xref: /petsc/include/petsclandau.h (revision f985a14370d3736da5ba205dce0ca3caa3955d88)
1a4963045SJacob Faibussowitsch #pragma once
2e0eea495SMark 
3e0eea495SMark #include <petscdmplex.h> /*I      "petscdmplex.h"    I*/
4e0eea495SMark #include <petscts.h>
5e0eea495SMark 
6ce78bad3SBarry Smith /* MANSEC = TS */
7ce78bad3SBarry Smith /* SUBMANSEC = LANDAU */
8ce78bad3SBarry Smith 
98594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauPrintNorms(Vec, PetscInt);
108594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauCreateVelocitySpace(MPI_Comm, PetscInt, const char[], Vec *, Mat *, DM *);
118594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauDestroyVelocitySpace(DM *);
12f53e7263SMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauAccess(DM, Vec, PetscErrorCode (*)(DM, Vec, PetscInt, PetscInt, PetscInt, void *), void *);
13c3e4dd79SMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauAddMaxwellians(DM, Vec, PetscReal, PetscReal[], PetscReal[], PetscInt, PetscInt, PetscInt, void *);
148594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauCreateMassMatrix(DM dm, Mat *Amat);
158594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauIFunction(TS, PetscReal, Vec, Vec, Vec, void *);
168594ddcfSMark Adams PETSC_EXTERN PetscErrorCode DMPlexLandauIJacobian(TS, PetscReal, Vec, Vec, PetscReal, Mat, Mat, void *);
17e0eea495SMark 
186497c311SBarry Smith typedef PetscInt LandauIdx;
19bfc784b7SMark Adams 
20e0eea495SMark /* the Fokker-Planck-Landau context */
21e0eea495SMark #if !defined(LANDAU_MAX_SPECIES)
22984ed092SMark Adams   #if defined(PETSC_USE_DMLANDAU_2D)
23e0eea495SMark     #define LANDAU_MAX_SPECIES 10
248a6f2e61SMark Adams     #define LANDAU_MAX_GRIDS   3
2552cdd6eaSMark   #else
26763ae2f8SMark Adams     #define LANDAU_MAX_SPECIES 10
27c3e4dd79SMark Adams     #define LANDAU_MAX_GRIDS   3
2852cdd6eaSMark   #endif
298a6f2e61SMark Adams #else
308a6f2e61SMark Adams   #define LANDAU_MAX_GRIDS 3
31e0eea495SMark #endif
32a587d139SMark 
33a587d139SMark #if !defined(LANDAU_MAX_Q)
34f44ea33cSmarkadams4   #if defined(LANDAU_MAX_NQND)
35f44ea33cSmarkadams4     #error "LANDAU_MAX_NQND but not LANDAU_MAX_Q. Use -DLANDAU_MAX_Q=4 for Q3 elements"
36a587d139SMark   #endif
37984ed092SMark Adams   #if defined(PETSC_USE_DMLANDAU_2D)
386b664d00Smarkadams4     #define LANDAU_MAX_Q 6
39c3e4dd79SMark Adams   #else
40a8b63cc6SMark Adams     #define LANDAU_MAX_Q 6
41a587d139SMark   #endif
42a587d139SMark #else
43f44ea33cSmarkadams4   #undef LANDAU_MAX_NQND
44e0eea495SMark #endif
45a587d139SMark 
46984ed092SMark Adams #if defined(PETSC_USE_DMLANDAU_2D)
47930e68a5SMark Adams   #define LANDAU_MAX_Q_FACE   LANDAU_MAX_Q
48f44ea33cSmarkadams4   #define LANDAU_MAX_NQND     (LANDAU_MAX_Q * LANDAU_MAX_Q)
493b65d667Smarkadams4   #define LANDAU_MAX_BATCH_SZ 1024
50984ed092SMark Adams   #define LANDAU_DIM          2
51a587d139SMark #else
52930e68a5SMark Adams   #define LANDAU_MAX_Q_FACE   (LANDAU_MAX_Q * LANDAU_MAX_Q)
53f44ea33cSmarkadams4   #define LANDAU_MAX_NQND     (LANDAU_MAX_Q * LANDAU_MAX_Q * LANDAU_MAX_Q)
543b65d667Smarkadams4   #define LANDAU_MAX_BATCH_SZ 64
55984ed092SMark Adams   #define LANDAU_DIM          3
56a587d139SMark #endif
57a587d139SMark 
589371c9d4SSatish Balay typedef enum {
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;
130*e1d0a54aSMark Adams   PetscBool map_sphere;
131d043ef4cSMark Adams   PetscReal sphere_inner_radius_90degree[LANDAU_MAX_GRIDS];
132d043ef4cSMark Adams   PetscReal sphere_inner_radius_45degree[LANDAU_MAX_GRIDS];
133f53e7263SMark Adams   PetscInt  cells0[3];
1348a6f2e61SMark Adams   /* AMR */
1358a6f2e61SMark Adams   PetscBool use_p4est;
1368a6f2e61SMark Adams   PetscInt  numRERefine;                     /* RE: refinement along v_perp=0, z > 0 */
1378a6f2e61SMark Adams   PetscInt  nZRefine1;                       /* RE: origin refinement after v_perp=0 refinement */
1388a6f2e61SMark Adams   PetscInt  nZRefine2;                       /* RE: origin refinement after origin AMR refinement */
1398a6f2e61SMark Adams   PetscInt  numAMRRefine[LANDAU_MAX_GRIDS];  /* normal AMR - refine from origin */
1408a6f2e61SMark Adams   PetscInt  postAMRRefine[LANDAU_MAX_GRIDS]; /* uniform refinement of AMR */
141cd27c6deSmarkadams4   PetscBool simplex;
142f44ea33cSmarkadams4   char      filename[PETSC_MAX_PATH_LEN];
143f44ea33cSmarkadams4   PetscReal thermal_speed[LANDAU_MAX_GRIDS];
144d043ef4cSMark Adams   PetscBool sphere_uniform_normal;
1458a6f2e61SMark Adams   /* relativistic */
146cefb98e8SMark Adams   PetscBool use_energy_tensor_trick;
147cefb98e8SMark Adams   PetscBool use_relativistic_corrections;
148e0eea495SMark   /* physics */
149e0eea495SMark   PetscReal thermal_temps[LANDAU_MAX_SPECIES];
150e0eea495SMark   PetscReal masses[LANDAU_MAX_SPECIES];  /* mass of each species  */
151e0eea495SMark   PetscReal charges[LANDAU_MAX_SPECIES]; /* charge of each species  */
152e0eea495SMark   PetscReal n[LANDAU_MAX_SPECIES];       /* number density of each species  */
153e0eea495SMark   PetscReal m_0;                         /* reference mass */
154e0eea495SMark   PetscReal v_0;                         /* reference velocity */
155e0eea495SMark   PetscReal n_0;                         /* reference number density */
156e0eea495SMark   PetscReal t_0;                         /* reference time */
157e0eea495SMark   PetscReal Ez;
158e0eea495SMark   PetscReal epsilon0;
159e0eea495SMark   PetscReal k;
160ae6bf4a8Smarkadams4   PetscReal lambdas[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS];
1618a6f2e61SMark Adams   PetscReal electronShift;
162e0eea495SMark   PetscInt  num_species;
1638a6f2e61SMark Adams   PetscInt  num_grids;
1648fdabdddSMark Adams   PetscInt  species_offset[LANDAU_MAX_GRIDS + 1]; // for each grid, but same for all batched vertices
1658fdabdddSMark Adams   PetscInt  mat_offset[LANDAU_MAX_GRIDS + 1];     // for each grid, but same for all batched vertices
166cb25d741SMark Adams   // batching
167cb25d741SMark Adams   PetscBool  jacobian_field_major_order; // this could be a type but lets not get pedantic
168cb25d741SMark Adams   VecScatter plex_batch;
169cb25d741SMark Adams   Vec        work_vec;
170cb25d741SMark Adams   IS         batch_is;
171cb25d741SMark Adams   PetscErrorCode (*seqaij_mult)(Mat, Vec, Vec);
172cb25d741SMark Adams   PetscErrorCode (*seqaij_multtranspose)(Mat, Vec, Vec);
173cb25d741SMark Adams   PetscErrorCode (*seqaij_solve)(Mat, Vec, Vec);
174cb25d741SMark Adams   PetscErrorCode (*seqaij_getdiagonal)(Mat, Vec);
175bfc784b7SMark Adams   /* COO */
176e0eea495SMark   Mat J;
177e0eea495SMark   Mat M;
178e0eea495SMark   Vec X;
179e0eea495SMark   /* derived type */
180e0eea495SMark   void *data;
181e0eea495SMark   /* computing */
182e0eea495SMark   LandauDeviceType deviceType;
1838a6f2e61SMark Adams   DM               pack;
1848a6f2e61SMark Adams   DM               plex[LANDAU_MAX_GRIDS];
1858fdabdddSMark Adams   LandauStaticData SData_d; /* static geometric data on device */
18654545eeeSMark Adams   /* diagnostics */
18754545eeeSMark Adams   PetscInt         verbose;
18854545eeeSMark Adams   PetscLogEvent    events[20];
189c751c0a2SMark Adams   PetscLogStage    stage;
190984ed092SMark Adams   PetscObjectState norm_state;
1918fdabdddSMark Adams   PetscInt         batch_sz;
1928fdabdddSMark Adams   PetscInt         batch_view_idx;
193e0eea495SMark } LandauCtx;
194e0eea495SMark 
1958fdabdddSMark Adams #define LANDAU_SPECIES_MAJOR
1968fdabdddSMark Adams #if !defined(LANDAU_SPECIES_MAJOR)
1978fdabdddSMark Adams   #define LAND_PACK_IDX(_b, _g)                         (_b * ctx->num_grids + _g)
1988fdabdddSMark Adams   #define LAND_MOFFSET(_b, _g, _nbch, _ngrid, _mat_off) (_b * _mat_off[_ngrid] + _mat_off[_g])
1998fdabdddSMark Adams #else
2008fdabdddSMark Adams   #define LAND_PACK_IDX(_b, _g)                         (_g * ctx->batch_sz + _b)
2018fdabdddSMark Adams   #define LAND_MOFFSET(_b, _g, _nbch, _ngrid, _mat_off) (_nbch * _mat_off[_g] + _b * (_mat_off[_g + 1] - _mat_off[_g]))
2028fdabdddSMark Adams #endif
2038fdabdddSMark Adams 
204a587d139SMark typedef struct {
205a587d139SMark   PetscReal scale;
206bfc784b7SMark Adams   LandauIdx gid; // Landau matrix index (<10,000)
207a587d139SMark } pointInterpolationP4est;
208a587d139SMark typedef struct _lP4estVertexMaps {
209f44ea33cSmarkadams4   LandauIdx (*gIdx)[LANDAU_MAX_SPECIES][LANDAU_MAX_NQND]; // #elems *  LANDAU_MAX_NQND
210a587d139SMark   LandauIdx        num_elements;
211a587d139SMark   LandauIdx        num_reduced;
212a587d139SMark   LandauIdx        num_face; // (Q or Q^2 for 3D)
213a587d139SMark   LandauDeviceType deviceType;
214a587d139SMark   PetscInt         Nf;
215a587d139SMark   pointInterpolationP4est (*c_maps)[LANDAU_MAX_Q_FACE];
2168a6f2e61SMark Adams   struct _lP4estVertexMaps *d_self;
217a587d139SMark   void                     *vp1, *vp2, *vp3;
2188a6f2e61SMark Adams   PetscInt                  numgrids;
219a587d139SMark } P4estVertexMaps;
220a587d139SMark 
221e0eea495SMark #if defined(PETSC_HAVE_KOKKOS)
222f44ea33cSmarkadams4 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);
223f44ea33cSmarkadams4 PETSC_EXTERN PetscErrorCode LandauKokkosCreateMatMaps(P4estVertexMaps *, pointInterpolationP4est (*)[LANDAU_MAX_Q_FACE], PetscInt[], PetscInt);
2248a6f2e61SMark Adams PETSC_EXTERN PetscErrorCode LandauKokkosDestroyMatMaps(P4estVertexMaps *, PetscInt);
225f44ea33cSmarkadams4 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 *);
2268a6f2e61SMark Adams PETSC_EXTERN PetscErrorCode LandauKokkosStaticDataClear(LandauStaticData *);
227e0eea495SMark #endif
228