Enable ssse3 bilinear tests

The code only has issues when xoffset == 0 and yoffset == 0 which
represents a simple copy. Presumably this case does not need to be
handled because the issue has existed since 2010.

BUG=webm:1287

Change-Id: Ic47e2653f3b729e99b40e53d8d2d8d1501edaaa9

2 files changed, 17 insertions, 20 deletions

diff --git a/vp8/common/filter.c b/vp8/common/filter.c
index a312efb6c..267498335 100644
--- a/vp8/common/filter.c
+++ b/vp8/common/filter.c
@@ -8,8 +8,9 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#include "filter.h"
+#include <assert.h>
 #include "./vp8_rtcd.h"
+#include "vp8/common/filter.h"
 
 DECLARE_ALIGNED(16, const short, vp8_bilinear_filters[8][2]) = {
   { 128, 0 }, { 112, 16 }, { 96, 32 }, { 80, 48 },
@@ -324,27 +325,11 @@ void vp8_bilinear_predict4x4_c(unsigned char *src_ptr, int src_pixels_per_line,
   const short *HFilter;
   const short *VFilter;
 
+  // This represents a copy and is not required to be handled by optimizations.
+  assert((xoffset | yoffset) != 0);
+
   HFilter = vp8_bilinear_filters[xoffset];
   VFilter = vp8_bilinear_filters[yoffset];
-#if 0
-    {
-        int i;
-        unsigned char temp1[16];
-        unsigned char temp2[16];
-
-        bilinear_predict4x4_mmx(src_ptr, src_pixels_per_line, xoffset, yoffset, temp1, 4);
-        filter_block2d_bil(src_ptr, temp2, src_pixels_per_line, 4, HFilter, VFilter, 4, 4);
-
-        for (i = 0; i < 16; ++i)
-        {
-            if (temp1[i] != temp2[i])
-            {
-                bilinear_predict4x4_mmx(src_ptr, src_pixels_per_line, xoffset, yoffset, temp1, 4);
-                filter_block2d_bil(src_ptr, temp2, src_pixels_per_line, 4, HFilter, VFilter, 4, 4);
-            }
-        }
-    }
-#endif
   filter_block2d_bil(src_ptr, dst_ptr, src_pixels_per_line, dst_pitch, HFilter,
                      VFilter, 4, 4);
 }
@@ -355,6 +340,8 @@ void vp8_bilinear_predict8x8_c(unsigned char *src_ptr, int src_pixels_per_line,
   const short *HFilter;
   const short *VFilter;
 
+  assert((xoffset | yoffset) != 0);
+
   HFilter = vp8_bilinear_filters[xoffset];
   VFilter = vp8_bilinear_filters[yoffset];
 
@@ -368,6 +355,8 @@ void vp8_bilinear_predict8x4_c(unsigned char *src_ptr, int src_pixels_per_line,
   const short *HFilter;
   const short *VFilter;
 
+  assert((xoffset | yoffset) != 0);
+
   HFilter = vp8_bilinear_filters[xoffset];
   VFilter = vp8_bilinear_filters[yoffset];
 
@@ -382,6 +371,8 @@ void vp8_bilinear_predict16x16_c(unsigned char *src_ptr,
   const short *HFilter;
   const short *VFilter;
 
+  assert((xoffset | yoffset) != 0);
+
   HFilter = vp8_bilinear_filters[xoffset];
   VFilter = vp8_bilinear_filters[yoffset];
 
diff --git a/vp8/common/x86/subpixel_ssse3.asm b/vp8/common/x86/subpixel_ssse3.asm
index c06f24556..1f6cbd1d1 100644
--- a/vp8/common/x86/subpixel_ssse3.asm
+++ b/vp8/common/x86/subpixel_ssse3.asm
@@ -1291,6 +1291,8 @@ sym(vp8_bilinear_predict8x8_ssse3):
         movq        xmm7,       XMMWORD PTR [rsp+96]
         punpcklbw   xmm5,       xmm6
 
+        ; Because the source register (xmm0) is always treated as signed by
+        ; pmaddubsw, the constant '128' is treated as '-128'.
         pmaddubsw   xmm1,       xmm0
         pmaddubsw   xmm2,       xmm0
 
@@ -1319,6 +1321,10 @@ sym(vp8_bilinear_predict8x8_ssse3):
         psraw       xmm5,       VP8_FILTER_SHIFT
 
         psraw       xmm6,       VP8_FILTER_SHIFT
+
+        ; Having multiplied everything by '-128' and obtained negative
+        ; numbers, the unsigned saturation truncates those values to 0,
+        ; resulting in incorrect handling of xoffset == 0 && yoffset == 0
         packuswb    xmm1,       xmm1
 
         packuswb    xmm2,       xmm2