44     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()  local
54       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
59       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
64       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
69       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
74       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
79       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
84       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
89       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c7));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
100       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
107         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
114           vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
121             vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
128               vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
135                 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
142                   vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
152     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
157     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
160     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
164 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v…  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()
168 …x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_…  in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup()