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 #include "land_tensors.h" 20 21 #if defined(PETSC_HAVE_OPENMP) 22 #include <omp.h> 23 #endif 24 25 static PetscErrorCode LandauGPUMapsDestroy(void *ptr) 26 { 27 P4estVertexMaps *maps = (P4estVertexMaps *)ptr; 28 PetscFunctionBegin; 29 // free device data 30 if (maps[0].deviceType != LANDAU_CPU) { 31 #if defined(PETSC_HAVE_KOKKOS) 32 if (maps[0].deviceType == LANDAU_KOKKOS) { 33 PetscCall(LandauKokkosDestroyMatMaps(maps, maps[0].numgrids)); // implies Kokkos does 34 } 35 #endif 36 } 37 // free host data 38 for (PetscInt grid = 0; grid < maps[0].numgrids; grid++) { 39 PetscCall(PetscFree(maps[grid].c_maps)); 40 PetscCall(PetscFree(maps[grid].gIdx)); 41 } 42 PetscCall(PetscFree(maps)); 43 44 PetscFunctionReturn(PETSC_SUCCESS); 45 } 46 static PetscErrorCode energy_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx) 47 { 48 PetscReal v2 = 0; 49 PetscFunctionBegin; 50 /* compute v^2 / 2 */ 51 for (int i = 0; i < dim; ++i) v2 += x[i] * x[i]; 52 /* evaluate the Maxwellian */ 53 u[0] = v2 / 2; 54 PetscFunctionReturn(PETSC_SUCCESS); 55 } 56 57 /* needs double */ 58 static PetscErrorCode gamma_m1_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx) 59 { 60 PetscReal *c2_0_arr = ((PetscReal *)actx); 61 double u2 = 0, c02 = (double)*c2_0_arr, xx; 62 63 PetscFunctionBegin; 64 /* compute u^2 / 2 */ 65 for (int i = 0; i < dim; ++i) u2 += x[i] * x[i]; 66 /* gamma - 1 = g_eps, for conditioning and we only take derivatives */ 67 xx = u2 / c02; 68 #if defined(PETSC_USE_DEBUG) 69 u[0] = PetscSqrtReal(1. + xx); 70 #else 71 u[0] = xx / (PetscSqrtReal(1. + xx) + 1.) - 1.; // better conditioned. -1 might help condition and only used for derivative 72 #endif 73 PetscFunctionReturn(PETSC_SUCCESS); 74 } 75 76 /* 77 LandauFormJacobian_Internal - Evaluates Jacobian matrix. 78 79 Input Parameters: 80 . globX - input vector 81 . actx - optional user-defined context 82 . dim - dimension 83 84 Output Parameter: 85 . J0acP - Jacobian matrix filled, not created 86 */ 87 static PetscErrorCode LandauFormJacobian_Internal(Vec a_X, Mat JacP, const PetscInt dim, PetscReal shift, void *a_ctx) 88 { 89 LandauCtx *ctx = (LandauCtx *)a_ctx; 90 PetscInt numCells[LANDAU_MAX_GRIDS], Nq, Nb; 91 PetscQuadrature quad; 92 PetscReal Eq_m[LANDAU_MAX_SPECIES]; // could be static data w/o quench (ex2) 93 PetscScalar *cellClosure = NULL; 94 const PetscScalar *xdata = NULL; 95 PetscDS prob; 96 PetscContainer container; 97 P4estVertexMaps *maps; 98 Mat subJ[LANDAU_MAX_GRIDS * LANDAU_MAX_BATCH_SZ]; 99 100 PetscFunctionBegin; 101 PetscValidHeaderSpecific(a_X, VEC_CLASSID, 1); 102 PetscValidHeaderSpecific(JacP, MAT_CLASSID, 2); 103 PetscAssertPointer(ctx, 5); 104 /* check for matrix container for GPU assembly. Support CPU assembly for debugging */ 105 PetscCheck(ctx->plex[0] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created"); 106 PetscCall(PetscLogEventBegin(ctx->events[10], 0, 0, 0, 0)); 107 PetscCall(DMGetDS(ctx->plex[0], &prob)); // same DS for all grids 108 PetscCall(PetscObjectQuery((PetscObject)JacP, "assembly_maps", (PetscObject *)&container)); 109 if (container) { 110 PetscCheck(ctx->gpu_assembly, ctx->comm, PETSC_ERR_ARG_WRONG, "maps but no GPU assembly"); 111 PetscCall(PetscContainerGetPointer(container, (void **)&maps)); 112 PetscCheck(maps, ctx->comm, PETSC_ERR_ARG_WRONG, "empty GPU matrix container"); 113 for (PetscInt i = 0; i < ctx->num_grids * ctx->batch_sz; i++) subJ[i] = NULL; 114 } else { 115 PetscCheck(!ctx->gpu_assembly, ctx->comm, PETSC_ERR_ARG_WRONG, "No maps but GPU assembly"); 116 for (PetscInt tid = 0; tid < ctx->batch_sz; tid++) { 117 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCreateMatrix(ctx->plex[grid], &subJ[LAND_PACK_IDX(tid, grid)])); 118 } 119 maps = NULL; 120 } 121 // 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) 122 PetscCall(PetscFEGetQuadrature(ctx->fe[0], &quad)); 123 PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, NULL)); 124 PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb)); 125 PetscCheck(Nq <= LANDAU_MAX_NQND, ctx->comm, PETSC_ERR_ARG_WRONG, "Order too high. Nq = %" PetscInt_FMT " > LANDAU_MAX_NQND (%d)", Nq, LANDAU_MAX_NQND); 126 PetscCheck(Nb <= LANDAU_MAX_NQND, ctx->comm, PETSC_ERR_ARG_WRONG, "Order too high. Nb = %" PetscInt_FMT " > LANDAU_MAX_NQND (%d)", Nb, LANDAU_MAX_NQND); 127 // get metadata for collecting dynamic data 128 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 129 PetscInt cStart, cEnd; 130 PetscCheck(ctx->plex[grid] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created"); 131 PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd)); 132 numCells[grid] = cEnd - cStart; // grids can have different topology 133 } 134 PetscCall(PetscLogEventEnd(ctx->events[10], 0, 0, 0, 0)); 135 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 */ 136 DM pack; 137 PetscCall(VecGetDM(a_X, &pack)); 138 PetscCheck(pack, PETSC_COMM_SELF, PETSC_ERR_PLIB, "pack has no DM"); 139 PetscCall(PetscLogEventBegin(ctx->events[1], 0, 0, 0, 0)); 140 for (PetscInt fieldA = 0; fieldA < ctx->num_species; fieldA++) { 141 Eq_m[fieldA] = ctx->Ez * ctx->t_0 * ctx->charges[fieldA] / (ctx->v_0 * ctx->masses[fieldA]); /* normalize dimensionless */ 142 if (dim == 2) Eq_m[fieldA] *= 2 * PETSC_PI; /* add the 2pi term that is not in Landau */ 143 } 144 if (!ctx->gpu_assembly) { 145 Vec *locXArray, *globXArray; 146 PetscScalar *cellClosure_it; 147 PetscInt cellClosure_sz = 0, nDMs, Nf[LANDAU_MAX_GRIDS]; 148 PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS]; 149 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 150 PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid])); 151 PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid])); 152 PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid])); 153 } 154 /* count cellClosure size */ 155 PetscCall(DMCompositeGetNumberDM(pack, &nDMs)); 156 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) cellClosure_sz += Nb * Nf[grid] * numCells[grid]; 157 PetscCall(PetscMalloc1(cellClosure_sz * ctx->batch_sz, &cellClosure)); 158 cellClosure_it = cellClosure; 159 PetscCall(PetscMalloc(sizeof(*locXArray) * nDMs, &locXArray)); 160 PetscCall(PetscMalloc(sizeof(*globXArray) * nDMs, &globXArray)); 161 PetscCall(DMCompositeGetLocalAccessArray(pack, a_X, nDMs, NULL, locXArray)); 162 PetscCall(DMCompositeGetAccessArray(pack, a_X, nDMs, NULL, globXArray)); 163 for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // OpenMP (once) 164 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 165 Vec locX = locXArray[LAND_PACK_IDX(b_id, grid)], globX = globXArray[LAND_PACK_IDX(b_id, grid)], locX2; 166 PetscInt cStart, cEnd, ei; 167 PetscCall(VecDuplicate(locX, &locX2)); 168 PetscCall(DMGlobalToLocalBegin(ctx->plex[grid], globX, INSERT_VALUES, locX2)); 169 PetscCall(DMGlobalToLocalEnd(ctx->plex[grid], globX, INSERT_VALUES, locX2)); 170 PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd)); 171 for (ei = cStart; ei < cEnd; ++ei) { 172 PetscScalar *coef = NULL; 173 PetscCall(DMPlexVecGetClosure(ctx->plex[grid], section[grid], locX2, ei, NULL, &coef)); 174 PetscCall(PetscMemcpy(cellClosure_it, coef, Nb * Nf[grid] * sizeof(*cellClosure_it))); /* change if LandauIPReal != PetscScalar */ 175 PetscCall(DMPlexVecRestoreClosure(ctx->plex[grid], section[grid], locX2, ei, NULL, &coef)); 176 cellClosure_it += Nb * Nf[grid]; 177 } 178 PetscCall(VecDestroy(&locX2)); 179 } 180 } 181 PetscCheck(cellClosure_it - cellClosure == cellClosure_sz * ctx->batch_sz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "iteration wrong %" PetscCount_FMT " != cellClosure_sz = %" PetscInt_FMT, (PetscCount)(cellClosure_it - cellClosure), 182 cellClosure_sz * ctx->batch_sz); 183 PetscCall(DMCompositeRestoreLocalAccessArray(pack, a_X, nDMs, NULL, locXArray)); 184 PetscCall(DMCompositeRestoreAccessArray(pack, a_X, nDMs, NULL, globXArray)); 185 PetscCall(PetscFree(locXArray)); 186 PetscCall(PetscFree(globXArray)); 187 xdata = NULL; 188 } else { 189 PetscMemType mtype; 190 if (ctx->jacobian_field_major_order) { // get data in batch ordering 191 PetscCall(VecScatterBegin(ctx->plex_batch, a_X, ctx->work_vec, INSERT_VALUES, SCATTER_FORWARD)); 192 PetscCall(VecScatterEnd(ctx->plex_batch, a_X, ctx->work_vec, INSERT_VALUES, SCATTER_FORWARD)); 193 PetscCall(VecGetArrayReadAndMemType(ctx->work_vec, &xdata, &mtype)); 194 } else { 195 PetscCall(VecGetArrayReadAndMemType(a_X, &xdata, &mtype)); 196 } 197 PetscCheck(mtype == PETSC_MEMTYPE_HOST || ctx->deviceType != LANDAU_CPU, ctx->comm, PETSC_ERR_ARG_WRONG, "CPU run with device data: use -mat_type aij"); 198 cellClosure = NULL; 199 } 200 PetscCall(PetscLogEventEnd(ctx->events[1], 0, 0, 0, 0)); 201 } else xdata = cellClosure = NULL; 202 203 /* do it */ 204 if (ctx->deviceType == LANDAU_KOKKOS) { 205 #if defined(PETSC_HAVE_KOKKOS) 206 PetscCall(LandauKokkosJacobian(ctx->plex, Nq, Nb, ctx->batch_sz, ctx->num_grids, numCells, Eq_m, cellClosure, xdata, &ctx->SData_d, shift, ctx->events, ctx->mat_offset, ctx->species_offset, subJ, JacP)); 207 #else 208 SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type %s not built", "kokkos"); 209 #endif 210 } else { /* CPU version */ 211 PetscTabulation *Tf; // used for CPU and print info. Same on all grids and all species 212 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]; 213 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; 214 PetscReal *nu_alpha = (PetscReal *)ctx->SData_d.alpha, *nu_beta = (PetscReal *)ctx->SData_d.beta, *invMass = (PetscReal *)ctx->SData_d.invMass; 215 PetscReal(*lambdas)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS] = (PetscReal(*)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS])ctx->SData_d.lambdas; 216 PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS]; 217 PetscScalar *coo_vals = NULL; 218 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 219 PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid])); 220 PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid])); 221 PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid])); 222 } 223 /* count IPf size, etc */ 224 PetscCall(PetscDSGetTabulation(prob, &Tf)); // Bf, &Df same for all grids 225 const PetscReal *const BB = Tf[0]->T[0], *const DD = Tf[0]->T[1]; 226 ip_offset[0] = ipf_offset[0] = elem_offset[0] = 0; 227 for (PetscInt grid = 0; grid < num_grids; grid++) { 228 PetscInt nfloc = ctx->species_offset[grid + 1] - ctx->species_offset[grid]; 229 elem_offset[grid + 1] = elem_offset[grid] + numCells[grid]; 230 ip_offset[grid + 1] = ip_offset[grid] + numCells[grid] * Nq; 231 ipf_offset[grid + 1] = ipf_offset[grid] + Nq * nfloc * numCells[grid]; 232 } 233 IPf_sz_glb = ipf_offset[num_grids]; 234 IPf_sz_tot = IPf_sz_glb * ctx->batch_sz; 235 // prep COO 236 { 237 PetscCall(PetscMalloc1(ctx->SData_d.coo_size, &coo_vals)); // allocate every time? 238 PetscCall(PetscInfo(ctx->plex[0], "COO Allocate %" PetscInt_FMT " values\n", (PetscInt)ctx->SData_d.coo_size)); 239 } 240 if (shift == 0.0) { /* compute dynamic data f and df and init data for Jacobian */ 241 #if defined(PETSC_HAVE_THREADSAFETY) 242 double starttime, endtime; 243 starttime = MPI_Wtime(); 244 #endif 245 PetscCall(PetscLogEventBegin(ctx->events[8], 0, 0, 0, 0)); 246 PetscCall(PetscMalloc4(IPf_sz_tot, &ff, IPf_sz_tot, &dudx, IPf_sz_tot, &dudy, dim == 3 ? IPf_sz_tot : 0, &dudz)); 247 // F df/dx 248 for (PetscInt tid = 0; tid < ctx->batch_sz * elem_offset[num_grids]; tid++) { // for each element 249 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 250 // find my grid: 251 PetscInt grid = 0; 252 while (b_elem_idx >= elem_offset[grid + 1]) grid++; // yuck search for grid 253 { 254 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]; 255 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]; 256 PetscScalar *coef, coef_buff[LANDAU_MAX_SPECIES * LANDAU_MAX_NQND]; 257 PetscReal *invJe = &invJ_a[(ip_offset[grid] + loc_elem * Nq) * dim * dim]; // ingJ is static data on batch 0 258 PetscInt b, f, q; 259 if (cellClosure) { 260 coef = &cellClosure[b_id * IPf_sz_glb + ipf_offset[grid] + loc_elem * Nb * loc_Nf]; // this is const 261 } else { 262 coef = coef_buff; 263 for (f = 0; f < loc_Nf; ++f) { 264 LandauIdx *const Idxs = &maps[grid].gIdx[loc_elem][f][0]; 265 for (b = 0; b < Nb; ++b) { 266 PetscInt idx = Idxs[b]; 267 if (idx >= 0) { 268 coef[f * Nb + b] = xdata[idx + moffset]; 269 } else { 270 idx = -idx - 1; 271 coef[f * Nb + b] = 0; 272 for (q = 0; q < maps[grid].num_face; q++) { 273 PetscInt id = maps[grid].c_maps[idx][q].gid; 274 PetscScalar scale = maps[grid].c_maps[idx][q].scale; 275 coef[f * Nb + b] += scale * xdata[id + moffset]; 276 } 277 } 278 } 279 } 280 } 281 /* get f and df */ 282 for (PetscInt qi = 0; qi < Nq; qi++) { 283 const PetscReal *invJ = &invJe[qi * dim * dim]; 284 const PetscReal *Bq = &BB[qi * Nb]; 285 const PetscReal *Dq = &DD[qi * Nb * dim]; 286 PetscReal u_x[LANDAU_DIM]; 287 /* get f & df */ 288 for (f = 0; f < loc_Nf; ++f) { 289 const PetscInt idx = b_id * IPf_sz_glb + ipf_offset[grid] + f * loc_nip + loc_elem * Nq + qi; 290 PetscInt b, e; 291 PetscReal refSpaceDer[LANDAU_DIM]; 292 ff[idx] = 0.0; 293 for (int d = 0; d < LANDAU_DIM; ++d) refSpaceDer[d] = 0.0; 294 for (b = 0; b < Nb; ++b) { 295 const PetscInt cidx = b; 296 ff[idx] += Bq[cidx] * PetscRealPart(coef[f * Nb + cidx]); 297 for (int d = 0; d < dim; ++d) refSpaceDer[d] += Dq[cidx * dim + d] * PetscRealPart(coef[f * Nb + cidx]); 298 } 299 for (int d = 0; d < LANDAU_DIM; ++d) { 300 for (e = 0, u_x[d] = 0.0; e < LANDAU_DIM; ++e) u_x[d] += invJ[e * dim + d] * refSpaceDer[e]; 301 } 302 dudx[idx] = u_x[0]; 303 dudy[idx] = u_x[1]; 304 #if LANDAU_DIM == 3 305 dudz[idx] = u_x[2]; 306 #endif 307 } 308 } // q 309 } // grid 310 } // grid*batch 311 PetscCall(PetscLogEventEnd(ctx->events[8], 0, 0, 0, 0)); 312 #if defined(PETSC_HAVE_THREADSAFETY) 313 endtime = MPI_Wtime(); 314 if (ctx->stage) ctx->times[LANDAU_F_DF] += (endtime - starttime); 315 #endif 316 } // Jacobian setup 317 // assemble Jacobian (or mass) 318 for (PetscInt tid = 0; tid < ctx->batch_sz * elem_offset[num_grids]; tid++) { // for each element 319 const PetscInt b_Nelem = elem_offset[num_grids]; 320 const PetscInt glb_elem_idx = tid % b_Nelem, b_id = tid / b_Nelem; 321 PetscInt grid = 0; 322 #if defined(PETSC_HAVE_THREADSAFETY) 323 double starttime, endtime; 324 starttime = MPI_Wtime(); 325 #endif 326 while (glb_elem_idx >= elem_offset[grid + 1]) grid++; 327 { 328 const PetscInt loc_Nf = ctx->species_offset[grid + 1] - ctx->species_offset[grid], loc_elem = glb_elem_idx - elem_offset[grid]; 329 const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset), totDim = loc_Nf * Nq, elemMatSize = totDim * totDim; 330 PetscScalar *elemMat; 331 const PetscReal *invJe = &invJ_a[(ip_offset[grid] + loc_elem * Nq) * dim * dim]; 332 PetscCall(PetscMalloc1(elemMatSize, &elemMat)); 333 PetscCall(PetscMemzero(elemMat, elemMatSize * sizeof(*elemMat))); 334 if (shift == 0.0) { // Jacobian 335 PetscCall(PetscLogEventBegin(ctx->events[4], 0, 0, 0, 0)); 336 } else { // mass 337 PetscCall(PetscLogEventBegin(ctx->events[16], 0, 0, 0, 0)); 338 } 339 for (PetscInt qj = 0; qj < Nq; ++qj) { 340 const PetscInt jpidx_glb = ip_offset[grid] + qj + loc_elem * Nq; 341 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 - 1 342 PetscInt d, d2, dp, d3, IPf_idx; 343 if (shift == 0.0) { // Jacobian 344 const PetscReal *const invJj = &invJe[qj * dim * dim]; 345 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]; 346 const PetscReal vj[3] = {xx[jpidx_glb], yy[jpidx_glb], zz ? zz[jpidx_glb] : 0}, wj = ww[jpidx_glb]; 347 // create g2 & g3 348 for (d = 0; d < LANDAU_DIM; d++) { // clear accumulation data D & K 349 gg2_temp[d] = 0; 350 for (d2 = 0; d2 < LANDAU_DIM; d2++) gg3_temp[d][d2] = 0; 351 } 352 /* inner beta reduction */ 353 IPf_idx = 0; 354 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 355 PetscInt nip_loc_r = numCells[grid_r] * Nq, Nfloc_r = Nf[grid_r]; 356 for (PetscInt ei_r = 0, loc_fdf_idx = 0; ei_r < numCells[grid_r]; ++ei_r) { 357 for (PetscInt qi = 0; qi < Nq; qi++, ipidx++, loc_fdf_idx++) { 358 const PetscReal wi = ww[ipidx], x = xx[ipidx], y = yy[ipidx]; 359 PetscReal temp1[3] = {0, 0, 0}, temp2 = 0; 360 #if LANDAU_DIM == 2 361 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.; 362 LandauTensor2D(vj, x, y, Ud, Uk, mask); 363 #else 364 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.; 365 if (ctx->use_relativistic_corrections) { 366 LandauTensor3DRelativistic(vj, x, y, z, U, mask, C_0(ctx->v_0)); 367 } else { 368 LandauTensor3D(vj, x, y, z, U, mask); 369 } 370 #endif 371 for (int f = 0; f < Nfloc_r; ++f) { 372 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; 373 temp1[0] += dudx[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r]; 374 temp1[1] += dudy[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r]; 375 #if LANDAU_DIM == 3 376 temp1[2] += dudz[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r]; 377 #endif 378 temp2 += ff[idx] * nu_beta[f + f_off] * (*lambdas)[grid][grid_r]; 379 } 380 temp1[0] *= wi; 381 temp1[1] *= wi; 382 #if LANDAU_DIM == 3 383 temp1[2] *= wi; 384 #endif 385 temp2 *= wi; 386 #if LANDAU_DIM == 2 387 for (d2 = 0; d2 < 2; d2++) { 388 for (d3 = 0; d3 < 2; ++d3) { 389 /* K = U * grad(f): g2=e: i,A */ 390 gg2_temp[d2] += Uk[d2][d3] * temp1[d3]; 391 /* D = -U * (I \kron (fx)): g3=f: i,j,A */ 392 gg3_temp[d2][d3] += Ud[d2][d3] * temp2; 393 } 394 } 395 #else 396 for (d2 = 0; d2 < 3; ++d2) { 397 for (d3 = 0; d3 < 3; ++d3) { 398 /* K = U * grad(f): g2 = e: i,A */ 399 gg2_temp[d2] += U[d2][d3] * temp1[d3]; 400 /* D = -U * (I \kron (fx)): g3 = f: i,j,A */ 401 gg3_temp[d2][d3] += U[d2][d3] * temp2; 402 } 403 } 404 #endif 405 } // qi 406 } // ei_r 407 IPf_idx += nip_loc_r * Nfloc_r; 408 } /* grid_r - IPs */ 409 PetscCheck(IPf_idx == IPf_sz_glb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "IPf_idx != IPf_sz %" PetscInt_FMT " %" PetscInt_FMT, IPf_idx, IPf_sz_glb); 410 // add alpha and put in gg2/3 411 for (PetscInt fieldA = 0, f_off = ctx->species_offset[grid]; fieldA < loc_Nf; ++fieldA) { 412 for (d2 = 0; d2 < LANDAU_DIM; d2++) { 413 gg2[fieldA][d2] = gg2_temp[d2] * nu_alpha[fieldA + f_off]; 414 for (d3 = 0; d3 < LANDAU_DIM; d3++) gg3[fieldA][d2][d3] = -gg3_temp[d2][d3] * nu_alpha[fieldA + f_off] * invMass[fieldA + f_off]; 415 } 416 } 417 /* add electric field term once per IP */ 418 for (PetscInt fieldA = 0, f_off = ctx->species_offset[grid]; fieldA < loc_Nf; ++fieldA) gg2[fieldA][LANDAU_DIM - 1] += Eq_m[fieldA + f_off]; 419 /* Jacobian transform - g2, g3 */ 420 for (PetscInt fieldA = 0; fieldA < loc_Nf; ++fieldA) { 421 for (d = 0; d < dim; ++d) { 422 g2[fieldA][d] = 0.0; 423 for (d2 = 0; d2 < dim; ++d2) { 424 g2[fieldA][d] += invJj[d * dim + d2] * gg2[fieldA][d2]; 425 g3[fieldA][d][d2] = 0.0; 426 for (d3 = 0; d3 < dim; ++d3) { 427 for (dp = 0; dp < dim; ++dp) g3[fieldA][d][d2] += invJj[d * dim + d3] * gg3[fieldA][d3][dp] * invJj[d2 * dim + dp]; 428 } 429 g3[fieldA][d][d2] *= wj; 430 } 431 g2[fieldA][d] *= wj; 432 } 433 } 434 } else { // mass 435 PetscReal wj = ww[jpidx_glb]; 436 /* Jacobian transform - g0 */ 437 for (PetscInt fieldA = 0; fieldA < loc_Nf; ++fieldA) { 438 if (dim == 2) { 439 g0[fieldA] = wj * shift * 2. * PETSC_PI; // move this to below and remove g0 440 } else { 441 g0[fieldA] = wj * shift; // move this to below and remove g0 442 } 443 } 444 } 445 /* FE matrix construction */ 446 { 447 PetscInt fieldA, d, f, d2, g; 448 const PetscReal *BJq = &BB[qj * Nb], *DIq = &DD[qj * Nb * dim]; 449 /* assemble - on the diagonal (I,I) */ 450 for (fieldA = 0; fieldA < loc_Nf; fieldA++) { 451 for (f = 0; f < Nb; f++) { 452 const PetscInt i = fieldA * Nb + f; /* Element matrix row */ 453 for (g = 0; g < Nb; ++g) { 454 const PetscInt j = fieldA * Nb + g; /* Element matrix column */ 455 const PetscInt fOff = i * totDim + j; 456 if (shift == 0.0) { 457 for (d = 0; d < dim; ++d) { 458 elemMat[fOff] += DIq[f * dim + d] * g2[fieldA][d] * BJq[g]; 459 for (d2 = 0; d2 < dim; ++d2) elemMat[fOff] += DIq[f * dim + d] * g3[fieldA][d][d2] * DIq[g * dim + d2]; 460 } 461 } else { // mass 462 elemMat[fOff] += BJq[f] * g0[fieldA] * BJq[g]; 463 } 464 } 465 } 466 } 467 } 468 } /* qj loop */ 469 if (shift == 0.0) { // Jacobian 470 PetscCall(PetscLogEventEnd(ctx->events[4], 0, 0, 0, 0)); 471 } else { 472 PetscCall(PetscLogEventEnd(ctx->events[16], 0, 0, 0, 0)); 473 } 474 #if defined(PETSC_HAVE_THREADSAFETY) 475 endtime = MPI_Wtime(); 476 if (ctx->stage) ctx->times[LANDAU_KERNEL] += (endtime - starttime); 477 #endif 478 /* assemble matrix */ 479 if (!container) { 480 PetscInt cStart; 481 PetscCall(PetscLogEventBegin(ctx->events[6], 0, 0, 0, 0)); 482 PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, NULL)); 483 PetscCall(DMPlexMatSetClosure(ctx->plex[grid], section[grid], globsection[grid], subJ[LAND_PACK_IDX(b_id, grid)], loc_elem + cStart, elemMat, ADD_VALUES)); 484 PetscCall(PetscLogEventEnd(ctx->events[6], 0, 0, 0, 0)); 485 } else { // GPU like assembly for debugging 486 PetscInt fieldA, q, f, g, d, nr, nc, rows0[LANDAU_MAX_Q_FACE] = {0}, cols0[LANDAU_MAX_Q_FACE] = {0}, rows[LANDAU_MAX_Q_FACE], cols[LANDAU_MAX_Q_FACE]; 487 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}; 488 LandauIdx *coo_elem_offsets = (LandauIdx *)ctx->SData_d.coo_elem_offsets, *coo_elem_fullNb = (LandauIdx *)ctx->SData_d.coo_elem_fullNb, (*coo_elem_point_offsets)[LANDAU_MAX_NQND + 1] = (LandauIdx(*)[LANDAU_MAX_NQND + 1]) ctx->SData_d.coo_elem_point_offsets; 489 /* assemble - from the diagonal (I,I) in this format for DMPlexMatSetClosure */ 490 for (fieldA = 0; fieldA < loc_Nf; fieldA++) { 491 LandauIdx *const Idxs = &maps[grid].gIdx[loc_elem][fieldA][0]; 492 for (f = 0; f < Nb; f++) { 493 PetscInt idx = Idxs[f]; 494 if (idx >= 0) { 495 nr = 1; 496 rows0[0] = idx; 497 row_scale[0] = 1.; 498 } else { 499 idx = -idx - 1; 500 for (q = 0, nr = 0; q < maps[grid].num_face; q++, nr++) { 501 if (maps[grid].c_maps[idx][q].gid < 0) break; 502 rows0[q] = maps[grid].c_maps[idx][q].gid; 503 row_scale[q] = maps[grid].c_maps[idx][q].scale; 504 } 505 } 506 for (g = 0; g < Nb; ++g) { 507 idx = Idxs[g]; 508 if (idx >= 0) { 509 nc = 1; 510 cols0[0] = idx; 511 col_scale[0] = 1.; 512 } else { 513 idx = -idx - 1; 514 nc = maps[grid].num_face; 515 for (q = 0, nc = 0; q < maps[grid].num_face; q++, nc++) { 516 if (maps[grid].c_maps[idx][q].gid < 0) break; 517 cols0[q] = maps[grid].c_maps[idx][q].gid; 518 col_scale[q] = maps[grid].c_maps[idx][q].scale; 519 } 520 } 521 const PetscInt i = fieldA * Nb + f; /* Element matrix row */ 522 const PetscInt j = fieldA * Nb + g; /* Element matrix column */ 523 const PetscScalar Aij = elemMat[i * totDim + j]; 524 if (coo_vals) { // mirror (i,j) in CreateStaticGPUData 525 const int fullNb = coo_elem_fullNb[glb_elem_idx], fullNb2 = fullNb * fullNb; 526 const int idx0 = b_id * coo_elem_offsets[elem_offset[num_grids]] + coo_elem_offsets[glb_elem_idx] + fieldA * fullNb2 + fullNb * coo_elem_point_offsets[glb_elem_idx][f] + nr * coo_elem_point_offsets[glb_elem_idx][g]; 527 for (int q = 0, idx2 = idx0; q < nr; q++) { 528 for (int d = 0; d < nc; d++, idx2++) coo_vals[idx2] = row_scale[q] * col_scale[d] * Aij; 529 } 530 } else { 531 for (q = 0; q < nr; q++) rows[q] = rows0[q] + moffset; 532 for (d = 0; d < nc; d++) cols[d] = cols0[d] + moffset; 533 for (q = 0; q < nr; q++) { 534 for (d = 0; d < nc; d++) vals[q * nc + d] = row_scale[q] * col_scale[d] * Aij; 535 } 536 PetscCall(MatSetValues(JacP, nr, rows, nc, cols, vals, ADD_VALUES)); 537 } 538 } 539 } 540 } 541 } 542 if (loc_elem == -1) { 543 PetscCall(PetscPrintf(ctx->comm, "CPU Element matrix\n")); 544 for (int d = 0; d < totDim; ++d) { 545 for (int f = 0; f < totDim; ++f) PetscCall(PetscPrintf(ctx->comm, " %12.5e", (double)PetscRealPart(elemMat[d * totDim + f]))); 546 PetscCall(PetscPrintf(ctx->comm, "\n")); 547 } 548 exit(12); 549 } 550 PetscCall(PetscFree(elemMat)); 551 } /* grid */ 552 } /* outer element & batch loop */ 553 if (shift == 0.0) { // mass 554 PetscCall(PetscFree4(ff, dudx, dudy, dudz)); 555 } 556 if (!container) { // 'CPU' assembly move nest matrix to global JacP 557 for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // OpenMP 558 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 559 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]; 560 PetscInt nloc, nzl, colbuf[1024], row; 561 const PetscInt *cols; 562 const PetscScalar *vals; 563 Mat B = subJ[LAND_PACK_IDX(b_id, grid)]; 564 PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY)); 565 PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY)); 566 PetscCall(MatGetSize(B, &nloc, NULL)); 567 for (int i = 0; i < nloc; i++) { 568 PetscCall(MatGetRow(B, i, &nzl, &cols, &vals)); 569 PetscCheck(nzl <= 1024, PetscObjectComm((PetscObject)B), PETSC_ERR_PLIB, "Row too big: %" PetscInt_FMT, nzl); 570 for (int j = 0; j < nzl; j++) colbuf[j] = moffset + cols[j]; 571 row = moffset + i; 572 PetscCall(MatSetValues(JacP, 1, &row, nzl, colbuf, vals, ADD_VALUES)); 573 PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals)); 574 } 575 PetscCall(MatDestroy(&B)); 576 } 577 } 578 } 579 if (coo_vals) { 580 PetscCall(MatSetValuesCOO(JacP, coo_vals, ADD_VALUES)); 581 PetscCall(PetscFree(coo_vals)); 582 } 583 } /* CPU version */ 584 PetscCall(MatAssemblyBegin(JacP, MAT_FINAL_ASSEMBLY)); 585 PetscCall(MatAssemblyEnd(JacP, MAT_FINAL_ASSEMBLY)); 586 /* clean up */ 587 if (cellClosure) PetscCall(PetscFree(cellClosure)); 588 if (xdata) PetscCall(VecRestoreArrayReadAndMemType(a_X, &xdata)); 589 PetscFunctionReturn(PETSC_SUCCESS); 590 } 591 592 static PetscErrorCode GeometryDMLandau(DM base, PetscInt point, PetscInt dim, const PetscReal abc[], PetscReal xyz[], void *a_ctx) 593 { 594 PetscReal r = abc[0], z = abc[1]; 595 596 PetscFunctionBegin; 597 xyz[0] = r; 598 xyz[1] = z; 599 if (dim == 3) xyz[2] = abc[2]; 600 601 PetscFunctionReturn(PETSC_SUCCESS); 602 } 603 604 /* create DMComposite of meshes for each species group */ 605 static PetscErrorCode LandauDMCreateVMeshes(MPI_Comm comm_self, const PetscInt dim, const char prefix[], LandauCtx *ctx, DM pack) 606 { 607 PetscFunctionBegin; 608 { /* p4est, quads */ 609 /* Create plex mesh of Landau domain */ 610 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 611 PetscReal par_radius = ctx->radius_par[grid], perp_radius = ctx->radius_perp[grid]; 612 if (!ctx->sphere && !ctx->simplex) { // 2 or 3D (only 3D option) 613 PetscReal lo[] = {-perp_radius, -par_radius, -par_radius}, hi[] = {perp_radius, par_radius, par_radius}; 614 DMBoundaryType periodicity[3] = {DM_BOUNDARY_NONE, dim == 2 ? DM_BOUNDARY_NONE : DM_BOUNDARY_NONE, DM_BOUNDARY_NONE}; 615 if (dim == 2) lo[0] = 0; 616 else { 617 lo[1] = -perp_radius; 618 hi[1] = perp_radius; // 3D y is a perp 619 } 620 PetscCall(DMPlexCreateBoxMesh(comm_self, dim, PETSC_FALSE, ctx->cells0, lo, hi, periodicity, PETSC_TRUE, &ctx->plex[grid])); // todo: make composite and create dm[grid] here 621 PetscCall(DMLocalizeCoordinates(ctx->plex[grid])); /* needed for periodic */ 622 if (dim == 3) PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "cube")); 623 else PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "half-plane")); 624 } else if (dim == 2) { 625 size_t len; 626 PetscCall(PetscStrlen(ctx->filename, &len)); 627 if (len) { 628 Vec coords; 629 PetscScalar *x; 630 PetscInt N; 631 char str[] = "-dm_landau_view_file_0"; 632 str[21] += grid; 633 PetscCall(DMPlexCreateFromFile(comm_self, ctx->filename, "plexland.c", PETSC_TRUE, &ctx->plex[grid])); 634 PetscCall(DMPlexOrient(ctx->plex[grid])); 635 PetscCall(DMGetCoordinatesLocal(ctx->plex[grid], &coords)); 636 PetscCall(VecGetSize(coords, &N)); 637 PetscCall(VecGetArray(coords, &x)); 638 /* scale by domain size */ 639 for (PetscInt i = 0; i < N; i += 2) { 640 x[i + 0] *= ctx->radius_perp[grid]; 641 x[i + 1] *= ctx->radius_par[grid]; 642 } 643 PetscCall(VecRestoreArray(coords, &x)); 644 PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], ctx->filename)); 645 PetscCall(PetscInfo(ctx->plex[grid], "%d) Read %s mesh file (%s)\n", (int)grid, ctx->filename, str)); 646 PetscCall(DMViewFromOptions(ctx->plex[grid], NULL, str)); 647 } else { 648 PetscInt numCells = ctx->simplex ? 12 : 6, cell_size = ctx->simplex ? 3 : 4, j; 649 const PetscInt numVerts = 11; 650 PetscInt cellsT[][4] = { 651 {0, 1, 6, 5 }, 652 {1, 2, 7, 6 }, 653 {2, 3, 8, 7 }, 654 {3, 4, 9, 8 }, 655 {5, 6, 7, 10}, 656 {10, 7, 8, 9 } 657 }; 658 PetscInt cellsS[][3] = { 659 {0, 1, 6 }, 660 {1, 2, 6 }, 661 {6, 2, 7 }, 662 {7, 2, 8 }, 663 {8, 2, 3 }, 664 {8, 3, 4 }, 665 {0, 6, 5 }, 666 {5, 6, 7 }, 667 {5, 7, 10}, 668 {10, 7, 9 }, 669 {9, 7, 8 }, 670 {9, 8, 4 } 671 }; 672 const PetscInt *pcell = (const PetscInt *)(ctx->simplex ? &cellsS[0][0] : &cellsT[0][0]); 673 PetscReal coords[11][2], *flatCoords = (PetscReal *)&coords[0][0]; 674 PetscReal rad = ctx->radius[grid]; 675 for (j = 0; j < 5; j++) { // outside edge 676 PetscReal z, r, theta = -PETSC_PI / 2 + (j % 5) * PETSC_PI / 4; 677 r = rad * PetscCosReal(theta); 678 coords[j][0] = r; 679 z = rad * PetscSinReal(theta); 680 coords[j][1] = z; 681 } 682 coords[j][0] = 0; 683 coords[j++][1] = -rad * ctx->sphere_inner_radius_90degree; 684 coords[j][0] = rad * ctx->sphere_inner_radius_45degree; 685 coords[j++][1] = -rad * ctx->sphere_inner_radius_45degree; 686 coords[j][0] = rad * ctx->sphere_inner_radius_90degree; 687 coords[j++][1] = 0; 688 coords[j][0] = rad * ctx->sphere_inner_radius_45degree; 689 coords[j++][1] = rad * ctx->sphere_inner_radius_45degree; 690 coords[j][0] = 0; 691 coords[j++][1] = rad * ctx->sphere_inner_radius_90degree; 692 coords[j][0] = 0; 693 coords[j++][1] = 0; 694 PetscCall(DMPlexCreateFromCellListPetsc(comm_self, 2, numCells, numVerts, cell_size, ctx->interpolate, pcell, 2, flatCoords, &ctx->plex[grid])); 695 PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "semi-circle")); 696 PetscCall(PetscInfo(ctx->plex[grid], "\t%" PetscInt_FMT ") Make circle %s mesh\n", grid, ctx->simplex ? "simplex" : "tensor")); 697 } 698 } else SETERRQ(ctx->comm, PETSC_ERR_PLIB, "Velocity space meshes does not support 3V cubed sphere or simplex"); 699 PetscCall(DMSetFromOptions(ctx->plex[grid])); 700 } // grid loop 701 PetscCall(PetscObjectSetOptionsPrefix((PetscObject)pack, prefix)); 702 { /* convert to p4est (or whatever), wait for discretization to create pack */ 703 char convType[256]; 704 PetscBool flg; 705 706 PetscOptionsBegin(ctx->comm, prefix, "Mesh conversion options", "DMPLEX"); 707 PetscCall(PetscOptionsFList("-dm_landau_type", "Convert DMPlex to another format (p4est)", "plexland.c", DMList, DMPLEX, convType, 256, &flg)); 708 PetscOptionsEnd(); 709 if (flg) { 710 ctx->use_p4est = PETSC_TRUE; /* flag for Forest */ 711 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 712 DM dmforest; 713 PetscCall(DMConvert(ctx->plex[grid], convType, &dmforest)); 714 if (dmforest) { 715 PetscBool isForest; 716 PetscCall(PetscObjectSetOptionsPrefix((PetscObject)dmforest, prefix)); 717 PetscCall(DMIsForest(dmforest, &isForest)); 718 if (isForest) { 719 if (ctx->sphere) PetscCall(DMForestSetBaseCoordinateMapping(dmforest, GeometryDMLandau, ctx)); 720 PetscCall(DMDestroy(&ctx->plex[grid])); 721 ctx->plex[grid] = dmforest; // Forest for adaptivity 722 } else SETERRQ(ctx->comm, PETSC_ERR_PLIB, "Converted to non Forest?"); 723 } else SETERRQ(ctx->comm, PETSC_ERR_PLIB, "Convert failed?"); 724 } 725 } else ctx->use_p4est = PETSC_FALSE; /* flag for Forest */ 726 } 727 } /* non-file */ 728 PetscCall(DMSetDimension(pack, dim)); 729 PetscCall(PetscObjectSetName((PetscObject)pack, "Mesh")); 730 PetscCall(DMSetApplicationContext(pack, ctx)); 731 732 PetscFunctionReturn(PETSC_SUCCESS); 733 } 734 735 static PetscErrorCode SetupDS(DM pack, PetscInt dim, PetscInt grid, LandauCtx *ctx) 736 { 737 PetscInt ii, i0; 738 char buf[256]; 739 PetscSection section; 740 741 PetscFunctionBegin; 742 for (ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) { 743 if (ii == 0) PetscCall(PetscSNPrintf(buf, sizeof(buf), "e")); 744 else PetscCall(PetscSNPrintf(buf, sizeof(buf), "i%" PetscInt_FMT, ii)); 745 /* Setup Discretization - FEM */ 746 PetscCall(PetscFECreateDefault(PETSC_COMM_SELF, dim, 1, ctx->simplex, NULL, PETSC_DECIDE, &ctx->fe[ii])); 747 PetscCall(PetscObjectSetName((PetscObject)ctx->fe[ii], buf)); 748 PetscCall(DMSetField(ctx->plex[grid], i0, NULL, (PetscObject)ctx->fe[ii])); 749 } 750 PetscCall(DMCreateDS(ctx->plex[grid])); 751 PetscCall(DMGetSection(ctx->plex[grid], §ion)); 752 for (PetscInt ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) { 753 if (ii == 0) PetscCall(PetscSNPrintf(buf, sizeof(buf), "se")); 754 else PetscCall(PetscSNPrintf(buf, sizeof(buf), "si%" PetscInt_FMT, ii)); 755 PetscCall(PetscSectionSetComponentName(section, i0, 0, buf)); 756 } 757 PetscFunctionReturn(PETSC_SUCCESS); 758 } 759 760 /* Define a Maxwellian function for testing out the operator. */ 761 762 /* Using cartesian velocity space coordinates, the particle */ 763 /* density, [1/m^3], is defined according to */ 764 765 /* $$ n=\int_{R^3} dv^3 \left(\frac{m}{2\pi T}\right)^{3/2}\exp [- mv^2/(2T)] $$ */ 766 767 /* Using some constant, c, we normalize the velocity vector into a */ 768 /* dimensionless variable according to v=c*x. Thus the density, $n$, becomes */ 769 770 /* $$ n=\int_{R^3} dx^3 \left(\frac{mc^2}{2\pi T}\right)^{3/2}\exp [- mc^2/(2T)*x^2] $$ */ 771 772 /* Defining $\theta=2T/mc^2$, we thus find that the probability density */ 773 /* for finding the particle within the interval in a box dx^3 around x is */ 774 775 /* f(x;\theta)=\left(\frac{1}{\pi\theta}\right)^{3/2} \exp [ -x^2/\theta ] */ 776 777 typedef struct { 778 PetscReal v_0; 779 PetscReal kT_m; 780 PetscReal n; 781 PetscReal shift; 782 } MaxwellianCtx; 783 784 static PetscErrorCode maxwellian(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx) 785 { 786 MaxwellianCtx *mctx = (MaxwellianCtx *)actx; 787 PetscInt i; 788 PetscReal v2 = 0, theta = 2 * mctx->kT_m / (mctx->v_0 * mctx->v_0), shift; /* theta = 2kT/mc^2 */ 789 PetscFunctionBegin; 790 /* compute the exponents, v^2 */ 791 for (i = 0; i < dim; ++i) v2 += x[i] * x[i]; 792 /* evaluate the Maxwellian */ 793 if (mctx->shift < 0) shift = -mctx->shift; 794 else { 795 u[0] = mctx->n * PetscPowReal(PETSC_PI * theta, -1.5) * (PetscExpReal(-v2 / theta)); 796 shift = mctx->shift; 797 } 798 if (shift != 0.) { 799 v2 = 0; 800 for (i = 0; i < dim - 1; ++i) v2 += x[i] * x[i]; 801 v2 += (x[dim - 1] - shift) * (x[dim - 1] - shift); 802 /* evaluate the shifted Maxwellian */ 803 u[0] += mctx->n * PetscPowReal(PETSC_PI * theta, -1.5) * (PetscExpReal(-v2 / theta)); 804 } 805 PetscFunctionReturn(PETSC_SUCCESS); 806 } 807 808 /*@ 809 DMPlexLandauAddMaxwellians - Add a Maxwellian distribution to a state 810 811 Collective 812 813 Input Parameters: 814 + dm - The mesh (local) 815 . time - Current time 816 . temps - Temperatures of each species (global) 817 . ns - Number density of each species (global) 818 . grid - index into current grid - just used for offset into `temp` and `ns` 819 . b_id - batch index 820 . n_batch - number of batches 821 - actx - Landau context 822 823 Output Parameter: 824 . X - The state (local to this grid) 825 826 Level: beginner 827 828 .seealso: `DMPlexLandauCreateVelocitySpace()` 829 @*/ 830 PetscErrorCode DMPlexLandauAddMaxwellians(DM dm, Vec X, PetscReal time, PetscReal temps[], PetscReal ns[], PetscInt grid, PetscInt b_id, PetscInt n_batch, void *actx) 831 { 832 LandauCtx *ctx = (LandauCtx *)actx; 833 PetscErrorCode (*initu[LANDAU_MAX_SPECIES])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *); 834 PetscInt dim; 835 MaxwellianCtx *mctxs[LANDAU_MAX_SPECIES], data[LANDAU_MAX_SPECIES]; 836 837 PetscFunctionBegin; 838 PetscCall(DMGetDimension(dm, &dim)); 839 if (!ctx) PetscCall(DMGetApplicationContext(dm, &ctx)); 840 for (PetscInt ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) { 841 mctxs[i0] = &data[i0]; 842 data[i0].v_0 = ctx->v_0; // v_0 same for all grids 843 data[i0].kT_m = ctx->k * temps[ii] / ctx->masses[ii]; /* kT/m */ 844 data[i0].n = ns[ii]; 845 initu[i0] = maxwellian; 846 data[i0].shift = 0; 847 } 848 data[0].shift = ctx->electronShift; 849 /* need to make ADD_ALL_VALUES work - TODO */ 850 PetscCall(DMProjectFunction(dm, time, initu, (void **)mctxs, INSERT_ALL_VALUES, X)); 851 PetscFunctionReturn(PETSC_SUCCESS); 852 } 853 854 /* 855 LandauSetInitialCondition - Adds Maxwellians with context 856 857 Collective 858 859 Input Parameters: 860 . dm - The mesh 861 - grid - index into current grid - just used for offset into temp and ns 862 . b_id - batch index 863 - n_batch - number of batches 864 + actx - Landau context with T and n 865 866 Output Parameter: 867 . X - The state 868 869 Level: beginner 870 871 .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauAddMaxwellians()` 872 */ 873 static PetscErrorCode LandauSetInitialCondition(DM dm, Vec X, PetscInt grid, PetscInt b_id, PetscInt n_batch, void *actx) 874 { 875 LandauCtx *ctx = (LandauCtx *)actx; 876 PetscFunctionBegin; 877 if (!ctx) PetscCall(DMGetApplicationContext(dm, &ctx)); 878 PetscCall(VecZeroEntries(X)); 879 PetscCall(DMPlexLandauAddMaxwellians(dm, X, 0.0, ctx->thermal_temps, ctx->n, grid, b_id, n_batch, ctx)); 880 PetscFunctionReturn(PETSC_SUCCESS); 881 } 882 883 // adapt a level once. Forest in/out 884 #if defined(PETSC_USE_INFO) 885 static const char *s_refine_names[] = {"RE", "Z1", "Origin", "Z2", "Uniform"}; 886 #endif 887 static PetscErrorCode adaptToleranceFEM(PetscFE fem, Vec sol, PetscInt type, PetscInt grid, LandauCtx *ctx, DM *newForest) 888 { 889 DM forest, plex, adaptedDM = NULL; 890 PetscDS prob; 891 PetscBool isForest; 892 PetscQuadrature quad; 893 PetscInt Nq, Nb, *Nb2, cStart, cEnd, c, dim, qj, k; 894 DMLabel adaptLabel = NULL; 895 896 PetscFunctionBegin; 897 forest = ctx->plex[grid]; 898 PetscCall(DMCreateDS(forest)); 899 PetscCall(DMGetDS(forest, &prob)); 900 PetscCall(DMGetDimension(forest, &dim)); 901 PetscCall(DMIsForest(forest, &isForest)); 902 PetscCheck(isForest, ctx->comm, PETSC_ERR_ARG_WRONG, "! Forest"); 903 PetscCall(DMConvert(forest, DMPLEX, &plex)); 904 PetscCall(DMPlexGetHeightStratum(plex, 0, &cStart, &cEnd)); 905 PetscCall(DMLabelCreate(PETSC_COMM_SELF, "adapt", &adaptLabel)); 906 PetscCall(PetscFEGetQuadrature(fem, &quad)); 907 PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, NULL)); 908 PetscCheck(Nq <= LANDAU_MAX_NQND, ctx->comm, PETSC_ERR_ARG_WRONG, "Order too high. Nq = %" PetscInt_FMT " > LANDAU_MAX_NQND (%d)", Nq, LANDAU_MAX_NQND); 909 PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb)); 910 PetscCall(PetscDSGetDimensions(prob, &Nb2)); 911 PetscCheck(Nb2[0] == Nb, ctx->comm, PETSC_ERR_ARG_WRONG, " Nb = %" PetscInt_FMT " != Nb (%d)", Nb, (int)Nb2[0]); 912 PetscCheck(Nb <= LANDAU_MAX_NQND, ctx->comm, PETSC_ERR_ARG_WRONG, "Order too high. Nb = %" PetscInt_FMT " > LANDAU_MAX_NQND (%d)", Nb, LANDAU_MAX_NQND); 913 PetscCall(PetscInfo(sol, "%" PetscInt_FMT ") Refine phase: %s\n", grid, s_refine_names[type])); 914 if (type == 4) { 915 for (c = cStart; c < cEnd; c++) PetscCall(DMLabelSetValue(adaptLabel, c, DM_ADAPT_REFINE)); 916 } else if (type == 2) { 917 PetscInt rCellIdx[8], nr = 0, nrmax = (dim == 3) ? 8 : 2; 918 PetscReal minRad = PETSC_INFINITY, r; 919 for (c = cStart; c < cEnd; c++) { 920 PetscReal tt, v0[LANDAU_MAX_NQND * 3], detJ[LANDAU_MAX_NQND]; 921 PetscCall(DMPlexComputeCellGeometryFEM(plex, c, quad, v0, NULL, NULL, detJ)); 922 for (qj = 0; qj < Nq; ++qj) { 923 tt = PetscSqr(v0[dim * qj + 0]) + PetscSqr(v0[dim * qj + 1]) + PetscSqr(((dim == 3) ? v0[dim * qj + 2] : 0)); 924 r = PetscSqrtReal(tt); 925 if (r < minRad - PETSC_SQRT_MACHINE_EPSILON * 10.) { 926 minRad = r; 927 nr = 0; 928 rCellIdx[nr++] = c; 929 PetscCall(PetscInfo(sol, "\t\t%" PetscInt_FMT ") Found first inner r=%e, cell %" PetscInt_FMT ", qp %" PetscInt_FMT "/%" PetscInt_FMT "\n", grid, (double)r, c, qj + 1, Nq)); 930 } else if ((r - minRad) < PETSC_SQRT_MACHINE_EPSILON * 100. && nr < nrmax) { 931 for (k = 0; k < nr; k++) 932 if (c == rCellIdx[k]) break; 933 if (k == nr) { 934 rCellIdx[nr++] = c; 935 PetscCall(PetscInfo(sol, "\t\t\t%" PetscInt_FMT ") Found another inner r=%e, cell %" PetscInt_FMT ", qp %" PetscInt_FMT "/%" PetscInt_FMT ", d=%e\n", grid, (double)r, c, qj + 1, Nq, (double)(r - minRad))); 936 } 937 } 938 } 939 } 940 for (k = 0; k < nr; k++) PetscCall(DMLabelSetValue(adaptLabel, rCellIdx[k], DM_ADAPT_REFINE)); 941 PetscCall(PetscInfo(sol, "\t\t\t%" PetscInt_FMT ") Refined %" PetscInt_FMT " origin cells %" PetscInt_FMT ",%" PetscInt_FMT " r=%g\n", grid, nr, rCellIdx[0], rCellIdx[1], (double)minRad)); 942 } else if (type == 0 || type == 1 || type == 3) { /* refine along r=0 axis */ 943 PetscScalar *coef = NULL; 944 Vec coords; 945 PetscInt csize, Nv, d, nz, nrefined = 0; 946 DM cdm; 947 PetscSection cs; 948 PetscCall(DMGetCoordinatesLocal(forest, &coords)); 949 PetscCall(DMGetCoordinateDM(forest, &cdm)); 950 PetscCall(DMGetLocalSection(cdm, &cs)); 951 for (c = cStart; c < cEnd; c++) { 952 PetscInt doit = 0, outside = 0; 953 PetscCall(DMPlexVecGetClosure(cdm, cs, coords, c, &csize, &coef)); 954 Nv = csize / dim; 955 for (nz = d = 0; d < Nv; d++) { 956 PetscReal z = PetscRealPart(coef[d * dim + (dim - 1)]), x = PetscSqr(PetscRealPart(coef[d * dim + 0])) + ((dim == 3) ? PetscSqr(PetscRealPart(coef[d * dim + 1])) : 0); 957 x = PetscSqrtReal(x); 958 if (type == 0) { 959 if (ctx->re_radius > PETSC_SQRT_MACHINE_EPSILON && (z < -PETSC_MACHINE_EPSILON * 10. || z > ctx->re_radius + PETSC_MACHINE_EPSILON * 10.)) outside++; /* first pass don't refine bottom */ 960 } else if (type == 1 && (z > ctx->vperp0_radius1 || z < -ctx->vperp0_radius1)) { 961 outside++; /* don't refine outside electron refine radius */ 962 PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") (debug) found %s cells\n", grid, s_refine_names[type])); 963 } else if (type == 3 && (z > ctx->vperp0_radius2 || z < -ctx->vperp0_radius2)) { 964 outside++; /* refine r=0 cells on refinement front */ 965 PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") (debug) found %s cells\n", grid, s_refine_names[type])); 966 } 967 if (x < PETSC_MACHINE_EPSILON * 10. && (type != 0 || ctx->re_radius > PETSC_SQRT_MACHINE_EPSILON)) nz++; 968 } 969 PetscCall(DMPlexVecRestoreClosure(cdm, cs, coords, c, &csize, &coef)); 970 if (doit || (outside < Nv && nz)) { 971 PetscCall(DMLabelSetValue(adaptLabel, c, DM_ADAPT_REFINE)); 972 nrefined++; 973 } 974 } 975 PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") Refined %" PetscInt_FMT " cells\n", grid, nrefined)); 976 } 977 PetscCall(DMDestroy(&plex)); 978 PetscCall(DMAdaptLabel(forest, adaptLabel, &adaptedDM)); 979 PetscCall(DMLabelDestroy(&adaptLabel)); 980 *newForest = adaptedDM; 981 if (adaptedDM) { 982 if (isForest) { 983 PetscCall(DMForestSetAdaptivityForest(adaptedDM, NULL)); // ???? 984 } 985 PetscCall(DMConvert(adaptedDM, DMPLEX, &plex)); 986 PetscCall(DMPlexGetHeightStratum(plex, 0, &cStart, &cEnd)); 987 PetscCall(PetscInfo(sol, "\t\t\t\t%" PetscInt_FMT ") %" PetscInt_FMT " cells, %" PetscInt_FMT " total quadrature points\n", grid, cEnd - cStart, Nq * (cEnd - cStart))); 988 PetscCall(DMDestroy(&plex)); 989 } else *newForest = NULL; 990 PetscFunctionReturn(PETSC_SUCCESS); 991 } 992 993 // forest goes in (ctx->plex[grid]), plex comes out 994 static PetscErrorCode adapt(PetscInt grid, LandauCtx *ctx, Vec *uu) 995 { 996 PetscInt adaptIter; 997 998 PetscFunctionBegin; 999 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]}; 1000 for (type = 0; type < 5; type++) { 1001 for (adaptIter = 0; adaptIter < limits[type]; adaptIter++) { 1002 DM newForest = NULL; 1003 PetscCall(adaptToleranceFEM(ctx->fe[0], *uu, type, grid, ctx, &newForest)); 1004 if (newForest) { 1005 PetscCall(DMDestroy(&ctx->plex[grid])); 1006 PetscCall(VecDestroy(uu)); 1007 PetscCall(DMCreateGlobalVector(newForest, uu)); 1008 PetscCall(PetscObjectSetName((PetscObject)*uu, "uAMR")); 1009 PetscCall(LandauSetInitialCondition(newForest, *uu, grid, 0, 1, ctx)); 1010 ctx->plex[grid] = newForest; 1011 } else { 1012 PetscCall(PetscInfo(*uu, "No refinement\n")); 1013 } 1014 } 1015 } 1016 PetscFunctionReturn(PETSC_SUCCESS); 1017 } 1018 1019 // make log(Lambdas) from NRL Plasma formulary 1020 static PetscErrorCode makeLambdas(LandauCtx *ctx) 1021 { 1022 PetscFunctionBegin; 1023 for (PetscInt gridi = 0; gridi < ctx->num_grids; gridi++) { 1024 int iii = ctx->species_offset[gridi]; 1025 PetscReal Ti_ev = (ctx->thermal_temps[iii] / 1.1604525e7) * 1000; // convert (back) to eV 1026 PetscReal ni = ctx->n[iii] * ctx->n_0; 1027 for (PetscInt gridj = gridi; gridj < ctx->num_grids; gridj++) { 1028 PetscInt jjj = ctx->species_offset[gridj]; 1029 PetscReal Zj = ctx->charges[jjj] / 1.6022e-19; 1030 if (gridi == 0) { 1031 if (gridj == 0) { // lam_ee 1032 ctx->lambdas[gridi][gridj] = 23.5 - PetscLogReal(PetscSqrtReal(ni) * PetscPowReal(Ti_ev, -1.25)) - PetscSqrtReal(1e-5 + PetscSqr(PetscLogReal(Ti_ev) - 2) / 16); 1033 } else { // lam_ei == lam_ie 1034 if (10 * Zj * Zj > Ti_ev) { 1035 ctx->lambdas[gridi][gridj] = ctx->lambdas[gridj][gridi] = 23 - PetscLogReal(PetscSqrtReal(ni) * Zj * PetscPowReal(Ti_ev, -1.5)); 1036 } else { 1037 ctx->lambdas[gridi][gridj] = ctx->lambdas[gridj][gridi] = 24 - PetscLogReal(PetscSqrtReal(ni) / Ti_ev); 1038 } 1039 } 1040 } else { // lam_ii' 1041 PetscReal mui = ctx->masses[iii] / 1.6720e-27, Zi = ctx->charges[iii] / 1.6022e-19; 1042 PetscReal Tj_ev = (ctx->thermal_temps[jjj] / 1.1604525e7) * 1000; // convert (back) to eV 1043 PetscReal muj = ctx->masses[jjj] / 1.6720e-27; 1044 PetscReal nj = ctx->n[jjj] * ctx->n_0; 1045 ctx->lambdas[gridi][gridj] = ctx->lambdas[gridj][gridi] = 23 - PetscLogReal(Zi * Zj * (mui + muj) / (mui * Tj_ev + muj * Ti_ev) * PetscSqrtReal(ni * Zi * Zi / Ti_ev + nj * Zj * Zj / Tj_ev)); 1046 } 1047 } 1048 } 1049 //PetscReal v0 = PetscSqrtReal(ctx->k * ctx->thermal_temps[iii] / ctx->masses[iii]); /* arbitrary units for non-dimensionalization: plasma formulary def */ 1050 PetscFunctionReturn(PETSC_SUCCESS); 1051 } 1052 1053 static PetscErrorCode ProcessOptions(LandauCtx *ctx, const char prefix[]) 1054 { 1055 PetscBool flg, fileflg; 1056 PetscInt ii, nt, nm, nc, num_species_grid[LANDAU_MAX_GRIDS], non_dim_grid; 1057 PetscReal lnLam = 10; 1058 DM dummy; 1059 1060 PetscFunctionBegin; 1061 PetscCall(DMCreate(ctx->comm, &dummy)); 1062 /* get options - initialize context */ 1063 ctx->verbose = 1; // should be 0 for silent compliance 1064 ctx->batch_sz = 1; 1065 ctx->batch_view_idx = 0; 1066 ctx->interpolate = PETSC_TRUE; 1067 ctx->gpu_assembly = PETSC_TRUE; 1068 ctx->norm_state = 0; 1069 ctx->electronShift = 0; 1070 ctx->M = NULL; 1071 ctx->J = NULL; 1072 /* geometry and grids */ 1073 ctx->sphere = PETSC_FALSE; 1074 ctx->use_p4est = PETSC_FALSE; 1075 ctx->simplex = PETSC_FALSE; 1076 for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) { 1077 ctx->radius[grid] = 5.; /* thermal radius (velocity) */ 1078 ctx->radius_perp[grid] = 5.; /* thermal radius (velocity) */ 1079 ctx->radius_par[grid] = 5.; /* thermal radius (velocity) */ 1080 ctx->numAMRRefine[grid] = 0; 1081 ctx->postAMRRefine[grid] = 0; 1082 ctx->species_offset[grid + 1] = 1; // one species default 1083 num_species_grid[grid] = 0; 1084 ctx->plex[grid] = NULL; /* cache as expensive to Convert */ 1085 } 1086 ctx->species_offset[0] = 0; 1087 ctx->re_radius = 0.; 1088 ctx->vperp0_radius1 = 0; 1089 ctx->vperp0_radius2 = 0; 1090 ctx->nZRefine1 = 0; 1091 ctx->nZRefine2 = 0; 1092 ctx->numRERefine = 0; 1093 num_species_grid[0] = 1; // one species default 1094 /* species - [0] electrons, [1] one ion species eg, duetarium, [2] heavy impurity ion, ... */ 1095 ctx->charges[0] = -1; /* electron charge (MKS) */ 1096 ctx->masses[0] = 1 / 1835.469965278441013; /* temporary value in proton mass */ 1097 ctx->n[0] = 1; 1098 ctx->v_0 = 1; /* thermal velocity, we could start with a scale != 1 */ 1099 ctx->thermal_temps[0] = 1; 1100 /* constants, etc. */ 1101 ctx->epsilon0 = 8.8542e-12; /* permittivity of free space (MKS) F/m */ 1102 ctx->k = 1.38064852e-23; /* Boltzmann constant (MKS) J/K */ 1103 ctx->n_0 = 1.e20; /* typical plasma n, but could set it to 1 */ 1104 ctx->Ez = 0; 1105 for (PetscInt grid = 0; grid < LANDAU_NUM_TIMERS; grid++) ctx->times[grid] = 0; 1106 for (PetscInt ii = 0; ii < LANDAU_DIM; ii++) ctx->cells0[ii] = 2; 1107 if (LANDAU_DIM == 2) ctx->cells0[0] = 1; 1108 ctx->use_matrix_mass = PETSC_FALSE; 1109 ctx->use_relativistic_corrections = PETSC_FALSE; 1110 ctx->use_energy_tensor_trick = PETSC_FALSE; /* Use Eero's trick for energy conservation v --> grad(v^2/2) */ 1111 ctx->SData_d.w = NULL; 1112 ctx->SData_d.x = NULL; 1113 ctx->SData_d.y = NULL; 1114 ctx->SData_d.z = NULL; 1115 ctx->SData_d.invJ = NULL; 1116 ctx->jacobian_field_major_order = PETSC_FALSE; 1117 ctx->SData_d.coo_elem_offsets = NULL; 1118 ctx->SData_d.coo_elem_point_offsets = NULL; 1119 ctx->SData_d.coo_elem_fullNb = NULL; 1120 ctx->SData_d.coo_size = 0; 1121 PetscOptionsBegin(ctx->comm, prefix, "Options for Fokker-Plank-Landau collision operator", "none"); 1122 #if defined(PETSC_HAVE_KOKKOS) 1123 ctx->deviceType = LANDAU_KOKKOS; 1124 PetscCall(PetscStrncpy(ctx->filename, "kokkos", sizeof(ctx->filename))); 1125 #else 1126 ctx->deviceType = LANDAU_CPU; 1127 PetscCall(PetscStrncpy(ctx->filename, "cpu", sizeof(ctx->filename))); 1128 #endif 1129 PetscCall(PetscOptionsString("-dm_landau_device_type", "Use kernels on 'cpu' 'kokkos'", "plexland.c", ctx->filename, ctx->filename, sizeof(ctx->filename), NULL)); 1130 PetscCall(PetscStrcmp("cpu", ctx->filename, &flg)); 1131 if (flg) { 1132 ctx->deviceType = LANDAU_CPU; 1133 } else { 1134 PetscCall(PetscStrcmp("kokkos", ctx->filename, &flg)); 1135 if (flg) ctx->deviceType = LANDAU_KOKKOS; 1136 else SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_device_type %s", ctx->filename); 1137 } 1138 ctx->filename[0] = '\0'; 1139 PetscCall(PetscOptionsString("-dm_landau_filename", "file to read mesh from", "plexland.c", ctx->filename, ctx->filename, sizeof(ctx->filename), &fileflg)); 1140 PetscCall(PetscOptionsReal("-dm_landau_electron_shift", "Shift in thermal velocity of electrons", "none", ctx->electronShift, &ctx->electronShift, NULL)); 1141 PetscCall(PetscOptionsInt("-dm_landau_verbose", "Level of verbosity output", "plexland.c", ctx->verbose, &ctx->verbose, NULL)); 1142 PetscCall(PetscOptionsInt("-dm_landau_batch_size", "Number of 'vertices' to batch", "ex2.c", ctx->batch_sz, &ctx->batch_sz, NULL)); 1143 PetscCheck(LANDAU_MAX_BATCH_SZ >= ctx->batch_sz, ctx->comm, PETSC_ERR_ARG_WRONG, "LANDAU_MAX_BATCH_SZ %" PetscInt_FMT " < ctx->batch_sz %" PetscInt_FMT, (PetscInt)LANDAU_MAX_BATCH_SZ, ctx->batch_sz); 1144 PetscCall(PetscOptionsInt("-dm_landau_batch_view_idx", "Index of batch for diagnostics like plotting", "ex2.c", ctx->batch_view_idx, &ctx->batch_view_idx, NULL)); 1145 PetscCheck(ctx->batch_view_idx < ctx->batch_sz, ctx->comm, PETSC_ERR_ARG_WRONG, "-ctx->batch_view_idx %" PetscInt_FMT " > ctx->batch_sz %" PetscInt_FMT, ctx->batch_view_idx, ctx->batch_sz); 1146 PetscCall(PetscOptionsReal("-dm_landau_Ez", "Initial parallel electric field in unites of Conner-Hastie critical field", "plexland.c", ctx->Ez, &ctx->Ez, NULL)); 1147 PetscCall(PetscOptionsReal("-dm_landau_n_0", "Normalization constant for number density", "plexland.c", ctx->n_0, &ctx->n_0, NULL)); 1148 PetscCall(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)); 1149 PetscCall(PetscOptionsBool("-dm_landau_use_relativistic_corrections", "Use relativistic corrections", "plexland.c", ctx->use_relativistic_corrections, &ctx->use_relativistic_corrections, NULL)); 1150 PetscCall(PetscOptionsBool("-dm_landau_simplex", "Use simplex elements", "plexland.c", ctx->simplex, &ctx->simplex, NULL)); 1151 if (LANDAU_DIM == 2 && ctx->use_relativistic_corrections) ctx->use_relativistic_corrections = PETSC_FALSE; // should warn 1152 PetscCall(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, 1153 &ctx->use_energy_tensor_trick, NULL)); 1154 1155 /* get num species with temperature, set defaults */ 1156 for (ii = 1; ii < LANDAU_MAX_SPECIES; ii++) { 1157 ctx->thermal_temps[ii] = 1; 1158 ctx->charges[ii] = 1; 1159 ctx->masses[ii] = 1; 1160 ctx->n[ii] = 1; 1161 } 1162 nt = LANDAU_MAX_SPECIES; 1163 PetscCall(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)); 1164 if (flg) { 1165 PetscCall(PetscInfo(dummy, "num_species set to number of thermal temps provided (%" PetscInt_FMT ")\n", nt)); 1166 ctx->num_species = nt; 1167 } else SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_thermal_temps ,t1,t2,.. must be provided to set the number of species"); 1168 for (ii = 0; ii < ctx->num_species; ii++) ctx->thermal_temps[ii] *= 1.1604525e7; /* convert to Kelvin */ 1169 nm = LANDAU_MAX_SPECIES - 1; 1170 PetscCall(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)); 1171 PetscCheck(!flg || nm == ctx->num_species - 1, ctx->comm, PETSC_ERR_ARG_WRONG, "num ion masses %" PetscInt_FMT " != num species %" PetscInt_FMT, nm, ctx->num_species - 1); 1172 nm = LANDAU_MAX_SPECIES; 1173 PetscCall(PetscOptionsRealArray("-dm_landau_n", "Number density of each species = n_s * n_0", "plexland.c", ctx->n, &nm, &flg)); 1174 PetscCheck(!flg || nm == ctx->num_species, ctx->comm, PETSC_ERR_ARG_WRONG, "wrong num n: %" PetscInt_FMT " != num species %" PetscInt_FMT, nm, ctx->num_species); 1175 for (ii = 0; ii < LANDAU_MAX_SPECIES; ii++) ctx->masses[ii] *= 1.6720e-27; /* scale by proton mass kg */ 1176 ctx->masses[0] = 9.10938356e-31; /* electron mass kg (should be about right already) */ 1177 nc = LANDAU_MAX_SPECIES - 1; 1178 PetscCall(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)); 1179 if (flg) PetscCheck(nc == ctx->num_species - 1, ctx->comm, PETSC_ERR_ARG_WRONG, "num charges %" PetscInt_FMT " != num species %" PetscInt_FMT, nc, ctx->num_species - 1); 1180 for (ii = 0; ii < LANDAU_MAX_SPECIES; ii++) ctx->charges[ii] *= 1.6022e-19; /* electron/proton charge (MKS) */ 1181 /* geometry and grids */ 1182 nt = LANDAU_MAX_GRIDS; 1183 PetscCall(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)); 1184 if (flg) { 1185 ctx->num_grids = nt; 1186 for (ii = nt = 0; ii < ctx->num_grids; ii++) nt += num_species_grid[ii]; 1187 PetscCheck(ctx->num_species == nt, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_num_species_grid: sum %" PetscInt_FMT " != num_species = %" PetscInt_FMT ". %" PetscInt_FMT " grids (check that number of grids <= LANDAU_MAX_GRIDS = %d)", nt, ctx->num_species, 1188 ctx->num_grids, LANDAU_MAX_GRIDS); 1189 } else { 1190 if (ctx->num_species > LANDAU_MAX_GRIDS) { 1191 num_species_grid[0] = 1; 1192 num_species_grid[1] = ctx->num_species - 1; 1193 ctx->num_grids = 2; 1194 } else { 1195 ctx->num_grids = ctx->num_species; 1196 for (ii = 0; ii < ctx->num_grids; ii++) num_species_grid[ii] = 1; 1197 } 1198 } 1199 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]; 1200 PetscCheck(ctx->species_offset[ctx->num_grids] == ctx->num_species, ctx->comm, PETSC_ERR_ARG_WRONG, "ctx->species_offset[ctx->num_grids] %" PetscInt_FMT " != ctx->num_species = %" PetscInt_FMT " ???????????", ctx->species_offset[ctx->num_grids], 1201 ctx->num_species); 1202 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 1203 int iii = ctx->species_offset[grid]; // normalize with first (arbitrary) species on grid 1204 ctx->thermal_speed[grid] = PetscSqrtReal(ctx->k * ctx->thermal_temps[iii] / ctx->masses[iii]); /* arbitrary units for non-dimensionalization: plasma formulary def */ 1205 } 1206 // get lambdas here because we need them for t_0 etc 1207 PetscCall(PetscOptionsReal("-dm_landau_ln_lambda", "Universal cross section parameter. Default uses NRL formulas", "plexland.c", lnLam, &lnLam, &flg)); 1208 if (flg) { 1209 for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) { 1210 for (PetscInt gridj = 0; gridj < LANDAU_MAX_GRIDS; gridj++) ctx->lambdas[gridj][grid] = lnLam; /* cross section ratio large - small angle collisions */ 1211 } 1212 } else { 1213 PetscCall(makeLambdas(ctx)); 1214 } 1215 non_dim_grid = 0; 1216 PetscCall(PetscOptionsInt("-dm_landau_normalization_grid", "Index of grid to use for setting v_0, m_0, t_0. (Not recommended)", "plexland.c", non_dim_grid, &non_dim_grid, &flg)); 1217 if (non_dim_grid != 0) PetscCall(PetscInfo(dummy, "Normalization grid set to %" PetscInt_FMT ", but non-default not well verified\n", non_dim_grid)); 1218 PetscCheck(non_dim_grid >= 0 && non_dim_grid < ctx->num_species, ctx->comm, PETSC_ERR_ARG_WRONG, "Normalization grid wrong: %" PetscInt_FMT, non_dim_grid); 1219 ctx->v_0 = ctx->thermal_speed[non_dim_grid]; /* arbitrary units for non dimensionalization: global mean velocity in 1D of electrons */ 1220 ctx->m_0 = ctx->masses[non_dim_grid]; /* arbitrary reference mass, electrons */ 1221 ctx->t_0 = 8 * PETSC_PI * PetscSqr(ctx->epsilon0 * ctx->m_0 / PetscSqr(ctx->charges[non_dim_grid])) / ctx->lambdas[non_dim_grid][non_dim_grid] / ctx->n_0 * PetscPowReal(ctx->v_0, 3); /* note, this t_0 makes nu[non_dim_grid,non_dim_grid]=1 */ 1222 /* domain */ 1223 nt = LANDAU_MAX_GRIDS; 1224 PetscCall(PetscOptionsRealArray("-dm_landau_domain_radius", "Phase space size in units of thermal velocity of grid", "plexland.c", ctx->radius, &nt, &flg)); 1225 if (flg) { 1226 PetscCheck(nt >= ctx->num_grids, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_domain_radius: given %" PetscInt_FMT " radius != number grids %" PetscInt_FMT, nt, ctx->num_grids); 1227 while (nt--) ctx->radius_par[nt] = ctx->radius_perp[nt] = ctx->radius[nt]; 1228 } else { 1229 nt = LANDAU_MAX_GRIDS; 1230 PetscCall(PetscOptionsRealArray("-dm_landau_domain_max_par", "Parallel velocity domain size in units of thermal velocity of grid", "plexland.c", ctx->radius_par, &nt, &flg)); 1231 if (flg) PetscCheck(nt >= ctx->num_grids, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_domain_max_par: given %" PetscInt_FMT " radius != number grids %" PetscInt_FMT, nt, ctx->num_grids); 1232 PetscCall(PetscOptionsRealArray("-dm_landau_domain_max_perp", "Perpendicular velocity domain size in units of thermal velocity of grid", "plexland.c", ctx->radius_perp, &nt, &flg)); 1233 if (flg) PetscCheck(nt >= ctx->num_grids, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_domain_max_perp: given %" PetscInt_FMT " radius != number grids %" PetscInt_FMT, nt, ctx->num_grids); 1234 } 1235 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 1236 if (flg && ctx->radius[grid] <= 0) { /* negative is ratio of c - need to set par and perp with this -- todo */ 1237 if (ctx->radius[grid] == 0) ctx->radius[grid] = 0.75; 1238 else ctx->radius[grid] = -ctx->radius[grid]; 1239 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) 1240 PetscCall(PetscInfo(dummy, "Change domain radius to %g for grid %" PetscInt_FMT "\n", (double)ctx->radius[grid], grid)); 1241 } 1242 ctx->radius[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0 1243 ctx->radius_perp[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0 1244 ctx->radius_par[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0 1245 } 1246 /* amr parameters */ 1247 if (!fileflg) { 1248 nt = LANDAU_MAX_GRIDS; 1249 PetscCall(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)); 1250 PetscCheck(!flg || nt >= ctx->num_grids, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_amr_levels_max: given %" PetscInt_FMT " != number grids %" PetscInt_FMT, nt, ctx->num_grids); 1251 nt = LANDAU_MAX_GRIDS; 1252 PetscCall(PetscOptionsIntArray("-dm_landau_amr_post_refine", "Number of levels to uniformly refine after AMR", "plexland.c", ctx->postAMRRefine, &nt, &flg)); 1253 for (ii = 1; ii < ctx->num_grids; ii++) ctx->postAMRRefine[ii] = ctx->postAMRRefine[0]; // all grids the same now 1254 PetscCall(PetscOptionsInt("-dm_landau_amr_re_levels", "Number of levels to refine along v_perp=0, z>0", "plexland.c", ctx->numRERefine, &ctx->numRERefine, &flg)); 1255 PetscCall(PetscOptionsInt("-dm_landau_amr_z_refine_pre", "Number of levels to refine along v_perp=0 before origin refine", "plexland.c", ctx->nZRefine1, &ctx->nZRefine1, &flg)); 1256 PetscCall(PetscOptionsInt("-dm_landau_amr_z_refine_post", "Number of levels to refine along v_perp=0 after origin refine", "plexland.c", ctx->nZRefine2, &ctx->nZRefine2, &flg)); 1257 PetscCall(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)); 1258 PetscCall(PetscOptionsReal("-dm_landau_z_radius_pre", "velocity range to refine r=0 axis (for electrons)", "plexland.c", ctx->vperp0_radius1, &ctx->vperp0_radius1, &flg)); 1259 PetscCall(PetscOptionsReal("-dm_landau_z_radius_post", "velocity range to refine r=0 axis (for electrons) after origin AMR", "plexland.c", ctx->vperp0_radius2, &ctx->vperp0_radius2, &flg)); 1260 /* spherical domain (not used) */ 1261 PetscCall(PetscOptionsBool("-dm_landau_sphere", "use sphere/semi-circle domain instead of rectangle", "plexland.c", ctx->sphere, &ctx->sphere, NULL)); 1262 if (ctx->sphere || ctx->simplex) { 1263 ctx->sphere_inner_radius_90degree = 0.40; 1264 ctx->sphere_inner_radius_45degree = 0.35; 1265 PetscCall(PetscOptionsReal("-dm_landau_sphere_inner_radius_90degree_scale", "Scaling of radius for inner circle on 90 degree grid", "plexland.c", ctx->sphere_inner_radius_90degree, &ctx->sphere_inner_radius_90degree, NULL)); 1266 PetscCall(PetscOptionsReal("-dm_landau_sphere_inner_radius_45degree_scale", "Scaling of radius for inner circle on 45 degree grid", "plexland.c", ctx->sphere_inner_radius_45degree, &ctx->sphere_inner_radius_45degree, NULL)); 1267 } else { 1268 nt = LANDAU_DIM; 1269 PetscCall(PetscOptionsIntArray("-dm_landau_num_cells", "Number of cells in each dimension of base grid", "plexland.c", ctx->cells0, &nt, &flg)); 1270 } 1271 } 1272 /* processing options */ 1273 PetscCall(PetscOptionsBool("-dm_landau_gpu_assembly", "Assemble Jacobian on GPU", "plexland.c", ctx->gpu_assembly, &ctx->gpu_assembly, NULL)); 1274 PetscCall(PetscOptionsBool("-dm_landau_jacobian_field_major_order", "Reorder Jacobian for GPU assembly with field major, or block diagonal, ordering (DEPRECATED)", "plexland.c", ctx->jacobian_field_major_order, &ctx->jacobian_field_major_order, NULL)); 1275 if (ctx->jacobian_field_major_order) PetscCheck(ctx->gpu_assembly, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_jacobian_field_major_order requires -dm_landau_gpu_assembly"); 1276 PetscCheck(!ctx->jacobian_field_major_order, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_jacobian_field_major_order DEPRECATED"); 1277 PetscOptionsEnd(); 1278 1279 for (ii = ctx->num_species; ii < LANDAU_MAX_SPECIES; ii++) ctx->masses[ii] = ctx->thermal_temps[ii] = ctx->charges[ii] = 0; 1280 if (ctx->verbose != 0) { 1281 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "masses: e=%10.3e; ions in proton mass units: %10.3e %10.3e ...\n", (double)ctx->masses[0], (double)(ctx->masses[1] / 1.6720e-27), (double)(ctx->num_species > 2 ? ctx->masses[2] / 1.6720e-27 : 0))); 1282 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "charges: e=%10.3e; charges in elementary units: %10.3e %10.3e\n", (double)ctx->charges[0], (double)(-ctx->charges[1] / ctx->charges[0]), (double)(ctx->num_species > 2 ? -ctx->charges[2] / ctx->charges[0] : 0))); 1283 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "n: e: %10.3e i: %10.3e %10.3e\n", (double)ctx->n[0], (double)ctx->n[1], (double)(ctx->num_species > 2 ? ctx->n[2] : 0))); 1284 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "thermal T (K): e=%10.3e i=%10.3e %10.3e. Normalization grid %d: v_0=%10.3e (%10.3ec) n_0=%10.3e t_0=%10.3e %" PetscInt_FMT " batched, view batch %" PetscInt_FMT "\n", (double)ctx->thermal_temps[0], 1285 (double)ctx->thermal_temps[1], (double)((ctx->num_species > 2) ? ctx->thermal_temps[2] : 0), (int)non_dim_grid, (double)ctx->v_0, (double)(ctx->v_0 / SPEED_OF_LIGHT), (double)ctx->n_0, (double)ctx->t_0, ctx->batch_sz, ctx->batch_view_idx)); 1286 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "Domain radius (AMR levels) grid %d: par=%10.3e perp=%10.3e (%" PetscInt_FMT ") ", 0, (double)ctx->radius_par[0], (double)ctx->radius_perp[0], ctx->numAMRRefine[0])); 1287 for (ii = 1; ii < ctx->num_grids; ii++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, ", %" PetscInt_FMT ": par=%10.3e perp=%10.3e (%" PetscInt_FMT ") ", ii, (double)ctx->radius_par[ii], (double)ctx->radius_perp[ii], ctx->numAMRRefine[ii])); 1288 if (ctx->use_relativistic_corrections) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nUse relativistic corrections\n")); 1289 else PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n")); 1290 } 1291 PetscCall(DMDestroy(&dummy)); 1292 { 1293 PetscMPIInt rank; 1294 PetscCallMPI(MPI_Comm_rank(PETSC_COMM_WORLD, &rank)); 1295 ctx->stage = 0; 1296 PetscCall(PetscLogEventRegister("Landau Create", DM_CLASSID, &ctx->events[13])); /* 13 */ 1297 PetscCall(PetscLogEventRegister(" GPU ass. setup", DM_CLASSID, &ctx->events[2])); /* 2 */ 1298 PetscCall(PetscLogEventRegister(" Build matrix", DM_CLASSID, &ctx->events[12])); /* 12 */ 1299 PetscCall(PetscLogEventRegister(" Assembly maps", DM_CLASSID, &ctx->events[15])); /* 15 */ 1300 PetscCall(PetscLogEventRegister("Landau Mass mat", DM_CLASSID, &ctx->events[14])); /* 14 */ 1301 PetscCall(PetscLogEventRegister("Landau Operator", DM_CLASSID, &ctx->events[11])); /* 11 */ 1302 PetscCall(PetscLogEventRegister("Landau Jacobian", DM_CLASSID, &ctx->events[0])); /* 0 */ 1303 PetscCall(PetscLogEventRegister("Landau Mass", DM_CLASSID, &ctx->events[9])); /* 9 */ 1304 PetscCall(PetscLogEventRegister(" Preamble", DM_CLASSID, &ctx->events[10])); /* 10 */ 1305 PetscCall(PetscLogEventRegister(" static IP Data", DM_CLASSID, &ctx->events[7])); /* 7 */ 1306 PetscCall(PetscLogEventRegister(" dynamic IP-Jac", DM_CLASSID, &ctx->events[1])); /* 1 */ 1307 PetscCall(PetscLogEventRegister(" Kernel-init", DM_CLASSID, &ctx->events[3])); /* 3 */ 1308 PetscCall(PetscLogEventRegister(" Jac-f-df (GPU)", DM_CLASSID, &ctx->events[8])); /* 8 */ 1309 PetscCall(PetscLogEventRegister(" J Kernel (GPU)", DM_CLASSID, &ctx->events[4])); /* 4 */ 1310 PetscCall(PetscLogEventRegister(" M Kernel (GPU)", DM_CLASSID, &ctx->events[16])); /* 16 */ 1311 PetscCall(PetscLogEventRegister(" Copy to CPU", DM_CLASSID, &ctx->events[5])); /* 5 */ 1312 PetscCall(PetscLogEventRegister(" CPU assemble", DM_CLASSID, &ctx->events[6])); /* 6 */ 1313 1314 if (rank) { /* turn off output stuff for duplicate runs - do we need to add the prefix to all this? */ 1315 PetscCall(PetscOptionsClearValue(NULL, "-snes_converged_reason")); 1316 PetscCall(PetscOptionsClearValue(NULL, "-ksp_converged_reason")); 1317 PetscCall(PetscOptionsClearValue(NULL, "-snes_monitor")); 1318 PetscCall(PetscOptionsClearValue(NULL, "-ksp_monitor")); 1319 PetscCall(PetscOptionsClearValue(NULL, "-ts_monitor")); 1320 PetscCall(PetscOptionsClearValue(NULL, "-ts_view")); 1321 PetscCall(PetscOptionsClearValue(NULL, "-ts_adapt_monitor")); 1322 PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_amr_dm_view")); 1323 PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_amr_vec_view")); 1324 PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mass_dm_view")); 1325 PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mass_view")); 1326 PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_jacobian_view")); 1327 PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mat_view")); 1328 PetscCall(PetscOptionsClearValue(NULL, "-pc_bjkokkos_ksp_converged_reason")); 1329 PetscCall(PetscOptionsClearValue(NULL, "-pc_bjkokkos_ksp_monitor")); 1330 PetscCall(PetscOptionsClearValue(NULL, "-")); 1331 PetscCall(PetscOptionsClearValue(NULL, "-info")); 1332 } 1333 } 1334 PetscFunctionReturn(PETSC_SUCCESS); 1335 } 1336 1337 static PetscErrorCode CreateStaticData(PetscInt dim, IS grid_batch_is_inv[], LandauCtx *ctx) 1338 { 1339 PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS]; 1340 PetscQuadrature quad; 1341 const PetscReal *quadWeights; 1342 PetscReal invMass[LANDAU_MAX_SPECIES], nu_alpha[LANDAU_MAX_SPECIES], nu_beta[LANDAU_MAX_SPECIES]; 1343 PetscInt numCells[LANDAU_MAX_GRIDS], Nq, Nb, Nf[LANDAU_MAX_GRIDS], ncellsTot = 0, MAP_BF_SIZE = 64 * LANDAU_DIM * LANDAU_DIM * LANDAU_MAX_Q_FACE * LANDAU_MAX_SPECIES; 1344 PetscTabulation *Tf; 1345 PetscDS prob; 1346 1347 PetscFunctionBegin; 1348 PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb)); 1349 PetscCheck(Nb <= LANDAU_MAX_NQND, ctx->comm, PETSC_ERR_ARG_WRONG, "Order too high. Nb = %" PetscInt_FMT " > LANDAU_MAX_NQND (%d)", Nb, LANDAU_MAX_NQND); 1350 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 1351 for (PetscInt ii = ctx->species_offset[grid]; ii < ctx->species_offset[grid + 1]; ii++) { 1352 invMass[ii] = ctx->m_0 / ctx->masses[ii]; 1353 nu_alpha[ii] = PetscSqr(ctx->charges[ii] / ctx->m_0) * ctx->m_0 / ctx->masses[ii]; 1354 nu_beta[ii] = PetscSqr(ctx->charges[ii] / ctx->epsilon0) / (8 * PETSC_PI) * ctx->t_0 * ctx->n_0 / PetscPowReal(ctx->v_0, 3); 1355 } 1356 } 1357 if (ctx->verbose == 4) { 1358 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "nu_alpha: ")); 1359 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 1360 int iii = ctx->species_offset[grid]; 1361 for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %e", (double)nu_alpha[ii])); 1362 } 1363 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nnu_beta: ")); 1364 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 1365 int iii = ctx->species_offset[grid]; 1366 for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %e", (double)nu_beta[ii])); 1367 } 1368 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nnu_alpha[i]*nu_beta[j]*lambda[i][j]:\n")); 1369 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 1370 int iii = ctx->species_offset[grid]; 1371 for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) { 1372 for (PetscInt gridj = 0; gridj < ctx->num_grids; gridj++) { 1373 int jjj = ctx->species_offset[gridj]; 1374 for (PetscInt jj = jjj; jj < ctx->species_offset[gridj + 1]; jj++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %14.9e", (double)(nu_alpha[ii] * nu_beta[jj] * ctx->lambdas[grid][gridj]))); 1375 } 1376 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n")); 1377 } 1378 } 1379 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "lambda[i][j]:\n")); 1380 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 1381 int iii = ctx->species_offset[grid]; 1382 for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) { 1383 for (PetscInt gridj = 0; gridj < ctx->num_grids; gridj++) { 1384 int jjj = ctx->species_offset[gridj]; 1385 for (PetscInt jj = jjj; jj < ctx->species_offset[gridj + 1]; jj++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %14.9e", (double)ctx->lambdas[grid][gridj])); 1386 } 1387 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n")); 1388 } 1389 } 1390 } 1391 PetscCall(DMGetDS(ctx->plex[0], &prob)); // same DS for all grids 1392 PetscCall(PetscDSGetTabulation(prob, &Tf)); // Bf, &Df same for all grids 1393 /* DS, Tab and quad is same on all grids */ 1394 PetscCheck(ctx->plex[0], ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created"); 1395 PetscCall(PetscFEGetQuadrature(ctx->fe[0], &quad)); 1396 PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, &quadWeights)); 1397 PetscCheck(Nq <= LANDAU_MAX_NQND, ctx->comm, PETSC_ERR_ARG_WRONG, "Order too high. Nq = %" PetscInt_FMT " > LANDAU_MAX_NQND (%d)", Nq, LANDAU_MAX_NQND); 1398 /* setup each grid */ 1399 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 1400 PetscInt cStart, cEnd; 1401 PetscCheck(ctx->plex[grid] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created"); 1402 PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd)); 1403 numCells[grid] = cEnd - cStart; // grids can have different topology 1404 PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid])); 1405 PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid])); 1406 PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid])); 1407 ncellsTot += numCells[grid]; 1408 } 1409 /* create GPU assembly data */ 1410 if (ctx->gpu_assembly) { /* we need GPU object with GPU assembly */ 1411 PetscContainer container; 1412 PetscScalar *elemMatrix, *elMat; 1413 pointInterpolationP4est(*pointMaps)[LANDAU_MAX_Q_FACE]; 1414 P4estVertexMaps *maps; 1415 const PetscInt *plex_batch = NULL, elMatSz = Nb * Nb * ctx->num_species * ctx->num_species; 1416 LandauIdx *coo_elem_offsets = NULL, *coo_elem_fullNb = NULL, (*coo_elem_point_offsets)[LANDAU_MAX_NQND + 1] = NULL; 1417 /* create GPU assembly data */ 1418 PetscCall(PetscInfo(ctx->plex[0], "Make GPU maps %d\n", 1)); 1419 PetscCall(PetscLogEventBegin(ctx->events[2], 0, 0, 0, 0)); 1420 PetscCall(PetscMalloc(sizeof(*maps) * ctx->num_grids, &maps)); 1421 PetscCall(PetscMalloc(sizeof(*pointMaps) * MAP_BF_SIZE, &pointMaps)); 1422 PetscCall(PetscMalloc(sizeof(*elemMatrix) * elMatSz, &elemMatrix)); 1423 1424 { // setup COO assembly -- put COO metadata directly in ctx->SData_d 1425 PetscCall(PetscMalloc3(ncellsTot + 1, &coo_elem_offsets, ncellsTot, &coo_elem_fullNb, ncellsTot, &coo_elem_point_offsets)); // array of integer pointers 1426 coo_elem_offsets[0] = 0; // finish later 1427 PetscCall(PetscInfo(ctx->plex[0], "COO initialization, %" PetscInt_FMT " cells\n", ncellsTot)); 1428 ctx->SData_d.coo_n_cellsTot = ncellsTot; 1429 ctx->SData_d.coo_elem_offsets = (void *)coo_elem_offsets; 1430 ctx->SData_d.coo_elem_fullNb = (void *)coo_elem_fullNb; 1431 ctx->SData_d.coo_elem_point_offsets = (void *)coo_elem_point_offsets; 1432 } 1433 1434 ctx->SData_d.coo_max_fullnb = 0; 1435 for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) { 1436 PetscInt cStart, cEnd, Nfloc = Nf[grid], totDim = Nfloc * Nb; 1437 if (grid_batch_is_inv[grid]) PetscCall(ISGetIndices(grid_batch_is_inv[grid], &plex_batch)); 1438 PetscCheck(!plex_batch, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_jacobian_field_major_order DEPRECATED"); 1439 PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd)); 1440 // make maps 1441 maps[grid].d_self = NULL; 1442 maps[grid].num_elements = numCells[grid]; 1443 maps[grid].num_face = (PetscInt)(pow(Nq, 1. / ((double)dim)) + .001); // Q 1444 maps[grid].num_face = (PetscInt)(pow(maps[grid].num_face, (double)(dim - 1)) + .001); // Q^2 1445 maps[grid].num_reduced = 0; 1446 maps[grid].deviceType = ctx->deviceType; 1447 maps[grid].numgrids = ctx->num_grids; 1448 // count reduced and get 1449 PetscCall(PetscMalloc(maps[grid].num_elements * sizeof(*maps[grid].gIdx), &maps[grid].gIdx)); 1450 for (int ej = cStart, eidx = 0; ej < cEnd; ++ej, ++eidx, glb_elem_idx++) { 1451 if (coo_elem_offsets) coo_elem_offsets[glb_elem_idx + 1] = coo_elem_offsets[glb_elem_idx]; // start with last one, then add 1452 for (int fieldA = 0; fieldA < Nf[grid]; fieldA++) { 1453 int fullNb = 0; 1454 for (int q = 0; q < Nb; ++q) { 1455 PetscInt numindices, *indices; 1456 PetscScalar *valuesOrig = elMat = elemMatrix; 1457 PetscCall(PetscArrayzero(elMat, totDim * totDim)); 1458 elMat[(fieldA * Nb + q) * totDim + fieldA * Nb + q] = 1; 1459 PetscCall(DMPlexGetClosureIndices(ctx->plex[grid], section[grid], globsection[grid], ej, PETSC_TRUE, &numindices, &indices, NULL, (PetscScalar **)&elMat)); 1460 if (ctx->simplex) { 1461 PetscCheck(numindices == Nb, ctx->comm, PETSC_ERR_ARG_WRONG, "numindices != Nb numindices=%d Nb=%d", (int)numindices, (int)Nb); 1462 for (int q = 0; q < numindices; ++q) { maps[grid].gIdx[eidx][fieldA][q] = (LandauIdx)indices[q]; } 1463 fullNb++; 1464 } else { 1465 for (PetscInt f = 0; f < numindices; ++f) { // look for a non-zero on the diagonal (is this too complicated for simplices?) 1466 if (PetscAbs(PetscRealPart(elMat[f * numindices + f])) > PETSC_MACHINE_EPSILON) { 1467 // found it 1468 if (PetscAbs(PetscRealPart(elMat[f * numindices + f] - 1.)) < PETSC_MACHINE_EPSILON) { // normal vertex 1.0 1469 if (plex_batch) { 1470 maps[grid].gIdx[eidx][fieldA][q] = (LandauIdx)plex_batch[indices[f]]; 1471 } else { 1472 maps[grid].gIdx[eidx][fieldA][q] = (LandauIdx)indices[f]; 1473 } 1474 fullNb++; 1475 } else { //found a constraint 1476 int jj = 0; 1477 PetscReal sum = 0; 1478 const PetscInt ff = f; 1479 maps[grid].gIdx[eidx][fieldA][q] = -maps[grid].num_reduced - 1; // store (-)index: id = -(idx+1): idx = -id - 1 1480 PetscCheck(!ctx->simplex, ctx->comm, PETSC_ERR_ARG_WRONG, "No constraints with simplex"); 1481 do { // constraints are continuous in Plex - exploit that here 1482 int ii; // get 'scale' 1483 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 1484 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 1485 pointMaps[maps[grid].num_reduced][jj].scale += PetscRealPart(elMat[f * numindices + ff + ii]); 1486 } 1487 } 1488 sum += pointMaps[maps[grid].num_reduced][jj].scale; // diagnostic 1489 // get 'gid' 1490 if (pointMaps[maps[grid].num_reduced][jj].scale == 0) pointMaps[maps[grid].num_reduced][jj].gid = -1; // 3D has Q and Q^2 interps 1491 else { 1492 if (plex_batch) { 1493 pointMaps[maps[grid].num_reduced][jj].gid = plex_batch[indices[f]]; 1494 } else { 1495 pointMaps[maps[grid].num_reduced][jj].gid = indices[f]; 1496 } 1497 fullNb++; 1498 } 1499 } while (++jj < maps[grid].num_face && ++f < numindices); // jj is incremented if we hit the end 1500 while (jj < maps[grid].num_face) { 1501 pointMaps[maps[grid].num_reduced][jj].scale = 0; 1502 pointMaps[maps[grid].num_reduced][jj].gid = -1; 1503 jj++; 1504 } 1505 if (PetscAbs(sum - 1.0) > 10 * PETSC_MACHINE_EPSILON) { // debug 1506 int d, f; 1507 PetscReal tmp = 0; 1508 PetscCall(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, (double)sum, LANDAU_MAX_Q_FACE, maps[grid].num_face)); 1509 for (d = 0, tmp = 0; d < numindices; ++d) { 1510 if (tmp != 0 && PetscAbs(tmp - 1.0) > 10 * PETSC_MACHINE_EPSILON) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%3d) %3" PetscInt_FMT ": ", d, indices[d])); 1511 for (f = 0; f < numindices; ++f) tmp += PetscRealPart(elMat[d * numindices + f]); 1512 if (tmp != 0) PetscCall(PetscPrintf(ctx->comm, " | %22.16e\n", (double)tmp)); 1513 } 1514 } 1515 maps[grid].num_reduced++; 1516 PetscCheck(maps[grid].num_reduced < MAP_BF_SIZE, PETSC_COMM_SELF, PETSC_ERR_PLIB, "maps[grid].num_reduced %d > %" PetscInt_FMT, maps[grid].num_reduced, MAP_BF_SIZE); 1517 } 1518 break; 1519 } 1520 } 1521 } // !simplex 1522 // cleanup 1523 PetscCall(DMPlexRestoreClosureIndices(ctx->plex[grid], section[grid], globsection[grid], ej, PETSC_TRUE, &numindices, &indices, NULL, (PetscScalar **)&elMat)); 1524 if (elMat != valuesOrig) PetscCall(DMRestoreWorkArray(ctx->plex[grid], numindices * numindices, MPIU_SCALAR, &elMat)); 1525 } 1526 { // setup COO assembly 1527 coo_elem_offsets[glb_elem_idx + 1] += fullNb * fullNb; // one species block, adds a block for each species, on this element in this grid 1528 if (fieldA == 0) { // cache full Nb for this element, on this grid per species 1529 coo_elem_fullNb[glb_elem_idx] = fullNb; 1530 if (fullNb > ctx->SData_d.coo_max_fullnb) ctx->SData_d.coo_max_fullnb = fullNb; 1531 } else PetscCheck(coo_elem_fullNb[glb_elem_idx] == fullNb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "full element size change with species %d %d", coo_elem_fullNb[glb_elem_idx], fullNb); 1532 } 1533 } // field 1534 } // cell 1535 // allocate and copy point data maps[grid].gIdx[eidx][field][q] 1536 PetscCall(PetscMalloc(maps[grid].num_reduced * sizeof(*maps[grid].c_maps), &maps[grid].c_maps)); 1537 for (int ej = 0; ej < maps[grid].num_reduced; ++ej) { 1538 for (int q = 0; q < maps[grid].num_face; ++q) { 1539 maps[grid].c_maps[ej][q].scale = pointMaps[ej][q].scale; 1540 maps[grid].c_maps[ej][q].gid = pointMaps[ej][q].gid; 1541 } 1542 } 1543 #if defined(PETSC_HAVE_KOKKOS) 1544 if (ctx->deviceType == LANDAU_KOKKOS) { 1545 PetscCall(LandauKokkosCreateMatMaps(maps, pointMaps, Nf, grid)); // implies Kokkos does 1546 } 1547 #endif 1548 if (plex_batch) { 1549 PetscCall(ISRestoreIndices(grid_batch_is_inv[grid], &plex_batch)); 1550 PetscCall(ISDestroy(&grid_batch_is_inv[grid])); // we are done with this 1551 } 1552 } /* grids */ 1553 // finish COO 1554 { // setup COO assembly 1555 PetscInt *oor, *ooc; 1556 ctx->SData_d.coo_size = coo_elem_offsets[ncellsTot] * ctx->batch_sz; 1557 PetscCall(PetscMalloc2(ctx->SData_d.coo_size, &oor, ctx->SData_d.coo_size, &ooc)); 1558 for (int i = 0; i < ctx->SData_d.coo_size; i++) oor[i] = ooc[i] = -1; 1559 // get 1560 for (int grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) { 1561 for (int ej = 0; ej < numCells[grid]; ++ej, glb_elem_idx++) { 1562 const int fullNb = coo_elem_fullNb[glb_elem_idx]; 1563 const LandauIdx *const Idxs = &maps[grid].gIdx[ej][0][0]; // just use field-0 maps, They should be the same but this is just for COO storage 1564 coo_elem_point_offsets[glb_elem_idx][0] = 0; 1565 for (int f = 0, cnt2 = 0; f < Nb; f++) { 1566 int idx = Idxs[f]; 1567 coo_elem_point_offsets[glb_elem_idx][f + 1] = coo_elem_point_offsets[glb_elem_idx][f]; // start at last 1568 if (idx >= 0) { 1569 cnt2++; 1570 coo_elem_point_offsets[glb_elem_idx][f + 1]++; // inc 1571 } else { 1572 idx = -idx - 1; 1573 for (int q = 0; q < maps[grid].num_face; q++) { 1574 if (maps[grid].c_maps[idx][q].gid < 0) break; 1575 cnt2++; 1576 coo_elem_point_offsets[glb_elem_idx][f + 1]++; // inc 1577 } 1578 } 1579 PetscCheck(cnt2 <= fullNb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "wrong count %d < %d", fullNb, cnt2); 1580 } 1581 PetscCheck(coo_elem_point_offsets[glb_elem_idx][Nb] == fullNb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "coo_elem_point_offsets size %d != fullNb=%d", coo_elem_point_offsets[glb_elem_idx][Nb], fullNb); 1582 } 1583 } 1584 // set 1585 for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { 1586 for (int grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) { 1587 const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset); 1588 for (int ej = 0; ej < numCells[grid]; ++ej, glb_elem_idx++) { 1589 const int fullNb = coo_elem_fullNb[glb_elem_idx], fullNb2 = fullNb * fullNb; 1590 // set (i,j) 1591 for (int fieldA = 0; fieldA < Nf[grid]; fieldA++) { 1592 const LandauIdx *const Idxs = &maps[grid].gIdx[ej][fieldA][0]; 1593 int rows[LANDAU_MAX_Q_FACE], cols[LANDAU_MAX_Q_FACE]; 1594 for (int f = 0; f < Nb; ++f) { 1595 const int nr = coo_elem_point_offsets[glb_elem_idx][f + 1] - coo_elem_point_offsets[glb_elem_idx][f]; 1596 if (nr == 1) rows[0] = Idxs[f]; 1597 else { 1598 const int idx = -Idxs[f] - 1; 1599 for (int q = 0; q < nr; q++) rows[q] = maps[grid].c_maps[idx][q].gid; 1600 } 1601 for (int g = 0; g < Nb; ++g) { 1602 const int nc = coo_elem_point_offsets[glb_elem_idx][g + 1] - coo_elem_point_offsets[glb_elem_idx][g]; 1603 if (nc == 1) cols[0] = Idxs[g]; 1604 else { 1605 const int idx = -Idxs[g] - 1; 1606 for (int q = 0; q < nc; q++) cols[q] = maps[grid].c_maps[idx][q].gid; 1607 } 1608 const int idx0 = b_id * coo_elem_offsets[ncellsTot] + coo_elem_offsets[glb_elem_idx] + fieldA * fullNb2 + fullNb * coo_elem_point_offsets[glb_elem_idx][f] + nr * coo_elem_point_offsets[glb_elem_idx][g]; 1609 for (int q = 0, idx = idx0; q < nr; q++) { 1610 for (int d = 0; d < nc; d++, idx++) { 1611 oor[idx] = rows[q] + moffset; 1612 ooc[idx] = cols[d] + moffset; 1613 } 1614 } 1615 } 1616 } 1617 } 1618 } // cell 1619 } // grid 1620 } // batch 1621 PetscCall(MatSetPreallocationCOO(ctx->J, ctx->SData_d.coo_size, oor, ooc)); 1622 PetscCall(PetscFree2(oor, ooc)); 1623 } 1624 PetscCall(PetscFree(pointMaps)); 1625 PetscCall(PetscFree(elemMatrix)); 1626 PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container)); 1627 PetscCall(PetscContainerSetPointer(container, (void *)maps)); 1628 PetscCall(PetscContainerSetUserDestroy(container, LandauGPUMapsDestroy)); 1629 PetscCall(PetscObjectCompose((PetscObject)ctx->J, "assembly_maps", (PetscObject)container)); 1630 PetscCall(PetscContainerDestroy(&container)); 1631 PetscCall(PetscLogEventEnd(ctx->events[2], 0, 0, 0, 0)); 1632 } // end GPU assembly 1633 { /* create static point data, Jacobian called first, only one vertex copy */ 1634 PetscReal *invJe, *ww, *xx, *yy, *zz = NULL, *invJ_a; 1635 PetscInt outer_ipidx, outer_ej, grid, nip_glb = 0; 1636 PetscFE fe; 1637 PetscCall(PetscLogEventBegin(ctx->events[7], 0, 0, 0, 0)); 1638 PetscCall(PetscInfo(ctx->plex[0], "Initialize static data\n")); 1639 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) nip_glb += Nq * numCells[grid]; 1640 /* collect f data, first time is for Jacobian, but make mass now */ 1641 if (ctx->verbose != 0) { 1642 PetscInt ncells = 0, N; 1643 PetscCall(MatGetSize(ctx->J, &N, NULL)); 1644 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ncells += numCells[grid]; 1645 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d) %s %" PetscInt_FMT " IPs, %" PetscInt_FMT " cells total, Nb=%" PetscInt_FMT ", Nq=%" PetscInt_FMT ", dim=%" PetscInt_FMT ", Tab: Nb=%" PetscInt_FMT " Nf=%" PetscInt_FMT " Np=%" PetscInt_FMT " cdim=%" PetscInt_FMT " N=%" PetscInt_FMT "\n", 0, "FormLandau", nip_glb, ncells, Nb, Nq, dim, Nb, 1646 ctx->num_species, Nb, dim, N)); 1647 } 1648 PetscCall(PetscMalloc4(nip_glb, &ww, nip_glb, &xx, nip_glb, &yy, nip_glb * dim * dim, &invJ_a)); 1649 if (dim == 3) PetscCall(PetscMalloc1(nip_glb, &zz)); 1650 if (ctx->use_energy_tensor_trick) { 1651 PetscCall(PetscFECreateDefault(PETSC_COMM_SELF, dim, 1, ctx->simplex, NULL, PETSC_DECIDE, &fe)); 1652 PetscCall(PetscObjectSetName((PetscObject)fe, "energy")); 1653 } 1654 /* init each grids static data - no batch */ 1655 for (grid = 0, outer_ipidx = 0, outer_ej = 0; grid < ctx->num_grids; grid++) { // OpenMP (once) 1656 Vec v2_2 = NULL; // projected function: v^2/2 for non-relativistic, gamma... for relativistic 1657 PetscSection e_section; 1658 DM dmEnergy; 1659 PetscInt cStart, cEnd, ej; 1660 1661 PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd)); 1662 // prep energy trick, get v^2 / 2 vector 1663 if (ctx->use_energy_tensor_trick) { 1664 PetscErrorCode (*energyf[1])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *) = {ctx->use_relativistic_corrections ? gamma_m1_f : energy_f}; 1665 Vec glob_v2; 1666 PetscReal *c2_0[1], data[1] = {PetscSqr(C_0(ctx->v_0))}; 1667 1668 PetscCall(DMClone(ctx->plex[grid], &dmEnergy)); 1669 PetscCall(PetscObjectSetName((PetscObject)dmEnergy, "energy")); 1670 PetscCall(DMSetField(dmEnergy, 0, NULL, (PetscObject)fe)); 1671 PetscCall(DMCreateDS(dmEnergy)); 1672 PetscCall(DMGetSection(dmEnergy, &e_section)); 1673 PetscCall(DMGetGlobalVector(dmEnergy, &glob_v2)); 1674 PetscCall(PetscObjectSetName((PetscObject)glob_v2, "trick")); 1675 c2_0[0] = &data[0]; 1676 PetscCall(DMProjectFunction(dmEnergy, 0., energyf, (void **)c2_0, INSERT_ALL_VALUES, glob_v2)); 1677 PetscCall(DMGetLocalVector(dmEnergy, &v2_2)); 1678 PetscCall(VecZeroEntries(v2_2)); /* zero BCs so don't set */ 1679 PetscCall(DMGlobalToLocalBegin(dmEnergy, glob_v2, INSERT_VALUES, v2_2)); 1680 PetscCall(DMGlobalToLocalEnd(dmEnergy, glob_v2, INSERT_VALUES, v2_2)); 1681 PetscCall(DMViewFromOptions(dmEnergy, NULL, "-energy_dm_view")); 1682 PetscCall(VecViewFromOptions(glob_v2, NULL, "-energy_vec_view")); 1683 PetscCall(DMRestoreGlobalVector(dmEnergy, &glob_v2)); 1684 } 1685 /* append part of the IP data for each grid */ 1686 for (ej = 0; ej < numCells[grid]; ++ej, ++outer_ej) { 1687 PetscScalar *coefs = NULL; 1688 PetscReal vj[LANDAU_MAX_NQND * LANDAU_DIM], detJj[LANDAU_MAX_NQND], Jdummy[LANDAU_MAX_NQND * LANDAU_DIM * LANDAU_DIM], c0 = C_0(ctx->v_0), c02 = PetscSqr(c0); 1689 invJe = invJ_a + outer_ej * Nq * dim * dim; 1690 PetscCall(DMPlexComputeCellGeometryFEM(ctx->plex[grid], ej + cStart, quad, vj, Jdummy, invJe, detJj)); 1691 if (ctx->use_energy_tensor_trick) PetscCall(DMPlexVecGetClosure(dmEnergy, e_section, v2_2, ej + cStart, NULL, &coefs)); 1692 /* create static point data */ 1693 for (PetscInt qj = 0; qj < Nq; qj++, outer_ipidx++) { 1694 const PetscInt gidx = outer_ipidx; 1695 const PetscReal *invJ = &invJe[qj * dim * dim]; 1696 ww[gidx] = detJj[qj] * quadWeights[qj]; 1697 if (dim == 2) ww[gidx] *= vj[qj * dim + 0]; /* cylindrical coordinate, w/o 2pi */ 1698 // get xx, yy, zz 1699 if (ctx->use_energy_tensor_trick) { 1700 double refSpaceDer[3], eGradPhi[3]; 1701 const PetscReal *const DD = Tf[0]->T[1]; 1702 const PetscReal *Dq = &DD[qj * Nb * dim]; 1703 for (int d = 0; d < 3; ++d) refSpaceDer[d] = eGradPhi[d] = 0.0; 1704 for (int b = 0; b < Nb; ++b) { 1705 for (int d = 0; d < dim; ++d) refSpaceDer[d] += Dq[b * dim + d] * PetscRealPart(coefs[b]); 1706 } 1707 xx[gidx] = 1e10; 1708 if (ctx->use_relativistic_corrections) { 1709 double dg2_c2 = 0; 1710 //for (int d = 0; d < dim; ++d) refSpaceDer[d] *= c02; 1711 for (int d = 0; d < dim; ++d) dg2_c2 += PetscSqr(refSpaceDer[d]); 1712 dg2_c2 *= (double)c02; 1713 if (dg2_c2 >= .999) { 1714 xx[gidx] = vj[qj * dim + 0]; /* coordinate */ 1715 yy[gidx] = vj[qj * dim + 1]; 1716 if (dim == 3) zz[gidx] = vj[qj * dim + 2]; 1717 PetscCall(PetscPrintf(ctx->comm, "Error: %12.5e %" PetscInt_FMT ".%" PetscInt_FMT ") dg2/c02 = %12.5e x= %12.5e %12.5e %12.5e\n", (double)PetscSqrtReal(xx[gidx] * xx[gidx] + yy[gidx] * yy[gidx] + zz[gidx] * zz[gidx]), ej, qj, dg2_c2, (double)xx[gidx], (double)yy[gidx], (double)zz[gidx])); 1718 } else { 1719 PetscReal fact = c02 / PetscSqrtReal(1. - dg2_c2); 1720 for (int d = 0; d < dim; ++d) refSpaceDer[d] *= fact; 1721 // could test with other point u' that (grad - grad') * U (refSpaceDer, refSpaceDer') == 0 1722 } 1723 } 1724 if (xx[gidx] == 1e10) { 1725 for (int d = 0; d < dim; ++d) { 1726 for (int e = 0; e < dim; ++e) eGradPhi[d] += invJ[e * dim + d] * refSpaceDer[e]; 1727 } 1728 xx[gidx] = eGradPhi[0]; 1729 yy[gidx] = eGradPhi[1]; 1730 if (dim == 3) zz[gidx] = eGradPhi[2]; 1731 } 1732 } else { 1733 xx[gidx] = vj[qj * dim + 0]; /* coordinate */ 1734 yy[gidx] = vj[qj * dim + 1]; 1735 if (dim == 3) zz[gidx] = vj[qj * dim + 2]; 1736 } 1737 } /* q */ 1738 if (ctx->use_energy_tensor_trick) PetscCall(DMPlexVecRestoreClosure(dmEnergy, e_section, v2_2, ej + cStart, NULL, &coefs)); 1739 } /* ej */ 1740 if (ctx->use_energy_tensor_trick) { 1741 PetscCall(DMRestoreLocalVector(dmEnergy, &v2_2)); 1742 PetscCall(DMDestroy(&dmEnergy)); 1743 } 1744 } /* grid */ 1745 if (ctx->use_energy_tensor_trick) PetscCall(PetscFEDestroy(&fe)); 1746 /* cache static data */ 1747 if (ctx->deviceType == LANDAU_KOKKOS) { 1748 #if defined(PETSC_HAVE_KOKKOS) 1749 PetscCall(LandauKokkosStaticDataSet(ctx->plex[0], Nq, Nb, ctx->batch_sz, ctx->num_grids, numCells, ctx->species_offset, ctx->mat_offset, nu_alpha, nu_beta, invMass, (PetscReal *)ctx->lambdas, invJ_a, xx, yy, zz, ww, &ctx->SData_d)); 1750 #else 1751 SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type kokkos not built"); 1752 #endif 1753 /* free */ 1754 PetscCall(PetscFree4(ww, xx, yy, invJ_a)); 1755 if (dim == 3) PetscCall(PetscFree(zz)); 1756 } else { /* CPU version, just copy in, only use part */ 1757 PetscReal *nu_alpha_p = (PetscReal *)ctx->SData_d.alpha, *nu_beta_p = (PetscReal *)ctx->SData_d.beta, *invMass_p = (PetscReal *)ctx->SData_d.invMass, *lambdas_p = NULL; // why set these ? 1758 ctx->SData_d.w = (void *)ww; 1759 ctx->SData_d.x = (void *)xx; 1760 ctx->SData_d.y = (void *)yy; 1761 ctx->SData_d.z = (void *)zz; 1762 ctx->SData_d.invJ = (void *)invJ_a; 1763 PetscCall(PetscMalloc4(ctx->num_species, &nu_alpha_p, ctx->num_species, &nu_beta_p, ctx->num_species, &invMass_p, LANDAU_MAX_GRIDS * LANDAU_MAX_GRIDS, &lambdas_p)); 1764 for (PetscInt ii = 0; ii < ctx->num_species; ii++) { 1765 nu_alpha_p[ii] = nu_alpha[ii]; 1766 nu_beta_p[ii] = nu_beta[ii]; 1767 invMass_p[ii] = invMass[ii]; 1768 } 1769 ctx->SData_d.alpha = (void *)nu_alpha_p; 1770 ctx->SData_d.beta = (void *)nu_beta_p; 1771 ctx->SData_d.invMass = (void *)invMass_p; 1772 ctx->SData_d.lambdas = (void *)lambdas_p; 1773 for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) { 1774 PetscReal(*lambdas)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS] = (PetscReal(*)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS])ctx->SData_d.lambdas; 1775 for (PetscInt gridj = 0; gridj < LANDAU_MAX_GRIDS; gridj++) { (*lambdas)[grid][gridj] = ctx->lambdas[grid][gridj]; } 1776 } 1777 } 1778 PetscCall(PetscLogEventEnd(ctx->events[7], 0, 0, 0, 0)); 1779 } // initialize 1780 PetscFunctionReturn(PETSC_SUCCESS); 1781 } 1782 1783 /* < v, u > */ 1784 static void g0_1(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, PetscReal u_tShift, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar g0[]) 1785 { 1786 g0[0] = 1.; 1787 } 1788 1789 /* < v, u > */ 1790 static void g0_fake(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, PetscReal u_tShift, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar g0[]) 1791 { 1792 static double ttt = 1e-12; 1793 g0[0] = ttt++; 1794 } 1795 1796 /* < v, u > */ 1797 static void g0_r(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, PetscReal u_tShift, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar g0[]) 1798 { 1799 g0[0] = 2. * PETSC_PI * x[0]; 1800 } 1801 1802 static PetscErrorCode MatrixNfDestroy(void *ptr) 1803 { 1804 PetscInt *nf = (PetscInt *)ptr; 1805 PetscFunctionBegin; 1806 PetscCall(PetscFree(nf)); 1807 PetscFunctionReturn(PETSC_SUCCESS); 1808 } 1809 1810 /* 1811 LandauCreateJacobianMatrix - creates ctx->J with without real data. Hard to keep sparse. 1812 - Like DMPlexLandauCreateMassMatrix. Should remove one and combine 1813 - has old support for field major ordering 1814 */ 1815 static PetscErrorCode LandauCreateJacobianMatrix(MPI_Comm comm, Vec X, IS grid_batch_is_inv[LANDAU_MAX_GRIDS], LandauCtx *ctx) 1816 { 1817 PetscInt *idxs = NULL; 1818 Mat subM[LANDAU_MAX_GRIDS]; 1819 1820 PetscFunctionBegin; 1821 if (!ctx->gpu_assembly) { /* we need GPU object with GPU assembly */ 1822 PetscFunctionReturn(PETSC_SUCCESS); 1823 } 1824 // get the RCM for this grid to separate out species into blocks -- create 'idxs' & 'ctx->batch_is' -- not used 1825 if (ctx->gpu_assembly && ctx->jacobian_field_major_order) PetscCall(PetscMalloc1(ctx->mat_offset[ctx->num_grids] * ctx->batch_sz, &idxs)); 1826 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 1827 const PetscInt *values, n = ctx->mat_offset[grid + 1] - ctx->mat_offset[grid]; 1828 Mat gMat; 1829 DM massDM; 1830 PetscDS prob; 1831 Vec tvec; 1832 // get "mass" matrix for reordering 1833 PetscCall(DMClone(ctx->plex[grid], &massDM)); 1834 PetscCall(DMCopyFields(ctx->plex[grid], massDM)); 1835 PetscCall(DMCreateDS(massDM)); 1836 PetscCall(DMGetDS(massDM, &prob)); 1837 for (int ix = 0, ii = ctx->species_offset[grid]; ii < ctx->species_offset[grid + 1]; ii++, ix++) PetscCall(PetscDSSetJacobian(prob, ix, ix, g0_fake, NULL, NULL, NULL)); 1838 PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only")); // this trick is need to both sparsify the matrix and avoid runtime error 1839 PetscCall(DMCreateMatrix(massDM, &gMat)); 1840 PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false")); 1841 PetscCall(MatSetOption(gMat, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE)); 1842 PetscCall(MatSetOption(gMat, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE)); 1843 PetscCall(DMCreateLocalVector(ctx->plex[grid], &tvec)); 1844 PetscCall(DMPlexSNESComputeJacobianFEM(massDM, tvec, gMat, gMat, ctx)); 1845 PetscCall(MatViewFromOptions(gMat, NULL, "-dm_landau_reorder_mat_view")); 1846 PetscCall(DMDestroy(&massDM)); 1847 PetscCall(VecDestroy(&tvec)); 1848 subM[grid] = gMat; 1849 if (ctx->gpu_assembly && ctx->jacobian_field_major_order) { 1850 MatOrderingType rtype = MATORDERINGRCM; 1851 IS isrow, isicol; 1852 PetscCall(MatGetOrdering(gMat, rtype, &isrow, &isicol)); 1853 PetscCall(ISInvertPermutation(isrow, PETSC_DECIDE, &grid_batch_is_inv[grid])); 1854 PetscCall(ISGetIndices(isrow, &values)); 1855 for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid 1856 #if !defined(LANDAU_SPECIES_MAJOR) 1857 PetscInt N = ctx->mat_offset[ctx->num_grids], n0 = ctx->mat_offset[grid] + b_id * N; 1858 for (int ii = 0; ii < n; ++ii) idxs[n0 + ii] = values[ii] + n0; 1859 #else 1860 PetscInt n0 = ctx->mat_offset[grid] * ctx->batch_sz + b_id * n; 1861 for (int ii = 0; ii < n; ++ii) idxs[n0 + ii] = values[ii] + n0; 1862 #endif 1863 } 1864 PetscCall(ISRestoreIndices(isrow, &values)); 1865 PetscCall(ISDestroy(&isrow)); 1866 PetscCall(ISDestroy(&isicol)); 1867 } 1868 } 1869 if (ctx->gpu_assembly && ctx->jacobian_field_major_order) PetscCall(ISCreateGeneral(comm, ctx->mat_offset[ctx->num_grids] * ctx->batch_sz, idxs, PETSC_OWN_POINTER, &ctx->batch_is)); 1870 // get a block matrix 1871 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 1872 Mat B = subM[grid]; 1873 PetscInt nloc, nzl, *colbuf, row, COL_BF_SIZE = 1024; 1874 PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf)); 1875 PetscCall(MatGetSize(B, &nloc, NULL)); 1876 for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { 1877 const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset); 1878 const PetscInt *cols; 1879 const PetscScalar *vals; 1880 for (int i = 0; i < nloc; i++) { 1881 PetscCall(MatGetRow(B, i, &nzl, NULL, NULL)); 1882 if (nzl > COL_BF_SIZE) { 1883 PetscCall(PetscFree(colbuf)); 1884 PetscCall(PetscInfo(ctx->plex[grid], "Realloc buffer %" PetscInt_FMT " to %" PetscInt_FMT " (row size %" PetscInt_FMT ") \n", COL_BF_SIZE, 2 * COL_BF_SIZE, nzl)); 1885 COL_BF_SIZE = nzl; 1886 PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf)); 1887 } 1888 PetscCall(MatGetRow(B, i, &nzl, &cols, &vals)); 1889 for (int j = 0; j < nzl; j++) colbuf[j] = cols[j] + moffset; 1890 row = i + moffset; 1891 PetscCall(MatSetValues(ctx->J, 1, &row, nzl, colbuf, vals, INSERT_VALUES)); 1892 PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals)); 1893 } 1894 } 1895 PetscCall(PetscFree(colbuf)); 1896 } 1897 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(MatDestroy(&subM[grid])); 1898 PetscCall(MatAssemblyBegin(ctx->J, MAT_FINAL_ASSEMBLY)); 1899 PetscCall(MatAssemblyEnd(ctx->J, MAT_FINAL_ASSEMBLY)); 1900 1901 // debug 1902 PetscCall(MatViewFromOptions(ctx->J, NULL, "-dm_landau_mat_view")); 1903 if (ctx->gpu_assembly && ctx->jacobian_field_major_order) { 1904 Mat mat_block_order; 1905 PetscCall(MatCreateSubMatrix(ctx->J, ctx->batch_is, ctx->batch_is, MAT_INITIAL_MATRIX, &mat_block_order)); // use MatPermute 1906 PetscCall(MatViewFromOptions(mat_block_order, NULL, "-dm_landau_mat_view")); 1907 PetscCall(MatDestroy(&mat_block_order)); 1908 PetscCall(VecScatterCreate(X, ctx->batch_is, X, NULL, &ctx->plex_batch)); 1909 PetscCall(VecDuplicate(X, &ctx->work_vec)); 1910 } 1911 1912 PetscFunctionReturn(PETSC_SUCCESS); 1913 } 1914 1915 PetscErrorCode DMPlexLandauCreateMassMatrix(DM pack, Mat *Amat); 1916 /*@C 1917 DMPlexLandauCreateVelocitySpace - Create a `DMPLEX` velocity space mesh 1918 1919 Collective 1920 1921 Input Parameters: 1922 + comm - The MPI communicator 1923 . dim - velocity space dimension (2 for axisymmetric, 3 for full 3X + 3V solver) 1924 - prefix - prefix for options (not tested) 1925 1926 Output Parameters: 1927 + pack - The `DM` object representing the mesh 1928 . X - A vector (user destroys) 1929 - J - Optional matrix (object destroys) 1930 1931 Level: beginner 1932 1933 .seealso: `DMPlexCreate()`, `DMPlexLandauDestroyVelocitySpace()` 1934 @*/ 1935 PetscErrorCode DMPlexLandauCreateVelocitySpace(MPI_Comm comm, PetscInt dim, const char prefix[], Vec *X, Mat *J, DM *pack) 1936 { 1937 LandauCtx *ctx; 1938 Vec Xsub[LANDAU_MAX_GRIDS]; 1939 IS grid_batch_is_inv[LANDAU_MAX_GRIDS]; 1940 1941 PetscFunctionBegin; 1942 PetscCheck(dim == 2 || dim == 3, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Only 2D and 3D supported"); 1943 PetscCheck(LANDAU_DIM == dim, PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " != LANDAU_DIM %d", dim, LANDAU_DIM); 1944 PetscCall(PetscNew(&ctx)); 1945 ctx->comm = comm; /* used for diagnostics and global errors */ 1946 /* process options */ 1947 PetscCall(ProcessOptions(ctx, prefix)); 1948 if (dim == 2) ctx->use_relativistic_corrections = PETSC_FALSE; 1949 /* Create Mesh */ 1950 PetscCall(DMCompositeCreate(PETSC_COMM_SELF, pack)); 1951 PetscCall(PetscLogEventBegin(ctx->events[13], 0, 0, 0, 0)); 1952 PetscCall(PetscLogEventBegin(ctx->events[15], 0, 0, 0, 0)); 1953 PetscCall(LandauDMCreateVMeshes(PETSC_COMM_SELF, dim, prefix, ctx, *pack)); // creates grids (Forest of AMR) 1954 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 1955 /* create FEM */ 1956 PetscCall(SetupDS(ctx->plex[grid], dim, grid, ctx)); 1957 /* set initial state */ 1958 PetscCall(DMCreateGlobalVector(ctx->plex[grid], &Xsub[grid])); 1959 PetscCall(PetscObjectSetName((PetscObject)Xsub[grid], "u_orig")); 1960 /* initial static refinement, no solve */ 1961 PetscCall(LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, 0, 1, ctx)); 1962 /* forest refinement - forest goes in (if forest), plex comes out */ 1963 if (ctx->use_p4est) { 1964 DM plex; 1965 PetscCall(adapt(grid, ctx, &Xsub[grid])); // forest goes in, plex comes out 1966 PetscCall(DMViewFromOptions(ctx->plex[grid], NULL, "-dm_landau_amr_dm_view")); // need to differentiate - todo 1967 PetscCall(VecViewFromOptions(Xsub[grid], NULL, "-dm_landau_amr_vec_view")); 1968 // convert to plex, all done with this level 1969 PetscCall(DMConvert(ctx->plex[grid], DMPLEX, &plex)); 1970 PetscCall(DMDestroy(&ctx->plex[grid])); 1971 ctx->plex[grid] = plex; 1972 } 1973 #if !defined(LANDAU_SPECIES_MAJOR) 1974 PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid])); 1975 #else 1976 for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid 1977 PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid])); 1978 } 1979 #endif 1980 PetscCall(DMSetApplicationContext(ctx->plex[grid], ctx)); 1981 } 1982 #if !defined(LANDAU_SPECIES_MAJOR) 1983 // stack the batched DMs, could do it all here!!! b_id=0 1984 for (PetscInt b_id = 1; b_id < ctx->batch_sz; b_id++) { 1985 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid])); 1986 } 1987 #endif 1988 // create ctx->mat_offset 1989 ctx->mat_offset[0] = 0; 1990 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 1991 PetscInt n; 1992 PetscCall(VecGetLocalSize(Xsub[grid], &n)); 1993 ctx->mat_offset[grid + 1] = ctx->mat_offset[grid] + n; 1994 } 1995 // creat DM & Jac 1996 PetscCall(DMSetApplicationContext(*pack, ctx)); 1997 PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only")); 1998 PetscCall(DMCreateMatrix(*pack, &ctx->J)); 1999 PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false")); 2000 PetscCall(MatSetOption(ctx->J, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE)); 2001 PetscCall(MatSetOption(ctx->J, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE)); 2002 PetscCall(PetscObjectSetName((PetscObject)ctx->J, "Jac")); 2003 // construct initial conditions in X 2004 PetscCall(DMCreateGlobalVector(*pack, X)); 2005 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 2006 PetscInt n; 2007 PetscCall(VecGetLocalSize(Xsub[grid], &n)); 2008 for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { 2009 PetscScalar const *values; 2010 const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset); 2011 PetscCall(LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, b_id, ctx->batch_sz, ctx)); 2012 PetscCall(VecGetArrayRead(Xsub[grid], &values)); // Drop whole grid in Plex ordering 2013 for (int i = 0, idx = moffset; i < n; i++, idx++) PetscCall(VecSetValue(*X, idx, values[i], INSERT_VALUES)); 2014 PetscCall(VecRestoreArrayRead(Xsub[grid], &values)); 2015 } 2016 } 2017 // cleanup 2018 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(VecDestroy(&Xsub[grid])); 2019 /* check for correct matrix type */ 2020 if (ctx->gpu_assembly) { /* we need GPU object with GPU assembly */ 2021 PetscBool flg; 2022 if (ctx->deviceType == LANDAU_KOKKOS) { 2023 PetscCall(PetscObjectTypeCompareAny((PetscObject)ctx->J, &flg, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, MATAIJKOKKOS, "")); 2024 #if defined(PETSC_HAVE_KOKKOS) 2025 PetscCheck(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'"); 2026 #else 2027 PetscCheck(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'"); 2028 #endif 2029 } 2030 } 2031 PetscCall(PetscLogEventEnd(ctx->events[15], 0, 0, 0, 0)); 2032 2033 // create field major ordering 2034 ctx->work_vec = NULL; 2035 ctx->plex_batch = NULL; 2036 ctx->batch_is = NULL; 2037 for (int i = 0; i < LANDAU_MAX_GRIDS; i++) grid_batch_is_inv[i] = NULL; 2038 PetscCall(PetscLogEventBegin(ctx->events[12], 0, 0, 0, 0)); 2039 PetscCall(LandauCreateJacobianMatrix(comm, *X, grid_batch_is_inv, ctx)); 2040 PetscCall(PetscLogEventEnd(ctx->events[12], 0, 0, 0, 0)); 2041 2042 // create AMR GPU assembly maps and static GPU data 2043 PetscCall(CreateStaticData(dim, grid_batch_is_inv, ctx)); 2044 2045 PetscCall(PetscLogEventEnd(ctx->events[13], 0, 0, 0, 0)); 2046 2047 // create mass matrix 2048 PetscCall(DMPlexLandauCreateMassMatrix(*pack, NULL)); 2049 2050 if (J) *J = ctx->J; 2051 2052 if (ctx->gpu_assembly && ctx->jacobian_field_major_order) { 2053 PetscContainer container; 2054 // cache ctx for KSP with batch/field major Jacobian ordering -ksp_type gmres/etc -dm_landau_jacobian_field_major_order 2055 PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container)); 2056 PetscCall(PetscContainerSetPointer(container, (void *)ctx)); 2057 PetscCall(PetscObjectCompose((PetscObject)ctx->J, "LandauCtx", (PetscObject)container)); 2058 PetscCall(PetscContainerDestroy(&container)); 2059 // batch solvers need to map -- can batch solvers work 2060 PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container)); 2061 PetscCall(PetscContainerSetPointer(container, (void *)ctx->plex_batch)); 2062 PetscCall(PetscObjectCompose((PetscObject)ctx->J, "plex_batch_is", (PetscObject)container)); 2063 PetscCall(PetscContainerDestroy(&container)); 2064 } 2065 // for batch solvers 2066 { 2067 PetscContainer container; 2068 PetscInt *pNf; 2069 PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container)); 2070 PetscCall(PetscMalloc1(sizeof(*pNf), &pNf)); 2071 *pNf = ctx->batch_sz; 2072 PetscCall(PetscContainerSetPointer(container, (void *)pNf)); 2073 PetscCall(PetscContainerSetUserDestroy(container, MatrixNfDestroy)); 2074 PetscCall(PetscObjectCompose((PetscObject)ctx->J, "batch size", (PetscObject)container)); 2075 PetscCall(PetscContainerDestroy(&container)); 2076 } 2077 2078 PetscFunctionReturn(PETSC_SUCCESS); 2079 } 2080 2081 /*@C 2082 DMPlexLandauAccess - Access to the distribution function with user callback 2083 2084 Collective 2085 2086 Input Parameters: 2087 + pack - the `DMCOMPOSITE` 2088 . func - call back function 2089 - user_ctx - user context 2090 2091 Input/Output Parameter: 2092 . X - Vector to data to 2093 2094 Level: advanced 2095 2096 .seealso: `DMPlexLandauCreateVelocitySpace()` 2097 @*/ 2098 PetscErrorCode DMPlexLandauAccess(DM pack, Vec X, PetscErrorCode (*func)(DM, Vec, PetscInt, PetscInt, PetscInt, void *), void *user_ctx) 2099 { 2100 LandauCtx *ctx; 2101 PetscFunctionBegin; 2102 PetscCall(DMGetApplicationContext(pack, &ctx)); // uses ctx->num_grids; ctx->plex[grid]; ctx->batch_sz; ctx->mat_offset 2103 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 2104 PetscInt dim, n; 2105 PetscCall(DMGetDimension(pack, &dim)); 2106 for (PetscInt sp = ctx->species_offset[grid], i0 = 0; sp < ctx->species_offset[grid + 1]; sp++, i0++) { 2107 Vec vec; 2108 PetscInt vf[1] = {i0}; 2109 IS vis; 2110 DM vdm; 2111 PetscCall(DMCreateSubDM(ctx->plex[grid], 1, vf, &vis, &vdm)); 2112 PetscCall(DMSetApplicationContext(vdm, ctx)); // the user might want this 2113 PetscCall(DMCreateGlobalVector(vdm, &vec)); 2114 PetscCall(VecGetSize(vec, &n)); 2115 for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { 2116 const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset); 2117 PetscCall(VecZeroEntries(vec)); 2118 /* Add your data with 'dm' for species 'sp' to 'vec' */ 2119 PetscCall(func(vdm, vec, i0, grid, b_id, user_ctx)); 2120 /* add to global */ 2121 PetscScalar const *values; 2122 const PetscInt *offsets; 2123 PetscCall(VecGetArrayRead(vec, &values)); 2124 PetscCall(ISGetIndices(vis, &offsets)); 2125 for (int i = 0; i < n; i++) PetscCall(VecSetValue(X, moffset + offsets[i], values[i], ADD_VALUES)); 2126 PetscCall(VecRestoreArrayRead(vec, &values)); 2127 PetscCall(ISRestoreIndices(vis, &offsets)); 2128 } // batch 2129 PetscCall(VecDestroy(&vec)); 2130 PetscCall(ISDestroy(&vis)); 2131 PetscCall(DMDestroy(&vdm)); 2132 } 2133 } // grid 2134 PetscFunctionReturn(PETSC_SUCCESS); 2135 } 2136 2137 /*@ 2138 DMPlexLandauDestroyVelocitySpace - Destroy a `DMPLEX` velocity space mesh 2139 2140 Collective 2141 2142 Input/Output Parameters: 2143 . dm - the `DM` to destroy 2144 2145 Level: beginner 2146 2147 .seealso: `DMPlexLandauCreateVelocitySpace()` 2148 @*/ 2149 PetscErrorCode DMPlexLandauDestroyVelocitySpace(DM *dm) 2150 { 2151 LandauCtx *ctx; 2152 PetscFunctionBegin; 2153 PetscCall(DMGetApplicationContext(*dm, &ctx)); 2154 PetscCall(MatDestroy(&ctx->M)); 2155 PetscCall(MatDestroy(&ctx->J)); 2156 for (PetscInt ii = 0; ii < ctx->num_species; ii++) PetscCall(PetscFEDestroy(&ctx->fe[ii])); 2157 PetscCall(ISDestroy(&ctx->batch_is)); 2158 PetscCall(VecDestroy(&ctx->work_vec)); 2159 PetscCall(VecScatterDestroy(&ctx->plex_batch)); 2160 if (ctx->deviceType == LANDAU_KOKKOS) { 2161 #if defined(PETSC_HAVE_KOKKOS) 2162 PetscCall(LandauKokkosStaticDataClear(&ctx->SData_d)); 2163 #else 2164 SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type %s not built", "kokkos"); 2165 #endif 2166 } else { 2167 if (ctx->SData_d.x) { /* in a CPU run */ 2168 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; 2169 LandauIdx *coo_elem_offsets = (LandauIdx *)ctx->SData_d.coo_elem_offsets, *coo_elem_fullNb = (LandauIdx *)ctx->SData_d.coo_elem_fullNb, (*coo_elem_point_offsets)[LANDAU_MAX_NQND + 1] = (LandauIdx(*)[LANDAU_MAX_NQND + 1]) ctx->SData_d.coo_elem_point_offsets; 2170 PetscCall(PetscFree4(ww, xx, yy, invJ)); 2171 if (zz) PetscCall(PetscFree(zz)); 2172 if (coo_elem_offsets) { 2173 PetscCall(PetscFree3(coo_elem_offsets, coo_elem_fullNb, coo_elem_point_offsets)); // could be NULL 2174 } 2175 PetscCall(PetscFree4(ctx->SData_d.alpha, ctx->SData_d.beta, ctx->SData_d.invMass, ctx->SData_d.lambdas)); 2176 } 2177 } 2178 2179 if (ctx->times[LANDAU_MATRIX_TOTAL] > 0) { // OMP timings 2180 PetscCall(PetscPrintf(ctx->comm, "TSStep N 1.0 %10.3e\n", ctx->times[LANDAU_EX2_TSSOLVE])); 2181 PetscCall(PetscPrintf(ctx->comm, "2: Solve: %10.3e with %" PetscInt_FMT " threads\n", ctx->times[LANDAU_EX2_TSSOLVE] - ctx->times[LANDAU_MATRIX_TOTAL], ctx->batch_sz)); 2182 PetscCall(PetscPrintf(ctx->comm, "3: Landau: %10.3e\n", ctx->times[LANDAU_MATRIX_TOTAL])); 2183 PetscCall(PetscPrintf(ctx->comm, "Landau Jacobian %" PetscInt_FMT " 1.0 %10.3e\n", (PetscInt)ctx->times[LANDAU_JACOBIAN_COUNT], ctx->times[LANDAU_JACOBIAN])); 2184 PetscCall(PetscPrintf(ctx->comm, "Landau Operator N 1.0 %10.3e\n", ctx->times[LANDAU_OPERATOR])); 2185 PetscCall(PetscPrintf(ctx->comm, "Landau Mass N 1.0 %10.3e\n", ctx->times[LANDAU_MASS])); 2186 PetscCall(PetscPrintf(ctx->comm, " Jac-f-df (GPU) N 1.0 %10.3e\n", ctx->times[LANDAU_F_DF])); 2187 PetscCall(PetscPrintf(ctx->comm, " Kernel (GPU) N 1.0 %10.3e\n", ctx->times[LANDAU_KERNEL])); 2188 PetscCall(PetscPrintf(ctx->comm, "MatLUFactorNum X 1.0 %10.3e\n", ctx->times[KSP_FACTOR])); 2189 PetscCall(PetscPrintf(ctx->comm, "MatSolve X 1.0 %10.3e\n", ctx->times[KSP_SOLVE])); 2190 } 2191 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMDestroy(&ctx->plex[grid])); 2192 PetscCall(PetscFree(ctx)); 2193 PetscCall(DMDestroy(dm)); 2194 PetscFunctionReturn(PETSC_SUCCESS); 2195 } 2196 2197 /* < v, ru > */ 2198 static void f0_s_den(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0) 2199 { 2200 PetscInt ii = (PetscInt)PetscRealPart(constants[0]); 2201 f0[0] = u[ii]; 2202 } 2203 2204 /* < v, ru > */ 2205 static void f0_s_mom(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0) 2206 { 2207 PetscInt ii = (PetscInt)PetscRealPart(constants[0]), jj = (PetscInt)PetscRealPart(constants[1]); 2208 f0[0] = x[jj] * u[ii]; /* x momentum */ 2209 } 2210 2211 static void f0_s_v2(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0) 2212 { 2213 PetscInt i, ii = (PetscInt)PetscRealPart(constants[0]); 2214 double tmp1 = 0.; 2215 for (i = 0; i < dim; ++i) tmp1 += x[i] * x[i]; 2216 f0[0] = tmp1 * u[ii]; 2217 } 2218 2219 static PetscErrorCode gamma_n_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf, PetscScalar *u, void *actx) 2220 { 2221 const PetscReal *c2_0_arr = ((PetscReal *)actx); 2222 const PetscReal c02 = c2_0_arr[0]; 2223 2224 PetscFunctionBegin; 2225 for (int s = 0; s < Nf; s++) { 2226 PetscReal tmp1 = 0.; 2227 for (int i = 0; i < dim; ++i) tmp1 += x[i] * x[i]; 2228 #if defined(PETSC_USE_DEBUG) 2229 u[s] = PetscSqrtReal(1. + tmp1 / c02); // u[0] = PetscSqrtReal(1. + xx); 2230 #else 2231 { 2232 PetscReal xx = tmp1 / c02; 2233 u[s] = xx / (PetscSqrtReal(1. + xx) + 1.); // better conditioned = xx/(PetscSqrtReal(1. + xx) + 1.) 2234 } 2235 #endif 2236 } 2237 PetscFunctionReturn(PETSC_SUCCESS); 2238 } 2239 2240 /* < v, ru > */ 2241 static void f0_s_rden(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0) 2242 { 2243 PetscInt ii = (PetscInt)PetscRealPart(constants[0]); 2244 f0[0] = 2. * PETSC_PI * x[0] * u[ii]; 2245 } 2246 2247 /* < v, ru > */ 2248 static void f0_s_rmom(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0) 2249 { 2250 PetscInt ii = (PetscInt)PetscRealPart(constants[0]); 2251 f0[0] = 2. * PETSC_PI * x[0] * x[1] * u[ii]; 2252 } 2253 2254 static void f0_s_rv2(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0) 2255 { 2256 PetscInt ii = (PetscInt)PetscRealPart(constants[0]); 2257 f0[0] = 2. * PETSC_PI * x[0] * (x[0] * x[0] + x[1] * x[1]) * u[ii]; 2258 } 2259 2260 /*@ 2261 DMPlexLandauPrintNorms - collects moments and prints them 2262 2263 Collective 2264 2265 Input Parameters: 2266 + X - the state 2267 - stepi - current step to print 2268 2269 Level: beginner 2270 2271 .seealso: `DMPlexLandauCreateVelocitySpace()` 2272 @*/ 2273 PetscErrorCode DMPlexLandauPrintNorms(Vec X, PetscInt stepi) 2274 { 2275 LandauCtx *ctx; 2276 PetscDS prob; 2277 DM pack; 2278 PetscInt cStart, cEnd, dim, ii, i0, nDMs; 2279 PetscScalar xmomentumtot = 0, ymomentumtot = 0, zmomentumtot = 0, energytot = 0, densitytot = 0, tt[LANDAU_MAX_SPECIES]; 2280 PetscScalar xmomentum[LANDAU_MAX_SPECIES], ymomentum[LANDAU_MAX_SPECIES], zmomentum[LANDAU_MAX_SPECIES], energy[LANDAU_MAX_SPECIES], density[LANDAU_MAX_SPECIES]; 2281 Vec *globXArray; 2282 2283 PetscFunctionBegin; 2284 PetscCall(VecGetDM(X, &pack)); 2285 PetscCheck(pack, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Vector has no DM"); 2286 PetscCall(DMGetDimension(pack, &dim)); 2287 PetscCheck(dim == 2 || dim == 3, PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " not in [2,3]", dim); 2288 PetscCall(DMGetApplicationContext(pack, &ctx)); 2289 PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 2290 /* print momentum and energy */ 2291 PetscCall(DMCompositeGetNumberDM(pack, &nDMs)); 2292 PetscCheck(nDMs == ctx->num_grids * ctx->batch_sz, PETSC_COMM_WORLD, PETSC_ERR_PLIB, "#DM wrong %" PetscInt_FMT " %" PetscInt_FMT, nDMs, ctx->num_grids * ctx->batch_sz); 2293 PetscCall(PetscMalloc(sizeof(*globXArray) * nDMs, &globXArray)); 2294 PetscCall(DMCompositeGetAccessArray(pack, X, nDMs, NULL, globXArray)); 2295 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 2296 Vec Xloc = globXArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)]; 2297 PetscCall(DMGetDS(ctx->plex[grid], &prob)); 2298 for (ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) { 2299 PetscScalar user[2] = {(PetscScalar)i0, (PetscScalar)ctx->charges[ii]}; 2300 PetscCall(PetscDSSetConstants(prob, 2, user)); 2301 if (dim == 2) { /* 2/3X + 3V (cylindrical coordinates) */ 2302 PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rden)); 2303 PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx)); 2304 density[ii] = tt[0] * ctx->n_0 * ctx->charges[ii]; 2305 PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rmom)); 2306 PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx)); 2307 zmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii]; 2308 PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rv2)); 2309 PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx)); 2310 energy[ii] = tt[0] * 0.5 * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ii]; 2311 zmomentumtot += zmomentum[ii]; 2312 energytot += energy[ii]; 2313 densitytot += density[ii]; 2314 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%3" PetscInt_FMT ") species-%" PetscInt_FMT ": charge density= %20.13e z-momentum= %20.13e energy= %20.13e", stepi, ii, (double)PetscRealPart(density[ii]), (double)PetscRealPart(zmomentum[ii]), (double)PetscRealPart(energy[ii]))); 2315 } else { /* 2/3Xloc + 3V */ 2316 PetscCall(PetscDSSetObjective(prob, 0, &f0_s_den)); 2317 PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx)); 2318 density[ii] = tt[0] * ctx->n_0 * ctx->charges[ii]; 2319 PetscCall(PetscDSSetObjective(prob, 0, &f0_s_mom)); 2320 user[1] = 0; 2321 PetscCall(PetscDSSetConstants(prob, 2, user)); 2322 PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx)); 2323 xmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii]; 2324 user[1] = 1; 2325 PetscCall(PetscDSSetConstants(prob, 2, user)); 2326 PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx)); 2327 ymomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii]; 2328 user[1] = 2; 2329 PetscCall(PetscDSSetConstants(prob, 2, user)); 2330 PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx)); 2331 zmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii]; 2332 if (ctx->use_relativistic_corrections) { 2333 /* gamma * M * f */ 2334 if (ii == 0 && grid == 0) { // do all at once 2335 Vec Mf, globGamma, *globMfArray, *globGammaArray; 2336 PetscErrorCode (*gammaf[1])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *) = {gamma_n_f}; 2337 PetscReal *c2_0[1], data[1]; 2338 2339 PetscCall(VecDuplicate(X, &globGamma)); 2340 PetscCall(VecDuplicate(X, &Mf)); 2341 PetscCall(PetscMalloc(sizeof(*globMfArray) * nDMs, &globMfArray)); 2342 PetscCall(PetscMalloc(sizeof(*globMfArray) * nDMs, &globGammaArray)); 2343 /* M * f */ 2344 PetscCall(MatMult(ctx->M, X, Mf)); 2345 /* gamma */ 2346 PetscCall(DMCompositeGetAccessArray(pack, globGamma, nDMs, NULL, globGammaArray)); 2347 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 2348 Vec v1 = globGammaArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)]; 2349 data[0] = PetscSqr(C_0(ctx->v_0)); 2350 c2_0[0] = &data[0]; 2351 PetscCall(DMProjectFunction(ctx->plex[grid], 0., gammaf, (void **)c2_0, INSERT_ALL_VALUES, v1)); 2352 } 2353 PetscCall(DMCompositeRestoreAccessArray(pack, globGamma, nDMs, NULL, globGammaArray)); 2354 /* gamma * Mf */ 2355 PetscCall(DMCompositeGetAccessArray(pack, globGamma, nDMs, NULL, globGammaArray)); 2356 PetscCall(DMCompositeGetAccessArray(pack, Mf, nDMs, NULL, globMfArray)); 2357 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { // yes a grid loop in a grid loop to print nice 2358 PetscInt Nf = ctx->species_offset[grid + 1] - ctx->species_offset[grid], N, bs; 2359 Vec Mfsub = globMfArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)], Gsub = globGammaArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)], v1, v2; 2360 // get each component 2361 PetscCall(VecGetSize(Mfsub, &N)); 2362 PetscCall(VecCreate(ctx->comm, &v1)); 2363 PetscCall(VecSetSizes(v1, PETSC_DECIDE, N / Nf)); 2364 PetscCall(VecCreate(ctx->comm, &v2)); 2365 PetscCall(VecSetSizes(v2, PETSC_DECIDE, N / Nf)); 2366 PetscCall(VecSetFromOptions(v1)); // ??? 2367 PetscCall(VecSetFromOptions(v2)); 2368 // get each component 2369 PetscCall(VecGetBlockSize(Gsub, &bs)); 2370 PetscCheck(bs == Nf, PETSC_COMM_SELF, PETSC_ERR_PLIB, "bs %" PetscInt_FMT " != num_species %" PetscInt_FMT " in Gsub", bs, Nf); 2371 PetscCall(VecGetBlockSize(Mfsub, &bs)); 2372 PetscCheck(bs == Nf, PETSC_COMM_SELF, PETSC_ERR_PLIB, "bs %" PetscInt_FMT " != num_species %" PetscInt_FMT, bs, Nf); 2373 for (int i = 0, ix = ctx->species_offset[grid]; i < Nf; i++, ix++) { 2374 PetscScalar val; 2375 PetscCall(VecStrideGather(Gsub, i, v1, INSERT_VALUES)); // this is not right -- TODO 2376 PetscCall(VecStrideGather(Mfsub, i, v2, INSERT_VALUES)); 2377 PetscCall(VecDot(v1, v2, &val)); 2378 energy[ix] = PetscRealPart(val) * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ix]; 2379 } 2380 PetscCall(VecDestroy(&v1)); 2381 PetscCall(VecDestroy(&v2)); 2382 } /* grids */ 2383 PetscCall(DMCompositeRestoreAccessArray(pack, globGamma, nDMs, NULL, globGammaArray)); 2384 PetscCall(DMCompositeRestoreAccessArray(pack, Mf, nDMs, NULL, globMfArray)); 2385 PetscCall(PetscFree(globGammaArray)); 2386 PetscCall(PetscFree(globMfArray)); 2387 PetscCall(VecDestroy(&globGamma)); 2388 PetscCall(VecDestroy(&Mf)); 2389 } 2390 } else { 2391 PetscCall(PetscDSSetObjective(prob, 0, &f0_s_v2)); 2392 PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx)); 2393 energy[ii] = 0.5 * tt[0] * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ii]; 2394 } 2395 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%3" PetscInt_FMT ") species %" PetscInt_FMT ": density=%20.13e, x-momentum=%20.13e, y-momentum=%20.13e, z-momentum=%20.13e, energy=%21.13e", stepi, ii, (double)PetscRealPart(density[ii]), (double)PetscRealPart(xmomentum[ii]), (double)PetscRealPart(ymomentum[ii]), (double)PetscRealPart(zmomentum[ii]), (double)PetscRealPart(energy[ii]))); 2396 xmomentumtot += xmomentum[ii]; 2397 ymomentumtot += ymomentum[ii]; 2398 zmomentumtot += zmomentum[ii]; 2399 energytot += energy[ii]; 2400 densitytot += density[ii]; 2401 } 2402 if (ctx->num_species > 1) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n")); 2403 } 2404 } 2405 PetscCall(DMCompositeRestoreAccessArray(pack, X, nDMs, NULL, globXArray)); 2406 PetscCall(PetscFree(globXArray)); 2407 /* totals */ 2408 PetscCall(DMPlexGetHeightStratum(ctx->plex[0], 0, &cStart, &cEnd)); 2409 if (ctx->num_species > 1) { 2410 if (dim == 2) { 2411 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\t%3" PetscInt_FMT ") Total: charge density=%21.13e, momentum=%21.13e, energy=%21.13e (m_i[0]/m_e = %g, %" PetscInt_FMT " cells on electron grid)", stepi, (double)PetscRealPart(densitytot), (double)PetscRealPart(zmomentumtot), (double)PetscRealPart(energytot), 2412 (double)(ctx->masses[1] / ctx->masses[0]), cEnd - cStart)); 2413 } else { 2414 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\t%3" PetscInt_FMT ") Total: charge density=%21.13e, x-momentum=%21.13e, y-momentum=%21.13e, z-momentum=%21.13e, energy=%21.13e (m_i[0]/m_e = %g, %" PetscInt_FMT " cells)", stepi, (double)PetscRealPart(densitytot), (double)PetscRealPart(xmomentumtot), (double)PetscRealPart(ymomentumtot), (double)PetscRealPart(zmomentumtot), (double)PetscRealPart(energytot), 2415 (double)(ctx->masses[1] / ctx->masses[0]), cEnd - cStart)); 2416 } 2417 } else PetscCall(PetscPrintf(PETSC_COMM_WORLD, " -- %" PetscInt_FMT " cells", cEnd - cStart)); 2418 PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n")); 2419 PetscFunctionReturn(PETSC_SUCCESS); 2420 } 2421 2422 /*@ 2423 DMPlexLandauCreateMassMatrix - Create mass matrix for Landau in Plex space (not field major order of Jacobian) 2424 - puts mass matrix into ctx->M 2425 2426 Collective 2427 2428 Input Parameter: 2429 . pack - the `DM` object. Puts matrix in Landau context M field 2430 2431 Output Parameter: 2432 . Amat - The mass matrix (optional), mass matrix is added to the `DM` context 2433 2434 Level: beginner 2435 2436 .seealso: `DMPlexLandauCreateVelocitySpace()` 2437 @*/ 2438 PetscErrorCode DMPlexLandauCreateMassMatrix(DM pack, Mat *Amat) 2439 { 2440 DM mass_pack, massDM[LANDAU_MAX_GRIDS]; 2441 PetscDS prob; 2442 PetscInt ii, dim, N1 = 1, N2; 2443 LandauCtx *ctx; 2444 Mat packM, subM[LANDAU_MAX_GRIDS]; 2445 2446 PetscFunctionBegin; 2447 PetscValidHeaderSpecific(pack, DM_CLASSID, 1); 2448 if (Amat) PetscAssertPointer(Amat, 2); 2449 PetscCall(DMGetApplicationContext(pack, &ctx)); 2450 PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 2451 PetscCall(PetscLogEventBegin(ctx->events[14], 0, 0, 0, 0)); 2452 PetscCall(DMGetDimension(pack, &dim)); 2453 PetscCall(DMCompositeCreate(PetscObjectComm((PetscObject)pack), &mass_pack)); 2454 /* create pack mass matrix */ 2455 for (PetscInt grid = 0, ix = 0; grid < ctx->num_grids; grid++) { 2456 PetscCall(DMClone(ctx->plex[grid], &massDM[grid])); 2457 PetscCall(DMCopyFields(ctx->plex[grid], massDM[grid])); 2458 PetscCall(DMCreateDS(massDM[grid])); 2459 PetscCall(DMGetDS(massDM[grid], &prob)); 2460 for (ix = 0, ii = ctx->species_offset[grid]; ii < ctx->species_offset[grid + 1]; ii++, ix++) { 2461 if (dim == 3) PetscCall(PetscDSSetJacobian(prob, ix, ix, g0_1, NULL, NULL, NULL)); 2462 else PetscCall(PetscDSSetJacobian(prob, ix, ix, g0_r, NULL, NULL, NULL)); 2463 } 2464 #if !defined(LANDAU_SPECIES_MAJOR) 2465 PetscCall(DMCompositeAddDM(mass_pack, massDM[grid])); 2466 #else 2467 for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid 2468 PetscCall(DMCompositeAddDM(mass_pack, massDM[grid])); 2469 } 2470 #endif 2471 PetscCall(DMCreateMatrix(massDM[grid], &subM[grid])); 2472 } 2473 #if !defined(LANDAU_SPECIES_MAJOR) 2474 // stack the batched DMs 2475 for (PetscInt b_id = 1; b_id < ctx->batch_sz; b_id++) { 2476 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCompositeAddDM(mass_pack, massDM[grid])); 2477 } 2478 #endif 2479 PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only")); 2480 PetscCall(DMCreateMatrix(mass_pack, &packM)); 2481 PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false")); 2482 PetscCall(MatSetOption(packM, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE)); 2483 PetscCall(MatSetOption(packM, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE)); 2484 PetscCall(DMDestroy(&mass_pack)); 2485 /* make mass matrix for each block */ 2486 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 2487 Vec locX; 2488 DM plex = massDM[grid]; 2489 PetscCall(DMGetLocalVector(plex, &locX)); 2490 /* Mass matrix is independent of the input, so no need to fill locX */ 2491 PetscCall(DMPlexSNESComputeJacobianFEM(plex, locX, subM[grid], subM[grid], ctx)); 2492 PetscCall(DMRestoreLocalVector(plex, &locX)); 2493 PetscCall(DMDestroy(&massDM[grid])); 2494 } 2495 PetscCall(MatGetSize(ctx->J, &N1, NULL)); 2496 PetscCall(MatGetSize(packM, &N2, NULL)); 2497 PetscCheck(N1 == N2, PetscObjectComm((PetscObject)pack), PETSC_ERR_PLIB, "Incorrect matrix sizes: |Jacobian| = %" PetscInt_FMT ", |Mass|=%" PetscInt_FMT, N1, N2); 2498 /* assemble block diagonals */ 2499 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { 2500 Mat B = subM[grid]; 2501 PetscInt nloc, nzl, *colbuf, COL_BF_SIZE = 1024, row; 2502 PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf)); 2503 PetscCall(MatGetSize(B, &nloc, NULL)); 2504 for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { 2505 const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset); 2506 const PetscInt *cols; 2507 const PetscScalar *vals; 2508 for (int i = 0; i < nloc; i++) { 2509 PetscCall(MatGetRow(B, i, &nzl, NULL, NULL)); 2510 if (nzl > COL_BF_SIZE) { 2511 PetscCall(PetscFree(colbuf)); 2512 PetscCall(PetscInfo(pack, "Realloc buffer %" PetscInt_FMT " to %" PetscInt_FMT " (row size %" PetscInt_FMT ") \n", COL_BF_SIZE, 2 * COL_BF_SIZE, nzl)); 2513 COL_BF_SIZE = nzl; 2514 PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf)); 2515 } 2516 PetscCall(MatGetRow(B, i, &nzl, &cols, &vals)); 2517 for (int j = 0; j < nzl; j++) colbuf[j] = cols[j] + moffset; 2518 row = i + moffset; 2519 PetscCall(MatSetValues(packM, 1, &row, nzl, colbuf, vals, INSERT_VALUES)); 2520 PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals)); 2521 } 2522 } 2523 PetscCall(PetscFree(colbuf)); 2524 } 2525 // cleanup 2526 for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(MatDestroy(&subM[grid])); 2527 PetscCall(MatAssemblyBegin(packM, MAT_FINAL_ASSEMBLY)); 2528 PetscCall(MatAssemblyEnd(packM, MAT_FINAL_ASSEMBLY)); 2529 PetscCall(PetscObjectSetName((PetscObject)packM, "mass")); 2530 PetscCall(MatViewFromOptions(packM, NULL, "-dm_landau_mass_view")); 2531 ctx->M = packM; 2532 if (Amat) *Amat = packM; 2533 PetscCall(PetscLogEventEnd(ctx->events[14], 0, 0, 0, 0)); 2534 PetscFunctionReturn(PETSC_SUCCESS); 2535 } 2536 2537 /*@ 2538 DMPlexLandauIFunction - `TS` residual calculation, confusingly this computes the Jacobian w/o mass 2539 2540 Collective 2541 2542 Input Parameters: 2543 + ts - The time stepping context 2544 . time_dummy - current time (not used) 2545 . X - Current state 2546 . X_t - Time derivative of current state 2547 - actx - Landau context 2548 2549 Output Parameter: 2550 . F - The residual 2551 2552 Level: beginner 2553 2554 .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauIJacobian()` 2555 @*/ 2556 PetscErrorCode DMPlexLandauIFunction(TS ts, PetscReal time_dummy, Vec X, Vec X_t, Vec F, void *actx) 2557 { 2558 LandauCtx *ctx = (LandauCtx *)actx; 2559 PetscInt dim; 2560 DM pack; 2561 #if defined(PETSC_HAVE_THREADSAFETY) 2562 double starttime, endtime; 2563 #endif 2564 PetscObjectState state; 2565 2566 PetscFunctionBegin; 2567 PetscCall(TSGetDM(ts, &pack)); 2568 PetscCall(DMGetApplicationContext(pack, &ctx)); 2569 PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 2570 if (ctx->stage) PetscCall(PetscLogStagePush(ctx->stage)); 2571 PetscCall(PetscLogEventBegin(ctx->events[11], 0, 0, 0, 0)); 2572 PetscCall(PetscLogEventBegin(ctx->events[0], 0, 0, 0, 0)); 2573 #if defined(PETSC_HAVE_THREADSAFETY) 2574 starttime = MPI_Wtime(); 2575 #endif 2576 PetscCall(DMGetDimension(pack, &dim)); 2577 PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state)); 2578 if (state != ctx->norm_state) { 2579 PetscCall(PetscInfo(ts, "Create Landau Jacobian t=%g J.state %" PetscInt64_FMT " --> %" PetscInt64_FMT "\n", (double)time_dummy, ctx->norm_state, state)); 2580 PetscCall(MatZeroEntries(ctx->J)); 2581 PetscCall(LandauFormJacobian_Internal(X, ctx->J, dim, 0.0, (void *)ctx)); 2582 PetscCall(MatViewFromOptions(ctx->J, NULL, "-dm_landau_jacobian_view")); 2583 PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state)); 2584 ctx->norm_state = state; 2585 } else { 2586 PetscCall(PetscInfo(ts, "WARNING Skip forming Jacobian, has not changed %" PetscInt64_FMT "\n", state)); 2587 } 2588 /* mat vec for op */ 2589 PetscCall(MatMult(ctx->J, X, F)); /* C*f */ 2590 /* add time term */ 2591 if (X_t) PetscCall(MatMultAdd(ctx->M, X_t, F, F)); 2592 #if defined(PETSC_HAVE_THREADSAFETY) 2593 if (ctx->stage) { 2594 endtime = MPI_Wtime(); 2595 ctx->times[LANDAU_OPERATOR] += (endtime - starttime); 2596 ctx->times[LANDAU_JACOBIAN] += (endtime - starttime); 2597 ctx->times[LANDAU_MATRIX_TOTAL] += (endtime - starttime); 2598 ctx->times[LANDAU_JACOBIAN_COUNT] += 1; 2599 } 2600 #endif 2601 PetscCall(PetscLogEventEnd(ctx->events[0], 0, 0, 0, 0)); 2602 PetscCall(PetscLogEventEnd(ctx->events[11], 0, 0, 0, 0)); 2603 if (ctx->stage) PetscCall(PetscLogStagePop()); 2604 PetscFunctionReturn(PETSC_SUCCESS); 2605 } 2606 2607 /*@ 2608 DMPlexLandauIJacobian - `TS` Jacobian construction, confusingly this adds mass 2609 2610 Collective 2611 2612 Input Parameters: 2613 + ts - The time stepping context 2614 . time_dummy - current time (not used) 2615 . X - Current state 2616 . U_tdummy - Time derivative of current state (not used) 2617 . shift - shift for du/dt term 2618 - actx - Landau context 2619 2620 Output Parameters: 2621 + Amat - Jacobian 2622 - Pmat - same as Amat 2623 2624 Level: beginner 2625 2626 .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauIFunction()` 2627 @*/ 2628 PetscErrorCode DMPlexLandauIJacobian(TS ts, PetscReal time_dummy, Vec X, Vec U_tdummy, PetscReal shift, Mat Amat, Mat Pmat, void *actx) 2629 { 2630 LandauCtx *ctx = NULL; 2631 PetscInt dim; 2632 DM pack; 2633 #if defined(PETSC_HAVE_THREADSAFETY) 2634 double starttime, endtime; 2635 #endif 2636 PetscObjectState state; 2637 2638 PetscFunctionBegin; 2639 PetscCall(TSGetDM(ts, &pack)); 2640 PetscCall(DMGetApplicationContext(pack, &ctx)); 2641 PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 2642 PetscCheck(Amat == Pmat && Amat == ctx->J, ctx->comm, PETSC_ERR_PLIB, "Amat!=Pmat || Amat!=ctx->J"); 2643 PetscCall(DMGetDimension(pack, &dim)); 2644 /* get collision Jacobian into A */ 2645 if (ctx->stage) PetscCall(PetscLogStagePush(ctx->stage)); 2646 PetscCall(PetscLogEventBegin(ctx->events[11], 0, 0, 0, 0)); 2647 PetscCall(PetscLogEventBegin(ctx->events[9], 0, 0, 0, 0)); 2648 #if defined(PETSC_HAVE_THREADSAFETY) 2649 starttime = MPI_Wtime(); 2650 #endif 2651 PetscCall(PetscInfo(ts, "Adding mass to Jacobian t=%g, shift=%g\n", (double)time_dummy, (double)shift)); 2652 PetscCheck(shift != 0.0, ctx->comm, PETSC_ERR_PLIB, "zero shift"); 2653 PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state)); 2654 PetscCheck(state == ctx->norm_state, ctx->comm, PETSC_ERR_PLIB, "wrong state, %" PetscInt64_FMT " %" PetscInt64_FMT "", ctx->norm_state, state); 2655 if (!ctx->use_matrix_mass) { 2656 PetscCall(LandauFormJacobian_Internal(X, ctx->J, dim, shift, (void *)ctx)); 2657 } else { /* add mass */ 2658 PetscCall(MatAXPY(Pmat, shift, ctx->M, SAME_NONZERO_PATTERN)); 2659 } 2660 #if defined(PETSC_HAVE_THREADSAFETY) 2661 if (ctx->stage) { 2662 endtime = MPI_Wtime(); 2663 ctx->times[LANDAU_OPERATOR] += (endtime - starttime); 2664 ctx->times[LANDAU_MASS] += (endtime - starttime); 2665 ctx->times[LANDAU_MATRIX_TOTAL] += (endtime - starttime); 2666 } 2667 #endif 2668 PetscCall(PetscLogEventEnd(ctx->events[9], 0, 0, 0, 0)); 2669 PetscCall(PetscLogEventEnd(ctx->events[11], 0, 0, 0, 0)); 2670 if (ctx->stage) PetscCall(PetscLogStagePop()); 2671 PetscFunctionReturn(PETSC_SUCCESS); 2672 } 2673