61     int32x4_t vacc1xCDEF = vacc0xCDEF;  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()  local
107         vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc0));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
129         vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc1));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
151         vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc2));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
173         vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc3));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
195         vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc4));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
217         vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc5));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
239         vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc6));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
261         vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc7));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
287         vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc0));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
310           vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc1));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
333             vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc2));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
356               vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc3));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
379                 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc4));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
402                   vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc5));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
425                     vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc6));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
450     vacc1xCDEF = vqrdmulhq_s32(vacc1xCDEF, vmultiplier);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
465     vacc1xCDEF = vsraq_n_s32(vacc1xCDEF, vbicq_s32(vacc1xCDEF, vzero_shift_mask), 31);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
478     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_shift);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
489 …const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), v…  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
500 …x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_…  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()