Home
last modified time | relevance | path

Searched refs:vi5x7456 (Results 1 – 11 of 11) sorted by relevance

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5p2-minmax-sse-2x4.c127 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
147 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
170 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
171 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
263 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
283 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
306 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
307 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
409 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
416 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
[all …]
D5x5p2-minmax-sse-2x4-acc3.c127 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
147 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3()
170 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3()
171 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3()
267 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
287 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3()
310 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3()
311 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3()
417 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
424 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3()
[all …]
D5x5p2-minmax-sse-2x4-acc2.c127 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
147 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
170 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
171 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
265 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
285 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
308 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
309 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
413 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
420 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
[all …]
D5x5p2-minmax-sse-3x4-acc2.c141 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
164 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
193 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
194 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
319 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
342 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
371 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
372 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
508 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
516 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
[all …]
D5x5p2-minmax-sse-3x4.c141 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
164 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
193 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
194 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
316 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
339 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
368 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
369 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
502 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
510 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
[all …]
D3x3p1-minmax-sse-4x4.c138 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
164 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4()
184 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4()
274 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
300 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4()
D3x3p1-minmax-sse-5x4.c151 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
182 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4()
207 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4()
311 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
342 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4()
D5x5p2-minmax-sse-4x4.c155 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
181 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
216 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
217 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
369 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
395 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
430 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
431 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
595 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
604 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
[all …]
D5x5p2-minmax-sse-4x4-acc2.c155 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
181 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
216 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
217 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
373 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
399 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
434 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
435 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
603 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
612 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
[all …]
D3x3p1-minmax-sse-6x4.c164 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
200 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
230 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
348 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
384 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
D5x5p2-minmax-sse-5x4.c169 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
198 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
239 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
240 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
422 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
451 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
492 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
493 vi5x3012 = vi5x7456; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
688 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
698 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
[all …]