Rewriting MatrixConvolution image filter with SSE and AVX2

src/opts/SkMatrixConvolutionImageFilter_opts.h

+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef SkMatrixConvolutionImageFilter_opts_DEFINED
+#define SkMatrixConvolutionImageFilter_opts_DEFINED
+
+#include "SkScalar.h"
+#include "SkSize.h"
+#include "SkPoint.h"
+
+
+namespace SK_OPTS_NS {
+
+static void SkMatrixConvolutionImageFilter_filterPixels_none(
+                        const SkBitmap& src,
+                        SkBitmap* result,
+                        const SkIRect& r,
+                        const SkIRect& bounds,
+                        bool convolveAlpha,
+                        SkScalar* kernel,
+                        const SkISize& kernelSize,
+                        const SkIPoint&  kernelOffset,
+                        SkScalar gain,
+                        SkScalar bias)
+{
+    SkIRect rect(r);
+    if (!rect.intersect(bounds)) {
+        return;
+    }
+    for (int y = rect.fTop; y < rect.fBottom; ++y) {
+        SkPMColor* dptr = result->getAddr32(rect.fLeft - bounds.fLeft, y - bounds.fTop);
+        for (int x = rect.fLeft; x < rect.fRight; ++x) {
+            SkScalar sumA = 0, sumR = 0, sumG = 0, sumB = 0;
+            for (int cy = 0; cy < kernelSize.fHeight; cy++) {
+                for (int cx = 0; cx < kernelSize.fWidth; cx++) {
+                    SkPMColor s = *src.getAddr32(x + cx - kernelOffset.fX,

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

Rather than maintain a no-opts and SSE2 flavour, another way to do this might be
to use the Sk4f classes, which give SSE2 and NEON automatically. The Sk4f
version could be done directly in SkMatrixConvolutionImageFilter, and the AVX2
version could be switched at runtime.

+                                                 y + cy - kernelOffset.fY);
+                    SkScalar k = kernel[cy * kernelSize.fWidth + cx];
+
+                    // Calculate sumA in spite of convolveAlpha value. It's more
+                    // likely a compiler does SLP than if convolveAlpha is a
+                    // template argument. if alpha isn't really necessary
+                    // (convolveAlpha == false), case will be handled after the loop.
+                    sumA += SkScalarMul(SkIntToScalar(SkGetPackedA32(s)), k);
+                    sumR += SkScalarMul(SkIntToScalar(SkGetPackedR32(s)), k);
+                    sumG += SkScalarMul(SkIntToScalar(SkGetPackedG32(s)), k);
+                    sumB += SkScalarMul(SkIntToScalar(SkGetPackedB32(s)), k);
+                }
+            }
+            int a = convolveAlpha
+                  ? SkClampMax(SkScalarFloorToInt(SkScalarMul(sumA, gain) + bias), 255)
+                  : 255;
+            int r = SkClampMax(SkScalarFloorToInt(SkScalarMul(sumR, gain) + bias), a);
+            int g = SkClampMax(SkScalarFloorToInt(SkScalarMul(sumG, gain) + bias), a);
+            int b = SkClampMax(SkScalarFloorToInt(SkScalarMul(sumB, gain) + bias), a);
+            if (!convolveAlpha) {
+                a = SkGetPackedA32(*src.getAddr32(x, y));
+                *dptr++ = SkPreMultiplyARGB(a, r, g, b);
+            } else {
+                *dptr++ = SkPackARGB32(a, r, g, b);
+            }
+        }
+    }
+}
+
+}
+#endif