Lines Matching refs:vget_low_f32
94 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
95 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
96 float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
110 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
111 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
112 vo2p0 = vmlaq_lane_f32(vo2p0, vi3x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
114 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
115 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
116 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
142 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
143 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
144 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
146 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
147 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
148 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
169 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
170 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
171 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
181 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
182 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
183 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
185 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
186 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
187 vo2p0 = vmlaq_lane_f32(vo2p0, vi6x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
197 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
198 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
199 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
201 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
202 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
203 vo2p0 = vmlaq_lane_f32(vo2p0, vi3x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
232 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
233 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
234 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
244 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
245 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
246 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
267 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
268 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
269 float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
291 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
292 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
293 vo2p0 = vmlaq_lane_f32(vo2p0, vi3x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
295 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
296 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
297 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
323 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
324 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
325 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
327 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
328 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
329 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
350 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
351 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
352 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
362 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
363 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
364 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
366 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
367 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
368 vo2p0 = vmlaq_lane_f32(vo2p0, vi6x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
378 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
379 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
380 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
382 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
383 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
384 vo2p0 = vmlaq_lane_f32(vo2p0, vi3x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
413 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
414 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
415 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
425 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
426 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
427 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
451 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
452 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
453 float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
467 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
468 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
469 vo2p0 = vmlaq_lane_f32(vo2p0, vi3x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
471 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
472 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
473 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
499 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
500 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
501 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
503 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
504 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
505 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
519 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
520 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
521 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
531 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
532 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
533 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
535 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
536 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
537 vo2p0 = vmlaq_lane_f32(vo2p0, vi6x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
548 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
549 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
550 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
552 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
553 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
554 vo2p0 = vmlaq_lane_f32(vo2p0, vi3x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
576 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
577 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
578 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
588 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
589 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
590 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
610 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
611 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
612 float32x2_t vo2_lo = vget_low_f32(vo2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()