72     int32x4_t vacc2xCDEF = vacc0xCDEF;  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()  local
108       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
130       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
152       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
174       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
197       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
219       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
241       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
263       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
295       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
318 …      vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
341 …      vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
364 …      vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
387 …     vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
410 …     vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
433 …     vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
458     vacc2xCDEF = vqrdmulhq_s32(vacc2xCDEF, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
477     vacc2xCDEF = vsraq_n_s32(vacc2xCDEF, vbicq_s32(vacc2xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
494     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
507 …const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), v…  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
521 …x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_…  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()