| /external/libaom/libaom/av1/common/x86/ |
| D | jnt_convolve_avx2.c | 39 uint8_t *dst0, int dst_stride0, int w, int h, in av1_dist_wtd_convolve_x_avx2() argument
113 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); in av1_dist_wtd_convolve_x_avx2()
115 (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1); in av1_dist_wtd_convolve_x_avx2()
117 *(uint32_t *)(&dst0[i * dst_stride0 + j]) = in av1_dist_wtd_convolve_x_avx2()
119 *(uint32_t *)(&dst0[i * dst_stride0 + j + dst_stride0]) = in av1_dist_wtd_convolve_x_avx2()
168 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); in av1_dist_wtd_convolve_x_avx2()
170 (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1); in av1_dist_wtd_convolve_x_avx2()
172 *(uint32_t *)(&dst0[i * dst_stride0 + j]) = in av1_dist_wtd_convolve_x_avx2()
174 *(uint32_t *)(&dst0[i * dst_stride0 + j + dst_stride0]) = in av1_dist_wtd_convolve_x_avx2()
191 uint8_t *dst0, int dst_stride0, int w, int h, in av1_dist_wtd_convolve_y_avx2() argument
[all …]
|
| D | highbd_jnt_convolve_avx2.c | 26 const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w, in av1_highbd_dist_wtd_convolve_2d_copy_avx2() argument
102 _mm256_store_si256((__m256i *)(&dst0[i * dst_stride0 + j]), res_clip); in av1_highbd_dist_wtd_convolve_2d_copy_avx2()
155 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); in av1_highbd_dist_wtd_convolve_2d_copy_avx2()
157 (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), res_1); in av1_highbd_dist_wtd_convolve_2d_copy_avx2()
212 _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); in av1_highbd_dist_wtd_convolve_2d_copy_avx2()
214 (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), res_1); in av1_highbd_dist_wtd_convolve_2d_copy_avx2()
232 const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w, in av1_highbd_dist_wtd_convolve_2d_avx2() argument
387 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); in av1_highbd_dist_wtd_convolve_2d_avx2()
389 (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), res_1); in av1_highbd_dist_wtd_convolve_2d_avx2()
440 _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); in av1_highbd_dist_wtd_convolve_2d_avx2()
[all …]
|
| D | highbd_jnt_convolve_sse4.c | 21 const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w, in av1_highbd_dist_wtd_convolve_y_sse4_1() argument
147 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), in av1_highbd_dist_wtd_convolve_y_sse4_1()
150 (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), in av1_highbd_dist_wtd_convolve_y_sse4_1()
225 _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), in av1_highbd_dist_wtd_convolve_y_sse4_1()
228 (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), in av1_highbd_dist_wtd_convolve_y_sse4_1()
263 const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w, in av1_highbd_dist_wtd_convolve_x_sse4_1() argument
350 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip); in av1_highbd_dist_wtd_convolve_x_sse4_1()
379 _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip); in av1_highbd_dist_wtd_convolve_x_sse4_1()
|
| D | highbd_convolve_2d_sse4.c | 25 const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w, in av1_highbd_dist_wtd_convolve_2d_copy_sse4_1() argument
99 _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip); in av1_highbd_dist_wtd_convolve_2d_copy_sse4_1()
152 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip); in av1_highbd_dist_wtd_convolve_2d_copy_sse4_1()
154 (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), res_1); in av1_highbd_dist_wtd_convolve_2d_copy_sse4_1()
172 const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w, in av1_highbd_dist_wtd_convolve_2d_sse4_1() argument
375 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
415 _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
|
| D | jnt_convolve_sse2.c | 20 uint8_t *dst0, int dst_stride0, int w, int h, in av1_dist_wtd_convolve_x_sse2() argument
92 dst0 += dst_stride0; in av1_dist_wtd_convolve_x_sse2()
143 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_x_sse2()
154 uint8_t *dst0, int dst_stride0, int w, int h, in av1_dist_wtd_convolve_y_sse2() argument
242 dst0 += dst_stride0; in av1_dist_wtd_convolve_y_sse2()
271 dst0 += dst_stride0; 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()
388 uint8_t *dst0, int dst_stride0, int w, int h, in av1_dist_wtd_convolve_2d_sse2() argument
605 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_2d_sse2()
[all …]
|
| D | jnt_convolve_ssse3.c | 20 const uint8_t *src, int src_stride, uint8_t *dst0, int dst_stride0, int w, in av1_dist_wtd_convolve_2d_ssse3() argument
221 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); in av1_dist_wtd_convolve_2d_ssse3()
223 *(uint32_t *)(&dst0[i * dst_stride0 + j]) = in av1_dist_wtd_convolve_2d_ssse3()
|
| D | convolve_2d_sse2.c | 358 const uint8_t *src, int src_stride, uint8_t *dst0, int dst_stride0, int w, in av1_dist_wtd_convolve_2d_copy_sse2() argument
436 dst0 += dst_stride0; in av1_dist_wtd_convolve_2d_copy_sse2()
468 dst0 += dst_stride0; in av1_dist_wtd_convolve_2d_copy_sse2()
|
| /external/libaom/libaom/aom_dsp/x86/ |
| D | convolve_avx2.h | 194 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0); \
196 (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1); \
223 *(uint32_t *)(&dst0[i * dst_stride0 + j]) = _mm_cvtsi128_si32(res_0); \
224 *(uint32_t *)(&dst0[i * dst_stride0 + j + dst_stride0]) = \
|