blob: 7535c25a4c2eaed1098be2b82fa6dae6377c263f [file] [log] [blame]
// Copyright 2019 The Chromium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#include "gpu/command_buffer/service/webgpu_decoder_impl.h"
#include <dawn/native/DawnNative.h>
#include <dawn/native/OpenGLBackend.h>
#include <dawn/platform/DawnPlatform.h>
#include <dawn/wire/WireServer.h>
#include <algorithm>
#include <memory>
#include <vector>
#include "base/bits.h"
#include "base/logging.h"
#include "base/memory/ptr_util.h"
#include "base/memory/raw_ptr.h"
#include "base/numerics/checked_math.h"
#include "base/trace_event/trace_event.h"
#include "build/build_config.h"
#include "components/viz/common/resources/resource_format_utils.h"
#include "gpu/command_buffer/common/mailbox.h"
#include "gpu/command_buffer/common/webgpu_cmd_format.h"
#include "gpu/command_buffer/service/command_buffer_service.h"
#include "gpu/command_buffer/service/dawn_instance.h"
#include "gpu/command_buffer/service/dawn_platform.h"
#include "gpu/command_buffer/service/dawn_service_memory_transfer_service.h"
#include "gpu/command_buffer/service/dawn_service_serializer.h"
#include "gpu/command_buffer/service/decoder_client.h"
#include "gpu/command_buffer/service/shared_context_state.h"
#include "gpu/command_buffer/service/shared_image/shared_image_factory.h"
#include "gpu/command_buffer/service/shared_image/shared_image_manager.h"
#include "gpu/command_buffer/service/shared_image/shared_image_representation.h"
#include "gpu/command_buffer/service/skia_utils.h"
#include "gpu/command_buffer/service/webgpu_decoder.h"
#include "gpu/config/gpu_preferences.h"
#include "gpu/webgpu/callback.h"
#include "third_party/skia/include/core/SkPromiseImageTexture.h"
#include "third_party/skia/include/gpu/GrBackendSemaphore.h"
#include "ui/gl/gl_context_egl.h"
#include "ui/gl/gl_surface_egl.h"
#if BUILDFLAG(IS_WIN)
#include <dawn/native/D3D12Backend.h>
#include <dawn/native/VulkanBackend.h>
#include "ui/gl/gl_angle_util_win.h"
#endif
namespace gpu {
namespace webgpu {
namespace {
static constexpr uint32_t kAllowedWritableMailboxTextureUsages =
static_cast<uint32_t>(WGPUTextureUsage_CopyDst |
WGPUTextureUsage_RenderAttachment |
WGPUTextureUsage_StorageBinding);
static constexpr uint32_t kAllowedReadableMailboxTextureUsages =
static_cast<uint32_t>(WGPUTextureUsage_CopySrc |
WGPUTextureUsage_TextureBinding);
static constexpr uint32_t kAllowedMailboxTextureUsages =
kAllowedWritableMailboxTextureUsages | kAllowedReadableMailboxTextureUsages;
WGPUAdapterType PowerPreferenceToDawnAdapterType(
WGPUPowerPreference power_preference) {
switch (power_preference) {
case WGPUPowerPreference_LowPower:
return WGPUAdapterType_IntegratedGPU;
case WGPUPowerPreference_HighPerformance:
// Currently for simplicity we always choose discrete GPU as the device
// related to default power preference.
case WGPUPowerPreference_Undefined:
return WGPUAdapterType_DiscreteGPU;
default:
NOTREACHED();
return WGPUAdapterType_CPU;
}
}
} // namespace
class WebGPUDecoderImpl final : public WebGPUDecoder {
public:
WebGPUDecoderImpl(DecoderClient* client,
CommandBufferServiceBase* command_buffer_service,
SharedImageManager* shared_image_manager,
MemoryTracker* memory_tracker,
gles2::Outputter* outputter,
const GpuPreferences& gpu_preferences,
scoped_refptr<SharedContextState> shared_context_state);
WebGPUDecoderImpl(const WebGPUDecoderImpl&) = delete;
WebGPUDecoderImpl& operator=(const WebGPUDecoderImpl&) = delete;
~WebGPUDecoderImpl() override;
// WebGPUDecoder implementation
ContextResult Initialize(const GpuFeatureInfo& gpu_feature_info) override;
// DecoderContext implementation.
base::WeakPtr<DecoderContext> AsWeakPtr() override {
return weak_ptr_factory_.GetWeakPtr();
}
const gles2::ContextState* GetContextState() override {
NOTREACHED();
return nullptr;
}
void Destroy(bool have_context) override;
bool MakeCurrent() override {
if (gl_context_.get()) {
gl_context_->MakeCurrent(gl_surface_.get());
}
return true;
}
gl::GLContext* GetGLContext() override { return nullptr; }
gl::GLSurface* GetGLSurface() override {
NOTREACHED();
return nullptr;
}
const gles2::FeatureInfo* GetFeatureInfo() const override {
NOTREACHED();
return nullptr;
}
Capabilities GetCapabilities() override { return {}; }
void RestoreGlobalState() const override { NOTREACHED(); }
void ClearAllAttributes() const override { NOTREACHED(); }
void RestoreAllAttributes() const override { NOTREACHED(); }
void RestoreState(const gles2::ContextState* prev_state) override {
NOTREACHED();
}
void RestoreActiveTexture() const override { NOTREACHED(); }
void RestoreAllTextureUnitAndSamplerBindings(
const gles2::ContextState* prev_state) const override {
NOTREACHED();
}
void RestoreActiveTextureUnitBinding(unsigned int target) const override {
NOTREACHED();
}
void RestoreBufferBinding(unsigned int target) override { NOTREACHED(); }
void RestoreBufferBindings() const override { NOTREACHED(); }
void RestoreFramebufferBindings() const override { NOTREACHED(); }
void RestoreRenderbufferBindings() override { NOTREACHED(); }
void RestoreProgramBindings() const override { NOTREACHED(); }
void RestoreTextureState(unsigned service_id) override { NOTREACHED(); }
void RestoreTextureUnitBindings(unsigned unit) const override {
NOTREACHED();
}
void RestoreVertexAttribArray(unsigned index) override { NOTREACHED(); }
void RestoreAllExternalTextureBindingsIfNeeded() override { NOTREACHED(); }
QueryManager* GetQueryManager() override {
NOTREACHED();
return nullptr;
}
void SetQueryCallback(unsigned int query_client_id,
base::OnceClosure callback) override {
NOTREACHED();
}
gles2::GpuFenceManager* GetGpuFenceManager() override {
NOTREACHED();
return nullptr;
}
bool HasPendingQueries() const override { return false; }
void ProcessPendingQueries(bool did_finish) override {}
bool HasMoreIdleWork() const override { return false; }
void PerformIdleWork() override {}
bool HasPollingWork() const override {
return has_polling_work_ || wire_serializer_->NeedsFlush();
}
void PerformPollingWork() override {
TRACE_EVENT0(TRACE_DISABLED_BY_DEFAULT("gpu.dawn"),
"WebGPUDecoderImpl::PerformPollingWork");
has_polling_work_ = false;
if (known_device_metadata_.empty()) {
wire_serializer_->Flush();
return;
}
// Call DeviceTick() on all known devices, and prune any that are no longer
// on the wire and do not need a tick.
for (auto it = known_device_metadata_.begin();
it != known_device_metadata_.end();) {
WGPUDevice device = it->first;
const bool known = wire_server_->IsDeviceKnown(device);
const bool needs_tick = dawn::native::DeviceTick(device);
if (needs_tick) {
has_polling_work_ = true;
}
if (!known) {
// The client has dropped all references and the device has been
// removed from the wire.
// Release the device and erase it from the map.
dawn::native::GetProcs().deviceRelease(device);
it = known_device_metadata_.erase(it);
} else {
++it;
}
}
wire_serializer_->Flush();
}
TextureBase* GetTextureBase(uint32_t client_id) override {
NOTREACHED();
return nullptr;
}
void SetLevelInfo(uint32_t client_id,
int level,
unsigned internal_format,
unsigned width,
unsigned height,
unsigned depth,
unsigned format,
unsigned type,
const gfx::Rect& cleared_rect) override {
NOTREACHED();
}
bool WasContextLost() const override {
NOTIMPLEMENTED();
return false;
}
bool WasContextLostByRobustnessExtension() const override { return false; }
void MarkContextLost(error::ContextLostReason reason) override {
NOTIMPLEMENTED();
}
bool CheckResetStatus() override {
NOTREACHED();
return false;
}
void BeginDecoding() override {}
void EndDecoding() override {}
const char* GetCommandName(unsigned int command_id) const;
error::Error DoCommands(unsigned int num_commands,
const volatile void* buffer,
int num_entries,
int* entries_processed) override;
base::StringPiece GetLogPrefix() override { return "WebGPUDecoderImpl"; }
void BindImage(uint32_t client_texture_id,
uint32_t texture_target,
gl::GLImage* image,
bool can_bind_to_sampler) override {
NOTREACHED();
}
gles2::ContextGroup* GetContextGroup() override { return nullptr; }
gles2::ErrorState* GetErrorState() override {
NOTREACHED();
return nullptr;
}
std::unique_ptr<gles2::AbstractTexture> CreateAbstractTexture(
GLenum target,
GLenum internal_format,
GLsizei width,
GLsizei height,
GLsizei depth,
GLint border,
GLenum format,
GLenum type) override {
NOTREACHED();
return nullptr;
}
bool IsCompressedTextureFormat(unsigned format) override {
NOTREACHED();
return false;
}
bool ClearLevel(gles2::Texture* texture,
unsigned target,
int level,
unsigned format,
unsigned type,
int xoffset,
int yoffset,
int width,
int height) override {
NOTREACHED();
return false;
}
bool ClearCompressedTextureLevel(gles2::Texture* texture,
unsigned target,
int level,
unsigned format,
int width,
int height) override {
NOTREACHED();
return false;
}
bool ClearCompressedTextureLevel3D(gles2::Texture* texture,
unsigned target,
int level,
unsigned format,
int width,
int height,
int depth) override {
NOTREACHED();
return false;
}
bool ClearLevel3D(gles2::Texture* texture,
unsigned target,
int level,
unsigned format,
unsigned type,
int width,
int height,
int depth) override {
NOTREACHED();
return false;
}
bool initialized() const override { return true; }
void SetLogCommands(bool log_commands) override { NOTIMPLEMENTED(); }
gles2::Outputter* outputter() const override {
NOTIMPLEMENTED();
return nullptr;
}
int GetRasterDecoderId() const override {
NOTREACHED();
return -1;
}
private:
typedef error::Error (WebGPUDecoderImpl::*CmdHandler)(
uint32_t immediate_data_size,
const volatile void* data);
// A struct to hold info about each command.
struct CommandInfo {
CmdHandler cmd_handler;
uint8_t arg_flags; // How to handle the arguments for this command
uint8_t cmd_flags; // How to handle this command
uint16_t arg_count; // How many arguments are expected for this command.
};
// A table of CommandInfo for all the commands.
static const CommandInfo command_info[kNumCommands - kFirstWebGPUCommand];
// Generate a member function prototype for each command in an automated and
// typesafe way.
#define WEBGPU_CMD_OP(name) \
Error Handle##name(uint32_t immediate_data_size, const volatile void* data);
WEBGPU_COMMAND_LIST(WEBGPU_CMD_OP)
#undef WEBGPU_CMD_OP
// The current decoder error communicates the decoder error through command
// processing functions that do not return the error value. Should be set
// only if not returning an error.
error::Error current_decoder_error_ = error::kNoError;
void DiscoverAdapters();
int32_t GetPreferredAdapterIndex(WGPUPowerPreference power_preference,
bool force_fallback) const;
bool IsFeatureExposed(WGPUFeatureName feature) const;
// Dawn wire uses procs which forward their calls to these methods.
void RequestAdapterImpl(WGPUInstance instance,
const WGPURequestAdapterOptions* options,
WGPURequestAdapterCallback callback,
void* userdata);
bool AdapterHasFeatureImpl(WGPUAdapter adapter, WGPUFeatureName feature);
size_t AdapterEnumerateFeaturesImpl(WGPUAdapter adapter,
WGPUFeatureName* features);
void RequestDeviceImpl(WGPUAdapter adapter,
const WGPUDeviceDescriptor* descriptor,
WGPURequestDeviceCallback callback,
void* userdata);
class SharedImageRepresentationAndAccess;
std::unique_ptr<SharedImageRepresentationAndAccess> AssociateMailboxDawn(
const Mailbox& mailbox,
MailboxFlags flags,
WGPUDevice device,
WGPUBackendType backendType,
WGPUTextureUsage usage);
std::unique_ptr<SharedImageRepresentationAndAccess>
AssociateMailboxUsingSkiaFallback(const Mailbox& mailbox,
MailboxFlags flags,
WGPUDevice device,
WGPUTextureUsage usage);
scoped_refptr<SharedContextState> shared_context_state_;
const GrContextType gr_context_type_;
std::unique_ptr<SharedImageRepresentationFactory>
shared_image_representation_factory_;
std::unique_ptr<dawn::platform::Platform> dawn_platform_;
std::unique_ptr<DawnInstance> dawn_instance_;
std::unique_ptr<DawnServiceMemoryTransferService> memory_transfer_service_;
std::vector<dawn::native::Adapter> dawn_adapters_;
bool enable_unsafe_webgpu_ = false;
WebGPUAdapterName use_webgpu_adapter_ = WebGPUAdapterName::kDefault;
std::vector<std::string> force_enabled_toggles_;
std::vector<std::string> force_disabled_toggles_;
bool allow_unsafe_apis_;
std::unique_ptr<dawn::wire::WireServer> wire_server_;
std::unique_ptr<DawnServiceSerializer> wire_serializer_;
// Helper class whose derived implementations holds a representation
// and its ScopedAccess, ensuring safe destruction order.
class SharedImageRepresentationAndAccess {
public:
virtual ~SharedImageRepresentationAndAccess() = default;
// Get an unowned reference to the WGPUTexture for the shared image.
virtual WGPUTexture texture() const = 0;
};
// Wraps a |SharedImageRepresentationDawn| as a WGPUTexture.
class SharedImageRepresentationAndAccessDawn
: public SharedImageRepresentationAndAccess {
public:
SharedImageRepresentationAndAccessDawn(
std::unique_ptr<SharedImageRepresentationDawn> representation,
std::unique_ptr<SharedImageRepresentationDawn::ScopedAccess> access)
: representation_(std::move(representation)),
access_(std::move(access)) {}
WGPUTexture texture() const override { return access_->texture(); }
private:
std::unique_ptr<SharedImageRepresentationDawn> representation_;
std::unique_ptr<SharedImageRepresentationDawn::ScopedAccess> access_;
};
// Wraps a |SharedImageRepresentationSkia| and exposes
// it as a WGPUTexture by performing CPU readbacks/uploads.
class SharedImageRepresentationAndAccessSkiaFallback
: public SharedImageRepresentationAndAccess {
public:
static std::unique_ptr<SharedImageRepresentationAndAccessSkiaFallback>
Create(scoped_refptr<SharedContextState> shared_context_state,
std::unique_ptr<SharedImageRepresentationSkia> representation,
const DawnProcTable& procs,
WGPUDevice device,
WGPUTextureUsage usage) {
viz::ResourceFormat format = representation->format();
// Include list of formats this is tested to work with.
// See gpu/command_buffer/tests/webgpu_mailbox_unittest.cc
switch (format) {
// TODO(crbug.com/1241369): Handle additional formats.
#if !BUILDFLAG(IS_MAC)
case viz::ResourceFormat::RGBA_8888:
#endif // !BUILDFLAG(IS_MAC)
case viz::ResourceFormat::BGRA_8888:
case viz::ResourceFormat::RGBA_F16:
break;
default:
return nullptr;
}
// Make sure we can create a WebGPU texture for this format
if (viz::ToWGPUFormat(format) == WGPUTextureFormat_Undefined) {
return nullptr;
}
const bool is_initialized = representation->IsCleared();
auto result =
base::WrapUnique(new SharedImageRepresentationAndAccessSkiaFallback(
std::move(shared_context_state), std::move(representation), procs,
device, usage));
if (is_initialized && !result->PopulateFromSkia()) {
return nullptr;
}
return result;
}
~SharedImageRepresentationAndAccessSkiaFallback() override {
// If we have write access, flush any writes by uploading
// into the SkSurface.
if ((usage_ & kAllowedWritableMailboxTextureUsages) != 0) {
if (UploadContentsToSkia()) {
// Upload to skia was successful. Mark the contents as initialized.
representation_->SetCleared();
} else {
DLOG(ERROR) << "Failed to write to SkSurface.";
}
}
procs_.textureDestroy(texture_);
procs_.textureRelease(texture_);
procs_.deviceRelease(device_);
}
WGPUTexture texture() const override { return texture_; }
private:
SharedImageRepresentationAndAccessSkiaFallback(
scoped_refptr<SharedContextState> shared_context_state,
std::unique_ptr<SharedImageRepresentationSkia> representation,
const DawnProcTable& procs,
WGPUDevice device,
WGPUTextureUsage usage)
: shared_context_state_(std::move(shared_context_state)),
representation_(std::move(representation)),
procs_(procs),
device_(device),
usage_(usage) {
// Create a WGPUTexture to hold the image contents.
// It should be internally copyable so Chrome can internally perform
// copies with it, but Javascript cannot (unless |usage| contains copy
// src/dst).
WGPUDawnTextureInternalUsageDescriptor internal_usage_desc = {
.chain = {.sType = WGPUSType_DawnTextureInternalUsageDescriptor},
.internalUsage =
static_cast<WGPUTextureUsageFlags>(WGPUTextureUsage_CopyDst) |
static_cast<WGPUTextureUsageFlags>(WGPUTextureUsage_CopySrc),
};
WGPUTextureDescriptor texture_desc = {
.nextInChain = &internal_usage_desc.chain,
.usage = static_cast<WGPUTextureUsageFlags>(usage),
.dimension = WGPUTextureDimension_2D,
.size = {static_cast<uint32_t>(representation_->size().width()),
static_cast<uint32_t>(representation_->size().height()), 1},
.format = viz::ToWGPUFormat(representation_->format()),
.mipLevelCount = 1,
.sampleCount = 1,
};
procs_.deviceReference(device_);
texture_ = procs_.deviceCreateTexture(device, &texture_desc);
DCHECK(texture_);
}
bool ComputeStagingBufferParams(const SkImageInfo& image_info,
uint32_t* bytes_per_row,
size_t* buffer_size) const {
DCHECK(bytes_per_row);
DCHECK(buffer_size);
base::CheckedNumeric<uint32_t> checked_bytes_per_row(
image_info.bytesPerPixel());
checked_bytes_per_row *= image_info.width();
uint32_t packed_bytes_per_row;
if (!checked_bytes_per_row.AssignIfValid(&packed_bytes_per_row)) {
return false;
}
// Align up to 256, required by WebGPU buffer->texture and texture->buffer
// copies.
checked_bytes_per_row =
base::bits::AlignUp(packed_bytes_per_row, uint32_t{256});
if (!checked_bytes_per_row.AssignIfValid(bytes_per_row)) {
return false;
}
if (*bytes_per_row < packed_bytes_per_row) {
// Overflow in AlignUp.
return false;
}
base::CheckedNumeric<size_t> checked_buffer_size = checked_bytes_per_row;
checked_buffer_size *= image_info.height();
return checked_buffer_size.AssignIfValid(buffer_size);
}
bool PopulateFromSkia() {
std::vector<GrBackendSemaphore> begin_semaphores;
std::vector<GrBackendSemaphore> end_semaphores;
auto scoped_read_access = representation_->BeginScopedReadAccess(
&begin_semaphores, &end_semaphores);
if (!scoped_read_access) {
DLOG(ERROR) << "PopulateFromSkia: Couldn't begin shared image access";
return false;
}
// Wait for any work that previously used the image.
WaitForSemaphores(std::move(begin_semaphores));
// Make an SkImage to read the image contents
auto sk_image = scoped_read_access->CreateSkImage(
shared_context_state_->gr_context());
if (!sk_image) {
DLOG(ERROR) << "Couldn't make SkImage";
return false;
}
uint32_t bytes_per_row;
size_t buffer_size;
if (!ComputeStagingBufferParams(sk_image->imageInfo(), &bytes_per_row,
&buffer_size)) {
return false;
}
// Create a staging buffer to hold pixel data which will be uploaded into
// a texture.
WGPUBufferDescriptor buffer_desc = {
.usage = WGPUBufferUsage_CopySrc,
.size = buffer_size,
.mappedAtCreation = true,
};
WGPUBuffer buffer = procs_.deviceCreateBuffer(device_, &buffer_desc);
// Read back the Skia image contents into the staging buffer.
void* dst_pointer =
procs_.bufferGetMappedRange(buffer, 0, WGPU_WHOLE_MAP_SIZE);
DCHECK(dst_pointer);
if (!sk_image->readPixels(shared_context_state_->gr_context(),
sk_image->imageInfo(), dst_pointer,
bytes_per_row, 0, 0)) {
procs_.bufferRelease(buffer);
DLOG(ERROR) << "Failed to read from SkImage";
return false;
}
// Unmap the buffer.
procs_.bufferUnmap(buffer);
// Transition the image back to the desired end state. This is used for
// transitioning the image to the external queue for Vulkan/GL interop.
if (auto end_state = scoped_read_access->TakeEndState()) {
if (!shared_context_state_->gr_context()->setBackendTextureState(
scoped_read_access->promise_image_texture()->backendTexture(),
*end_state)) {
DLOG(ERROR) << "setBackendTextureState() failed.";
return false;
}
}
// ReadPixels finished; signal the semaphores.
SignalSemaphores(std::move(end_semaphores));
// Copy from the staging WGPUBuffer into the WGPUTexture.
WGPUDawnEncoderInternalUsageDescriptor internal_usage_desc = {
.chain = {.sType = WGPUSType_DawnEncoderInternalUsageDescriptor},
.useInternalUsages = true,
};
WGPUCommandEncoderDescriptor command_encoder_desc = {
.nextInChain = &internal_usage_desc.chain,
};
WGPUCommandEncoder encoder =
procs_.deviceCreateCommandEncoder(device_, &command_encoder_desc);
WGPUImageCopyBuffer buffer_copy = {
.layout =
{
.bytesPerRow = bytes_per_row,
.rowsPerImage = WGPU_COPY_STRIDE_UNDEFINED,
},
.buffer = buffer,
};
WGPUImageCopyTexture texture_copy = {
.texture = texture_,
};
WGPUExtent3D extent = {
static_cast<uint32_t>(representation_->size().width()),
static_cast<uint32_t>(representation_->size().height()), 1};
procs_.commandEncoderCopyBufferToTexture(encoder, &buffer_copy,
&texture_copy, &extent);
WGPUCommandBuffer commandBuffer =
procs_.commandEncoderFinish(encoder, nullptr);
procs_.commandEncoderRelease(encoder);
WGPUQueue queue = procs_.deviceGetQueue(device_);
procs_.queueSubmit(queue, 1, &commandBuffer);
procs_.commandBufferRelease(commandBuffer);
procs_.queueRelease(queue);
procs_.bufferRelease(buffer);
return true;
}
bool UploadContentsToSkia() {
std::vector<GrBackendSemaphore> begin_semaphores;
std::vector<GrBackendSemaphore> end_semaphores;
auto scoped_write_access = representation_->BeginScopedWriteAccess(
&begin_semaphores, &end_semaphores,
SharedImageRepresentation::AllowUnclearedAccess::kYes);
if (!scoped_write_access) {
DLOG(ERROR)
<< "UploadContentsToSkia: Couldn't begin shared image access";
return false;
}
auto* surface = scoped_write_access->surface();
uint32_t bytes_per_row;
size_t buffer_size;
if (!ComputeStagingBufferParams(surface->imageInfo(), &bytes_per_row,
&buffer_size)) {
return false;
}
// Create a staging buffer to read back from the texture.
WGPUBufferDescriptor buffer_desc = {
.usage = WGPUBufferUsage_CopyDst | WGPUBufferUsage_MapRead,
.size = buffer_size,
};
WGPUBuffer buffer = procs_.deviceCreateBuffer(device_, &buffer_desc);
WGPUImageCopyTexture texture_copy = {
.texture = texture_,
};
WGPUImageCopyBuffer buffer_copy = {
.layout =
{
.bytesPerRow = bytes_per_row,
.rowsPerImage = WGPU_COPY_STRIDE_UNDEFINED,
},
.buffer = buffer,
};
WGPUExtent3D extent = {
static_cast<uint32_t>(representation_->size().width()),
static_cast<uint32_t>(representation_->size().height()), 1};
// Copy from the texture into the staging buffer.
WGPUDawnEncoderInternalUsageDescriptor internal_usage_desc = {
.chain = {.sType = WGPUSType_DawnEncoderInternalUsageDescriptor},
.useInternalUsages = true,
};
WGPUCommandEncoderDescriptor command_encoder_desc = {
.nextInChain = &internal_usage_desc.chain,
};
WGPUCommandEncoder encoder =
procs_.deviceCreateCommandEncoder(device_, &command_encoder_desc);
procs_.commandEncoderCopyTextureToBuffer(encoder, &texture_copy,
&buffer_copy, &extent);
WGPUCommandBuffer commandBuffer =
procs_.commandEncoderFinish(encoder, nullptr);
procs_.commandEncoderRelease(encoder);
WGPUQueue queue = procs_.deviceGetQueue(device_);
procs_.queueSubmit(queue, 1, &commandBuffer);
procs_.commandBufferRelease(commandBuffer);
procs_.queueRelease(queue);
struct Userdata {
bool map_complete = false;
WGPUBufferMapAsyncStatus status;
} userdata;
// Map the staging buffer for read.
procs_.bufferMapAsync(
buffer, WGPUMapMode_Read, 0, WGPU_WHOLE_MAP_SIZE,
[](WGPUBufferMapAsyncStatus status, void* void_userdata) {
Userdata* userdata = static_cast<Userdata*>(void_userdata);
userdata->status = status;
userdata->map_complete = true;
},
&userdata);
// While the map is in flight, wait for any work that previously used the
// image.
WaitForSemaphores(std::move(begin_semaphores));
// Poll for the map to complete.
while (!userdata.map_complete) {
base::PlatformThread::Sleep(base::Milliseconds(1));
procs_.deviceTick(device_);
}
if (userdata.status != WGPUBufferMapAsyncStatus_Success) {
procs_.bufferRelease(buffer);
return false;
}
const void* data =
procs_.bufferGetConstMappedRange(buffer, 0, WGPU_WHOLE_MAP_SIZE);
DCHECK(data);
surface->writePixels(SkPixmap(surface->imageInfo(), data, bytes_per_row),
/*x*/ 0, /*y*/ 0);
procs_.bufferRelease(buffer);
// Transition the image back to the desired end state. This is used for
// transitioning the image to the external queue for Vulkan/GL interop.
if (auto end_state = scoped_write_access->TakeEndState()) {
// It's ok to pass in empty GrFlushInfo here since SignalSemaphores()
// will populate it with semaphores and call GrDirectContext::flush.
scoped_write_access->surface()->flush(/*info=*/{}, end_state.get());
}
SignalSemaphores(std::move(end_semaphores));
return true;
}
void WaitForSemaphores(std::vector<GrBackendSemaphore> semaphores) {
if (semaphores.empty())
return;
bool wait_result = shared_context_state_->gr_context()->wait(
semaphores.size(), semaphores.data(),
/*deleteSemaphoresAfterWait=*/false);
DCHECK(wait_result);
}
void SignalSemaphores(std::vector<GrBackendSemaphore> semaphores) {
if (semaphores.empty())
return;
GrFlushInfo flush_info = {
.fNumSemaphores = semaphores.size(),
.fSignalSemaphores = semaphores.data(),
};
// Note: this is a no-op if vk_context_provider is null.
AddVulkanCleanupTaskForSkiaFlush(
shared_context_state_->vk_context_provider(), &flush_info);
auto flush_result =
shared_context_state_->gr_context()->flush(flush_info);
DCHECK(flush_result == GrSemaphoresSubmitted::kYes);
shared_context_state_->gr_context()->submit();
}
scoped_refptr<SharedContextState> shared_context_state_;
std::unique_ptr<SharedImageRepresentationSkia> representation_;
const DawnProcTable& procs_;
WGPUDevice device_;
WGPUTexture texture_;
WGPUTextureUsage usage_;
};
// Map from the <ID, generation> pair for a wire texture to the shared image
// representation and access for it.
base::flat_map<std::tuple<uint32_t, uint32_t>,
std::unique_ptr<SharedImageRepresentationAndAccess>>
associated_shared_image_map_;
// A container of devices that we've seen on the wire, and their associated
// metadata. Not all of them may be valid, so it gets pruned when
// iterating through it in PerformPollingWork.
struct DeviceMetadata {
WGPUAdapterType adapterType;
WGPUBackendType backendType;
};
std::unordered_map<WGPUDevice, DeviceMetadata> known_device_metadata_;
bool has_polling_work_ = false;
bool destroyed_ = false;
scoped_refptr<gl::GLContext> gl_context_;
scoped_refptr<gl::GLSurface> gl_surface_;
base::WeakPtrFactory<WebGPUDecoderImpl> weak_ptr_factory_{this};
};
constexpr WebGPUDecoderImpl::CommandInfo WebGPUDecoderImpl::command_info[] = {
#define WEBGPU_CMD_OP(name) \
{ \
&WebGPUDecoderImpl::Handle##name, \
cmds::name::kArgFlags, \
cmds::name::cmd_flags, \
sizeof(cmds::name) / sizeof(CommandBufferEntry) - 1, \
}, /* NOLINT */
WEBGPU_COMMAND_LIST(WEBGPU_CMD_OP)
#undef WEBGPU_CMD_OP
};
// DawnWireServer is a wrapper around dawn::wire::WireServer which allows
// overriding some of the WGPU procs the server delegates calls to.
// It has a special feature that around HandleDawnCommands, its owning
// WebGPUDecoderImpl is stored in thread-local storage. This enables
// some of the overridden procs to be overridden with a WebGPUDecoderImpl
// member function. The proc will be set to a plain-old C function pointer,
// which loads the WebGPUDecoderImpl from thread-local storage and forwards
// the call to the member function.
class DawnWireServer : public dawn::wire::WireServer {
// This variable is set to DawnWireServer's parent decoder during
// execution of HandleCommands. It is cleared to nullptr after.
thread_local static WebGPUDecoderImpl* tls_parent_decoder;
// Base template for specialization.
template <auto Method>
struct ProcForDecoderMethodImpl;
public:
// Helper function for making a Proc from a WebGPUDecoderImpl method.
template <auto Method>
static auto ProcForDecoderMethod() {
return ProcForDecoderMethodImpl<Method>{}();
}
template <typename... Procs>
static std::unique_ptr<DawnWireServer> Create(
WebGPUDecoderImpl* decoder,
dawn::wire::CommandSerializer* serializer,
dawn::wire::server::MemoryTransferService* memory_transfer_service,
const DawnProcTable& procs) {
dawn::wire::WireServerDescriptor descriptor = {};
descriptor.procs = &procs;
descriptor.serializer = serializer;
descriptor.memoryTransferService = memory_transfer_service;
return base::WrapUnique(new DawnWireServer(decoder, descriptor));
}
~DawnWireServer() override = default;
// Handle Dawn commands. Forward the call to the base class, but
// set |tls_parent_decoder| around it.
const volatile char* HandleCommands(const volatile char* commands,
size_t size) override {
tls_parent_decoder = decoder_;
const volatile char* rv =
dawn::wire::WireServer::HandleCommands(commands, size);
tls_parent_decoder = nullptr;
return rv;
}
private:
DawnWireServer(WebGPUDecoderImpl* decoder,
const dawn::wire::WireServerDescriptor& desc)
: dawn::wire::WireServer(desc), decoder_(decoder) {}
// Specialization where |Method| is a WebGPUDecoderImpl member function.
// Returns a proc which loads the decoder from thread-local storage
// and forwards the call to the member function.
template <typename R,
typename... Args,
R (WebGPUDecoderImpl::*Method)(Args...)>
struct ProcForDecoderMethodImpl<Method> {
using Proc = R (*)(Args...);
Proc operator()() {
return [](Args... args) -> R {
WebGPUDecoderImpl* decoder = tls_parent_decoder;
DCHECK(decoder);
return (decoder->*Method)(std::forward<Args>(args)...);
};
}
};
raw_ptr<WebGPUDecoderImpl> decoder_;
};
thread_local WebGPUDecoderImpl* DawnWireServer::tls_parent_decoder = nullptr;
WebGPUDecoder* CreateWebGPUDecoderImpl(
DecoderClient* client,
CommandBufferServiceBase* command_buffer_service,
SharedImageManager* shared_image_manager,
MemoryTracker* memory_tracker,
gles2::Outputter* outputter,
const GpuPreferences& gpu_preferences,
scoped_refptr<SharedContextState> shared_context_state) {
return new WebGPUDecoderImpl(
client, command_buffer_service, shared_image_manager, memory_tracker,
outputter, gpu_preferences, std::move(shared_context_state));
}
WebGPUDecoderImpl::WebGPUDecoderImpl(
DecoderClient* client,
CommandBufferServiceBase* command_buffer_service,
SharedImageManager* shared_image_manager,
MemoryTracker* memory_tracker,
gles2::Outputter* outputter,
const GpuPreferences& gpu_preferences,
scoped_refptr<SharedContextState> shared_context_state)
: WebGPUDecoder(client, command_buffer_service, outputter),
shared_context_state_(std::move(shared_context_state)),
gr_context_type_(gpu_preferences.gr_context_type),
shared_image_representation_factory_(
std::make_unique<SharedImageRepresentationFactory>(
shared_image_manager,
memory_tracker)),
dawn_platform_(new DawnPlatform()),
dawn_instance_(
DawnInstance::Create(dawn_platform_.get(), gpu_preferences)),
memory_transfer_service_(new DawnServiceMemoryTransferService(this)),
wire_serializer_(new DawnServiceSerializer(client)) {
enable_unsafe_webgpu_ = gpu_preferences.enable_unsafe_webgpu;
use_webgpu_adapter_ = gpu_preferences.use_webgpu_adapter;
force_enabled_toggles_ = gpu_preferences.enabled_dawn_features_list;
force_disabled_toggles_ = gpu_preferences.disabled_dawn_features_list;
// Only allow unsafe APIs if the disallow_unsafe_apis toggle is explicitly
// disabled.
allow_unsafe_apis_ =
std::find(force_disabled_toggles_.begin(), force_disabled_toggles_.end(),
"disallow_unsafe_apis") != force_disabled_toggles_.end();
DawnProcTable wire_procs = dawn::native::GetProcs();
wire_procs.createInstance =
[](const WGPUInstanceDescriptor*) -> WGPUInstance {
CHECK(false);
return nullptr;
};
wire_procs.instanceRequestAdapter = DawnWireServer::ProcForDecoderMethod<
&WebGPUDecoderImpl::RequestAdapterImpl>();
wire_procs.adapterHasFeature = DawnWireServer::ProcForDecoderMethod<
&WebGPUDecoderImpl::AdapterHasFeatureImpl>();
wire_procs.adapterEnumerateFeatures = DawnWireServer::ProcForDecoderMethod<
&WebGPUDecoderImpl::AdapterEnumerateFeaturesImpl>();
wire_procs.adapterRequestDevice = DawnWireServer::ProcForDecoderMethod<
&WebGPUDecoderImpl::RequestDeviceImpl>();
wire_server_ = DawnWireServer::Create(
this, wire_serializer_.get(), memory_transfer_service_.get(), wire_procs);
wire_server_->InjectInstance(dawn_instance_->Get(), 1, 0);
}
WebGPUDecoderImpl::~WebGPUDecoderImpl() {
Destroy(false);
}
void WebGPUDecoderImpl::Destroy(bool have_context) {
associated_shared_image_map_.clear();
for (auto [device, _] : known_device_metadata_) {
dawn::native::GetProcs().deviceRelease(device);
}
known_device_metadata_.clear();
wire_server_ = nullptr;
destroyed_ = true;
}
ContextResult WebGPUDecoderImpl::Initialize(
const GpuFeatureInfo& gpu_feature_info) {
if (kGpuFeatureStatusSoftware ==
gpu_feature_info.status_values[GPU_FEATURE_TYPE_ACCELERATED_WEBGPU]) {
use_webgpu_adapter_ = WebGPUAdapterName::kSwiftShader;
}
if (use_webgpu_adapter_ == WebGPUAdapterName::kCompat) {
gl_surface_ = new gl::SurfacelessEGL(gl::GLSurfaceEGL::GetGLDisplayEGL(),
gfx::Size(1, 1));
gl::GLContextAttribs attribs;
attribs.client_major_es_version = 3;
attribs.client_minor_es_version = 1;
gl_context_ = new gl::GLContextEGL(nullptr);
gl_context_->Initialize(gl_surface_.get(), attribs);
gl_context_->MakeCurrent(gl_surface_.get());
}
DiscoverAdapters();
return ContextResult::kSuccess;
}
bool WebGPUDecoderImpl::IsFeatureExposed(WGPUFeatureName feature) const {
switch (feature) {
case WGPUFeatureName_TimestampQuery:
case WGPUFeatureName_PipelineStatisticsQuery:
case WGPUFeatureName_ChromiumExperimentalDp4a:
case WGPUFeatureName_DawnMultiPlanarFormats:
case WGPUFeatureName_DepthClipControl:
return allow_unsafe_apis_;
case WGPUFeatureName_Depth32FloatStencil8:
case WGPUFeatureName_TextureCompressionBC:
case WGPUFeatureName_TextureCompressionETC2:
case WGPUFeatureName_TextureCompressionASTC:
case WGPUFeatureName_IndirectFirstInstance:
return true;
default:
return false;
}
}
void WebGPUDecoderImpl::RequestAdapterImpl(
WGPUInstance instance,
const WGPURequestAdapterOptions* options,
WGPURequestAdapterCallback callback,
void* userdata) {
WGPURequestAdapterOptions default_options;
if (options == nullptr) {
default_options = {};
options = &default_options;
}
bool force_fallback_adapter = options->forceFallbackAdapter;
if (use_webgpu_adapter_ == WebGPUAdapterName::kSwiftShader) {
force_fallback_adapter = true;
}
if (gr_context_type_ != GrContextType::kVulkan &&
use_webgpu_adapter_ != WebGPUAdapterName::kCompat) {
#if BUILDFLAG(IS_LINUX)
callback(WGPURequestAdapterStatus_Unavailable, nullptr,
"WebGPU on Linux requires command-line flag "
"--enable-features=Vulkan",
userdata);
return;
#endif // BUILDFLAG(IS_LINUX)
}
int32_t requested_adapter_index = GetPreferredAdapterIndex(
options->powerPreference, force_fallback_adapter);
if (requested_adapter_index < 0) {
// There are no adapters to return since webgpu is not supported here
callback(WGPURequestAdapterStatus_Unavailable, nullptr,
"No available adapters.", userdata);
return;
}
// Currently we treat the index of the adapter in
// dawn_adapters_ as the id of the adapter in the server side.
DCHECK_LT(static_cast<size_t>(requested_adapter_index),
dawn_adapters_.size());
const dawn::native::Adapter& adapter =
dawn_adapters_[requested_adapter_index];
// TODO(crbug.com/1266550): Hide CPU adapters until WebGPU fallback adapters
// are fully tested.
WGPUAdapterProperties properties = {};
adapter.GetProperties(&properties);
if (properties.adapterType == WGPUAdapterType_CPU && !enable_unsafe_webgpu_) {
callback(WGPURequestAdapterStatus_Unavailable, nullptr,
"No available adapters.", userdata);
return;
}
// Callback takes ownership of the reference. Add a ref to pass to the
// callback.
dawn::native::GetProcs().adapterReference(adapter.Get());
callback(WGPURequestAdapterStatus_Success, adapter.Get(), nullptr, userdata);
}
bool WebGPUDecoderImpl::AdapterHasFeatureImpl(WGPUAdapter adapter,
WGPUFeatureName feature) {
if (!dawn::native::GetProcs().adapterHasFeature(adapter, feature)) {
return false;
}
return IsFeatureExposed(feature);
}
size_t WebGPUDecoderImpl::AdapterEnumerateFeaturesImpl(
WGPUAdapter adapter,
WGPUFeatureName* features_out) {
size_t count =
dawn::native::GetProcs().adapterEnumerateFeatures(adapter, nullptr);
std::vector<WGPUFeatureName> features(count);
dawn::native::GetProcs().adapterEnumerateFeatures(adapter, features.data());
auto it =
std::partition(features.begin(), features.end(),
[&](WGPUFeatureName f) { return IsFeatureExposed(f); });
count = std::distance(features.begin(), it);
if (features_out != nullptr) {
memcpy(features_out, features.data(), sizeof(WGPUFeatureName) * count);
}
return count;
}
void WebGPUDecoderImpl::RequestDeviceImpl(
WGPUAdapter adapter,
const WGPUDeviceDescriptor* descriptor,
WGPURequestDeviceCallback callback,
void* userdata) {
// Copy the descriptor so we can modify it.
WGPUDeviceDescriptor desc =
descriptor != nullptr ? *descriptor : WGPUDeviceDescriptor{};
DCHECK_EQ(desc.nextInChain, nullptr);
std::vector<WGPUFeatureName> required_features;
if (desc.requiredFeaturesCount) {
required_features = {
desc.requiredFeatures,
desc.requiredFeatures + desc.requiredFeaturesCount,
};
// Check that no disallowed features were requested. They should be hidden
// by AdapterEnumerateFeaturesImpl.
for (const WGPUFeatureName& f : required_features) {
if (!IsFeatureExposed(f)) {
callback(WGPURequestDeviceStatus_Error, nullptr,
"Disallowed feature requested.", userdata);
return;
}
}
}
// We need to request internal usage to be able to do operations with internal
// SharedImage / interop methods that would need specific usages.
required_features.push_back(WGPUFeatureName_DawnInternalUsages);
// Always enable "multi-planar-formats" as long as available.
if (dawn::native::GetProcs().adapterHasFeature(
adapter, WGPUFeatureName_DawnMultiPlanarFormats)) {
required_features.push_back(WGPUFeatureName_DawnMultiPlanarFormats);
}
desc.requiredFeatures = required_features.data();
desc.requiredFeaturesCount = required_features.size();
// If a new toggle is added here, ForceDawnTogglesForWebGPU() which collects
// info for about:gpu should be updated as well.
WGPUDawnTogglesDeviceDescriptor dawn_toggles = {};
std::vector<const char*> force_enabled_toggles;
std::vector<const char*> force_disabled_toggles;
// Disallows usage of SPIR-V by default for security (we only ensure that WGSL
// is secure), unless --enable-unsafe-webgpu is used.
if (!enable_unsafe_webgpu_) {
force_enabled_toggles.push_back("disallow_spirv");
}
for (const std::string& toggles : force_enabled_toggles_) {
force_enabled_toggles.push_back(toggles.c_str());
}
for (const std::string& toggles : force_disabled_toggles_) {
force_disabled_toggles.push_back(toggles.c_str());
}
dawn_toggles.forceEnabledToggles = force_enabled_toggles.data();
dawn_toggles.forceEnabledTogglesCount = force_enabled_toggles.size();
dawn_toggles.forceDisabledToggles = force_disabled_toggles.data();
dawn_toggles.forceDisabledTogglesCount = force_disabled_toggles.size();
dawn_toggles.chain.sType = WGPUSType_DawnTogglesDeviceDescriptor;
desc.nextInChain = reinterpret_cast<WGPUChainedStruct*>(&dawn_toggles);
bool called = false;
auto* bound_callback = BindWGPUOnceCallback(
[](
// bound arguments
bool* called, //
WebGPUDecoderImpl* decoder, //
WGPUAdapter adapter, //
WGPURequestDeviceCallback original_callback, //
void* original_userdata, //
// callback arguments, minus userdata
WGPURequestDeviceStatus status, //
WGPUDevice device, //
const char* message //
) {
*called = true;
// Forward to the original callback.
original_callback(status, device, message, original_userdata);
if (device) {
// Intercept the response so we can add a device ref to the list of
// known devices that we may need to call DeviceTick on.
WGPUAdapterProperties properties{};
dawn::native::GetProcs().adapterGetProperties(adapter, &properties);
dawn::native::GetProcs().deviceReference(device);
decoder->known_device_metadata_.emplace(
device,
DeviceMetadata{properties.adapterType, properties.backendType});
}
},
&called, this, adapter, callback, userdata);
dawn::native::GetProcs().adapterRequestDevice(
adapter, &desc, bound_callback->UnboundCallback(),
bound_callback->AsUserdata());
// The callback must have been called synchronously. We could allow async
// here, but it would require careful handling of the decoder lifetime.
CHECK(called);
}
void WebGPUDecoderImpl::DiscoverAdapters() {
#if BUILDFLAG(DAWN_ENABLE_BACKEND_OPENGLES)
if (use_webgpu_adapter_ == WebGPUAdapterName::kCompat) {
auto getProc = [](const char* pname) {
return reinterpret_cast<void*>(eglGetProcAddress(pname));
};
dawn::native::opengl::AdapterDiscoveryOptionsES optionsES;
optionsES.getProc = getProc;
dawn_instance_->DiscoverAdapters(&optionsES);
}
#endif
#if BUILDFLAG(IS_WIN)
Microsoft::WRL::ComPtr<ID3D11Device> d3d11_device =
gl::QueryD3D11DeviceObjectFromANGLE();
if (!d3d11_device) {
// In the case where the d3d11 device is nullptr, we want to return a null
// adapter
return;
}
Microsoft::WRL::ComPtr<IDXGIDevice> dxgi_device;
d3d11_device.As(&dxgi_device);
Microsoft::WRL::ComPtr<IDXGIAdapter> dxgi_adapter;
dxgi_device->GetAdapter(&dxgi_adapter);
dawn::native::d3d12::AdapterDiscoveryOptions options(std::move(dxgi_adapter));
dawn_instance_->DiscoverAdapters(&options);
// Also discover the SwiftShader adapter. It will be discovered by default
// for other OSes in DiscoverDefaultAdapters.
dawn::native::vulkan::AdapterDiscoveryOptions swiftShaderOptions;
swiftShaderOptions.forceSwiftShader = true;
dawn_instance_->DiscoverAdapters(&swiftShaderOptions);
#else
dawn_instance_->DiscoverDefaultAdapters();
#endif
std::vector<dawn::native::Adapter> adapters = dawn_instance_->GetAdapters();
for (dawn::native::Adapter& adapter : adapters) {
adapter.SetUseTieredLimits(true);
WGPUAdapterProperties adapterProperties = {};
adapter.GetProperties(&adapterProperties);
const bool is_fallback_adapter =
adapterProperties.adapterType == WGPUAdapterType_CPU &&
adapterProperties.vendorID == 0x1AE0 &&
adapterProperties.deviceID == 0xC0DE;
// The adapter must be able to import external images, or it must be a
// SwiftShader adapter. For SwiftShader, we will perform a manual
// upload/readback to/from shared images.
if (!(adapter.SupportsExternalImages() || is_fallback_adapter)) {
continue;
}
if (use_webgpu_adapter_ == WebGPUAdapterName::kCompat) {
if (adapterProperties.backendType == WGPUBackendType_OpenGLES) {
dawn_adapters_.push_back(adapter);
}
} else if (adapterProperties.backendType != WGPUBackendType_Null &&
adapterProperties.backendType != WGPUBackendType_OpenGL) {
dawn_adapters_.push_back(adapter);
}
}
}
int32_t WebGPUDecoderImpl::GetPreferredAdapterIndex(
WGPUPowerPreference power_preference,
bool force_fallback) const {
WGPUAdapterType preferred_adapter_type =
PowerPreferenceToDawnAdapterType(power_preference);
int32_t discrete_gpu_adapter_index = -1;
int32_t integrated_gpu_adapter_index = -1;
int32_t cpu_adapter_index = -1;
int32_t unknown_adapter_index = -1;
for (int32_t i = 0; i < static_cast<int32_t>(dawn_adapters_.size()); ++i) {
const dawn::native::Adapter& adapter = dawn_adapters_[i];
WGPUAdapterProperties adapterProperties = {};
adapter.GetProperties(&adapterProperties);
if (force_fallback &&
adapterProperties.adapterType != WGPUAdapterType_CPU) {
continue;
}
if (adapterProperties.adapterType == preferred_adapter_type) {
return i;
}
switch (adapterProperties.adapterType) {
case WGPUAdapterType_DiscreteGPU:
discrete_gpu_adapter_index = i;
break;
case WGPUAdapterType_IntegratedGPU:
integrated_gpu_adapter_index = i;
break;
case WGPUAdapterType_CPU:
cpu_adapter_index = i;
break;
case WGPUAdapterType_Unknown:
unknown_adapter_index = i;
break;
default:
NOTREACHED();
break;
}
}
// For now, we always prefer the discrete GPU
if (discrete_gpu_adapter_index >= 0) {
return discrete_gpu_adapter_index;
}
if (integrated_gpu_adapter_index >= 0) {
return integrated_gpu_adapter_index;
}
if (cpu_adapter_index >= 0) {
return cpu_adapter_index;
}
if (unknown_adapter_index >= 0) {
return unknown_adapter_index;
}
return -1;
}
const char* WebGPUDecoderImpl::GetCommandName(unsigned int command_id) const {
if (command_id >= kFirstWebGPUCommand && command_id < kNumCommands) {
return webgpu::GetCommandName(static_cast<CommandId>(command_id));
}
return GetCommonCommandName(static_cast<cmd::CommandId>(command_id));
}
error::Error WebGPUDecoderImpl::DoCommands(unsigned int num_commands,
const volatile void* buffer,
int num_entries,
int* entries_processed) {
DCHECK(entries_processed);
int commands_to_process = num_commands;
error::Error result = error::kNoError;
const volatile CommandBufferEntry* cmd_data =
static_cast<const volatile CommandBufferEntry*>(buffer);
int process_pos = 0;
CommandId command = static_cast<CommandId>(0);
while (process_pos < num_entries && result == error::kNoError &&
commands_to_process--) {
const unsigned int size = cmd_data->value_header.size;
command = static_cast<CommandId>(cmd_data->value_header.command);
if (size == 0) {
result = error::kInvalidSize;
break;
}
if (static_cast<int>(size) + process_pos > num_entries) {
result = error::kOutOfBounds;
break;
}
const unsigned int arg_count = size - 1;
unsigned int command_index = command - kFirstWebGPUCommand;
if (command_index < std::size(command_info)) {
// Prevent all further WebGPU commands from being processed if the server
// is destroyed.
if (destroyed_) {
result = error::kLostContext;
break;
}
const CommandInfo& info = command_info[command_index];
unsigned int info_arg_count = static_cast<unsigned int>(info.arg_count);
if ((info.arg_flags == cmd::kFixed && arg_count == info_arg_count) ||
(info.arg_flags == cmd::kAtLeastN && arg_count >= info_arg_count)) {
uint32_t immediate_data_size = (arg_count - info_arg_count) *
sizeof(CommandBufferEntry); // NOLINT
result = (this->*info.cmd_handler)(immediate_data_size, cmd_data);
} else {
result = error::kInvalidArguments;
}
} else {
result = DoCommonCommand(command, arg_count, cmd_data);
}
if (result == error::kNoError &&
current_decoder_error_ != error::kNoError) {
result = current_decoder_error_;
current_decoder_error_ = error::kNoError;
}
if (result != error::kDeferCommandUntilLater) {
process_pos += size;
cmd_data += size;
}
}
*entries_processed = process_pos;
if (error::IsError(result)) {
LOG(ERROR) << "Error: " << result << " for Command "
<< GetCommandName(command);
}
return result;
}
error::Error WebGPUDecoderImpl::HandleDawnCommands(
uint32_t immediate_data_size,
const volatile void* cmd_data) {
const volatile webgpu::cmds::DawnCommands& c =
*static_cast<const volatile webgpu::cmds::DawnCommands*>(cmd_data);
uint32_t size = static_cast<uint32_t>(c.size);
uint32_t commands_shm_id = static_cast<uint32_t>(c.commands_shm_id);
uint32_t commands_shm_offset = static_cast<uint32_t>(c.commands_shm_offset);
const volatile char* shm_commands = GetSharedMemoryAs<const volatile char*>(
commands_shm_id, commands_shm_offset, size);
if (shm_commands == nullptr) {
return error::kOutOfBounds;
}
TRACE_EVENT_WITH_FLOW0(
TRACE_DISABLED_BY_DEFAULT("gpu.dawn"), "DawnCommands",
(static_cast<uint64_t>(commands_shm_id) << 32) + commands_shm_offset,
TRACE_EVENT_FLAG_FLOW_IN);
TRACE_EVENT1(TRACE_DISABLED_BY_DEFAULT("gpu.dawn"),
"WebGPUDecoderImpl::HandleDawnCommands", "bytes", size);
if (!wire_server_->HandleCommands(shm_commands, size)) {
return error::kLostContext;
}
// TODO(crbug.com/1174145): This is O(N) where N is the number of devices.
// Multiple submits would be O(N*M). We should find a way to more
// intelligently poll for work on only the devices that need it.
PerformPollingWork();
return error::kNoError;
}
std::unique_ptr<WebGPUDecoderImpl::SharedImageRepresentationAndAccess>
WebGPUDecoderImpl::AssociateMailboxDawn(const Mailbox& mailbox,
MailboxFlags flags,
WGPUDevice device,
WGPUBackendType backendType,
WGPUTextureUsage usage) {
std::unique_ptr<SharedImageRepresentationDawn> shared_image =
shared_image_representation_factory_->ProduceDawn(mailbox, device,
backendType);
if (!shared_image) {
DLOG(ERROR) << "AssociateMailbox: Couldn't produce shared image";
return nullptr;
}
#if !BUILDFLAG(IS_WIN) && !BUILDFLAG(IS_CHROMEOS)
if (usage & WGPUTextureUsage_StorageBinding) {
DLOG(ERROR) << "AssociateMailbox: WGPUTextureUsage_StorageBinding is NOT "
"supported yet.";
return nullptr;
}
#endif
if (flags & WEBGPU_MAILBOX_DISCARD) {
// Set contents to uncleared.
shared_image->SetClearedRect(gfx::Rect());
}
std::unique_ptr<SharedImageRepresentationDawn::ScopedAccess> scoped_access =
shared_image->BeginScopedAccess(
usage, SharedImageRepresentation::AllowUnclearedAccess::kYes);
if (!scoped_access) {
DLOG(ERROR) << "AssociateMailbox: Couldn't begin shared image access";
return nullptr;
}
return std::make_unique<SharedImageRepresentationAndAccessDawn>(
std::move(shared_image), std::move(scoped_access));
}
std::unique_ptr<WebGPUDecoderImpl::SharedImageRepresentationAndAccess>
WebGPUDecoderImpl::AssociateMailboxUsingSkiaFallback(const Mailbox& mailbox,
MailboxFlags flags,
WGPUDevice device,
WGPUTextureUsage usage) {
// Produce a Skia image from the mailbox.
std::unique_ptr<SharedImageRepresentationSkia> shared_image =
shared_image_representation_factory_->ProduceSkia(
mailbox, shared_context_state_.get());
if (!shared_image) {
DLOG(ERROR) << "AssociateMailbox: Couldn't produce shared image";
return nullptr;
}
if (flags & WEBGPU_MAILBOX_DISCARD) {
// Set contents to uncleared.
shared_image->SetClearedRect(gfx::Rect());
}
return SharedImageRepresentationAndAccessSkiaFallback::Create(
shared_context_state_, std::move(shared_image), dawn::native::GetProcs(),
device, usage);
}
error::Error WebGPUDecoderImpl::HandleAssociateMailboxImmediate(
uint32_t immediate_data_size,
const volatile void* cmd_data) {
const volatile webgpu::cmds::AssociateMailboxImmediate& c =
*static_cast<const volatile webgpu::cmds::AssociateMailboxImmediate*>(
cmd_data);
uint32_t device_id = static_cast<uint32_t>(c.device_id);
uint32_t device_generation = static_cast<uint32_t>(c.device_generation);
uint32_t id = static_cast<uint32_t>(c.id);
uint32_t generation = static_cast<uint32_t>(c.generation);
WGPUTextureUsage usage = static_cast<WGPUTextureUsage>(c.usage);
MailboxFlags flags = static_cast<MailboxFlags>(c.flags);
// Unpack the mailbox
if (sizeof(Mailbox) > immediate_data_size) {
return error::kOutOfBounds;
}
volatile const GLbyte* mailbox_bytes =
gles2::GetImmediateDataAs<volatile const GLbyte*>(c, sizeof(Mailbox),
immediate_data_size);
if (mailbox_bytes == nullptr) {
return error::kOutOfBounds;
}
Mailbox mailbox = Mailbox::FromVolatile(
*reinterpret_cast<const volatile Mailbox*>(mailbox_bytes));
DLOG_IF(ERROR, !mailbox.Verify())
<< "AssociateMailbox was passed an invalid mailbox";
if (usage & ~kAllowedMailboxTextureUsages) {
DLOG(ERROR) << "AssociateMailbox: Invalid usage";
return error::kInvalidArguments;
}
WGPUDevice device = wire_server_->GetDevice(device_id, device_generation);
if (device == nullptr) {
return error::kInvalidArguments;
}
std::unique_ptr<SharedImageRepresentationAndAccess> representation_and_access;
auto it = known_device_metadata_.find(device);
DCHECK(it != known_device_metadata_.end());
if (it->second.adapterType == WGPUAdapterType_CPU) {
representation_and_access =
AssociateMailboxUsingSkiaFallback(mailbox, flags, device, usage);
} else {
representation_and_access = AssociateMailboxDawn(
mailbox, flags, device, it->second.backendType, usage);
}
if (!representation_and_access) {
return error::kInvalidArguments;
}
// Inject the texture in the dawn::wire::Server and remember which shared
// image it is associated with.
if (!wire_server_->InjectTexture(representation_and_access->texture(), id,
generation, device_id, device_generation)) {
DLOG(ERROR) << "AssociateMailbox: Invalid texture ID";
return error::kInvalidArguments;
}
std::tuple<uint32_t, uint32_t> id_and_generation{id, generation};
auto insertion = associated_shared_image_map_.emplace(
id_and_generation, std::move(representation_and_access));
// InjectTexture already validated that the (ID, generation) can't have been
// registered before.
DCHECK(insertion.second);
return error::kNoError;
}
error::Error WebGPUDecoderImpl::HandleDissociateMailbox(
uint32_t immediate_data_size,
const volatile void* cmd_data) {
const volatile webgpu::cmds::DissociateMailbox& c =
*static_cast<const volatile webgpu::cmds::DissociateMailbox*>(cmd_data);
uint32_t texture_id = static_cast<uint32_t>(c.texture_id);
uint32_t texture_generation = static_cast<uint32_t>(c.texture_generation);
std::tuple<uint32_t, uint32_t> id_and_generation{texture_id,
texture_generation};
auto it = associated_shared_image_map_.find(id_and_generation);
if (it == associated_shared_image_map_.end()) {
DLOG(ERROR) << "DissociateMailbox: Invalid texture ID";
return error::kInvalidArguments;
}
associated_shared_image_map_.erase(it);
return error::kNoError;
}
error::Error WebGPUDecoderImpl::HandleDissociateMailboxForPresent(
uint32_t immediate_data_size,
const volatile void* cmd_data) {
const volatile webgpu::cmds::DissociateMailboxForPresent& c =
*static_cast<const volatile webgpu::cmds::DissociateMailboxForPresent*>(
cmd_data);
uint32_t device_id = static_cast<uint32_t>(c.device_id);
uint32_t device_generation = static_cast<uint32_t>(c.device_generation);
uint32_t texture_id = static_cast<uint32_t>(c.texture_id);
uint32_t texture_generation = static_cast<uint32_t>(c.texture_generation);
std::tuple<uint32_t, uint32_t> id_and_generation{texture_id,
texture_generation};
auto it = associated_shared_image_map_.find(id_and_generation);
if (it == associated_shared_image_map_.end()) {
DLOG(ERROR) << "DissociateMailbox: Invalid texture ID";
return error::kInvalidArguments;
}
WGPUDevice device = wire_server_->GetDevice(device_id, device_generation);
if (device == nullptr) {
return error::kInvalidArguments;
}
WGPUTexture texture = it->second->texture();
DCHECK(texture);
if (!dawn::native::IsTextureSubresourceInitialized(texture, 0, 1, 0, 1)) {
// The compositor renders uninitialized textures as red. If the texture is
// not initialized, we need to explicitly clear its contents to black.
// TODO(crbug.com/1242712): Use the C++ WebGPU API.
const auto& procs = dawn::native::GetProcs();
WGPUTextureView view = procs.textureCreateView(texture, nullptr);
WGPURenderPassColorAttachment color_attachment = {};
color_attachment.view = view;
color_attachment.loadOp = WGPULoadOp_Clear;
color_attachment.storeOp = WGPUStoreOp_Store;
color_attachment.clearValue = {0.0, 0.0, 0.0, 0.0};
WGPURenderPassDescriptor render_pass_descriptor = {};
render_pass_descriptor.colorAttachmentCount = 1;
render_pass_descriptor.colorAttachments = &color_attachment;
WGPUDawnEncoderInternalUsageDescriptor internal_usage_desc = {
.chain = {.sType = WGPUSType_DawnEncoderInternalUsageDescriptor},
.useInternalUsages = true,
};
WGPUCommandEncoderDescriptor command_encoder_desc = {
.nextInChain = &internal_usage_desc.chain,
};
WGPUCommandEncoder encoder =
procs.deviceCreateCommandEncoder(device, &command_encoder_desc);
WGPURenderPassEncoder pass =
procs.commandEncoderBeginRenderPass(encoder, &render_pass_descriptor);
procs.renderPassEncoderEndPass(pass);
WGPUCommandBuffer command_buffer =
procs.commandEncoderFinish(encoder, nullptr);
WGPUQueue queue = procs.deviceGetQueue(device);
procs.queueSubmit(queue, 1, &command_buffer);
procs.queueRelease(queue);
procs.commandBufferRelease(command_buffer);
procs.renderPassEncoderRelease(pass);
procs.commandEncoderRelease(encoder);
procs.textureViewRelease(view);
DCHECK(dawn::native::IsTextureSubresourceInitialized(texture, 0, 1, 0, 1));
}
associated_shared_image_map_.erase(it);
return error::kNoError;
}
} // namespace webgpu
} // namespace gpu