16524c165SJacob Faibussowitsch #ifndef PETSCLANDAU_H 2e0eea495SMark #define PETSCLANDAU_H 3e0eea495SMark 4e0eea495SMark #include <petscdmplex.h> /*I "petscdmplex.h" I*/ 5e0eea495SMark #include <petscts.h> 6e0eea495SMark 78594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauPrintNorms(Vec, PetscInt); 88594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauCreateVelocitySpace(MPI_Comm, PetscInt, const char[], Vec *, Mat *, DM *); 98594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauDestroyVelocitySpace(DM *); 10f53e7263SMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauAccess(DM, Vec, PetscErrorCode (*)(DM, Vec, PetscInt, PetscInt, PetscInt, void *), void *); 11c3e4dd79SMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauAddMaxwellians(DM, Vec, PetscReal, PetscReal[], PetscReal[], PetscInt, PetscInt, PetscInt, void *); 128594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauCreateMassMatrix(DM dm, Mat *Amat); 138594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauIFunction(TS, PetscReal, Vec, Vec, Vec, void *); 148594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauIJacobian(TS, PetscReal, Vec, Vec, PetscReal, Mat, Mat, void *); 15e0eea495SMark 16bfc784b7SMark Adams typedef int LandauIdx; 17bfc784b7SMark Adams 18e0eea495SMark /* the Fokker-Planck-Landau context */ 19e0eea495SMark #if !defined(LANDAU_MAX_SPECIES) 20984ed092SMark Adams #if defined(PETSC_USE_DMLANDAU_2D) 21e0eea495SMark #define LANDAU_MAX_SPECIES 10 228a6f2e61SMark Adams #define LANDAU_MAX_GRIDS 3 2352cdd6eaSMark #else 24763ae2f8SMark Adams #define LANDAU_MAX_SPECIES 10 25c3e4dd79SMark Adams #define LANDAU_MAX_GRIDS 3 2652cdd6eaSMark #endif 278a6f2e61SMark Adams #else 288a6f2e61SMark Adams #define LANDAU_MAX_GRIDS 3 29e0eea495SMark #endif 30a587d139SMark 31a587d139SMark #if !defined(LANDAU_MAX_Q) 32a587d139SMark #if defined(LANDAU_MAX_NQ) 33a587d139SMark #error "LANDAU_MAX_NQ but not LANDAU_MAX_Q. Use -DLANDAU_MAX_Q=4 for Q3 elements" 34a587d139SMark #endif 35984ed092SMark Adams #if defined(PETSC_USE_DMLANDAU_2D) 366b664d00Smarkadams4 #define LANDAU_MAX_Q 6 37c3e4dd79SMark Adams #else 383b65d667Smarkadams4 // 3D CUDA fails with > 3 (KK-CUDA is OK) 396b664d00Smarkadams4 #define LANDAU_MAX_Q 4 40a587d139SMark #endif 41a587d139SMark #else 42a587d139SMark #undef LANDAU_MAX_NQ 43e0eea495SMark #endif 44a587d139SMark 45984ed092SMark Adams #if defined(PETSC_USE_DMLANDAU_2D) 46930e68a5SMark Adams #define LANDAU_MAX_Q_FACE LANDAU_MAX_Q 47a587d139SMark #define LANDAU_MAX_NQ (LANDAU_MAX_Q * LANDAU_MAX_Q) 483b65d667Smarkadams4 #define LANDAU_MAX_BATCH_SZ 1024 49984ed092SMark Adams #define LANDAU_DIM 2 50a587d139SMark #else 51930e68a5SMark Adams #define LANDAU_MAX_Q_FACE (LANDAU_MAX_Q * LANDAU_MAX_Q) 52a587d139SMark #define LANDAU_MAX_NQ (LANDAU_MAX_Q * LANDAU_MAX_Q * LANDAU_MAX_Q) 533b65d667Smarkadams4 #define LANDAU_MAX_BATCH_SZ 64 54984ed092SMark Adams #define LANDAU_DIM 3 55a587d139SMark #endif 56a587d139SMark 579371c9d4SSatish Balay typedef enum { 589371c9d4SSatish Balay LANDAU_CUDA, 599371c9d4SSatish Balay LANDAU_KOKKOS, 609371c9d4SSatish Balay LANDAU_CPU 619371c9d4SSatish Balay } LandauDeviceType; 62e8d2b73aSMark Adams 638fdabdddSMark Adams // static data - will be "device" data 64e8d2b73aSMark Adams typedef struct { 65e8d2b73aSMark Adams void *invJ; // nip*dim*dim 66e8d2b73aSMark Adams void *D; // nq*nb*dim 67e8d2b73aSMark Adams void *B; // nq*nb 68e8d2b73aSMark Adams void *alpha; // ns 69e8d2b73aSMark Adams void *beta; // ns 70e8d2b73aSMark Adams void *invMass; // ns 71e8d2b73aSMark Adams void *w; // nip 72e8d2b73aSMark Adams void *x; // nip 73e8d2b73aSMark Adams void *y; // nip 74e8d2b73aSMark Adams void *z; // nip 75e8d2b73aSMark Adams void *Eq_m; // ns - dynamic 76e8d2b73aSMark Adams void *f; // nip*Nf - dynamic (IP) 77e8d2b73aSMark Adams void *dfdx; // nip*Nf - dynamic (IP) 78e8d2b73aSMark Adams void *dfdy; // nip*Nf - dynamic (IP) 79e8d2b73aSMark Adams void *dfdz; // nip*Nf - dynamic (IP) 80e8d2b73aSMark Adams int dim_, ns_, nip_, nq_, nb_; 81f53e7263SMark Adams void *NCells; // remove and use elem_offset - TODO 828fdabdddSMark Adams void *species_offset; // for each grid, but same for all batched vertices 838fdabdddSMark Adams void *mat_offset; // for each grid, but same for all batched vertices 848fdabdddSMark Adams void *elem_offset; // for each grid, but same for all batched vertices 858fdabdddSMark Adams void *ip_offset; // for each grid, but same for all batched vertices 868fdabdddSMark Adams void *ipf_offset; // for each grid, but same for all batched vertices 878fdabdddSMark Adams void *ipfdf_data; // for each grid, but same for all batched vertices 888fdabdddSMark Adams void *maps; // for each grid, but same for all batched vertices 89bfc784b7SMark Adams // COO 90cada7fc7SMark Adams void *coo_elem_offsets; 91cada7fc7SMark Adams void *coo_elem_point_offsets; 92cada7fc7SMark Adams void *coo_elem_fullNb; 93a31f6053SMark Adams void *coo_vals; 94ae6bf4a8Smarkadams4 void *lambdas; 95cada7fc7SMark Adams LandauIdx coo_n_cellsTot; 96bfc784b7SMark Adams LandauIdx coo_size; 9713241b68SMark Adams LandauIdx coo_max_fullnb; 988a6f2e61SMark Adams } LandauStaticData; 99e8d2b73aSMark Adams 1009371c9d4SSatish Balay typedef enum { 1019371c9d4SSatish Balay LANDAU_EX2_TSSOLVE, 1029371c9d4SSatish Balay LANDAU_MATRIX_TOTAL, 1039371c9d4SSatish Balay LANDAU_OPERATOR, 1049371c9d4SSatish Balay LANDAU_JACOBIAN_COUNT, 1059371c9d4SSatish Balay LANDAU_JACOBIAN, 1069371c9d4SSatish Balay LANDAU_MASS, 1079371c9d4SSatish Balay LANDAU_F_DF, 1089371c9d4SSatish Balay LANDAU_KERNEL, 1099371c9d4SSatish Balay KSP_FACTOR, 1109371c9d4SSatish Balay KSP_SOLVE, 1119371c9d4SSatish Balay LANDAU_NUM_TIMERS 1129371c9d4SSatish Balay } LandauOMPTimers; 1138fdabdddSMark Adams 114e0eea495SMark typedef struct { 115e0eea495SMark PetscBool interpolate; /* Generate intermediate mesh elements */ 116a587d139SMark PetscBool gpu_assembly; 1178a6f2e61SMark Adams MPI_Comm comm; /* global communicator to use for errors and diagnostics */ 1188fdabdddSMark Adams double times[LANDAU_NUM_TIMERS]; 1198a6f2e61SMark Adams PetscBool use_matrix_mass; 1208a6f2e61SMark Adams /* FE */ 121e0eea495SMark PetscFE fe[LANDAU_MAX_SPECIES]; 122e0eea495SMark /* geometry */ 1238a6f2e61SMark Adams PetscReal radius[LANDAU_MAX_GRIDS]; 124531b49fdSmarkadams4 PetscReal radius_par[LANDAU_MAX_GRIDS]; 125531b49fdSmarkadams4 PetscReal radius_perp[LANDAU_MAX_GRIDS]; 1268a6f2e61SMark Adams PetscReal re_radius; /* RE: radius of refinement along v_perp=0, z>0 */ 1278a6f2e61SMark Adams PetscReal vperp0_radius1; /* RE: radius of refinement along v_perp=0 */ 1288a6f2e61SMark Adams PetscReal vperp0_radius2; /* RE: radius of refinement along v_perp=0 after origin AMR refinement */ 129e04ae51bSMark Adams PetscBool sphere; 130e04ae51bSMark Adams PetscReal sphere_inner_radius_90degree; 131e04ae51bSMark Adams PetscReal sphere_inner_radius_45degree; 132f53e7263SMark Adams PetscInt cells0[3]; 1338a6f2e61SMark Adams /* AMR */ 1348a6f2e61SMark Adams PetscBool use_p4est; 1358a6f2e61SMark Adams PetscInt numRERefine; /* RE: refinement along v_perp=0, z > 0 */ 1368a6f2e61SMark Adams PetscInt nZRefine1; /* RE: origin refinement after v_perp=0 refinement */ 1378a6f2e61SMark Adams PetscInt nZRefine2; /* RE: origin refinement after origin AMR refinement */ 1388a6f2e61SMark Adams PetscInt numAMRRefine[LANDAU_MAX_GRIDS]; /* normal AMR - refine from origin */ 1398a6f2e61SMark Adams PetscInt postAMRRefine[LANDAU_MAX_GRIDS]; /* uniform refinement of AMR */ 140*cd27c6deSmarkadams4 PetscBool simplex; 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 */ 172aab769cbSMark Adams PetscBool coo_assembly; 173e0eea495SMark /* cache */ 174e0eea495SMark Mat J; 175e0eea495SMark Mat M; 176e0eea495SMark Vec X; 177e0eea495SMark /* derived type */ 178e0eea495SMark void *data; 179e0eea495SMark /* computing */ 180e0eea495SMark LandauDeviceType deviceType; 1818a6f2e61SMark Adams DM pack; 1828a6f2e61SMark Adams DM plex[LANDAU_MAX_GRIDS]; 1838fdabdddSMark Adams LandauStaticData SData_d; /* static geometric data on device */ 18454545eeeSMark Adams /* diagnostics */ 18554545eeeSMark Adams PetscInt verbose; 18654545eeeSMark Adams PetscLogEvent events[20]; 187c751c0a2SMark Adams PetscLogStage stage; 188984ed092SMark Adams PetscObjectState norm_state; 1898fdabdddSMark Adams PetscInt batch_sz; 1908fdabdddSMark Adams PetscInt batch_view_idx; 191e0eea495SMark } LandauCtx; 192e0eea495SMark 1938fdabdddSMark Adams #define LANDAU_SPECIES_MAJOR 1948fdabdddSMark Adams #if !defined(LANDAU_SPECIES_MAJOR) 1958fdabdddSMark Adams #define LAND_PACK_IDX(_b, _g) (_b * ctx->num_grids + _g) 1968fdabdddSMark Adams #define LAND_MOFFSET(_b, _g, _nbch, _ngrid, _mat_off) (_b * _mat_off[_ngrid] + _mat_off[_g]) 1978fdabdddSMark Adams #else 1988fdabdddSMark Adams #define LAND_PACK_IDX(_b, _g) (_g * ctx->batch_sz + _b) 1998fdabdddSMark Adams #define LAND_MOFFSET(_b, _g, _nbch, _ngrid, _mat_off) (_nbch * _mat_off[_g] + _b * (_mat_off[_g + 1] - _mat_off[_g])) 2008fdabdddSMark Adams #endif 2018fdabdddSMark Adams 202a587d139SMark typedef struct { 203a587d139SMark PetscReal scale; 204bfc784b7SMark Adams LandauIdx gid; // Landau matrix index (<10,000) 205a587d139SMark } pointInterpolationP4est; 206a587d139SMark typedef struct _lP4estVertexMaps { 207a587d139SMark LandauIdx (*gIdx)[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ]; // #elems * LANDAU_MAX_NQ (spoof for max , Nb) on device, 208a587d139SMark LandauIdx num_elements; 209a587d139SMark LandauIdx num_reduced; 210a587d139SMark LandauIdx num_face; // (Q or Q^2 for 3D) 211a587d139SMark LandauDeviceType deviceType; 212a587d139SMark PetscInt Nf; 213a587d139SMark pointInterpolationP4est (*c_maps)[LANDAU_MAX_Q_FACE]; 2148a6f2e61SMark Adams struct _lP4estVertexMaps *d_self; 215a587d139SMark void *vp1, *vp2, *vp3; 2168a6f2e61SMark Adams PetscInt numgrids; 217a587d139SMark } P4estVertexMaps; 218a587d139SMark 219e0eea495SMark PETSC_EXTERN PetscErrorCode LandauCreateColoring(Mat, DM, PetscContainer *); 220e0eea495SMark #if defined(PETSC_HAVE_CUDA) 2219371c9d4SSatish Balay PETSC_EXTERN PetscErrorCode LandauCUDAJacobian(DM[], const PetscInt, const PetscInt, const PetscInt, const PetscInt[], PetscReal[], PetscScalar[], const PetscScalar[], const LandauStaticData *, const PetscReal, const PetscLogEvent[], const PetscInt[], const PetscInt[], Mat[], Mat); 2228a6f2e61SMark Adams PETSC_EXTERN PetscErrorCode LandauCUDACreateMatMaps(P4estVertexMaps *, pointInterpolationP4est (*)[LANDAU_MAX_Q_FACE], PetscInt[], PetscInt, PetscInt); 2238a6f2e61SMark Adams PETSC_EXTERN PetscErrorCode LandauCUDADestroyMatMaps(P4estVertexMaps *, PetscInt); 224ae6bf4a8Smarkadams4 PETSC_EXTERN PetscErrorCode LandauCUDAStaticDataSet(DM, const PetscInt, const PetscInt, const PetscInt, PetscInt[], PetscInt[], PetscInt[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], LandauStaticData *); 2258a6f2e61SMark Adams PETSC_EXTERN PetscErrorCode LandauCUDAStaticDataClear(LandauStaticData *); 226e0eea495SMark #endif 227e0eea495SMark #if defined(PETSC_HAVE_KOKKOS) 2289371c9d4SSatish Balay PETSC_EXTERN PetscErrorCode LandauKokkosJacobian(DM[], const PetscInt, const PetscInt, const PetscInt, const PetscInt[], PetscReal[], PetscScalar[], const PetscScalar[], const LandauStaticData *, const PetscReal, const PetscLogEvent[], const PetscInt[], const PetscInt[], Mat[], Mat); 2298a6f2e61SMark Adams PETSC_EXTERN PetscErrorCode LandauKokkosCreateMatMaps(P4estVertexMaps *, pointInterpolationP4est (*)[LANDAU_MAX_Q_FACE], PetscInt[], PetscInt, PetscInt); 2308a6f2e61SMark Adams PETSC_EXTERN PetscErrorCode LandauKokkosDestroyMatMaps(P4estVertexMaps *, PetscInt); 231ae6bf4a8Smarkadams4 PETSC_EXTERN PetscErrorCode LandauKokkosStaticDataSet(DM, const PetscInt, const PetscInt, const PetscInt, PetscInt[], PetscInt[], PetscInt[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], LandauStaticData *); 2328a6f2e61SMark Adams PETSC_EXTERN PetscErrorCode LandauKokkosStaticDataClear(LandauStaticData *); 233e0eea495SMark #endif 234e0eea495SMark 235e0eea495SMark #endif /* PETSCLANDAU_H */ 236