89     float32x4_t vi5x4567 = vld1q_f32(i5); i5 += 4;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()  local
120       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
123       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
131       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
164       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
165       vi5x0123 = vi5x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
194       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
227       const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
228       vi5x4567 = vi5x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
301       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
304       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
312       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
345       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
346       vi5x0123 = vi5x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
375       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
408       const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
409       vi5x4567 = vi5x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
460       vi5x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x4567)));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
477       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
480       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
488       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
516       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
545       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()