151       const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));  in xnn_q8_gemm_ukernel_4x8__neon()  local
153       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
154       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
155       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
156       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
157       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
158       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
159       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
160       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
265 …      const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));  in xnn_q8_gemm_ukernel_4x8__neon()  local
267 …      vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
268 …     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
269 …      vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
270 …     vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
271 …      vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
272 …     vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
273 …      vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);  in xnn_q8_gemm_ukernel_4x8__neon()
274 …     vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);  in xnn_q8_gemm_ukernel_4x8__neon()