Optimize CSS box-shadow performance

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 "SkCachedData.h"
 #include "SkMath.h"
+#include "SkResourceCache.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
@@ skipping 450 matching lines @@
 
 ///////////////////////////////////////////////////////////////////////////////
 
 // we use a local function to wrap the class static method to work around
 // a bug in gcc98
 void SkMask_FreeImage(uint8_t* image);
 void SkMask_FreeImage(uint8_t* image) {
     SkMask::FreeImage(image);
 }
 
-bool SkBlurMask::BoxBlur(SkMask* dst, const SkMask& src,
+bool SkBlurMask::BoxBlur(SkMask* dst, SkCachedData** data, const SkMask& src,
                          SkScalar sigma, SkBlurStyle style, SkBlurQuality quality,
                          SkIPoint* margin, bool force_quality) {
 
     if (src.fFormat != SkMask::kA8_Format) {
         return false;
     }
 
     // Force high quality off for small radii (performance)
     if (!force_quality && sigma <= SkIntToScalar(2)) {
         quality = kLow_SkBlurQuality;
@@ skipping 41 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;
+        SkCachedData* tmpData = NULL;
+        if (NULL == data) {

[reed1, 2014/10/21 21:32:55]

we have to repeat this pattern pretty often. This kinda suggest a way to
encapsulate the mask and its allocator (alluded to that in an earlier review).
Did you find any pattern that might have worked for that?

+            dp = SkMask::AllocImage(dstSize);
+        } else {
+            tmpData = SkResourceCache::NewCachedData(dstSize);
+            dp = (uint8_t*)(tmpData->data());
+        }
 
         // 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 19 matching lines @@
                 h = boxBlurInterp(tp, h,             dp, ry, h, w, false, outerWeight);
                 h = boxBlurInterp(dp, h,             tp, ry, h, w, false, outerWeight);
                 h = boxBlurInterp(tp, h,             dp, ry, h, w, true, outerWeight);
             } else {
                 w = boxBlurInterp(sp, src.fRowBytes, tp, rx, w, h, true, outerWeight);
                 h = boxBlurInterp(tp, h,             dp, ry, h, w, true, outerWeight);
             }
         }
 
         dst->fImage = dp;
+        if (data) {
+            *data = tmpData;
+            (*data)->ref();
+        }
         // 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 (NULL == data) {
+                    SkMask::FreeImage(dp);
+                } else {
+                    tmpData->unref();
+                    (*data)->unref();
+                }
                 return false;   // too big to allocate, abort
             }
-            dst->fImage = SkMask::AllocImage(srcSize);
+            if (NULL == data) {
+                dst->fImage = SkMask::AllocImage(srcSize);
+            } else {
+                (*data)->unref();
+                (*data) = SkResourceCache::NewCachedData(srcSize);
+                dst->fImage = (uint8_t*)((*data)->data());
+            }
             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 (NULL == data) {
+                SkMask::FreeImage(dp);

[reed1, 2014/10/21 21:32:55]

We don't have the if/else pattern here. How is this balanced when data != NULL?
This is hard to see (for me).

+            }
         } 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 (data) {
+          tmpData->unref();
+        }
     }
 
     if (style == kInner_SkBlurStyle) {
         dst->fBounds = src.fBounds; // restore trimmed bounds
         dst->fRowBytes = src.fRowBytes;
     }
 
     return true;
 }
 
@@ skipping 116 matching lines @@
        if (profile_size <= sw) {
            pixels[x] = ProfileLookup(profile, x, width, w);
        } else {
            float span = float(sw)/(2*sigma);
            float giX = 1.5f - (x+.5f)/(2*sigma);
            pixels[x] = (uint8_t) (255 * (gaussianIntegral(giX) - gaussianIntegral(giX + span)));
        }
     }
 }
 
-bool SkBlurMask::BlurRect(SkScalar sigma, SkMask *dst,
+bool SkBlurMask::BlurRect(SkScalar sigma, SkMask *dst, SkCachedData** data,
                           const SkRect &src, SkBlurStyle style,
                           SkIPoint *margin, SkMask::CreateMode createMode) {
     int profile_size = SkScalarCeilToInt(6*sigma);
 
     int pad = profile_size/2;
     if (margin) {
         margin->set( pad, pad );
     }
 
     dst->fBounds.set(SkScalarRoundToInt(src.fLeft - pad),
@@ skipping 21 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;
+    SkCachedData* tmpData = NULL;
+    if (NULL == data) {
+        dp = SkMask::AllocImage(dstSize);
+    } else {
+        tmpData = SkResourceCache::NewCachedData(dstSize);
+        dp = (uint8_t*)(tmpData->data());
+    }
 
     dst->fImage = dp;
+    if (data) {
+        *data = tmpData;
+        (*data)->ref();
+    }
 
     int dstHeight = dst->fBounds.height();
     int dstWidth = dst->fBounds.width();
 
     uint8_t *outptr = dp;
 
     SkAutoTMalloc<uint8_t> horizontalScanline(dstWidth);
     SkAutoTMalloc<uint8_t> verticalScanline(dstHeight);
 
     ComputeBlurredScanline(horizontalScanline, profile, dstWidth, sigma);
     ComputeBlurredScanline(verticalScanline, profile, dstHeight, sigma);
 
     for (int y = 0 ; y < dstHeight ; ++y) {
         for (int x = 0 ; x < dstWidth ; x++) {
             unsigned int maskval = SkMulDiv255Round(horizontalScanline[x], verticalScanline[y]);
             *(outptr++) = maskval;
         }
     }
 
     if (style == kInner_SkBlurStyle) {
         // now we allocate the "real" dst, mirror the size of src
         size_t srcSize = (size_t)(src.width() * src.height());
         if (0 == srcSize) {
+            if (NULL == data) {
+                SkMask::FreeImage(dp);
+            } else {
+                tmpData->unref();
+                (*data)->unref();
+            }
             return false;   // too big to allocate, abort
         }
-        dst->fImage = SkMask::AllocImage(srcSize);
+        if (NULL == data) {
+            dst->fImage = SkMask::AllocImage(srcSize);
+        } else {
+            (*data)->unref();
+            (*data) = SkResourceCache::NewCachedData(srcSize);
+            dst->fImage = (uint8_t*)((*data)->data());
+        }
         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 (NULL == data) {
+            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;
             memset(dst_scanline, 0, sw);
         }
     } else if (style == kSolid_SkBlurStyle) {
         for (int y = pad ; y < dstHeight-pad ; y++) {
             uint8_t *dst_scanline = dp + y*dstWidth + pad;
             memset(dst_scanline, 0xff, sw);
         }
     }
+    if (data) {
+        tmpData->unref();
+    }
     // normal and solid styles are the same for analytic rect blurs, so don't
     // need to handle solid specially.
 
     return true;
 }
 
-bool SkBlurMask::BlurRRect(SkScalar sigma, SkMask *dst,
+bool SkBlurMask::BlurRRect(SkScalar sigma, SkMask *dst, SkCachedData** data,
                            const SkRRect &src, SkBlurStyle style,
                            SkIPoint *margin, SkMask::CreateMode createMode) {
     // Temporary for now -- always fail, should cause caller to fall back
     // to old path.  Plumbing just to land API and parallelize effort.
 
     return false;
 }
 
 // The "simple" blur is a direct implementation of separable convolution with a discrete
 // gaussian kernel.  It's "ground truth" in a sense; too slow to be used, but very
@@ skipping 135 matching lines @@
         (void)autoCall.detach();
     }
 
     if (style == kInner_SkBlurStyle) {
         dst->fBounds = src.fBounds; // restore trimmed bounds
         dst->fRowBytes = src.fRowBytes;
     }
 
     return true;
 }