45     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()  local
60       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
70       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
80       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
90       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
100       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
110       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
120       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
130       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc7));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
145       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
156         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
167           vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
178             vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
189               vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
200                 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
211                   vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
223     vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
230     vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
235     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
241 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v…  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()
246 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0x…  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup()