97     float32x4_t vi6x4567 = vld1q_f32(i6); i6 += 4;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()  local
134       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
138       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
147       const float32x4_t vi6x3456 = vextq_f32(vi6x0123, vi6x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
187       const float32x4_t vi6x2345 = vextq_f32(vi6x0123, vi6x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
188       vi6x0123 = vi6x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
223       const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
263       const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
264       vi6x4567 = vi6x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
352       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
356       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
365       const float32x4_t vi6x3456 = vextq_f32(vi6x0123, vi6x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
405       const float32x4_t vi6x2345 = vextq_f32(vi6x0123, vi6x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
406       vi6x0123 = vi6x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
441       const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
481       const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
482       vi6x4567 = vi6x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
543       vi6x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi6x4567)));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
564       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
568       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
577       const float32x4_t vi6x3456 = vextq_f32(vi6x0123, vi6x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
611       const float32x4_t vi6x2345 = vextq_f32(vi6x0123, vi6x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
646       const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()