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

Unroll 64-wide AVX2 SAD by 2

Browse source
This commit is contained in:
Pauli Oikkonen 2019-01-16 21:19:09 +02:00
parent 84cf771dea
commit cbca3347b5
1 changed files with 36 additions and 9 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

45
src/strategies/avx2/reg_sad_pow2_widths-avx2.h
View file

@ -61,17 +61,44 @@ static INLINE uint32_t reg_sad_w64(const kvz_pixel * const data1, const kvz_pixe
__m256i avx_inc = _mm256_setzero_si256();
int32_t y;
for (y = 0; y < height; y++) {
__m256i a = _mm256_loadu_si256((const __m256i *)(data1 + y * stride1));
__m256i b = _mm256_loadu_si256((const __m256i *)(data2 + y * stride2));
__m256i c = _mm256_loadu_si256((const __m256i *)(data1 + y * stride1 + 32));
__m256i d = _mm256_loadu_si256((const __m256i *)(data2 + y * stride2 + 32));
const int32_t height_twoline_groups = height & ~1;
const int32_t height_residual_lines = height & 1;
__m256i curr_sads_1 = _mm256_sad_epu8(a, b);
__m256i curr_sads_2 = _mm256_sad_epu8(c, d);
avx_inc = _mm256_add_epi64(avx_inc, curr_sads_1);
avx_inc = _mm256_add_epi64(avx_inc, curr_sads_2);
for (y = 0; y < height_twoline_groups; y += 2) {
__m256i a = _mm256_loadu_si256((const __m256i *)(data1 + (y + 0) * stride1));
__m256i b = _mm256_loadu_si256((const __m256i *)(data2 + (y + 0) * stride2));
__m256i c = _mm256_loadu_si256((const __m256i *)(data1 + (y + 0) * stride1 + 32));
__m256i d = _mm256_loadu_si256((const __m256i *)(data2 + (y + 0) * stride2 + 32));
__m256i e = _mm256_loadu_si256((const __m256i *)(data1 + (y + 1) * stride1));
__m256i f = _mm256_loadu_si256((const __m256i *)(data2 + (y + 1) * stride2));
__m256i g = _mm256_loadu_si256((const __m256i *)(data1 + (y + 1) * stride1 + 32));
__m256i h = _mm256_loadu_si256((const __m256i *)(data2 + (y + 1) * stride2 + 32));
__m256i curr_sads_ab = _mm256_sad_epu8(a, b);
__m256i curr_sads_cd = _mm256_sad_epu8(c, d);
__m256i curr_sads_ef = _mm256_sad_epu8(e, f);
__m256i curr_sads_gh = _mm256_sad_epu8(g, h);
avx_inc = _mm256_add_epi64(avx_inc, curr_sads_ab);
avx_inc = _mm256_add_epi64(avx_inc, curr_sads_cd);
avx_inc = _mm256_add_epi64(avx_inc, curr_sads_ef);
avx_inc = _mm256_add_epi64(avx_inc, curr_sads_gh);
}
if (height_residual_lines) {
for (; y < height; y++) {
__m256i a = _mm256_loadu_si256((const __m256i *)(data1 + (y + 0) * stride1));
__m256i b = _mm256_loadu_si256((const __m256i *)(data2 + (y + 0) * stride2));
__m256i c = _mm256_loadu_si256((const __m256i *)(data1 + (y + 0) * stride1 + 32));
__m256i d = _mm256_loadu_si256((const __m256i *)(data2 + (y + 0) * stride2 + 32));
__m256i curr_sads_ab = _mm256_sad_epu8(a, b);
__m256i curr_sads_cd = _mm256_sad_epu8(c, d);
avx_inc = _mm256_add_epi64(avx_inc, curr_sads_ab);
avx_inc = _mm256_add_epi64(avx_inc, curr_sads_cd);
}
}
__m128i inchi = _mm256_extracti128_si256(avx_inc, 1);
__m128i inclo = _mm256_castsi256_si128 (avx_inc);