Start from scratch on a faster SkFlatDictionary.

src/core/SkPictureFlat.h

 
 /*
  * Copyright 2011 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 #ifndef SkPictureFlat_DEFINED
 #define SkPictureFlat_DEFINED
 
@@ skipping 133 matching lines @@
 //
 // The following templated classes provide an efficient way to store and compare
 // objects that have been flattened (i.e. serialized in an ordered binary
 // format).
 //
 // SkFlatData:       is a simple indexable container for the flattened data
 //                   which is agnostic to the type of data is is indexing. It is
 //                   also responsible for flattening/unflattening objects but
 //                   details of that operation are hidden in the provided procs
 // SkFlatDictionary: is an abstract templated dictionary that maintains a
-//                   searchable set of SkFlataData objects of type T.
+//                   searchable set of SkFlatData objects of type T.
 // SkFlatController: is an interface provided to SkFlatDictionary which handles
-//                   allocation and unallocation in some cases. It also holds
+//                   allocation (and unallocation in some cases). It also holds
 //                   ref count recorders and the like.
 //
 // NOTE: any class that wishes to be used in conjunction with SkFlatDictionary
 // must subclass the dictionary and provide the necessary flattening procs.
 // The end of this header contains dictionary subclasses for some common classes
 // like SkBitmap, SkMatrix, SkPaint, and SkRegion. SkFlatController must also
 // be implemented, or SkChunkFlatController can be used to use an
 // SkChunkAllocator and never do replacements.
 //
 //
 ///////////////////////////////////////////////////////////////////////////////
 
 class SkFlatData;
 
 class SkFlatController : public SkRefCnt {
 public:
     SK_DECLARE_INST_COUNT(SkFlatController)
 
     SkFlatController();
     virtual ~SkFlatController();
     /**
-     * Provide a new block of memory for the SkFlatDictionary to use.
+     * Return a new block of memory for the SkFlatDictionary to use.
+     * This memory is owned by the controller and has the same lifetime unless you
+     * call unalloc(), in which case it may be freed early.
      */
     virtual void* allocThrow(size_t bytes) = 0;
 
     /**
-     * Unallocate a previously allocated block, returned by allocThrow.
-     * Implementation should at least perform an unallocation if passed the last
-     * pointer returned by allocThrow. If findAndReplace() is intended to be
-     * used, unalloc should also be able to unallocate the SkFlatData that is
-     * provided.
+     * Hint that this block, which was allocated with allocThrow, is no longer needed.
+     * The implementation may choose to free this memory any time beteween now and destruction.
      */
     virtual void unalloc(void* ptr) = 0;
 
     /**
      * Used during creation and unflattening of SkFlatData objects. If the
      * objects being flattened contain bitmaps they are stored in this heap
      * and the flattenable stores the index to the bitmap on the heap.
      * This should be set by the protected setBitmapHeap.
      */
     SkBitmapHeap* getBitmapHeap() { return fBitmapHeap; }
@@ skipping 195 matching lines @@
     }
 
     enum {
         kInCache_Sentinel = 0,
         kCandidate_Sentinel = ~0U,
     };
     void setSentinel(uint32_t value) {
         SkASSERT(SkIsAlign4(fFlatSize));
         this->data32()[fFlatSize >> 2] = value;
     }
+
+    // This does not modify the payload flat data, in case it's already been written.
+    void stampHeaderAndSentinel(int index, int32_t size);
+    template <class T> friend class SkFlatDictionary;  // For stampHeaderAndSentinel().
 };
 
 template <class T>
 class SkFlatDictionary {
+    static const size_t kWriteBufferGrowthBytes = 1024;
+
 public:
-    SkFlatDictionary(SkFlatController* controller)
-    : fController(controller) {
-        fFlattenProc = NULL;
-        fUnflattenProc = NULL;
-        SkASSERT(controller);
-        fController->ref();
-        // set to 1 since returning a zero from find() indicates failure
-        fNextIndex = 1;
+    SkFlatDictionary(SkFlatController* controller, size_t scratchSizeGuess = 0)
+    : fFlattenProc(NULL)
+    , fUnflattenProc(NULL)
+    , fController(SkRef(controller))
+    , fScratchSize(scratchSizeGuess)
+    , fScratch(AllocScratch(fScratchSize))
+    , fWriteBuffer(kWriteBufferGrowthBytes)
+    , fWriteBufferReady(false)
+    , fNextIndex(1)  { // set to 1 since returning a zero from find() indicates failure
         sk_bzero(fHash, sizeof(fHash));
         // index 0 is always empty since it is used as a signal that find failed
         fIndexedData.push(NULL);
     }
 
-    virtual ~SkFlatDictionary() {
-        fController->unref();
+    ~SkFlatDictionary() {
+        sk_free(fScratch);
     }
 
     int count() const {
         SkASSERT(fIndexedData.count() == fSortedData.count()+1);
         return fSortedData.count();
     }
 
     const SkFlatData*  operator[](int index) const {
         SkASSERT(index >= 0 && index < fSortedData.count());
         return fSortedData[index];
@@ skipping 92 matching lines @@
         SkASSERT(fIndexedData.count() == fSortedData.count()+1);
         const SkFlatData* element = fIndexedData[index];
         SkASSERT(index == element->index());
 
         T* dst = new T;
         this->unflatten(dst, element);
         return dst;
     }
 
     const SkFlatData* findAndReturnFlat(const T& element) {
-        SkFlatData* flat = SkFlatData::Create(fController, &element, fNextIndex, fFlattenProc);
+        // Only valid until the next call to resetScratch().
+        const SkFlatData& scratch = this->resetScratch(element, fNextIndex);
 
-        int hashIndex = ChecksumToHashIndex(flat->checksum());
+        // See if we have it in the hash?
+        const int hashIndex = ChecksumToHashIndex(scratch.checksum());
         const SkFlatData* candidate = fHash[hashIndex];
-        if (candidate && !SkFlatData::Compare(*flat, *candidate)) {
-            fController->unalloc(flat);
+        if (candidate != NULL && SkFlatData::Compare(scratch, *candidate) == 0) {
             return candidate;
         }
 
-        int index = SkTSearch<const SkFlatData,
-                              SkFlatData::Less>((const SkFlatData**) fSortedData.begin(),
-                                                fSortedData.count(), flat, sizeof(flat));
+        // See if we have it at all?
+        const int index = SkTSearch<const SkFlatData, SkFlatData::Less>(fSortedData.begin(),
+                                                                        fSortedData.count(),
+                                                                        &scratch,
+                                                                        sizeof(&scratch));
         if (index >= 0) {
-            fController->unalloc(flat);
+            // Found.  Update hash before we return.
             fHash[hashIndex] = fSortedData[index];
             return fSortedData[index];
         }
 
-        index = ~index;
-        *fSortedData.insert(index) = flat;
-        *fIndexedData.insert(flat->index()) = flat;
+        // We don't have it.  Add it.
+        SkFlatData* detached = this->detachScratch();
+        // detached will live beyond the next call to resetScratch(), but is owned by fController.
+        *fSortedData.insert(~index) = detached;  // SkTSearch returned bit-not of where to insert.
+        *fIndexedData.insert(detached->index()) = detached;
+        fHash[hashIndex] = detached;
+
+        SkASSERT(detached->index() == fNextIndex);
         SkASSERT(fSortedData.count() == fNextIndex);
+        SkASSERT(fIndexedData.count() == fNextIndex+1);
         fNextIndex++;
-        flat->setSentinelInCache();
-        fHash[hashIndex] = flat;
-        SkASSERT(fIndexedData.count() == fSortedData.count()+1);
-        return flat;
+
+        return detached;
     }
 
 protected:
     void (*fFlattenProc)(SkOrderedWriteBuffer&, const void*);
     void (*fUnflattenProc)(SkOrderedReadBuffer&, void*);
 
 private:
+    // Layout: [ SkFlatData header, 20 bytes ] [ data ..., 4-byte aligned ] [ sentinel, 4 bytes]
+    static size_t SizeWithPadding(size_t flatDataSize) {
+        SkASSERT(SkIsAlign4(flatDataSize));
+        return sizeof(SkFlatData) + flatDataSize + sizeof(uint32_t);
+    }
+
+    // Allocate a new scratch SkFlatData.  Must be sk_freed.
+    static SkFlatData* AllocScratch(size_t scratchSize) {
+        return (SkFlatData*) sk_malloc_throw(SizeWithPadding(scratchSize));
+    }
+
+    // We have to delay fWriteBuffer's initialization until its first use; fController might not
+    // be fully set up by the time we get it in the constructor.
+    void lazyWriteBufferInit() {
+        if (fWriteBufferReady) {
+            return;
+        }
+        // Without a bitmap heap, we'll flatten bitmaps into paints.  That's never what you want.
+        SkASSERT(fController->getBitmapHeap() != NULL);
+        fWriteBuffer.setBitmapHeap(fController->getBitmapHeap());
+        fWriteBuffer.setTypefaceRecorder(fController->getTypefaceSet());
+        fWriteBuffer.setNamedFactoryRecorder(fController->getNamedFactorySet());
+        fWriteBuffer.setFlags(fController->getWriteBufferFlags());
+        fWriteBufferReady = true;
+    }
+
+    // This reference is valid only until the next call to resetScratch() or detachScratch().
+    const SkFlatData& resetScratch(const T& element, int index) {
+        this->lazyWriteBufferInit();
+
+        // Flatten element into fWriteBuffer (using fScratch as storage).
+        fWriteBuffer.reset(fScratch->data(), fScratchSize);
+        fFlattenProc(fWriteBuffer, &element);
+        const size_t bytesWritten = fWriteBuffer.bytesWritten();
+
+        // If all the flattened bytes fit into fScratch, we can skip a call to writeToMemory.
+        if (!fWriteBuffer.wroteOnlyToStorage()) {
+            SkASSERT(bytesWritten > fScratchSize);
+            // It didn't all fit.  Copy into a larger replacement SkFlatData.
+            // We can't just realloc because it might move the pointer and confuse writeToMemory.
+            SkFlatData* larger = AllocScratch(bytesWritten);
+            fWriteBuffer.writeToMemory(larger->data());
+
+            // Carry on with this larger scratch to minimize the likelihood of future resizing.
+            sk_free(fScratch);
+            fScratchSize = bytesWritten;
+            fScratch = larger;
+        }
+
+        // The data is in fScratch now, but we need to stamp its header and trailing sentinel.
+        fScratch->stampHeaderAndSentinel(index, bytesWritten);
+        return *fScratch;
+    }
+
+    // This result is owned by fController and lives as long as it does (unless unalloc'd).
+    SkFlatData* detachScratch() {
+        // Allocate a new SkFlatData exactly big enough to hold our current scratch.
+        // We use the controller for this allocation to extend the allocation's lifetime and allow
+        // the controller to do whatever memory management it wants.
+        const size_t paddedSize = SizeWithPadding(fScratch->flatSize());
+        SkFlatData* detached = (SkFlatData*)fController->allocThrow(paddedSize);
+
+        // Copy scratch into the new SkFlatData, setting the sentinel for cache storage.
+        memcpy(detached, fScratch, paddedSize);
+        detached->setSentinelInCache();
+
+        // We can now reuse fScratch, and detached will live until fController dies.
+        return detached;
+    }
+
     void unflatten(T* dst, const SkFlatData* element) const {
         element->unflatten(dst, fUnflattenProc,
                            fController->getBitmapHeap(),
                            fController->getTypefacePlayback());
     }
 
     void unflattenIntoArray(T* array) const {
         const int count = fSortedData.count();
         SkASSERT(fIndexedData.count() == fSortedData.count()+1);
         const SkFlatData* const* iter = fSortedData.begin();
         for (int i = 0; i < count; ++i) {
             const SkFlatData* element = iter[i];
             int index = element->index() - 1;
             SkASSERT((unsigned)index < (unsigned)count);
             unflatten(&array[index], element);
         }
     }
 
-    SkFlatController * const     fController;
-    int                          fNextIndex;
+    SkAutoTUnref<SkFlatController> fController;
+    size_t fScratchSize;  // How many bytes fScratch has allocated for data itself.
+    SkFlatData* fScratch;  // Owned, must be freed with sk_free.
+    SkOrderedWriteBuffer fWriteBuffer;
+    bool fWriteBufferReady;
 
     // SkFlatDictionary has two copies of the data one indexed by the
     // SkFlatData's index and the other sorted. The sorted data is used
     // for finding and uniquification while the indexed copy is used
     // for standard array-style lookups based on the SkFlatData's index
     // (as in 'unflatten').
+    int fNextIndex;
     SkTDArray<const SkFlatData*> fIndexedData;
     // fSortedData is sorted by checksum/size/data.
     SkTDArray<const SkFlatData*> fSortedData;
 
     enum {
         // Determined by trying diff values on picture-recording benchmarks
         // (e.g. PictureRecordBench.cpp), choosing the smallest value that
         // showed a big improvement. Even better would be to benchmark diff
         // values on recording representative web-pages or other "real" content.
         HASH_BITS   = 7,
@@ skipping 33 matching lines @@
 
 class SkChunkFlatController : public SkFlatController {
 public:
     SkChunkFlatController(size_t minSize)
     : fHeap(minSize)
     , fTypefaceSet(SkNEW(SkRefCntSet)) {
         this->setTypefaceSet(fTypefaceSet);
         this->setTypefacePlayback(&fTypefacePlayback);
     }
 
-    ~SkChunkFlatController() {
-        fTypefaceSet->unref();
-    }
-
     virtual void* allocThrow(size_t bytes) SK_OVERRIDE {
-        return fHeap.allocThrow(bytes);
+        fLastAllocated = fHeap.allocThrow(bytes);
+        return fLastAllocated;
     }
 
     virtual void unalloc(void* ptr) SK_OVERRIDE {
-        (void) fHeap.unalloc(ptr);
+        // fHeap can only free a pointer if it was the last one allocated.  Otherwise, we'll just
+        // have to wait until fHeap is destroyed.
+        if (ptr == fLastAllocated) (void)fHeap.unalloc(ptr);
     }
 
     void setupPlaybacks() const {
-        fTypefacePlayback.reset(fTypefaceSet);
+        fTypefacePlayback.reset(fTypefaceSet.get());
     }
 
     void setBitmapStorage(SkBitmapHeap* heap) {
         this->setBitmapHeap(heap);
     }
 
 private:
     SkChunkAlloc               fHeap;
-    SkRefCntSet*               fTypefaceSet;
+    SkAutoTUnref<SkRefCntSet>  fTypefaceSet;
+    void*                      fLastAllocated;

[scroggo, 2013/07/30 21:21:12]

This should never be a problem in practice, but shouldn't we set fLastAllocated
to NULL in the constructor?

     mutable SkTypefacePlayback fTypefacePlayback;
 };
 
-class SkBitmapDictionary : public SkFlatDictionary<SkBitmap> {
-public:
-    SkBitmapDictionary(SkFlatController* controller)
-    : SkFlatDictionary<SkBitmap>(controller) {
-        fFlattenProc = &SkFlattenObjectProc<SkBitmap>;
-        fUnflattenProc = &SkUnflattenObjectProc<SkBitmap>;
-    }
-};
-
 class SkMatrixDictionary : public SkFlatDictionary<SkMatrix> {
  public:
+    // All matrices fit in 36 bytes.
     SkMatrixDictionary(SkFlatController* controller)
-    : SkFlatDictionary<SkMatrix>(controller) {
+    : SkFlatDictionary<SkMatrix>(controller, 36) {
         fFlattenProc = &flattenMatrix;
         fUnflattenProc = &unflattenMatrix;
     }
 
     static void flattenMatrix(SkOrderedWriteBuffer& buffer, const void* obj) {
         buffer.getWriter32()->writeMatrix(*(SkMatrix*)obj);
     }
 
     static void unflattenMatrix(SkOrderedReadBuffer& buffer, void* obj) {
         buffer.getReader32()->readMatrix((SkMatrix*)obj);
     }
 };
 
 class SkPaintDictionary : public SkFlatDictionary<SkPaint> {
  public:
+    // The largest paint across ~60 .skps was 500 bytes.
     SkPaintDictionary(SkFlatController* controller)
-    : SkFlatDictionary<SkPaint>(controller) {
+    : SkFlatDictionary<SkPaint>(controller, 512) {
         fFlattenProc = &SkFlattenObjectProc<SkPaint>;
         fUnflattenProc = &SkUnflattenObjectProc<SkPaint>;
     }
 };
 
 class SkRegionDictionary : public SkFlatDictionary<SkRegion> {
  public:
     SkRegionDictionary(SkFlatController* controller)
     : SkFlatDictionary<SkRegion>(controller) {
         fFlattenProc = &flattenRegion;
         fUnflattenProc = &unflattenRegion;
     }
 
     static void flattenRegion(SkOrderedWriteBuffer& buffer, const void* obj) {
         buffer.getWriter32()->writeRegion(*(SkRegion*)obj);
     }
 
     static void unflattenRegion(SkOrderedReadBuffer& buffer, void* obj) {
         buffer.getReader32()->readRegion((SkRegion*)obj);
     }
 };
 
 #endif