hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-11-27 19:24:06 +00:00
Code Issues Actions Projects Releases Wiki Activity

Remove 16-pixel wide AVX2 SAD implementation

Browse source 
At least on Skylake, it's noticeably slower than the very simple
version using SSE4.1
This commit is contained in:
Pauli Oikkonen 2019-01-16 00:13:31 +02:00
parent 4cb371184b
commit 6504145cce
2 changed files with 1 additions and 39 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/strategies/avx2/picture-avx2.c
View file

@ -61,7 +61,7 @@ uint32_t kvz_reg_sad_avx2(const kvz_pixel * const data1, const kvz_pixel * const
if (width == 12)
return reg_sad_w12(data1, data2, height, stride1, stride2);
if (width == 16)
return reg_sad_w16_avx2(data1, data2, height, stride1, stride2);
return reg_sad_w16(data1, data2, height, stride1, stride2);
if (width == 24)
return reg_sad_w24(data1, data2, height, stride1, stride2);
if (width == 32)

38
src/strategies/avx2/reg_sad_pow2_widths-avx2.h
View file

@ -4,44 +4,6 @@
#include "strategies/sse41/reg_sad_pow2_widths-sse41.h"
#include "kvazaar.h"
static uint32_t reg_sad_w16_avx2(const kvz_pixel * const data1, const kvz_pixel * const data2,
const int32_t height, const uint32_t stride1,
const uint32_t stride2)
{
__m256i avx_inc = _mm256_setzero_si256();
int32_t y;
const int32_t height_ymm_bytes = height & ~1;
const int32_t height_parity = height & 1;
for (y = 0; y < height_ymm_bytes; y += 2) {
__m128i a_up = _mm_loadu_si128((const __m128i *)(data1 + (y + 0) * stride1));
__m128i b_up = _mm_loadu_si128((const __m128i *)(data2 + (y + 0) * stride2));
__m128i a_dn = _mm_loadu_si128((const __m128i *)(data1 + (y + 1) * stride1));
__m128i b_dn = _mm_loadu_si128((const __m128i *)(data2 + (y + 1) * stride2));
__m256i a = _mm256_inserti128_si256(_mm256_castsi128_si256(a_up), a_dn, 1);
__m256i b = _mm256_inserti128_si256(_mm256_castsi128_si256(b_up), b_dn, 1);
__m256i curr_sads = _mm256_sad_epu8(a, b);
avx_inc = _mm256_add_epi64(avx_inc, curr_sads);
}
__m128i inchi = _mm256_extracti128_si256(avx_inc, 1);
__m128i inclo = _mm256_castsi256_si128 (avx_inc);
if (height_parity) {
__m128i a = _mm_loadu_si128 ((__m128i *)(data1 + y * stride1));
__m128i b = _mm_loadu_si128 ((__m128i *)(data2 + y * stride2));
__m128i sads = _mm_sad_epu8 (a, b);
inclo = _mm_add_epi64(inclo, sads);
}
__m128i sum_1 = _mm_add_epi64 (inclo, inchi);
__m128i sum_2 = _mm_shuffle_epi32(sum_1, _MM_SHUFFLE(1, 0, 3, 2));
__m128i sad = _mm_add_epi64 (sum_1, sum_2);
return _mm_cvtsi128_si32(sad);
}
static uint32_t reg_sad_w32(const kvz_pixel * const data1, const kvz_pixel * const data2,
const int32_t height, const uint32_t stride1,
const uint32_t stride2)