/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3p1-minmax-sse-2x4.c | 108 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local 196 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local
|
D | 3x3p1-minmax-sse-2x4-acc2.c | 108 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 198 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-1x4.c | 111 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 208 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 314 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
|
D | 5x5p2-minmax-sse-1x4-acc2.c | 111 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 209 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 316 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
|
D | 3x3p1-minmax-sse-3x4.c | 121 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local 233 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local
|
D | 5x5p2-minmax-sse-1x4-acc4.c | 111 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 211 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 320 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local
|
D | 5x5p2-minmax-sse-1x4-acc3.c | 111 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 210 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 318 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local
|
D | 5x5p2-minmax-sse-1x4-acc5.c | 111 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 212 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 322 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local
|
D | 3x3p1-minmax-sse-4x4.c | 134 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local 270 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
|
D | 5x5p2-minmax-sse-2x4.c | 125 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 261 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 407 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
|
D | 5x5p2-minmax-sse-2x4-acc2.c | 125 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 263 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 411 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-2x4-acc3.c | 125 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 265 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 415 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
|
D | 3x3p1-minmax-sse-5x4.c | 147 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local 307 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
|
D | 5x5p2-minmax-sse-3x4.c | 139 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 314 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 500 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
|
D | 5x5p2-minmax-sse-3x4-acc2.c | 139 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 317 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 506 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
|
D | 3x3p1-minmax-sse-6x4.c | 160 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local 344 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
|
D | 5x5p2-minmax-sse-4x4.c | 153 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 367 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 593 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|
D | 5x5p2-minmax-sse-4x4-acc2.c | 153 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 371 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 601 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
|
D | 5x5p2-minmax-sse-5x4.c | 167 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 420 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 686 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
|