65       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
67       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
72       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
74       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
79       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
81       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
86       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
88       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
93       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
94       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
95       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
96       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
100       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
101       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
102       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
103       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
107       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
108       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
109       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
110       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
114       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
115       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
116       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
117       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
131       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
133       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
140 …      vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
142 …      vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
149 …      vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
151 …      vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
158 …      vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
160 …      vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
166 …      vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
167 …     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
168 …      vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
169 …     vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
175 …      vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
176 …     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
177 …      vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
178 …     vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
184 …      vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
185 …     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
186 …      vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
187 …     vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()