72     int32x4_t vacc3x89AB = vacc0x89AB;  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()  local
126         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
148         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
170         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
192         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
215 …      vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
237 …      vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
259 …      vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
281 …      vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
315         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
338 …       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
361 …       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
384 …       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
407 …      vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
430 …      vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
453 …      vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
480     vacc3x89AB = vqrdmulhq_s32(vacc3x89AB, vmultiplier);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
499     vacc3x89AB = vsraq_n_s32(vacc3x89AB, vbicq_s32(vacc3x89AB, vzero_shift_mask), 31);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
516     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_shift);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
528 …const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), v…  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
542 …const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3x…  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()