mirror of
https://github.com/ultravideo/uvg266.git
synced 2024-11-24 02:24:07 +00:00
Implement 32-wide hor_sad on AVX2
This commit is contained in:
parent
ff70c8a5ec
commit
dd7d989a39
|
@ -1309,8 +1309,8 @@ static uint32_t hor_sad_avx2(const kvz_pixel *pic_data, const kvz_pixel *ref_dat
|
|||
return hor_sad_sse41_w16(pic_data, ref_data, width, height,
|
||||
pic_stride, ref_stride, left, right);
|
||||
if (width == 32)
|
||||
return hor_sad_sse41_w32(pic_data, ref_data, width, height,
|
||||
pic_stride, ref_stride, left, right);
|
||||
return hor_sad_avx2_w32(pic_data, ref_data, width, height,
|
||||
pic_stride, ref_stride, left, right);
|
||||
else
|
||||
return hor_sad_sse41_arbitrary(pic_data, ref_data, width, height,
|
||||
pic_stride, ref_stride, left, right);
|
||||
|
|
|
@ -109,4 +109,86 @@ static INLINE uint32_t reg_sad_w64(const kvz_pixel * const data1, const kvz_pixe
|
|||
return _mm_cvtsi128_si32(sad);
|
||||
}
|
||||
|
||||
static uint32_t hor_sad_avx2_w32(const kvz_pixel *pic_data, const kvz_pixel *ref_data,
|
||||
int32_t width, int32_t height, uint32_t pic_stride,
|
||||
uint32_t ref_stride, uint32_t left, uint32_t right)
|
||||
{
|
||||
const int32_t height_fourline_groups = height & ~3;
|
||||
const int32_t height_residual_lines = height & 3;
|
||||
|
||||
__m256i ns = _mm256_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7,
|
||||
8, 9, 10, 11, 12, 13, 14, 15,
|
||||
16, 17, 18, 19, 20, 21, 22, 23,
|
||||
24, 25, 26, 27, 28, 29, 30, 31);
|
||||
__m256i epol_mask;
|
||||
int32_t border_pix_off;
|
||||
|
||||
if (left) {
|
||||
border_pix_off = left;
|
||||
__m256i first_valid_idx = _mm256_set1_epi8(left);
|
||||
|
||||
epol_mask = _mm256_cmpgt_epi8(first_valid_idx, ns);
|
||||
} else {
|
||||
border_pix_off = width - (right + 1);
|
||||
__m256i last_valid_idx = _mm256_set1_epi8(border_pix_off);
|
||||
|
||||
epol_mask = _mm256_cmpgt_epi8(ns, last_valid_idx);
|
||||
}
|
||||
|
||||
__m256i avx_inc = _mm256_setzero_si256();
|
||||
int32_t y;
|
||||
for (y = 0; y < height_fourline_groups; y += 4) {
|
||||
__m256i a = _mm256_loadu_si256((__m256i *)(pic_data + (y + 0) * pic_stride));
|
||||
__m256i b = _mm256_loadu_si256((__m256i *)(ref_data + (y + 0) * ref_stride));
|
||||
__m256i c = _mm256_loadu_si256((__m256i *)(pic_data + (y + 1) * pic_stride));
|
||||
__m256i d = _mm256_loadu_si256((__m256i *)(ref_data + (y + 1) * ref_stride));
|
||||
__m256i e = _mm256_loadu_si256((__m256i *)(pic_data + (y + 2) * pic_stride));
|
||||
__m256i f = _mm256_loadu_si256((__m256i *)(ref_data + (y + 2) * ref_stride));
|
||||
__m256i g = _mm256_loadu_si256((__m256i *)(pic_data + (y + 3) * pic_stride));
|
||||
__m256i h = _mm256_loadu_si256((__m256i *)(ref_data + (y + 3) * ref_stride));
|
||||
|
||||
__m256i border_px_b = _mm256_set1_epi8 (*(uint8_t *)(ref_data + (y + 0) * ref_stride + border_pix_off));
|
||||
__m256i border_px_d = _mm256_set1_epi8 (*(uint8_t *)(ref_data + (y + 1) * ref_stride + border_pix_off));
|
||||
__m256i border_px_f = _mm256_set1_epi8 (*(uint8_t *)(ref_data + (y + 2) * ref_stride + border_pix_off));
|
||||
__m256i border_px_h = _mm256_set1_epi8 (*(uint8_t *)(ref_data + (y + 3) * ref_stride + border_pix_off));
|
||||
|
||||
__m256i b_epol = _mm256_blendv_epi8(b, border_px_b, epol_mask);
|
||||
__m256i d_epol = _mm256_blendv_epi8(d, border_px_d, epol_mask);
|
||||
__m256i f_epol = _mm256_blendv_epi8(f, border_px_f, epol_mask);
|
||||
__m256i h_epol = _mm256_blendv_epi8(h, border_px_h, epol_mask);
|
||||
|
||||
__m256i curr_sads_ab = _mm256_sad_epu8(a, b_epol);
|
||||
__m256i curr_sads_cd = _mm256_sad_epu8(c, d_epol);
|
||||
__m256i curr_sads_ef = _mm256_sad_epu8(e, f_epol);
|
||||
__m256i curr_sads_gh = _mm256_sad_epu8(g, h_epol);
|
||||
|
||||
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((__m256i *)(pic_data + y * pic_stride));
|
||||
__m256i b = _mm256_loadu_si256((__m256i *)(ref_data + y * ref_stride));
|
||||
|
||||
__m256i border_px_b = _mm256_set1_epi8 (*(uint8_t *)(ref_data + y * ref_stride + border_pix_off));
|
||||
__m256i b_epol = _mm256_blendv_epi8(b, border_px_b, epol_mask);
|
||||
|
||||
__m256i curr_sads_ab = _mm256_sad_epu8(a, b_epol);
|
||||
|
||||
avx_inc = _mm256_add_epi64(avx_inc, curr_sads_ab);
|
||||
}
|
||||
}
|
||||
|
||||
__m128i inchi = _mm256_extracti128_si256(avx_inc, 1);
|
||||
__m128i inclo = _mm256_castsi256_si128 (avx_inc);
|
||||
|
||||
__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);
|
||||
}
|
||||
|
||||
#endif
|
||||
|
|
Loading…
Reference in a new issue