51     int32x4_t vacc1x0123 = vacc0x0123;  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()  local
66       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
73       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
80       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
87       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);  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()
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()
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()
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()
132       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
141         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
150 …       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
159 …       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);  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()
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()
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()
199     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
206     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
211     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
217 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v…  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
222 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x…  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()