Extract most of the mutable state of SkShader into a separate Context object.

src/effects/gradients/SkGradientShader.cpp

 /*
  * Copyright 2006 The Android Open Source Project
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include "SkGradientShaderPriv.h"
 #include "SkLinearGradient.h"
 #include "SkRadialGradient.h"
 #include "SkTwoPointRadialGradient.h"
 #include "SkTwoPointConicalGradient.h"
 #include "SkSweepGradient.h"
 
 SkGradientShaderBase::SkGradientShaderBase(const Descriptor& desc) {
     SkASSERT(desc.fCount > 1);
 
-    fCacheAlpha = 256;  // init to a value that paint.getAlpha() can't return
-
     fMapper = desc.fMapper;
     SkSafeRef(fMapper);
     fGradFlags = SkToU8(desc.fFlags);
 
     SkASSERT((unsigned)desc.fTileMode < SkShader::kTileModeCount);
     SkASSERT(SkShader::kTileModeCount == SK_ARRAY_COUNT(gTileProcs));
     fTileMode = desc.fTileMode;
     fTileProc = gTileProcs[desc.fTileMode];
 
-    fCache16 = fCache16Storage = NULL;
-    fCache32 = NULL;
-    fCache32PixelRef = NULL;
-
     /*  Note: we let the caller skip the first and/or last position.
         i.e. pos[0] = 0.3, pos[1] = 0.7
         In these cases, we insert dummy entries to ensure that the final data
         will be bracketed by [0, 1].
         i.e. our_pos[0] = 0, our_pos[1] = 0.3, our_pos[2] = 0.7, our_pos[3] = 1
 
         Thus colorCount (the caller's value, and fColorCount (our value) may
         differ by up to 2. In the above example:
             colorCount = 2
             fColorCount = 4
@@ skipping 94 matching lines @@
 
 static SkShader::TileMode unpack_mode(uint32_t packed) {
     return (SkShader::TileMode)(packed & 0xF);
 }
 
 static uint32_t unpack_flags(uint32_t packed) {
     return packed >> 4;
 }
 
 SkGradientShaderBase::SkGradientShaderBase(SkReadBuffer& buffer) : INHERITED(buffer) {
-    fCacheAlpha = 256;
-
     fMapper = buffer.readUnitMapper();
 
-    fCache16 = fCache16Storage = NULL;
-    fCache32 = NULL;
-    fCache32PixelRef = NULL;
-
     int colorCount = fColorCount = buffer.getArrayCount();
     if (colorCount > kColorStorageCount) {
         size_t allocSize = (sizeof(SkColor) + sizeof(SkPMColor) + sizeof(Rec)) * colorCount;
         if (buffer.validateAvailable(allocSize)) {
             fOrigColors = reinterpret_cast<SkColor*>(sk_malloc_throw(allocSize));
         } else {
             fOrigColors =  NULL;
             colorCount = fColorCount = 0;
         }
     } else {
@@ skipping 14 matching lines @@
         for (int i = 1; i < colorCount; i++) {
             recs[i].fPos = buffer.readInt();
             recs[i].fScale = buffer.readUInt();
         }
     }
     buffer.readMatrix(&fPtsToUnit);
     this->initCommon();
 }
 
 SkGradientShaderBase::~SkGradientShaderBase() {
-    if (fCache16Storage) {
-        sk_free(fCache16Storage);
-    }
-    SkSafeUnref(fCache32PixelRef);
     if (fOrigColors != fStorage) {
         sk_free(fOrigColors);
     }
     SkSafeUnref(fMapper);
 }
 
 void SkGradientShaderBase::initCommon() {
-    fFlags = 0;
     unsigned colorAlpha = 0xFF;
     for (int i = 0; i < fColorCount; i++) {
         colorAlpha &= SkColorGetA(fOrigColors[i]);
     }
     fColorsAreOpaque = colorAlpha == 0xFF;
 }
 
 void SkGradientShaderBase::flatten(SkWriteBuffer& buffer) const {
     this->INHERITED::flatten(buffer);
     buffer.writeFlattenable(fMapper);
     buffer.writeColorArray(fOrigColors, fColorCount);
     buffer.writeUInt(pack_mode_flags(fTileMode, fGradFlags));
     if (fColorCount > 2) {
         Rec* recs = fRecs;
         for (int i = 1; i < fColorCount; i++) {
             buffer.writeInt(recs[i].fPos);
             buffer.writeUInt(recs[i].fScale);
         }
     }
     buffer.writeMatrix(fPtsToUnit);
 }
 
 bool SkGradientShaderBase::isOpaque() const {
     return fColorsAreOpaque;
 }
 
-bool SkGradientShaderBase::setContext(const SkBitmap& device,
-                                 const SkPaint& paint,
-                                 const SkMatrix& matrix) {
-    if (!this->INHERITED::setContext(device, paint, matrix)) {
-        return false;
-    }
-
+SkGradientShaderBase::GradientShaderBaseContext::GradientShaderBaseContext(
+        const SkGradientShaderBase& shader, const SkBitmap& device,
+        const SkPaint& paint, const SkMatrix& matrix)
+    : INHERITED(shader, device, paint, matrix)
+    , fCache(shader.getCache(this->getPaintAlpha()))
+{
     const SkMatrix& inverse = this->getTotalInverse();
 
-    fDstToIndex.setConcat(fPtsToUnit, inverse);
+    fDstToIndex.setConcat(shader.fPtsToUnit, inverse);
+
     fDstToIndexProc = fDstToIndex.getMapXYProc();
-    fDstToIndexClass = (uint8_t)SkShader::ComputeMatrixClass(fDstToIndex);
+    fDstToIndexClass = (uint8_t)SkShader::Context::ComputeMatrixClass(fDstToIndex);
 
     // now convert our colors in to PMColors
     unsigned paintAlpha = this->getPaintAlpha();
 
     fFlags = this->INHERITED::getFlags();
-    if (fColorsAreOpaque && paintAlpha == 0xFF) {
+    if (shader.fColorsAreOpaque && paintAlpha == 0xFF) {
         fFlags |= kOpaqueAlpha_Flag;
     }
     // we can do span16 as long as our individual colors are opaque,
     // regardless of the paint's alpha
-    if (fColorsAreOpaque) {
+    if (shader.fColorsAreOpaque) {
         fFlags |= kHasSpan16_Flag;
     }
-
-    this->setCacheAlpha(paintAlpha);
-    return true;
-}
-
-void SkGradientShaderBase::setCacheAlpha(U8CPU alpha) const {
-    // if the new alpha differs from the previous time we were called, inval our cache
-    // this will trigger the cache to be rebuilt.
-    // we don't care about the first time, since the cache ptrs will already be NULL
-    if (fCacheAlpha != alpha) {
-        fCache16 = NULL;            // inval the cache
-        fCache32 = NULL;            // inval the cache
-        fCacheAlpha = alpha;        // record the new alpha
-        // inform our subclasses
-        if (fCache32PixelRef) {
-            fCache32PixelRef->notifyPixelsChanged();
-        }
-    }
 }
 
+SkGradientShaderBase::GradientShaderCache::GradientShaderCache(
+        U8CPU alpha, const SkGradientShaderBase& shader)
+    : fShader(shader)
+    , fCache16Inited(false)
+    , fCache32Inited(false)
+{
+    // Only initialize the cache in getCache16/32.
+    fCache16 = NULL;
+    fCache32 = NULL;
+    fCache16Storage = NULL;
+    fCache32PixelRef = NULL;
+    fCacheAlpha = alpha;
+}
+
+SkGradientShaderBase::GradientShaderCache::~GradientShaderCache() {
+    sk_free(fCache16Storage);
+    SkSafeUnref(fCache32PixelRef);
+}
+
 #define Fixed_To_Dot8(x)        (((x) + 0x80) >> 8)
 
 /** We take the original colors, not our premultiplied PMColors, since we can
     build a 16bit table as long as the original colors are opaque, even if the
     paint specifies a non-opaque alpha.
 */
-void SkGradientShaderBase::Build16bitCache(uint16_t cache[], SkColor c0, SkColor c1,
-                                      int count) {
+void SkGradientShaderBase::GradientShaderCache::Build16bitCache(
+        uint16_t cache[], SkColor c0, SkColor c1, int count) {
     SkASSERT(count > 1);
     SkASSERT(SkColorGetA(c0) == 0xFF);
     SkASSERT(SkColorGetA(c1) == 0xFF);
 
     SkFixed r = SkColorGetR(c0);
     SkFixed g = SkColorGetG(c0);
     SkFixed b = SkColorGetB(c0);
 
     SkFixed dr = SkIntToFixed(SkColorGetR(c1) - r) / (count - 1);
     SkFixed dg = SkIntToFixed(SkColorGetG(c1) - g) / (count - 1);
@@ skipping 27 matching lines @@
  *      2. change SkPackARGB32 to + its (a << SK_A32_SHIFT) value instead
  *         of using |
  *
  *  We chose #1 just because it was more localized.
  *  See http://code.google.com/p/skia/issues/detail?id=1113
  *
  *  The type SkUFixed encapsulate this need for unsigned, but logically Fixed.
  */
 typedef uint32_t SkUFixed;
 
-void SkGradientShaderBase::Build32bitCache(SkPMColor cache[], SkColor c0, SkColor c1,
-                                      int count, U8CPU paintAlpha, uint32_t gradFlags) {
+void SkGradientShaderBase::GradientShaderCache::Build32bitCache(
+        SkPMColor cache[], SkColor c0, SkColor c1,
+        int count, U8CPU paintAlpha, uint32_t gradFlags) {
     SkASSERT(count > 1);
 
     // need to apply paintAlpha to our two endpoints
     uint32_t a0 = SkMulDiv255Round(SkColorGetA(c0), paintAlpha);
     uint32_t a1 = SkMulDiv255Round(SkColorGetA(c1), paintAlpha);
 
 
     const bool interpInPremul = SkToBool(gradFlags &
                            SkGradientShader::kInterpolateColorsInPremul_Flag);
 
@@ skipping 122 matching lines @@
     if (6 == bits) {
         return (x << 10) | (x << 4) | (x >> 2);
     }
     if (8 == bits) {
         return (x << 8) | x;
     }
     sk_throw();
     return 0;
 }
 
-const uint16_t* SkGradientShaderBase::getCache16() const {
-    if (fCache16 == NULL) {
-        // double the count for dither entries
-        const int entryCount = kCache16Count * 2;
-        const size_t allocSize = sizeof(uint16_t) * entryCount;
-
-        if (fCache16Storage == NULL) { // set the storage and our working ptr
-            fCache16Storage = (uint16_t*)sk_malloc_throw(allocSize);
-        }
-        fCache16 = fCache16Storage;
-        if (fColorCount == 2) {
-            Build16bitCache(fCache16, fOrigColors[0], fOrigColors[1],
-                            kCache16Count);
-        } else {
-            Rec* rec = fRecs;
-            int prevIndex = 0;
-            for (int i = 1; i < fColorCount; i++) {
-                int nextIndex = SkFixedToFFFF(rec[i].fPos) >> kCache16Shift;
-                SkASSERT(nextIndex < kCache16Count);
-
-                if (nextIndex > prevIndex)
-                    Build16bitCache(fCache16 + prevIndex, fOrigColors[i-1], fOrigColors[i], nextIndex - prevIndex + 1);
-                prevIndex = nextIndex;
-            }
-        }
-
-        if (fMapper) {
-            fCache16Storage = (uint16_t*)sk_malloc_throw(allocSize);
-            uint16_t* linear = fCache16;         // just computed linear data
-            uint16_t* mapped = fCache16Storage;  // storage for mapped data
-            SkUnitMapper* map = fMapper;
-            for (int i = 0; i < kCache16Count; i++) {
-                int index = map->mapUnit16(bitsTo16(i, kCache16Bits)) >> kCache16Shift;
-                mapped[i] = linear[index];
-                mapped[i + kCache16Count] = linear[index + kCache16Count];
-            }
-            sk_free(fCache16);
-            fCache16 = fCache16Storage;
-        }
-    }
+const uint16_t* SkGradientShaderBase::GradientShaderCache::getCache16() {
+    SkOnce(&fCache16Inited, &fCache16Mutex, SkGradientShaderBase::GradientShaderCache::initCache16,
+           this);
+    SkASSERT(fCache16);
     return fCache16;
 }
 
-const SkPMColor* SkGradientShaderBase::getCache32() const {
-    if (fCache32 == NULL) {
-        SkImageInfo info;
-        info.fWidth = kCache32Count;
-        info.fHeight = 4;   // for our 4 dither rows
-        info.fAlphaType = kPremul_SkAlphaType;
-        info.fColorType = kPMColor_SkColorType;
+void SkGradientShaderBase::GradientShaderCache::initCache16(GradientShaderCache* cache) {
+    // double the count for dither entries
+    const int entryCount = kCache16Count * 2;
+    const size_t allocSize = sizeof(uint16_t) * entryCount;
 
-        if (NULL == fCache32PixelRef) {
-            fCache32PixelRef = SkMallocPixelRef::NewAllocate(info, 0, NULL);
-        }
-        fCache32 = (SkPMColor*)fCache32PixelRef->getAddr();
-        if (fColorCount == 2) {
-            Build32bitCache(fCache32, fOrigColors[0], fOrigColors[1],
-                            kCache32Count, fCacheAlpha, fGradFlags);
-        } else {
-            Rec* rec = fRecs;
-            int prevIndex = 0;
-            for (int i = 1; i < fColorCount; i++) {
-                int nextIndex = SkFixedToFFFF(rec[i].fPos) >> kCache32Shift;
-                SkASSERT(nextIndex < kCache32Count);
+    if (cache->fCache16Storage == NULL) { // set the storage and our working ptr
+        cache->fCache16Storage = (uint16_t*)sk_malloc_throw(allocSize);
+    }
+    cache->fCache16 = cache->fCache16Storage;
+    if (cache->fShader.fColorCount == 2) {
+        Build16bitCache(cache->fCache16, cache->fShader.fOrigColors[0],
+                        cache->fShader.fOrigColors[1], kCache16Count);
+    } else {
+        Rec* rec = cache->fShader.fRecs;
+        int prevIndex = 0;
+        for (int i = 1; i < cache->fShader.fColorCount; i++) {
+            int nextIndex = SkFixedToFFFF(rec[i].fPos) >> kCache16Shift;
+            SkASSERT(nextIndex < kCache16Count);
 
-                if (nextIndex > prevIndex)
-                    Build32bitCache(fCache32 + prevIndex, fOrigColors[i-1],
-                                    fOrigColors[i], nextIndex - prevIndex + 1,
-                                    fCacheAlpha, fGradFlags);
-                prevIndex = nextIndex;
-            }
-        }
-
-        if (fMapper) {
-            SkMallocPixelRef* newPR = SkMallocPixelRef::NewAllocate(info, 0, NULL);
-            SkPMColor* linear = fCache32;           // just computed linear data
-            SkPMColor* mapped = (SkPMColor*)newPR->getAddr();    // storage for mapped data
-            SkUnitMapper* map = fMapper;
-            for (int i = 0; i < kCache32Count; i++) {
-                int index = map->mapUnit16((i << 8) | i) >> 8;
-                mapped[i + kCache32Count*0] = linear[index + kCache32Count*0];
-                mapped[i + kCache32Count*1] = linear[index + kCache32Count*1];
-                mapped[i + kCache32Count*2] = linear[index + kCache32Count*2];
-                mapped[i + kCache32Count*3] = linear[index + kCache32Count*3];
-            }
-            fCache32PixelRef->unref();
-            fCache32PixelRef = newPR;
-            fCache32 = (SkPMColor*)newPR->getAddr();
+            if (nextIndex > prevIndex)
+                Build16bitCache(cache->fCache16 + prevIndex, cache->fShader.fOrigColors[i-1],
+                                cache->fShader.fOrigColors[i], nextIndex - prevIndex + 1);
+            prevIndex = nextIndex;
         }
     }
+
+    if (cache->fShader.fMapper) {
+        cache->fCache16Storage = (uint16_t*)sk_malloc_throw(allocSize);
+        uint16_t* linear = cache->fCache16;         // just computed linear data
+        uint16_t* mapped = cache->fCache16Storage;  // storage for mapped data
+        SkUnitMapper* map = cache->fShader.fMapper;
+        for (int i = 0; i < kCache16Count; i++) {
+            int index = map->mapUnit16(bitsTo16(i, kCache16Bits)) >> kCache16Shift;
+            mapped[i] = linear[index];
+            mapped[i + kCache16Count] = linear[index + kCache16Count];
+        }
+        sk_free(cache->fCache16);
+        cache->fCache16 = cache->fCache16Storage;
+    }
+}
+
+const SkPMColor* SkGradientShaderBase::GradientShaderCache::getCache32() {
+    SkOnce(&fCache32Inited, &fCache32Mutex, SkGradientShaderBase::GradientShaderCache::initCache32,
+           this);
+    SkASSERT(fCache32);
     return fCache32;
 }
 
+void SkGradientShaderBase::GradientShaderCache::initCache32(GradientShaderCache* cache) {
+    SkImageInfo info;
+    info.fWidth = kCache32Count;
+    info.fHeight = 4;   // for our 4 dither rows
+    info.fAlphaType = kPremul_SkAlphaType;
+    info.fColorType = kPMColor_SkColorType;
+
+    if (NULL == cache->fCache32PixelRef) {
+        cache->fCache32PixelRef = SkMallocPixelRef::NewAllocate(info, 0, NULL);
+    }
+    cache->fCache32 = (SkPMColor*)cache->fCache32PixelRef->getAddr();
+    if (cache->fShader.fColorCount == 2) {
+        Build32bitCache(cache->fCache32, cache->fShader.fOrigColors[0],
+                        cache->fShader.fOrigColors[1], kCache32Count, cache->fCacheAlpha,
+                        cache->fShader.fGradFlags);
+    } else {
+        Rec* rec = cache->fShader.fRecs;
+        int prevIndex = 0;
+        for (int i = 1; i < cache->fShader.fColorCount; i++) {
+            int nextIndex = SkFixedToFFFF(rec[i].fPos) >> kCache32Shift;
+            SkASSERT(nextIndex < kCache32Count);
+
+            if (nextIndex > prevIndex)
+                Build32bitCache(cache->fCache32 + prevIndex, cache->fShader.fOrigColors[i-1],
+                                cache->fShader.fOrigColors[i], nextIndex - prevIndex + 1,
+                                cache->fCacheAlpha, cache->fShader.fGradFlags);
+            prevIndex = nextIndex;
+        }
+    }
+
+    if (cache->fShader.fMapper) {
+        SkMallocPixelRef* newPR = SkMallocPixelRef::NewAllocate(info, 0, NULL);
+        SkPMColor* linear = cache->fCache32;           // just computed linear data
+        SkPMColor* mapped = (SkPMColor*)newPR->getAddr();    // storage for mapped data
+        SkUnitMapper* map = cache->fShader.fMapper;
+        for (int i = 0; i < kCache32Count; i++) {
+            int index = map->mapUnit16((i << 8) | i) >> 8;
+            mapped[i + kCache32Count*0] = linear[index + kCache32Count*0];
+            mapped[i + kCache32Count*1] = linear[index + kCache32Count*1];
+            mapped[i + kCache32Count*2] = linear[index + kCache32Count*2];
+            mapped[i + kCache32Count*3] = linear[index + kCache32Count*3];
+        }
+        cache->fCache32PixelRef->unref();
+        cache->fCache32PixelRef = newPR;
+        cache->fCache32 = (SkPMColor*)newPR->getAddr();
+    }
+}
+
 /*
+ *  The gradient holds a cache for the most recent value of alpha. Successive
+ *  callers with the same alpha value will share the same cache.
+ */
+SkGradientShaderBase::GradientShaderCache* SkGradientShaderBase::getCache(U8CPU alpha) const {
+    SK_DECLARE_STATIC_MUTEX(cacheMutex);

[scroggo, 2014/04/14 21:11:13]

Does this need to be static?

[Dominik Grewe, 2014/04/15 11:46:32]

No. Made it a class member.

+    SkAutoMutexAcquire ama(cacheMutex);
+    if (!fCache || fCache->getAlpha() != alpha) {

[scroggo, 2014/04/14 21:11:13]

Does "!fCache" do the right thing here? Or do you need to call "!fCache.get()"?

[Dominik Grewe, 2014/04/15 11:46:32]

It seems to be doing the right thing. Stepping through it in GDB you can see
that fCache gets converted to a pointer using
operator T*() { return fObj; }
because it doesn't have it's own "operator!".

+        fCache.reset(SkNEW_ARGS(GradientShaderCache, (alpha, *this)));
+    }
+    fCache.get()->ref();

[scroggo, 2014/04/14 21:11:13]

As in the picture shader, it would be nice to add a comment explaining why the
function needs to call ref, rather than the caller.

[Dominik Grewe, 2014/04/15 11:46:32]

Done.

+    return fCache;
+}
+
+/*
  *  Because our caller might rebuild the same (logically the same) gradient
  *  over and over, we'd like to return exactly the same "bitmap" if possible,
  *  allowing the client to utilize a cache of our bitmap (e.g. with a GPU).
  *  To do that, we maintain a private cache of built-bitmaps, based on our
  *  colors and positions. Note: we don't try to flatten the fMapper, so if one
  *  is present, we skip the cache for now.
  */
 void SkGradientShaderBase::getGradientTableBitmap(SkBitmap* bitmap) const {
     // our caller assumes no external alpha, so we ensure that our cache is
     // built with 0xFF
-    this->setCacheAlpha(0xFF);
+    GradientShaderCache* cache = this->getCache(0xFF);
 
     // don't have a way to put the mapper into our cache-key yet
     if (fMapper) {
-        // force our cahce32pixelref to be built
-        (void)this->getCache32();
+        // force our cache32pixelref to be built
+        (void)cache->getCache32();
         bitmap->setConfig(SkImageInfo::MakeN32Premul(kCache32Count, 1));
-        bitmap->setPixelRef(fCache32PixelRef);
+        bitmap->setPixelRef(cache->getCache32PixelRef());
         return;
     }
 
     // build our key: [numColors + colors[] + {positions[]} + flags ]
     int count = 1 + fColorCount + 1;
     if (fColorCount > 2) {
         count += fColorCount - 1;    // fRecs[].fPos
     }
 
     SkAutoSTMalloc<16, int32_t> storage(count);
@@ skipping 18 matching lines @@
     static const int MAX_NUM_CACHED_GRADIENT_BITMAPS = 32;
     SkAutoMutexAcquire ama(gMutex);
 
     if (NULL == gCache) {
         gCache = SkNEW_ARGS(SkBitmapCache, (MAX_NUM_CACHED_GRADIENT_BITMAPS));
     }
     size_t size = count * sizeof(int32_t);
 
     if (!gCache->find(storage.get(), size, bitmap)) {
         // force our cahce32pixelref to be built
-        (void)this->getCache32();
+        (void)cache->getCache32();
         bitmap->setConfig(SkImageInfo::MakeN32Premul(kCache32Count, 1));
-        bitmap->setPixelRef(fCache32PixelRef);
+        bitmap->setPixelRef(cache->getCache32PixelRef());
 
         gCache->add(storage.get(), size, *bitmap);
     }
 }
 
 void SkGradientShaderBase::commonAsAGradient(GradientInfo* info) const {
     if (info) {
         if (info->fColorCount >= fColorCount) {
             if (info->fColors) {
                 memcpy(info->fColors, fOrigColors, fColorCount * sizeof(SkColor));
@@ skipping 499 matching lines @@
             (*stops)[i] = stop;
             stop = i < outColors - 1 ? stop + random->nextUScalar1() * (1.f - stop) : 1.f;
         }
     }
     *tm = static_cast<SkShader::TileMode>(random->nextULessThan(SkShader::kTileModeCount));
 
     return outColors;
 }
 
 #endif