Merge "Add vp8_sub_pixel_variance16x8_ssse3 function"

3 files changed, 56 insertions, 2 deletions

diff --git a/vp8/encoder/x86/variance_ssse3.c b/vp8/encoder/x86/variance_ssse3.c
index d50ae3ade..eb5d486bf 100644
--- a/vp8/encoder/x86/variance_ssse3.c
+++ b/vp8/encoder/x86/variance_ssse3.c
@@ -76,8 +76,8 @@ unsigned int vp8_sub_pixel_variance16x16_ssse3
     unsigned int *sse
 )
 {
-    int xsum0, xsum1;
-    unsigned int xxsum0, xxsum1;
+    int xsum0;
+    unsigned int xxsum0;
 
     // note we could avoid these if statements if the calling function
     // just called the appropriate functions inside.
@@ -114,3 +114,52 @@ unsigned int vp8_sub_pixel_variance16x16_ssse3
     *sse = xxsum0;
     return (xxsum0 - ((xsum0 * xsum0) >> 8));
 }
+
+unsigned int vp8_sub_pixel_variance16x8_ssse3
+(
+    const unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    const unsigned char *dst_ptr,
+    int dst_pixels_per_line,
+    unsigned int *sse
+
+)
+{
+    int xsum0;
+    unsigned int xxsum0;
+
+    if (xoffset == 4 && yoffset == 0)
+    {
+        vp8_half_horiz_variance16x_h_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 8,
+            &xsum0, &xxsum0);
+    }
+    else if (xoffset == 0 && yoffset == 4)
+    {
+        vp8_half_vert_variance16x_h_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 8,
+            &xsum0, &xxsum0);
+    }
+    else if (xoffset == 4 && yoffset == 4)
+    {
+        vp8_half_horiz_vert_variance16x_h_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 8,
+            &xsum0, &xxsum0);
+    }
+    else
+    {
+        vp8_filter_block2d_bil_var_ssse3(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 8,
+            xoffset, yoffset,
+            &xsum0, &xxsum0);
+    }
+
+    *sse = xxsum0;
+    return (xxsum0 - ((xsum0 * xsum0) >> 7));
+}
diff --git a/vp8/encoder/x86/variance_x86.h b/vp8/encoder/x86/variance_x86.h
index 1e2fb3490..3560f7413 100644
--- a/vp8/encoder/x86/variance_x86.h
+++ b/vp8/encoder/x86/variance_x86.h
@@ -286,6 +286,7 @@ extern prototype_sad_multi_dif_address(vp8_sad4x4x4d_sse3);
 #if HAVE_SSSE3
 extern prototype_sad_multi_same_address(vp8_sad16x16x3_ssse3);
 extern prototype_sad_multi_same_address(vp8_sad16x8x3_ssse3);
+extern prototype_subpixvariance(vp8_sub_pixel_variance16x8_ssse3);
 extern prototype_subpixvariance(vp8_sub_pixel_variance16x16_ssse3);
 
 #if !CONFIG_RUNTIME_CPU_DETECT
@@ -295,6 +296,9 @@ extern prototype_subpixvariance(vp8_sub_pixel_variance16x16_ssse3);
 #undef  vp8_variance_sad16x8x3
 #define vp8_variance_sad16x8x3 vp8_sad16x8x3_ssse3
 
+#undef  vp8_variance_subpixvar16x8
+#define vp8_variance_subpixvar16x8 vp8_sub_pixel_variance16x8_ssse3
+
 #undef  vp8_variance_subpixvar16x16
 #define vp8_variance_subpixvar16x16 vp8_sub_pixel_variance16x16_ssse3
 
diff --git a/vp8/encoder/x86/x86_csystemdependent.c b/vp8/encoder/x86/x86_csystemdependent.c
index c7639a7e4..3158ac12b 100644
--- a/vp8/encoder/x86/x86_csystemdependent.c
+++ b/vp8/encoder/x86/x86_csystemdependent.c
@@ -334,6 +334,7 @@ void vp8_arch_x86_encoder_init(VP8_COMP *cpi)
         cpi->rtcd.variance.sad16x16x3            = vp8_sad16x16x3_ssse3;
         cpi->rtcd.variance.sad16x8x3             = vp8_sad16x8x3_ssse3;
 
+        cpi->rtcd.variance.subpixvar16x8         = vp8_sub_pixel_variance16x8_ssse3;
         cpi->rtcd.variance.subpixvar16x16        = vp8_sub_pixel_variance16x16_ssse3;
 
         cpi->rtcd.quantize.fastquantb            = vp8_fast_quantize_b_ssse3;