95     int32x4_t vacc6x4567 = vacc0x4567;  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()  local
134       vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa6), 0);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
154       vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa6), 1);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
174       vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa6), 2);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
194       vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa6), 3);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
214       vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa6), 0);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
234       vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa6), 1);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
254       vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa6), 2);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
274       vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa6), 3);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
314       vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa6), 0);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
335 …      vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa6), 1);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
356 …      vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa6), 2);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
377 …      vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa6), 3);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
398 …     vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa6), 0);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
419 …     vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa6), 1);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
440 …     vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa6), 2);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
465     vacc6x4567 = vqrdmulhq_s32(vacc6x4567, vmultiplier);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
484     vacc6x4567 = vsraq_n_s32(vacc6x4567, vbicq_s32(vacc6x4567, vzero_shift_mask), 31);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
501     vacc6x4567 = vrshlq_s32(vacc6x4567, vright_shift);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
513 …const int16x8_t vacc6x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x0123), vacc6x4567), v…  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
534       vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x0123), vqmovn_s32(vacc6x4567)), voutput_zero_point);  in xnn_qu8_gemm_minmax_ukernel_8x8__neon()