1a4963045SJacob Faibussowitsch #pragma once 2e0eea495SMark 3e0eea495SMark #include <petscdmplex.h> /*I "petscdmplex.h" I*/ 4e0eea495SMark #include <petscts.h> 5e0eea495SMark 68594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauPrintNorms(Vec, PetscInt); 78594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauCreateVelocitySpace(MPI_Comm, PetscInt, const char[], Vec *, Mat *, DM *); 88594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauDestroyVelocitySpace(DM *); 9f53e7263SMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauAccess(DM, Vec, PetscErrorCode (*)(DM, Vec, PetscInt, PetscInt, PetscInt, void *), void *); 10c3e4dd79SMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauAddMaxwellians(DM, Vec, PetscReal, PetscReal[], PetscReal[], PetscInt, PetscInt, PetscInt, void *); 118594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauCreateMassMatrix(DM dm, Mat *Amat); 128594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauIFunction(TS, PetscReal, Vec, Vec, Vec, void *); 138594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauIJacobian(TS, PetscReal, Vec, Vec, PetscReal, Mat, Mat, void *); 14e0eea495SMark 15bfc784b7SMark Adams typedef int LandauIdx; 16bfc784b7SMark Adams 17e0eea495SMark /* the Fokker-Planck-Landau context */ 18e0eea495SMark #if !defined(LANDAU_MAX_SPECIES) 19984ed092SMark Adams #if defined(PETSC_USE_DMLANDAU_2D) 20e0eea495SMark #define LANDAU_MAX_SPECIES 10 218a6f2e61SMark Adams #define LANDAU_MAX_GRIDS 3 2252cdd6eaSMark #else 23763ae2f8SMark Adams #define LANDAU_MAX_SPECIES 10 24c3e4dd79SMark Adams #define LANDAU_MAX_GRIDS 3 2552cdd6eaSMark #endif 268a6f2e61SMark Adams #else 278a6f2e61SMark Adams #define LANDAU_MAX_GRIDS 3 28e0eea495SMark #endif 29a587d139SMark 30a587d139SMark #if !defined(LANDAU_MAX_Q) 31f44ea33cSmarkadams4 #if defined(LANDAU_MAX_NQND) 32f44ea33cSmarkadams4 #error "LANDAU_MAX_NQND but not LANDAU_MAX_Q. Use -DLANDAU_MAX_Q=4 for Q3 elements" 33a587d139SMark #endif 34984ed092SMark Adams #if defined(PETSC_USE_DMLANDAU_2D) 356b664d00Smarkadams4 #define LANDAU_MAX_Q 6 36c3e4dd79SMark Adams #else 37*d043ef4cSMark Adams #define LANDAU_MAX_Q 5 38a587d139SMark #endif 39a587d139SMark #else 40f44ea33cSmarkadams4 #undef LANDAU_MAX_NQND 41e0eea495SMark #endif 42a587d139SMark 43984ed092SMark Adams #if defined(PETSC_USE_DMLANDAU_2D) 44930e68a5SMark Adams #define LANDAU_MAX_Q_FACE LANDAU_MAX_Q 45f44ea33cSmarkadams4 #define LANDAU_MAX_NQND (LANDAU_MAX_Q * LANDAU_MAX_Q) 463b65d667Smarkadams4 #define LANDAU_MAX_BATCH_SZ 1024 47984ed092SMark Adams #define LANDAU_DIM 2 48a587d139SMark #else 49930e68a5SMark Adams #define LANDAU_MAX_Q_FACE (LANDAU_MAX_Q * LANDAU_MAX_Q) 50f44ea33cSmarkadams4 #define LANDAU_MAX_NQND (LANDAU_MAX_Q * LANDAU_MAX_Q * LANDAU_MAX_Q) 513b65d667Smarkadams4 #define LANDAU_MAX_BATCH_SZ 64 52984ed092SMark Adams #define LANDAU_DIM 3 53a587d139SMark #endif 54a587d139SMark 559371c9d4SSatish Balay typedef enum { 569371c9d4SSatish Balay LANDAU_KOKKOS, 579371c9d4SSatish Balay LANDAU_CPU 589371c9d4SSatish Balay } LandauDeviceType; 59e8d2b73aSMark Adams 608fdabdddSMark Adams // static data - will be "device" data 61e8d2b73aSMark Adams typedef struct { 62e8d2b73aSMark Adams void *invJ; // nip*dim*dim 63e8d2b73aSMark Adams void *D; // nq*nb*dim 64e8d2b73aSMark Adams void *B; // nq*nb 65e8d2b73aSMark Adams void *alpha; // ns 66e8d2b73aSMark Adams void *beta; // ns 67e8d2b73aSMark Adams void *invMass; // ns 68e8d2b73aSMark Adams void *w; // nip 69e8d2b73aSMark Adams void *x; // nip 70e8d2b73aSMark Adams void *y; // nip 71e8d2b73aSMark Adams void *z; // nip 72e8d2b73aSMark Adams void *Eq_m; // ns - dynamic 73e8d2b73aSMark Adams void *f; // nip*Nf - dynamic (IP) 74e8d2b73aSMark Adams void *dfdx; // nip*Nf - dynamic (IP) 75e8d2b73aSMark Adams void *dfdy; // nip*Nf - dynamic (IP) 76e8d2b73aSMark Adams void *dfdz; // nip*Nf - dynamic (IP) 77e8d2b73aSMark Adams int dim_, ns_, nip_, nq_, nb_; 78f53e7263SMark Adams void *NCells; // remove and use elem_offset - TODO 798fdabdddSMark Adams void *species_offset; // for each grid, but same for all batched vertices 808fdabdddSMark Adams void *mat_offset; // for each grid, but same for all batched vertices 818fdabdddSMark Adams void *elem_offset; // for each grid, but same for all batched vertices 828fdabdddSMark Adams void *ip_offset; // for each grid, but same for all batched vertices 838fdabdddSMark Adams void *ipf_offset; // for each grid, but same for all batched vertices 848fdabdddSMark Adams void *ipfdf_data; // for each grid, but same for all batched vertices 858fdabdddSMark Adams void *maps; // for each grid, but same for all batched vertices 86bfc784b7SMark Adams // COO 87cada7fc7SMark Adams void *coo_elem_offsets; 88cada7fc7SMark Adams void *coo_elem_point_offsets; 89cada7fc7SMark Adams void *coo_elem_fullNb; 90a31f6053SMark Adams void *coo_vals; 91ae6bf4a8Smarkadams4 void *lambdas; 92cada7fc7SMark Adams LandauIdx coo_n_cellsTot; 93bfc784b7SMark Adams LandauIdx coo_size; 9413241b68SMark Adams LandauIdx coo_max_fullnb; 958a6f2e61SMark Adams } LandauStaticData; 96e8d2b73aSMark Adams 979371c9d4SSatish Balay typedef enum { 989371c9d4SSatish Balay LANDAU_EX2_TSSOLVE, 999371c9d4SSatish Balay LANDAU_MATRIX_TOTAL, 1009371c9d4SSatish Balay LANDAU_OPERATOR, 1019371c9d4SSatish Balay LANDAU_JACOBIAN_COUNT, 1029371c9d4SSatish Balay LANDAU_JACOBIAN, 1039371c9d4SSatish Balay LANDAU_MASS, 1049371c9d4SSatish Balay LANDAU_F_DF, 1059371c9d4SSatish Balay LANDAU_KERNEL, 1069371c9d4SSatish Balay KSP_FACTOR, 1079371c9d4SSatish Balay KSP_SOLVE, 1089371c9d4SSatish Balay LANDAU_NUM_TIMERS 1099371c9d4SSatish Balay } LandauOMPTimers; 1108fdabdddSMark Adams 111e0eea495SMark typedef struct { 112e0eea495SMark PetscBool interpolate; /* Generate intermediate mesh elements */ 113a587d139SMark PetscBool gpu_assembly; 1148a6f2e61SMark Adams MPI_Comm comm; /* global communicator to use for errors and diagnostics */ 1158fdabdddSMark Adams double times[LANDAU_NUM_TIMERS]; 1168a6f2e61SMark Adams PetscBool use_matrix_mass; 1178a6f2e61SMark Adams /* FE */ 118e0eea495SMark PetscFE fe[LANDAU_MAX_SPECIES]; 119e0eea495SMark /* geometry */ 1208a6f2e61SMark Adams PetscReal radius[LANDAU_MAX_GRIDS]; 121531b49fdSmarkadams4 PetscReal radius_par[LANDAU_MAX_GRIDS]; 122531b49fdSmarkadams4 PetscReal radius_perp[LANDAU_MAX_GRIDS]; 1238a6f2e61SMark Adams PetscReal re_radius; /* RE: radius of refinement along v_perp=0, z>0 */ 1248a6f2e61SMark Adams PetscReal vperp0_radius1; /* RE: radius of refinement along v_perp=0 */ 1258a6f2e61SMark Adams PetscReal vperp0_radius2; /* RE: radius of refinement along v_perp=0 after origin AMR refinement */ 126e04ae51bSMark Adams PetscBool sphere; 127*d043ef4cSMark Adams PetscReal sphere_inner_radius_90degree[LANDAU_MAX_GRIDS]; 128*d043ef4cSMark Adams PetscReal sphere_inner_radius_45degree[LANDAU_MAX_GRIDS]; 129f53e7263SMark Adams PetscInt cells0[3]; 1308a6f2e61SMark Adams /* AMR */ 1318a6f2e61SMark Adams PetscBool use_p4est; 1328a6f2e61SMark Adams PetscInt numRERefine; /* RE: refinement along v_perp=0, z > 0 */ 1338a6f2e61SMark Adams PetscInt nZRefine1; /* RE: origin refinement after v_perp=0 refinement */ 1348a6f2e61SMark Adams PetscInt nZRefine2; /* RE: origin refinement after origin AMR refinement */ 1358a6f2e61SMark Adams PetscInt numAMRRefine[LANDAU_MAX_GRIDS]; /* normal AMR - refine from origin */ 1368a6f2e61SMark Adams PetscInt postAMRRefine[LANDAU_MAX_GRIDS]; /* uniform refinement of AMR */ 137cd27c6deSmarkadams4 PetscBool simplex; 138f44ea33cSmarkadams4 char filename[PETSC_MAX_PATH_LEN]; 139f44ea33cSmarkadams4 PetscReal thermal_speed[LANDAU_MAX_GRIDS]; 140*d043ef4cSMark Adams PetscBool sphere_uniform_normal; 1418a6f2e61SMark Adams /* relativistic */ 142cefb98e8SMark Adams PetscBool use_energy_tensor_trick; 143cefb98e8SMark Adams PetscBool use_relativistic_corrections; 144e0eea495SMark /* physics */ 145e0eea495SMark PetscReal thermal_temps[LANDAU_MAX_SPECIES]; 146e0eea495SMark PetscReal masses[LANDAU_MAX_SPECIES]; /* mass of each species */ 147e0eea495SMark PetscReal charges[LANDAU_MAX_SPECIES]; /* charge of each species */ 148e0eea495SMark PetscReal n[LANDAU_MAX_SPECIES]; /* number density of each species */ 149e0eea495SMark PetscReal m_0; /* reference mass */ 150e0eea495SMark PetscReal v_0; /* reference velocity */ 151e0eea495SMark PetscReal n_0; /* reference number density */ 152e0eea495SMark PetscReal t_0; /* reference time */ 153e0eea495SMark PetscReal Ez; 154e0eea495SMark PetscReal epsilon0; 155e0eea495SMark PetscReal k; 156ae6bf4a8Smarkadams4 PetscReal lambdas[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS]; 1578a6f2e61SMark Adams PetscReal electronShift; 158e0eea495SMark PetscInt num_species; 1598a6f2e61SMark Adams PetscInt num_grids; 1608fdabdddSMark Adams PetscInt species_offset[LANDAU_MAX_GRIDS + 1]; // for each grid, but same for all batched vertices 1618fdabdddSMark Adams PetscInt mat_offset[LANDAU_MAX_GRIDS + 1]; // for each grid, but same for all batched vertices 162cb25d741SMark Adams // batching 163cb25d741SMark Adams PetscBool jacobian_field_major_order; // this could be a type but lets not get pedantic 164cb25d741SMark Adams VecScatter plex_batch; 165cb25d741SMark Adams Vec work_vec; 166cb25d741SMark Adams IS batch_is; 167cb25d741SMark Adams PetscErrorCode (*seqaij_mult)(Mat, Vec, Vec); 168cb25d741SMark Adams PetscErrorCode (*seqaij_multtranspose)(Mat, Vec, Vec); 169cb25d741SMark Adams PetscErrorCode (*seqaij_solve)(Mat, Vec, Vec); 170cb25d741SMark Adams PetscErrorCode (*seqaij_getdiagonal)(Mat, Vec); 171bfc784b7SMark Adams /* COO */ 172e0eea495SMark Mat J; 173e0eea495SMark Mat M; 174e0eea495SMark Vec X; 175e0eea495SMark /* derived type */ 176e0eea495SMark void *data; 177e0eea495SMark /* computing */ 178e0eea495SMark LandauDeviceType deviceType; 1798a6f2e61SMark Adams DM pack; 1808a6f2e61SMark Adams DM plex[LANDAU_MAX_GRIDS]; 1818fdabdddSMark Adams LandauStaticData SData_d; /* static geometric data on device */ 18254545eeeSMark Adams /* diagnostics */ 18354545eeeSMark Adams PetscInt verbose; 18454545eeeSMark Adams PetscLogEvent events[20]; 185c751c0a2SMark Adams PetscLogStage stage; 186984ed092SMark Adams PetscObjectState norm_state; 1878fdabdddSMark Adams PetscInt batch_sz; 1888fdabdddSMark Adams PetscInt batch_view_idx; 189e0eea495SMark } LandauCtx; 190e0eea495SMark 1918fdabdddSMark Adams #define LANDAU_SPECIES_MAJOR 1928fdabdddSMark Adams #if !defined(LANDAU_SPECIES_MAJOR) 1938fdabdddSMark Adams #define LAND_PACK_IDX(_b, _g) (_b * ctx->num_grids + _g) 1948fdabdddSMark Adams #define LAND_MOFFSET(_b, _g, _nbch, _ngrid, _mat_off) (_b * _mat_off[_ngrid] + _mat_off[_g]) 1958fdabdddSMark Adams #else 1968fdabdddSMark Adams #define LAND_PACK_IDX(_b, _g) (_g * ctx->batch_sz + _b) 1978fdabdddSMark Adams #define LAND_MOFFSET(_b, _g, _nbch, _ngrid, _mat_off) (_nbch * _mat_off[_g] + _b * (_mat_off[_g + 1] - _mat_off[_g])) 1988fdabdddSMark Adams #endif 1998fdabdddSMark Adams 200a587d139SMark typedef struct { 201a587d139SMark PetscReal scale; 202bfc784b7SMark Adams LandauIdx gid; // Landau matrix index (<10,000) 203a587d139SMark } pointInterpolationP4est; 204a587d139SMark typedef struct _lP4estVertexMaps { 205f44ea33cSmarkadams4 LandauIdx (*gIdx)[LANDAU_MAX_SPECIES][LANDAU_MAX_NQND]; // #elems * LANDAU_MAX_NQND 206a587d139SMark LandauIdx num_elements; 207a587d139SMark LandauIdx num_reduced; 208a587d139SMark LandauIdx num_face; // (Q or Q^2 for 3D) 209a587d139SMark LandauDeviceType deviceType; 210a587d139SMark PetscInt Nf; 211a587d139SMark pointInterpolationP4est (*c_maps)[LANDAU_MAX_Q_FACE]; 2128a6f2e61SMark Adams struct _lP4estVertexMaps *d_self; 213a587d139SMark void *vp1, *vp2, *vp3; 2148a6f2e61SMark Adams PetscInt numgrids; 215a587d139SMark } P4estVertexMaps; 216a587d139SMark 217e0eea495SMark #if defined(PETSC_HAVE_KOKKOS) 218f44ea33cSmarkadams4 PETSC_EXTERN PetscErrorCode LandauKokkosJacobian(DM[], const PetscInt, const PetscInt, const PetscInt, const PetscInt, const PetscInt[], PetscReal[], PetscScalar[], const PetscScalar[], const LandauStaticData *, const PetscReal, const PetscLogEvent[], const PetscInt[], const PetscInt[], Mat[], Mat); 219f44ea33cSmarkadams4 PETSC_EXTERN PetscErrorCode LandauKokkosCreateMatMaps(P4estVertexMaps *, pointInterpolationP4est (*)[LANDAU_MAX_Q_FACE], PetscInt[], PetscInt); 2208a6f2e61SMark Adams PETSC_EXTERN PetscErrorCode LandauKokkosDestroyMatMaps(P4estVertexMaps *, PetscInt); 221f44ea33cSmarkadams4 PETSC_EXTERN PetscErrorCode LandauKokkosStaticDataSet(DM, const PetscInt, const PetscInt, const PetscInt, const PetscInt, PetscInt[], PetscInt[], PetscInt[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], LandauStaticData *); 2228a6f2e61SMark Adams PETSC_EXTERN PetscErrorCode LandauKokkosStaticDataClear(LandauStaticData *); 223e0eea495SMark #endif 224