90     float32x4_t vi6x4567 = vld1q_f32(i6); i6 += 4;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()  local
124       vo2p0 = vfmaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
132       const float32x4_t vi6x3456 = vextq_f32(vi6x0123, vi6x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
166       const float32x4_t vi6x2345 = vextq_f32(vi6x0123, vi6x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
167       vi6x0123 = vi6x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
195       const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
229       const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
230       vi6x4567 = vi6x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
308       vo2p0 = vfmaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
316       const float32x4_t vi6x3456 = vextq_f32(vi6x0123, vi6x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
350       const float32x4_t vi6x2345 = vextq_f32(vi6x0123, vi6x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
351       vi6x0123 = vi6x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
379       const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
413       const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
414       vi6x4567 = vi6x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
467       vi6x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi6x4567)));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
487       vo2p0 = vfmaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
495       const float32x4_t vi6x3456 = vextq_f32(vi6x0123, vi6x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
523       const float32x4_t vi6x2345 = vextq_f32(vi6x0123, vi6x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
552       const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()