/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3p1-minmax-sse-1x4.c | 119 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4() local 180 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4() local
|
D | 3x3p1-minmax-sse-1x4-acc3.c | 119 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3() local 182 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3() local
|
D | 3x3p1-minmax-sse-1x4-acc2.c | 119 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2() local 181 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2() local
|
D | 3x3p1-minmax-sse-1x4-acc4.c | 119 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local 183 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local
|
D | 3x3p1-minmax-sse-2x4.c | 143 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local 227 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local
|
D | 3x3p1-minmax-sse-2x4-acc2.c | 143 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 229 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-1x4.c | 152 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 249 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 338 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
|
D | 5x5p2-minmax-sse-1x4-acc2.c | 152 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 250 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 340 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
|
D | 3x3p1-minmax-sse-3x4.c | 167 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local 274 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local
|
D | 5x5p2-minmax-sse-1x4-acc4.c | 152 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 252 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 344 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local
|
D | 5x5p2-minmax-sse-1x4-acc3.c | 152 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 251 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 342 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local
|
D | 5x5p2-minmax-sse-1x4-acc5.c | 152 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 253 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 346 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local
|
D | 3x3p1-minmax-sse-4x4.c | 191 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local 321 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
|
D | 5x5p2-minmax-sse-2x4.c | 177 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 313 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 439 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
|
D | 5x5p2-minmax-sse-2x4-acc2.c | 177 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 315 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 443 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-2x4-acc3.c | 177 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 317 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 447 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
|
D | 3x3p1-minmax-sse-5x4.c | 215 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local 368 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
|
D | 5x5p2-minmax-sse-3x4.c | 202 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 377 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 540 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
|
D | 5x5p2-minmax-sse-3x4-acc2.c | 202 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 380 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 546 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
|
D | 3x3p1-minmax-sse-6x4.c | 239 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local 415 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
|
D | 5x5p2-minmax-sse-4x4.c | 227 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 441 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 641 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|
D | 5x5p2-minmax-sse-4x4-acc2.c | 227 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 445 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 649 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
|
D | 5x5p2-minmax-sse-5x4.c | 252 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 505 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 742 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
|