60     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()  local
64     int32x4_t vacc1x89AB = vacc0x89AB;  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
68     int32x4_t vacc2x89AB = vacc0x89AB;  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
72     int32x4_t vacc3x89AB = vacc0x89AB;  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
119         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
147         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
175         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
203         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
231         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
259         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
287         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
315         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc7));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
342         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
371           vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
400             vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
429               vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
458                 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
487                   vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
516                     vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
549     vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
568     vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
585     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
603 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v…  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
617 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0x…  in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()