87       const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()  local
106       vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
107       vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
108       vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
109       vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
122       vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
123       vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
124       vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
125       vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
141       const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()  local
156       vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
157       vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
158       vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
159       vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()