| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-sse-1x4.c | 115 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4() local 176 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4() local
|
| D | 3x3p1-minmax-sse-1x4-acc2.c | 115 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2() local 177 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2() local
|
| D | 3x3p1-minmax-sse-1x4-acc3.c | 115 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3() local 178 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3() local
|
| D | 3x3p1-minmax-sse-1x4-acc4.c | 115 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local 179 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local
|
| D | 3x3p1-minmax-sse-2x4.c | 139 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local 223 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local
|
| D | 3x3p1-minmax-sse-2x4-acc2.c | 139 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 225 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
|
| D | 5x5p2-minmax-sse-1x4.c | 148 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 245 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 336 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
|
| D | 5x5p2-minmax-sse-1x4-acc2.c | 148 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 246 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 338 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
|
| D | 3x3p1-minmax-sse-3x4.c | 163 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local 270 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local
|
| D | 5x5p2-minmax-sse-1x4-acc3.c | 148 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 247 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 340 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local
|
| D | 5x5p2-minmax-sse-1x4-acc4.c | 148 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 248 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 342 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local
|
| D | 5x5p2-minmax-sse-1x4-acc5.c | 148 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 249 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 344 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local
|
| D | 5x5p2-minmax-sse-2x4.c | 173 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 309 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 437 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
|
| D | 3x3p1-minmax-sse-4x4.c | 187 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local 317 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
|
| D | 5x5p2-minmax-sse-2x4-acc2.c | 173 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 311 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 441 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
|
| D | 5x5p2-minmax-sse-2x4-acc3.c | 173 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 313 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 445 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
|
| D | 3x3p1-minmax-sse-5x4.c | 211 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local 364 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
|
| D | 5x5p2-minmax-sse-3x4.c | 198 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 373 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 538 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
|
| D | 5x5p2-minmax-sse-3x4-acc2.c | 198 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 376 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 544 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
|
| D | 3x3p1-minmax-sse-6x4.c | 235 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local 411 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
|
| D | 5x5p2-minmax-sse-4x4.c | 223 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 437 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 639 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|
| D | 5x5p2-minmax-sse-4x4-acc2.c | 223 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 441 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 647 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
|
| D | 5x5p2-minmax-sse-5x4.c | 248 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 501 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 740 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
|
| /external/XNNPACK/src/amalgam/ |
| D | sse.c | 2560 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 2646 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 3103 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 3317 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 3519 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|