Hash :
dae3c851
Author :
Date :
2025-03-14T11:44:53
Vulkan: Bake non-shader state into linked pipeline When using VK_EXT_graphics_pipeline_library, previously ANGLE would create three pipelines libraries: * The Shaders library was created based on the GL program's shaders + a few static states. This typically hit the program's own pipeline's cache that was warmed up during link. * The VertexInput and FragmentOutput libraries were created at draw time, which used the global pipeline cache At draw time, immediately after creating the non-Shaders libraries, the three libraries were linked into the final pipeline to be used by the draw call. This caused an inefficiency; because the non-Shaders libraries were created independently from the Shaders library, they had to be compiled pessimistically, for example because they could not be optimized to take into account the precision of the fragment shader's outputs or whether any value is const (typically alpha being set to one). Given the creation of VertexInput and FragmentOutput libraries is typically quite fast (the former being no-op and dynamic state anyway), this change removes the need for creating those libraries, and directly specifies the vertex input and fragment output state when creating the final pipeline out of the Shaders library. In this way, the same fragment output state can be tailored to the exact shaders it is being used with and incur a smaller overhead. In this change, the linked pipeline is cached in the GL program's pipeline cache, which is never synced to the blob cache as producing it is assumed to be fast already. Bug: angleproject:42265839 Change-Id: I8496ea37771555522bdc9de94043a1b56fa5967e Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/6354205 Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org> Reviewed-by: mohan maiya <m.maiya@samsung.com> Reviewed-by: Charlie Lao <cclao@google.com>
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158
//
// Copyright 2023 The ANGLE Project Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
//
// ShareGroupVk.h:
// Defines the class interface for ShareGroupVk, implementing ShareGroupImpl.
//
#ifndef LIBANGLE_RENDERER_VULKAN_SHAREGROUPVK_H_
#define LIBANGLE_RENDERER_VULKAN_SHAREGROUPVK_H_
#include "libANGLE/renderer/ShareGroupImpl.h"
#include "libANGLE/renderer/vulkan/vk_cache_utils.h"
#include "libANGLE/renderer/vulkan/vk_helpers.h"
#include "libANGLE/renderer/vulkan/vk_resource.h"
#include "libANGLE/renderer/vulkan/vk_utils.h"
namespace rx
{
constexpr VkDeviceSize kMaxTotalEmptyBufferBytes = 16 * 1024 * 1024;
class TextureUpload
{
public:
TextureUpload() { mPrevUploadedMutableTexture = nullptr; }
~TextureUpload() { resetPrevTexture(); }
angle::Result onMutableTextureUpload(ContextVk *contextVk, TextureVk *newTexture);
void onTextureRelease(TextureVk *textureVk);
void resetPrevTexture() { mPrevUploadedMutableTexture = nullptr; }
private:
// Keep track of the previously stored texture. Used to flush mutable textures.
TextureVk *mPrevUploadedMutableTexture;
};
class ShareGroupVk : public ShareGroupImpl
{
public:
ShareGroupVk(const egl::ShareGroupState &state, vk::Renderer *renderer);
void onDestroy(const egl::Display *display) override;
void onContextAdd() override;
FramebufferCache &getFramebufferCache() { return mFramebufferCache; }
bool hasAnyContextWithRobustness() const { return mState.hasAnyContextWithRobustness(); }
// PipelineLayoutCache and DescriptorSetLayoutCache can be shared between multiple threads
// accessing them via shared contexts. The ShareGroup locks around gl entrypoints ensuring
// synchronous update to the caches.
PipelineLayoutCache &getPipelineLayoutCache() { return mPipelineLayoutCache; }
DescriptorSetLayoutCache &getDescriptorSetLayoutCache() { return mDescriptorSetLayoutCache; }
const egl::ContextMap &getContexts() const { return mState.getContexts(); }
vk::DescriptorSetArray<vk::MetaDescriptorPool> &getMetaDescriptorPools()
{
return mMetaDescriptorPools;
}
// Used to flush the mutable textures more often.
angle::Result onMutableTextureUpload(ContextVk *contextVk, TextureVk *newTexture);
vk::BufferPool *getDefaultBufferPool(VkDeviceSize size,
uint32_t memoryTypeIndex,
BufferUsageType usageType);
void pruneDefaultBufferPools();
void calculateTotalBufferCount(size_t *bufferCount, VkDeviceSize *totalSize) const;
void logBufferPools() const;
// Temporary workaround until VkSemaphore(s) will be used between different priorities.
angle::Result unifyContextsPriority(ContextVk *newContextVk);
// Temporary workaround until VkSemaphore(s) will be used between different priorities.
angle::Result lockDefaultContextsPriority(ContextVk *contextVk);
UpdateDescriptorSetsBuilder *getUpdateDescriptorSetsBuilder()
{
return &mUpdateDescriptorSetsBuilder;
}
void onTextureRelease(TextureVk *textureVk);
angle::Result scheduleMonolithicPipelineCreationTask(
ContextVk *contextVk,
vk::WaitableMonolithicPipelineCreationTask *taskOut);
void waitForCurrentMonolithicPipelineCreationTask();
vk::RefCountedEventsGarbageRecycler *getRefCountedEventsGarbageRecycler()
{
return &mRefCountedEventsGarbageRecycler;
}
void cleanupRefCountedEventGarbage() { mRefCountedEventsGarbageRecycler.cleanup(mRenderer); }
void cleanupExcessiveRefCountedEventGarbage()
{
// TODO: b/336844257 needs tune.
constexpr size_t kExcessiveGarbageCountThreshold = 256;
if (mRefCountedEventsGarbageRecycler.getGarbageCount() > kExcessiveGarbageCountThreshold)
{
mRefCountedEventsGarbageRecycler.cleanup(mRenderer);
}
}
void onFramebufferBoundary();
uint32_t getCurrentFrameCount() const { return mCurrentFrameCount; }
private:
angle::Result updateContextsPriority(ContextVk *contextVk, egl::ContextPriority newPriority);
bool isDueForBufferPoolPrune();
vk::Renderer *mRenderer;
// Tracks the total number of frames rendered.
uint32_t mCurrentFrameCount;
// VkFramebuffer caches
FramebufferCache mFramebufferCache;
void resetPrevTexture() { mTextureUpload.resetPrevTexture(); }
// ANGLE uses a PipelineLayout cache to store compatible pipeline layouts.
PipelineLayoutCache mPipelineLayoutCache;
// DescriptorSetLayouts are also managed in a cache.
DescriptorSetLayoutCache mDescriptorSetLayoutCache;
// Descriptor set caches
vk::DescriptorSetArray<vk::MetaDescriptorPool> mMetaDescriptorPools;
// Priority of all Contexts in the context set
egl::ContextPriority mContextsPriority;
bool mIsContextsPriorityLocked;
// Storage for vkUpdateDescriptorSets
UpdateDescriptorSetsBuilder mUpdateDescriptorSetsBuilder;
// The per shared group buffer pools that all buffers should sub-allocate from.
vk::BufferPoolPointerArray mDefaultBufferPools;
// The system time when last pruneEmptyBuffer gets called.
double mLastPruneTime;
// The system time when the last monolithic pipeline creation job was launched. This is
// rate-limited to avoid hogging all cores and interfering with the application threads. A
// single pipeline creation job is currently supported.
double mLastMonolithicPipelineJobTime;
std::shared_ptr<angle::WaitableEvent> mMonolithicPipelineCreationEvent;
// Texture update manager used to flush uploaded mutable textures.
TextureUpload mTextureUpload;
// Holds RefCountedEvent that are free and ready to reuse
vk::RefCountedEventsGarbageRecycler mRefCountedEventsGarbageRecycler;
};
} // namespace rx
#endif // LIBANGLE_RENDERER_VULKAN_SHAREGROUPVK_H_