilities
2018-07-12 21:51:18 +00:00
builder.addExtension("SPV_EXT_descriptor_indexing");
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
builder.addCapability(spv::CapabilityStorageTexelBufferArrayDynamicIndexingEXT);
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
} else if (baseType.isTexture() && baseType.getSampler().isBuffer()) {
Add missing OpExtension SPV_EXT_descriptor_indexing for dynamic indexing capabilities
2018-07-12 21:51:18 +00:00
builder.addExtension("SPV_EXT_descriptor_indexing");
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
builder.addCapability(spv::CapabilityUniformTexelBufferArrayDynamicIndexingEXT);
Add missing OpExtension SPV_EXT_descriptor_indexing for dynamic indexing capabilities
2018-07-12 21:51:18 +00:00
}
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
}
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
}
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
// Return whether or not the given type is something that should be tied to a
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
// descriptor set.
bool IsDescriptorResource(const glslang::TType& type)
{
SPV: Issue #180: push_constants don't have descriptor sets.
2016-03-09 04:36:22 +00:00
// uniform and buffer blocks are included, unless it is a push_constant
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
if (type.getBasicType() == glslang::EbtBlock)
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
return type.getQualifier().isUniformOrBuffer() &&
Web: Remove/rationalize a set of *_EXTENSIONS, using GLSLANG_WEB. Focus was on the front end (not SPIR-V), minus the grammar. Reduces #ifdef count by around 320 and makes the web build 270K smaller, which is about 90% the target size. The grammar and scanner will be another step, as will the SPIR-V backend. This makes heavy use of methods #ifdef'd to return false as a global way of turning off code, relying on C++ DCE to do the rest.
2019-08-01 09:28:08 +00:00
! type.getQualifier().isShaderRecordNV() &&
! type.getQualifier().isPushConstant();
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
// non block...
// basically samplerXXX/subpass/sampler/texture are all included
// if they are the global-scope-class, not the function parameter
// (or local, if they ever exist) class.
if (type.getBasicType() == glslang::EbtSampler)
return type.getQualifier().isUniformOrBuffer();
// None of the above.
return false;
}
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
void InheritQualifiers(glslang::TQualifier& child, const glslang::TQualifier& parent)
{
if (child.layoutMatrix == glslang::ElmNone)
child.layoutMatrix = parent.layoutMatrix;
if (parent.invariant)
child.invariant = true;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
if (parent.flat)
child.flat = true;
if (parent.centroid)
child.centroid = true;
Web: Remove/rationalize a set of *_EXTENSIONS, using GLSLANG_WEB. Focus was on the front end (not SPIR-V), minus the grammar. Reduces #ifdef count by around 320 and makes the web build 270K smaller, which is about 90% the target size. The grammar and scanner will be another step, as will the SPIR-V backend. This makes heavy use of methods #ifdef'd to return false as a global way of turning off code, relying on C++ DCE to do the rest.
2019-08-01 09:28:08 +00:00
#ifndef GLSLANG_WEB
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
if (parent.nopersp)
child.nopersp = true;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
if (parent.explicitInterp)
child.explicitInterp = true;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
if (parent.perPrimitiveNV)
child.perPrimitiveNV = true;
if (parent.perViewNV)
child.perViewNV = true;
if (parent.perTaskNV)
child.perTaskNV = true;
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
if (parent.patch)
child.patch = true;
if (parent.sample)
child.sample = true;
SPV: Add support for memory qualifiers.
2016-02-21 12:59:01 +00:00
if (parent.coherent)
child.coherent = true;
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
if (parent.devicecoherent)
child.devicecoherent = true;
if (parent.queuefamilycoherent)
child.queuefamilycoherent = true;
if (parent.workgroupcoherent)
child.workgroupcoherent = true;
if (parent.subgroupcoherent)
child.subgroupcoherent = true;
if (parent.nonprivate)
child.nonprivate = true;
SPV: Add support for memory qualifiers.
2016-02-21 12:59:01 +00:00
if (parent.volatil)
child.volatil = true;
if (parent.restrict)
child.restrict = true;
if (parent.readonly)
child.readonly = true;
if (parent.writeonly)
child.writeonly = true;
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
#endif
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
}
SPV: Tighten up number of struct-types declared based on decoration. Takes some pressure off of issue #304. Structures don't inherit locations and then explicitly decorate members with them, so removed this reason to have another instance of a structure type.
2016-09-01 23:05:23 +00:00
bool HasNonLayoutQualifiers(const glslang::TType& type, const glslang::TQualifier& qualifier)
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
{
SPV: Add recursive distrubition of 'location' across structure members.
2016-01-22 01:56:57 +00:00
// This should list qualifiers that simultaneous satisfy:
SPV: Tighten up number of struct-types declared based on decoration. Takes some pressure off of issue #304. Structures don't inherit locations and then explicitly decorate members with them, so removed this reason to have another instance of a structure type.
2016-09-01 23:05:23 +00:00
// - struct members might inherit from a struct declaration
// (note that non-block structs don't explicitly inherit,
// only implicitly, meaning no decoration involved)
// - affect decorations on the struct members
// (note smooth does not, and expecting something like volatile
// to effect the whole object)
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
// - are not part of the offset/st430/etc or row/column-major layout
SPV: Tighten up number of struct-types declared based on decoration. Takes some pressure off of issue #304. Structures don't inherit locations and then explicitly decorate members with them, so removed this reason to have another instance of a structure type.
2016-09-01 23:05:23 +00:00
return qualifier.invariant || (qualifier.hasLocation() && type.getBasicType() == glslang::EbtBlock);
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
//
// Implement the TGlslangToSpvTraverser class.
//
SPV Version: Emit the requested SPIR-V version, not the header version. Fixes #1236.
2018-02-01 01:35:56 +00:00
TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const glslang::TIntermediate* glslangIntermediate,
SPV: Add OpSource shader source code and file name.
2017-05-31 23:11:16 +00:00
spv::SpvBuildLogger* buildLogger, glslang::SpvOptions& options)
: TIntermTraverser(true, false, true),
options(options),
shaderEntry(nullptr), currentFunction(nullptr),
HLSL: Do structure conversion for return type struct-punning on non-entry-point functions.
2016-10-06 18:59:51 +00:00
sequenceDepth(0), logger(buildLogger),
SPV Version: Emit the requested SPIR-V version, not the header version. Fixes #1236.
2018-02-01 01:35:56 +00:00
builder(spvVersion, (glslang::GetKhronosToolId() << 16) | glslang::GetSpirvGeneratorVersion(), logger),
Non-functional: Rename some entry-point variables to entryPoint, not main.
2016-11-26 20:31:47 +00:00
inEntryPoint(false), entryPointTerminated(false), linkageOnly(false),
SPV: Add a switch for favoring non-NaN operands in min, max, and clamp.
2019-06-18 05:33:09 +00:00
glslangIntermediate(glslangIntermediate),
nanMinMaxClamp(glslangIntermediate->getNanMinMaxClamp())
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
spv::ExecutionModel executionModel = TranslateExecutionModel(glslangIntermediate->getStage());
builder.clearAccessChain();
SPV reflection: Add OpModuleProcessed for compile options.
2017-07-21 02:00:36 +00:00
builder.setSource(TranslateSourceLanguage(glslangIntermediate->getSource(), glslangIntermediate->getProfile()),
glslangIntermediate->getVersion());
SPV: Add OpSource shader source code and file name.
2017-05-31 23:11:16 +00:00
if (options.generateDebugInfo) {
SPV: Debug output: Include OpLine information for execution path. Note that declaratives are not handled, only procedurals.
2017-06-01 00:50:53 +00:00
builder.setEmitOpLines();
SPV reflection: Add OpModuleProcessed for compile options.
2017-07-21 02:00:36 +00:00
builder.setSourceFile(glslangIntermediate->getSourceFile());
// Set the source shader's text. If for SPV version 1.0, include
// a preamble in comments stating the OpModuleProcessed instructions.
// Otherwise, emit those as actual instructions.
std::string text;
const std::vector<std::string>& processes = glslangIntermediate->getProcesses();
for (int p = 0; p < (int)processes.size(); ++p) {
SPV/Standalone: Support specifying arbitrary versions of SPIR-V.
2018-10-26 16:12:32 +00:00
if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_1) {
SPV reflection: Add OpModuleProcessed for compile options.
2017-07-21 02:00:36 +00:00
text.append("// OpModuleProcessed ");
text.append(processes[p]);
text.append("\n");
} else
builder.addModuleProcessed(processes[p]);
}
SPV/Standalone: Support specifying arbitrary versions of SPIR-V.
2018-10-26 16:12:32 +00:00
if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_1 && (int)processes.size() > 0)
SPV reflection: Add OpModuleProcessed for compile options.
2017-07-21 02:00:36 +00:00
text.append("#line 1\n");
text.append(glslangIntermediate->getSourceText());
builder.setSourceText(text);
Create separate OpSource for each included file
2018-12-06 18:13:15 +00:00
// Pass name and text for all included files
const std::map<std::string, std::string>& include_txt = glslangIntermediate->getIncludeText();
for (auto iItr = include_txt.begin(); iItr != include_txt.end(); ++iItr)
builder.addInclude(iItr->first, iItr->second);
SPV: Add OpSource shader source code and file name.
2017-05-31 23:11:16 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
stdBuiltins = builder.import("GLSL.std.450");
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
spv::AddressingModel addressingModel = spv::AddressingModelLogical;
spv::MemoryModel memoryModel = spv::MemoryModelGLSL450;
if (glslangIntermediate->usingPhysicalStorageBuffer()) {
addressingModel = spv::AddressingModelPhysicalStorageBuffer64EXT;
builder.addExtension(spv::E_SPV_EXT_physical_storage_buffer);
builder.addCapability(spv::CapabilityPhysicalStorageBufferAddressesEXT);
};
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
if (glslangIntermediate->usingVulkanMemoryModel()) {
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
memoryModel = spv::MemoryModelVulkanKHR;
builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
builder.addExtension(spv::E_SPV_KHR_vulkan_memory_model);
}
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
builder.setMemoryModel(addressingModel, memoryModel);
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
if (glslangIntermediate->usingVariablePointers()) {
builder.addCapability(spv::CapabilityVariablePointers);
}
Track separate entry-point names and mangled names... ... and use each in the correct way at consumption sites. This completes issue #513.
2016-09-20 00:09:30 +00:00
shaderEntry = builder.makeEntryPoint(glslangIntermediate->getEntryPointName().c_str());
entryPoint = builder.addEntryPoint(executionModel, shaderEntry, glslangIntermediate->getEntryPointName().c_str());
Final round for line endings.
2015-06-26 22:58:36 +00:00
// Add the source extensions
Performance: Simple upgrade/cleanup of stl c++11 containers (10% perf. increase).
2015-07-19 04:34:27 +00:00
const auto& sourceExtensions = glslangIntermediate->getRequestedExtensions();
for (auto it = sourceExtensions.begin(); it != sourceExtensions.end(); ++it)
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.addSourceExtension(it->c_str());
// Add the top-level modes for this shader.
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
if (glslangIntermediate->getXfbMode()) {
builder.addCapability(spv::CapabilityTransformFeedback);
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.addExecutionMode(shaderEntry, spv::ExecutionModeXfb);
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
unsigned int mode;
switch (glslangIntermediate->getStage()) {
case EShLangVertex:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
builder.addCapability(spv::CapabilityShader);
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case EShLangFragment:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
builder.addCapability(spv::CapabilityShader);
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (glslangIntermediate->getPixelCenterInteger())
builder.addExecutionMode(shaderEntry, spv::ExecutionModePixelCenterInteger);
SPV: Add modes for spacing, depth, depth replacing, winding, and point mode.
2015-10-13 22:29:02 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (glslangIntermediate->getOriginUpperLeft())
builder.addExecutionMode(shaderEntry, spv::ExecutionModeOriginUpperLeft);
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
else
builder.addExecutionMode(shaderEntry, spv::ExecutionModeOriginLowerLeft);
SPV: Add modes for spacing, depth, depth replacing, winding, and point mode.
2015-10-13 22:29:02 +00:00
if (glslangIntermediate->getEarlyFragmentTests())
builder.addExecutionMode(shaderEntry, spv::ExecutionModeEarlyFragmentTests);
Implement SPV_KHR_post_depth_coverage Added support for both extension GL_ARB_post_depth_coverage and GL_EXT_post_depth_coverage.
2017-07-01 00:14:30 +00:00
if (glslangIntermediate->getPostDepthCoverage()) {
builder.addCapability(spv::CapabilitySampleMaskPostDepthCoverage);
builder.addExecutionMode(shaderEntry, spv::ExecutionModePostDepthCoverage);
builder.addExtension(spv::E_SPV_KHR_post_depth_coverage);
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
if (glslangIntermediate->getDepth() != glslang::EldUnchanged && glslangIntermediate->isDepthReplacing())
builder.addExecutionMode(shaderEntry, spv::ExecutionModeDepthReplacing);
#ifndef GLSLANG_WEB
SPV: Add modes for spacing, depth, depth replacing, winding, and point mode.
2015-10-13 22:29:02 +00:00
switch(glslangIntermediate->getDepth()) {
case glslang::EldGreater: mode = spv::ExecutionModeDepthGreater; break;
case glslang::EldLess: mode = spv::ExecutionModeDepthLess; break;
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
default: mode = spv::ExecutionModeMax; break;
SPV: Add modes for spacing, depth, depth replacing, winding, and point mode.
2015-10-13 22:29:02 +00:00
}
Build: Switch to Rev. 6 of headers, removing "BadValue" enum stuff. Note: Technically, this is glslang's modified private copy of official headers. Official headers will appear in other places, like https://github.com/KhronosGroup/SPIRV-Headers.
2016-07-15 17:53:56 +00:00
if (mode != spv::ExecutionModeMax)
SPV: Add modes for spacing, depth, depth replacing, winding, and point mode.
2015-10-13 22:29:02 +00:00
builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
Support GL_ARB_fragment_shader_interlock
2019-06-03 16:33:50 +00:00
switch (glslangIntermediate->getInterlockOrdering()) {
case glslang::EioPixelInterlockOrdered: mode = spv::ExecutionModePixelInterlockOrderedEXT; break;
case glslang::EioPixelInterlockUnordered: mode = spv::ExecutionModePixelInterlockUnorderedEXT; break;
case glslang::EioSampleInterlockOrdered: mode = spv::ExecutionModeSampleInterlockOrderedEXT; break;
case glslang::EioSampleInterlockUnordered: mode = spv::ExecutionModeSampleInterlockUnorderedEXT; break;
case glslang::EioShadingRateInterlockOrdered: mode = spv::ExecutionModeShadingRateInterlockOrderedEXT; break;
case glslang::EioShadingRateInterlockUnordered: mode = spv::ExecutionModeShadingRateInterlockUnorderedEXT; break;
default: mode = spv::ExecutionModeMax; break;
}
if (mode != spv::ExecutionModeMax) {
builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
if (mode == spv::ExecutionModeShadingRateInterlockOrderedEXT ||
mode == spv::ExecutionModeShadingRateInterlockUnorderedEXT) {
builder.addCapability(spv::CapabilityFragmentShaderShadingRateInterlockEXT);
} else if (mode == spv::ExecutionModePixelInterlockOrderedEXT ||
mode == spv::ExecutionModePixelInterlockUnorderedEXT) {
builder.addCapability(spv::CapabilityFragmentShaderPixelInterlockEXT);
} else {
builder.addCapability(spv::CapabilityFragmentShaderSampleInterlockEXT);
}
builder.addExtension(spv::E_SPV_EXT_fragment_shader_interlock);
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Final round for line endings.
2015-06-26 22:58:36 +00:00
case EShLangCompute:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
builder.addCapability(spv::CapabilityShader);
SPV: Generate the LocalSize Execution Mode for compute shaders.
2015-09-16 22:04:05 +00:00
builder.addExecutionMode(shaderEntry, spv::ExecutionModeLocalSize, glslangIntermediate->getLocalSize(0),
glslangIntermediate->getLocalSize(1),
glslangIntermediate->getLocalSize(2));
Add-support-for-SPV_NV_compute_shader_derivatives
2018-09-19 18:40:45 +00:00
if (glslangIntermediate->getLayoutDerivativeModeNone() == glslang::LayoutDerivativeGroupQuads) {
builder.addCapability(spv::CapabilityComputeDerivativeGroupQuadsNV);
builder.addExecutionMode(shaderEntry, spv::ExecutionModeDerivativeGroupQuadsNV);
builder.addExtension(spv::E_SPV_NV_compute_shader_derivatives);
} else if (glslangIntermediate->getLayoutDerivativeModeNone() == glslang::LayoutDerivativeGroupLinear) {
builder.addCapability(spv::CapabilityComputeDerivativeGroupLinearNV);
builder.addExecutionMode(shaderEntry, spv::ExecutionModeDerivativeGroupLinearNV);
builder.addExtension(spv::E_SPV_NV_compute_shader_derivatives);
}
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
break;
case EShLangTessEvaluation:
case EShLangTessControl:
builder.addCapability(spv::CapabilityTessellation);
glslang::TLayoutGeometry primitive;
if (glslangIntermediate->getStage() == EShLangTessControl) {
builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices());
primitive = glslangIntermediate->getOutputPrimitive();
} else {
primitive = glslangIntermediate->getInputPrimitive();
}
switch (primitive) {
case glslang::ElgTriangles: mode = spv::ExecutionModeTriangles; break;
case glslang::ElgQuads: mode = spv::ExecutionModeQuads; break;
case glslang::ElgIsolines: mode = spv::ExecutionModeIsolines; break;
default: mode = spv::ExecutionModeMax; break;
}
if (mode != spv::ExecutionModeMax)
builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
switch (glslangIntermediate->getVertexSpacing()) {
case glslang::EvsEqual: mode = spv::ExecutionModeSpacingEqual; break;
case glslang::EvsFractionalEven: mode = spv::ExecutionModeSpacingFractionalEven; break;
case glslang::EvsFractionalOdd: mode = spv::ExecutionModeSpacingFractionalOdd; break;
default: mode = spv::ExecutionModeMax; break;
}
if (mode != spv::ExecutionModeMax)
builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
switch (glslangIntermediate->getVertexOrder()) {
case glslang::EvoCw: mode = spv::ExecutionModeVertexOrderCw; break;
case glslang::EvoCcw: mode = spv::ExecutionModeVertexOrderCcw; break;
default: mode = spv::ExecutionModeMax; break;
}
if (mode != spv::ExecutionModeMax)
builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
if (glslangIntermediate->getPointMode())
builder.addExecutionMode(shaderEntry, spv::ExecutionModePointMode);
break;
case EShLangGeometry:
builder.addCapability(spv::CapabilityGeometry);
switch (glslangIntermediate->getInputPrimitive()) {
case glslang::ElgPoints: mode = spv::ExecutionModeInputPoints; break;
case glslang::ElgLines: mode = spv::ExecutionModeInputLines; break;
case glslang::ElgLinesAdjacency: mode = spv::ExecutionModeInputLinesAdjacency; break;
case glslang::ElgTriangles: mode = spv::ExecutionModeTriangles; break;
case glslang::ElgTrianglesAdjacency: mode = spv::ExecutionModeInputTrianglesAdjacency; break;
default: mode = spv::ExecutionModeMax; break;
}
if (mode != spv::ExecutionModeMax)
builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
builder.addExecutionMode(shaderEntry, spv::ExecutionModeInvocations, glslangIntermediate->getInvocations());
switch (glslangIntermediate->getOutputPrimitive()) {
case glslang::ElgPoints: mode = spv::ExecutionModeOutputPoints; break;
case glslang::ElgLineStrip: mode = spv::ExecutionModeOutputLineStrip; break;
case glslang::ElgTriangleStrip: mode = spv::ExecutionModeOutputTriangleStrip; break;
default: mode = spv::ExecutionModeMax; break;
}
if (mode != spv::ExecutionModeMax)
builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices());
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case EShLangRayGenNV:
case EShLangIntersectNV:
case EShLangAnyHitNV:
case EShLangClosestHitNV:
case EShLangMissNV:
case EShLangCallableNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
builder.addCapability(spv::CapabilityRayTracingNV);
builder.addExtension("SPV_NV_ray_tracing");
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
break;
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
case EShLangTaskNV:
case EShLangMeshNV:
builder.addCapability(spv::CapabilityMeshShadingNV);
builder.addExtension(spv::E_SPV_NV_mesh_shader);
builder.addExecutionMode(shaderEntry, spv::ExecutionModeLocalSize, glslangIntermediate->getLocalSize(0),
glslangIntermediate->getLocalSize(1),
glslangIntermediate->getLocalSize(2));
if (glslangIntermediate->getStage() == EShLangMeshNV) {
builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputVertices, glslangIntermediate->getVertices());
builder.addExecutionMode(shaderEntry, spv::ExecutionModeOutputPrimitivesNV, glslangIntermediate->getPrimitives());
switch (glslangIntermediate->getOutputPrimitive()) {
case glslang::ElgPoints: mode = spv::ExecutionModeOutputPoints; break;
case glslang::ElgLines: mode = spv::ExecutionModeOutputLinesNV; break;
case glslang::ElgTriangles: mode = spv::ExecutionModeOutputTrianglesNV; break;
default: mode = spv::ExecutionModeMax; break;
}
if (mode != spv::ExecutionModeMax)
builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
}
break;
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
break;
}
}
Always correctly terminate main. Issue #588, PR #600.
2016-11-26 20:23:20 +00:00
// Finish creating SPV, after the traversal is complete.
void TGlslangToSpvTraverser::finishSpv()
SPV: Make OpEntryPoint I/O list include all static uses. Before, it was only including explicit interface, sufficient for IO-Block-declared oriented interface, but not sufficient for all modes GLSL might be used with SPIR-V. Two things are accomplished now: 1) each id will appear exactly once 2) the OpEntryPoint list will union static use with declarations
2015-12-21 00:37:07 +00:00
{
Non-Functional: Add postprocess as more robust way to add capabilities When capabilities are needed for specific SPIR-V instructions, it is fragile to do so based on GLSL/AST usage; it should be based on actual instructions they got translated to.
2018-08-03 21:56:12 +00:00
// Finish the entry point function
Non-functional: Rename some entry-point variables to entryPoint, not main.
2016-11-26 20:31:47 +00:00
if (! entryPointTerminated) {
Always correctly terminate main. Issue #588, PR #600.
2016-11-26 20:23:20 +00:00
builder.setBuildPoint(shaderEntry->getLastBlock());
builder.leaveFunction();
}
SPV: Make OpEntryPoint I/O list include all static uses. Before, it was only including explicit interface, sufficient for IO-Block-declared oriented interface, but not sufficient for all modes GLSL might be used with SPIR-V. Two things are accomplished now: 1) each id will appear exactly once 2) the OpEntryPoint list will union static use with declarations
2015-12-21 00:37:07 +00:00
// finish off the entry-point SPV instruction by adding the Input/Output <id>
Fix compilation issues with MSVC 2010 (mostly by eliminating use of range-based for loops and std::tie)
2016-02-23 21:17:38 +00:00
for (auto it = iOSet.cbegin(); it != iOSet.cend(); ++it)
entryPoint->addIdOperand(*it);
SPV: Make OpEntryPoint I/O list include all static uses. Before, it was only including explicit interface, sufficient for IO-Block-declared oriented interface, but not sufficient for all modes GLSL might be used with SPIR-V. Two things are accomplished now: 1) each id will appear exactly once 2) the OpEntryPoint list will union static use with declarations
2015-12-21 00:37:07 +00:00
Web: Selectively remove a few key features, using #ifndef GLSLANG_WEB Save about 100K. N.B.: This is done by eliminating a function call, at a high level, not by #ifdef'ing a bunch of code. Also, removed no longer needed *_EXTENSION #ifdef in the code not needed by GLSLANG_WEB.
2019-07-28 08:12:10 +00:00
#ifndef GLSLANG_WEB
Non-Functional: Add postprocess as more robust way to add capabilities When capabilities are needed for specific SPIR-V instructions, it is fragile to do so based on GLSL/AST usage; it should be based on actual instructions they got translated to.
2018-08-03 21:56:12 +00:00
// Add capabilities, extensions, remove unneeded decorations, etc.,
// based on the resulting SPIR-V.
builder.postProcess();
Web: Selectively remove a few key features, using #ifndef GLSLANG_WEB Save about 100K. N.B.: This is done by eliminating a function call, at a high level, not by #ifdef'ing a bunch of code. Also, removed no longer needed *_EXTENSION #ifdef in the code not needed by GLSLANG_WEB.
2019-07-28 08:12:10 +00:00
#endif
SPV: Make OpEntryPoint I/O list include all static uses. Before, it was only including explicit interface, sufficient for IO-Block-declared oriented interface, but not sufficient for all modes GLSL might be used with SPIR-V. Two things are accomplished now: 1) each id will appear exactly once 2) the OpEntryPoint list will union static use with declarations
2015-12-21 00:37:07 +00:00
}
Always correctly terminate main. Issue #588, PR #600.
2016-11-26 20:23:20 +00:00
// Write the SPV into 'out'.
void TGlslangToSpvTraverser::dumpSpv(std::vector<unsigned int>& out)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
Always correctly terminate main. Issue #588, PR #600.
2016-11-26 20:23:20 +00:00
builder.dump(out);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
//
// Implement the traversal functions.
//
// Return true from interior nodes to have the external traversal
// continue on to children. Return false if children were
// already processed.
//
//
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
// Symbols can turn into
Final round for line endings.
2015-06-26 22:58:36 +00:00
// - uniform/input reads
// - output writes
// - complex lvalue base setups: foo.bar[3].... , where we see foo and start up an access chain
// - something simple that degenerates into the last bullet
//
void TGlslangToSpvTraverser::visitSymbol(glslang::TIntermSymbol* symbol)
{
add SpecConstantOpModeGuard to GlslangToSpvTraverser::visitSymbol()
2016-04-06 18:42:01 +00:00
SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder);
if (symbol->getType().getQualifier().isSpecConstant())
spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
Final round for line endings.
2015-06-26 22:58:36 +00:00
// getSymbolId() will set up all the IO decorations on the first call.
// Formal function parameters were mapped during makeFunctions().
spv::Id id = getSymbolId(symbol);
SPV: Make OpEntryPoint I/O list include all static uses. Before, it was only including explicit interface, sufficient for IO-Block-declared oriented interface, but not sufficient for all modes GLSL might be used with SPIR-V. Two things are accomplished now: 1) each id will appear exactly once 2) the OpEntryPoint list will union static use with declarations
2015-12-21 00:37:07 +00:00
if (builder.isPointer(id)) {
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
// Include all "static use" and "linkage only" interface variables on the OpEntryPoint instruction
SPV 1.4: Generate all globals on OpEntryPoint interface list.
2019-01-04 09:47:06 +00:00
// Consider adding to the OpEntryPoint interface list.
// Only looking at structures if they have at least one member.
if (!symbol->getType().isStruct() || symbol->getType().getStruct()->size() > 0) {
spv::StorageClass sc = builder.getStorageClass(id);
// Before SPIR-V 1.4, we only want to include Input and Output.
// Starting with SPIR-V 1.4, we want all globals.
if ((glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4 && sc != spv::StorageClassFunction) ||
(sc == spv::StorageClassInput || sc == spv::StorageClassOutput)) {
HLSL: Fix #1064: Don't include empty structures on OpEntryPoint
2017-09-19 17:09:59 +00:00
iOSet.insert(id);
SPV 1.4: Generate all globals on OpEntryPoint interface list.
2019-01-04 09:47:06 +00:00
}
HLSL: Fix #1064: Don't include empty structures on OpEntryPoint
2017-09-19 17:09:59 +00:00
}
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
// If the SPIR-V type is required to be different than the AST type,
// translate now from the SPIR-V type to the AST type, for the consuming
// operation.
// Note this turns it from an l-value to an r-value.
// Currently, all symbols needing this are inputs; avoid the map lookup when non-input.
if (symbol->getType().getQualifier().storage == glslang::EvqVaryingIn)
id = translateForcedType(id);
SPV: Make OpEntryPoint I/O list include all static uses. Before, it was only including explicit interface, sufficient for IO-Block-declared oriented interface, but not sufficient for all modes GLSL might be used with SPIR-V. Two things are accomplished now: 1) each id will appear exactly once 2) the OpEntryPoint list will union static use with declarations
2015-12-21 00:37:07 +00:00
}
// Only process non-linkage-only nodes for generating actual static uses
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
if (! linkageOnly || symbol->getQualifier().isSpecConstant()) {
Final round for line endings.
2015-06-26 22:58:36 +00:00
// Prepare to generate code for the access
// L-value chains will be computed left to right. We're on the symbol now,
// which is the left-most part of the access chain, so now is "clear" time,
// followed by setting the base.
builder.clearAccessChain();
// For now, we consider all user variables as being in memory, so they are pointers,
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
// except for
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
// A) R-Value arguments to a function, which are an intermediate object.
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
// See comments in handleUserFunctionCall().
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
// B) Specialization constants (normal constants don't even come in as a variable),
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
// These are also pure R-values.
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
// C) R-Values from type translation, see above call to translateForcedType()
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
glslang::TQualifier qualifier = symbol->getQualifier();
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
if (qualifier.isSpecConstant() || rValueParameters.find(symbol->getId()) != rValueParameters.end() ||
!builder.isPointerType(builder.getTypeId(id)))
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.setAccessChainRValue(id);
else
builder.setAccessChainLValue(id);
}
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifdef ENABLE_HLSL
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
// Process linkage-only nodes for any special additional interface work.
if (linkageOnly) {
if (glslangIntermediate->getHlslFunctionality1()) {
// Map implicit counter buffers to their originating buffers, which should have been
// seen by now, given earlier pruning of unused counters, and preservation of order
// of declaration.
if (symbol->getType().getQualifier().isUniformOrBuffer()) {
if (!glslangIntermediate->hasCounterBufferName(symbol->getName())) {
// Save possible originating buffers for counter buffers, keyed by
// making the potential counter-buffer name.
std::string keyName = symbol->getName().c_str();
keyName = glslangIntermediate->addCounterBufferName(keyName);
counterOriginator[keyName] = symbol;
} else {
// Handle a counter buffer, by finding the saved originating buffer.
std::string keyName = symbol->getName().c_str();
auto it = counterOriginator.find(keyName);
if (it != counterOriginator.end()) {
id = getSymbolId(it->second);
if (id != spv::NoResult) {
spv::Id counterId = getSymbolId(symbol);
Fix #1331: Emit SPV_GOOGLE_hlsl_functionality1 for counters. It was missing when there were no semantics.
2018-04-06 01:35:38 +00:00
if (counterId != spv::NoResult) {
builder.addExtension("SPV_GOOGLE_hlsl_functionality1");
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
builder.addDecorationId(id, spv::DecorationHlslCounterBufferGOOGLE, counterId);
Fix #1331: Emit SPV_GOOGLE_hlsl_functionality1 for counters. It was missing when there were no semantics.
2018-04-06 01:35:38 +00:00
}
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
}
}
}
}
}
}
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::TIntermBinary* node)
{
SPV: Fix #1575, fix #1593: Support HLSL #line SPIR-V OpLines now contain filenames from HLSL-style #lines.
2018-12-08 00:36:33 +00:00
builder.setLine(node->getLoc().line, node->getLoc().getFilename());
SPV: Debug output: Include OpLine information for execution path. Note that declaratives are not handled, only procedurals.
2017-06-01 00:50:53 +00:00
Turn on SpecConstantOpMode based on node qualifier Move SpecConstantOpModeGuard from makeSpvConstantFromConstSubTree() to visitbinary() and visitunary(). Checking if the visiting node is a spec constants, if so, turn on the SpecConstantOpMode, otherwise, stay in the normal mode.
2016-04-04 02:20:42 +00:00
SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder);
if (node->getType().getQualifier().isSpecConstant())
spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
Final round for line endings.
2015-06-26 22:58:36 +00:00
// First, handle special cases
switch (node->getOp()) {
case glslang::EOpAssign:
case glslang::EOpAddAssign:
case glslang::EOpSubAssign:
case glslang::EOpMulAssign:
case glslang::EOpVectorTimesMatrixAssign:
case glslang::EOpVectorTimesScalarAssign:
case glslang::EOpMatrixTimesScalarAssign:
case glslang::EOpMatrixTimesMatrixAssign:
case glslang::EOpDivAssign:
case glslang::EOpModAssign:
case glslang::EOpAndAssign:
case glslang::EOpInclusiveOrAssign:
case glslang::EOpExclusiveOrAssign:
case glslang::EOpLeftShiftAssign:
case glslang::EOpRightShiftAssign:
// A bin-op assign "a += b" means the same thing as "a = a + b"
// where a is evaluated before b. For a simple assignment, GLSL
// says to evaluate the left before the right. So, always, left
// node then right node.
{
// get the left l-value, save it away
builder.clearAccessChain();
node->getLeft()->traverse(this);
spv::Builder::AccessChain lValue = builder.getAccessChain();
// evaluate the right
builder.clearAccessChain();
node->getRight()->traverse(this);
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
spv::Id rValue = accessChainLoad(node->getRight()->getType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (node->getOp() != glslang::EOpAssign) {
// the left is also an r-value
builder.setAccessChain(lValue);
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
spv::Id leftRValue = accessChainLoad(node->getLeft()->getType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
// do the operation
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()),
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
TranslateNoContractionDecoration(node->getType().getQualifier()),
TranslateNonUniformDecoration(node->getType().getQualifier()) };
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
rValue = createBinaryOperation(node->getOp(), decorations,
Final round for line endings.
2015-06-26 22:58:36 +00:00
convertGlslangToSpvType(node->getType()), leftRValue, rValue,
node->getType().getBasicType());
// these all need their counterparts in createBinaryOperation()
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(rValue != spv::NoResult);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
// store the result
builder.setAccessChain(lValue);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
multiTypeStore(node->getLeft()->getType(), rValue);
Final round for line endings.
2015-06-26 22:58:36 +00:00
// assignments are expressions having an rValue after they are evaluated...
builder.clearAccessChain();
builder.setAccessChainRValue(rValue);
}
return false;
case glslang::EOpIndexDirect:
case glslang::EOpIndexDirectStruct:
{
SPV 1.4: Lookup tables: Use variable initializer and NonWritable... ...when doing a variable lookup on an array of constants.
2019-02-07 15:04:12 +00:00
// Structure, array, matrix, or vector indirection with statically known index.
Final round for line endings.
2015-06-26 22:58:36 +00:00
// Get the left part of the access chain.
node->getLeft()->traverse(this);
// Add the next element in the chain
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
const int glslangIndex = node->getRight()->getAsConstantUnion()->getConstArray()[0].getIConst();
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (! node->getLeft()->getType().isArray() &&
node->getLeft()->getType().isVector() &&
node->getOp() == glslang::EOpIndexDirect) {
// This is essentially a hard-coded vector swizzle of size 1,
// so short circuit the access-chain stuff with a swizzle.
std::vector<unsigned> swizzle;
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
swizzle.push_back(glslangIndex);
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
int dummySize;
builder.accessChainPushSwizzle(swizzle, convertGlslangToSpvType(node->getLeft()->getType()),
TranslateCoherent(node->getLeft()->getType()),
glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize));
Final round for line endings.
2015-06-26 22:58:36 +00:00
} else {
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
// Load through a block reference is performed with a dot operator that
// is mapped to EOpIndexDirectStruct. When we get to the actual reference,
// do a load and reset the access chain.
Web: Remove/rationalize a set of *_EXTENSIONS, using GLSLANG_WEB. Focus was on the front end (not SPIR-V), minus the grammar. Reduces #ifdef count by around 320 and makes the web build 270K smaller, which is about 90% the target size. The grammar and scanner will be another step, as will the SPIR-V backend. This makes heavy use of methods #ifdef'd to return false as a global way of turning off code, relying on C++ DCE to do the rest.
2019-08-01 09:28:08 +00:00
if (node->getLeft()->isReference() &&
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
!node->getLeft()->getType().isArray() &&
node->getOp() == glslang::EOpIndexDirectStruct)
{
spv::Id left = accessChainLoad(node->getLeft()->getType());
builder.clearAccessChain();
builder.setAccessChainLValue(left);
}
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
int spvIndex = glslangIndex;
if (node->getLeft()->getBasicType() == glslang::EbtBlock &&
node->getOp() == glslang::EOpIndexDirectStruct)
{
// This may be, e.g., an anonymous block-member selection, which generally need
// index remapping due to hidden members in anonymous blocks.
std::vector<int>& remapper = memberRemapper[node->getLeft()->getType().getStruct()];
assert(remapper.size() > 0);
spvIndex = remapper[glslangIndex];
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
// normal case for indexing array or structure or block
Move getBufferReferenceAlignment to be a method of TType This is a better place for it logically, since it is not specific to glsl->spirv translation. And in a future change I want to use it outside of glslangtospv.
2019-03-06 19:34:10 +00:00
builder.accessChainPush(builder.makeIntConstant(spvIndex), TranslateCoherent(node->getLeft()->getType()), node->getLeft()->getType().getBufferReferenceAlignment());
SPV: Don't add clip/cull distance capabilities unless used. These capabalities were added on declaration of the members, but that is considered too aggressive, as those members are automatically declared in some shaders that don't use them. Now, actual access is needed to make the capabalities be declared.
2016-05-17 01:22:05 +00:00
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
// Add capabilities here for accessing PointSize and clip/cull distance.
// We have deferred generation of associated capabilities until now.
SPV: Don't add clip/cull distance capabilities unless used. These capabalities were added on declaration of the members, but that is considered too aggressive, as those members are automatically declared in some shaders that don't use them. Now, actual access is needed to make the capabalities be declared.
2016-05-17 01:22:05 +00:00
if (node->getLeft()->getType().isStruct() && ! node->getLeft()->getType().isArray())
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
declareUseOfStructMember(*(node->getLeft()->getType().getStruct()), glslangIndex);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
}
return false;
case glslang::EOpIndexIndirect:
{
SPV 1.4: Lookup tables: Use variable initializer and NonWritable... ...when doing a variable lookup on an array of constants.
2019-02-07 15:04:12 +00:00
// Array, matrix, or vector indirection with variable index.
// Will use native SPIR-V access-chain for and array indirection;
Final round for line endings.
2015-06-26 22:58:36 +00:00
// matrices are arrays of vectors, so will also work for a matrix.
// Will use the access chain's 'component' for variable index into a vector.
// This adapter is building access chains left to right.
// Set up the access chain to the left.
node->getLeft()->traverse(this);
// save it so that computing the right side doesn't trash it
spv::Builder::AccessChain partial = builder.getAccessChain();
// compute the next index in the chain
builder.clearAccessChain();
node->getRight()->traverse(this);
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
spv::Id index = accessChainLoad(node->getRight()->getType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
addIndirectionIndexCapabilities(node->getLeft()->getType(), node->getRight()->getType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
// restore the saved access chain
builder.setAccessChain(partial);
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
if (! node->getLeft()->getType().isArray() && node->getLeft()->getType().isVector()) {
int dummySize;
builder.accessChainPushComponent(index, convertGlslangToSpvType(node->getLeft()->getType()),
TranslateCoherent(node->getLeft()->getType()),
glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize));
} else
Move getBufferReferenceAlignment to be a method of TType This is a better place for it logically, since it is not specific to glsl->spirv translation. And in a future change I want to use it outside of glslangtospv.
2019-03-06 19:34:10 +00:00
builder.accessChainPush(index, TranslateCoherent(node->getLeft()->getType()), node->getLeft()->getType().getBufferReferenceAlignment());
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
return false;
case glslang::EOpVectorSwizzle:
{
node->getLeft()->traverse(this);
std::vector<unsigned> swizzle;
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
convertSwizzle(*node->getRight()->getAsAggregate(), swizzle);
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
int dummySize;
builder.accessChainPushSwizzle(swizzle, convertGlslangToSpvType(node->getLeft()->getType()),
TranslateCoherent(node->getLeft()->getType()),
glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(), dummySize));
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
return false;
HLSL: Add EOpMatrixSwizzle, selectively decomposed to other ops, for issue #670. Since EOpMatrixSwizzle is a new op, existing back-ends only work when the front end first decomposes it to other operations. So far, this is only being done for simple assignment into matrix swizzles.
2017-01-13 19:27:52 +00:00
case glslang::EOpMatrixSwizzle:
logger->missingFunctionality("matrix swizzle");
return true;
SPV: Implement short circuiting of && and || when emitting SPIR-V.
2015-10-15 19:29:11 +00:00
case glslang::EOpLogicalOr:
case glslang::EOpLogicalAnd:
{
// These may require short circuiting, but can sometimes be done as straight
// binary operations. The right operand must be short circuited if it has
// side effects, and should probably be if it is complex.
if (isTrivial(node->getRight()->getAsTyped()))
break; // handle below as a normal binary operation
// otherwise, we need to do dynamic short circuiting on the right operand
spv::Id result = createShortCircuit(node->getOp(), *node->getLeft()->getAsTyped(), *node->getRight()->getAsTyped());
builder.clearAccessChain();
builder.setAccessChainRValue(result);
}
return false;
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
break;
}
// Assume generic binary op...
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
// get right operand
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.clearAccessChain();
node->getLeft()->traverse(this);
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
spv::Id left = accessChainLoad(node->getLeft()->getType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
// get left operand
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.clearAccessChain();
node->getRight()->traverse(this);
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
spv::Id right = accessChainLoad(node->getRight()->getType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
// get result
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()),
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
TranslateNoContractionDecoration(node->getType().getQualifier()),
TranslateNonUniformDecoration(node->getType().getQualifier()) };
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id result = createBinaryOperation(node->getOp(), decorations,
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
convertGlslangToSpvType(node->getType()), left, right,
node->getLeft()->getType().getBasicType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
SPV: Don't make MissingFunctionality() exit anymore.
2015-12-22 04:21:11 +00:00
builder.clearAccessChain();
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (! result) {
Create a new logger class for all messages w.r.t. SPIR-V build.
2016-05-04 19:55:59 +00:00
logger->missingFunctionality("unknown glslang binary operation");
SPV: Don't make MissingFunctionality() exit anymore.
2015-12-22 04:21:11 +00:00
return true; // pick up a child as the place-holder result
Final round for line endings.
2015-06-26 22:58:36 +00:00
} else {
builder.setAccessChainRValue(result);
return false;
}
}
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
// Figure out what, if any, type changes are needed when accessing a specific built-in.
// Returns <the type SPIR-V requires for declarion, the type to translate to on use>.
// Also see comment for 'forceType', regarding tracking SPIR-V-required types.
std::pair<spv::Id, spv::Id> TGlslangToSpvTraverser::getForcedType(spv::BuiltIn builtIn,
const glslang::TType& glslangType)
{
switch(builtIn)
{
case spv::BuiltInSubgroupEqMask:
case spv::BuiltInSubgroupGeMask:
case spv::BuiltInSubgroupGtMask:
case spv::BuiltInSubgroupLeMask:
case spv::BuiltInSubgroupLtMask: {
// these require changing a 64-bit scaler -> a vector of 32-bit components
if (glslangType.isVector())
break;
std::pair<spv::Id, spv::Id> ret(builder.makeVectorType(builder.makeUintType(32), 4),
builder.makeUintType(64));
return ret;
}
default:
break;
}
std::pair<spv::Id, spv::Id> ret(spv::NoType, spv::NoType);
return ret;
}
// For an object previously identified (see getForcedType() and forceType)
// as needing type translations, do the translation needed for a load, turning
// an L-value into in R-value.
spv::Id TGlslangToSpvTraverser::translateForcedType(spv::Id object)
{
const auto forceIt = forceType.find(object);
if (forceIt == forceType.end())
return object;
spv::Id desiredTypeId = forceIt->second;
spv::Id objectTypeId = builder.getTypeId(object);
assert(builder.isPointerType(objectTypeId));
objectTypeId = builder.getContainedTypeId(objectTypeId);
if (builder.isVectorType(objectTypeId) &&
builder.getScalarTypeWidth(builder.getContainedTypeId(objectTypeId)) == 32) {
if (builder.getScalarTypeWidth(desiredTypeId) == 64) {
// handle 32-bit v.xy* -> 64-bit
builder.clearAccessChain();
builder.setAccessChainLValue(object);
object = builder.accessChainLoad(spv::NoPrecision, spv::DecorationMax, objectTypeId);
std::vector<spv::Id> components;
components.push_back(builder.createCompositeExtract(object, builder.getContainedTypeId(objectTypeId), 0));
components.push_back(builder.createCompositeExtract(object, builder.getContainedTypeId(objectTypeId), 1));
spv::Id vecType = builder.makeVectorType(builder.getContainedTypeId(objectTypeId), 2);
return builder.createUnaryOp(spv::OpBitcast, desiredTypeId,
builder.createCompositeConstruct(vecType, components));
} else {
logger->missingFunctionality("forcing 32-bit vector type to non 64-bit scalar");
}
} else {
logger->missingFunctionality("forcing non 32-bit vector type");
}
return object;
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
bool TGlslangToSpvTraverser::visitUnary(glslang::TVisit /* visit */, glslang::TIntermUnary* node)
{
SPV: Fix #1575, fix #1593: Support HLSL #line SPIR-V OpLines now contain filenames from HLSL-style #lines.
2018-12-08 00:36:33 +00:00
builder.setLine(node->getLoc().line, node->getLoc().getFilename());
SPV: Debug output: Include OpLine information for execution path. Note that declaratives are not handled, only procedurals.
2017-06-01 00:50:53 +00:00
Turn on SpecConstantOpMode based on node qualifier Move SpecConstantOpModeGuard from makeSpvConstantFromConstSubTree() to visitbinary() and visitunary(). Checking if the visiting node is a spec constants, if so, turn on the SpecConstantOpMode, otherwise, stay in the normal mode.
2016-04-04 02:20:42 +00:00
SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder);
if (node->getType().getQualifier().isSpecConstant())
spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
spv::Id result = spv::NoResult;
// try texturing first
result = createImageTextureFunctionCall(node);
if (result != spv::NoResult) {
builder.clearAccessChain();
builder.setAccessChainRValue(result);
return false; // done with this node
}
// Non-texturing.
SPV arrays: Add support for runtime-sized array types and arrays of arrays. This includes run-time block.member.length() (OpArrayLength).
2015-09-12 18:17:44 +00:00
if (node->getOp() == glslang::EOpArrayLength) {
// Quite special; won't want to evaluate the operand.
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
// Currently, the front-end does not allow .length() on an array until it is sized,
// except for the last block membeor of an SSBO.
// TODO: If this changes, link-time sized arrays might show up here, and need their
// size extracted.
SPV arrays: Add support for runtime-sized array types and arrays of arrays. This includes run-time block.member.length() (OpArrayLength).
2015-09-12 18:17:44 +00:00
// Normal .length() would have been constant folded by the front-end.
// So, this has to be block.lastMember.length().
SPV: Fix missing 'Member' operand to OpArrayLength.
2015-09-22 03:50:29 +00:00
// SPV wants "block" and member number as the operands, go get them.
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
spv::Id length;
if (node->getOperand()->getType().isCoopMat()) {
spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
spv::Id typeId = convertGlslangToSpvType(node->getOperand()->getType());
assert(builder.isCooperativeMatrixType(typeId));
length = builder.createCooperativeMatrixLength(typeId);
} else {
glslang::TIntermTyped* block = node->getOperand()->getAsBinaryNode()->getLeft();
block->traverse(this);
unsigned int member = node->getOperand()->getAsBinaryNode()->getRight()->getAsConstantUnion()->getConstArray()[0].getUConst();
length = builder.createArrayLength(builder.accessChainGetLValue(), member);
}
SPV arrays: Add support for runtime-sized array types and arrays of arrays. This includes run-time block.member.length() (OpArrayLength).
2015-09-12 18:17:44 +00:00
SPV: For GLSL only, bitcast OpArrayLength result back to int. This makes the rest of the AST consuming the result match GLSL semantics and hence get complete type matching.
2018-11-28 14:01:37 +00:00
// GLSL semantics say the result of .length() is an int, while SPIR-V says
// signedness must be 0. So, convert from SPIR-V unsigned back to GLSL's
// AST expectation of a signed result.
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
if (glslangIntermediate->getSource() == glslang::EShSourceGlsl) {
if (builder.isInSpecConstCodeGenMode()) {
length = builder.createBinOp(spv::OpIAdd, builder.makeIntType(32), length, builder.makeIntConstant(0));
} else {
length = builder.createUnaryOp(spv::OpBitcast, builder.makeIntType(32), length);
}
}
SPV: For GLSL only, bitcast OpArrayLength result back to int. This makes the rest of the AST consuming the result match GLSL semantics and hence get complete type matching.
2018-11-28 14:01:37 +00:00
SPV arrays: Add support for runtime-sized array types and arrays of arrays. This includes run-time block.member.length() (OpArrayLength).
2015-09-12 18:17:44 +00:00
builder.clearAccessChain();
builder.setAccessChainRValue(length);
return false;
}
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
// Start by evaluating the operand
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
// Does it need a swizzle inversion? If so, evaluation is inverted;
// operate first on the swizzle base, then apply the swizzle.
spv::Id invertedType = spv::NoType;
auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ? invertedType : convertGlslangToSpvType(node->getType()); };
if (node->getOp() == glslang::EOpInterpolateAtCentroid)
invertedType = getInvertedSwizzleType(*node->getOperand());
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.clearAccessChain();
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
TIntermNode *operandNode;
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
if (invertedType != spv::NoType)
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
operandNode = node->getOperand()->getAsBinaryNode()->getLeft();
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
else
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
operandNode = node->getOperand();
operandNode->traverse(this);
Merge master branch from upstream
2015-09-14 02:38:56 +00:00
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
spv::Id operand = spv::NoResult;
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
spv::Builder::AccessChain::CoherentFlags lvalueCoherentFlags;
Web: Use isEsProfile() instead of run-time testing; remove more atomics Saves 2.5K, and design is better.
2019-08-09 09:49:15 +00:00
#ifndef GLSLANG_WEB
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
if (node->getOp() == glslang::EOpAtomicCounterIncrement ||
node->getOp() == glslang::EOpAtomicCounterDecrement ||
SPV: Implement GLSL interpolation built-in functions.
2015-12-08 09:12:09 +00:00
node->getOp() == glslang::EOpAtomicCounter ||
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
node->getOp() == glslang::EOpInterpolateAtCentroid) {
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
operand = builder.accessChainGetLValue(); // Special case l-value operands
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
lvalueCoherentFlags = builder.getAccessChain().coherentFlags;
lvalueCoherentFlags |= TranslateCoherent(operandNode->getAsTyped()->getType());
} else
Web: Use isEsProfile() instead of run-time testing; remove more atomics Saves 2.5K, and design is better.
2019-08-09 09:49:15 +00:00
#endif
{
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
operand = accessChainLoad(node->getOperand()->getType());
Web: Use isEsProfile() instead of run-time testing; remove more atomics Saves 2.5K, and design is better.
2019-08-09 09:49:15 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()),
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
TranslateNoContractionDecoration(node->getType().getQualifier()),
TranslateNonUniformDecoration(node->getType().getQualifier()) };
Final round for line endings.
2015-06-26 22:58:36 +00:00
// it could be a conversion
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
if (! result)
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
result = createConversion(node->getOp(), decorations, resultType(), operand, node->getOperand()->getBasicType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
// if not, then possibly an operation
if (! result)
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
result = createUnaryOperation(node->getOp(), decorations, resultType(), operand, node->getOperand()->getBasicType(), lvalueCoherentFlags);
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (result) {
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
if (invertedType) {
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
result = createInvertedSwizzle(decorations.precision, *node->getOperand(), result);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
decorations.addNonUniform(builder, result);
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
}
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.clearAccessChain();
builder.setAccessChainRValue(result);
return false; // done with this node
}
// it must be a special case, check...
switch (node->getOp()) {
case glslang::EOpPostIncrement:
case glslang::EOpPostDecrement:
case glslang::EOpPreIncrement:
case glslang::EOpPreDecrement:
{
// we need the integer value "1" or the floating point "1.0" to add/subtract
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
spv::Id one = 0;
if (node->getBasicType() == glslang::EbtFloat)
one = builder.makeFloatConstant(1.0F);
Web: Remove unnecessary GLSL numeric types, and some collateral. This saves another 40K of x86 binary, which is about 13% of the target size.
2019-08-08 16:35:51 +00:00
#ifndef GLSLANG_WEB
SPV: Fix an issue of ++/-- for dvec/dmat.
2016-07-29 08:13:04 +00:00
else if (node->getBasicType() == glslang::EbtDouble)
one = builder.makeDoubleConstant(1.0);
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
else if (node->getBasicType() == glslang::EbtFloat16)
one = builder.makeFloat16Constant(1.0F);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
else if (node->getBasicType() == glslang::EbtInt8 || node->getBasicType() == glslang::EbtUint8)
one = builder.makeInt8Constant(1);
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
else if (node->getBasicType() == glslang::EbtInt16 || node->getBasicType() == glslang::EbtUint16)
one = builder.makeInt16Constant(1);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
else if (node->getBasicType() == glslang::EbtInt64 || node->getBasicType() == glslang::EbtUint64)
one = builder.makeInt64Constant(1);
Web: Remove unnecessary GLSL numeric types, and some collateral. This saves another 40K of x86 binary, which is about 13% of the target size.
2019-08-08 16:35:51 +00:00
#endif
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
else
one = builder.makeIntConstant(1);
Final round for line endings.
2015-06-26 22:58:36 +00:00
glslang::TOperator op;
if (node->getOp() == glslang::EOpPreIncrement ||
node->getOp() == glslang::EOpPostIncrement)
op = glslang::EOpAdd;
else
op = glslang::EOpSub;
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id result = createBinaryOperation(op, decorations,
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
convertGlslangToSpvType(node->getType()), operand, one,
node->getType().getBasicType());
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(result != spv::NoResult);
Final round for line endings.
2015-06-26 22:58:36 +00:00
// The result of operation is always stored, but conditionally the
// consumed result. The consumed result is always an r-value.
builder.accessChainStore(result);
builder.clearAccessChain();
if (node->getOp() == glslang::EOpPreIncrement ||
node->getOp() == glslang::EOpPreDecrement)
builder.setAccessChainRValue(result);
else
builder.setAccessChainRValue(operand);
}
return false;
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#ifndef GLSLANG_WEB
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpEmitStreamVertex:
builder.createNoResultOp(spv::OpEmitStreamVertex, operand);
return false;
case glslang::EOpEndStreamPrimitive:
builder.createNoResultOp(spv::OpEndStreamPrimitive, operand);
return false;
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
Create a new logger class for all messages w.r.t. SPIR-V build.
2016-05-04 19:55:59 +00:00
logger->missingFunctionality("unknown glslang unary");
SPV: Don't make MissingFunctionality() exit anymore.
2015-12-22 04:21:11 +00:00
return true; // pick up operand as placeholder result
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
}
Handle SPIR-V type mismatch when constructing a composite
2019-06-25 18:31:10 +00:00
// Construct a composite object, recursively copying members if their types don't match
spv::Id TGlslangToSpvTraverser::createCompositeConstruct(spv::Id resultTypeId, std::vector<spv::Id> constituents)
{
for (int c = 0; c < (int)constituents.size(); ++c) {
spv::Id& constituent = constituents[c];
spv::Id lType = builder.getContainedTypeId(resultTypeId, c);
spv::Id rType = builder.getTypeId(constituent);
if (lType != rType) {
if (glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4) {
constituent = builder.createUnaryOp(spv::OpCopyLogical, lType, constituent);
} else if (builder.isStructType(rType)) {
std::vector<spv::Id> rTypeConstituents;
int numrTypeConstituents = builder.getNumTypeConstituents(rType);
for (int i = 0; i < numrTypeConstituents; ++i) {
rTypeConstituents.push_back(builder.createCompositeExtract(constituent, builder.getContainedTypeId(rType, i), i));
}
constituents[c] = createCompositeConstruct(lType, rTypeConstituents);
} else {
assert(builder.isArrayType(rType));
std::vector<spv::Id> rTypeConstituents;
int numrTypeConstituents = builder.getNumTypeConstituents(rType);
spv::Id elementRType = builder.getContainedTypeId(rType);
for (int i = 0; i < numrTypeConstituents; ++i) {
rTypeConstituents.push_back(builder.createCompositeExtract(constituent, elementRType, i));
}
constituents[c] = createCompositeConstruct(lType, rTypeConstituents);
}
}
}
return builder.createCompositeConstruct(resultTypeId, constituents);
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TIntermAggregate* node)
{
Fix spec const construtor for matrix and vector Fix issue: #237 1. The code generated for matrix constructor should 1) build column vectors first, 2) build matrix with the vectors. 2. When there is only one scalar type constituent in vector's constructor, we should populate the constituent to fill all the slots in the vector. As for matrix, the single constituent should be populated to the diagonal positions (top-left to bottom-right diagonal). remove createSpvConstantFromConstSubTree()
2016-04-14 20:40:20 +00:00
SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder);
if (node->getType().getQualifier().isSpecConstant())
spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
spv::Id result = spv::NoResult;
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
spv::Id invertedType = spv::NoType; // to use to override the natural type of the node
auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ? invertedType : convertGlslangToSpvType(node->getType()); };
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
// try texturing
result = createImageTextureFunctionCall(node);
if (result != spv::NoResult) {
builder.clearAccessChain();
builder.setAccessChainRValue(result);
return false;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
}
#ifndef GLSLANG_WEB
else if (node->getOp() == glslang::EOpImageStore ||
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
node->getOp() == glslang::EOpImageStoreLod ||
node->getOp() == glslang::EOpImageAtomicStore) {
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
// "imageStore" is a special case, which has no result
return false;
}
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#endif
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
glslang::TOperator binOp = glslang::EOpNull;
bool reduceComparison = true;
bool isMatrix = false;
bool noReturnValue = false;
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
bool atomic = false;
Final round for line endings.
2015-06-26 22:58:36 +00:00
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
spv::Builder::AccessChain::CoherentFlags lvalueCoherentFlags;
Final round for line endings.
2015-06-26 22:58:36 +00:00
assert(node->getOp());
Front-end: Implement 2nd task of issue #400; precision of result and operation. From the ES spec + Bugzilla 15931 and GL_KHR_vulkan_glsl: - Update precision qualifiers for all built-in function prototypes. - Implement the new algorithm used to distinguish built-in function operation precisions from result precisions. Also add tracking of separate result and operation precisions, and use that in generating SPIR-V. (SPIR-V cares about precision of operation, while the front-end cares about precision of result, for propagation.)
2016-08-02 01:44:00 +00:00
spv::Decoration precision = TranslatePrecisionDecoration(node->getOperationPrecision());
Final round for line endings.
2015-06-26 22:58:36 +00:00
switch (node->getOp()) {
case glslang::EOpSequence:
{
if (preVisit)
++sequenceDepth;
else
--sequenceDepth;
if (sequenceDepth == 1) {
// If this is the parent node of all the functions, we want to see them
// early, so all call points have actual SPIR-V functions to reference.
// In all cases, still let the traverser visit the children for us.
makeFunctions(node->getAsAggregate()->getSequence());
Non-functional: Sweep through the stack for consistent with "main" and entry point. Partially addresses issue #513.
2016-09-19 22:01:41 +00:00
// Also, we want all globals initializers to go into the beginning of the entry point, before
Final round for line endings.
2015-06-26 22:58:36 +00:00
// anything else gets there, so visit out of order, doing them all now.
makeGlobalInitializers(node->getAsAggregate()->getSequence());
Linker: Walk the call graph to report an error on missing bodies.
2016-12-09 04:01:59 +00:00
// Initializers are done, don't want to visit again, but functions and link objects need to be processed,
Final round for line endings.
2015-06-26 22:58:36 +00:00
// so do them manually.
visitFunctions(node->getAsAggregate()->getSequence());
return false;
}
return true;
}
case glslang::EOpLinkerObjects:
{
if (visit == glslang::EvPreVisit)
linkageOnly = true;
else
linkageOnly = false;
return true;
}
case glslang::EOpComma:
{
// processing from left to right naturally leaves the right-most
// lying around in the access chain
glslang::TIntermSequence& glslangOperands = node->getSequence();
for (int i = 0; i < (int)glslangOperands.size(); ++i)
glslangOperands[i]->traverse(this);
return false;
}
case glslang::EOpFunction:
if (visit == glslang::EvPreVisit) {
Non-functional: Sweep through the stack for consistent with "main" and entry point. Partially addresses issue #513.
2016-09-19 22:01:41 +00:00
if (isShaderEntryPoint(node)) {
Non-functional: Rename some entry-point variables to entryPoint, not main.
2016-11-26 20:31:47 +00:00
inEntryPoint = true;
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.setBuildPoint(shaderEntry->getLastBlock());
HLSL: Do structure conversion for return type struct-punning on non-entry-point functions.
2016-10-06 18:59:51 +00:00
currentFunction = shaderEntry;
Final round for line endings.
2015-06-26 22:58:36 +00:00
} else {
handleFunctionEntry(node);
}
} else {
Non-functional: Rename some entry-point variables to entryPoint, not main.
2016-11-26 20:31:47 +00:00
if (inEntryPoint)
entryPointTerminated = true;
SPV return from main: Simplify a legacy design such returns are not jumps to exit block. Structured control-flow rules allow leaving the middle of a construct through a return, but not through a jump to a block that does a return. Addresses issue #58.
2015-09-15 02:58:02 +00:00
builder.leaveFunction();
Non-functional: Rename some entry-point variables to entryPoint, not main.
2016-11-26 20:31:47 +00:00
inEntryPoint = false;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
return true;
case glslang::EOpParameters:
// Parameters will have been consumed by EOpFunction processing, but not
// the body, so we still visited the function node's children, making this
// child redundant.
return false;
case glslang::EOpFunctionCall:
{
SPV: Fix #1575, fix #1593: Support HLSL #line SPIR-V OpLines now contain filenames from HLSL-style #lines.
2018-12-08 00:36:33 +00:00
builder.setLine(node->getLoc().line, node->getLoc().getFilename());
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (node->isUserDefined())
result = handleUserFunctionCall(node);
Non-functional: White space after "//", mostly for copyrights.
2017-01-06 19:34:14 +00:00
// assert(result); // this can happen for bad shaders because the call graph completeness checking is not yet done
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
if (result) {
builder.clearAccessChain();
builder.setAccessChainRValue(result);
} else
Create a new logger class for all messages w.r.t. SPIR-V build.
2016-05-04 19:55:59 +00:00
logger->missingFunctionality("missing user function; linker needs to catch that");
Final round for line endings.
2015-06-26 22:58:36 +00:00
return false;
}
case glslang::EOpConstructMat2x2:
case glslang::EOpConstructMat2x3:
case glslang::EOpConstructMat2x4:
case glslang::EOpConstructMat3x2:
case glslang::EOpConstructMat3x3:
case glslang::EOpConstructMat3x4:
case glslang::EOpConstructMat4x2:
case glslang::EOpConstructMat4x3:
case glslang::EOpConstructMat4x4:
case glslang::EOpConstructDMat2x2:
case glslang::EOpConstructDMat2x3:
case glslang::EOpConstructDMat2x4:
case glslang::EOpConstructDMat3x2:
case glslang::EOpConstructDMat3x3:
case glslang::EOpConstructDMat3x4:
case glslang::EOpConstructDMat4x2:
case glslang::EOpConstructDMat4x3:
case glslang::EOpConstructDMat4x4:
HLSL: Add imat, umat, and bmat constructors Fixes #894
2017-05-21 03:40:27 +00:00
case glslang::EOpConstructIMat2x2:
case glslang::EOpConstructIMat2x3:
case glslang::EOpConstructIMat2x4:
case glslang::EOpConstructIMat3x2:
case glslang::EOpConstructIMat3x3:
case glslang::EOpConstructIMat3x4:
case glslang::EOpConstructIMat4x2:
case glslang::EOpConstructIMat4x3:
case glslang::EOpConstructIMat4x4:
case glslang::EOpConstructUMat2x2:
case glslang::EOpConstructUMat2x3:
case glslang::EOpConstructUMat2x4:
case glslang::EOpConstructUMat3x2:
case glslang::EOpConstructUMat3x3:
case glslang::EOpConstructUMat3x4:
case glslang::EOpConstructUMat4x2:
case glslang::EOpConstructUMat4x3:
case glslang::EOpConstructUMat4x4:
case glslang::EOpConstructBMat2x2:
case glslang::EOpConstructBMat2x3:
case glslang::EOpConstructBMat2x4:
case glslang::EOpConstructBMat3x2:
case glslang::EOpConstructBMat3x3:
case glslang::EOpConstructBMat3x4:
case glslang::EOpConstructBMat4x2:
case glslang::EOpConstructBMat4x3:
case glslang::EOpConstructBMat4x4:
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
case glslang::EOpConstructF16Mat2x2:
case glslang::EOpConstructF16Mat2x3:
case glslang::EOpConstructF16Mat2x4:
case glslang::EOpConstructF16Mat3x2:
case glslang::EOpConstructF16Mat3x3:
case glslang::EOpConstructF16Mat3x4:
case glslang::EOpConstructF16Mat4x2:
case glslang::EOpConstructF16Mat4x3:
case glslang::EOpConstructF16Mat4x4:
Final round for line endings.
2015-06-26 22:58:36 +00:00
isMatrix = true;
// fall through
case glslang::EOpConstructFloat:
case glslang::EOpConstructVec2:
case glslang::EOpConstructVec3:
case glslang::EOpConstructVec4:
case glslang::EOpConstructDouble:
case glslang::EOpConstructDVec2:
case glslang::EOpConstructDVec3:
case glslang::EOpConstructDVec4:
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
case glslang::EOpConstructFloat16:
case glslang::EOpConstructF16Vec2:
case glslang::EOpConstructF16Vec3:
case glslang::EOpConstructF16Vec4:
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpConstructBool:
case glslang::EOpConstructBVec2:
case glslang::EOpConstructBVec3:
case glslang::EOpConstructBVec4:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConstructInt8:
case glslang::EOpConstructI8Vec2:
case glslang::EOpConstructI8Vec3:
case glslang::EOpConstructI8Vec4:
case glslang::EOpConstructUint8:
case glslang::EOpConstructU8Vec2:
case glslang::EOpConstructU8Vec3:
case glslang::EOpConstructU8Vec4:
case glslang::EOpConstructInt16:
case glslang::EOpConstructI16Vec2:
case glslang::EOpConstructI16Vec3:
case glslang::EOpConstructI16Vec4:
case glslang::EOpConstructUint16:
case glslang::EOpConstructU16Vec2:
case glslang::EOpConstructU16Vec3:
case glslang::EOpConstructU16Vec4:
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpConstructInt:
case glslang::EOpConstructIVec2:
case glslang::EOpConstructIVec3:
case glslang::EOpConstructIVec4:
case glslang::EOpConstructUint:
case glslang::EOpConstructUVec2:
case glslang::EOpConstructUVec3:
case glslang::EOpConstructUVec4:
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
case glslang::EOpConstructInt64:
case glslang::EOpConstructI64Vec2:
case glslang::EOpConstructI64Vec3:
case glslang::EOpConstructI64Vec4:
case glslang::EOpConstructUint64:
case glslang::EOpConstructU64Vec2:
case glslang::EOpConstructU64Vec3:
case glslang::EOpConstructU64Vec4:
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpConstructStruct:
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
case glslang::EOpConstructTextureSampler:
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
case glslang::EOpConstructReference:
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
case glslang::EOpConstructCooperativeMatrix:
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
SPV: Fix #1575, fix #1593: Support HLSL #line SPIR-V OpLines now contain filenames from HLSL-style #lines.
2018-12-08 00:36:33 +00:00
builder.setLine(node->getLoc().line, node->getLoc().getFilename());
Final round for line endings.
2015-06-26 22:58:36 +00:00
std::vector<spv::Id> arguments;
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
translateArguments(*node, arguments, lvalueCoherentFlags);
Final round for line endings.
2015-06-26 22:58:36 +00:00
spv::Id constructed;
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
if (node->getOp() == glslang::EOpConstructTextureSampler)
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
constructed = builder.createOp(spv::OpSampledImage, resultType(), arguments);
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
else if (node->getOp() == glslang::EOpConstructStruct ||
node->getOp() == glslang::EOpConstructCooperativeMatrix ||
node->getType().isArray()) {
Final round for line endings.
2015-06-26 22:58:36 +00:00
std::vector<spv::Id> constituents;
for (int c = 0; c < (int)arguments.size(); ++c)
constituents.push_back(arguments[c]);
Handle SPIR-V type mismatch when constructing a composite
2019-06-25 18:31:10 +00:00
constructed = createCompositeConstruct(resultType(), constituents);
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
} else if (isMatrix)
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
constructed = builder.createMatrixConstructor(precision, arguments, resultType());
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
else
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
constructed = builder.createConstructor(precision, arguments, resultType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.clearAccessChain();
builder.setAccessChainRValue(constructed);
return false;
}
// These six are component-wise compares with component-wise results.
// Forward on to createBinaryOperation(), requesting a vector result.
case glslang::EOpLessThan:
case glslang::EOpGreaterThan:
case glslang::EOpLessThanEqual:
case glslang::EOpGreaterThanEqual:
case glslang::EOpVectorEqual:
case glslang::EOpVectorNotEqual:
{
// Map the operation to a binary
binOp = node->getOp();
reduceComparison = false;
switch (node->getOp()) {
case glslang::EOpVectorEqual: binOp = glslang::EOpVectorEqual; break;
case glslang::EOpVectorNotEqual: binOp = glslang::EOpVectorNotEqual; break;
default: binOp = node->getOp(); break;
}
break;
}
case glslang::EOpMul:
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
// component-wise matrix multiply
Final round for line endings.
2015-06-26 22:58:36 +00:00
binOp = glslang::EOpMul;
break;
case glslang::EOpOuterProduct:
// two vectors multiplied to make a matrix
binOp = glslang::EOpOuterProduct;
break;
case glslang::EOpDot:
{
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
// for scalar dot product, use multiply
Final round for line endings.
2015-06-26 22:58:36 +00:00
glslang::TIntermSequence& glslangOperands = node->getSequence();
Full stack: distinguish between a scalar and a vector of size 1. There have been GLSL extensions considering this, and HLSL does it. This is a fully backward compatible change that allows this distinction.
2016-05-20 18:06:03 +00:00
if (glslangOperands[0]->getAsTyped()->getVectorSize() == 1)
Final round for line endings.
2015-06-26 22:58:36 +00:00
binOp = glslang::EOpMul;
break;
}
case glslang::EOpMod:
// when an aggregate, this is the floating-point mod built-in function,
// which can be emitted by the one in createBinaryOperation()
binOp = glslang::EOpMod;
break;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpEmitVertex:
case glslang::EOpEndPrimitive:
case glslang::EOpBarrier:
case glslang::EOpMemoryBarrier:
case glslang::EOpMemoryBarrierAtomicCounter:
case glslang::EOpMemoryBarrierBuffer:
case glslang::EOpMemoryBarrierImage:
case glslang::EOpMemoryBarrierShared:
case glslang::EOpGroupMemoryBarrier:
SPV: HLSL: Move to correct HLSL barrier semantics, per Khronos recommendation.
2017-12-13 05:50:53 +00:00
case glslang::EOpDeviceMemoryBarrier:
Add HLSL memory barrier intrinsics, fix dst, add lit & EvaluateAttributeSnapped
2016-06-15 15:50:24 +00:00
case glslang::EOpAllMemoryBarrierWithGroupSync:
SPV: HLSL: Move to correct HLSL barrier semantics, per Khronos recommendation.
2017-12-13 05:50:53 +00:00
case glslang::EOpDeviceMemoryBarrierWithGroupSync:
Add HLSL memory barrier intrinsics, fix dst, add lit & EvaluateAttributeSnapped
2016-06-15 15:50:24 +00:00
case glslang::EOpWorkgroupMemoryBarrier:
case glslang::EOpWorkgroupMemoryBarrierWithGroupSync:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpSubgroupBarrier:
case glslang::EOpSubgroupMemoryBarrier:
case glslang::EOpSubgroupMemoryBarrierBuffer:
case glslang::EOpSubgroupMemoryBarrierImage:
case glslang::EOpSubgroupMemoryBarrierShared:
Final round for line endings.
2015-06-26 22:58:36 +00:00
noReturnValue = true;
// These all have 0 operands and will naturally finish up in the code below for 0 operands
break;
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
case glslang::EOpAtomicStore:
noReturnValue = true;
// fallthrough
case glslang::EOpAtomicLoad:
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
case glslang::EOpAtomicAdd:
case glslang::EOpAtomicMin:
case glslang::EOpAtomicMax:
case glslang::EOpAtomicAnd:
case glslang::EOpAtomicOr:
case glslang::EOpAtomicXor:
case glslang::EOpAtomicExchange:
case glslang::EOpAtomicCompSwap:
atomic = true;
break;
GLSL 4.6: Implement atomic counter ops and SPV_KHR_shader_atomic_counter_ops.
2017-07-23 22:08:26 +00:00
case glslang::EOpAtomicCounterAdd:
case glslang::EOpAtomicCounterSubtract:
case glslang::EOpAtomicCounterMin:
case glslang::EOpAtomicCounterMax:
case glslang::EOpAtomicCounterAnd:
case glslang::EOpAtomicCounterOr:
case glslang::EOpAtomicCounterXor:
case glslang::EOpAtomicCounterExchange:
case glslang::EOpAtomicCounterCompSwap:
builder.addExtension("SPV_KHR_shader_atomic_counter_ops");
builder.addCapability(spv::CapabilityAtomicStorageOps);
atomic = true;
break;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EOpIgnoreIntersectionNV:
case glslang::EOpTerminateRayNV:
case glslang::EOpTraceNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
case glslang::EOpExecuteCallableNV:
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
case glslang::EOpWritePackedPrimitiveIndices4x8NV:
noReturnValue = true;
break;
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
case glslang::EOpCooperativeMatrixLoad:
case glslang::EOpCooperativeMatrixStore:
noReturnValue = true;
break;
Support GL_ARB_fragment_shader_interlock
2019-06-03 16:33:50 +00:00
case glslang::EOpBeginInvocationInterlock:
case glslang::EOpEndInvocationInterlock:
builder.addExtension(spv::E_SPV_EXT_fragment_shader_interlock);
noReturnValue = true;
break;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#endif
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
break;
}
//
// See if it maps to a regular operation.
//
if (binOp != glslang::EOpNull) {
glslang::TIntermTyped* left = node->getSequence()[0]->getAsTyped();
glslang::TIntermTyped* right = node->getSequence()[1]->getAsTyped();
assert(left && right);
builder.clearAccessChain();
left->traverse(this);
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
spv::Id leftId = accessChainLoad(left->getType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.clearAccessChain();
right->traverse(this);
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
spv::Id rightId = accessChainLoad(right->getType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
SPV: Fix #1575, fix #1593: Support HLSL #line SPIR-V OpLines now contain filenames from HLSL-style #lines.
2018-12-08 00:36:33 +00:00
builder.setLine(node->getLoc().line, node->getLoc().getFilename());
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
OpDecorations decorations = { precision,
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
TranslateNoContractionDecoration(node->getType().getQualifier()),
TranslateNonUniformDecoration(node->getType().getQualifier()) };
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
result = createBinaryOperation(binOp, decorations,
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
resultType(), leftId, rightId,
Final round for line endings.
2015-06-26 22:58:36 +00:00
left->getType().getBasicType(), reduceComparison);
// code above should only make binOp that exists in createBinaryOperation
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(result != spv::NoResult);
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.clearAccessChain();
builder.setAccessChainRValue(result);
return false;
}
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
//
// Create the list of operands.
//
Final round for line endings.
2015-06-26 22:58:36 +00:00
glslang::TIntermSequence& glslangOperands = node->getSequence();
std::vector<spv::Id> operands;
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
std::vector<spv::IdImmediate> memoryAccessOperands;
Final round for line endings.
2015-06-26 22:58:36 +00:00
for (int arg = 0; arg < (int)glslangOperands.size(); ++arg) {
// special case l-value operands; there are just a few
bool lvalue = false;
switch (node->getOp()) {
case glslang::EOpModf:
if (arg == 1)
lvalue = true;
break;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
case glslang::EOpFrexp:
if (arg == 1)
lvalue = true;
break;
SPV: Implement GLSL interpolation built-in functions.
2015-12-08 09:12:09 +00:00
case glslang::EOpInterpolateAtSample:
case glslang::EOpInterpolateAtOffset:
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
case glslang::EOpInterpolateAtVertex:
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
if (arg == 0) {
SPV: Implement GLSL interpolation built-in functions.
2015-12-08 09:12:09 +00:00
lvalue = true;
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
// Does it need a swizzle inversion? If so, evaluation is inverted;
// operate first on the swizzle base, then apply the swizzle.
Non-Functional: Whitespace, comments, replace accidentally deleted comment. - fixed ParseHelper.cpp newlines (crlf -> lf) - removed trailing white space in most source files - fix some spelling issues - extra blank lines - tabs to spaces - replace #include comment about no location
2017-01-06 07:34:48 +00:00
if (glslangOperands[0]->getAsOperator() &&
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
glslangOperands[0]->getAsOperator()->getOp() == glslang::EOpVectorSwizzle)
invertedType = convertGlslangToSpvType(glslangOperands[0]->getAsBinaryNode()->getLeft()->getType());
}
SPV: Implement GLSL interpolation built-in functions.
2015-12-08 09:12:09 +00:00
break;
SPIRV: Token layout of OpAtomicXXX instructions is incorrect #70
2015-09-06 08:30:11 +00:00
case glslang::EOpAtomicAdd:
case glslang::EOpAtomicMin:
case glslang::EOpAtomicMax:
case glslang::EOpAtomicAnd:
case glslang::EOpAtomicOr:
case glslang::EOpAtomicXor:
case glslang::EOpAtomicExchange:
case glslang::EOpAtomicCompSwap:
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
case glslang::EOpAtomicLoad:
case glslang::EOpAtomicStore:
GLSL 4.6: Implement atomic counter ops and SPV_KHR_shader_atomic_counter_ops.
2017-07-23 22:08:26 +00:00
case glslang::EOpAtomicCounterAdd:
case glslang::EOpAtomicCounterSubtract:
case glslang::EOpAtomicCounterMin:
case glslang::EOpAtomicCounterMax:
case glslang::EOpAtomicCounterAnd:
case glslang::EOpAtomicCounterOr:
case glslang::EOpAtomicCounterXor:
case glslang::EOpAtomicCounterExchange:
case glslang::EOpAtomicCounterCompSwap:
SPIRV: Token layout of OpAtomicXXX instructions is incorrect #70
2015-09-06 08:30:11 +00:00
if (arg == 0)
lvalue = true;
break;
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
case glslang::EOpAddCarry:
case glslang::EOpSubBorrow:
if (arg == 2)
lvalue = true;
break;
case glslang::EOpUMulExtended:
case glslang::EOpIMulExtended:
if (arg >= 2)
lvalue = true;
break;
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
case glslang::EOpCooperativeMatrixLoad:
if (arg == 0 || arg == 1)
lvalue = true;
break;
case glslang::EOpCooperativeMatrixStore:
if (arg == 1)
lvalue = true;
break;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
break;
}
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
builder.clearAccessChain();
if (invertedType != spv::NoType && arg == 0)
glslangOperands[0]->getAsBinaryNode()->getLeft()->traverse(this);
else
glslangOperands[arg]->traverse(this);
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
if (node->getOp() == glslang::EOpCooperativeMatrixLoad ||
node->getOp() == glslang::EOpCooperativeMatrixStore) {
if (arg == 1) {
// fold "element" parameter into the access chain
spv::Builder::AccessChain save = builder.getAccessChain();
builder.clearAccessChain();
glslangOperands[2]->traverse(this);
spv::Id elementId = accessChainLoad(glslangOperands[2]->getAsTyped()->getType());
builder.setAccessChain(save);
// Point to the first element of the array.
builder.accessChainPush(elementId, TranslateCoherent(glslangOperands[arg]->getAsTyped()->getType()),
Move getBufferReferenceAlignment to be a method of TType This is a better place for it logically, since it is not specific to glsl->spirv translation. And in a future change I want to use it outside of glslangtospv.
2019-03-06 19:34:10 +00:00
glslangOperands[arg]->getAsTyped()->getType().getBufferReferenceAlignment());
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
spv::Builder::AccessChain::CoherentFlags coherentFlags = builder.getAccessChain().coherentFlags;
unsigned int alignment = builder.getAccessChain().alignment;
int memoryAccess = TranslateMemoryAccess(coherentFlags);
if (node->getOp() == glslang::EOpCooperativeMatrixLoad)
memoryAccess &= ~spv::MemoryAccessMakePointerAvailableKHRMask;
if (node->getOp() == glslang::EOpCooperativeMatrixStore)
memoryAccess &= ~spv::MemoryAccessMakePointerVisibleKHRMask;
if (builder.getStorageClass(builder.getAccessChain().base) == spv::StorageClassPhysicalStorageBufferEXT) {
memoryAccess = (spv::MemoryAccessMask)(memoryAccess | spv::MemoryAccessAlignedMask);
}
memoryAccessOperands.push_back(spv::IdImmediate(false, memoryAccess));
if (memoryAccess & spv::MemoryAccessAlignedMask) {
memoryAccessOperands.push_back(spv::IdImmediate(false, alignment));
}
if (memoryAccess & (spv::MemoryAccessMakePointerAvailableKHRMask | spv::MemoryAccessMakePointerVisibleKHRMask)) {
memoryAccessOperands.push_back(spv::IdImmediate(true, builder.makeUintConstant(TranslateMemoryScope(coherentFlags))));
}
} else if (arg == 2) {
continue;
}
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
if (lvalue) {
Final round for line endings.
2015-06-26 22:58:36 +00:00
operands.push_back(builder.accessChainGetLValue());
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
lvalueCoherentFlags = builder.getAccessChain().coherentFlags;
lvalueCoherentFlags |= TranslateCoherent(glslangOperands[arg]->getAsTyped()->getType());
} else {
SPV: Fix #1575, fix #1593: Support HLSL #line SPIR-V OpLines now contain filenames from HLSL-style #lines.
2018-12-08 00:36:33 +00:00
builder.setLine(node->getLoc().line, node->getLoc().getFilename());
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
operands.push_back(accessChainLoad(glslangOperands[arg]->getAsTyped()->getType()));
SPV: Debug output: Include OpLine information for execution path. Note that declaratives are not handled, only procedurals.
2017-06-01 00:50:53 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
SPV: Fix #1575, fix #1593: Support HLSL #line SPIR-V OpLines now contain filenames from HLSL-style #lines.
2018-12-08 00:36:33 +00:00
builder.setLine(node->getLoc().line, node->getLoc().getFilename());
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
if (node->getOp() == glslang::EOpCooperativeMatrixLoad) {
std::vector<spv::IdImmediate> idImmOps;
idImmOps.push_back(spv::IdImmediate(true, operands[1])); // buf
idImmOps.push_back(spv::IdImmediate(true, operands[2])); // stride
idImmOps.push_back(spv::IdImmediate(true, operands[3])); // colMajor
idImmOps.insert(idImmOps.end(), memoryAccessOperands.begin(), memoryAccessOperands.end());
// get the pointee type
spv::Id typeId = builder.getContainedTypeId(builder.getTypeId(operands[0]));
assert(builder.isCooperativeMatrixType(typeId));
// do the op
spv::Id result = builder.createOp(spv::OpCooperativeMatrixLoadNV, typeId, idImmOps);
// store the result to the pointer (out param 'm')
builder.createStore(result, operands[0]);
result = 0;
} else if (node->getOp() == glslang::EOpCooperativeMatrixStore) {
std::vector<spv::IdImmediate> idImmOps;
idImmOps.push_back(spv::IdImmediate(true, operands[1])); // buf
idImmOps.push_back(spv::IdImmediate(true, operands[0])); // object
idImmOps.push_back(spv::IdImmediate(true, operands[2])); // stride
idImmOps.push_back(spv::IdImmediate(true, operands[3])); // colMajor
idImmOps.insert(idImmOps.end(), memoryAccessOperands.begin(), memoryAccessOperands.end());
builder.createNoResultOp(spv::OpCooperativeMatrixStoreNV, idImmOps);
result = 0;
} else if (atomic) {
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
// Handle all atomics
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
result = createAtomicOperation(node->getOp(), precision, resultType(), operands, node->getBasicType(), lvalueCoherentFlags);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
} else
#endif
{
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
// Pass through to generic operations.
switch (glslangOperands.size()) {
case 0:
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
result = createNoArgOperation(node->getOp(), precision, resultType());
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
break;
case 1:
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
{
OpDecorations decorations = { precision,
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
TranslateNoContractionDecoration(node->getType().getQualifier()),
TranslateNonUniformDecoration(node->getType().getQualifier()) };
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
result = createUnaryOperation(
node->getOp(), decorations,
resultType(), operands.front(),
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
glslangOperands[0]->getAsTyped()->getBasicType(), lvalueCoherentFlags);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
}
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
break;
default:
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
result = createMiscOperation(node->getOp(), precision, resultType(), operands, node->getBasicType());
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
break;
}
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
if (invertedType)
result = createInvertedSwizzle(precision, *glslangOperands[0]->getAsBinaryNode(), result);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
if (noReturnValue)
return false;
if (! result) {
Create a new logger class for all messages w.r.t. SPIR-V build.
2016-05-04 19:55:59 +00:00
logger->missingFunctionality("unknown glslang aggregate");
SPV: Don't make MissingFunctionality() exit anymore.
2015-12-22 04:21:11 +00:00
return true; // pick up a child as a placeholder operand
Final round for line endings.
2015-06-26 22:58:36 +00:00
} else {
builder.clearAccessChain();
builder.setAccessChainRValue(result);
return false;
}
}
SPV: Emit OpSelect when a selection node is simple enough. Also, ensures it has a type, no disallowed side effects, or performance trade offs.
2017-01-27 03:31:11 +00:00
// This path handles both if-then-else and ?:
// The if-then-else has a node type of void, while
// ?: has either a void or a non-void node type
//
// Leaving the result, when not void:
// GLSL only has r-values as the result of a :?, but
// if we have an l-value, that can be more efficient if it will
// become the base of a complex r-value expression, because the
// next layer copies r-values into memory to use the access-chain mechanism
Final round for line endings.
2015-06-26 22:58:36 +00:00
bool TGlslangToSpvTraverser::visitSelection(glslang::TVisit /* visit */, glslang::TIntermSelection* node)
{
SPV 1.4: Use OpSelect for trivial typed non-scalar/vector expressions.
2019-01-10 11:23:06 +00:00
// see if OpSelect can handle it
const auto isOpSelectable = [&]() {
if (node->getBasicType() == glslang::EbtVoid)
return false;
// OpSelect can do all other types starting with SPV 1.4
if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_4) {
// pre-1.4, only scalars and vectors can be handled
if ((!node->getType().isScalar() && !node->getType().isVector()))
return false;
}
return true;
};
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// See if it simple and safe, or required, to execute both sides.
// Crucially, side effects must be either semantically required or avoided,
// and there are performance trade-offs.
// Return true if required or a good idea (and safe) to execute both sides,
// false otherwise.
const auto bothSidesPolicy = [&]() -> bool {
// do we have both sides?
SPV: Emit OpSelect when a selection node is simple enough. Also, ensures it has a type, no disallowed side effects, or performance trade offs.
2017-01-27 03:31:11 +00:00
if (node->getTrueBlock() == nullptr ||
node->getFalseBlock() == nullptr)
return false;
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// required? (unless we write additional code to look for side effects
// and make performance trade-offs if none are present)
if (!node->getShortCircuit())
return true;
// if not required to execute both, decide based on performance/practicality...
SPV 1.4: Use OpSelect for trivial typed non-scalar/vector expressions.
2019-01-10 11:23:06 +00:00
if (!isOpSelectable())
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
return false;
SPV: Emit OpSelect when a selection node is simple enough. Also, ensures it has a type, no disallowed side effects, or performance trade offs.
2017-01-27 03:31:11 +00:00
assert(node->getType() == node->getTrueBlock() ->getAsTyped()->getType() &&
node->getType() == node->getFalseBlock()->getAsTyped()->getType());
// return true if a single operand to ? : is okay for OpSelect
const auto operandOkay = [](glslang::TIntermTyped* node) {
SPV: Implement specialization constants for ?:.
2017-01-29 02:29:42 +00:00
return node->getAsSymbolNode() || node->getType().getQualifier().isConstant();
SPV: Emit OpSelect when a selection node is simple enough. Also, ensures it has a type, no disallowed side effects, or performance trade offs.
2017-01-27 03:31:11 +00:00
};
return operandOkay(node->getTrueBlock() ->getAsTyped()) &&
operandOkay(node->getFalseBlock()->getAsTyped());
};
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
spv::Id result = spv::NoResult; // upcoming result selecting between trueValue and falseValue
// emit the condition before doing anything with selection
node->getCondition()->traverse(this);
spv::Id condition = accessChainLoad(node->getCondition()->getType());
// Find a way of executing both sides and selecting the right result.
const auto executeBothSides = [&]() -> void {
// execute both sides
SPV: Emit OpSelect when a selection node is simple enough. Also, ensures it has a type, no disallowed side effects, or performance trade offs.
2017-01-27 03:31:11 +00:00
node->getTrueBlock()->traverse(this);
spv::Id trueValue = accessChainLoad(node->getTrueBlock()->getAsTyped()->getType());
node->getFalseBlock()->traverse(this);
spv::Id falseValue = accessChainLoad(node->getTrueBlock()->getAsTyped()->getType());
SPV: Fix #1575, fix #1593: Support HLSL #line SPIR-V OpLines now contain filenames from HLSL-style #lines.
2018-12-08 00:36:33 +00:00
builder.setLine(node->getLoc().line, node->getLoc().getFilename());
SPV: Debug output: Include OpLine information for execution path. Note that declaratives are not handled, only procedurals.
2017-06-01 00:50:53 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// done if void
if (node->getBasicType() == glslang::EbtVoid)
return;
Fix #809: smear scalar condition in OpSelect for selecting vector operands.
2017-03-30 16:09:28 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// emit code to select between trueValue and falseValue
SPV: Emit OpSelect when a selection node is simple enough. Also, ensures it has a type, no disallowed side effects, or performance trade offs.
2017-01-27 03:31:11 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// see if OpSelect can handle it
SPV 1.4: Use OpSelect for trivial typed non-scalar/vector expressions.
2019-01-10 11:23:06 +00:00
if (isOpSelectable()) {
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// Emit OpSelect for this selection.
SPV: Emit OpSelect when a selection node is simple enough. Also, ensures it has a type, no disallowed side effects, or performance trade offs.
2017-01-27 03:31:11 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// smear condition to vector, if necessary (AST is always scalar)
SPV 1.4: Use OpSelect for trivial typed non-scalar/vector expressions.
2019-01-10 11:23:06 +00:00
// Before 1.4, smear like for mix(), starting with 1.4, keep it scalar
if (glslangIntermediate->getSpv().spv < glslang::EShTargetSpv_1_4 && builder.isVector(trueValue)) {
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
condition = builder.smearScalar(spv::NoPrecision, condition,
builder.makeVectorType(builder.makeBoolType(),
builder.getNumComponents(trueValue)));
SPV 1.4: Use OpSelect for trivial typed non-scalar/vector expressions.
2019-01-10 11:23:06 +00:00
}
SPV: Implement specialization constants for ?:.
2017-01-29 02:29:42 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// OpSelect
result = builder.createTriOp(spv::OpSelect,
convertGlslangToSpvType(node->getType()), condition,
trueValue, falseValue);
Final round for line endings.
2015-06-26 22:58:36 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
builder.clearAccessChain();
builder.setAccessChainRValue(result);
} else {
// We need control flow to select the result.
// TODO: Once SPIR-V OpSelect allows arbitrary types, eliminate this path.
result = builder.createVariable(spv::StorageClassFunction, convertGlslangToSpvType(node->getType()));
SPV: Emit OpSelect when a selection node is simple enough. Also, ensures it has a type, no disallowed side effects, or performance trade offs.
2017-01-27 03:31:11 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// Selection control:
const spv::SelectionControlMask control = TranslateSelectionControl(*node);
Final round for line endings.
2015-06-26 22:58:36 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// make an "if" based on the value created by the condition
spv::Builder::If ifBuilder(condition, control, builder);
HLSL: Translate directive [flatten] and [branch] to SPV control mask.
2017-07-04 15:23:40 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// emit the "then" statement
builder.createStore(trueValue, result);
ifBuilder.makeBeginElse();
// emit the "else" statement
builder.createStore(falseValue, result);
Final round for line endings.
2015-06-26 22:58:36 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// finish off the control flow
ifBuilder.makeEndIf();
Final round for line endings.
2015-06-26 22:58:36 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
builder.clearAccessChain();
builder.setAccessChainLValue(result);
}
};
Final round for line endings.
2015-06-26 22:58:36 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// Execute the one side needed, as per the condition
const auto executeOneSide = [&]() {
// Always emit control flow.
if (node->getBasicType() != glslang::EbtVoid)
result = builder.createVariable(spv::StorageClassFunction, convertGlslangToSpvType(node->getType()));
Final round for line endings.
2015-06-26 22:58:36 +00:00
HLSL: Fix #1249: Always execute both sides of ternary "?:". This is semantically required by HLSL, and frequently results in using OpSelect instead of control flow.
2018-02-21 04:29:05 +00:00
// Selection control:
const spv::SelectionControlMask control = TranslateSelectionControl(*node);
// make an "if" based on the value created by the condition
spv::Builder::If ifBuilder(condition, control, builder);
// emit the "then" statement
if (node->getTrueBlock() != nullptr) {
node->getTrueBlock()->traverse(this);
if (result != spv::NoResult)
builder.createStore(accessChainLoad(node->getTrueBlock()->getAsTyped()->getType()), result);
}
if (node->getFalseBlock() != nullptr) {
ifBuilder.makeBeginElse();
// emit the "else" statement
node->getFalseBlock()->traverse(this);
if (result != spv::NoResult)
builder.createStore(accessChainLoad(node->getFalseBlock()->getAsTyped()->getType()), result);
}
// finish off the control flow
ifBuilder.makeEndIf();
if (result != spv::NoResult) {
builder.clearAccessChain();
builder.setAccessChainLValue(result);
}
};
// Try for OpSelect (or a requirement to execute both sides)
if (bothSidesPolicy()) {
SpecConstantOpModeGuard spec_constant_op_mode_setter(&builder);
if (node->getType().getQualifier().isSpecConstant())
spec_constant_op_mode_setter.turnOnSpecConstantOpMode();
executeBothSides();
} else
executeOneSide();
Final round for line endings.
2015-06-26 22:58:36 +00:00
return false;
}
bool TGlslangToSpvTraverser::visitSwitch(glslang::TVisit /* visit */, glslang::TIntermSwitch* node)
{
// emit and get the condition before doing anything with switch
node->getCondition()->traverse(this);
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
spv::Id selector = accessChainLoad(node->getCondition()->getAsTyped()->getType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
HLSL: Translate directive [flatten] and [branch] to SPV control mask.
2017-07-04 15:23:40 +00:00
// Selection control:
HLSL: Refactor attribute implementation. - make it sharable with GLSL - correct the case insensitivity - remove the map; queries are not needed, all entries need processing - make it easier to build bottom up (will help GLSL parsing) - support semantic checking and reporting - allow front-end dependent semantics and attribute name mapping
2018-01-30 18:01:39 +00:00
const spv::SelectionControlMask control = TranslateSwitchControl(*node);
HLSL: Translate directive [flatten] and [branch] to SPV control mask.
2017-07-04 15:23:40 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
// browse the children to sort out code segments
int defaultSegment = -1;
std::vector<TIntermNode*> codeSegments;
glslang::TIntermSequence& sequence = node->getBody()->getSequence();
std::vector<int> caseValues;
std::vector<int> valueIndexToSegment(sequence.size()); // note: probably not all are used, it is an overestimate
for (glslang::TIntermSequence::iterator c = sequence.begin(); c != sequence.end(); ++c) {
TIntermNode* child = *c;
if (child->getAsBranchNode() && child->getAsBranchNode()->getFlowOp() == glslang::EOpDefault)
on x64, size_t is 64-bit and int is 32-bit, so needs an explicit cast
2015-07-12 09:32:58 +00:00
defaultSegment = (int)codeSegments.size();
Final round for line endings.
2015-06-26 22:58:36 +00:00
else if (child->getAsBranchNode() && child->getAsBranchNode()->getFlowOp() == glslang::EOpCase) {
on x64, size_t is 64-bit and int is 32-bit, so needs an explicit cast
2015-07-12 09:32:58 +00:00
valueIndexToSegment[caseValues.size()] = (int)codeSegments.size();
Final round for line endings.
2015-06-26 22:58:36 +00:00
caseValues.push_back(child->getAsBranchNode()->getExpression()->getAsConstantUnion()->getConstArray()[0].getIConst());
} else
codeSegments.push_back(child);
}
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
// handle the case where the last code segment is missing, due to no code
Final round for line endings.
2015-06-26 22:58:36 +00:00
// statements between the last case and the end of the switch statement
if ((caseValues.size() && (int)codeSegments.size() == valueIndexToSegment[caseValues.size() - 1]) ||
(int)codeSegments.size() == defaultSegment)
codeSegments.push_back(nullptr);
// make the switch statement
std::vector<spv::Block*> segmentBlocks; // returned, as the blocks allocated in the call
HLSL: Translate directive [flatten] and [branch] to SPV control mask.
2017-07-04 15:23:40 +00:00
builder.makeSwitch(selector, control, (int)codeSegments.size(), caseValues, valueIndexToSegment, defaultSegment, segmentBlocks);
Final round for line endings.
2015-06-26 22:58:36 +00:00
// emit all the code in the segments
breakForLoop.push(false);
for (unsigned int s = 0; s < codeSegments.size(); ++s) {
builder.nextSwitchSegment(segmentBlocks, s);
if (codeSegments[s])
codeSegments[s]->traverse(this);
else
builder.addSwitchBreak();
}
breakForLoop.pop();
builder.endSwitch(segmentBlocks);
return false;
}
void TGlslangToSpvTraverser::visitConstantUnion(glslang::TIntermConstantUnion* node)
{
int nextConst = 0;
Support specialization composite constants Fix issue #163, support creation and reference of composite type specialization constants. e.g.: ``` layout(constant_id = 200) const float myfloat = 1.25; layout(constant_id = 201) const int myint = 14; struct structtype { float f; int i; }; const structtype outer_struct_var = {myfloat, myint}; void main(){} ``` generated code (use glslangValidator): ``` // Module Version 10000 // Generated by (magic number): 80001 // Id's are bound by 12 Capability Shader 1: ExtInstImport "GLSL.std.450" MemoryModel Logical GLSL450 EntryPoint Vertex 4 "main" Source GLSL 450 Name 4 "main" Name 10 "structtype" MemberName 10(structtype) 0 "f" MemberName 10(structtype) 1 "i" Decorate 7 SpecId 200 Decorate 9 SpecId 201 2: TypeVoid 3: TypeFunction 2 6: TypeFloat 32 7: 6(float) SpecConstant 1067450368 8: TypeInt 32 1 9: 8(int) SpecConstant 14 10(structtype): TypeStruct 6(float) 8(int) 11:10(structtype) SpecConstantComposite 7 9 4(main): 2 Function None 3 5: Label Return FunctionEnd ``` Rname two function names to match their functionalities. 1) Rename `GlslangToSpvTraverser::createSpvSpecConstant()` to `createSpvConstant()`; 2) Rename `GlslangToSpvTraverser::createSpvConstant()` to `createSpvConstantFromConstUnionArray()` Add function `GlslangToSpvTraverser::createSpvConstantFromSubTree()` to handle constant creation from sub trees (e.g.: specialization constants). Related PR: #208
2016-03-21 13:51:37 +00:00
spv::Id constant = createSpvConstantFromConstUnionArray(node->getType(), node->getConstArray(), nextConst, false);
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.clearAccessChain();
builder.setAccessChainRValue(constant);
}
bool TGlslangToSpvTraverser::visitLoop(glslang::TVisit /* visit */, glslang::TIntermLoop* node)
{
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
auto blocks = builder.makeNewLoop();
Fix back-branch target for do-while loops. To ensure back branches always go to a header block, create a header block even for !testFirst loops. Then unify common code between the testFirst/!testFirst cases. Generate the header-block code first, so update golden files. Realize that certain infinite loops generate invalid SPIR-V, so put a TODO to instead abort code generation in such cases. Change-Id: I1e173c8f73daad186cfc666b7d72bd563ed7665d
2016-01-11 20:57:11 +00:00
builder.createBranch(&blocks.head);
HLSL: implement [unroll] and [loop] attributes This adds infrastructure suitable for any front end to create SPIR-V loop control flags. The only current front end doing so is HLSL. [unroll] turns into spv::LoopControlUnrollMask [loop] turns into spv::LoopControlDontUnrollMask no specification means spv::LoopControlMaskNone
2017-05-03 02:14:50 +00:00
// Loop control:
SPV 1.4: Implement the 5 new loop controls.
2019-01-12 10:31:41 +00:00
std::vector<unsigned int> operands;
const spv::LoopControlMask control = TranslateLoopControl(*node, operands);
HLSL: implement [unroll] and [loop] attributes This adds infrastructure suitable for any front end to create SPIR-V loop control flags. The only current front end doing so is HLSL. [unroll] turns into spv::LoopControlUnrollMask [loop] turns into spv::LoopControlDontUnrollMask no specification means spv::LoopControlMaskNone
2017-05-03 02:14:50 +00:00
Split loop header from condition testing for for/while loops.
2016-01-20 16:51:43 +00:00
// Spec requires back edges to target header blocks, and every header block
// must dominate its merge block. Make a header block first to ensure these
// conditions are met. By definition, it will contain OpLoopMerge, followed
// by a block-ending branch. But we don't want to put any other body/test
// instructions in it, since the body/test may have arbitrary instructions,
// including merges of its own.
SPV: Fix #1575, fix #1593: Support HLSL #line SPIR-V OpLines now contain filenames from HLSL-style #lines.
2018-12-08 00:36:33 +00:00
builder.setLine(node->getLoc().line, node->getLoc().getFilename());
Split loop header from condition testing for for/while loops.
2016-01-20 16:51:43 +00:00
builder.setBuildPoint(&blocks.head);
SPV 1.4: Implement the 5 new loop controls.
2019-01-12 10:31:41 +00:00
builder.createLoopMerge(&blocks.merge, &blocks.continue_target, control, operands);
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
if (node->testFirst() && node->getTest()) {
Split loop header from condition testing for for/while loops.
2016-01-20 16:51:43 +00:00
spv::Block& test = builder.makeNewBlock();
builder.createBranch(&test);
builder.setBuildPoint(&test);
Final round for line endings.
2015-06-26 22:58:36 +00:00
node->getTest()->traverse(this);
SPV: Debug output: Include OpLine information for execution path. Note that declaratives are not handled, only procedurals.
2017-06-01 00:50:53 +00:00
spv::Id condition = accessChainLoad(node->getTest()->getType());
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
builder.createConditionalBranch(condition, &blocks.body, &blocks.merge);
builder.setBuildPoint(&blocks.body);
Fix back-branch target for do-while loops. To ensure back branches always go to a header block, create a header block even for !testFirst loops. Then unify common code between the testFirst/!testFirst cases. Generate the header-block code first, so update golden files. Realize that certain infinite loops generate invalid SPIR-V, so put a TODO to instead abort code generation in such cases. Change-Id: I1e173c8f73daad186cfc666b7d72bd563ed7665d
2016-01-11 20:57:11 +00:00
breakForLoop.push(true);
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
if (node->getBody())
Fix unreachable-block removal. Add a test for loop without a condition. Change-Id: Idd7fc462218a84b1e745207e2975a3f2897d30a0
2016-01-11 00:37:00 +00:00
node->getBody()->traverse(this);
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
builder.createBranch(&blocks.continue_target);
Update .out files with new loop SPIR-V. Also update breakForLoop and call builder.closeLoop() in code generation. Remove dead code. Change-Id: Ic8ce5b208530f8787989ef45a2efa405f1b45310
2016-01-11 19:48:36 +00:00
breakForLoop.pop();
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
builder.setBuildPoint(&blocks.continue_target);
if (node->getTerminal())
node->getTerminal()->traverse(this);
Fix back-branch target for do-while loops. To ensure back branches always go to a header block, create a header block even for !testFirst loops. Then unify common code between the testFirst/!testFirst cases. Generate the header-block code first, so update golden files. Realize that certain infinite loops generate invalid SPIR-V, so put a TODO to instead abort code generation in such cases. Change-Id: I1e173c8f73daad186cfc666b7d72bd563ed7665d
2016-01-11 20:57:11 +00:00
builder.createBranch(&blocks.head);
Generate correctly structured do-while loops. The loop test is always emitted before the loop body. For do-while loops, use a phi node to track whether we're on the first loop iteration, and only check the loop test on the second and subsequent iterations. For do-while loops, the loop test branch no longer occurs at the top of the loop, so it must get its own selection merge instruction. A block can't be the target of more than one merge instruction. So when the loop test executes after the body (as in do-while in GLSL) we need to introduce a dummy block to be the target of the selection merge just before the loop test conditional branch. The other arm of the branch exits the loop and hence is the "break block" exception in the structured control flow rules.
2015-07-15 20:21:26 +00:00
} else {
SPV: Fix #1575, fix #1593: Support HLSL #line SPIR-V OpLines now contain filenames from HLSL-style #lines.
2018-12-08 00:36:33 +00:00
builder.setLine(node->getLoc().line, node->getLoc().getFilename());
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
builder.createBranch(&blocks.body);
Update .out files with new loop SPIR-V. Also update breakForLoop and call builder.closeLoop() in code generation. Remove dead code. Change-Id: Ic8ce5b208530f8787989ef45a2efa405f1b45310
2016-01-11 19:48:36 +00:00
breakForLoop.push(true);
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
builder.setBuildPoint(&blocks.body);
if (node->getBody())
Fix unreachable-block removal. Add a test for loop without a condition. Change-Id: Idd7fc462218a84b1e745207e2975a3f2897d30a0
2016-01-11 00:37:00 +00:00
node->getBody()->traverse(this);
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
builder.createBranch(&blocks.continue_target);
Update .out files with new loop SPIR-V. Also update breakForLoop and call builder.closeLoop() in code generation. Remove dead code. Change-Id: Ic8ce5b208530f8787989ef45a2efa405f1b45310
2016-01-11 19:48:36 +00:00
breakForLoop.pop();
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
builder.setBuildPoint(&blocks.continue_target);
if (node->getTerminal())
node->getTerminal()->traverse(this);
if (node->getTest()) {
node->getTest()->traverse(this);
spv::Id condition =
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
accessChainLoad(node->getTest()->getType());
Fix back-branch target for do-while loops. To ensure back branches always go to a header block, create a header block even for !testFirst loops. Then unify common code between the testFirst/!testFirst cases. Generate the header-block code first, so update golden files. Realize that certain infinite loops generate invalid SPIR-V, so put a TODO to instead abort code generation in such cases. Change-Id: I1e173c8f73daad186cfc666b7d72bd563ed7665d
2016-01-11 20:57:11 +00:00
builder.createConditionalBranch(condition, &blocks.head, &blocks.merge);
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
} else {
Don't remove SPIR-V blocks before codegen. A removed block releases its instructions, so Module::idToInstruction suddenly contains dangling references. The original motivation for block removal was to skip some unreachable blocks, but that's already achieved by InReadableOrder.cpp. Also updated stale comments.
2016-01-20 02:13:38 +00:00
// TODO: unless there was a break/return/discard instruction
// somewhere in the body, this is an infinite loop, so we should
// issue a warning.
Fix back-branch target for do-while loops. To ensure back branches always go to a header block, create a header block even for !testFirst loops. Then unify common code between the testFirst/!testFirst cases. Generate the header-block code first, so update golden files. Realize that certain infinite loops generate invalid SPIR-V, so put a TODO to instead abort code generation in such cases. Change-Id: I1e173c8f73daad186cfc666b7d72bd563ed7665d
2016-01-11 20:57:11 +00:00
builder.createBranch(&blocks.head);
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
First cut at new loop codegen. Change-Id: Id3bdf8b7a5606e7ce5d856ef225d5ddbe59a584b
2016-01-10 17:15:13 +00:00
builder.setBuildPoint(&blocks.merge);
Update .out files with new loop SPIR-V. Also update breakForLoop and call builder.closeLoop() in code generation. Remove dead code. Change-Id: Ic8ce5b208530f8787989ef45a2efa405f1b45310
2016-01-11 19:48:36 +00:00
builder.closeLoop();
Final round for line endings.
2015-06-26 22:58:36 +00:00
return false;
}
bool TGlslangToSpvTraverser::visitBranch(glslang::TVisit /* visit */, glslang::TIntermBranch* node)
{
if (node->getExpression())
node->getExpression()->traverse(this);
SPV: Fix #1575, fix #1593: Support HLSL #line SPIR-V OpLines now contain filenames from HLSL-style #lines.
2018-12-08 00:36:33 +00:00
builder.setLine(node->getLoc().line, node->getLoc().getFilename());
SPV: Debug output: Include OpLine information for execution path. Note that declaratives are not handled, only procedurals.
2017-06-01 00:50:53 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
switch (node->getFlowOp()) {
case glslang::EOpKill:
builder.makeDiscard();
break;
case glslang::EOpBreak:
if (breakForLoop.top())
builder.createLoopExit();
else
builder.addSwitchBreak();
break;
case glslang::EOpContinue:
builder.createLoopContinue();
break;
case glslang::EOpReturn:
HLSL: Do structure conversion for return type struct-punning on non-entry-point functions.
2016-10-06 18:59:51 +00:00
if (node->getExpression()) {
const glslang::TType& glslangReturnType = node->getExpression()->getType();
spv::Id returnId = accessChainLoad(glslangReturnType);
if (builder.getTypeId(returnId) != currentFunction->getReturnType()) {
builder.clearAccessChain();
spv::Id copyId = builder.createVariable(spv::StorageClassFunction, currentFunction->getReturnType());
builder.setAccessChainLValue(copyId);
multiTypeStore(glslangReturnType, returnId);
returnId = builder.createLoad(copyId);
}
builder.makeReturn(false, returnId);
} else
SPV return from main: Simplify a legacy design such returns are not jumps to exit block. Structured control-flow rules allow leaving the middle of a construct through a return, but not through a jump to a block that does a return. Addresses issue #58.
2015-09-15 02:58:02 +00:00
builder.makeReturn(false);
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.clearAccessChain();
break;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
Implement GL_EXT_demote_to_helper_invocation
2019-07-01 14:23:23 +00:00
case glslang::EOpDemote:
builder.createNoResultOp(spv::OpDemoteToHelperInvocationEXT);
builder.addExtension(spv::E_SPV_EXT_demote_to_helper_invocation);
builder.addCapability(spv::CapabilityDemoteToHelperInvocationEXT);
break;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
Implement GL_EXT_demote_to_helper_invocation
2019-07-01 14:23:23 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(0);
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
}
return false;
}
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
spv::Id TGlslangToSpvTraverser::createSpvVariable(const glslang::TIntermSymbol* node, spv::Id forcedType)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
// First, steer off constants, which are not SPIR-V variables, but
Final round for line endings.
2015-06-26 22:58:36 +00:00
// can still have a mapping to a SPIR-V Id.
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
// This includes specialization constants.
Front-end infrastructure: Encapsulate semantic-level questions/actions about const/temp. Much about const or temp is mechanical, about actual declaration, while much is semantic, about something higher level. This commit checks every use everywhere, and for the high-level ones, substitutes an encapsulated version instead.
2016-03-20 06:46:02 +00:00
if (node->getQualifier().isConstant()) {
Review feedback
2018-11-13 14:17:44 +00:00
spv::Id result = createSpvConstant(*node);
if (result != spv::NoResult)
return result;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
// Now, handle actual variables
SPV: Emit extension for using StorageClassStorageBuffer.
2017-05-05 11:09:58 +00:00
spv::StorageClass storageClass = TranslateStorageClass(node->getType());
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
spv::Id spvType = forcedType == spv::NoType ? convertGlslangToSpvType(node->getType())
: forcedType;
Final round for line endings.
2015-06-26 22:58:36 +00:00
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
const bool contains16BitType = node->getType().contains16BitFloat() ||
node->getType().contains16BitInt();
Implement the extension SPV_KHR_16bit_storage
2017-04-07 15:22:33 +00:00
if (contains16BitType) {
SPV: Fix #1374: For an SSBO 16-bit access, only emit StorageUniformBufferBlock16 capability.
2018-05-15 04:08:53 +00:00
switch (storageClass) {
case spv::StorageClassInput:
case spv::StorageClassOutput:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
addPre13Extension(spv::E_SPV_KHR_16bit_storage);
Implement the extension SPV_KHR_16bit_storage
2017-04-07 15:22:33 +00:00
builder.addCapability(spv::CapabilityStorageInputOutput16);
SPV: Fix #1374: For an SSBO 16-bit access, only emit StorageUniformBufferBlock16 capability.
2018-05-15 04:08:53 +00:00
break;
case spv::StorageClassUniform:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
addPre13Extension(spv::E_SPV_KHR_16bit_storage);
Implement the extension SPV_KHR_16bit_storage
2017-04-07 15:22:33 +00:00
if (node->getType().getQualifier().storage == glslang::EvqBuffer)
builder.addCapability(spv::CapabilityStorageUniformBufferBlock16);
SPV: Fix #1374: For an SSBO 16-bit access, only emit StorageUniformBufferBlock16 capability.
2018-05-15 04:08:53 +00:00
else
builder.addCapability(spv::CapabilityStorageUniform16);
break;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
case spv::StorageClassPushConstant:
addPre13Extension(spv::E_SPV_KHR_16bit_storage);
builder.addCapability(spv::CapabilityStoragePushConstant16);
break;
SPV: Fix #1374: For an SSBO 16-bit access, only emit StorageUniformBufferBlock16 capability.
2018-05-15 04:08:53 +00:00
case spv::StorageClassStorageBuffer:
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
case spv::StorageClassPhysicalStorageBufferEXT:
SPV: Fix #1374: For an SSBO 16-bit access, only emit StorageUniformBufferBlock16 capability.
2018-05-15 04:08:53 +00:00
addPre13Extension(spv::E_SPV_KHR_16bit_storage);
builder.addCapability(spv::CapabilityStorageUniformBufferBlock16);
break;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
SPV: Fix #1374: For an SSBO 16-bit access, only emit StorageUniformBufferBlock16 capability.
2018-05-15 04:08:53 +00:00
default:
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
if (node->getType().contains16BitFloat())
Add Float16/Int8/Int16 capabilities for private variables and function parameters
2019-02-18 04:49:28 +00:00
builder.addCapability(spv::CapabilityFloat16);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
if (node->getType().contains16BitInt())
Add Float16/Int8/Int16 capabilities for private variables and function parameters
2019-02-18 04:49:28 +00:00
builder.addCapability(spv::CapabilityInt16);
SPV: Fix #1374: For an SSBO 16-bit access, only emit StorageUniformBufferBlock16 capability.
2018-05-15 04:08:53 +00:00
break;
Implement the extension SPV_KHR_16bit_storage
2017-04-07 15:22:33 +00:00
}
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
if (node->getType().contains8BitInt()) {
Implement GL_EXT_shader_16bit_storage and GL_EXT_shader_8bit_storage extensions. These introduce limited support for 8/16-bit types such that they can only be accessed in buffer memory and converted to/from 32-bit types. Contributed from Khronos-internal work.
2018-07-03 19:19:51 +00:00
if (storageClass == spv::StorageClassPushConstant) {
builder.addExtension(spv::E_SPV_KHR_8bit_storage);
builder.addCapability(spv::CapabilityStoragePushConstant8);
} else if (storageClass == spv::StorageClassUniform) {
builder.addExtension(spv::E_SPV_KHR_8bit_storage);
builder.addCapability(spv::CapabilityUniformAndStorageBuffer8BitAccess);
Fix 8-bit storage nearly always using the UniformAndStorageBuffer8BitAccess capability. When using the 8-bit storage extension it basically always used the `UniformAndStorageBuffer8BitAccess` capability, even in cases where it wasn't required. For instance if we are targeting Vulkan 1.1 (SPIR-V 1.3 or higher), and we are only using 8-bit types in an SSBO, we only need the `StorageBuffer8BitAccess` capability. I fixed this by enabling storage buffer use in Vulkan 1.1 / SPIR-V 1.3 or higher, and then changing the logic to match. I also added some tests that will output different capabilities when run on Vulkan 1.0 and 1.1, thus they are added twice to the test list (one for each version). Fixes #1539
2018-10-23 14:02:29 +00:00
} else if (storageClass == spv::StorageClassStorageBuffer) {
builder.addExtension(spv::E_SPV_KHR_8bit_storage);
builder.addCapability(spv::CapabilityStorageBuffer8BitAccess);
Add Float16/Int8/Int16 capabilities for private variables and function parameters
2019-02-18 04:49:28 +00:00
} else {
builder.addCapability(spv::CapabilityInt8);
Implement GL_EXT_shader_16bit_storage and GL_EXT_shader_8bit_storage extensions. These introduce limited support for 8/16-bit types such that they can only be accessed in buffer memory and converted to/from 32-bit types. Contributed from Khronos-internal work.
2018-07-03 19:19:51 +00:00
}
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
const char* name = node->getName().c_str();
if (glslang::IsAnonymous(name))
name = "";
return builder.createVariable(storageClass, spvType, name);
}
// Return type Id of the sampled type.
spv::Id TGlslangToSpvTraverser::getSampledType(const glslang::TSampler& sampler)
{
switch (sampler.type) {
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
case glslang::EbtInt: return builder.makeIntType(32);
case glslang::EbtUint: return builder.makeUintType(32);
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EbtFloat: return builder.makeFloatType(32);
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Implement the extension GL_AMD_gpu_shader_half_float_fetch - Support new opaque types: f16sampler*, f16image*, f16subpassInput*. - Add new built-in GLSL texture/image functions.
2016-11-29 09:36:31 +00:00
case glslang::EbtFloat16:
builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float_fetch);
builder.addCapability(spv::CapabilityFloat16ImageAMD);
return builder.makeFloatType(16);
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(0);
Final round for line endings.
2015-06-26 22:58:36 +00:00
return builder.makeFloatType(32);
}
}
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
// If node is a swizzle operation, return the type that should be used if
// the swizzle base is first consumed by another operation, before the swizzle
// is applied.
spv::Id TGlslangToSpvTraverser::getInvertedSwizzleType(const glslang::TIntermTyped& node)
{
Non-Functional: Whitespace, comments, replace accidentally deleted comment. - fixed ParseHelper.cpp newlines (crlf -> lf) - removed trailing white space in most source files - fix some spelling issues - extra blank lines - tabs to spaces - replace #include comment about no location
2017-01-06 07:34:48 +00:00
if (node.getAsOperator() &&
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
node.getAsOperator()->getOp() == glslang::EOpVectorSwizzle)
return convertGlslangToSpvType(node.getAsBinaryNode()->getLeft()->getType());
else
return spv::NoType;
}
// When inverting a swizzle with a parent op, this function
// will apply the swizzle operation to a completed parent operation.
spv::Id TGlslangToSpvTraverser::createInvertedSwizzle(spv::Decoration precision, const glslang::TIntermTyped& node, spv::Id parentResult)
{
std::vector<unsigned> swizzle;
convertSwizzle(*node.getAsBinaryNode()->getRight()->getAsAggregate(), swizzle);
return builder.createRvalueSwizzle(precision, convertGlslangToSpvType(node.getType()), parentResult, swizzle);
}
// Convert a glslang AST swizzle node to a swizzle vector for building SPIR-V.
void TGlslangToSpvTraverser::convertSwizzle(const glslang::TIntermAggregate& node, std::vector<unsigned>& swizzle)
{
const glslang::TIntermSequence& swizzleSequence = node.getSequence();
for (int i = 0; i < (int)swizzleSequence.size(); ++i)
swizzle.push_back(swizzleSequence[i]->getAsConstantUnion()->getConstArray()[0].getIConst());
}
SPV: recursively propagate row/col majorness through nested structures. This includes doing structure uniqueness modulo majorness, for shared nested structures.
2015-12-20 18:29:16 +00:00
// Convert from a glslang type to an SPV type, by calling into a
// recursive version of this function. This establishes the inherited
// layout state rooted from the top-level type.
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
spv::Id TGlslangToSpvTraverser::convertGlslangToSpvType(const glslang::TType& type, bool forwardReferenceOnly)
SPV: Only decorate array and matrix strides for transparent types requiring explicit layouts.
2015-09-09 23:51:38 +00:00
{
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
return convertGlslangToSpvType(type, getExplicitLayout(type), type.getQualifier(), false, forwardReferenceOnly);
SPV: Only decorate array and matrix strides for transparent types requiring explicit layouts.
2015-09-09 23:51:38 +00:00
}
// Do full recursive conversion of an arbitrary glslang type to a SPIR-V Id.
SPV: Add recursive distrubition of 'location' across structure members.
2016-01-22 01:56:57 +00:00
// explicitLayout can be kept the same throughout the hierarchical recursive walk.
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
// Mutually recursive with convertGlslangStructToSpvType().
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id TGlslangToSpvTraverser::convertGlslangToSpvType(const glslang::TType& type,
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
glslang::TLayoutPacking explicitLayout, const glslang::TQualifier& qualifier,
bool lastBufferBlockMember, bool forwardReferenceOnly)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
spv::Id spvType = spv::NoResult;
Final round for line endings.
2015-06-26 22:58:36 +00:00
switch (type.getBasicType()) {
case glslang::EbtVoid:
spvType = builder.makeVoidType();
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert (! type.isArray());
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
case glslang::EbtBool:
// "transparent" bool doesn't exist in SPIR-V. The GLSL convention is
// a 32-bit int where non-0 means true.
if (explicitLayout != glslang::ElpNone)
spvType = builder.makeUintType(32);
else
spvType = builder.makeBoolType();
break;
case glslang::EbtInt:
spvType = builder.makeIntType(32);
break;
case glslang::EbtUint:
spvType = builder.makeUintType(32);
break;
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EbtFloat:
spvType = builder.makeFloatType(32);
break;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EbtDouble:
spvType = builder.makeFloatType(64);
break;
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
case glslang::EbtFloat16:
spvType = builder.makeFloatType(16);
break;
SPV: only declare the pure 8/16-bit capabilities when needed. Only when operations stray outside the 8/16-bit storage-capabilities are the general (pure 8/16-bit) capabilities needed.
2018-08-15 19:54:09 +00:00
case glslang::EbtInt8:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
spvType = builder.makeIntType(8);
break;
case glslang::EbtUint8:
spvType = builder.makeUintType(8);
break;
SPV: only declare the pure 8/16-bit capabilities when needed. Only when operations stray outside the 8/16-bit storage-capabilities are the general (pure 8/16-bit) capabilities needed.
2018-08-15 19:54:09 +00:00
case glslang::EbtInt16:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
spvType = builder.makeIntType(16);
break;
case glslang::EbtUint16:
spvType = builder.makeUintType(16);
break;
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
case glslang::EbtInt64:
spvType = builder.makeIntType(64);
break;
case glslang::EbtUint64:
spvType = builder.makeUintType(64);
break;
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
case glslang::EbtAtomicUint:
SPV: Turn on atomic-storage functionality for SPIR-V. This is used by OpenGL, but not Vulkan. Includes: - atomicCounter, atomicIncrement, atomicCounterDecrement - atomic_uint layout-offset checking - AtomicStorage capability
2016-07-07 19:20:00 +00:00
builder.addCapability(spv::CapabilityAtomicStorage);
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
spvType = builder.makeUintType(32);
break;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EbtAccStructNV:
spvType = builder.makeAccelerationStructureNVType();
break;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
case glslang::EbtReference:
{
// Make the forward pointer, then recurse to convert the structure type, then
// patch up the forward pointer with a real pointer type.
if (forwardPointers.find(type.getReferentType()) == forwardPointers.end()) {
spv::Id forwardId = builder.makeForwardPointer(spv::StorageClassPhysicalStorageBufferEXT);
forwardPointers[type.getReferentType()] = forwardId;
}
spvType = forwardPointers[type.getReferentType()];
if (!forwardReferenceOnly) {
spv::Id referentType = convertGlslangToSpvType(*type.getReferentType());
builder.makePointerFromForwardPointer(spv::StorageClassPhysicalStorageBufferEXT,
forwardPointers[type.getReferentType()],
referentType);
}
}
break;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EbtSampler:
{
const glslang::TSampler& sampler = type.getSampler();
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
if (sampler.isPureSampler()) {
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
spvType = builder.makeSamplerType();
} else {
// an image is present, make its type
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
spvType = builder.makeImageType(getSampledType(sampler), TranslateDimensionality(sampler),
sampler.isShadow(), sampler.isArrayed(), sampler.isMultiSample(),
sampler.isImageClass() ? 2 : 1, TranslateImageFormat(type));
if (sampler.isCombined()) {
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
// already has both image and sampler, make the combined type
spvType = builder.makeSampledImageType(spvType);
}
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
}
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EbtStruct:
case glslang::EbtBlock:
{
// If we've seen this struct type, return it
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
const glslang::TTypeList* glslangMembers = type.getStruct();
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
// Try to share structs for different layouts, but not yet for other
// kinds of qualification (primarily not yet including interpolant qualification).
SPV: Tighten up number of struct-types declared based on decoration. Takes some pressure off of issue #304. Structures don't inherit locations and then explicitly decorate members with them, so removed this reason to have another instance of a structure type.
2016-09-01 23:05:23 +00:00
if (! HasNonLayoutQualifiers(type, qualifier))
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
spvType = structMap[explicitLayout][qualifier.layoutMatrix][glslangMembers];
SPV: Generalize multiple struct-type instances for interpolation/invariant qualifiers.
2015-12-24 17:30:13 +00:00
if (spvType != spv::NoResult)
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
// else, we haven't seen it...
if (type.getBasicType() == glslang::EbtBlock)
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
memberRemapper[glslangMembers].resize(glslangMembers->size());
spvType = convertGlslangStructToSpvType(type, glslangMembers, explicitLayout, qualifier);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
break;
default:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(0);
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
}
if (type.isMatrix())
spvType = builder.makeMatrixType(spvType, type.getMatrixCols(), type.getMatrixRows());
else {
// If this variable has a vector element count greater than 1, create a SPIR-V vector
if (type.getVectorSize() > 1)
spvType = builder.makeVectorType(spvType, type.getVectorSize());
}
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
if (type.isCoopMat()) {
builder.addCapability(spv::CapabilityCooperativeMatrixNV);
builder.addExtension(spv::E_SPV_NV_cooperative_matrix);
if (type.getBasicType() == glslang::EbtFloat16)
builder.addCapability(spv::CapabilityFloat16);
spv::Id scope = makeArraySizeId(*type.getTypeParameters(), 1);
spv::Id rows = makeArraySizeId(*type.getTypeParameters(), 2);
spv::Id cols = makeArraySizeId(*type.getTypeParameters(), 3);
spvType = builder.makeCooperativeMatrixType(spvType, scope, rows, cols);
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (type.isArray()) {
SPV: Handle stride decorations for arrays of arrays, and using multiple type instances when strides are used.
2015-12-30 04:27:24 +00:00
int stride = 0; // keep this 0 unless doing an explicit layout; 0 will mean no decoration, no stride
SPV arrays: Add support for runtime-sized array types and arrays of arrays. This includes run-time block.member.length() (OpArrayLength).
2015-09-12 18:17:44 +00:00
// Do all but the outer dimension
SPV: Handle stride decorations for arrays of arrays, and using multiple type instances when strides are used.
2015-12-30 04:27:24 +00:00
if (type.getArraySizes()->getNumDims() > 1) {
SPV: Don't do stride decorations for arrays of arrays of blocks.
2016-01-05 02:22:56 +00:00
// We need to decorate array strides for types needing explicit layout, except blocks.
if (explicitLayout != glslang::ElpNone && type.getBasicType() != glslang::EbtBlock) {
SPV: Handle stride decorations for arrays of arrays, and using multiple type instances when strides are used.
2015-12-30 04:27:24 +00:00
// Use a dummy glslang type for querying internal strides of
// arrays of arrays, but using just a one-dimensional array.
glslang::TType simpleArrayType(type, 0); // deference type of the array
Types: Fix #1290: Rationalize and correct "mixed" style array dimensioning. There a couple functional problems, which when reduced down also led to some good simplifications and rationalization. So, this commit: - corrects "mixed" functionality: int[A] f[B] -> f[B][A] - correct multi-identifier decls: int[A] f[B], g[C] -> f and g are independently sized. - increases symmetry between different places in the code that do this - makes fewer ways to do the same thing; several methods are just gone now - makes more clear when something is copied or shared
2018-03-26 06:38:53 +00:00
while (simpleArrayType.getArraySizes()->getNumDims() > 1)
simpleArrayType.getArraySizes()->dereference();
SPV: Handle stride decorations for arrays of arrays, and using multiple type instances when strides are used.
2015-12-30 04:27:24 +00:00
// Will compute the higher-order strides here, rather than making a whole
// pile of types and doing repetitive recursion on their contents.
stride = getArrayStride(simpleArrayType, explicitLayout, qualifier.layoutMatrix);
}
SPV: Don't do stride decorations for arrays of arrays of blocks.
2016-01-05 02:22:56 +00:00
// make the arrays
SPV: Handle stride decorations for arrays of arrays, and using multiple type instances when strides are used.
2015-12-30 04:27:24 +00:00
for (int dim = type.getArraySizes()->getNumDims() - 1; dim > 0; --dim) {
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
spvType = builder.makeArrayType(spvType, makeArraySizeId(*type.getArraySizes(), dim), stride);
SPV: Handle stride decorations for arrays of arrays, and using multiple type instances when strides are used.
2015-12-30 04:27:24 +00:00
if (stride > 0)
builder.addDecoration(spvType, spv::DecorationArrayStride, stride);
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
stride *= type.getArraySizes()->getDimSize(dim);
SPV: Handle stride decorations for arrays of arrays, and using multiple type instances when strides are used.
2015-12-30 04:27:24 +00:00
}
} else {
// single-dimensional array, and don't yet have stride
SPV: Don't do stride decorations for arrays of arrays of blocks.
2016-01-05 02:22:56 +00:00
// We need to decorate array strides for types needing explicit layout, except blocks.
SPV: Handle stride decorations for arrays of arrays, and using multiple type instances when strides are used.
2015-12-30 04:27:24 +00:00
if (explicitLayout != glslang::ElpNone && type.getBasicType() != glslang::EbtBlock)
stride = getArrayStride(type, explicitLayout, qualifier.layoutMatrix);
SPV arrays: Add support for runtime-sized array types and arrays of arrays. This includes run-time block.member.length() (OpArrayLength).
2015-09-12 18:17:44 +00:00
}
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
// Do the outer dimension, which might not be known for a runtime-sized array.
// (Unsized arrays that survive through linking will be runtime-sized arrays)
if (type.isSizedArray())
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
spvType = builder.makeArrayType(spvType, makeArraySizeId(*type.getArraySizes(), 0), stride);
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
else {
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
if (!lastBufferBlockMember) {
builder.addExtension("SPV_EXT_descriptor_indexing");
builder.addCapability(spv::CapabilityRuntimeDescriptorArrayEXT);
}
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spvType = builder.makeRuntimeArray(spvType);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
}
SPV: Handle stride decorations for arrays of arrays, and using multiple type instances when strides are used.
2015-12-30 04:27:24 +00:00
if (stride > 0)
builder.addDecoration(spvType, spv::DecorationArrayStride, stride);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
return spvType;
}
Fix #790: Don't emit NV-specific interface members unless enabled by extension.
2017-03-25 00:38:16 +00:00
// TODO: this functionality should exist at a higher level, in creating the AST
//
// Identify interface members that don't have their required extension turned on.
//
bool TGlslangToSpvTraverser::filterMember(const glslang::TType& member)
{
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Fix #790: Don't emit NV-specific interface members unless enabled by extension.
2017-03-25 00:38:16 +00:00
auto& extensions = glslangIntermediate->getRequestedExtensions();
Don't emit NV-specific interface members if NV extensions are disabled.
2017-03-29 15:01:36 +00:00
if (member.getFieldName() == "gl_SecondaryViewportMaskNV" &&
extensions.find("GL_NV_stereo_view_rendering") == extensions.end())
return true;
Fix #790: Don't emit NV-specific interface members unless enabled by extension.
2017-03-25 00:38:16 +00:00
if (member.getFieldName() == "gl_SecondaryPositionNV" &&
extensions.find("GL_NV_stereo_view_rendering") == extensions.end())
return true;
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
if (glslangIntermediate->getStage() != EShLangMeshNV) {
if (member.getFieldName() == "gl_ViewportMask" &&
extensions.find("GL_NV_viewport_array2") == extensions.end())
return true;
if (member.getFieldName() == "gl_PositionPerViewNV" &&
extensions.find("GL_NVX_multiview_per_view_attributes") == extensions.end())
return true;
if (member.getFieldName() == "gl_ViewportMaskPerViewNV" &&
extensions.find("GL_NVX_multiview_per_view_attributes") == extensions.end())
return true;
}
#endif
Fix #790: Don't emit NV-specific interface members unless enabled by extension.
2017-03-25 00:38:16 +00:00
return false;
};
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
// Do full recursive conversion of a glslang structure (or block) type to a SPIR-V Id.
// explicitLayout can be kept the same throughout the hierarchical recursive walk.
// Mutually recursive with convertGlslangToSpvType().
spv::Id TGlslangToSpvTraverser::convertGlslangStructToSpvType(const glslang::TType& type,
const glslang::TTypeList* glslangMembers,
glslang::TLayoutPacking explicitLayout,
const glslang::TQualifier& qualifier)
{
// Create a vector of struct types for SPIR-V to consume
std::vector<spv::Id> spvMembers;
int memberDelta = 0; // how much the member's index changes from glslang to SPIR-V, normally 0, except sometimes for blocks
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
std::vector<std::pair<glslang::TType*, glslang::TQualifier> > deferredForwardPointers;
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
for (int i = 0; i < (int)glslangMembers->size(); i++) {
glslang::TType& glslangMember = *(*glslangMembers)[i].type;
if (glslangMember.hiddenMember()) {
++memberDelta;
if (type.getBasicType() == glslang::EbtBlock)
memberRemapper[glslangMembers][i] = -1;
} else {
Fix #790: Don't emit NV-specific interface members unless enabled by extension.
2017-03-25 00:38:16 +00:00
if (type.getBasicType() == glslang::EbtBlock) {
Fix bugs in missing Builtin decoration for some NV builtins for tessellation control shaders. Fix bug in member remapping.
2019-07-22 19:36:38 +00:00
if (filterMember(glslangMember)) {
memberDelta++;
memberRemapper[glslangMembers][i] = -1;
Fix #790: Don't emit NV-specific interface members unless enabled by extension.
2017-03-25 00:38:16 +00:00
continue;
Fix bugs in missing Builtin decoration for some NV builtins for tessellation control shaders. Fix bug in member remapping.
2019-07-22 19:36:38 +00:00
}
memberRemapper[glslangMembers][i] = i - memberDelta;
Fix #790: Don't emit NV-specific interface members unless enabled by extension.
2017-03-25 00:38:16 +00:00
}
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
// modify just this child's view of the qualifier
glslang::TQualifier memberQualifier = glslangMember.getQualifier();
InheritQualifiers(memberQualifier, qualifier);
Replace #422: Remove the redundant location setting in AST->SPIR-V. This was redundant in two ways: 1) it replicated algorithms owned in the front end, and 2) it sometimes left location information on both a block and its members.
2017-06-04 19:22:39 +00:00
// manually inherit location
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
if (! memberQualifier.hasLocation() && qualifier.hasLocation())
Replace #422: Remove the redundant location setting in AST->SPIR-V. This was redundant in two ways: 1) it replicated algorithms owned in the front end, and 2) it sometimes left location information on both a block and its members.
2017-06-04 19:22:39 +00:00
memberQualifier.layoutLocation = qualifier.layoutLocation;
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
// recurse
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
bool lastBufferBlockMember = qualifier.storage == glslang::EvqBuffer &&
i == (int)glslangMembers->size() - 1;
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
// Make forward pointers for any pointer members, and create a list of members to
// convert to spirv types after creating the struct.
Web: Remove/rationalize a set of *_EXTENSIONS, using GLSLANG_WEB. Focus was on the front end (not SPIR-V), minus the grammar. Reduces #ifdef count by around 320 and makes the web build 270K smaller, which is about 90% the target size. The grammar and scanner will be another step, as will the SPIR-V backend. This makes heavy use of methods #ifdef'd to return false as a global way of turning off code, relying on C++ DCE to do the rest.
2019-08-01 09:28:08 +00:00
if (glslangMember.isReference()) {
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
if (forwardPointers.find(glslangMember.getReferentType()) == forwardPointers.end()) {
deferredForwardPointers.push_back(std::make_pair(&glslangMember, memberQualifier));
}
spvMembers.push_back(
convertGlslangToSpvType(glslangMember, explicitLayout, memberQualifier, lastBufferBlockMember, true));
} else {
spvMembers.push_back(
convertGlslangToSpvType(glslangMember, explicitLayout, memberQualifier, lastBufferBlockMember, false));
}
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
}
}
// Make the SPIR-V type
spv::Id spvType = builder.makeStructType(spvMembers, type.getTypeName().c_str());
SPV: Tighten up number of struct-types declared based on decoration. Takes some pressure off of issue #304. Structures don't inherit locations and then explicitly decorate members with them, so removed this reason to have another instance of a structure type.
2016-09-01 23:05:23 +00:00
if (! HasNonLayoutQualifiers(type, qualifier))
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
structMap[explicitLayout][qualifier.layoutMatrix][glslangMembers] = spvType;
// Decorate it
decorateStructType(type, glslangMembers, explicitLayout, qualifier, spvType);
Build: fix warnings. Fixes #1653.
2019-01-16 07:55:37 +00:00
for (int i = 0; i < (int)deferredForwardPointers.size(); ++i) {
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
auto it = deferredForwardPointers[i];
convertGlslangToSpvType(*it.first, explicitLayout, it.second, false);
}
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
return spvType;
}
void TGlslangToSpvTraverser::decorateStructType(const glslang::TType& type,
const glslang::TTypeList* glslangMembers,
glslang::TLayoutPacking explicitLayout,
const glslang::TQualifier& qualifier,
spv::Id spvType)
{
// Name and decorate the non-hidden members
int offset = -1;
int locationOffset = 0; // for use within the members of this struct
for (int i = 0; i < (int)glslangMembers->size(); i++) {
glslang::TType& glslangMember = *(*glslangMembers)[i].type;
int member = i;
Fix #790: Don't emit NV-specific interface members unless enabled by extension.
2017-03-25 00:38:16 +00:00
if (type.getBasicType() == glslang::EbtBlock) {
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
member = memberRemapper[glslangMembers][i];
Fix #790: Don't emit NV-specific interface members unless enabled by extension.
2017-03-25 00:38:16 +00:00
if (filterMember(glslangMember))
continue;
}
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
// modify just this child's view of the qualifier
glslang::TQualifier memberQualifier = glslangMember.getQualifier();
InheritQualifiers(memberQualifier, qualifier);
// using -1 above to indicate a hidden member
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
if (member < 0)
continue;
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
builder.addMemberName(spvType, member, glslangMember.getFieldName().c_str());
builder.addMemberDecoration(spvType, member,
TranslateLayoutDecoration(glslangMember, memberQualifier.layoutMatrix));
builder.addMemberDecoration(spvType, member, TranslatePrecisionDecoration(glslangMember));
// Add interpolation and auxiliary storage decorations only to
// top-level members of Input and Output storage classes
if (type.getQualifier().storage == glslang::EvqVaryingIn ||
type.getQualifier().storage == glslang::EvqVaryingOut) {
if (type.getBasicType() == glslang::EbtBlock ||
glslangIntermediate->getSource() == glslang::EShSourceHlsl) {
builder.addMemberDecoration(spvType, member, TranslateInterpolationDecoration(memberQualifier));
builder.addMemberDecoration(spvType, member, TranslateAuxiliaryStorageDecoration(memberQualifier));
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
addMeshNVDecoration(spvType, member, memberQualifier);
#endif
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
}
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
}
builder.addMemberDecoration(spvType, member, TranslateInvariantDecoration(memberQualifier));
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
if (type.getBasicType() == glslang::EbtBlock &&
qualifier.storage == glslang::EvqBuffer) {
// Add memory decorations only to top-level members of shader storage block
std::vector<spv::Decoration> memory;
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
TranslateMemoryDecoration(memberQualifier, memory, glslangIntermediate->usingVulkanMemoryModel());
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
for (unsigned int i = 0; i < memory.size(); ++i)
builder.addMemberDecoration(spvType, member, memory[i]);
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
// Location assignment was already completed correctly by the front end,
// just track whether a member needs to be decorated.
// Ignore member locations if the container is an array, as that's
// ill-specified and decisions have been made to not allow this.
if (! type.isArray() && memberQualifier.hasLocation())
builder.addMemberDecoration(spvType, member, spv::DecorationLocation, memberQualifier.layoutLocation);
if (qualifier.hasLocation()) // track for upcoming inheritance
locationOffset += glslangIntermediate->computeTypeLocationSize(
glslangMember, glslangIntermediate->getStage());
// component, XFB, others
if (glslangMember.getQualifier().hasComponent())
builder.addMemberDecoration(spvType, member, spv::DecorationComponent,
glslangMember.getQualifier().layoutComponent);
if (glslangMember.getQualifier().hasXfbOffset())
builder.addMemberDecoration(spvType, member, spv::DecorationOffset,
glslangMember.getQualifier().layoutXfbOffset);
else if (explicitLayout != glslang::ElpNone) {
// figure out what to do with offset, which is accumulating
int nextOffset;
updateMemberOffset(type, glslangMember, offset, nextOffset, explicitLayout, memberQualifier.layoutMatrix);
if (offset >= 0)
builder.addMemberDecoration(spvType, member, spv::DecorationOffset, offset);
offset = nextOffset;
}
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
if (glslangMember.isMatrix() && explicitLayout != glslang::ElpNone)
builder.addMemberDecoration(spvType, member, spv::DecorationMatrixStride,
getMatrixStride(glslangMember, explicitLayout, memberQualifier.layoutMatrix));
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
// built-in variable decorations
spv::BuiltIn builtIn = TranslateBuiltInDecoration(glslangMember.getQualifier().builtIn, true);
if (builtIn != spv::BuiltInMax)
builder.addMemberDecoration(spvType, member, spv::DecorationBuiltIn, (int)builtIn);
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
// nonuniform
builder.addMemberDecoration(spvType, member, TranslateNonUniformDecoration(glslangMember.getQualifier()));
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
if (glslangIntermediate->getHlslFunctionality1() && memberQualifier.semanticName != nullptr) {
builder.addExtension("SPV_GOOGLE_hlsl_functionality1");
builder.addMemberDecoration(spvType, member, (spv::Decoration)spv::DecorationHlslSemanticGOOGLE,
memberQualifier.semanticName);
}
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
if (builtIn == spv::BuiltInLayer) {
// SPV_NV_viewport_array2 extension
if (glslangMember.getQualifier().layoutViewportRelative){
builder.addMemberDecoration(spvType, member, (spv::Decoration)spv::DecorationViewportRelativeNV);
builder.addCapability(spv::CapabilityShaderViewportMaskNV);
builder.addExtension(spv::E_SPV_NV_viewport_array2);
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
}
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
if (glslangMember.getQualifier().layoutSecondaryViewportRelativeOffset != -2048){
builder.addMemberDecoration(spvType, member,
(spv::Decoration)spv::DecorationSecondaryViewportRelativeNV,
glslangMember.getQualifier().layoutSecondaryViewportRelativeOffset);
builder.addCapability(spv::CapabilityShaderStereoViewNV);
builder.addExtension(spv::E_SPV_NV_stereo_view_rendering);
Implement NVX_multiview_per_view_attributes
2017-02-14 22:52:34 +00:00
}
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
}
if (glslangMember.getQualifier().layoutPassthrough) {
builder.addMemberDecoration(spvType, member, (spv::Decoration)spv::DecorationPassthroughNV);
builder.addCapability(spv::CapabilityGeometryShaderPassthroughNV);
builder.addExtension(spv::E_SPV_NV_geometry_shader_passthrough);
}
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
#endif
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
}
// Decorate the structure
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
builder.addDecoration(spvType, TranslateLayoutDecoration(type, qualifier.layoutMatrix));
builder.addDecoration(spvType, TranslateBlockDecoration(type, glslangIntermediate->usingStorageBuffer()));
SPV: Non-functional: break up big function into three smaller ones. Also, improve variable names. This grew to be overly hard to understand.
2016-07-01 03:18:02 +00:00
}
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
// Turn the expression forming the array size into an id.
// This is not quite trivial, because of specialization constants.
// Sometimes, a raw constant is turned into an Id, and sometimes
// a specialization constant expression is.
spv::Id TGlslangToSpvTraverser::makeArraySizeId(const glslang::TArraySizes& arraySizes, int dim)
{
// First, see if this is sized with a node, meaning a specialization constant:
glslang::TIntermTyped* specNode = arraySizes.getDimNode(dim);
if (specNode != nullptr) {
builder.clearAccessChain();
specNode->traverse(this);
return accessChainLoad(specNode->getAsTyped()->getType());
}
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
// Otherwise, need a compile-time (front end) size, get it:
int size = arraySizes.getDimSize(dim);
assert(size > 0);
return builder.makeUintConstant(size);
}
SPV: Handle GLSL bool loads from a uniform buffer as a conversion from int -> bool. SPIR-V bool is abstract; it has no bit pattern for storage with transparent memory. OpenGL's convention is a bool in a uniform buffer is 32-bit int with non-0 being 'true'.
2016-02-09 04:38:15 +00:00
// Wrap the builder's accessChainLoad to:
// - localize handling of RelaxedPrecision
// - use the SPIR-V inferred type instead of another conversion of the glslang type
// (avoids unnecessary work and possible type punning for structures)
// - do conversion of concrete to abstract type
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
spv::Id TGlslangToSpvTraverser::accessChainLoad(const glslang::TType& type)
{
SPV: Handle GLSL bool loads from a uniform buffer as a conversion from int -> bool. SPIR-V bool is abstract; it has no bit pattern for storage with transparent memory. OpenGL's convention is a bool in a uniform buffer is 32-bit int with non-0 being 'true'.
2016-02-09 04:38:15 +00:00
spv::Id nominalTypeId = builder.accessChainGetInferredType();
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spv::Builder::AccessChain::CoherentFlags coherentFlags = builder.getAccessChain().coherentFlags;
coherentFlags |= TranslateCoherent(type);
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
unsigned int alignment = builder.getAccessChain().alignment;
Move getBufferReferenceAlignment to be a method of TType This is a better place for it logically, since it is not specific to glsl->spirv translation. And in a future change I want to use it outside of glslangtospv.
2019-03-06 19:34:10 +00:00
alignment |= type.getBufferReferenceAlignment();
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
spv::Id loadedId = builder.accessChainLoad(TranslatePrecisionDecoration(type),
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
TranslateNonUniformDecoration(type.getQualifier()),
nominalTypeId,
spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & ~spv::MemoryAccessMakePointerAvailableKHRMask),
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
TranslateMemoryScope(coherentFlags),
alignment);
SPV: Handle GLSL bool loads from a uniform buffer as a conversion from int -> bool. SPIR-V bool is abstract; it has no bit pattern for storage with transparent memory. OpenGL's convention is a bool in a uniform buffer is 32-bit int with non-0 being 'true'.
2016-02-09 04:38:15 +00:00
// Need to convert to abstract types when necessary
SPV: Fix an issue caused by bool-to-uint32 conversion. This issue is related with the commit 103bef9d74d768f0690ed53f52681baead384d1e.
2016-02-23 09:51:09 +00:00
if (type.getBasicType() == glslang::EbtBool) {
if (builder.isScalarType(nominalTypeId)) {
// Conversion for bool
spv::Id boolType = builder.makeBoolType();
if (nominalTypeId != boolType)
loadedId = builder.createBinOp(spv::OpINotEqual, boolType, loadedId, builder.makeUintConstant(0));
} else if (builder.isVectorType(nominalTypeId)) {
// Conversion for bvec
int vecSize = builder.getNumTypeComponents(nominalTypeId);
spv::Id bvecType = builder.makeVectorType(builder.makeBoolType(), vecSize);
if (nominalTypeId != bvecType)
loadedId = builder.createBinOp(spv::OpINotEqual, bvecType, loadedId, makeSmearedConstant(builder.makeUintConstant(0), vecSize));
}
}
SPV: Handle GLSL bool loads from a uniform buffer as a conversion from int -> bool. SPIR-V bool is abstract; it has no bit pattern for storage with transparent memory. OpenGL's convention is a bool in a uniform buffer is 32-bit int with non-0 being 'true'.
2016-02-09 04:38:15 +00:00
return loadedId;
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
}
SPV: Fix an issue caused by bool-to-uint32 conversion. This issue is related with the commit 103bef9d74d768f0690ed53f52681baead384d1e.
2016-02-23 09:51:09 +00:00
// Wrap the builder's accessChainStore to:
// - do conversion of concrete to abstract type
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
//
// Implicitly uses the existing builder.accessChain as the storage target.
SPV: Fix an issue caused by bool-to-uint32 conversion. This issue is related with the commit 103bef9d74d768f0690ed53f52681baead384d1e.
2016-02-23 09:51:09 +00:00
void TGlslangToSpvTraverser::accessChainStore(const glslang::TType& type, spv::Id rvalue)
{
// Need to convert to abstract types when necessary
if (type.getBasicType() == glslang::EbtBool) {
spv::Id nominalTypeId = builder.accessChainGetInferredType();
if (builder.isScalarType(nominalTypeId)) {
// Conversion for bool
spv::Id boolType = builder.makeBoolType();
Fix C++ portability, non-deterministic order of argument evaluation.
2017-05-20 05:29:50 +00:00
if (nominalTypeId != boolType) {
// keep these outside arguments, for determinant order-of-evaluation
spv::Id one = builder.makeUintConstant(1);
spv::Id zero = builder.makeUintConstant(0);
rvalue = builder.createTriOp(spv::OpSelect, nominalTypeId, rvalue, one, zero);
} else if (builder.getTypeId(rvalue) != boolType)
Fix #857: Convert uniform int to local bool for struct alias assignment. This was done for one direction, but not both directions, so this commit picks up the other direction.
2017-05-20 05:00:13 +00:00
rvalue = builder.createBinOp(spv::OpINotEqual, boolType, rvalue, builder.makeUintConstant(0));
SPV: Fix an issue caused by bool-to-uint32 conversion. This issue is related with the commit 103bef9d74d768f0690ed53f52681baead384d1e.
2016-02-23 09:51:09 +00:00
} else if (builder.isVectorType(nominalTypeId)) {
// Conversion for bvec
int vecSize = builder.getNumTypeComponents(nominalTypeId);
spv::Id bvecType = builder.makeVectorType(builder.makeBoolType(), vecSize);
Fix C++ portability, non-deterministic order of argument evaluation.
2017-05-20 05:29:50 +00:00
if (nominalTypeId != bvecType) {
// keep these outside arguments, for determinant order-of-evaluation
More non-determinism fixed.
2017-05-20 05:44:51 +00:00
spv::Id one = makeSmearedConstant(builder.makeUintConstant(1), vecSize);
spv::Id zero = makeSmearedConstant(builder.makeUintConstant(0), vecSize);
rvalue = builder.createTriOp(spv::OpSelect, nominalTypeId, rvalue, one, zero);
Fix C++ portability, non-deterministic order of argument evaluation.
2017-05-20 05:29:50 +00:00
} else if (builder.getTypeId(rvalue) != bvecType)
Fix #857: Convert uniform int to local bool for struct alias assignment. This was done for one direction, but not both directions, so this commit picks up the other direction.
2017-05-20 05:00:13 +00:00
rvalue = builder.createBinOp(spv::OpINotEqual, bvecType, rvalue,
makeSmearedConstant(builder.makeUintConstant(0), vecSize));
SPV: Fix an issue caused by bool-to-uint32 conversion. This issue is related with the commit 103bef9d74d768f0690ed53f52681baead384d1e.
2016-02-23 09:51:09 +00:00
}
}
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spv::Builder::AccessChain::CoherentFlags coherentFlags = builder.getAccessChain().coherentFlags;
coherentFlags |= TranslateCoherent(type);
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
unsigned int alignment = builder.getAccessChain().alignment;
Move getBufferReferenceAlignment to be a method of TType This is a better place for it logically, since it is not specific to glsl->spirv translation. And in a future change I want to use it outside of glslangtospv.
2019-03-06 19:34:10 +00:00
alignment |= type.getBufferReferenceAlignment();
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
builder.accessChainStore(rvalue,
spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & ~spv::MemoryAccessMakePointerVisibleKHRMask),
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
TranslateMemoryScope(coherentFlags), alignment);
SPV: Fix an issue caused by bool-to-uint32 conversion. This issue is related with the commit 103bef9d74d768f0690ed53f52681baead384d1e.
2016-02-23 09:51:09 +00:00
}
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
// For storing when types match at the glslang level, but not might match at the
// SPIR-V level.
//
// This especially happens when a single glslang type expands to multiple
HLSL: Do structure conversion for return type struct-punning on non-entry-point functions.
2016-10-06 18:59:51 +00:00
// SPIR-V types, like a struct that is used in a member-undecorated way as well
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
// as in a member-decorated way.
//
// NOTE: This function can handle any store request; if it's not special it
// simplifies to a simple OpStore.
//
// Implicitly uses the existing builder.accessChain as the storage target.
void TGlslangToSpvTraverser::multiTypeStore(const glslang::TType& type, spv::Id rValue)
{
SPV: Fix issue #506: generalize struct deep copy to include arrays.
2016-09-11 18:33:43 +00:00
// we only do the complex path here if it's an aggregate
if (! type.isStruct() && ! type.isArray()) {
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
accessChainStore(type, rValue);
return;
}
SPV: Fix issue #506: generalize struct deep copy to include arrays.
2016-09-11 18:33:43 +00:00
// and, it has to be a case of type aliasing
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
spv::Id rType = builder.getTypeId(rValue);
spv::Id lValue = builder.accessChainGetLValue();
spv::Id lType = builder.getContainedTypeId(builder.getTypeId(lValue));
if (lType == rType) {
accessChainStore(type, rValue);
return;
}
SPV: Fix issue #506: generalize struct deep copy to include arrays.
2016-09-11 18:33:43 +00:00
// Recursively (as needed) copy an aggregate type to a different aggregate type,
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
// where the two types were the same type in GLSL. This requires member
// by member copy, recursively.
SPV 1.4: Add support for OpCopyLogical, careful of Boolean differences.
2019-01-15 14:48:27 +00:00
// SPIR-V 1.4 added an instruction to do help do this.
if (glslangIntermediate->getSpv().spv >= glslang::EShTargetSpv_1_4) {
// However, bool in uniform space is changed to int, so
// OpCopyLogical does not work for that.
// TODO: It would be more robust to do a full recursive verification of the types satisfying SPIR-V rules.
bool rBool = builder.containsType(builder.getTypeId(rValue), spv::OpTypeBool, 0);
bool lBool = builder.containsType(lType, spv::OpTypeBool, 0);
if (lBool == rBool) {
spv::Id logicalCopy = builder.createUnaryOp(spv::OpCopyLogical, lType, rValue);
accessChainStore(type, logicalCopy);
return;
}
}
SPV: Fix issue #506: generalize struct deep copy to include arrays.
2016-09-11 18:33:43 +00:00
// If an array, copy element by element.
if (type.isArray()) {
glslang::TType glslangElementType(type, 0);
spv::Id elementRType = builder.getContainedTypeId(rType);
for (int index = 0; index < type.getOuterArraySize(); ++index) {
// get the source member
spv::Id elementRValue = builder.createCompositeExtract(rValue, elementRType, index);
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
SPV: Fix issue #506: generalize struct deep copy to include arrays.
2016-09-11 18:33:43 +00:00
// set up the target storage
builder.clearAccessChain();
builder.setAccessChainLValue(lValue);
Move getBufferReferenceAlignment to be a method of TType This is a better place for it logically, since it is not specific to glsl->spirv translation. And in a future change I want to use it outside of glslangtospv.
2019-03-06 19:34:10 +00:00
builder.accessChainPush(builder.makeIntConstant(index), TranslateCoherent(type), type.getBufferReferenceAlignment());
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
SPV: Fix issue #506: generalize struct deep copy to include arrays.
2016-09-11 18:33:43 +00:00
// store the member
multiTypeStore(glslangElementType, elementRValue);
}
} else {
assert(type.isStruct());
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
SPV: Fix issue #506: generalize struct deep copy to include arrays.
2016-09-11 18:33:43 +00:00
// loop over structure members
const glslang::TTypeList& members = *type.getStruct();
for (int m = 0; m < (int)members.size(); ++m) {
const glslang::TType& glslangMemberType = *members[m].type;
// get the source member
spv::Id memberRType = builder.getContainedTypeId(rType, m);
spv::Id memberRValue = builder.createCompositeExtract(rValue, memberRType, m);
// set up the target storage
builder.clearAccessChain();
builder.setAccessChainLValue(lValue);
Move getBufferReferenceAlignment to be a method of TType This is a better place for it logically, since it is not specific to glsl->spirv translation. And in a future change I want to use it outside of glslangtospv.
2019-03-06 19:34:10 +00:00
builder.accessChainPush(builder.makeIntConstant(m), TranslateCoherent(type), type.getBufferReferenceAlignment());
SPV: Fix issue #506: generalize struct deep copy to include arrays.
2016-09-11 18:33:43 +00:00
// store the member
multiTypeStore(glslangMemberType, memberRValue);
}
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
}
}
SPV: Nested layouts: Recursively send around the top-level std140 or std430 packing. This also distinguishes between the same struct included in multiple parent packings.
2015-12-19 20:57:10 +00:00
// Decide whether or not this type should be
// decorated with offsets and strides, and if so
// whether std140 or std430 rules should be applied.
glslang::TLayoutPacking TGlslangToSpvTraverser::getExplicitLayout(const glslang::TType& type) const
SPV: Only decorate array and matrix strides for transparent types requiring explicit layouts.
2015-09-09 23:51:38 +00:00
{
SPV: Nested layouts: Recursively send around the top-level std140 or std430 packing. This also distinguishes between the same struct included in multiple parent packings.
2015-12-19 20:57:10 +00:00
// has to be a block
if (type.getBasicType() != glslang::EbtBlock)
return glslang::ElpNone;
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
// has to be a uniform or buffer block or task in/out blocks
SPV: Nested layouts: Recursively send around the top-level std140 or std430 packing. This also distinguishes between the same struct included in multiple parent packings.
2015-12-19 20:57:10 +00:00
if (type.getQualifier().storage != glslang::EvqUniform &&
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
type.getQualifier().storage != glslang::EvqBuffer &&
!type.getQualifier().isTaskMemory())
SPV: Nested layouts: Recursively send around the top-level std140 or std430 packing. This also distinguishes between the same struct included in multiple parent packings.
2015-12-19 20:57:10 +00:00
return glslang::ElpNone;
// return the layout to use
switch (type.getQualifier().layoutPacking) {
case glslang::ElpStd140:
case glslang::ElpStd430:
Implement GL_EXT_scalar_block_layout
2018-11-14 15:30:53 +00:00
case glslang::ElpScalar:
SPV: Nested layouts: Recursively send around the top-level std140 or std430 packing. This also distinguishes between the same struct included in multiple parent packings.
2015-12-19 20:57:10 +00:00
return type.getQualifier().layoutPacking;
default:
return glslang::ElpNone;
}
SPV: Only decorate array and matrix strides for transparent types requiring explicit layouts.
2015-09-09 23:51:38 +00:00
}
SPIRV: Decorate matrices and arrays with their strides
2015-09-05 16:50:58 +00:00
// Given an array type, returns the integer stride required for that array
SPV: recursively propagate row/col majorness through nested structures. This includes doing structure uniqueness modulo majorness, for shared nested structures.
2015-12-20 18:29:16 +00:00
int TGlslangToSpvTraverser::getArrayStride(const glslang::TType& arrayType, glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout)
SPIRV: Decorate matrices and arrays with their strides
2015-09-05 16:50:58 +00:00
{
int size;
SPV: Fix array strides by explicitly computing them in the getBaseAlignment() algorithm.
2015-12-30 00:11:44 +00:00
int stride;
Implement GL_EXT_scalar_block_layout
2018-11-14 15:30:53 +00:00
glslangIntermediate->getMemberAlignment(arrayType, size, stride, explicitLayout, matrixLayout == glslang::ElmRowMajor);
SPV: Distinguish between std140 and std430 layouts, and correctly use alignments as the base for strides.
2015-12-07 02:17:49 +00:00
return stride;
SPIRV: Decorate matrices and arrays with their strides
2015-09-05 16:50:58 +00:00
}
SPV: Fix array strides by explicitly computing them in the getBaseAlignment() algorithm.
2015-12-30 00:11:44 +00:00
// Given a matrix type, or array (of array) of matrixes type, returns the integer stride required for that matrix
SPIRV: Decorate matrices and arrays with their strides
2015-09-05 16:50:58 +00:00
// when used as a member of an interface block
SPV: recursively propagate row/col majorness through nested structures. This includes doing structure uniqueness modulo majorness, for shared nested structures.
2015-12-20 18:29:16 +00:00
int TGlslangToSpvTraverser::getMatrixStride(const glslang::TType& matrixType, glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout)
SPIRV: Decorate matrices and arrays with their strides
2015-09-05 16:50:58 +00:00
{
SPV: Fix array strides by explicitly computing them in the getBaseAlignment() algorithm.
2015-12-30 00:11:44 +00:00
glslang::TType elementType;
elementType.shallowCopy(matrixType);
elementType.clearArraySizes();
SPIRV: Decorate matrices and arrays with their strides
2015-09-05 16:50:58 +00:00
int size;
SPV: Fix array strides by explicitly computing them in the getBaseAlignment() algorithm.
2015-12-30 00:11:44 +00:00
int stride;
Implement GL_EXT_scalar_block_layout
2018-11-14 15:30:53 +00:00
glslangIntermediate->getMemberAlignment(elementType, size, stride, explicitLayout, matrixLayout == glslang::ElmRowMajor);
SPV: Fix array strides by explicitly computing them in the getBaseAlignment() algorithm.
2015-12-30 00:11:44 +00:00
return stride;
SPIRV: Decorate matrices and arrays with their strides
2015-09-05 16:50:58 +00:00
}
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
// Given a member type of a struct, realign the current offset for it, and compute
// the next (not yet aligned) offset for the next member, which will get aligned
// on the next call.
// 'currentOffset' should be passed in already initialized, ready to modify, and reflecting
// the migration of data from nextOffset -> currentOffset. It should be -1 on the first call.
// -1 means a non-forced member offset (no decoration needed).
Address part A of #982: $Global will use std140 instead of HLSL offsets. From comment about this: Adjust alignment for HLSL rules TODO: make this consistent in early phases of code: adjusting this late means inconsistencies with earlier code, which for reflection is an issue. Until reflection is brought in sync with these adjustments, don't apply to $Global, which is the most likely to rely on reflection, and least likely to rely implicit layouts.
2017-07-13 17:39:16 +00:00
void TGlslangToSpvTraverser::updateMemberOffset(const glslang::TType& structType, const glslang::TType& memberType, int& currentOffset, int& nextOffset,
SPV: recursively propagate row/col majorness through nested structures. This includes doing structure uniqueness modulo majorness, for shared nested structures.
2015-12-20 18:29:16 +00:00
glslang::TLayoutPacking explicitLayout, glslang::TLayoutMatrix matrixLayout)
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
{
// this will get a positive value when deemed necessary
nextOffset = -1;
// override anything in currentOffset with user-set offset
if (memberType.getQualifier().hasOffset())
currentOffset = memberType.getQualifier().layoutOffset;
// It could be that current linker usage in glslang updated all the layoutOffset,
// in which case the following code does not matter. But, that's not quite right
// once cross-compilation unit GLSL validation is done, as the original user
// settings are needed in layoutOffset, and then the following will come into play.
SPV: Nested layouts: Recursively send around the top-level std140 or std430 packing. This also distinguishes between the same struct included in multiple parent packings.
2015-12-19 20:57:10 +00:00
if (explicitLayout == glslang::ElpNone) {
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
if (! memberType.getQualifier().hasOffset())
currentOffset = -1;
return;
}
SPV: Nested layouts: Recursively send around the top-level std140 or std430 packing. This also distinguishes between the same struct included in multiple parent packings.
2015-12-19 20:57:10 +00:00
// Getting this far means we need explicit offsets
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
if (currentOffset < 0)
currentOffset = 0;
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
// Now, currentOffset is valid (either 0, or from a previous nextOffset),
// but possibly not yet correctly aligned.
int memberSize;
SPV: Fix array strides by explicitly computing them in the getBaseAlignment() algorithm.
2015-12-30 00:11:44 +00:00
int dummyStride;
Implement GL_EXT_scalar_block_layout
2018-11-14 15:30:53 +00:00
int memberAlignment = glslangIntermediate->getMemberAlignment(memberType, memberSize, dummyStride, explicitLayout, matrixLayout == glslang::ElmRowMajor);
SPV: Fix #807: use --hlsl-offsets to allow hlsl-style offsets in a buffer. Corresponds to the EShMsgHlslOffsets flag in messages. Works for both GLSL and HLSL.
2017-04-05 23:38:20 +00:00
// Adjust alignment for HLSL rules
Address part A of #982: $Global will use std140 instead of HLSL offsets. From comment about this: Adjust alignment for HLSL rules TODO: make this consistent in early phases of code: adjusting this late means inconsistencies with earlier code, which for reflection is an issue. Until reflection is brought in sync with these adjustments, don't apply to $Global, which is the most likely to rely on reflection, and least likely to rely implicit layouts.
2017-07-13 17:39:16 +00:00
// TODO: make this consistent in early phases of code:
// adjusting this late means inconsistencies with earlier code, which for reflection is an issue
// Until reflection is brought in sync with these adjustments, don't apply to $Global,
// which is the most likely to rely on reflection, and least likely to rely implicit layouts
Non-functional: Rationalize some existing use of SPIRV-Tools.
2018-08-22 23:12:46 +00:00
if (glslangIntermediate->usingHlslOffsets() &&
Address part A of #982: $Global will use std140 instead of HLSL offsets. From comment about this: Adjust alignment for HLSL rules TODO: make this consistent in early phases of code: adjusting this late means inconsistencies with earlier code, which for reflection is an issue. Until reflection is brought in sync with these adjustments, don't apply to $Global, which is the most likely to rely on reflection, and least likely to rely implicit layouts.
2017-07-13 17:39:16 +00:00
! memberType.isArray() && memberType.isVector() && structType.getTypeName().compare("$Global") != 0) {
SPV: Fix #807: use --hlsl-offsets to allow hlsl-style offsets in a buffer. Corresponds to the EShMsgHlslOffsets flag in messages. Works for both GLSL and HLSL.
2017-04-05 23:38:20 +00:00
int dummySize;
int componentAlignment = glslangIntermediate->getBaseAlignmentScalar(memberType, dummySize);
if (componentAlignment <= 4)
memberAlignment = componentAlignment;
}
// Bump up to member alignment
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
glslang::RoundToPow2(currentOffset, memberAlignment);
SPV: Fix #807: use --hlsl-offsets to allow hlsl-style offsets in a buffer. Corresponds to the EShMsgHlslOffsets flag in messages. Works for both GLSL and HLSL.
2017-04-05 23:38:20 +00:00
// Bump up to vec4 if there is a bad straddle
Implement GL_EXT_scalar_block_layout
2018-11-14 15:30:53 +00:00
if (explicitLayout != glslang::ElpScalar && glslangIntermediate->improperStraddle(memberType, memberSize, currentOffset))
SPV: Fix #807: use --hlsl-offsets to allow hlsl-style offsets in a buffer. Corresponds to the EShMsgHlslOffsets flag in messages. Works for both GLSL and HLSL.
2017-04-05 23:38:20 +00:00
glslang::RoundToPow2(currentOffset, 16);
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
nextOffset = currentOffset + memberSize;
}
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
void TGlslangToSpvTraverser::declareUseOfStructMember(const glslang::TTypeList& members, int glslangMember)
SPV: Don't add clip/cull distance capabilities unless used. These capabalities were added on declaration of the members, but that is considered too aggressive, as those members are automatically declared in some shaders that don't use them. Now, actual access is needed to make the capabalities be declared.
2016-05-17 01:22:05 +00:00
{
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
const glslang::TBuiltInVariable glslangBuiltIn = members[glslangMember].type->getQualifier().builtIn;
switch (glslangBuiltIn)
{
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
case glslang::EbvPointSize:
#ifndef GLSLANG_WEB
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
case glslang::EbvClipDistance:
case glslang::EbvCullDistance:
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
case glslang::EbvViewportMaskNV:
case glslang::EbvSecondaryPositionNV:
case glslang::EbvSecondaryViewportMaskNV:
Implement NVX_multiview_per_view_attributes
2017-02-14 22:52:34 +00:00
case glslang::EbvPositionPerViewNV:
case glslang::EbvViewportMaskPerViewNV:
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
case glslang::EbvTaskCountNV:
case glslang::EbvPrimitiveCountNV:
case glslang::EbvPrimitiveIndicesNV:
case glslang::EbvClipDistancePerViewNV:
case glslang::EbvCullDistancePerViewNV:
case glslang::EbvLayerPerViewNV:
case glslang::EbvMeshViewCountNV:
case glslang::EbvMeshViewIndicesNV:
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
#endif
Defer capability decl for ClipDistance, CullDistance, PointSize until actual use The compiler will mark struct members with those builtins, but won't declare the capability until that member is accessed by some executable instruction. Test changes: - spv.430.vert: was missing ClipDistance capability. - spv.precise.tese: remove TessellationPointSize capability.
2016-06-08 13:11:40 +00:00
// Generate the associated capability. Delegate to TranslateBuiltInDecoration.
// Alternately, we could just call this for any glslang built-in, since the
// capability already guards against duplicates.
TranslateBuiltInDecoration(glslangBuiltIn, false);
break;
default:
// Capabilities were already generated when the struct was declared.
break;
}
SPV: Don't add clip/cull distance capabilities unless used. These capabalities were added on declaration of the members, but that is considered too aggressive, as those members are automatically declared in some shaders that don't use them. Now, actual access is needed to make the capabalities be declared.
2016-05-17 01:22:05 +00:00
}
Non-functional: Sweep through the stack for consistent with "main" and entry point. Partially addresses issue #513.
2016-09-19 22:01:41 +00:00
bool TGlslangToSpvTraverser::isShaderEntryPoint(const glslang::TIntermAggregate* node)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
Track separate entry-point names and mangled names... ... and use each in the correct way at consumption sites. This completes issue #513.
2016-09-20 00:09:30 +00:00
return node->getName().compare(glslangIntermediate->getEntryPointMangledName().c_str()) == 0;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
SPV: rationalize parameter handling for "original" and "writable" parameters.
2017-09-10 21:21:05 +00:00
// Does parameter need a place to keep writes, separate from the original?
HLSL: Fix #1163: treat buffers as references when calling functions. This continues to prevent writing output buffers (out from a function), but fixes the problem where the copy-in/out was not getting done. Making everything work will require knowing both in/out-ness and bufferness, but these are currently mutually exclusive, because both are storage qualifiers.
2017-12-04 09:48:10 +00:00
// Assumes called after originalParam(), which filters out block/buffer/opaque-based
// qualifiers such that we should have only in/out/inout/constreadonly here.
Fix #944: Convert argument type to match formal parameter type.
2018-05-04 17:43:03 +00:00
bool TGlslangToSpvTraverser::writableParam(glslang::TStorageQualifier qualifier) const
SPV: rationalize parameter handling for "original" and "writable" parameters.
2017-09-10 21:21:05 +00:00
{
HLSL: Fix #1163: treat buffers as references when calling functions. This continues to prevent writing output buffers (out from a function), but fixes the problem where the copy-in/out was not getting done. Making everything work will require knowing both in/out-ness and bufferness, but these are currently mutually exclusive, because both are storage qualifiers.
2017-12-04 09:48:10 +00:00
assert(qualifier == glslang::EvqIn ||
qualifier == glslang::EvqOut ||
qualifier == glslang::EvqInOut ||
qualifier == glslang::EvqConstReadOnly);
SPV: rationalize parameter handling for "original" and "writable" parameters.
2017-09-10 21:21:05 +00:00
return qualifier != glslang::EvqConstReadOnly;
}
// Is parameter pass-by-original?
bool TGlslangToSpvTraverser::originalParam(glslang::TStorageQualifier qualifier, const glslang::TType& paramType,
bool implicitThisParam)
{
if (implicitThisParam) // implicit this
return true;
if (glslangIntermediate->getSource() == glslang::EShSourceHlsl)
HLSL: Fix #1163: treat buffers as references when calling functions. This continues to prevent writing output buffers (out from a function), but fixes the problem where the copy-in/out was not getting done. Making everything work will require knowing both in/out-ness and bufferness, but these are currently mutually exclusive, because both are storage qualifiers.
2017-12-04 09:48:10 +00:00
return paramType.getBasicType() == glslang::EbtBlock;
SPV: rationalize parameter handling for "original" and "writable" parameters.
2017-09-10 21:21:05 +00:00
return paramType.containsOpaque() || // sampler, etc.
(paramType.getBasicType() == glslang::EbtBlock && qualifier == glslang::EvqBuffer); // SSBO
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
// Make all the functions, skeletally, without actually visiting their bodies.
void TGlslangToSpvTraverser::makeFunctions(const glslang::TIntermSequence& glslFunctions)
{
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
const auto getParamDecorations = [&](std::vector<spv::Decoration>& decorations, const glslang::TType& type, bool useVulkanMemoryModel) {
SPV: Non-functional: support lists of decorations per parameter.
2017-07-18 08:35:46 +00:00
spv::Decoration paramPrecision = TranslatePrecisionDecoration(type);
if (paramPrecision != spv::NoPrecision)
decorations.push_back(paramPrecision);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
TranslateMemoryDecoration(type.getQualifier(), decorations, useVulkanMemoryModel);
Web: Remove/rationalize a set of *_EXTENSIONS, using GLSLANG_WEB. Focus was on the front end (not SPIR-V), minus the grammar. Reduces #ifdef count by around 320 and makes the web build 270K smaller, which is about 90% the target size. The grammar and scanner will be another step, as will the SPIR-V backend. This makes heavy use of methods #ifdef'd to return false as a global way of turning off code, relying on C++ DCE to do the rest.
2019-08-01 09:28:08 +00:00
if (type.isReference()) {
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
// Original and non-writable params pass the pointer directly and
// use restrict/aliased, others are stored to a pointer in Function
// memory and use RestrictPointer/AliasedPointer.
if (originalParam(type.getQualifier().storage, type, false) ||
!writableParam(type.getQualifier().storage)) {
decorations.push_back(type.getQualifier().restrict ? spv::DecorationRestrict : spv::DecorationAliased);
} else {
decorations.push_back(type.getQualifier().restrict ? spv::DecorationRestrictPointerEXT : spv::DecorationAliasedPointerEXT);
}
}
SPV: Non-functional: support lists of decorations per parameter.
2017-07-18 08:35:46 +00:00
};
Final round for line endings.
2015-06-26 22:58:36 +00:00
for (int f = 0; f < (int)glslFunctions.size(); ++f) {
glslang::TIntermAggregate* glslFunction = glslFunctions[f]->getAsAggregate();
Non-functional: Sweep through the stack for consistent with "main" and entry point. Partially addresses issue #513.
2016-09-19 22:01:41 +00:00
if (! glslFunction || glslFunction->getOp() != glslang::EOpFunction || isShaderEntryPoint(glslFunction))
Final round for line endings.
2015-06-26 22:58:36 +00:00
continue;
// We're on a user function. Set up the basic interface for the function now,
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
// so that it's available to call. Translating the body will happen later.
Final round for line endings.
2015-06-26 22:58:36 +00:00
//
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
// Typically (except for a "const in" parameter), an address will be passed to the
Final round for line endings.
2015-06-26 22:58:36 +00:00
// function. What it is an address of varies:
//
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
// - "in" parameters not marked as "const" can be written to without modifying the calling
// argument so that write needs to be to a copy, hence the address of a copy works.
Final round for line endings.
2015-06-26 22:58:36 +00:00
//
// - "const in" parameters can just be the r-value, as no writes need occur.
//
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
// - "out" and "inout" arguments can't be done as pointers to the calling argument, because
// GLSL has copy-in/copy-out semantics. They can be handled though with a pointer to a copy.
Final round for line endings.
2015-06-26 22:58:36 +00:00
std::vector<spv::Id> paramTypes;
SPV: Non-functional: support lists of decorations per parameter.
2017-07-18 08:35:46 +00:00
std::vector<std::vector<spv::Decoration>> paramDecorations; // list of decorations per parameter
Final round for line endings.
2015-06-26 22:58:36 +00:00
glslang::TIntermSequence& parameters = glslFunction->getSequence()[0]->getAsAggregate()->getSequence();
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#ifdef ENABLE_HLSL
SPV: Non-functional: support lists of decorations per parameter.
2017-07-18 08:35:46 +00:00
bool implicitThis = (int)parameters.size() > 0 && parameters[0]->getAsSymbolNode()->getName() ==
glslangIntermediate->implicitThisName;
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#else
bool implicitThis = false;
#endif
HLSL: non-static member functions: track and find active anonymous 'this' scopes and members. Thanks to @steve-lunarg for his input and discussions on handling member functions.
2017-03-22 05:56:40 +00:00
SPV: Non-functional: support lists of decorations per parameter.
2017-07-18 08:35:46 +00:00
paramDecorations.resize(parameters.size());
Final round for line endings.
2015-06-26 22:58:36 +00:00
for (int p = 0; p < (int)parameters.size(); ++p) {
const glslang::TType& paramType = parameters[p]->getAsTyped()->getType();
spv::Id typeId = convertGlslangToSpvType(paramType);
SPV: rationalize parameter handling for "original" and "writable" parameters.
2017-09-10 21:21:05 +00:00
if (originalParam(paramType.getQualifier().storage, paramType, implicitThis && p == 0))
SPV: Emit extension for using StorageClassStorageBuffer.
2017-05-05 11:09:58 +00:00
typeId = builder.makePointer(TranslateStorageClass(paramType), typeId);
SPV: rationalize parameter handling for "original" and "writable" parameters.
2017-09-10 21:21:05 +00:00
else if (writableParam(paramType.getQualifier().storage))
Final round for line endings.
2015-06-26 22:58:36 +00:00
typeId = builder.makePointer(spv::StorageClassFunction, typeId);
else
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
rValueParameters.insert(parameters[p]->getAsSymbolNode()->getId());
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
getParamDecorations(paramDecorations[p], paramType, glslangIntermediate->usingVulkanMemoryModel());
Final round for line endings.
2015-06-26 22:58:36 +00:00
paramTypes.push_back(typeId);
}
spv::Block* functionBlock;
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
spv::Function *function = builder.makeFunctionEntry(TranslatePrecisionDecoration(glslFunction->getType()),
convertGlslangToSpvType(glslFunction->getType()),
SPV: Non-functional: support lists of decorations per parameter.
2017-07-18 08:35:46 +00:00
glslFunction->getName().c_str(), paramTypes,
paramDecorations, &functionBlock);
HLSL: non-static member functions: track and find active anonymous 'this' scopes and members. Thanks to @steve-lunarg for his input and discussions on handling member functions.
2017-03-22 05:56:40 +00:00
if (implicitThis)
function->setImplicitThis();
Final round for line endings.
2015-06-26 22:58:36 +00:00
// Track function to emit/call later
functionMap[glslFunction->getName().c_str()] = function;
// Set the parameter id's
for (int p = 0; p < (int)parameters.size(); ++p) {
symbolValues[parameters[p]->getAsSymbolNode()->getId()] = function->getParamId(p);
// give a name too
builder.addName(function->getParamId(p), parameters[p]->getAsSymbolNode()->getName().c_str());
Add Float16/Int8/Int16 capabilities for private variables and function parameters
2019-02-18 04:49:28 +00:00
const glslang::TType& paramType = parameters[p]->getAsTyped()->getType();
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
if (paramType.contains8BitInt())
Add Float16/Int8/Int16 capabilities for private variables and function parameters
2019-02-18 04:49:28 +00:00
builder.addCapability(spv::CapabilityInt8);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
if (paramType.contains16BitInt())
Add Float16/Int8/Int16 capabilities for private variables and function parameters
2019-02-18 04:49:28 +00:00
builder.addCapability(spv::CapabilityInt16);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
if (paramType.contains16BitFloat())
Add Float16/Int8/Int16 capabilities for private variables and function parameters
2019-02-18 04:49:28 +00:00
builder.addCapability(spv::CapabilityFloat16);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
}
}
// Process all the initializers, while skipping the functions and link objects
void TGlslangToSpvTraverser::makeGlobalInitializers(const glslang::TIntermSequence& initializers)
{
builder.setBuildPoint(shaderEntry->getLastBlock());
for (int i = 0; i < (int)initializers.size(); ++i) {
glslang::TIntermAggregate* initializer = initializers[i]->getAsAggregate();
if (initializer && initializer->getOp() != glslang::EOpFunction && initializer->getOp() != glslang::EOpLinkerObjects) {
// We're on a top-level node that's not a function. Treat as an initializer, whose
Non-functional: Sweep through the stack for consistent with "main" and entry point. Partially addresses issue #513.
2016-09-19 22:01:41 +00:00
// code goes into the beginning of the entry point.
Final round for line endings.
2015-06-26 22:58:36 +00:00
initializer->traverse(this);
}
}
}
// Process all the functions, while skipping initializers.
void TGlslangToSpvTraverser::visitFunctions(const glslang::TIntermSequence& glslFunctions)
{
for (int f = 0; f < (int)glslFunctions.size(); ++f) {
glslang::TIntermAggregate* node = glslFunctions[f]->getAsAggregate();
Linker: Walk the call graph to report an error on missing bodies.
2016-12-09 04:01:59 +00:00
if (node && (node->getOp() == glslang::EOpFunction || node->getOp() == glslang::EOpLinkerObjects))
Final round for line endings.
2015-06-26 22:58:36 +00:00
node->traverse(this);
}
}
void TGlslangToSpvTraverser::handleFunctionEntry(const glslang::TIntermAggregate* node)
{
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
// SPIR-V functions should already be in the functionMap from the prepass
Final round for line endings.
2015-06-26 22:58:36 +00:00
// that called makeFunctions().
HLSL: Do structure conversion for return type struct-punning on non-entry-point functions.
2016-10-06 18:59:51 +00:00
currentFunction = functionMap[node->getName().c_str()];
spv::Block* functionBlock = currentFunction->getEntryBlock();
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.setBuildPoint(functionBlock);
}
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate& node, std::vector<spv::Id>& arguments, spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
const glslang::TIntermSequence& glslangArguments = node.getSequence();
Parser & SPV: Implement two extensions regarding GLSL sparse texture. Implement extension "GL_ARB_sparse_texture2". Implement extension "GL_ARB_sparse_texture_clamp".
2015-12-31 08:11:41 +00:00
glslang::TSampler sampler = {};
bool cubeCompare = false;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Implement the extension GL_AMD_gpu_shader_half_float_fetch - Support new opaque types: f16sampler*, f16image*, f16subpassInput*. - Add new built-in GLSL texture/image functions.
2016-11-29 09:36:31 +00:00
bool f16ShadowCompare = false;
#endif
SPV: Support the new OpCode - OpImageSparseRead
2016-02-19 14:24:03 +00:00
if (node.isTexture() || node.isImage()) {
Parser & SPV: Implement two extensions regarding GLSL sparse texture. Implement extension "GL_ARB_sparse_texture2". Implement extension "GL_ARB_sparse_texture_clamp".
2015-12-31 08:11:41 +00:00
sampler = glslangArguments[0]->getAsTyped()->getType().getSampler();
cubeCompare = sampler.dim == glslang::EsdCube && sampler.arrayed && sampler.shadow;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Implement the extension GL_AMD_gpu_shader_half_float_fetch - Support new opaque types: f16sampler*, f16image*, f16subpassInput*. - Add new built-in GLSL texture/image functions.
2016-11-29 09:36:31 +00:00
f16ShadowCompare = sampler.shadow && glslangArguments[1]->getAsTyped()->getType().getBasicType() == glslang::EbtFloat16;
#endif
Parser & SPV: Implement two extensions regarding GLSL sparse texture. Implement extension "GL_ARB_sparse_texture2". Implement extension "GL_ARB_sparse_texture_clamp".
2015-12-31 08:11:41 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
for (int i = 0; i < (int)glslangArguments.size(); ++i) {
builder.clearAccessChain();
glslangArguments[i]->traverse(this);
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
// Special case l-value operands
bool lvalue = false;
switch (node.getOp()) {
case glslang::EOpImageAtomicAdd:
case glslang::EOpImageAtomicMin:
case glslang::EOpImageAtomicMax:
case glslang::EOpImageAtomicAnd:
case glslang::EOpImageAtomicOr:
case glslang::EOpImageAtomicXor:
case glslang::EOpImageAtomicExchange:
case glslang::EOpImageAtomicCompSwap:
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
case glslang::EOpImageAtomicLoad:
case glslang::EOpImageAtomicStore:
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
if (i == 0)
lvalue = true;
break;
SPV: Support the new OpCode - OpImageSparseRead
2016-02-19 14:24:03 +00:00
case glslang::EOpSparseImageLoad:
if ((sampler.ms && i == 3) || (! sampler.ms && i == 2))
lvalue = true;
break;
Implement the extension GL_AMD_gpu_shader_half_float_fetch - Support new opaque types: f16sampler*, f16image*, f16subpassInput*. - Add new built-in GLSL texture/image functions.
2016-11-29 09:36:31 +00:00
case glslang::EOpSparseTexture:
if (((cubeCompare || f16ShadowCompare) && i == 3) || (! (cubeCompare || f16ShadowCompare) && i == 2))
lvalue = true;
break;
case glslang::EOpSparseTextureClamp:
if (((cubeCompare || f16ShadowCompare) && i == 4) || (! (cubeCompare || f16ShadowCompare) && i == 3))
lvalue = true;
break;
case glslang::EOpSparseTextureLod:
case glslang::EOpSparseTextureOffset:
if ((f16ShadowCompare && i == 4) || (! f16ShadowCompare && i == 3))
lvalue = true;
break;
Parser & SPV: Implement two extensions regarding GLSL sparse texture. Implement extension "GL_ARB_sparse_texture2". Implement extension "GL_ARB_sparse_texture_clamp".
2015-12-31 08:11:41 +00:00
case glslang::EOpSparseTextureFetch:
if ((sampler.dim != glslang::EsdRect && i == 3) || (sampler.dim == glslang::EsdRect && i == 2))
lvalue = true;
break;
case glslang::EOpSparseTextureFetchOffset:
if ((sampler.dim != glslang::EsdRect && i == 4) || (sampler.dim == glslang::EsdRect && i == 3))
lvalue = true;
break;
Implement the extension GL_AMD_gpu_shader_half_float_fetch - Support new opaque types: f16sampler*, f16image*, f16subpassInput*. - Add new built-in GLSL texture/image functions.
2016-11-29 09:36:31 +00:00
case glslang::EOpSparseTextureLodOffset:
case glslang::EOpSparseTextureGrad:
case glslang::EOpSparseTextureOffsetClamp:
if ((f16ShadowCompare && i == 5) || (! f16ShadowCompare && i == 4))
lvalue = true;
break;
case glslang::EOpSparseTextureGradOffset:
case glslang::EOpSparseTextureGradClamp:
if ((f16ShadowCompare && i == 6) || (! f16ShadowCompare && i == 5))
lvalue = true;
break;
case glslang::EOpSparseTextureGradOffsetClamp:
if ((f16ShadowCompare && i == 7) || (! f16ShadowCompare && i == 6))
lvalue = true;
break;
Implement the extension GL_AMD_texture_gather_bias_lod
2016-11-17 09:47:59 +00:00
case glslang::EOpSparseTextureGather:
Parser & SPV: Implement two extensions regarding GLSL sparse texture. Implement extension "GL_ARB_sparse_texture2". Implement extension "GL_ARB_sparse_texture_clamp".
2015-12-31 08:11:41 +00:00
if ((sampler.shadow && i == 3) || (! sampler.shadow && i == 2))
lvalue = true;
break;
case glslang::EOpSparseTextureGatherOffset:
case glslang::EOpSparseTextureGatherOffsets:
if ((sampler.shadow && i == 4) || (! sampler.shadow && i == 3))
lvalue = true;
break;
Implement the extension GL_AMD_texture_gather_bias_lod
2016-11-17 09:47:59 +00:00
case glslang::EOpSparseTextureGatherLod:
if (i == 3)
lvalue = true;
break;
case glslang::EOpSparseTextureGatherLodOffset:
case glslang::EOpSparseTextureGatherLodOffsets:
if (i == 4)
lvalue = true;
break;
Implement extension GL_AMD_shader_image_load_store_lod
2017-07-05 09:23:28 +00:00
case glslang::EOpSparseImageLoadLod:
if (i == 3)
lvalue = true;
break;
0003-Add-support-for-SPV_NV_shader_image_footprint
2018-09-19 18:41:27 +00:00
case glslang::EOpImageSampleFootprintNV:
if (i == 4)
lvalue = true;
break;
case glslang::EOpImageSampleFootprintClampNV:
case glslang::EOpImageSampleFootprintLodNV:
if (i == 5)
lvalue = true;
break;
case glslang::EOpImageSampleFootprintGradNV:
if (i == 6)
lvalue = true;
break;
case glslang::EOpImageSampleFootprintGradClampNV:
if (i == 7)
lvalue = true;
break;
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
default:
break;
}
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
if (lvalue) {
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
arguments.push_back(builder.accessChainGetLValue());
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
lvalueCoherentFlags = builder.getAccessChain().coherentFlags;
lvalueCoherentFlags |= TranslateCoherent(glslangArguments[i]->getAsTyped()->getType());
} else
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#endif
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
arguments.push_back(accessChainLoad(glslangArguments[i]->getAsTyped()->getType()));
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
}
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
void TGlslangToSpvTraverser::translateArguments(glslang::TIntermUnary& node, std::vector<spv::Id>& arguments)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
builder.clearAccessChain();
node.getOperand()->traverse(this);
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
arguments.push_back(accessChainLoad(node.getOperand()->getType()));
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermOperator* node)
{
SPV: Debug output: Include OpLine information for execution path. Note that declaratives are not handled, only procedurals.
2017-06-01 00:50:53 +00:00
if (! node->isImage() && ! node->isTexture())
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
return spv::NoResult;
SPV: Debug output: Include OpLine information for execution path. Note that declaratives are not handled, only procedurals.
2017-06-01 00:50:53 +00:00
SPV: Fix #1575, fix #1593: Support HLSL #line SPIR-V OpLines now contain filenames from HLSL-style #lines.
2018-12-08 00:36:33 +00:00
builder.setLine(node->getLoc().line, node->getLoc().getFilename());
SPV: Debug output: Include OpLine information for execution path. Note that declaratives are not handled, only procedurals.
2017-06-01 00:50:53 +00:00
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
// Process a GLSL texturing op (will be SPV image)
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
const glslang::TType &imageType = node->getAsAggregate()
? node->getAsAggregate()->getSequence()[0]->getAsTyped()->getType()
: node->getAsUnaryNode()->getOperand()->getAsTyped()->getType();
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
const glslang::TSampler sampler = imageType.getSampler();
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifdef GLSLANG_WEB
const bool f16ShadowCompare = false;
#else
Implement the extension GL_AMD_gpu_shader_half_float_fetch - Support new opaque types: f16sampler*, f16image*, f16subpassInput*. - Add new built-in GLSL texture/image functions.
2016-11-29 09:36:31 +00:00
bool f16ShadowCompare = (sampler.shadow && node->getAsAggregate())
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
? node->getAsAggregate()->getSequence()[1]->getAsTyped()->getType().getBasicType() == glslang::EbtFloat16
: false;
Implement the extension GL_AMD_gpu_shader_half_float_fetch - Support new opaque types: f16sampler*, f16image*, f16subpassInput*. - Add new built-in GLSL texture/image functions.
2016-11-29 09:36:31 +00:00
#endif
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
const auto signExtensionMask = [&]() {
if (builder.getSpvVersion() >= spv::Spv_1_4) {
if (sampler.type == glslang::EbtUint)
return spv::ImageOperandsZeroExtendMask;
else if (sampler.type == glslang::EbtInt)
return spv::ImageOperandsSignExtendMask;
}
return spv::ImageOperandsMaskNone;
};
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
spv::Builder::AccessChain::CoherentFlags lvalueCoherentFlags;
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
std::vector<spv::Id> arguments;
if (node->getAsAggregate())
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
translateArguments(*node->getAsAggregate(), arguments, lvalueCoherentFlags);
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
else
translateArguments(*node->getAsUnaryNode(), arguments);
Front-end: Implement 2nd task of issue #400; precision of result and operation. From the ES spec + Bugzilla 15931 and GL_KHR_vulkan_glsl: - Update precision qualifiers for all built-in function prototypes. - Implement the new algorithm used to distinguish built-in function operation precisions from result precisions. Also add tracking of separate result and operation precisions, and use that in generating SPIR-V. (SPIR-V cares about precision of operation, while the front-end cares about precision of result, for propagation.)
2016-08-02 01:44:00 +00:00
spv::Decoration precision = TranslatePrecisionDecoration(node->getOperationPrecision());
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
spv::Builder::TextureParameters params = { };
params.sampler = arguments[0];
Fix issues from review comments and those relevant to texelFetch
2015-09-16 03:44:02 +00:00
glslang::TCrackedTextureOp cracked;
node->crackTexture(sampler, cracked);
Implement the extension GL_AMD_shader_fragment_mask
2017-06-13 20:53:02 +00:00
const bool isUnsignedResult = node->getType().getBasicType() == glslang::EbtUint;
Preserve signedness in SPV image query ops The AST->SPIRV translation of image queries was dropping signedness, causing some validation troubles.
2017-03-10 19:45:50 +00:00
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
// Check for queries
if (cracked.query) {
SPV: Use SampledImage with OpImageQueryLod Khronos SPIR-V issue #74
2016-10-12 13:40:37 +00:00
// OpImageQueryLod works on a sampled image, for other queries the image has to be extracted first
if (node->getOp() != glslang::EOpTextureQueryLod && builder.isSampledImage(params.sampler))
SPV: Implement OpImage functionality (emit before query).
2015-12-09 02:32:47 +00:00
params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler);
SPV: Use SampledImage with OpImageQueryLod Khronos SPIR-V issue #74
2016-10-12 13:40:37 +00:00
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
switch (node->getOp()) {
case glslang::EOpImageQuerySize:
case glslang::EOpTextureQuerySize:
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (arguments.size() > 1) {
params.lod = arguments[1];
Preserve signedness in SPV image query ops The AST->SPIRV translation of image queries was dropping signedness, causing some validation troubles.
2017-03-10 19:45:50 +00:00
return builder.createTextureQueryCall(spv::OpImageQuerySizeLod, params, isUnsignedResult);
Final round for line endings.
2015-06-26 22:58:36 +00:00
} else
Preserve signedness in SPV image query ops The AST->SPIRV translation of image queries was dropping signedness, causing some validation troubles.
2017-03-10 19:45:50 +00:00
return builder.createTextureQueryCall(spv::OpImageQuerySize, params, isUnsignedResult);
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
case glslang::EOpImageQuerySamples:
case glslang::EOpTextureQuerySamples:
Preserve signedness in SPV image query ops The AST->SPIRV translation of image queries was dropping signedness, causing some validation troubles.
2017-03-10 19:45:50 +00:00
return builder.createTextureQueryCall(spv::OpImageQuerySamples, params, isUnsignedResult);
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
case glslang::EOpTextureQueryLod:
params.coords = arguments[1];
Preserve signedness in SPV image query ops The AST->SPIRV translation of image queries was dropping signedness, causing some validation troubles.
2017-03-10 19:45:50 +00:00
return builder.createTextureQueryCall(spv::OpImageQueryLod, params, isUnsignedResult);
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
case glslang::EOpTextureQueryLevels:
Preserve signedness in SPV image query ops The AST->SPIRV translation of image queries was dropping signedness, causing some validation troubles.
2017-03-10 19:45:50 +00:00
return builder.createTextureQueryCall(spv::OpImageQueryLevels, params, isUnsignedResult);
Parser & SPV: Implement two extensions regarding GLSL sparse texture. Implement extension "GL_ARB_sparse_texture2". Implement extension "GL_ARB_sparse_texture_clamp".
2015-12-31 08:11:41 +00:00
case glslang::EOpSparseTexelsResident:
return builder.createUnaryOp(spv::OpImageSparseTexelsResident, builder.makeBoolType(), arguments[0]);
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#endif
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
default:
assert(0);
break;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
HLSL: Add conversions for image ops during SPV construction HLSL allows image and texture types to be templatized on sub-vec4 types, or even structures. This was mostly handled already during creation of sampling operations. However, for operator[] which can generate image loads, this wasn't happening. It also isn't very easy to do at that point in time, because operator[] does not know where the results it produces will end up. They may be an lvalue or an rvalue, and there's a post-process to convert loads to stores. They may end up in atomic ops. To bypass that difficulty, GlslangToSpv now looks for this case and adds the appropriate conversion. LIMITATION: this only works for cases for which a simple conversion opcode suffices. That is to say, it will not work if the type is templatized on a struct.
2018-02-18 18:40:01 +00:00
int components = node->getType().getVectorSize();
if (node->getOp() == glslang::EOpTextureFetch) {
// These must produce 4 components, per SPIR-V spec. We'll add a conversion constructor if needed.
// This will only happen through the HLSL path for operator[], so we do not have to handle e.g.
// the EOpTexture/Proj/Lod/etc family. It would be harmless to do so, but would need more logic
// here around e.g. which ones return scalars or other types.
components = 4;
}
glslang::TType returnType(node->getType().getBasicType(), glslang::EvqTemporary, components);
auto resultType = [&returnType,this]{ return convertGlslangToSpvType(returnType); };
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
// Check for image functions other than queries
if (node->isImage()) {
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
std::vector<spv::IdImmediate> operands;
SPV: Minor corrections to previous big merge for images and atomics.
2015-09-16 16:54:31 +00:00
auto opIt = arguments.begin();
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::IdImmediate image = { true, *(opIt++) };
operands.push_back(image);
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
// Handle subpass operations
// TODO: GLSL should change to have the "MS" only on the type rather than the
// built-in function.
if (cracked.subpass) {
// add on the (0,0) coordinate
spv::Id zero = builder.makeIntConstant(0);
std::vector<spv::Id> comps;
comps.push_back(zero);
comps.push_back(zero);
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::IdImmediate coord = { true,
builder.makeCompositeConstant(builder.makeVectorType(builder.makeIntType(32), 2), comps) };
operands.push_back(coord);
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
spv::IdImmediate imageOperands = { false, spv::ImageOperandsMaskNone };
imageOperands.word = imageOperands.word | signExtensionMask();
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
if (sampler.isMultiSample()) {
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
imageOperands.word = imageOperands.word | spv::ImageOperandsSampleMask;
}
if (imageOperands.word != spv::ImageOperandsMaskNone) {
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
operands.push_back(imageOperands);
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
if (sampler.isMultiSample()) {
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
spv::IdImmediate imageOperand = { true, *(opIt++) };
operands.push_back(imageOperand);
}
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
}
GLSL->SPIR-V: Put precision decorations on imageLoad().
2017-10-19 08:07:30 +00:00
spv::Id result = builder.createOp(spv::OpImageRead, resultType(), operands);
builder.setPrecision(result, precision);
return result;
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
}
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::IdImmediate coord = { true, *(opIt++) };
operands.push_back(coord);
Implement extension GL_AMD_shader_image_load_store_lod
2017-07-05 09:23:28 +00:00
if (node->getOp() == glslang::EOpImageLoad || node->getOp() == glslang::EOpImageLoadLod) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone;
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
if (sampler.isMultiSample()) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
mask = mask | spv::ImageOperandsSampleMask;
}
if (cracked.lod) {
Implement extension GL_AMD_shader_image_load_store_lod
2017-07-05 09:23:28 +00:00
builder.addExtension(spv::E_SPV_AMD_shader_image_load_store_lod);
builder.addCapability(spv::CapabilityImageReadWriteLodAMD);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
mask = mask | spv::ImageOperandsLodMask;
}
mask = mask | TranslateImageOperands(TranslateCoherent(imageType));
mask = (spv::ImageOperandsMask)(mask & ~spv::ImageOperandsMakeTexelAvailableKHRMask);
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
mask = mask | signExtensionMask();
Build: Fix 3 warnings.
2019-05-10 12:47:00 +00:00
if (mask != spv::ImageOperandsMaskNone) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spv::IdImmediate imageOperands = { false, (unsigned int)mask };
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
operands.push_back(imageOperands);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
}
if (mask & spv::ImageOperandsSampleMask) {
spv::IdImmediate imageOperand = { true, *opIt++ };
operands.push_back(imageOperand);
}
if (mask & spv::ImageOperandsLodMask) {
spv::IdImmediate imageOperand = { true, *opIt++ };
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
operands.push_back(imageOperand);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
}
if (mask & spv::ImageOperandsMakeTexelVisibleKHRMask) {
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
spv::IdImmediate imageOperand = { true,
builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) };
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
operands.push_back(imageOperand);
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
}
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
if (builder.getImageTypeFormat(builder.getImageType(operands.front().word)) == spv::ImageFormatUnknown)
SPV Capabilities: StorageImageExtendedFormats, StorageImageReadWithoutFormat, and StorageImageWriteWithoutFormat.
2016-02-15 18:57:00 +00:00
builder.addCapability(spv::CapabilityStorageImageReadWithoutFormat);
GLSL->SPIR-V: Put precision decorations on imageLoad().
2017-10-19 08:07:30 +00:00
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
std::vector<spv::Id> result(1, builder.createOp(spv::OpImageRead, resultType(), operands));
HLSL: Add conversions for image ops during SPV construction HLSL allows image and texture types to be templatized on sub-vec4 types, or even structures. This was mostly handled already during creation of sampling operations. However, for operator[] which can generate image loads, this wasn't happening. It also isn't very easy to do at that point in time, because operator[] does not know where the results it produces will end up. They may be an lvalue or an rvalue, and there's a post-process to convert loads to stores. They may end up in atomic ops. To bypass that difficulty, GlslangToSpv now looks for this case and adds the appropriate conversion. LIMITATION: this only works for cases for which a simple conversion opcode suffices. That is to say, it will not work if the type is templatized on a struct.
2018-02-18 18:40:01 +00:00
builder.setPrecision(result[0], precision);
// If needed, add a conversion constructor to the proper size.
if (components != node->getType().getVectorSize())
result[0] = builder.createConstructor(precision, result, convertGlslangToSpvType(node->getType()));
return result[0];
Implement extension GL_AMD_shader_image_load_store_lod
2017-07-05 09:23:28 +00:00
} else if (node->getOp() == glslang::EOpImageStore || node->getOp() == glslang::EOpImageStoreLod) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
// Push the texel value before the operands
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
if (sampler.isMultiSample() || cracked.lod) {
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::IdImmediate texel = { true, *(opIt + 1) };
operands.push_back(texel);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
} else {
spv::IdImmediate texel = { true, *opIt };
operands.push_back(texel);
}
spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone;
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
if (sampler.isMultiSample()) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
mask = mask | spv::ImageOperandsSampleMask;
}
if (cracked.lod) {
Implement extension GL_AMD_shader_image_load_store_lod
2017-07-05 09:23:28 +00:00
builder.addExtension(spv::E_SPV_AMD_shader_image_load_store_lod);
builder.addCapability(spv::CapabilityImageReadWriteLodAMD);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
mask = mask | spv::ImageOperandsLodMask;
}
mask = mask | TranslateImageOperands(TranslateCoherent(imageType));
mask = (spv::ImageOperandsMask)(mask & ~spv::ImageOperandsMakeTexelVisibleKHRMask);
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
mask = mask | signExtensionMask();
Build: Fix 3 warnings.
2019-05-10 12:47:00 +00:00
if (mask != spv::ImageOperandsMaskNone) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spv::IdImmediate imageOperands = { false, (unsigned int)mask };
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
operands.push_back(imageOperands);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
}
if (mask & spv::ImageOperandsSampleMask) {
spv::IdImmediate imageOperand = { true, *opIt++ };
operands.push_back(imageOperand);
}
if (mask & spv::ImageOperandsLodMask) {
spv::IdImmediate imageOperand = { true, *opIt++ };
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
operands.push_back(imageOperand);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
}
if (mask & spv::ImageOperandsMakeTexelAvailableKHRMask) {
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
spv::IdImmediate imageOperand = { true,
builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) };
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
operands.push_back(imageOperand);
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
}
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
SPV: Minor corrections to previous big merge for images and atomics.
2015-09-16 16:54:31 +00:00
builder.createNoResultOp(spv::OpImageWrite, operands);
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
if (builder.getImageTypeFormat(builder.getImageType(operands.front().word)) == spv::ImageFormatUnknown)
SPV Capabilities: StorageImageExtendedFormats, StorageImageReadWithoutFormat, and StorageImageWriteWithoutFormat.
2016-02-15 18:57:00 +00:00
builder.addCapability(spv::CapabilityStorageImageWriteWithoutFormat);
SPV: Minor corrections to previous big merge for images and atomics.
2015-09-16 16:54:31 +00:00
return spv::NoResult;
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
} else if (node->getOp() == glslang::EOpSparseImageLoad ||
node->getOp() == glslang::EOpSparseImageLoadLod) {
SPV: Support the new OpCode - OpImageSparseRead
2016-02-19 14:24:03 +00:00
builder.addCapability(spv::CapabilitySparseResidency);
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
if (builder.getImageTypeFormat(builder.getImageType(operands.front().word)) == spv::ImageFormatUnknown)
SPV: Support the new OpCode - OpImageSparseRead
2016-02-19 14:24:03 +00:00
builder.addCapability(spv::CapabilityStorageImageReadWithoutFormat);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spv::ImageOperandsMask mask = spv::ImageOperandsMaskNone;
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
if (sampler.isMultiSample()) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
mask = mask | spv::ImageOperandsSampleMask;
}
if (cracked.lod) {
Implement extension GL_AMD_shader_image_load_store_lod
2017-07-05 09:23:28 +00:00
builder.addExtension(spv::E_SPV_AMD_shader_image_load_store_lod);
builder.addCapability(spv::CapabilityImageReadWriteLodAMD);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
mask = mask | spv::ImageOperandsLodMask;
}
mask = mask | TranslateImageOperands(TranslateCoherent(imageType));
mask = (spv::ImageOperandsMask)(mask & ~spv::ImageOperandsMakeTexelAvailableKHRMask);
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
mask = mask | signExtensionMask();
Build: Fix 3 warnings.
2019-05-10 12:47:00 +00:00
if (mask != spv::ImageOperandsMaskNone) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spv::IdImmediate imageOperands = { false, (unsigned int)mask };
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
operands.push_back(imageOperands);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
}
if (mask & spv::ImageOperandsSampleMask) {
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::IdImmediate imageOperand = { true, *opIt++ };
operands.push_back(imageOperand);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
}
if (mask & spv::ImageOperandsLodMask) {
spv::IdImmediate imageOperand = { true, *opIt++ };
operands.push_back(imageOperand);
}
if (mask & spv::ImageOperandsMakeTexelVisibleKHRMask) {
spv::IdImmediate imageOperand = { true, builder.makeUintConstant(TranslateMemoryScope(TranslateCoherent(imageType))) };
operands.push_back(imageOperand);
SPV: Support the new OpCode - OpImageSparseRead
2016-02-19 14:24:03 +00:00
}
// Create the return type that was a special structure
spv::Id texelOut = *opIt;
SPV: For AST interpolateAt* ops consuming a swizzle, invert the order. Apply the interpolation first, then apply the swizzle to the result, the inverse of the order requested by the AST. This fixes issue #411.
2016-07-26 18:50:38 +00:00
spv::Id typeId0 = resultType();
SPV: Support the new OpCode - OpImageSparseRead
2016-02-19 14:24:03 +00:00
spv::Id typeId1 = builder.getDerefTypeId(texelOut);
spv::Id resultTypeId = builder.makeStructResultType(typeId0, typeId1);
spv::Id resultId = builder.createOp(spv::OpImageSparseRead, resultTypeId, operands);
// Decode the return type
builder.createStore(builder.createCompositeExtract(resultId, typeId1, 1), texelOut);
return builder.createCompositeExtract(resultId, typeId0, 0);
SPV: the OpImageTexelPointer sample 0 should be <id> of 0, not literal 0.
2016-01-22 16:54:12 +00:00
} else {
Add new test case for image functions and fix issues caught by this test
2015-09-16 09:48:22 +00:00
// Process image atomic operations
// GLSL "IMAGE_PARAMS" will involve in constructing an image texel pointer and this pointer,
// as the first source operand, is required by SPIR-V atomic operations.
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
// For non-MS, the sample value should be 0
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
spv::IdImmediate sample = { true, sampler.isMultiSample() ? *(opIt++) : builder.makeUintConstant(0) };
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
operands.push_back(sample);
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spv::Id resultTypeId;
// imageAtomicStore has a void return type so base the pointer type on
// the type of the value operand.
if (node->getOp() == glslang::EOpImageAtomicStore) {
resultTypeId = builder.makePointer(spv::StorageClassImage, builder.getTypeId(operands[2].word));
} else {
resultTypeId = builder.makePointer(spv::StorageClassImage, resultType());
}
SPV: Minor corrections to previous big merge for images and atomics.
2015-09-16 16:54:31 +00:00
spv::Id pointer = builder.createOp(spv::OpImageTexelPointer, resultTypeId, operands);
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
std::vector<spv::Id> operands;
operands.push_back(pointer);
for (; opIt != arguments.end(); ++opIt)
operands.push_back(*opIt);
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
return createAtomicOperation(node->getOp(), precision, resultType(), operands, node->getBasicType(), lvalueCoherentFlags);
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
}
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Implement the extension GL_AMD_shader_fragment_mask
2017-06-13 20:53:02 +00:00
// Check for fragment mask functions other than queries
if (cracked.fragMask) {
assert(sampler.ms);
auto opIt = arguments.begin();
std::vector<spv::Id> operands;
// Extract the image if necessary
if (builder.isSampledImage(params.sampler))
params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler);
operands.push_back(params.sampler);
++opIt;
if (sampler.isSubpass()) {
// add on the (0,0) coordinate
spv::Id zero = builder.makeIntConstant(0);
std::vector<spv::Id> comps;
comps.push_back(zero);
comps.push_back(zero);
operands.push_back(builder.makeCompositeConstant(builder.makeVectorType(builder.makeIntType(32), 2), comps));
}
for (; opIt != arguments.end(); ++opIt)
operands.push_back(*opIt);
spv::Op fragMaskOp = spv::OpNop;
if (node->getOp() == glslang::EOpFragmentMaskFetch)
fragMaskOp = spv::OpFragmentMaskFetchAMD;
else if (node->getOp() == glslang::EOpFragmentFetch)
fragMaskOp = spv::OpFragmentFetchAMD;
builder.addExtension(spv::E_SPV_AMD_shader_fragment_mask);
builder.addCapability(spv::CapabilityFragmentMaskAMD);
return builder.createOp(fragMaskOp, resultType(), operands);
}
#endif
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
// Check for texture functions other than queries
Parser & SPV: Implement two extensions regarding GLSL sparse texture. Implement extension "GL_ARB_sparse_texture2". Implement extension "GL_ARB_sparse_texture_clamp".
2015-12-31 08:11:41 +00:00
bool sparse = node->isSparseTexture();
0003-Add-support-for-SPV_NV_shader_image_footprint
2018-09-19 18:41:27 +00:00
bool imageFootprint = node->isImageFootprint();
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
bool cubeCompare = sampler.dim == glslang::EsdCube && sampler.isArrayed() && sampler.isShadow();
SPV: Fix an issue related with using CubeArrayShadow for texture()
2015-11-11 07:35:47 +00:00
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
// check for bias argument
bool bias = false;
Implement the extension GL_AMD_texture_gather_bias_lod
2016-11-17 09:47:59 +00:00
if (! cracked.lod && ! cracked.grad && ! cracked.fetch && ! cubeCompare) {
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
int nonBiasArgCount = 2;
Implement the extension GL_AMD_texture_gather_bias_lod
2016-11-17 09:47:59 +00:00
if (cracked.gather)
++nonBiasArgCount; // comp argument should be present when bias argument is present
Implement the extension GL_AMD_gpu_shader_half_float_fetch - Support new opaque types: f16sampler*, f16image*, f16subpassInput*. - Add new built-in GLSL texture/image functions.
2016-11-29 09:36:31 +00:00
if (f16ShadowCompare)
++nonBiasArgCount;
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
if (cracked.offset)
++nonBiasArgCount;
Implement the extension GL_AMD_texture_gather_bias_lod
2016-11-17 09:47:59 +00:00
else if (cracked.offsets)
++nonBiasArgCount;
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
if (cracked.grad)
nonBiasArgCount += 2;
Parser & SPV: Implement two extensions regarding GLSL sparse texture. Implement extension "GL_ARB_sparse_texture2". Implement extension "GL_ARB_sparse_texture_clamp".
2015-12-31 08:11:41 +00:00
if (cracked.lodClamp)
++nonBiasArgCount;
if (sparse)
++nonBiasArgCount;
0003-Add-support-for-SPV_NV_shader_image_footprint
2018-09-19 18:41:27 +00:00
if (imageFootprint)
//Following three extra arguments
// int granularity, bool coarse, out gl_TextureFootprint2DNV footprint
nonBiasArgCount += 3;
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
if ((int)arguments.size() > nonBiasArgCount)
bias = true;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
SPV: Fix issue #320: Fetch needs to extract the image. GLSL takes a traditional sampler, but SPIR-V wants just the sampled image, not the combined sampler and image.
2016-06-03 05:45:21 +00:00
// See if the sampler param should really be just the SPV image part
if (cracked.fetch) {
// a fetch needs to have the image extracted first
if (builder.isSampledImage(params.sampler))
params.sampler = builder.createUnaryOp(spv::OpImage, builder.getImageType(params.sampler), params.sampler);
}
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Implement the extension GL_AMD_texture_gather_bias_lod
2016-11-17 09:47:59 +00:00
if (cracked.gather) {
const auto& sourceExtensions = glslangIntermediate->getRequestedExtensions();
if (bias || cracked.lod ||
sourceExtensions.find(glslang::E_GL_AMD_texture_gather_bias_lod) != sourceExtensions.end()) {
builder.addExtension(spv::E_SPV_AMD_texture_gather_bias_lod);
SPV: Fix an typo of SPV_AMD_texture_gather_bias_lod
2017-06-14 15:09:39 +00:00
builder.addCapability(spv::CapabilityImageGatherBiasLodAMD);
Implement the extension GL_AMD_texture_gather_bias_lod
2016-11-17 09:47:59 +00:00
}
}
#endif
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
// set the rest of the arguments
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
params.coords = arguments[1];
int extraArgs = 0;
SPV: Fix issue #159: use ExplicitLod forms for non-fragment stages.
2016-02-15 22:40:42 +00:00
bool noImplicitLod = false;
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
// sort out where Dref is coming from
Implement the extension GL_AMD_gpu_shader_half_float_fetch - Support new opaque types: f16sampler*, f16image*, f16subpassInput*. - Add new built-in GLSL texture/image functions.
2016-11-29 09:36:31 +00:00
if (cubeCompare || f16ShadowCompare) {
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
params.Dref = arguments[2];
Parser & SPV: Implement two extensions regarding GLSL sparse texture. Implement extension "GL_ARB_sparse_texture2". Implement extension "GL_ARB_sparse_texture_clamp".
2015-12-31 08:11:41 +00:00
++extraArgs;
} else if (sampler.shadow && cracked.gather) {
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
params.Dref = arguments[2];
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
++extraArgs;
} else if (sampler.shadow) {
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
std::vector<spv::Id> indexes;
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
int dRefComp;
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
if (cracked.proj)
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
dRefComp = 2; // "The resulting 3rd component of P in the shadow forms is used as Dref"
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
else
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
dRefComp = builder.getNumComponents(params.coords) - 1;
indexes.push_back(dRefComp);
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
params.Dref = builder.createCompositeExtract(params.coords, builder.getScalarTypeId(builder.getTypeId(params.coords)), indexes);
}
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
// lod
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
if (cracked.lod) {
Fix dref explicit LOD form of sample with cube texture arrays The dref parameter was being used as the LOD. Now it it's properly the dref.
2017-07-25 00:45:37 +00:00
params.lod = arguments[2 + extraArgs];
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
++extraArgs;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
} else if (glslangIntermediate->getStage() != EShLangFragment &&
!(glslangIntermediate->getStage() == EShLangCompute &&
glslangIntermediate->hasLayoutDerivativeModeNone())) {
SPV: Fix issue #159: use ExplicitLod forms for non-fragment stages.
2016-02-15 22:40:42 +00:00
// we need to invent the default lod for an explicit lod instruction for a non-fragment stage
noImplicitLod = true;
}
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
// multisample
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
if (sampler.isMultiSample()) {
Fix dref explicit LOD form of sample with cube texture arrays The dref parameter was being used as the LOD. Now it it's properly the dref.
2017-07-25 00:45:37 +00:00
params.sample = arguments[2 + extraArgs]; // For MS, "sample" should be specified
Fix issues from review comments and those relevant to texelFetch
2015-09-16 03:44:02 +00:00
++extraArgs;
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
}
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
// gradient
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
if (cracked.grad) {
params.gradX = arguments[2 + extraArgs];
params.gradY = arguments[3 + extraArgs];
extraArgs += 2;
}
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
// offset and offsets
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
if (cracked.offset) {
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
params.offset = arguments[2 + extraArgs];
++extraArgs;
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
} else if (cracked.offsets) {
params.offsets = arguments[2 + extraArgs];
++extraArgs;
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
}
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
Web: Tighten up sampling code and interfaces. Saves about 9K.
2019-08-08 07:15:24 +00:00
#ifndef GLSLANG_WEB
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
// lod clamp
Parser & SPV: Implement two extensions regarding GLSL sparse texture. Implement extension "GL_ARB_sparse_texture2". Implement extension "GL_ARB_sparse_texture_clamp".
2015-12-31 08:11:41 +00:00
if (cracked.lodClamp) {
params.lodClamp = arguments[2 + extraArgs];
++extraArgs;
}
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
// sparse
Parser & SPV: Implement two extensions regarding GLSL sparse texture. Implement extension "GL_ARB_sparse_texture2". Implement extension "GL_ARB_sparse_texture_clamp".
2015-12-31 08:11:41 +00:00
if (sparse) {
params.texelOut = arguments[2 + extraArgs];
++extraArgs;
}
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
// gather component
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
if (cracked.gather && ! sampler.shadow) {
// default component is 0, if missing, otherwise an argument
if (2 + extraArgs < (int)arguments.size()) {
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
params.component = arguments[2 + extraArgs];
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
++extraArgs;
Implement the extension GL_AMD_texture_gather_bias_lod
2016-11-17 09:47:59 +00:00
} else
SPV non-functional: minor readability improvements for texturing.
2016-06-16 18:43:23 +00:00
params.component = builder.makeIntConstant(0);
Implement the extension GL_AMD_texture_gather_bias_lod
2016-11-17 09:47:59 +00:00
}
0003-Add-support-for-SPV_NV_shader_image_footprint
2018-09-19 18:41:27 +00:00
spv::Id resultStruct = spv::NoResult;
if (imageFootprint) {
//Following three extra arguments
// int granularity, bool coarse, out gl_TextureFootprint2DNV footprint
params.granularity = arguments[2 + extraArgs];
params.coarse = arguments[3 + extraArgs];
resultStruct = arguments[4 + extraArgs];
extraArgs += 3;
}
#endif
Implement the extension GL_AMD_texture_gather_bias_lod
2016-11-17 09:47:59 +00:00
// bias
if (bias) {
params.bias = arguments[2 + extraArgs];
++extraArgs;
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
0003-Add-support-for-SPV_NV_shader_image_footprint
2018-09-19 18:41:27 +00:00
if (imageFootprint) {
builder.addExtension(spv::E_SPV_NV_shader_image_footprint);
builder.addCapability(spv::CapabilityImageFootprintNV);
//resultStructType(OpenGL type) contains 5 elements:
//struct gl_TextureFootprint2DNV {
// uvec2 anchor;
// uvec2 offset;
// uvec2 mask;
// uint lod;
// uint granularity;
//};
//or
//struct gl_TextureFootprint3DNV {
// uvec3 anchor;
// uvec3 offset;
// uvec2 mask;
// uint lod;
// uint granularity;
//};
spv::Id resultStructType = builder.getContainedTypeId(builder.getTypeId(resultStruct));
assert(builder.isStructType(resultStructType));
//resType (SPIR-V type) contains 6 elements:
//Member 0 must be a Boolean type scalar(LOD),
//Member 1 must be a vector of integer type, whose Signedness operand is 0(anchor),
//Member 2 must be a vector of integer type, whose Signedness operand is 0(offset),
//Member 3 must be a vector of integer type, whose Signedness operand is 0(mask),
//Member 4 must be a scalar of integer type, whose Signedness operand is 0(lod),
//Member 5 must be a scalar of integer type, whose Signedness operand is 0(granularity).
std::vector<spv::Id> members;
members.push_back(resultType());
for (int i = 0; i < 5; i++) {
members.push_back(builder.getContainedTypeId(resultStructType, i));
}
spv::Id resType = builder.makeStructType(members, "ResType");
//call ImageFootprintNV
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
spv::Id res = builder.createTextureCall(precision, resType, sparse, cracked.fetch, cracked.proj,
cracked.gather, noImplicitLod, params, signExtensionMask());
0003-Add-support-for-SPV_NV_shader_image_footprint
2018-09-19 18:41:27 +00:00
//copy resType (SPIR-V type) to resultStructType(OpenGL type)
for (int i = 0; i < 5; i++) {
builder.clearAccessChain();
builder.setAccessChainLValue(resultStruct);
//Accessing to a struct we created, no coherent flag is set
spv::Builder::AccessChain::CoherentFlags flags;
flags.clear();
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
builder.accessChainPush(builder.makeIntConstant(i), flags, 0);
0003-Add-support-for-SPV_NV_shader_image_footprint
2018-09-19 18:41:27 +00:00
builder.accessChainStore(builder.createCompositeExtract(res, builder.getContainedTypeId(resType, i+1), i+1));
}
return builder.createCompositeExtract(res, resultType(), 0);
}
#endif
SPV: projective texturing: move the project coordinate. SPV doesn't allow gaps in the components of the texturing coordinate. It also removes the shadow reference. So, close up the components so all used components are together.
2016-06-16 20:06:26 +00:00
// projective component (might not to move)
// GLSL: "The texture coordinates consumed from P, not including the last component of P,
// are divided by the last component of P."
// SPIR-V: "... (u [, v] [, w], q)... It may be a vector larger than needed, but all
// unused components will appear after all used components."
if (cracked.proj) {
int projSourceComp = builder.getNumComponents(params.coords) - 1;
int projTargetComp;
switch (sampler.dim) {
case glslang::Esd1D: projTargetComp = 1; break;
case glslang::Esd2D: projTargetComp = 2; break;
case glslang::EsdRect: projTargetComp = 2; break;
default: projTargetComp = projSourceComp; break;
}
// copy the projective coordinate if we have to
if (projTargetComp != projSourceComp) {
Non-Functional: Whitespace, comments, replace accidentally deleted comment. - fixed ParseHelper.cpp newlines (crlf -> lf) - removed trailing white space in most source files - fix some spelling issues - extra blank lines - tabs to spaces - replace #include comment about no location
2017-01-06 07:34:48 +00:00
spv::Id projComp = builder.createCompositeExtract(params.coords,
SPV: projective texturing: move the project coordinate. SPV doesn't allow gaps in the components of the texturing coordinate. It also removes the shadow reference. So, close up the components so all used components are together.
2016-06-16 20:06:26 +00:00
builder.getScalarTypeId(builder.getTypeId(params.coords)),
projSourceComp);
params.coords = builder.createCompositeInsert(projComp, params.coords,
builder.getTypeId(params.coords), projTargetComp);
}
}
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
// nonprivate
if (imageType.getQualifier().nonprivate) {
params.nonprivate = true;
}
// volatile
if (imageType.getQualifier().volatil) {
params.volatil = true;
}
fix further MSVC 2012 build issues MSVC 2012 does not allow initializer lists - replaced with corrresponding CTor calls.
2018-04-09 19:08:14 +00:00
std::vector<spv::Id> result( 1,
SPV 1.4: Emit SignExtend and ZeroExtend for integer image reads/writes.
2019-03-31 16:51:57 +00:00
builder.createTextureCall(precision, resultType(), sparse, cracked.fetch, cracked.proj, cracked.gather,
noImplicitLod, params, signExtensionMask())
fix further MSVC 2012 build issues MSVC 2012 does not allow initializer lists - replaced with corrresponding CTor calls.
2018-04-09 19:08:14 +00:00
);
HLSL: Add conversions for image ops during SPV construction HLSL allows image and texture types to be templatized on sub-vec4 types, or even structures. This was mostly handled already during creation of sampling operations. However, for operator[] which can generate image loads, this wasn't happening. It also isn't very easy to do at that point in time, because operator[] does not know where the results it produces will end up. They may be an lvalue or an rvalue, and there's a post-process to convert loads to stores. They may end up in atomic ops. To bypass that difficulty, GlslangToSpv now looks for this case and adds the appropriate conversion. LIMITATION: this only works for cases for which a simple conversion opcode suffices. That is to say, it will not work if the type is templatized on a struct.
2018-02-18 18:40:01 +00:00
if (components != node->getType().getVectorSize())
result[0] = builder.createConstructor(precision, result, convertGlslangToSpvType(node->getType()));
return result[0];
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
spv::Id TGlslangToSpvTraverser::handleUserFunctionCall(const glslang::TIntermAggregate* node)
{
// Grab the function's pointer from the previously created function
spv::Function* function = functionMap[node->getName().c_str()];
if (! function)
return 0;
const glslang::TIntermSequence& glslangArgs = node->getSequence();
const glslang::TQualifierList& qualifiers = node->getQualifierList();
// See comments in makeFunctions() for details about the semantics for parameter passing.
//
// These imply we need a four step process:
// 1. Evaluate the arguments
// 2. Allocate and make copies of in, out, and inout arguments
// 3. Make the call
// 4. Copy back the results
Fix #944: Convert argument type to match formal parameter type.
2018-05-04 17:43:03 +00:00
// 1. Evaluate the arguments and their types
Final round for line endings.
2015-06-26 22:58:36 +00:00
std::vector<spv::Builder::AccessChain> lValues;
std::vector<spv::Id> rValues;
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
std::vector<const glslang::TType*> argTypes;
Final round for line endings.
2015-06-26 22:58:36 +00:00
for (int a = 0; a < (int)glslangArgs.size(); ++a) {
Fix #944: Convert argument type to match formal parameter type.
2018-05-04 17:43:03 +00:00
argTypes.push_back(&glslangArgs[a]->getAsTyped()->getType());
Final round for line endings.
2015-06-26 22:58:36 +00:00
// build l-value
builder.clearAccessChain();
glslangArgs[a]->traverse(this);
SPV: rationalize parameter handling for "original" and "writable" parameters.
2017-09-10 21:21:05 +00:00
// keep outputs and pass-by-originals as l-values, evaluate others as r-values
Fix #944: Convert argument type to match formal parameter type.
2018-05-04 17:43:03 +00:00
if (originalParam(qualifiers[a], *argTypes[a], function->hasImplicitThis() && a == 0) ||
HLSL: Fix #1163: treat buffers as references when calling functions. This continues to prevent writing output buffers (out from a function), but fixes the problem where the copy-in/out was not getting done. Making everything work will require knowing both in/out-ness and bufferness, but these are currently mutually exclusive, because both are storage qualifiers.
2017-12-04 09:48:10 +00:00
writableParam(qualifiers[a])) {
Final round for line endings.
2015-06-26 22:58:36 +00:00
// save l-value
lValues.push_back(builder.getAccessChain());
} else {
// process r-value
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
rValues.push_back(accessChainLoad(*argTypes.back()));
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
}
// 2. Allocate space for anything needing a copy, and if it's "in" or "inout"
// copy the original into that space.
//
// Also, build up the list of actual arguments to pass in for the call
int lValueCount = 0;
int rValueCount = 0;
std::vector<spv::Id> spvArgs;
for (int a = 0; a < (int)glslangArgs.size(); ++a) {
spv::Id arg;
Fix #944: Convert argument type to match formal parameter type.
2018-05-04 17:43:03 +00:00
if (originalParam(qualifiers[a], *argTypes[a], function->hasImplicitThis() && a == 0)) {
SPV: Don't shadow sampler parameters when performing function calls Fixes #179
2016-05-25 18:50:21 +00:00
builder.setAccessChain(lValues[lValueCount]);
arg = builder.accessChainGetLValue();
++lValueCount;
SPV: rationalize parameter handling for "original" and "writable" parameters.
2017-09-10 21:21:05 +00:00
} else if (writableParam(qualifiers[a])) {
Final round for line endings.
2015-06-26 22:58:36 +00:00
// need space to hold the copy
Fix #944: Convert argument type to match formal parameter type.
2018-05-04 17:43:03 +00:00
arg = builder.createVariable(spv::StorageClassFunction, builder.getContainedTypeId(function->getParamType(a)), "param");
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (qualifiers[a] == glslang::EvqIn || qualifiers[a] == glslang::EvqInOut) {
// need to copy the input into output space
builder.setAccessChain(lValues[lValueCount]);
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
spv::Id copy = accessChainLoad(*argTypes[a]);
SPV: Flatten structs for copy when they are GLSL type aliases. Addresses issue #304 and issue #307 by replacing unmatched type OpStores with per-member copies. Covers assignment statements and most argument passing, but does not yet cover r-value-based argument passing.
2016-09-02 17:20:21 +00:00
builder.clearAccessChain();
builder.setAccessChainLValue(arg);
Fix #944: Convert argument type to match formal parameter type.
2018-05-04 17:43:03 +00:00
multiTypeStore(*argTypes[a], copy);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
++lValueCount;
} else {
Fix #944: Convert argument type to match formal parameter type.
2018-05-04 17:43:03 +00:00
// process r-value, which involves a copy for a type mismatch
if (function->getParamType(a) != convertGlslangToSpvType(*argTypes[a])) {
spv::Id argCopy = builder.createVariable(spv::StorageClassFunction, function->getParamType(a), "arg");
builder.clearAccessChain();
builder.setAccessChainLValue(argCopy);
multiTypeStore(*argTypes[a], rValues[rValueCount]);
arg = builder.createLoad(argCopy);
} else
arg = rValues[rValueCount];
Final round for line endings.
2015-06-26 22:58:36 +00:00
++rValueCount;
}
spvArgs.push_back(arg);
}
// 3. Make the call.
spv::Id result = builder.createFunctionCall(function, spvArgs);
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
builder.setPrecision(result, TranslatePrecisionDecoration(node->getType()));
Final round for line endings.
2015-06-26 22:58:36 +00:00
// 4. Copy back out an "out" arguments.
lValueCount = 0;
for (int a = 0; a < (int)glslangArgs.size(); ++a) {
Fix #944: Convert argument type to match formal parameter type.
2018-05-04 17:43:03 +00:00
if (originalParam(qualifiers[a], *argTypes[a], function->hasImplicitThis() && a == 0))
SPV: rationalize parameter handling for "original" and "writable" parameters.
2017-09-10 21:21:05 +00:00
++lValueCount;
else if (writableParam(qualifiers[a])) {
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (qualifiers[a] == glslang::EvqOut || qualifiers[a] == glslang::EvqInOut) {
spv::Id copy = builder.createLoad(spvArgs[a]);
builder.setAccessChain(lValues[lValueCount]);
Fix #944: Convert argument type to match formal parameter type.
2018-05-04 17:43:03 +00:00
multiTypeStore(*argTypes[a], copy);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
++lValueCount;
}
}
return result;
}
// Translate AST operation to SPV operation, already having SPV-based operands/types.
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id TGlslangToSpvTraverser::createBinaryOperation(glslang::TOperator op, OpDecorations& decorations,
Final round for line endings.
2015-06-26 22:58:36 +00:00
spv::Id typeId, spv::Id left, spv::Id right,
glslang::TBasicType typeProxy, bool reduceComparison)
{
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
bool isUnsigned = isTypeUnsignedInt(typeProxy);
bool isFloat = isTypeFloat(typeProxy);
SPV: Use OpLogicalEqual/OpLogicalNotEqual for boolean type comparison.
2016-04-27 00:15:37 +00:00
bool isBool = typeProxy == glslang::EbtBool;
Final round for line endings.
2015-06-26 22:58:36 +00:00
spv::Op binOp = spv::OpNop;
glslang -> SPIR-V: smear scalars for integer (scalar * vector). OpVectorTimesScalar is only for floats.
2015-07-04 23:17:31 +00:00
bool needMatchingVectors = true; // for non-matrix ops, would a scalar need to smear to match a vector?
Final round for line endings.
2015-06-26 22:58:36 +00:00
bool comparison = false;
switch (op) {
case glslang::EOpAdd:
case glslang::EOpAddAssign:
if (isFloat)
binOp = spv::OpFAdd;
else
binOp = spv::OpIAdd;
break;
case glslang::EOpSub:
case glslang::EOpSubAssign:
if (isFloat)
binOp = spv::OpFSub;
else
binOp = spv::OpISub;
break;
case glslang::EOpMul:
case glslang::EOpMulAssign:
if (isFloat)
binOp = spv::OpFMul;
else
binOp = spv::OpIMul;
break;
case glslang::EOpVectorTimesScalar:
case glslang::EOpVectorTimesScalarAssign:
Full stack: distinguish between a scalar and a vector of size 1. There have been GLSL extensions considering this, and HLSL does it. This is a fully backward compatible change that allows this distinction.
2016-05-20 18:06:03 +00:00
if (isFloat && (builder.isVector(left) || builder.isVector(right))) {
glslang -> SPIR-V: smear scalars for integer (scalar * vector). OpVectorTimesScalar is only for floats.
2015-07-04 23:17:31 +00:00
if (builder.isVector(right))
std::swap(left, right);
assert(builder.isScalar(right));
needMatchingVectors = false;
binOp = spv::OpVectorTimesScalar;
HLSL: fix invalid spir-v vector * scalar for float1 Missing check for float results in always emitting OpIMul even for float types
2018-11-14 12:04:39 +00:00
} else if (isFloat)
binOp = spv::OpFMul;
else
glslang -> SPIR-V: smear scalars for integer (scalar * vector). OpVectorTimesScalar is only for floats.
2015-07-04 23:17:31 +00:00
binOp = spv::OpIMul;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpVectorTimesMatrix:
case glslang::EOpVectorTimesMatrixAssign:
binOp = spv::OpVectorTimesMatrix;
break;
case glslang::EOpMatrixTimesVector:
binOp = spv::OpMatrixTimesVector;
break;
case glslang::EOpMatrixTimesScalar:
case glslang::EOpMatrixTimesScalarAssign:
binOp = spv::OpMatrixTimesScalar;
break;
case glslang::EOpMatrixTimesMatrix:
case glslang::EOpMatrixTimesMatrixAssign:
binOp = spv::OpMatrixTimesMatrix;
break;
case glslang::EOpOuterProduct:
binOp = spv::OpOuterProduct;
glslang -> SPIR-V: smear scalars for integer (scalar * vector). OpVectorTimesScalar is only for floats.
2015-07-04 23:17:31 +00:00
needMatchingVectors = false;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpDiv:
case glslang::EOpDivAssign:
if (isFloat)
binOp = spv::OpFDiv;
else if (isUnsigned)
binOp = spv::OpUDiv;
else
binOp = spv::OpSDiv;
break;
case glslang::EOpMod:
case glslang::EOpModAssign:
if (isFloat)
binOp = spv::OpFMod;
else if (isUnsigned)
binOp = spv::OpUMod;
else
binOp = spv::OpSMod;
break;
case glslang::EOpRightShift:
case glslang::EOpRightShiftAssign:
if (isUnsigned)
binOp = spv::OpShiftRightLogical;
else
binOp = spv::OpShiftRightArithmetic;
break;
case glslang::EOpLeftShift:
case glslang::EOpLeftShiftAssign:
binOp = spv::OpShiftLeftLogical;
break;
case glslang::EOpAnd:
case glslang::EOpAndAssign:
binOp = spv::OpBitwiseAnd;
break;
case glslang::EOpLogicalAnd:
glslang -> SPIR-V: smear scalars for integer (scalar * vector). OpVectorTimesScalar is only for floats.
2015-07-04 23:17:31 +00:00
needMatchingVectors = false;
Final round for line endings.
2015-06-26 22:58:36 +00:00
binOp = spv::OpLogicalAnd;
break;
case glslang::EOpInclusiveOr:
case glslang::EOpInclusiveOrAssign:
binOp = spv::OpBitwiseOr;
break;
case glslang::EOpLogicalOr:
glslang -> SPIR-V: smear scalars for integer (scalar * vector). OpVectorTimesScalar is only for floats.
2015-07-04 23:17:31 +00:00
needMatchingVectors = false;
Final round for line endings.
2015-06-26 22:58:36 +00:00
binOp = spv::OpLogicalOr;
break;
case glslang::EOpExclusiveOr:
case glslang::EOpExclusiveOrAssign:
binOp = spv::OpBitwiseXor;
break;
case glslang::EOpLogicalXor:
glslang -> SPIR-V: smear scalars for integer (scalar * vector). OpVectorTimesScalar is only for floats.
2015-07-04 23:17:31 +00:00
needMatchingVectors = false;
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
binOp = spv::OpLogicalNotEqual;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpLessThan:
case glslang::EOpGreaterThan:
case glslang::EOpLessThanEqual:
case glslang::EOpGreaterThanEqual:
case glslang::EOpEqual:
case glslang::EOpNotEqual:
case glslang::EOpVectorEqual:
case glslang::EOpVectorNotEqual:
comparison = true;
break;
default:
break;
}
SPV: Implement short circuiting of && and || when emitting SPIR-V.
2015-10-15 19:29:11 +00:00
// handle mapped binary operations (should be non-comparison)
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (binOp != spv::OpNop) {
SPV: Implement short circuiting of && and || when emitting SPIR-V.
2015-10-15 19:29:11 +00:00
assert(comparison == false);
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
if (builder.isMatrix(left) || builder.isMatrix(right) ||
builder.isCooperativeMatrix(left) || builder.isCooperativeMatrix(right))
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
return createBinaryMatrixOperation(binOp, decorations, typeId, left, right);
Final round for line endings.
2015-06-26 22:58:36 +00:00
// No matrix involved; make both operands be the same number of components, if needed
glslang -> SPIR-V: smear scalars for integer (scalar * vector). OpVectorTimesScalar is only for floats.
2015-07-04 23:17:31 +00:00
if (needMatchingVectors)
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
builder.promoteScalar(decorations.precision, left, right);
Final round for line endings.
2015-06-26 22:58:36 +00:00
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
spv::Id result = builder.createBinOp(binOp, typeId, left, right);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
decorations.addNoContraction(builder, result);
decorations.addNonUniform(builder, result);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
return builder.setPrecision(result, decorations.precision);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
if (! comparison)
return 0;
SPV: Implement short circuiting of && and || when emitting SPIR-V.
2015-10-15 19:29:11 +00:00
// Handle comparison instructions
Final round for line endings.
2015-06-26 22:58:36 +00:00
HLSL: Smear scalars to match vectors for relational operations. Yield a vector relational compare and a vector result.
2016-08-08 01:14:22 +00:00
if (reduceComparison && (op == glslang::EOpEqual || op == glslang::EOpNotEqual)
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
&& (builder.isVector(left) || builder.isMatrix(left) || builder.isAggregate(left))) {
spv::Id result = builder.createCompositeCompare(decorations.precision, left, right, op == glslang::EOpEqual);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
decorations.addNonUniform(builder, result);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
return result;
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
switch (op) {
case glslang::EOpLessThan:
if (isFloat)
binOp = spv::OpFOrdLessThan;
else if (isUnsigned)
binOp = spv::OpULessThan;
else
binOp = spv::OpSLessThan;
break;
case glslang::EOpGreaterThan:
if (isFloat)
binOp = spv::OpFOrdGreaterThan;
else if (isUnsigned)
binOp = spv::OpUGreaterThan;
else
binOp = spv::OpSGreaterThan;
break;
case glslang::EOpLessThanEqual:
if (isFloat)
binOp = spv::OpFOrdLessThanEqual;
else if (isUnsigned)
binOp = spv::OpULessThanEqual;
else
binOp = spv::OpSLessThanEqual;
break;
case glslang::EOpGreaterThanEqual:
if (isFloat)
binOp = spv::OpFOrdGreaterThanEqual;
else if (isUnsigned)
binOp = spv::OpUGreaterThanEqual;
else
binOp = spv::OpSGreaterThanEqual;
break;
case glslang::EOpEqual:
case glslang::EOpVectorEqual:
if (isFloat)
binOp = spv::OpFOrdEqual;
SPV: Use OpLogicalEqual/OpLogicalNotEqual for boolean type comparison.
2016-04-27 00:15:37 +00:00
else if (isBool)
binOp = spv::OpLogicalEqual;
Final round for line endings.
2015-06-26 22:58:36 +00:00
else
binOp = spv::OpIEqual;
break;
case glslang::EOpNotEqual:
case glslang::EOpVectorNotEqual:
if (isFloat)
binOp = spv::OpFOrdNotEqual;
SPV: Use OpLogicalEqual/OpLogicalNotEqual for boolean type comparison.
2016-04-27 00:15:37 +00:00
else if (isBool)
binOp = spv::OpLogicalNotEqual;
Final round for line endings.
2015-06-26 22:58:36 +00:00
else
binOp = spv::OpINotEqual;
break;
default:
break;
}
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
if (binOp != spv::OpNop) {
spv::Id result = builder.createBinOp(binOp, typeId, left, right);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
decorations.addNoContraction(builder, result);
decorations.addNonUniform(builder, result);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
return builder.setPrecision(result, decorations.precision);
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
return 0;
}
SPV: Implement all matrix operators {+,-,*,/} for {matrix,scalar,vector}.
2015-12-12 19:28:14 +00:00
//
// Translate AST matrix operation to SPV operation, already having SPV-based operands/types.
// These can be any of:
//
// matrix * scalar
// scalar * matrix
// matrix * matrix linear algebraic
// matrix * vector
// vector * matrix
// matrix * matrix componentwise
// matrix op matrix op in {+, -, /}
// matrix op scalar op in {+, -, /}
// scalar op matrix op in {+, -, /}
//
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id TGlslangToSpvTraverser::createBinaryMatrixOperation(spv::Op op, OpDecorations& decorations, spv::Id typeId,
spv::Id left, spv::Id right)
SPV: Implement all matrix operators {+,-,*,/} for {matrix,scalar,vector}.
2015-12-12 19:28:14 +00:00
{
bool firstClass = true;
// First, handle first-class matrix operations (* and matrix/scalar)
switch (op) {
case spv::OpFDiv:
if (builder.isMatrix(left) && builder.isScalar(right)) {
// turn matrix / scalar into a multiply...
Use the correct type for the constant for matrix/scalar division When a matrix is divided by a scalar it tries to take the reciprocal of the scalar to convert the operation into a multiply. However it was always doing this by making a 32-bit constant. If the scalar is a double then this would end up making an FDiv instruction with different types in the operands. This patch adds a helper method called makeFpConstant which makes a floating-point constant of the given type. The code to take the reciprocal now uses it to make the same type as the result. Fixes https://github.com/KhronosGroup/glslang/issues/1278
2018-03-13 09:57:59 +00:00
spv::Id resultType = builder.getTypeId(right);
right = builder.createBinOp(spv::OpFDiv, resultType, builder.makeFpConstant(resultType, 1.0), right);
SPV: Implement all matrix operators {+,-,*,/} for {matrix,scalar,vector}.
2015-12-12 19:28:14 +00:00
op = spv::OpMatrixTimesScalar;
} else
firstClass = false;
break;
case spv::OpMatrixTimesScalar:
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
if (builder.isMatrix(right) || builder.isCooperativeMatrix(right))
SPV: Implement all matrix operators {+,-,*,/} for {matrix,scalar,vector}.
2015-12-12 19:28:14 +00:00
std::swap(left, right);
assert(builder.isScalar(right));
break;
case spv::OpVectorTimesMatrix:
assert(builder.isVector(left));
assert(builder.isMatrix(right));
break;
case spv::OpMatrixTimesVector:
assert(builder.isMatrix(left));
assert(builder.isVector(right));
break;
case spv::OpMatrixTimesMatrix:
assert(builder.isMatrix(left));
assert(builder.isMatrix(right));
break;
default:
firstClass = false;
break;
}
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
if (builder.isCooperativeMatrix(left) || builder.isCooperativeMatrix(right))
firstClass = true;
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
if (firstClass) {
spv::Id result = builder.createBinOp(op, typeId, left, right);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
decorations.addNoContraction(builder, result);
decorations.addNonUniform(builder, result);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
return builder.setPrecision(result, decorations.precision);
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
}
SPV: Implement all matrix operators {+,-,*,/} for {matrix,scalar,vector}.
2015-12-12 19:28:14 +00:00
Add decompositions for some HLSL intrinsics.
2016-06-09 14:57:35 +00:00
// Handle component-wise +, -, *, %, and / for all combinations of type.
SPV: Implement all matrix operators {+,-,*,/} for {matrix,scalar,vector}.
2015-12-12 19:28:14 +00:00
// The result type of all of them is the same type as the (a) matrix operand.
// The algorithm is to:
// - break the matrix(es) into vectors
// - smear any scalar to a vector
// - do vector operations
// - make a matrix out the vector results
switch (op) {
case spv::OpFAdd:
case spv::OpFSub:
case spv::OpFDiv:
Add decompositions for some HLSL intrinsics.
2016-06-09 14:57:35 +00:00
case spv::OpFMod:
SPV: Implement all matrix operators {+,-,*,/} for {matrix,scalar,vector}.
2015-12-12 19:28:14 +00:00
case spv::OpFMul:
{
// one time set up...
bool leftMat = builder.isMatrix(left);
bool rightMat = builder.isMatrix(right);
unsigned int numCols = leftMat ? builder.getNumColumns(left) : builder.getNumColumns(right);
int numRows = leftMat ? builder.getNumRows(left) : builder.getNumRows(right);
spv::Id scalarType = builder.getScalarTypeId(typeId);
spv::Id vecType = builder.makeVectorType(scalarType, numRows);
std::vector<spv::Id> results;
spv::Id smearVec = spv::NoResult;
if (builder.isScalar(left))
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
smearVec = builder.smearScalar(decorations.precision, left, vecType);
SPV: Implement all matrix operators {+,-,*,/} for {matrix,scalar,vector}.
2015-12-12 19:28:14 +00:00
else if (builder.isScalar(right))
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
smearVec = builder.smearScalar(decorations.precision, right, vecType);
SPV: Implement all matrix operators {+,-,*,/} for {matrix,scalar,vector}.
2015-12-12 19:28:14 +00:00
// do each vector op
for (unsigned int c = 0; c < numCols; ++c) {
std::vector<unsigned int> indexes;
indexes.push_back(c);
spv::Id leftVec = leftMat ? builder.createCompositeExtract( left, vecType, indexes) : smearVec;
spv::Id rightVec = rightMat ? builder.createCompositeExtract(right, vecType, indexes) : smearVec;
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
spv::Id result = builder.createBinOp(op, vecType, leftVec, rightVec);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
decorations.addNoContraction(builder, result);
decorations.addNonUniform(builder, result);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
results.push_back(builder.setPrecision(result, decorations.precision));
SPV: Implement all matrix operators {+,-,*,/} for {matrix,scalar,vector}.
2015-12-12 19:28:14 +00:00
}
// put the pieces together
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id result = builder.setPrecision(builder.createCompositeConstruct(typeId, results), decorations.precision);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
decorations.addNonUniform(builder, result);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
return result;
SPV: Implement all matrix operators {+,-,*,/} for {matrix,scalar,vector}.
2015-12-12 19:28:14 +00:00
}
default:
assert(0);
return spv::NoResult;
}
}
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id TGlslangToSpvTraverser::createUnaryOperation(glslang::TOperator op, OpDecorations& decorations, spv::Id typeId,
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
spv::Id operand, glslang::TBasicType typeProxy, const spv::Builder::AccessChain::CoherentFlags &lvalueCoherentFlags)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
spv::Op unaryOp = spv::OpNop;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
int extBuiltins = -1;
Final round for line endings.
2015-06-26 22:58:36 +00:00
int libCall = -1;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
bool isUnsigned = isTypeUnsignedInt(typeProxy);
bool isFloat = isTypeFloat(typeProxy);
Final round for line endings.
2015-06-26 22:58:36 +00:00
switch (op) {
case glslang::EOpNegative:
SPV: Add unary-matrix operations, operating at vector level.
2016-01-20 18:19:27 +00:00
if (isFloat) {
Final round for line endings.
2015-06-26 22:58:36 +00:00
unaryOp = spv::OpFNegate;
SPV: Add unary-matrix operations, operating at vector level.
2016-01-20 18:19:27 +00:00
if (builder.isMatrixType(typeId))
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
return createUnaryMatrixOperation(unaryOp, decorations, typeId, operand, typeProxy);
SPV: Add unary-matrix operations, operating at vector level.
2016-01-20 18:19:27 +00:00
} else
Final round for line endings.
2015-06-26 22:58:36 +00:00
unaryOp = spv::OpSNegate;
break;
case glslang::EOpLogicalNot:
case glslang::EOpVectorLogicalNot:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
unaryOp = spv::OpLogicalNot;
break;
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpBitwiseNot:
unaryOp = spv::OpNot;
break;
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpDeterminant:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Determinant;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpMatrixInverse:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450MatrixInverse;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpTranspose:
unaryOp = spv::OpTranspose;
break;
case glslang::EOpRadians:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Radians;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpDegrees:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Degrees;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpSin:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Sin;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpCos:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Cos;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpTan:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Tan;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpAcos:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Acos;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpAsin:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Asin;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpAtan:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Atan;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpAcosh:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Acosh;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpAsinh:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Asinh;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpAtanh:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Atanh;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpTanh:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Tanh;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpCosh:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Cosh;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpSinh:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Sinh;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpLength:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Length;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpNormalize:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Normalize;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpExp:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Exp;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpLog:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Log;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpExp2:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Exp2;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpLog2:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Log2;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpSqrt:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Sqrt;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpInverseSqrt:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450InverseSqrt;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpFloor:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Floor;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpTrunc:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Trunc;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpRound:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Round;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpRoundEven:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450RoundEven;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpCeil:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Ceil;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpFract:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Fract;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpIsNan:
unaryOp = spv::OpIsNan;
break;
case glslang::EOpIsInf:
unaryOp = spv::OpIsInf;
break;
Add decompositions for some HLSL intrinsics.
2016-06-09 14:57:35 +00:00
case glslang::EOpIsFinite:
unaryOp = spv::OpIsFinite;
break;
Final round for line endings.
2015-06-26 22:58:36 +00:00
SPV: Implement GLSL built-in bitcast functions.
2015-12-15 08:03:10 +00:00
case glslang::EOpFloatBitsToInt:
case glslang::EOpFloatBitsToUint:
case glslang::EOpIntBitsToFloat:
case glslang::EOpUintBitsToFloat:
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
case glslang::EOpDoubleBitsToInt64:
case glslang::EOpDoubleBitsToUint64:
case glslang::EOpInt64BitsToDouble:
case glslang::EOpUint64BitsToDouble:
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
case glslang::EOpFloat16BitsToInt16:
case glslang::EOpFloat16BitsToUint16:
case glslang::EOpInt16BitsToFloat16:
case glslang::EOpUint16BitsToFloat16:
SPV: Implement GLSL built-in bitcast functions.
2015-12-15 08:03:10 +00:00
unaryOp = spv::OpBitcast;
break;
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpPackSnorm2x16:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450PackSnorm2x16;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpUnpackSnorm2x16:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450UnpackSnorm2x16;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpPackUnorm2x16:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450PackUnorm2x16;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpUnpackUnorm2x16:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450UnpackUnorm2x16;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpPackHalf2x16:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450PackHalf2x16;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpUnpackHalf2x16:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450UnpackHalf2x16;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
case glslang::EOpPackSnorm4x8:
libCall = spv::GLSLstd450PackSnorm4x8;
break;
case glslang::EOpUnpackSnorm4x8:
libCall = spv::GLSLstd450UnpackSnorm4x8;
break;
case glslang::EOpPackUnorm4x8:
libCall = spv::GLSLstd450PackUnorm4x8;
break;
case glslang::EOpUnpackUnorm4x8:
libCall = spv::GLSLstd450UnpackUnorm4x8;
break;
case glslang::EOpPackDouble2x32:
libCall = spv::GLSLstd450PackDouble2x32;
break;
case glslang::EOpUnpackDouble2x32:
libCall = spv::GLSLstd450UnpackDouble2x32;
break;
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
case glslang::EOpPackInt2x32:
case glslang::EOpUnpackInt2x32:
case glslang::EOpPackUint2x32:
case glslang::EOpUnpackUint2x32:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpPack16:
case glslang::EOpPack32:
case glslang::EOpPack64:
case glslang::EOpUnpack32:
case glslang::EOpUnpack16:
case glslang::EOpUnpack8:
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
case glslang::EOpPackInt2x16:
case glslang::EOpUnpackInt2x16:
case glslang::EOpPackUint2x16:
case glslang::EOpUnpackUint2x16:
case glslang::EOpPackInt4x16:
case glslang::EOpUnpackInt4x16:
case glslang::EOpPackUint4x16:
case glslang::EOpUnpackUint4x16:
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
case glslang::EOpPackFloat2x16:
case glslang::EOpUnpackFloat2x16:
unaryOp = spv::OpBitcast;
break;
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpDPdx:
unaryOp = spv::OpDPdx;
break;
case glslang::EOpDPdy:
unaryOp = spv::OpDPdy;
break;
case glslang::EOpFwidth:
unaryOp = spv::OpFwidth;
break;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
case glslang::EOpAny:
unaryOp = spv::OpAny;
break;
case glslang::EOpAll:
unaryOp = spv::OpAll;
break;
case glslang::EOpAbs:
if (isFloat)
libCall = spv::GLSLstd450FAbs;
else
libCall = spv::GLSLstd450SAbs;
break;
case glslang::EOpSign:
if (isFloat)
libCall = spv::GLSLstd450FSign;
else
libCall = spv::GLSLstd450SSign;
break;
#ifndef GLSLANG_WEB
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpDPdxFine:
unaryOp = spv::OpDPdxFine;
break;
case glslang::EOpDPdyFine:
unaryOp = spv::OpDPdyFine;
break;
case glslang::EOpFwidthFine:
unaryOp = spv::OpFwidthFine;
break;
case glslang::EOpDPdxCoarse:
unaryOp = spv::OpDPdxCoarse;
break;
case glslang::EOpDPdyCoarse:
unaryOp = spv::OpDPdyCoarse;
break;
case glslang::EOpFwidthCoarse:
unaryOp = spv::OpFwidthCoarse;
break;
SPV: Implement GLSL interpolation built-in functions.
2015-12-08 09:12:09 +00:00
case glslang::EOpInterpolateAtCentroid:
Fix issue #1376: SPV_AMD_gpu_shader_half_float is not emitted under vulkan1.1 Interpolation functions, using 16-bit interpolants, should always emit this extension string regardless of vulkan enviornment version.
2018-05-17 05:51:28 +00:00
if (typeProxy == glslang::EbtFloat16)
builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float);
SPV: Implement GLSL interpolation built-in functions.
2015-12-08 09:12:09 +00:00
libCall = spv::GLSLstd450InterpolateAtCentroid;
break;
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
case glslang::EOpAtomicCounterIncrement:
case glslang::EOpAtomicCounterDecrement:
case glslang::EOpAtomicCounter:
{
// Handle all of the atomics in one place, in createAtomicOperation()
std::vector<spv::Id> operands;
operands.push_back(operand);
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
return createAtomicOperation(op, decorations.precision, typeId, operands, typeProxy, lvalueCoherentFlags);
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
}
case glslang::EOpBitFieldReverse:
unaryOp = spv::OpBitReverse;
break;
case glslang::EOpBitCount:
unaryOp = spv::OpBitCount;
break;
case glslang::EOpFindLSB:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
libCall = spv::GLSLstd450FindILsb;
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
break;
case glslang::EOpFindMSB:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
if (isUnsigned)
libCall = spv::GLSLstd450FindUMsb;
else
libCall = spv::GLSLstd450FindSMsb;
Backward incompatible: Turn on PureOperatorBuiltins: use only enum-based built-in functions in the AST. If this breaks your AST consumer, best is to modify it to test against the enum values instead of doing string comparisons on built-in function names. This is the reason the change was made. If you need the old behavior, you should be able to get it back by changing PureOperatorBuiltins to be false instead of true. This path will work for a while, but is marked deprecated. Also, the old behavior is tagged as release 2.4.
2015-08-19 19:34:18 +00:00
break;
Implement the extension GL_ARB_shader_ballot Add new built-in variables and functions to the parser (SPIR-V tokens are missing).
2016-04-14 08:53:07 +00:00
case glslang::EOpBallot:
case glslang::EOpReadFirstInvocation:
Implement the extension GL_ARB_shader_group_vote.
2016-05-05 12:38:33 +00:00
case glslang::EOpAnyInvocation:
case glslang::EOpAllInvocations:
case glslang::EOpAllInvocationsEqual:
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
case glslang::EOpMinInvocations:
case glslang::EOpMaxInvocations:
case glslang::EOpAddInvocations:
case glslang::EOpMinInvocationsNonUniform:
case glslang::EOpMaxInvocationsNonUniform:
case glslang::EOpAddInvocationsNonUniform:
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
case glslang::EOpMinInvocationsInclusiveScan:
case glslang::EOpMaxInvocationsInclusiveScan:
case glslang::EOpAddInvocationsInclusiveScan:
case glslang::EOpMinInvocationsInclusiveScanNonUniform:
case glslang::EOpMaxInvocationsInclusiveScanNonUniform:
case glslang::EOpAddInvocationsInclusiveScanNonUniform:
case glslang::EOpMinInvocationsExclusiveScan:
case glslang::EOpMaxInvocationsExclusiveScan:
case glslang::EOpAddInvocationsExclusiveScan:
case glslang::EOpMinInvocationsExclusiveScanNonUniform:
case glslang::EOpMaxInvocationsExclusiveScanNonUniform:
case glslang::EOpAddInvocationsExclusiveScanNonUniform:
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
{
std::vector<spv::Id> operands;
operands.push_back(operand);
return createInvocationsOperation(op, typeId, operands, typeProxy);
}
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpSubgroupAll:
case glslang::EOpSubgroupAny:
case glslang::EOpSubgroupAllEqual:
case glslang::EOpSubgroupBroadcastFirst:
case glslang::EOpSubgroupBallot:
case glslang::EOpSubgroupInverseBallot:
case glslang::EOpSubgroupBallotBitCount:
case glslang::EOpSubgroupBallotInclusiveBitCount:
case glslang::EOpSubgroupBallotExclusiveBitCount:
case glslang::EOpSubgroupBallotFindLSB:
case glslang::EOpSubgroupBallotFindMSB:
case glslang::EOpSubgroupAdd:
case glslang::EOpSubgroupMul:
case glslang::EOpSubgroupMin:
case glslang::EOpSubgroupMax:
case glslang::EOpSubgroupAnd:
case glslang::EOpSubgroupOr:
case glslang::EOpSubgroupXor:
case glslang::EOpSubgroupInclusiveAdd:
case glslang::EOpSubgroupInclusiveMul:
case glslang::EOpSubgroupInclusiveMin:
case glslang::EOpSubgroupInclusiveMax:
case glslang::EOpSubgroupInclusiveAnd:
case glslang::EOpSubgroupInclusiveOr:
case glslang::EOpSubgroupInclusiveXor:
case glslang::EOpSubgroupExclusiveAdd:
case glslang::EOpSubgroupExclusiveMul:
case glslang::EOpSubgroupExclusiveMin:
case glslang::EOpSubgroupExclusiveMax:
case glslang::EOpSubgroupExclusiveAnd:
case glslang::EOpSubgroupExclusiveOr:
case glslang::EOpSubgroupExclusiveXor:
case glslang::EOpSubgroupQuadSwapHorizontal:
case glslang::EOpSubgroupQuadSwapVertical:
case glslang::EOpSubgroupQuadSwapDiagonal: {
std::vector<spv::Id> operands;
operands.push_back(operand);
return createSubgroupOperation(op, typeId, operands, typeProxy);
}
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
case glslang::EOpMbcnt:
extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot);
libCall = spv::MbcntAMD;
break;
case glslang::EOpCubeFaceIndex:
extBuiltins = getExtBuiltins(spv::E_SPV_AMD_gcn_shader);
libCall = spv::CubeFaceIndexAMD;
break;
case glslang::EOpCubeFaceCoord:
extBuiltins = getExtBuiltins(spv::E_SPV_AMD_gcn_shader);
libCall = spv::CubeFaceCoordAMD;
break;
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
case glslang::EOpSubgroupPartition:
unaryOp = spv::OpGroupNonUniformPartitionNV;
break;
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
case glslang::EOpConstructReference:
unaryOp = spv::OpBitcast;
break;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#endif
For nonuniformEXT constructor, make a copy of the node to decorate
2019-05-08 15:24:46 +00:00
case glslang::EOpCopyObject:
unaryOp = spv::OpCopyObject;
break;
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
return 0;
}
spv::Id id;
if (libCall >= 0) {
std::vector<spv::Id> args;
args.push_back(operand);
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
id = builder.createBuiltinCall(typeId, extBuiltins >= 0 ? extBuiltins : stdBuiltins, libCall, args);
Implement the extension GL_ARB_shader_group_vote.
2016-05-05 12:38:33 +00:00
} else {
SPV: Non-functional: modularize createInvocationsOperation() from the previous commit.
2016-05-05 22:45:40 +00:00
id = builder.createUnaryOp(unaryOp, typeId, operand);
Implement the extension GL_ARB_shader_group_vote.
2016-05-05 12:38:33 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
decorations.addNoContraction(builder, id);
decorations.addNonUniform(builder, id);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
return builder.setPrecision(id, decorations.precision);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
SPV: Add unary-matrix operations, operating at vector level.
2016-01-20 18:19:27 +00:00
// Create a unary operation on a matrix
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id TGlslangToSpvTraverser::createUnaryMatrixOperation(spv::Op op, OpDecorations& decorations, spv::Id typeId,
spv::Id operand, glslang::TBasicType /* typeProxy */)
SPV: Add unary-matrix operations, operating at vector level.
2016-01-20 18:19:27 +00:00
{
// Handle unary operations vector by vector.
// The result type is the same type as the original type.
// The algorithm is to:
// - break the matrix into vectors
// - apply the operation to each vector
// - make a matrix out the vector results
// get the types sorted out
int numCols = builder.getNumColumns(operand);
int numRows = builder.getNumRows(operand);
SPV: Correct an issue in createUnaryMatrixOperation(). Type of the source is not necessarily the same as that of the destination.
2016-05-17 10:57:18 +00:00
spv::Id srcVecType = builder.makeVectorType(builder.getScalarTypeId(builder.getTypeId(operand)), numRows);
spv::Id destVecType = builder.makeVectorType(builder.getScalarTypeId(typeId), numRows);
SPV: Add unary-matrix operations, operating at vector level.
2016-01-20 18:19:27 +00:00
std::vector<spv::Id> results;
// do each vector op
for (int c = 0; c < numCols; ++c) {
std::vector<unsigned int> indexes;
indexes.push_back(c);
SPV: Correct an issue in createUnaryMatrixOperation(). Type of the source is not necessarily the same as that of the destination.
2016-05-17 10:57:18 +00:00
spv::Id srcVec = builder.createCompositeExtract(operand, srcVecType, indexes);
spv::Id destVec = builder.createUnaryOp(op, destVecType, srcVec);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
decorations.addNoContraction(builder, destVec);
decorations.addNonUniform(builder, destVec);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
results.push_back(builder.setPrecision(destVec, decorations.precision));
SPV: Add unary-matrix operations, operating at vector level.
2016-01-20 18:19:27 +00:00
}
// put the pieces together
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id result = builder.setPrecision(builder.createCompositeConstruct(typeId, results), decorations.precision);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
decorations.addNonUniform(builder, result);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
return result;
SPV: Add unary-matrix operations, operating at vector level.
2016-01-20 18:19:27 +00:00
}
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
// For converting integers where both the bitwidth and the signedness could
// change, but only do the width change here. The caller is still responsible
// for the signedness conversion.
spv::Id TGlslangToSpvTraverser::createIntWidthConversion(glslang::TOperator op, spv::Id operand, int vectorSize)
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
{
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
// Get the result type width, based on the type to convert to.
int width = 32;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
switch(op) {
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
case glslang::EOpConvInt16ToUint8:
case glslang::EOpConvIntToUint8:
case glslang::EOpConvInt64ToUint8:
case glslang::EOpConvUint16ToInt8:
case glslang::EOpConvUintToInt8:
case glslang::EOpConvUint64ToInt8:
width = 8;
break;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvInt8ToUint16:
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
case glslang::EOpConvIntToUint16:
case glslang::EOpConvInt64ToUint16:
case glslang::EOpConvUint8ToInt16:
case glslang::EOpConvUintToInt16:
case glslang::EOpConvUint64ToInt16:
width = 16;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
case glslang::EOpConvInt8ToUint:
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
case glslang::EOpConvInt16ToUint:
case glslang::EOpConvInt64ToUint:
case glslang::EOpConvUint8ToInt:
case glslang::EOpConvUint16ToInt:
case glslang::EOpConvUint64ToInt:
width = 32;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
case glslang::EOpConvInt8ToUint64:
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
case glslang::EOpConvInt16ToUint64:
case glslang::EOpConvIntToUint64:
case glslang::EOpConvUint8ToInt64:
case glslang::EOpConvUint16ToInt64:
case glslang::EOpConvUintToInt64:
width = 64;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
default:
assert(false && "Default missing");
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
}
// Get the conversion operation and result type,
// based on the target width, but the source type.
spv::Id type = spv::NoType;
spv::Op convOp = spv::OpNop;
switch(op) {
case glslang::EOpConvInt8ToUint16:
case glslang::EOpConvInt8ToUint:
case glslang::EOpConvInt8ToUint64:
case glslang::EOpConvInt16ToUint8:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvInt16ToUint:
case glslang::EOpConvInt16ToUint64:
case glslang::EOpConvIntToUint8:
case glslang::EOpConvIntToUint16:
case glslang::EOpConvIntToUint64:
case glslang::EOpConvInt64ToUint8:
case glslang::EOpConvInt64ToUint16:
case glslang::EOpConvInt64ToUint:
convOp = spv::OpSConvert;
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
type = builder.makeIntType(width);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
default:
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
convOp = spv::OpUConvert;
type = builder.makeUintType(width);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
}
if (vectorSize > 0)
type = builder.makeVectorType(type, vectorSize);
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
return builder.createUnaryOp(convOp, type, operand);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
}
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
spv::Id TGlslangToSpvTraverser::createConversion(glslang::TOperator op, OpDecorations& decorations, spv::Id destType,
spv::Id operand, glslang::TBasicType typeProxy)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
spv::Op convOp = spv::OpNop;
spv::Id zero = 0;
spv::Id one = 0;
int vectorSize = builder.isVectorType(destType) ? builder.getNumTypeComponents(destType) : 0;
switch (op) {
case glslang::EOpConvIntToBool:
case glslang::EOpConvUintToBool:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
zero = builder.makeUintConstant(0);
zero = makeSmearedConstant(zero, vectorSize);
return builder.createBinOp(spv::OpINotEqual, destType, operand, zero);
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpConvFloatToBool:
zero = builder.makeFloatConstant(0.0F);
zero = makeSmearedConstant(zero, vectorSize);
return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero);
case glslang::EOpConvBoolToFloat:
convOp = spv::OpSelect;
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
zero = builder.makeFloatConstant(0.0F);
one = builder.makeFloatConstant(1.0F);
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpConvBoolToInt:
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
case glslang::EOpConvBoolToInt64:
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
if (op == glslang::EOpConvBoolToInt64) {
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
zero = builder.makeInt64Constant(0);
one = builder.makeInt64Constant(1);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
} else
#endif
{
zero = builder.makeIntConstant(0);
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
one = builder.makeIntConstant(1);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
}
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
convOp = spv::OpSelect;
break;
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpConvBoolToUint:
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
case glslang::EOpConvBoolToUint64:
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
if (op == glslang::EOpConvBoolToUint64) {
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
zero = builder.makeUint64Constant(0);
one = builder.makeUint64Constant(1);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
} else
#endif
{
zero = builder.makeUintConstant(0);
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
one = builder.makeUintConstant(1);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
}
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
convOp = spv::OpSelect;
break;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvInt8ToFloat16:
case glslang::EOpConvInt8ToFloat:
case glslang::EOpConvInt8ToDouble:
case glslang::EOpConvInt16ToFloat16:
case glslang::EOpConvInt16ToFloat:
case glslang::EOpConvInt16ToDouble:
case glslang::EOpConvIntToFloat16:
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpConvIntToFloat:
case glslang::EOpConvIntToDouble:
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
case glslang::EOpConvInt64ToFloat:
case glslang::EOpConvInt64ToDouble:
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
case glslang::EOpConvInt64ToFloat16:
Final round for line endings.
2015-06-26 22:58:36 +00:00
convOp = spv::OpConvertSToF;
break;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvUint8ToFloat16:
case glslang::EOpConvUint8ToFloat:
case glslang::EOpConvUint8ToDouble:
case glslang::EOpConvUint16ToFloat16:
case glslang::EOpConvUint16ToFloat:
case glslang::EOpConvUint16ToDouble:
case glslang::EOpConvUintToFloat16:
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpConvUintToFloat:
case glslang::EOpConvUintToDouble:
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
case glslang::EOpConvUint64ToFloat:
case glslang::EOpConvUint64ToDouble:
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
case glslang::EOpConvUint64ToFloat16:
Final round for line endings.
2015-06-26 22:58:36 +00:00
convOp = spv::OpConvertUToF;
break;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvFloat16ToInt8:
case glslang::EOpConvFloatToInt8:
case glslang::EOpConvDoubleToInt8:
case glslang::EOpConvFloat16ToInt16:
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
case glslang::EOpConvFloatToInt16:
case glslang::EOpConvDoubleToInt16:
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
case glslang::EOpConvFloat16ToInt:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvFloatToInt:
case glslang::EOpConvDoubleToInt:
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
case glslang::EOpConvFloat16ToInt64:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvFloatToInt64:
case glslang::EOpConvDoubleToInt64:
Final round for line endings.
2015-06-26 22:58:36 +00:00
convOp = spv::OpConvertFToS;
break;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvUint8ToInt8:
case glslang::EOpConvInt8ToUint8:
case glslang::EOpConvUint16ToInt16:
case glslang::EOpConvInt16ToUint16:
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpConvUintToInt:
case glslang::EOpConvIntToUint:
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
case glslang::EOpConvUint64ToInt64:
case glslang::EOpConvInt64ToUint64:
SpecOp bool->uint/int and uint<->int conversion Bool -> uint/int with OpSpecConstantOp OpSelect instruction. uint <-> int conversion with OpSpecConstantOp OpIAdd instruction. Note, implicit conversion: `const uint = an_int_spec_constant` is not supported. Explicit type casting is required: `const uint = uint(an_int_spec_constant)`
2016-04-07 19:40:27 +00:00
if (builder.isInSpecConstCodeGenMode()) {
// Build zero scalar or vector for OpIAdd.
Web: Remove unnecessary GLSL numeric types, and some collateral. This saves another 40K of x86 binary, which is about 13% of the target size.
2019-08-08 16:35:51 +00:00
#ifndef GLSLANG_WEB
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
if(op == glslang::EOpConvUint8ToInt8 || op == glslang::EOpConvInt8ToUint8) {
zero = builder.makeUint8Constant(0);
} else if (op == glslang::EOpConvUint16ToInt16 || op == glslang::EOpConvInt16ToUint16) {
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
zero = builder.makeUint16Constant(0);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
} else if (op == glslang::EOpConvUint64ToInt64 || op == glslang::EOpConvInt64ToUint64) {
zero = builder.makeUint64Constant(0);
Web: Remove unnecessary GLSL numeric types, and some collateral. This saves another 40K of x86 binary, which is about 13% of the target size.
2019-08-08 16:35:51 +00:00
} else
#endif
{
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
zero = builder.makeUintConstant(0);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
}
Refine the code and address comments
2016-04-13 03:16:20 +00:00
zero = makeSmearedConstant(zero, vectorSize);
SpecOp bool->uint/int and uint<->int conversion Bool -> uint/int with OpSpecConstantOp OpSelect instruction. uint <-> int conversion with OpSpecConstantOp OpIAdd instruction. Note, implicit conversion: `const uint = an_int_spec_constant` is not supported. Explicit type casting is required: `const uint = uint(an_int_spec_constant)`
2016-04-07 19:40:27 +00:00
// Use OpIAdd, instead of OpBitcast to do the conversion when
// generating for OpSpecConstantOp instruction.
return builder.createBinOp(spv::OpIAdd, destType, operand, zero);
}
// For normal run-time conversion instruction, use OpBitcast.
Final round for line endings.
2015-06-26 22:58:36 +00:00
convOp = spv::OpBitcast;
break;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvFloat16ToUint8:
case glslang::EOpConvFloatToUint8:
case glslang::EOpConvDoubleToUint8:
case glslang::EOpConvFloat16ToUint16:
case glslang::EOpConvFloatToUint16:
case glslang::EOpConvDoubleToUint16:
case glslang::EOpConvFloat16ToUint:
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EOpConvFloatToUint:
case glslang::EOpConvDoubleToUint:
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
case glslang::EOpConvFloatToUint64:
case glslang::EOpConvDoubleToUint64:
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
case glslang::EOpConvFloat16ToUint64:
Final round for line endings.
2015-06-26 22:58:36 +00:00
convOp = spv::OpConvertFToU;
break;
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
Web: Remove unnecessary GLSL numeric types, and some collateral. This saves another 40K of x86 binary, which is about 13% of the target size.
2019-08-08 16:35:51 +00:00
#ifndef GLSLANG_WEB
case glslang::EOpConvInt8ToBool:
case glslang::EOpConvUint8ToBool:
zero = builder.makeUint8Constant(0);
zero = makeSmearedConstant(zero, vectorSize);
return builder.createBinOp(spv::OpINotEqual, destType, operand, zero);
case glslang::EOpConvInt16ToBool:
case glslang::EOpConvUint16ToBool:
zero = builder.makeUint16Constant(0);
zero = makeSmearedConstant(zero, vectorSize);
return builder.createBinOp(spv::OpINotEqual, destType, operand, zero);
case glslang::EOpConvInt64ToBool:
case glslang::EOpConvUint64ToBool:
zero = builder.makeUint64Constant(0);
zero = makeSmearedConstant(zero, vectorSize);
return builder.createBinOp(spv::OpINotEqual, destType, operand, zero);
case glslang::EOpConvDoubleToBool:
zero = builder.makeDoubleConstant(0.0);
zero = makeSmearedConstant(zero, vectorSize);
return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero);
case glslang::EOpConvFloat16ToBool:
zero = builder.makeFloat16Constant(0.0F);
zero = makeSmearedConstant(zero, vectorSize);
return builder.createBinOp(spv::OpFOrdNotEqual, destType, operand, zero);
case glslang::EOpConvBoolToDouble:
convOp = spv::OpSelect;
zero = builder.makeDoubleConstant(0.0);
one = builder.makeDoubleConstant(1.0);
break;
case glslang::EOpConvBoolToFloat16:
convOp = spv::OpSelect;
zero = builder.makeFloat16Constant(0.0F);
one = builder.makeFloat16Constant(1.0F);
break;
case glslang::EOpConvBoolToInt8:
zero = builder.makeInt8Constant(0);
one = builder.makeInt8Constant(1);
convOp = spv::OpSelect;
break;
case glslang::EOpConvBoolToUint8:
zero = builder.makeUint8Constant(0);
one = builder.makeUint8Constant(1);
convOp = spv::OpSelect;
break;
case glslang::EOpConvBoolToInt16:
zero = builder.makeInt16Constant(0);
one = builder.makeInt16Constant(1);
convOp = spv::OpSelect;
break;
case glslang::EOpConvBoolToUint16:
zero = builder.makeUint16Constant(0);
one = builder.makeUint16Constant(1);
convOp = spv::OpSelect;
break;
case glslang::EOpConvDoubleToFloat:
case glslang::EOpConvFloatToDouble:
case glslang::EOpConvDoubleToFloat16:
case glslang::EOpConvFloat16ToDouble:
case glslang::EOpConvFloatToFloat16:
case glslang::EOpConvFloat16ToFloat:
convOp = spv::OpFConvert;
if (builder.isMatrixType(destType))
return createUnaryMatrixOperation(convOp, decorations, destType, operand, typeProxy);
break;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvInt8ToInt16:
case glslang::EOpConvInt8ToInt:
case glslang::EOpConvInt8ToInt64:
case glslang::EOpConvInt16ToInt8:
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
case glslang::EOpConvInt16ToInt:
case glslang::EOpConvInt16ToInt64:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvIntToInt8:
case glslang::EOpConvIntToInt16:
case glslang::EOpConvIntToInt64:
case glslang::EOpConvInt64ToInt8:
case glslang::EOpConvInt64ToInt16:
case glslang::EOpConvInt64ToInt:
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
convOp = spv::OpSConvert;
break;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvUint8ToUint16:
case glslang::EOpConvUint8ToUint:
case glslang::EOpConvUint8ToUint64:
case glslang::EOpConvUint16ToUint8:
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
case glslang::EOpConvUint16ToUint:
case glslang::EOpConvUint16ToUint64:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvUintToUint8:
case glslang::EOpConvUintToUint16:
case glslang::EOpConvUintToUint64:
case glslang::EOpConvUint64ToUint8:
case glslang::EOpConvUint64ToUint16:
case glslang::EOpConvUint64ToUint:
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
convOp = spv::OpUConvert;
break;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvInt8ToUint16:
case glslang::EOpConvInt8ToUint:
case glslang::EOpConvInt8ToUint64:
case glslang::EOpConvInt16ToUint8:
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
case glslang::EOpConvInt16ToUint:
case glslang::EOpConvInt16ToUint64:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvIntToUint8:
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
case glslang::EOpConvIntToUint16:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvIntToUint64:
case glslang::EOpConvInt64ToUint8:
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
case glslang::EOpConvInt64ToUint16:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpConvInt64ToUint:
case glslang::EOpConvUint8ToInt16:
case glslang::EOpConvUint8ToInt:
case glslang::EOpConvUint8ToInt64:
case glslang::EOpConvUint16ToInt8:
case glslang::EOpConvUint16ToInt:
case glslang::EOpConvUint16ToInt64:
case glslang::EOpConvUintToInt8:
case glslang::EOpConvUintToInt16:
case glslang::EOpConvUintToInt64:
case glslang::EOpConvUint64ToInt8:
case glslang::EOpConvUint64ToInt16:
case glslang::EOpConvUint64ToInt:
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
// OpSConvert/OpUConvert + OpBitCast
Fix #1360: uint->int width conversions must still be typed as uint.
2018-06-05 01:11:25 +00:00
operand = createIntWidthConversion(op, operand, vectorSize);
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
if (builder.isInSpecConstCodeGenMode()) {
// Build zero scalar or vector for OpIAdd.
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
switch(op) {
case glslang::EOpConvInt16ToUint8:
case glslang::EOpConvIntToUint8:
case glslang::EOpConvInt64ToUint8:
case glslang::EOpConvUint16ToInt8:
case glslang::EOpConvUintToInt8:
case glslang::EOpConvUint64ToInt8:
zero = builder.makeUint8Constant(0);
break;
case glslang::EOpConvInt8ToUint16:
case glslang::EOpConvIntToUint16:
case glslang::EOpConvInt64ToUint16:
case glslang::EOpConvUint8ToInt16:
case glslang::EOpConvUintToInt16:
case glslang::EOpConvUint64ToInt16:
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
zero = builder.makeUint16Constant(0);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
case glslang::EOpConvInt8ToUint:
case glslang::EOpConvInt16ToUint:
case glslang::EOpConvInt64ToUint:
case glslang::EOpConvUint8ToInt:
case glslang::EOpConvUint16ToInt:
case glslang::EOpConvUint64ToInt:
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
zero = builder.makeUintConstant(0);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
case glslang::EOpConvInt8ToUint64:
case glslang::EOpConvInt16ToUint64:
case glslang::EOpConvIntToUint64:
case glslang::EOpConvUint8ToInt64:
case glslang::EOpConvUint16ToInt64:
case glslang::EOpConvUintToInt64:
Implement extension GL_AMD_gpu_shader_int16 - Add int16 types (int16_t, uint16_t, i16vec, u16vec). - Add int16 support to GLSL operators. - Add int16 type conversions (to int16, from int16). - Add int16 built-in functions.
2017-03-24 05:41:14 +00:00
zero = builder.makeUint64Constant(0);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
default:
assert(false && "Default missing");
break;
}
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
zero = makeSmearedConstant(zero, vectorSize);
// Use OpIAdd, instead of OpBitcast to do the conversion when
// generating for OpSpecConstantOp instruction.
return builder.createBinOp(spv::OpIAdd, destType, operand, zero);
}
// For normal run-time conversion instruction, use OpBitcast.
convOp = spv::OpBitcast;
break;
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
case glslang::EOpConvUint64ToPtr:
convOp = spv::OpConvertUToPtr;
break;
case glslang::EOpConvPtrToUint64:
convOp = spv::OpConvertPtrToU;
break;
Web: Remove unnecessary GLSL numeric types, and some collateral. This saves another 40K of x86 binary, which is about 13% of the target size.
2019-08-08 16:35:51 +00:00
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
break;
}
spv::Id result = 0;
if (convOp == spv::OpNop)
return result;
if (convOp == spv::OpSelect) {
zero = makeSmearedConstant(zero, vectorSize);
one = makeSmearedConstant(one, vectorSize);
result = builder.createTriOp(convOp, destType, operand, one, zero);
} else
result = builder.createUnaryOp(convOp, destType, operand);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
result = builder.setPrecision(result, decorations.precision);
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
decorations.addNonUniform(builder, result);
Non-functional. Rationalizations enabling future generalizations: - Use much simpler method to update implicit array sizes. The previous overly complicated method was error prone. - Rationalize all use of unsized arrays. - Combine decorations when generating SPIR-V, to simplify adding extensions.
2018-03-29 00:01:20 +00:00
return result;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
spv::Id TGlslangToSpvTraverser::makeSmearedConstant(spv::Id constant, int vectorSize)
{
if (vectorSize == 0)
return constant;
spv::Id vectorTypeId = builder.makeVectorType(builder.getTypeId(constant), vectorSize);
std::vector<spv::Id> components;
for (int c = 0; c < vectorSize; ++c)
components.push_back(constant);
return builder.makeCompositeConstant(vectorTypeId, components);
}
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
// For glslang ops that map to SPV atomic opCodes
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
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)
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
{
spv::Op opCode = spv::OpNop;
switch (op) {
case glslang::EOpAtomicAdd:
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
case glslang::EOpImageAtomicAdd:
GLSL 4.6: Implement atomic counter ops and SPV_KHR_shader_atomic_counter_ops.
2017-07-23 22:08:26 +00:00
case glslang::EOpAtomicCounterAdd:
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
opCode = spv::OpAtomicIAdd;
break;
GLSL 4.6: Implement atomic counter ops and SPV_KHR_shader_atomic_counter_ops.
2017-07-23 22:08:26 +00:00
case glslang::EOpAtomicCounterSubtract:
opCode = spv::OpAtomicISub;
break;
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
case glslang::EOpAtomicMin:
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
case glslang::EOpImageAtomicMin:
GLSL 4.6: Implement atomic counter ops and SPV_KHR_shader_atomic_counter_ops.
2017-07-23 22:08:26 +00:00
case glslang::EOpAtomicCounterMin:
Implement extension GL_NV_shader_atomic_int64
2017-09-26 07:42:56 +00:00
opCode = (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) ? spv::OpAtomicUMin : spv::OpAtomicSMin;
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
break;
case glslang::EOpAtomicMax:
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
case glslang::EOpImageAtomicMax:
GLSL 4.6: Implement atomic counter ops and SPV_KHR_shader_atomic_counter_ops.
2017-07-23 22:08:26 +00:00
case glslang::EOpAtomicCounterMax:
Implement extension GL_NV_shader_atomic_int64
2017-09-26 07:42:56 +00:00
opCode = (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) ? spv::OpAtomicUMax : spv::OpAtomicSMax;
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
break;
case glslang::EOpAtomicAnd:
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
case glslang::EOpImageAtomicAnd:
GLSL 4.6: Implement atomic counter ops and SPV_KHR_shader_atomic_counter_ops.
2017-07-23 22:08:26 +00:00
case glslang::EOpAtomicCounterAnd:
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
opCode = spv::OpAtomicAnd;
break;
case glslang::EOpAtomicOr:
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
case glslang::EOpImageAtomicOr:
GLSL 4.6: Implement atomic counter ops and SPV_KHR_shader_atomic_counter_ops.
2017-07-23 22:08:26 +00:00
case glslang::EOpAtomicCounterOr:
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
opCode = spv::OpAtomicOr;
break;
case glslang::EOpAtomicXor:
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
case glslang::EOpImageAtomicXor:
GLSL 4.6: Implement atomic counter ops and SPV_KHR_shader_atomic_counter_ops.
2017-07-23 22:08:26 +00:00
case glslang::EOpAtomicCounterXor:
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
opCode = spv::OpAtomicXor;
break;
case glslang::EOpAtomicExchange:
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
case glslang::EOpImageAtomicExchange:
GLSL 4.6: Implement atomic counter ops and SPV_KHR_shader_atomic_counter_ops.
2017-07-23 22:08:26 +00:00
case glslang::EOpAtomicCounterExchange:
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
opCode = spv::OpAtomicExchange;
break;
case glslang::EOpAtomicCompSwap:
SPIRV: Add the support of missing image functions #72
2015-09-09 08:42:49 +00:00
case glslang::EOpImageAtomicCompSwap:
GLSL 4.6: Implement atomic counter ops and SPV_KHR_shader_atomic_counter_ops.
2017-07-23 22:08:26 +00:00
case glslang::EOpAtomicCounterCompSwap:
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
opCode = spv::OpAtomicCompareExchange;
break;
case glslang::EOpAtomicCounterIncrement:
opCode = spv::OpAtomicIIncrement;
break;
case glslang::EOpAtomicCounterDecrement:
opCode = spv::OpAtomicIDecrement;
break;
case glslang::EOpAtomicCounter:
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
case glslang::EOpImageAtomicLoad:
case glslang::EOpAtomicLoad:
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
opCode = spv::OpAtomicLoad;
break;
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
case glslang::EOpAtomicStore:
case glslang::EOpImageAtomicStore:
opCode = spv::OpAtomicStore;
break;
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
default:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(0);
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
break;
}
Implement extension GL_NV_shader_atomic_int64
2017-09-26 07:42:56 +00:00
if (typeProxy == glslang::EbtInt64 || typeProxy == glslang::EbtUint64)
builder.addCapability(spv::CapabilityInt64Atomics);
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
// Sort out the operands
// - mapping from glslang -> SPV
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
// - there are extra SPV operands that are optional in glslang
SPV OpAtomicCompareSwap: Generate correct operand order and number of operands.
2015-09-15 04:45:16 +00:00
// - compare-exchange swaps the value and comparator
// - compare-exchange has an extra memory semantics
SPV: Correct semantics of atomic-counter decrement. The semantics differ between GLSL/HLSL and SPIR-V. Translate between these.
2017-10-07 03:21:48 +00:00
// - EOpAtomicCounterDecrement needs a post decrement
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spv::Id pointerId = 0, compareId = 0, valueId = 0;
// scope defaults to Device in the old model, QueueFamilyKHR in the new model
spv::Id scopeId;
if (glslangIntermediate->usingVulkanMemoryModel()) {
scopeId = builder.makeUintConstant(spv::ScopeQueueFamilyKHR);
} else {
scopeId = builder.makeUintConstant(spv::ScopeDevice);
}
// semantics default to relaxed
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
spv::Id semanticsId = builder.makeUintConstant(lvalueCoherentFlags.isVolatile() ?
spv::MemorySemanticsVolatileMask :
spv::MemorySemanticsMaskNone);
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spv::Id semanticsId2 = semanticsId;
pointerId = operands[0];
if (opCode == spv::OpAtomicIIncrement || opCode == spv::OpAtomicIDecrement) {
// no additional operands
} else if (opCode == spv::OpAtomicCompareExchange) {
compareId = operands[1];
valueId = operands[2];
if (operands.size() > 3) {
scopeId = operands[3];
semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[4]) | builder.getConstantScalar(operands[5]));
semanticsId2 = builder.makeUintConstant(builder.getConstantScalar(operands[6]) | builder.getConstantScalar(operands[7]));
}
} else if (opCode == spv::OpAtomicLoad) {
if (operands.size() > 1) {
scopeId = operands[1];
semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[2]) | builder.getConstantScalar(operands[3]));
}
} else {
// atomic store or RMW
valueId = operands[1];
if (operands.size() > 2) {
scopeId = operands[2];
semanticsId = builder.makeUintConstant(builder.getConstantScalar(operands[3]) | builder.getConstantScalar(operands[4]));
}
}
// Check for capabilities
unsigned semanticsImmediate = builder.getConstantScalar(semanticsId) | builder.getConstantScalar(semanticsId2);
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
if (semanticsImmediate & (spv::MemorySemanticsMakeAvailableKHRMask |
spv::MemorySemanticsMakeVisibleKHRMask |
spv::MemorySemanticsOutputMemoryKHRMask |
spv::MemorySemanticsVolatileMask)) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
}
if (glslangIntermediate->usingVulkanMemoryModel() && builder.getConstantScalar(scopeId) == spv::ScopeDevice) {
builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR);
}
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
std::vector<spv::Id> spvAtomicOperands; // hold the spv operands
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spvAtomicOperands.push_back(pointerId);
spvAtomicOperands.push_back(scopeId);
spvAtomicOperands.push_back(semanticsId);
Fix issues from review comments and those relevant to texelFetch
2015-09-16 03:44:02 +00:00
if (opCode == spv::OpAtomicCompareExchange) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
spvAtomicOperands.push_back(semanticsId2);
spvAtomicOperands.push_back(valueId);
spvAtomicOperands.push_back(compareId);
} else if (opCode != spv::OpAtomicLoad && opCode != spv::OpAtomicIIncrement && opCode != spv::OpAtomicIDecrement) {
spvAtomicOperands.push_back(valueId);
Fix issues from review comments and those relevant to texelFetch
2015-09-16 03:44:02 +00:00
}
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
if (opCode == spv::OpAtomicStore) {
builder.createNoResultOp(opCode, spvAtomicOperands);
return 0;
} else {
spv::Id resultId = builder.createOp(opCode, typeId, spvAtomicOperands);
SPV: Correct semantics of atomic-counter decrement. The semantics differ between GLSL/HLSL and SPIR-V. Translate between these.
2017-10-07 03:21:48 +00:00
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
// GLSL and HLSL atomic-counter decrement return post-decrement value,
// while SPIR-V returns pre-decrement value. Translate between these semantics.
if (op == glslang::EOpAtomicCounterDecrement)
resultId = builder.createBinOp(spv::OpISub, typeId, resultId, builder.makeIntConstant(1));
SPV: Correct semantics of atomic-counter decrement. The semantics differ between GLSL/HLSL and SPIR-V. Translate between these.
2017-10-07 03:21:48 +00:00
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
return resultId;
}
AST -> SPV: Add basic atomic_uint and atomic*() built-in function functionality.
2015-07-23 16:22:48 +00:00
}
SPV: Non-functional: modularize createInvocationsOperation() from the previous commit.
2016-05-05 22:45:40 +00:00
// Create group invocation operations.
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
spv::Id TGlslangToSpvTraverser::createInvocationsOperation(glslang::TOperator op, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy)
SPV: Non-functional: modularize createInvocationsOperation() from the previous commit.
2016-05-05 22:45:40 +00:00
{
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
bool isUnsigned = isTypeUnsignedInt(typeProxy);
bool isFloat = isTypeFloat(typeProxy);
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
spv::Op opCode = spv::OpNop;
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
std::vector<spv::IdImmediate> spvGroupOperands;
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
spv::GroupOperation groupOperation = spv::GroupOperationMax;
Modify shader ballot extension by adding OpSubgroupReadInvocationKHR
2016-12-20 20:44:35 +00:00
if (op == glslang::EOpBallot || op == glslang::EOpReadFirstInvocation ||
op == glslang::EOpReadInvocation) {
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
builder.addExtension(spv::E_SPV_KHR_shader_ballot);
builder.addCapability(spv::CapabilitySubgroupBallotKHR);
Implement SPV_KHR_subgroup_vote doc.cpp: Add capabilities, scope to the opcodes. Add opcode and capability strings. GLSL.ext.KHR.h: Add extension string. GlslangToSpv.cpp: Fix handling of opcodes to generate appropriate SPIR-V. spirv.hpp: Add capability and opcode enums. spv.shaderGroupVote.comp.out: Update SPIR-V output for test shader.
2017-01-18 22:16:49 +00:00
} else if (op == glslang::EOpAnyInvocation ||
op == glslang::EOpAllInvocations ||
op == glslang::EOpAllInvocationsEqual) {
builder.addExtension(spv::E_SPV_KHR_subgroup_vote);
builder.addCapability(spv::CapabilitySubgroupVoteKHR);
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
} else {
builder.addCapability(spv::CapabilityGroups);
SPV: Add missing extension declarations for certain AMD extensions.
2016-10-14 09:41:45 +00:00
if (op == glslang::EOpMinInvocationsNonUniform ||
op == glslang::EOpMaxInvocationsNonUniform ||
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
op == glslang::EOpAddInvocationsNonUniform ||
op == glslang::EOpMinInvocationsInclusiveScanNonUniform ||
op == glslang::EOpMaxInvocationsInclusiveScanNonUniform ||
op == glslang::EOpAddInvocationsInclusiveScanNonUniform ||
op == glslang::EOpMinInvocationsExclusiveScanNonUniform ||
op == glslang::EOpMaxInvocationsExclusiveScanNonUniform ||
op == glslang::EOpAddInvocationsExclusiveScanNonUniform)
SPV: Add missing extension declarations for certain AMD extensions.
2016-10-14 09:41:45 +00:00
builder.addExtension(spv::E_SPV_AMD_shader_ballot);
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
switch (op) {
case glslang::EOpMinInvocations:
case glslang::EOpMaxInvocations:
case glslang::EOpAddInvocations:
case glslang::EOpMinInvocationsNonUniform:
case glslang::EOpMaxInvocationsNonUniform:
case glslang::EOpAddInvocationsNonUniform:
groupOperation = spv::GroupOperationReduce;
break;
case glslang::EOpMinInvocationsInclusiveScan:
case glslang::EOpMaxInvocationsInclusiveScan:
case glslang::EOpAddInvocationsInclusiveScan:
case glslang::EOpMinInvocationsInclusiveScanNonUniform:
case glslang::EOpMaxInvocationsInclusiveScanNonUniform:
case glslang::EOpAddInvocationsInclusiveScanNonUniform:
groupOperation = spv::GroupOperationInclusiveScan;
break;
case glslang::EOpMinInvocationsExclusiveScan:
case glslang::EOpMaxInvocationsExclusiveScan:
case glslang::EOpAddInvocationsExclusiveScan:
case glslang::EOpMinInvocationsExclusiveScanNonUniform:
case glslang::EOpMaxInvocationsExclusiveScanNonUniform:
case glslang::EOpAddInvocationsExclusiveScanNonUniform:
groupOperation = spv::GroupOperationExclusiveScan;
break;
Fix for not-handled-in-switch warnings Added default to the switch() for AMD_EXTENSIONS to avoid a spew of warning messages.
2017-01-20 20:34:10 +00:00
default:
break;
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
}
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::IdImmediate scope = { true, builder.makeUintConstant(spv::ScopeSubgroup) };
spvGroupOperands.push_back(scope);
if (groupOperation != spv::GroupOperationMax) {
Build: Fix some warnings.
2018-09-18 09:43:30 +00:00
spv::IdImmediate groupOp = { false, (unsigned)groupOperation };
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spvGroupOperands.push_back(groupOp);
}
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
}
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
for (auto opIt = operands.begin(); opIt != operands.end(); ++opIt) {
spv::IdImmediate op = { true, *opIt };
spvGroupOperands.push_back(op);
}
SPV: Non-functional: modularize createInvocationsOperation() from the previous commit.
2016-05-05 22:45:40 +00:00
switch (op) {
case glslang::EOpAnyInvocation:
Implement SPV_KHR_subgroup_vote doc.cpp: Add capabilities, scope to the opcodes. Add opcode and capability strings. GLSL.ext.KHR.h: Add extension string. GlslangToSpv.cpp: Fix handling of opcodes to generate appropriate SPIR-V. spirv.hpp: Add capability and opcode enums. spv.shaderGroupVote.comp.out: Update SPIR-V output for test shader.
2017-01-18 22:16:49 +00:00
opCode = spv::OpSubgroupAnyKHR;
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
break;
SPV: Non-functional: modularize createInvocationsOperation() from the previous commit.
2016-05-05 22:45:40 +00:00
case glslang::EOpAllInvocations:
Implement SPV_KHR_subgroup_vote doc.cpp: Add capabilities, scope to the opcodes. Add opcode and capability strings. GLSL.ext.KHR.h: Add extension string. GlslangToSpv.cpp: Fix handling of opcodes to generate appropriate SPIR-V. spirv.hpp: Add capability and opcode enums. spv.shaderGroupVote.comp.out: Update SPIR-V output for test shader.
2017-01-18 22:16:49 +00:00
opCode = spv::OpSubgroupAllKHR;
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
break;
SPV: Non-functional: modularize createInvocationsOperation() from the previous commit.
2016-05-05 22:45:40 +00:00
case glslang::EOpAllInvocationsEqual:
Implement SPV_KHR_subgroup_vote doc.cpp: Add capabilities, scope to the opcodes. Add opcode and capability strings. GLSL.ext.KHR.h: Add extension string. GlslangToSpv.cpp: Fix handling of opcodes to generate appropriate SPIR-V. spirv.hpp: Add capability and opcode enums. spv.shaderGroupVote.comp.out: Update SPIR-V output for test shader.
2017-01-18 22:16:49 +00:00
opCode = spv::OpSubgroupAllEqualKHR;
break;
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
case glslang::EOpReadInvocation:
Modify shader ballot extension by adding OpSubgroupReadInvocationKHR
2016-12-20 20:44:35 +00:00
opCode = spv::OpSubgroupReadInvocationKHR;
SPV: OpGroupBroadcast is unable to handle vector operand.
2016-09-26 07:53:40 +00:00
if (builder.isVectorType(typeId))
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
return CreateInvocationsVectorOperation(opCode, groupOperation, typeId, operands);
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
break;
case glslang::EOpReadFirstInvocation:
opCode = spv::OpSubgroupFirstInvocationKHR;
break;
case glslang::EOpBallot:
{
// NOTE: According to the spec, the result type of "OpSubgroupBallotKHR" must be a 4 component vector of 32
// bit integer types. The GLSL built-in function "ballotARB()" assumes the maximum number of invocations in
// a subgroup is 64. Thus, we have to convert uvec4.xy to uint64_t as follow:
//
// result = Bitcast(SubgroupBallotKHR(Predicate).xy)
//
spv::Id uintType = builder.makeUintType(32);
spv::Id uvec4Type = builder.makeVectorType(uintType, 4);
spv::Id result = builder.createOp(spv::OpSubgroupBallotKHR, uvec4Type, spvGroupOperands);
std::vector<spv::Id> components;
components.push_back(builder.createCompositeExtract(result, uintType, 0));
components.push_back(builder.createCompositeExtract(result, uintType, 1));
spv::Id uvec2Type = builder.makeVectorType(uintType, 2);
return builder.createUnaryOp(spv::OpBitcast, typeId,
builder.createCompositeConstruct(uvec2Type, components));
}
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
case glslang::EOpMinInvocations:
case glslang::EOpMaxInvocations:
case glslang::EOpAddInvocations:
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
case glslang::EOpMinInvocationsInclusiveScan:
case glslang::EOpMaxInvocationsInclusiveScan:
case glslang::EOpAddInvocationsInclusiveScan:
case glslang::EOpMinInvocationsExclusiveScan:
case glslang::EOpMaxInvocationsExclusiveScan:
case glslang::EOpAddInvocationsExclusiveScan:
if (op == glslang::EOpMinInvocations ||
op == glslang::EOpMinInvocationsInclusiveScan ||
op == glslang::EOpMinInvocationsExclusiveScan) {
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
if (isFloat)
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupFMin;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
else {
if (isUnsigned)
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupUMin;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
else
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupSMin;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
}
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
} else if (op == glslang::EOpMaxInvocations ||
op == glslang::EOpMaxInvocationsInclusiveScan ||
op == glslang::EOpMaxInvocationsExclusiveScan) {
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
if (isFloat)
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupFMax;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
else {
if (isUnsigned)
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupUMax;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
else
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupSMax;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
}
} else {
if (isFloat)
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupFAdd;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
else
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupIAdd;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
}
SPV: Some group invocation operations should be applied to scalars.
2016-08-23 07:41:05 +00:00
if (builder.isVectorType(typeId))
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
return CreateInvocationsVectorOperation(opCode, groupOperation, typeId, operands);
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
break;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
case glslang::EOpMinInvocationsNonUniform:
case glslang::EOpMaxInvocationsNonUniform:
case glslang::EOpAddInvocationsNonUniform:
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
case glslang::EOpMinInvocationsInclusiveScanNonUniform:
case glslang::EOpMaxInvocationsInclusiveScanNonUniform:
case glslang::EOpAddInvocationsInclusiveScanNonUniform:
case glslang::EOpMinInvocationsExclusiveScanNonUniform:
case glslang::EOpMaxInvocationsExclusiveScanNonUniform:
case glslang::EOpAddInvocationsExclusiveScanNonUniform:
if (op == glslang::EOpMinInvocationsNonUniform ||
op == glslang::EOpMinInvocationsInclusiveScanNonUniform ||
op == glslang::EOpMinInvocationsExclusiveScanNonUniform) {
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
if (isFloat)
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupFMinNonUniformAMD;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
else {
if (isUnsigned)
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupUMinNonUniformAMD;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
else
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupSMinNonUniformAMD;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
}
}
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
else if (op == glslang::EOpMaxInvocationsNonUniform ||
op == glslang::EOpMaxInvocationsInclusiveScanNonUniform ||
op == glslang::EOpMaxInvocationsExclusiveScanNonUniform) {
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
if (isFloat)
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupFMaxNonUniformAMD;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
else {
if (isUnsigned)
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupUMaxNonUniformAMD;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
else
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupSMaxNonUniformAMD;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
}
}
else {
if (isFloat)
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupFAddNonUniformAMD;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
else
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
opCode = spv::OpGroupIAddNonUniformAMD;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
}
SPV: Some group invocation operations should be applied to scalars.
2016-08-23 07:41:05 +00:00
if (builder.isVectorType(typeId))
Implement new revision of extension GL_AMD_shader_ballot - Add support for invocation functions with "InclusiveScan" and "ExclusiveScan" modes. - Add support for invocation functions taking int64/uint64/doube/float16 as inout data types.
2016-10-14 09:22:23 +00:00
return CreateInvocationsVectorOperation(opCode, groupOperation, typeId, operands);
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
break;
SPV: Non-functional: modularize createInvocationsOperation() from the previous commit.
2016-05-05 22:45:40 +00:00
default:
logger->missingFunctionality("invocation operation");
return spv::NoResult;
}
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
assert(opCode != spv::OpNop);
return builder.createOp(opCode, typeId, spvGroupOperands);
SPV: Non-functional: modularize createInvocationsOperation() from the previous commit.
2016-05-05 22:45:40 +00:00
}
SPV: Some group invocation operations should be applied to scalars.
2016-08-23 07:41:05 +00:00
// Create group invocation operations on a vector
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::Id TGlslangToSpvTraverser::CreateInvocationsVectorOperation(spv::Op op, spv::GroupOperation groupOperation,
spv::Id typeId, std::vector<spv::Id>& operands)
SPV: Some group invocation operations should be applied to scalars.
2016-08-23 07:41:05 +00:00
{
assert(op == spv::OpGroupFMin || op == spv::OpGroupUMin || op == spv::OpGroupSMin ||
op == spv::OpGroupFMax || op == spv::OpGroupUMax || op == spv::OpGroupSMax ||
SPV: OpGroupBroadcast is unable to handle vector operand.
2016-09-26 07:53:40 +00:00
op == spv::OpGroupFAdd || op == spv::OpGroupIAdd || op == spv::OpGroupBroadcast ||
Modify shader ballot extension by adding OpSubgroupReadInvocationKHR
2016-12-20 20:44:35 +00:00
op == spv::OpSubgroupReadInvocationKHR ||
SPV: Some group invocation operations should be applied to scalars.
2016-08-23 07:41:05 +00:00
op == spv::OpGroupFMinNonUniformAMD || op == spv::OpGroupUMinNonUniformAMD || op == spv::OpGroupSMinNonUniformAMD ||
op == spv::OpGroupFMaxNonUniformAMD || op == spv::OpGroupUMaxNonUniformAMD || op == spv::OpGroupSMaxNonUniformAMD ||
op == spv::OpGroupFAddNonUniformAMD || op == spv::OpGroupIAddNonUniformAMD);
// Handle group invocation operations scalar by scalar.
// The result type is the same type as the original type.
// The algorithm is to:
// - break the vector into scalars
// - apply the operation to each scalar
// - make a vector out the scalar results
// get the types sorted out
SPV: OpGroupBroadcast is unable to handle vector operand.
2016-09-26 07:53:40 +00:00
int numComponents = builder.getNumComponents(operands[0]);
spv::Id scalarType = builder.getScalarTypeId(builder.getTypeId(operands[0]));
SPV: Some group invocation operations should be applied to scalars.
2016-08-23 07:41:05 +00:00
std::vector<spv::Id> results;
// do each scalar op
for (int comp = 0; comp < numComponents; ++comp) {
std::vector<unsigned int> indexes;
indexes.push_back(comp);
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::IdImmediate scalar = { true, builder.createCompositeExtract(operands[0], scalarType, indexes) };
std::vector<spv::IdImmediate> spvGroupOperands;
Modify shader ballot extension by adding OpSubgroupReadInvocationKHR
2016-12-20 20:44:35 +00:00
if (op == spv::OpSubgroupReadInvocationKHR) {
spvGroupOperands.push_back(scalar);
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::IdImmediate operand = { true, operands[1] };
spvGroupOperands.push_back(operand);
Modify shader ballot extension by adding OpSubgroupReadInvocationKHR
2016-12-20 20:44:35 +00:00
} else if (op == spv::OpGroupBroadcast) {
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::IdImmediate scope = { true, builder.makeUintConstant(spv::ScopeSubgroup) };
spvGroupOperands.push_back(scope);
SPV: OpGroupBroadcast is unable to handle vector operand.
2016-09-26 07:53:40 +00:00
spvGroupOperands.push_back(scalar);
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::IdImmediate operand = { true, operands[1] };
spvGroupOperands.push_back(operand);
SPV: OpGroupBroadcast is unable to handle vector operand.
2016-09-26 07:53:40 +00:00
} else {
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::IdImmediate scope = { true, builder.makeUintConstant(spv::ScopeSubgroup) };
spvGroupOperands.push_back(scope);
Build: Fix some warnings.
2018-09-18 09:43:30 +00:00
spv::IdImmediate groupOp = { false, (unsigned)groupOperation };
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spvGroupOperands.push_back(groupOp);
SPV: OpGroupBroadcast is unable to handle vector operand.
2016-09-26 07:53:40 +00:00
spvGroupOperands.push_back(scalar);
}
SPV: Some group invocation operations should be applied to scalars.
2016-08-23 07:41:05 +00:00
SPV: OpGroupBroadcast is unable to handle vector operand.
2016-09-26 07:53:40 +00:00
results.push_back(builder.createOp(op, scalarType, spvGroupOperands));
SPV: Some group invocation operations should be applied to scalars.
2016-08-23 07:41:05 +00:00
}
// put the pieces together
return builder.createCompositeConstruct(typeId, results);
}
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
// Create subgroup invocation operations.
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::Id TGlslangToSpvTraverser::createSubgroupOperation(glslang::TOperator op, spv::Id typeId,
std::vector<spv::Id>& operands, glslang::TBasicType typeProxy)
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
{
// Add the required capabilities.
switch (op) {
case glslang::EOpSubgroupElect:
builder.addCapability(spv::CapabilityGroupNonUniform);
break;
case glslang::EOpSubgroupAll:
case glslang::EOpSubgroupAny:
case glslang::EOpSubgroupAllEqual:
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformVote);
break;
case glslang::EOpSubgroupBroadcast:
case glslang::EOpSubgroupBroadcastFirst:
case glslang::EOpSubgroupBallot:
case glslang::EOpSubgroupInverseBallot:
case glslang::EOpSubgroupBallotBitExtract:
case glslang::EOpSubgroupBallotBitCount:
case glslang::EOpSubgroupBallotInclusiveBitCount:
case glslang::EOpSubgroupBallotExclusiveBitCount:
case glslang::EOpSubgroupBallotFindLSB:
case glslang::EOpSubgroupBallotFindMSB:
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformBallot);
break;
case glslang::EOpSubgroupShuffle:
case glslang::EOpSubgroupShuffleXor:
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformShuffle);
break;
case glslang::EOpSubgroupShuffleUp:
case glslang::EOpSubgroupShuffleDown:
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformShuffleRelative);
break;
case glslang::EOpSubgroupAdd:
case glslang::EOpSubgroupMul:
case glslang::EOpSubgroupMin:
case glslang::EOpSubgroupMax:
case glslang::EOpSubgroupAnd:
case glslang::EOpSubgroupOr:
case glslang::EOpSubgroupXor:
case glslang::EOpSubgroupInclusiveAdd:
case glslang::EOpSubgroupInclusiveMul:
case glslang::EOpSubgroupInclusiveMin:
case glslang::EOpSubgroupInclusiveMax:
case glslang::EOpSubgroupInclusiveAnd:
case glslang::EOpSubgroupInclusiveOr:
case glslang::EOpSubgroupInclusiveXor:
case glslang::EOpSubgroupExclusiveAdd:
case glslang::EOpSubgroupExclusiveMul:
case glslang::EOpSubgroupExclusiveMin:
case glslang::EOpSubgroupExclusiveMax:
case glslang::EOpSubgroupExclusiveAnd:
case glslang::EOpSubgroupExclusiveOr:
case glslang::EOpSubgroupExclusiveXor:
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformArithmetic);
break;
case glslang::EOpSubgroupClusteredAdd:
case glslang::EOpSubgroupClusteredMul:
case glslang::EOpSubgroupClusteredMin:
case glslang::EOpSubgroupClusteredMax:
case glslang::EOpSubgroupClusteredAnd:
case glslang::EOpSubgroupClusteredOr:
case glslang::EOpSubgroupClusteredXor:
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformClustered);
break;
case glslang::EOpSubgroupQuadBroadcast:
case glslang::EOpSubgroupQuadSwapHorizontal:
case glslang::EOpSubgroupQuadSwapVertical:
case glslang::EOpSubgroupQuadSwapDiagonal:
builder.addCapability(spv::CapabilityGroupNonUniform);
builder.addCapability(spv::CapabilityGroupNonUniformQuad);
break;
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
case glslang::EOpSubgroupPartitionedAdd:
case glslang::EOpSubgroupPartitionedMul:
case glslang::EOpSubgroupPartitionedMin:
case glslang::EOpSubgroupPartitionedMax:
case glslang::EOpSubgroupPartitionedAnd:
case glslang::EOpSubgroupPartitionedOr:
case glslang::EOpSubgroupPartitionedXor:
case glslang::EOpSubgroupPartitionedInclusiveAdd:
case glslang::EOpSubgroupPartitionedInclusiveMul:
case glslang::EOpSubgroupPartitionedInclusiveMin:
case glslang::EOpSubgroupPartitionedInclusiveMax:
case glslang::EOpSubgroupPartitionedInclusiveAnd:
case glslang::EOpSubgroupPartitionedInclusiveOr:
case glslang::EOpSubgroupPartitionedInclusiveXor:
case glslang::EOpSubgroupPartitionedExclusiveAdd:
case glslang::EOpSubgroupPartitionedExclusiveMul:
case glslang::EOpSubgroupPartitionedExclusiveMin:
case glslang::EOpSubgroupPartitionedExclusiveMax:
case glslang::EOpSubgroupPartitionedExclusiveAnd:
case glslang::EOpSubgroupPartitionedExclusiveOr:
case glslang::EOpSubgroupPartitionedExclusiveXor:
builder.addExtension(spv::E_SPV_NV_shader_subgroup_partitioned);
builder.addCapability(spv::CapabilityGroupNonUniformPartitionedNV);
break;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
default: assert(0 && "Unhandled subgroup operation!");
}
const bool isUnsigned = typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64;
const bool isFloat = typeProxy == glslang::EbtFloat || typeProxy == glslang::EbtDouble;
const bool isBool = typeProxy == glslang::EbtBool;
spv::Op opCode = spv::OpNop;
// Figure out which opcode to use.
switch (op) {
case glslang::EOpSubgroupElect: opCode = spv::OpGroupNonUniformElect; break;
case glslang::EOpSubgroupAll: opCode = spv::OpGroupNonUniformAll; break;
case glslang::EOpSubgroupAny: opCode = spv::OpGroupNonUniformAny; break;
case glslang::EOpSubgroupAllEqual: opCode = spv::OpGroupNonUniformAllEqual; break;
case glslang::EOpSubgroupBroadcast: opCode = spv::OpGroupNonUniformBroadcast; break;
case glslang::EOpSubgroupBroadcastFirst: opCode = spv::OpGroupNonUniformBroadcastFirst; break;
case glslang::EOpSubgroupBallot: opCode = spv::OpGroupNonUniformBallot; break;
case glslang::EOpSubgroupInverseBallot: opCode = spv::OpGroupNonUniformInverseBallot; break;
case glslang::EOpSubgroupBallotBitExtract: opCode = spv::OpGroupNonUniformBallotBitExtract; break;
case glslang::EOpSubgroupBallotBitCount:
case glslang::EOpSubgroupBallotInclusiveBitCount:
case glslang::EOpSubgroupBallotExclusiveBitCount: opCode = spv::OpGroupNonUniformBallotBitCount; break;
case glslang::EOpSubgroupBallotFindLSB: opCode = spv::OpGroupNonUniformBallotFindLSB; break;
case glslang::EOpSubgroupBallotFindMSB: opCode = spv::OpGroupNonUniformBallotFindMSB; break;
case glslang::EOpSubgroupShuffle: opCode = spv::OpGroupNonUniformShuffle; break;
case glslang::EOpSubgroupShuffleXor: opCode = spv::OpGroupNonUniformShuffleXor; break;
case glslang::EOpSubgroupShuffleUp: opCode = spv::OpGroupNonUniformShuffleUp; break;
case glslang::EOpSubgroupShuffleDown: opCode = spv::OpGroupNonUniformShuffleDown; break;
case glslang::EOpSubgroupAdd:
case glslang::EOpSubgroupInclusiveAdd:
case glslang::EOpSubgroupExclusiveAdd:
case glslang::EOpSubgroupClusteredAdd:
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
case glslang::EOpSubgroupPartitionedAdd:
case glslang::EOpSubgroupPartitionedInclusiveAdd:
case glslang::EOpSubgroupPartitionedExclusiveAdd:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
if (isFloat) {
opCode = spv::OpGroupNonUniformFAdd;
} else {
opCode = spv::OpGroupNonUniformIAdd;
}
break;
case glslang::EOpSubgroupMul:
case glslang::EOpSubgroupInclusiveMul:
case glslang::EOpSubgroupExclusiveMul:
case glslang::EOpSubgroupClusteredMul:
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
case glslang::EOpSubgroupPartitionedMul:
case glslang::EOpSubgroupPartitionedInclusiveMul:
case glslang::EOpSubgroupPartitionedExclusiveMul:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
if (isFloat) {
opCode = spv::OpGroupNonUniformFMul;
} else {
opCode = spv::OpGroupNonUniformIMul;
}
break;
case glslang::EOpSubgroupMin:
case glslang::EOpSubgroupInclusiveMin:
case glslang::EOpSubgroupExclusiveMin:
case glslang::EOpSubgroupClusteredMin:
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
case glslang::EOpSubgroupPartitionedMin:
case glslang::EOpSubgroupPartitionedInclusiveMin:
case glslang::EOpSubgroupPartitionedExclusiveMin:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
if (isFloat) {
opCode = spv::OpGroupNonUniformFMin;
} else if (isUnsigned) {
opCode = spv::OpGroupNonUniformUMin;
} else {
opCode = spv::OpGroupNonUniformSMin;
}
break;
case glslang::EOpSubgroupMax:
case glslang::EOpSubgroupInclusiveMax:
case glslang::EOpSubgroupExclusiveMax:
case glslang::EOpSubgroupClusteredMax:
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
case glslang::EOpSubgroupPartitionedMax:
case glslang::EOpSubgroupPartitionedInclusiveMax:
case glslang::EOpSubgroupPartitionedExclusiveMax:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
if (isFloat) {
opCode = spv::OpGroupNonUniformFMax;
} else if (isUnsigned) {
opCode = spv::OpGroupNonUniformUMax;
} else {
opCode = spv::OpGroupNonUniformSMax;
}
break;
case glslang::EOpSubgroupAnd:
case glslang::EOpSubgroupInclusiveAnd:
case glslang::EOpSubgroupExclusiveAnd:
case glslang::EOpSubgroupClusteredAnd:
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
case glslang::EOpSubgroupPartitionedAnd:
case glslang::EOpSubgroupPartitionedInclusiveAnd:
case glslang::EOpSubgroupPartitionedExclusiveAnd:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
if (isBool) {
opCode = spv::OpGroupNonUniformLogicalAnd;
} else {
opCode = spv::OpGroupNonUniformBitwiseAnd;
}
break;
case glslang::EOpSubgroupOr:
case glslang::EOpSubgroupInclusiveOr:
case glslang::EOpSubgroupExclusiveOr:
case glslang::EOpSubgroupClusteredOr:
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
case glslang::EOpSubgroupPartitionedOr:
case glslang::EOpSubgroupPartitionedInclusiveOr:
case glslang::EOpSubgroupPartitionedExclusiveOr:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
if (isBool) {
opCode = spv::OpGroupNonUniformLogicalOr;
} else {
opCode = spv::OpGroupNonUniformBitwiseOr;
}
break;
case glslang::EOpSubgroupXor:
case glslang::EOpSubgroupInclusiveXor:
case glslang::EOpSubgroupExclusiveXor:
case glslang::EOpSubgroupClusteredXor:
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
case glslang::EOpSubgroupPartitionedXor:
case glslang::EOpSubgroupPartitionedInclusiveXor:
case glslang::EOpSubgroupPartitionedExclusiveXor:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
if (isBool) {
opCode = spv::OpGroupNonUniformLogicalXor;
} else {
opCode = spv::OpGroupNonUniformBitwiseXor;
}
break;
case glslang::EOpSubgroupQuadBroadcast: opCode = spv::OpGroupNonUniformQuadBroadcast; break;
case glslang::EOpSubgroupQuadSwapHorizontal:
case glslang::EOpSubgroupQuadSwapVertical:
case glslang::EOpSubgroupQuadSwapDiagonal: opCode = spv::OpGroupNonUniformQuadSwap; break;
default: assert(0 && "Unhandled subgroup operation!");
}
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
// get the right Group Operation
spv::GroupOperation groupOperation = spv::GroupOperationMax;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
switch (op) {
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
default:
break;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpSubgroupBallotBitCount:
case glslang::EOpSubgroupAdd:
case glslang::EOpSubgroupMul:
case glslang::EOpSubgroupMin:
case glslang::EOpSubgroupMax:
case glslang::EOpSubgroupAnd:
case glslang::EOpSubgroupOr:
case glslang::EOpSubgroupXor:
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
groupOperation = spv::GroupOperationReduce;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
case glslang::EOpSubgroupBallotInclusiveBitCount:
case glslang::EOpSubgroupInclusiveAdd:
case glslang::EOpSubgroupInclusiveMul:
case glslang::EOpSubgroupInclusiveMin:
case glslang::EOpSubgroupInclusiveMax:
case glslang::EOpSubgroupInclusiveAnd:
case glslang::EOpSubgroupInclusiveOr:
case glslang::EOpSubgroupInclusiveXor:
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
groupOperation = spv::GroupOperationInclusiveScan;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
case glslang::EOpSubgroupBallotExclusiveBitCount:
case glslang::EOpSubgroupExclusiveAdd:
case glslang::EOpSubgroupExclusiveMul:
case glslang::EOpSubgroupExclusiveMin:
case glslang::EOpSubgroupExclusiveMax:
case glslang::EOpSubgroupExclusiveAnd:
case glslang::EOpSubgroupExclusiveOr:
case glslang::EOpSubgroupExclusiveXor:
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
groupOperation = spv::GroupOperationExclusiveScan;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
case glslang::EOpSubgroupClusteredAdd:
case glslang::EOpSubgroupClusteredMul:
case glslang::EOpSubgroupClusteredMin:
case glslang::EOpSubgroupClusteredMax:
case glslang::EOpSubgroupClusteredAnd:
case glslang::EOpSubgroupClusteredOr:
case glslang::EOpSubgroupClusteredXor:
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
groupOperation = spv::GroupOperationClusteredReduce;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
break;
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
case glslang::EOpSubgroupPartitionedAdd:
case glslang::EOpSubgroupPartitionedMul:
case glslang::EOpSubgroupPartitionedMin:
case glslang::EOpSubgroupPartitionedMax:
case glslang::EOpSubgroupPartitionedAnd:
case glslang::EOpSubgroupPartitionedOr:
case glslang::EOpSubgroupPartitionedXor:
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
groupOperation = spv::GroupOperationPartitionedReduceNV;
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
break;
case glslang::EOpSubgroupPartitionedInclusiveAdd:
case glslang::EOpSubgroupPartitionedInclusiveMul:
case glslang::EOpSubgroupPartitionedInclusiveMin:
case glslang::EOpSubgroupPartitionedInclusiveMax:
case glslang::EOpSubgroupPartitionedInclusiveAnd:
case glslang::EOpSubgroupPartitionedInclusiveOr:
case glslang::EOpSubgroupPartitionedInclusiveXor:
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
groupOperation = spv::GroupOperationPartitionedInclusiveScanNV;
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
break;
case glslang::EOpSubgroupPartitionedExclusiveAdd:
case glslang::EOpSubgroupPartitionedExclusiveMul:
case glslang::EOpSubgroupPartitionedExclusiveMin:
case glslang::EOpSubgroupPartitionedExclusiveMax:
case glslang::EOpSubgroupPartitionedExclusiveAnd:
case glslang::EOpSubgroupPartitionedExclusiveOr:
case glslang::EOpSubgroupPartitionedExclusiveXor:
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
groupOperation = spv::GroupOperationPartitionedExclusiveScanNV;
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
break;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
}
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
// build the instruction
std::vector<spv::IdImmediate> spvGroupOperands;
// Every operation begins with the Execution Scope operand.
spv::IdImmediate executionScope = { true, builder.makeUintConstant(spv::ScopeSubgroup) };
spvGroupOperands.push_back(executionScope);
// Next, for all operations that use a Group Operation, push that as an operand.
if (groupOperation != spv::GroupOperationMax) {
Build: Fix some warnings.
2018-09-18 09:43:30 +00:00
spv::IdImmediate groupOperand = { false, (unsigned)groupOperation };
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spvGroupOperands.push_back(groupOperand);
}
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
// Push back the operands next.
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
for (auto opIt = operands.cbegin(); opIt != operands.cend(); ++opIt) {
spv::IdImmediate operand = { true, *opIt };
spvGroupOperands.push_back(operand);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
}
// Some opcodes have additional operands.
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
spv::Id directionId = spv::NoResult;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
switch (op) {
default: break;
SPV: More corrections of <id> versus "immediate" operands.
2018-08-14 19:31:43 +00:00
case glslang::EOpSubgroupQuadSwapHorizontal: directionId = builder.makeUintConstant(0); break;
case glslang::EOpSubgroupQuadSwapVertical: directionId = builder.makeUintConstant(1); break;
case glslang::EOpSubgroupQuadSwapDiagonal: directionId = builder.makeUintConstant(2); break;
}
if (directionId != spv::NoResult) {
spv::IdImmediate direction = { true, directionId };
spvGroupOperands.push_back(direction);
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
}
return builder.createOp(opCode, typeId, spvGroupOperands);
}
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
bool isUnsigned = isTypeUnsignedInt(typeProxy);
bool isFloat = isTypeFloat(typeProxy);
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
spv::Op opCode = spv::OpNop;
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
int extBuiltins = -1;
Final round for line endings.
2015-06-26 22:58:36 +00:00
int libCall = -1;
Fix several build warnings/error encountered with VS2013 This also fixes the newlines for spirv.hpp to be consistent with the rest of the files.
2016-01-06 18:41:02 +00:00
size_t consumedOperands = operands.size();
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
spv::Id typeId0 = 0;
if (consumedOperands > 0)
typeId0 = builder.getTypeId(operands[0]);
HLSL: Fix an issue of frexp(). The "exp" parameter is floating-point type in HLSL intrinsic while it is integer type in GLSL built-in function.
2017-03-29 09:12:40 +00:00
spv::Id typeId1 = 0;
if (consumedOperands > 1)
typeId1 = builder.getTypeId(operands[1]);
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
spv::Id frexpIntType = 0;
Final round for line endings.
2015-06-26 22:58:36 +00:00
switch (op) {
case glslang::EOpMin:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
if (isFloat)
SPV: Add a switch for favoring non-NaN operands in min, max, and clamp.
2019-06-18 05:33:09 +00:00
libCall = nanMinMaxClamp ? spv::GLSLstd450NMin : spv::GLSLstd450FMin;
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
else if (isUnsigned)
libCall = spv::GLSLstd450UMin;
else
libCall = spv::GLSLstd450SMin;
SPV GLSL.std.450: Add needed smear operations for min, max, clamp, mix, step, and smoothstep.
2015-12-13 20:34:37 +00:00
builder.promoteScalar(precision, operands.front(), operands.back());
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpModf:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Modf;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpMax:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
if (isFloat)
SPV: Add a switch for favoring non-NaN operands in min, max, and clamp.
2019-06-18 05:33:09 +00:00
libCall = nanMinMaxClamp ? spv::GLSLstd450NMax : spv::GLSLstd450FMax;
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
else if (isUnsigned)
libCall = spv::GLSLstd450UMax;
else
libCall = spv::GLSLstd450SMax;
SPV GLSL.std.450: Add needed smear operations for min, max, clamp, mix, step, and smoothstep.
2015-12-13 20:34:37 +00:00
builder.promoteScalar(precision, operands.front(), operands.back());
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpPow:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Pow;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpDot:
opCode = spv::OpDot;
break;
case glslang::EOpAtan:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Atan2;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpClamp:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
if (isFloat)
SPV: Add a switch for favoring non-NaN operands in min, max, and clamp.
2019-06-18 05:33:09 +00:00
libCall = nanMinMaxClamp ? spv::GLSLstd450NClamp : spv::GLSLstd450FClamp;
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
else if (isUnsigned)
libCall = spv::GLSLstd450UClamp;
else
libCall = spv::GLSLstd450SClamp;
SPV GLSL.std.450: Add needed smear operations for min, max, clamp, mix, step, and smoothstep.
2015-12-13 20:34:37 +00:00
builder.promoteScalar(precision, operands.front(), operands[1]);
builder.promoteScalar(precision, operands.front(), operands[2]);
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpMix:
SPV: Implement boolean selection for mix(). Use OpSelect to implement boolean selection of mix(). FMix is applicable to linear blending of mix().
2016-03-15 04:08:31 +00:00
if (! builder.isBoolType(builder.getScalarTypeId(builder.getTypeId(operands.back())))) {
assert(isFloat);
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
libCall = spv::GLSLstd450FMix;
SPV: Implement boolean selection for mix(). Use OpSelect to implement boolean selection of mix(). FMix is applicable to linear blending of mix().
2016-03-15 04:08:31 +00:00
} else {
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
opCode = spv::OpSelect;
SPV: Implement boolean selection for mix(). Use OpSelect to implement boolean selection of mix(). FMix is applicable to linear blending of mix().
2016-03-15 04:08:31 +00:00
std::swap(operands.front(), operands.back());
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
}
SPV GLSL.std.450: Add needed smear operations for min, max, clamp, mix, step, and smoothstep.
2015-12-13 20:34:37 +00:00
builder.promoteScalar(precision, operands.front(), operands.back());
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpStep:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Step;
SPV GLSL.std.450: Add needed smear operations for min, max, clamp, mix, step, and smoothstep.
2015-12-13 20:34:37 +00:00
builder.promoteScalar(precision, operands.front(), operands.back());
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpSmoothStep:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450SmoothStep;
SPV GLSL.std.450: Add needed smear operations for min, max, clamp, mix, step, and smoothstep.
2015-12-13 20:34:37 +00:00
builder.promoteScalar(precision, operands[0], operands[2]);
builder.promoteScalar(precision, operands[1], operands[2]);
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpDistance:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Distance;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpCross:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Cross;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpFaceForward:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450FaceForward;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpReflect:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Reflect;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
case glslang::EOpRefract:
Move to revision 31 of SPIR-V.
2015-08-07 04:53:06 +00:00
libCall = spv::GLSLstd450Refract;
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
SPV: Implement GLSL interpolation built-in functions.
2015-12-08 09:12:09 +00:00
case glslang::EOpInterpolateAtSample:
Fix issue #1376: SPV_AMD_gpu_shader_half_float is not emitted under vulkan1.1 Interpolation functions, using 16-bit interpolants, should always emit this extension string regardless of vulkan enviornment version.
2018-05-17 05:51:28 +00:00
if (typeProxy == glslang::EbtFloat16)
builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float);
SPV: Implement GLSL interpolation built-in functions.
2015-12-08 09:12:09 +00:00
libCall = spv::GLSLstd450InterpolateAtSample;
break;
case glslang::EOpInterpolateAtOffset:
Fix issue #1376: SPV_AMD_gpu_shader_half_float is not emitted under vulkan1.1 Interpolation functions, using 16-bit interpolants, should always emit this extension string regardless of vulkan enviornment version.
2018-05-17 05:51:28 +00:00
if (typeProxy == glslang::EbtFloat16)
builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float);
SPV: Implement GLSL interpolation built-in functions.
2015-12-08 09:12:09 +00:00
libCall = spv::GLSLstd450InterpolateAtOffset;
break;
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
case glslang::EOpAddCarry:
opCode = spv::OpIAddCarry;
typeId = builder.makeStructResultType(typeId0, typeId0);
consumedOperands = 2;
break;
case glslang::EOpSubBorrow:
opCode = spv::OpISubBorrow;
typeId = builder.makeStructResultType(typeId0, typeId0);
consumedOperands = 2;
break;
case glslang::EOpUMulExtended:
opCode = spv::OpUMulExtended;
typeId = builder.makeStructResultType(typeId0, typeId0);
consumedOperands = 2;
break;
case glslang::EOpIMulExtended:
opCode = spv::OpSMulExtended;
typeId = builder.makeStructResultType(typeId0, typeId0);
consumedOperands = 2;
break;
case glslang::EOpBitfieldExtract:
if (isUnsigned)
opCode = spv::OpBitFieldUExtract;
else
opCode = spv::OpBitFieldSExtract;
break;
case glslang::EOpBitfieldInsert:
opCode = spv::OpBitFieldInsert;
break;
case glslang::EOpFma:
libCall = spv::GLSLstd450Fma;
break;
case glslang::EOpFrexp:
HLSL: Fix an issue of frexp(). The "exp" parameter is floating-point type in HLSL intrinsic while it is integer type in GLSL built-in function.
2017-03-29 09:12:40 +00:00
{
libCall = spv::GLSLstd450FrexpStruct;
assert(builder.isPointerType(typeId1));
typeId1 = builder.getContainedTypeId(typeId1);
int width = builder.getScalarTypeWidth(typeId1);
Specify SPV_AMD_gpu_shader_int16 extension string when frexp() takes 16-bit exp operand
2018-04-11 08:56:50 +00:00
if (width == 16)
// Using 16-bit exp operand, enable extension SPV_AMD_gpu_shader_int16
builder.addExtension(spv::E_SPV_AMD_gpu_shader_int16);
HLSL: Fix an issue of frexp(). The "exp" parameter is floating-point type in HLSL intrinsic while it is integer type in GLSL built-in function.
2017-03-29 09:12:40 +00:00
if (builder.getNumComponents(operands[0]) == 1)
frexpIntType = builder.makeIntegerType(width, true);
else
frexpIntType = builder.makeVectorType(builder.makeIntegerType(width, true), builder.getNumComponents(operands[0]));
typeId = builder.makeStructResultType(typeId0, frexpIntType);
consumedOperands = 1;
}
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
break;
case glslang::EOpLdexp:
libCall = spv::GLSLstd450Ldexp;
break;
Implement the extension GL_ARB_shader_ballot Add new built-in variables and functions to the parser (SPIR-V tokens are missing).
2016-04-14 08:53:07 +00:00
case glslang::EOpReadInvocation:
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
return createInvocationsOperation(op, typeId, operands, typeProxy);
Implement the extension GL_ARB_shader_ballot Add new built-in variables and functions to the parser (SPIR-V tokens are missing).
2016-04-14 08:53:07 +00:00
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpSubgroupBroadcast:
case glslang::EOpSubgroupBallotBitExtract:
case glslang::EOpSubgroupShuffle:
case glslang::EOpSubgroupShuffleXor:
case glslang::EOpSubgroupShuffleUp:
case glslang::EOpSubgroupShuffleDown:
case glslang::EOpSubgroupClusteredAdd:
case glslang::EOpSubgroupClusteredMul:
case glslang::EOpSubgroupClusteredMin:
case glslang::EOpSubgroupClusteredMax:
case glslang::EOpSubgroupClusteredAnd:
case glslang::EOpSubgroupClusteredOr:
case glslang::EOpSubgroupClusteredXor:
case glslang::EOpSubgroupQuadBroadcast:
Implement GL_NV_shader_subgroup_partitioned
2018-03-30 03:52:17 +00:00
case glslang::EOpSubgroupPartitionedAdd:
case glslang::EOpSubgroupPartitionedMul:
case glslang::EOpSubgroupPartitionedMin:
case glslang::EOpSubgroupPartitionedMax:
case glslang::EOpSubgroupPartitionedAnd:
case glslang::EOpSubgroupPartitionedOr:
case glslang::EOpSubgroupPartitionedXor:
case glslang::EOpSubgroupPartitionedInclusiveAdd:
case glslang::EOpSubgroupPartitionedInclusiveMul:
case glslang::EOpSubgroupPartitionedInclusiveMin:
case glslang::EOpSubgroupPartitionedInclusiveMax:
case glslang::EOpSubgroupPartitionedInclusiveAnd:
case glslang::EOpSubgroupPartitionedInclusiveOr:
case glslang::EOpSubgroupPartitionedInclusiveXor:
case glslang::EOpSubgroupPartitionedExclusiveAdd:
case glslang::EOpSubgroupPartitionedExclusiveMul:
case glslang::EOpSubgroupPartitionedExclusiveMin:
case glslang::EOpSubgroupPartitionedExclusiveMax:
case glslang::EOpSubgroupPartitionedExclusiveAnd:
case glslang::EOpSubgroupPartitionedExclusiveOr:
case glslang::EOpSubgroupPartitionedExclusiveXor:
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
return createSubgroupOperation(op, typeId, operands, typeProxy);
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
case glslang::EOpSwizzleInvocations:
extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot);
libCall = spv::SwizzleInvocationsAMD;
break;
case glslang::EOpSwizzleInvocationsMasked:
extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot);
libCall = spv::SwizzleInvocationsMaskedAMD;
break;
case glslang::EOpWriteInvocation:
extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot);
libCall = spv::WriteInvocationAMD;
break;
case glslang::EOpMin3:
extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax);
if (isFloat)
libCall = spv::FMin3AMD;
else {
if (isUnsigned)
libCall = spv::UMin3AMD;
else
libCall = spv::SMin3AMD;
}
break;
case glslang::EOpMax3:
extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax);
if (isFloat)
libCall = spv::FMax3AMD;
else {
if (isUnsigned)
libCall = spv::UMax3AMD;
else
libCall = spv::SMax3AMD;
}
break;
case glslang::EOpMid3:
extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax);
if (isFloat)
libCall = spv::FMid3AMD;
else {
if (isUnsigned)
libCall = spv::UMid3AMD;
else
libCall = spv::SMid3AMD;
}
break;
case glslang::EOpInterpolateAtVertex:
Fix issue #1376: SPV_AMD_gpu_shader_half_float is not emitted under vulkan1.1 Interpolation functions, using 16-bit interpolants, should always emit this extension string regardless of vulkan enviornment version.
2018-05-17 05:51:28 +00:00
if (typeProxy == glslang::EbtFloat16)
builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float);
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
libCall = spv::InterpolateAtVertexAMD;
break;
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
case glslang::EOpBarrier:
{
// This is for the extended controlBarrier function, with four operands.
// The unextended barrier() goes through createNoArgOperation.
assert(operands.size() == 4);
unsigned int executionScope = builder.getConstantScalar(operands[0]);
unsigned int memoryScope = builder.getConstantScalar(operands[1]);
unsigned int semantics = builder.getConstantScalar(operands[2]) | builder.getConstantScalar(operands[3]);
builder.createControlBarrier((spv::Scope)executionScope, (spv::Scope)memoryScope, (spv::MemorySemanticsMask)semantics);
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
if (semantics & (spv::MemorySemanticsMakeAvailableKHRMask |
spv::MemorySemanticsMakeVisibleKHRMask |
spv::MemorySemanticsOutputMemoryKHRMask |
spv::MemorySemanticsVolatileMask)) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
}
if (glslangIntermediate->usingVulkanMemoryModel() && (executionScope == spv::ScopeDevice || memoryScope == spv::ScopeDevice)) {
builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR);
}
return 0;
}
break;
case glslang::EOpMemoryBarrier:
{
// This is for the extended memoryBarrier function, with three operands.
// The unextended memoryBarrier() goes through createNoArgOperation.
assert(operands.size() == 3);
unsigned int memoryScope = builder.getConstantScalar(operands[0]);
unsigned int semantics = builder.getConstantScalar(operands[1]) | builder.getConstantScalar(operands[2]);
builder.createMemoryBarrier((spv::Scope)memoryScope, (spv::MemorySemanticsMask)semantics);
Add gl_SemanticsVolatile to GL_KHR_memory_scope_semantics, and make volatile-qualified atomics generate MemorySemanticsVolatile when using the Vulkan memory model
2019-06-14 14:56:28 +00:00
if (semantics & (spv::MemorySemanticsMakeAvailableKHRMask |
spv::MemorySemanticsMakeVisibleKHRMask |
spv::MemorySemanticsOutputMemoryKHRMask |
spv::MemorySemanticsVolatileMask)) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
}
if (glslangIntermediate->usingVulkanMemoryModel() && memoryScope == spv::ScopeDevice) {
builder.addCapability(spv::CapabilityVulkanMemoryModelDeviceScopeKHR);
}
return 0;
}
break;
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EOpReportIntersectionNV:
{
typeId = builder.makeBoolType();
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
opCode = spv::OpReportIntersectionNV;
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
}
break;
case glslang::EOpTraceNV:
{
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
builder.createNoResultOp(spv::OpTraceNV, operands);
return 0;
}
break;
case glslang::EOpExecuteCallableNV:
{
builder.createNoResultOp(spv::OpExecuteCallableNV, operands);
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
return 0;
}
break;
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
case glslang::EOpWritePackedPrimitiveIndices4x8NV:
builder.createNoResultOp(spv::OpWritePackedPrimitiveIndices4x8NV, operands);
return 0;
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
case glslang::EOpCooperativeMatrixMulAdd:
opCode = spv::OpCooperativeMatrixMulAddNV;
break;
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#endif // GLSLANG_WEB
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
return 0;
}
spv::Id id = 0;
SPV: Fix extra operands when using the "struct" form of a built-in taking an 'out' param.
2015-12-07 02:29:11 +00:00
if (libCall >= 0) {
Avoid read past end of operands vector for EOpFrexp When emitting SPIR-V code for frexp, avoid access beyond the end of the operands vector. When constructing the OpExtInst, construct a new arguments vector instead of modifying the existing operands vector. In the case of OpFrexp, well need that last operand later on to generate the store. Fixes https://github.com/KhronosGroup/glslang/issues/110 Change-Id: Ibc380fadf5e600ac491932e9ecef7afe2d72fd7f
2015-12-07 21:17:06 +00:00
// Use an extended instruction from the standard library.
// Construct the call arguments, without modifying the original operands vector.
// We might need the remaining arguments, e.g. in the EOpFrexp case.
std::vector<spv::Id> callArguments(operands.begin(), operands.begin() + consumedOperands);
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
id = builder.createBuiltinCall(typeId, extBuiltins >= 0 ? extBuiltins : stdBuiltins, libCall, callArguments);
[HLSL/Spir-V] fix for incorrect spir-v on int dot(int, int) Decomposes OpDot into corresponding multiplies and additions.
2018-11-15 09:21:36 +00:00
} else if (opCode == spv::OpDot && !isFloat) {
// int dot(int, int)
// NOTE: never called for scalar/vector1, this is turned into simple mul before this can be reached
const int componentCount = builder.getNumComponents(operands[0]);
spv::Id mulOp = builder.createBinOp(spv::OpIMul, builder.getTypeId(operands[0]), operands[0], operands[1]);
builder.setPrecision(mulOp, precision);
id = builder.createCompositeExtract(mulOp, typeId, 0);
for (int i = 1; i < componentCount; ++i) {
builder.setPrecision(id, precision);
id = builder.createBinOp(spv::OpIAdd, typeId, id, builder.createCompositeExtract(operands[0], typeId, i));
}
SPV: Fix extra operands when using the "struct" form of a built-in taking an 'out' param.
2015-12-07 02:29:11 +00:00
} else {
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
switch (consumedOperands) {
Final round for line endings.
2015-06-26 22:58:36 +00:00
case 0:
// should all be handled by visitAggregate and createNoArgOperation
assert(0);
return 0;
case 1:
// should all be handled by createUnaryOperation
assert(0);
return 0;
case 2:
id = builder.createBinOp(opCode, typeId, operands[0], operands[1]);
break;
default:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
// anything 3 or over doesn't have l-value operands, so all should be consumed
assert(consumedOperands == operands.size());
id = builder.createOp(opCode, typeId, operands);
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
}
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#ifndef GLSLANG_WEB
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
// Decode the return types that were structures
switch (op) {
case glslang::EOpAddCarry:
case glslang::EOpSubBorrow:
builder.createStore(builder.createCompositeExtract(id, typeId0, 1), operands[2]);
id = builder.createCompositeExtract(id, typeId0, 0);
break;
case glslang::EOpUMulExtended:
case glslang::EOpIMulExtended:
builder.createStore(builder.createCompositeExtract(id, typeId0, 0), operands[3]);
builder.createStore(builder.createCompositeExtract(id, typeId0, 1), operands[2]);
break;
case glslang::EOpFrexp:
HLSL: Fix an issue of frexp(). The "exp" parameter is floating-point type in HLSL intrinsic while it is integer type in GLSL built-in function.
2017-03-29 09:12:40 +00:00
{
assert(operands.size() == 2);
if (builder.isFloatType(builder.getScalarTypeId(typeId1))) {
// "exp" is floating-point type (from HLSL intrinsic)
spv::Id member1 = builder.createCompositeExtract(id, frexpIntType, 1);
member1 = builder.createUnaryOp(spv::OpConvertSToF, typeId1, member1);
builder.createStore(member1, operands[1]);
} else
// "exp" is integer type (from GLSL built-in function)
builder.createStore(builder.createCompositeExtract(id, frexpIntType, 1), operands[1]);
id = builder.createCompositeExtract(id, typeId0, 0);
}
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
break;
default:
break;
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
SPV: RelaxedPrecision: Plumb this through the full AST->SPV translator.
2016-02-02 19:37:46 +00:00
return builder.setPrecision(id, precision);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
// Intrinsics with no arguments (or no return value, and no precision).
spv::Id TGlslangToSpvTraverser::createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#ifndef GLSLANG_WEB
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
// GLSL memory barriers use queuefamily scope in new model, device scope in old model
spv::Scope memoryBarrierScope = glslangIntermediate->usingVulkanMemoryModel() ? spv::ScopeQueueFamilyKHR : spv::ScopeDevice;
Final round for line endings.
2015-06-26 22:58:36 +00:00
switch (op) {
case glslang::EOpEmitVertex:
builder.createNoResultOp(spv::OpEmitVertex);
return 0;
case glslang::EOpEndPrimitive:
builder.createNoResultOp(spv::OpEndPrimitive);
return 0;
case glslang::EOpBarrier:
SPV: Change barrier emission to conform to Khronos decisions. The memory model group agreed to these definitions for how to map GLSL barrier, memoryBarrier, etc. With HLSL following suit.
2017-12-11 11:02:24 +00:00
if (glslangIntermediate->getStage() == EShLangTessControl) {
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
if (glslangIntermediate->usingVulkanMemoryModel()) {
builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup,
spv::MemorySemanticsOutputMemoryKHRMask |
spv::MemorySemanticsAcquireReleaseMask);
builder.addCapability(spv::CapabilityVulkanMemoryModelKHR);
} else {
builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeInvocation, spv::MemorySemanticsMaskNone);
}
SPV: Change barrier emission to conform to Khronos decisions. The memory model group agreed to these definitions for how to map GLSL barrier, memoryBarrier, etc. With HLSL following suit.
2017-12-11 11:02:24 +00:00
} else {
builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup,
spv::MemorySemanticsWorkgroupMemoryMask |
spv::MemorySemanticsAcquireReleaseMask);
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
return 0;
case glslang::EOpMemoryBarrier:
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsAllMemory |
spv::MemorySemanticsAcquireReleaseMask);
Final round for line endings.
2015-06-26 22:58:36 +00:00
return 0;
case glslang::EOpMemoryBarrierAtomicCounter:
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsAtomicCounterMemoryMask |
spv::MemorySemanticsAcquireReleaseMask);
Final round for line endings.
2015-06-26 22:58:36 +00:00
return 0;
case glslang::EOpMemoryBarrierBuffer:
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsUniformMemoryMask |
spv::MemorySemanticsAcquireReleaseMask);
Final round for line endings.
2015-06-26 22:58:36 +00:00
return 0;
case glslang::EOpMemoryBarrierImage:
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsImageMemoryMask |
spv::MemorySemanticsAcquireReleaseMask);
Final round for line endings.
2015-06-26 22:58:36 +00:00
return 0;
case glslang::EOpMemoryBarrierShared:
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
builder.createMemoryBarrier(memoryBarrierScope, spv::MemorySemanticsWorkgroupMemoryMask |
spv::MemorySemanticsAcquireReleaseMask);
Final round for line endings.
2015-06-26 22:58:36 +00:00
return 0;
case glslang::EOpGroupMemoryBarrier:
SPV: Change barrier emission to conform to Khronos decisions. The memory model group agreed to these definitions for how to map GLSL barrier, memoryBarrier, etc. With HLSL following suit.
2017-12-11 11:02:24 +00:00
builder.createMemoryBarrier(spv::ScopeWorkgroup, spv::MemorySemanticsAllMemory |
spv::MemorySemanticsAcquireReleaseMask);
Final round for line endings.
2015-06-26 22:58:36 +00:00
return 0;
Add HLSL memory barrier intrinsics, fix dst, add lit & EvaluateAttributeSnapped
2016-06-15 15:50:24 +00:00
case glslang::EOpAllMemoryBarrierWithGroupSync:
SPV: HLSL: Move to correct HLSL barrier semantics, per Khronos recommendation.
2017-12-13 05:50:53 +00:00
builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeDevice,
SPV: Change barrier emission to conform to Khronos decisions. The memory model group agreed to these definitions for how to map GLSL barrier, memoryBarrier, etc. With HLSL following suit.
2017-12-11 11:02:24 +00:00
spv::MemorySemanticsAllMemory |
SPV: HLSL: Move to correct HLSL barrier semantics, per Khronos recommendation.
2017-12-13 05:50:53 +00:00
spv::MemorySemanticsAcquireReleaseMask);
Add HLSL memory barrier intrinsics, fix dst, add lit & EvaluateAttributeSnapped
2016-06-15 15:50:24 +00:00
return 0;
SPV: HLSL: Move to correct HLSL barrier semantics, per Khronos recommendation.
2017-12-13 05:50:53 +00:00
case glslang::EOpDeviceMemoryBarrier:
builder.createMemoryBarrier(spv::ScopeDevice, spv::MemorySemanticsUniformMemoryMask |
spv::MemorySemanticsImageMemoryMask |
spv::MemorySemanticsAcquireReleaseMask);
return 0;
case glslang::EOpDeviceMemoryBarrierWithGroupSync:
builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeDevice, spv::MemorySemanticsUniformMemoryMask |
spv::MemorySemanticsImageMemoryMask |
spv::MemorySemanticsAcquireReleaseMask);
Add HLSL memory barrier intrinsics, fix dst, add lit & EvaluateAttributeSnapped
2016-06-15 15:50:24 +00:00
return 0;
case glslang::EOpWorkgroupMemoryBarrier:
SPV: HLSL: Move to correct HLSL barrier semantics, per Khronos recommendation.
2017-12-13 05:50:53 +00:00
builder.createMemoryBarrier(spv::ScopeWorkgroup, spv::MemorySemanticsWorkgroupMemoryMask |
spv::MemorySemanticsAcquireReleaseMask);
Add HLSL memory barrier intrinsics, fix dst, add lit & EvaluateAttributeSnapped
2016-06-15 15:50:24 +00:00
return 0;
case glslang::EOpWorkgroupMemoryBarrierWithGroupSync:
SPV: HLSL: Move to correct HLSL barrier semantics, per Khronos recommendation.
2017-12-13 05:50:53 +00:00
builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup,
spv::MemorySemanticsWorkgroupMemoryMask |
spv::MemorySemanticsAcquireReleaseMask);
Add HLSL memory barrier intrinsics, fix dst, add lit & EvaluateAttributeSnapped
2016-06-15 15:50:24 +00:00
return 0;
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EOpSubgroupBarrier:
builder.createControlBarrier(spv::ScopeSubgroup, spv::ScopeSubgroup, spv::MemorySemanticsAllMemory |
spv::MemorySemanticsAcquireReleaseMask);
return spv::NoResult;
case glslang::EOpSubgroupMemoryBarrier:
builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsAllMemory |
spv::MemorySemanticsAcquireReleaseMask);
return spv::NoResult;
case glslang::EOpSubgroupMemoryBarrierBuffer:
builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsUniformMemoryMask |
spv::MemorySemanticsAcquireReleaseMask);
return spv::NoResult;
case glslang::EOpSubgroupMemoryBarrierImage:
builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsImageMemoryMask |
spv::MemorySemanticsAcquireReleaseMask);
return spv::NoResult;
case glslang::EOpSubgroupMemoryBarrierShared:
builder.createMemoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsWorkgroupMemoryMask |
spv::MemorySemanticsAcquireReleaseMask);
return spv::NoResult;
case glslang::EOpSubgroupElect: {
std::vector<spv::Id> operands;
return createSubgroupOperation(op, typeId, operands, glslang::EbtVoid);
}
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
case glslang::EOpTime:
{
std::vector<spv::Id> args; // Dummy arguments
spv::Id id = builder.createBuiltinCall(typeId, getExtBuiltins(spv::E_SPV_AMD_gcn_shader), spv::TimeAMD, args);
return builder.setPrecision(id, precision);
}
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
case glslang::EOpIgnoreIntersectionNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
builder.createNoResultOp(spv::OpIgnoreIntersectionNV);
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
return 0;
case glslang::EOpTerminateRayNV:
Multiple features and misc fixes for final raytracing non experimental version.
2018-10-22 23:41:44 +00:00
builder.createNoResultOp(spv::OpTerminateRayNV);
Add-support-for-SPV_NVX_raytracing
2018-09-19 18:42:24 +00:00
return 0;
Support GL_ARB_fragment_shader_interlock
2019-06-03 16:33:50 +00:00
case glslang::EOpBeginInvocationInterlock:
builder.createNoResultOp(spv::OpBeginInvocationInterlockEXT);
return 0;
case glslang::EOpEndInvocationInterlock:
builder.createNoResultOp(spv::OpEndInvocationInterlockEXT);
return 0;
Implement GL_EXT_demote_to_helper_invocation
2019-07-01 14:23:23 +00:00
case glslang::EOpIsHelperInvocation:
{
std::vector<spv::Id> args; // Dummy arguments
OpIsHelperInvocationEXT should declare relevant SPV extension and capability
2019-07-15 06:57:20 +00:00
builder.addExtension(spv::E_SPV_EXT_demote_to_helper_invocation);
builder.addCapability(spv::CapabilityDemoteToHelperInvocationEXT);
return builder.createOp(spv::OpIsHelperInvocationEXT, typeId, args);
Implement GL_EXT_demote_to_helper_invocation
2019-07-01 14:23:23 +00:00
}
Add support for SPV_KHR_shader_clock
2019-07-11 01:14:38 +00:00
case glslang::EOpReadClockSubgroupKHR: {
std::vector<spv::Id> args;
args.push_back(builder.makeUintConstant(spv::ScopeSubgroup));
builder.addExtension(spv::E_SPV_KHR_shader_clock);
builder.addCapability(spv::CapabilityShaderClockKHR);
return builder.createOp(spv::OpReadClockKHR, typeId, args);
}
case glslang::EOpReadClockDeviceKHR: {
std::vector<spv::Id> args;
args.push_back(builder.makeUintConstant(spv::ScopeDevice));
builder.addExtension(spv::E_SPV_KHR_shader_clock);
builder.addCapability(spv::CapabilityShaderClockKHR);
return builder.createOp(spv::OpReadClockKHR, typeId, args);
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
break;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#endif
logger->missingFunctionality("unknown operation with no arguments");
return 0;
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
spv::Id TGlslangToSpvTraverser::getSymbolId(const glslang::TIntermSymbol* symbol)
{
Performance: Simple upgrade/cleanup of stl c++11 containers (10% perf. increase).
2015-07-19 04:34:27 +00:00
auto iter = symbolValues.find(symbol->getId());
Final round for line endings.
2015-06-26 22:58:36 +00:00
spv::Id id;
if (symbolValues.end() != iter) {
id = iter->second;
return id;
}
// it was not found, create it
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
spv::BuiltIn builtIn = TranslateBuiltInDecoration(symbol->getQualifier().builtIn, false);
auto forcedType = getForcedType(builtIn, symbol->getType());
id = createSpvVariable(symbol, forcedType.first);
Final round for line endings.
2015-06-26 22:58:36 +00:00
symbolValues[symbol->getId()] = id;
AST/SPV: Fix #930: translate uvec4 <-> uint64 for SubgroupGeMask et. al. On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask, SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V is defined as vectors of 32-bit ints. The declaration type has to be translated in the opposite direction.
2019-06-17 14:38:35 +00:00
if (forcedType.second != spv::NoType)
forceType[id] = forcedType.second;
Final round for line endings.
2015-06-26 22:58:36 +00:00
SPV: Certain decorations are missing for structure-typed in/out variables.
2016-06-29 07:03:44 +00:00
if (symbol->getBasicType() != glslang::EbtBlock) {
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
builder.addDecoration(id, TranslatePrecisionDecoration(symbol->getType()));
builder.addDecoration(id, TranslateInterpolationDecoration(symbol->getType().getQualifier()));
builder.addDecoration(id, TranslateAuxiliaryStorageDecoration(symbol->getType().getQualifier()));
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
addMeshNVDecoration(id, /*member*/ -1, symbol->getType().getQualifier());
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
if (symbol->getQualifier().hasComponent())
builder.addDecoration(id, spv::DecorationComponent, symbol->getQualifier().layoutComponent);
if (symbol->getQualifier().hasIndex())
builder.addDecoration(id, spv::DecorationIndex, symbol->getQualifier().layoutIndex);
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
#endif
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
if (symbol->getType().getQualifier().hasSpecConstantId())
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
builder.addDecoration(id, spv::DecorationSpecId, symbol->getType().getQualifier().layoutSpecConstantId);
SPV: Further refine OpenGL vs. Vulkan differences for SPIR-V. Includes adding test cases to verify the differences.
2016-07-07 23:46:42 +00:00
// atomic counters use this:
if (symbol->getQualifier().hasOffset())
builder.addDecoration(id, spv::DecorationOffset, symbol->getQualifier().layoutOffset);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
Fix missing location decoration for structures put directly on input/output interfaces Spec for decorating the OpVariable: "The remaining variables listed by OpEntryPoint with the Input or Output storage class form the user-defined variable interface. These variables must be identified with a Location decoration" Spec for decorating struct type: "The layout of a structure type used as an Input or Output depends on whether it is also a Block (i.e. has a Block decoration). If it is a not a Block, then the structure type must have a Location decoration"
2016-05-18 16:09:17 +00:00
if (symbol->getQualifier().hasLocation())
builder.addDecoration(id, spv::DecorationLocation, symbol->getQualifier().layoutLocation);
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
builder.addDecoration(id, TranslateInvariantDecoration(symbol->getType().getQualifier()));
SPV: Use heuristic to avoid geometry multi-streams when possible.
2016-03-04 05:29:11 +00:00
if (symbol->getQualifier().hasStream() && glslangIntermediate->isMultiStream()) {
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
builder.addCapability(spv::CapabilityGeometryStreams);
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.addDecoration(id, spv::DecorationStream, symbol->getQualifier().layoutStream);
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (symbol->getQualifier().hasSet())
builder.addDecoration(id, spv::DecorationDescriptorSet, symbol->getQualifier().layoutSet);
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
else if (IsDescriptorResource(symbol->getType())) {
// default to 0
builder.addDecoration(id, spv::DecorationDescriptorSet, 0);
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (symbol->getQualifier().hasBinding())
builder.addDecoration(id, spv::DecorationBinding, symbol->getQualifier().layoutBinding);
Update to newest SPIRV-Tools. It checks the Vulkan rule that all resources must have a set and binding decoration, which many tests fail. So add code to set a default value of zero. Also disable PCH for Ninja builds.
2018-12-12 02:53:59 +00:00
else if (IsDescriptorResource(symbol->getType())) {
// default to 0
builder.addDecoration(id, spv::DecorationBinding, 0);
}
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
if (symbol->getQualifier().hasAttachment())
builder.addDecoration(id, spv::DecorationInputAttachmentIndex, symbol->getQualifier().layoutAttachment);
Final round for line endings.
2015-06-26 22:58:36 +00:00
if (glslangIntermediate->getXfbMode()) {
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
builder.addCapability(spv::CapabilityTransformFeedback);
SPV: Plumb through XFB buffer and stride information. Also, only emit this XFB information where the SPIR-V spec says it should be emitted: essentially, on objects. This and the previous commit together fix #1185.
2017-12-15 13:21:46 +00:00
if (symbol->getQualifier().hasXfbBuffer()) {
Final round for line endings.
2015-06-26 22:58:36 +00:00
builder.addDecoration(id, spv::DecorationXfbBuffer, symbol->getQualifier().layoutXfbBuffer);
SPV: Plumb through XFB buffer and stride information. Also, only emit this XFB information where the SPIR-V spec says it should be emitted: essentially, on objects. This and the previous commit together fix #1185.
2017-12-15 13:21:46 +00:00
unsigned stride = glslangIntermediate->getXfbStride(symbol->getQualifier().layoutXfbBuffer);
if (stride != glslang::TQualifier::layoutXfbStrideEnd)
builder.addDecoration(id, spv::DecorationXfbStride, stride);
}
if (symbol->getQualifier().hasXfbOffset())
builder.addDecoration(id, spv::DecorationOffset, symbol->getQualifier().layoutXfbOffset);
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
// add built-in variable decoration
if (builtIn != spv::BuiltInMax) {
builder.addDecoration(id, spv::DecorationBuiltIn, (int)builtIn);
}
#ifndef GLSLANG_WEB
SPV: Add support for memory qualifiers.
2016-02-21 12:59:01 +00:00
if (symbol->getType().isImage()) {
std::vector<spv::Decoration> memory;
GL_KHR_memory_scope_semantics
2018-09-05 15:11:41 +00:00
TranslateMemoryDecoration(symbol->getType().getQualifier(), memory, glslangIntermediate->usingVulkanMemoryModel());
SPV: Add support for memory qualifiers.
2016-02-21 12:59:01 +00:00
for (unsigned int i = 0; i < memory.size(); ++i)
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
builder.addDecoration(id, memory[i]);
SPV: Add support for memory qualifiers.
2016-02-21 12:59:01 +00:00
}
GLSL/SPV: Implement SPV_EXT_descriptor_indexing and GL_EXT_nonuniform_qualifier
2018-04-05 17:25:02 +00:00
// nonuniform
builder.addDecoration(id, TranslateNonUniformDecoration(symbol->getType().getQualifier()));
Add support for SPV_NV_sample_mask_override_coverage
2016-12-20 00:29:34 +00:00
if (builtIn == spv::BuiltInSampleMask) {
spv::Decoration decoration;
// GL_NV_sample_mask_override_coverage extension
if (glslangIntermediate->getLayoutOverrideCoverage())
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
decoration = (spv::Decoration)spv::DecorationOverrideCoverageNV;
Add support for SPV_NV_sample_mask_override_coverage
2016-12-20 00:29:34 +00:00
else
decoration = (spv::Decoration)spv::DecorationMax;
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
builder.addDecoration(id, decoration);
Add support for SPV_NV_sample_mask_override_coverage
2016-12-20 00:29:34 +00:00
if (decoration != spv::DecorationMax) {
spirv: Generate missing SampleMaskOverrideCoverageNV capability op The SampleMaskOverrideCoverageNV capability should be declared when gl_SampleMask is used.
2019-07-12 21:33:02 +00:00
builder.addCapability(spv::CapabilitySampleMaskOverrideCoverageNV);
Add support for SPV_NV_sample_mask_override_coverage
2016-12-20 00:29:34 +00:00
builder.addExtension(spv::E_SPV_NV_sample_mask_override_coverage);
}
}
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
else if (builtIn == spv::BuiltInLayer) {
// SPV_NV_viewport_array2 extension
SPV: 1st pass implementation of SPV_EXT_shader_viewport_index_layer.
2017-08-11 19:07:17 +00:00
if (symbol->getQualifier().layoutViewportRelative) {
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
builder.addDecoration(id, (spv::Decoration)spv::DecorationViewportRelativeNV);
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
builder.addCapability(spv::CapabilityShaderViewportMaskNV);
builder.addExtension(spv::E_SPV_NV_viewport_array2);
}
SPV: 1st pass implementation of SPV_EXT_shader_viewport_index_layer.
2017-08-11 19:07:17 +00:00
if (symbol->getQualifier().layoutSecondaryViewportRelativeOffset != -2048) {
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
builder.addDecoration(id, (spv::Decoration)spv::DecorationSecondaryViewportRelativeNV,
symbol->getQualifier().layoutSecondaryViewportRelativeOffset);
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
builder.addCapability(spv::CapabilityShaderStereoViewNV);
builder.addExtension(spv::E_SPV_NV_stereo_view_rendering);
}
}
Add support for SPV_NV_geometry_shader_passthrough
2016-12-20 21:28:52 +00:00
if (symbol->getQualifier().layoutPassthrough) {
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
builder.addDecoration(id, spv::DecorationPassthroughNV);
support SPV_NV_viewport_array2 and SPV_NV_stereo_view_rendering
2017-01-13 09:10:53 +00:00
builder.addCapability(spv::CapabilityGeometryShaderPassthroughNV);
Add support for SPV_NV_geometry_shader_passthrough
2016-12-20 21:28:52 +00:00
builder.addExtension(spv::E_SPV_NV_geometry_shader_passthrough);
}
Add-support-for-SPV_NV_fragment_shader_barycentric
2018-09-19 18:39:56 +00:00
if (symbol->getQualifier().pervertexNV) {
builder.addDecoration(id, spv::DecorationPerVertexNV);
builder.addCapability(spv::CapabilityFragmentBarycentricNV);
builder.addExtension(spv::E_SPV_NV_fragment_shader_barycentric);
}
Add support for SPV_NV_sample_mask_override_coverage
2016-12-20 00:29:34 +00:00
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
if (glslangIntermediate->getHlslFunctionality1() && symbol->getType().getQualifier().semanticName != nullptr) {
builder.addExtension("SPV_GOOGLE_hlsl_functionality1");
builder.addDecoration(id, (spv::Decoration)spv::DecorationHlslSemanticGOOGLE,
symbol->getType().getQualifier().semanticName);
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
Web: Remove/rationalize a set of *_EXTENSIONS, using GLSLANG_WEB. Focus was on the front end (not SPIR-V), minus the grammar. Reduces #ifdef count by around 320 and makes the web build 270K smaller, which is about 90% the target size. The grammar and scanner will be another step, as will the SPIR-V backend. This makes heavy use of methods #ifdef'd to return false as a global way of turning off code, relying on C++ DCE to do the rest.
2019-08-01 09:28:08 +00:00
if (symbol->isReference()) {
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
builder.addDecoration(id, symbol->getType().getQualifier().restrict ? spv::DecorationRestrictPointerEXT : spv::DecorationAliasedPointerEXT);
}
Web: Make switched methods all be non-virtual, more web-dependent code, added a few more HLSL flag tests. This was mostly focused on the SPV generator. Saves about 17K.
2019-08-11 13:41:45 +00:00
#endif
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
return id;
SPV: Address superset of issue #151: missing OpCapability instructions. This commit adds: CapabilityGeometryPointSize CapabilityTessellationPointSize CapabilityClipDistance CapabilityCullDistance CapabilityMultiViewport CapabilityTransformFeedback CapabilityGeometryStreams CapabilityDerivativeControl CapabilityInterpolationFunction
2016-02-01 20:45:25 +00:00
}
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
// add per-primitive, per-view. per-task decorations to a struct member (member >= 0) or an object
void TGlslangToSpvTraverser::addMeshNVDecoration(spv::Id id, int member, const glslang::TQualifier& qualifier)
{
if (member >= 0) {
Allow GL_NV_mesh_shader in fragment shaders for perprimitiveNV - Emit relevant capability/extension for use of perprimitiveNV in fragment shader - Remove redundant checks for mesh shader qualifiers in glslang.y - Add profile version check for use of extension GL_NV_mesh_shader - Add a new gtest for use of perprimitiveNV in fragment shader
2018-10-26 06:50:59 +00:00
if (qualifier.perPrimitiveNV) {
// Need to add capability/extension for fragment shader.
// Mesh shader already adds this by default.
if (glslangIntermediate->getStage() == EShLangFragment) {
builder.addCapability(spv::CapabilityMeshShadingNV);
builder.addExtension(spv::E_SPV_NV_mesh_shader);
}
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerPrimitiveNV);
Allow GL_NV_mesh_shader in fragment shaders for perprimitiveNV - Emit relevant capability/extension for use of perprimitiveNV in fragment shader - Remove redundant checks for mesh shader qualifiers in glslang.y - Add profile version check for use of extension GL_NV_mesh_shader - Add a new gtest for use of perprimitiveNV in fragment shader
2018-10-26 06:50:59 +00:00
}
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
if (qualifier.perViewNV)
builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerViewNV);
if (qualifier.perTaskNV)
builder.addMemberDecoration(id, (unsigned)member, spv::DecorationPerTaskNV);
} else {
Allow GL_NV_mesh_shader in fragment shaders for perprimitiveNV - Emit relevant capability/extension for use of perprimitiveNV in fragment shader - Remove redundant checks for mesh shader qualifiers in glslang.y - Add profile version check for use of extension GL_NV_mesh_shader - Add a new gtest for use of perprimitiveNV in fragment shader
2018-10-26 06:50:59 +00:00
if (qualifier.perPrimitiveNV) {
// Need to add capability/extension for fragment shader.
// Mesh shader already adds this by default.
if (glslangIntermediate->getStage() == EShLangFragment) {
builder.addCapability(spv::CapabilityMeshShadingNV);
builder.addExtension(spv::E_SPV_NV_mesh_shader);
}
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
builder.addDecoration(id, spv::DecorationPerPrimitiveNV);
Allow GL_NV_mesh_shader in fragment shaders for perprimitiveNV - Emit relevant capability/extension for use of perprimitiveNV in fragment shader - Remove redundant checks for mesh shader qualifiers in glslang.y - Add profile version check for use of extension GL_NV_mesh_shader - Add a new gtest for use of perprimitiveNV in fragment shader
2018-10-26 06:50:59 +00:00
}
Add-support-for-SPV_NV_mesh_shader
2018-09-19 18:41:59 +00:00
if (qualifier.perViewNV)
builder.addDecoration(id, spv::DecorationPerViewNV);
if (qualifier.perTaskNV)
builder.addDecoration(id, spv::DecorationPerTaskNV);
}
}
#endif
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
// Make a full tree of instructions to build a SPIR-V specialization constant,
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
// or regular constant if possible.
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
//
// TBD: this is not yet done, nor verified to be the best design, it does do the leaf symbols though
//
// Recursively walk the nodes. The nodes form a tree whose leaves are
// regular constants, which themselves are trees that createSpvConstant()
// recursively walks. So, this function walks the "top" of the tree:
// - emit specialization constant-building instructions for specConstant
// - when running into a non-spec-constant, switch to createSpvConstant()
Support specialization composite constants Fix issue #163, support creation and reference of composite type specialization constants. e.g.: ``` layout(constant_id = 200) const float myfloat = 1.25; layout(constant_id = 201) const int myint = 14; struct structtype { float f; int i; }; const structtype outer_struct_var = {myfloat, myint}; void main(){} ``` generated code (use glslangValidator): ``` // Module Version 10000 // Generated by (magic number): 80001 // Id's are bound by 12 Capability Shader 1: ExtInstImport "GLSL.std.450" MemoryModel Logical GLSL450 EntryPoint Vertex 4 "main" Source GLSL 450 Name 4 "main" Name 10 "structtype" MemberName 10(structtype) 0 "f" MemberName 10(structtype) 1 "i" Decorate 7 SpecId 200 Decorate 9 SpecId 201 2: TypeVoid 3: TypeFunction 2 6: TypeFloat 32 7: 6(float) SpecConstant 1067450368 8: TypeInt 32 1 9: 8(int) SpecConstant 14 10(structtype): TypeStruct 6(float) 8(int) 11:10(structtype) SpecConstantComposite 7 9 4(main): 2 Function None 3 5: Label Return FunctionEnd ``` Rname two function names to match their functionalities. 1) Rename `GlslangToSpvTraverser::createSpvSpecConstant()` to `createSpvConstant()`; 2) Rename `GlslangToSpvTraverser::createSpvConstant()` to `createSpvConstantFromConstUnionArray()` Add function `GlslangToSpvTraverser::createSpvConstantFromSubTree()` to handle constant creation from sub trees (e.g.: specialization constants). Related PR: #208
2016-03-21 13:51:37 +00:00
spv::Id TGlslangToSpvTraverser::createSpvConstant(const glslang::TIntermTyped& node)
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
{
Front-end infrastructure: Encapsulate semantic-level questions/actions about const/temp. Much about const or temp is mechanical, about actual declaration, while much is semantic, about something higher level. This commit checks every use everywhere, and for the high-level ones, substitutes an encapsulated version instead.
2016-03-20 06:46:02 +00:00
assert(node.getQualifier().isConstant());
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
Built-in values declared as specialization constant Support declaring built-in values as spec constants. Refine the code in createSpvConstant().
2016-04-04 03:55:17 +00:00
// Handle front-end constants first (non-specialization constants).
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
if (! node.getQualifier().specConstant) {
// hand off to the non-spec-constant path
assert(node.getAsConstantUnion() != nullptr || node.getAsSymbolNode() != nullptr);
int nextConst = 0;
Support specialization composite constants Fix issue #163, support creation and reference of composite type specialization constants. e.g.: ``` layout(constant_id = 200) const float myfloat = 1.25; layout(constant_id = 201) const int myint = 14; struct structtype { float f; int i; }; const structtype outer_struct_var = {myfloat, myint}; void main(){} ``` generated code (use glslangValidator): ``` // Module Version 10000 // Generated by (magic number): 80001 // Id's are bound by 12 Capability Shader 1: ExtInstImport "GLSL.std.450" MemoryModel Logical GLSL450 EntryPoint Vertex 4 "main" Source GLSL 450 Name 4 "main" Name 10 "structtype" MemberName 10(structtype) 0 "f" MemberName 10(structtype) 1 "i" Decorate 7 SpecId 200 Decorate 9 SpecId 201 2: TypeVoid 3: TypeFunction 2 6: TypeFloat 32 7: 6(float) SpecConstant 1067450368 8: TypeInt 32 1 9: 8(int) SpecConstant 14 10(structtype): TypeStruct 6(float) 8(int) 11:10(structtype) SpecConstantComposite 7 9 4(main): 2 Function None 3 5: Label Return FunctionEnd ``` Rname two function names to match their functionalities. 1) Rename `GlslangToSpvTraverser::createSpvSpecConstant()` to `createSpvConstant()`; 2) Rename `GlslangToSpvTraverser::createSpvConstant()` to `createSpvConstantFromConstUnionArray()` Add function `GlslangToSpvTraverser::createSpvConstantFromSubTree()` to handle constant creation from sub trees (e.g.: specialization constants). Related PR: #208
2016-03-21 13:51:37 +00:00
return createSpvConstantFromConstUnionArray(node.getType(), node.getAsConstantUnion() ? node.getAsConstantUnion()->getConstArray() : node.getAsSymbolNode()->getConstArray(),
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
nextConst, false);
}
// We now know we have a specialization constant to build
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#ifndef GLSLANG_WEB
Front-end: Complete GL_ARB_compute_shader implementation. Core compute shaders were working, but the extension wasn't implemented.
2016-05-31 01:29:40 +00:00
// gl_WorkGroupSize is a special case until the front-end handles hierarchical specialization constants,
Built-in values declared as specialization constant Support declaring built-in values as spec constants. Refine the code in createSpvConstant().
2016-04-04 03:55:17 +00:00
// even then, it's specialization ids are handled by special case syntax in GLSL: layout(local_size_x = ...
if (node.getType().getQualifier().builtIn == glslang::EbvWorkGroupSize) {
std::vector<spv::Id> dimConstId;
for (int dim = 0; dim < 3; ++dim) {
bool specConst = (glslangIntermediate->getLocalSizeSpecId(dim) != glslang::TQualifier::layoutNotSet);
dimConstId.push_back(builder.makeUintConstant(glslangIntermediate->getLocalSize(dim), specConst));
Implement SPV_GOOGLE_hlsl_functionality1. Enabled via -fhlsl_functionality1
2018-03-08 01:05:55 +00:00
if (specConst) {
builder.addDecoration(dimConstId.back(), spv::DecorationSpecId,
glslangIntermediate->getLocalSizeSpecId(dim));
}
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
}
Built-in values declared as specialization constant Support declaring built-in values as spec constants. Refine the code in createSpvConstant().
2016-04-04 03:55:17 +00:00
return builder.makeCompositeConstant(builder.makeVectorType(builder.makeUintType(32), 3), dimConstId, true);
SPV: Implement Vulkan version of GLSL (KHR_vulkan_glsl).
2016-02-16 03:58:50 +00:00
}
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#endif
Built-in values declared as specialization constant Support declaring built-in values as spec constants. Refine the code in createSpvConstant().
2016-04-04 03:55:17 +00:00
// An AST node labelled as specialization constant should be a symbol node.
// Its initializer should either be a sub tree with constant nodes, or a constant union array.
if (auto* sn = node.getAsSymbolNode()) {
Add names for composite constants in SPIR-V Consider the following code: layout(constant_id=0) const int Y = 1; layout(constant_id=1) const int Z = 2; layout(constant_id=3) const int X = Y + Z; Previously, it would produce SPIR-V decorations like this: Decorate 21(Y) SpecId 1 Decorate 22 SpecId 3 Decorate 33(Z) SpecId 0 This seems inaccurate, since the spec constant `X` that is dependent on the two others did not get a name in the SPIR-V decorations. This behavior may potentially negatively affect shader introspection capabilities. This change alters the behavior to always add a name, which results in the code above producing the following decorations: Decorate 21(Y) SpecId 1 Decorate 22(X) SpecId 3 Decorate 33(Z) SpecId 0
2018-10-30 05:56:44 +00:00
spv::Id result;
Built-in values declared as specialization constant Support declaring built-in values as spec constants. Refine the code in createSpvConstant().
2016-04-04 03:55:17 +00:00
if (auto* sub_tree = sn->getConstSubtree()) {
Fix spec const construtor for matrix and vector Fix issue: #237 1. The code generated for matrix constructor should 1) build column vectors first, 2) build matrix with the vectors. 2. When there is only one scalar type constituent in vector's constructor, we should populate the constituent to fill all the slots in the vector. As for matrix, the single constituent should be populated to the diagonal positions (top-left to bottom-right diagonal). remove createSpvConstantFromConstSubTree()
2016-04-14 20:40:20 +00:00
// Traverse the constant constructor sub tree like generating normal run-time instructions.
// During the AST traversal, if the node is marked as 'specConstant', SpecConstantOpModeGuard
// will set the builder into spec constant op instruction generating mode.
sub_tree->traverse(this);
Add names for composite constants in SPIR-V Consider the following code: layout(constant_id=0) const int Y = 1; layout(constant_id=1) const int Z = 2; layout(constant_id=3) const int X = Y + Z; Previously, it would produce SPIR-V decorations like this: Decorate 21(Y) SpecId 1 Decorate 22 SpecId 3 Decorate 33(Z) SpecId 0 This seems inaccurate, since the spec constant `X` that is dependent on the two others did not get a name in the SPIR-V decorations. This behavior may potentially negatively affect shader introspection capabilities. This change alters the behavior to always add a name, which results in the code above producing the following decorations: Decorate 21(Y) SpecId 1 Decorate 22(X) SpecId 3 Decorate 33(Z) SpecId 0
2018-10-30 05:56:44 +00:00
result = accessChainLoad(sub_tree->getType());
} else if (auto* const_union_array = &sn->getConstArray()) {
Built-in values declared as specialization constant Support declaring built-in values as spec constants. Refine the code in createSpvConstant().
2016-04-04 03:55:17 +00:00
int nextConst = 0;
Add names for composite constants in SPIR-V Consider the following code: layout(constant_id=0) const int Y = 1; layout(constant_id=1) const int Z = 2; layout(constant_id=3) const int X = Y + Z; Previously, it would produce SPIR-V decorations like this: Decorate 21(Y) SpecId 1 Decorate 22 SpecId 3 Decorate 33(Z) SpecId 0 This seems inaccurate, since the spec constant `X` that is dependent on the two others did not get a name in the SPIR-V decorations. This behavior may potentially negatively affect shader introspection capabilities. This change alters the behavior to always add a name, which results in the code above producing the following decorations: Decorate 21(Y) SpecId 1 Decorate 22(X) SpecId 3 Decorate 33(Z) SpecId 0
2018-10-30 05:56:44 +00:00
result = createSpvConstantFromConstUnionArray(sn->getType(), *const_union_array, nextConst, true);
Handle potentially uninitialized result
2018-11-12 18:56:52 +00:00
} else {
logger->missingFunctionality("Invalid initializer for spec onstant.");
return spv::NoResult;
Built-in values declared as specialization constant Support declaring built-in values as spec constants. Refine the code in createSpvConstant().
2016-04-04 03:55:17 +00:00
}
Add names for composite constants in SPIR-V Consider the following code: layout(constant_id=0) const int Y = 1; layout(constant_id=1) const int Z = 2; layout(constant_id=3) const int X = Y + Z; Previously, it would produce SPIR-V decorations like this: Decorate 21(Y) SpecId 1 Decorate 22 SpecId 3 Decorate 33(Z) SpecId 0 This seems inaccurate, since the spec constant `X` that is dependent on the two others did not get a name in the SPIR-V decorations. This behavior may potentially negatively affect shader introspection capabilities. This change alters the behavior to always add a name, which results in the code above producing the following decorations: Decorate 21(Y) SpecId 1 Decorate 22(X) SpecId 3 Decorate 33(Z) SpecId 0
2018-10-30 05:56:44 +00:00
builder.addName(result, sn->getName().c_str());
return result;
Built-in values declared as specialization constant Support declaring built-in values as spec constants. Refine the code in createSpvConstant().
2016-04-04 03:55:17 +00:00
}
// Neither a front-end constant node, nor a specialization constant node with constant union array or
// constant sub tree as initializer.
Create a new logger class for all messages w.r.t. SPIR-V build.
2016-05-04 19:55:59 +00:00
logger->missingFunctionality("Neither a front-end constant nor a spec constant.");
Built-in values declared as specialization constant Support declaring built-in values as spec constants. Refine the code in createSpvConstant().
2016-04-04 03:55:17 +00:00
return spv::NoResult;
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
// Use 'consts' as the flattened glslang source of scalar constants to recursively
// build the aggregate SPIR-V constant.
//
// If there are not enough elements present in 'consts', 0 will be substituted;
// an empty 'consts' can be used to create a fully zeroed SPIR-V constant.
//
Support specialization composite constants Fix issue #163, support creation and reference of composite type specialization constants. e.g.: ``` layout(constant_id = 200) const float myfloat = 1.25; layout(constant_id = 201) const int myint = 14; struct structtype { float f; int i; }; const structtype outer_struct_var = {myfloat, myint}; void main(){} ``` generated code (use glslangValidator): ``` // Module Version 10000 // Generated by (magic number): 80001 // Id's are bound by 12 Capability Shader 1: ExtInstImport "GLSL.std.450" MemoryModel Logical GLSL450 EntryPoint Vertex 4 "main" Source GLSL 450 Name 4 "main" Name 10 "structtype" MemberName 10(structtype) 0 "f" MemberName 10(structtype) 1 "i" Decorate 7 SpecId 200 Decorate 9 SpecId 201 2: TypeVoid 3: TypeFunction 2 6: TypeFloat 32 7: 6(float) SpecConstant 1067450368 8: TypeInt 32 1 9: 8(int) SpecConstant 14 10(structtype): TypeStruct 6(float) 8(int) 11:10(structtype) SpecConstantComposite 7 9 4(main): 2 Function None 3 5: Label Return FunctionEnd ``` Rname two function names to match their functionalities. 1) Rename `GlslangToSpvTraverser::createSpvSpecConstant()` to `createSpvConstant()`; 2) Rename `GlslangToSpvTraverser::createSpvConstant()` to `createSpvConstantFromConstUnionArray()` Add function `GlslangToSpvTraverser::createSpvConstantFromSubTree()` to handle constant creation from sub trees (e.g.: specialization constants). Related PR: #208
2016-03-21 13:51:37 +00:00
spv::Id TGlslangToSpvTraverser::createSpvConstantFromConstUnionArray(const glslang::TType& glslangType, const glslang::TConstUnionArray& consts, int& nextConst, bool specConstant)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
// vector of constants for SPIR-V
std::vector<spv::Id> spvConsts;
// Type is used for struct and array constants
spv::Id typeId = convertGlslangToSpvType(glslangType);
if (glslangType.isArray()) {
Array of array: Implement the core functionality: types, constructors, operations. There will be subsequent commits to refine semantics, esp. version-specific semantics, as well as I/O functionality and restrictions. Note: I'm getting white-space differences in the preprocessor test results, which I'm not checking in. I think they need to be tagged as binary or something.
2015-08-10 23:08:55 +00:00
glslang::TType elementType(glslangType, 0);
for (int i = 0; i < glslangType.getOuterArraySize(); ++i)
Support specialization composite constants Fix issue #163, support creation and reference of composite type specialization constants. e.g.: ``` layout(constant_id = 200) const float myfloat = 1.25; layout(constant_id = 201) const int myint = 14; struct structtype { float f; int i; }; const structtype outer_struct_var = {myfloat, myint}; void main(){} ``` generated code (use glslangValidator): ``` // Module Version 10000 // Generated by (magic number): 80001 // Id's are bound by 12 Capability Shader 1: ExtInstImport "GLSL.std.450" MemoryModel Logical GLSL450 EntryPoint Vertex 4 "main" Source GLSL 450 Name 4 "main" Name 10 "structtype" MemberName 10(structtype) 0 "f" MemberName 10(structtype) 1 "i" Decorate 7 SpecId 200 Decorate 9 SpecId 201 2: TypeVoid 3: TypeFunction 2 6: TypeFloat 32 7: 6(float) SpecConstant 1067450368 8: TypeInt 32 1 9: 8(int) SpecConstant 14 10(structtype): TypeStruct 6(float) 8(int) 11:10(structtype) SpecConstantComposite 7 9 4(main): 2 Function None 3 5: Label Return FunctionEnd ``` Rname two function names to match their functionalities. 1) Rename `GlslangToSpvTraverser::createSpvSpecConstant()` to `createSpvConstant()`; 2) Rename `GlslangToSpvTraverser::createSpvConstant()` to `createSpvConstantFromConstUnionArray()` Add function `GlslangToSpvTraverser::createSpvConstantFromSubTree()` to handle constant creation from sub trees (e.g.: specialization constants). Related PR: #208
2016-03-21 13:51:37 +00:00
spvConsts.push_back(createSpvConstantFromConstUnionArray(elementType, consts, nextConst, false));
Final round for line endings.
2015-06-26 22:58:36 +00:00
} else if (glslangType.isMatrix()) {
Array of array: Implement the core functionality: types, constructors, operations. There will be subsequent commits to refine semantics, esp. version-specific semantics, as well as I/O functionality and restrictions. Note: I'm getting white-space differences in the preprocessor test results, which I'm not checking in. I think they need to be tagged as binary or something.
2015-08-10 23:08:55 +00:00
glslang::TType vectorType(glslangType, 0);
Final round for line endings.
2015-06-26 22:58:36 +00:00
for (int col = 0; col < glslangType.getMatrixCols(); ++col)
Support specialization composite constants Fix issue #163, support creation and reference of composite type specialization constants. e.g.: ``` layout(constant_id = 200) const float myfloat = 1.25; layout(constant_id = 201) const int myint = 14; struct structtype { float f; int i; }; const structtype outer_struct_var = {myfloat, myint}; void main(){} ``` generated code (use glslangValidator): ``` // Module Version 10000 // Generated by (magic number): 80001 // Id's are bound by 12 Capability Shader 1: ExtInstImport "GLSL.std.450" MemoryModel Logical GLSL450 EntryPoint Vertex 4 "main" Source GLSL 450 Name 4 "main" Name 10 "structtype" MemberName 10(structtype) 0 "f" MemberName 10(structtype) 1 "i" Decorate 7 SpecId 200 Decorate 9 SpecId 201 2: TypeVoid 3: TypeFunction 2 6: TypeFloat 32 7: 6(float) SpecConstant 1067450368 8: TypeInt 32 1 9: 8(int) SpecConstant 14 10(structtype): TypeStruct 6(float) 8(int) 11:10(structtype) SpecConstantComposite 7 9 4(main): 2 Function None 3 5: Label Return FunctionEnd ``` Rname two function names to match their functionalities. 1) Rename `GlslangToSpvTraverser::createSpvSpecConstant()` to `createSpvConstant()`; 2) Rename `GlslangToSpvTraverser::createSpvConstant()` to `createSpvConstantFromConstUnionArray()` Add function `GlslangToSpvTraverser::createSpvConstantFromSubTree()` to handle constant creation from sub trees (e.g.: specialization constants). Related PR: #208
2016-03-21 13:51:37 +00:00
spvConsts.push_back(createSpvConstantFromConstUnionArray(vectorType, consts, nextConst, false));
Implement GL_NV_cooperative_matrix
2019-02-19 19:10:32 +00:00
} else if (glslangType.isCoopMat()) {
glslang::TType componentType(glslangType.getBasicType());
spvConsts.push_back(createSpvConstantFromConstUnionArray(componentType, consts, nextConst, false));
GL_EXT_buffer_reference
2019-01-06 23:58:04 +00:00
} else if (glslangType.isStruct()) {
Final round for line endings.
2015-06-26 22:58:36 +00:00
glslang::TVector<glslang::TTypeLoc>::const_iterator iter;
for (iter = glslangType.getStruct()->begin(); iter != glslangType.getStruct()->end(); ++iter)
Support specialization composite constants Fix issue #163, support creation and reference of composite type specialization constants. e.g.: ``` layout(constant_id = 200) const float myfloat = 1.25; layout(constant_id = 201) const int myint = 14; struct structtype { float f; int i; }; const structtype outer_struct_var = {myfloat, myint}; void main(){} ``` generated code (use glslangValidator): ``` // Module Version 10000 // Generated by (magic number): 80001 // Id's are bound by 12 Capability Shader 1: ExtInstImport "GLSL.std.450" MemoryModel Logical GLSL450 EntryPoint Vertex 4 "main" Source GLSL 450 Name 4 "main" Name 10 "structtype" MemberName 10(structtype) 0 "f" MemberName 10(structtype) 1 "i" Decorate 7 SpecId 200 Decorate 9 SpecId 201 2: TypeVoid 3: TypeFunction 2 6: TypeFloat 32 7: 6(float) SpecConstant 1067450368 8: TypeInt 32 1 9: 8(int) SpecConstant 14 10(structtype): TypeStruct 6(float) 8(int) 11:10(structtype) SpecConstantComposite 7 9 4(main): 2 Function None 3 5: Label Return FunctionEnd ``` Rname two function names to match their functionalities. 1) Rename `GlslangToSpvTraverser::createSpvSpecConstant()` to `createSpvConstant()`; 2) Rename `GlslangToSpvTraverser::createSpvConstant()` to `createSpvConstantFromConstUnionArray()` Add function `GlslangToSpvTraverser::createSpvConstantFromSubTree()` to handle constant creation from sub trees (e.g.: specialization constants). Related PR: #208
2016-03-21 13:51:37 +00:00
spvConsts.push_back(createSpvConstantFromConstUnionArray(*iter->type, consts, nextConst, false));
Full stack: distinguish between a scalar and a vector of size 1. There have been GLSL extensions considering this, and HLSL does it. This is a fully backward compatible change that allows this distinction.
2016-05-20 18:06:03 +00:00
} else if (glslangType.getVectorSize() > 1) {
Final round for line endings.
2015-06-26 22:58:36 +00:00
for (unsigned int i = 0; i < (unsigned int)glslangType.getVectorSize(); ++i) {
bool zero = nextConst >= consts.size();
switch (glslangType.getBasicType()) {
Web: Remove unnecessary GLSL numeric types, and some collateral. This saves another 40K of x86 binary, which is about 13% of the target size.
2019-08-08 16:35:51 +00:00
case glslang::EbtInt:
spvConsts.push_back(builder.makeIntConstant(zero ? 0 : consts[nextConst].getIConst()));
break;
case glslang::EbtUint:
spvConsts.push_back(builder.makeUintConstant(zero ? 0 : consts[nextConst].getUConst()));
break;
case glslang::EbtFloat:
spvConsts.push_back(builder.makeFloatConstant(zero ? 0.0F : (float)consts[nextConst].getDConst()));
break;
case glslang::EbtBool:
spvConsts.push_back(builder.makeBoolConstant(zero ? false : consts[nextConst].getBConst()));
break;
#ifndef GLSLANG_WEB
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EbtInt8:
spvConsts.push_back(builder.makeInt8Constant(zero ? 0 : consts[nextConst].getI8Const()));
break;
case glslang::EbtUint8:
spvConsts.push_back(builder.makeUint8Constant(zero ? 0 : consts[nextConst].getU8Const()));
break;
case glslang::EbtInt16:
spvConsts.push_back(builder.makeInt16Constant(zero ? 0 : consts[nextConst].getI16Const()));
break;
case glslang::EbtUint16:
spvConsts.push_back(builder.makeUint16Constant(zero ? 0 : consts[nextConst].getU16Const()));
break;
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
case glslang::EbtInt64:
spvConsts.push_back(builder.makeInt64Constant(zero ? 0 : consts[nextConst].getI64Const()));
break;
case glslang::EbtUint64:
spvConsts.push_back(builder.makeUint64Constant(zero ? 0 : consts[nextConst].getU64Const()));
break;
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EbtDouble:
spvConsts.push_back(builder.makeDoubleConstant(zero ? 0.0 : consts[nextConst].getDConst()));
break;
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
case glslang::EbtFloat16:
spvConsts.push_back(builder.makeFloat16Constant(zero ? 0.0F : (float)consts[nextConst].getDConst()));
break;
Web: Remove unnecessary GLSL numeric types, and some collateral. This saves another 40K of x86 binary, which is about 13% of the target size.
2019-08-08 16:35:51 +00:00
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(0);
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
}
++nextConst;
}
} else {
// we have a non-aggregate (scalar) constant
bool zero = nextConst >= consts.size();
spv::Id scalar = 0;
switch (glslangType.getBasicType()) {
Web: Remove unnecessary GLSL numeric types, and some collateral. This saves another 40K of x86 binary, which is about 13% of the target size.
2019-08-08 16:35:51 +00:00
case glslang::EbtInt:
scalar = builder.makeIntConstant(zero ? 0 : consts[nextConst].getIConst(), specConstant);
break;
case glslang::EbtUint:
scalar = builder.makeUintConstant(zero ? 0 : consts[nextConst].getUConst(), specConstant);
break;
case glslang::EbtFloat:
scalar = builder.makeFloatConstant(zero ? 0.0F : (float)consts[nextConst].getDConst(), specConstant);
break;
case glslang::EbtBool:
scalar = builder.makeBoolConstant(zero ? false : consts[nextConst].getBConst(), specConstant);
break;
#ifndef GLSLANG_WEB
SPV: Implement Vulkan 1.1 features and extensions.
2018-03-06 23:12:04 +00:00
case glslang::EbtInt8:
scalar = builder.makeInt8Constant(zero ? 0 : consts[nextConst].getI8Const(), specConstant);
break;
case glslang::EbtUint8:
scalar = builder.makeUint8Constant(zero ? 0 : consts[nextConst].getU8Const(), specConstant);
break;
case glslang::EbtInt16:
scalar = builder.makeInt16Constant(zero ? 0 : consts[nextConst].getI16Const(), specConstant);
break;
case glslang::EbtUint16:
scalar = builder.makeUint16Constant(zero ? 0 : consts[nextConst].getU16Const(), specConstant);
break;
Implement the extension GL_ARB_gpu_shader_int64 - Add new keyword int64_t/uint64_t/i64vec/u64vec. - Support 64-bit integer literals (dec/hex/oct). - Support built-in operators for 64-bit integer type. - Add implicit and explicit type conversion for 64-bit integer type. - Add new built-in functions defined in this extension.
2016-04-22 08:51:45 +00:00
case glslang::EbtInt64:
scalar = builder.makeInt64Constant(zero ? 0 : consts[nextConst].getI64Const(), specConstant);
break;
case glslang::EbtUint64:
scalar = builder.makeUint64Constant(zero ? 0 : consts[nextConst].getU64Const(), specConstant);
break;
Final round for line endings.
2015-06-26 22:58:36 +00:00
case glslang::EbtDouble:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
scalar = builder.makeDoubleConstant(zero ? 0.0 : consts[nextConst].getDConst(), specConstant);
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
Parser: Implement extension GL_AMD_gpu_shader_half_float. - Add built-in types: float16_t, f16vec, f16mat. - Add support of half float constant: hf, HF. - Extend built-in floating-point operators: +, -, *, /, ++, --, +=, -=, *=, /=, ==, !=, >=, <=, >, <. - Add support of type conversions: float16_t -> XXX, XXX -> float16_t. - Add new built-in functions.
2016-07-29 08:00:05 +00:00
case glslang::EbtFloat16:
scalar = builder.makeFloat16Constant(zero ? 0.0F : (float)consts[nextConst].getDConst(), specConstant);
break;
Improved fix for buffer reference constants This is an alternate fix for the issue described in commit be63facd, whose solution didn't work if there were non-trivial operations involved in computing a constant initializer which caused the 'constant unfolding' code to kick in (addConstantReferenceConversion). Instead, this change does the 'unfolding' later in createSpvConstantFromConstUnionArray. If a reference-type constant has survived that long, then folding is already done, this must be a 'real' (inside a function) use of the constant, and it should be safe to unfold and apply the bitcast.
2019-03-06 05:27:09 +00:00
case glslang::EbtReference:
scalar = builder.makeUint64Constant(zero ? 0 : consts[nextConst].getU64Const(), specConstant);
scalar = builder.createUnaryOp(spv::OpBitcast, typeId, scalar);
break;
Web: Remove unnecessary GLSL numeric types, and some collateral. This saves another 40K of x86 binary, which is about 13% of the target size.
2019-08-08 16:35:51 +00:00
#endif
Final round for line endings.
2015-06-26 22:58:36 +00:00
default:
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
assert(0);
Final round for line endings.
2015-06-26 22:58:36 +00:00
break;
}
++nextConst;
return scalar;
}
return builder.makeCompositeConstant(typeId, spvConsts);
}
SPV: Implement short circuiting of && and || when emitting SPIR-V.
2015-10-15 19:29:11 +00:00
// Return true if the node is a constant or symbol whose reading has no
// non-trivial observable cost or effect.
bool TGlslangToSpvTraverser::isTrivialLeaf(const glslang::TIntermTyped* node)
{
// don't know what this is
if (node == nullptr)
return false;
// a constant is safe
if (node->getAsConstantUnion() != nullptr)
return true;
// not a symbol means non-trivial
if (node->getAsSymbolNode() == nullptr)
return false;
// a symbol, depends on what's being read
switch (node->getType().getQualifier().storage) {
case glslang::EvqTemporary:
case glslang::EvqGlobal:
case glslang::EvqIn:
case glslang::EvqInOut:
case glslang::EvqConst:
case glslang::EvqConstReadOnly:
case glslang::EvqUniform:
return true;
default:
return false;
}
Resolve comments 1. Sink adding noContraction decoration to createBinaryOperation() and createUnaryOperation(). 2. Fix comments. 3. Remove the #define of my delimiter, use global constant char.
2016-05-06 21:25:16 +00:00
}
SPV: Implement short circuiting of && and || when emitting SPIR-V.
2015-10-15 19:29:11 +00:00
// A node is trivial if it is a single operation with no side effects.
HLSL: Fix #903: Don't short-circuit && or ||.
2017-05-24 22:44:47 +00:00
// HLSL (and/or vectors) are always trivial, as it does not short circuit.
HLSL: Don't do logical short-circuits when the operands are bool-vectors. This seems a bit ill-defined, and was generating code that made OpPhi of two operands that were Boolean vectors result in a scalar bool.
2017-05-20 02:19:00 +00:00
// Otherwise, error on the side of saying non-trivial.
SPV: Implement short circuiting of && and || when emitting SPIR-V.
2015-10-15 19:29:11 +00:00
// Return true if trivial.
bool TGlslangToSpvTraverser::isTrivial(const glslang::TIntermTyped* node)
{
if (node == nullptr)
return false;
HLSL: Fix #903: Don't short-circuit && or ||.
2017-05-24 22:44:47 +00:00
// count non scalars as trivial, as well as anything coming from HLSL
if (! node->getType().isScalarOrVec1() || glslangIntermediate->getSource() == glslang::EShSourceHlsl)
HLSL: Don't do logical short-circuits when the operands are bool-vectors. This seems a bit ill-defined, and was generating code that made OpPhi of two operands that were Boolean vectors result in a scalar bool.
2017-05-20 02:19:00 +00:00
return true;
SPV: Implement short circuiting of && and || when emitting SPIR-V.
2015-10-15 19:29:11 +00:00
// symbols and constants are trivial
if (isTrivialLeaf(node))
return true;
// otherwise, it needs to be a simple operation or one or two leaf nodes
// not a simple operation
const glslang::TIntermBinary* binaryNode = node->getAsBinaryNode();
const glslang::TIntermUnary* unaryNode = node->getAsUnaryNode();
if (binaryNode == nullptr && unaryNode == nullptr)
return false;
// not on leaf nodes
if (binaryNode && (! isTrivialLeaf(binaryNode->getLeft()) || ! isTrivialLeaf(binaryNode->getRight())))
return false;
if (unaryNode && ! isTrivialLeaf(unaryNode->getOperand())) {
return false;
}
switch (node->getAsOperator()->getOp()) {
case glslang::EOpLogicalNot:
case glslang::EOpConvIntToBool:
case glslang::EOpConvUintToBool:
case glslang::EOpConvFloatToBool:
case glslang::EOpConvDoubleToBool:
case glslang::EOpEqual:
case glslang::EOpNotEqual:
case glslang::EOpLessThan:
case glslang::EOpGreaterThan:
case glslang::EOpLessThanEqual:
case glslang::EOpGreaterThanEqual:
case glslang::EOpIndexDirect:
case glslang::EOpIndexDirectStruct:
case glslang::EOpLogicalXor:
case glslang::EOpAny:
case glslang::EOpAll:
return true;
default:
return false;
}
}
// Emit short-circuiting code, where 'right' is never evaluated unless
// the left side is true (for &&) or false (for ||).
spv::Id TGlslangToSpvTraverser::createShortCircuit(glslang::TOperator op, glslang::TIntermTyped& left, glslang::TIntermTyped& right)
{
spv::Id boolTypeId = builder.makeBoolType();
// emit left operand
builder.clearAccessChain();
left.traverse(this);
SPV: Continue to fix the issue of bool -> uint32 For short-circuit operator (&& and ||), the conversion is missing.
2016-03-03 06:38:51 +00:00
spv::Id leftId = accessChainLoad(left.getType());
SPV: Implement short circuiting of && and || when emitting SPIR-V.
2015-10-15 19:29:11 +00:00
// Operands to accumulate OpPhi operands
std::vector<spv::Id> phiOperands;
// accumulate left operand's phi information
phiOperands.push_back(leftId);
phiOperands.push_back(builder.getBuildPoint()->getId());
// Make the two kinds of operation symmetric with a "!"
// || => emit "if (! left) result = right"
// && => emit "if ( left) result = right"
//
// TODO: this runtime "not" for || could be avoided by adding functionality
// to 'builder' to have an "else" without an "then"
if (op == glslang::EOpLogicalOr)
leftId = builder.createUnaryOp(spv::OpLogicalNot, boolTypeId, leftId);
// make an "if" based on the left value
HLSL: Translate directive [flatten] and [branch] to SPV control mask.
2017-07-04 15:23:40 +00:00
spv::Builder::If ifBuilder(leftId, spv::SelectionControlMaskNone, builder);
SPV: Implement short circuiting of && and || when emitting SPIR-V.
2015-10-15 19:29:11 +00:00
// emit right operand as the "then" part of the "if"
builder.clearAccessChain();
right.traverse(this);
SPV: Continue to fix the issue of bool -> uint32 For short-circuit operator (&& and ||), the conversion is missing.
2016-03-03 06:38:51 +00:00
spv::Id rightId = accessChainLoad(right.getType());
SPV: Implement short circuiting of && and || when emitting SPIR-V.
2015-10-15 19:29:11 +00:00
// accumulate left operand's phi information
phiOperands.push_back(rightId);
phiOperands.push_back(builder.getBuildPoint()->getId());
// finish the "if"
ifBuilder.makeEndIf();
// phi together the two results
return builder.createOp(spv::OpPhi, boolTypeId, phiOperands);
}
Web: Generalize _EXTENSIONS* in SPIR-V back-end. About 50 fewer #ifdefs. About 14K smaller. Note, the base size is ill-defined due to optimizer settings (size vs. performance), compression, and target architecture. Some recent %'s are accidentally reported as 3X the real savings. Early %'s were accurate. What matters though is that each step got worthwhile gains, and what the final size ends up being.
2019-08-06 13:00:58 +00:00
#ifndef GLSLANG_WEB
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
// Return type Id of the imported set of extended instructions corresponds to the name.
// Import this set if it has not been imported yet.
spv::Id TGlslangToSpvTraverser::getExtBuiltins(const char* name)
{
if (extBuiltinMap.find(name) != extBuiltinMap.end())
return extBuiltinMap[name];
else {
SPV: Implement the extension SPV_KHR_shader_ballot
2016-09-21 10:56:12 +00:00
builder.addExtension(name);
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
spv::Id extBuiltins = builder.import(name);
extBuiltinMap[name] = extBuiltins;
return extBuiltins;
}
}
Fix unused function warning. TGlslangToSpvTraverser::getExtBuiltins is only used when AMD_EXTENSIONS is defined, so only define it in that case to avoid an unused function warning.
2018-01-16 05:18:26 +00:00
#endif
Implement 4 AMD-specific extensions. - Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot). - Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax). - Support GL_AMD_shader_explicit_vertex_parameter (SPV_AMD_shader_explicit_vertex_parameter). - Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
2016-05-05 04:30:44 +00:00
Final round for line endings.
2015-06-26 22:58:36 +00:00
}; // end anonymous namespace
namespace glslang {
Updated command-line options, adding -o for saving binaries, -G for OpenGL SPIR-V validation, -v etc. Old uses should still work as they did before. Also encapsulated use of these flags during parsing, for the parse context. Added SPIR-V version to -v.
2015-07-13 01:28:10 +00:00
void GetSpirvVersion(std::string& version)
{
Include specification revision as part of the version string.
2015-07-15 16:03:39 +00:00
const int bufSize = 100;
Fix incorrect used of snprintf()
2015-07-13 01:39:51 +00:00
char buf[bufSize];
SPIR-V: Move from Version .99 Rev 31 to Version 1.0, Rev 2.
2015-11-16 04:33:39 +00:00
snprintf(buf, bufSize, "0x%08x, Revision %d", spv::Version, spv::Revision);
Updated command-line options, adding -o for saving binaries, -G for OpenGL SPIR-V validation, -v etc. Old uses should still work as they did before. Also encapsulated use of these flags during parsing, for the parse context. Added SPIR-V version to -v.
2015-07-13 01:28:10 +00:00
version = buf;
}
Versioning: Update some version strings. Still need better overall version.
2017-11-03 04:32:14 +00:00
// For low-order part of the generator's magic number. Bump up
// when there is a change in the style (e.g., if SSA form changes,
// or a different instruction sequence to do something gets used).
int GetSpirvGeneratorVersion()
{
SPV: Document history of the generator version number.
2017-12-17 06:46:37 +00:00
// return 1; // start
// return 2; // EOpAtomicCounterDecrement gets a post decrement, to map between GLSL -> SPIR-V
SPV: Bump up generator number, because previous commit changes code gen slightly.
2018-02-06 15:06:36 +00:00
// return 3; // change/correct barrier-instruction operands, to match memory model group decisions
SPV: Bump up the generator number.
2018-03-03 18:47:07 +00:00
// return 4; // some deeper access chains: for dynamic vector component, and local Boolean component
Revert the commits that change OpArrayLength type and bumped the version number. Now, version 5.* is all connected to making the uint type, which doesn't quite work. Generator versions 4 and 6 do not do this.
2018-03-07 18:24:50 +00:00
// return 5; // make OpArrayLength result type be an int with signedness of 0
Change the major revision number for next commit.
2018-06-04 21:33:31 +00:00
// return 6; // revert version 5 change, which makes a different (new) kind of incorrect code,
// versions 4 and 6 each generate OpArrayLength as it has long been done
return 7; // GLSL volatile keyword maps to both SPIR-V decorations Volatile and Coherent
Versioning: Update some version strings. Still need better overall version.
2017-11-03 04:32:14 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
// Write SPIR-V out to a binary file
Added -x option to save SPIR-V as 32-bit hexadecimal numbers to a text file.
2016-05-27 17:55:53 +00:00
void OutputSpvBin(const std::vector<unsigned int>& spirv, const char* baseName)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
std::ofstream out;
Updated command-line options, adding -o for saving binaries, -G for OpenGL SPIR-V validation, -v etc. Old uses should still work as they did before. Also encapsulated use of these flags during parsing, for the parse context. Added SPIR-V version to -v.
2015-07-13 01:28:10 +00:00
out.open(baseName, std::ios::binary | std::ios::out);
Fix issue #676: emit error message on failure to open spv file.
2017-02-18 16:45:40 +00:00
if (out.fail())
printf("ERROR: Failed to open file: %s\n", baseName);
Final round for line endings.
2015-06-26 22:58:36 +00:00
for (int i = 0; i < (int)spirv.size(); ++i) {
unsigned int word = spirv[i];
out.write((const char*)&word, 4);
}
out.close();
}
Added -x option to save SPIR-V as 32-bit hexadecimal numbers to a text file.
2016-05-27 17:55:53 +00:00
// Write SPIR-V out to a text file with 32-bit hexadecimal words
Added --vn option to generate a C header file containing a variable assigned to the hex representation of the shader. This is a standard feature on Microsoft's HLSL compiler and it allows developers to include pre-compiled shaders directly into the code. This option enables "Hex output", so it is NOT required to specify -x as well. The output file name is preserved, so no ".h" extension is added. If you want the output file to have ".h" extension then you have to specify it on the output file name. The generated header file uses the "#pragma once" pragma to avoid multiple inclusions.
2017-02-06 19:46:35 +00:00
void OutputSpvHex(const std::vector<unsigned int>& spirv, const char* baseName, const char* varName)
Added -x option to save SPIR-V as 32-bit hexadecimal numbers to a text file.
2016-05-27 17:55:53 +00:00
{
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#ifndef GLSLANG_WEB
Added -x option to save SPIR-V as 32-bit hexadecimal numbers to a text file.
2016-05-27 17:55:53 +00:00
std::ofstream out;
out.open(baseName, std::ios::binary | std::ios::out);
Fix issue #676: emit error message on failure to open spv file.
2017-02-18 16:45:40 +00:00
if (out.fail())
printf("ERROR: Failed to open file: %s\n", baseName);
Versioning: Address #1255: Move to semantic versioning. Still missing is automation and final decisions on mechanisms in general.
2018-03-06 05:23:17 +00:00
out << "\t// " <<
SPV: Turn off the default validation done by the SPIRV-Tools optimizer. Also use the glslang namespace qualifier consistently and validate after legalization, not before. (But most tests don't legalize.)
2018-08-30 22:56:59 +00:00
GetSpirvGeneratorVersion() << "." << GLSLANG_MINOR_VERSION << "." << GLSLANG_PATCH_LEVEL <<
Versioning: Address #1255: Move to semantic versioning. Still missing is automation and final decisions on mechanisms in general.
2018-03-06 05:23:17 +00:00
std::endl;
Removed tabs and replaced with spaces. Changed layout for "else if"
2017-02-15 22:29:33 +00:00
if (varName != nullptr) {
out << "\t #pragma once" << std::endl;
out << "const uint32_t " << varName << "[] = {" << std::endl;
}
Added -x option to save SPIR-V as 32-bit hexadecimal numbers to a text file.
2016-05-27 17:55:53 +00:00
const int WORDS_PER_LINE = 8;
for (int i = 0; i < (int)spirv.size(); i += WORDS_PER_LINE) {
out << "\t";
for (int j = 0; j < WORDS_PER_LINE && i + j < (int)spirv.size(); ++j) {
const unsigned int word = spirv[i + j];
out << "0x" << std::hex << std::setw(8) << std::setfill('0') << word;
if (i + j + 1 < (int)spirv.size()) {
out << ",";
}
}
out << std::endl;
}
Removed tabs and replaced with spaces. Changed layout for "else if"
2017-02-15 22:29:33 +00:00
if (varName != nullptr) {
out << "};";
}
Added -x option to save SPIR-V as 32-bit hexadecimal numbers to a text file.
2016-05-27 17:55:53 +00:00
out.close();
Web: Remove unused stage functionality, SPIR-V logger, and hex_utils Saves another 20K.
2019-08-09 05:29:20 +00:00
#endif
Added -x option to save SPIR-V as 32-bit hexadecimal numbers to a text file.
2016-05-27 17:55:53 +00:00
}
Final round for line endings.
2015-06-26 22:58:36 +00:00
//
// Set up the glslang traversal
//
SPV: Turn off the default validation done by the SPIRV-Tools optimizer. Also use the glslang namespace qualifier consistently and validate after legalization, not before. (But most tests don't legalize.)
2018-08-30 22:56:59 +00:00
void GlslangToSpv(const TIntermediate& intermediate, std::vector<unsigned int>& spirv, SpvOptions* options)
Avoid printing to stdout directly in library functions. Previously GlslangToSpv() reported missing/TBD functionalities by directly writing to stdout using printf. That could cause problems to callers of GlslangToSpv(). This patch cleans up the error reporting logic in GlslangToSpv(), TGlslangToSpvTraverser, and spv::Builder a little bit to use ostringstream. Also fixed the usage of GlslangToSpv() in GTest fixtures to capture warnings/errors reported when translating AST to SPIR-V.
2016-05-02 22:11:54 +00:00
{
Create a new logger class for all messages w.r.t. SPIR-V build.
2016-05-04 19:55:59 +00:00
spv::SpvBuildLogger logger;
SPV: Add OpSource shader source code and file name.
2017-05-31 23:11:16 +00:00
GlslangToSpv(intermediate, spirv, &logger, options);
Avoid printing to stdout directly in library functions. Previously GlslangToSpv() reported missing/TBD functionalities by directly writing to stdout using printf. That could cause problems to callers of GlslangToSpv(). This patch cleans up the error reporting logic in GlslangToSpv(), TGlslangToSpvTraverser, and spv::Builder a little bit to use ostringstream. Also fixed the usage of GlslangToSpv() in GTest fixtures to capture warnings/errors reported when translating AST to SPIR-V.
2016-05-02 22:11:54 +00:00
}
SPV: Turn off the default validation done by the SPIRV-Tools optimizer. Also use the glslang namespace qualifier consistently and validate after legalization, not before. (But most tests don't legalize.)
2018-08-30 22:56:59 +00:00
void GlslangToSpv(const TIntermediate& intermediate, std::vector<unsigned int>& spirv,
SPV: Add OpSource shader source code and file name.
2017-05-31 23:11:16 +00:00
spv::SpvBuildLogger* logger, SpvOptions* options)
Final round for line endings.
2015-06-26 22:58:36 +00:00
{
TIntermNode* root = intermediate.getTreeRoot();
if (root == 0)
return;
SPV: Turn off the default validation done by the SPIRV-Tools optimizer. Also use the glslang namespace qualifier consistently and validate after legalization, not before. (But most tests don't legalize.)
2018-08-30 22:56:59 +00:00
SpvOptions defaultOptions;
SPV: Add OpSource shader source code and file name.
2017-05-31 23:11:16 +00:00
if (options == nullptr)
options = &defaultOptions;
SPV: Turn off the default validation done by the SPIRV-Tools optimizer. Also use the glslang namespace qualifier consistently and validate after legalization, not before. (But most tests don't legalize.)
2018-08-30 22:56:59 +00:00
GetThreadPoolAllocator().push();
Final round for line endings.
2015-06-26 22:58:36 +00:00
SPV Version: Emit the requested SPIR-V version, not the header version. Fixes #1236.
2018-02-01 01:35:56 +00:00
TGlslangToSpvTraverser it(intermediate.getSpv().spv, &intermediate, logger, *options);
Final round for line endings.
2015-06-26 22:58:36 +00:00
root->traverse(&it);
Always correctly terminate main. Issue #588, PR #600.
2016-11-26 20:23:20 +00:00
it.finishSpv();
Final round for line endings.
2015-06-26 22:58:36 +00:00
it.dumpSpv(spirv);
Only print legalization warning if optimizer disabled
2018-03-29 17:49:14 +00:00
#if ENABLE_OPT
Enable HLSL legalization Also added known-good mechanism to fetch latest validated spirv-tools. Also added -Od and -Os to disable optimizer and optimize for size. Fetching spirv-tools is optional for both glsl and hlsl. Legalization of hlsl is done by default if spirv-opt is present at cmake time. Optimization for glsl is currently done through the option -Os. Legalization testing is currently only done on four existing shaders. A separate baseLegalResults directory holds those results. All previous testing is done with the optimizer disabled.
2017-09-22 00:40:22 +00:00
// If from HLSL, run spirv-opt to "legalize" the SPIR-V for Vulkan
// eg. forward and remove memory writes of opaque types.
Use spvValidatorOptionsSetBeforeHlslLegalization for pre-legalized HLSL
2019-06-07 19:42:08 +00:00
bool prelegalization = intermediate.getSource() == EShSourceHlsl;
if ((intermediate.getSource() == EShSourceHlsl || options->optimizeSize) && !options->disableOptimizer) {
Non-functional: Rationalize some existing use of SPIRV-Tools.
2018-08-22 23:12:46 +00:00
SpirvToolsLegalize(intermediate, spirv, logger, options);
Use spvValidatorOptionsSetBeforeHlslLegalization for pre-legalized HLSL
2019-06-07 19:42:08 +00:00
prelegalization = false;
}
SPV: Isolate SPIRV-tools glue to its own file.
2018-08-23 21:17:10 +00:00
SPV: Turn off the default validation done by the SPIRV-Tools optimizer. Also use the glslang namespace qualifier consistently and validate after legalization, not before. (But most tests don't legalize.)
2018-08-30 22:56:59 +00:00
if (options->validate)
Use spvValidatorOptionsSetBeforeHlslLegalization for pre-legalized HLSL
2019-06-07 19:42:08 +00:00
SpirvToolsValidate(intermediate, spirv, logger, prelegalization);
SPV: Turn off the default validation done by the SPIRV-Tools optimizer. Also use the glslang namespace qualifier consistently and validate after legalization, not before. (But most tests don't legalize.)
2018-08-30 22:56:59 +00:00
SPV: Isolate SPIRV-tools glue to its own file.
2018-08-23 21:17:10 +00:00
if (options->disassemble)
SPV: Turn off the default validation done by the SPIRV-Tools optimizer. Also use the glslang namespace qualifier consistently and validate after legalization, not before. (But most tests don't legalize.)
2018-08-30 22:56:59 +00:00
SpirvToolsDisassemble(std::cout, spirv);
SPV: Isolate SPIRV-tools glue to its own file.
2018-08-23 21:17:10 +00:00
Enable HLSL legalization Also added known-good mechanism to fetch latest validated spirv-tools. Also added -Od and -Os to disable optimizer and optimize for size. Fetching spirv-tools is optional for both glsl and hlsl. Legalization of hlsl is done by default if spirv-opt is present at cmake time. Optimization for glsl is currently done through the option -Os. Legalization testing is currently only done on four existing shaders. A separate baseLegalResults directory holds those results. All previous testing is done with the optimizer disabled.
2017-09-22 00:40:22 +00:00
#endif
SPV: Turn off the default validation done by the SPIRV-Tools optimizer. Also use the glslang namespace qualifier consistently and validate after legalization, not before. (But most tests don't legalize.)
2018-08-30 22:56:59 +00:00
GetThreadPoolAllocator().pop();
Final round for line endings.
2015-06-26 22:58:36 +00:00
}
}; // end namespace glslang
Reference in New Issue Copy Permalink