| /external/libaom/libaom/av1/common/x86/ |
| D | wiener_convolve_avx2.c | 189 const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round); in av1_wiener_convolve_add_src_avx2() local
191 const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit); in av1_wiener_convolve_add_src_avx2()
233 const __m128i res_16bit = _mm_packus_epi32(reslo, reshi); in av1_wiener_convolve_add_src_avx2() local
236 const __m128i res_8b = _mm_packus_epi16(res_16bit, res_16bit); in av1_wiener_convolve_add_src_avx2()
|
| D | highbd_convolve_2d_avx2.c | 149 __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round); in av1_highbd_convolve_2d_sr_avx2() local
150 res_16bit = _mm256_min_epi16(res_16bit, clip_pixel); in av1_highbd_convolve_2d_sr_avx2()
151 res_16bit = _mm256_max_epi16(res_16bit, zero); in av1_highbd_convolve_2d_sr_avx2()
154 _mm256_castsi256_si128(res_16bit)); in av1_highbd_convolve_2d_sr_avx2()
156 _mm256_extracti128_si256(res_16bit, 1)); in av1_highbd_convolve_2d_sr_avx2()
|
| D | highbd_wiener_convolve_ssse3.c | 194 __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round); in av1_highbd_wiener_convolve_add_src_ssse3() local
195 res_16bit = _mm_min_epi16(_mm_max_epi16(res_16bit, zero), maxval); in av1_highbd_wiener_convolve_add_src_ssse3()
198 _mm_storeu_si128(p, res_16bit); in av1_highbd_wiener_convolve_add_src_ssse3()
|
| D | wiener_convolve_sse2.c | 191 const __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round); in av1_wiener_convolve_add_src_sse2() local
192 __m128i res_8bit = _mm_packus_epi16(res_16bit, res_16bit); in av1_wiener_convolve_add_src_sse2()
|
| D | convolve_2d_avx2.c | 156 const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round); in av1_convolve_2d_sr_avx2() local
158 const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit); in av1_convolve_2d_sr_avx2()
|
| D | highbd_warp_plane_sse4.c | 611 __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round); in av1_highbd_warp_affine_sse4_1() local
615 res_16bit = _mm_max_epi16(_mm_min_epi16(res_16bit, max_val), zero); in av1_highbd_warp_affine_sse4_1()
624 _mm_storel_epi64(p, res_16bit); in av1_highbd_warp_affine_sse4_1()
626 _mm_storeu_si128(p, res_16bit); in av1_highbd_warp_affine_sse4_1()
|
| D | highbd_wiener_convolve_avx2.c | 235 const __m256i res_16bit = in av1_highbd_wiener_convolve_add_src_avx2() local
238 _mm256_max_epi16(res_16bit, clamp_low), clamp_high); in av1_highbd_wiener_convolve_add_src_avx2()
|
| D | warp_plane_sse4.c | 660 const __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round); in store_vertical_filter_output() local
661 __m128i res_8bit = _mm_packus_epi16(res_16bit, res_16bit); in store_vertical_filter_output()
|
| D | warp_plane_avx2.c | 760 const __m256i res_16bit = _mm256_packs_epi32(res_lo_round, res_hi_round); in store_vertical_filter_output_avx2() local
761 const __m256i res_8bit = _mm256_packus_epi16(res_16bit, res_16bit); in store_vertical_filter_output_avx2()
|
| /external/libaom/libaom/aom_dsp/x86/ |
| D | highbd_convolve_avx2.c | 223 __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round); in av1_highbd_convolve_y_sr_avx2() local
224 res_16bit = _mm256_min_epi16(res_16bit, clip_pixel); in av1_highbd_convolve_y_sr_avx2()
225 res_16bit = _mm256_max_epi16(res_16bit, zero); in av1_highbd_convolve_y_sr_avx2()
228 _mm256_castsi256_si128(res_16bit)); in av1_highbd_convolve_y_sr_avx2()
230 _mm256_extracti128_si256(res_16bit, 1)); in av1_highbd_convolve_y_sr_avx2()
1112 __m256i res_16bit = _mm256_min_epi32(res_a_round, clip_pixel); in aom_highbd_filter_block1d4_v4_avx2() local
1113 res_16bit = _mm256_max_epi32(res_16bit, zero); in aom_highbd_filter_block1d4_v4_avx2()
1114 res_16bit = _mm256_packs_epi32(res_16bit, res_16bit); in aom_highbd_filter_block1d4_v4_avx2()
1117 _mm256_castsi256_si128(res_16bit)); in aom_highbd_filter_block1d4_v4_avx2()
1119 _mm256_extracti128_si256(res_16bit, 1)); in aom_highbd_filter_block1d4_v4_avx2()
[all …]
|
| D | convolve_avx2.h | 126 const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round); \
127 const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit); \
|