/dports/graphics/dssim/dssim-3.1.2/cargo-crates/libaom-sys-0.9.1/vendor/av1/common/x86/ |
H A D | jnt_convolve_avx2.c | 105 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_x_avx2() 160 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_x_avx2() 296 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 298 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_y_avx2() 366 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 368 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_y_avx2() 475 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 545 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 721 const __m256i res_8 = in av1_dist_wtd_convolve_2d_avx2() local 752 const __m256i res_8 = in av1_dist_wtd_convolve_2d_avx2() local [all …]
|
H A D | jnt_convolve_sse2.c | 81 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_x_sse2() local 82 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_x_sse2() 138 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_x_sse2() local 139 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_x_sse2() 225 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_y_sse2() local 226 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_y_sse2() 254 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_y_sse2() local 255 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_y_sse2() 332 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_y_sse2() 361 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_y_sse2() [all …]
|
H A D | convolve_2d_sse2.c | 74 const __m128i res_8 = _mm_madd_epi16(src_8, coeffs[4]); in av1_convolve_2d_sr_12tap_sse2() local 82 _mm_add_epi32(_mm_add_epi32(res_8, res_10), res_0246); in av1_convolve_2d_sr_12tap_sse2() 157 const __m128i res_8 = _mm_madd_epi16(src_8, coeffs[4]); in av1_convolve_2d_sr_12tap_sse2() local 163 _mm_add_epi32(_mm_add_epi32(res_8, res_10), res_0246); in av1_convolve_2d_sr_12tap_sse2() 501 const __m128i res_8 = in av1_dist_wtd_convolve_2d_copy_sse2() local 504 _mm_store_si128((__m128i *)(&dst0[j]), res_8); in av1_dist_wtd_convolve_2d_copy_sse2() 532 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_2d_copy_sse2() local 535 _mm_storel_epi64((__m128i *)(&dst0[j]), res_8); in av1_dist_wtd_convolve_2d_copy_sse2() 537 *(uint32_t *)(&dst0[j]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_2d_copy_sse2()
|
H A D | reconinter_sse4.c | 141 const __m128i res_8 = _mm_packus_epi16(diff_const_16, diff_const_16); in av1_build_compound_diffwtd_mask_d16_sse4_1() local 147 _mm_storel_epi64(dst, res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1() 149 *(uint32_t *)dst = _mm_cvtsi128_si32(res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1()
|
/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_avx2.c | 109 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_x_avx2() 164 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_x_avx2() 305 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 307 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_y_avx2() 375 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 377 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_y_avx2() 484 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 554 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 730 const __m256i res_8 = in av1_dist_wtd_convolve_2d_avx2() local 761 const __m256i res_8 = in av1_dist_wtd_convolve_2d_avx2() local [all …]
|
H A D | jnt_convolve_sse2.c | 85 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_x_sse2() local 86 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_x_sse2() 142 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_x_sse2() local 143 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_x_sse2() 233 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_y_sse2() local 234 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_y_sse2() 262 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_y_sse2() local 263 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_y_sse2() 340 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_y_sse2() 369 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_y_sse2() [all …]
|
H A D | reconinter_sse4.c | 141 const __m128i res_8 = _mm_packus_epi16(diff_const_16, diff_const_16); in av1_build_compound_diffwtd_mask_d16_sse4_1() local 147 _mm_storel_epi64(dst, res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1() 149 *(uint32_t *)dst = _mm_cvtsi128_si32(res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1()
|
/dports/www/chromium-legacy/chromium-88.0.4324.182/third_party/libaom/source/libaom/av1/common/x86/ |
H A D | jnt_convolve_avx2.c | 105 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_x_avx2() 160 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_x_avx2() 296 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 298 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_y_avx2() 366 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 368 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_y_avx2() 475 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 545 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 721 const __m256i res_8 = in av1_dist_wtd_convolve_2d_avx2() local 752 const __m256i res_8 = in av1_dist_wtd_convolve_2d_avx2() local [all …]
|
H A D | jnt_convolve_sse2.c | 81 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_x_sse2() local 82 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_x_sse2() 138 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_x_sse2() local 139 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_x_sse2() 225 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_y_sse2() local 226 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_y_sse2() 254 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_y_sse2() local 255 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_y_sse2() 332 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_y_sse2() 361 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_y_sse2() [all …]
|
H A D | reconinter_sse4.c | 141 const __m128i res_8 = _mm_packus_epi16(diff_const_16, diff_const_16); in av1_build_compound_diffwtd_mask_d16_sse4_1() local 147 _mm_storel_epi64(dst, res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1() 149 *(uint32_t *)dst = _mm_cvtsi128_si32(res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1()
|
/dports/multimedia/aom/aom-3.2.0/av1/common/x86/ |
H A D | jnt_convolve_avx2.c | 105 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_x_avx2() 160 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_x_avx2() 296 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 298 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_y_avx2() 366 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 368 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_dist_wtd_convolve_y_avx2() 475 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 545 const __m256i res_8 = in av1_dist_wtd_convolve_y_avx2() local 721 const __m256i res_8 = in av1_dist_wtd_convolve_2d_avx2() local 752 const __m256i res_8 = in av1_dist_wtd_convolve_2d_avx2() local [all …]
|
H A D | jnt_convolve_sse2.c | 81 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_x_sse2() local 82 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_x_sse2() 138 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_x_sse2() local 139 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_x_sse2() 225 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_y_sse2() local 226 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_y_sse2() 254 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_y_sse2() local 255 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_y_sse2() 332 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_y_sse2() 361 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_y_sse2() [all …]
|
H A D | convolve_2d_sse2.c | 74 const __m128i res_8 = _mm_madd_epi16(src_8, coeffs[4]); in av1_convolve_2d_sr_12tap_sse2() local 82 _mm_add_epi32(_mm_add_epi32(res_8, res_10), res_0246); in av1_convolve_2d_sr_12tap_sse2() 157 const __m128i res_8 = _mm_madd_epi16(src_8, coeffs[4]); in av1_convolve_2d_sr_12tap_sse2() local 163 _mm_add_epi32(_mm_add_epi32(res_8, res_10), res_0246); in av1_convolve_2d_sr_12tap_sse2() 501 const __m128i res_8 = in av1_dist_wtd_convolve_2d_copy_sse2() local 504 _mm_store_si128((__m128i *)(&dst0[j]), res_8); in av1_dist_wtd_convolve_2d_copy_sse2() 532 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_dist_wtd_convolve_2d_copy_sse2() local 535 _mm_storel_epi64((__m128i *)(&dst0[j]), res_8); in av1_dist_wtd_convolve_2d_copy_sse2() 537 *(uint32_t *)(&dst0[j]) = _mm_cvtsi128_si32(res_8); in av1_dist_wtd_convolve_2d_copy_sse2()
|
H A D | reconinter_sse4.c | 141 const __m128i res_8 = _mm_packus_epi16(diff_const_16, diff_const_16); in av1_build_compound_diffwtd_mask_d16_sse4_1() local 147 _mm_storel_epi64(dst, res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1() 149 *(uint32_t *)dst = _mm_cvtsi128_si32(res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1()
|
/dports/www/firefox-esr/firefox-91.8.0/third_party/aom/av1/common/x86/ |
H A D | jnt_convolve_avx2.c | 106 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_x_avx2() 107 const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); in av1_jnt_convolve_x_avx2() 240 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_y_avx2() 241 const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); in av1_jnt_convolve_y_avx2() 304 const __m256i res_8 = in av1_jnt_convolve_y_avx2() local 306 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_y_avx2() 457 const __m256i res_8 = in av1_jnt_convolve_2d_avx2() local 459 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_2d_avx2() 488 const __m256i res_8 = in av1_jnt_convolve_2d_avx2() local 490 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_2d_avx2() [all …]
|
H A D | jnt_convolve_sse2.c | 85 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_x_sse2() local 86 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_jnt_convolve_x_sse2() 142 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_x_sse2() local 143 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_jnt_convolve_x_sse2() 233 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_y_sse2() local 234 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_jnt_convolve_y_sse2() 262 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_y_sse2() local 263 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_jnt_convolve_y_sse2() 339 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_y_sse2() local 340 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_jnt_convolve_y_sse2() [all …]
|
H A D | reconinter_sse4.c | 141 const __m128i res_8 = _mm_packus_epi16(diff_const_16, diff_const_16); in av1_build_compound_diffwtd_mask_d16_sse4_1() local 147 _mm_storel_epi64(dst, res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1() 149 *(uint32_t *)dst = _mm_cvtsi128_si32(res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1()
|
/dports/www/firefox/firefox-99.0/third_party/aom/av1/common/x86/ |
H A D | jnt_convolve_avx2.c | 106 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_x_avx2() 107 const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); in av1_jnt_convolve_x_avx2() 240 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_y_avx2() 241 const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); in av1_jnt_convolve_y_avx2() 304 const __m256i res_8 = in av1_jnt_convolve_y_avx2() local 306 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_y_avx2() 457 const __m256i res_8 = in av1_jnt_convolve_2d_avx2() local 459 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_2d_avx2() 488 const __m256i res_8 = in av1_jnt_convolve_2d_avx2() local 490 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_2d_avx2() [all …]
|
H A D | jnt_convolve_sse2.c | 85 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_x_sse2() local 86 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_jnt_convolve_x_sse2() 142 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_x_sse2() local 143 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_jnt_convolve_x_sse2() 233 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_y_sse2() local 234 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_jnt_convolve_y_sse2() 262 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_y_sse2() local 263 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_jnt_convolve_y_sse2() 339 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_y_sse2() local 340 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_jnt_convolve_y_sse2() [all …]
|
H A D | reconinter_sse4.c | 141 const __m128i res_8 = _mm_packus_epi16(diff_const_16, diff_const_16); in av1_build_compound_diffwtd_mask_d16_sse4_1() local 147 _mm_storel_epi64(dst, res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1() 149 *(uint32_t *)dst = _mm_cvtsi128_si32(res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1()
|
/dports/mail/thunderbird/thunderbird-91.8.0/third_party/aom/av1/common/x86/ |
H A D | jnt_convolve_avx2.c | 106 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_x_avx2() 107 const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); in av1_jnt_convolve_x_avx2() 240 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_y_avx2() 241 const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); in av1_jnt_convolve_y_avx2() 304 const __m256i res_8 = in av1_jnt_convolve_y_avx2() local 306 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_y_avx2() 457 const __m256i res_8 = in av1_jnt_convolve_2d_avx2() local 459 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_2d_avx2() 488 const __m256i res_8 = in av1_jnt_convolve_2d_avx2() local 490 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_2d_avx2() [all …]
|
H A D | jnt_convolve_sse2.c | 85 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_x_sse2() local 86 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_jnt_convolve_x_sse2() 142 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_x_sse2() local 143 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_jnt_convolve_x_sse2() 233 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_y_sse2() local 234 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_jnt_convolve_y_sse2() 262 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_y_sse2() local 263 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_jnt_convolve_y_sse2() 339 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_y_sse2() local 340 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_jnt_convolve_y_sse2() [all …]
|
H A D | reconinter_sse4.c | 141 const __m128i res_8 = _mm_packus_epi16(diff_const_16, diff_const_16); in av1_build_compound_diffwtd_mask_d16_sse4_1() local 147 _mm_storel_epi64(dst, res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1() 149 *(uint32_t *)dst = _mm_cvtsi128_si32(res_8); in av1_build_compound_diffwtd_mask_d16_sse4_1()
|
/dports/lang/spidermonkey78/firefox-78.9.0/third_party/aom/av1/common/x86/ |
H A D | jnt_convolve_avx2.c | 106 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_x_avx2() 107 const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); in av1_jnt_convolve_x_avx2() 240 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_y_avx2() 241 const __m128i res_1 = _mm256_extracti128_si256(res_8, 1); in av1_jnt_convolve_y_avx2() 304 const __m256i res_8 = in av1_jnt_convolve_y_avx2() local 306 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_y_avx2() 457 const __m256i res_8 = in av1_jnt_convolve_2d_avx2() local 459 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_2d_avx2() 488 const __m256i res_8 = in av1_jnt_convolve_2d_avx2() local 490 const __m128i res_0 = _mm256_castsi256_si128(res_8); in av1_jnt_convolve_2d_avx2() [all …]
|
H A D | jnt_convolve_sse2.c | 85 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_x_sse2() local 86 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_jnt_convolve_x_sse2() 142 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_x_sse2() local 143 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_jnt_convolve_x_sse2() 233 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_y_sse2() local 234 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_jnt_convolve_y_sse2() 262 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_y_sse2() local 263 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); in av1_jnt_convolve_y_sse2() 339 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); in av1_jnt_convolve_y_sse2() local 340 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_jnt_convolve_y_sse2() [all …]
|