/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-sse-1x4-acc2.c | 156 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 254 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 342 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
|
D | 5x5p2-minmax-sse-1x4.c | 156 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 253 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 340 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
|
D | 3x3p1-minmax-sse-3x4.c | 171 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local 278 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local
|
D | 5x5p2-minmax-sse-1x4-acc5.c | 156 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 257 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 348 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local
|
D | 5x5p2-minmax-sse-1x4-acc4.c | 156 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 256 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 346 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local
|
D | 5x5p2-minmax-sse-1x4-acc3.c | 156 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 255 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 344 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local
|
D | 3x3p1-minmax-sse-4x4.c | 195 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local 325 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
|
D | 5x5p2-minmax-sse-2x4.c | 181 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 317 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 441 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
|
D | 5x5p2-minmax-sse-2x4-acc3.c | 181 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 321 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 449 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
|
D | 5x5p2-minmax-sse-2x4-acc2.c | 181 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 319 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 445 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
|
D | 3x3p1-minmax-sse-5x4.c | 219 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local 372 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
|
D | 5x5p2-minmax-sse-3x4.c | 206 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 381 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 542 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
|
D | 5x5p2-minmax-sse-3x4-acc2.c | 206 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 384 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 548 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
|
D | 3x3p1-minmax-sse-6x4.c | 243 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local 419 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
|
D | 5x5p2-minmax-sse-4x4.c | 231 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 445 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 643 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|
D | 5x5p2-minmax-sse-4x4-acc2.c | 231 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 449 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 651 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
|
D | 5x5p2-minmax-sse-5x4.c | 256 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 509 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 744 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
|