/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3p1-minmax-sse-1x4.c | 117 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4() local 178 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4() local
|
D | 3x3p1-minmax-sse-1x4-acc2.c | 117 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2() local 179 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2() local
|
D | 3x3p1-minmax-sse-1x4-acc3.c | 117 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3() local 180 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3() local
|
D | 3x3p1-minmax-sse-1x4-acc4.c | 117 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local 181 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local
|
D | 3x3p1-minmax-sse-2x4.c | 141 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local 225 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local
|
D | 3x3p1-minmax-sse-2x4-acc2.c | 141 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 227 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-1x4.c | 150 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 247 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 337 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
|
D | 5x5p2-minmax-sse-1x4-acc2.c | 150 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 248 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 339 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
|
D | 3x3p1-minmax-sse-3x4.c | 165 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local 272 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local
|
D | 5x5p2-minmax-sse-1x4-acc3.c | 150 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 249 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 341 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local
|
D | 5x5p2-minmax-sse-1x4-acc4.c | 150 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 250 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 343 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local
|
D | 5x5p2-minmax-sse-1x4-acc5.c | 150 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 251 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 345 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local
|
D | 5x5p2-minmax-sse-2x4.c | 175 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 311 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 438 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
|
D | 3x3p1-minmax-sse-4x4.c | 189 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local 319 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
|
D | 5x5p2-minmax-sse-2x4-acc2.c | 175 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 313 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 442 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-2x4-acc3.c | 175 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 315 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 446 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
|
D | 3x3p1-minmax-sse-5x4.c | 213 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local 366 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
|
D | 5x5p2-minmax-sse-3x4.c | 200 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 375 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 539 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
|
D | 5x5p2-minmax-sse-3x4-acc2.c | 200 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 378 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 545 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
|
D | 3x3p1-minmax-sse-6x4.c | 237 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local 413 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
|
D | 5x5p2-minmax-sse-4x4.c | 225 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 439 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 640 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|
D | 5x5p2-minmax-sse-4x4-acc2.c | 225 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 443 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 648 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
|
D | 5x5p2-minmax-sse-5x4.c | 250 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 503 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 741 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
|
/external/XNNPACK/src/amalgam/ |
D | sse.c | 2562 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 2648 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 3105 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 3319 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 3520 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|