68     int32x4_t vacc3x4567 = vacc0x4567;  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()  local
90       vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c0));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
104       vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c1));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
118       vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c2));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
132       vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c3));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
146       vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c4));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
160       vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c5));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
174       vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c6));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
188       vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c7));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
211       vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c0));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
227         vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c1));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
243           vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c2));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
259             vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c3));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
275               vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c4));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
291                 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c5));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
307                   vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c6));  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
323     vacc3x4567 = vqrdmulhq_s32(vacc3x4567, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
334     vacc3x4567 = vsraq_n_s32(vacc3x4567, vbicq_s32(vacc3x4567, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
343     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
350 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v…  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()
358 …x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_…  in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup()