shithub: libvpx

Info  •  Files  •  Log  •  Branches

Merge "Scale the normalization factor depending on the block size"

--- a/vp9/encoder/vp9_avg.c

+++ b/vp9/encoder/vp9_avg.c

@@ -32,12 +32,13 @@

 void vp9_int_pro_row_c(int16_t *hbuf, uint8_t const *ref,

                        const int ref_stride, const int height) {

   int idx;

+  const int norm_factor = MAX(8, height >> 1);

   for (idx = 0; idx < 16; ++idx) {

     int i;

     hbuf[idx] = 0;

     for (i = 0; i < height; ++i)

       hbuf[idx] += ref[i * ref_stride];

-    hbuf[idx] /= 32;

+    hbuf[idx] /= norm_factor;

     ++ref;

   }

 }

@@ -45,9 +46,10 @@

 int16_t vp9_int_pro_col_c(uint8_t const *ref, const int width) {

   int idx;

   int16_t sum = 0;

+  const int norm_factor = MAX(8, width >> 1);

   for (idx = 0; idx < width; ++idx)

     sum += ref[idx];

-  return sum / 32;

+  return sum / norm_factor;

 }

 int vp9_vector_var_c(int16_t const *ref, int16_t const *src,

--- a/vp9/encoder/x86/vp9_avg_intrin_sse2.c

+++ b/vp9/encoder/x86/vp9_avg_intrin_sse2.c

@@ -90,8 +90,16 @@

   s0 = _mm_adds_epu16(s0, t0);

   s1 = _mm_adds_epu16(s1, t1);

-  s0 = _mm_srai_epi16(s0, 5);

-  s1 = _mm_srai_epi16(s1, 5);

+  if (height == 64) {

+    s0 = _mm_srai_epi16(s0, 5);

+    s1 = _mm_srai_epi16(s1, 5);

+  } else if (height == 32) {

+    s0 = _mm_srai_epi16(s0, 4);

+    s1 = _mm_srai_epi16(s1, 4);

+  } else {

+    s0 = _mm_srai_epi16(s0, 3);

+    s1 = _mm_srai_epi16(s1, 3);

+  }

   _mm_store_si128((__m128i *)hbuf, s0);

   hbuf += 8;

@@ -104,6 +112,7 @@

   __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;

@@ -115,7 +124,7 @@

   s1 = _mm_srli_si128(s0, 8);

   s0 = _mm_adds_epu16(s0, s1);

-  return (_mm_extract_epi16(s0, 0)) >> 5;

+  return _mm_extract_epi16(s0, 0) >> norm_factor;

 }

 int vp9_vector_var_sse2(int16_t const *ref, int16_t const *src,