libvpx: Pull from upstream

source/libvpx/vp9/encoder/x86/vp9_avg_intrin_sse2.c

Index: source/libvpx/vp9/encoder/x86/vp9_avg_intrin_sse2.c
diff --git a/source/libvpx/vp9/encoder/x86/vp9_avg_intrin_sse2.c b/source/libvpx/vp9/encoder/x86/vp9_avg_intrin_sse2.c
index 0a105629f067478a5754101879b38364e07af345..f49949940395e9ef29ec8c76470e4122c843879f 100644
--- a/source/libvpx/vp9/encoder/x86/vp9_avg_intrin_sse2.c
+++ b/source/libvpx/vp9/encoder/x86/vp9_avg_intrin_sse2.c
@@ -90,6 +90,17 @@ void vp9_int_pro_row_sse2(int16_t *hbuf, uint8_t const*ref,
   s0 = _mm_adds_epu16(s0, t0);
   s1 = _mm_adds_epu16(s1, t1);
 
+  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;
   _mm_store_si128((__m128i *)hbuf, s1);
@@ -100,73 +111,62 @@ int16_t vp9_int_pro_col_sse2(uint8_t const *ref, const int width) {
   __m128i src_line = _mm_load_si128((const __m128i *)ref);
   __m128i s0 = _mm_sad_epu8(src_line, zero);
   __m128i s1;
-  (void) width;  // width = 64
-
-  ref += 16;
-  src_line = _mm_load_si128((const __m128i *)ref);
-  s1 = _mm_sad_epu8(src_line, zero);
-  s0 = _mm_adds_epu16(s0, s1);
+  int i;
+  const int norm_factor = 3 + (width >> 5);
 
-  ref += 16;
-  src_line = _mm_load_si128((const __m128i *)ref);
-  s1 = _mm_sad_epu8(src_line, zero);
-  s0 = _mm_adds_epu16(s0, s1);
-
-  ref += 16;
-  src_line = _mm_load_si128((const __m128i *)ref);
-  s1 = _mm_sad_epu8(src_line, zero);
-  s0 = _mm_adds_epu16(s0, s1);
+  for (i = 16; i < width; i += 16) {
+    ref += 16;
+    src_line = _mm_load_si128((const __m128i *)ref);
+    s1 = _mm_sad_epu8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, s1);
+  }
 
   s1 = _mm_srli_si128(s0, 8);
   s0 = _mm_adds_epu16(s0, s1);
 
-  return _mm_extract_epi16(s0, 0);
+  return _mm_extract_epi16(s0, 0) >> norm_factor;
 }
 
-int vp9_vector_sad_sse2(int16_t const *ref, int16_t const *src,
-                        const int width) {
+int vp9_vector_var_sse2(int16_t const *ref, int16_t const *src,
+                        const int bwl) {
   int idx;
-  __m128i zero = _mm_setzero_si128();
-  __m128i sum;
+  int width = 4 << bwl;
+  int16_t mean;
   __m128i v0 = _mm_loadu_si128((const __m128i *)ref);
   __m128i v1 = _mm_load_si128((const __m128i *)src);
   __m128i diff = _mm_subs_epi16(v0, v1);
-  __m128i sign = _mm_srai_epi16(diff, 15);
-
-  diff = _mm_xor_si128(diff, sign);
-  sum = _mm_sub_epi16(diff, sign);
-
-  (void) width;  // width = 64;
+  __m128i sum = diff;
+  __m128i sse = _mm_madd_epi16(diff, diff);
 
   ref += 8;
   src += 8;
 
-  v0 = _mm_unpacklo_epi16(sum, zero);
-  v1 = _mm_unpackhi_epi16(sum, zero);
-  sum = _mm_add_epi32(v0, v1);
-
-  for (idx = 1; idx < 8; ++idx) {
+  for (idx = 8; idx < width; idx += 8) {
     v0 = _mm_loadu_si128((const __m128i *)ref);
     v1 = _mm_load_si128((const __m128i *)src);
     diff = _mm_subs_epi16(v0, v1);
-    sign = _mm_srai_epi16(diff, 15);
-    diff = _mm_xor_si128(diff, sign);
-    diff = _mm_sub_epi16(diff, sign);
-
-    v0 = _mm_unpacklo_epi16(diff, zero);
-    v1 = _mm_unpackhi_epi16(diff, zero);
 
-    sum = _mm_add_epi32(sum, v0);
-    sum = _mm_add_epi32(sum, v1);
+    sum = _mm_add_epi16(sum, diff);
+    v0  = _mm_madd_epi16(diff, diff);
+    sse = _mm_add_epi32(sse, v0);
 
     ref += 8;
     src += 8;
   }
 
-  v0 = _mm_srli_si128(sum, 8);
-  sum = _mm_add_epi32(sum, v0);
-  v0 = _mm_srli_epi64(sum, 32);
-  sum = _mm_add_epi32(sum, v0);
+  v0  = _mm_srli_si128(sum, 8);
+  sum = _mm_add_epi16(sum, v0);
+  v0  = _mm_srli_epi64(sum, 32);
+  sum = _mm_add_epi16(sum, v0);
+  v0  = _mm_srli_epi32(sum, 16);
+  sum = _mm_add_epi16(sum, v0);
+
+  v1  = _mm_srli_si128(sse, 8);
+  sse = _mm_add_epi32(sse, v1);
+  v1  = _mm_srli_epi64(sse, 32);
+  sse = _mm_add_epi32(sse, v1);
+
+  mean = _mm_extract_epi16(sum, 0);
 
-  return _mm_cvtsi128_si32(sum);
+  return _mm_cvtsi128_si32(sse) - ((mean * mean) >> (bwl + 2));
 }