96       float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()  local
105       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
118       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
124       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x3456, vget_low_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
130       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x3456, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
149       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x2345, vget_high_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
155       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x2345, vget_low_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
168       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x5678, vget_low_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
174       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x5678, vget_high_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
180       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x5678, vwOP, 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
199       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x6789, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
205       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x6789, vget_low_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
211       vo0p0 = vaddq_f32(vo0p0, vo0p1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
242       float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()  local
251       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
264       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
270       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x3456, vget_low_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
276       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x3456, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
295       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x2345, vget_high_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
301       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x2345, vget_low_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
314       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x5678, vget_low_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
320       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x5678, vget_high_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
326       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x5678, vwOP, 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
345       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x6789, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
351       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x6789, vget_low_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
357       vo0p0 = vaddq_f32(vo0p0, vo0p1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
384       float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()  local
393       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
406       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
412       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x3456, vget_low_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
418       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x3456, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
431       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x2345, vget_high_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
437       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x2345, vget_low_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
451       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x5678, vget_low_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
457       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x5678, vget_high_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
463       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x5678, vwOP, 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
476       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x6789, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
482       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x6789, vget_low_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
488       vo0p0 = vaddq_f32(vo0p0, vo0p1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()