Move 32-wide SSE4.1 hor_sad to picture-sse41.c

It's not used by picture-avx2.c that also includes the header, so
it should not be in the header
This commit is contained in:
Pauli Oikkonen 2019-01-31 21:04:30 +02:00
parent dd7d989a39
commit 57db234d95
2 changed files with 79 additions and 79 deletions

View file

@ -83,6 +83,85 @@ static uint32_t ver_sad_sse41(const kvz_pixel *pic_data, const kvz_pixel *ref_da
return ver_sad_arbitrary(pic_data, ref_data, width, height, stride);
}
static uint32_t hor_sad_sse41_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_twoline_groups = height & ~1;
const int32_t height_residual_lines = height & 1;
__m128i nslo = _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7,
8, 9, 10, 11, 12, 13, 14, 15);
__m128i nshi = _mm_setr_epi8(16, 17, 18, 19, 20, 21, 22, 23,
24, 25, 26, 27, 28, 29, 30, 31);
__m128i epol_masklo, epol_maskhi;
int32_t border_pix_off;
if (left) {
border_pix_off = left;
__m128i first_valid_idx = _mm_set1_epi8(left);
epol_masklo = _mm_cmpgt_epi8(first_valid_idx, nslo);
epol_maskhi = _mm_cmpgt_epi8(first_valid_idx, nshi);
} else {
border_pix_off = width - (right + 1);
__m128i last_valid_idx = _mm_set1_epi8(border_pix_off);
epol_masklo = _mm_cmpgt_epi8(nslo, last_valid_idx);
epol_maskhi = _mm_cmpgt_epi8(nshi, last_valid_idx);
}
__m128i sse_inc = _mm_setzero_si128();
int32_t y;
for (y = 0; y < height_twoline_groups; y += 2) {
__m128i a = _mm_loadu_si128((__m128i *)(pic_data + (y + 0) * pic_stride + 0));
__m128i b = _mm_loadu_si128((__m128i *)(ref_data + (y + 0) * ref_stride + 0));
__m128i c = _mm_loadu_si128((__m128i *)(pic_data + (y + 0) * pic_stride + 16));
__m128i d = _mm_loadu_si128((__m128i *)(ref_data + (y + 0) * ref_stride + 16));
__m128i e = _mm_loadu_si128((__m128i *)(pic_data + (y + 1) * pic_stride + 0));
__m128i f = _mm_loadu_si128((__m128i *)(ref_data + (y + 1) * ref_stride + 0));
__m128i g = _mm_loadu_si128((__m128i *)(pic_data + (y + 1) * pic_stride + 16));
__m128i h = _mm_loadu_si128((__m128i *)(ref_data + (y + 1) * ref_stride + 16));
__m128i border_px_lo = _mm_set1_epi8 (*(uint8_t *)(ref_data + (y + 0) * ref_stride + border_pix_off));
__m128i border_px_hi = _mm_set1_epi8 (*(uint8_t *)(ref_data + (y + 1) * ref_stride + border_pix_off));
__m128i b_epol = _mm_blendv_epi8(b, border_px_lo, epol_masklo);
__m128i d_epol = _mm_blendv_epi8(d, border_px_lo, epol_maskhi);
__m128i f_epol = _mm_blendv_epi8(f, border_px_hi, epol_masklo);
__m128i h_epol = _mm_blendv_epi8(h, border_px_hi, epol_maskhi);
__m128i curr_sads_ab = _mm_sad_epu8(a, b_epol);
__m128i curr_sads_cd = _mm_sad_epu8(c, d_epol);
__m128i curr_sads_ef = _mm_sad_epu8(e, f_epol);
__m128i curr_sads_gh = _mm_sad_epu8(g, h_epol);
sse_inc = _mm_add_epi64(sse_inc, curr_sads_ab);
sse_inc = _mm_add_epi64(sse_inc, curr_sads_cd);
sse_inc = _mm_add_epi64(sse_inc, curr_sads_ef);
sse_inc = _mm_add_epi64(sse_inc, curr_sads_gh);
}
if (height_residual_lines) {
__m128i a = _mm_loadu_si128((__m128i *)(pic_data + (y + 0) * pic_stride + 0));
__m128i b = _mm_loadu_si128((__m128i *)(ref_data + (y + 0) * ref_stride + 0));
__m128i c = _mm_loadu_si128((__m128i *)(pic_data + (y + 0) * pic_stride + 16));
__m128i d = _mm_loadu_si128((__m128i *)(ref_data + (y + 0) * ref_stride + 16));
__m128i border_px = _mm_set1_epi8 (*(uint8_t *)(ref_data + (y + 0) * ref_stride + border_pix_off));
__m128i b_epol = _mm_blendv_epi8(b, border_px, epol_masklo);
__m128i d_epol = _mm_blendv_epi8(d, border_px, epol_maskhi);
__m128i curr_sads_ab = _mm_sad_epu8(a, b_epol);
__m128i curr_sads_cd = _mm_sad_epu8(c, d_epol);
sse_inc = _mm_add_epi64(sse_inc, curr_sads_ab);
sse_inc = _mm_add_epi64(sse_inc, curr_sads_cd);
}
__m128i sse_inc_2 = _mm_shuffle_epi32(sse_inc, _MM_SHUFFLE(1, 0, 3, 2));
__m128i sad = _mm_add_epi64 (sse_inc, sse_inc_2);
return _mm_cvtsi128_si32(sad);
}
static uint32_t hor_sad_sse41(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)

View file

@ -788,85 +788,6 @@ static uint32_t hor_sad_sse41_w16(const kvz_pixel *pic_data, const kvz_pixel *re
return _mm_cvtsi128_si32(sad);
}
static uint32_t hor_sad_sse41_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_twoline_groups = height & ~1;
const int32_t height_residual_lines = height & 1;
__m128i nslo = _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7,
8, 9, 10, 11, 12, 13, 14, 15);
__m128i nshi = _mm_setr_epi8(16, 17, 18, 19, 20, 21, 22, 23,
24, 25, 26, 27, 28, 29, 30, 31);
__m128i epol_masklo, epol_maskhi;
int32_t border_pix_off;
if (left) {
border_pix_off = left;
__m128i first_valid_idx = _mm_set1_epi8(left);
epol_masklo = _mm_cmpgt_epi8(first_valid_idx, nslo);
epol_maskhi = _mm_cmpgt_epi8(first_valid_idx, nshi);
} else {
border_pix_off = width - (right + 1);
__m128i last_valid_idx = _mm_set1_epi8(border_pix_off);
epol_masklo = _mm_cmpgt_epi8(nslo, last_valid_idx);
epol_maskhi = _mm_cmpgt_epi8(nshi, last_valid_idx);
}
__m128i sse_inc = _mm_setzero_si128();
int32_t y;
for (y = 0; y < height_twoline_groups; y += 2) {
__m128i a = _mm_loadu_si128((__m128i *)(pic_data + (y + 0) * pic_stride + 0));
__m128i b = _mm_loadu_si128((__m128i *)(ref_data + (y + 0) * ref_stride + 0));
__m128i c = _mm_loadu_si128((__m128i *)(pic_data + (y + 0) * pic_stride + 16));
__m128i d = _mm_loadu_si128((__m128i *)(ref_data + (y + 0) * ref_stride + 16));
__m128i e = _mm_loadu_si128((__m128i *)(pic_data + (y + 1) * pic_stride + 0));
__m128i f = _mm_loadu_si128((__m128i *)(ref_data + (y + 1) * ref_stride + 0));
__m128i g = _mm_loadu_si128((__m128i *)(pic_data + (y + 1) * pic_stride + 16));
__m128i h = _mm_loadu_si128((__m128i *)(ref_data + (y + 1) * ref_stride + 16));
__m128i border_px_lo = _mm_set1_epi8 (*(uint8_t *)(ref_data + (y + 0) * ref_stride + border_pix_off));
__m128i border_px_hi = _mm_set1_epi8 (*(uint8_t *)(ref_data + (y + 1) * ref_stride + border_pix_off));
__m128i b_epol = _mm_blendv_epi8(b, border_px_lo, epol_masklo);
__m128i d_epol = _mm_blendv_epi8(d, border_px_lo, epol_maskhi);
__m128i f_epol = _mm_blendv_epi8(f, border_px_hi, epol_masklo);
__m128i h_epol = _mm_blendv_epi8(h, border_px_hi, epol_maskhi);
__m128i curr_sads_ab = _mm_sad_epu8(a, b_epol);
__m128i curr_sads_cd = _mm_sad_epu8(c, d_epol);
__m128i curr_sads_ef = _mm_sad_epu8(e, f_epol);
__m128i curr_sads_gh = _mm_sad_epu8(g, h_epol);
sse_inc = _mm_add_epi64(sse_inc, curr_sads_ab);
sse_inc = _mm_add_epi64(sse_inc, curr_sads_cd);
sse_inc = _mm_add_epi64(sse_inc, curr_sads_ef);
sse_inc = _mm_add_epi64(sse_inc, curr_sads_gh);
}
if (height_residual_lines) {
__m128i a = _mm_loadu_si128((__m128i *)(pic_data + (y + 0) * pic_stride + 0));
__m128i b = _mm_loadu_si128((__m128i *)(ref_data + (y + 0) * ref_stride + 0));
__m128i c = _mm_loadu_si128((__m128i *)(pic_data + (y + 0) * pic_stride + 16));
__m128i d = _mm_loadu_si128((__m128i *)(ref_data + (y + 0) * ref_stride + 16));
__m128i border_px = _mm_set1_epi8 (*(uint8_t *)(ref_data + (y + 0) * ref_stride + border_pix_off));
__m128i b_epol = _mm_blendv_epi8(b, border_px, epol_masklo);
__m128i d_epol = _mm_blendv_epi8(d, border_px, epol_maskhi);
__m128i curr_sads_ab = _mm_sad_epu8(a, b_epol);
__m128i curr_sads_cd = _mm_sad_epu8(c, d_epol);
sse_inc = _mm_add_epi64(sse_inc, curr_sads_ab);
sse_inc = _mm_add_epi64(sse_inc, curr_sads_cd);
}
__m128i sse_inc_2 = _mm_shuffle_epi32(sse_inc, _MM_SHUFFLE(1, 0, 3, 2));
__m128i sad = _mm_add_epi64 (sse_inc, sse_inc_2);
return _mm_cvtsi128_si32(sad);
}
static uint32_t hor_sad_sse41_arbitrary(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)