Lines Matching refs:__m512
33 const __m512 vmax = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max)); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
34 const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min)); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
166 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
167 __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
170 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
171 const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
174 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
175 const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
179 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
180 const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
183 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
184 const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
185 __m512 vacc0123456789ABCDEFp1 = _mm512_mul_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
186 __m512 vaccGHIJKLMNOPQRSTUVp1 = _mm512_mul_ps(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
188 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
189 const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
192 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
193 const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
197 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
198 const __m512 vi3xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i3 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
201 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
202 const __m512 vk3xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
206 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
207 const __m512 vi4xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i4 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
210 const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
211 const __m512 vk4xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
215 const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
216 const __m512 vi5xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i5 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
219 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
220 const __m512 vk5xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 208); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
224 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
225 const __m512 vi6xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i6 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
228 const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
229 const __m512 vk6xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 240); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
233 const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
234 const __m512 vi7xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i7 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
237 const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 256); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
238 const __m512 vk7xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
242 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
243 const __m512 vi8xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i8 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
246 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
247 const __m512 vk8xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 304); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
251 const __m512 vi9x0123456789ABCDEF = _mm512_loadu_ps(i9); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
252 const __m512 vi9xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i9 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
255 const __m512 vk9x0123456789ABCDEF = _mm512_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
256 const __m512 vk9xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 336); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
260 const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
261 const __m512 vi10xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i10 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
264 const __m512 vk10x0123456789ABCDEF = _mm512_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
265 const __m512 vk10xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
269 const __m512 vi11x0123456789ABCDEF = _mm512_loadu_ps(i11); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
270 const __m512 vi11xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i11 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
273 const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 384); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
274 const __m512 vk11xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 400); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
278 const __m512 vi12x0123456789ABCDEF = _mm512_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
279 const __m512 vi12xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i12 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
282 const __m512 vk12x0123456789ABCDEF = _mm512_load_ps(w + 416); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
283 const __m512 vk12xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 432); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
287 const __m512 vi13x0123456789ABCDEF = _mm512_loadu_ps(i13); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
288 const __m512 vi13xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i13 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
291 const __m512 vk13x0123456789ABCDEF = _mm512_load_ps(w + 448); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
292 const __m512 vk13xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 464); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
296 const __m512 vi14x0123456789ABCDEF = _mm512_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
297 const __m512 vi14xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i14 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
300 const __m512 vk14x0123456789ABCDEF = _mm512_load_ps(w + 480); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
301 const __m512 vk14xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 496); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
305 const __m512 vi15x0123456789ABCDEF = _mm512_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
306 const __m512 vi15xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i15 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
309 const __m512 vk15x0123456789ABCDEF = _mm512_load_ps(w + 512); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
310 const __m512 vk15xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 528); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
314 const __m512 vi16x0123456789ABCDEF = _mm512_loadu_ps(i16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
315 const __m512 vi16xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i16 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
318 const __m512 vk16x0123456789ABCDEF = _mm512_load_ps(w + 544); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
319 const __m512 vk16xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 560); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
323 const __m512 vi17x0123456789ABCDEF = _mm512_loadu_ps(i17); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
324 const __m512 vi17xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i17 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
327 const __m512 vk17x0123456789ABCDEF = _mm512_load_ps(w + 576); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
328 const __m512 vk17xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 592); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
332 const __m512 vi18x0123456789ABCDEF = _mm512_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
333 const __m512 vi18xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i18 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
336 const __m512 vk18x0123456789ABCDEF = _mm512_load_ps(w + 608); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
337 const __m512 vk18xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 624); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
341 const __m512 vi19x0123456789ABCDEF = _mm512_loadu_ps(i19); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
342 const __m512 vi19xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i19 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
345 const __m512 vk19x0123456789ABCDEF = _mm512_load_ps(w + 640); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
346 const __m512 vk19xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 656); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
350 const __m512 vi20x0123456789ABCDEF = _mm512_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
351 const __m512 vi20xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i20 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
354 const __m512 vk20x0123456789ABCDEF = _mm512_load_ps(w + 672); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
355 const __m512 vk20xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 688); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
359 const __m512 vi21x0123456789ABCDEF = _mm512_loadu_ps(i21); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
360 const __m512 vi21xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i21 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
363 const __m512 vk21x0123456789ABCDEF = _mm512_load_ps(w + 704); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
364 const __m512 vk21xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 720); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
368 const __m512 vi22x0123456789ABCDEF = _mm512_loadu_ps(i22); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
369 const __m512 vi22xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i22 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
372 const __m512 vk22x0123456789ABCDEF = _mm512_load_ps(w + 736); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
373 const __m512 vk22xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 752); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
377 const __m512 vi23x0123456789ABCDEF = _mm512_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
378 const __m512 vi23xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i23 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
381 const __m512 vk23x0123456789ABCDEF = _mm512_load_ps(w + 768); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
382 const __m512 vk23xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 784); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
386 const __m512 vi24x0123456789ABCDEF = _mm512_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
387 const __m512 vi24xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i24 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
390 const __m512 vk24x0123456789ABCDEF = _mm512_load_ps(w + 800); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
391 const __m512 vk24xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 816); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
401 __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
402 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUVp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
411 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
413 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
416 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
419 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
422 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
423 __m512 vacc0123456789ABCDEFp1 = _mm512_mul_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
425 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
428 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
431 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
434 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
437 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
440 const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
443 const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
446 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
449 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
452 const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
455 const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
458 const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 256); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
461 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
464 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
467 const __m512 vi9x0123456789ABCDEF = _mm512_loadu_ps(i9); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
470 const __m512 vk9x0123456789ABCDEF = _mm512_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
473 const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
476 const __m512 vk10x0123456789ABCDEF = _mm512_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
479 const __m512 vi11x0123456789ABCDEF = _mm512_loadu_ps(i11); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
482 const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 384); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
485 const __m512 vi12x0123456789ABCDEF = _mm512_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
488 const __m512 vk12x0123456789ABCDEF = _mm512_load_ps(w + 416); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
491 const __m512 vi13x0123456789ABCDEF = _mm512_loadu_ps(i13); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
494 const __m512 vk13x0123456789ABCDEF = _mm512_load_ps(w + 448); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
497 const __m512 vi14x0123456789ABCDEF = _mm512_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
500 const __m512 vk14x0123456789ABCDEF = _mm512_load_ps(w + 480); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
503 const __m512 vi15x0123456789ABCDEF = _mm512_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
506 const __m512 vk15x0123456789ABCDEF = _mm512_load_ps(w + 512); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
509 const __m512 vi16x0123456789ABCDEF = _mm512_loadu_ps(i16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
512 const __m512 vk16x0123456789ABCDEF = _mm512_load_ps(w + 544); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
515 const __m512 vi17x0123456789ABCDEF = _mm512_loadu_ps(i17); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
518 const __m512 vk17x0123456789ABCDEF = _mm512_load_ps(w + 576); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
521 const __m512 vi18x0123456789ABCDEF = _mm512_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
524 const __m512 vk18x0123456789ABCDEF = _mm512_load_ps(w + 608); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
527 const __m512 vi19x0123456789ABCDEF = _mm512_loadu_ps(i19); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
530 const __m512 vk19x0123456789ABCDEF = _mm512_load_ps(w + 640); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
533 const __m512 vi20x0123456789ABCDEF = _mm512_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
536 const __m512 vk20x0123456789ABCDEF = _mm512_load_ps(w + 672); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
539 const __m512 vi21x0123456789ABCDEF = _mm512_loadu_ps(i21); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
542 const __m512 vk21x0123456789ABCDEF = _mm512_load_ps(w + 704); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
545 const __m512 vi22x0123456789ABCDEF = _mm512_loadu_ps(i22); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
548 const __m512 vk22x0123456789ABCDEF = _mm512_load_ps(w + 736); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
551 const __m512 vi23x0123456789ABCDEF = _mm512_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
554 const __m512 vk23x0123456789ABCDEF = _mm512_load_ps(w + 768); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
557 const __m512 vi24x0123456789ABCDEF = _mm512_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
560 const __m512 vk24x0123456789ABCDEF = _mm512_load_ps(w + 800); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
568 __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
580 __m512 vacc0123456789ABCDEFp0 = _mm512_maskz_loadu_ps(vmask, w); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
582 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
583 const __m512 vk0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 32); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
586 const __m512 vi1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i1); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
587 const __m512 vk1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 64); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
588 __m512 vacc0123456789ABCDEFp1 = _mm512_mul_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
590 const __m512 vi2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i2); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
591 const __m512 vk2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 96); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
594 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
595 const __m512 vk3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 128); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
598 const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
599 const __m512 vk4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 160); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
602 const __m512 vi5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i5); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
603 const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
606 const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
607 const __m512 vk6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 224); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
610 const __m512 vi7x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i7); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
611 const __m512 vk7x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 256); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
614 const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
615 const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 288); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
618 const __m512 vi9x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i9); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
619 const __m512 vk9x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 320); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
622 const __m512 vi10x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i10); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
623 const __m512 vk10x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 352); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
626 const __m512 vi11x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i11); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
627 const __m512 vk11x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 384); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
630 const __m512 vi12x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i12); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
631 const __m512 vk12x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 416); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
634 const __m512 vi13x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i13); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
635 const __m512 vk13x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 448); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
638 const __m512 vi14x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i14); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
639 const __m512 vk14x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 480); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
642 const __m512 vi15x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i15); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
643 const __m512 vk15x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 512); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
646 const __m512 vi16x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
647 const __m512 vk16x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 544); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
650 const __m512 vi17x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i17); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
651 const __m512 vk17x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 576); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
654 const __m512 vi18x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i18); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
655 const __m512 vk18x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 608); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
658 const __m512 vi19x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i19); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
659 const __m512 vk19x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 640); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
662 const __m512 vi20x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i20); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
663 const __m512 vk20x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 672); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
666 const __m512 vi21x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i21); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
667 const __m512 vk21x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 704); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
670 const __m512 vi22x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i22); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
671 const __m512 vk22x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 736); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
674 const __m512 vi23x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i23); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
675 const __m512 vk23x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 768); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
678 const __m512 vi24x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i24); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
679 const __m512 vk24x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 800); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
685 __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()