SSE2 implementation of RGBA box blurs. This yields ~2X perf improvement on

src/opts/SkBlurImage_opts_SSE2.cpp

Index: src/opts/SkBlurImage_opts_SSE2.cpp
diff --git a/src/opts/SkBlurImage_opts_SSE2.cpp b/src/opts/SkBlurImage_opts_SSE2.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..3a4c2ff00b0aef63a989ff84df657f739cffded5
--- /dev/null
+++ b/src/opts/SkBlurImage_opts_SSE2.cpp
@@ -0,0 +1,91 @@
+/*
+ * Copyright 2013 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 "SkBitmap.h"
+#include "SkColorPriv.h"
+#include "SkBlurImage_opts_SSE2.h"
+#include "SkRect.h"
+
+#include <emmintrin.h>
+
+template<SkBlurDirection srcDirection, SkBlurDirection dstDirection>
+void SkBoxBlur_SSE2(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
+                         int leftOffset, int rightOffset, int width, int height)
+{
+    int rightBorder = SkMin32(rightOffset + 1, width);
+    int srcStrideX = srcDirection == kX_BlurDirection ? 1 : srcStride;
+    int dstStrideX = dstDirection == kX_BlurDirection ? 1 : height;
+    int srcStrideY = srcDirection == kX_BlurDirection ? srcStride : 1;
+    int dstStrideY = dstDirection == kX_BlurDirection ? width : 1;
+    __m128i scale = _mm_set1_epi32((1 << 24) / kernelSize);
+    __m128i half = _mm_set1_epi32(1 << 23);
+    __m128i zero = _mm_setzero_si128();
+    for (int y = 0; y < height; ++y) {
+        __m128i sum = zero;
+        const SkPMColor* p = src;
+        for (int i = 0; i < rightBorder; ++i) {
+            __m128i expanded = _mm_cvtsi32_si128(*p);
+            expanded = _mm_unpacklo_epi8(expanded, zero);
+            expanded = _mm_unpacklo_epi16(expanded, zero);
+            sum = _mm_add_epi32(sum, expanded);
+            p += srcStrideX;
+        }
+
+        const SkPMColor* sptr = src;
+        SkColor* dptr = dst;
+        for (int x = 0; x < width; ++x) {
+            __m128i tmp1 = _mm_mul_epu32(sum, scale);
+            __m128i tmp2 = _mm_mul_epu32(_mm_srli_si128(sum, 4),
+                                         _mm_srli_si128(scale, 4));
+            __m128i result = _mm_unpacklo_epi32(_mm_shuffle_epi32(tmp1, _MM_SHUFFLE(0,0,2,0)),
+                                                _mm_shuffle_epi32(tmp2, _MM_SHUFFLE(0,0,2,0)));
+            result = _mm_add_epi32(result, half);
+            result = _mm_srli_epi32(result, 24);
+            result = _mm_packs_epi32(result, zero);
+            result = _mm_packus_epi16(result, zero);
+            *dptr = _mm_cvtsi128_si32(result);
+            if (x >= leftOffset) {
+                SkColor l = *(sptr - leftOffset * srcStrideX);
+                __m128i expanded = _mm_cvtsi32_si128(l);
+                expanded = _mm_unpacklo_epi8(expanded, zero);
+                expanded = _mm_unpacklo_epi16(expanded, zero);
+                sum = _mm_sub_epi32(sum, expanded);
+            }
+            if (x + rightOffset + 1 < width) {
+                SkColor r = *(sptr + (rightOffset + 1) * srcStrideX);
+                __m128i expanded = _mm_cvtsi32_si128(r);
+                expanded = _mm_unpacklo_epi8(expanded, zero);
+                expanded = _mm_unpacklo_epi16(expanded, zero);
+                sum = _mm_add_epi32(sum, expanded);
+            }
+            sptr += srcStrideX;
+            if (srcDirection == kY_BlurDirection) {
+                SK_PREFETCH(sptr + (rightOffset + 1) * srcStrideX);

[mtklein, 2013/11/08 02:11:03]

This reminds me, >=xmmintrin.h has _mm_prefetch, suspiciously similar to
SK_PREFETCH but more portable (i.e. works on Windows).  I think the right thing
to do is actually update SK_PREFETCH to use _mm_prefetch when (SSE && !GCC), but
just wanted to mention it here in case you prefer to use it directly.

[Stephen White, 2013/11/08 15:47:47]

Cool! Done. (I think it's better to do it explicitly, since we don't compile
skia with SSE2 enabled for all files, so SK_PREFETCH couldn't use it without a
CPUID check).

[mtklein, 2013/11/08 18:42:33]

Right.  I was thinking rather the other way, that we just #ifdef __SSE__ and
only do _mm_prefetch in SK_PREFETCH when we know we've got SSE globally.

+            }
+            dptr += dstStrideX;
+        }
+        src += srcStrideY;
+        dst += dstStrideY;
+    }
+}
+
+template void SkBoxBlur_SSE2<kX_BlurDirection, kX_BlurDirection>(
+    const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
+    int leftOffset, int rightOffset, int width, int height);
+
+template void SkBoxBlur_SSE2<kX_BlurDirection, kY_BlurDirection>(
+    const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
+    int leftOffset, int rightOffset, int width, int height);
+
+template void SkBoxBlur_SSE2<kY_BlurDirection, kX_BlurDirection>(
+    const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
+    int leftOffset, int rightOffset, int width, int height);
+
+template void SkBoxBlur_SSE2<kY_BlurDirection, kY_BlurDirection>(
+    const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
+    int leftOffset, int rightOffset, int width, int height);