63     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()  local
67     int32x4_t vacc1x89AB = vacc0x89AB;  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
71     int32x4_t vacc2x89AB = vacc0x89AB;  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
75     int32x4_t vacc3x89AB = vacc0x89AB;  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
102       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
130       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
158       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
186       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
214       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
242       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
270       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
298       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc7));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
325       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
354         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
383           vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
412             vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
441               vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
470                 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
499                   vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
529     vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
548     vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
565     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
583 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v…  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
597 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0x…  in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()