change scale down by 4 to use rounding.

source/scale_gcc.cc

Index: source/scale_gcc.cc
diff --git a/source/scale_gcc.cc b/source/scale_gcc.cc
index 9424eceddb4339eaa651bcd1695993f1d53f3c5d..5d10a01fa3831867f45d961f9fb203695a335d4c 100644
--- a/source/scale_gcc.cc
+++ b/source/scale_gcc.cc
@@ -286,7 +286,7 @@ void ScaleRowDown2Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
 }
 #endif  // HAS_SCALEROWDOWN2_AVX2
 
-void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+void ScaleRowDown4_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
                         uint8* dst_ptr, int dst_width) {
   asm volatile (
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
@@ -314,12 +314,15 @@ void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
   );
 }
 
-void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+void ScaleRowDown4Box_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
                            uint8* dst_ptr, int dst_width) {
   intptr_t stridex3 = 0;
   asm volatile (
-    "pcmpeqb   %%xmm7,%%xmm7                   \n"
-    "psrlw     $0x8,%%xmm7                     \n"
+    "pcmpeqb    %%xmm4,%%xmm4                  \n"
+    "psrlw      $0xf,%%xmm4                    \n"
+    "movdqa     %%xmm4,%%xmm5                  \n"
+    "packuswb   %%xmm4,%%xmm4                  \n"
+    "psllw      $0x3,%%xmm5                    \n"
     "lea       " MEMLEA4(0x00,4,4,2) ",%3      \n"
 
     LABELALIGN
@@ -328,31 +331,29 @@ void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
     "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
     MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
-    "pavgb     %%xmm2,%%xmm0                   \n"
-    "pavgb     %%xmm3,%%xmm1                   \n"
+    "pmaddubsw  %%xmm4,%%xmm0                  \n"
+    "pmaddubsw  %%xmm4,%%xmm1                  \n"
+    "pmaddubsw  %%xmm4,%%xmm2                  \n"
+    "pmaddubsw  %%xmm4,%%xmm3                  \n"
+    "paddw      %%xmm2,%%xmm0                  \n"
+    "paddw      %%xmm3,%%xmm1                  \n"
     MEMOPREG(movdqu,0x00,0,4,2,xmm2)           //  movdqu  (%0,%4,2),%%xmm2
     MEMOPREG(movdqu,0x10,0,4,2,xmm3)           //  movdqu  0x10(%0,%4,2),%%xmm3
-    MEMOPREG(movdqu,0x00,0,3,1,xmm4)           //  movdqu  (%0,%3,1),%%xmm4
-    MEMOPREG(movdqu,0x10,0,3,1,xmm5)           //  movdqu  0x10(%0,%3,1),%%xmm5
+    "pmaddubsw  %%xmm4,%%xmm2                  \n"
+    "pmaddubsw  %%xmm4,%%xmm3                  \n"
+    "paddw      %%xmm2,%%xmm0                  \n"
+    "paddw      %%xmm3,%%xmm1                  \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu  (%0,%3,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu  0x10(%0,%3,1),%%xmm3
     "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pavgb     %%xmm4,%%xmm2                   \n"
-    "pavgb     %%xmm2,%%xmm0                   \n"
-    "pavgb     %%xmm5,%%xmm3                   \n"
-    "pavgb     %%xmm3,%%xmm1                   \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "movdqa    %%xmm1,%%xmm3                   \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "pand      %%xmm7,%%xmm2                   \n"
-    "pand      %%xmm7,%%xmm3                   \n"
-    "pavgw     %%xmm2,%%xmm0                   \n"
-    "pavgw     %%xmm3,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "pand      %%xmm7,%%xmm2                   \n"
-    "pavgw     %%xmm2,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
+    "pmaddubsw  %%xmm4,%%xmm2                  \n"
+    "pmaddubsw  %%xmm4,%%xmm3                  \n"
+    "paddw      %%xmm2,%%xmm0                  \n"
+    "paddw      %%xmm3,%%xmm1                  \n"
+    "phaddw     %%xmm1,%%xmm0                  \n"
+    "paddw      %%xmm5,%%xmm0                  \n"
+    "psrlw      $0x4,%%xmm0                    \n"
+    "packuswb   %%xmm0,%%xmm0                  \n"
     "movq      %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x8,1) ",%1            \n"
     "sub       $0x8,%2                         \n"
@@ -363,7 +364,7 @@ void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
     "+r"(stridex3)     // %3
   : "r"((intptr_t)(src_stride))    // %4
   : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm7"
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }