65     int32x4_t vacc3x4567 = vacc0x4567;  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()  local
107         vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c0));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
121         vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c1));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
135         vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c2));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
149         vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c3));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
163         vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c4));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
177         vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c5));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
191         vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c6));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
205         vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c7));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
228         vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c0));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
244           vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c1));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
260             vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c2));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
276               vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c3));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
292                 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c4));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
308                   vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c5));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
324                     vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c6));  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
343     vacc3x4567 = vqrdmulhq_s32(vacc3x4567, vmultiplier);  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
354     vacc3x4567 = vsraq_n_s32(vacc3x4567, vbicq_s32(vacc3x4567, vzero_shift_mask), 31);  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
363     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_shift);  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
370 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v…  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()
378 …x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_…  in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mull_addw_dup()