vpx/vp9/encoder/x86/vp9_variance_impl_intrin_avx2.c
levytamar82 357b65369f AVX2 Variance Optimization
Optimizing the variance functions: vp9_variance16x16, vp9_variance32x32,
vp9_variance64x64, vp9_variance32x16, vp9_variance64x32,
vp9_mse16x16 by migrating to AVX2
some of the functions were optimized by processing 32 elements instead of 16.
some of the functions were optimized by processing 2 loop strides of 16
elements in a single 256 bit register
This optimization gives between 2.4% - 2.7% user level performance gain
and 42% function level gain.

Change-Id: I265ae08a2b0196057a224a86450153ef3aebd85d
2014-01-08 12:05:53 -07:00

214 lines
8.4 KiB
C

/*
* 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> // AVX2
void vp9_get16x16var_avx2(const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *SSE,
int *Sum) {
__m256i src, src_expand_low, src_expand_high, ref, ref_expand_low;
__m256i ref_expand_high, madd_low, madd_high;
unsigned int i, src_2strides, ref_2strides;
__m256i zero_reg = _mm256_set1_epi16(0);
__m256i sum_ref_src = _mm256_set1_epi16(0);
__m256i madd_ref_src = _mm256_set1_epi16(0);
// processing two strides in a 256 bit register reducing the number
// of loop stride by half (comparing to the sse2 code)
src_2strides = source_stride << 1;
ref_2strides = recon_stride << 1;
for (i = 0; i < 8; i++) {
src = _mm256_castsi128_si256(
_mm_loadu_si128((__m128i const *) (src_ptr)));
src = _mm256_inserti128_si256(src,
_mm_loadu_si128((__m128i const *)(src_ptr+source_stride)), 1);
ref =_mm256_castsi128_si256(
_mm_loadu_si128((__m128i const *) (ref_ptr)));
ref = _mm256_inserti128_si256(ref,
_mm_loadu_si128((__m128i const *)(ref_ptr+recon_stride)), 1);
// expanding to 16 bit each lane
src_expand_low = _mm256_unpacklo_epi8(src, zero_reg);
src_expand_high = _mm256_unpackhi_epi8(src, zero_reg);
ref_expand_low = _mm256_unpacklo_epi8(ref, zero_reg);
ref_expand_high = _mm256_unpackhi_epi8(ref, zero_reg);
// src-ref
src_expand_low = _mm256_sub_epi16(src_expand_low, ref_expand_low);
src_expand_high = _mm256_sub_epi16(src_expand_high, ref_expand_high);
// madd low (src - ref)
madd_low = _mm256_madd_epi16(src_expand_low, src_expand_low);
// add high to low
src_expand_low = _mm256_add_epi16(src_expand_low, src_expand_high);
// madd high (src - ref)
madd_high = _mm256_madd_epi16(src_expand_high, src_expand_high);
sum_ref_src = _mm256_add_epi16(sum_ref_src, src_expand_low);
// add high to low
madd_ref_src = _mm256_add_epi32(madd_ref_src,
_mm256_add_epi32(madd_low, madd_high));
src_ptr+= src_2strides;
ref_ptr+= ref_2strides;
}
{
__m128i sum_res, madd_res;
__m128i expand_sum_low, expand_sum_high, expand_sum;
__m128i expand_madd_low, expand_madd_high, expand_madd;
__m128i ex_expand_sum_low, ex_expand_sum_high, ex_expand_sum;
// extract the low lane and add it to the high lane
sum_res = _mm_add_epi16(_mm256_castsi256_si128(sum_ref_src),
_mm256_extractf128_si256(sum_ref_src, 1));
madd_res = _mm_add_epi32(_mm256_castsi256_si128(madd_ref_src),
_mm256_extractf128_si256(madd_ref_src, 1));
// padding each 2 bytes with another 2 zeroed bytes
expand_sum_low = _mm_unpacklo_epi16(_mm256_castsi256_si128(zero_reg),
sum_res);
expand_sum_high = _mm_unpackhi_epi16(_mm256_castsi256_si128(zero_reg),
sum_res);
// shifting the sign 16 bits right
expand_sum_low = _mm_srai_epi32(expand_sum_low, 16);
expand_sum_high = _mm_srai_epi32(expand_sum_high, 16);
expand_sum = _mm_add_epi32(expand_sum_low, expand_sum_high);
// expand each 32 bits of the madd result to 64 bits
expand_madd_low = _mm_unpacklo_epi32(madd_res,
_mm256_castsi256_si128(zero_reg));
expand_madd_high = _mm_unpackhi_epi32(madd_res,
_mm256_castsi256_si128(zero_reg));
expand_madd = _mm_add_epi32(expand_madd_low, expand_madd_high);
ex_expand_sum_low = _mm_unpacklo_epi32(expand_sum,
_mm256_castsi256_si128(zero_reg));
ex_expand_sum_high = _mm_unpackhi_epi32(expand_sum,
_mm256_castsi256_si128(zero_reg));
ex_expand_sum = _mm_add_epi32(ex_expand_sum_low, ex_expand_sum_high);
// shift 8 bytes eight
madd_res = _mm_srli_si128(expand_madd, 8);
sum_res = _mm_srli_si128(ex_expand_sum, 8);
madd_res = _mm_add_epi32(madd_res, expand_madd);
sum_res = _mm_add_epi32(sum_res, ex_expand_sum);
*((int*)SSE)= _mm_cvtsi128_si32(madd_res);
*((int*)Sum)= _mm_cvtsi128_si32(sum_res);
}
}
void vp9_get32x32var_avx2(const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *SSE,
int *Sum) {
__m256i src, src_expand_low, src_expand_high, ref, ref_expand_low;
__m256i ref_expand_high, madd_low, madd_high;
unsigned int i;
__m256i zero_reg = _mm256_set1_epi16(0);
__m256i sum_ref_src = _mm256_set1_epi16(0);
__m256i madd_ref_src = _mm256_set1_epi16(0);
// processing 32 elements in parallel
for (i = 0; i < 16; i++) {
src = _mm256_loadu_si256((__m256i const *) (src_ptr));
ref = _mm256_loadu_si256((__m256i const *) (ref_ptr));
// expanding to 16 bit each lane
src_expand_low = _mm256_unpacklo_epi8(src, zero_reg);
src_expand_high = _mm256_unpackhi_epi8(src, zero_reg);
ref_expand_low = _mm256_unpacklo_epi8(ref, zero_reg);
ref_expand_high = _mm256_unpackhi_epi8(ref, zero_reg);
// src-ref
src_expand_low = _mm256_sub_epi16(src_expand_low, ref_expand_low);
src_expand_high = _mm256_sub_epi16(src_expand_high, ref_expand_high);
// madd low (src - ref)
madd_low = _mm256_madd_epi16(src_expand_low, src_expand_low);
// add high to low
src_expand_low = _mm256_add_epi16(src_expand_low, src_expand_high);
// madd high (src - ref)
madd_high = _mm256_madd_epi16(src_expand_high, src_expand_high);
sum_ref_src = _mm256_add_epi16(sum_ref_src, src_expand_low);
// add high to low
madd_ref_src = _mm256_add_epi32(madd_ref_src,
_mm256_add_epi32(madd_low, madd_high));
src_ptr+= source_stride;
ref_ptr+= recon_stride;
}
{
__m256i expand_sum_low, expand_sum_high, expand_sum;
__m256i expand_madd_low, expand_madd_high, expand_madd;
__m256i ex_expand_sum_low, ex_expand_sum_high, ex_expand_sum;
// padding each 2 bytes with another 2 zeroed bytes
expand_sum_low = _mm256_unpacklo_epi16(zero_reg, sum_ref_src);
expand_sum_high = _mm256_unpackhi_epi16(zero_reg, sum_ref_src);
// shifting the sign 16 bits right
expand_sum_low = _mm256_srai_epi32(expand_sum_low, 16);
expand_sum_high = _mm256_srai_epi32(expand_sum_high, 16);
expand_sum = _mm256_add_epi32(expand_sum_low, expand_sum_high);
// expand each 32 bits of the madd result to 64 bits
expand_madd_low = _mm256_unpacklo_epi32(madd_ref_src, zero_reg);
expand_madd_high = _mm256_unpackhi_epi32(madd_ref_src, zero_reg);
expand_madd = _mm256_add_epi32(expand_madd_low, expand_madd_high);
ex_expand_sum_low = _mm256_unpacklo_epi32(expand_sum, zero_reg);
ex_expand_sum_high = _mm256_unpackhi_epi32(expand_sum, zero_reg);
ex_expand_sum = _mm256_add_epi32(ex_expand_sum_low, ex_expand_sum_high);
// shift 8 bytes eight
madd_ref_src = _mm256_srli_si256(expand_madd, 8);
sum_ref_src = _mm256_srli_si256(ex_expand_sum, 8);
madd_ref_src = _mm256_add_epi32(madd_ref_src, expand_madd);
sum_ref_src = _mm256_add_epi32(sum_ref_src, ex_expand_sum);
// extract the low lane and the high lane and add the results
*((int*)SSE)= _mm_cvtsi128_si32(_mm256_castsi256_si128(madd_ref_src)) +
_mm_cvtsi128_si32(_mm256_extractf128_si256(madd_ref_src, 1));
*((int*)Sum)= _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_ref_src)) +
_mm_cvtsi128_si32(_mm256_extractf128_si256(sum_ref_src, 1));
}
}