/dports/lang/spidermonkey60/firefox-60.9.0/third_party/aom/av1/common/x86/ |
H A D | highbd_convolve_2d_ssse3.c | 31 int im_stride = MAX_SB_SIZE; in av1_highbd_convolve_2d_ssse3() local 134 const int16_t *data = &im_block[i * im_stride + j]; in av1_highbd_convolve_2d_ssse3() 136 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_highbd_convolve_2d_ssse3() 139 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_highbd_convolve_2d_ssse3() 142 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_highbd_convolve_2d_ssse3() 145 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_highbd_convolve_2d_ssse3() 158 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_highbd_convolve_2d_ssse3() 161 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_highbd_convolve_2d_ssse3() 164 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), in av1_highbd_convolve_2d_ssse3() 212 int im_stride = MAX_SB_SIZE; in av1_highbd_convolve_2d_ssse3() local [all …]
|
H A D | convolve_2d_sse2.c | 30 int im_stride = MAX_SB_SIZE; in av1_convolve_2d_sse2() local 133 const uint8_t *data = &im_block[i * im_stride + j]; in av1_convolve_2d_sse2() 135 _mm_loadl_epi64((__m128i *)(data + 0 * im_stride)), in av1_convolve_2d_sse2() 136 _mm_loadl_epi64((__m128i *)(data + 1 * im_stride))); in av1_convolve_2d_sse2() 138 _mm_loadl_epi64((__m128i *)(data + 2 * im_stride)), in av1_convolve_2d_sse2() 139 _mm_loadl_epi64((__m128i *)(data + 3 * im_stride))); in av1_convolve_2d_sse2() 141 _mm_loadl_epi64((__m128i *)(data + 4 * im_stride)), in av1_convolve_2d_sse2() 142 _mm_loadl_epi64((__m128i *)(data + 5 * im_stride))); in av1_convolve_2d_sse2() 144 _mm_loadl_epi64((__m128i *)(data + 6 * im_stride)), in av1_convolve_2d_sse2() 210 int im_stride = MAX_SB_SIZE; in av1_convolve_2d_sse2() local [all …]
|
/dports/graphics/dssim/dssim-3.1.2/cargo-crates/libaom-sys-0.9.1/vendor/av1/common/x86/ |
H A D | convolve_2d_sse2.c | 33 int im_stride = w; in av1_convolve_2d_sr_12tap_sse2() local 133 const int16_t *data = &im_block[i * im_stride + j]; in av1_convolve_2d_sr_12tap_sse2() 135 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 138 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 141 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 144 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 147 _mm_unpacklo_epi16(*(__m128i *)(data + 8 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 167 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 170 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 244 int im_stride = MAX_SB_SIZE; in av1_convolve_2d_sr_sse2() local [all …]
|
H A D | jnt_convolve_ssse3.c | 31 int im_stride = MAX_SB_SIZE; in av1_dist_wtd_convolve_2d_ssse3() local 151 const int16_t *data = &im_block[i * im_stride + j]; in av1_dist_wtd_convolve_2d_ssse3() 153 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 154 *(__m128i *)(data + 1 * im_stride)); in av1_dist_wtd_convolve_2d_ssse3() 156 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 159 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 162 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 175 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 178 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 181 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() [all …]
|
H A D | wiener_convolve_avx2.c | 57 int im_stride = 8; in av1_wiener_convolve_add_src_avx2() local 58 memset(im_block + (im_h * im_stride), 0, MAX_SB_SIZE); in av1_wiener_convolve_add_src_avx2() 147 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res_clamped); in av1_wiener_convolve_add_src_avx2() 152 __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); in av1_wiener_convolve_add_src_avx2() 153 __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); in av1_wiener_convolve_add_src_avx2() 169 const int16_t *data = &im_block[i * im_stride]; in av1_wiener_convolve_add_src_avx2() 172 _mm256_loadu_si256((__m256i *)(data + 6 * im_stride)); in av1_wiener_convolve_add_src_avx2() 174 _mm256_loadu_si256((__m256i *)(data + 7 * im_stride)); in av1_wiener_convolve_add_src_avx2() 216 const int16_t *data = &im_block[i * im_stride]; in av1_wiener_convolve_add_src_avx2() 217 const __m128i s6_ = _mm_loadu_si128((__m128i *)(data + 6 * im_stride)); in av1_wiener_convolve_add_src_avx2() [all …]
|
H A D | highbd_convolve_2d_ssse3.c | 30 int im_stride = 8; in av1_highbd_convolve_2d_sr_ssse3() local 101 _mm_store_si128((__m128i *)&im_block[i * im_stride], res); in av1_highbd_convolve_2d_sr_ssse3() 107 __m128i s0 = _mm_loadu_si128((__m128i *)(im_block + 0 * im_stride)); in av1_highbd_convolve_2d_sr_ssse3() 108 __m128i s1 = _mm_loadu_si128((__m128i *)(im_block + 1 * im_stride)); in av1_highbd_convolve_2d_sr_ssse3() 109 __m128i s2 = _mm_loadu_si128((__m128i *)(im_block + 2 * im_stride)); in av1_highbd_convolve_2d_sr_ssse3() 110 __m128i s3 = _mm_loadu_si128((__m128i *)(im_block + 3 * im_stride)); in av1_highbd_convolve_2d_sr_ssse3() 111 __m128i s4 = _mm_loadu_si128((__m128i *)(im_block + 4 * im_stride)); in av1_highbd_convolve_2d_sr_ssse3() 112 __m128i s5 = _mm_loadu_si128((__m128i *)(im_block + 5 * im_stride)); in av1_highbd_convolve_2d_sr_ssse3() 144 const int16_t *data = &im_block[i * im_stride]; in av1_highbd_convolve_2d_sr_ssse3() 286 _mm_store_si128((__m128i *)&im_block[i * im_stride], res); in av1_highbd_convolve_2d_sr_ssse3() [all …]
|
/dports/multimedia/aom/aom-3.2.0/av1/common/x86/ |
H A D | convolve_2d_sse2.c | 33 int im_stride = w; in av1_convolve_2d_sr_12tap_sse2() local 133 const int16_t *data = &im_block[i * im_stride + j]; in av1_convolve_2d_sr_12tap_sse2() 135 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 138 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 141 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 144 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 147 _mm_unpacklo_epi16(*(__m128i *)(data + 8 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 167 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 170 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_convolve_2d_sr_12tap_sse2() 244 int im_stride = MAX_SB_SIZE; in av1_convolve_2d_sr_sse2() local [all …]
|
H A D | jnt_convolve_ssse3.c | 31 int im_stride = MAX_SB_SIZE; in av1_dist_wtd_convolve_2d_ssse3() local 151 const int16_t *data = &im_block[i * im_stride + j]; in av1_dist_wtd_convolve_2d_ssse3() 153 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 154 *(__m128i *)(data + 1 * im_stride)); in av1_dist_wtd_convolve_2d_ssse3() 156 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 159 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 162 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 175 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 178 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 181 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() [all …]
|
H A D | wiener_convolve_avx2.c | 57 int im_stride = 8; in av1_wiener_convolve_add_src_avx2() local 58 memset(im_block + (im_h * im_stride), 0, MAX_SB_SIZE); in av1_wiener_convolve_add_src_avx2() 147 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res_clamped); in av1_wiener_convolve_add_src_avx2() 152 __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); in av1_wiener_convolve_add_src_avx2() 153 __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); in av1_wiener_convolve_add_src_avx2() 169 const int16_t *data = &im_block[i * im_stride]; in av1_wiener_convolve_add_src_avx2() 172 _mm256_loadu_si256((__m256i *)(data + 6 * im_stride)); in av1_wiener_convolve_add_src_avx2() 174 _mm256_loadu_si256((__m256i *)(data + 7 * im_stride)); in av1_wiener_convolve_add_src_avx2() 216 const int16_t *data = &im_block[i * im_stride]; in av1_wiener_convolve_add_src_avx2() 217 const __m128i s6_ = _mm_loadu_si128((__m128i *)(data + 6 * im_stride)); in av1_wiener_convolve_add_src_avx2() [all …]
|
H A D | highbd_convolve_2d_ssse3.c | 30 int im_stride = 8; in av1_highbd_convolve_2d_sr_ssse3() local 101 _mm_store_si128((__m128i *)&im_block[i * im_stride], res); in av1_highbd_convolve_2d_sr_ssse3() 107 __m128i s0 = _mm_loadu_si128((__m128i *)(im_block + 0 * im_stride)); in av1_highbd_convolve_2d_sr_ssse3() 108 __m128i s1 = _mm_loadu_si128((__m128i *)(im_block + 1 * im_stride)); in av1_highbd_convolve_2d_sr_ssse3() 109 __m128i s2 = _mm_loadu_si128((__m128i *)(im_block + 2 * im_stride)); in av1_highbd_convolve_2d_sr_ssse3() 110 __m128i s3 = _mm_loadu_si128((__m128i *)(im_block + 3 * im_stride)); in av1_highbd_convolve_2d_sr_ssse3() 111 __m128i s4 = _mm_loadu_si128((__m128i *)(im_block + 4 * im_stride)); in av1_highbd_convolve_2d_sr_ssse3() 112 __m128i s5 = _mm_loadu_si128((__m128i *)(im_block + 5 * im_stride)); in av1_highbd_convolve_2d_sr_ssse3() 144 const int16_t *data = &im_block[i * im_stride]; in av1_highbd_convolve_2d_sr_ssse3() 286 _mm_store_si128((__m128i *)&im_block[i * im_stride], res); in av1_highbd_convolve_2d_sr_ssse3() [all …]
|
/dports/graphics/dssim/dssim-3.1.2/cargo-crates/libaom-sys-0.9.1/vendor/aom_dsp/x86/ |
H A D | convolve_avx2.h | 80 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); 84 __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); \ 85 __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); \ 86 __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride)); \ 87 __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride)); \ 95 const int16_t *data = &im_block[i * im_stride]; \ 159 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); 245 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); 349 _mm_storel_epi64((__m128i *)&im_block[i * im_stride + im_stride], \ 357 im_block[i * im_stride + im_stride] = (uint16_t)horiz_2; \ [all …]
|
/dports/multimedia/aom/aom-3.2.0/aom_dsp/x86/ |
H A D | convolve_avx2.h | 80 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); 84 __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); \ 85 __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); \ 86 __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride)); \ 87 __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride)); \ 95 const int16_t *data = &im_block[i * im_stride]; \ 159 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); 245 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); 349 _mm_storel_epi64((__m128i *)&im_block[i * im_stride + im_stride], \ 357 im_block[i * im_stride + im_stride] = (uint16_t)horiz_2; \ [all …]
|
/dports/www/firefox-esr/firefox-91.8.0/third_party/aom/av1/common/x86/ |
H A D | jnt_convolve_ssse3.c | 32 int im_stride = MAX_SB_SIZE; in av1_jnt_convolve_2d_ssse3() local 152 const int16_t *data = &im_block[i * im_stride + j]; in av1_jnt_convolve_2d_ssse3() 154 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_jnt_convolve_2d_ssse3() 155 *(__m128i *)(data + 1 * im_stride)); in av1_jnt_convolve_2d_ssse3() 157 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_jnt_convolve_2d_ssse3() 160 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_jnt_convolve_2d_ssse3() 163 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_jnt_convolve_2d_ssse3() 176 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_jnt_convolve_2d_ssse3() 179 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_jnt_convolve_2d_ssse3() 182 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), in av1_jnt_convolve_2d_ssse3() [all …]
|
H A D | highbd_convolve_2d_sse4.c | 179 int im_stride = MAX_SB_SIZE; in av1_highbd_jnt_convolve_2d_sse4_1() local 301 const int16_t *data = &im_block[i * im_stride + j]; in av1_highbd_jnt_convolve_2d_sse4_1() 303 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_highbd_jnt_convolve_2d_sse4_1() 304 *(__m128i *)(data + 1 * im_stride)); in av1_highbd_jnt_convolve_2d_sse4_1() 306 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_highbd_jnt_convolve_2d_sse4_1() 309 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_highbd_jnt_convolve_2d_sse4_1() 312 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_highbd_jnt_convolve_2d_sse4_1() 325 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_highbd_jnt_convolve_2d_sse4_1() 328 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_highbd_jnt_convolve_2d_sse4_1() 331 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), in av1_highbd_jnt_convolve_2d_sse4_1() [all …]
|
/dports/www/qt5-webengine/qtwebengine-everywhere-src-5.15.2/src/3rdparty/chromium/third_party/libaom/source/libaom/av1/common/x86/ |
H A D | jnt_convolve_ssse3.c | 31 int im_stride = MAX_SB_SIZE; in av1_dist_wtd_convolve_2d_ssse3() local 151 const int16_t *data = &im_block[i * im_stride + j]; in av1_dist_wtd_convolve_2d_ssse3() 153 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 154 *(__m128i *)(data + 1 * im_stride)); in av1_dist_wtd_convolve_2d_ssse3() 156 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 159 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 162 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 175 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 178 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 181 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() [all …]
|
H A D | wiener_convolve_avx2.c | 57 int im_stride = 8; in av1_wiener_convolve_add_src_avx2() local 58 memset(im_block + (im_h * im_stride), 0, MAX_SB_SIZE); in av1_wiener_convolve_add_src_avx2() 147 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res_clamped); in av1_wiener_convolve_add_src_avx2() 152 __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); in av1_wiener_convolve_add_src_avx2() 153 __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); in av1_wiener_convolve_add_src_avx2() 169 const int16_t *data = &im_block[i * im_stride]; in av1_wiener_convolve_add_src_avx2() 172 _mm256_loadu_si256((__m256i *)(data + 6 * im_stride)); in av1_wiener_convolve_add_src_avx2() 174 _mm256_loadu_si256((__m256i *)(data + 7 * im_stride)); in av1_wiener_convolve_add_src_avx2() 216 const int16_t *data = &im_block[i * im_stride]; in av1_wiener_convolve_add_src_avx2() 217 const __m128i s6_ = _mm_loadu_si128((__m128i *)(data + 6 * im_stride)); in av1_wiener_convolve_add_src_avx2() [all …]
|
/dports/www/firefox/firefox-99.0/third_party/aom/av1/common/x86/ |
H A D | jnt_convolve_ssse3.c | 32 int im_stride = MAX_SB_SIZE; in av1_jnt_convolve_2d_ssse3() local 152 const int16_t *data = &im_block[i * im_stride + j]; in av1_jnt_convolve_2d_ssse3() 154 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_jnt_convolve_2d_ssse3() 155 *(__m128i *)(data + 1 * im_stride)); in av1_jnt_convolve_2d_ssse3() 157 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_jnt_convolve_2d_ssse3() 160 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_jnt_convolve_2d_ssse3() 163 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_jnt_convolve_2d_ssse3() 176 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_jnt_convolve_2d_ssse3() 179 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_jnt_convolve_2d_ssse3() 182 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), in av1_jnt_convolve_2d_ssse3() [all …]
|
/dports/mail/thunderbird/thunderbird-91.8.0/third_party/aom/av1/common/x86/ |
H A D | jnt_convolve_ssse3.c | 32 int im_stride = MAX_SB_SIZE; in av1_jnt_convolve_2d_ssse3() local 152 const int16_t *data = &im_block[i * im_stride + j]; in av1_jnt_convolve_2d_ssse3() 154 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_jnt_convolve_2d_ssse3() 155 *(__m128i *)(data + 1 * im_stride)); in av1_jnt_convolve_2d_ssse3() 157 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_jnt_convolve_2d_ssse3() 160 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_jnt_convolve_2d_ssse3() 163 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_jnt_convolve_2d_ssse3() 176 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_jnt_convolve_2d_ssse3() 179 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_jnt_convolve_2d_ssse3() 182 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), in av1_jnt_convolve_2d_ssse3() [all …]
|
/dports/www/chromium-legacy/chromium-88.0.4324.182/third_party/libaom/source/libaom/av1/common/x86/ |
H A D | jnt_convolve_ssse3.c | 31 int im_stride = MAX_SB_SIZE; in av1_dist_wtd_convolve_2d_ssse3() local 151 const int16_t *data = &im_block[i * im_stride + j]; in av1_dist_wtd_convolve_2d_ssse3() 153 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 154 *(__m128i *)(data + 1 * im_stride)); in av1_dist_wtd_convolve_2d_ssse3() 156 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 159 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 162 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 175 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 178 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() 181 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), in av1_dist_wtd_convolve_2d_ssse3() [all …]
|
H A D | wiener_convolve_avx2.c | 57 int im_stride = 8; in av1_wiener_convolve_add_src_avx2() local 58 memset(im_block + (im_h * im_stride), 0, MAX_SB_SIZE); in av1_wiener_convolve_add_src_avx2() 147 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res_clamped); in av1_wiener_convolve_add_src_avx2() 152 __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); in av1_wiener_convolve_add_src_avx2() 153 __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); in av1_wiener_convolve_add_src_avx2() 169 const int16_t *data = &im_block[i * im_stride]; in av1_wiener_convolve_add_src_avx2() 172 _mm256_loadu_si256((__m256i *)(data + 6 * im_stride)); in av1_wiener_convolve_add_src_avx2() 174 _mm256_loadu_si256((__m256i *)(data + 7 * im_stride)); in av1_wiener_convolve_add_src_avx2() 216 const int16_t *data = &im_block[i * im_stride]; in av1_wiener_convolve_add_src_avx2() 217 const __m128i s6_ = _mm_loadu_si128((__m128i *)(data + 6 * im_stride)); in av1_wiener_convolve_add_src_avx2() [all …]
|
H A D | convolve_2d_sse2.c | 32 int im_stride = MAX_SB_SIZE; in av1_convolve_2d_sr_sse2() local 143 const int16_t *data = &im_block[i * im_stride + j]; in av1_convolve_2d_sr_sse2() 145 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_convolve_2d_sr_sse2() 146 *(__m128i *)(data + 1 * im_stride)); in av1_convolve_2d_sr_sse2() 148 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_convolve_2d_sr_sse2() 151 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_convolve_2d_sr_sse2() 154 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_convolve_2d_sr_sse2() 167 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_convolve_2d_sr_sse2() 170 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_convolve_2d_sr_sse2() 173 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), in av1_convolve_2d_sr_sse2() [all …]
|
H A D | highbd_convolve_2d_avx2.c | 32 int im_stride = 8; in av1_highbd_convolve_2d_sr_avx2() local 102 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); in av1_highbd_convolve_2d_sr_avx2() 108 __m256i s0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); in av1_highbd_convolve_2d_sr_avx2() 109 __m256i s1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); in av1_highbd_convolve_2d_sr_avx2() 110 __m256i s2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride)); in av1_highbd_convolve_2d_sr_avx2() 111 __m256i s3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride)); in av1_highbd_convolve_2d_sr_avx2() 112 __m256i s4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride)); in av1_highbd_convolve_2d_sr_avx2() 113 __m256i s5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride)); in av1_highbd_convolve_2d_sr_avx2() 124 const int16_t *data = &im_block[i * im_stride]; in av1_highbd_convolve_2d_sr_avx2() 127 _mm256_loadu_si256((__m256i *)(data + 6 * im_stride)); in av1_highbd_convolve_2d_sr_avx2() [all …]
|
/dports/lang/spidermonkey78/firefox-78.9.0/third_party/aom/av1/common/x86/ |
H A D | jnt_convolve_ssse3.c | 32 int im_stride = MAX_SB_SIZE; in av1_jnt_convolve_2d_ssse3() local 152 const int16_t *data = &im_block[i * im_stride + j]; in av1_jnt_convolve_2d_ssse3() 154 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), in av1_jnt_convolve_2d_ssse3() 155 *(__m128i *)(data + 1 * im_stride)); in av1_jnt_convolve_2d_ssse3() 157 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), in av1_jnt_convolve_2d_ssse3() 160 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), in av1_jnt_convolve_2d_ssse3() 163 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), in av1_jnt_convolve_2d_ssse3() 176 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), in av1_jnt_convolve_2d_ssse3() 179 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), in av1_jnt_convolve_2d_ssse3() 182 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), in av1_jnt_convolve_2d_ssse3() [all …]
|
/dports/www/chromium-legacy/chromium-88.0.4324.182/third_party/libaom/source/libaom/aom_dsp/x86/ |
H A D | convolve_avx2.h | 74 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); \ 80 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); 84 __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); \ 85 __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); \ 86 __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride)); \ 87 __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride)); \ 95 const int16_t *data = &im_block[i * im_stride]; \ 96 const __m256i s4 = _mm256_loadu_si256((__m256i *)(data + 4 * im_stride)); \ 97 const __m256i s5 = _mm256_loadu_si256((__m256i *)(data + 5 * im_stride)); \ 159 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); [all …]
|
/dports/www/qt5-webengine/qtwebengine-everywhere-src-5.15.2/src/3rdparty/chromium/third_party/libaom/source/libaom/aom_dsp/x86/ |
H A D | convolve_avx2.h | 75 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); \ 85 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); 88 __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); \ 89 __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); \ 90 __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride)); \ 91 __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride)); \ 92 __m256i src_4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride)); \ 93 __m256i src_5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride)); \ 105 const int16_t *data = &im_block[i * im_stride]; \ 107 const __m256i s6 = _mm256_loadu_si256((__m256i *)(data + 6 * im_stride)); \ [all …]
|