62 template <int BitShift, bool InputLessThanMod>
70   if (BitShift == 32) {  in FwdButterfly()
75   } else if (BitShift == 52) {  in FwdButterfly()
76     __m512i Q = _mm512_hexl_mulhi_epi<BitShift>(W_precon, *Y);  in FwdButterfly()
77     __m512i W_Y = _mm512_hexl_mullo_epi<BitShift>(W, *Y);  in FwdButterfly()
78     T = _mm512_hexl_mullo_add_lo_epi<BitShift>(W_Y, Q, neg_modulus);  in FwdButterfly()
79   } else if (BitShift == 64) {  in FwdButterfly()
82     __m512i Q = _mm512_hexl_mulhi_approx_epi<BitShift>(W_precon, *Y);  in FwdButterfly()
83     __m512i W_Y = _mm512_hexl_mullo_epi<BitShift>(W, *Y);  in FwdButterfly()
85     T = _mm512_hexl_mullo_add_lo_epi<BitShift>(W_Y, Q, neg_modulus);  in FwdButterfly()
89     HEXL_CHECK(false, "Invalid BitShift " << BitShift);  in FwdButterfly()
97 template <int BitShift>
116     FwdButterfly<BitShift, false>(&v_X, &v_Y, v_W, v_W_precon, v_neg_modulus,  in FwdT1()
124 template <int BitShift>
148     FwdButterfly<BitShift, false>(&v_X, &v_Y, v_W, v_W_precon, v_neg_modulus,  in FwdT2()
158 template <int BitShift>
177     FwdButterfly<BitShift, false>(&v_X, &v_Y, v_W, v_W_precon, v_neg_modulus,  in FwdT4()
188 template <int BitShift, bool InputLessThanMod>
219       FwdButterfly<BitShift, InputLessThanMod>(&v_X, &v_Y, v_W, v_W_precon,  in FwdT8()
233 template <int BitShift>
241   HEXL_CHECK(modulus < NTT::s_max_fwd_modulus(BitShift),  in ForwardTransformToBitReverseAVX512()
242              "modulus " << modulus << " too large for BitShift " << BitShift  in ForwardTransformToBitReverseAVX512()
244                         << NTT::s_max_fwd_modulus(BitShift));  in ForwardTransformToBitReverseAVX512()
245   HEXL_CHECK_BOUNDS(precon_root_of_unity_powers, n, MaximumValue(BitShift),  in ForwardTransformToBitReverseAVX512()
247   HEXL_CHECK_BOUNDS(operand, n, MaximumValue(BitShift), "operand too large");  in ForwardTransformToBitReverseAVX512()
292         FwdT8<BitShift, true>(result, result, v_neg_modulus, v_twice_mod, t, m,  in ForwardTransformToBitReverseAVX512()
295         FwdT8<BitShift, false>(result, result, v_neg_modulus, v_twice_mod, t, m,  in ForwardTransformToBitReverseAVX512()
306       FwdT8<BitShift, false>(result, result, v_neg_modulus, v_twice_mod, t, m,  in ForwardTransformToBitReverseAVX512()
348       FwdT4<BitShift>(result, v_neg_modulus, v_twice_mod, m, W, W_precon);  in ForwardTransformToBitReverseAVX512()
355       FwdT2<BitShift>(result, v_neg_modulus, v_twice_mod, m, W, W_precon);  in ForwardTransformToBitReverseAVX512()
362       FwdT1<BitShift>(result, v_neg_modulus, v_twice_mod, m, W, W_precon);  in ForwardTransformToBitReverseAVX512()
391     FwdT8<BitShift, false>(result, operand, v_neg_modulus, v_twice_mod, t, 1, W,  in ForwardTransformToBitReverseAVX512()
394     ForwardTransformToBitReverseAVX512<BitShift>(  in ForwardTransformToBitReverseAVX512()
399     ForwardTransformToBitReverseAVX512<BitShift>(  in ForwardTransformToBitReverseAVX512()