#pragma once template<typename T> void compute_flux_gpu(const T *ws, const T *dt, const T *st, const T *dq, const T *cflh, const T *z0, T *zl, T *ri, T *re, T *lnzuzt, T *zu, T *zt, T *rith, T *cm, T *ch, T *ct, T *ckt, const T g, const T Pr_m, const T kappa, const T Pr_t_0_inv, const T Pr_t_inf_inv, const T alpha_m, const T alpha_h, const T b4, const T alfam, const T betam, const T an, const T h1, const T x8, const T an1, const T an2, const T g0, const T r0, const int it, const int lu_indx, const int grid_size);