105       const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()  local
107       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()
108       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()
109       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()
110       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()
111       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()
112       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()
256         const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()  local
260         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()
261 …      vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()
264         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()
265 …      vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()
268         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()
269 …      vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);  in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane()