SAD32xh and SAD64xh for AVX2

All sad function that process above 32 consecutive elements are optimized
for AVX2:
vp9_sad64x64
vp9_sad64x32
vp9_sad32x64
vp9_sad32x32
vp9_sad32x16
vp9_sad64x64_avg
vp9_sad64x32_avg
vp9_sad32x64_avg
vp9_sad32x32_avg
vp9_sad32x16_avg
The functions that appeared as a hotspot is vp9_sad32x32 and vp9_sad64x64
vp9_sad32x32 was optimized by 68% and vp9_sad64x64 was optimized by 90%
both of them gave and overall ~2.3% user level gain

Change-Id: Iccf86b375a2b54c5fbbe685902ead0c9a561b9fd

diff --git a/test/sad_test.cc b/test/sad_test.cc
index 5377c1e..c7042fe 100644 (file)
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -625,6 +625,20 @@ INSTANTIATE_TEST_CASE_P(SSE3, SADTest, ::testing::Values(
 
 #if HAVE_AVX2
 #if CONFIG_VP9_ENCODER
+const SadMxNVp9Func sad_64x64_avx2_vp9 = vp9_sad64x64_avx2;
+const SadMxNVp9Func sad_64x32_avx2_vp9 = vp9_sad64x32_avx2;
+const SadMxNVp9Func sad_32x64_avx2_vp9 = vp9_sad32x64_avx2;
+const SadMxNVp9Func sad_32x32_avx2_vp9 = vp9_sad32x32_avx2;
+const SadMxNVp9Func sad_32x16_avx2_vp9 = vp9_sad32x16_avx2;
+const SadMxNVp9Param avx2_vp9_tests[] = {
+  make_tuple(64, 64, sad_64x64_avx2_vp9),
+  make_tuple(64, 32, sad_64x32_avx2_vp9),
+  make_tuple(32, 64, sad_32x64_avx2_vp9),
+  make_tuple(32, 32, sad_32x32_avx2_vp9),
+  make_tuple(32, 16, sad_32x16_avx2_vp9),
+};
+INSTANTIATE_TEST_CASE_P(AVX2, SADVP9Test, ::testing::ValuesIn(avx2_vp9_tests));
+
 const SadMxNx4Func sad_64x64x4d_avx2 = vp9_sad64x64x4d_avx2;
 const SadMxNx4Func sad_32x32x4d_avx2 = vp9_sad32x32x4d_avx2;
 INSTANTIATE_TEST_CASE_P(AVX2, SADx4Test, ::testing::Values(

diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 838ec44..a5d3ef2 100644 (file)
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -873,22 +873,22 @@ add_proto qw/unsigned int vp9_sub_pixel_avg_variance4x4/, "const uint8_t *src_pt
 specialize qw/vp9_sub_pixel_avg_variance4x4/, "$sse_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sad64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
-specialize qw/vp9_sad64x64 neon/, "$sse2_x86inc";
+specialize qw/vp9_sad64x64 neon avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vp9_sad32x64/, "$sse2_x86inc";
+specialize qw/vp9_sad32x64 avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vp9_sad64x32/, "$sse2_x86inc";
+specialize qw/vp9_sad64x32 avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vp9_sad32x16/, "$sse2_x86inc";
+specialize qw/vp9_sad32x16 avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
 specialize qw/vp9_sad16x32/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
-specialize qw/vp9_sad32x32 neon/, "$sse2_x86inc";
+specialize qw/vp9_sad32x32 neon avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
 specialize qw/vp9_sad16x16 neon/, "$sse2_x86inc";
@@ -912,22 +912,22 @@ add_proto qw/unsigned int vp9_sad4x4/, "const uint8_t *src_ptr, int source_strid
 specialize qw/vp9_sad4x4/, "$sse_x86inc";
 
 add_proto qw/unsigned int vp9_sad64x64_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
-specialize qw/vp9_sad64x64_avg/, "$sse2_x86inc";
+specialize qw/vp9_sad64x64_avg avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad32x64_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vp9_sad32x64_avg/, "$sse2_x86inc";
+specialize qw/vp9_sad32x64_avg avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad64x32_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vp9_sad64x32_avg/, "$sse2_x86inc";
+specialize qw/vp9_sad64x32_avg avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad32x16_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vp9_sad32x16_avg/, "$sse2_x86inc";
+specialize qw/vp9_sad32x16_avg avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad16x32_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad16x32_avg/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad32x32_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
-specialize qw/vp9_sad32x32_avg/, "$sse2_x86inc";
+specialize qw/vp9_sad32x32_avg avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad16x16_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad16x16_avg/, "$sse2_x86inc";

diff --git a/vp9/encoder/x86/vp9_sad_intrin_avx2.c b/vp9/encoder/x86/vp9_sad_intrin_avx2.c
new file mode 100644 (file)
index 0000000..1131930
--- /dev/null
+++ b/vp9/encoder/x86/vp9_sad_intrin_avx2.c
@@ -0,0 +1,180 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <immintrin.h>
+#include "vpx_ports/mem.h"
+
+#define FSAD64_H(h) \
+unsigned int vp9_sad64x##h##_avx2(const uint8_t *src_ptr, \
+                                  int src_stride, \
+                                  const uint8_t *ref_ptr, \
+                                  int ref_stride) { \
+  int i, res; \
+  __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \
+  __m256i sum_sad = _mm256_setzero_si256(); \
+  __m256i sum_sad_h; \
+  __m128i sum_sad128; \
+  for (i = 0 ; i < h ; i++) { \
+    ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \
+    ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32)); \
+    sad1_reg = _mm256_sad_epu8(ref1_reg, \
+               _mm256_loadu_si256((__m256i const *)src_ptr)); \
+    sad2_reg = _mm256_sad_epu8(ref2_reg, \
+               _mm256_loadu_si256((__m256i const *)(src_ptr + 32))); \
+    sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \
+    ref_ptr+= ref_stride; \
+    src_ptr+= src_stride; \
+  } \
+  sum_sad_h = _mm256_srli_si256(sum_sad, 8); \
+  sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \
+  sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \
+  sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \
+  res = _mm_cvtsi128_si32(sum_sad128); \
+  return res; \
+}
+
+#define FSAD32_H(h) \
+unsigned int vp9_sad32x##h##_avx2(const uint8_t *src_ptr, \
+                                  int src_stride, \
+                                  const uint8_t *ref_ptr, \
+                                  int ref_stride) { \
+  int i, res; \
+  __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \
+  __m256i sum_sad = _mm256_setzero_si256(); \
+  __m256i sum_sad_h; \
+  __m128i sum_sad128; \
+  int ref2_stride = ref_stride << 1; \
+  int src2_stride = src_stride << 1; \
+  int max = h >> 1; \
+  for (i = 0 ; i < max ; i++) { \
+    ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \
+    ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); \
+    sad1_reg = _mm256_sad_epu8(ref1_reg, \
+               _mm256_loadu_si256((__m256i const *)src_ptr)); \
+    sad2_reg = _mm256_sad_epu8(ref2_reg, \
+               _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride))); \
+    sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \
+    ref_ptr+= ref2_stride; \
+    src_ptr+= src2_stride; \
+  } \
+  sum_sad_h = _mm256_srli_si256(sum_sad, 8); \
+  sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \
+  sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \
+  sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \
+  res = _mm_cvtsi128_si32(sum_sad128); \
+  return res; \
+}
+
+#define FSAD64 \
+FSAD64_H(64); \
+FSAD64_H(32);
+
+#define FSAD32 \
+FSAD32_H(64); \
+FSAD32_H(32); \
+FSAD32_H(16);
+
+FSAD64;
+FSAD32;
+
+#undef FSAD64
+#undef FSAD32
+#undef FSAD64_H
+#undef FSAD32_H
+
+#define FSADAVG64_H(h) \
+unsigned int vp9_sad64x##h##_avg_avx2(const uint8_t *src_ptr, \
+                                      int src_stride, \
+                                      const uint8_t *ref_ptr, \
+                                      int  ref_stride, \
+                                      const uint8_t *second_pred) { \
+  int i, res; \
+  __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \
+  __m256i sum_sad = _mm256_setzero_si256(); \
+  __m256i sum_sad_h; \
+  __m128i sum_sad128; \
+  for (i = 0 ; i < h ; i++) { \
+    ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \
+    ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32)); \
+    ref1_reg = _mm256_avg_epu8(ref1_reg, \
+               _mm256_loadu_si256((__m256i const *)second_pred)); \
+    ref2_reg = _mm256_avg_epu8(ref2_reg, \
+               _mm256_loadu_si256((__m256i const *)(second_pred +32))); \
+    sad1_reg = _mm256_sad_epu8(ref1_reg, \
+               _mm256_loadu_si256((__m256i const *)src_ptr)); \
+    sad2_reg = _mm256_sad_epu8(ref2_reg, \
+               _mm256_loadu_si256((__m256i const *)(src_ptr + 32))); \
+    sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \
+    ref_ptr+= ref_stride; \
+    src_ptr+= src_stride; \
+    second_pred+= 64; \
+  } \
+  sum_sad_h = _mm256_srli_si256(sum_sad, 8); \
+  sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \
+  sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \
+  sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \
+  res = _mm_cvtsi128_si32(sum_sad128); \
+  return res; \
+}
+
+#define FSADAVG32_H(h) \
+unsigned int vp9_sad32x##h##_avg_avx2(const uint8_t *src_ptr, \
+                                      int src_stride, \
+                                      const uint8_t *ref_ptr, \
+                                      int  ref_stride, \
+                                      const uint8_t *second_pred) { \
+  int i, res; \
+  __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \
+  __m256i sum_sad = _mm256_setzero_si256(); \
+  __m256i sum_sad_h; \
+  __m128i sum_sad128; \
+  int ref2_stride = ref_stride << 1; \
+  int src2_stride = src_stride << 1; \
+  int max = h >> 1; \
+  for (i = 0 ; i < max ; i++) { \
+    ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \
+    ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); \
+    ref1_reg = _mm256_avg_epu8(ref1_reg, \
+               _mm256_loadu_si256((__m256i const *)second_pred)); \
+    ref2_reg = _mm256_avg_epu8(ref2_reg, \
+               _mm256_loadu_si256((__m256i const *)(second_pred +32))); \
+    sad1_reg = _mm256_sad_epu8(ref1_reg, \
+               _mm256_loadu_si256((__m256i const *)src_ptr)); \
+    sad2_reg = _mm256_sad_epu8(ref2_reg, \
+               _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride))); \
+    sum_sad = _mm256_add_epi32(sum_sad, \
+              _mm256_add_epi32(sad1_reg, sad2_reg)); \
+    ref_ptr+= ref2_stride; \
+    src_ptr+= src2_stride; \
+    second_pred+= 64; \
+  } \
+  sum_sad_h = _mm256_srli_si256(sum_sad, 8); \
+  sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \
+  sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \
+  sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \
+  res = _mm_cvtsi128_si32(sum_sad128); \
+  return res; \
+}
+
+#define FSADAVG64 \
+FSADAVG64_H(64); \
+FSADAVG64_H(32);
+
+#define FSADAVG32 \
+FSADAVG32_H(64); \
+FSADAVG32_H(32); \
+FSADAVG32_H(16);
+
+FSADAVG64;
+FSADAVG32;
+
+#undef FSADAVG64
+#undef FSADAVG32
+#undef FSADAVG64_H
+#undef FSADAVG32_H

diff --git a/vp9/vp9cx.mk b/vp9/vp9cx.mk
index e450f7b..b1903c2 100644 (file)
--- a/vp9/vp9cx.mk
+++ b/vp9/vp9cx.mk
@@ -115,6 +115,7 @@ VP9_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/vp9_dct_ssse3_x86_64.asm
 endif
 VP9_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/vp9_sad_ssse3.asm
 VP9_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/vp9_sad_sse4.asm
+VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_sad_intrin_avx2.c
 VP9_CX_SRCS-$(ARCH_X86_64) += encoder/x86/vp9_ssim_opt_x86_64.asm
 
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_dct_sse2.c