Home
last modified time | relevance | path

Searched refs:vi6x7456 (Results 1 – 7 of 7) sorted by relevance

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5p2-minmax-sse-3x4-acc2.c142 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
165 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
195 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
196 vi6x3012 = vi6x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
320 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
343 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
373 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
374 vi6x3012 = vi6x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
509 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
517 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
[all …]
D5x5p2-minmax-sse-3x4.c142 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
165 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
195 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
196 vi6x3012 = vi6x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
317 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
340 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
370 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
371 vi6x3012 = vi6x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
503 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
511 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
[all …]
D3x3p1-minmax-sse-5x4.c153 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
184 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4()
208 vi6x3012 = vi6x7456; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4()
313 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
344 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4()
D5x5p2-minmax-sse-4x4.c156 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
182 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
218 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
219 vi6x3012 = vi6x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
370 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
396 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
432 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
433 vi6x3012 = vi6x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
596 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
605 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
[all …]
D5x5p2-minmax-sse-4x4-acc2.c156 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
182 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
218 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
219 vi6x3012 = vi6x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
374 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
400 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
436 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
437 vi6x3012 = vi6x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
604 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
613 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
[all …]
D3x3p1-minmax-sse-6x4.c166 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
202 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
231 vi6x3012 = vi6x7456; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
350 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
386 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
D5x5p2-minmax-sse-5x4.c170 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
199 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
241 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
242 vi6x3012 = vi6x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
423 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
452 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
494 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
495 vi6x3012 = vi6x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
689 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
699 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
[all …]