257       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()  local
271       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x6789, vget_low_f32(vw4567), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
275       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x6789, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
279       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x6789, vget_high_f32(vwCDEF), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
283       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
475       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()  local
489       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x6789, vget_low_f32(vw4567), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
493       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x6789, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
497       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x6789, vget_high_f32(vwCDEF), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
501       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
677       const float32x4_t vi3x6789 = vextq_f32(vi3x5678, vzero, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()  local
686       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x6789, vget_low_f32(vw4567), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
690       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x6789, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
694       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x6789, vget_high_f32(vwCDEF), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
698       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()