102       const int16x8_t vxb89ABCDEFc0 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc0, vb_zero_point));  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()  local
104       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
105       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
106       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
107       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
108       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
109       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
110       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
111       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
284       const int16x8_t vxb89ABCDEFc0 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc0, vb_zero_point));  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()  local
288       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
289       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
292       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
293       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
296       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
297       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
300       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
301       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);  in xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()