66     int32x4_t vacc2xCDEF = vacc0xCDEF;  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()  local
95       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
117       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
139       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
161       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
183       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
205       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
227       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
249       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc7));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
278       vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
301         vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
324           vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
347             vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
370               vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
393                 vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
416                   vacc2xCDEF = vaddw_s16(vacc2xCDEF, vget_high_s16(vprod2x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
436     vacc2xCDEF = vqrdmulhq_s32(vacc2xCDEF, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
451     vacc2xCDEF = vsraq_n_s32(vacc2xCDEF, vbicq_s32(vacc2xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
464     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
473 …const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), v…  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
484 …x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_…  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()