90     float32x4_t vi6x4567 = vld1q_f32(i6); i6 += 4;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()  local
124       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
132       const float32x4_t vi6x3456 = vextq_f32(vi6x0123, vi6x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
166       const float32x4_t vi6x2345 = vextq_f32(vi6x0123, vi6x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
167       vi6x0123 = vi6x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
195       const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
229       const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
230       vi6x4567 = vi6x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
305       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
313       const float32x4_t vi6x3456 = vextq_f32(vi6x0123, vi6x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
347       const float32x4_t vi6x2345 = vextq_f32(vi6x0123, vi6x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
348       vi6x0123 = vi6x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
376       const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
410       const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
411       vi6x4567 = vi6x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
461       vi6x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi6x4567)));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
481       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
489       const float32x4_t vi6x3456 = vextq_f32(vi6x0123, vi6x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
517       const float32x4_t vi6x2345 = vextq_f32(vi6x0123, vi6x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
546       const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()