1 #include <../src/mat/impls/aij/seq/aij.h> 2 #include <petsc/private/dmpleximpl.h> /*I "petscdmplex.h" I*/ 3 #include <petsclandau.h> /*I "petsclandau.h" I*/ 4 #include <petscts.h> 5 #include <petscdmforest.h> 6 #include <petscdmcomposite.h> 7 8 /* Landau collision operator */ 9 10 /* relativistic terms */ 11 #if defined(PETSC_USE_REAL_SINGLE) 12 #define SPEED_OF_LIGHT 2.99792458e8F 13 #define C_0(v0) (SPEED_OF_LIGHT/v0) /* needed for relativistic tensor on all architectures */ 14 #else 15 #define SPEED_OF_LIGHT 2.99792458e8 16 #define C_0(v0) (SPEED_OF_LIGHT/v0) /* needed for relativistic tensor on all architectures */ 17 #endif 18 19 #define PETSC_THREAD_SYNC 20 #include "land_tensors.h" 21 22 #if defined(PETSC_HAVE_OPENMP) 23 #include <omp.h> 24 #endif 25 26 /* vector padding not supported */ 27 #define LANDAU_VL 1 28 29 static PetscErrorCode LandauMatMult(Mat A, Vec x, Vec y) 30 { 31 PetscErrorCode ierr; 32 LandauCtx *ctx; 33 PetscContainer container; 34 35 PetscFunctionBegin; 36 ierr = PetscObjectQuery((PetscObject) A, "LandauCtx", (PetscObject *) &container);CHKERRQ(ierr); 37 if (container) { 38 ierr = PetscContainerGetPointer(container, (void **) &ctx);CHKERRQ(ierr); 39 ierr = VecScatterBegin(ctx->plex_batch,x,ctx->work_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 40 ierr = VecScatterEnd(ctx->plex_batch,x,ctx->work_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 41 ierr = (*ctx->seqaij_mult)(A,ctx->work_vec,y);CHKERRQ(ierr); 42 ierr = VecCopy(y, ctx->work_vec);CHKERRQ(ierr); 43 ierr = VecScatterBegin(ctx->plex_batch,ctx->work_vec,y,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 44 ierr = VecScatterEnd(ctx->plex_batch,ctx->work_vec,y,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 45 PetscFunctionReturn(0); 46 } 47 ierr = MatMult(A,x,y);CHKERRQ(ierr); 48 PetscFunctionReturn(0); 49 } 50 51 // Computes v3 = v2 + A * v1. 52 static PetscErrorCode LandauMatMultAdd(Mat A,Vec v1,Vec v2,Vec v3) 53 { 54 PetscErrorCode ierr; 55 56 PetscFunctionBegin; 57 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "?????"); 58 ierr = LandauMatMult(A,v1,v3);CHKERRQ(ierr); 59 ierr = VecAYPX(v3,1,v2);CHKERRQ(ierr); 60 PetscFunctionReturn(0); 61 } 62 63 static PetscErrorCode LandauMatMultTranspose(Mat A, Vec x, Vec y) 64 { 65 PetscErrorCode ierr; 66 LandauCtx *ctx; 67 PetscContainer container; 68 69 PetscFunctionBegin; 70 ierr = PetscObjectQuery((PetscObject) A, "LandauCtx", (PetscObject *) &container);CHKERRQ(ierr); 71 if (container) { 72 ierr = PetscContainerGetPointer(container, (void **) &ctx);CHKERRQ(ierr); 73 ierr = VecScatterBegin(ctx->plex_batch,x,ctx->work_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 74 ierr = VecScatterEnd(ctx->plex_batch,x,ctx->work_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 75 ierr = (*ctx->seqaij_multtranspose)(A,ctx->work_vec,y);CHKERRQ(ierr); 76 ierr = VecCopy(y, ctx->work_vec);CHKERRQ(ierr); 77 ierr = VecScatterBegin(ctx->plex_batch,ctx->work_vec,y,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 78 ierr = VecScatterEnd(ctx->plex_batch,ctx->work_vec,y,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 79 PetscFunctionReturn(0); 80 } 81 ierr = MatMultTranspose(A,x,y);CHKERRQ(ierr); 82 PetscFunctionReturn(0); 83 } 84 85 static PetscErrorCode LandauMatGetDiagonal(Mat A,Vec x) 86 { 87 PetscErrorCode ierr; 88 LandauCtx *ctx; 89 PetscContainer container; 90 91 PetscFunctionBegin; 92 ierr = PetscObjectQuery((PetscObject) A, "LandauCtx", (PetscObject *) &container);CHKERRQ(ierr); 93 if (container) { 94 ierr = PetscContainerGetPointer(container, (void **) &ctx);CHKERRQ(ierr); 95 ierr = (*ctx->seqaij_getdiagonal)(A,ctx->work_vec);CHKERRQ(ierr); 96 ierr = VecScatterBegin(ctx->plex_batch,ctx->work_vec,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 97 ierr = VecScatterEnd(ctx->plex_batch,ctx->work_vec,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 98 PetscFunctionReturn(0); 99 } 100 ierr = MatGetDiagonal(A, x);CHKERRQ(ierr); 101 PetscFunctionReturn(0); 102 } 103 104 static PetscErrorCode LandauGPUMapsDestroy(void *ptr) 105 { 106 P4estVertexMaps *maps = (P4estVertexMaps*)ptr; 107 PetscErrorCode ierr; 108 PetscFunctionBegin; 109 // free device data 110 if (maps[0].deviceType != LANDAU_CPU) { 111 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 112 if (maps[0].deviceType == LANDAU_KOKKOS) { 113 ierr = LandauKokkosDestroyMatMaps(maps, maps[0].numgrids);CHKERRQ(ierr); // imples Kokkos does 114 } // else could be CUDA 115 #elif defined(PETSC_HAVE_CUDA) 116 if (maps[0].deviceType == LANDAU_CUDA) { 117 ierr = LandauCUDADestroyMatMaps(maps, maps[0].numgrids);CHKERRQ(ierr); 118 } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "maps->deviceType %" PetscInt_FMT " ?????",maps->deviceType); 119 #endif 120 } 121 // free host data 122 for (PetscInt grid=0 ; grid < maps[0].numgrids ; grid++) { 123 ierr = PetscFree(maps[grid].c_maps);CHKERRQ(ierr); 124 ierr = PetscFree(maps[grid].gIdx);CHKERRQ(ierr); 125 } 126 ierr = PetscFree(maps);CHKERRQ(ierr); 127 128 PetscFunctionReturn(0); 129 } 130 static PetscErrorCode energy_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx) 131 { 132 PetscReal v2 = 0; 133 PetscFunctionBegin; 134 /* compute v^2 / 2 */ 135 for (int i = 0; i < dim; ++i) v2 += x[i]*x[i]; 136 /* evaluate the Maxwellian */ 137 u[0] = v2/2; 138 PetscFunctionReturn(0); 139 } 140 141 /* needs double */ 142 static PetscErrorCode gamma_m1_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx) 143 { 144 PetscReal *c2_0_arr = ((PetscReal*)actx); 145 double u2 = 0, c02 = (double)*c2_0_arr, xx; 146 147 PetscFunctionBegin; 148 /* compute u^2 / 2 */ 149 for (int i = 0; i < dim; ++i) u2 += x[i]*x[i]; 150 /* gamma - 1 = g_eps, for conditioning and we only take derivatives */ 151 xx = u2/c02; 152 #if defined(PETSC_USE_DEBUG) 153 u[0] = PetscSqrtReal(1. + xx); 154 #else 155 u[0] = xx/(PetscSqrtReal(1. + xx) + 1.) - 1.; // better conditioned. -1 might help condition and only used for derivative 156 #endif 157 PetscFunctionReturn(0); 158 } 159 160 /* 161 LandauFormJacobian_Internal - Evaluates Jacobian matrix. 162 163 Input Parameters: 164 . globX - input vector 165 . actx - optional user-defined context 166 . dim - dimension 167 168 Output Parameters: 169 . J0acP - Jacobian matrix filled, not created 170 */ 171 static PetscErrorCode LandauFormJacobian_Internal(Vec a_X, Mat JacP, const PetscInt dim, PetscReal shift, void *a_ctx) 172 { 173 LandauCtx *ctx = (LandauCtx*)a_ctx; 174 PetscErrorCode ierr; 175 PetscInt numCells[LANDAU_MAX_GRIDS],Nq,Nb; 176 PetscQuadrature quad; 177 PetscReal Eq_m[LANDAU_MAX_SPECIES]; // could be static data w/o quench (ex2) 178 PetscScalar *cellClosure=NULL; 179 const PetscScalar *xdata=NULL; 180 PetscDS prob; 181 PetscContainer container; 182 P4estVertexMaps *maps; 183 Mat subJ[LANDAU_MAX_GRIDS*LANDAU_MAX_BATCH_SZ]; 184 185 PetscFunctionBegin; 186 PetscValidHeaderSpecific(a_X,VEC_CLASSID,1); 187 PetscValidHeaderSpecific(JacP,MAT_CLASSID,2); 188 PetscValidPointer(ctx,5); 189 /* check for matrix container for GPU assembly. Support CPU assembly for debugging */ 190 PetscCheckFalse(ctx->plex[0] == NULL,ctx->comm,PETSC_ERR_ARG_WRONG,"Plex not created"); 191 ierr = PetscLogEventBegin(ctx->events[10],0,0,0,0);CHKERRQ(ierr); 192 ierr = DMGetDS(ctx->plex[0], &prob);CHKERRQ(ierr); // same DS for all grids 193 ierr = PetscObjectQuery((PetscObject) JacP, "assembly_maps", (PetscObject *) &container);CHKERRQ(ierr); 194 if (container) { 195 PetscCheck(ctx->gpu_assembly,ctx->comm,PETSC_ERR_ARG_WRONG,"maps but no GPU assembly"); 196 ierr = PetscContainerGetPointer(container, (void **) &maps);CHKERRQ(ierr); 197 PetscCheckFalse(!maps,ctx->comm,PETSC_ERR_ARG_WRONG,"empty GPU matrix container"); 198 for (PetscInt i=0;i<ctx->num_grids*ctx->batch_sz;i++) subJ[i] = NULL; 199 } else { 200 PetscCheck(!ctx->gpu_assembly,ctx->comm,PETSC_ERR_ARG_WRONG,"No maps but GPU assembly"); 201 for (PetscInt tid=0 ; tid<ctx->batch_sz ; tid++) { 202 for (PetscInt grid=0;grid<ctx->num_grids;grid++) { 203 ierr = DMCreateMatrix(ctx->plex[grid], &subJ[ LAND_PACK_IDX(tid,grid) ]);CHKERRQ(ierr); 204 } 205 } 206 maps = NULL; 207 } 208 // get dynamic data (Eq is odd, for quench and Spitzer test) for CPU assembly and raw data for Jacobian GPU assembly. Get host numCells[], Nq (yuck) 209 ierr = PetscFEGetQuadrature(ctx->fe[0], &quad);CHKERRQ(ierr); 210 ierr = PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, NULL);CHKERRQ(ierr); Nb = Nq; 211 PetscCheckFalse(Nq >LANDAU_MAX_NQ,ctx->comm,PETSC_ERR_ARG_WRONG,"Order too high. Nq = %" PetscInt_FMT " > LANDAU_MAX_NQ (%" PetscInt_FMT ")",Nq,LANDAU_MAX_NQ); 212 // get metadata for collecting dynamic data 213 for (PetscInt grid=0;grid<ctx->num_grids;grid++) { 214 PetscInt cStart, cEnd; 215 PetscCheckFalse(ctx->plex[grid] == NULL,ctx->comm,PETSC_ERR_ARG_WRONG,"Plex not created"); 216 ierr = DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd);CHKERRQ(ierr); 217 numCells[grid] = cEnd - cStart; // grids can have different topology 218 } 219 ierr = PetscLogEventEnd(ctx->events[10],0,0,0,0);CHKERRQ(ierr); 220 if (shift==0) { /* create dynamic point data: f_alpha for closure of each cell (cellClosure[nbatch,ngrids,ncells[g],f[Nb,ns[g]]]) or xdata */ 221 DM pack; 222 ierr = VecGetDM(a_X, &pack);CHKERRQ(ierr); 223 PetscCheckFalse(!pack,PETSC_COMM_SELF, PETSC_ERR_PLIB, "pack has no DM"); 224 ierr = PetscLogEventBegin(ctx->events[1],0,0,0,0);CHKERRQ(ierr); 225 ierr = MatZeroEntries(JacP);CHKERRQ(ierr); 226 for (PetscInt fieldA=0;fieldA<ctx->num_species;fieldA++) { 227 Eq_m[fieldA] = ctx->Ez * ctx->t_0 * ctx->charges[fieldA] / (ctx->v_0 * ctx->masses[fieldA]); /* normalize dimensionless */ 228 if (dim==2) Eq_m[fieldA] *= 2 * PETSC_PI; /* add the 2pi term that is not in Landau */ 229 } 230 if (!ctx->gpu_assembly) { 231 Vec *locXArray,*globXArray; 232 PetscScalar *cellClosure_it; 233 PetscInt cellClosure_sz=0,nDMs,Nf[LANDAU_MAX_GRIDS]; 234 PetscSection section[LANDAU_MAX_GRIDS],globsection[LANDAU_MAX_GRIDS]; 235 for (PetscInt grid=0;grid<ctx->num_grids;grid++) { 236 ierr = DMGetLocalSection(ctx->plex[grid], §ion[grid]);CHKERRQ(ierr); 237 ierr = DMGetGlobalSection(ctx->plex[grid], &globsection[grid]);CHKERRQ(ierr); 238 ierr = PetscSectionGetNumFields(section[grid], &Nf[grid]);CHKERRQ(ierr); 239 } 240 /* count cellClosure size */ 241 ierr = DMCompositeGetNumberDM(pack,&nDMs);CHKERRQ(ierr); 242 for (PetscInt grid=0 ; grid<ctx->num_grids ; grid++) cellClosure_sz += Nb*Nf[grid]*numCells[grid]; 243 ierr = PetscMalloc1(cellClosure_sz*ctx->batch_sz,&cellClosure);CHKERRQ(ierr); 244 cellClosure_it = cellClosure; 245 ierr = PetscMalloc(sizeof(*locXArray)*nDMs, &locXArray);CHKERRQ(ierr); 246 ierr = PetscMalloc(sizeof(*globXArray)*nDMs, &globXArray);CHKERRQ(ierr); 247 ierr = DMCompositeGetLocalAccessArray(pack, a_X, nDMs, NULL, locXArray);CHKERRQ(ierr); 248 ierr = DMCompositeGetAccessArray(pack, a_X, nDMs, NULL, globXArray);CHKERRQ(ierr); 249 for (PetscInt b_id = 0 ; b_id < ctx->batch_sz ; b_id++) { // OpenMP (once) 250 for (PetscInt grid=0 ; grid<ctx->num_grids ; grid++) { 251 Vec locX = locXArray[ LAND_PACK_IDX(b_id,grid) ], globX = globXArray[ LAND_PACK_IDX(b_id,grid) ], locX2; 252 PetscInt cStart, cEnd, ei; 253 ierr = VecDuplicate(locX,&locX2);CHKERRQ(ierr); 254 ierr = DMGlobalToLocalBegin(ctx->plex[grid], globX, INSERT_VALUES, locX2);CHKERRQ(ierr); 255 ierr = DMGlobalToLocalEnd (ctx->plex[grid], globX, INSERT_VALUES, locX2);CHKERRQ(ierr); 256 ierr = DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd);CHKERRQ(ierr); 257 for (ei = cStart ; ei < cEnd; ++ei) { 258 PetscScalar *coef = NULL; 259 ierr = DMPlexVecGetClosure(ctx->plex[grid], section[grid], locX2, ei, NULL, &coef);CHKERRQ(ierr); 260 ierr = PetscMemcpy(cellClosure_it,coef,Nb*Nf[grid]*sizeof(*cellClosure_it));CHKERRQ(ierr); /* change if LandauIPReal != PetscScalar */ 261 ierr = DMPlexVecRestoreClosure(ctx->plex[grid], section[grid], locX2, ei, NULL, &coef);CHKERRQ(ierr); 262 cellClosure_it += Nb*Nf[grid]; 263 } 264 ierr = VecDestroy(&locX2);CHKERRQ(ierr); 265 } 266 } 267 PetscCheck(cellClosure_it-cellClosure == cellClosure_sz*ctx->batch_sz,PETSC_COMM_SELF, PETSC_ERR_PLIB, "iteration wrong %" PetscInt_FMT " != cellClosure_sz = %" PetscInt_FMT,cellClosure_it-cellClosure,cellClosure_sz*ctx->batch_sz); 268 ierr = DMCompositeRestoreLocalAccessArray(pack, a_X, nDMs, NULL, locXArray);CHKERRQ(ierr); 269 ierr = DMCompositeRestoreAccessArray(pack, a_X, nDMs, NULL, globXArray);CHKERRQ(ierr); 270 ierr = PetscFree(locXArray);CHKERRQ(ierr); 271 ierr = PetscFree(globXArray);CHKERRQ(ierr); 272 xdata = NULL; 273 } else { 274 PetscMemType mtype; 275 if (ctx->jacobian_field_major_order) { // get data in batch ordering 276 ierr = VecScatterBegin(ctx->plex_batch,a_X,ctx->work_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 277 ierr = VecScatterEnd(ctx->plex_batch,a_X,ctx->work_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 278 ierr = VecGetArrayReadAndMemType(ctx->work_vec,&xdata,&mtype);CHKERRQ(ierr); 279 } else { 280 ierr = VecGetArrayReadAndMemType(a_X,&xdata,&mtype);CHKERRQ(ierr); 281 } 282 if (mtype!=PETSC_MEMTYPE_HOST && ctx->deviceType == LANDAU_CPU) { 283 SETERRQ(ctx->comm,PETSC_ERR_ARG_WRONG,"CPU run with device data: use -mat_type aij"); 284 } 285 cellClosure = NULL; 286 } 287 ierr = PetscLogEventEnd(ctx->events[1],0,0,0,0);CHKERRQ(ierr); 288 } else xdata = cellClosure = NULL; 289 290 /* do it */ 291 if (ctx->deviceType == LANDAU_CUDA || ctx->deviceType == LANDAU_KOKKOS) { 292 if (ctx->deviceType == LANDAU_CUDA) { 293 #if defined(PETSC_HAVE_CUDA) 294 ierr = LandauCUDAJacobian(ctx->plex,Nq,ctx->batch_sz,ctx->num_grids,numCells,Eq_m,cellClosure,xdata,&ctx->SData_d,shift,ctx->events,ctx->mat_offset, ctx->species_offset, subJ, JacP);CHKERRQ(ierr); 295 #else 296 SETERRQ(ctx->comm,PETSC_ERR_ARG_WRONG,"-landau_device_type %s not built","cuda"); 297 #endif 298 } else if (ctx->deviceType == LANDAU_KOKKOS) { 299 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 300 ierr = LandauKokkosJacobian(ctx->plex,Nq,ctx->batch_sz,ctx->num_grids,numCells,Eq_m,cellClosure,xdata,&ctx->SData_d,shift,ctx->events,ctx->mat_offset, ctx->species_offset, subJ,JacP);CHKERRQ(ierr); 301 #else 302 SETERRQ(ctx->comm,PETSC_ERR_ARG_WRONG,"-landau_device_type %s not built","kokkos"); 303 #endif 304 } 305 } else { /* CPU version */ 306 PetscTabulation *Tf; // used for CPU and print info. Same on all grids and all species 307 PetscInt ip_offset[LANDAU_MAX_GRIDS+1], ipf_offset[LANDAU_MAX_GRIDS+1], elem_offset[LANDAU_MAX_GRIDS+1],IPf_sz_glb,IPf_sz_tot,num_grids=ctx->num_grids,Nf[LANDAU_MAX_GRIDS]; 308 PetscReal *ff, *dudx, *dudy, *dudz, *invJ_a = (PetscReal*)ctx->SData_d.invJ, *xx = (PetscReal*)ctx->SData_d.x, *yy = (PetscReal*)ctx->SData_d.y, *zz = (PetscReal*)ctx->SData_d.z, *ww = (PetscReal*)ctx->SData_d.w; 309 PetscReal Eq_m[LANDAU_MAX_SPECIES], invMass[LANDAU_MAX_SPECIES], nu_alpha[LANDAU_MAX_SPECIES], nu_beta[LANDAU_MAX_SPECIES]; 310 PetscSection section[LANDAU_MAX_GRIDS],globsection[LANDAU_MAX_GRIDS]; 311 PetscScalar *coo_vals=NULL; 312 for (PetscInt grid=0;grid<ctx->num_grids;grid++) { 313 ierr = DMGetLocalSection(ctx->plex[grid], §ion[grid]);CHKERRQ(ierr); 314 ierr = DMGetGlobalSection(ctx->plex[grid], &globsection[grid]);CHKERRQ(ierr); 315 ierr = PetscSectionGetNumFields(section[grid], &Nf[grid]);CHKERRQ(ierr); 316 } 317 /* count IPf size, etc */ 318 ierr = PetscDSGetTabulation(prob, &Tf);CHKERRQ(ierr); // Bf, &Df same for all grids 319 const PetscReal *const BB = Tf[0]->T[0], * const DD = Tf[0]->T[1]; 320 ip_offset[0] = ipf_offset[0] = elem_offset[0] = 0; 321 for (PetscInt grid=0 ; grid<num_grids ; grid++) { 322 PetscInt nfloc = ctx->species_offset[grid+1] - ctx->species_offset[grid]; 323 elem_offset[grid+1] = elem_offset[grid] + numCells[grid]; 324 ip_offset[grid+1] = ip_offset[grid] + numCells[grid]*Nq; 325 ipf_offset[grid+1] = ipf_offset[grid] + Nq*nfloc*numCells[grid]; 326 } 327 IPf_sz_glb = ipf_offset[num_grids]; 328 IPf_sz_tot = IPf_sz_glb*ctx->batch_sz; 329 // prep COO 330 if (ctx->use_coo_assembly) { 331 ierr = PetscMalloc1(ctx->SData_d.coo_size,&coo_vals);CHKERRQ(ierr); // allocate every time? 332 ierr = PetscInfo(ctx->plex[0], "COO Allocate %" PetscInt_FMT " values\n",ctx->SData_d.coo_size);CHKERRQ(ierr); 333 } 334 if (shift==0.0) { /* compute dynamic data f and df and init data for Jacobian */ 335 #if defined(PETSC_HAVE_THREADSAFETY) 336 double starttime, endtime; 337 starttime = MPI_Wtime(); 338 #endif 339 ierr = PetscLogEventBegin(ctx->events[8],0,0,0,0);CHKERRQ(ierr); 340 for (PetscInt fieldA=0;fieldA<ctx->num_species;fieldA++) { 341 invMass[fieldA] = ctx->m_0/ctx->masses[fieldA]; 342 Eq_m[fieldA] = ctx->Ez * ctx->t_0 * ctx->charges[fieldA] / (ctx->v_0 * ctx->masses[fieldA]); /* normalize dimensionless */ 343 if (dim==2) Eq_m[fieldA] *= 2 * PETSC_PI; /* add the 2pi term that is not in Landau */ 344 nu_alpha[fieldA] = PetscSqr(ctx->charges[fieldA]/ctx->m_0)*ctx->m_0/ctx->masses[fieldA]; 345 nu_beta[fieldA] = PetscSqr(ctx->charges[fieldA]/ctx->epsilon0)*ctx->lnLam / (8*PETSC_PI) * ctx->t_0*ctx->n_0/PetscPowReal(ctx->v_0,3); 346 } 347 ierr = PetscMalloc4(IPf_sz_tot, &ff, IPf_sz_tot, &dudx, IPf_sz_tot, &dudy, dim==3 ? IPf_sz_tot : 0, &dudz);CHKERRQ(ierr); 348 // F df/dx 349 for (PetscInt tid = 0 ; tid < ctx->batch_sz*elem_offset[num_grids] ; tid++) { // for each element 350 const PetscInt b_Nelem = elem_offset[num_grids], b_elem_idx = tid%b_Nelem, b_id = tid/b_Nelem; // b_id == OMP thd_id in batch 351 // find my grid: 352 PetscInt grid = 0; 353 while (b_elem_idx >= elem_offset[grid+1]) grid++; // yuck search for grid 354 { 355 const PetscInt loc_nip = numCells[grid]*Nq, loc_Nf = ctx->species_offset[grid+1] - ctx->species_offset[grid], loc_elem = b_elem_idx - elem_offset[grid]; 356 const PetscInt moffset = LAND_MOFFSET(b_id,grid,ctx->batch_sz,ctx->num_grids,ctx->mat_offset); //b_id*b_N + ctx->mat_offset[grid]; 357 PetscScalar *coef, coef_buff[LANDAU_MAX_SPECIES*LANDAU_MAX_NQ]; 358 PetscReal *invJe = &invJ_a[(ip_offset[grid] + loc_elem*Nq)*dim*dim]; // ingJ is static data on batch 0 359 PetscInt b,f,q; 360 if (cellClosure) { 361 coef = &cellClosure[b_id*IPf_sz_glb + ipf_offset[grid] + loc_elem*Nb*loc_Nf]; // this is const 362 } else { 363 coef = coef_buff; 364 for (f = 0; f < loc_Nf; ++f) { 365 LandauIdx *const Idxs = &maps[grid].gIdx[loc_elem][f][0]; 366 for (b = 0; b < Nb; ++b) { 367 PetscInt idx = Idxs[b]; 368 if (idx >= 0) { 369 coef[f*Nb+b] = xdata[idx+moffset]; 370 } else { 371 idx = -idx - 1; 372 coef[f*Nb+b] = 0; 373 for (q = 0; q < maps[grid].num_face; q++) { 374 PetscInt id = maps[grid].c_maps[idx][q].gid; 375 PetscScalar scale = maps[grid].c_maps[idx][q].scale; 376 coef[f*Nb+b] += scale*xdata[id+moffset]; 377 } 378 } 379 } 380 } 381 } 382 /* get f and df */ 383 for (PetscInt qi = 0; qi < Nq; qi++) { 384 const PetscReal *invJ = &invJe[qi*dim*dim]; 385 const PetscReal *Bq = &BB[qi*Nb]; 386 const PetscReal *Dq = &DD[qi*Nb*dim]; 387 PetscReal u_x[LANDAU_DIM]; 388 /* get f & df */ 389 for (f = 0; f < loc_Nf; ++f) { 390 const PetscInt idx = b_id*IPf_sz_glb + ipf_offset[grid] + f*loc_nip + loc_elem*Nq + qi; 391 PetscInt b, e; 392 PetscReal refSpaceDer[LANDAU_DIM]; 393 ff[idx] = 0.0; 394 for (int d = 0; d < LANDAU_DIM; ++d) refSpaceDer[d] = 0.0; 395 for (b = 0; b < Nb; ++b) { 396 const PetscInt cidx = b; 397 ff[idx] += Bq[cidx]*PetscRealPart(coef[f*Nb+cidx]); 398 for (int d = 0; d < dim; ++d) { 399 refSpaceDer[d] += Dq[cidx*dim+d]*PetscRealPart(coef[f*Nb+cidx]); 400 } 401 } 402 for (int d = 0; d < LANDAU_DIM; ++d) { 403 for (e = 0, u_x[d] = 0.0; e < LANDAU_DIM; ++e) { 404 u_x[d] += invJ[e*dim+d]*refSpaceDer[e]; 405 } 406 } 407 dudx[idx] = u_x[0]; 408 dudy[idx] = u_x[1]; 409 #if LANDAU_DIM==3 410 dudz[idx] = u_x[2]; 411 #endif 412 } 413 } // q 414 } // grid 415 } // grid*batch 416 ierr = PetscLogEventEnd(ctx->events[8],0,0,0,0);CHKERRQ(ierr); 417 #if defined(PETSC_HAVE_THREADSAFETY) 418 endtime = MPI_Wtime(); 419 if (ctx->stage) ctx->times[LANDAU_F_DF] += (endtime - starttime); 420 #endif 421 } // Jacobian setup 422 // assemble Jacobian (or mass) 423 for (PetscInt tid = 0 ; tid < ctx->batch_sz*elem_offset[num_grids] ; tid++) { // for each element 424 const PetscInt b_Nelem = elem_offset[num_grids]; 425 const PetscInt glb_elem_idx = tid%b_Nelem, b_id = tid/b_Nelem; 426 PetscInt grid = 0; 427 #if defined(PETSC_HAVE_THREADSAFETY) 428 double starttime, endtime; 429 starttime = MPI_Wtime(); 430 #endif 431 while (glb_elem_idx >= elem_offset[grid+1]) grid++; 432 { 433 const PetscInt loc_Nf = ctx->species_offset[grid+1] - ctx->species_offset[grid], loc_elem = glb_elem_idx - elem_offset[grid]; 434 const PetscInt moffset = LAND_MOFFSET(b_id,grid,ctx->batch_sz,ctx->num_grids,ctx->mat_offset), totDim = loc_Nf*Nq, elemMatSize = totDim*totDim; 435 PetscScalar *elemMat; 436 const PetscReal *invJe = &invJ_a[(ip_offset[grid] + loc_elem*Nq)*dim*dim]; 437 ierr = PetscMalloc1(elemMatSize, &elemMat);CHKERRQ(ierr); 438 ierr = PetscMemzero(elemMat, elemMatSize*sizeof(*elemMat));CHKERRQ(ierr); 439 if (shift==0.0) { // Jacobian 440 ierr = PetscLogEventBegin(ctx->events[4],0,0,0,0);CHKERRQ(ierr); 441 } else { // mass 442 ierr = PetscLogEventBegin(ctx->events[16],0,0,0,0);CHKERRQ(ierr); 443 } 444 for (PetscInt qj = 0; qj < Nq; ++qj) { 445 const PetscInt jpidx_glb = ip_offset[grid] + qj + loc_elem * Nq; 446 PetscReal g0[LANDAU_MAX_SPECIES], g2[LANDAU_MAX_SPECIES][LANDAU_DIM], g3[LANDAU_MAX_SPECIES][LANDAU_DIM][LANDAU_DIM]; // could make a LANDAU_MAX_SPECIES_GRID ~ number of ions - 1 447 PetscInt d,d2,dp,d3,IPf_idx; 448 if (shift==0.0) { // Jacobian 449 const PetscReal * const invJj = &invJe[qj*dim*dim]; 450 PetscReal gg2[LANDAU_MAX_SPECIES][LANDAU_DIM],gg3[LANDAU_MAX_SPECIES][LANDAU_DIM][LANDAU_DIM], gg2_temp[LANDAU_DIM], gg3_temp[LANDAU_DIM][LANDAU_DIM]; 451 const PetscReal vj[3] = {xx[jpidx_glb], yy[jpidx_glb], zz ? zz[jpidx_glb] : 0}, wj = ww[jpidx_glb]; 452 // create g2 & g3 453 for (d=0;d<LANDAU_DIM;d++) { // clear accumulation data D & K 454 gg2_temp[d] = 0; 455 for (d2=0;d2<LANDAU_DIM;d2++) gg3_temp[d][d2] = 0; 456 } 457 /* inner beta reduction */ 458 IPf_idx = 0; 459 for (PetscInt grid_r = 0, f_off = 0, ipidx = 0; grid_r < ctx->num_grids ; grid_r++, f_off = ctx->species_offset[grid_r]) { // IPf_idx += nip_loc_r*Nfloc_r 460 PetscInt nip_loc_r = numCells[grid_r]*Nq, Nfloc_r = Nf[grid_r]; 461 for (PetscInt ei_r = 0, loc_fdf_idx = 0; ei_r < numCells[grid_r]; ++ei_r) { 462 for (PetscInt qi = 0; qi < Nq; qi++, ipidx++, loc_fdf_idx++) { 463 const PetscReal wi = ww[ipidx], x = xx[ipidx], y = yy[ipidx]; 464 PetscReal temp1[3] = {0, 0, 0}, temp2 = 0; 465 #if LANDAU_DIM==2 466 PetscReal Ud[2][2], Uk[2][2], mask = (PetscAbs(vj[0]-x) < 100*PETSC_SQRT_MACHINE_EPSILON && PetscAbs(vj[1]-y) < 100*PETSC_SQRT_MACHINE_EPSILON) ? 0. : 1.; 467 LandauTensor2D(vj, x, y, Ud, Uk, mask); 468 #else 469 PetscReal U[3][3], z = zz[ipidx], mask = (PetscAbs(vj[0]-x) < 100*PETSC_SQRT_MACHINE_EPSILON && PetscAbs(vj[1]-y) < 100*PETSC_SQRT_MACHINE_EPSILON && PetscAbs(vj[2]-z) < 100*PETSC_SQRT_MACHINE_EPSILON) ? 0. : 1.; 470 if (ctx->use_relativistic_corrections) { 471 LandauTensor3DRelativistic(vj, x, y, z, U, mask, C_0(ctx->v_0)); 472 } else { 473 LandauTensor3D(vj, x, y, z, U, mask); 474 } 475 #endif 476 for (int f = 0; f < Nfloc_r ; ++f) { 477 const PetscInt idx = b_id*IPf_sz_glb + ipf_offset[grid_r] + f*nip_loc_r + ei_r*Nq + qi; // IPf_idx + f*nip_loc_r + loc_fdf_idx; 478 temp1[0] += dudx[idx]*nu_beta[f+f_off]*invMass[f+f_off]; 479 temp1[1] += dudy[idx]*nu_beta[f+f_off]*invMass[f+f_off]; 480 #if LANDAU_DIM==3 481 temp1[2] += dudz[idx]*nu_beta[f+f_off]*invMass[f+f_off]; 482 #endif 483 temp2 += ff[idx]*nu_beta[f+f_off]; 484 } 485 temp1[0] *= wi; 486 temp1[1] *= wi; 487 #if LANDAU_DIM==3 488 temp1[2] *= wi; 489 #endif 490 temp2 *= wi; 491 #if LANDAU_DIM==2 492 for (d2 = 0; d2 < 2; d2++) { 493 for (d3 = 0; d3 < 2; ++d3) { 494 /* K = U * grad(f): g2=e: i,A */ 495 gg2_temp[d2] += Uk[d2][d3]*temp1[d3]; 496 /* D = -U * (I \kron (fx)): g3=f: i,j,A */ 497 gg3_temp[d2][d3] += Ud[d2][d3]*temp2; 498 } 499 } 500 #else 501 for (d2 = 0; d2 < 3; ++d2) { 502 for (d3 = 0; d3 < 3; ++d3) { 503 /* K = U * grad(f): g2 = e: i,A */ 504 gg2_temp[d2] += U[d2][d3]*temp1[d3]; 505 /* D = -U * (I \kron (fx)): g3 = f: i,j,A */ 506 gg3_temp[d2][d3] += U[d2][d3]*temp2; 507 } 508 } 509 #endif 510 } // qi 511 } // ei_r 512 IPf_idx += nip_loc_r*Nfloc_r; 513 } /* grid_r - IPs */ 514 PetscCheck(IPf_idx == IPf_sz_glb,PETSC_COMM_SELF, PETSC_ERR_PLIB, "IPf_idx != IPf_sz %" PetscInt_FMT " %" PetscInt_FMT,IPf_idx,IPf_sz_glb); 515 // add alpha and put in gg2/3 516 for (PetscInt fieldA = 0, f_off = ctx->species_offset[grid]; fieldA < loc_Nf; ++fieldA) { 517 for (d2 = 0; d2 < dim; d2++) { 518 gg2[fieldA][d2] = gg2_temp[d2]*nu_alpha[fieldA+f_off]; 519 for (d3 = 0; d3 < dim; d3++) { 520 gg3[fieldA][d2][d3] = -gg3_temp[d2][d3]*nu_alpha[fieldA+f_off]*invMass[fieldA+f_off]; 521 } 522 } 523 } 524 /* add electric field term once per IP */ 525 for (PetscInt fieldA = 0, f_off = ctx->species_offset[grid] ; fieldA < loc_Nf; ++fieldA) { 526 gg2[fieldA][dim-1] += Eq_m[fieldA+f_off]; 527 } 528 /* Jacobian transform - g2, g3 */ 529 for (PetscInt fieldA = 0; fieldA < loc_Nf; ++fieldA) { 530 for (d = 0; d < dim; ++d) { 531 g2[fieldA][d] = 0.0; 532 for (d2 = 0; d2 < dim; ++d2) { 533 g2[fieldA][d] += invJj[d*dim+d2]*gg2[fieldA][d2]; 534 g3[fieldA][d][d2] = 0.0; 535 for (d3 = 0; d3 < dim; ++d3) { 536 for (dp = 0; dp < dim; ++dp) { 537 g3[fieldA][d][d2] += invJj[d*dim + d3]*gg3[fieldA][d3][dp]*invJj[d2*dim + dp]; 538 } 539 } 540 g3[fieldA][d][d2] *= wj; 541 } 542 g2[fieldA][d] *= wj; 543 } 544 } 545 } else { // mass 546 PetscReal wj = ww[jpidx_glb]; 547 /* Jacobian transform - g0 */ 548 for (PetscInt fieldA = 0; fieldA < loc_Nf ; ++fieldA) { 549 if (dim==2) { 550 g0[fieldA] = wj * shift * 2. * PETSC_PI; // move this to below and remove g0 551 } else { 552 g0[fieldA] = wj * shift; // move this to below and remove g0 553 } 554 } 555 } 556 /* FE matrix construction */ 557 { 558 PetscInt fieldA,d,f,d2,g; 559 const PetscReal *BJq = &BB[qj*Nb], *DIq = &DD[qj*Nb*dim]; 560 /* assemble - on the diagonal (I,I) */ 561 for (fieldA = 0; fieldA < loc_Nf ; fieldA++) { 562 for (f = 0; f < Nb ; f++) { 563 const PetscInt i = fieldA*Nb + f; /* Element matrix row */ 564 for (g = 0; g < Nb; ++g) { 565 const PetscInt j = fieldA*Nb + g; /* Element matrix column */ 566 const PetscInt fOff = i*totDim + j; 567 if (shift==0.0) { 568 for (d = 0; d < dim; ++d) { 569 elemMat[fOff] += DIq[f*dim+d]*g2[fieldA][d]*BJq[g]; 570 for (d2 = 0; d2 < dim; ++d2) { 571 elemMat[fOff] += DIq[f*dim + d]*g3[fieldA][d][d2]*DIq[g*dim + d2]; 572 } 573 } 574 } else { // mass 575 elemMat[fOff] += BJq[f]*g0[fieldA]*BJq[g]; 576 } 577 } 578 } 579 } 580 } 581 } /* qj loop */ 582 if (shift==0.0) { // Jacobian 583 ierr = PetscLogEventEnd(ctx->events[4],0,0,0,0);CHKERRQ(ierr); 584 } else { 585 ierr = PetscLogEventEnd(ctx->events[16],0,0,0,0);CHKERRQ(ierr); 586 } 587 #if defined(PETSC_HAVE_THREADSAFETY) 588 endtime = MPI_Wtime(); 589 if (ctx->stage) ctx->times[LANDAU_KERNEL] += (endtime - starttime); 590 #endif 591 /* assemble matrix */ 592 if (!container) { 593 PetscInt cStart; 594 ierr = PetscLogEventBegin(ctx->events[6],0,0,0,0);CHKERRQ(ierr); 595 ierr = DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, NULL);CHKERRQ(ierr); 596 ierr = DMPlexMatSetClosure(ctx->plex[grid], section[grid], globsection[grid], subJ[ LAND_PACK_IDX(b_id,grid) ], loc_elem + cStart, elemMat, ADD_VALUES);CHKERRQ(ierr); 597 ierr = PetscLogEventEnd(ctx->events[6],0,0,0,0);CHKERRQ(ierr); 598 } else { // GPU like assembly for debugging 599 PetscInt fieldA,q,f,g,d,nr,nc,rows0[LANDAU_MAX_Q_FACE]={0},cols0[LANDAU_MAX_Q_FACE]={0},rows[LANDAU_MAX_Q_FACE],cols[LANDAU_MAX_Q_FACE]; 600 PetscScalar vals[LANDAU_MAX_Q_FACE*LANDAU_MAX_Q_FACE]={0},row_scale[LANDAU_MAX_Q_FACE]={0},col_scale[LANDAU_MAX_Q_FACE]={0}; 601 /* assemble - from the diagonal (I,I) in this format for DMPlexMatSetClosure */ 602 for (fieldA = 0; fieldA < loc_Nf ; fieldA++) { 603 LandauIdx *const Idxs = &maps[grid].gIdx[loc_elem][fieldA][0]; 604 for (f = 0; f < Nb ; f++) { 605 PetscInt idx = Idxs[f]; 606 if (idx >= 0) { 607 nr = 1; 608 rows0[0] = idx; 609 row_scale[0] = 1.; 610 } else { 611 idx = -idx - 1; 612 for (q = 0, nr = 0; q < maps[grid].num_face; q++, nr++) { 613 if (maps[grid].c_maps[idx][q].gid < 0) break; 614 rows0[q] = maps[grid].c_maps[idx][q].gid; 615 row_scale[q] = maps[grid].c_maps[idx][q].scale; 616 } 617 } 618 for (g = 0; g < Nb; ++g) { 619 idx = Idxs[g]; 620 if (idx >= 0) { 621 nc = 1; 622 cols0[0] = idx; 623 col_scale[0] = 1.; 624 } else { 625 idx = -idx - 1; 626 nc = maps[grid].num_face; 627 for (q = 0, nc = 0; q < maps[grid].num_face; q++, nc++) { 628 if (maps[grid].c_maps[idx][q].gid < 0) break; 629 cols0[q] = maps[grid].c_maps[idx][q].gid; 630 col_scale[q] = maps[grid].c_maps[idx][q].scale; 631 } 632 } 633 const PetscInt i = fieldA*Nb + f; /* Element matrix row */ 634 const PetscInt j = fieldA*Nb + g; /* Element matrix column */ 635 const PetscScalar Aij = elemMat[i*totDim + j]; 636 if (coo_vals) { // mirror (i,j) in CreateStaticGPUData 637 const int fullNb = ctx->SData_d.coo_elem_fullNb[glb_elem_idx],fullNb2=fullNb*fullNb; 638 const int idx0 = b_id*ctx->SData_d.coo_elem_offsets[elem_offset[num_grids]] + ctx->SData_d.coo_elem_offsets[glb_elem_idx] + fieldA*fullNb2 + fullNb * ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][f] + nr * ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][g]; 639 for (int q = 0, idx2 = idx0; q < nr; q++) { 640 for (int d = 0; d < nc; d++, idx2++) { 641 coo_vals[idx2] = row_scale[q]*col_scale[d]*Aij; 642 } 643 } 644 } else { 645 for (q = 0; q < nr; q++) rows[q] = rows0[q] + moffset; 646 for (d = 0; d < nc; d++) cols[d] = cols0[d] + moffset; 647 for (q = 0; q < nr; q++) { 648 for (d = 0; d < nc; d++) { 649 vals[q*nc + d] = row_scale[q]*col_scale[d]*Aij; 650 } 651 } 652 ierr = MatSetValues(JacP,nr,rows,nc,cols,vals,ADD_VALUES);CHKERRQ(ierr); 653 } 654 } 655 } 656 } 657 } 658 if (loc_elem==-1) { 659 PetscErrorCode ierr2; 660 ierr2 = PetscPrintf(ctx->comm,"CPU Element matrix\n");CHKERRQ(ierr2); 661 for (int d = 0; d < totDim; ++d) { 662 for (int f = 0; f < totDim; ++f) {ierr2 = PetscPrintf(ctx->comm," %12.5e", PetscRealPart(elemMat[d*totDim + f]));CHKERRQ(ierr2);} 663 ierr2 = PetscPrintf(ctx->comm,"\n");CHKERRQ(ierr2); 664 } 665 exit(12); 666 } 667 ierr = PetscFree(elemMat);CHKERRQ(ierr); 668 } /* grid */ 669 } /* outer element & batch loop */ 670 if (shift==0.0) { // mass 671 ierr = PetscFree4(ff, dudx, dudy, dudz);CHKERRQ(ierr); 672 } 673 if (!container) { // 'CPU' assembly move nest matrix to global JacP 674 for (PetscInt b_id = 0 ; b_id < ctx->batch_sz ; b_id++) { // OpenMP 675 for (PetscInt grid=0 ; grid<ctx->num_grids ; grid++) { 676 const PetscInt moffset = LAND_MOFFSET(b_id,grid,ctx->batch_sz,ctx->num_grids,ctx->mat_offset); // b_id*b_N + ctx->mat_offset[grid]; 677 PetscInt nloc, nzl, colbuf[1024], row; 678 const PetscInt *cols; 679 const PetscScalar *vals; 680 Mat B = subJ[ LAND_PACK_IDX(b_id,grid) ]; 681 ierr = MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 682 ierr = MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 683 ierr = MatGetSize(B, &nloc, NULL);CHKERRQ(ierr); 684 for (int i=0 ; i<nloc ; i++) { 685 ierr = MatGetRow(B,i,&nzl,&cols,&vals);CHKERRQ(ierr); 686 PetscCheck(nzl<=1024,PetscObjectComm((PetscObject) B), PETSC_ERR_PLIB, "Row too big: %" PetscInt_FMT,nzl); 687 for (int j=0; j<nzl; j++) colbuf[j] = moffset + cols[j]; 688 row = moffset + i; 689 ierr = MatSetValues(JacP,1,&row,nzl,colbuf,vals,ADD_VALUES);CHKERRQ(ierr); 690 ierr = MatRestoreRow(B,i,&nzl,&cols,&vals);CHKERRQ(ierr); 691 } 692 ierr = MatDestroy(&B);CHKERRQ(ierr); 693 } 694 } 695 } 696 if (coo_vals) { 697 ierr = MatSetValuesCOO(JacP,coo_vals,ADD_VALUES);CHKERRQ(ierr); 698 ierr = PetscFree(coo_vals);CHKERRQ(ierr); 699 } 700 } /* CPU version */ 701 ierr = MatAssemblyBegin(JacP, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 702 ierr = MatAssemblyEnd(JacP, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 703 /* clean up */ 704 if (cellClosure) { 705 ierr = PetscFree(cellClosure);CHKERRQ(ierr); 706 } 707 if (xdata) { 708 ierr = VecRestoreArrayReadAndMemType(a_X,&xdata);CHKERRQ(ierr); 709 } 710 PetscFunctionReturn(0); 711 } 712 713 #if defined(LANDAU_ADD_BCS) 714 static void zero_bc(PetscInt dim, PetscInt Nf, PetscInt NfAux, 715 const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], 716 const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], 717 PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar uexact[]) 718 { 719 uexact[0] = 0; 720 } 721 #endif 722 723 #define MATVEC2(__a,__x,__p) {int i,j; for (i=0.; i<2; i++) {__p[i] = 0; for (j=0.; j<2; j++) __p[i] += __a[i][j]*__x[j]; }} 724 static void CircleInflate(PetscReal r1, PetscReal r2, PetscReal r0, PetscInt num_sections, PetscReal x, PetscReal y, 725 PetscReal *outX, PetscReal *outY) 726 { 727 PetscReal rr = PetscSqrtReal(x*x + y*y), outfact, efact; 728 if (rr < r1 + PETSC_SQRT_MACHINE_EPSILON) { 729 *outX = x; *outY = y; 730 } else { 731 const PetscReal xy[2] = {x,y}, sinphi=y/rr, cosphi=x/rr; 732 PetscReal cth,sth,xyprime[2],Rth[2][2],rotcos,newrr; 733 if (num_sections==2) { 734 rotcos = 0.70710678118654; 735 outfact = 1.5; efact = 2.5; 736 /* rotate normalized vector into [-pi/4,pi/4) */ 737 if (sinphi >= 0.) { /* top cell, -pi/2 */ 738 cth = 0.707106781186548; sth = -0.707106781186548; 739 } else { /* bottom cell -pi/8 */ 740 cth = 0.707106781186548; sth = .707106781186548; 741 } 742 } else if (num_sections==3) { 743 rotcos = 0.86602540378443; 744 outfact = 1.5; efact = 2.5; 745 /* rotate normalized vector into [-pi/6,pi/6) */ 746 if (sinphi >= 0.5) { /* top cell, -pi/3 */ 747 cth = 0.5; sth = -0.866025403784439; 748 } else if (sinphi >= -.5) { /* mid cell 0 */ 749 cth = 1.; sth = .0; 750 } else { /* bottom cell +pi/3 */ 751 cth = 0.5; sth = 0.866025403784439; 752 } 753 } else if (num_sections==4) { 754 rotcos = 0.9238795325112; 755 outfact = 1.5; efact = 3; 756 /* rotate normalized vector into [-pi/8,pi/8) */ 757 if (sinphi >= 0.707106781186548) { /* top cell, -3pi/8 */ 758 cth = 0.38268343236509; sth = -0.923879532511287; 759 } else if (sinphi >= 0.) { /* mid top cell -pi/8 */ 760 cth = 0.923879532511287; sth = -.38268343236509; 761 } else if (sinphi >= -0.707106781186548) { /* mid bottom cell + pi/8 */ 762 cth = 0.923879532511287; sth = 0.38268343236509; 763 } else { /* bottom cell + 3pi/8 */ 764 cth = 0.38268343236509; sth = .923879532511287; 765 } 766 } else { 767 cth = 0.; sth = 0.; rotcos = 0; efact = 0; 768 } 769 Rth[0][0] = cth; Rth[0][1] =-sth; 770 Rth[1][0] = sth; Rth[1][1] = cth; 771 MATVEC2(Rth,xy,xyprime); 772 if (num_sections==2) { 773 newrr = xyprime[0]/rotcos; 774 } else { 775 PetscReal newcosphi=xyprime[0]/rr, rin = r1, rout = rr - rin; 776 PetscReal routmax = r0*rotcos/newcosphi - rin, nroutmax = r0 - rin, routfrac = rout/routmax; 777 newrr = rin + routfrac*nroutmax; 778 } 779 *outX = cosphi*newrr; *outY = sinphi*newrr; 780 /* grade */ 781 PetscReal fact,tt,rs,re, rr = PetscSqrtReal(PetscSqr(*outX) + PetscSqr(*outY)); 782 if (rr > r2) { rs = r2; re = r0; fact = outfact;} /* outer zone */ 783 else { rs = r1; re = r2; fact = efact;} /* electron zone */ 784 tt = (rs + PetscPowReal((rr - rs)/(re - rs),fact) * (re-rs)) / rr; 785 *outX *= tt; 786 *outY *= tt; 787 } 788 } 789 790 static PetscErrorCode GeometryDMLandau(DM base, PetscInt point, PetscInt dim, const PetscReal abc[], PetscReal xyz[], void *a_ctx) 791 { 792 LandauCtx *ctx = (LandauCtx*)a_ctx; 793 PetscReal r = abc[0], z = abc[1]; 794 if (ctx->inflate) { 795 PetscReal absR, absZ; 796 absR = PetscAbs(r); 797 absZ = PetscAbs(z); 798 CircleInflate(ctx->i_radius[0],ctx->e_radius,ctx->radius[0],ctx->num_sections,absR,absZ,&absR,&absZ); // wrong: how do I know what grid I am on? 799 r = (r > 0) ? absR : -absR; 800 z = (z > 0) ? absZ : -absZ; 801 } 802 xyz[0] = r; 803 xyz[1] = z; 804 if (dim==3) xyz[2] = abc[2]; 805 806 PetscFunctionReturn(0); 807 } 808 809 /* create DMComposite of meshes for each species group */ 810 static PetscErrorCode LandauDMCreateVMeshes(MPI_Comm comm_self, const PetscInt dim, const char prefix[], LandauCtx *ctx, DM pack) 811 { 812 PetscErrorCode ierr; 813 814 PetscFunctionBegin; 815 { /* p4est, quads */ 816 /* Create plex mesh of Landau domain */ 817 for (PetscInt grid=0;grid<ctx->num_grids;grid++) { 818 PetscReal radius = ctx->radius[grid]; 819 if (!ctx->sphere) { 820 PetscInt cells[] = {2,2,2}; 821 PetscReal lo[] = {-radius,-radius,-radius}, hi[] = {radius,radius,radius}; 822 DMBoundaryType periodicity[3] = {DM_BOUNDARY_NONE, dim==2 ? DM_BOUNDARY_NONE : DM_BOUNDARY_NONE, DM_BOUNDARY_NONE}; 823 if (dim==2) { lo[0] = 0; cells[0] /* = cells[1] */ = 1; } 824 ierr = DMPlexCreateBoxMesh(comm_self, dim, PETSC_FALSE, cells, lo, hi, periodicity, PETSC_TRUE, &ctx->plex[grid]);CHKERRQ(ierr); // todo: make composite and create dm[grid] here 825 ierr = DMLocalizeCoordinates(ctx->plex[grid]);CHKERRQ(ierr); /* needed for periodic */ 826 if (dim==3) {ierr = PetscObjectSetName((PetscObject) ctx->plex[grid], "cube");CHKERRQ(ierr);} 827 else {ierr = PetscObjectSetName((PetscObject) ctx->plex[grid], "half-plane");CHKERRQ(ierr);} 828 } else if (dim==2) { // sphere is all wrong. should just have one inner radius 829 PetscInt numCells,cells[16][4],i,j; 830 PetscInt numVerts; 831 PetscReal inner_radius1 = ctx->i_radius[grid], inner_radius2 = ctx->e_radius; 832 PetscReal *flatCoords = NULL; 833 PetscInt *flatCells = NULL, *pcell; 834 if (ctx->num_sections==2) { 835 #if 1 836 numCells = 5; 837 numVerts = 10; 838 int cells2[][4] = { {0,1,4,3}, 839 {1,2,5,4}, 840 {3,4,7,6}, 841 {4,5,8,7}, 842 {6,7,8,9} }; 843 for (i = 0; i < numCells; i++) for (j = 0; j < 4; j++) cells[i][j] = cells2[i][j]; 844 ierr = PetscMalloc2(numVerts * 2, &flatCoords, numCells * 4, &flatCells);CHKERRQ(ierr); 845 { 846 PetscReal (*coords)[2] = (PetscReal (*) [2]) flatCoords; 847 for (j = 0; j < numVerts-1; j++) { 848 PetscReal z, r, theta = -PETSC_PI/2 + (j%3) * PETSC_PI/2; 849 PetscReal rad = (j >= 6) ? inner_radius1 : (j >= 3) ? inner_radius2 : ctx->radius[grid]; 850 z = rad * PetscSinReal(theta); 851 coords[j][1] = z; 852 r = rad * PetscCosReal(theta); 853 coords[j][0] = r; 854 } 855 coords[numVerts-1][0] = coords[numVerts-1][1] = 0; 856 } 857 #else 858 numCells = 4; 859 numVerts = 8; 860 static int cells2[][4] = {{0,1,2,3}, 861 {4,5,1,0}, 862 {5,6,2,1}, 863 {6,7,3,2}}; 864 for (i = 0; i < numCells; i++) for (j = 0; j < 4; j++) cells[i][j] = cells2[i][j]; 865 ierr = loc2(numVerts * 2, &flatCoords, numCells * 4, &flatCells);CHKERRQ(ierr); 866 { 867 PetscReal (*coords)[2] = (PetscReal (*) [2]) flatCoords; 868 PetscInt j; 869 for (j = 0; j < 8; j++) { 870 PetscReal z, r; 871 PetscReal theta = -PETSC_PI/2 + (j%4) * PETSC_PI/3.; 872 PetscReal rad = ctx->radius[grid] * ((j < 4) ? 0.5 : 1.0); 873 z = rad * PetscSinReal(theta); 874 coords[j][1] = z; 875 r = rad * PetscCosReal(theta); 876 coords[j][0] = r; 877 } 878 } 879 #endif 880 } else if (ctx->num_sections==3) { 881 numCells = 7; 882 numVerts = 12; 883 int cells2[][4] = { {0,1,5,4}, 884 {1,2,6,5}, 885 {2,3,7,6}, 886 {4,5,9,8}, 887 {5,6,10,9}, 888 {6,7,11,10}, 889 {8,9,10,11} }; 890 for (i = 0; i < numCells; i++) for (j = 0; j < 4; j++) cells[i][j] = cells2[i][j]; 891 ierr = PetscMalloc2(numVerts * 2, &flatCoords, numCells * 4, &flatCells);CHKERRQ(ierr); 892 { 893 PetscReal (*coords)[2] = (PetscReal (*) [2]) flatCoords; 894 for (j = 0; j < numVerts; j++) { 895 PetscReal z, r, theta = -PETSC_PI/2 + (j%4) * PETSC_PI/3; 896 PetscReal rad = (j >= 8) ? inner_radius1 : (j >= 4) ? inner_radius2 : ctx->radius[grid]; 897 z = rad * PetscSinReal(theta); 898 coords[j][1] = z; 899 r = rad * PetscCosReal(theta); 900 coords[j][0] = r; 901 } 902 } 903 } else if (ctx->num_sections==4) { 904 numCells = 10; 905 numVerts = 16; 906 int cells2[][4] = { {0,1,6,5}, 907 {1,2,7,6}, 908 {2,3,8,7}, 909 {3,4,9,8}, 910 {5,6,11,10}, 911 {6,7,12,11}, 912 {7,8,13,12}, 913 {8,9,14,13}, 914 {10,11,12,15}, 915 {12,13,14,15}}; 916 for (i = 0; i < numCells; i++) for (j = 0; j < 4; j++) cells[i][j] = cells2[i][j]; 917 ierr = PetscMalloc2(numVerts * 2, &flatCoords, numCells * 4, &flatCells);CHKERRQ(ierr); 918 { 919 PetscReal (*coords)[2] = (PetscReal (*) [2]) flatCoords; 920 for (j = 0; j < numVerts-1; j++) { 921 PetscReal z, r, theta = -PETSC_PI/2 + (j%5) * PETSC_PI/4; 922 PetscReal rad = (j >= 10) ? inner_radius1 : (j >= 5) ? inner_radius2 : ctx->radius[grid]; 923 z = rad * PetscSinReal(theta); 924 coords[j][1] = z; 925 r = rad * PetscCosReal(theta); 926 coords[j][0] = r; 927 } 928 coords[numVerts-1][0] = coords[numVerts-1][1] = 0; 929 } 930 } else { 931 numCells = 0; 932 numVerts = 0; 933 } 934 for (j = 0, pcell = flatCells; j < numCells; j++, pcell += 4) { 935 pcell[0] = cells[j][0]; pcell[1] = cells[j][1]; 936 pcell[2] = cells[j][2]; pcell[3] = cells[j][3]; 937 } 938 ierr = DMPlexCreateFromCellListPetsc(comm_self,2,numCells,numVerts,4,ctx->interpolate,flatCells,2,flatCoords,&ctx->plex[grid]);CHKERRQ(ierr); 939 ierr = PetscFree2(flatCoords,flatCells);CHKERRQ(ierr); 940 ierr = PetscObjectSetName((PetscObject) ctx->plex[grid], "semi-circle");CHKERRQ(ierr); 941 } else SETERRQ(ctx->comm, PETSC_ERR_PLIB, "Velocity space meshes does not support cubed sphere"); 942 943 ierr = DMSetFromOptions(ctx->plex[grid]);CHKERRQ(ierr); 944 } // grid loop 945 ierr = PetscObjectSetOptionsPrefix((PetscObject)pack,prefix);CHKERRQ(ierr); 946 ierr = DMSetFromOptions(pack);CHKERRQ(ierr); 947 948 { /* convert to p4est (or whatever), wait for discretization to create pack */ 949 char convType[256]; 950 PetscBool flg; 951 ierr = PetscOptionsBegin(ctx->comm, prefix, "Mesh conversion options", "DMPLEX");CHKERRQ(ierr); 952 ierr = PetscOptionsFList("-dm_landau_type","Convert DMPlex to another format (p4est)","plexland.c",DMList,DMPLEX,convType,256,&flg);CHKERRQ(ierr); 953 ierr = PetscOptionsEnd();CHKERRQ(ierr); 954 if (flg) { 955 ctx->use_p4est = PETSC_TRUE; /* flag for Forest */ 956 for (PetscInt grid=0;grid<ctx->num_grids;grid++) { 957 DM dmforest; 958 ierr = DMConvert(ctx->plex[grid],convType,&dmforest);CHKERRQ(ierr); 959 if (dmforest) { 960 PetscBool isForest; 961 ierr = PetscObjectSetOptionsPrefix((PetscObject)dmforest,prefix);CHKERRQ(ierr); 962 ierr = DMIsForest(dmforest,&isForest);CHKERRQ(ierr); 963 if (isForest) { 964 if (ctx->sphere && ctx->inflate) { 965 ierr = DMForestSetBaseCoordinateMapping(dmforest,GeometryDMLandau,ctx);CHKERRQ(ierr); 966 } 967 ierr = DMDestroy(&ctx->plex[grid]);CHKERRQ(ierr); 968 ctx->plex[grid] = dmforest; // Forest for adaptivity 969 } else SETERRQ(ctx->comm, PETSC_ERR_PLIB, "Converted to non Forest?"); 970 } else SETERRQ(ctx->comm, PETSC_ERR_PLIB, "Convert failed?"); 971 } 972 } else ctx->use_p4est = PETSC_FALSE; /* flag for Forest */ 973 } 974 } /* non-file */ 975 ierr = DMSetDimension(pack, dim);CHKERRQ(ierr); 976 ierr = PetscObjectSetName((PetscObject) pack, "Mesh");CHKERRQ(ierr); 977 ierr = DMSetApplicationContext(pack, ctx);CHKERRQ(ierr); 978 979 PetscFunctionReturn(0); 980 } 981 982 static PetscErrorCode SetupDS(DM pack, PetscInt dim, PetscInt grid, LandauCtx *ctx) 983 { 984 PetscErrorCode ierr; 985 PetscInt ii,i0; 986 char buf[256]; 987 PetscSection section; 988 989 PetscFunctionBegin; 990 for (ii = ctx->species_offset[grid], i0 = 0 ; ii < ctx->species_offset[grid+1] ; ii++, i0++) { 991 if (ii==0) ierr = PetscSNPrintf(buf, 256, "e"); 992 else {ierr = PetscSNPrintf(buf, sizeof(buf), "i%" PetscInt_FMT, ii);CHKERRQ(ierr);} 993 /* Setup Discretization - FEM */ 994 ierr = PetscFECreateDefault(PETSC_COMM_SELF, dim, 1, PETSC_FALSE, NULL, PETSC_DECIDE, &ctx->fe[ii]);CHKERRQ(ierr); 995 ierr = PetscObjectSetName((PetscObject) ctx->fe[ii], buf);CHKERRQ(ierr); 996 ierr = DMSetField(ctx->plex[grid], i0, NULL, (PetscObject) ctx->fe[ii]);CHKERRQ(ierr); 997 } 998 ierr = DMCreateDS(ctx->plex[grid]);CHKERRQ(ierr); 999 ierr = DMGetSection(ctx->plex[grid], §ion);CHKERRQ(ierr); 1000 for (PetscInt ii = ctx->species_offset[grid], i0 = 0 ; ii < ctx->species_offset[grid+1] ; ii++, i0++) { 1001 if (ii==0) {ierr = PetscSNPrintf(buf, sizeof(buf), "se");CHKERRQ(ierr);} 1002 else {ierr = PetscSNPrintf(buf, sizeof(buf), "si%" PetscInt_FMT, ii);CHKERRQ(ierr);} 1003 ierr = PetscSectionSetComponentName(section, i0, 0, buf);CHKERRQ(ierr); 1004 } 1005 PetscFunctionReturn(0); 1006 } 1007 1008 /* Define a Maxwellian function for testing out the operator. */ 1009 1010 /* Using cartesian velocity space coordinates, the particle */ 1011 /* density, [1/m^3], is defined according to */ 1012 1013 /* $$ n=\int_{R^3} dv^3 \left(\frac{m}{2\pi T}\right)^{3/2}\exp [- mv^2/(2T)] $$ */ 1014 1015 /* Using some constant, c, we normalize the velocity vector into a */ 1016 /* dimensionless variable according to v=c*x. Thus the density, $n$, becomes */ 1017 1018 /* $$ n=\int_{R^3} dx^3 \left(\frac{mc^2}{2\pi T}\right)^{3/2}\exp [- mc^2/(2T)*x^2] $$ */ 1019 1020 /* Defining $\theta=2T/mc^2$, we thus find that the probability density */ 1021 /* for finding the particle within the interval in a box dx^3 around x is */ 1022 1023 /* f(x;\theta)=\left(\frac{1}{\pi\theta}\right)^{3/2} \exp [ -x^2/\theta ] */ 1024 1025 typedef struct { 1026 PetscReal v_0; 1027 PetscReal kT_m; 1028 PetscReal n; 1029 PetscReal shift; 1030 } MaxwellianCtx; 1031 1032 static PetscErrorCode maxwellian(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx) 1033 { 1034 MaxwellianCtx *mctx = (MaxwellianCtx*)actx; 1035 PetscInt i; 1036 PetscReal v2 = 0, theta = 2*mctx->kT_m/(mctx->v_0*mctx->v_0); /* theta = 2kT/mc^2 */ 1037 PetscFunctionBegin; 1038 /* compute the exponents, v^2 */ 1039 for (i = 0; i < dim; ++i) v2 += x[i]*x[i]; 1040 /* evaluate the Maxwellian */ 1041 u[0] = mctx->n*PetscPowReal(PETSC_PI*theta,-1.5)*(PetscExpReal(-v2/theta)); 1042 if (mctx->shift!=0.) { 1043 v2 = 0; 1044 for (i = 0; i < dim-1; ++i) v2 += x[i]*x[i]; 1045 v2 += (x[dim-1]-mctx->shift)*(x[dim-1]-mctx->shift); 1046 /* evaluate the shifted Maxwellian */ 1047 u[0] += mctx->n*PetscPowReal(PETSC_PI*theta,-1.5)*(PetscExpReal(-v2/theta)); 1048 } 1049 PetscFunctionReturn(0); 1050 } 1051 1052 /*@ 1053 LandauAddMaxwellians - Add a Maxwellian distribution to a state 1054 1055 Collective on X 1056 1057 Input Parameters: 1058 . dm - The mesh (local) 1059 + time - Current time 1060 - temps - Temperatures of each species (global) 1061 . ns - Number density of each species (global) 1062 - grid - index into current grid - just used for offset into temp and ns 1063 + actx - Landau context 1064 1065 Output Parameter: 1066 . X - The state (local to this grid) 1067 1068 Level: beginner 1069 1070 .keywords: mesh 1071 .seealso: LandauCreateVelocitySpace() 1072 @*/ 1073 PetscErrorCode LandauAddMaxwellians(DM dm, Vec X, PetscReal time, PetscReal temps[], PetscReal ns[], PetscInt grid, PetscInt b_id, void *actx) 1074 { 1075 LandauCtx *ctx = (LandauCtx*)actx; 1076 PetscErrorCode (*initu[LANDAU_MAX_SPECIES])(PetscInt, PetscReal, const PetscReal [], PetscInt, PetscScalar [], void *); 1077 PetscErrorCode ierr,ii,i0; 1078 PetscInt dim; 1079 MaxwellianCtx *mctxs[LANDAU_MAX_SPECIES], data[LANDAU_MAX_SPECIES]; 1080 1081 PetscFunctionBegin; 1082 ierr = DMGetDimension(dm, &dim);CHKERRQ(ierr); 1083 if (!ctx) { ierr = DMGetApplicationContext(dm, &ctx);CHKERRQ(ierr); } 1084 for (ii = ctx->species_offset[grid], i0 = 0 ; ii < ctx->species_offset[grid+1] ; ii++, i0++) { 1085 mctxs[i0] = &data[i0]; 1086 data[i0].v_0 = ctx->v_0; // v_0 same for all grids 1087 data[i0].kT_m = ctx->k*temps[ii]/ctx->masses[ii]; /* kT/m */ 1088 data[i0].n = ns[ii] * (1+(double)b_id/100.0); // make solves a little different to mimic application, n[0] use for Conner-Hastie 1089 initu[i0] = maxwellian; 1090 data[i0].shift = 0; 1091 } 1092 data[0].shift = ctx->electronShift; 1093 /* need to make ADD_ALL_VALUES work - TODO */ 1094 ierr = DMProjectFunction(dm, time, initu, (void**)mctxs, INSERT_ALL_VALUES, X);CHKERRQ(ierr); 1095 PetscFunctionReturn(0); 1096 } 1097 1098 /* 1099 LandauSetInitialCondition - Addes Maxwellians with context 1100 1101 Collective on X 1102 1103 Input Parameters: 1104 . dm - The mesh 1105 - grid - index into current grid - just used for offset into temp and ns 1106 + actx - Landau context with T and n 1107 1108 Output Parameter: 1109 . X - The state 1110 1111 Level: beginner 1112 1113 .keywords: mesh 1114 .seealso: LandauCreateVelocitySpace(), LandauAddMaxwellians() 1115 */ 1116 static PetscErrorCode LandauSetInitialCondition(DM dm, Vec X, PetscInt grid, PetscInt b_id, void *actx) 1117 { 1118 LandauCtx *ctx = (LandauCtx*)actx; 1119 PetscErrorCode ierr; 1120 PetscFunctionBegin; 1121 if (!ctx) { ierr = DMGetApplicationContext(dm, &ctx);CHKERRQ(ierr); } 1122 ierr = VecZeroEntries(X);CHKERRQ(ierr); 1123 ierr = LandauAddMaxwellians(dm, X, 0.0, ctx->thermal_temps, ctx->n, grid, b_id, ctx);CHKERRQ(ierr); 1124 PetscFunctionReturn(0); 1125 } 1126 1127 // adapt a level once. Forest in/out 1128 static PetscErrorCode adaptToleranceFEM(PetscFE fem, Vec sol, PetscInt type, PetscInt grid, LandauCtx *ctx, DM *newForest) 1129 { 1130 DM forest, plex, adaptedDM = NULL; 1131 PetscDS prob; 1132 PetscBool isForest; 1133 PetscQuadrature quad; 1134 PetscInt Nq, *Nb, cStart, cEnd, c, dim, qj, k; 1135 DMLabel adaptLabel = NULL; 1136 PetscErrorCode ierr; 1137 1138 PetscFunctionBegin; 1139 forest = ctx->plex[grid]; 1140 ierr = DMCreateDS(forest);CHKERRQ(ierr); 1141 ierr = DMGetDS(forest, &prob);CHKERRQ(ierr); 1142 ierr = DMGetDimension(forest, &dim);CHKERRQ(ierr); 1143 ierr = DMIsForest(forest, &isForest);CHKERRQ(ierr); 1144 PetscCheckFalse(!isForest,ctx->comm,PETSC_ERR_ARG_WRONG,"! Forest"); 1145 ierr = DMConvert(forest, DMPLEX, &plex);CHKERRQ(ierr); 1146 ierr = DMPlexGetHeightStratum(plex,0,&cStart,&cEnd);CHKERRQ(ierr); 1147 ierr = DMLabelCreate(PETSC_COMM_SELF,"adapt",&adaptLabel);CHKERRQ(ierr); 1148 ierr = PetscFEGetQuadrature(fem, &quad);CHKERRQ(ierr); 1149 ierr = PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, NULL);CHKERRQ(ierr); 1150 PetscCheckFalse(Nq >LANDAU_MAX_NQ,ctx->comm,PETSC_ERR_ARG_WRONG,"Order too high. Nq = %" PetscInt_FMT " > LANDAU_MAX_NQ (%" PetscInt_FMT ")",Nq,LANDAU_MAX_NQ); 1151 ierr = PetscDSGetDimensions(prob, &Nb);CHKERRQ(ierr); 1152 if (type==4) { 1153 for (c = cStart; c < cEnd; c++) { 1154 ierr = DMLabelSetValue(adaptLabel, c, DM_ADAPT_REFINE);CHKERRQ(ierr); 1155 } 1156 ierr = PetscInfo(sol, "Phase:%s: Uniform refinement\n","adaptToleranceFEM");CHKERRQ(ierr); 1157 } else if (type==2) { 1158 PetscInt rCellIdx[8], eCellIdx[64], iCellIdx[64], eMaxIdx = -1, iMaxIdx = -1, nr = 0, nrmax = (dim==3) ? 8 : 2; 1159 PetscReal minRad = PETSC_INFINITY, r, eMinRad = PETSC_INFINITY, iMinRad = PETSC_INFINITY; 1160 for (c = 0; c < 64; c++) { eCellIdx[c] = iCellIdx[c] = -1; } 1161 for (c = cStart; c < cEnd; c++) { 1162 PetscReal tt, v0[LANDAU_MAX_NQ*3], detJ[LANDAU_MAX_NQ]; 1163 ierr = DMPlexComputeCellGeometryFEM(plex, c, quad, v0, NULL, NULL, detJ);CHKERRQ(ierr); 1164 for (qj = 0; qj < Nq; ++qj) { 1165 tt = PetscSqr(v0[dim*qj+0]) + PetscSqr(v0[dim*qj+1]) + PetscSqr(((dim==3) ? v0[dim*qj+2] : 0)); 1166 r = PetscSqrtReal(tt); 1167 if (r < minRad - PETSC_SQRT_MACHINE_EPSILON*10.) { 1168 minRad = r; 1169 nr = 0; 1170 rCellIdx[nr++]= c; 1171 ierr = PetscInfo(sol, "\t\tPhase: adaptToleranceFEM Found first inner r=%e, cell %" PetscInt_FMT ", qp %" PetscInt_FMT "/%" PetscInt_FMT "\n", r, c, qj+1, Nq);CHKERRQ(ierr); 1172 } else if ((r-minRad) < PETSC_SQRT_MACHINE_EPSILON*100. && nr < nrmax) { 1173 for (k=0;k<nr;k++) if (c == rCellIdx[k]) break; 1174 if (k==nr) { 1175 rCellIdx[nr++]= c; 1176 ierr = PetscInfo(sol, "\t\t\tPhase: adaptToleranceFEM Found another inner r=%e, cell %" PetscInt_FMT ", qp %" PetscInt_FMT "/%" PetscInt_FMT ", d=%e\n", r, c, qj+1, Nq, r-minRad);CHKERRQ(ierr); 1177 } 1178 } 1179 if (ctx->sphere) { 1180 if ((tt=r-ctx->e_radius) > 0) { 1181 ierr = PetscInfo(sol, "\t\t\t %" PetscInt_FMT " cell r=%g\n",c,tt);CHKERRQ(ierr); 1182 if (tt < eMinRad - PETSC_SQRT_MACHINE_EPSILON*100.) { 1183 eMinRad = tt; 1184 eMaxIdx = 0; 1185 eCellIdx[eMaxIdx++] = c; 1186 } else if (eMaxIdx > 0 && (tt-eMinRad) <= PETSC_SQRT_MACHINE_EPSILON && c != eCellIdx[eMaxIdx-1]) { 1187 eCellIdx[eMaxIdx++] = c; 1188 } 1189 } 1190 if ((tt=r-ctx->i_radius[grid]) > 0) { 1191 if (tt < iMinRad - 1.e-5) { 1192 iMinRad = tt; 1193 iMaxIdx = 0; 1194 iCellIdx[iMaxIdx++] = c; 1195 } else if (iMaxIdx > 0 && (tt-iMinRad) <= PETSC_SQRT_MACHINE_EPSILON && c != iCellIdx[iMaxIdx-1]) { 1196 iCellIdx[iMaxIdx++] = c; 1197 } 1198 } 1199 } 1200 } 1201 } 1202 for (k=0;k<nr;k++) { 1203 ierr = DMLabelSetValue(adaptLabel, rCellIdx[k], DM_ADAPT_REFINE);CHKERRQ(ierr); 1204 } 1205 if (ctx->sphere) { 1206 for (c = 0; c < eMaxIdx; c++) { 1207 ierr = DMLabelSetValue(adaptLabel, eCellIdx[c], DM_ADAPT_REFINE);CHKERRQ(ierr); 1208 ierr = PetscInfo(sol, "\t\tPhase:%s: refine sphere e cell %" PetscInt_FMT " r=%g\n","adaptToleranceFEM",eCellIdx[c],eMinRad);CHKERRQ(ierr); 1209 } 1210 for (c = 0; c < iMaxIdx; c++) { 1211 ierr = DMLabelSetValue(adaptLabel, iCellIdx[c], DM_ADAPT_REFINE);CHKERRQ(ierr); 1212 ierr = PetscInfo(sol, "\t\tPhase:%s: refine sphere i cell %" PetscInt_FMT " r=%g\n","adaptToleranceFEM",iCellIdx[c],iMinRad);CHKERRQ(ierr); 1213 } 1214 } 1215 ierr = PetscInfo(sol, "Phase:%s: Adaptive refine origin cells %" PetscInt_FMT ",%" PetscInt_FMT " r=%g\n","adaptToleranceFEM",rCellIdx[0],rCellIdx[1],minRad);CHKERRQ(ierr); 1216 } else if (type==0 || type==1 || type==3) { /* refine along r=0 axis */ 1217 PetscScalar *coef = NULL; 1218 Vec coords; 1219 PetscInt csize,Nv,d,nz; 1220 DM cdm; 1221 PetscSection cs; 1222 ierr = DMGetCoordinatesLocal(forest, &coords);CHKERRQ(ierr); 1223 ierr = DMGetCoordinateDM(forest, &cdm);CHKERRQ(ierr); 1224 ierr = DMGetLocalSection(cdm, &cs);CHKERRQ(ierr); 1225 for (c = cStart; c < cEnd; c++) { 1226 PetscInt doit = 0, outside = 0; 1227 ierr = DMPlexVecGetClosure(cdm, cs, coords, c, &csize, &coef);CHKERRQ(ierr); 1228 Nv = csize/dim; 1229 for (nz = d = 0; d < Nv; d++) { 1230 PetscReal z = PetscRealPart(coef[d*dim + (dim-1)]), x = PetscSqr(PetscRealPart(coef[d*dim + 0])) + ((dim==3) ? PetscSqr(PetscRealPart(coef[d*dim + 1])) : 0); 1231 x = PetscSqrtReal(x); 1232 if (x < PETSC_MACHINE_EPSILON*10. && PetscAbs(z)<PETSC_MACHINE_EPSILON*10.) doit = 1; /* refine origin */ 1233 else if (type==0 && (z < -PETSC_MACHINE_EPSILON*10. || z > ctx->re_radius+PETSC_MACHINE_EPSILON*10.)) outside++; /* first pass don't refine bottom */ 1234 else if (type==1 && (z > ctx->vperp0_radius1 || z < -ctx->vperp0_radius1)) outside++; /* don't refine outside electron refine radius */ 1235 else if (type==3 && (z > ctx->vperp0_radius2 || z < -ctx->vperp0_radius2)) outside++; /* don't refine outside ion refine radius */ 1236 if (x < PETSC_MACHINE_EPSILON*10.) nz++; 1237 } 1238 ierr = DMPlexVecRestoreClosure(cdm, cs, coords, c, &csize, &coef);CHKERRQ(ierr); 1239 if (doit || (outside<Nv && nz)) { 1240 ierr = DMLabelSetValue(adaptLabel, c, DM_ADAPT_REFINE);CHKERRQ(ierr); 1241 } 1242 } 1243 ierr = PetscInfo(sol, "Phase:%s: RE refinement\n","adaptToleranceFEM");CHKERRQ(ierr); 1244 } 1245 ierr = DMDestroy(&plex);CHKERRQ(ierr); 1246 ierr = DMAdaptLabel(forest, adaptLabel, &adaptedDM);CHKERRQ(ierr); 1247 ierr = DMLabelDestroy(&adaptLabel);CHKERRQ(ierr); 1248 *newForest = adaptedDM; 1249 if (adaptedDM) { 1250 if (isForest) { 1251 ierr = DMForestSetAdaptivityForest(adaptedDM,NULL);CHKERRQ(ierr); // ???? 1252 } else exit(33); // ??????? 1253 ierr = DMConvert(adaptedDM, DMPLEX, &plex);CHKERRQ(ierr); 1254 ierr = DMPlexGetHeightStratum(plex,0,&cStart,&cEnd);CHKERRQ(ierr); 1255 ierr = PetscInfo(sol, "\tPhase: adaptToleranceFEM: %" PetscInt_FMT " cells, %" PetscInt_FMT " total quadrature points\n",cEnd-cStart,Nq*(cEnd-cStart));CHKERRQ(ierr); 1256 ierr = DMDestroy(&plex);CHKERRQ(ierr); 1257 } else *newForest = NULL; 1258 PetscFunctionReturn(0); 1259 } 1260 1261 // forest goes in (ctx->plex[grid]), plex comes out 1262 static PetscErrorCode adapt(PetscInt grid, LandauCtx *ctx, Vec *uu) 1263 { 1264 PetscErrorCode ierr; 1265 PetscInt adaptIter; 1266 1267 PetscFunctionBegin; 1268 PetscInt type, limits[5] = {(grid==0) ? ctx->numRERefine : 0, (grid==0) ? ctx->nZRefine1 : 0, ctx->numAMRRefine[grid], (grid==0) ? ctx->nZRefine2 : 0,ctx->postAMRRefine[grid]}; 1269 for (type=0;type<5;type++) { 1270 for (adaptIter = 0; adaptIter<limits[type];adaptIter++) { 1271 DM newForest = NULL; 1272 ierr = adaptToleranceFEM(ctx->fe[0], *uu, type, grid, ctx, &newForest);CHKERRQ(ierr); 1273 if (newForest) { 1274 ierr = DMDestroy(&ctx->plex[grid]);CHKERRQ(ierr); 1275 ierr = VecDestroy(uu);CHKERRQ(ierr); 1276 ierr = DMCreateGlobalVector(newForest,uu);CHKERRQ(ierr); 1277 ierr = PetscObjectSetName((PetscObject) *uu, "uAMR");CHKERRQ(ierr); 1278 ierr = LandauSetInitialCondition(newForest, *uu, grid, 0, ctx);CHKERRQ(ierr); 1279 ctx->plex[grid] = newForest; 1280 } else { 1281 exit(4); // can happen with no AMR and post refinement 1282 } 1283 } 1284 } 1285 PetscFunctionReturn(0); 1286 } 1287 1288 static PetscErrorCode ProcessOptions(LandauCtx *ctx, const char prefix[]) 1289 { 1290 PetscErrorCode ierr; 1291 PetscBool flg, sph_flg; 1292 PetscInt ii,nt,nm,nc,num_species_grid[LANDAU_MAX_GRIDS]; 1293 PetscReal v0_grid[LANDAU_MAX_GRIDS]; 1294 DM dummy; 1295 1296 PetscFunctionBegin; 1297 ierr = DMCreate(ctx->comm,&dummy);CHKERRQ(ierr); 1298 /* get options - initialize context */ 1299 ctx->verbose = 1; // should be 0 for silent compliance 1300 #if defined(PETSC_HAVE_THREADSAFETY) 1301 ctx->batch_sz = PetscNumOMPThreads; 1302 #else 1303 ctx->batch_sz = 1; 1304 #endif 1305 ctx->batch_view_idx = 0; 1306 ctx->interpolate = PETSC_TRUE; 1307 ctx->gpu_assembly = PETSC_TRUE; 1308 ctx->aux_bool = PETSC_FALSE; 1309 ctx->electronShift = 0; 1310 ctx->M = NULL; 1311 ctx->J = NULL; 1312 /* geometry and grids */ 1313 ctx->sphere = PETSC_FALSE; 1314 ctx->inflate = PETSC_FALSE; 1315 ctx->aux_bool = PETSC_FALSE; 1316 ctx->use_p4est = PETSC_FALSE; 1317 ctx->num_sections = 3; /* 2, 3 or 4 */ 1318 for (PetscInt grid=0;grid<LANDAU_MAX_GRIDS;grid++) { 1319 ctx->radius[grid] = 5.; /* thermal radius (velocity) */ 1320 ctx->numAMRRefine[grid] = 5; 1321 ctx->postAMRRefine[grid] = 0; 1322 ctx->species_offset[grid+1] = 1; // one species default 1323 num_species_grid[grid] = 0; 1324 ctx->plex[grid] = NULL; /* cache as expensive to Convert */ 1325 } 1326 ctx->species_offset[0] = 0; 1327 ctx->re_radius = 0.; 1328 ctx->vperp0_radius1 = 0; 1329 ctx->vperp0_radius2 = 0; 1330 ctx->nZRefine1 = 0; 1331 ctx->nZRefine2 = 0; 1332 ctx->numRERefine = 0; 1333 num_species_grid[0] = 1; // one species default 1334 /* species - [0] electrons, [1] one ion species eg, duetarium, [2] heavy impurity ion, ... */ 1335 ctx->charges[0] = -1; /* electron charge (MKS) */ 1336 ctx->masses[0] = 1/1835.469965278441013; /* temporary value in proton mass */ 1337 ctx->n[0] = 1; 1338 ctx->v_0 = 1; /* thermal velocity, we could start with a scale != 1 */ 1339 ctx->thermal_temps[0] = 1; 1340 /* constants, etc. */ 1341 ctx->epsilon0 = 8.8542e-12; /* permittivity of free space (MKS) F/m */ 1342 ctx->k = 1.38064852e-23; /* Boltzmann constant (MKS) J/K */ 1343 ctx->lnLam = 10; /* cross section ratio large - small angle collisions */ 1344 ctx->n_0 = 1.e20; /* typical plasma n, but could set it to 1 */ 1345 ctx->Ez = 0; 1346 for (PetscInt grid=0;grid<LANDAU_NUM_TIMERS;grid++) ctx->times[grid] = 0; 1347 ctx->use_matrix_mass = PETSC_FALSE; 1348 ctx->use_relativistic_corrections = PETSC_FALSE; 1349 ctx->use_energy_tensor_trick = PETSC_FALSE; /* Use Eero's trick for energy conservation v --> grad(v^2/2) */ 1350 ctx->SData_d.w = NULL; 1351 ctx->SData_d.x = NULL; 1352 ctx->SData_d.y = NULL; 1353 ctx->SData_d.z = NULL; 1354 ctx->SData_d.invJ = NULL; 1355 ctx->jacobian_field_major_order = PETSC_FALSE; 1356 ctx->SData_d.coo_elem_offsets = NULL; 1357 ctx->SData_d.coo_elem_point_offsets = NULL; 1358 ctx->use_coo_assembly = PETSC_FALSE; 1359 ctx->SData_d.coo_elem_fullNb = NULL; 1360 ctx->SData_d.coo_size = 0; 1361 ierr = PetscOptionsBegin(ctx->comm, prefix, "Options for Fokker-Plank-Landau collision operator", "none");CHKERRQ(ierr); 1362 { 1363 char opstring[256]; 1364 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 1365 ctx->deviceType = LANDAU_KOKKOS; 1366 ierr = PetscStrcpy(opstring,"kokkos");CHKERRQ(ierr); 1367 #elif defined(PETSC_HAVE_CUDA) 1368 ctx->deviceType = LANDAU_CUDA; 1369 ierr = PetscStrcpy(opstring,"cuda");CHKERRQ(ierr); 1370 #else 1371 ctx->deviceType = LANDAU_CPU; 1372 ierr = PetscStrcpy(opstring,"cpu");CHKERRQ(ierr); 1373 #endif 1374 ierr = PetscOptionsString("-dm_landau_device_type","Use kernels on 'cpu', 'cuda', or 'kokkos'","plexland.c",opstring,opstring,sizeof(opstring),NULL);CHKERRQ(ierr); 1375 ierr = PetscStrcmp("cpu",opstring,&flg);CHKERRQ(ierr); 1376 if (flg) { 1377 ctx->deviceType = LANDAU_CPU; 1378 } else { 1379 ierr = PetscStrcmp("cuda",opstring,&flg);CHKERRQ(ierr); 1380 if (flg) { 1381 ctx->deviceType = LANDAU_CUDA; 1382 } else { 1383 ierr = PetscStrcmp("kokkos",opstring,&flg);CHKERRQ(ierr); 1384 if (flg) ctx->deviceType = LANDAU_KOKKOS; 1385 else SETERRQ(ctx->comm,PETSC_ERR_ARG_WRONG,"-dm_landau_device_type %s",opstring); 1386 } 1387 } 1388 } 1389 ierr = PetscOptionsReal("-dm_landau_electron_shift","Shift in thermal velocity of electrons","none",ctx->electronShift,&ctx->electronShift, NULL);CHKERRQ(ierr); 1390 ierr = PetscOptionsInt("-dm_landau_verbose", "Level of verbosity output", "plexland.c", ctx->verbose, &ctx->verbose, NULL);CHKERRQ(ierr); 1391 ierr = PetscOptionsInt("-dm_landau_batch_size", "Number of 'vertices' to batch", "ex2.c", ctx->batch_sz, &ctx->batch_sz, NULL);CHKERRQ(ierr); 1392 PetscCheck(LANDAU_MAX_BATCH_SZ >= ctx->batch_sz,ctx->comm,PETSC_ERR_ARG_WRONG,"LANDAU_MAX_BATCH_SZ %" PetscInt_FMT " < ctx->batch_sz %" PetscInt_FMT,LANDAU_MAX_BATCH_SZ,ctx->batch_sz); 1393 ierr = PetscOptionsInt("-dm_landau_batch_view_idx", "Index of batch for diagnostics like plotting", "ex2.c", ctx->batch_view_idx, &ctx->batch_view_idx, NULL);CHKERRQ(ierr); 1394 PetscCheck(ctx->batch_view_idx < ctx->batch_sz,ctx->comm,PETSC_ERR_ARG_WRONG,"-ctx->batch_view_idx %" PetscInt_FMT " > ctx->batch_sz %" PetscInt_FMT,ctx->batch_view_idx,ctx->batch_sz); 1395 ierr = PetscOptionsReal("-dm_landau_Ez","Initial parallel electric field in unites of Conner-Hastie critical field","plexland.c",ctx->Ez,&ctx->Ez, NULL);CHKERRQ(ierr); 1396 ierr = PetscOptionsReal("-dm_landau_n_0","Normalization constant for number density","plexland.c",ctx->n_0,&ctx->n_0, NULL);CHKERRQ(ierr); 1397 ierr = PetscOptionsReal("-dm_landau_ln_lambda","Cross section parameter","plexland.c",ctx->lnLam,&ctx->lnLam, NULL);CHKERRQ(ierr); 1398 ierr = PetscOptionsBool("-dm_landau_use_mataxpy_mass", "Use fast but slightly fragile MATAXPY to add mass term", "plexland.c", ctx->use_matrix_mass, &ctx->use_matrix_mass, NULL);CHKERRQ(ierr); 1399 ierr = PetscOptionsBool("-dm_landau_use_relativistic_corrections", "Use relativistic corrections", "plexland.c", ctx->use_relativistic_corrections, &ctx->use_relativistic_corrections, NULL);CHKERRQ(ierr); 1400 ierr = PetscOptionsBool("-dm_landau_use_energy_tensor_trick", "Use Eero's trick of using grad(v^2/2) instead of v as args to Landau tensor to conserve energy with relativistic corrections and Q1 elements", "plexland.c", ctx->use_energy_tensor_trick, &ctx->use_energy_tensor_trick, NULL);CHKERRQ(ierr); 1401 1402 /* get num species with temperature, set defaults */ 1403 for (ii=1;ii<LANDAU_MAX_SPECIES;ii++) { 1404 ctx->thermal_temps[ii] = 1; 1405 ctx->charges[ii] = 1; 1406 ctx->masses[ii] = 1; 1407 ctx->n[ii] = 1; 1408 } 1409 nt = LANDAU_MAX_SPECIES; 1410 ierr = PetscOptionsRealArray("-dm_landau_thermal_temps", "Temperature of each species [e,i_0,i_1,...] in keV (must be set to set number of species)", "plexland.c", ctx->thermal_temps, &nt, &flg);CHKERRQ(ierr); 1411 if (flg) { 1412 ierr = PetscInfo(dummy, "num_species set to number of thermal temps provided (%" PetscInt_FMT ")\n",nt);CHKERRQ(ierr); 1413 ctx->num_species = nt; 1414 } else SETERRQ(ctx->comm,PETSC_ERR_ARG_WRONG,"-dm_landau_thermal_temps ,t1,t2,.. must be provided to set the number of species"); 1415 for (ii=0;ii<ctx->num_species;ii++) ctx->thermal_temps[ii] *= 1.1604525e7; /* convert to Kelvin */ 1416 nm = LANDAU_MAX_SPECIES-1; 1417 ierr = PetscOptionsRealArray("-dm_landau_ion_masses", "Mass of each species in units of proton mass [i_0=2,i_1=40...]", "plexland.c", &ctx->masses[1], &nm, &flg);CHKERRQ(ierr); 1418 if (flg && nm != ctx->num_species-1) { 1419 SETERRQ(ctx->comm,PETSC_ERR_ARG_WRONG,"num ion masses %" PetscInt_FMT " != num species %" PetscInt_FMT "",nm,ctx->num_species-1); 1420 } 1421 nm = LANDAU_MAX_SPECIES; 1422 ierr = PetscOptionsRealArray("-dm_landau_n", "Number density of each species = n_s * n_0", "plexland.c", ctx->n, &nm, &flg);CHKERRQ(ierr); 1423 PetscCheckFalse(flg && nm != ctx->num_species,ctx->comm,PETSC_ERR_ARG_WRONG,"wrong num n: %" PetscInt_FMT " != num species %" PetscInt_FMT "",nm,ctx->num_species); 1424 for (ii=0;ii<LANDAU_MAX_SPECIES;ii++) ctx->masses[ii] *= 1.6720e-27; /* scale by proton mass kg */ 1425 ctx->masses[0] = 9.10938356e-31; /* electron mass kg (should be about right already) */ 1426 ctx->m_0 = ctx->masses[0]; /* arbitrary reference mass, electrons */ 1427 nc = LANDAU_MAX_SPECIES-1; 1428 ierr = PetscOptionsRealArray("-dm_landau_ion_charges", "Charge of each species in units of proton charge [i_0=2,i_1=18,...]", "plexland.c", &ctx->charges[1], &nc, &flg);CHKERRQ(ierr); 1429 if (flg) PetscCheck(nc == ctx->num_species-1,ctx->comm,PETSC_ERR_ARG_WRONG,"num charges %" PetscInt_FMT " != num species %" PetscInt_FMT,nc,ctx->num_species-1); 1430 for (ii=0;ii<LANDAU_MAX_SPECIES;ii++) ctx->charges[ii] *= 1.6022e-19; /* electron/proton charge (MKS) */ 1431 /* geometry and grids */ 1432 nt = LANDAU_MAX_GRIDS; 1433 ierr = PetscOptionsIntArray("-dm_landau_num_species_grid","Number of species on each grid: [ 1, ....] or [S, 0 ....] for single grid","plexland.c", num_species_grid, &nt, &flg);CHKERRQ(ierr); 1434 if (flg) { 1435 ctx->num_grids = nt; 1436 for (ii=nt=0;ii<ctx->num_grids;ii++) nt += num_species_grid[ii]; 1437 PetscCheck(ctx->num_species == nt,ctx->comm,PETSC_ERR_ARG_WRONG,"-dm_landau_num_species_grid: sum %" PetscInt_FMT " != num_species = %" PetscInt_FMT ". %" PetscInt_FMT " grids (check that number of grids <= LANDAU_MAX_GRIDS = %" PetscInt_FMT ")",nt,ctx->num_species,ctx->num_grids,LANDAU_MAX_GRIDS); 1438 } else { 1439 ctx->num_grids = 1; // go back to a single grid run 1440 num_species_grid[0] = ctx->num_species; 1441 } 1442 for (ctx->species_offset[0] = ii = 0; ii < ctx->num_grids ; ii++) ctx->species_offset[ii+1] = ctx->species_offset[ii] + num_species_grid[ii]; 1443 PetscCheck(ctx->species_offset[ctx->num_grids] == ctx->num_species,ctx->comm,PETSC_ERR_ARG_WRONG,"ctx->species_offset[ctx->num_grids] %" PetscInt_FMT " != ctx->num_species = %" PetscInt_FMT " ???????????",ctx->species_offset[ctx->num_grids],ctx->num_species); 1444 for (PetscInt grid = 0; grid < ctx->num_grids ; grid++) { 1445 int iii = ctx->species_offset[grid]; // normalize with first (arbitrary) species on grid 1446 v0_grid[grid] = PetscSqrtReal(ctx->k*ctx->thermal_temps[iii]/ctx->masses[iii]); /* arbitrary units for non-dimensionalization: mean velocity in 1D of first species on grid */ 1447 } 1448 ii = 0; 1449 ierr = PetscOptionsInt("-dm_landau_v0_grid", "Index of grid to use for setting v_0 (electrons are default). Not recommended to change", "plexland.c", ii, &ii, NULL);CHKERRQ(ierr); 1450 ctx->v_0 = v0_grid[ii]; /* arbitrary units for non dimensionalization: global mean velocity in 1D of electrons */ 1451 ctx->t_0 = 8*PETSC_PI*PetscSqr(ctx->epsilon0*ctx->m_0/PetscSqr(ctx->charges[0]))/ctx->lnLam/ctx->n_0*PetscPowReal(ctx->v_0,3); /* note, this t_0 makes nu[0,0]=1 */ 1452 /* domain */ 1453 nt = LANDAU_MAX_GRIDS; 1454 ierr = PetscOptionsRealArray("-dm_landau_domain_radius","Phase space size in units of thermal velocity of grid","plexland.c",ctx->radius,&nt, &flg);CHKERRQ(ierr); 1455 if (flg) PetscCheck(nt >= ctx->num_grids,ctx->comm,PETSC_ERR_ARG_WRONG,"-dm_landau_domain_radius: given %" PetscInt_FMT " radius != number grids %" PetscInt_FMT,nt,ctx->num_grids); 1456 for (PetscInt grid = 0; grid < ctx->num_grids ; grid++) { 1457 if (flg && ctx->radius[grid] <= 0) { /* negative is ratio of c */ 1458 if (ctx->radius[grid] == 0) ctx->radius[grid] = 0.75; 1459 else ctx->radius[grid] = -ctx->radius[grid]; 1460 ctx->radius[grid] = ctx->radius[grid]*SPEED_OF_LIGHT/ctx->v_0; // use any species on grid to normalize (v_0 same for all on grid) 1461 ierr = PetscInfo(dummy, "Change domain radius to %g for grid %" PetscInt_FMT "\n",ctx->radius[grid],grid);CHKERRQ(ierr); 1462 } 1463 ctx->radius[grid] *= v0_grid[grid]/ctx->v_0; // scale domain by thermal radius relative to v_0 1464 } 1465 /* amr parametres */ 1466 nt = LANDAU_MAX_GRIDS; 1467 ierr = PetscOptionsIntArray("-dm_landau_amr_levels_max", "Number of AMR levels of refinement around origin, after (RE) refinements along z", "plexland.c", ctx->numAMRRefine, &nt, &flg);CHKERRQ(ierr); 1468 PetscCheckFalse(flg && nt < ctx->num_grids,ctx->comm,PETSC_ERR_ARG_WRONG,"-dm_landau_amr_levels_max: given %" PetscInt_FMT " != number grids %" PetscInt_FMT,nt,ctx->num_grids); 1469 nt = LANDAU_MAX_GRIDS; 1470 ierr = PetscOptionsIntArray("-dm_landau_amr_post_refine", "Number of levels to uniformly refine after AMR", "plexland.c", ctx->postAMRRefine, &nt, &flg);CHKERRQ(ierr); 1471 for (ii=1;ii<ctx->num_grids;ii++) ctx->postAMRRefine[ii] = ctx->postAMRRefine[0]; // all grids the same now 1472 ierr = PetscOptionsInt("-dm_landau_amr_re_levels", "Number of levels to refine along v_perp=0, z>0", "plexland.c", ctx->numRERefine, &ctx->numRERefine, &flg);CHKERRQ(ierr); 1473 ierr = PetscOptionsInt("-dm_landau_amr_z_refine1", "Number of levels to refine along v_perp=0", "plexland.c", ctx->nZRefine1, &ctx->nZRefine1, &flg);CHKERRQ(ierr); 1474 ierr = PetscOptionsInt("-dm_landau_amr_z_refine2", "Number of levels to refine along v_perp=0", "plexland.c", ctx->nZRefine2, &ctx->nZRefine2, &flg);CHKERRQ(ierr); 1475 ierr = PetscOptionsReal("-dm_landau_re_radius","velocity range to refine on positive (z>0) r=0 axis for runaways","plexland.c",ctx->re_radius,&ctx->re_radius, &flg);CHKERRQ(ierr); 1476 ierr = PetscOptionsReal("-dm_landau_z_radius1","velocity range to refine r=0 axis (for electrons)","plexland.c",ctx->vperp0_radius1,&ctx->vperp0_radius1, &flg);CHKERRQ(ierr); 1477 ierr = PetscOptionsReal("-dm_landau_z_radius2","velocity range to refine r=0 axis (for ions) after origin AMR","plexland.c",ctx->vperp0_radius2, &ctx->vperp0_radius2, &flg);CHKERRQ(ierr); 1478 /* spherical domain (not used) */ 1479 ierr = PetscOptionsInt("-dm_landau_num_sections", "Number of tangential section in (2D) grid, 2, 3, of 4", "plexland.c", ctx->num_sections, &ctx->num_sections, NULL);CHKERRQ(ierr); 1480 ierr = PetscOptionsBool("-dm_landau_sphere", "use sphere/semi-circle domain instead of rectangle", "plexland.c", ctx->sphere, &ctx->sphere, &sph_flg);CHKERRQ(ierr); 1481 ierr = PetscOptionsBool("-dm_landau_inflate", "With sphere, inflate for curved edges", "plexland.c", ctx->inflate, &ctx->inflate, &flg);CHKERRQ(ierr); 1482 ierr = PetscOptionsReal("-dm_landau_e_radius","Electron thermal velocity, used for circular meshes","plexland.c",ctx->e_radius, &ctx->e_radius, &flg);CHKERRQ(ierr); 1483 if (flg && !sph_flg) ctx->sphere = PETSC_TRUE; /* you gave me an e radius but did not set sphere, user error really */ 1484 if (!flg) { 1485 ctx->e_radius = 1.5*PetscSqrtReal(8*ctx->k*ctx->thermal_temps[0]/ctx->masses[0]/PETSC_PI)/ctx->v_0; 1486 } 1487 nt = LANDAU_MAX_GRIDS; 1488 ierr = PetscOptionsRealArray("-dm_landau_i_radius","Ion thermal velocity, used for circular meshes","plexland.c",ctx->i_radius, &nt, &flg);CHKERRQ(ierr); 1489 if (flg && !sph_flg) ctx->sphere = PETSC_TRUE; 1490 if (!flg) { 1491 ctx->i_radius[0] = 1.5*PetscSqrtReal(8*ctx->k*ctx->thermal_temps[1]/ctx->masses[1]/PETSC_PI)/ctx->v_0; // need to correct for ion grid domain 1492 } 1493 if (flg) PetscCheck(ctx->num_grids == nt,ctx->comm,PETSC_ERR_ARG_WRONG,"-dm_landau_i_radius: %" PetscInt_FMT " != num_species = %" PetscInt_FMT,nt,ctx->num_grids); 1494 if (ctx->sphere) PetscCheck(ctx->e_radius > ctx->i_radius[0],ctx->comm,PETSC_ERR_ARG_WRONG,"bad radii: %g < %g < %g",ctx->i_radius[0],ctx->e_radius,ctx->radius[0]); 1495 /* processing options */ 1496 ierr = PetscOptionsBool("-dm_landau_gpu_assembly", "Assemble Jacobian on GPU", "plexland.c", ctx->gpu_assembly, &ctx->gpu_assembly, NULL);CHKERRQ(ierr); 1497 if (ctx->deviceType == LANDAU_CPU || ctx->deviceType == LANDAU_KOKKOS) { // make Kokkos 1498 ierr = PetscOptionsBool("-dm_landau_use_coo_assembly", "Assemble Jacobian with Kokkos on 'device'", "plexland.c", ctx->use_coo_assembly, &ctx->use_coo_assembly, NULL);CHKERRQ(ierr); 1499 if (ctx->use_coo_assembly) PetscCheck(ctx->gpu_assembly,ctx->comm,PETSC_ERR_ARG_WRONG,"COO assembly requires 'gpu assembly' even if Kokkos 'CPU' back-end %d",ctx->use_coo_assembly); 1500 } 1501 ierr = PetscOptionsBool("-dm_landau_jacobian_field_major_order", "Reorder Jacobian for GPU assembly with field major, or block diagonal, ordering", "plexland.c", ctx->jacobian_field_major_order, &ctx->jacobian_field_major_order, NULL);CHKERRQ(ierr); 1502 if (ctx->jacobian_field_major_order) PetscCheck(ctx->gpu_assembly,ctx->comm,PETSC_ERR_ARG_WRONG,"-dm_landau_jacobian_field_major_order requires -dm_landau_gpu_assembly"); 1503 ierr = PetscOptionsEnd();CHKERRQ(ierr); 1504 1505 for (ii=ctx->num_species;ii<LANDAU_MAX_SPECIES;ii++) ctx->masses[ii] = ctx->thermal_temps[ii] = ctx->charges[ii] = 0; 1506 if (ctx->verbose > 0) { 1507 ierr = PetscPrintf(ctx->comm, "masses: e=%10.3e; ions in proton mass units: %10.3e %10.3e ...\n",ctx->masses[0],ctx->masses[1]/1.6720e-27,ctx->num_species>2 ? ctx->masses[2]/1.6720e-27 : 0);CHKERRQ(ierr); 1508 ierr = PetscPrintf(ctx->comm, "charges: e=%10.3e; charges in elementary units: %10.3e %10.3e\n", ctx->charges[0],-ctx->charges[1]/ctx->charges[0],ctx->num_species>2 ? -ctx->charges[2]/ctx->charges[0] : 0);CHKERRQ(ierr); 1509 ierr = PetscPrintf(ctx->comm, "n: e: %10.3e i: %10.3e %10.3e\n", ctx->n[0],ctx->n[1],ctx->num_species>2 ? ctx->n[2] : 0);CHKERRQ(ierr); 1510 ierr = PetscPrintf(ctx->comm, "thermal T (K): e=%10.3e i=%10.3e %10.3e. v_0=%10.3e (%10.3ec) n_0=%10.3e t_0=%10.3e, %s, %s, %" PetscInt_FMT " batched\n", ctx->thermal_temps[0], ctx->thermal_temps[1], (ctx->num_species>2) ? ctx->thermal_temps[2] : 0, ctx->v_0, ctx->v_0/SPEED_OF_LIGHT, ctx->n_0, ctx->t_0, ctx->use_relativistic_corrections ? "relativistic" : "classical", ctx->use_energy_tensor_trick ? "Use trick" : "Intuitive",ctx->batch_sz);CHKERRQ(ierr); 1511 ierr = PetscPrintf(ctx->comm, "Domain radius (AMR levels) grid %" PetscInt_FMT ": %10.3e (%" PetscInt_FMT ") ",0,ctx->radius[0],ctx->numAMRRefine[0]);CHKERRQ(ierr); 1512 for (ii=1;ii<ctx->num_grids;ii++) {ierr = PetscPrintf(ctx->comm, ", %" PetscInt_FMT ": %10.3e (%" PetscInt_FMT ") ",ii,ctx->radius[ii],ctx->numAMRRefine[ii]);CHKERRQ(ierr);} 1513 ierr = PetscPrintf(ctx->comm,"\n");CHKERRQ(ierr); 1514 if (ctx->jacobian_field_major_order) { 1515 ierr = PetscPrintf(ctx->comm,"Using field major order for GPU Jacobian\n");CHKERRQ(ierr); 1516 } else { 1517 ierr = PetscPrintf(ctx->comm,"Using default Plex order for all matrices\n");CHKERRQ(ierr); 1518 } 1519 } 1520 ierr = DMDestroy(&dummy);CHKERRQ(ierr); 1521 { 1522 PetscMPIInt rank; 1523 ierr = MPI_Comm_rank(ctx->comm, &rank);CHKERRMPI(ierr); 1524 ctx->stage = 0; 1525 ierr = PetscLogEventRegister("Landau Create", DM_CLASSID, &ctx->events[13]);CHKERRQ(ierr); /* 13 */ 1526 ierr = PetscLogEventRegister(" GPU ass. setup", DM_CLASSID, &ctx->events[2]);CHKERRQ(ierr); /* 2 */ 1527 ierr = PetscLogEventRegister(" Build matrix", DM_CLASSID, &ctx->events[12]);CHKERRQ(ierr); /* 12 */ 1528 ierr = PetscLogEventRegister(" Assembly maps", DM_CLASSID, &ctx->events[15]);CHKERRQ(ierr); /* 15 */ 1529 ierr = PetscLogEventRegister("Landau Mass mat", DM_CLASSID, &ctx->events[14]);CHKERRQ(ierr); /* 14 */ 1530 ierr = PetscLogEventRegister("Landau Operator", DM_CLASSID, &ctx->events[11]);CHKERRQ(ierr); /* 11 */ 1531 ierr = PetscLogEventRegister("Landau Jacobian", DM_CLASSID, &ctx->events[0]);CHKERRQ(ierr); /* 0 */ 1532 ierr = PetscLogEventRegister("Landau Mass", DM_CLASSID, &ctx->events[9]);CHKERRQ(ierr); /* 9 */ 1533 ierr = PetscLogEventRegister(" Preamble", DM_CLASSID, &ctx->events[10]);CHKERRQ(ierr); /* 10 */ 1534 ierr = PetscLogEventRegister(" static IP Data", DM_CLASSID, &ctx->events[7]);CHKERRQ(ierr); /* 7 */ 1535 ierr = PetscLogEventRegister(" dynamic IP-Jac", DM_CLASSID, &ctx->events[1]);CHKERRQ(ierr); /* 1 */ 1536 ierr = PetscLogEventRegister(" Kernel-init", DM_CLASSID, &ctx->events[3]);CHKERRQ(ierr); /* 3 */ 1537 ierr = PetscLogEventRegister(" Jac-f-df (GPU)", DM_CLASSID, &ctx->events[8]);CHKERRQ(ierr); /* 8 */ 1538 ierr = PetscLogEventRegister(" J Kernel (GPU)", DM_CLASSID, &ctx->events[4]);CHKERRQ(ierr); /* 4 */ 1539 ierr = PetscLogEventRegister(" M Kernel (GPU)", DM_CLASSID, &ctx->events[16]);CHKERRQ(ierr); /* 16 */ 1540 ierr = PetscLogEventRegister(" Copy to CPU", DM_CLASSID, &ctx->events[5]);CHKERRQ(ierr); /* 5 */ 1541 ierr = PetscLogEventRegister(" CPU assemble", DM_CLASSID, &ctx->events[6]);CHKERRQ(ierr); /* 6 */ 1542 1543 if (rank) { /* turn off output stuff for duplicate runs - do we need to add the prefix to all this? */ 1544 ierr = PetscOptionsClearValue(NULL,"-snes_converged_reason");CHKERRQ(ierr); 1545 ierr = PetscOptionsClearValue(NULL,"-ksp_converged_reason");CHKERRQ(ierr); 1546 ierr = PetscOptionsClearValue(NULL,"-snes_monitor");CHKERRQ(ierr); 1547 ierr = PetscOptionsClearValue(NULL,"-ksp_monitor");CHKERRQ(ierr); 1548 ierr = PetscOptionsClearValue(NULL,"-ts_monitor");CHKERRQ(ierr); 1549 ierr = PetscOptionsClearValue(NULL,"-ts_view");CHKERRQ(ierr); 1550 ierr = PetscOptionsClearValue(NULL,"-ts_adapt_monitor");CHKERRQ(ierr); 1551 ierr = PetscOptionsClearValue(NULL,"-dm_landau_amr_dm_view");CHKERRQ(ierr); 1552 ierr = PetscOptionsClearValue(NULL,"-dm_landau_amr_vec_view");CHKERRQ(ierr); 1553 ierr = PetscOptionsClearValue(NULL,"-dm_landau_mass_dm_view");CHKERRQ(ierr); 1554 ierr = PetscOptionsClearValue(NULL,"-dm_landau_mass_view");CHKERRQ(ierr); 1555 ierr = PetscOptionsClearValue(NULL,"-dm_landau_jacobian_view");CHKERRQ(ierr); 1556 ierr = PetscOptionsClearValue(NULL,"-dm_landau_mat_view");CHKERRQ(ierr); 1557 ierr = PetscOptionsClearValue(NULL,"-");CHKERRQ(ierr); 1558 ierr = PetscOptionsClearValue(NULL,"-info");CHKERRQ(ierr); 1559 } 1560 } 1561 PetscFunctionReturn(0); 1562 } 1563 1564 static PetscErrorCode CreateStaticGPUData(PetscInt dim, IS grid_batch_is_inv[], LandauCtx *ctx) 1565 { 1566 PetscErrorCode ierr; 1567 PetscSection section[LANDAU_MAX_GRIDS],globsection[LANDAU_MAX_GRIDS]; 1568 PetscQuadrature quad; 1569 const PetscReal *quadWeights; 1570 PetscInt numCells[LANDAU_MAX_GRIDS],Nq,Nf[LANDAU_MAX_GRIDS], ncellsTot=0; 1571 PetscTabulation *Tf; 1572 PetscDS prob; 1573 1574 PetscFunctionBegin; 1575 ierr = DMGetDS(ctx->plex[0], &prob);CHKERRQ(ierr); // same DS for all grids 1576 ierr = PetscDSGetTabulation(prob, &Tf);CHKERRQ(ierr); // Bf, &Df same for all grids 1577 /* DS, Tab and quad is same on all grids */ 1578 PetscCheck(ctx->plex[0],ctx->comm,PETSC_ERR_ARG_WRONG,"Plex not created"); 1579 ierr = PetscFEGetQuadrature(ctx->fe[0], &quad);CHKERRQ(ierr); 1580 ierr = PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, &quadWeights);CHKERRQ(ierr); 1581 PetscCheck(Nq <= LANDAU_MAX_NQ,ctx->comm,PETSC_ERR_ARG_WRONG,"Order too high. Nq = %" PetscInt_FMT " > LANDAU_MAX_NQ (%" PetscInt_FMT ")",Nq,LANDAU_MAX_NQ); 1582 /* setup each grid */ 1583 for (PetscInt grid=0;grid<ctx->num_grids;grid++) { 1584 PetscInt cStart, cEnd; 1585 PetscCheckFalse(ctx->plex[grid] == NULL,ctx->comm,PETSC_ERR_ARG_WRONG,"Plex not created"); 1586 ierr = DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd);CHKERRQ(ierr); 1587 numCells[grid] = cEnd - cStart; // grids can have different topology 1588 ierr = DMGetLocalSection(ctx->plex[grid], §ion[grid]);CHKERRQ(ierr); 1589 ierr = DMGetGlobalSection(ctx->plex[grid], &globsection[grid]);CHKERRQ(ierr); 1590 ierr = PetscSectionGetNumFields(section[grid], &Nf[grid]);CHKERRQ(ierr); 1591 ncellsTot += numCells[grid]; 1592 } 1593 #define MAP_BF_SIZE (64*LANDAU_DIM*LANDAU_DIM*LANDAU_MAX_Q_FACE*LANDAU_MAX_SPECIES) 1594 /* create GPU assembly data */ 1595 if (ctx->gpu_assembly) { /* we need GPU object with GPU assembly */ 1596 PetscContainer container; 1597 PetscScalar elemMatrix[LANDAU_MAX_NQ*LANDAU_MAX_NQ*LANDAU_MAX_SPECIES*LANDAU_MAX_SPECIES], *elMat; 1598 pointInterpolationP4est pointMaps[MAP_BF_SIZE][LANDAU_MAX_Q_FACE]; 1599 P4estVertexMaps *maps; 1600 const PetscInt *plex_batch=NULL,Nb=Nq; // tensor elements; 1601 1602 /* create GPU asssembly data */ 1603 ierr = PetscInfo(ctx->plex[0], "Make GPU maps %d\n",1);CHKERRQ(ierr); 1604 ierr = PetscLogEventBegin(ctx->events[2],0,0,0,0);CHKERRQ(ierr); 1605 ierr = PetscMalloc(sizeof(*maps)*ctx->num_grids, &maps);CHKERRQ(ierr); 1606 1607 if (ctx->use_coo_assembly) { // setup COO assembly -- put COO metadata directly in ctx->SData_d 1608 ierr = PetscMalloc3(ncellsTot+1,&ctx->SData_d.coo_elem_offsets,ncellsTot,&ctx->SData_d.coo_elem_fullNb,ncellsTot, &ctx->SData_d.coo_elem_point_offsets);CHKERRQ(ierr); // array of integer pointers 1609 ctx->SData_d.coo_elem_offsets[0] = 0; // finish later 1610 ierr = PetscInfo(ctx->plex[0], "COO initialization, %" PetscInt_FMT " cells\n",ncellsTot);CHKERRQ(ierr); 1611 ctx->SData_d.n_coo_cellsTot = ncellsTot; 1612 } else { 1613 ctx->SData_d.coo_elem_offsets = ctx->SData_d.coo_elem_fullNb = NULL; 1614 ctx->SData_d.coo_elem_point_offsets = NULL; 1615 ctx->SData_d.n_coo_cellsTot = 0; 1616 } 1617 1618 for (PetscInt grid=0,glb_elem_idx=0;grid<ctx->num_grids;grid++) { 1619 PetscInt cStart, cEnd, Nfloc = Nf[grid], totDim = Nfloc*Nq; 1620 if (grid_batch_is_inv[grid]) { 1621 ierr = ISGetIndices(grid_batch_is_inv[grid], &plex_batch);CHKERRQ(ierr); 1622 } 1623 ierr = DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd);CHKERRQ(ierr); 1624 // make maps 1625 maps[grid].d_self = NULL; 1626 maps[grid].num_elements = numCells[grid]; 1627 maps[grid].num_face = (PetscInt)(pow(Nq,1./((double)dim))+.001); // Q 1628 maps[grid].num_face = (PetscInt)(pow(maps[grid].num_face,(double)(dim-1))+.001); // Q^2 1629 maps[grid].num_reduced = 0; 1630 maps[grid].deviceType = ctx->deviceType; 1631 maps[grid].numgrids = ctx->num_grids; 1632 // count reduced and get 1633 ierr = PetscMalloc(maps[grid].num_elements * sizeof(*maps[grid].gIdx), &maps[grid].gIdx);CHKERRQ(ierr); 1634 for (int ej = cStart, eidx = 0 ; ej < cEnd; ++ej, ++eidx, glb_elem_idx++) { 1635 if (ctx->SData_d.coo_elem_offsets) ctx->SData_d.coo_elem_offsets[glb_elem_idx+1] = ctx->SData_d.coo_elem_offsets[glb_elem_idx]; // start with last one, then add 1636 for (int fieldA=0;fieldA<Nf[grid];fieldA++) { 1637 int fullNb = 0; 1638 for (int q = 0; q < Nb; ++q) { 1639 PetscInt numindices,*indices; 1640 PetscScalar *valuesOrig = elMat = elemMatrix; 1641 ierr = PetscArrayzero(elMat, totDim*totDim);CHKERRQ(ierr); 1642 elMat[ (fieldA*Nb + q)*totDim + fieldA*Nb + q] = 1; 1643 ierr = DMPlexGetClosureIndices(ctx->plex[grid], section[grid], globsection[grid], ej, PETSC_TRUE, &numindices, &indices, NULL, (PetscScalar **) &elMat);CHKERRQ(ierr); 1644 for (PetscInt f = 0 ; f < numindices ; ++f) { // look for a non-zero on the diagonal 1645 if (PetscAbs(PetscRealPart(elMat[f*numindices + f])) > PETSC_MACHINE_EPSILON) { 1646 // found it 1647 if (PetscAbs(PetscRealPart(elMat[f*numindices + f] - 1.)) < PETSC_MACHINE_EPSILON) { // normal vertex 1.0 1648 if (plex_batch) { 1649 maps[grid].gIdx[eidx][fieldA][q] = (LandauIdx) plex_batch[indices[f]]; 1650 } else { 1651 maps[grid].gIdx[eidx][fieldA][q] = (LandauIdx)indices[f]; 1652 } 1653 fullNb++; 1654 } else { //found a constraint 1655 int jj = 0; 1656 PetscReal sum = 0; 1657 const PetscInt ff = f; 1658 maps[grid].gIdx[eidx][fieldA][q] = -maps[grid].num_reduced - 1; // store (-)index: id = -(idx+1): idx = -id - 1 1659 1660 do { // constraints are continuous in Plex - exploit that here 1661 int ii; // get 'scale' 1662 for (ii = 0, pointMaps[maps[grid].num_reduced][jj].scale = 0; ii < maps[grid].num_face; ii++) { // sum row of outer product to recover vector value 1663 if (ff + ii < numindices) { // 3D has Q and Q^2 interps so might run off end. We could test that elMat[f*numindices + ff + ii] > 0, and break if not 1664 pointMaps[maps[grid].num_reduced][jj].scale += PetscRealPart(elMat[f*numindices + ff + ii]); 1665 } 1666 } 1667 sum += pointMaps[maps[grid].num_reduced][jj].scale; // diagnostic 1668 // get 'gid' 1669 if (pointMaps[maps[grid].num_reduced][jj].scale == 0) pointMaps[maps[grid].num_reduced][jj].gid = -1; // 3D has Q and Q^2 interps 1670 else { 1671 if (plex_batch) { 1672 pointMaps[maps[grid].num_reduced][jj].gid = plex_batch[indices[f]]; 1673 } else { 1674 pointMaps[maps[grid].num_reduced][jj].gid = indices[f]; 1675 } 1676 fullNb++; 1677 } 1678 } while (++jj < maps[grid].num_face && ++f < numindices); // jj is incremented if we hit the end 1679 while (jj < maps[grid].num_face) { 1680 pointMaps[maps[grid].num_reduced][jj].scale = 0; 1681 pointMaps[maps[grid].num_reduced][jj].gid = -1; 1682 jj++; 1683 } 1684 if (PetscAbs(sum-1.0) > 10*PETSC_MACHINE_EPSILON) { // debug 1685 int d,f; 1686 PetscReal tmp = 0; 1687 ierr = PetscPrintf(PETSC_COMM_SELF,"\t\t%" PetscInt_FMT ".%" PetscInt_FMT ".%" PetscInt_FMT ") ERROR total I = %22.16e (LANDAU_MAX_Q_FACE=%d, #face=%" PetscInt_FMT ")\n",eidx,q,fieldA,sum,LANDAU_MAX_Q_FACE,maps[grid].num_face);CHKERRQ(ierr); 1688 for (d = 0, tmp = 0; d < numindices; ++d) { 1689 if (tmp!=0 && PetscAbs(tmp-1.0) > 10*PETSC_MACHINE_EPSILON) {ierr = PetscPrintf(PETSC_COMM_WORLD,"%3" PetscInt_FMT ") %3" PetscInt_FMT ": ",d,indices[d]);CHKERRQ(ierr);} 1690 for (f = 0; f < numindices; ++f) { 1691 tmp += PetscRealPart(elMat[d*numindices + f]); 1692 } 1693 if (tmp!=0) {ierr = PetscPrintf(ctx->comm," | %22.16e\n",tmp);CHKERRQ(ierr);} 1694 } 1695 } 1696 maps[grid].num_reduced++; 1697 PetscCheckFalse(maps[grid].num_reduced>=MAP_BF_SIZE,PETSC_COMM_SELF, PETSC_ERR_PLIB, "maps[grid].num_reduced %d > %d",maps[grid].num_reduced,MAP_BF_SIZE); 1698 } 1699 break; 1700 } 1701 } 1702 // cleanup 1703 ierr = DMPlexRestoreClosureIndices(ctx->plex[grid], section[grid], globsection[grid], ej, PETSC_TRUE, &numindices, &indices, NULL, (PetscScalar **) &elMat);CHKERRQ(ierr); 1704 if (elMat != valuesOrig) {ierr = DMRestoreWorkArray(ctx->plex[grid], numindices*numindices, MPIU_SCALAR, &elMat);CHKERRQ(ierr);} 1705 } 1706 if (ctx->use_coo_assembly) { // setup COO assembly 1707 ctx->SData_d.coo_elem_offsets[glb_elem_idx+1] += fullNb*fullNb; // one species block, adds a block for each species, on this element in this grid 1708 if (fieldA==0) { // cache full Nb for this element, on this grid per species 1709 ctx->SData_d.coo_elem_fullNb[glb_elem_idx] = fullNb; 1710 } else PetscCheck(ctx->SData_d.coo_elem_fullNb[glb_elem_idx] == fullNb,PETSC_COMM_SELF, PETSC_ERR_PLIB, "full element size change with species %" PetscInt_FMT " %" PetscInt_FMT,ctx->SData_d.coo_elem_fullNb[glb_elem_idx],fullNb); 1711 } 1712 } // field 1713 } // cell 1714 // allocate and copy point data maps[grid].gIdx[eidx][field][q] 1715 ierr = PetscMalloc(maps[grid].num_reduced * sizeof(*maps[grid].c_maps), &maps[grid].c_maps);CHKERRQ(ierr); 1716 for (int ej = 0; ej < maps[grid].num_reduced; ++ej) { 1717 for (int q = 0; q < maps[grid].num_face; ++q) { 1718 maps[grid].c_maps[ej][q].scale = pointMaps[ej][q].scale; 1719 maps[grid].c_maps[ej][q].gid = pointMaps[ej][q].gid; 1720 } 1721 } 1722 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 1723 if (ctx->deviceType == LANDAU_KOKKOS) { 1724 ierr = LandauKokkosCreateMatMaps(maps, pointMaps, Nf, Nq, grid);CHKERRQ(ierr); // imples Kokkos does 1725 } // else could be CUDA 1726 #endif 1727 #if defined(PETSC_HAVE_CUDA) 1728 if (ctx->deviceType == LANDAU_CUDA) { 1729 ierr = LandauCUDACreateMatMaps(maps, pointMaps, Nf, Nq, grid);CHKERRQ(ierr); 1730 } 1731 #endif 1732 if (plex_batch) { 1733 ierr = ISRestoreIndices(grid_batch_is_inv[grid], &plex_batch);CHKERRQ(ierr); 1734 ierr = ISDestroy(&grid_batch_is_inv[grid]);CHKERRQ(ierr); // we are done with this 1735 } 1736 } /* grids */ 1737 // finish COO 1738 if (ctx->use_coo_assembly) { // setup COO assembly 1739 PetscInt *oor, *ooc; 1740 ctx->SData_d.coo_size = ctx->SData_d.coo_elem_offsets[ncellsTot]*ctx->batch_sz; 1741 ierr = PetscMalloc2(ctx->SData_d.coo_size,&oor,ctx->SData_d.coo_size,&ooc);CHKERRQ(ierr); 1742 for (int i=0;i<ctx->SData_d.coo_size;i++) oor[i] = ooc[i] = -1; 1743 // get 1744 for (int grid=0,glb_elem_idx=0;grid<ctx->num_grids;grid++) { 1745 for (int ej = 0 ; ej < numCells[grid] ; ++ej, glb_elem_idx++) { 1746 const int fullNb = ctx->SData_d.coo_elem_fullNb[glb_elem_idx]; 1747 const LandauIdx *const Idxs = &maps[grid].gIdx[ej][0][0]; // just use field-0 maps, They should be the same but this is just for COO storage 1748 ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][0] = 0; 1749 for (int f=0, cnt2=0;f<Nb;f++) { 1750 int idx = Idxs[f]; 1751 ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][f+1] = ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][f]; // start at last 1752 if (idx >= 0) { 1753 cnt2++; 1754 ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][f+1]++; // inc 1755 } else { 1756 idx = -idx - 1; 1757 for (int q = 0 ; q < maps[grid].num_face; q++) { 1758 if (maps[grid].c_maps[idx][q].gid < 0) break; 1759 cnt2++; 1760 ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][f+1]++; // inc 1761 } 1762 } 1763 PetscCheck(cnt2 <= fullNb,PETSC_COMM_SELF, PETSC_ERR_PLIB, "wrong count %d < %d",fullNb,cnt2); 1764 } 1765 PetscCheck(ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][Nb]==fullNb,PETSC_COMM_SELF, PETSC_ERR_PLIB, "coo_elem_point_offsets size %d != fullNb=%d",ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][Nb],fullNb); 1766 } 1767 } 1768 // set 1769 for (PetscInt b_id = 0 ; b_id < ctx->batch_sz ; b_id++) { 1770 for (int grid=0,glb_elem_idx=0;grid<ctx->num_grids;grid++) { 1771 const PetscInt moffset = LAND_MOFFSET(b_id,grid,ctx->batch_sz,ctx->num_grids,ctx->mat_offset); 1772 for (int ej = 0 ; ej < numCells[grid] ; ++ej, glb_elem_idx++) { 1773 const int fullNb = ctx->SData_d.coo_elem_fullNb[glb_elem_idx],fullNb2=fullNb*fullNb; 1774 // set (i,j) 1775 for (int fieldA=0;fieldA<Nf[grid];fieldA++) { 1776 const LandauIdx *const Idxs = &maps[grid].gIdx[ej][fieldA][0]; 1777 int rows[LANDAU_MAX_Q_FACE],cols[LANDAU_MAX_Q_FACE]; 1778 for (int f = 0; f < Nb; ++f) { 1779 const int nr = ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][f+1] - ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][f]; 1780 if (nr==1) rows[0] = Idxs[f]; 1781 else { 1782 const int idx = -Idxs[f] - 1; 1783 for (int q = 0; q < nr; q++) { 1784 rows[q] = maps[grid].c_maps[idx][q].gid; 1785 } 1786 } 1787 for (int g = 0; g < Nb; ++g) { 1788 const int nc = ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][g+1] - ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][g]; 1789 if (nc==1) cols[0] = Idxs[g]; 1790 else { 1791 const int idx = -Idxs[g] - 1; 1792 for (int q = 0; q < nc; q++) { 1793 cols[q] = maps[grid].c_maps[idx][q].gid; 1794 } 1795 } 1796 const int idx0 = b_id*ctx->SData_d.coo_elem_offsets[ncellsTot] + ctx->SData_d.coo_elem_offsets[glb_elem_idx] + fieldA*fullNb2 + fullNb * ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][f] + nr * ctx->SData_d.coo_elem_point_offsets[glb_elem_idx][g]; 1797 for (int q = 0, idx = idx0; q < nr; q++) { 1798 for (int d = 0; d < nc; d++, idx++) { 1799 oor[idx] = rows[q] + moffset; 1800 ooc[idx] = cols[d] + moffset; 1801 } 1802 } 1803 } 1804 } 1805 } 1806 } // cell 1807 } // grid 1808 } // batch 1809 ierr = MatSetPreallocationCOO(ctx->J,ctx->SData_d.coo_size,oor,ooc);CHKERRQ(ierr); 1810 ierr = PetscFree2(oor,ooc);CHKERRQ(ierr); 1811 } 1812 ierr = PetscContainerCreate(PETSC_COMM_SELF, &container);CHKERRQ(ierr); 1813 ierr = PetscContainerSetPointer(container, (void *)maps);CHKERRQ(ierr); 1814 ierr = PetscContainerSetUserDestroy(container, LandauGPUMapsDestroy);CHKERRQ(ierr); 1815 ierr = PetscObjectCompose((PetscObject) ctx->J, "assembly_maps", (PetscObject) container);CHKERRQ(ierr); 1816 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 1817 ierr = PetscLogEventEnd(ctx->events[2],0,0,0,0);CHKERRQ(ierr); 1818 } // end GPU assembly 1819 { /* create static point data, Jacobian called first, only one vertex copy */ 1820 PetscReal *invJe,*ww,*xx,*yy,*zz=NULL,*invJ_a; 1821 PetscInt outer_ipidx, outer_ej,grid, nip_glb = 0; 1822 PetscFE fe; 1823 const PetscInt Nb = Nq; 1824 ierr = PetscLogEventBegin(ctx->events[7],0,0,0,0);CHKERRQ(ierr); 1825 ierr = PetscInfo(ctx->plex[0], "Initialize static data\n");CHKERRQ(ierr); 1826 for (PetscInt grid=0;grid<ctx->num_grids;grid++) nip_glb += Nq*numCells[grid]; 1827 /* collect f data, first time is for Jacobian, but make mass now */ 1828 if (ctx->verbose > 0) { 1829 PetscInt ncells = 0, N; 1830 ierr = MatGetSize(ctx->J,&N,NULL);CHKERRQ(ierr); 1831 for (PetscInt grid=0;grid<ctx->num_grids;grid++) ncells += numCells[grid]; 1832 ierr = PetscPrintf(ctx->comm,"%" PetscInt_FMT ") %s %" PetscInt_FMT " IPs, %" PetscInt_FMT " cells total, Nb=%" PetscInt_FMT ", Nq=%" PetscInt_FMT ", dim=%" PetscInt_FMT ", Tab: Nb=%" PetscInt_FMT " Nf=%" PetscInt_FMT " Np=%" PetscInt_FMT " cdim=%" PetscInt_FMT " N=%" PetscInt_FMT "\n", 1833 0,"FormLandau",nip_glb,ncells, Nb, Nq, dim, Nb, ctx->num_species, Nb, dim, N);CHKERRQ(ierr); 1834 } 1835 ierr = PetscMalloc4(nip_glb,&ww,nip_glb,&xx,nip_glb,&yy,nip_glb*dim*dim,&invJ_a);CHKERRQ(ierr); 1836 if (dim==3) { 1837 ierr = PetscMalloc1(nip_glb,&zz);CHKERRQ(ierr); 1838 } 1839 if (ctx->use_energy_tensor_trick) { 1840 ierr = PetscFECreateDefault(PETSC_COMM_SELF, dim, 1, PETSC_FALSE, NULL, PETSC_DECIDE, &fe);CHKERRQ(ierr); 1841 ierr = PetscObjectSetName((PetscObject) fe, "energy");CHKERRQ(ierr); 1842 } 1843 /* init each grids static data - no batch */ 1844 for (grid=0, outer_ipidx=0, outer_ej=0 ; grid < ctx->num_grids ; grid++) { // OpenMP (once) 1845 Vec v2_2 = NULL; // projected function: v^2/2 for non-relativistic, gamma... for relativistic 1846 PetscSection e_section; 1847 DM dmEnergy; 1848 PetscInt cStart, cEnd, ej; 1849 1850 ierr = DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd);CHKERRQ(ierr); 1851 // prep energy trick, get v^2 / 2 vector 1852 if (ctx->use_energy_tensor_trick) { 1853 PetscErrorCode (*energyf[1])(PetscInt, PetscReal, const PetscReal [], PetscInt, PetscScalar [], void *) = {ctx->use_relativistic_corrections ? gamma_m1_f : energy_f}; 1854 Vec glob_v2; 1855 PetscReal *c2_0[1], data[1] = {PetscSqr(C_0(ctx->v_0))}; 1856 1857 ierr = DMClone(ctx->plex[grid], &dmEnergy);CHKERRQ(ierr); 1858 ierr = PetscObjectSetName((PetscObject) dmEnergy, "energy");CHKERRQ(ierr); 1859 ierr = DMSetField(dmEnergy, 0, NULL, (PetscObject)fe);CHKERRQ(ierr); 1860 ierr = DMCreateDS(dmEnergy);CHKERRQ(ierr); 1861 ierr = DMGetSection(dmEnergy, &e_section);CHKERRQ(ierr); 1862 ierr = DMGetGlobalVector(dmEnergy,&glob_v2);CHKERRQ(ierr); 1863 ierr = PetscObjectSetName((PetscObject) glob_v2, "trick");CHKERRQ(ierr); 1864 c2_0[0] = &data[0]; 1865 ierr = DMProjectFunction(dmEnergy, 0., energyf, (void**)c2_0, INSERT_ALL_VALUES, glob_v2);CHKERRQ(ierr); 1866 ierr = DMGetLocalVector(dmEnergy, &v2_2);CHKERRQ(ierr); 1867 ierr = VecZeroEntries(v2_2);CHKERRQ(ierr); /* zero BCs so don't set */ 1868 ierr = DMGlobalToLocalBegin(dmEnergy, glob_v2, INSERT_VALUES, v2_2);CHKERRQ(ierr); 1869 ierr = DMGlobalToLocalEnd (dmEnergy, glob_v2, INSERT_VALUES, v2_2);CHKERRQ(ierr); 1870 ierr = DMViewFromOptions(dmEnergy,NULL, "-energy_dm_view");CHKERRQ(ierr); 1871 ierr = VecViewFromOptions(glob_v2,NULL, "-energy_vec_view");CHKERRQ(ierr); 1872 ierr = DMRestoreGlobalVector(dmEnergy, &glob_v2);CHKERRQ(ierr); 1873 } 1874 /* append part of the IP data for each grid */ 1875 for (ej = 0 ; ej < numCells[grid]; ++ej, ++outer_ej) { 1876 PetscScalar *coefs = NULL; 1877 PetscReal vj[LANDAU_MAX_NQ*LANDAU_DIM],detJj[LANDAU_MAX_NQ], Jdummy[LANDAU_MAX_NQ*LANDAU_DIM*LANDAU_DIM], c0 = C_0(ctx->v_0), c02 = PetscSqr(c0); 1878 invJe = invJ_a + outer_ej*Nq*dim*dim; 1879 ierr = DMPlexComputeCellGeometryFEM(ctx->plex[grid], ej+cStart, quad, vj, Jdummy, invJe, detJj);CHKERRQ(ierr); 1880 if (ctx->use_energy_tensor_trick) { 1881 ierr = DMPlexVecGetClosure(dmEnergy, e_section, v2_2, ej+cStart, NULL, &coefs);CHKERRQ(ierr); 1882 } 1883 /* create static point data */ 1884 for (PetscInt qj = 0; qj < Nq; qj++, outer_ipidx++) { 1885 const PetscInt gidx = outer_ipidx; 1886 const PetscReal *invJ = &invJe[qj*dim*dim]; 1887 ww [gidx] = detJj[qj] * quadWeights[qj]; 1888 if (dim==2) ww [gidx] *= vj[qj * dim + 0]; /* cylindrical coordinate, w/o 2pi */ 1889 // get xx, yy, zz 1890 if (ctx->use_energy_tensor_trick) { 1891 double refSpaceDer[3],eGradPhi[3]; 1892 const PetscReal * const DD = Tf[0]->T[1]; 1893 const PetscReal *Dq = &DD[qj*Nb*dim]; 1894 for (int d = 0; d < 3; ++d) refSpaceDer[d] = eGradPhi[d] = 0.0; 1895 for (int b = 0; b < Nb; ++b) { 1896 for (int d = 0; d < dim; ++d) refSpaceDer[d] += Dq[b*dim+d]*PetscRealPart(coefs[b]); 1897 } 1898 xx[gidx] = 1e10; 1899 if (ctx->use_relativistic_corrections) { 1900 double dg2_c2 = 0; 1901 //for (int d = 0; d < dim; ++d) refSpaceDer[d] *= c02; 1902 for (int d = 0; d < dim; ++d) dg2_c2 += PetscSqr(refSpaceDer[d]); 1903 dg2_c2 *= (double)c02; 1904 if (dg2_c2 >= .999) { 1905 xx[gidx] = vj[qj * dim + 0]; /* coordinate */ 1906 yy[gidx] = vj[qj * dim + 1]; 1907 if (dim==3) zz[gidx] = vj[qj * dim + 2]; 1908 ierr = PetscPrintf(ctx->comm,"Error: %12.5e %" PetscInt_FMT ".%" PetscInt_FMT ") dg2/c02 = %12.5e x= %12.5e %12.5e %12.5e\n",PetscSqrtReal(xx[gidx]*xx[gidx] + yy[gidx]*yy[gidx] + zz[gidx]*zz[gidx]), ej, qj, dg2_c2, xx[gidx],yy[gidx],zz[gidx]);CHKERRQ(ierr); 1909 } else { 1910 PetscReal fact = c02/PetscSqrtReal(1. - dg2_c2); 1911 for (int d = 0; d < dim; ++d) refSpaceDer[d] *= fact; 1912 // could test with other point u' that (grad - grad') * U (refSpaceDer, refSpaceDer') == 0 1913 } 1914 } 1915 if (xx[gidx] == 1e10) { 1916 for (int d = 0; d < dim; ++d) { 1917 for (int e = 0 ; e < dim; ++e) { 1918 eGradPhi[d] += invJ[e*dim+d]*refSpaceDer[e]; 1919 } 1920 } 1921 xx[gidx] = eGradPhi[0]; 1922 yy[gidx] = eGradPhi[1]; 1923 if (dim==3) zz[gidx] = eGradPhi[2]; 1924 } 1925 } else { 1926 xx[gidx] = vj[qj * dim + 0]; /* coordinate */ 1927 yy[gidx] = vj[qj * dim + 1]; 1928 if (dim==3) zz[gidx] = vj[qj * dim + 2]; 1929 } 1930 } /* q */ 1931 if (ctx->use_energy_tensor_trick) { 1932 ierr = DMPlexVecRestoreClosure(dmEnergy, e_section, v2_2, ej+cStart, NULL, &coefs);CHKERRQ(ierr); 1933 } 1934 } /* ej */ 1935 if (ctx->use_energy_tensor_trick) { 1936 ierr = DMRestoreLocalVector(dmEnergy, &v2_2);CHKERRQ(ierr); 1937 ierr = DMDestroy(&dmEnergy);CHKERRQ(ierr); 1938 } 1939 } /* grid */ 1940 if (ctx->use_energy_tensor_trick) { 1941 ierr = PetscFEDestroy(&fe);CHKERRQ(ierr); 1942 } 1943 /* cache static data */ 1944 if (ctx->deviceType == LANDAU_CUDA || ctx->deviceType == LANDAU_KOKKOS) { 1945 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_KOKKOS_KERNELS) 1946 PetscReal invMass[LANDAU_MAX_SPECIES],nu_alpha[LANDAU_MAX_SPECIES], nu_beta[LANDAU_MAX_SPECIES]; 1947 for (PetscInt grid = 0; grid < ctx->num_grids ; grid++) { 1948 for (PetscInt ii=ctx->species_offset[grid];ii<ctx->species_offset[grid+1];ii++) { 1949 invMass[ii] = ctx->m_0/ctx->masses[ii]; 1950 nu_alpha[ii] = PetscSqr(ctx->charges[ii]/ctx->m_0)*ctx->m_0/ctx->masses[ii]; 1951 nu_beta[ii] = PetscSqr(ctx->charges[ii]/ctx->epsilon0)*ctx->lnLam / (8*PETSC_PI) * ctx->t_0*ctx->n_0/PetscPowReal(ctx->v_0,3); 1952 } 1953 } 1954 if (ctx->deviceType == LANDAU_CUDA) { 1955 #if defined(PETSC_HAVE_CUDA) 1956 ierr = LandauCUDAStaticDataSet(ctx->plex[0], Nq, ctx->batch_sz, ctx->num_grids, numCells, ctx->species_offset, ctx->mat_offset, 1957 nu_alpha, nu_beta, invMass, invJ_a, xx, yy, zz, ww, &ctx->SData_d);CHKERRQ(ierr); 1958 #else 1959 SETERRQ(ctx->comm,PETSC_ERR_ARG_WRONG,"-landau_device_type %s not built","cuda"); 1960 #endif 1961 } else if (ctx->deviceType == LANDAU_KOKKOS) { 1962 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 1963 ierr = LandauKokkosStaticDataSet(ctx->plex[0], Nq, ctx->batch_sz, ctx->num_grids, numCells, ctx->species_offset, ctx->mat_offset, 1964 nu_alpha, nu_beta, invMass,invJ_a,xx,yy,zz,ww,&ctx->SData_d);CHKERRQ(ierr); 1965 #else 1966 SETERRQ(ctx->comm,PETSC_ERR_ARG_WRONG,"-landau_device_type %s not built","kokkos"); 1967 #endif 1968 } 1969 #endif 1970 /* free */ 1971 ierr = PetscFree4(ww,xx,yy,invJ_a);CHKERRQ(ierr); 1972 if (dim==3) { 1973 ierr = PetscFree(zz);CHKERRQ(ierr); 1974 } 1975 } else { /* CPU version, just copy in, only use part */ 1976 ctx->SData_d.w = (void*)ww; 1977 ctx->SData_d.x = (void*)xx; 1978 ctx->SData_d.y = (void*)yy; 1979 ctx->SData_d.z = (void*)zz; 1980 ctx->SData_d.invJ = (void*)invJ_a; 1981 } 1982 ierr = PetscLogEventEnd(ctx->events[7],0,0,0,0);CHKERRQ(ierr); 1983 } // initialize 1984 PetscFunctionReturn(0); 1985 } 1986 1987 /* < v, u > */ 1988 static void g0_1(PetscInt dim, PetscInt Nf, PetscInt NfAux, 1989 const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], 1990 const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], 1991 PetscReal t, PetscReal u_tShift, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar g0[]) 1992 { 1993 g0[0] = 1.; 1994 } 1995 1996 /* < v, u > */ 1997 static void g0_fake(PetscInt dim, PetscInt Nf, PetscInt NfAux, 1998 const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], 1999 const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], 2000 PetscReal t, PetscReal u_tShift, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar g0[]) 2001 { 2002 static double ttt = 1; 2003 g0[0] = ttt++; 2004 } 2005 2006 /* < v, u > */ 2007 static void g0_r(PetscInt dim, PetscInt Nf, PetscInt NfAux, 2008 const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], 2009 const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], 2010 PetscReal t, PetscReal u_tShift, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar g0[]) 2011 { 2012 g0[0] = 2.*PETSC_PI*x[0]; 2013 } 2014 2015 static PetscErrorCode MatrixNfDestroy(void *ptr) 2016 { 2017 PetscInt *nf = (PetscInt *)ptr; 2018 PetscErrorCode ierr; 2019 PetscFunctionBegin; 2020 ierr = PetscFree(nf);CHKERRQ(ierr); 2021 PetscFunctionReturn(0); 2022 } 2023 2024 static PetscErrorCode LandauCreateMatrix(MPI_Comm comm, Vec X, IS grid_batch_is_inv[LANDAU_MAX_GRIDS], LandauCtx *ctx) 2025 { 2026 PetscErrorCode ierr; 2027 PetscInt *idxs=NULL; 2028 Mat subM[LANDAU_MAX_GRIDS]; 2029 2030 PetscFunctionBegin; 2031 if (!ctx->gpu_assembly) { /* we need GPU object with GPU assembly */ 2032 PetscFunctionReturn(0); 2033 } 2034 // get the RCM for this grid to separate out species into blocks -- create 'idxs' & 'ctx->batch_is' 2035 if (ctx->gpu_assembly && ctx->jacobian_field_major_order) { 2036 ierr = PetscMalloc1(ctx->mat_offset[ctx->num_grids]*ctx->batch_sz, &idxs);CHKERRQ(ierr); 2037 } 2038 for (PetscInt grid=0 ; grid < ctx->num_grids ; grid++) { 2039 const PetscInt *values, n = ctx->mat_offset[grid+1] - ctx->mat_offset[grid]; 2040 Mat gMat; 2041 DM massDM; 2042 PetscDS prob; 2043 Vec tvec; 2044 // get "mass" matrix for reordering 2045 ierr = DMClone(ctx->plex[grid], &massDM);CHKERRQ(ierr); 2046 ierr = DMCopyFields(ctx->plex[grid], massDM);CHKERRQ(ierr); 2047 ierr = DMCreateDS(massDM);CHKERRQ(ierr); 2048 ierr = DMGetDS(massDM, &prob);CHKERRQ(ierr); 2049 for (int ix=0, ii=ctx->species_offset[grid];ii<ctx->species_offset[grid+1];ii++,ix++) { 2050 ierr = PetscDSSetJacobian(prob, ix, ix, g0_fake, NULL, NULL, NULL);CHKERRQ(ierr); 2051 } 2052 ierr = PetscOptionsInsertString(NULL,"-dm_preallocate_only"); 2053 ierr = DMSetFromOptions(massDM);CHKERRQ(ierr); 2054 ierr = DMCreateMatrix(massDM, &gMat);CHKERRQ(ierr); 2055 ierr = PetscOptionsInsertString(NULL,"-dm_preallocate_only false"); 2056 ierr = MatSetOption(gMat,MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE);CHKERRQ(ierr); 2057 ierr = MatSetOption(gMat,MAT_IGNORE_ZERO_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 2058 ierr = DMCreateLocalVector(ctx->plex[grid],&tvec);CHKERRQ(ierr); 2059 ierr = DMPlexSNESComputeJacobianFEM(massDM, tvec, gMat, gMat, ctx);CHKERRQ(ierr); 2060 ierr = MatViewFromOptions(gMat, NULL, "-dm_landau_reorder_mat_view");CHKERRQ(ierr); 2061 ierr = DMDestroy(&massDM);CHKERRQ(ierr); 2062 ierr = VecDestroy(&tvec);CHKERRQ(ierr); 2063 subM[grid] = gMat; 2064 if (ctx->gpu_assembly && ctx->jacobian_field_major_order) { 2065 MatOrderingType rtype = MATORDERINGRCM; 2066 IS isrow,isicol; 2067 ierr = MatGetOrdering(gMat,rtype,&isrow,&isicol);CHKERRQ(ierr); 2068 ierr = ISInvertPermutation(isrow,PETSC_DECIDE,&grid_batch_is_inv[grid]);CHKERRQ(ierr); 2069 ierr = ISGetIndices(isrow, &values);CHKERRQ(ierr); 2070 for (PetscInt b_id=0 ; b_id < ctx->batch_sz ; b_id++) { // add batch size DMs for this species grid 2071 #if !defined(LANDAU_SPECIES_MAJOR) 2072 PetscInt N = ctx->mat_offset[ctx->num_grids], n0 = ctx->mat_offset[grid] + b_id*N; 2073 for (int ii = 0; ii < n; ++ii) idxs[n0+ii] = values[ii] + n0; 2074 #else 2075 PetscInt n0 = ctx->mat_offset[grid]*ctx->batch_sz + b_id*n; 2076 for (int ii = 0; ii < n; ++ii) idxs[n0+ii] = values[ii] + n0; 2077 #endif 2078 } 2079 ierr = ISRestoreIndices(isrow, &values);CHKERRQ(ierr); 2080 ierr = ISDestroy(&isrow);CHKERRQ(ierr); 2081 ierr = ISDestroy(&isicol);CHKERRQ(ierr); 2082 } 2083 } 2084 if (ctx->gpu_assembly && ctx->jacobian_field_major_order) { 2085 ierr = ISCreateGeneral(comm,ctx->mat_offset[ctx->num_grids]*ctx->batch_sz,idxs,PETSC_OWN_POINTER,&ctx->batch_is);CHKERRQ(ierr); 2086 } 2087 // get a block matrix 2088 for (PetscInt grid=0 ; grid<ctx->num_grids ; grid++) { 2089 Mat B = subM[grid]; 2090 PetscInt nloc, nzl, colbuf[1024], row; 2091 ierr = MatGetSize(B, &nloc, NULL);CHKERRQ(ierr); 2092 for (PetscInt b_id = 0 ; b_id < ctx->batch_sz ; b_id++) { 2093 const PetscInt moffset = LAND_MOFFSET(b_id,grid,ctx->batch_sz,ctx->num_grids,ctx->mat_offset); 2094 const PetscInt *cols; 2095 const PetscScalar *vals; 2096 for (int i=0 ; i<nloc ; i++) { 2097 ierr = MatGetRow(B,i,&nzl,&cols,&vals);CHKERRQ(ierr); 2098 PetscCheck(nzl<=1024,comm, PETSC_ERR_PLIB, "Row too big: %" PetscInt_FMT,nzl); 2099 for (int j=0; j<nzl; j++) colbuf[j] = cols[j] + moffset; 2100 row = i + moffset; 2101 ierr = MatSetValues(ctx->J,1,&row,nzl,colbuf,vals,INSERT_VALUES);CHKERRQ(ierr); 2102 ierr = MatRestoreRow(B,i,&nzl,&cols,&vals);CHKERRQ(ierr); 2103 } 2104 } 2105 } 2106 for (PetscInt grid=0 ; grid<ctx->num_grids ; grid++) { 2107 ierr = MatDestroy(&subM[grid]);CHKERRQ(ierr); 2108 } 2109 ierr = MatAssemblyBegin(ctx->J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2110 ierr = MatAssemblyEnd(ctx->J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2111 2112 if (ctx->gpu_assembly && ctx->jacobian_field_major_order) { 2113 Mat mat_block_order; 2114 ierr = MatCreateSubMatrix(ctx->J,ctx->batch_is,ctx->batch_is,MAT_INITIAL_MATRIX,&mat_block_order);CHKERRQ(ierr); // use MatPermute 2115 ierr = MatViewFromOptions(mat_block_order, NULL, "-dm_landau_field_major_mat_view");CHKERRQ(ierr); 2116 ierr = MatDestroy(&ctx->J);CHKERRQ(ierr); 2117 ctx->J = mat_block_order; 2118 // override ops to make KSP work in field major space 2119 ctx->seqaij_mult = mat_block_order->ops->mult; 2120 mat_block_order->ops->mult = LandauMatMult; 2121 mat_block_order->ops->multadd = LandauMatMultAdd; 2122 ctx->seqaij_solve = NULL; 2123 ctx->seqaij_getdiagonal = mat_block_order->ops->getdiagonal; 2124 mat_block_order->ops->getdiagonal = LandauMatGetDiagonal; 2125 ctx->seqaij_multtranspose = mat_block_order->ops->multtranspose; 2126 mat_block_order->ops->multtranspose = LandauMatMultTranspose; 2127 ierr = VecDuplicate(X,&ctx->work_vec);CHKERRQ(ierr); 2128 ierr = VecScatterCreate(X, ctx->batch_is, ctx->work_vec, NULL, &ctx->plex_batch);CHKERRQ(ierr); 2129 } 2130 2131 PetscFunctionReturn(0); 2132 } 2133 2134 PetscErrorCode LandauCreateMassMatrix(DM pack, Mat *Amat); 2135 /*@C 2136 LandauCreateVelocitySpace - Create a DMPlex velocity space mesh 2137 2138 Collective on comm 2139 2140 Input Parameters: 2141 + comm - The MPI communicator 2142 . dim - velocity space dimension (2 for axisymmetric, 3 for full 3X + 3V solver) 2143 - prefix - prefix for options (not tested) 2144 2145 Output Parameter: 2146 . pack - The DM object representing the mesh 2147 + X - A vector (user destroys) 2148 - J - Optional matrix (object destroys) 2149 2150 Level: beginner 2151 2152 .keywords: mesh 2153 .seealso: DMPlexCreate(), LandauDestroyVelocitySpace() 2154 @*/ 2155 PetscErrorCode LandauCreateVelocitySpace(MPI_Comm comm, PetscInt dim, const char prefix[], Vec *X, Mat *J, DM *pack) 2156 { 2157 PetscErrorCode ierr; 2158 LandauCtx *ctx; 2159 Vec Xsub[LANDAU_MAX_GRIDS]; 2160 IS grid_batch_is_inv[LANDAU_MAX_GRIDS]; 2161 2162 PetscFunctionBegin; 2163 PetscCheckFalse(dim!=2 && dim!=3,PETSC_COMM_SELF, PETSC_ERR_PLIB, "Only 2D and 3D supported"); 2164 PetscCheck(LANDAU_DIM == dim,PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " != LANDAU_DIM %d",dim,LANDAU_DIM); 2165 ierr = PetscNew(&ctx);CHKERRQ(ierr); 2166 ctx->comm = comm; /* used for diagnostics and global errors */ 2167 /* process options */ 2168 ierr = ProcessOptions(ctx,prefix);CHKERRQ(ierr); 2169 if (dim==2) ctx->use_relativistic_corrections = PETSC_FALSE; 2170 /* Create Mesh */ 2171 ierr = DMCompositeCreate(PETSC_COMM_SELF,pack);CHKERRQ(ierr); 2172 ierr = PetscLogEventBegin(ctx->events[13],0,0,0,0);CHKERRQ(ierr); 2173 ierr = PetscLogEventBegin(ctx->events[15],0,0,0,0);CHKERRQ(ierr); 2174 ierr = LandauDMCreateVMeshes(PETSC_COMM_SELF, dim, prefix, ctx, *pack);CHKERRQ(ierr); // creates grids (Forest of AMR) 2175 for (PetscInt grid=0;grid<ctx->num_grids;grid++) { 2176 /* create FEM */ 2177 ierr = SetupDS(ctx->plex[grid],dim,grid,ctx);CHKERRQ(ierr); 2178 /* set initial state */ 2179 ierr = DMCreateGlobalVector(ctx->plex[grid],&Xsub[grid]);CHKERRQ(ierr); 2180 ierr = PetscObjectSetName((PetscObject) Xsub[grid], "u_orig");CHKERRQ(ierr); 2181 /* initial static refinement, no solve */ 2182 ierr = LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, 0, ctx);CHKERRQ(ierr); 2183 /* forest refinement - forest goes in (if forest), plex comes out */ 2184 if (ctx->use_p4est) { 2185 DM plex; 2186 ierr = adapt(grid,ctx,&Xsub[grid]);CHKERRQ(ierr); // forest goes in, plex comes out 2187 ierr = DMViewFromOptions(ctx->plex[grid],NULL,"-dm_landau_amr_dm_view");CHKERRQ(ierr); // need to differentiate - todo 2188 ierr = VecViewFromOptions(Xsub[grid], NULL, "-dm_landau_amr_vec_view");CHKERRQ(ierr); 2189 // convert to plex, all done with this level 2190 ierr = DMConvert(ctx->plex[grid], DMPLEX, &plex);CHKERRQ(ierr); 2191 ierr = DMDestroy(&ctx->plex[grid]);CHKERRQ(ierr); 2192 ctx->plex[grid] = plex; 2193 } 2194 #if !defined(LANDAU_SPECIES_MAJOR) 2195 ierr = DMCompositeAddDM(*pack,ctx->plex[grid]);CHKERRQ(ierr); 2196 #else 2197 for (PetscInt b_id=0;b_id<ctx->batch_sz;b_id++) { // add batch size DMs for this species grid 2198 ierr = DMCompositeAddDM(*pack,ctx->plex[grid]);CHKERRQ(ierr); 2199 } 2200 #endif 2201 ierr = DMSetApplicationContext(ctx->plex[grid], ctx);CHKERRQ(ierr); 2202 } 2203 #if !defined(LANDAU_SPECIES_MAJOR) 2204 // stack the batched DMs, could do it all here!!! b_id=0 2205 for (PetscInt b_id=1;b_id<ctx->batch_sz;b_id++) { 2206 for (PetscInt grid=0;grid<ctx->num_grids;grid++) { 2207 ierr = DMCompositeAddDM(*pack,ctx->plex[grid]);CHKERRQ(ierr); 2208 } 2209 } 2210 #endif 2211 // create ctx->mat_offset 2212 ctx->mat_offset[0] = 0; 2213 for (PetscInt grid=0 ; grid < ctx->num_grids ; grid++) { 2214 PetscInt n; 2215 ierr = VecGetLocalSize(Xsub[grid],&n);CHKERRQ(ierr); 2216 ctx->mat_offset[grid+1] = ctx->mat_offset[grid] + n; 2217 } 2218 // creat DM & Jac 2219 ierr = DMSetApplicationContext(*pack, ctx);CHKERRQ(ierr); 2220 ierr = PetscOptionsInsertString(NULL,"-dm_preallocate_only"); 2221 ierr = DMSetFromOptions(*pack);CHKERRQ(ierr); 2222 ierr = DMCreateMatrix(*pack, &ctx->J);CHKERRQ(ierr); 2223 ierr = PetscOptionsInsertString(NULL,"-dm_preallocate_only false"); 2224 ierr = MatSetOption(ctx->J,MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE);CHKERRQ(ierr); 2225 ierr = MatSetOption(ctx->J,MAT_IGNORE_ZERO_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 2226 ierr = PetscObjectSetName((PetscObject)ctx->J, "Jac");CHKERRQ(ierr); 2227 // construct initial conditions in X 2228 ierr = DMCreateGlobalVector(*pack,X);CHKERRQ(ierr); 2229 for (PetscInt grid=0 ; grid < ctx->num_grids ; grid++) { 2230 PetscInt n; 2231 ierr = VecGetLocalSize(Xsub[grid],&n);CHKERRQ(ierr); 2232 for (PetscInt b_id = 0 ; b_id < ctx->batch_sz ; b_id++) { 2233 PetscScalar const *values; 2234 const PetscInt moffset = LAND_MOFFSET(b_id,grid,ctx->batch_sz,ctx->num_grids,ctx->mat_offset); 2235 ierr = LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, b_id, ctx);CHKERRQ(ierr); 2236 ierr = VecGetArrayRead(Xsub[grid],&values);CHKERRQ(ierr); 2237 for (int i=0, idx = moffset; i<n; i++, idx++) { 2238 ierr = VecSetValue(*X,idx,values[i],INSERT_VALUES);CHKERRQ(ierr); 2239 } 2240 ierr = VecRestoreArrayRead(Xsub[grid],&values);CHKERRQ(ierr); 2241 } 2242 } 2243 // cleanup 2244 for (PetscInt grid=0 ; grid < ctx->num_grids ; grid++) { 2245 ierr = VecDestroy(&Xsub[grid]);CHKERRQ(ierr); 2246 } 2247 /* check for correct matrix type */ 2248 if (ctx->gpu_assembly) { /* we need GPU object with GPU assembly */ 2249 PetscBool flg; 2250 if (ctx->deviceType == LANDAU_CUDA) { 2251 ierr = PetscObjectTypeCompareAny((PetscObject)ctx->J,&flg,MATSEQAIJCUSPARSE,MATMPIAIJCUSPARSE,MATAIJCUSPARSE,"");CHKERRQ(ierr); 2252 PetscCheckFalse(!flg,ctx->comm,PETSC_ERR_ARG_WRONG,"must use '-dm_mat_type aijcusparse -dm_vec_type cuda' for GPU assembly and Cuda or use '-dm_landau_device_type cpu'"); 2253 } else if (ctx->deviceType == LANDAU_KOKKOS) { 2254 ierr = PetscObjectTypeCompareAny((PetscObject)ctx->J,&flg,MATSEQAIJKOKKOS,MATMPIAIJKOKKOS,MATAIJKOKKOS,"");CHKERRQ(ierr); 2255 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 2256 PetscCheckFalse(!flg,ctx->comm,PETSC_ERR_ARG_WRONG,"must use '-dm_mat_type aijkokkos -dm_vec_type kokkos' for GPU assembly and Kokkos or use '-dm_landau_device_type cpu'"); 2257 #else 2258 PetscCheckFalse(!flg,ctx->comm,PETSC_ERR_ARG_WRONG,"must configure with '--download-kokkos-kernels' for GPU assembly and Kokkos or use '-dm_landau_device_type cpu'"); 2259 #endif 2260 } 2261 } 2262 ierr = PetscLogEventEnd(ctx->events[15],0,0,0,0);CHKERRQ(ierr); 2263 // create field major ordering 2264 2265 ctx->work_vec = NULL; 2266 ctx->plex_batch = NULL; 2267 ctx->batch_is = NULL; 2268 for (int i=0;i<LANDAU_MAX_GRIDS;i++) grid_batch_is_inv[i] = NULL; 2269 ierr = PetscLogEventBegin(ctx->events[12],0,0,0,0);CHKERRQ(ierr); 2270 ierr = LandauCreateMatrix(comm, *X, grid_batch_is_inv, ctx);CHKERRQ(ierr); 2271 ierr = PetscLogEventEnd(ctx->events[12],0,0,0,0);CHKERRQ(ierr); 2272 2273 // create AMR GPU assembly maps and static GPU data 2274 ierr = CreateStaticGPUData(dim,grid_batch_is_inv,ctx);CHKERRQ(ierr); 2275 2276 ierr = PetscLogEventEnd(ctx->events[13],0,0,0,0);CHKERRQ(ierr); 2277 2278 // create mass matrix 2279 ierr = LandauCreateMassMatrix(*pack, NULL);CHKERRQ(ierr); 2280 2281 if (J) *J = ctx->J; 2282 2283 if (ctx->gpu_assembly && ctx->jacobian_field_major_order) { 2284 PetscContainer container; 2285 // cache ctx for KSP with batch/field major Jacobian ordering -ksp_type gmres/etc -dm_landau_jacobian_field_major_order 2286 ierr = PetscContainerCreate(PETSC_COMM_SELF, &container);CHKERRQ(ierr); 2287 ierr = PetscContainerSetPointer(container, (void *)ctx);CHKERRQ(ierr); 2288 ierr = PetscObjectCompose((PetscObject) ctx->J, "LandauCtx", (PetscObject) container);CHKERRQ(ierr); 2289 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 2290 // batch solvers need to map -- can batch solvers work 2291 ierr = PetscContainerCreate(PETSC_COMM_SELF, &container);CHKERRQ(ierr); 2292 ierr = PetscContainerSetPointer(container, (void *)ctx->plex_batch);CHKERRQ(ierr); 2293 ierr = PetscObjectCompose((PetscObject) ctx->J, "plex_batch_is", (PetscObject) container);CHKERRQ(ierr); 2294 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 2295 } 2296 // for batch solvers 2297 { 2298 PetscContainer container; 2299 PetscInt *pNf; 2300 ierr = PetscContainerCreate(PETSC_COMM_SELF, &container);CHKERRQ(ierr); 2301 ierr = PetscMalloc1(sizeof(*pNf), &pNf);CHKERRQ(ierr); 2302 *pNf = ctx->batch_sz; 2303 ierr = PetscContainerSetPointer(container, (void *)pNf);CHKERRQ(ierr); 2304 ierr = PetscContainerSetUserDestroy(container, MatrixNfDestroy);CHKERRQ(ierr); 2305 ierr = PetscObjectCompose((PetscObject)ctx->J, "batch size", (PetscObject) container);CHKERRQ(ierr); 2306 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 2307 } 2308 2309 PetscFunctionReturn(0); 2310 } 2311 2312 /*@ 2313 LandauDestroyVelocitySpace - Destroy a DMPlex velocity space mesh 2314 2315 Collective on dm 2316 2317 Input/Output Parameters: 2318 . dm - the dm to destroy 2319 2320 Level: beginner 2321 2322 .keywords: mesh 2323 .seealso: LandauCreateVelocitySpace() 2324 @*/ 2325 PetscErrorCode LandauDestroyVelocitySpace(DM *dm) 2326 { 2327 PetscErrorCode ierr,ii; 2328 LandauCtx *ctx; 2329 PetscFunctionBegin; 2330 ierr = DMGetApplicationContext(*dm, &ctx);CHKERRQ(ierr); 2331 ierr = MatDestroy(&ctx->M);CHKERRQ(ierr); 2332 ierr = MatDestroy(&ctx->J);CHKERRQ(ierr); 2333 for (ii=0;ii<ctx->num_species;ii++) { 2334 ierr = PetscFEDestroy(&ctx->fe[ii]);CHKERRQ(ierr); 2335 } 2336 ierr = ISDestroy(&ctx->batch_is);CHKERRQ(ierr); 2337 ierr = VecDestroy(&ctx->work_vec);CHKERRQ(ierr); 2338 ierr = VecScatterDestroy(&ctx->plex_batch);CHKERRQ(ierr); 2339 if (ctx->deviceType == LANDAU_CUDA) { 2340 #if defined(PETSC_HAVE_CUDA) 2341 ierr = LandauCUDAStaticDataClear(&ctx->SData_d);CHKERRQ(ierr); 2342 #else 2343 SETERRQ(ctx->comm,PETSC_ERR_ARG_WRONG,"-landau_device_type %s not built","cuda"); 2344 #endif 2345 } else if (ctx->deviceType == LANDAU_KOKKOS) { 2346 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 2347 ierr = LandauKokkosStaticDataClear(&ctx->SData_d);CHKERRQ(ierr); 2348 #else 2349 SETERRQ(ctx->comm,PETSC_ERR_ARG_WRONG,"-landau_device_type %s not built","kokkos"); 2350 #endif 2351 } else { 2352 if (ctx->SData_d.x) { /* in a CPU run */ 2353 PetscReal *invJ = (PetscReal*)ctx->SData_d.invJ, *xx = (PetscReal*)ctx->SData_d.x, *yy = (PetscReal*)ctx->SData_d.y, *zz = (PetscReal*)ctx->SData_d.z, *ww = (PetscReal*)ctx->SData_d.w; 2354 ierr = PetscFree4(ww,xx,yy,invJ);CHKERRQ(ierr); 2355 if (zz) { 2356 ierr = PetscFree(zz);CHKERRQ(ierr); 2357 } 2358 ierr = PetscFree3(ctx->SData_d.coo_elem_offsets,ctx->SData_d.coo_elem_fullNb,ctx->SData_d.coo_elem_point_offsets);CHKERRQ(ierr); // could be NULL 2359 } 2360 } 2361 2362 if (ctx->times[LANDAU_MATRIX_TOTAL] > 0) { // OMP timings 2363 ierr = PetscPrintf(ctx->comm, "TSStep N 1.0 %10.3e\n",ctx->times[LANDAU_EX2_TSSOLVE]);CHKERRQ(ierr); 2364 ierr = PetscPrintf(ctx->comm, "2: Solve: %10.3e with %" PetscInt_FMT " threads\n",ctx->times[LANDAU_EX2_TSSOLVE] - ctx->times[LANDAU_MATRIX_TOTAL],ctx->batch_sz);CHKERRQ(ierr); 2365 ierr = PetscPrintf(ctx->comm, "3: Landau: %10.3e\n",ctx->times[LANDAU_MATRIX_TOTAL]);CHKERRQ(ierr); 2366 ierr = PetscPrintf(ctx->comm, "Landau Jacobian %" PetscInt_FMT " 1.0 %10.3e\n",(PetscInt)ctx->times[LANDAU_JACOBIAN_COUNT],ctx->times[LANDAU_JACOBIAN]);CHKERRQ(ierr); 2367 ierr = PetscPrintf(ctx->comm, "Landau Operator N 1.0 %10.3e\n",ctx->times[LANDAU_OPERATOR]);CHKERRQ(ierr); 2368 ierr = PetscPrintf(ctx->comm, "Landau Mass N 1.0 %10.3e\n",ctx->times[LANDAU_MASS]);CHKERRQ(ierr); 2369 ierr = PetscPrintf(ctx->comm, " Jac-f-df (GPU) N 1.0 %10.3e\n",ctx->times[LANDAU_F_DF]);CHKERRQ(ierr); 2370 ierr = PetscPrintf(ctx->comm, " Kernel (GPU) N 1.0 %10.3e\n",ctx->times[LANDAU_KERNEL]);CHKERRQ(ierr); 2371 ierr = PetscPrintf(ctx->comm, "MatLUFactorNum X 1.0 %10.3e\n",ctx->times[KSP_FACTOR]);CHKERRQ(ierr); 2372 ierr = PetscPrintf(ctx->comm, "MatSolve X 1.0 %10.3e\n",ctx->times[KSP_SOLVE]);CHKERRQ(ierr); 2373 } 2374 for (PetscInt grid=0 ; grid < ctx->num_grids ; grid++) { 2375 ierr = DMDestroy(&ctx->plex[grid]);CHKERRQ(ierr); 2376 } 2377 PetscFree(ctx); 2378 ierr = DMDestroy(dm);CHKERRQ(ierr); 2379 PetscFunctionReturn(0); 2380 } 2381 2382 /* < v, ru > */ 2383 static void f0_s_den(PetscInt dim, PetscInt Nf, PetscInt NfAux, 2384 const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], 2385 const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], 2386 PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0) 2387 { 2388 PetscInt ii = (PetscInt)PetscRealPart(constants[0]); 2389 f0[0] = u[ii]; 2390 } 2391 2392 /* < v, ru > */ 2393 static void f0_s_mom(PetscInt dim, PetscInt Nf, PetscInt NfAux, 2394 const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], 2395 const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], 2396 PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0) 2397 { 2398 PetscInt ii = (PetscInt)PetscRealPart(constants[0]), jj = (PetscInt)PetscRealPart(constants[1]); 2399 f0[0] = x[jj]*u[ii]; /* x momentum */ 2400 } 2401 2402 static void f0_s_v2(PetscInt dim, PetscInt Nf, PetscInt NfAux, 2403 const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], 2404 const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], 2405 PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0) 2406 { 2407 PetscInt i, ii = (PetscInt)PetscRealPart(constants[0]); 2408 double tmp1 = 0.; 2409 for (i = 0; i < dim; ++i) tmp1 += x[i]*x[i]; 2410 f0[0] = tmp1*u[ii]; 2411 } 2412 2413 static PetscErrorCode gamma_n_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf, PetscScalar *u, void *actx) 2414 { 2415 const PetscReal *c2_0_arr = ((PetscReal*)actx); 2416 const PetscReal c02 = c2_0_arr[0]; 2417 2418 PetscFunctionBegin; 2419 for (int s = 0 ; s < Nf ; s++) { 2420 PetscReal tmp1 = 0.; 2421 for (int i = 0; i < dim; ++i) tmp1 += x[i]*x[i]; 2422 #if defined(PETSC_USE_DEBUG) 2423 u[s] = PetscSqrtReal(1. + tmp1/c02);// u[0] = PetscSqrtReal(1. + xx); 2424 #else 2425 { 2426 PetscReal xx = tmp1/c02; 2427 u[s] = xx/(PetscSqrtReal(1. + xx) + 1.); // better conditioned = xx/(PetscSqrtReal(1. + xx) + 1.) 2428 } 2429 #endif 2430 } 2431 PetscFunctionReturn(0); 2432 } 2433 2434 /* < v, ru > */ 2435 static void f0_s_rden(PetscInt dim, PetscInt Nf, PetscInt NfAux, 2436 const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], 2437 const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], 2438 PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0) 2439 { 2440 PetscInt ii = (PetscInt)PetscRealPart(constants[0]); 2441 f0[0] = 2.*PETSC_PI*x[0]*u[ii]; 2442 } 2443 2444 /* < v, ru > */ 2445 static void f0_s_rmom(PetscInt dim, PetscInt Nf, PetscInt NfAux, 2446 const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], 2447 const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], 2448 PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0) 2449 { 2450 PetscInt ii = (PetscInt)PetscRealPart(constants[0]); 2451 f0[0] = 2.*PETSC_PI*x[0]*x[1]*u[ii]; 2452 } 2453 2454 static void f0_s_rv2(PetscInt dim, PetscInt Nf, PetscInt NfAux, 2455 const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], 2456 const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], 2457 PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0) 2458 { 2459 PetscInt ii = (PetscInt)PetscRealPart(constants[0]); 2460 f0[0] = 2.*PETSC_PI*x[0]*(x[0]*x[0] + x[1]*x[1])*u[ii]; 2461 } 2462 2463 /*@ 2464 LandauPrintNorms - collects moments and prints them 2465 2466 Collective on dm 2467 2468 Input Parameters: 2469 + X - the state 2470 - stepi - current step to print 2471 2472 Level: beginner 2473 2474 .keywords: mesh 2475 .seealso: LandauCreateVelocitySpace() 2476 @*/ 2477 PetscErrorCode LandauPrintNorms(Vec X, PetscInt stepi) 2478 { 2479 PetscErrorCode ierr; 2480 LandauCtx *ctx; 2481 PetscDS prob; 2482 DM pack; 2483 PetscInt cStart, cEnd, dim, ii, i0, nDMs; 2484 PetscScalar xmomentumtot=0, ymomentumtot=0, zmomentumtot=0, energytot=0, densitytot=0, tt[LANDAU_MAX_SPECIES]; 2485 PetscScalar xmomentum[LANDAU_MAX_SPECIES], ymomentum[LANDAU_MAX_SPECIES], zmomentum[LANDAU_MAX_SPECIES], energy[LANDAU_MAX_SPECIES], density[LANDAU_MAX_SPECIES]; 2486 Vec *globXArray; 2487 2488 PetscFunctionBegin; 2489 ierr = VecGetDM(X, &pack);CHKERRQ(ierr); 2490 PetscCheckFalse(!pack,PETSC_COMM_SELF, PETSC_ERR_PLIB, "Vector has no DM"); 2491 ierr = DMGetDimension(pack, &dim);CHKERRQ(ierr); 2492 PetscCheck(dim == 2 || dim == 3,PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " not in [2,3]",dim); 2493 ierr = DMGetApplicationContext(pack, &ctx);CHKERRQ(ierr); 2494 PetscCheck(ctx,PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 2495 /* print momentum and energy */ 2496 ierr = DMCompositeGetNumberDM(pack,&nDMs);CHKERRQ(ierr); 2497 PetscCheck(nDMs == ctx->num_grids*ctx->batch_sz,PETSC_COMM_WORLD, PETSC_ERR_PLIB, "#DM wrong %" PetscInt_FMT " %" PetscInt_FMT,nDMs,ctx->num_grids*ctx->batch_sz); 2498 ierr = PetscMalloc(sizeof(*globXArray)*nDMs, &globXArray);CHKERRQ(ierr); 2499 ierr = DMCompositeGetAccessArray(pack, X, nDMs, NULL, globXArray);CHKERRQ(ierr); 2500 for (PetscInt grid = 0; grid < ctx->num_grids ; grid++) { 2501 Vec Xloc = globXArray[ LAND_PACK_IDX(ctx->batch_view_idx,grid) ]; 2502 ierr = DMGetDS(ctx->plex[grid], &prob);CHKERRQ(ierr); 2503 for (ii=ctx->species_offset[grid],i0=0;ii<ctx->species_offset[grid+1];ii++,i0++) { 2504 PetscScalar user[2] = { (PetscScalar)i0, (PetscScalar)ctx->charges[ii]}; 2505 ierr = PetscDSSetConstants(prob, 2, user);CHKERRQ(ierr); 2506 if (dim==2) { /* 2/3X + 3V (cylindrical coordinates) */ 2507 ierr = PetscDSSetObjective(prob, 0, &f0_s_rden);CHKERRQ(ierr); 2508 ierr = DMPlexComputeIntegralFEM(ctx->plex[grid],Xloc,tt,ctx);CHKERRQ(ierr); 2509 density[ii] = tt[0]*ctx->n_0*ctx->charges[ii]; 2510 ierr = PetscDSSetObjective(prob, 0, &f0_s_rmom);CHKERRQ(ierr); 2511 ierr = DMPlexComputeIntegralFEM(ctx->plex[grid],Xloc,tt,ctx);CHKERRQ(ierr); 2512 zmomentum[ii] = tt[0]*ctx->n_0*ctx->v_0*ctx->masses[ii]; 2513 ierr = PetscDSSetObjective(prob, 0, &f0_s_rv2);CHKERRQ(ierr); 2514 ierr = DMPlexComputeIntegralFEM(ctx->plex[grid],Xloc,tt,ctx);CHKERRQ(ierr); 2515 energy[ii] = tt[0]*0.5*ctx->n_0*ctx->v_0*ctx->v_0*ctx->masses[ii]; 2516 zmomentumtot += zmomentum[ii]; 2517 energytot += energy[ii]; 2518 densitytot += density[ii]; 2519 ierr = PetscPrintf(ctx->comm, "%3D) species-%" PetscInt_FMT ": charge density= %20.13e z-momentum= %20.13e energy= %20.13e",stepi,ii,PetscRealPart(density[ii]),PetscRealPart(zmomentum[ii]),PetscRealPart(energy[ii]));CHKERRQ(ierr); 2520 } else { /* 2/3Xloc + 3V */ 2521 ierr = PetscDSSetObjective(prob, 0, &f0_s_den);CHKERRQ(ierr); 2522 ierr = DMPlexComputeIntegralFEM(ctx->plex[grid],Xloc,tt,ctx);CHKERRQ(ierr); 2523 density[ii] = tt[0]*ctx->n_0*ctx->charges[ii]; 2524 ierr = PetscDSSetObjective(prob, 0, &f0_s_mom);CHKERRQ(ierr); 2525 user[1] = 0; 2526 ierr = PetscDSSetConstants(prob, 2, user);CHKERRQ(ierr); 2527 ierr = DMPlexComputeIntegralFEM(ctx->plex[grid],Xloc,tt,ctx);CHKERRQ(ierr); 2528 xmomentum[ii] = tt[0]*ctx->n_0*ctx->v_0*ctx->masses[ii]; 2529 user[1] = 1; 2530 ierr = PetscDSSetConstants(prob, 2, user);CHKERRQ(ierr); 2531 ierr = DMPlexComputeIntegralFEM(ctx->plex[grid],Xloc,tt,ctx);CHKERRQ(ierr); 2532 ymomentum[ii] = tt[0]*ctx->n_0*ctx->v_0*ctx->masses[ii]; 2533 user[1] = 2; 2534 ierr = PetscDSSetConstants(prob, 2, user);CHKERRQ(ierr); 2535 ierr = DMPlexComputeIntegralFEM(ctx->plex[grid],Xloc,tt,ctx);CHKERRQ(ierr); 2536 zmomentum[ii] = tt[0]*ctx->n_0*ctx->v_0*ctx->masses[ii]; 2537 if (ctx->use_relativistic_corrections) { 2538 /* gamma * M * f */ 2539 if (ii==0 && grid==0) { // do all at once 2540 Vec Mf, globGamma, *globMfArray, *globGammaArray; 2541 PetscErrorCode (*gammaf[1])(PetscInt, PetscReal, const PetscReal [], PetscInt, PetscScalar [], void *) = {gamma_n_f}; 2542 PetscReal *c2_0[1], data[1]; 2543 2544 ierr = VecDuplicate(X,&globGamma);CHKERRQ(ierr); 2545 ierr = VecDuplicate(X,&Mf);CHKERRQ(ierr); 2546 ierr = PetscMalloc(sizeof(*globMfArray)*nDMs, &globMfArray);CHKERRQ(ierr); 2547 ierr = PetscMalloc(sizeof(*globMfArray)*nDMs, &globGammaArray);CHKERRQ(ierr); 2548 /* M * f */ 2549 ierr = MatMult(ctx->M,X,Mf);CHKERRQ(ierr); 2550 /* gamma */ 2551 ierr = DMCompositeGetAccessArray(pack, globGamma, nDMs, NULL, globGammaArray);CHKERRQ(ierr); 2552 for (PetscInt grid = 0; grid < ctx->num_grids ; grid++) { // yes a grid loop in a grid loop to print nice, need to fix for batching 2553 Vec v1 = globGammaArray[ LAND_PACK_IDX(ctx->batch_view_idx,grid) ]; 2554 data[0] = PetscSqr(C_0(ctx->v_0)); 2555 c2_0[0] = &data[0]; 2556 ierr = DMProjectFunction(ctx->plex[grid], 0., gammaf, (void**)c2_0, INSERT_ALL_VALUES, v1);CHKERRQ(ierr); 2557 } 2558 ierr = DMCompositeRestoreAccessArray(pack, globGamma, nDMs, NULL, globGammaArray);CHKERRQ(ierr); 2559 /* gamma * Mf */ 2560 ierr = DMCompositeGetAccessArray(pack, globGamma, nDMs, NULL, globGammaArray);CHKERRQ(ierr); 2561 ierr = DMCompositeGetAccessArray(pack, Mf, nDMs, NULL, globMfArray);CHKERRQ(ierr); 2562 for (PetscInt grid = 0; grid < ctx->num_grids ; grid++) { // yes a grid loop in a grid loop to print nice 2563 PetscInt Nf = ctx->species_offset[grid+1] - ctx->species_offset[grid], N, bs; 2564 Vec Mfsub = globMfArray[ LAND_PACK_IDX(ctx->batch_view_idx,grid) ], Gsub = globGammaArray[ LAND_PACK_IDX(ctx->batch_view_idx,grid) ], v1, v2; 2565 // get each component 2566 ierr = VecGetSize(Mfsub,&N);CHKERRQ(ierr); 2567 ierr = VecCreate(ctx->comm,&v1);CHKERRQ(ierr); 2568 ierr = VecSetSizes(v1,PETSC_DECIDE,N/Nf);CHKERRQ(ierr); 2569 ierr = VecCreate(ctx->comm,&v2);CHKERRQ(ierr); 2570 ierr = VecSetSizes(v2,PETSC_DECIDE,N/Nf);CHKERRQ(ierr); 2571 ierr = VecSetFromOptions(v1);CHKERRQ(ierr); // ??? 2572 ierr = VecSetFromOptions(v2);CHKERRQ(ierr); 2573 // get each component 2574 ierr = VecGetBlockSize(Gsub,&bs);CHKERRQ(ierr); 2575 PetscCheck(bs == Nf,PETSC_COMM_SELF, PETSC_ERR_PLIB, "bs %" PetscInt_FMT " != num_species %" PetscInt_FMT " in Gsub",bs,Nf); 2576 ierr = VecGetBlockSize(Mfsub,&bs);CHKERRQ(ierr); 2577 PetscCheck(bs == Nf,PETSC_COMM_SELF, PETSC_ERR_PLIB, "bs %" PetscInt_FMT " != num_species %" PetscInt_FMT,bs,Nf); 2578 for (int i=0, ix=ctx->species_offset[grid] ; i<Nf ; i++, ix++) { 2579 PetscScalar val; 2580 ierr = VecStrideGather(Gsub,i,v1,INSERT_VALUES);CHKERRQ(ierr); 2581 ierr = VecStrideGather(Mfsub,i,v2,INSERT_VALUES);CHKERRQ(ierr); 2582 ierr = VecDot(v1,v2,&val);CHKERRQ(ierr); 2583 energy[ix] = PetscRealPart(val)*ctx->n_0*ctx->v_0*ctx->v_0*ctx->masses[ix]; 2584 } 2585 ierr = VecDestroy(&v1);CHKERRQ(ierr); 2586 ierr = VecDestroy(&v2);CHKERRQ(ierr); 2587 } /* grids */ 2588 ierr = DMCompositeRestoreAccessArray(pack, globGamma, nDMs, NULL, globGammaArray);CHKERRQ(ierr); 2589 ierr = DMCompositeRestoreAccessArray(pack, Mf, nDMs, NULL, globMfArray);CHKERRQ(ierr); 2590 ierr = PetscFree(globGammaArray);CHKERRQ(ierr); 2591 ierr = PetscFree(globMfArray);CHKERRQ(ierr); 2592 ierr = VecDestroy(&globGamma);CHKERRQ(ierr); 2593 ierr = VecDestroy(&Mf);CHKERRQ(ierr); 2594 } 2595 } else { 2596 ierr = PetscDSSetObjective(prob, 0, &f0_s_v2);CHKERRQ(ierr); 2597 ierr = DMPlexComputeIntegralFEM(ctx->plex[grid],Xloc,tt,ctx);CHKERRQ(ierr); 2598 energy[ii] = 0.5*tt[0]*ctx->n_0*ctx->v_0*ctx->v_0*ctx->masses[ii]; 2599 } 2600 ierr = PetscPrintf(ctx->comm, "%3" PetscInt_FMT ") species %" PetscInt_FMT ": density=%20.13e, x-momentum=%20.13e, y-momentum=%20.13e, z-momentum=%20.13e, energy=%21.13e", 2601 stepi,ii,PetscRealPart(density[ii]),PetscRealPart(xmomentum[ii]),PetscRealPart(ymomentum[ii]),PetscRealPart(zmomentum[ii]),PetscRealPart(energy[ii]));CHKERRQ(ierr); 2602 xmomentumtot += xmomentum[ii]; 2603 ymomentumtot += ymomentum[ii]; 2604 zmomentumtot += zmomentum[ii]; 2605 energytot += energy[ii]; 2606 densitytot += density[ii]; 2607 } 2608 if (ctx->num_species>1) PetscPrintf(ctx->comm, "\n"); 2609 } 2610 } 2611 ierr = DMCompositeRestoreAccessArray(pack, X, nDMs, NULL, globXArray);CHKERRQ(ierr); 2612 ierr = PetscFree(globXArray);CHKERRQ(ierr); 2613 /* totals */ 2614 ierr = DMPlexGetHeightStratum(ctx->plex[0],0,&cStart,&cEnd);CHKERRQ(ierr); 2615 if (ctx->num_species>1) { 2616 if (dim==2) { 2617 ierr = PetscPrintf(ctx->comm, "\t%3" PetscInt_FMT ") Total: charge density=%21.13e, momentum=%21.13e, energy=%21.13e (m_i[0]/m_e = %g, %" PetscInt_FMT " cells on electron grid)", 2618 stepi,(double)PetscRealPart(densitytot),(double)PetscRealPart(zmomentumtot),(double)PetscRealPart(energytot),(double)(ctx->masses[1]/ctx->masses[0]),cEnd-cStart);CHKERRQ(ierr); 2619 } else { 2620 ierr = PetscPrintf(ctx->comm, "\t%3" PetscInt_FMT ") Total: charge density=%21.13e, x-momentum=%21.13e, y-momentum=%21.13e, z-momentum=%21.13e, energy=%21.13e (m_i[0]/m_e = %g, %" PetscInt_FMT " cells)", 2621 stepi,(double)PetscRealPart(densitytot),(double)PetscRealPart(xmomentumtot),(double)PetscRealPart(ymomentumtot),(double)PetscRealPart(zmomentumtot),(double)PetscRealPart(energytot),(double)(ctx->masses[1]/ctx->masses[0]),cEnd-cStart);CHKERRQ(ierr); 2622 } 2623 } else { 2624 ierr = PetscPrintf(ctx->comm, " -- %" PetscInt_FMT " cells",cEnd-cStart);CHKERRQ(ierr); 2625 } 2626 ierr = PetscPrintf(ctx->comm,"\n");CHKERRQ(ierr); 2627 PetscFunctionReturn(0); 2628 } 2629 2630 /*@ 2631 LandauCreateMassMatrix - Create mass matrix for Landau in Plex space (not field major order of Jacobian) 2632 2633 Collective on pack 2634 2635 Input Parameters: 2636 . pack - the DM object 2637 2638 Output Parameters: 2639 . Amat - The mass matrix (optional), mass matrix is added to the DM context 2640 2641 Level: beginner 2642 2643 .keywords: mesh 2644 .seealso: LandauCreateVelocitySpace() 2645 @*/ 2646 PetscErrorCode LandauCreateMassMatrix(DM pack, Mat *Amat) 2647 { 2648 DM mass_pack,massDM[LANDAU_MAX_GRIDS]; 2649 PetscDS prob; 2650 PetscInt ii,dim,N1=1,N2; 2651 PetscErrorCode ierr; 2652 LandauCtx *ctx; 2653 Mat packM,subM[LANDAU_MAX_GRIDS]; 2654 2655 PetscFunctionBegin; 2656 PetscValidHeaderSpecific(pack,DM_CLASSID,1); 2657 if (Amat) PetscValidPointer(Amat,2); 2658 ierr = DMGetApplicationContext(pack, &ctx);CHKERRQ(ierr); 2659 PetscCheck(ctx,PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 2660 ierr = PetscLogEventBegin(ctx->events[14],0,0,0,0);CHKERRQ(ierr); 2661 ierr = DMGetDimension(pack, &dim);CHKERRQ(ierr); 2662 ierr = DMCompositeCreate(PetscObjectComm((PetscObject) pack),&mass_pack);CHKERRQ(ierr); 2663 /* create pack mass matrix */ 2664 for (PetscInt grid=0, ix=0 ; grid<ctx->num_grids ; grid++) { 2665 ierr = DMClone(ctx->plex[grid], &massDM[grid]);CHKERRQ(ierr); 2666 ierr = DMCopyFields(ctx->plex[grid], massDM[grid]);CHKERRQ(ierr); 2667 ierr = DMCreateDS(massDM[grid]);CHKERRQ(ierr); 2668 ierr = DMGetDS(massDM[grid], &prob);CHKERRQ(ierr); 2669 for (ix=0, ii=ctx->species_offset[grid];ii<ctx->species_offset[grid+1];ii++,ix++) { 2670 if (dim==3) {ierr = PetscDSSetJacobian(prob, ix, ix, g0_1, NULL, NULL, NULL);CHKERRQ(ierr);} 2671 else {ierr = PetscDSSetJacobian(prob, ix, ix, g0_r, NULL, NULL, NULL);CHKERRQ(ierr);} 2672 } 2673 #if !defined(LANDAU_SPECIES_MAJOR) 2674 ierr = DMCompositeAddDM(mass_pack,massDM[grid]);CHKERRQ(ierr); 2675 #else 2676 for (PetscInt b_id=0;b_id<ctx->batch_sz;b_id++) { // add batch size DMs for this species grid 2677 ierr = DMCompositeAddDM(mass_pack,massDM[grid]);CHKERRQ(ierr); 2678 } 2679 #endif 2680 ierr = DMCreateMatrix(massDM[grid], &subM[grid]);CHKERRQ(ierr); 2681 } 2682 #if !defined(LANDAU_SPECIES_MAJOR) 2683 // stack the batched DMs 2684 for (PetscInt b_id=1;b_id<ctx->batch_sz;b_id++) { 2685 for (PetscInt grid=0;grid<ctx->num_grids;grid++) { 2686 ierr = DMCompositeAddDM(mass_pack, massDM[grid]);CHKERRQ(ierr); 2687 } 2688 } 2689 #endif 2690 ierr = PetscOptionsInsertString(NULL,"-dm_preallocate_only"); 2691 ierr = DMSetFromOptions(mass_pack);CHKERRQ(ierr); 2692 ierr = DMCreateMatrix(mass_pack, &packM);CHKERRQ(ierr); 2693 ierr = PetscOptionsInsertString(NULL,"-dm_preallocate_only false"); 2694 ierr = MatSetOption(packM,MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE);CHKERRQ(ierr); 2695 ierr = MatSetOption(packM,MAT_IGNORE_ZERO_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 2696 ierr = DMDestroy(&mass_pack);CHKERRQ(ierr); 2697 /* make mass matrix for each block */ 2698 for (PetscInt grid=0;grid<ctx->num_grids;grid++) { 2699 Vec locX; 2700 DM plex = massDM[grid]; 2701 ierr = DMGetLocalVector(plex, &locX);CHKERRQ(ierr); 2702 /* Mass matrix is independent of the input, so no need to fill locX */ 2703 ierr = DMPlexSNESComputeJacobianFEM(plex, locX, subM[grid], subM[grid], ctx);CHKERRQ(ierr); 2704 ierr = DMRestoreLocalVector(plex, &locX);CHKERRQ(ierr); 2705 ierr = DMDestroy(&massDM[grid]);CHKERRQ(ierr); 2706 } 2707 ierr = MatGetSize(ctx->J, &N1, NULL);CHKERRQ(ierr); 2708 ierr = MatGetSize(packM, &N2, NULL);CHKERRQ(ierr); 2709 PetscCheck(N1 == N2,PetscObjectComm((PetscObject) pack), PETSC_ERR_PLIB, "Incorrect matrix sizes: |Jacobian| = %" PetscInt_FMT ", |Mass|=%" PetscInt_FMT,N1,N2); 2710 /* assemble block diagonals */ 2711 for (PetscInt grid=0 ; grid<ctx->num_grids ; grid++) { 2712 Mat B = subM[grid]; 2713 PetscInt nloc, nzl, colbuf[1024], row; 2714 ierr = MatGetSize(B, &nloc, NULL);CHKERRQ(ierr); 2715 for (PetscInt b_id = 0 ; b_id < ctx->batch_sz ; b_id++) { 2716 const PetscInt moffset = LAND_MOFFSET(b_id,grid,ctx->batch_sz,ctx->num_grids,ctx->mat_offset); 2717 const PetscInt *cols; 2718 const PetscScalar *vals; 2719 for (int i=0 ; i<nloc ; i++) { 2720 ierr = MatGetRow(B,i,&nzl,&cols,&vals);CHKERRQ(ierr); 2721 PetscCheck(nzl<=1024,PetscObjectComm((PetscObject) pack), PETSC_ERR_PLIB, "Row too big: %" PetscInt_FMT,nzl); 2722 for (int j=0; j<nzl; j++) colbuf[j] = cols[j] + moffset; 2723 row = i + moffset; 2724 ierr = MatSetValues(packM,1,&row,nzl,colbuf,vals,INSERT_VALUES);CHKERRQ(ierr); 2725 ierr = MatRestoreRow(B,i,&nzl,&cols,&vals);CHKERRQ(ierr); 2726 } 2727 } 2728 } 2729 // cleanup 2730 for (PetscInt grid=0 ; grid<ctx->num_grids ; grid++) { 2731 ierr = MatDestroy(&subM[grid]);CHKERRQ(ierr); 2732 } 2733 ierr = MatAssemblyBegin(packM,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2734 ierr = MatAssemblyEnd(packM,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2735 ierr = PetscObjectSetName((PetscObject)packM, "mass");CHKERRQ(ierr); 2736 ierr = MatViewFromOptions(packM,NULL,"-dm_landau_mass_view");CHKERRQ(ierr); 2737 ctx->M = packM; 2738 if (Amat) *Amat = packM; 2739 ierr = PetscLogEventEnd(ctx->events[14],0,0,0,0);CHKERRQ(ierr); 2740 PetscFunctionReturn(0); 2741 } 2742 2743 /*@ 2744 LandauIFunction - TS residual calculation 2745 2746 Collective on ts 2747 2748 Input Parameters: 2749 + TS - The time stepping context 2750 . time_dummy - current time (not used) 2751 - X - Current state 2752 + X_t - Time derivative of current state 2753 . actx - Landau context 2754 2755 Output Parameter: 2756 . F - The residual 2757 2758 Level: beginner 2759 2760 .keywords: mesh 2761 .seealso: LandauCreateVelocitySpace(), LandauIJacobian() 2762 @*/ 2763 PetscErrorCode LandauIFunction(TS ts, PetscReal time_dummy, Vec X, Vec X_t, Vec F, void *actx) 2764 { 2765 PetscErrorCode ierr; 2766 LandauCtx *ctx=(LandauCtx*)actx; 2767 PetscInt dim; 2768 DM pack; 2769 #if defined(PETSC_HAVE_THREADSAFETY) 2770 double starttime, endtime; 2771 #endif 2772 2773 PetscFunctionBegin; 2774 ierr = TSGetDM(ts,&pack);CHKERRQ(ierr); 2775 ierr = DMGetApplicationContext(pack, &ctx);CHKERRQ(ierr); 2776 PetscCheckFalse(!ctx,PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 2777 if (ctx->stage) { 2778 ierr = PetscLogStagePush(ctx->stage);CHKERRQ(ierr); 2779 } 2780 ierr = PetscLogEventBegin(ctx->events[11],0,0,0,0);CHKERRQ(ierr); 2781 ierr = PetscLogEventBegin(ctx->events[0],0,0,0,0);CHKERRQ(ierr); 2782 #if defined(PETSC_HAVE_THREADSAFETY) 2783 starttime = MPI_Wtime(); 2784 #endif 2785 ierr = DMGetDimension(pack, &dim);CHKERRQ(ierr); 2786 if (!ctx->aux_bool) { 2787 ierr = PetscInfo(ts, "Create Landau Jacobian t=%g X=%p %s\n",time_dummy,X_t,ctx->aux_bool ? " -- seems to be in line search" : "");CHKERRQ(ierr); 2788 ierr = LandauFormJacobian_Internal(X,ctx->J,dim,0.0,(void*)ctx);CHKERRQ(ierr); 2789 ierr = MatViewFromOptions(ctx->J, NULL, "-dm_landau_jacobian_view");CHKERRQ(ierr); 2790 ctx->aux_bool = PETSC_TRUE; 2791 } else { 2792 ierr = PetscInfo(ts, "Skip forming Jacobian, has not changed (should check norm)\n");CHKERRQ(ierr); 2793 } 2794 /* mat vec for op */ 2795 ierr = MatMult(ctx->J,X,F);CHKERRQ(ierr);CHKERRQ(ierr); /* C*f */ 2796 /* add time term */ 2797 if (X_t) { 2798 ierr = MatMultAdd(ctx->M,X_t,F,F);CHKERRQ(ierr); 2799 } 2800 #if defined(PETSC_HAVE_THREADSAFETY) 2801 if (ctx->stage) { 2802 endtime = MPI_Wtime(); 2803 ctx->times[LANDAU_OPERATOR] += (endtime - starttime); 2804 ctx->times[LANDAU_JACOBIAN] += (endtime - starttime); 2805 ctx->times[LANDAU_JACOBIAN_COUNT] += 1; 2806 } 2807 #endif 2808 ierr = PetscLogEventEnd(ctx->events[0],0,0,0,0);CHKERRQ(ierr); 2809 ierr = PetscLogEventEnd(ctx->events[11],0,0,0,0);CHKERRQ(ierr); 2810 if (ctx->stage) { 2811 ierr = PetscLogStagePop();CHKERRQ(ierr); 2812 #if defined(PETSC_HAVE_THREADSAFETY) 2813 ctx->times[LANDAU_MATRIX_TOTAL] += (endtime - starttime); 2814 #endif 2815 } 2816 PetscFunctionReturn(0); 2817 } 2818 2819 /*@ 2820 LandauIJacobian - TS Jacobian construction 2821 2822 Collective on ts 2823 2824 Input Parameters: 2825 + TS - The time stepping context 2826 . time_dummy - current time (not used) 2827 - X - Current state 2828 + U_tdummy - Time derivative of current state (not used) 2829 . shift - shift for du/dt term 2830 - actx - Landau context 2831 2832 Output Parameter: 2833 . Amat - Jacobian 2834 + Pmat - same as Amat 2835 2836 Level: beginner 2837 2838 .keywords: mesh 2839 .seealso: LandauCreateVelocitySpace(), LandauIFunction() 2840 @*/ 2841 PetscErrorCode LandauIJacobian(TS ts, PetscReal time_dummy, Vec X, Vec U_tdummy, PetscReal shift, Mat Amat, Mat Pmat, void *actx) 2842 { 2843 PetscErrorCode ierr; 2844 LandauCtx *ctx=NULL; 2845 PetscInt dim; 2846 DM pack; 2847 #if defined(PETSC_HAVE_THREADSAFETY) 2848 double starttime, endtime; 2849 #endif 2850 PetscFunctionBegin; 2851 ierr = TSGetDM(ts,&pack);CHKERRQ(ierr); 2852 ierr = DMGetApplicationContext(pack, &ctx);CHKERRQ(ierr); 2853 PetscCheckFalse(!ctx,PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 2854 PetscCheckFalse(Amat!=Pmat || Amat!=ctx->J,ctx->comm, PETSC_ERR_PLIB, "Amat!=Pmat || Amat!=ctx->J"); 2855 ierr = DMGetDimension(pack, &dim);CHKERRQ(ierr); 2856 /* get collision Jacobian into A */ 2857 if (ctx->stage) { 2858 ierr = PetscLogStagePush(ctx->stage);CHKERRQ(ierr); 2859 } 2860 ierr = PetscLogEventBegin(ctx->events[11],0,0,0,0);CHKERRQ(ierr); 2861 ierr = PetscLogEventBegin(ctx->events[9],0,0,0,0);CHKERRQ(ierr); 2862 #if defined(PETSC_HAVE_THREADSAFETY) 2863 starttime = MPI_Wtime(); 2864 #endif 2865 ierr = PetscInfo(ts, "Adding just mass to Jacobian t=%g, shift=%g\n",(double)time_dummy,(double)shift);CHKERRQ(ierr); 2866 PetscCheckFalse(shift==0.0,ctx->comm, PETSC_ERR_PLIB, "zero shift"); 2867 PetscCheckFalse(!ctx->aux_bool,ctx->comm, PETSC_ERR_PLIB, "wrong state"); 2868 if (!ctx->use_matrix_mass) { 2869 ierr = LandauFormJacobian_Internal(X,ctx->J,dim,shift,(void*)ctx);CHKERRQ(ierr); 2870 ierr = MatViewFromOptions(ctx->J, NULL, "-dm_landau_mat_view");CHKERRQ(ierr); 2871 } else { /* add mass */ 2872 ierr = MatAXPY(Pmat,shift,ctx->M,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2873 } 2874 ctx->aux_bool = PETSC_FALSE; 2875 #if defined(PETSC_HAVE_THREADSAFETY) 2876 if (ctx->stage) { 2877 endtime = MPI_Wtime(); 2878 ctx->times[LANDAU_OPERATOR] += (endtime - starttime); 2879 ctx->times[LANDAU_MASS] += (endtime - starttime); 2880 } 2881 #endif 2882 ierr = PetscLogEventEnd(ctx->events[9],0,0,0,0);CHKERRQ(ierr); 2883 ierr = PetscLogEventEnd(ctx->events[11],0,0,0,0);CHKERRQ(ierr); 2884 if (ctx->stage) { 2885 ierr = PetscLogStagePop();CHKERRQ(ierr); 2886 #if defined(PETSC_HAVE_THREADSAFETY) 2887 ctx->times[LANDAU_MATRIX_TOTAL] += (endtime - starttime); 2888 #endif 2889 } 2890 PetscFunctionReturn(0); 2891 } 2892