72     int32x4_t vacc2xCDEF = vacc0xCDEF;  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()  local
109       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
137       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
165       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
193       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
221       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
249       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
277       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
305       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc7));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
338       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
367         vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
396           vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
425             vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
454               vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
483                 vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
512                   vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
538     vacc2xCDEF = vqrdmulhq_s32(vacc2xCDEF, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
557     vacc2xCDEF = vsraq_n_s32(vacc2xCDEF, vbicq_s32(vacc2xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
574     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
587 …const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), v…  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
601 …x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_…  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()