51     int32x4_t vacc1x0123 = vacc0x0123;  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()  local
65       vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
73       vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
81       vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
89       vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
97       vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
105       vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
113       vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
121       vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c7));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
136       vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
146         vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
156           vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
166             vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
176               vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
186                 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
196                   vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
208     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
215     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
220     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
226 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v…  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()
231 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x…  in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup()