Refactor GrBufferAllocPools to use resource cache

Committed: https://skia.googlesource.com/skia/+/e935f1a0e2351373c33600b8388492ce1218014a

Review URL: https://codereview.chromium.org/1139753002

11 files changed, 161 insertions, 143 deletions

diff --git a/src/gpu/GrBatchTarget.cpp b/src/gpu/GrBatchTarget.cpp
index 31b4cc9f47..b6dadd699b 100644
--- a/src/gpu/GrBatchTarget.cpp
+++ b/src/gpu/GrBatchTarget.cpp
@@ -10,27 +10,16 @@
 #include "GrBatchAtlas.h"
 #include "GrPipeline.h"
 
-static const size_t DRAW_BUFFER_VBPOOL_BUFFER_SIZE = 1 << 15;
-static const int DRAW_BUFFER_VBPOOL_PREALLOC_BUFFERS = 4;
-
-static const size_t DRAW_BUFFER_IBPOOL_BUFFER_SIZE = 1 << 11;
-static const int DRAW_BUFFER_IBPOOL_PREALLOC_BUFFERS = 4;
-
 GrBatchTarget::GrBatchTarget(GrGpu* gpu)
     : fGpu(gpu)
+    , fVertexPool(gpu)
+    , fIndexPool(gpu)
     , fFlushBuffer(kFlushBufferInitialSizeInBytes)
     , fIter(fFlushBuffer)
     , fNumberOfDraws(0)
     , fCurrentToken(0)
     , fLastFlushedToken(0)
     , fInlineUpdatesIndex(0) {
-
-    fVertexPool.reset(SkNEW_ARGS(GrVertexBufferAllocPool, (fGpu,
-                                             DRAW_BUFFER_VBPOOL_BUFFER_SIZE,
-                                             DRAW_BUFFER_VBPOOL_PREALLOC_BUFFERS)));
-    fIndexPool.reset(SkNEW_ARGS(GrIndexBufferAllocPool, (fGpu,
-                                            DRAW_BUFFER_IBPOOL_BUFFER_SIZE,
-                                            DRAW_BUFFER_IBPOOL_PREALLOC_BUFFERS)));
 }
 
 void GrBatchTarget::flushNext(int n)  {
@@ -65,11 +54,11 @@ void GrBatchTarget::flushNext(int n)  {
 
 void* GrBatchTarget::makeVertSpace(size_t vertexSize, int vertexCount,
                     const GrVertexBuffer** buffer, int* startVertex) {
-    return fVertexPool->makeSpace(vertexSize, vertexCount, buffer, startVertex);
+    return fVertexPool.makeSpace(vertexSize, vertexCount, buffer, startVertex);
 }
 
 uint16_t* GrBatchTarget::makeIndexSpace(int indexCount,
                                         const GrIndexBuffer** buffer, int* startIndex) {
-    return reinterpret_cast<uint16_t*>(fIndexPool->makeSpace(indexCount, buffer, startIndex));
+    return reinterpret_cast<uint16_t*>(fIndexPool.makeSpace(indexCount, buffer, startIndex));
 }
 
diff --git a/src/gpu/GrBatchTarget.h b/src/gpu/GrBatchTarget.h
index 80a8b9bad3..19cfb09cb7 100644
--- a/src/gpu/GrBatchTarget.h
+++ b/src/gpu/GrBatchTarget.h
@@ -115,26 +115,26 @@ public:
                              const GrIndexBuffer** buffer, int* startIndex);
 
     // A helper for draws which overallocate and then return data to the pool
-    void putBackIndices(size_t indices) { fIndexPool->putBack(indices * sizeof(uint16_t)); }
+    void putBackIndices(size_t indices) { fIndexPool.putBack(indices * sizeof(uint16_t)); }
 
     void putBackVertices(size_t vertices, size_t vertexStride) {
-        fVertexPool->putBack(vertices * vertexStride);
+        fVertexPool.putBack(vertices * vertexStride);
     }
 
     void reset() {
-        fVertexPool->reset();
-        fIndexPool->reset();    
+        fVertexPool.reset();
+        fIndexPool.reset();    
     }
 
 private:
     void unmapVertexAndIndexBuffers() {
-        fVertexPool->unmap();
-        fIndexPool->unmap();
+        fVertexPool.unmap();
+        fIndexPool.unmap();
     }
 
     GrGpu* fGpu;
-    SkAutoTDelete<GrVertexBufferAllocPool> fVertexPool;
-    SkAutoTDelete<GrIndexBufferAllocPool> fIndexPool;
+    GrVertexBufferAllocPool fVertexPool;
+    GrIndexBufferAllocPool fIndexPool;
 
     typedef void* TBufferAlign; // This wouldn't be enough align if a command used long double.
 
diff --git a/src/gpu/GrBufferAllocPool.cpp b/src/gpu/GrBufferAllocPool.cpp
index be88b32cb1..e9172108f3 100644
--- a/src/gpu/GrBufferAllocPool.cpp
+++ b/src/gpu/GrBufferAllocPool.cpp
@@ -9,8 +9,10 @@
 
 #include "GrBufferAllocPool.h"
 #include "GrCaps.h"
+#include "GrContext.h"
 #include "GrGpu.h"
 #include "GrIndexBuffer.h"
+#include "GrResourceProvider.h"
 #include "GrTypes.h"
 #include "GrVertexBuffer.h"
 
@@ -22,6 +24,9 @@
     static void VALIDATE(bool = false) {}
 #endif
 
+static const size_t MIN_VERTEX_BUFFER_SIZE = 1 << 15;
+static const size_t MIN_INDEX_BUFFER_SIZE = 1 << 12;
+
 // page size
 #define GrBufferAllocPool_MIN_BLOCK_SIZE ((size_t)1 << 15)
 
@@ -37,9 +42,8 @@ do {
 
 GrBufferAllocPool::GrBufferAllocPool(GrGpu* gpu,
                                      BufferType bufferType,
-                                     size_t blockSize,
-                                     int preallocBufferCnt)
-    : fBlocks(SkTMax(8, 2*preallocBufferCnt)) {
+                                     size_t blockSize)
+    : fBlocks(8) {
 
     fGpu = SkRef(gpu);
 
@@ -49,19 +53,10 @@ GrBufferAllocPool::GrBufferAllocPool(GrGpu* gpu,
 
     fBytesInUse = 0;
 
-    fPreallocBuffersInUse = 0;
-    fPreallocBufferStartIdx = 0;
-    for (int i = 0; i < preallocBufferCnt; ++i) {
-        GrGeometryBuffer* buffer = this->createBuffer(fMinBlockSize);
-        if (buffer) {
-            *fPreallocBuffers.append() = buffer;
-        }
-    }
     fGeometryBufferMapThreshold = gpu->caps()->geometryBufferMapThreshold();
 }
 
-GrBufferAllocPool::~GrBufferAllocPool() {
-    VALIDATE();
+void GrBufferAllocPool::deleteBlocks() {
     if (fBlocks.count()) {
         GrGeometryBuffer* buffer = fBlocks.back().fBuffer;
         if (buffer->isMapped()) {
@@ -71,34 +66,22 @@ GrBufferAllocPool::~GrBufferAllocPool() {
     while (!fBlocks.empty()) {
         this->destroyBlock();
     }
-    fPreallocBuffers.unrefAll();
+    SkASSERT(!fBufferPtr);
+}
+
+GrBufferAllocPool::~GrBufferAllocPool() {
+    VALIDATE();
+    this->deleteBlocks();
     fGpu->unref();
 }
 
 void GrBufferAllocPool::reset() {
     VALIDATE();
     fBytesInUse = 0;
-    if (fBlocks.count()) {
-        GrGeometryBuffer* buffer = fBlocks.back().fBuffer;
-        if (buffer->isMapped()) {
-            UNMAP_BUFFER(fBlocks.back());
-        }
-    }
-    // fPreallocBuffersInUse will be decremented down to zero in the while loop
-    int preallocBuffersInUse = fPreallocBuffersInUse;
-    while (!fBlocks.empty()) {
-        this->destroyBlock();
-    }
-    if (fPreallocBuffers.count()) {
-        // must set this after above loop.
-        fPreallocBufferStartIdx = (fPreallocBufferStartIdx +
-                                   preallocBuffersInUse) %
-                                  fPreallocBuffers.count();
-    }
+    this->deleteBlocks();
     // we may have created a large cpu mirror of a large VB. Reset the size
-    // to match our pre-allocated VBs.
+    // to match our minimum.
     fCpuData.reset(fMinBlockSize);
-    SkASSERT(0 == fPreallocBuffersInUse);
     VALIDATE();
 }
 
@@ -170,8 +153,7 @@ void* GrBufferAllocPool::makeSpace(size_t size,
     if (fBufferPtr) {
         BufferBlock& back = fBlocks.back();
         size_t usedBytes = back.fBuffer->gpuMemorySize() - back.fBytesFree;
-        size_t pad = GrSizeAlignUpPad(usedBytes,
-                                      alignment);
+        size_t pad = GrSizeAlignUpPad(usedBytes, alignment);
         if ((size + pad) <= back.fBytesFree) {
             memset((void*)(reinterpret_cast<intptr_t>(fBufferPtr) + usedBytes), 0, pad);
             usedBytes += pad;
@@ -209,12 +191,6 @@ void* GrBufferAllocPool::makeSpace(size_t size,
 void GrBufferAllocPool::putBack(size_t bytes) {
     VALIDATE();
 
-    // if the putBack unwinds all the preallocated buffers then we will
-    // advance the starting index. As blocks are destroyed fPreallocBuffersInUse
-    // will be decremented. I will reach zero if all blocks using preallocated
-    // buffers are released.
-    int preallocBuffersInUse = fPreallocBuffersInUse;
-
     while (bytes) {
         // caller shouldn't try to put back more than they've taken
         SkASSERT(!fBlocks.empty());
@@ -236,11 +212,7 @@ void GrBufferAllocPool::putBack(size_t bytes) {
             break;
         }
     }
-    if (!fPreallocBuffersInUse && fPreallocBuffers.count()) {
-            fPreallocBufferStartIdx = (fPreallocBufferStartIdx +
-                                       preallocBuffersInUse) %
-                                      fPreallocBuffers.count();
-    }
+
     VALIDATE();
 }
 
@@ -253,24 +225,13 @@ bool GrBufferAllocPool::createBlock(size_t requestSize) {
 
     BufferBlock& block = fBlocks.push_back();
 
-    if (size == fMinBlockSize &&
-        fPreallocBuffersInUse < fPreallocBuffers.count()) {
-
-        uint32_t nextBuffer = (fPreallocBuffersInUse +
-                               fPreallocBufferStartIdx) %
-                              fPreallocBuffers.count();
-        block.fBuffer = fPreallocBuffers[nextBuffer];
-        block.fBuffer->ref();
-        ++fPreallocBuffersInUse;
-    } else {
-        block.fBuffer = this->createBuffer(size);
-        if (NULL == block.fBuffer) {
-            fBlocks.pop_back();
-            return false;
-        }
+    block.fBuffer = this->getBuffer(size);
+    if (NULL == block.fBuffer) {
+        fBlocks.pop_back();
+        return false;
     }
 
-    block.fBytesFree = size;
+    block.fBytesFree = block.fBuffer->gpuMemorySize();
     if (fBufferPtr) {
         SkASSERT(fBlocks.count() > 1);
         BufferBlock& prev = fBlocks.fromBack(1);
@@ -297,7 +258,7 @@ bool GrBufferAllocPool::createBlock(size_t requestSize) {
     }
 
     if (NULL == fBufferPtr) {
-        fBufferPtr = fCpuData.reset(size);
+        fBufferPtr = fCpuData.reset(block.fBytesFree);
     }
 
     VALIDATE(true);
@@ -309,15 +270,7 @@ void GrBufferAllocPool::destroyBlock() {
     SkASSERT(!fBlocks.empty());
 
     BufferBlock& block = fBlocks.back();
-    if (fPreallocBuffersInUse > 0) {
-        uint32_t prevPreallocBuffer = (fPreallocBuffersInUse +
-                                       fPreallocBufferStartIdx +
-                                       (fPreallocBuffers.count() - 1)) %
-                                      fPreallocBuffers.count();
-        if (block.fBuffer == fPreallocBuffers[prevPreallocBuffer]) {
-            --fPreallocBuffersInUse;
-        }
-    }
+
     SkASSERT(!block.fBuffer->isMapped());
     block.fBuffer->unref();
     fBlocks.pop_back();
@@ -345,24 +298,22 @@ void GrBufferAllocPool::flushCpuData(const BufferBlock& block, size_t flushSize)
     VALIDATE(true);
 }
 
-GrGeometryBuffer* GrBufferAllocPool::createBuffer(size_t size) {
+GrGeometryBuffer* GrBufferAllocPool::getBuffer(size_t size) {
+
+    GrResourceProvider* rp = fGpu->getContext()->resourceProvider();
+
     if (kIndex_BufferType == fBufferType) {
-        return fGpu->createIndexBuffer(size, true);
+        return rp->getIndexBuffer(size, /* dynamic = */ true, /* duringFlush = */ true);
     } else {
         SkASSERT(kVertex_BufferType == fBufferType);
-        return fGpu->createVertexBuffer(size, true);
+        return rp->getVertexBuffer(size, /* dynamic = */ true, /* duringFlush = */ true);
     }
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 
-GrVertexBufferAllocPool::GrVertexBufferAllocPool(GrGpu* gpu,
-                                                 size_t bufferSize,
-                                                 int preallocBufferCnt)
-    : GrBufferAllocPool(gpu,
-                        kVertex_BufferType,
-                        bufferSize,
-                        preallocBufferCnt) {
+GrVertexBufferAllocPool::GrVertexBufferAllocPool(GrGpu* gpu)
+    : GrBufferAllocPool(gpu, kVertex_BufferType, MIN_VERTEX_BUFFER_SIZE) {
 }
 
 void* GrVertexBufferAllocPool::makeSpace(size_t vertexSize,
@@ -389,13 +340,8 @@ void* GrVertexBufferAllocPool::makeSpace(size_t vertexSize,
 
 ////////////////////////////////////////////////////////////////////////////////
 
-GrIndexBufferAllocPool::GrIndexBufferAllocPool(GrGpu* gpu,
-                                               size_t bufferSize,
-                                               int preallocBufferCnt)
-    : GrBufferAllocPool(gpu,
-                        kIndex_BufferType,
-                        bufferSize,
-                        preallocBufferCnt) {
+GrIndexBufferAllocPool::GrIndexBufferAllocPool(GrGpu* gpu)
+    : GrBufferAllocPool(gpu, kIndex_BufferType, MIN_INDEX_BUFFER_SIZE) {
 }
 
 void* GrIndexBufferAllocPool::makeSpace(int indexCount,
diff --git a/src/gpu/GrBufferAllocPool.h b/src/gpu/GrBufferAllocPool.h
index cd5d1971da..bbcb8a5d7c 100644
--- a/src/gpu/GrBufferAllocPool.h
+++ b/src/gpu/GrBufferAllocPool.h
@@ -64,16 +64,12 @@ protected:
      * @param bufferSize            The minimum size of created buffers.
      *                              This value will be clamped to some
      *                              reasonable minimum.
-     * @param preallocBufferCnt     The pool will allocate this number of
-     *                              buffers at bufferSize and keep them until it
-     *                              is destroyed.
      */
      GrBufferAllocPool(GrGpu* gpu,
                        BufferType bufferType,
-                       size_t   bufferSize = 0,
-                       int preallocBufferCnt = 0);
+                       size_t   bufferSize = 0);
 
-    virtual ~GrBufferAllocPool();
+     virtual ~GrBufferAllocPool();
 
     /**
      * Returns a block of memory to hold data. A buffer designated to hold the
@@ -99,7 +95,7 @@ protected:
                     const GrGeometryBuffer** buffer,
                     size_t* offset);
 
-    GrGeometryBuffer* createBuffer(size_t size);
+    GrGeometryBuffer* getBuffer(size_t size);
 
 private:
     struct BufferBlock {
@@ -109,6 +105,7 @@ private:
 
     bool createBlock(size_t requestSize);
     void destroyBlock();
+    void deleteBlocks();
     void flushCpuData(const BufferBlock& block, size_t flushSize);
 #ifdef SK_DEBUG
     void validate(bool unusedBlockAllowed = false) const;
@@ -117,15 +114,10 @@ private:
     size_t                          fBytesInUse;
 
     GrGpu*                          fGpu;
-    SkTDArray<GrGeometryBuffer*>    fPreallocBuffers;
     size_t                          fMinBlockSize;
     BufferType                      fBufferType;
 
     SkTArray<BufferBlock>           fBlocks;
-    int                             fPreallocBuffersInUse;
-    // We attempt to cycle through the preallocated buffers rather than
-    // always starting from the first.
-    int                             fPreallocBufferStartIdx;
     SkAutoMalloc                    fCpuData;
     void*                           fBufferPtr;
     size_t                          fGeometryBufferMapThreshold;
@@ -142,13 +134,8 @@ public:
      * Constructor
      *
      * @param gpu                   The GrGpu used to create the vertex buffers.
-     * @param bufferSize            The minimum size of created VBs. This value
-     *                              will be clamped to some reasonable minimum.
-     * @param preallocBufferCnt     The pool will allocate this number of VBs at
-     *                              bufferSize and keep them until it is
-     *                              destroyed.
      */
-    GrVertexBufferAllocPool(GrGpu* gpu, size_t bufferSize = 0, int preallocBufferCnt = 0);
+    GrVertexBufferAllocPool(GrGpu* gpu);
 
     /**
      * Returns a block of memory to hold vertices. A buffer designated to hold
@@ -191,15 +178,8 @@ public:
      * Constructor
      *
      * @param gpu                   The GrGpu used to create the index buffers.
-     * @param bufferSize            The minimum size of created IBs. This value
-     *                              will be clamped to some reasonable minimum.
-     * @param preallocBufferCnt     The pool will allocate this number of VBs at
-     *                              bufferSize and keep them until it is
-     *                              destroyed.
      */
-    GrIndexBufferAllocPool(GrGpu* gpu,
-                           size_t bufferSize = 0,
-                           int preallocBufferCnt = 0);
+    GrIndexBufferAllocPool(GrGpu* gpu);
 
     /**
      * Returns a block of memory to hold indices. A buffer designated to hold
diff --git a/src/gpu/GrCaps.cpp b/src/gpu/GrCaps.cpp
index f2e88b01c4..900bd65cc2 100644
--- a/src/gpu/GrCaps.cpp
+++ b/src/gpu/GrCaps.cpp
@@ -86,6 +86,7 @@ GrCaps::GrCaps(const GrContextOptions& options) {
     fStencilWrapOpsSupport = false;
     fDiscardRenderTargetSupport = false;
     fReuseScratchTextures = true;
+    fReuseScratchBuffers = true;
     fGpuTracingSupport = false;
     fCompressedTexSubImageSupport = false;
     fOversizedStencilSupport = false;
@@ -146,6 +147,7 @@ SkString GrCaps::dump() const {
     r.appendf("Stencil Wrap Ops  Support          : %s\n", gNY[fStencilWrapOpsSupport]);
     r.appendf("Discard Render Target Support      : %s\n", gNY[fDiscardRenderTargetSupport]);
     r.appendf("Reuse Scratch Textures             : %s\n", gNY[fReuseScratchTextures]);
+    r.appendf("Reuse Scratch Buffers              : %s\n", gNY[fReuseScratchBuffers]);
     r.appendf("Gpu Tracing Support                : %s\n", gNY[fGpuTracingSupport]);
     r.appendf("Compressed Update Support          : %s\n", gNY[fCompressedTexSubImageSupport]);
     r.appendf("Oversized Stencil Support          : %s\n", gNY[fOversizedStencilSupport]);
diff --git a/src/gpu/GrGpu.cpp b/src/gpu/GrGpu.cpp
index d46d09d5c7..728fb88062 100644
--- a/src/gpu/GrGpu.cpp
+++ b/src/gpu/GrGpu.cpp
@@ -199,12 +199,20 @@ GrRenderTarget* GrGpu::wrapBackendRenderTarget(const GrBackendRenderTargetDesc&
 
 GrVertexBuffer* GrGpu::createVertexBuffer(size_t size, bool dynamic) {
     this->handleDirtyContext();
-    return this->onCreateVertexBuffer(size, dynamic);
+    GrVertexBuffer* vb = this->onCreateVertexBuffer(size, dynamic);
+    if (!this->caps()->reuseScratchBuffers()) {
+        vb->resourcePriv().removeScratchKey();
+    }
+    return vb;
 }
 
 GrIndexBuffer* GrGpu::createIndexBuffer(size_t size, bool dynamic) {
     this->handleDirtyContext();
-    return this->onCreateIndexBuffer(size, dynamic);
+    GrIndexBuffer* ib = this->onCreateIndexBuffer(size, dynamic);
+    if (!this->caps()->reuseScratchBuffers()) {
+        ib->resourcePriv().removeScratchKey();
+    }
+    return ib;
 }
 
 void GrGpu::clear(const SkIRect* rect,
diff --git a/src/gpu/GrIndexBuffer.h b/src/gpu/GrIndexBuffer.h
index 4dfd1c1b07..bf64ff87d2 100644
--- a/src/gpu/GrIndexBuffer.h
+++ b/src/gpu/GrIndexBuffer.h
@@ -13,8 +13,18 @@
 
 #include "GrGeometryBuffer.h"
 
+
 class GrIndexBuffer : public GrGeometryBuffer {
 public:
+    static void ComputeScratchKey(size_t size, bool dynamic, GrScratchKey* key) {
+        static const GrScratchKey::ResourceType kType = GrScratchKey::GenerateResourceType();
+
+        GrScratchKey::Builder builder(key, kType, 2);
+
+        builder[0] = SkToUInt(size);
+        builder[1] = dynamic ? 1 : 0;
+    }
+
     /**
      * Retrieves the maximum number of quads that could be rendered
      * from the index buffer (using kTriangles_GrPrimitiveType).
@@ -25,7 +35,12 @@ public:
     }
 protected:
     GrIndexBuffer(GrGpu* gpu, size_t gpuMemorySize, bool dynamic, bool cpuBacked)
-        : INHERITED(gpu, gpuMemorySize, dynamic, cpuBacked) {}
+        : INHERITED(gpu, gpuMemorySize, dynamic, cpuBacked) {
+        GrScratchKey key;
+        ComputeScratchKey(gpuMemorySize, dynamic, &key);
+        this->setScratchKey(key);
+    }
+
 private:
     typedef GrGeometryBuffer INHERITED;
 };
diff --git a/src/gpu/GrResourceProvider.cpp b/src/gpu/GrResourceProvider.cpp
index 4b9d265129..3c447bded6 100644
--- a/src/gpu/GrResourceProvider.cpp
+++ b/src/gpu/GrResourceProvider.cpp
@@ -28,7 +28,7 @@ const GrIndexBuffer* GrResourceProvider::createInstancedIndexBuffer(const uint16
                                                                     const GrUniqueKey& key) {
     size_t bufferSize = patternSize * reps * sizeof(uint16_t);
 
-    GrIndexBuffer* buffer = this->gpu()->createIndexBuffer(bufferSize, /* dynamic = */ false);
+    GrIndexBuffer* buffer = this->getIndexBuffer(bufferSize, /* dynamic = */ false, true);
     if (!buffer) {
         return NULL;
     }
@@ -83,3 +83,58 @@ GrPathRange* GrResourceProvider::createGlyphs(const SkTypeface* tf, const SkDesc
     return this->gpu()->pathRendering()->createGlyphs(tf, desc, stroke);
 }
 
+GrIndexBuffer* GrResourceProvider::getIndexBuffer(size_t size, bool dynamic,
+                                                  bool calledDuringFlush) {
+    if (this->isAbandoned()) {
+        return NULL;
+    }
+
+    if (dynamic) {
+        // bin by pow2 with a reasonable min
+        static const uint32_t MIN_SIZE = 1 << 12;
+        size = SkTMax(MIN_SIZE, GrNextPow2(SkToUInt(size)));
+
+        GrScratchKey key;
+        GrIndexBuffer::ComputeScratchKey(size, dynamic, &key);
+        uint32_t scratchFlags = 0;
+        if (calledDuringFlush) {
+            scratchFlags = GrResourceCache::kRequireNoPendingIO_ScratchFlag;
+        } else {
+            scratchFlags = GrResourceCache::kPreferNoPendingIO_ScratchFlag;
+        }
+        GrGpuResource* resource = this->cache()->findAndRefScratchResource(key, scratchFlags);
+        if (resource) {
+            return static_cast<GrIndexBuffer*>(resource);
+        }
+    }
+
+    return this->gpu()->createIndexBuffer(size, dynamic);    
+}
+
+GrVertexBuffer* GrResourceProvider::getVertexBuffer(size_t size, bool dynamic, 
+                                                    bool calledDuringFlush) {
+    if (this->isAbandoned()) {
+        return NULL;
+    }
+
+    if (dynamic) {
+        // bin by pow2 with a reasonable min
+        static const uint32_t MIN_SIZE = 1 << 15;
+        size = SkTMax(MIN_SIZE, GrNextPow2(SkToUInt(size)));
+
+        GrScratchKey key;
+        GrVertexBuffer::ComputeScratchKey(size, dynamic, &key);
+        uint32_t scratchFlags = 0;
+        if (calledDuringFlush) {
+            scratchFlags = GrResourceCache::kRequireNoPendingIO_ScratchFlag;
+        } else {
+            scratchFlags = GrResourceCache::kPreferNoPendingIO_ScratchFlag;
+        }
+        GrGpuResource* resource = this->cache()->findAndRefScratchResource(key, scratchFlags);
+        if (resource) {
+            return static_cast<GrVertexBuffer*>(resource);
+        }
+    }
+
+    return this->gpu()->createVertexBuffer(size, dynamic);
+}
diff --git a/src/gpu/GrResourceProvider.h b/src/gpu/GrResourceProvider.h
index 2962ad3788..f0b79b97f5 100644
--- a/src/gpu/GrResourceProvider.h
+++ b/src/gpu/GrResourceProvider.h
@@ -86,6 +86,9 @@ public:
     using GrTextureProvider::findAndRefResourceByUniqueKey;
     using GrTextureProvider::abandon;
 
+    GrIndexBuffer* getIndexBuffer(size_t size, bool dynamic, bool calledDuringFlush);
+    GrVertexBuffer* getVertexBuffer(size_t size, bool dynamic, bool calledDuringFlush);
+
 private:
     const GrIndexBuffer* createInstancedIndexBuffer(const uint16_t* pattern,
                                                     int patternSize,
diff --git a/src/gpu/GrVertexBuffer.h b/src/gpu/GrVertexBuffer.h
index 3f2ada2a2c..3c12cd76d2 100644
--- a/src/gpu/GrVertexBuffer.h
+++ b/src/gpu/GrVertexBuffer.h
@@ -14,9 +14,24 @@
 #include "GrGeometryBuffer.h"
 
 class GrVertexBuffer : public GrGeometryBuffer {
+public:
+    static void ComputeScratchKey(size_t size, bool dynamic, GrScratchKey* key) {
+        static const GrScratchKey::ResourceType kType = GrScratchKey::GenerateResourceType();
+
+        GrScratchKey::Builder builder(key, kType, 2);
+
+        builder[0] = SkToUInt(size);
+        builder[1] = dynamic ? 1 : 0;
+    }
+
 protected:
     GrVertexBuffer(GrGpu* gpu, size_t gpuMemorySize, bool dynamic, bool cpuBacked)
-        : INHERITED(gpu, gpuMemorySize, dynamic, cpuBacked) {}
+        : INHERITED(gpu, gpuMemorySize, dynamic, cpuBacked) {
+        GrScratchKey key;
+        ComputeScratchKey(gpuMemorySize, dynamic, &key);
+        this->setScratchKey(key);
+    }
+
 private:
     typedef GrGeometryBuffer INHERITED;
 };
diff --git a/src/gpu/gl/GrGLCaps.cpp b/src/gpu/gl/GrGLCaps.cpp
index b23619d8b6..0c53006bd2 100644
--- a/src/gpu/gl/GrGLCaps.cpp
+++ b/src/gpu/gl/GrGLCaps.cpp
@@ -414,6 +414,11 @@ void GrGLCaps::init(const GrContextOptions& contextOptions,
     fReuseScratchTextures = kARM_GrGLVendor != ctxInfo.vendor() &&
                             kQualcomm_GrGLVendor != ctxInfo.vendor();
 
+#if 0
+    fReuseScratchBuffers = kARM_GrGLVendor != ctxInfo.vendor() &&
+                           kQualcomm_GrGLVendor != ctxInfo.vendor();
+#endif
+
     if (GrGLCaps::kES_IMG_MsToTexture_MSFBOType == fMSFBOType) {
         GR_GL_GetIntegerv(gli, GR_GL_MAX_SAMPLES_IMG, &fMaxSampleCount);
     } else if (GrGLCaps::kNone_MSFBOType != fMSFBOType) {