128       const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()  local
154       vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
155       vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
156       vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
157       vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
186       vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
187       vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
188       vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
189       vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
218       const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()  local
240       vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
241       vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
242       vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
243       vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()