Rewriting MatrixConvolution image filter with SSE and AVX2

src/effects/SkMatrixConvolutionImageFilter.cpp

Index: src/effects/SkMatrixConvolutionImageFilter.cpp
diff --git a/src/effects/SkMatrixConvolutionImageFilter.cpp b/src/effects/SkMatrixConvolutionImageFilter.cpp
index 2a32fd5cdcebcfa9ac58e227a2af53c49104d419..20bee4ccd74f080cf06b1931c4531d60418a6026 100644
--- a/src/effects/SkMatrixConvolutionImageFilter.cpp
+++ b/src/effects/SkMatrixConvolutionImageFilter.cpp
@@ -13,6 +13,7 @@
 #include "SkWriteBuffer.h"
 #include "SkRect.h"
 #include "SkUnPreMultiply.h"
+#include "SkOpts.h"
 
 #if SK_SUPPORT_GPU
 #include "effects/GrMatrixConvolutionEffect.h"
@@ -117,13 +118,6 @@ SkMatrixConvolutionImageFilter::~SkMatrixConvolutionImageFilter() {
     delete[] fKernel;
 }
 
-class UncheckedPixelFetcher {
-public:
-    static inline SkPMColor fetch(const SkBitmap& src, int x, int y, const SkIRect& bounds) {
-        return *src.getAddr32(x, y);
-    }
-};
-
 class ClampPixelFetcher {
 public:
     static inline SkPMColor fetch(const SkBitmap& src, int x, int y, const SkIRect& bounds) {
@@ -159,7 +153,7 @@ public:
     }
 };
 
-template<class PixelFetcher, bool convolveAlpha>
+template<class PixelFetcher>
 void SkMatrixConvolutionImageFilter::filterPixels(const SkBitmap& src,
                                                   SkBitmap* result,
                                                   const SkIRect& r,
@@ -179,7 +173,7 @@ void SkMatrixConvolutionImageFilter::filterPixels(const SkBitmap& src,
                                                       y + cy - fKernelOffset.fY,
                                                       bounds);
                     SkScalar k = fKernel[cy * fKernelSize.fWidth + cx];
-                    if (convolveAlpha) {
+                    if (fConvolveAlpha) {

[Stephen White, 2016/04/12 20:01:58]

IIRC, this did actually have a measurable impact when I original wrote it
(hence, the template param). Is that no longer the case?

                         sumA += SkScalarMul(SkIntToScalar(SkGetPackedA32(s)), k);
                     }
                     sumR += SkScalarMul(SkIntToScalar(SkGetPackedR32(s)), k);
@@ -187,13 +181,13 @@ void SkMatrixConvolutionImageFilter::filterPixels(const SkBitmap& src,
                     sumB += SkScalarMul(SkIntToScalar(SkGetPackedB32(s)), k);
                 }
             }
-            int a = convolveAlpha
+            int a = fConvolveAlpha
                   ? SkClampMax(SkScalarFloorToInt(SkScalarMul(sumA, fGain) + fBias), 255)
                   : 255;
             int r = SkClampMax(SkScalarFloorToInt(SkScalarMul(sumR, fGain) + fBias), a);
             int g = SkClampMax(SkScalarFloorToInt(SkScalarMul(sumG, fGain) + fBias), a);
             int b = SkClampMax(SkScalarFloorToInt(SkScalarMul(sumB, fGain) + fBias), a);
-            if (!convolveAlpha) {
+            if (!fConvolveAlpha) {
                 a = SkGetPackedA32(PixelFetcher::fetch(src, x, y, bounds));
                 *dptr++ = SkPreMultiplyARGB(a, r, g, b);
             } else {
@@ -203,23 +197,22 @@ void SkMatrixConvolutionImageFilter::filterPixels(const SkBitmap& src,
     }
 }
 
-template<class PixelFetcher>
-void SkMatrixConvolutionImageFilter::filterPixels(const SkBitmap& src,
-                                                  SkBitmap* result,
-                                                  const SkIRect& rect,
-                                                  const SkIRect& bounds) const {
-    if (fConvolveAlpha) {
-        filterPixels<PixelFetcher, true>(src, result, rect, bounds);
-    } else {
-        filterPixels<PixelFetcher, false>(src, result, rect, bounds);
-    }
-}
 
 void SkMatrixConvolutionImageFilter::filterInteriorPixels(const SkBitmap& src,
                                                           SkBitmap* result,
                                                           const SkIRect& rect,
                                                           const SkIRect& bounds) const {
-    filterPixels<UncheckedPixelFetcher>(src, result, rect, bounds);
+    SkOpts::matrix_convolution_image_filter_filter_pixels(
+	src,

[Stephen White, 2016/04/12 20:01:58]

Please use spaces instead of tabs.

+	result,
+	rect,
+	bounds,
+	fConvolveAlpha,
+	fKernel,
+	fKernelSize,
+	fKernelOffset,
+	fGain,
+	fBias);
 }
 
 void SkMatrixConvolutionImageFilter::filterBorderPixels(const SkBitmap& src,