64 template <int BitShift, bool InputLessThanMod>
79   if (BitShift == 32) {  in InvButterfly()
84   } else if (BitShift == 52) {  in InvButterfly()
85     __m512i Q = _mm512_hexl_mulhi_epi<BitShift>(W_precon, T);  in InvButterfly()
86     __m512i Q_p = _mm512_hexl_mullo_epi<BitShift>(Q, neg_modulus);  in InvButterfly()
87     *Y = _mm512_hexl_mullo_add_lo_epi<BitShift>(Q_p, W, T);  in InvButterfly()
88   } else if (BitShift == 64) {  in InvButterfly()
91     __m512i Q = _mm512_hexl_mulhi_approx_epi<BitShift>(W_precon, T);  in InvButterfly()
92     __m512i Q_p = _mm512_hexl_mullo_epi<BitShift>(Q, neg_modulus);  in InvButterfly()
94     *Y = _mm512_hexl_mullo_add_lo_epi<BitShift>(Q_p, W, T);  in InvButterfly()
98     HEXL_CHECK(false, "Invalid BitShift " << BitShift);  in InvButterfly()
102 template <int BitShift, bool InputLessThanMod>
122     InvButterfly<BitShift, InputLessThanMod>(&v_X, &v_Y, v_W, v_W_precon,  in InvT1()
132 template <int BitShift>
147     InvButterfly<BitShift, false>(&v_X, &v_Y, v_W, v_W_precon, v_neg_modulus,  in InvT2()
159 template <int BitShift>
176     InvButterfly<BitShift, false>(&v_X, &v_Y, v_W, v_W_precon, v_neg_modulus,  in InvT4()
187 template <int BitShift>
209       InvButterfly<BitShift, false>(&v_X, &v_Y, v_W, v_W_precon, v_neg_modulus,  in InvT8()
219 template <int BitShift>
231   HEXL_CHECK(modulus < NTT::s_max_inv_modulus(BitShift),  in InverseTransformFromBitReverseAVX512()
232              "modulus " << modulus << " too large for BitShift " << BitShift  in InverseTransformFromBitReverseAVX512()
234                         << NTT::s_max_inv_modulus(BitShift));  in InverseTransformFromBitReverseAVX512()
235   HEXL_CHECK_BOUNDS(precon_inv_root_of_unity_powers, n, MaximumValue(BitShift),  in InverseTransformFromBitReverseAVX512()
237   HEXL_CHECK_BOUNDS(operand, n, MaximumValue(BitShift), "operand too large");  in InverseTransformFromBitReverseAVX512()
270         InvT1<BitShift, true>(result, v_neg_modulus, v_twice_mod, m, W,  in InverseTransformFromBitReverseAVX512()
273         InvT1<BitShift, false>(result, v_neg_modulus, v_twice_mod, m, W,  in InverseTransformFromBitReverseAVX512()
286       InvT2<BitShift>(result, v_neg_modulus, v_twice_mod, m, W, W_precon);  in InverseTransformFromBitReverseAVX512()
296       InvT4<BitShift>(result, v_neg_modulus, v_twice_mod, m, W, W_precon);  in InverseTransformFromBitReverseAVX512()
306         InvT8<BitShift>(result, v_neg_modulus, v_twice_mod, t, m, W, W_precon);  in InverseTransformFromBitReverseAVX512()
314     InverseTransformFromBitReverseAVX512<BitShift>(  in InverseTransformFromBitReverseAVX512()
318     InverseTransformFromBitReverseAVX512<BitShift>(  in InverseTransformFromBitReverseAVX512()
334       InvT8<BitShift>(result, v_neg_modulus, v_twice_mod, t, m, W, W_precon);  in InverseTransformFromBitReverseAVX512()
348     MultiplyFactor mf_inv_n(InverseMod(n, modulus), BitShift, modulus);  in InverseTransformFromBitReverseAVX512()
352     MultiplyFactor mf_inv_n_w(MultiplyMod(inv_n, W, modulus), BitShift,  in InverseTransformFromBitReverseAVX512()
386       if (BitShift == 32) {  in InverseTransformFromBitReverseAVX512()
401             _mm512_hexl_mulhi_epi<BitShift>(v_inv_n_prime, X_plus_Y_mod2q);  in InverseTransformFromBitReverseAVX512()
404             _mm512_hexl_mullo_epi<BitShift>(v_inv_n, X_plus_Y_mod2q);  in InverseTransformFromBitReverseAVX512()
406             _mm512_hexl_mullo_add_lo_epi<BitShift>(inv_N_tx, Q1, v_neg_modulus);  in InverseTransformFromBitReverseAVX512()
408         __m512i Q2 = _mm512_hexl_mulhi_epi<BitShift>(v_inv_n_w_prime, T);  in InverseTransformFromBitReverseAVX512()
410         __m512i inv_N_W_T = _mm512_hexl_mullo_epi<BitShift>(v_inv_n_w, T);  in InverseTransformFromBitReverseAVX512()
411         v_Y = _mm512_hexl_mullo_add_lo_epi<BitShift>(inv_N_W_T, Q2,  in InverseTransformFromBitReverseAVX512()