generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

AVX2 SAD Optimization:

Browse Source 
2 functions were optimized for avx2 by using full 256 bit register
In order to handle 32 elements in parallel instead of only 16 in parallel:
1. vp9_sad32x32x4d
2. vp9_sad64x64x4d

The function level gain is 66% and the user level gain is ~1%.

Change-Id: I4efbb3bc7d8bc03b64b6c98f5cd5c4a9dd3212cb
This commit is contained in:
levytamar82 2014-03-13 14:47:30 -07:00 committed by Yunqing Wang
parent 3967435f65
commit 0fa8b668c1
3 changed files with 170 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vp9/common/vp9_rtcd_defs.pl
View File

@ -636,7 +636,7 @@ add_proto qw/void vp9_sad4x4x8/, "const uint8_t *src_ptr, int src_stride, const
specialize qw/vp9_sad4x4x8 sse4/;
add_proto qw/void vp9_sad64x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array";
specialize qw/vp9_sad64x64x4d sse2/;
specialize qw/vp9_sad64x64x4d sse2 avx2/;
add_proto qw/void vp9_sad32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array";
specialize qw/vp9_sad32x64x4d sse2/;
@ -651,7 +651,7 @@ add_proto qw/void vp9_sad16x32x4d/, "const uint8_t *src_ptr, int src_stride, co
specialize qw/vp9_sad16x32x4d sse2/;
add_proto qw/void vp9_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array";
specialize qw/vp9_sad32x32x4d sse2/;
specialize qw/vp9_sad32x32x4d sse2 avx2/;
add_proto qw/void vp9_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array";
specialize qw/vp9_sad16x16x4d sse2/;

167
vp9/encoder/x86/vp9_sad4d_intrin_avx2.c Normal file
View File

@ -0,0 +1,167 @@
/*
* Copyright (c) 2014 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> // AVX2
#include "vpx/vpx_integer.h"
void vp9_sad32x32x4d_avx2(uint8_t *src,
int src_stride,
uint8_t *ref[4],
int ref_stride,
unsigned int res[4]) {
__m256i src_reg, ref0_reg, ref1_reg, ref2_reg, ref3_reg;
__m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3;
__m256i sum_mlow, sum_mhigh;
int i;
uint8_t *ref0, *ref1, *ref2, *ref3;
ref0 = ref[0];
ref1 = ref[1];
ref2 = ref[2];
ref3 = ref[3];
sum_ref0 = _mm256_set1_epi16(0);
sum_ref1 = _mm256_set1_epi16(0);
sum_ref2 = _mm256_set1_epi16(0);
sum_ref3 = _mm256_set1_epi16(0);
for (i = 0; i < 32 ; i++) {
// load src and all refs
src_reg = _mm256_load_si256((__m256i *)(src));
ref0_reg = _mm256_loadu_si256((__m256i *) (ref0));
ref1_reg = _mm256_loadu_si256((__m256i *) (ref1));
ref2_reg = _mm256_loadu_si256((__m256i *) (ref2));
ref3_reg = _mm256_loadu_si256((__m256i *) (ref3));
// sum of the absolute differences between every ref-i to src
ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg);
// sum every ref-i
sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg);
src+= src_stride;
ref0+= ref_stride;
ref1+= ref_stride;
ref2+= ref_stride;
ref3+= ref_stride;
}
{
__m128i sum;
// in sum_ref-i the result is saved in the first 4 bytes
// the other 4 bytes are zeroed.
// sum_ref1 and sum_ref3 are shifted left by 4 bytes
sum_ref1 = _mm256_slli_si256(sum_ref1, 4);
sum_ref3 = _mm256_slli_si256(sum_ref3, 4);
// merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3
sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1);
sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3);
// merge every 64 bit from each sum_ref-i
sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2);
sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2);
// add the low 64 bit to the high 64 bit
sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh);
// add the low 128 bit to the high 128 bit
sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow),
_mm256_extractf128_si256(sum_mlow, 1));
_mm_storeu_si128((__m128i *)(res), sum);
}
}
void vp9_sad64x64x4d_avx2(uint8_t *src,
int src_stride,
uint8_t *ref[4],
int ref_stride,
unsigned int res[4]) {
__m256i src_reg, srcnext_reg, ref0_reg, ref0next_reg;
__m256i ref1_reg, ref1next_reg, ref2_reg, ref2next_reg;
__m256i ref3_reg, ref3next_reg;
__m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3;
__m256i sum_mlow, sum_mhigh;
int i;
uint8_t *ref0, *ref1, *ref2, *ref3;
ref0 = ref[0];
ref1 = ref[1];
ref2 = ref[2];
ref3 = ref[3];
sum_ref0 = _mm256_set1_epi16(0);
sum_ref1 = _mm256_set1_epi16(0);
sum_ref2 = _mm256_set1_epi16(0);
sum_ref3 = _mm256_set1_epi16(0);
for (i = 0; i < 64 ; i++) {
// load 64 bytes from src and all refs
src_reg = _mm256_load_si256((__m256i *)(src));
srcnext_reg = _mm256_load_si256((__m256i *)(src + 32));
ref0_reg = _mm256_loadu_si256((__m256i *) (ref0));
ref0next_reg = _mm256_loadu_si256((__m256i *) (ref0 + 32));
ref1_reg = _mm256_loadu_si256((__m256i *) (ref1));
ref1next_reg = _mm256_loadu_si256((__m256i *) (ref1 + 32));
ref2_reg = _mm256_loadu_si256((__m256i *) (ref2));
ref2next_reg = _mm256_loadu_si256((__m256i *) (ref2 + 32));
ref3_reg = _mm256_loadu_si256((__m256i *) (ref3));
ref3next_reg = _mm256_loadu_si256((__m256i *) (ref3 + 32));
// sum of the absolute differences between every ref-i to src
ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg);
ref0next_reg = _mm256_sad_epu8(ref0next_reg, srcnext_reg);
ref1next_reg = _mm256_sad_epu8(ref1next_reg, srcnext_reg);
ref2next_reg = _mm256_sad_epu8(ref2next_reg, srcnext_reg);
ref3next_reg = _mm256_sad_epu8(ref3next_reg, srcnext_reg);
// sum every ref-i
sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg);
sum_ref0 = _mm256_add_epi32(sum_ref0, ref0next_reg);
sum_ref1 = _mm256_add_epi32(sum_ref1, ref1next_reg);
sum_ref2 = _mm256_add_epi32(sum_ref2, ref2next_reg);
sum_ref3 = _mm256_add_epi32(sum_ref3, ref3next_reg);
src+= src_stride;
ref0+= ref_stride;
ref1+= ref_stride;
ref2+= ref_stride;
ref3+= ref_stride;
}
{
__m128i sum;
// in sum_ref-i the result is saved in the first 4 bytes
// the other 4 bytes are zeroed.
// sum_ref1 and sum_ref3 are shifted left by 4 bytes
sum_ref1 = _mm256_slli_si256(sum_ref1, 4);
sum_ref3 = _mm256_slli_si256(sum_ref3, 4);
// merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3
sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1);
sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3);
// merge every 64 bit from each sum_ref-i
sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2);
sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2);
// add the low 64 bit to the high 64 bit
sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh);
// add the low 128 bit to the high 128 bit
sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow),
_mm256_extractf128_si256(sum_mlow, 1));
_mm_storeu_si128((__m128i *)(res), sum);
}
}

1
vp9/vp9cx.mk
View File

@ -89,6 +89,7 @@ VP9_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp9_sad_mmx.asm
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_variance_impl_sse2.asm
VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_variance_impl_intrin_avx2.c
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_sad4d_sse2.asm
VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_sad4d_intrin_avx2.c
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_subpel_variance_impl_sse2.asm
VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_subpel_variance_impl_intrin_avx2.c
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_temporal_filter_apply_sse2.asm