62     int32x4_t vacc1xCDEF = vacc0xCDEF;  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()  local
92       vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
114       vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
136       vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
158       vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
180       vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
202       vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
224       vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
246       vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc7));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
272       vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
295         vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
318           vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
341             vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
364               vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
387                 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
410                   vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
432     vacc1xCDEF = vqrdmulhq_s32(vacc1xCDEF, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
447     vacc1xCDEF = vsraq_n_s32(vacc1xCDEF, vbicq_s32(vacc1xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
460     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
471 …const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), v…  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
482 …x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_…  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()