88       float32x4_t vo0p2 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()  local
92       vo0p2 = vmlaq_lane_f32(vo0p2, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
104       vo0p2 = vmlaq_lane_f32(vo0p2, vi1x3456, vget_high_f32(vw4567), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
110       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x3456, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
127       vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
139       vo0p2 = vmlaq_lane_f32(vo0p2, vi0x5678, vget_low_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
145       vo0p2 = vmlaq_lane_f32(vo0p2, vi3x5678, vget_high_f32(vwGHIJ), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
162       vo0p2 = vmlaq_lane_f32(vo0p2, vi1x6789, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
168       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x6789, vwOP, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
171       vo0p0 = vaddq_f32(vo0p0, vo0p2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
197       float32x4_t vo0p2 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()  local
201       vo0p2 = vmlaq_lane_f32(vo0p2, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
213       vo0p2 = vmlaq_lane_f32(vo0p2, vi1x3456, vget_high_f32(vw4567), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
219       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x3456, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
236       vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
248       vo0p2 = vmlaq_lane_f32(vo0p2, vi0x5678, vget_low_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
254       vo0p2 = vmlaq_lane_f32(vo0p2, vi3x5678, vget_high_f32(vwGHIJ), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
271       vo0p2 = vmlaq_lane_f32(vo0p2, vi1x6789, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
277       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x6789, vwOP, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
280       vo0p0 = vaddq_f32(vo0p0, vo0p2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
303       float32x4_t vo0p2 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()  local
307       vo0p2 = vmlaq_lane_f32(vo0p2, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
319       vo0p2 = vmlaq_lane_f32(vo0p2, vi1x3456, vget_high_f32(vw4567), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
325       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x3456, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
337       vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
350       vo0p2 = vmlaq_lane_f32(vo0p2, vi0x5678, vget_low_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
356       vo0p2 = vmlaq_lane_f32(vo0p2, vi3x5678, vget_high_f32(vwGHIJ), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
368       vo0p2 = vmlaq_lane_f32(vo0p2, vi1x6789, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
374       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x6789, vwOP, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
377       vo0p0 = vaddq_f32(vo0p0, vo0p2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()