/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3p1-minmax-sse-1x4.c | 93 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4() local 157 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4() local
|
D | 3x3p1-minmax-sse-1x4-acc3.c | 93 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3() local 159 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3() local
|
D | 3x3p1-minmax-sse-1x4-acc2.c | 93 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2() local 158 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2() local
|
D | 3x3p1-minmax-sse-1x4-acc4.c | 93 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local 160 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local
|
D | 3x3p1-minmax-sse-2x4.c | 106 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local 194 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local
|
D | 3x3p1-minmax-sse-2x4-acc2.c | 106 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 196 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-1x4.c | 110 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 207 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 313 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
|
D | 5x5p2-minmax-sse-1x4-acc2.c | 110 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 208 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 315 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
|
D | 3x3p1-minmax-sse-3x4.c | 119 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local 231 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local
|
D | 5x5p2-minmax-sse-1x4-acc4.c | 110 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 210 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 319 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local
|
D | 5x5p2-minmax-sse-1x4-acc3.c | 110 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 209 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 317 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local
|
D | 5x5p2-minmax-sse-1x4-acc5.c | 110 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 211 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 321 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local
|
D | 3x3p1-minmax-sse-4x4.c | 132 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local 268 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
|
D | 5x5p2-minmax-sse-2x4.c | 124 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 260 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 406 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
|
D | 5x5p2-minmax-sse-2x4-acc2.c | 124 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 262 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 410 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-2x4-acc3.c | 124 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 264 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 414 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
|
D | 3x3p1-minmax-sse-5x4.c | 145 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local 305 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
|
D | 5x5p2-minmax-sse-3x4.c | 138 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 313 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 499 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
|
D | 5x5p2-minmax-sse-3x4-acc2.c | 138 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 316 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 505 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
|
D | 3x3p1-minmax-sse-6x4.c | 158 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local 342 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
|
D | 5x5p2-minmax-sse-4x4.c | 152 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 366 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 592 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|
D | 5x5p2-minmax-sse-4x4-acc2.c | 152 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 370 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 600 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
|
D | 5x5p2-minmax-sse-5x4.c | 166 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 419 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 685 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
|