Set up cache in vulkan to reuse GrVkPrograms (aka VkPipelines)

src/gpu/vk/GrVkPipelineState.h

+/*
+ * Copyright 2015 Google Inc.

[bsalomon, 2016/03/23 15:04:36]

2016

[egdaniel, 2016/03/23 19:12:07]

Done.

+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+
+#ifndef GrVkPipelineState_DEFINED
+#define GrVkPipelineState_DEFINED
+
+#include "GrVkImage.h"
+#include "GrVkProgramDesc.h"
+#include "GrVkPipelineStateDataManager.h"
+#include "glsl/GrGLSLProgramBuilder.h"
+
+#include "vulkan/vulkan.h"
+
+class GrPipeline;
+class GrVkCommandBuffer;
+class GrVkDescriptorPool;
+class GrVkGpu;
+class GrVkImageView;
+class GrVkPipeline;
+class GrVkSampler;
+class GrVkUniformBuffer;
+
+class GrVkPipelineState : public SkRefCnt {

[bsalomon, 2016/03/23 15:04:37]

brief comment on what this is? 

[egdaniel, 2016/03/23 19:12:07]

Done.

+public:
+    typedef GrGLSLProgramBuilder::BuiltinUniformHandles BuiltinUniformHandles;
+
+    ~GrVkPipelineState();
+
+    GrVkPipeline* vkPipeline() const { return fPipeline; }
+
+    void setData(GrVkGpu*, const GrPrimitiveProcessor&, const GrPipeline&);
+
+    void bind(const GrVkGpu* gpu, GrVkCommandBuffer* commandBuffer);
+
+    void addUniformResources(GrVkCommandBuffer&);
+
+    void freeGPUResources(const GrVkGpu* gpu);
+
+    // This releases resources that only a given instance of a GrVkPipelineState needs to hold onto
+    // and don't need to survive across new uses of the GrVkPipelineState.
+    void freeTempResources(const GrVkGpu* gpu);
+
+    void abandonGPUResources();
+
+    // The key is composed of two parts:
+    // 1. uint32_t for total key length
+    // 2. Pipeline state data
+    enum StateKeyOffsets {
+        // Part 1.
+        kLength_StateKeyOffset = 0,
+        // Part 2.
+        kData_StateKeyOffset = kLength_StateKeyOffset + sizeof(uint32_t),
+    };
+    static void BuildStateKey(const GrPipeline&, GrPrimitiveType primitiveType,
+                               SkTArray<unsigned char, true>* key);
+
+    /**
+     * For Vulkan we want to cache the entire VkPipeline for reuse of draws. The Desc here holds all
+     * the information needed to differentiate one pipeline from another.
+     *
+     * The GrVkProgramDesc contains all the information need to create the actual shaders for the
+     * pipeline.
+     *
+     * The fStateKey is used to store all the inputs for the rest of the state stored on the
+     * pipeline. This includes stencil settings, blending information, render pass format, draw face
+     * information, and primitive type. Note that some state is set dynamically on the pipeline for
+     * each draw  and thus is not included in this descriptor. This includes the viewport, scissor,
+     * and blend constant.
+     *
+     * A checksum which includes the fProgramDesc and fStateKey is included at the top of the Desc
+     * for caching purposes and faster equality checks.
+     */
+    struct Desc {
+        uint32_t                fChecksum;
+        GrVkProgramDesc         fProgramDesc;
+
+        enum {
+            kRenderPassKeyAlloc = 12, // This is typical color attachment with no stencil or msaa
+            kStencilKeyAlloc = sizeof(GrStencilSettings),
+            kDrawFaceKeyAlloc = 4,
+            kBlendingKeyAlloc = 4,
+            kPrimitiveTypeKeyAlloc = 4,
+            kPreAllocSize = kData_StateKeyOffset + kRenderPassKeyAlloc + kStencilKeyAlloc +
+                            kDrawFaceKeyAlloc + kBlendingKeyAlloc + kPrimitiveTypeKeyAlloc,
+        };
+        SkSTArray<kPreAllocSize, uint8_t, true> fStateKey;
+
+        bool operator== (const Desc& that) const {
+            if (fChecksum != that.fChecksum || fProgramDesc != that.fProgramDesc) {
+                return false;
+            }
+            // We store the keyLength at the start of fVkKey. Thus we don't have to worry about
+            // different length keys since we will fail on the comparison immediately. Therefore we
+            // just use this PipelineDesc to get the length to iterate over.
+            int keyLength = fStateKey.count();
+            SkASSERT(SkIsAlign4(keyLength));
+            int l = keyLength >> 2;
+            const uint32_t* aKey = reinterpret_cast<const uint32_t*>(fStateKey.begin());
+            const uint32_t* bKey = reinterpret_cast<const uint32_t*>(that.fStateKey.begin());
+            for (int i = 0; i < l; ++i) {
+                if (aKey[i] != bKey[i]) {
+                    return false;
+                }
+            }
+            return true;
+        }
+
+        static bool Less(const Desc& a, const Desc& b) {
+            if (a.fChecksum != b.fChecksum) {
+                return a.fChecksum < b.fChecksum ? true : false;
+            }
+            bool progDescLess = GrProgramDesc::Less(a.fProgramDesc, b.fProgramDesc);
+            if (progDescLess || a.fProgramDesc != b.fProgramDesc) {
+                return progDescLess;
+            }
+
+            int keyLength = a.fStateKey.count();
+            SkASSERT(SkIsAlign4(keyLength));
+            int l = keyLength >> 2;
+            const uint32_t* aKey = reinterpret_cast<const uint32_t*>(a.fStateKey.begin());
+            const uint32_t* bKey = reinterpret_cast<const uint32_t*>(b.fStateKey.begin());
+            for (int i = 0; i < l; ++i) {
+                if (aKey[i] != bKey[i]) {
+                    return aKey[i] < bKey[i] ? true : false;
+                }
+            }
+            return false;
+        }
+    };
+
+    const Desc& getDesc() { return fDesc; }
+
+private:
+    typedef GrVkPipelineStateDataManager::UniformInfoArray UniformInfoArray;
+    typedef GrGLSLProgramDataManager::UniformHandle UniformHandle;
+
+    GrVkPipelineState(GrVkGpu* gpu,
+                      const GrVkPipelineState::Desc&,
+                      GrVkPipeline* pipeline,
+                      VkPipelineLayout layout,
+                      VkDescriptorSetLayout dsLayout[2],
+                      const BuiltinUniformHandles& builtinUniformHandles,
+                      const UniformInfoArray& uniforms,
+                      uint32_t vertexUniformSize,
+                      uint32_t fragmentUniformSize,
+                      uint32_t numSamplers,
+                      GrGLSLPrimitiveProcessor* geometryProcessor,
+                      GrGLSLXferProcessor* xferProcessor,
+                      const GrGLSLFragProcs& fragmentProcessors);
+
+    // Each pool will manage one type of descriptor. Thus each descriptor set we use will all be of
+    // one VkDescriptorType.
+    struct DescriptorPoolManager {
+        DescriptorPoolManager(VkDescriptorSetLayout layout, VkDescriptorType type,
+                              uint32_t descCount, GrVkGpu* gpu)
+            : fDescLayout(layout)
+            , fDescType(type)
+            , fCurrentDescriptorSet(0)
+            , fPool(nullptr) {
+            SkASSERT(descCount < (SK_MaxU32 >> 2));
+            fMaxDescriptorSets = descCount << 2;
+            this->getNewPool(gpu);
+        }
+
+        ~DescriptorPoolManager() {
+            SkASSERT(!fDescLayout);
+            SkASSERT(!fPool);
+        }
+
+        void getNewDescriptorSet(GrVkGpu* gpu, VkDescriptorSet* ds);
+
+        void freeGPUResources(const GrVkGpu* gpu);
+        void abandonGPUResources();
+
+        VkDescriptorSetLayout  fDescLayout;
+        VkDescriptorType       fDescType;
+        uint32_t               fMaxDescriptorSets;
+        uint32_t               fCurrentDescriptorSet;
+        GrVkDescriptorPool*    fPool;
+
+    private:
+        void getNewPool(GrVkGpu* gpu);
+    };
+
+    void writeUniformBuffers(const GrVkGpu* gpu);
+
+    void writeSamplers(GrVkGpu* gpu, const SkTArray<const GrTextureAccess*>& textureBindings);
+
+

[bsalomon, 2016/03/23 15:04:37]

rm \n?

[egdaniel, 2016/03/23 19:12:07]

Done.

+    /**
+    * We use the RT's size and origin to adjust from Skia device space to OpenGL normalized device

[bsalomon, 2016/03/23 15:04:37]

OpenGL?

[egdaniel, 2016/03/23 19:12:07]

Done.

+    * space and to make device space positions have the correct origin for processors that require
+    * them.
+    */
+    struct RenderTargetState {
+        SkISize         fRenderTargetSize;
+        GrSurfaceOrigin fRenderTargetOrigin;
+
+        RenderTargetState() { this->invalidate(); }
+        void invalidate() {
+            fRenderTargetSize.fWidth = -1;
+            fRenderTargetSize.fHeight = -1;
+            fRenderTargetOrigin = (GrSurfaceOrigin)-1;
+        }
+
+        /**
+        * Gets a vec4 that adjusts the position from Skia device coords to GL's normalized device

[bsalomon, 2016/03/23 15:04:36]

GL

[egdaniel, 2016/03/23 19:12:06]

Done.

+        * coords. Assuming the transformed position, pos, is a homogeneous vec3, the vec, v, is
+        * applied as such:
+        * pos.x = dot(v.xy, pos.xz)
+        * pos.y = dot(v.zw, pos.yz)
+        */
+        void getRTAdjustmentVec(float* destVec) {
+            destVec[0] = 2.f / fRenderTargetSize.fWidth;
+            destVec[1] = -1.f;
+            if (kBottomLeft_GrSurfaceOrigin == fRenderTargetOrigin) {
+                destVec[2] = -2.f / fRenderTargetSize.fHeight;
+                destVec[3] = 1.f;
+            } else {
+                destVec[2] = 2.f / fRenderTargetSize.fHeight;
+                destVec[3] = -1.f;
+            }
+        }
+    };
+
+    // Helper for setData() that sets the view matrix and loads the render target height uniform
+    void setRenderTargetState(const GrPipeline&);
+
+    // GrVkResources
+    GrVkPipeline*       fPipeline;
+
+    // Used for binding DescriptorSets to the command buffer but does not need to survive during
+    // command buffer execution. Thus this is not need to be a GrVkResource.
+    VkPipelineLayout fPipelineLayout;
+
+    // The DescriptorSets need to survive until the gpu has finished all draws that use them.
+    // However, they will only be freed by the descriptor pool. Thus by simply keeping the
+    // descriptor pool alive through the draw, the descritor sets will also stay alive. Thus we do
+    // not need a GrVkResource versions of VkDescriptorSet. We hold on to these in the
+    // GrVkPipelineState since we update the descriptor sets and bind them at separate times;
+    VkDescriptorSet       fDescriptorSets[2];
+
+    // Meta data so we know which descriptor sets we are using and need to bind.
+    int                   fStartDS;
+    int                   fDSCount;
+
+    SkAutoTDelete<GrVkUniformBuffer>    fVertexUniformBuffer;
+    SkAutoTDelete<GrVkUniformBuffer>    fFragmentUniformBuffer;
+
+    // GrVkResources used for sampling textures
+    SkTDArray<GrVkSampler*>                fSamplers;
+    SkTDArray<const GrVkImageView*>        fTextureViews;
+    SkTDArray<const GrVkImage::Resource*>  fTextures;
+
+    // Tracks the current render target uniforms stored in the vertex buffer.
+    RenderTargetState fRenderTargetState;
+    BuiltinUniformHandles fBuiltinUniformHandles;
+
+    // Processors in the GrVkPipelineState
+    SkAutoTDelete<GrGLSLPrimitiveProcessor> fGeometryProcessor;
+    SkAutoTDelete<GrGLSLXferProcessor>      fXferProcessor;
+    GrGLSLFragProcs                         fFragmentProcessors;
+
+    Desc                         fDesc;
+
+    GrVkPipelineStateDataManager fDataManager;
+
+    DescriptorPoolManager        fSamplerPoolManager;
+    DescriptorPoolManager        fUniformPoolManager;
+

[bsalomon, 2016/03/23 15:04:37]

Inconsistent alignment in these members... I'd say either give up on aligning or
align all of them.

+    int fNumSamplers;
+
+    friend class GrVkPipelineStateBuilder;
+};
+
+#endif