57     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()  local
61     int32x4_t vacc1x89AB = vacc0x89AB;  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
65     int32x4_t vacc2x89AB = vacc0x89AB;  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
88       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
110       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
132       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
154       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
176       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
198       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
220       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
242       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc7));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
265       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
288         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
311           vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
334             vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
357               vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
380                 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
403                   vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
427     vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
442     vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
455     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
469 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v…  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()
480 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0x…  in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup()