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