src/gpu/vk/GrVkResourceProvider.cpp

/*
* Copyright 2016 Google Inc.
*
* Use of this source code is governed by a BSD-style license that can be
* found in the LICENSE file.
*/

#include "src/gpu/vk/GrVkResourceProvider.h"

#include "src/core/SkTaskGroup.h"
#include "src/gpu/GrContextPriv.h"
#include "src/gpu/GrSamplerState.h"
#include "src/gpu/vk/GrVkCommandBuffer.h"
#include "src/gpu/vk/GrVkCommandPool.h"
#include "src/gpu/vk/GrVkGpu.h"
#include "src/gpu/vk/GrVkPipeline.h"
#include "src/gpu/vk/GrVkRenderTarget.h"
#include "src/gpu/vk/GrVkUniformBuffer.h"
#include "src/gpu/vk/GrVkUtil.h"

#ifdef SK_TRACE_VK_RESOURCES
std::atomic<uint32_t> GrVkResource::fKeyCounter{0};
#endif

GrVkResourceProvider::GrVkResourceProvider(GrVkGpu* gpu)
    : fGpu(gpu)
    , fPipelineCache(VK_NULL_HANDLE) {
    fPipelineStateCache = new PipelineStateCache(gpu);
}

GrVkResourceProvider::~GrVkResourceProvider() {
    SkASSERT(0 == fRenderPassArray.count());
    SkASSERT(0 == fExternalRenderPasses.count());
    SkASSERT(VK_NULL_HANDLE == fPipelineCache);
    delete fPipelineStateCache;
}

VkPipelineCache GrVkResourceProvider::pipelineCache() {
    if (fPipelineCache == VK_NULL_HANDLE) {
        VkPipelineCacheCreateInfo createInfo;
        memset(&createInfo, 0, sizeof(VkPipelineCacheCreateInfo));
        createInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO;
        createInfo.pNext = nullptr;
        createInfo.flags = 0;

        auto persistentCache = fGpu->getContext()->priv().getPersistentCache();
        sk_sp<SkData> cached;
        if (persistentCache) {
            uint32_t key = GrVkGpu::kPipelineCache_PersistentCacheKeyType;
            sk_sp<SkData> keyData = SkData::MakeWithoutCopy(&key, sizeof(uint32_t));
            cached = persistentCache->load(*keyData);
        }
        bool usedCached = false;
        if (cached) {
            uint32_t* cacheHeader = (uint32_t*)cached->data();
            if (cacheHeader[1] == VK_PIPELINE_CACHE_HEADER_VERSION_ONE) {
                // For version one of the header, the total header size is 16 bytes plus
                // VK_UUID_SIZE bytes. See Section 9.6 (Pipeline Cache) in the vulkan spec to see
                // the breakdown of these bytes.
                SkASSERT(cacheHeader[0] == 16 + VK_UUID_SIZE);
                const VkPhysicalDeviceProperties& devProps = fGpu->physicalDeviceProperties();
                const uint8_t* supportedPipelineCacheUUID = devProps.pipelineCacheUUID;
                if (cacheHeader[2] == devProps.vendorID && cacheHeader[3] == devProps.deviceID &&
                    !memcmp(&cacheHeader[4], supportedPipelineCacheUUID, VK_UUID_SIZE)) {
                    createInfo.initialDataSize = cached->size();
                    createInfo.pInitialData = cached->data();
                    usedCached = true;
                }
            }
        }
        if (!usedCached) {
            createInfo.initialDataSize = 0;
            createInfo.pInitialData = nullptr;
        }

        VkResult result;
        GR_VK_CALL_RESULT(fGpu, result, CreatePipelineCache(fGpu->device(), &createInfo, nullptr,
                                                            &fPipelineCache));
        if (VK_SUCCESS != result) {
            fPipelineCache = VK_NULL_HANDLE;
        }
    }
    return fPipelineCache;
}

void GrVkResourceProvider::init() {
    // Init uniform descriptor objects
    GrVkDescriptorSetManager* dsm = GrVkDescriptorSetManager::CreateUniformManager(fGpu);
    fDescriptorSetManagers.emplace_back(dsm);
    SkASSERT(1 == fDescriptorSetManagers.count());
    fUniformDSHandle = GrVkDescriptorSetManager::Handle(0);
}

GrVkPipeline* GrVkResourceProvider::createPipeline(const GrProgramInfo& programInfo,
                                                   VkPipelineShaderStageCreateInfo* shaderStageInfo,
                                                   int shaderStageCount,
                                                   VkRenderPass compatibleRenderPass,
                                                   VkPipelineLayout layout) {
    return GrVkPipeline::Create(fGpu, programInfo, shaderStageInfo,
                                shaderStageCount, compatibleRenderPass, layout,
                                this->pipelineCache());
}

// To create framebuffers, we first need to create a simple RenderPass that is
// only used for framebuffer creation. When we actually render we will create
// RenderPasses as needed that are compatible with the framebuffer.
const GrVkRenderPass*
GrVkResourceProvider::findCompatibleRenderPass(const GrVkRenderTarget& target,
                                               CompatibleRPHandle* compatibleHandle) {
    for (int i = 0; i < fRenderPassArray.count(); ++i) {
        if (fRenderPassArray[i].isCompatible(target)) {
            const GrVkRenderPass* renderPass = fRenderPassArray[i].getCompatibleRenderPass();
            renderPass->ref();
            if (compatibleHandle) {
                *compatibleHandle = CompatibleRPHandle(i);
            }
            return renderPass;
        }
    }

    GrVkRenderPass* renderPass = GrVkRenderPass::CreateSimple(fGpu, target);
    if (!renderPass) {
        return nullptr;
    }
    fRenderPassArray.emplace_back(renderPass);

    if (compatibleHandle) {
        *compatibleHandle = CompatibleRPHandle(fRenderPassArray.count() - 1);
    }
    return renderPass;
}

const GrVkRenderPass*
GrVkResourceProvider::findCompatibleRenderPass(const CompatibleRPHandle& compatibleHandle) {
    SkASSERT(compatibleHandle.isValid() && compatibleHandle.toIndex() < fRenderPassArray.count());
    int index = compatibleHandle.toIndex();
    const GrVkRenderPass* renderPass = fRenderPassArray[index].getCompatibleRenderPass();
    SkASSERT(renderPass);
    renderPass->ref();
    return renderPass;
}

const GrVkRenderPass* GrVkResourceProvider::findCompatibleExternalRenderPass(
        VkRenderPass renderPass, uint32_t colorAttachmentIndex) {
    for (int i = 0; i < fExternalRenderPasses.count(); ++i) {
        if (fExternalRenderPasses[i]->isCompatibleExternalRP(renderPass)) {
            fExternalRenderPasses[i]->ref();
#ifdef SK_DEBUG
            uint32_t cachedColorIndex;
            SkASSERT(fExternalRenderPasses[i]->colorAttachmentIndex(&cachedColorIndex));
            SkASSERT(cachedColorIndex == colorAttachmentIndex);
#endif
            return fExternalRenderPasses[i];
        }
    }

    const GrVkRenderPass* newRenderPass = new GrVkRenderPass(renderPass, colorAttachmentIndex);
    fExternalRenderPasses.push_back(newRenderPass);
    newRenderPass->ref();
    return newRenderPass;
}

const GrVkRenderPass* GrVkResourceProvider::findRenderPass(
                                                     GrVkRenderTarget* target,
                                                     const GrVkRenderPass::LoadStoreOps& colorOps,
                                                     const GrVkRenderPass::LoadStoreOps& stencilOps,
                                                     CompatibleRPHandle* compatibleHandle) {
    GrVkResourceProvider::CompatibleRPHandle tempRPHandle;
    GrVkResourceProvider::CompatibleRPHandle* pRPHandle = compatibleHandle ? compatibleHandle
                                                                           : &tempRPHandle;
    *pRPHandle = target->compatibleRenderPassHandle();
    if (!pRPHandle->isValid()) {
        return nullptr;
    }

    return this->findRenderPass(*pRPHandle, colorOps, stencilOps);
}

const GrVkRenderPass*
GrVkResourceProvider::findRenderPass(const CompatibleRPHandle& compatibleHandle,
                                     const GrVkRenderPass::LoadStoreOps& colorOps,
                                     const GrVkRenderPass::LoadStoreOps& stencilOps) {
    SkASSERT(compatibleHandle.isValid() && compatibleHandle.toIndex() < fRenderPassArray.count());
    CompatibleRenderPassSet& compatibleSet = fRenderPassArray[compatibleHandle.toIndex()];
    const GrVkRenderPass* renderPass = compatibleSet.getRenderPass(fGpu,
                                                                   colorOps,
                                                                   stencilOps);
    if (!renderPass) {
        return nullptr;
    }
    renderPass->ref();
    return renderPass;
}

GrVkDescriptorPool* GrVkResourceProvider::findOrCreateCompatibleDescriptorPool(
                                                            VkDescriptorType type, uint32_t count) {
    return GrVkDescriptorPool::Create(fGpu, type, count);
}

GrVkSampler* GrVkResourceProvider::findOrCreateCompatibleSampler(
        GrSamplerState params, const GrVkYcbcrConversionInfo& ycbcrInfo) {
    GrVkSampler* sampler = fSamplers.find(GrVkSampler::GenerateKey(params, ycbcrInfo));
    if (!sampler) {
        sampler = GrVkSampler::Create(fGpu, params, ycbcrInfo);
        if (!sampler) {
            return nullptr;
        }
        fSamplers.add(sampler);
    }
    SkASSERT(sampler);
    sampler->ref();
    return sampler;
}

GrVkSamplerYcbcrConversion* GrVkResourceProvider::findOrCreateCompatibleSamplerYcbcrConversion(
        const GrVkYcbcrConversionInfo& ycbcrInfo) {
    GrVkSamplerYcbcrConversion* ycbcrConversion =
            fYcbcrConversions.find(GrVkSamplerYcbcrConversion::GenerateKey(ycbcrInfo));
    if (!ycbcrConversion) {
        ycbcrConversion = GrVkSamplerYcbcrConversion::Create(fGpu, ycbcrInfo);
        if (!ycbcrConversion) {
            return nullptr;
        }
        fYcbcrConversions.add(ycbcrConversion);
    }
    SkASSERT(ycbcrConversion);
    ycbcrConversion->ref();
    return ycbcrConversion;
}

GrVkPipelineState* GrVkResourceProvider::findOrCreateCompatiblePipelineState(
        GrRenderTarget* renderTarget,
        const GrProgramInfo& programInfo,
        VkRenderPass compatibleRenderPass) {
    return fPipelineStateCache->findOrCreatePipelineState(renderTarget, programInfo,
                                                          compatibleRenderPass);
}

void GrVkResourceProvider::getSamplerDescriptorSetHandle(VkDescriptorType type,
                                                         const GrVkUniformHandler& uniformHandler,
                                                         GrVkDescriptorSetManager::Handle* handle) {
    SkASSERT(handle);
    SkASSERT(VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER == type ||
             VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER == type);
    for (int i = 0; i < fDescriptorSetManagers.count(); ++i) {
        if (fDescriptorSetManagers[i]->isCompatible(type, &uniformHandler)) {
           *handle = GrVkDescriptorSetManager::Handle(i);
           return;
        }
    }

    GrVkDescriptorSetManager* dsm = GrVkDescriptorSetManager::CreateSamplerManager(fGpu, type,
                                                                                   uniformHandler);
    fDescriptorSetManagers.emplace_back(dsm);
    *handle = GrVkDescriptorSetManager::Handle(fDescriptorSetManagers.count() - 1);
}

void GrVkResourceProvider::getSamplerDescriptorSetHandle(VkDescriptorType type,
                                                         const SkTArray<uint32_t>& visibilities,
                                                         GrVkDescriptorSetManager::Handle* handle) {
    SkASSERT(handle);
    SkASSERT(VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER == type ||
             VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER == type);
    for (int i = 0; i < fDescriptorSetManagers.count(); ++i) {
        if (fDescriptorSetManagers[i]->isCompatible(type, visibilities)) {
            *handle = GrVkDescriptorSetManager::Handle(i);
            return;
        }
    }

    GrVkDescriptorSetManager* dsm = GrVkDescriptorSetManager::CreateSamplerManager(fGpu, type,
                                                                                   visibilities);
    fDescriptorSetManagers.emplace_back(dsm);
    *handle = GrVkDescriptorSetManager::Handle(fDescriptorSetManagers.count() - 1);
}

VkDescriptorSetLayout GrVkResourceProvider::getUniformDSLayout() const {
    SkASSERT(fUniformDSHandle.isValid());
    return fDescriptorSetManagers[fUniformDSHandle.toIndex()]->layout();
}

VkDescriptorSetLayout GrVkResourceProvider::getSamplerDSLayout(
        const GrVkDescriptorSetManager::Handle& handle) const {
    SkASSERT(handle.isValid());
    return fDescriptorSetManagers[handle.toIndex()]->layout();
}

const GrVkDescriptorSet* GrVkResourceProvider::getUniformDescriptorSet() {
    SkASSERT(fUniformDSHandle.isValid());
    return fDescriptorSetManagers[fUniformDSHandle.toIndex()]->getDescriptorSet(fGpu,
                                                                                fUniformDSHandle);
}

const GrVkDescriptorSet* GrVkResourceProvider::getSamplerDescriptorSet(
        const GrVkDescriptorSetManager::Handle& handle) {
    SkASSERT(handle.isValid());
    return fDescriptorSetManagers[handle.toIndex()]->getDescriptorSet(fGpu, handle);
}

void GrVkResourceProvider::recycleDescriptorSet(const GrVkDescriptorSet* descSet,
                                                const GrVkDescriptorSetManager::Handle& handle) {
    SkASSERT(descSet);
    SkASSERT(handle.isValid());
    int managerIdx = handle.toIndex();
    SkASSERT(managerIdx < fDescriptorSetManagers.count());
    fDescriptorSetManagers[managerIdx]->recycleDescriptorSet(descSet);
}

GrVkCommandPool* GrVkResourceProvider::findOrCreateCommandPool() {
    std::unique_lock<std::recursive_mutex> lock(fBackgroundMutex);
    GrVkCommandPool* result;
    if (fAvailableCommandPools.count()) {
        result = fAvailableCommandPools.back();
        fAvailableCommandPools.pop_back();
    } else {
        result = GrVkCommandPool::Create(fGpu);
        if (!result) {
            return nullptr;
        }
    }
    SkASSERT(result->unique());
    SkDEBUGCODE(
        for (const GrVkCommandPool* pool : fActiveCommandPools) {
            SkASSERT(pool != result);
        }
        for (const GrVkCommandPool* pool : fAvailableCommandPools) {
            SkASSERT(pool != result);
        }
    )
    fActiveCommandPools.push_back(result);
    result->ref();
    return result;
}

void GrVkResourceProvider::checkCommandBuffers() {
    for (int i = fActiveCommandPools.count() - 1; i >= 0; --i) {
        GrVkCommandPool* pool = fActiveCommandPools[i];
        if (!pool->isOpen()) {
            GrVkPrimaryCommandBuffer* buffer = pool->getPrimaryCommandBuffer();
            if (buffer->finished(fGpu)) {
                fActiveCommandPools.removeShuffle(i);
                this->backgroundReset(pool);
            }
        }
    }
}

void GrVkResourceProvider::addFinishedProcToActiveCommandBuffers(
        GrGpuFinishedProc finishedProc, GrGpuFinishedContext finishedContext) {
    sk_sp<GrRefCntedCallback> procRef(new GrRefCntedCallback(finishedProc, finishedContext));
    for (int i = 0; i < fActiveCommandPools.count(); ++i) {
        GrVkCommandPool* pool = fActiveCommandPools[i];
        if (!pool->isOpen()) {
            GrVkPrimaryCommandBuffer* buffer = pool->getPrimaryCommandBuffer();
            buffer->addFinishedProc(procRef);
        }
    }
}

const GrVkResource* GrVkResourceProvider::findOrCreateStandardUniformBufferResource() {
    const GrVkResource* resource = nullptr;
    int count = fAvailableUniformBufferResources.count();
    if (count > 0) {
        resource = fAvailableUniformBufferResources[count - 1];
        fAvailableUniformBufferResources.removeShuffle(count - 1);
    } else {
        resource = GrVkUniformBuffer::CreateResource(fGpu, GrVkUniformBuffer::kStandardSize);
    }
    return resource;
}

void GrVkResourceProvider::recycleStandardUniformBufferResource(const GrVkResource* resource) {
    fAvailableUniformBufferResources.push_back(resource);
}

void GrVkResourceProvider::destroyResources(bool deviceLost) {
    SkTaskGroup* taskGroup = fGpu->getContext()->priv().getTaskGroup();
    if (taskGroup) {
        taskGroup->wait();
    }

    // loop over all render pass sets to make sure we destroy all the internal VkRenderPasses
    for (int i = 0; i < fRenderPassArray.count(); ++i) {
        fRenderPassArray[i].releaseResources(fGpu);
    }
    fRenderPassArray.reset();

    for (int i = 0; i < fExternalRenderPasses.count(); ++i) {
        fExternalRenderPasses[i]->unref(fGpu);
    }
    fExternalRenderPasses.reset();

    // Iterate through all store GrVkSamplers and unref them before resetting the hash.
    for (decltype(fSamplers)::Iter iter(&fSamplers); !iter.done(); ++iter) {
        (*iter).unref(fGpu);
    }
    fSamplers.reset();

    for (decltype(fYcbcrConversions)::Iter iter(&fYcbcrConversions); !iter.done(); ++iter) {
        (*iter).unref(fGpu);
    }
    fYcbcrConversions.reset();

    fPipelineStateCache->release();

    GR_VK_CALL(fGpu->vkInterface(), DestroyPipelineCache(fGpu->device(), fPipelineCache, nullptr));
    fPipelineCache = VK_NULL_HANDLE;

    for (GrVkCommandPool* pool : fActiveCommandPools) {
        SkASSERT(pool->unique());
        pool->unref(fGpu);
    }
    fActiveCommandPools.reset();

    for (GrVkCommandPool* pool : fAvailableCommandPools) {
        SkASSERT(pool->unique());
        pool->unref(fGpu);
    }
    fAvailableCommandPools.reset();

    // We must release/destroy all command buffers and pipeline states before releasing the
    // GrVkDescriptorSetManagers
    for (int i = 0; i < fDescriptorSetManagers.count(); ++i) {
        fDescriptorSetManagers[i]->release(fGpu);
    }
    fDescriptorSetManagers.reset();

    // release our uniform buffers
    for (int i = 0; i < fAvailableUniformBufferResources.count(); ++i) {
        SkASSERT(fAvailableUniformBufferResources[i]->unique());
        fAvailableUniformBufferResources[i]->unref(fGpu);
    }
    fAvailableUniformBufferResources.reset();
}

void GrVkResourceProvider::backgroundReset(GrVkCommandPool* pool) {
    TRACE_EVENT0("skia.gpu", TRACE_FUNC);
    SkASSERT(pool->unique());
    pool->releaseResources(fGpu);
    SkTaskGroup* taskGroup = fGpu->getContext()->priv().getTaskGroup();
    if (taskGroup) {
        taskGroup->add([this, pool]() {
            this->reset(pool);
        });
    } else {
        this->reset(pool);
    }
}

void GrVkResourceProvider::reset(GrVkCommandPool* pool) {
    TRACE_EVENT0("skia.gpu", TRACE_FUNC);
    SkASSERT(pool->unique());
    pool->reset(fGpu);
    std::unique_lock<std::recursive_mutex> providerLock(fBackgroundMutex);
    fAvailableCommandPools.push_back(pool);
}

void GrVkResourceProvider::storePipelineCacheData() {
    if (this->pipelineCache() == VK_NULL_HANDLE) {
        return;
    }
    size_t dataSize = 0;
    VkResult result;
    GR_VK_CALL_RESULT(fGpu, result, GetPipelineCacheData(fGpu->device(), this->pipelineCache(),
                                                         &dataSize, nullptr));
    if (result != VK_SUCCESS) {
        return;
    }

    std::unique_ptr<uint8_t[]> data(new uint8_t[dataSize]);

    GR_VK_CALL_RESULT(fGpu, result, GetPipelineCacheData(fGpu->device(), this->pipelineCache(),
                                                         &dataSize, (void*)data.get()));
    if (result != VK_SUCCESS) {
        return;
    }

    uint32_t key = GrVkGpu::kPipelineCache_PersistentCacheKeyType;
    sk_sp<SkData> keyData = SkData::MakeWithoutCopy(&key, sizeof(uint32_t));

    fGpu->getContext()->priv().getPersistentCache()->store(
            *keyData, *SkData::MakeWithoutCopy(data.get(), dataSize));
}

////////////////////////////////////////////////////////////////////////////////

GrVkResourceProvider::CompatibleRenderPassSet::CompatibleRenderPassSet(GrVkRenderPass* renderPass)
        : fLastReturnedIndex(0) {
    renderPass->ref();
    fRenderPasses.push_back(renderPass);
}

bool GrVkResourceProvider::CompatibleRenderPassSet::isCompatible(
        const GrVkRenderTarget& target) const {
    // The first GrVkRenderpass should always exists since we create the basic load store
    // render pass on create
    SkASSERT(fRenderPasses[0]);
    return fRenderPasses[0]->isCompatible(target);
}

GrVkRenderPass* GrVkResourceProvider::CompatibleRenderPassSet::getRenderPass(
        GrVkGpu* gpu,
        const GrVkRenderPass::LoadStoreOps& colorOps,
        const GrVkRenderPass::LoadStoreOps& stencilOps) {
    for (int i = 0; i < fRenderPasses.count(); ++i) {
        int idx = (i + fLastReturnedIndex) % fRenderPasses.count();
        if (fRenderPasses[idx]->equalLoadStoreOps(colorOps, stencilOps)) {
            fLastReturnedIndex = idx;
            return fRenderPasses[idx];
        }
    }
    GrVkRenderPass* renderPass = GrVkRenderPass::Create(gpu, *this->getCompatibleRenderPass(),
                                                        colorOps, stencilOps);
    if (!renderPass) {
        return nullptr;
    }
    fRenderPasses.push_back(renderPass);
    fLastReturnedIndex = fRenderPasses.count() - 1;
    return renderPass;
}

void GrVkResourceProvider::CompatibleRenderPassSet::releaseResources(GrVkGpu* gpu) {
    for (int i = 0; i < fRenderPasses.count(); ++i) {
        if (fRenderPasses[i]) {
            fRenderPasses[i]->unref(gpu);
            fRenderPasses[i] = nullptr;
        }
    }
}