Refactor column integral projection computation

Move the scaling factor outside column projection. This avoids
repeated calculation of the same scaling factor. Profiling shows
that the percentage of vp9_int_pro_col_sse2 of overall cycles
goes from 2.29% down to 1.88%.

Change-Id: I5ac4e324ab2d7f33ba2de66dd2a12e04e04dfd66

diff --git a/vp9/encoder/vp9_mcomp.c b/vp9/encoder/vp9_mcomp.c
index 88c5ca0c3757b507bfae98f211ac151a948e3ef2..9602eb56869aff614696c5435444fb5262965a4a 100644 (file)
--- a/vp9/encoder/vp9_mcomp.c
+++ b/vp9/encoder/vp9_mcomp.c
@@ -1805,6 +1805,7 @@ unsigned int vp9_int_pro_motion_estimation(const VP9_COMP *cpi, MACROBLOCK *x,
   MV *tmp_mv = &xd->mi[0].src_mi->mbmi.mv[0].as_mv;
   unsigned int best_sad, tmp_sad, this_sad[4];
   MV this_mv;
+  const int norm_factor = 3 + (bw >> 5);
 
 #if CONFIG_VP9_HIGHBITDEPTH
   tmp_mv->row = 0;
@@ -1822,7 +1823,7 @@ unsigned int vp9_int_pro_motion_estimation(const VP9_COMP *cpi, MACROBLOCK *x,
 
   ref_buf = xd->plane[0].pre[0].buf - (bh >> 1) * ref_stride;
   for (idx = 0; idx < search_height; ++idx) {
-    vbuf[idx] = vp9_int_pro_col(ref_buf, bw);
+    vbuf[idx] = vp9_int_pro_col(ref_buf, bw) >> norm_factor;
     ref_buf += ref_stride;
   }
 
@@ -1834,7 +1835,7 @@ unsigned int vp9_int_pro_motion_estimation(const VP9_COMP *cpi, MACROBLOCK *x,
 
   src_buf = x->plane[0].src.buf;
   for (idx = 0; idx < bh; ++idx) {
-    src_vbuf[idx] = vp9_int_pro_col(src_buf, bw);
+    src_vbuf[idx] = vp9_int_pro_col(src_buf, bw) >> norm_factor;
     src_buf += src_stride;
   }
 

diff --git a/vp9/encoder/x86/vp9_avg_intrin_sse2.c b/vp9/encoder/x86/vp9_avg_intrin_sse2.c
index 618b5f73db66433b26226da7b8ca2cdba5a3a743..4e80b255e88fb318e31f25b504d380d46fb6afe7 100644 (file)
--- a/vp9/encoder/x86/vp9_avg_intrin_sse2.c
+++ b/vp9/encoder/x86/vp9_avg_intrin_sse2.c
@@ -112,7 +112,6 @@ int16_t vp9_int_pro_col_sse2(uint8_t const *ref, const int width) {
   __m128i s0 = _mm_sad_epu8(src_line, zero);
   __m128i s1;
   int i;
-  const int norm_factor = 3 + (width >> 5);
 
   for (i = 16; i < width; i += 16) {
     ref += 16;
@@ -124,7 +123,7 @@ int16_t vp9_int_pro_col_sse2(uint8_t const *ref, const int width) {
   s1 = _mm_srli_si128(s0, 8);
   s0 = _mm_adds_epu16(s0, s1);
 
-  return _mm_extract_epi16(s0, 0) >> norm_factor;
+  return _mm_extract_epi16(s0, 0);
 }
 
 int vp9_vector_var_sse2(int16_t const *ref, int16_t const *src,