73     float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()  local
92       vo0p2 = vfmaq_lane_f32(vo0p2, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
99       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
118       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
119       vi3x0123 = vi3x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
136       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
155       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
156       vi3x4567 = vi3x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
201       vo0p2 = vfmaq_lane_f32(vo0p2, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
208       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
227       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
228       vi3x0123 = vi3x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
245       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
264       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
265       vi3x4567 = vi3x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
298       vi3x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x4567)));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
307       vo0p2 = vfmaq_lane_f32(vo0p2, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
314       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
330       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
347       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()