75     int32x4_t vacc3x89AB = vacc0x89AB;  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()  local
111       vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
139       vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
167       vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
195       vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
223       vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
251       vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
279       vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
307       vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc7));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
343       vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
372         vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
401           vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
430             vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
459               vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
488                 vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
517                   vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
541     vacc3x89AB = vqrdmulhq_s32(vacc3x89AB, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
560     vacc3x89AB = vsraq_n_s32(vacc3x89AB, vbicq_s32(vacc3x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
577     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
589 …const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), v…  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
603 …const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3x…  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()