Optimize CSS box-shadow performance by caching the SkMask of the blur effect.

src/effects/SkBlurMask.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 "SkBlurMask.h"
 #include "SkMath.h"
+#include "SkScaledImageCache.h"
 #include "SkTemplates.h"
 #include "SkEndian.h"
 
 
 // This constant approximates the scaling done in the software path's
 // "high quality" mode, in SkBlurMask::Blur() (1 / sqrt(3)).
 // IMHO, it actually should be 1:  we blur "less" than we should do
 // according to the CSS and canvas specs, simply because Safari does the same.
 // Firefox used to do the same too, until 4.0 where they fixed it.  So at some
 // point we should probably get rid of these scaling constants and rebaseline
 // all the blur tests.
 static const SkScalar kBLUR_SIGMA_SCALE = 0.57735f;
+std::map<SkMask*, SkDiscardableMemoryMask*> SkBlurMask::fDiscardableMemoryMaskMap;
+
+SK_DECLARE_STATIC_MUTEX(gMapMutex);
+
+bool SkBlurMask::addDiscardableMemoryMaskToMap(SkMask* mask, SkDiscardableMemoryMask* dmMask) {
+    SkAutoMutexAcquire am(gMapMutex);
+    fDiscardableMemoryMaskMap.insert(std::pair<SkMask*, SkDiscardableMemoryMask*>(mask, dmMask));
+    return true;
+}
+
+bool SkBlurMask::removeDiscardableMemoryMaskFromMap(SkMask* mask) {
+    SkAutoMutexAcquire am(gMapMutex);
+    std::map<SkMask*, SkDiscardableMemoryMask*>::iterator it;
+    it = fDiscardableMemoryMaskMap.find(mask);
+    if (it != fDiscardableMemoryMaskMap.end())
+        fDiscardableMemoryMaskMap.erase(it);
+    return true;
+}
+
+SkDiscardableMemoryMask* SkBlurMask::getDiscardableMemoryMaskFromMap(SkMask* mask) {
+    SkAutoMutexAcquire am(gMapMutex);
+    std::map<SkMask*, SkDiscardableMemoryMask*>::iterator it;
+    it = fDiscardableMemoryMaskMap.find(mask);
+    if (it != fDiscardableMemoryMaskMap.end())
+        return it->second;
+    return NULL;
+}
 
 SkScalar SkBlurMask::ConvertRadiusToSigma(SkScalar radius) {
     return radius > 0 ? kBLUR_SIGMA_SCALE * radius + 0.5f : 0.0f;
 }
 
 SkScalar SkBlurMask::ConvertSigmaToRadius(SkScalar sigma) {
     return sigma > 0.5f ? (sigma - 0.5f) / kBLUR_SIGMA_SCALE : 0.0f;
 }
 
 #define UNROLL_SEPARABLE_LOOPS
@@ skipping 500 matching lines @@
 
     if (src.fImage) {
         size_t dstSize = dst->computeImageSize();
         if (0 == dstSize) {
             return false;   // too big to allocate, abort
         }
 
         int             sw = src.fBounds.width();
         int             sh = src.fBounds.height();
         const uint8_t*  sp = src.fImage;
-        uint8_t*        dp = SkMask::AllocImage(dstSize);
-        SkAutoTCallVProc<uint8_t, SkMask_FreeImage> autoCall(dp);
+        uint8_t* dp = NULL;
+        SkDiscardableMemoryMask* dm = SkBlurMask::getDiscardableMemoryMaskFromMap(dst);
+        if (dm) {
+            SkBitmap bitmap;
+            SkBitmap::Allocator* allocator = SkScaledImageCache::GetAllocator();
+            bitmap.setConfig(SkBitmap::kA8_Config, dst->fBounds.width(),
+                dst->fBounds.height(), dst->fRowBytes);
+            bitmap.allocPixels(allocator, NULL);
+            if (!bitmap.readyToDraw()) {
+                return false;
+            }
+            dm->fPixelGenerationID = bitmap.getGenerationID();
+            dm->fCacheId = SkScaledImageCache::AddAndLock(dm->fPixelGenerationID,
+                                                          bitmap.width(),
+                                                          bitmap.height(),
+                                                          bitmap);
+            SkASSERT(dm->fCacheId != NULL);
+            dp = static_cast<uint8_t*>(bitmap.getPixels());
+        } else
+            dp = SkMask::AllocImage(dstSize);
 
         // build the blurry destination
         SkAutoTMalloc<uint8_t>  tmpBuffer(dstSize);
         uint8_t*                tp = tmpBuffer.get();
         int w = sw, h = sh;
 
         if (outerWeight == 255) {
             int loRadius, hiRadius;
             get_adjusted_radii(passRadius, &loRadius, &hiRadius);
             if (kHigh_SkBlurQuality == quality) {
@@ skipping 25 matching lines @@
             }
         }
 
         dst->fImage = dp;
         // if need be, alloc the "real" dst (same size as src) and copy/merge
         // the blur into it (applying the src)
         if (style == kInner_SkBlurStyle) {
             // now we allocate the "real" dst, mirror the size of src
             size_t srcSize = src.computeImageSize();
             if (0 == srcSize) {
+                if (dm) {
+                    SkScaledImageCache::Unlock(static_cast<SkScaledImageCache::ID*>(dm->fCacheId));

[Stephen White, 2014/06/09 18:03:57]

This 5-line stanza is repeated 3 times. Could we refactor this into some kind of
guard class that does either scaled image caching or regular allocations, so we
don't lose the RAII? Just a thought.

+                } else if (dp) {
+                    SkMask::FreeImage(dp);
+                }
                 return false;   // too big to allocate, abort
             }
             dst->fImage = SkMask::AllocImage(srcSize);
             merge_src_with_blur(dst->fImage, src.fRowBytes,
                                 sp, src.fRowBytes,
                                 dp + passCount * (rx + ry * dst->fRowBytes),
                                 dst->fRowBytes, sw, sh);
-            SkMask::FreeImage(dp);
+            if (dm) {
+                SkScaledImageCache::Unlock(static_cast<SkScaledImageCache::ID*>(dm->fCacheId));
+            } else if (dp) {
+                SkMask::FreeImage(dp);
+            }
         } else if (style != kNormal_SkBlurStyle) {
             clamp_with_orig(dp + passCount * (rx + ry * dst->fRowBytes),
                             dst->fRowBytes, sp, src.fRowBytes, sw, sh, style);
         }
-        (void)autoCall.detach();
     }
 
     if (style == kInner_SkBlurStyle) {
         dst->fBounds = src.fBounds; // restore trimmed bounds
         dst->fRowBytes = src.fRowBytes;
     }
 
     return true;
 }
 
@@ skipping 158 matching lines @@
     uint8_t *profile = NULL;
 
     ComputeBlurProfile(sigma, &profile);
     SkAutoTDeleteArray<uint8_t> ada(profile);
 
     size_t dstSize = dst->computeImageSize();
     if (0 == dstSize) {
         return false;   // too big to allocate, abort
     }
 
-    uint8_t*        dp = SkMask::AllocImage(dstSize);
+    uint8_t* dp = NULL;
+    SkDiscardableMemoryMask* dm = SkBlurMask::getDiscardableMemoryMaskFromMap(dst);
+    if (dm) {
+        SkBitmap bitmap;
+        SkBitmap::Allocator* allocator = SkScaledImageCache::GetAllocator();
+        bitmap.setConfig(SkBitmap::kA8_Config, dst->fBounds.width(),
+            dst->fBounds.height(), dst->fRowBytes);
+        bitmap.allocPixels(allocator, NULL);
+        if (!bitmap.readyToDraw()) {
+            return false;
+        }
+        dm->fPixelGenerationID = bitmap.getGenerationID();
+        dm->fCacheId = SkScaledImageCache::AddAndLock(dm->fPixelGenerationID,
+                                                      bitmap.width(),
+                                                      bitmap.height(),
+                                                      bitmap);
+        SkASSERT(dm->fCacheId != NULL);
+        dp = static_cast<uint8_t*>(bitmap.getPixels());
+    } else
+        dp = SkMask::AllocImage(dstSize);
 
     dst->fImage = dp;
 
     int dstHeight = dst->fBounds.height();
     int dstWidth = dst->fBounds.width();
 
     uint8_t *outptr = dp;
 
     SkAutoTMalloc<uint8_t> horizontalScanline(dstWidth);
     SkAutoTMalloc<uint8_t> verticalScanline(dstHeight);
@@ skipping 13 matching lines @@
         size_t srcSize = (size_t)(src.width() * src.height());
         if (0 == srcSize) {
             return false;   // too big to allocate, abort
         }
         dst->fImage = SkMask::AllocImage(srcSize);
         for (int y = 0 ; y < sh ; y++) {
             uint8_t *blur_scanline = dp + (y+pad)*dstWidth + pad;
             uint8_t *inner_scanline = dst->fImage + y*sw;
             memcpy(inner_scanline, blur_scanline, sw);
         }
-        SkMask::FreeImage(dp);
+        if (dm) {
+            SkScaledImageCache::Unlock(static_cast<SkScaledImageCache::ID*>(dm->fCacheId));
+        } else if (dp) {
+            SkMask::FreeImage(dp);
+        }
 
         dst->fBounds.set(SkScalarRoundToInt(src.fLeft),
                          SkScalarRoundToInt(src.fTop),
                          SkScalarRoundToInt(src.fRight),
                          SkScalarRoundToInt(src.fBottom)); // restore trimmed bounds
         dst->fRowBytes = sw;
 
     } else if (style == kOuter_SkBlurStyle) {
         for (int y = pad ; y < dstHeight-pad ; y++) {
             uint8_t *dst_scanline = dp + y*dstWidth + pad;
@@ skipping 160 matching lines @@
         (void)autoCall.detach();
     }
 
     if (style == kInner_SkBlurStyle) {
         dst->fBounds = src.fBounds; // restore trimmed bounds
         dst->fRowBytes = src.fRowBytes;
     }
 
     return true;
 }