21 template <int BitShift = 64>
40   constexpr int64_t alpha = BitShift - 2;  in EltwiseReduceModAVX512()
47       MultiplyFactor(uint64_t(1) << (ceil_log_mod + alpha - BitShift), BitShift,  in EltwiseReduceModAVX512()
53   if (BitShift == 64) {  in EltwiseReduceModAVX512()
81         v_op = _mm512_hexl_barrett_reduce64<BitShift, 2>(  in EltwiseReduceModAVX512()
92         v_op = _mm512_hexl_barrett_reduce64<BitShift, 1>(  in EltwiseReduceModAVX512()
154 template <int BitShift, int r>
171   if (BitShift == 64) {  in EltwiseMontReduceModAVX512()
185       result[i] = MontgomeryReduce<BitShift>(T_hi, T_lo, modulus, r, mod_R_mask,  in EltwiseMontReduceModAVX512()
204     __m512i v_T_hi = _mm512_hexl_mulhi_epi<BitShift>(v_a_op, v_b_op);  in EltwiseMontReduceModAVX512()
205     __m512i v_T_lo = _mm512_hexl_mullo_epi<BitShift>(v_a_op, v_b_op);  in EltwiseMontReduceModAVX512()
207     if (BitShift == 64) {  in EltwiseMontReduceModAVX512()
214     __m512i v_c = _mm512_hexl_montgomery_reduce<BitShift, r>(  in EltwiseMontReduceModAVX512()
237 template <int BitShift, int r>
253   if (BitShift == 64) {  in EltwiseMontgomeryFormAVX512()
267       result[i] = MontgomeryReduce<BitShift>(T_hi, T_lo, modulus, r, mod_R_mask,  in EltwiseMontgomeryFormAVX512()
284     __m512i v_T_hi = _mm512_hexl_mulhi_epi<BitShift>(v_a_op, v_b);  in EltwiseMontgomeryFormAVX512()
285     __m512i v_T_lo = _mm512_hexl_mullo_epi<BitShift>(v_a_op, v_b);  in EltwiseMontgomeryFormAVX512()
287     if (BitShift == 64) {  in EltwiseMontgomeryFormAVX512()
294     __m512i v_c = _mm512_hexl_montgomery_reduce<BitShift, r>(  in EltwiseMontgomeryFormAVX512()