123       const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()  local
125       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()
126       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()
127       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_rndnu_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_rndnu_ukernel_4x16__neon_mlal_lane()
129       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);  in xnn_qs8_gemm_minmax_rndnu_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_rndnu_ukernel_4x16__neon_mlal_lane()
131       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()
132       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()
306         const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()  local
310         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()
311 …      vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()
314         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()
315 …      vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()
318         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()
319 …      vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()
322         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()
323 …      vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane()