140       const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()  local
150       vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
154       vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
158       vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
162       vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
166       vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
170       vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
174       vacc6xCDEF = vdotq_lane_s32(vacc6xCDEF, vb0123xCDEF, va6x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
178       vacc7xCDEF = vdotq_lane_s32(vacc7xCDEF, vb0123xCDEF, va7x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
230       const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()  local
236       vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
240       vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
244       vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
248       vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
252       vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
256       vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
260       vacc6xCDEF = vdotq_lane_s32(vacc6xCDEF, vb0123xCDEF, va6x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
264       vacc7xCDEF = vdotq_lane_s32(vacc7xCDEF, vb0123xCDEF, va7x01234567, 0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()