68     int32x4_t vacc1xCDEF = vacc0xCDEF;  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()  local
106       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
128       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
150       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
172       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
195       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
217       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
239       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
261       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
291       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
314 …      vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
337 …      vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
360 …      vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
383 …     vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
406 …     vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
429 …     vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
454     vacc1xCDEF = vqrdmulhq_s32(vacc1xCDEF, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
473     vacc1xCDEF = vsraq_n_s32(vacc1xCDEF, vbicq_s32(vacc1xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
490     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
505 …const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), v…  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
519 …x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_…  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()