50     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()  local
52     int32x4_t vacc1x4567 = vacc0x4567;  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
65       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
72       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
79       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
86       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
94       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
101       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
108       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
115       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
131       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
140 …      vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
149 …      vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
158 …      vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
167 …     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
176 …     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
185 …     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
198     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
205     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
210     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
216 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v…  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
221 …x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_…  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()