86       float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()  local
92       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
102       vo0p1 = vfmaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
106       vo0p1 = vfmaq_lane_f32(vo0p1, vi2x3456, vget_low_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
110       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x3456, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
125       vo0p1 = vfmaq_lane_f32(vo0p1, vi1x2345, vget_high_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
129       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x2345, vget_low_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
139       vo0p1 = vfmaq_lane_f32(vo0p1, vi0x5678, vget_low_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
143       vo0p1 = vfmaq_lane_f32(vo0p1, vi2x5678, vget_high_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
147       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x5678, vwOP, 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
162       vo0p1 = vfmaq_lane_f32(vo0p1, vi1x6789, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
166       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x6789, vget_low_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
170       vo0p0 = vaddq_f32(vo0p0, vo0p1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
194       float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()  local
200       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
210       vo0p1 = vfmaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
214       vo0p1 = vfmaq_lane_f32(vo0p1, vi2x3456, vget_low_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
218       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x3456, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
233       vo0p1 = vfmaq_lane_f32(vo0p1, vi1x2345, vget_high_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
237       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x2345, vget_low_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
247       vo0p1 = vfmaq_lane_f32(vo0p1, vi0x5678, vget_low_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
251       vo0p1 = vfmaq_lane_f32(vo0p1, vi2x5678, vget_high_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
255       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x5678, vwOP, 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
270       vo0p1 = vfmaq_lane_f32(vo0p1, vi1x6789, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
274       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x6789, vget_low_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
278       vo0p0 = vaddq_f32(vo0p0, vo0p1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
299       float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()  local
305       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
315       vo0p1 = vfmaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
319       vo0p1 = vfmaq_lane_f32(vo0p1, vi2x3456, vget_low_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
323       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x3456, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
333       vo0p1 = vfmaq_lane_f32(vo0p1, vi1x2345, vget_high_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
337       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x2345, vget_low_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
348       vo0p1 = vfmaq_lane_f32(vo0p1, vi0x5678, vget_low_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
352       vo0p1 = vfmaq_lane_f32(vo0p1, vi2x5678, vget_high_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
356       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x5678, vwOP, 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
366       vo0p1 = vfmaq_lane_f32(vo0p1, vi1x6789, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
370       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x6789, vget_low_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
374       vo0p0 = vaddq_f32(vo0p0, vo0p1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()