68     int32x4_t vacc3x4567 = vacc0x4567;  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()  local
91       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
102       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
113       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
124       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
136       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
147       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
158       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
169       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
193       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
206 …      vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
219 …      vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
232 …      vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
245 …     vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
258 …     vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
271 …     vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
288     vacc3x4567 = vqrdmulhq_s32(vacc3x4567, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
299     vacc3x4567 = vsraq_n_s32(vacc3x4567, vbicq_s32(vacc3x4567, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
308     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
315 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v…  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
323 …x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_…  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()