91       const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));  in xnn_q8_gemm_ukernel_4x8__neon()  local
93       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
94       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
95       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
96       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
97       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
98       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
99       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
100       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
200 …      const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));  in xnn_q8_gemm_ukernel_4x8__neon()  local
202         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
203 …      vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
204         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
205 …      vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
206         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
207 …      vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
208         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);  in xnn_q8_gemm_ukernel_4x8__neon()
209 …      vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);  in xnn_q8_gemm_ukernel_4x8__neon()