encounter/dawn-cmake
1
0
Fork 0
mirror of https://github.com/encounter/dawn-cmake.git synced 2025-05-16 04:11:25 +00:00
Code Issues Packages Projects Releases Wiki Activity
dawn-cmake/src/dawn_native/vulkan/DeviceVk.cpp
Corentin Wallez fefb452f29 Reland: Vulkan: Enforce fixed subgroup size for compute shaders. 
Reland after a refactor of how the extension handling work in the Vulkan
backend.

The original author is David Turner <david.turner.dev@gmail.com>.

This CL ensures that, on architectures with a varying subgroup size,
compute shaders are always compiled with a fixed subgroup size to
avoid consistency issues when one shader writes data in a subgroup-size
dependent layout to GPU memory, to be read by another shader in a
future dispatch.

At the moment, only Intel ICDs are known to implement this [1],
and the code uses a heuristics to chose the size of 16, which seems to
be the sweet spot according to Intel engineers.

+ Update the PNextChainBuilder class to deal with the fact that
  VkComputePipelineCreateInfo::pNext is defined as a const void*,
  which created compiler errors in the previous implementation.

[1] https://bugs.freedesktop.org/show_bug.cgi?id=108875

Bug: dawn:464
Change-Id: I035ee06084fcc964742f0bff4c54cff257c742ae
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/23202
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
Reviewed-by: Stephen White <senorblanco@chromium.org>
2020-06-16 09:17:48 +00:00

857 lines
36 KiB
C++
Raw Blame History

// 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
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "dawn_native/vulkan/DeviceVk.h"
#include "common/Platform.h"
#include "dawn_native/BackendConnection.h"
#include "dawn_native/Commands.h"
#include "dawn_native/Error.h"
#include "dawn_native/ErrorData.h"
#include "dawn_native/VulkanBackend.h"
#include "dawn_native/vulkan/AdapterVk.h"
#include "dawn_native/vulkan/BackendVk.h"
#include "dawn_native/vulkan/BindGroupLayoutVk.h"
#include "dawn_native/vulkan/BindGroupVk.h"
#include "dawn_native/vulkan/BufferVk.h"
#include "dawn_native/vulkan/CommandBufferVk.h"
#include "dawn_native/vulkan/ComputePipelineVk.h"
#include "dawn_native/vulkan/FencedDeleter.h"
#include "dawn_native/vulkan/PipelineLayoutVk.h"
#include "dawn_native/vulkan/QueueVk.h"
#include "dawn_native/vulkan/RenderPassCache.h"
#include "dawn_native/vulkan/RenderPipelineVk.h"
#include "dawn_native/vulkan/ResourceMemoryAllocatorVk.h"
#include "dawn_native/vulkan/SamplerVk.h"
#include "dawn_native/vulkan/ShaderModuleVk.h"
#include "dawn_native/vulkan/StagingBufferVk.h"
#include "dawn_native/vulkan/SwapChainVk.h"
#include "dawn_native/vulkan/TextureVk.h"
#include "dawn_native/vulkan/UtilsVulkan.h"
#include "dawn_native/vulkan/VulkanError.h"
namespace dawn_native { namespace vulkan {
// static
ResultOrError<Device*> Device::Create(Adapter* adapter, const DeviceDescriptor* descriptor) {
Ref<Device> device = AcquireRef(new Device(adapter, descriptor));
DAWN_TRY(device->Initialize());
return device.Detach();
}
Device::Device(Adapter* adapter, const DeviceDescriptor* descriptor)
: DeviceBase(adapter, descriptor) {
InitTogglesFromDriver();
}
MaybeError Device::Initialize() {
// Copy the adapter's device info to the device so that we can change the "knobs"
mDeviceInfo = ToBackend(GetAdapter())->GetDeviceInfo();
// Initialize the "instance" procs of our local function table.
VulkanFunctions* functions = GetMutableFunctions();
*functions = ToBackend(GetAdapter())->GetBackend()->GetFunctions();
// Two things are crucial if device initialization fails: the function pointers to destroy
// objects, and the fence deleter that calls these functions. Do not do anything before
// these two are set up, so that a failed initialization doesn't cause a crash in
// ShutDownImpl()
{
VkPhysicalDevice physicalDevice = ToBackend(GetAdapter())->GetPhysicalDevice();
VulkanDeviceKnobs usedDeviceKnobs = {};
DAWN_TRY_ASSIGN(usedDeviceKnobs, CreateDevice(physicalDevice));
*static_cast<VulkanDeviceKnobs*>(&mDeviceInfo) = usedDeviceKnobs;
DAWN_TRY(functions->LoadDeviceProcs(mVkDevice, mDeviceInfo));
// The queue can be loaded before the fenced deleter because their lifetime is tied to
// the device.
GatherQueueFromDevice();
mDeleter = std::make_unique<FencedDeleter>(this);
}
mRenderPassCache = std::make_unique<RenderPassCache>(this);
mResourceMemoryAllocator = std::make_unique<ResourceMemoryAllocator>(this);
mExternalMemoryService = std::make_unique<external_memory::Service>(this);
mExternalSemaphoreService = std::make_unique<external_semaphore::Service>(this);
DAWN_TRY(PrepareRecordingContext());
// The environment can request to use D32S8 or D24S8 when it's not available. Override
// the decision if it is not applicable.
ApplyDepth24PlusS8Toggle();
return DeviceBase::Initialize(new Queue(this));
}
Device::~Device() {
ShutDownBase();
}
ResultOrError<BindGroupBase*> Device::CreateBindGroupImpl(
const BindGroupDescriptor* descriptor) {
return BindGroup::Create(this, descriptor);
}
ResultOrError<BindGroupLayoutBase*> Device::CreateBindGroupLayoutImpl(
const BindGroupLayoutDescriptor* descriptor) {
return BindGroupLayout::Create(this, descriptor);
}
ResultOrError<BufferBase*> Device::CreateBufferImpl(const BufferDescriptor* descriptor) {
return Buffer::Create(this, descriptor);
}
CommandBufferBase* Device::CreateCommandBuffer(CommandEncoder* encoder,
const CommandBufferDescriptor* descriptor) {
return CommandBuffer::Create(encoder, descriptor);
}
ResultOrError<ComputePipelineBase*> Device::CreateComputePipelineImpl(
const ComputePipelineDescriptor* descriptor) {
return ComputePipeline::Create(this, descriptor);
}
ResultOrError<PipelineLayoutBase*> Device::CreatePipelineLayoutImpl(
const PipelineLayoutDescriptor* descriptor) {
return PipelineLayout::Create(this, descriptor);
}
ResultOrError<QuerySetBase*> Device::CreateQuerySetImpl(const QuerySetDescriptor* descriptor) {
return DAWN_UNIMPLEMENTED_ERROR("Waiting for implementation");
}
ResultOrError<RenderPipelineBase*> Device::CreateRenderPipelineImpl(
const RenderPipelineDescriptor* descriptor) {
return RenderPipeline::Create(this, descriptor);
}
ResultOrError<SamplerBase*> Device::CreateSamplerImpl(const SamplerDescriptor* descriptor) {
return Sampler::Create(this, descriptor);
}
ResultOrError<ShaderModuleBase*> Device::CreateShaderModuleImpl(
const ShaderModuleDescriptor* descriptor) {
return ShaderModule::Create(this, descriptor);
}
ResultOrError<SwapChainBase*> Device::CreateSwapChainImpl(
const SwapChainDescriptor* descriptor) {
return SwapChain::Create(this, descriptor);
}
ResultOrError<NewSwapChainBase*> Device::CreateSwapChainImpl(
Surface* surface,
NewSwapChainBase* previousSwapChain,
const SwapChainDescriptor* descriptor) {
return DAWN_VALIDATION_ERROR("New swapchains not implemented.");
}
ResultOrError<Ref<TextureBase>> Device::CreateTextureImpl(const TextureDescriptor* descriptor) {
return Texture::Create(this, descriptor);
}
ResultOrError<TextureViewBase*> Device::CreateTextureViewImpl(
TextureBase* texture,
const TextureViewDescriptor* descriptor) {
return TextureView::Create(texture, descriptor);
}
MaybeError Device::TickImpl() {
RecycleCompletedCommands();
Serial completedSerial = GetCompletedCommandSerial();
for (Ref<BindGroupLayout>& bgl :
mBindGroupLayoutsPendingDeallocation.IterateUpTo(completedSerial)) {
bgl->FinishDeallocation(completedSerial);
}
mBindGroupLayoutsPendingDeallocation.ClearUpTo(completedSerial);
mResourceMemoryAllocator->Tick(completedSerial);
mDeleter->Tick(completedSerial);
if (mRecordingContext.used) {
DAWN_TRY(SubmitPendingCommands());
}
return {};
}
VkInstance Device::GetVkInstance() const {
return ToBackend(GetAdapter())->GetBackend()->GetVkInstance();
}
const VulkanDeviceInfo& Device::GetDeviceInfo() const {
return mDeviceInfo;
}
VkDevice Device::GetVkDevice() const {
return mVkDevice;
}
uint32_t Device::GetGraphicsQueueFamily() const {
return mQueueFamily;
}
VkQueue Device::GetQueue() const {
return mQueue;
}
FencedDeleter* Device::GetFencedDeleter() const {
return mDeleter.get();
}
RenderPassCache* Device::GetRenderPassCache() const {
return mRenderPassCache.get();
}
void Device::EnqueueDeferredDeallocation(BindGroupLayout* bindGroupLayout) {
mBindGroupLayoutsPendingDeallocation.Enqueue(bindGroupLayout, GetPendingCommandSerial());
}
CommandRecordingContext* Device::GetPendingRecordingContext() {
ASSERT(mRecordingContext.commandBuffer != VK_NULL_HANDLE);
mRecordingContext.used = true;
return &mRecordingContext;
}
MaybeError Device::SubmitPendingCommands() {
if (!mRecordingContext.used) {
return {};
}
DAWN_TRY(CheckVkSuccess(fn.EndCommandBuffer(mRecordingContext.commandBuffer),
"vkEndCommandBuffer"));
std::vector<VkPipelineStageFlags> dstStageMasks(mRecordingContext.waitSemaphores.size(),
VK_PIPELINE_STAGE_ALL_COMMANDS_BIT);
VkSubmitInfo submitInfo;
submitInfo.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
submitInfo.pNext = nullptr;
submitInfo.waitSemaphoreCount =
static_cast<uint32_t>(mRecordingContext.waitSemaphores.size());
submitInfo.pWaitSemaphores = AsVkArray(mRecordingContext.waitSemaphores.data());
submitInfo.pWaitDstStageMask = dstStageMasks.data();
submitInfo.commandBufferCount = 1;
submitInfo.pCommandBuffers = &mRecordingContext.commandBuffer;
submitInfo.signalSemaphoreCount =
static_cast<uint32_t>(mRecordingContext.signalSemaphores.size());
submitInfo.pSignalSemaphores = AsVkArray(mRecordingContext.signalSemaphores.data());
VkFence fence = VK_NULL_HANDLE;
DAWN_TRY_ASSIGN(fence, GetUnusedFence());
DAWN_TRY(CheckVkSuccess(fn.QueueSubmit(mQueue, 1, &submitInfo, fence), "vkQueueSubmit"));
// Enqueue the semaphores before incrementing the serial, so that they can be deleted as
// soon as the current submission is finished.
for (VkSemaphore semaphore : mRecordingContext.waitSemaphores) {
mDeleter->DeleteWhenUnused(semaphore);
}
for (VkSemaphore semaphore : mRecordingContext.signalSemaphores) {
mDeleter->DeleteWhenUnused(semaphore);
}
IncrementLastSubmittedCommandSerial();
Serial lastSubmittedSerial = GetLastSubmittedCommandSerial();
mFencesInFlight.emplace(fence, lastSubmittedSerial);
CommandPoolAndBuffer submittedCommands = {mRecordingContext.commandPool,
mRecordingContext.commandBuffer};
mCommandsInFlight.Enqueue(submittedCommands, lastSubmittedSerial);
mRecordingContext = CommandRecordingContext();
DAWN_TRY(PrepareRecordingContext());
return {};
}
ResultOrError<VulkanDeviceKnobs> Device::CreateDevice(VkPhysicalDevice physicalDevice) {
VulkanDeviceKnobs usedKnobs = {};
// Default to asking for all avilable known extensions.
usedKnobs.extensions = mDeviceInfo.extensions;
// However only request the extensions that haven't been promoted in the device's apiVersion
std::vector<const char*> extensionNames;
for (uint32_t ext : IterateBitSet(usedKnobs.extensions.extensionBitSet)) {
const DeviceExtInfo& info = GetDeviceExtInfo(static_cast<DeviceExt>(ext));
if (info.versionPromoted > mDeviceInfo.properties.apiVersion) {
extensionNames.push_back(info.name);
}
}
// Some device features can only be enabled using a VkPhysicalDeviceFeatures2 struct, which
// is supported by the VK_EXT_get_physical_properties2 instance extension, which was
// promoted as a core API in Vulkan 1.1.
//
// Prepare a VkPhysicalDeviceFeatures2 struct for this use case, it will only be populated
// if HasExt(DeviceExt::GetPhysicalDeviceProperties2) is true.
VkPhysicalDeviceFeatures2 features2 = {};
features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2;
PNextChainBuilder featuresChain(&features2);
// Always require independentBlend because it is a core Dawn feature
usedKnobs.features.independentBlend = VK_TRUE;
// Always require imageCubeArray because it is a core Dawn feature
usedKnobs.features.imageCubeArray = VK_TRUE;
// Always require fragmentStoresAndAtomics because it is required by end2end tests.
usedKnobs.features.fragmentStoresAndAtomics = VK_TRUE;
if (mDeviceInfo.HasExt(DeviceExt::SubgroupSizeControl)) {
ASSERT(usedKnobs.HasExt(DeviceExt::SubgroupSizeControl));
// Always request all the features from VK_EXT_subgroup_size_control when available.
usedKnobs.subgroupSizeControlFeatures = mDeviceInfo.subgroupSizeControlFeatures;
featuresChain.Add(&usedKnobs.subgroupSizeControlFeatures);
mComputeSubgroupSize = FindComputeSubgroupSize();
}
if (IsExtensionEnabled(Extension::TextureCompressionBC)) {
ASSERT(ToBackend(GetAdapter())->GetDeviceInfo().features.textureCompressionBC ==
VK_TRUE);
usedKnobs.features.textureCompressionBC = VK_TRUE;
}
if (IsExtensionEnabled(Extension::ShaderFloat16)) {
const VulkanDeviceInfo& deviceInfo = ToBackend(GetAdapter())->GetDeviceInfo();
ASSERT(deviceInfo.HasExt(DeviceExt::ShaderFloat16Int8) &&
deviceInfo.shaderFloat16Int8Features.shaderFloat16 == VK_TRUE &&
deviceInfo.HasExt(DeviceExt::_16BitStorage) &&
deviceInfo._16BitStorageFeatures.uniformAndStorageBuffer16BitAccess == VK_TRUE);
usedKnobs.shaderFloat16Int8Features.shaderFloat16 = VK_TRUE;
usedKnobs._16BitStorageFeatures.uniformAndStorageBuffer16BitAccess = VK_TRUE;
}
// Find a universal queue family
{
// Note that GRAPHICS and COMPUTE imply TRANSFER so we don't need to check for it.
constexpr uint32_t kUniversalFlags = VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT;
int universalQueueFamily = -1;
for (unsigned int i = 0; i < mDeviceInfo.queueFamilies.size(); ++i) {
if ((mDeviceInfo.queueFamilies[i].queueFlags & kUniversalFlags) ==
kUniversalFlags) {
universalQueueFamily = i;
break;
}
}
if (universalQueueFamily == -1) {
return DAWN_INTERNAL_ERROR("No universal queue family");
}
mQueueFamily = static_cast<uint32_t>(universalQueueFamily);
}
// Choose to create a single universal queue
std::vector<VkDeviceQueueCreateInfo> queuesToRequest;
float zero = 0.0f;
{
VkDeviceQueueCreateInfo queueCreateInfo;
queueCreateInfo.sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO;
queueCreateInfo.pNext = nullptr;
queueCreateInfo.flags = 0;
queueCreateInfo.queueFamilyIndex = static_cast<uint32_t>(mQueueFamily);
queueCreateInfo.queueCount = 1;
queueCreateInfo.pQueuePriorities = &zero;
queuesToRequest.push_back(queueCreateInfo);
}
VkDeviceCreateInfo createInfo;
createInfo.sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO;
createInfo.pNext = nullptr;
createInfo.flags = 0;
createInfo.queueCreateInfoCount = static_cast<uint32_t>(queuesToRequest.size());
createInfo.pQueueCreateInfos = queuesToRequest.data();
createInfo.enabledLayerCount = 0;
createInfo.ppEnabledLayerNames = nullptr;
createInfo.enabledExtensionCount = static_cast<uint32_t>(extensionNames.size());
createInfo.ppEnabledExtensionNames = extensionNames.data();
// When we have DeviceExt::GetPhysicalDeviceProperties2, use features2 so that features not
// covered by VkPhysicalDeviceFeatures can be enabled.
if (mDeviceInfo.HasExt(DeviceExt::GetPhysicalDeviceProperties2)) {
features2.features = usedKnobs.features;
createInfo.pNext = &features2;
createInfo.pEnabledFeatures = nullptr;
} else {
ASSERT(features2.pNext == nullptr);
createInfo.pEnabledFeatures = &usedKnobs.features;
}
DAWN_TRY(CheckVkSuccess(fn.CreateDevice(physicalDevice, &createInfo, nullptr, &mVkDevice),
"vkCreateDevice"));
return usedKnobs;
}
uint32_t Device::FindComputeSubgroupSize() const {
if (!mDeviceInfo.HasExt(DeviceExt::SubgroupSizeControl)) {
return 0;
}
const VkPhysicalDeviceSubgroupSizeControlPropertiesEXT& ext =
mDeviceInfo.subgroupSizeControlProperties;
if (ext.minSubgroupSize == ext.maxSubgroupSize) {
return 0;
}
// At the moment, only Intel devices support varying subgroup sizes and 16, which is the
// next value after the minimum of 8, is the sweet spot according to [1]. Hence the
// following heuristics, which may need to be adjusted in the future for other
// architectures, or if a specific API is added to let client code select the size..
//
// [1] https://bugs.freedesktop.org/show_bug.cgi?id=108875
uint32_t subgroupSize = ext.minSubgroupSize * 2;
if (subgroupSize <= ext.maxSubgroupSize) {
return subgroupSize;
} else {
return ext.minSubgroupSize;
}
}
void Device::GatherQueueFromDevice() {
fn.GetDeviceQueue(mVkDevice, mQueueFamily, 0, &mQueue);
}
void Device::InitTogglesFromDriver() {
// TODO(jiawei.shao@intel.com): tighten this workaround when this issue is fixed in both
// Vulkan SPEC and drivers.
SetToggle(Toggle::UseTemporaryBufferInCompressedTextureToTextureCopy, true);
// By default try to use D32S8 for Depth24PlusStencil8
SetToggle(Toggle::VulkanUseD32S8, true);
}
void Device::ApplyDepth24PlusS8Toggle() {
VkPhysicalDevice physicalDevice = ToBackend(GetAdapter())->GetPhysicalDevice();
bool supportsD32s8 = false;
{
VkFormatProperties properties;
fn.GetPhysicalDeviceFormatProperties(physicalDevice, VK_FORMAT_D32_SFLOAT_S8_UINT,
&properties);
supportsD32s8 =
properties.optimalTilingFeatures & VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT;
}
bool supportsD24s8 = false;
{
VkFormatProperties properties;
fn.GetPhysicalDeviceFormatProperties(physicalDevice, VK_FORMAT_D24_UNORM_S8_UINT,
&properties);
supportsD24s8 =
properties.optimalTilingFeatures & VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT;
}
ASSERT(supportsD32s8 || supportsD24s8);
if (!supportsD24s8) {
ForceSetToggle(Toggle::VulkanUseD32S8, true);
}
if (!supportsD32s8) {
ForceSetToggle(Toggle::VulkanUseD32S8, false);
}
}
VulkanFunctions* Device::GetMutableFunctions() {
return const_cast<VulkanFunctions*>(&fn);
}
ResultOrError<VkFence> Device::GetUnusedFence() {
if (!mUnusedFences.empty()) {
VkFence fence = mUnusedFences.back();
DAWN_TRY(CheckVkSuccess(fn.ResetFences(mVkDevice, 1, &*fence), "vkResetFences"));
mUnusedFences.pop_back();
return fence;
}
VkFenceCreateInfo createInfo;
createInfo.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO;
createInfo.pNext = nullptr;
createInfo.flags = 0;
VkFence fence = VK_NULL_HANDLE;
DAWN_TRY(CheckVkSuccess(fn.CreateFence(mVkDevice, &createInfo, nullptr, &*fence),
"vkCreateFence"));
return fence;
}
Serial Device::CheckAndUpdateCompletedSerials() {
Serial fenceSerial = 0;
while (!mFencesInFlight.empty()) {
VkFence fence = mFencesInFlight.front().first;
Serial tentativeSerial = mFencesInFlight.front().second;
VkResult result = VkResult::WrapUnsafe(
INJECT_ERROR_OR_RUN(fn.GetFenceStatus(mVkDevice, fence), VK_ERROR_DEVICE_LOST));
// TODO: Handle DeviceLost error.
ASSERT(result == VK_SUCCESS || result == VK_NOT_READY);
// Fence are added in order, so we can stop searching as soon
// as we see one that's not ready.
if (result == VK_NOT_READY) {
return fenceSerial;
}
// Update fenceSerial since fence is ready.
fenceSerial = tentativeSerial;
mUnusedFences.push_back(fence);
ASSERT(fenceSerial > GetCompletedCommandSerial());
mFencesInFlight.pop();
}
return fenceSerial;
}
MaybeError Device::PrepareRecordingContext() {
ASSERT(!mRecordingContext.used);
ASSERT(mRecordingContext.commandBuffer == VK_NULL_HANDLE);
ASSERT(mRecordingContext.commandPool == VK_NULL_HANDLE);
// First try to recycle unused command pools.
if (!mUnusedCommands.empty()) {
CommandPoolAndBuffer commands = mUnusedCommands.back();
mUnusedCommands.pop_back();
DAWN_TRY(CheckVkSuccess(fn.ResetCommandPool(mVkDevice, commands.pool, 0),
"vkResetCommandPool"));
mRecordingContext.commandBuffer = commands.commandBuffer;
mRecordingContext.commandPool = commands.pool;
} else {
// Create a new command pool for our commands and allocate the command buffer.
VkCommandPoolCreateInfo createInfo;
createInfo.sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO;
createInfo.pNext = nullptr;
createInfo.flags = VK_COMMAND_POOL_CREATE_TRANSIENT_BIT;
createInfo.queueFamilyIndex = mQueueFamily;
DAWN_TRY(CheckVkSuccess(fn.CreateCommandPool(mVkDevice, &createInfo, nullptr,
&*mRecordingContext.commandPool),
"vkCreateCommandPool"));
VkCommandBufferAllocateInfo allocateInfo;
allocateInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;
allocateInfo.pNext = nullptr;
allocateInfo.commandPool = mRecordingContext.commandPool;
allocateInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
allocateInfo.commandBufferCount = 1;
DAWN_TRY(CheckVkSuccess(fn.AllocateCommandBuffers(mVkDevice, &allocateInfo,
&mRecordingContext.commandBuffer),
"vkAllocateCommandBuffers"));
}
// Start the recording of commands in the command buffer.
VkCommandBufferBeginInfo beginInfo;
beginInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
beginInfo.pNext = nullptr;
beginInfo.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
beginInfo.pInheritanceInfo = nullptr;
return CheckVkSuccess(fn.BeginCommandBuffer(mRecordingContext.commandBuffer, &beginInfo),
"vkBeginCommandBuffer");
}
void Device::RecycleCompletedCommands() {
for (auto& commands : mCommandsInFlight.IterateUpTo(GetCompletedCommandSerial())) {
mUnusedCommands.push_back(commands);
}
mCommandsInFlight.ClearUpTo(GetCompletedCommandSerial());
}
ResultOrError<std::unique_ptr<StagingBufferBase>> Device::CreateStagingBuffer(size_t size) {
std::unique_ptr<StagingBufferBase> stagingBuffer =
std::make_unique<StagingBuffer>(size, this);
DAWN_TRY(stagingBuffer->Initialize());
return std::move(stagingBuffer);
}
MaybeError Device::CopyFromStagingToBuffer(StagingBufferBase* source,
uint64_t sourceOffset,
BufferBase* destination,
uint64_t destinationOffset,
uint64_t size) {
// It is a validation error to do a 0-sized copy in Vulkan, check it is skipped prior to
// calling this function.
ASSERT(size != 0);
// Insert memory barrier to ensure host write operations are made visible before
// copying from the staging buffer. However, this barrier can be removed (see note below).
//
// Note: Depending on the spec understanding, an explicit barrier may not be required when
// used with HOST_COHERENT as vkQueueSubmit does an implicit barrier between host and
// device. See "Availability, Visibility, and Domain Operations" in Vulkan spec for details.
// Insert pipeline barrier to ensure correct ordering with previous memory operations on the
// buffer.
CommandRecordingContext* recordingContext = GetPendingRecordingContext();
ToBackend(destination)->TransitionUsageNow(recordingContext, wgpu::BufferUsage::CopyDst);
VkBufferCopy copy;
copy.srcOffset = sourceOffset;
copy.dstOffset = destinationOffset;
copy.size = size;
this->fn.CmdCopyBuffer(recordingContext->commandBuffer,
ToBackend(source)->GetBufferHandle(),
ToBackend(destination)->GetHandle(), 1, &copy);
return {};
}
MaybeError Device::ImportExternalImage(const ExternalImageDescriptor* descriptor,
ExternalMemoryHandle memoryHandle,
VkImage image,
const std::vector<ExternalSemaphoreHandle>& waitHandles,
VkSemaphore* outSignalSemaphore,
VkDeviceMemory* outAllocation,
std::vector<VkSemaphore>* outWaitSemaphores) {
const TextureDescriptor* textureDescriptor =
reinterpret_cast<const TextureDescriptor*>(descriptor->cTextureDescriptor);
// TODO(dawn:22): Remove once migration from GPUTextureDescriptor.arrayLayerCount to
// GPUTextureDescriptor.size.depth is done.
TextureDescriptor fixedDescriptor;
DAWN_TRY_ASSIGN(fixedDescriptor, FixTextureDescriptor(this, textureDescriptor));
textureDescriptor = &fixedDescriptor;
// Check services support this combination of handle type / image info
if (!mExternalSemaphoreService->Supported()) {
return DAWN_VALIDATION_ERROR("External semaphore usage not supported");
}
if (!mExternalMemoryService->SupportsImportMemory(
VulkanImageFormat(this, textureDescriptor->format), VK_IMAGE_TYPE_2D,
VK_IMAGE_TILING_OPTIMAL,
VulkanImageUsage(textureDescriptor->usage,
GetValidInternalFormat(textureDescriptor->format)),
VK_IMAGE_CREATE_ALIAS_BIT_KHR)) {
return DAWN_VALIDATION_ERROR("External memory usage not supported");
}
// Create an external semaphore to signal when the texture is done being used
DAWN_TRY_ASSIGN(*outSignalSemaphore,
mExternalSemaphoreService->CreateExportableSemaphore());
// Import the external image's memory
external_memory::MemoryImportParams importParams;
DAWN_TRY_ASSIGN(importParams,
mExternalMemoryService->GetMemoryImportParams(descriptor, image));
DAWN_TRY_ASSIGN(*outAllocation,
mExternalMemoryService->ImportMemory(memoryHandle, importParams, image));
// Import semaphores we have to wait on before using the texture
for (const ExternalSemaphoreHandle& handle : waitHandles) {
VkSemaphore semaphore = VK_NULL_HANDLE;
DAWN_TRY_ASSIGN(semaphore, mExternalSemaphoreService->ImportSemaphore(handle));
outWaitSemaphores->push_back(semaphore);
}
return {};
}
MaybeError Device::SignalAndExportExternalTexture(Texture* texture,
ExternalSemaphoreHandle* outHandle) {
DAWN_TRY(ValidateObject(texture));
VkSemaphore outSignalSemaphore;
DAWN_TRY(texture->SignalAndDestroy(&outSignalSemaphore));
// This has to happen right after SignalAndDestroy, since the semaphore will be
// deleted when the fenced deleter runs after the queue submission
DAWN_TRY_ASSIGN(*outHandle, mExternalSemaphoreService->ExportSemaphore(outSignalSemaphore));
return {};
}
TextureBase* Device::CreateTextureWrappingVulkanImage(
const ExternalImageDescriptor* descriptor,
ExternalMemoryHandle memoryHandle,
const std::vector<ExternalSemaphoreHandle>& waitHandles) {
const TextureDescriptor* textureDescriptor =
reinterpret_cast<const TextureDescriptor*>(descriptor->cTextureDescriptor);
// Initial validation
if (ConsumedError(ValidateTextureDescriptor(this, textureDescriptor))) {
return nullptr;
}
if (ConsumedError(ValidateVulkanImageCanBeWrapped(this, textureDescriptor))) {
return nullptr;
}
VkSemaphore signalSemaphore = VK_NULL_HANDLE;
VkDeviceMemory allocation = VK_NULL_HANDLE;
std::vector<VkSemaphore> waitSemaphores;
waitSemaphores.reserve(waitHandles.size());
// Cleanup in case of a failure, the image creation doesn't acquire the external objects
// if a failure happems.
Texture* result = nullptr;
// TODO(crbug.com/1026480): Consolidate this into a single CreateFromExternal call.
if (ConsumedError(Texture::CreateFromExternal(this, descriptor, textureDescriptor,
mExternalMemoryService.get()),
&result) ||
ConsumedError(ImportExternalImage(descriptor, memoryHandle, result->GetHandle(),
waitHandles, &signalSemaphore, &allocation,
&waitSemaphores)) ||
ConsumedError(result->BindExternalMemory(descriptor, signalSemaphore, allocation,
waitSemaphores))) {
// Delete the Texture if it was created
if (result != nullptr) {
result->Release();
}
// Clear the signal semaphore
fn.DestroySemaphore(GetVkDevice(), signalSemaphore, nullptr);
// Clear image memory
fn.FreeMemory(GetVkDevice(), allocation, nullptr);
// Clear any wait semaphores we were able to import
for (VkSemaphore semaphore : waitSemaphores) {
fn.DestroySemaphore(GetVkDevice(), semaphore, nullptr);
}
return nullptr;
}
return result;
}
ResultOrError<ResourceMemoryAllocation> Device::AllocateMemory(
VkMemoryRequirements requirements,
bool mappable) {
return mResourceMemoryAllocator->Allocate(requirements, mappable);
}
void Device::DeallocateMemory(ResourceMemoryAllocation* allocation) {
mResourceMemoryAllocator->Deallocate(allocation);
}
int Device::FindBestMemoryTypeIndex(VkMemoryRequirements requirements, bool mappable) {
return mResourceMemoryAllocator->FindBestTypeIndex(requirements, mappable);
}
ResourceMemoryAllocator* Device::GetResourceMemoryAllocatorForTesting() const {
return mResourceMemoryAllocator.get();
}
uint32_t Device::GetComputeSubgroupSize() const {
return mComputeSubgroupSize;
}
MaybeError Device::WaitForIdleForDestruction() {
// Immediately tag the recording context as unused so we don't try to submit it in Tick.
// Move the mRecordingContext.used to mUnusedCommands so it can be cleaned up in
// ShutDownImpl
if (mRecordingContext.used) {
CommandPoolAndBuffer commands = {mRecordingContext.commandPool,
mRecordingContext.commandBuffer};
mUnusedCommands.push_back(commands);
mRecordingContext = CommandRecordingContext();
}
VkResult waitIdleResult = VkResult::WrapUnsafe(fn.QueueWaitIdle(mQueue));
// Ignore the result of QueueWaitIdle: it can return OOM which we can't really do anything
// about, Device lost, which means workloads running on the GPU are no longer accessible
// (so they are as good as waited on) or success.
DAWN_UNUSED(waitIdleResult);
// Make sure all fences are complete by explicitly waiting on them all
while (!mFencesInFlight.empty()) {
VkFence fence = mFencesInFlight.front().first;
Serial fenceSerial = mFencesInFlight.front().second;
ASSERT(fenceSerial > GetCompletedCommandSerial());
VkResult result = VkResult::WrapUnsafe(VK_TIMEOUT);
do {
result = VkResult::WrapUnsafe(
INJECT_ERROR_OR_RUN(fn.WaitForFences(mVkDevice, 1, &*fence, true, UINT64_MAX),
VK_ERROR_DEVICE_LOST));
} while (result == VK_TIMEOUT);
// TODO: Handle errors
ASSERT(result == VK_SUCCESS);
fn.DestroyFence(mVkDevice, fence, nullptr);
mFencesInFlight.pop();
}
return {};
}
void Device::ShutDownImpl() {
ASSERT(GetState() == State::Disconnected);
// We failed during initialization so early that we don't even have a VkDevice. There is
// nothing to do.
if (mVkDevice == VK_NULL_HANDLE) {
return;
}
// The deleter is the second thing we initialize. If it is not present, it means that
// only the VkDevice was created and nothing else. Destroy the device and do nothing else
// because the function pointers might not have been loaded (and there is nothing to
// destroy anyway).
if (mDeleter == nullptr) {
fn.DestroyDevice(mVkDevice, nullptr);
mVkDevice = VK_NULL_HANDLE;
return;
}
// Enough of the Device's initialization happened that we can now do regular robust
// deinitialization.
// Immediately tag the recording context as unused so we don't try to submit it in Tick.
mRecordingContext.used = false;
fn.DestroyCommandPool(mVkDevice, mRecordingContext.commandPool, nullptr);
for (VkSemaphore semaphore : mRecordingContext.waitSemaphores) {
fn.DestroySemaphore(mVkDevice, semaphore, nullptr);
}
mRecordingContext.waitSemaphores.clear();
for (VkSemaphore semaphore : mRecordingContext.signalSemaphores) {
fn.DestroySemaphore(mVkDevice, semaphore, nullptr);
}
mRecordingContext.signalSemaphores.clear();
ASSERT(mCommandsInFlight.Empty());
for (const CommandPoolAndBuffer& commands : mUnusedCommands) {
fn.DestroyCommandPool(mVkDevice, commands.pool, nullptr);
}
mUnusedCommands.clear();
for (VkFence fence : mUnusedFences) {
fn.DestroyFence(mVkDevice, fence, nullptr);
}
mUnusedFences.clear();
// Releasing the uploader enqueues buffers to be released.
// Call Tick() again to clear them before releasing the deleter.
mDeleter->Tick(GetCompletedCommandSerial());
// The VkRenderPasses in the cache can be destroyed immediately since all commands referring
// to them are guaranteed to be finished executing.
mRenderPassCache = nullptr;
// We need handle deleting all child objects by calling Tick() again with a large serial to
// force all operations to look as if they were completed, and delete all objects before
// destroying the Deleter and vkDevice.
ASSERT(mDeleter != nullptr);
mDeleter->Tick(std::numeric_limits<Serial>::max());
mDeleter = nullptr;
// VkQueues are destroyed when the VkDevice is destroyed
// The VkDevice is needed to destroy child objects, so it must be destroyed last after all
// child objects have been deleted.
ASSERT(mVkDevice != VK_NULL_HANDLE);
fn.DestroyDevice(mVkDevice, nullptr);
mVkDevice = VK_NULL_HANDLE;
}
}} // namespace dawn_native::vulkan
Reference in New Issue View Git Blame Copy Permalink