blob: dba7b33000ddb7ec4afd41d68fbd5f0f0a57c563 [file] [log] [blame]
// Copyright 2017 The Dawn Authors
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// See the License for the specific language governing permissions and
// limitations under the License.
#include "dawn_native/opengl/ShaderModuleGL.h"
#include "common/Assert.h"
#include "common/Platform.h"
#include "dawn_native/BindGroupLayout.h"
#include "dawn_native/SpirvValidation.h"
#include "dawn_native/TintUtils.h"
#include "dawn_native/opengl/DeviceGL.h"
#include "dawn_native/opengl/PipelineLayoutGL.h"
#include "dawn_native/opengl/SpirvUtils.h"
#include <spirv_glsl.hpp>
// Tint include must be after spirv_glsl.hpp, because spirv-cross has its own
// version of spirv_headers. We also need to undef SPV_REVISION because SPIRV-Cross
// is at 3 while spirv-headers is at 4.
#include <tint/tint.h>
#include <spirv-tools/libspirv.hpp>
#include <sstream>
namespace dawn_native { namespace opengl {
std::string GetBindingName(BindGroupIndex group, BindingNumber bindingNumber) {
std::ostringstream o;
o << "dawn_binding_" << static_cast<uint32_t>(group) << "_"
<< static_cast<uint32_t>(bindingNumber);
return o.str();
bool operator<(const BindingLocation& a, const BindingLocation& b) {
return std::tie(, a.binding) < std::tie(, b.binding);
bool operator<(const CombinedSampler& a, const CombinedSampler& b) {
return std::tie(a.useDummySampler, a.samplerLocation, a.textureLocation) <
std::tie(b.useDummySampler, a.samplerLocation, b.textureLocation);
std::string CombinedSampler::GetName() const {
std::ostringstream o;
o << "dawn_combined";
if (useDummySampler) {
o << "_dummy_sampler";
} else {
o << "_" << static_cast<uint32_t>( << "_"
<< static_cast<uint32_t>(samplerLocation.binding);
o << "_with_" << static_cast<uint32_t>( << "_"
<< static_cast<uint32_t>(textureLocation.binding);
return o.str();
ResultOrError<std::unique_ptr<BindingInfoArray>> ExtractSpirvInfo(
const DeviceBase* device,
const spirv_cross::Compiler& compiler,
const std::string& entryPointName,
SingleShaderStage stage) {
const auto& resources = compiler.get_shader_resources();
// Fill in bindingInfo with the SPIRV bindings
auto ExtractResourcesBinding =
[](const DeviceBase* device,
const spirv_cross::SmallVector<spirv_cross::Resource>& resources,
const spirv_cross::Compiler& compiler, BindingInfoType bindingType,
BindingInfoArray* bindings, bool isStorageBuffer = false) -> MaybeError {
for (const auto& resource : resources) {
"No Binding decoration set for resource");
"No Descriptor Decoration set for resource");
BindingNumber bindingNumber(
compiler.get_decoration(, spv::DecorationBinding));
BindGroupIndex bindGroupIndex(
compiler.get_decoration(, spv::DecorationDescriptorSet));
DAWN_INVALID_IF(bindGroupIndex >= kMaxBindGroupsTyped,
"Bind group index over limits in the SPIRV");
const auto& it =
(*bindings)[bindGroupIndex].emplace(bindingNumber, ShaderBindingInfo{});
DAWN_INVALID_IF(!it.second, "Shader has duplicate bindings");
ShaderBindingInfo* info = &it.first->second;
info->id =;
info->base_type_id = resource.base_type_id;
info->bindingType = bindingType;
switch (bindingType) {
case BindingInfoType::Texture: {
spirv_cross::SPIRType::ImageType imageType =
spirv_cross::SPIRType::BaseType textureComponentType =
info->texture.viewDimension =
SpirvDimToTextureViewDimension(imageType.dim, imageType.arrayed);
info->texture.multisampled =;
info->texture.compatibleSampleTypes =
if (imageType.depth) {
(info->texture.compatibleSampleTypes & SampleTypeBit::Float) == 0,
"Depth textures must have a float type");
info->texture.compatibleSampleTypes = SampleTypeBit::Depth;
DAWN_INVALID_IF( && imageType.arrayed,
"Multisampled array textures aren't supported");
case BindingInfoType::Buffer: {
// Determine buffer size, with a minimum of 1 element in the runtime
// array
spirv_cross::SPIRType type = compiler.get_type(info->base_type_id);
info->buffer.minBindingSize =
compiler.get_declared_struct_size_runtime_array(type, 1);
// Differentiate between readonly storage bindings and writable ones
// based on the NonWritable decoration.
// TODO(dawn:527): Could isStorageBuffer be determined by calling
// compiler.get_storage_class(
if (isStorageBuffer) {
spirv_cross::Bitset flags =
if (flags.get(spv::DecorationNonWritable)) {
info->buffer.type = wgpu::BufferBindingType::ReadOnlyStorage;
} else {
info->buffer.type = wgpu::BufferBindingType::Storage;
} else {
info->buffer.type = wgpu::BufferBindingType::Uniform;
case BindingInfoType::StorageTexture: {
spirv_cross::Bitset flags = compiler.get_decoration_bitset(;
"Read-write storage textures are not supported.");
info->storageTexture.access = wgpu::StorageTextureAccess::WriteOnly;
spirv_cross::SPIRType::ImageType imageType =
wgpu::TextureFormat storageTextureFormat =
DAWN_INVALID_IF(storageTextureFormat == wgpu::TextureFormat::Undefined,
"Invalid image format declaration on storage image.");
const Format& format = device->GetValidInternalFormat(storageTextureFormat);
"The storage texture format (%s) is not supported.",
"Multisampled storage textures aren't supported.");
"Depth storage textures aren't supported.");
info->storageTexture.format = storageTextureFormat;
info->storageTexture.viewDimension =
SpirvDimToTextureViewDimension(imageType.dim, imageType.arrayed);
case BindingInfoType::Sampler: {
info->sampler.isComparison = false;
case BindingInfoType::ExternalTexture: {
return DAWN_FORMAT_VALIDATION_ERROR("External textures are not supported.");
return {};
std::unique_ptr<BindingInfoArray> resultBindings = std::make_unique<BindingInfoArray>();
BindingInfoArray* bindings = resultBindings.get();
DAWN_TRY(ExtractResourcesBinding(device, resources.uniform_buffers, compiler,
BindingInfoType::Buffer, bindings));
DAWN_TRY(ExtractResourcesBinding(device, resources.separate_images, compiler,
BindingInfoType::Texture, bindings));
DAWN_TRY(ExtractResourcesBinding(device, resources.separate_samplers, compiler,
BindingInfoType::Sampler, bindings));
DAWN_TRY(ExtractResourcesBinding(device, resources.storage_buffers, compiler,
BindingInfoType::Buffer, bindings, true));
// ReadonlyStorageTexture is used as a tag to do general storage texture handling.
DAWN_TRY(ExtractResourcesBinding(device, resources.storage_images, compiler,
BindingInfoType::StorageTexture, resultBindings.get()));
return {std::move(resultBindings)};
// static
ResultOrError<Ref<ShaderModule>> ShaderModule::Create(Device* device,
const ShaderModuleDescriptor* descriptor,
ShaderModuleParseResult* parseResult) {
Ref<ShaderModule> module = AcquireRef(new ShaderModule(device, descriptor));
return module;
ShaderModule::ShaderModule(Device* device, const ShaderModuleDescriptor* descriptor)
: ShaderModuleBase(device, descriptor) {
// static
ResultOrError<BindingInfoArrayTable> ShaderModule::ReflectShaderUsingSPIRVCross(
DeviceBase* device,
const std::vector<uint32_t>& spirv) {
BindingInfoArrayTable result;
spirv_cross::Compiler compiler(spirv);
for (const spirv_cross::EntryPoint& entryPoint : compiler.get_entry_points_and_stages()) {
ASSERT(result.count( == 0);
SingleShaderStage stage = ExecutionModelToShaderStage(entryPoint.execution_model);
compiler.set_entry_point(, entryPoint.execution_model);
std::unique_ptr<BindingInfoArray> bindings;
DAWN_TRY_ASSIGN(bindings, ExtractSpirvInfo(device, compiler,, stage));
result[] = std::move(bindings);
return std::move(result);
MaybeError ShaderModule::Initialize(ShaderModuleParseResult* parseResult) {
ScopedTintICEHandler scopedICEHandler(GetDevice());
// Tint currently does not support emitting GLSL, so when provided a Tint program need to
// generate SPIRV and SPIRV-Cross reflection data to be used in this backend.
tint::writer::spirv::Options options;
options.disable_workgroup_init = GetDevice()->IsToggleEnabled(Toggle::DisableWorkgroupInit);
auto result = tint::writer::spirv::Generate(GetTintProgram(), options);
DAWN_INVALID_IF(!result.success, "An error occured while generating SPIR-V: %s.",
DAWN_TRY_ASSIGN(mGLBindings, ReflectShaderUsingSPIRVCross(GetDevice(), result.spirv));
return {};
ResultOrError<std::string> ShaderModule::TranslateToGLSL(const char* entryPointName,
SingleShaderStage stage,
CombinedSamplerInfo* combinedSamplers,
const PipelineLayout* layout,
bool* needsDummySampler) const {
tint::transform::SingleEntryPoint singleEntryPointTransform;
tint::transform::DataMap transformInputs;
tint::Program program;
DAWN_TRY_ASSIGN(program, RunTransforms(&singleEntryPointTransform, GetTintProgram(),
transformInputs, nullptr, nullptr));
tint::writer::spirv::Options tintOptions;
tintOptions.disable_workgroup_init =
auto result = tint::writer::spirv::Generate(&program, tintOptions);
DAWN_INVALID_IF(!result.success, "An error occured while generating SPIR-V: %s.",
std::vector<uint32_t> spirv = std::move(result.spirv);
ValidateSpirv(GetDevice(), spirv, GetDevice()->IsToggleEnabled(Toggle::DumpShaders)));
// If these options are changed, the values in DawnSPIRVCrossGLSLFastFuzzer.cpp need to
// be updated.
spirv_cross::CompilerGLSL::Options options;
// The range of Z-coordinate in the clipping volume of OpenGL is [-w, w], while it is
// [0, w] in D3D12, Metal and Vulkan, so we should normalize it in shaders in all
// backends. See the documentation of
// spirv_cross::CompilerGLSL::Options::vertex::fixup_clipspace for more details.
options.vertex.flip_vert_y = true;
options.vertex.fixup_clipspace = true;
const OpenGLVersion& version = ToBackend(GetDevice())->gl.GetVersion();
if (version.IsDesktop()) {
// The computation of GLSL version below only works for 3.3 and above.
ASSERT(version.IsAtLeast(3, 3));
} = version.IsES();
options.version = version.GetMajor() * 100 + version.GetMinor() * 10;
spirv_cross::CompilerGLSL compiler(std::move(spirv));
compiler.set_entry_point(entryPointName, ShaderStageToExecutionModel(stage));
// Analyzes all OpImageFetch opcodes and checks if there are instances where
// said instruction is used without a combined image sampler.
// GLSL does not support texelFetch without a sampler.
// To workaround this, we must inject a dummy sampler which can be used to form a sampler2D
// at the call-site of texelFetch as necessary.
spirv_cross::VariableID dummySamplerId = compiler.build_dummy_sampler_for_combined_images();
// Extract bindings names so that it can be used to get its location in program.
// Now translate the separate sampler / textures into combined ones and store their info. We
// need to do this before removing the set and binding decorations.
for (const auto& combined : compiler.get_combined_image_samplers()) {
CombinedSampler* info = &combinedSamplers->back();
if (combined.sampler_id == dummySamplerId) {
*needsDummySampler = true;
info->useDummySampler = true;
info->samplerLocation = {};
} else {
info->useDummySampler = false;
info-> = BindGroupIndex(
compiler.get_decoration(combined.sampler_id, spv::DecorationDescriptorSet));
info->samplerLocation.binding = BindingNumber(
compiler.get_decoration(combined.sampler_id, spv::DecorationBinding));
info-> = BindGroupIndex(
compiler.get_decoration(combined.image_id, spv::DecorationDescriptorSet));
info->textureLocation.binding =
BindingNumber(compiler.get_decoration(combined.image_id, spv::DecorationBinding));
compiler.set_name(combined.combined_id, info->GetName());
const BindingInfoArray& bindingInfo = *(;
// Change binding names to be "dawn_binding_<group>_<binding>".
// Also unsets the SPIRV "Binding" decoration as it outputs "layout(binding=)" which
// isn't supported on OSX's OpenGL.
const PipelineLayout::BindingIndexInfo& indices = layout->GetBindingIndexInfo();
// Modify the decoration of variables so that SPIRV-Cross outputs only
// layout(binding=<index>) for interface variables.
// Tint is used for the reflection of bindings for the implicit pipeline layout and
// pipeline/layout validation, but bindingInfo is set to mGLEntryPoints which is the
// SPIRV-Cross reflection. Tint reflects bindings used more precisely than SPIRV-Cross so
// some bindings in bindingInfo might not exist in the layout and querying the layout for
// them would cause an ASSERT. That's why we defensively check that bindings are in the
// layout before modifying them. This slight hack is ok because in the long term we will use
// Tint to produce GLSL.
for (BindGroupIndex group : IterateBitSet(layout->GetBindGroupLayoutsMask())) {
for (const auto& it : bindingInfo[group]) {
const BindGroupLayoutBase* bgl = layout->GetBindGroupLayout(group);
BindingNumber bindingNumber = it.first;
const auto& info = it.second;
if (!bgl->HasBinding(bindingNumber)) {
// Remove the name of the base type. This works around an issue where if the SPIRV
// has two uniform/storage interface variables that point to the same base type,
// then SPIRV-Cross would emit two bindings with type names that conflict:
// layout(binding=0) uniform Buf {...} binding0;
// layout(binding=1) uniform Buf {...} binding1;
compiler.set_name(info.base_type_id, "");
BindingIndex bindingIndex = bgl->GetBindingIndex(bindingNumber);
compiler.unset_decoration(, spv::DecorationDescriptorSet);
compiler.set_decoration(, spv::DecorationBinding,
std::string glsl = compiler.compile();
if (GetDevice()->IsToggleEnabled(Toggle::DumpShaders)) {
std::ostringstream dumpedMsg;
dumpedMsg << "/* Dumped generated GLSL */" << std::endl << glsl;
GetDevice()->EmitLog(WGPULoggingType_Info, dumpedMsg.str().c_str());
return glsl;
}} // namespace dawn_native::opengl