106     float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()  local
154       vo4p0 = vfmaq_lane_f32(vo4p0, vi8x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
164       const float32x4_t vi8x3456 = vextq_f32(vi8x0123, vi8x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
212       const float32x4_t vi8x2345 = vextq_f32(vi8x0123, vi8x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
213       vi8x0123 = vi8x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
253       const float32x4_t vi8x5678 = vextq_f32(vi8x4567, vi8x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
301       const float32x4_t vi8x6789 = vextq_f32(vi8x4567, vi8x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
302       vi8x4567 = vi8x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
409       vo4p0 = vfmaq_lane_f32(vo4p0, vi8x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
419       const float32x4_t vi8x3456 = vextq_f32(vi8x0123, vi8x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
467       const float32x4_t vi8x2345 = vextq_f32(vi8x0123, vi8x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
468       vi8x0123 = vi8x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
508       const float32x4_t vi8x5678 = vextq_f32(vi8x4567, vi8x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
556       const float32x4_t vi8x6789 = vextq_f32(vi8x4567, vi8x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
557       vi8x4567 = vi8x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
627       vi8x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi8x4567)));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
657       vo4p0 = vfmaq_lane_f32(vo4p0, vi8x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
667       const float32x4_t vi8x3456 = vextq_f32(vi8x0123, vi8x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
707       const float32x4_t vi8x2345 = vextq_f32(vi8x0123, vi8x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
748       const float32x4_t vi8x5678 = vextq_f32(vi8x4567, vzero, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()