Lines Matching refs:vget_low_f32
110 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
111 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
112 float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
113 float32x4_t vo3p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
114 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
132 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
133 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
134 vo2p0 = vmlaq_lane_f32(vo2p0, vi3x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
135 vo3p0 = vmlaq_lane_f32(vo3p0, vi4x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
136 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
138 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
139 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
140 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
141 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
142 vo4p0 = vmlaq_lane_f32(vo4p0, vi6x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
178 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
179 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
180 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
181 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
182 vo4p0 = vmlaq_lane_f32(vo4p0, vi6x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
184 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
185 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
186 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
187 vo3p0 = vmlaq_lane_f32(vo3p0, vi6x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
188 vo4p0 = vmlaq_lane_f32(vo4p0, vi7x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
215 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
216 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
217 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
218 vo3p0 = vmlaq_lane_f32(vo3p0, vi3x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
219 vo4p0 = vmlaq_lane_f32(vo4p0, vi4x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
233 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
234 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
235 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
236 vo3p0 = vmlaq_lane_f32(vo3p0, vi6x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
237 vo4p0 = vmlaq_lane_f32(vo4p0, vi7x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
239 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
240 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
241 vo2p0 = vmlaq_lane_f32(vo2p0, vi6x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
242 vo3p0 = vmlaq_lane_f32(vo3p0, vi7x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
243 vo4p0 = vmlaq_lane_f32(vo4p0, vi8x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
255 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
256 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
257 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
258 vo3p0 = vmlaq_lane_f32(vo3p0, vi3x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
259 vo4p0 = vmlaq_lane_f32(vo4p0, vi4x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
261 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
262 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
263 vo2p0 = vmlaq_lane_f32(vo2p0, vi3x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
264 vo3p0 = vmlaq_lane_f32(vo3p0, vi4x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
265 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
304 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
305 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
306 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
307 vo3p0 = vmlaq_lane_f32(vo3p0, vi3x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
308 vo4p0 = vmlaq_lane_f32(vo4p0, vi4x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
322 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
323 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
324 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
325 vo3p0 = vmlaq_lane_f32(vo3p0, vi6x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
326 vo4p0 = vmlaq_lane_f32(vo4p0, vi7x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
355 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
356 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
357 float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
358 float32x4_t vo3p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
359 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
387 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
388 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
389 vo2p0 = vmlaq_lane_f32(vo2p0, vi3x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
390 vo3p0 = vmlaq_lane_f32(vo3p0, vi4x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
391 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
393 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
394 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
395 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
396 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
397 vo4p0 = vmlaq_lane_f32(vo4p0, vi6x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
433 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
434 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
435 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
436 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
437 vo4p0 = vmlaq_lane_f32(vo4p0, vi6x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
439 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
440 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
441 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
442 vo3p0 = vmlaq_lane_f32(vo3p0, vi6x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
443 vo4p0 = vmlaq_lane_f32(vo4p0, vi7x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
470 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
471 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
472 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
473 vo3p0 = vmlaq_lane_f32(vo3p0, vi3x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
474 vo4p0 = vmlaq_lane_f32(vo4p0, vi4x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
488 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
489 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
490 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
491 vo3p0 = vmlaq_lane_f32(vo3p0, vi6x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
492 vo4p0 = vmlaq_lane_f32(vo4p0, vi7x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
494 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
495 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
496 vo2p0 = vmlaq_lane_f32(vo2p0, vi6x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
497 vo3p0 = vmlaq_lane_f32(vo3p0, vi7x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
498 vo4p0 = vmlaq_lane_f32(vo4p0, vi8x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
510 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
511 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
512 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
513 vo3p0 = vmlaq_lane_f32(vo3p0, vi3x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
514 vo4p0 = vmlaq_lane_f32(vo4p0, vi4x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
516 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
517 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
518 vo2p0 = vmlaq_lane_f32(vo2p0, vi3x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
519 vo3p0 = vmlaq_lane_f32(vo3p0, vi4x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
520 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
559 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
560 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
561 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
562 vo3p0 = vmlaq_lane_f32(vo3p0, vi3x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
563 vo4p0 = vmlaq_lane_f32(vo4p0, vi4x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
577 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
578 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
579 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
580 vo3p0 = vmlaq_lane_f32(vo3p0, vi6x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
581 vo4p0 = vmlaq_lane_f32(vo4p0, vi7x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
613 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
614 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
615 float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
616 float32x4_t vo3p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
617 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
635 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
636 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
637 vo2p0 = vmlaq_lane_f32(vo2p0, vi3x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
638 vo3p0 = vmlaq_lane_f32(vo3p0, vi4x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
639 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
641 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
642 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
643 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
644 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
645 vo4p0 = vmlaq_lane_f32(vo4p0, vi6x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
681 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
682 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
683 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
684 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
685 vo4p0 = vmlaq_lane_f32(vo4p0, vi6x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
687 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
688 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
689 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
690 vo3p0 = vmlaq_lane_f32(vo3p0, vi6x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
691 vo4p0 = vmlaq_lane_f32(vo4p0, vi7x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
709 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
710 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
711 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
712 vo3p0 = vmlaq_lane_f32(vo3p0, vi3x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
713 vo4p0 = vmlaq_lane_f32(vo4p0, vi4x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
727 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
728 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
729 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
730 vo3p0 = vmlaq_lane_f32(vo3p0, vi6x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
731 vo4p0 = vmlaq_lane_f32(vo4p0, vi7x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
733 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
734 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
735 vo2p0 = vmlaq_lane_f32(vo2p0, vi6x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
736 vo3p0 = vmlaq_lane_f32(vo3p0, vi7x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
737 vo4p0 = vmlaq_lane_f32(vo4p0, vi8x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
750 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
751 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
752 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
753 vo3p0 = vmlaq_lane_f32(vo3p0, vi3x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
754 vo4p0 = vmlaq_lane_f32(vo4p0, vi4x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
756 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
757 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
758 vo2p0 = vmlaq_lane_f32(vo2p0, vi3x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
759 vo3p0 = vmlaq_lane_f32(vo3p0, vi4x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
760 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
790 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
791 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
792 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
793 vo3p0 = vmlaq_lane_f32(vo3p0, vi3x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
794 vo4p0 = vmlaq_lane_f32(vo4p0, vi4x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
808 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
809 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
810 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
811 vo3p0 = vmlaq_lane_f32(vo3p0, vi6x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
812 vo4p0 = vmlaq_lane_f32(vo4p0, vi7x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
840 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
841 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
842 float32x2_t vo2_lo = vget_low_f32(vo2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
843 float32x2_t vo3_lo = vget_low_f32(vo3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
844 float32x2_t vo4_lo = vget_low_f32(vo4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()