Added sse2 instrinsic version of vp8_sad3x16

1.6% boost in decoder performance for the clip used.

Change-Id: I91f3c4573fd3d10afbf18930f279af7ae2223e3a

diff --git a/vp8/common/rtcd_defs.sh b/vp8/common/rtcd_defs.sh
index 90debf5..d7e5b5b 100644 (file)
--- a/vp8/common/rtcd_defs.sh
+++ b/vp8/common/rtcd_defs.sh
@@ -182,7 +182,7 @@ prototype unsigned int vp8_sad16x3 "const unsigned char *src_ptr, int  src_strid
 specialize vp8_sad16x3 sse2
 
 prototype unsigned int vp8_sad3x16 "const unsigned char *src_ptr, int  src_stride, const unsigned char *ref_ptr, int ref_stride, int max_sad"
-specialize vp8_sad3x16
+specialize vp8_sad3x16 sse2
 fi
 
 #

diff --git a/vp8/common/x86/sadmxn_x86.c b/vp8/common/x86/sadmxn_x86.c
index 5057703..46f3882 100644 (file)
--- a/vp8/common/x86/sadmxn_x86.c
+++ b/vp8/common/x86/sadmxn_x86.c
@@ -44,6 +44,48 @@ unsigned int vp8_sad16x3_sse2(
 
   return _mm_cvtsi128_si32(sad);
 }
+
+unsigned int vp8_sad3x16_sse2(
+  const unsigned char *src_ptr,
+  int  src_stride,
+  const unsigned char *ref_ptr,
+  int  ref_stride,
+  int max_sad) {
+  int r;
+  __m128i s0, s1, s2, s3;
+  __m128i r0, r1, r2, r3;
+  __m128i sad = _mm_set1_epi16(0);
+  for (r = 0; r < 16; r += 4) {
+    s0 = _mm_cvtsi32_si128 (*(const int *)(src_ptr + 0 * src_stride));
+    s1 = _mm_cvtsi32_si128 (*(const int *)(src_ptr + 1 * src_stride));
+    s2 = _mm_cvtsi32_si128 (*(const int *)(src_ptr + 2 * src_stride));
+    s3 = _mm_cvtsi32_si128 (*(const int *)(src_ptr + 3 * src_stride));
+    r0 = _mm_cvtsi32_si128 (*(const int *)(ref_ptr + 0 * src_stride));
+    r1 = _mm_cvtsi32_si128 (*(const int *)(ref_ptr + 1 * src_stride));
+    r2 = _mm_cvtsi32_si128 (*(const int *)(ref_ptr + 2 * src_stride));
+    r3 = _mm_cvtsi32_si128 (*(const int *)(ref_ptr + 3 * src_stride));
+
+    s0 = _mm_unpacklo_epi8(s0, s1);
+    r0 = _mm_unpacklo_epi8(r0, r1);
+    s2 = _mm_unpacklo_epi8(s2, s3);
+    r2 = _mm_unpacklo_epi8(r2, r3);
+    s0 = _mm_unpacklo_epi64(s0, s2);
+    r0 = _mm_unpacklo_epi64(r0, r2);
+
+    // throw out byte 3
+    s0 = _mm_slli_epi64(s0, 16);
+    r0 = _mm_slli_epi64(r0, 16);
+
+    sad = _mm_add_epi16(sad, _mm_sad_epu8(s0, r0));
+
+    src_ptr += src_stride*4;
+    ref_ptr += ref_stride*4;
+  }
+
+  sad = _mm_add_epi16(sad,  _mm_srli_si128(sad, 8));
+  return _mm_cvtsi128_si32(sad);
+}
+
 #endif