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

src/gpu/vk/GrVkPipelineStateCache.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 "GrVkResourceProvider.h"
+
+#include "GrVkGpu.h"
+#include "GrProcessor.h"
+#include "GrVkPipelineState.h"
+#include "GrVkPipelineStateBuilder.h"
+#include "SkRTConf.h"
+#include "SkTSearch.h"
+#include "glsl/GrGLSLFragmentProcessor.h"
+#include "glsl/GrGLSLProgramDataManager.h"
+
+#ifdef PIPELINE_STATE_CACHE_STATS
+SK_CONF_DECLARE(bool, c_DisplayCache, "gpu.displayCache", false,
+                "Display pipeline state cache usage.");
+#endif
+
+typedef GrGLSLProgramDataManager::UniformHandle UniformHandle;
+
+struct GrVkResourceProvider::PipelineStateCache::Entry {
+
+    Entry() : fPipelineState(nullptr), fLRUStamp(0) {}
+
+    SkAutoTUnref<GrVkPipelineState> fPipelineState;
+    unsigned int                    fLRUStamp;
+};
+
+struct GrVkResourceProvider::PipelineStateCache::PipelineDescLess {
+    bool operator() (const GrVkPipelineState::Desc& desc, const Entry* entry) {
+        SkASSERT(entry->fPipelineState.get());
+        return GrVkPipelineState::Desc::Less(desc, entry->fPipelineState->getDesc());
+    }
+
+    bool operator() (const Entry* entry, const GrVkPipelineState::Desc& desc) {
+        SkASSERT(entry->fPipelineState.get());
+        return GrVkPipelineState::Desc::Less(entry->fPipelineState->getDesc(), desc);
+    }
+};
+
+GrVkResourceProvider::PipelineStateCache::PipelineStateCache(GrVkGpu* gpu)
+    : fCount(0)
+    , fCurrLRUStamp(0)
+    , fGpu(gpu)
+#ifdef PIPELINE_STATE_CACHE_STATS
+    , fTotalRequests(0)
+    , fCacheMisses(0)
+    , fHashMisses(0)
+#endif
+{
+    for (int i = 0; i < 1 << kHashBits; ++i) {
+        fHashTable[i] = nullptr;
+    }
+}
+
+GrVkResourceProvider::PipelineStateCache::~PipelineStateCache() {
+    SkASSERT(0 == fCount);
+    // dump stats
+#ifdef PIPELINE_STATE_CACHE_STATS
+    if (c_DisplayCache) {
+        SkDebugf("--- Pipeline State Cache ---\n");
+        SkDebugf("Total requests: %d\n", fTotalRequests);
+        SkDebugf("Cache misses: %d\n", fCacheMisses);
+        SkDebugf("Cache miss %%: %f\n", (fTotalRequests > 0) ?
+                 100.f * fCacheMisses / fTotalRequests :
+                 0.f);
+        int cacheHits = fTotalRequests - fCacheMisses;
+        SkDebugf("Hash miss %%: %f\n", (cacheHits > 0) ? 100.f * fHashMisses / cacheHits : 0.f);
+        SkDebugf("---------------------\n");
+    }
+#endif
+}
+
+void GrVkResourceProvider::PipelineStateCache::reset() {
+    for (int i = 0; i < fCount; ++i) {
+        delete fEntries[i];
+        fEntries[i] = nullptr;
+    }
+    fCount = 0;
+
+    // zero out hash table
+    for (int i = 0; i < 1 << kHashBits; i++) {
+        fHashTable[i] = nullptr;
+    }
+
+    fCurrLRUStamp = 0;
+}
+
+void GrVkResourceProvider::PipelineStateCache::abandon() {
+    for (int i = 0; i < fCount; ++i) {
+        SkASSERT(fEntries[i]->fPipelineState.get());
+        fEntries[i]->fPipelineState->abandonGPUResources();
+    }
+    this->reset();
+}
+
+void GrVkResourceProvider::PipelineStateCache::release() {
+    for (int i = 0; i < fCount; ++i) {
+        SkASSERT(fEntries[i]->fPipelineState.get());
+        fEntries[i]->fPipelineState->freeGPUResources(fGpu);
+    }
+    this->reset();
+}
+
+int GrVkResourceProvider::PipelineStateCache::search(const GrVkPipelineState::Desc& desc) const {
+    PipelineDescLess less;
+    return SkTSearch(fEntries, fCount, desc, sizeof(Entry*), less);
+}
+
+GrVkPipelineState* GrVkResourceProvider::PipelineStateCache::refPipelineState(
+                                                               const GrPipeline& pipeline,
+                                                               const GrPrimitiveProcessor& primProc,
+                                                               GrPrimitiveType primiteType,
+                                                               const GrVkRenderPass& renderPass) {
+#ifdef PIPELINE_STATE_CACHE_STATS
+    ++fTotalRequests;
+#endif
+
+    Entry* entry = nullptr;
+
+    // Get GrVkProgramDesc
+    GrVkPipelineState::Desc desc;
+    if (!GrVkProgramDescBuilder::Build(&desc.fProgramDesc,
+                                       primProc,
+                                       pipeline,
+                                       *fGpu->vkCaps().glslCaps())) {
+        GrCapsDebugf(fGpu->caps(), "Failed to build vk program descriptor!\n");
+        return false;
+    }
+
+    // Get vulkan specific descriptor key
+    GrVkPipelineState::BuildStateKey(pipeline, primiteType, &desc.fStateKey);
+    // Get checksum of entire PipelineDesc
+    int keyLength = desc.fStateKey.count();
+    SkASSERT(0 == (keyLength % 4));
+    // Seed the checksum with the checksum of the programDesc then add the vulkan key to it.
+    desc.fChecksum = SkChecksum::Murmur3(desc.fStateKey.begin(), keyLength,
+                                         desc.fProgramDesc.getChecksum());
+
+    uint32_t hashIdx = desc.fChecksum;
+    hashIdx ^= hashIdx >> 16;
+    if (kHashBits <= 8) {
+        hashIdx ^= hashIdx >> 8;
+    }
+    hashIdx &= ((1 << kHashBits) - 1);
+    Entry* hashedEntry = fHashTable[hashIdx];
+    if (hashedEntry && hashedEntry->fPipelineState->getDesc() == desc) {
+        SkASSERT(hashedEntry->fPipelineState);
+        entry = hashedEntry;
+    }
+
+    int entryIdx;
+    if (nullptr == entry) {
+        entryIdx = this->search(desc);
+        if (entryIdx >= 0) {
+            entry = fEntries[entryIdx];
+#ifdef PIPELINE_STATE_CACHE_STATS
+            ++fHashMisses;
+#endif
+        }
+    }
+
+    if (nullptr == entry) {
+        // We have a cache miss
+#ifdef PIPELINE_STATE_CACHE_STATS
+        ++fCacheMisses;
+#endif
+        GrVkPipelineState* pipelineState =
+            GrVkPipelineStateBuilder::CreatePipelineState(fGpu,
+                                                          pipeline,
+                                                          primProc,
+                                                          primiteType,
+                                                          desc,
+                                                          renderPass);
+        if (nullptr == pipelineState) {
+            return nullptr;
+        }
+        int purgeIdx = 0;
+        if (fCount < kMaxEntries) {
+            entry = new Entry;
+            purgeIdx = fCount++;
+            fEntries[purgeIdx] = entry;
+        } else {
+            SkASSERT(fCount == kMaxEntries);
+            purgeIdx = 0;
+            for (int i = 1; i < kMaxEntries; ++i) {
+                if (fEntries[i]->fLRUStamp < fEntries[purgeIdx]->fLRUStamp) {
+                    purgeIdx = i;
+                }
+            }
+            entry = fEntries[purgeIdx];
+            int purgedHashIdx = entry->fPipelineState->getDesc().fChecksum & ((1 << kHashBits) - 1);
+            if (fHashTable[purgedHashIdx] == entry) {
+                fHashTable[purgedHashIdx] = nullptr;
+            }
+            entry->fPipelineState->freeGPUResources(fGpu);
+        }
+        SkASSERT(fEntries[purgeIdx] == entry);
+        entry->fPipelineState.reset(pipelineState);
+        // We need to shift fEntries around so that the entry currently at purgeIdx is placed
+        // just before the entry at ~entryIdx (in order to keep fEntries sorted by descriptor).
+        entryIdx = ~entryIdx;
+        if (entryIdx < purgeIdx) {
+            //  Let E and P be the entries at index entryIdx and purgeIdx, respectively.
+            //  If the entries array looks like this:
+            //       aaaaEbbbbbPccccc
+            //  we rearrange it to look like this:
+            //       aaaaPEbbbbbccccc
+            size_t copySize = (purgeIdx - entryIdx) * sizeof(Entry*);
+            memmove(fEntries + entryIdx + 1, fEntries + entryIdx, copySize);
+            fEntries[entryIdx] = entry;
+        } else if (purgeIdx < entryIdx) {
+            //  If the entries array looks like this:
+            //       aaaaPbbbbbEccccc
+            //  we rearrange it to look like this:
+            //       aaaabbbbbPEccccc
+            size_t copySize = (entryIdx - purgeIdx - 1) * sizeof(Entry*);
+            memmove(fEntries + purgeIdx, fEntries + purgeIdx + 1, copySize);
+            fEntries[entryIdx - 1] = entry;
+        }
+#ifdef SK_DEBUG
+        SkASSERT(fEntries[0]->fPipelineState.get());
+        for (int i = 0; i < fCount - 1; ++i) {
+            SkASSERT(fEntries[i + 1]->fPipelineState.get());
+            const GrVkPipelineState::Desc& a = fEntries[i]->fPipelineState->getDesc();
+            const GrVkPipelineState::Desc& b = fEntries[i + 1]->fPipelineState->getDesc();
+            SkASSERT(GrVkPipelineState::Desc::Less(a, b));
+            SkASSERT(!GrVkPipelineState::Desc::Less(b, a));
+        }
+#endif
+    }
+
+    fHashTable[hashIdx] = entry;
+    entry->fLRUStamp = fCurrLRUStamp;
+
+    if (SK_MaxU32 == fCurrLRUStamp) {
+        // wrap around! just trash our LRU, one time hit.
+        for (int i = 0; i < fCount; ++i) {
+            fEntries[i]->fLRUStamp = 0;
+        }
+    }
+    ++fCurrLRUStamp;
+    return SkRef(entry->fPipelineState.get());
+}