56       const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()  local
69       vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
70       vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
71       vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
72       vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
73       vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
74       vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
75       vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
76       vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
83       const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()  local
92       vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
93       vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
94       vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
95       vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);  in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()