Xfermode: SSE2 implementation of multiply_modeproc

This patch implements basics for Xfermode SSE optimization. Based on
these basics, SSE2 implementation of multiply_modeproc is provided. SSE2
implementation for other modes will come in future. With this patch
performance of Xfermode_Multiply will improve about 45%. Here are the
data on desktop i7-3770.
before:
Xfermode_Multiply   8888:  cmsecs =     33.30   565:  cmsecs =     45.65
after:
Xfermode_Multiply   8888:  cmsecs =     17.18   565:  cmsecs =     24.87

BUG=

Committed: http://code.google.com/p/skia/source/detail?r=14006

Committed: http://code.google.com/p/skia/source/detail?r=14050

R=mtklein@google.com, robertphillips@google.com

Author: qiankun.miao@intel.com

Review URL: https://codereview.chromium.org/202903004

git-svn-id: http://skia.googlecode.com/svn/trunk@14107 2bbb7eff-a529-9590-31e7-b0007b416f81

6 files changed, 473 insertions, 28 deletions

diff --git a/src/core/SkXfermode.cpp b/src/core/SkXfermode.cpp
index 78b85daa3e..7914fdc034 100644
--- a/src/core/SkXfermode.cpp
+++ b/src/core/SkXfermode.cpp
@@ -8,6 +8,7 @@
 
 
 #include "SkXfermode.h"
+#include "SkXfermode_opts_SSE2.h"
 #include "SkXfermode_proccoeff.h"
 #include "SkColorPriv.h"
 #include "SkMathPriv.h"
@@ -1996,4 +1997,5 @@ SK_DEFINE_FLATTENABLE_REGISTRAR_GROUP_START(SkXfermode)
 #if !SK_ARM_NEON_IS_NONE
     SK_DEFINE_FLATTENABLE_REGISTRAR_ENTRY(SkNEONProcCoeffXfermode)
 #endif
+    SK_DEFINE_FLATTENABLE_REGISTRAR_ENTRY(SkSSE2ProcCoeffXfermode)
 SK_DEFINE_FLATTENABLE_REGISTRAR_GROUP_END
diff --git a/src/opts/SkBlitRow_opts_SSE2.cpp b/src/opts/SkBlitRow_opts_SSE2.cpp
index 4aa08e8165..d1474f4a67 100644
--- a/src/opts/SkBlitRow_opts_SSE2.cpp
+++ b/src/opts/SkBlitRow_opts_SSE2.cpp
@@ -914,7 +914,7 @@ void S32_D565_Opaque_SSE2(uint16_t* SK_RESTRICT dst,
             __m128i b = _mm_packs_epi32(b1, b2);
 
             // Store 8 16-bit colors in dst.
-            __m128i d_pixel = SkPackRGB16_SSE(r, g, b);
+            __m128i d_pixel = SkPackRGB16_SSE2(r, g, b);
             _mm_store_si128(d++, d_pixel);
             count -= 8;
         }
@@ -983,54 +983,54 @@ void S32A_D565_Opaque_SSE2(uint16_t* SK_RESTRICT dst,
             __m128i dst_pixel = _mm_load_si128(d);
 
             // Extract A from src.
-            __m128i sa1 = _mm_slli_epi32(src_pixel1,(24 - SK_A32_SHIFT));
+            __m128i sa1 = _mm_slli_epi32(src_pixel1, (24 - SK_A32_SHIFT));
             sa1 = _mm_srli_epi32(sa1, 24);
-            __m128i sa2 = _mm_slli_epi32(src_pixel2,(24 - SK_A32_SHIFT));
+            __m128i sa2 = _mm_slli_epi32(src_pixel2, (24 - SK_A32_SHIFT));
             sa2 = _mm_srli_epi32(sa2, 24);
             __m128i sa = _mm_packs_epi32(sa1, sa2);
 
             // Extract R from src.
-            __m128i sr1 = _mm_slli_epi32(src_pixel1,(24 - SK_R32_SHIFT));
+            __m128i sr1 = _mm_slli_epi32(src_pixel1, (24 - SK_R32_SHIFT));
             sr1 = _mm_srli_epi32(sr1, 24);
-            __m128i sr2 = _mm_slli_epi32(src_pixel2,(24 - SK_R32_SHIFT));
+            __m128i sr2 = _mm_slli_epi32(src_pixel2, (24 - SK_R32_SHIFT));
             sr2 = _mm_srli_epi32(sr2, 24);
             __m128i sr = _mm_packs_epi32(sr1, sr2);
 
             // Extract G from src.
-            __m128i sg1 = _mm_slli_epi32(src_pixel1,(24 - SK_G32_SHIFT));
+            __m128i sg1 = _mm_slli_epi32(src_pixel1, (24 - SK_G32_SHIFT));
             sg1 = _mm_srli_epi32(sg1, 24);
-            __m128i sg2 = _mm_slli_epi32(src_pixel2,(24 - SK_G32_SHIFT));
+            __m128i sg2 = _mm_slli_epi32(src_pixel2, (24 - SK_G32_SHIFT));
             sg2 = _mm_srli_epi32(sg2, 24);
             __m128i sg = _mm_packs_epi32(sg1, sg2);
 
             // Extract B from src.
-            __m128i sb1 = _mm_slli_epi32(src_pixel1,(24 - SK_B32_SHIFT));
+            __m128i sb1 = _mm_slli_epi32(src_pixel1, (24 - SK_B32_SHIFT));
             sb1 = _mm_srli_epi32(sb1, 24);
-            __m128i sb2 = _mm_slli_epi32(src_pixel2,(24 - SK_B32_SHIFT));
+            __m128i sb2 = _mm_slli_epi32(src_pixel2, (24 - SK_B32_SHIFT));
             sb2 = _mm_srli_epi32(sb2, 24);
             __m128i sb = _mm_packs_epi32(sb1, sb2);
 
             // Extract R G B from dst.
-            __m128i dr = _mm_srli_epi16(dst_pixel,SK_R16_SHIFT);
+            __m128i dr = _mm_srli_epi16(dst_pixel, SK_R16_SHIFT);
             dr = _mm_and_si128(dr, r16_mask);
-            __m128i dg = _mm_srli_epi16(dst_pixel,SK_G16_SHIFT);
+            __m128i dg = _mm_srli_epi16(dst_pixel, SK_G16_SHIFT);
             dg = _mm_and_si128(dg, g16_mask);
-            __m128i db = _mm_srli_epi16(dst_pixel,SK_B16_SHIFT);
+            __m128i db = _mm_srli_epi16(dst_pixel, SK_B16_SHIFT);
             db = _mm_and_si128(db, b16_mask);
 
             __m128i isa = _mm_sub_epi16(var255, sa); // 255 -sa
 
             // Calculate R G B of result.
             // Original algorithm is in SkSrcOver32To16().
-            dr = _mm_add_epi16(sr, SkMul16ShiftRound_SSE(dr, isa, SK_R16_BITS));
+            dr = _mm_add_epi16(sr, SkMul16ShiftRound_SSE2(dr, isa, SK_R16_BITS));
             dr = _mm_srli_epi16(dr, 8 - SK_R16_BITS);
-            dg = _mm_add_epi16(sg, SkMul16ShiftRound_SSE(dg, isa, SK_G16_BITS));
+            dg = _mm_add_epi16(sg, SkMul16ShiftRound_SSE2(dg, isa, SK_G16_BITS));
             dg = _mm_srli_epi16(dg, 8 - SK_G16_BITS);
-            db = _mm_add_epi16(sb, SkMul16ShiftRound_SSE(db, isa, SK_B16_BITS));
+            db = _mm_add_epi16(sb, SkMul16ShiftRound_SSE2(db, isa, SK_B16_BITS));
             db = _mm_srli_epi16(db, 8 - SK_B16_BITS);
 
             // Pack R G B into 16-bit color.
-            __m128i d_pixel = SkPackRGB16_SSE(dr, dg, db);
+            __m128i d_pixel = SkPackRGB16_SSE2(dr, dg, db);
 
             // Store 8 16-bit colors in dst.
             _mm_store_si128(d++, d_pixel);
@@ -1143,7 +1143,7 @@ void S32_D565_Opaque_Dither_SSE2(uint16_t* SK_RESTRICT dst,
             sb = _mm_srli_epi16(sb, SK_B32_BITS - SK_B16_BITS);
 
             // Pack and store 16-bit dst pixel.
-            __m128i d_pixel = SkPackRGB16_SSE(sr, sg, sb);
+            __m128i d_pixel = SkPackRGB16_SSE2(sr, sg, sb);
             _mm_store_si128(d++, d_pixel);
 
             count -= 8;
@@ -1242,9 +1242,9 @@ void S32A_D565_Opaque_Dither_SSE2(uint16_t* SK_RESTRICT dst,
             __m128i dst_pixel = _mm_load_si128(d);
 
             // Extract A from src.
-            __m128i sa1 = _mm_slli_epi32(src_pixel1,(24 - SK_A32_SHIFT));
+            __m128i sa1 = _mm_slli_epi32(src_pixel1, (24 - SK_A32_SHIFT));
             sa1 = _mm_srli_epi32(sa1, 24);
-            __m128i sa2 = _mm_slli_epi32(src_pixel2,(24 - SK_A32_SHIFT));
+            __m128i sa2 = _mm_slli_epi32(src_pixel2, (24 - SK_A32_SHIFT));
             sa2 = _mm_srli_epi32(sa2, 24);
             __m128i sa = _mm_packs_epi32(sa1, sa2);
 
@@ -1323,7 +1323,7 @@ void S32A_D565_Opaque_Dither_SSE2(uint16_t* SK_RESTRICT dst,
             db = _mm_srli_epi16(db, 5);
 
             // Package and store dst pixel.
-            __m128i d_pixel = SkPackRGB16_SSE(dr, dg, db);
+            __m128i d_pixel = SkPackRGB16_SSE2(dr, dg, db);
             _mm_store_si128(d++, d_pixel);
 
             count -= 8;
diff --git a/src/opts/SkColor_opts_SSE2.h b/src/opts/SkColor_opts_SSE2.h
index 13a5be5655..24ab6f73c8 100644
--- a/src/opts/SkColor_opts_SSE2.h
+++ b/src/opts/SkColor_opts_SSE2.h
@@ -10,7 +10,39 @@
 
 #include <emmintrin.h>
 
-static inline __m128i SkMul16ShiftRound_SSE(__m128i a, __m128i b, int shift) {
+// See #define SkAlphaMulAlpha(a, b)  SkMulDiv255Round(a, b) in SkXfermode.cpp.
+static inline __m128i SkAlphaMulAlpha_SSE2(const __m128i& a,
+                                           const __m128i& b) {
+    __m128i prod = _mm_mullo_epi16(a, b);
+    prod = _mm_add_epi32(prod, _mm_set1_epi32(128));
+    prod = _mm_add_epi32(prod, _mm_srli_epi32(prod, 8));
+    prod = _mm_srli_epi32(prod, 8);
+
+    return prod;
+}
+
+static inline __m128i SkGetPackedA32_SSE2(const __m128i& src) {
+    __m128i a = _mm_slli_epi32(src, (24 - SK_A32_SHIFT));
+    return _mm_srli_epi32(a, 24);
+}
+
+static inline __m128i SkGetPackedR32_SSE2(const __m128i& src) {
+    __m128i r = _mm_slli_epi32(src, (24 - SK_R32_SHIFT));
+    return _mm_srli_epi32(r, 24);
+}
+
+static inline __m128i SkGetPackedG32_SSE2(const __m128i& src) {
+    __m128i g = _mm_slli_epi32(src, (24 - SK_G32_SHIFT));
+    return _mm_srli_epi32(g, 24);
+}
+
+static inline __m128i SkGetPackedB32_SSE2(const __m128i& src) {
+    __m128i b = _mm_slli_epi32(src, (24 - SK_B32_SHIFT));
+    return _mm_srli_epi32(b, 24);
+}
+
+static inline __m128i SkMul16ShiftRound_SSE2(const __m128i& a,
+                                             const __m128i& b, int shift) {
     __m128i prod = _mm_mullo_epi16(a, b);
     prod = _mm_add_epi16(prod, _mm_set1_epi16(1 << (shift - 1)));
     prod = _mm_add_epi16(prod, _mm_srli_epi16(prod, shift));
@@ -19,13 +51,96 @@ static inline __m128i SkMul16ShiftRound_SSE(__m128i a, __m128i b, int shift) {
     return prod;
 }
 
-static inline __m128i SkPackRGB16_SSE(__m128i r, __m128i g, __m128i b) {
-    r = _mm_slli_epi16(r, SK_R16_SHIFT);
-    g = _mm_slli_epi16(g, SK_G16_SHIFT);
-    b = _mm_slli_epi16(b, SK_B16_SHIFT);
+static inline __m128i SkPackRGB16_SSE2(const __m128i& r,
+                                       const __m128i& g, const __m128i& b) {
+    __m128i dr = _mm_slli_epi16(r, SK_R16_SHIFT);
+    __m128i dg = _mm_slli_epi16(g, SK_G16_SHIFT);
+    __m128i db = _mm_slli_epi16(b, SK_B16_SHIFT);
+
+    __m128i c = _mm_or_si128(dr, dg);
+    return _mm_or_si128(c, db);
+}
+
+static inline __m128i SkPackARGB32_SSE2(const __m128i& a, const __m128i& r,
+                                        const __m128i& g, const __m128i& b) {
+    __m128i da = _mm_slli_epi32(a, SK_A32_SHIFT);
+    __m128i dr = _mm_slli_epi32(r, SK_R32_SHIFT);
+    __m128i dg = _mm_slli_epi32(g, SK_G32_SHIFT);
+    __m128i db = _mm_slli_epi32(b, SK_B32_SHIFT);
+
+    __m128i c = _mm_or_si128(da, dr);
+    c = _mm_or_si128(c, dg);
+    return _mm_or_si128(c, db);
+}
+
+static inline __m128i SkPacked16ToR32_SSE2(const __m128i& src) {
+    __m128i r = _mm_srli_epi32(src, SK_R16_SHIFT);
+    r = _mm_and_si128(r, _mm_set1_epi32(SK_R16_MASK));
+    r = _mm_or_si128(_mm_slli_epi32(r, (8 - SK_R16_BITS)),
+                     _mm_srli_epi32(r, (2 * SK_R16_BITS - 8)));
+
+    return r;
+}
+
+static inline __m128i SkPacked16ToG32_SSE2(const __m128i& src) {
+    __m128i g = _mm_srli_epi32(src, SK_G16_SHIFT);
+    g = _mm_and_si128(g, _mm_set1_epi32(SK_G16_MASK));
+    g = _mm_or_si128(_mm_slli_epi32(g, (8 - SK_G16_BITS)),
+                     _mm_srli_epi32(g, (2 * SK_G16_BITS - 8)));
+
+    return g;
+}
+
+static inline __m128i SkPacked16ToB32_SSE2(const __m128i& src) {
+    __m128i b = _mm_srli_epi32(src, SK_B16_SHIFT);
+    b = _mm_and_si128(b, _mm_set1_epi32(SK_B16_MASK));
+    b = _mm_or_si128(_mm_slli_epi32(b, (8 - SK_B16_BITS)),
+                     _mm_srli_epi32(b, (2 * SK_B16_BITS - 8)));
+
+    return b;
+}
+
+static inline __m128i SkPixel16ToPixel32_SSE2(const __m128i& src) {
+    __m128i r = SkPacked16ToR32_SSE2(src);
+    __m128i g = SkPacked16ToG32_SSE2(src);
+    __m128i b = SkPacked16ToB32_SSE2(src);
+
+    return SkPackARGB32_SSE2(_mm_set1_epi32(0xFF), r, g, b);
+}
+
+static inline __m128i SkPixel32ToPixel16_ToU16_SSE2(const __m128i& src_pixel1,
+                                                    const __m128i& src_pixel2) {
+    // Calculate result r.
+    __m128i r1 = _mm_srli_epi32(src_pixel1,
+                                SK_R32_SHIFT + (8 - SK_R16_BITS));
+    r1 = _mm_and_si128(r1, _mm_set1_epi32(SK_R16_MASK));
+    __m128i r2 = _mm_srli_epi32(src_pixel2,
+                                SK_R32_SHIFT + (8 - SK_R16_BITS));
+    r2 = _mm_and_si128(r2, _mm_set1_epi32(SK_R16_MASK));
+    __m128i r = _mm_packs_epi32(r1, r2);
+
+    // Calculate result g.
+    __m128i g1 = _mm_srli_epi32(src_pixel1,
+                                SK_G32_SHIFT + (8 - SK_G16_BITS));
+    g1 = _mm_and_si128(g1, _mm_set1_epi32(SK_G16_MASK));
+    __m128i g2 = _mm_srli_epi32(src_pixel2,
+                                SK_G32_SHIFT + (8 - SK_G16_BITS));
+    g2 = _mm_and_si128(g2, _mm_set1_epi32(SK_G16_MASK));
+    __m128i g = _mm_packs_epi32(g1, g2);
+
+    // Calculate result b.
+    __m128i b1 = _mm_srli_epi32(src_pixel1,
+                                SK_B32_SHIFT + (8 - SK_B16_BITS));
+    b1 = _mm_and_si128(b1, _mm_set1_epi32(SK_B16_MASK));
+    __m128i b2 = _mm_srli_epi32(src_pixel2,
+                                SK_B32_SHIFT + (8 - SK_B16_BITS));
+    b2 = _mm_and_si128(b2, _mm_set1_epi32(SK_B16_MASK));
+    __m128i b = _mm_packs_epi32(b1, b2);
+
+    // Store 8 16-bit colors in dst.
+    __m128i d_pixel = SkPackRGB16_SSE2(r, g, b);
 
-    __m128i c = _mm_or_si128(r, g);
-    return _mm_or_si128(c, b);
+    return d_pixel;
 }
 
-#endif//SkColor_opts_SSE2_DEFINED
+#endif // SkColor_opts_SSE2_DEFINED
diff --git a/src/opts/SkXfermode_opts_SSE2.cpp b/src/opts/SkXfermode_opts_SSE2.cpp
new file mode 100644
index 0000000000..01c782b9ef
--- /dev/null
+++ b/src/opts/SkXfermode_opts_SSE2.cpp
@@ -0,0 +1,267 @@
+#include "SkColorPriv.h"
+#include "SkColor_opts_SSE2.h"
+#include "SkMathPriv.h"
+#include "SkXfermode.h"
+#include "SkXfermode_opts_SSE2.h"
+#include "SkXfermode_proccoeff.h"
+
+////////////////////////////////////////////////////////////////////////////////
+// 4 pixels SSE2 version functions
+////////////////////////////////////////////////////////////////////////////////
+
+static inline __m128i SkDiv255Round_SSE2(const __m128i& a) {
+    __m128i prod = _mm_add_epi32(a, _mm_set1_epi32(128)); // prod += 128;
+    prod = _mm_add_epi32(prod, _mm_srli_epi32(prod, 8));  // prod + (prod >> 8)
+    prod = _mm_srli_epi32(prod, 8);                       // >> 8
+
+    return prod;
+}
+
+static inline __m128i clamp_div255round_SSE2(const __m128i& prod) {
+    // test if > 0
+    __m128i cmp1 = _mm_cmpgt_epi32(prod, _mm_setzero_si128());
+    // test if < 255*255
+    __m128i cmp2 = _mm_cmplt_epi32(prod, _mm_set1_epi32(255*255));
+
+    __m128i ret = _mm_setzero_si128();
+
+    // if value >= 255*255, value = 255
+    ret = _mm_andnot_si128(cmp2,  _mm_set1_epi32(255));
+
+    __m128i div = SkDiv255Round_SSE2(prod);
+
+    // test if > 0 && < 255*255
+    __m128i cmp = _mm_and_si128(cmp1, cmp2);
+
+    ret = _mm_or_si128(_mm_and_si128(cmp, div), _mm_andnot_si128(cmp, ret));
+
+    return ret;
+}
+
+static inline __m128i srcover_byte_SSE2(const __m128i& a, const __m128i& b) {
+    // a + b - SkAlphaMulAlpha(a, b);
+    return _mm_sub_epi32(_mm_add_epi32(a, b), SkAlphaMulAlpha_SSE2(a, b));
+
+}
+
+static inline __m128i blendfunc_multiply_byte_SSE2(const __m128i& sc, const __m128i& dc,
+                                                   const __m128i& sa, const __m128i& da) {
+    // sc * (255 - da)
+    __m128i ret1 = _mm_sub_epi32(_mm_set1_epi32(255), da);
+    ret1 = _mm_mullo_epi16(sc, ret1);
+
+    // dc * (255 - sa)
+    __m128i ret2 = _mm_sub_epi32(_mm_set1_epi32(255), sa);
+    ret2 = _mm_mullo_epi16(dc, ret2);
+
+    // sc * dc
+    __m128i ret3 = _mm_mullo_epi16(sc, dc);
+
+    __m128i ret = _mm_add_epi32(ret1, ret2);
+    ret = _mm_add_epi32(ret, ret3);
+
+    return clamp_div255round_SSE2(ret);
+}
+
+static __m128i multiply_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
+    __m128i sa = SkGetPackedA32_SSE2(src);
+    __m128i da = SkGetPackedA32_SSE2(dst);
+    __m128i a = srcover_byte_SSE2(sa, da);
+
+    __m128i sr = SkGetPackedR32_SSE2(src);
+    __m128i dr = SkGetPackedR32_SSE2(dst);
+    __m128i r = blendfunc_multiply_byte_SSE2(sr, dr, sa, da);
+
+    __m128i sg = SkGetPackedG32_SSE2(src);
+    __m128i dg = SkGetPackedG32_SSE2(dst);
+    __m128i g = blendfunc_multiply_byte_SSE2(sg, dg, sa, da);
+
+
+    __m128i sb = SkGetPackedB32_SSE2(src);
+    __m128i db = SkGetPackedB32_SSE2(dst);
+    __m128i b = blendfunc_multiply_byte_SSE2(sb, db, sa, da);
+
+    return SkPackARGB32_SSE2(a, r, g, b);
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+typedef __m128i (*SkXfermodeProcSIMD)(const __m128i& src, const __m128i& dst);
+
+extern SkXfermodeProcSIMD gSSE2XfermodeProcs[];
+
+SkSSE2ProcCoeffXfermode::SkSSE2ProcCoeffXfermode(SkReadBuffer& buffer)
+    : INHERITED(buffer) {
+    fProcSIMD = reinterpret_cast<void*>(gSSE2XfermodeProcs[this->getMode()]);
+}
+
+void SkSSE2ProcCoeffXfermode::xfer32(SkPMColor dst[], const SkPMColor src[],
+                                     int count, const SkAlpha aa[]) const {
+    SkASSERT(dst && src && count >= 0);
+
+    SkXfermodeProc proc = this->getProc();
+    SkXfermodeProcSIMD procSIMD = reinterpret_cast<SkXfermodeProcSIMD>(fProcSIMD);
+    SkASSERT(procSIMD != NULL);
+
+    if (NULL == aa) {
+        if (count >= 4) {
+            while (((size_t)dst & 0x0F) != 0) {
+                *dst = proc(*src, *dst);
+                dst++;
+                src++;
+                count--;
+            }
+
+            const __m128i* s = reinterpret_cast<const __m128i*>(src);
+            __m128i* d = reinterpret_cast<__m128i*>(dst);
+
+            while (count >= 4) {
+                __m128i src_pixel = _mm_loadu_si128(s++);
+                __m128i dst_pixel = _mm_load_si128(d);
+
+                dst_pixel = procSIMD(src_pixel, dst_pixel);
+                _mm_store_si128(d++, dst_pixel);
+                count -= 4;
+            }
+
+            src = reinterpret_cast<const SkPMColor*>(s);
+            dst = reinterpret_cast<SkPMColor*>(d);
+        }
+
+        for (int i = count - 1; i >= 0; --i) {
+            *dst = proc(*src, *dst);
+            dst++;
+            src++;
+        }
+    } else {
+        for (int i = count - 1; i >= 0; --i) {
+            unsigned a = aa[i];
+            if (0 != a) {
+                SkPMColor dstC = dst[i];
+                SkPMColor C = proc(src[i], dstC);
+                if (a != 0xFF) {
+                    C = SkFourByteInterp(C, dstC, a);
+                }
+                dst[i] = C;
+            }
+        }
+    }
+}
+
+void SkSSE2ProcCoeffXfermode::xfer16(uint16_t dst[], const SkPMColor src[],
+                                     int count, const SkAlpha aa[]) const {
+    SkASSERT(dst && src && count >= 0);
+
+    SkXfermodeProc proc = this->getProc();
+    SkXfermodeProcSIMD procSIMD = reinterpret_cast<SkXfermodeProcSIMD>(fProcSIMD);
+    SkASSERT(procSIMD != NULL);
+
+    if (NULL == aa) {
+        if (count >= 8) {
+            while (((size_t)dst & 0x0F) != 0) {
+                SkPMColor dstC = SkPixel16ToPixel32(*dst);
+                *dst = SkPixel32ToPixel16_ToU16(proc(*src, dstC));
+                dst++;
+                src++;
+                count--;
+            }
+
+            const __m128i* s = reinterpret_cast<const __m128i*>(src);
+            __m128i* d = reinterpret_cast<__m128i*>(dst);
+
+            while (count >= 8) {
+                __m128i src_pixel1 = _mm_loadu_si128(s++);
+                __m128i src_pixel2 = _mm_loadu_si128(s++);
+                __m128i dst_pixel = _mm_load_si128(d);
+
+                __m128i dst_pixel1 = _mm_unpacklo_epi16(dst_pixel, _mm_setzero_si128());
+                __m128i dst_pixel2 = _mm_unpackhi_epi16(dst_pixel, _mm_setzero_si128());
+
+                __m128i dstC1 = SkPixel16ToPixel32_SSE2(dst_pixel1);
+                __m128i dstC2 = SkPixel16ToPixel32_SSE2(dst_pixel2);
+
+                dst_pixel1 = procSIMD(src_pixel1, dstC1);
+                dst_pixel2 = procSIMD(src_pixel2, dstC2);
+                dst_pixel = SkPixel32ToPixel16_ToU16_SSE2(dst_pixel1, dst_pixel2);
+
+                _mm_store_si128(d++, dst_pixel);
+                count -= 8;
+            }
+
+            src = reinterpret_cast<const SkPMColor*>(s);
+            dst = reinterpret_cast<uint16_t*>(d);
+        }
+
+        for (int i = count - 1; i >= 0; --i) {
+            SkPMColor dstC = SkPixel16ToPixel32(*dst);
+            *dst = SkPixel32ToPixel16_ToU16(proc(*src, dstC));
+            dst++;
+            src++;
+        }
+    } else {
+        for (int i = count - 1; i >= 0; --i) {
+            unsigned a = aa[i];
+            if (0 != a) {
+                SkPMColor dstC = SkPixel16ToPixel32(dst[i]);
+                SkPMColor C = proc(src[i], dstC);
+                if (0xFF != a) {
+                    C = SkFourByteInterp(C, dstC, a);
+                }
+                dst[i] = SkPixel32ToPixel16_ToU16(C);
+            }
+        }
+    }
+}
+
+#ifndef SK_IGNORE_TO_STRING
+void SkSSE2ProcCoeffXfermode::toString(SkString* str) const {
+    this->INHERITED::toString(str);
+}
+#endif
+
+////////////////////////////////////////////////////////////////////////////////
+
+// 4 pixels modeprocs with SSE2
+SkXfermodeProcSIMD gSSE2XfermodeProcs[] = {
+    NULL, // kClear_Mode
+    NULL, // kSrc_Mode
+    NULL, // kDst_Mode
+    NULL, // kSrcOver_Mode
+    NULL, // kDstOver_Mode
+    NULL, // kSrcIn_Mode
+    NULL, // kDstIn_Mode
+    NULL, // kSrcOut_Mode
+    NULL, // kDstOut_Mode
+    NULL, // kSrcATop_Mode
+    NULL, // kDstATop_Mode
+    NULL, // kXor_Mode
+    NULL, // kPlus_Mode
+    NULL, // kModulate_Mode
+    NULL, // kScreen_Mode
+
+    NULL, // kOverlay_Mode
+    NULL, // kDarken_Mode
+    NULL, // kLighten_Mode
+    NULL, // kColorDodge_Mode
+    NULL, // kColorBurn_Mode
+    NULL, // kHardLight_Mode
+    NULL, // kSoftLight_Mode
+    NULL, // kDifference_Mode
+    NULL, // kExclusion_Mode
+    multiply_modeproc_SSE2,
+
+    NULL, // kHue_Mode
+    NULL, // kSaturation_Mode
+    NULL, // kColor_Mode
+    NULL, // kLuminosity_Mode
+};
+
+SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl_SSE2(const ProcCoeff& rec,
+                                                         SkXfermode::Mode mode) {
+    void* procSIMD = reinterpret_cast<void*>(gSSE2XfermodeProcs[mode]);
+
+    if (procSIMD != NULL) {
+        return SkNEW_ARGS(SkSSE2ProcCoeffXfermode, (rec, mode, procSIMD));
+    }
+    return NULL;
+}
diff --git a/src/opts/SkXfermode_opts_SSE2.h b/src/opts/SkXfermode_opts_SSE2.h
new file mode 100644
index 0000000000..f2226fa4fd
--- /dev/null
+++ b/src/opts/SkXfermode_opts_SSE2.h
@@ -0,0 +1,30 @@
+#ifndef SkXfermode_opts_SSE2_DEFINED
+#define SkXfermode_opts_SSE2_DEFINED
+
+#include "SkXfermode_proccoeff.h"
+
+class SkSSE2ProcCoeffXfermode : public SkProcCoeffXfermode {
+public:
+    SkSSE2ProcCoeffXfermode(const ProcCoeff& rec, SkXfermode::Mode mode,
+                            void* procSIMD)
+        : INHERITED(rec, mode), fProcSIMD(procSIMD) {}
+
+    virtual void xfer32(SkPMColor dst[], const SkPMColor src[], int count,
+                        const SkAlpha aa[]) const SK_OVERRIDE;
+    virtual void xfer16(uint16_t dst[], const SkPMColor src[],
+                        int count, const SkAlpha aa[]) const SK_OVERRIDE;
+
+    SK_TO_STRING_OVERRIDE()
+    SK_DECLARE_PUBLIC_FLATTENABLE_DESERIALIZATION_PROCS(SkSSE2ProcCoeffXfermode)
+
+private:
+    SkSSE2ProcCoeffXfermode(SkReadBuffer& buffer);
+
+    void* fProcSIMD;
+    typedef SkProcCoeffXfermode INHERITED;
+};
+
+SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl_SSE2(const ProcCoeff& rec,
+                                                         SkXfermode::Mode mode);
+
+#endif // SkXfermode_opts_SSE2_DEFINED
diff --git a/src/opts/opts_check_SSE2.cpp b/src/opts/opts_check_SSE2.cpp
index b63a2ffa0f..e7677220cb 100644
--- a/src/opts/opts_check_SSE2.cpp
+++ b/src/opts/opts_check_SSE2.cpp
@@ -17,6 +17,8 @@
 #include "SkUtils.h"
 #include "SkMorphology_opts.h"
 #include "SkMorphology_opts_SSE2.h"
+#include "SkXfermode.h"
+#include "SkXfermode_proccoeff.h"
 
 #include "SkRTConf.h"
 
@@ -307,3 +309,32 @@ SkBlitRow::ColorRectProc PlatformColorRectProcFactory() {
         return NULL;
     }
 }
+
+extern SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl_SSE2(const ProcCoeff& rec,
+                                                                SkXfermode::Mode mode);
+
+SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl(const ProcCoeff& rec,
+                                                    SkXfermode::Mode mode);
+
+SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl(const ProcCoeff& rec,
+                                                    SkXfermode::Mode mode) {
+    return NULL;
+}
+
+SkProcCoeffXfermode* SkPlatformXfermodeFactory(const ProcCoeff& rec,
+                                               SkXfermode::Mode mode);
+
+SkProcCoeffXfermode* SkPlatformXfermodeFactory(const ProcCoeff& rec,
+                                               SkXfermode::Mode mode) {
+    if (cachedHasSSE2()) {
+        return SkPlatformXfermodeFactory_impl_SSE2(rec, mode);
+    } else {
+        return SkPlatformXfermodeFactory_impl(rec, mode);
+    }
+}
+
+SkXfermodeProc SkPlatformXfermodeProcFactory(SkXfermode::Mode mode);
+
+SkXfermodeProc SkPlatformXfermodeProcFactory(SkXfermode::Mode mode) {
+    return NULL;
+}