/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3p1-minmax-sse-2x4.c | 145 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local 229 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local
|
D | 3x3p1-minmax-sse-2x4-acc2.c | 145 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 231 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-1x4.c | 154 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 251 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 339 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
|
D | 5x5p2-minmax-sse-1x4-acc2.c | 154 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 252 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 341 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
|
D | 3x3p1-minmax-sse-3x4.c | 169 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local 276 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local
|
D | 5x5p2-minmax-sse-1x4-acc4.c | 154 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 254 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 345 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local
|
D | 5x5p2-minmax-sse-1x4-acc3.c | 154 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 253 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 343 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local
|
D | 5x5p2-minmax-sse-1x4-acc5.c | 154 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 255 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 347 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local
|
D | 3x3p1-minmax-sse-4x4.c | 193 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local 323 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
|
D | 5x5p2-minmax-sse-2x4.c | 179 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 315 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 440 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
|
D | 5x5p2-minmax-sse-2x4-acc2.c | 179 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 317 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 444 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-2x4-acc3.c | 179 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 319 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 448 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
|
D | 3x3p1-minmax-sse-5x4.c | 217 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local 370 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
|
D | 5x5p2-minmax-sse-3x4.c | 204 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 379 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 541 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
|
D | 5x5p2-minmax-sse-3x4-acc2.c | 204 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 382 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 547 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
|
D | 3x3p1-minmax-sse-6x4.c | 241 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local 417 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
|
D | 5x5p2-minmax-sse-4x4.c | 229 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 443 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 642 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|
D | 5x5p2-minmax-sse-4x4-acc2.c | 229 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 447 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 650 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
|
D | 5x5p2-minmax-sse-5x4.c | 254 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 507 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 743 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
|