88       const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()  local
90       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
91       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
92       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
93       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
94       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
95       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
238       const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()  local
242       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
243       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
246       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
247       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
250       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
251       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);  in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()