/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3s2p1-minmax-sse-1x4.c | 95 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4() local 144 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4() local
|
D | 3x3s2p1-minmax-sse-1x4-acc3.c | 95 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3() local 146 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3() local
|
D | 3x3s2p1-minmax-sse-1x4-acc2.c | 95 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2() local 145 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2() local
|
D | 3x3s2p1-minmax-sse-1x4-acc4.c | 95 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4() local 147 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4() local
|
D | 3x3s2p1-minmax-sse-2x4.c | 121 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4() local 197 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4() local
|
D | 3x3s2p1-minmax-sse-2x4-acc2.c | 121 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2() local 199 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2() local
|
D | 3x3s2p1-minmax-sse-3x4.c | 146 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4() local 249 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4() local
|
D | 5x5s2p2-minmax-sse-1x4.c | 161 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() local 292 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() local
|
D | 5x5s2p2-minmax-sse-1x4-acc2.c | 161 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() local 293 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() local
|
D | 5x5s2p2-minmax-sse-1x4-acc5.c | 161 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() local 296 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() local
|
D | 5x5s2p2-minmax-sse-1x4-acc4.c | 161 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() local 295 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() local
|
D | 3x3s2p1-minmax-sse-4x4.c | 171 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4() local 301 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4() local
|
D | 5x5s2p2-minmax-sse-1x4-acc3.c | 161 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() local 294 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() local
|
D | 3x3s2p1-minmax-sse-5x4.c | 196 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4() local 353 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4() local
|
D | 5x5s2p2-minmax-sse-2x4-acc2.c | 202 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4_acc2() local 396 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4_acc2() local
|
D | 5x5s2p2-minmax-sse-2x4-acc3.c | 202 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4_acc3() local 398 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4_acc3() local
|
D | 5x5s2p2-minmax-sse-2x4.c | 202 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4() local 394 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4() local
|
D | 3x3s2p1-minmax-sse-6x4.c | 221 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local 405 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local
|
D | 5x5s2p2-minmax-sse-3x4.c | 242 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4() local 495 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4() local
|
D | 5x5s2p2-minmax-sse-3x4-acc2.c | 242 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2() local 498 const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2() local
|