fix vp9_satd_sse2

accumulate satd in 32-bits
+ add unit test

Change-Id: I6748183df3662ddb9d635f9641f9586f2fd38ad5

test/vp9_avg_test.cc

@@ -356,6 +356,14 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(16, &vp9_int_pro_col_sse2, &vp9_int_pro_col_c),
         make_tuple(32, &vp9_int_pro_col_sse2, &vp9_int_pro_col_c),
         make_tuple(64, &vp9_int_pro_col_sse2, &vp9_int_pro_col_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, SatdTest,
+    ::testing::Values(
+        make_tuple(16, &vp9_satd_sse2),
+        make_tuple(64, &vp9_satd_sse2),
+        make_tuple(256, &vp9_satd_sse2),
+        make_tuple(1024, &vp9_satd_sse2)));
 #endif
 
 #if HAVE_NEON

vp9/common/vp9_rtcd_defs.pl

@@ -210,7 +210,7 @@ add_proto qw/void vp9_hadamard_16x16/, "int16_t const *src_diff, int src_stride,
 specialize qw/vp9_hadamard_16x16 sse2/;
 
 add_proto qw/int vp9_satd/, "const int16_t *coeff, int length";
-specialize qw/vp9_satd/;
+specialize qw/vp9_satd sse2/;
 
 add_proto qw/void vp9_int_pro_row/, "int16_t *hbuf, uint8_t const *ref, const int ref_stride, const int height";
 specialize qw/vp9_int_pro_row sse2 neon/;

vp9/encoder/x86/vp9_avg_intrin_sse2.c

@@ -283,35 +283,31 @@ void vp9_hadamard_16x16_sse2(int16_t const *src_diff, int src_stride,
   }
 }
 
-// TODO(jingning): the sum needs to be accumulated in 32-bits to avoid rollover.
-#if 0
 int vp9_satd_sse2(const int16_t *coeff, int length) {
   int i;
-  __m128i sum = _mm_load_si128((const __m128i *)coeff);
-  __m128i sign = _mm_srai_epi16(sum, 15);
-  __m128i val = _mm_xor_si128(sum, sign);
-  sum = _mm_sub_epi16(val, sign);
-  coeff += 8;
+  const __m128i zero = _mm_setzero_si128();
+  __m128i accum = zero;
 
-  for (i = 8; i < length; i += 8) {
-    __m128i src_line = _mm_load_si128((const __m128i *)coeff);
-    sign = _mm_srai_epi16(src_line, 15);
-    val = _mm_xor_si128(src_line, sign);
-    val = _mm_sub_epi16(val, sign);
-    sum = _mm_add_epi16(sum, val);
+  for (i = 0; i < length; i += 8) {
+    const __m128i src_line = _mm_load_si128((const __m128i *)coeff);
+    const __m128i inv = _mm_sub_epi16(zero, src_line);
+    const __m128i abs = _mm_max_epi16(src_line, inv);  // abs(src_line)
+    const __m128i abs_lo = _mm_unpacklo_epi16(abs, zero);
+    const __m128i abs_hi = _mm_unpackhi_epi16(abs, zero);
+    const __m128i sum = _mm_add_epi32(abs_lo, abs_hi);
+    accum = _mm_add_epi32(accum, sum);
     coeff += 8;
   }
 
-  val = _mm_srli_si128(sum, 8);
-  sum = _mm_add_epi16(sum, val);
-  val = _mm_srli_epi64(sum, 32);
-  sum = _mm_add_epi16(sum, val);
-  val = _mm_srli_epi32(sum, 16);
-  sum = _mm_add_epi16(sum, val);
+  {  // cascading summation of accum
+    __m128i hi = _mm_srli_si128(accum, 8);
+    accum = _mm_add_epi32(accum, hi);
+    hi = _mm_srli_epi64(accum, 32);
+    accum = _mm_add_epi32(accum, hi);
+  }
 
-  return _mm_extract_epi16(sum, 0);
+  return _mm_cvtsi128_si32(accum);
 }
-#endif
 
 void vp9_int_pro_row_sse2(int16_t *hbuf, uint8_t const*ref,
                           const int ref_stride, const int height) {