1 // Copyright (c) 2017, Lawrence Livermore National Security, LLC. Produced at 2 // the Lawrence Livermore National Laboratory. LLNL-CODE-734707. All Rights 3 // reserved. See files LICENSE and NOTICE for details. 4 // 5 // This file is part of CEED, a collection of benchmarks, miniapps, software 6 // libraries and APIs for efficient high-order finite element and spectral 7 // element discretizations for exascale applications. For more information and 8 // source code availability see http://github.com/ceed. 9 // 10 // The CEED research is supported by the Exascale Computing Project 17-SC-20-SC, 11 // a collaborative effort of two U.S. Department of Energy organizations (Office 12 // of Science and the National Nuclear Security Administration) responsible for 13 // the planning and preparation of a capable exascale ecosystem, including 14 // software, applications, hardware, advanced system engineering and early 15 // testbed platforms, in support of the nation's exascale computing imperative. 16 17 // libCEED + PETSc Example: CEED BPs 18 // 19 // This example demonstrates a simple usage of libCEED with PETSc to solve the 20 // CEED BP benchmark problems, see http://ceed.exascaleproject.org/bps. 21 // 22 // The code is intentionally "raw", using only low-level communication 23 // primitives. 24 // 25 // Build with: 26 // 27 // make bpsraw [PETSC_DIR=</path/to/petsc>] [CEED_DIR=</path/to/libceed>] 28 // 29 // Sample runs: 30 // 31 // ./bpsraw -problem bp1 32 // ./bpsraw -problem bp2 -ceed /cpu/self 33 // ./bpsraw -problem bp3 -ceed /gpu/occa 34 // ./bpsraw -problem bp4 -ceed /cpu/occa 35 // ./bpsraw -problem bp5 -ceed /omp/occa 36 // ./bpsraw -problem bp6 -ceed /ocl/occa 37 // 38 //TESTARGS -ceed {ceed_resource} -test -problem bp2 -degree 5 -qextra 5 39 40 /// @file 41 /// CEED BPs example using PETSc 42 /// See bps.c for an implementation using DMPlex unstructured grids. 43 const char help[] = "Solve CEED BPs using PETSc\n"; 44 45 #include <stdbool.h> 46 #include <string.h> 47 #include <petscksp.h> 48 #include <ceed.h> 49 #include "qfunctions/bps/common.h" 50 #include "qfunctions/bps/bp1.h" 51 #include "qfunctions/bps/bp2.h" 52 #include "qfunctions/bps/bp3.h" 53 #include "qfunctions/bps/bp4.h" 54 55 #include <petscsys.h> 56 #if PETSC_VERSION_LT(3,12,0) 57 #ifdef PETSC_HAVE_CUDA 58 #include <petsccuda.h> 59 // Note: With PETSc prior to version 3.12.0, providing the source path to 60 // include 'cublas_v2.h' will be needed to use 'petsccuda.h'. 61 #endif 62 #endif 63 64 static void Split3(PetscInt size, PetscInt m[3], bool reverse) { 65 for (PetscInt d=0,sizeleft=size; d<3; d++) { 66 PetscInt try = (PetscInt)PetscCeilReal(PetscPowReal(sizeleft, 1./(3 - d))); 67 while (try * (sizeleft / try) != sizeleft) try++; 68 m[reverse ? 2-d : d] = try; 69 sizeleft /= try; 70 } 71 } 72 73 static PetscInt Max3(const PetscInt a[3]) { 74 return PetscMax(a[0], PetscMax(a[1], a[2])); 75 } 76 static PetscInt Min3(const PetscInt a[3]) { 77 return PetscMin(a[0], PetscMin(a[1], a[2])); 78 } 79 static void GlobalNodes(const PetscInt p[3], const PetscInt irank[3], 80 PetscInt degree, const PetscInt melem[3], 81 PetscInt mnodes[3]) { 82 for (int d=0; d<3; d++) 83 mnodes[d] = degree*melem[d] + (irank[d] == p[d]-1); 84 } 85 static PetscInt GlobalStart(const PetscInt p[3], const PetscInt irank[3], 86 PetscInt degree, const PetscInt melem[3]) { 87 PetscInt start = 0; 88 // Dumb brute-force is easier to read 89 for (PetscInt i=0; i<p[0]; i++) { 90 for (PetscInt j=0; j<p[1]; j++) { 91 for (PetscInt k=0; k<p[2]; k++) { 92 PetscInt mnodes[3], ijkrank[] = {i,j,k}; 93 if (i == irank[0] && j == irank[1] && k == irank[2]) return start; 94 GlobalNodes(p, ijkrank, degree, melem, mnodes); 95 start += mnodes[0] * mnodes[1] * mnodes[2]; 96 } 97 } 98 } 99 return -1; 100 } 101 static int CreateRestriction(Ceed ceed, CeedInterlaceMode imode, 102 const CeedInt melem[3], CeedInt P, CeedInt ncomp, 103 CeedElemRestriction *Erestrict) { 104 const PetscInt nelem = melem[0]*melem[1]*melem[2]; 105 PetscInt mnodes[3], *idx, *idxp; 106 107 // Get indicies 108 for (int d=0; d<3; d++) mnodes[d] = melem[d]*(P-1) + 1; 109 idxp = idx = malloc(nelem*P*P*P*sizeof idx[0]); 110 for (CeedInt i=0; i<melem[0]; i++) 111 for (CeedInt j=0; j<melem[1]; j++) 112 for (CeedInt k=0; k<melem[2]; k++,idxp += P*P*P) 113 for (CeedInt ii=0; ii<P; ii++) 114 for (CeedInt jj=0; jj<P; jj++) 115 for (CeedInt kk=0; kk<P; kk++) { 116 if (0) { // This is the C-style (i,j,k) ordering that I prefer 117 idxp[(ii*P+jj)*P+kk] = (((i*(P-1)+ii)*mnodes[1] 118 + (j*(P-1)+jj))*mnodes[2] 119 + (k*(P-1)+kk)); 120 } else { // (k,j,i) ordering for consistency with MFEM example 121 idxp[ii+P*(jj+P*kk)] = (((i*(P-1)+ii)*mnodes[1] 122 + (j*(P-1)+jj))*mnodes[2] 123 + (k*(P-1)+kk)); 124 } 125 } 126 127 // Setup CEED restriction 128 CeedElemRestrictionCreate(ceed, imode, nelem, P*P*P, 129 mnodes[0]*mnodes[1]*mnodes[2], ncomp, 130 CEED_MEM_HOST, CEED_OWN_POINTER, idx, Erestrict); 131 132 PetscFunctionReturn(0); 133 } 134 135 // Data for PETSc 136 typedef struct User_ *User; 137 struct User_ { 138 MPI_Comm comm; 139 VecScatter ltog; // Scatter for all entries 140 VecScatter ltog0; // Skip Dirichlet values 141 VecScatter gtogD; // global-to-global; only Dirichlet values 142 Vec Xloc, Yloc; 143 CeedVector xceed, yceed; 144 CeedOperator op; 145 CeedVector qdata; 146 Ceed ceed; 147 CeedMemType memtype; 148 int (*VecGetArray)(Vec, PetscScalar **); 149 int (*VecGetArrayRead)(Vec, const PetscScalar **); 150 int (*VecRestoreArray)(Vec, PetscScalar **); 151 int (*VecRestoreArrayRead)(Vec, const PetscScalar **); 152 }; 153 154 // MemType Options 155 static const char *const memTypes[] = {"host","device","memType", 156 "CEED_MEM_",0 157 }; 158 159 // BP Options 160 typedef enum { 161 CEED_BP1 = 0, CEED_BP2 = 1, CEED_BP3 = 2, 162 CEED_BP4 = 3, CEED_BP5 = 4, CEED_BP6 = 5 163 } bpType; 164 static const char *const bpTypes[] = {"bp1","bp2","bp3","bp4","bp5","bp6", 165 "bpType","CEED_BP",0 166 }; 167 168 // BP specific data 169 typedef struct { 170 CeedInt ncompu, qdatasize, qextra; 171 CeedQFunctionUser setupgeo, setuprhs, apply, error; 172 const char *setupgeofname, *setuprhsfname, *applyfname, *errorfname; 173 CeedEvalMode inmode, outmode; 174 CeedQuadMode qmode; 175 } bpData; 176 177 bpData bpOptions[6] = { 178 [CEED_BP1] = { 179 .ncompu = 1, 180 .qdatasize = 1, 181 .qextra = 1, 182 .setupgeo = SetupMassGeo, 183 .setuprhs = SetupMassRhs, 184 .apply = Mass, 185 .error = Error, 186 .setupgeofname = SetupMassGeo_loc, 187 .setuprhsfname = SetupMassRhs_loc, 188 .applyfname = Mass_loc, 189 .errorfname = Error_loc, 190 .inmode = CEED_EVAL_INTERP, 191 .outmode = CEED_EVAL_INTERP, 192 .qmode = CEED_GAUSS 193 }, 194 [CEED_BP2] = { 195 .ncompu = 3, 196 .qdatasize = 1, 197 .qextra = 1, 198 .setupgeo = SetupMassGeo, 199 .setuprhs = SetupMassRhs3, 200 .apply = Mass3, 201 .error = Error3, 202 .setupgeofname = SetupMassGeo_loc, 203 .setuprhsfname = SetupMassRhs3_loc, 204 .applyfname = Mass3_loc, 205 .errorfname = Error3_loc, 206 .inmode = CEED_EVAL_INTERP, 207 .outmode = CEED_EVAL_INTERP, 208 .qmode = CEED_GAUSS 209 }, 210 [CEED_BP3] = { 211 .ncompu = 1, 212 .qdatasize = 6, 213 .qextra = 1, 214 .setupgeo = SetupDiffGeo, 215 .setuprhs = SetupDiffRhs, 216 .apply = Diff, 217 .error = Error, 218 .setupgeofname = SetupDiffGeo_loc, 219 .setuprhsfname = SetupDiffRhs_loc, 220 .applyfname = Diff_loc, 221 .errorfname = Error_loc, 222 .inmode = CEED_EVAL_GRAD, 223 .outmode = CEED_EVAL_GRAD, 224 .qmode = CEED_GAUSS 225 }, 226 [CEED_BP4] = { 227 .ncompu = 3, 228 .qdatasize = 6, 229 .qextra = 1, 230 .setupgeo = SetupDiffGeo, 231 .setuprhs = SetupDiffRhs3, 232 .apply = Diff3, 233 .error = Error3, 234 .setupgeofname = SetupDiffGeo_loc, 235 .setuprhsfname = SetupDiffRhs3_loc, 236 .applyfname = Diff_loc, 237 .errorfname = Error3_loc, 238 .inmode = CEED_EVAL_GRAD, 239 .outmode = CEED_EVAL_GRAD, 240 .qmode = CEED_GAUSS 241 }, 242 [CEED_BP5] = { 243 .ncompu = 1, 244 .qdatasize = 6, 245 .qextra = 0, 246 .setupgeo = SetupDiffGeo, 247 .setuprhs = SetupDiffRhs, 248 .apply = Diff, 249 .error = Error, 250 .setupgeofname = SetupDiffGeo_loc, 251 .setuprhsfname = SetupDiffRhs_loc, 252 .applyfname = Diff_loc, 253 .errorfname = Error_loc, 254 .inmode = CEED_EVAL_GRAD, 255 .outmode = CEED_EVAL_GRAD, 256 .qmode = CEED_GAUSS_LOBATTO 257 }, 258 [CEED_BP6] = { 259 .ncompu = 3, 260 .qdatasize = 6, 261 .qextra = 0, 262 .setupgeo = SetupDiffGeo, 263 .setuprhs = SetupDiffRhs3, 264 .apply = Diff3, 265 .error = Error3, 266 .setupgeofname = SetupDiffGeo_loc, 267 .setuprhsfname = SetupDiffRhs3_loc, 268 .applyfname = Diff_loc, 269 .errorfname = Error3_loc, 270 .inmode = CEED_EVAL_GRAD, 271 .outmode = CEED_EVAL_GRAD, 272 .qmode = CEED_GAUSS_LOBATTO 273 } 274 }; 275 276 // This function uses libCEED to compute the action of the mass matrix 277 static PetscErrorCode MatMult_Mass(Mat A, Vec X, Vec Y) { 278 PetscErrorCode ierr; 279 User user; 280 PetscScalar *x, *y; 281 282 PetscFunctionBeginUser; 283 284 ierr = MatShellGetContext(A, &user); CHKERRQ(ierr); 285 286 // Global-to-local 287 ierr = VecScatterBegin(user->ltog, X, user->Xloc, INSERT_VALUES, 288 SCATTER_REVERSE); CHKERRQ(ierr); 289 ierr = VecScatterEnd(user->ltog, X, user->Xloc, INSERT_VALUES, 290 SCATTER_REVERSE); CHKERRQ(ierr); 291 ierr = VecZeroEntries(user->Yloc); CHKERRQ(ierr); 292 293 // Setup libCEED vectors 294 ierr = user->VecGetArrayRead(user->Xloc, (const PetscScalar **)&x); 295 CHKERRQ(ierr); 296 ierr = user->VecGetArray(user->Yloc, &y); CHKERRQ(ierr); 297 CeedVectorSetArray(user->xceed, user->memtype, CEED_USE_POINTER, x); 298 CeedVectorSetArray(user->yceed, user->memtype, CEED_USE_POINTER, y); 299 300 // Apply libCEED operator 301 CeedOperatorApply(user->op, user->xceed, user->yceed, 302 CEED_REQUEST_IMMEDIATE); 303 CeedVectorSyncArray(user->yceed, user->memtype); 304 305 // Restore PETSc vectors 306 ierr = user->VecRestoreArrayRead(user->Xloc, (const PetscScalar **)&x); 307 CHKERRQ(ierr); 308 ierr = user->VecRestoreArray(user->Yloc, &y); CHKERRQ(ierr); 309 310 // Local-to-global 311 if (Y) { 312 ierr = VecZeroEntries(Y); CHKERRQ(ierr); 313 ierr = VecScatterBegin(user->ltog, user->Yloc, Y, ADD_VALUES, 314 SCATTER_FORWARD); CHKERRQ(ierr); 315 ierr = VecScatterEnd(user->ltog, user->Yloc, Y, ADD_VALUES, 316 SCATTER_FORWARD); CHKERRQ(ierr); 317 } 318 PetscFunctionReturn(0); 319 } 320 321 // This function uses libCEED to compute the action of the Laplacian with 322 // Dirichlet boundary conditions 323 static PetscErrorCode MatMult_Diff(Mat A, Vec X, Vec Y) { 324 PetscErrorCode ierr; 325 User user; 326 PetscScalar *x, *y; 327 328 PetscFunctionBeginUser; 329 330 ierr = MatShellGetContext(A, &user); CHKERRQ(ierr); 331 332 // Global-to-local 333 ierr = VecScatterBegin(user->ltog0, X, user->Xloc, INSERT_VALUES, 334 SCATTER_REVERSE); CHKERRQ(ierr); 335 ierr = VecScatterEnd(user->ltog0, X, user->Xloc, INSERT_VALUES, 336 SCATTER_REVERSE); 337 CHKERRQ(ierr); 338 ierr = VecZeroEntries(user->Yloc); CHKERRQ(ierr); 339 340 // Setup libCEED vectors 341 ierr = user->VecGetArrayRead(user->Xloc, (const PetscScalar **)&x); 342 CHKERRQ(ierr); 343 ierr = user->VecGetArray(user->Yloc, &y); CHKERRQ(ierr); 344 CeedVectorSetArray(user->xceed, user->memtype, CEED_USE_POINTER, x); 345 CeedVectorSetArray(user->yceed, user->memtype, CEED_USE_POINTER, y); 346 347 // Apply libCEED operator 348 CeedOperatorApply(user->op, user->xceed, user->yceed, 349 CEED_REQUEST_IMMEDIATE); 350 CeedVectorSyncArray(user->yceed, user->memtype); 351 352 // Restore PETSc vectors 353 ierr = user->VecRestoreArrayRead(user->Xloc, (const PetscScalar **)&x); 354 CHKERRQ(ierr); 355 ierr = user->VecRestoreArray(user->Yloc, &y); CHKERRQ(ierr); 356 357 // Local-to-global 358 ierr = VecZeroEntries(Y); CHKERRQ(ierr); 359 ierr = VecScatterBegin(user->gtogD, X, Y, INSERT_VALUES, SCATTER_FORWARD); 360 CHKERRQ(ierr); 361 ierr = VecScatterEnd(user->gtogD, X, Y, INSERT_VALUES, SCATTER_FORWARD); 362 CHKERRQ(ierr); 363 ierr = VecScatterBegin(user->ltog0, user->Yloc, Y, ADD_VALUES, 364 SCATTER_FORWARD); CHKERRQ(ierr); 365 ierr = VecScatterEnd(user->ltog0, user->Yloc, Y, ADD_VALUES, SCATTER_FORWARD); 366 CHKERRQ(ierr); 367 368 PetscFunctionReturn(0); 369 } 370 371 // This function calculates the error in the final solution 372 static PetscErrorCode ComputeErrorMax(User user, CeedOperator operror, Vec X, 373 CeedVector target, PetscReal *maxerror) { 374 PetscErrorCode ierr; 375 PetscScalar *x; 376 CeedVector collocated_error; 377 CeedInt length; 378 379 PetscFunctionBeginUser; 380 381 CeedVectorGetLength(target, &length); 382 CeedVectorCreate(user->ceed, length, &collocated_error); 383 384 // Global-to-local 385 ierr = VecScatterBegin(user->ltog, X, user->Xloc, INSERT_VALUES, 386 SCATTER_REVERSE); CHKERRQ(ierr); 387 ierr = VecScatterEnd(user->ltog, X, user->Xloc, INSERT_VALUES, 388 SCATTER_REVERSE); CHKERRQ(ierr); 389 390 // Setup libCEED vector 391 ierr = user->VecGetArrayRead(user->Xloc, (const PetscScalar **)&x); 392 CHKERRQ(ierr); 393 CeedVectorSetArray(user->xceed, user->memtype, CEED_USE_POINTER, x); 394 395 // Apply libCEED operator 396 CeedOperatorApply(operror, user->xceed, collocated_error, 397 CEED_REQUEST_IMMEDIATE); 398 399 // Restore PETSc vector 400 ierr = user->VecRestoreArrayRead(user->Xloc, (const PetscScalar **)&x); 401 CHKERRQ(ierr); 402 403 // Reduce max error 404 *maxerror = 0; 405 const CeedScalar *e; 406 CeedVectorGetArrayRead(collocated_error, CEED_MEM_HOST, &e); 407 for (CeedInt i=0; i<length; i++) { 408 *maxerror = PetscMax(*maxerror, PetscAbsScalar(e[i])); 409 } 410 CeedVectorRestoreArrayRead(collocated_error, &e); 411 ierr = MPI_Allreduce(MPI_IN_PLACE, maxerror, 1, MPIU_REAL, MPIU_MAX, 412 user->comm); CHKERRQ(ierr); 413 414 // Cleanup 415 CeedVectorDestroy(&collocated_error); 416 417 PetscFunctionReturn(0); 418 } 419 420 int main(int argc, char **argv) { 421 PetscInt ierr; 422 MPI_Comm comm; 423 char ceedresource[PETSC_MAX_PATH_LEN] = "/cpu/self"; 424 double my_rt_start, my_rt, rt_min, rt_max; 425 PetscInt degree, qextra, localnodes, localelem, melem[3], mnodes[3], p[3], 426 irank[3], lnodes[3], lsize, ncompu = 1; 427 PetscScalar *r; 428 PetscBool test_mode, benchmark_mode, write_solution; 429 PetscMPIInt size, rank; 430 PetscLogStage solvestage; 431 Vec X, Xloc, rhs, rhsloc; 432 Mat mat; 433 KSP ksp; 434 VecScatter ltog, ltog0, gtogD; 435 User user; 436 Ceed ceed; 437 CeedBasis basisx, basisu; 438 CeedElemRestriction Erestrictx, Erestrictu, Erestrictui, Erestrictqdi; 439 CeedQFunction qfsetupgeo, qfsetuprhs, qfapply, qferror; 440 CeedOperator opsetupgeo, opsetuprhs, opapply, operror; 441 CeedVector xcoord, qdata, rhsceed, target; 442 CeedMemType memtyperequested; 443 CeedInt P, Q; 444 const CeedInt dim = 3, ncompx = 3; 445 bpType bpchoice; 446 447 // Check for PETSc CUDA availability 448 PetscBool petschavecuda, setmemtyperequest = PETSC_FALSE; 449 // *INDENT-OFF* 450 #ifdef PETSC_HAVE_CUDA 451 petschavecuda = PETSC_TRUE; 452 #else 453 petschavecuda = PETSC_FALSE; 454 #endif 455 // *INDENT-ON* 456 457 ierr = PetscInitialize(&argc, &argv, NULL, help); 458 if (ierr) return ierr; 459 comm = PETSC_COMM_WORLD; 460 461 // Read command line options 462 ierr = PetscOptionsBegin(comm, NULL, "CEED BPs in PETSc", NULL); CHKERRQ(ierr); 463 bpchoice = CEED_BP1; 464 ierr = PetscOptionsEnum("-problem", 465 "CEED benchmark problem to solve", NULL, 466 bpTypes, (PetscEnum)bpchoice, (PetscEnum *)&bpchoice, 467 NULL); CHKERRQ(ierr); 468 ncompu = bpOptions[bpchoice].ncompu; 469 test_mode = PETSC_FALSE; 470 ierr = PetscOptionsBool("-test", 471 "Testing mode (do not print unless error is large)", 472 NULL, test_mode, &test_mode, NULL); CHKERRQ(ierr); 473 benchmark_mode = PETSC_FALSE; 474 ierr = PetscOptionsBool("-benchmark", 475 "Benchmarking mode (prints benchmark statistics)", 476 NULL, benchmark_mode, &benchmark_mode, NULL); 477 CHKERRQ(ierr); 478 write_solution = PETSC_FALSE; 479 ierr = PetscOptionsBool("-write_solution", 480 "Write solution for visualization", 481 NULL, write_solution, &write_solution, NULL); 482 CHKERRQ(ierr); 483 degree = test_mode ? 3 : 1; 484 ierr = PetscOptionsInt("-degree", "Polynomial degree of tensor product basis", 485 NULL, degree, °ree, NULL); CHKERRQ(ierr); 486 qextra = bpOptions[bpchoice].qextra; 487 ierr = PetscOptionsInt("-qextra", "Number of extra quadrature points", 488 NULL, qextra, &qextra, NULL); CHKERRQ(ierr); 489 ierr = PetscOptionsString("-ceed", "CEED resource specifier", 490 NULL, ceedresource, ceedresource, 491 sizeof(ceedresource), NULL); CHKERRQ(ierr); 492 localnodes = 1000; 493 ierr = PetscOptionsInt("-local", 494 "Target number of locally owned nodes per process", 495 NULL, localnodes, &localnodes, NULL); CHKERRQ(ierr); 496 memtyperequested = petschavecuda ? CEED_MEM_DEVICE : CEED_MEM_HOST; 497 ierr = PetscOptionsEnum("-memtype", 498 "CEED MemType requested", NULL, 499 memTypes, (PetscEnum)memtyperequested, 500 (PetscEnum *)&memtyperequested, &setmemtyperequest); 501 CHKERRQ(ierr); 502 ierr = PetscOptionsEnd(); CHKERRQ(ierr); 503 P = degree + 1; 504 Q = P + qextra; 505 506 // Set up libCEED 507 CeedInit(ceedresource, &ceed); 508 CeedMemType memtypebackend; 509 CeedGetPreferredMemType(ceed, &memtypebackend); 510 511 // Check memtype compatibility 512 if (!setmemtyperequest) 513 memtyperequested = memtypebackend; 514 else if (!petschavecuda && memtyperequested == CEED_MEM_DEVICE) 515 SETERRQ1(PETSC_COMM_WORLD, PETSC_ERR_SUP_SYS, 516 "PETSc was not built with CUDA. " 517 "Requested MemType CEED_MEM_DEVICE is not supported.", NULL); 518 519 // Determine size of process grid 520 ierr = MPI_Comm_size(comm, &size); CHKERRQ(ierr); 521 Split3(size, p, false); 522 523 // Find a nicely composite number of elements no less than localnodes 524 for (localelem = PetscMax(1, localnodes / (degree*degree*degree)); ; 525 localelem++) { 526 Split3(localelem, melem, true); 527 if (Max3(melem) / Min3(melem) <= 2) break; 528 } 529 530 // Find my location in the process grid 531 ierr = MPI_Comm_rank(comm, &rank); CHKERRQ(ierr); 532 for (int d=0,rankleft=rank; d<dim; d++) { 533 const int pstride[3] = {p[1] *p[2], p[2], 1}; 534 irank[d] = rankleft / pstride[d]; 535 rankleft -= irank[d] * pstride[d]; 536 } 537 538 GlobalNodes(p, irank, degree, melem, mnodes); 539 540 // Setup global vector 541 ierr = VecCreate(comm, &X); CHKERRQ(ierr); 542 if (memtyperequested == CEED_MEM_DEVICE) { 543 ierr = VecSetType(X, VECCUDA); CHKERRQ(ierr); 544 } 545 ierr = VecSetSizes(X, mnodes[0]*mnodes[1]*mnodes[2]*ncompu, PETSC_DECIDE); 546 CHKERRQ(ierr); 547 ierr = VecSetUp(X); CHKERRQ(ierr); 548 549 // Set up libCEED 550 CeedInit(ceedresource, &ceed); 551 552 // Print summary 553 CeedInt gsize; 554 ierr = VecGetSize(X, &gsize); CHKERRQ(ierr); 555 if (!test_mode) { 556 const char *usedresource; 557 CeedGetResource(ceed, &usedresource); 558 559 VecType vectype; 560 ierr = VecGetType(X, &vectype); CHKERRQ(ierr); 561 562 ierr = PetscPrintf(comm, 563 "\n-- CEED Benchmark Problem %d -- libCEED + PETSc --\n" 564 " PETSc:\n" 565 " PETSc Vec Type : %s\n" 566 " libCEED:\n" 567 " libCEED Backend : %s\n" 568 " libCEED Backend MemType : %s\n" 569 " libCEED User Requested MemType : %s\n" 570 " Mesh:\n" 571 " Number of 1D Basis Nodes (p) : %d\n" 572 " Number of 1D Quadrature Points (q) : %d\n" 573 " Global nodes : %D\n" 574 " Process Decomposition : %D %D %D\n" 575 " Local Elements : %D = %D %D %D\n" 576 " Owned nodes : %D = %D %D %D\n" 577 " DoF per node : %D\n", 578 bpchoice+1, vectype, usedresource, 579 CeedMemTypes[memtypebackend], 580 (setmemtyperequest) ? 581 CeedMemTypes[memtyperequested] : "none", 582 P, Q, gsize/ncompu, p[0], p[1], p[2], localelem, 583 melem[0], melem[1], melem[2], 584 mnodes[0]*mnodes[1]*mnodes[2], mnodes[0], mnodes[1], 585 mnodes[2], ncompu); CHKERRQ(ierr); 586 } 587 588 { 589 lsize = 1; 590 for (int d=0; d<dim; d++) { 591 lnodes[d] = melem[d]*degree + 1; 592 lsize *= lnodes[d]; 593 } 594 ierr = VecCreate(PETSC_COMM_SELF, &Xloc); CHKERRQ(ierr); 595 if (memtyperequested == CEED_MEM_DEVICE) { 596 ierr = VecSetType(Xloc, VECCUDA); CHKERRQ(ierr); 597 } 598 ierr = VecSetSizes(Xloc, lsize*ncompu, PETSC_DECIDE); CHKERRQ(ierr); 599 ierr = VecSetUp(Xloc); CHKERRQ(ierr); 600 601 // Create local-to-global scatter 602 PetscInt *ltogind, *ltogind0, *locind, l0count; 603 IS ltogis, ltogis0, locis; 604 PetscInt gstart[2][2][2], gmnodes[2][2][2][dim]; 605 606 for (int i=0; i<2; i++) { 607 for (int j=0; j<2; j++) { 608 for (int k=0; k<2; k++) { 609 PetscInt ijkrank[3] = {irank[0]+i, irank[1]+j, irank[2]+k}; 610 gstart[i][j][k] = GlobalStart(p, ijkrank, degree, melem); 611 GlobalNodes(p, ijkrank, degree, melem, gmnodes[i][j][k]); 612 } 613 } 614 } 615 616 ierr = PetscMalloc1(lsize, <ogind); CHKERRQ(ierr); 617 ierr = PetscMalloc1(lsize, <ogind0); CHKERRQ(ierr); 618 ierr = PetscMalloc1(lsize, &locind); CHKERRQ(ierr); 619 l0count = 0; 620 for (PetscInt i=0,ir,ii; ir=i>=mnodes[0], ii=i-ir*mnodes[0], i<lnodes[0]; i++) 621 for (PetscInt j=0,jr,jj; jr=j>=mnodes[1], jj=j-jr*mnodes[1], j<lnodes[1]; j++) 622 for (PetscInt k=0,kr,kk; kr=k>=mnodes[2], kk=k-kr*mnodes[2], k<lnodes[2]; k++) { 623 PetscInt here = (i*lnodes[1]+j)*lnodes[2]+k; 624 ltogind[here] = 625 gstart[ir][jr][kr] + (ii*gmnodes[ir][jr][kr][1]+jj)*gmnodes[ir][jr][kr][2]+kk; 626 if ((irank[0] == 0 && i == 0) 627 || (irank[1] == 0 && j == 0) 628 || (irank[2] == 0 && k == 0) 629 || (irank[0]+1 == p[0] && i+1 == lnodes[0]) 630 || (irank[1]+1 == p[1] && j+1 == lnodes[1]) 631 || (irank[2]+1 == p[2] && k+1 == lnodes[2])) 632 continue; 633 ltogind0[l0count] = ltogind[here]; 634 locind[l0count++] = here; 635 } 636 ierr = ISCreateBlock(comm, ncompu, lsize, ltogind, PETSC_OWN_POINTER, 637 <ogis); CHKERRQ(ierr); 638 ierr = VecScatterCreate(Xloc, NULL, X, ltogis, <og); CHKERRQ(ierr); 639 CHKERRQ(ierr); 640 ierr = ISCreateBlock(comm, ncompu, l0count, ltogind0, PETSC_OWN_POINTER, 641 <ogis0); CHKERRQ(ierr); 642 ierr = ISCreateBlock(comm, ncompu, l0count, locind, PETSC_OWN_POINTER, 643 &locis); CHKERRQ(ierr); 644 ierr = VecScatterCreate(Xloc, locis, X, ltogis0, <og0); CHKERRQ(ierr); 645 { 646 // Create global-to-global scatter for Dirichlet values (everything not in 647 // ltogis0, which is the range of ltog0) 648 PetscInt xstart, xend, *indD, countD = 0; 649 IS isD; 650 const PetscScalar *x; 651 ierr = VecZeroEntries(Xloc); CHKERRQ(ierr); 652 ierr = VecSet(X, 1.0); CHKERRQ(ierr); 653 ierr = VecScatterBegin(ltog0, Xloc, X, INSERT_VALUES, SCATTER_FORWARD); 654 CHKERRQ(ierr); 655 ierr = VecScatterEnd(ltog0, Xloc, X, INSERT_VALUES, SCATTER_FORWARD); 656 CHKERRQ(ierr); 657 ierr = VecGetOwnershipRange(X, &xstart, &xend); CHKERRQ(ierr); 658 ierr = PetscMalloc1(xend-xstart, &indD); CHKERRQ(ierr); 659 ierr = VecGetArrayRead(X, &x); CHKERRQ(ierr); 660 for (PetscInt i=0; i<xend-xstart; i++) { 661 if (x[i] == 1.) indD[countD++] = xstart + i; 662 } 663 ierr = VecRestoreArrayRead(X, &x); CHKERRQ(ierr); 664 ierr = ISCreateGeneral(comm, countD, indD, PETSC_COPY_VALUES, &isD); 665 CHKERRQ(ierr); 666 ierr = PetscFree(indD); CHKERRQ(ierr); 667 ierr = VecScatterCreate(X, isD, X, isD, >ogD); CHKERRQ(ierr); 668 ierr = ISDestroy(&isD); CHKERRQ(ierr); 669 } 670 ierr = ISDestroy(<ogis); CHKERRQ(ierr); 671 ierr = ISDestroy(<ogis0); CHKERRQ(ierr); 672 ierr = ISDestroy(&locis); CHKERRQ(ierr); 673 } 674 675 // CEED bases 676 CeedBasisCreateTensorH1Lagrange(ceed, dim, ncompu, P, Q, 677 bpOptions[bpchoice].qmode, &basisu); 678 CeedBasisCreateTensorH1Lagrange(ceed, dim, ncompx, 2, Q, 679 bpOptions[bpchoice].qmode, &basisx); 680 681 // CEED restrictions 682 CreateRestriction(ceed, CEED_INTERLACED, melem, P, ncompu, &Erestrictu); 683 CreateRestriction(ceed, CEED_NONINTERLACED, melem, 2, dim, &Erestrictx); 684 CeedInt nelem = melem[0]*melem[1]*melem[2]; 685 CeedElemRestrictionCreateStrided(ceed, nelem, Q*Q*Q, nelem*Q*Q*Q, ncompu, 686 CEED_STRIDES_BACKEND, &Erestrictui); 687 CeedElemRestrictionCreateStrided(ceed, nelem, Q*Q*Q, nelem*Q*Q*Q, 688 bpOptions[bpchoice].qdatasize, 689 CEED_STRIDES_BACKEND, &Erestrictqdi); 690 { 691 CeedScalar *xloc; 692 CeedInt shape[3] = {melem[0]+1, melem[1]+1, melem[2]+1}, len = 693 shape[0]*shape[1]*shape[2]; 694 xloc = malloc(len*ncompx*sizeof xloc[0]); 695 for (CeedInt i=0; i<shape[0]; i++) { 696 for (CeedInt j=0; j<shape[1]; j++) { 697 for (CeedInt k=0; k<shape[2]; k++) { 698 xloc[((i*shape[1]+j)*shape[2]+k) + 0*len] = 1.*(irank[0]*melem[0]+i) / 699 (p[0]*melem[0]); 700 xloc[((i*shape[1]+j)*shape[2]+k) + 1*len] = 1.*(irank[1]*melem[1]+j) / 701 (p[1]*melem[1]); 702 xloc[((i*shape[1]+j)*shape[2]+k) + 2*len] = 1.*(irank[2]*melem[2]+k) / 703 (p[2]*melem[2]); 704 } 705 } 706 } 707 CeedVectorCreate(ceed, len*ncompx, &xcoord); 708 CeedVectorSetArray(xcoord, CEED_MEM_HOST, CEED_OWN_POINTER, xloc); 709 } 710 711 // Create the Qfunction that builds the operator quadrature data 712 CeedQFunctionCreateInterior(ceed, 1, bpOptions[bpchoice].setupgeo, 713 bpOptions[bpchoice].setupgeofname, &qfsetupgeo); 714 CeedQFunctionAddInput(qfsetupgeo, "dx", ncompx*dim, CEED_EVAL_GRAD); 715 CeedQFunctionAddInput(qfsetupgeo, "weight", 1, CEED_EVAL_WEIGHT); 716 CeedQFunctionAddOutput(qfsetupgeo, "qdata", bpOptions[bpchoice].qdatasize, 717 CEED_EVAL_NONE); 718 719 // Create the Qfunction that sets up the RHS and true solution 720 CeedQFunctionCreateInterior(ceed, 1, bpOptions[bpchoice].setuprhs, 721 bpOptions[bpchoice].setuprhsfname, &qfsetuprhs); 722 CeedQFunctionAddInput(qfsetuprhs, "x", ncompx, CEED_EVAL_INTERP); 723 CeedQFunctionAddInput(qfsetuprhs, "dx", ncompx*dim, CEED_EVAL_GRAD); 724 CeedQFunctionAddInput(qfsetuprhs, "weight", 1, CEED_EVAL_WEIGHT); 725 CeedQFunctionAddOutput(qfsetuprhs, "true_soln", ncompu, CEED_EVAL_NONE); 726 CeedQFunctionAddOutput(qfsetuprhs, "rhs", ncompu, CEED_EVAL_INTERP); 727 728 // Set up PDE operator 729 CeedQFunctionCreateInterior(ceed, 1, bpOptions[bpchoice].apply, 730 bpOptions[bpchoice].applyfname, &qfapply); 731 // Add inputs and outputs 732 CeedInt inscale = bpOptions[bpchoice].inmode==CEED_EVAL_GRAD ? 3 : 1; 733 CeedInt outscale = bpOptions[bpchoice].outmode==CEED_EVAL_GRAD ? 3 : 1; 734 CeedQFunctionAddInput(qfapply, "u", ncompu*inscale, 735 bpOptions[bpchoice].inmode); 736 CeedQFunctionAddInput(qfapply, "qdata", bpOptions[bpchoice].qdatasize, 737 CEED_EVAL_NONE); 738 CeedQFunctionAddOutput(qfapply, "v", ncompu*outscale, 739 bpOptions[bpchoice].outmode); 740 741 // Create the error qfunction 742 CeedQFunctionCreateInterior(ceed, 1, bpOptions[bpchoice].error, 743 bpOptions[bpchoice].errorfname, &qferror); 744 CeedQFunctionAddInput(qferror, "u", ncompu, CEED_EVAL_INTERP); 745 CeedQFunctionAddInput(qferror, "true_soln", ncompu, CEED_EVAL_NONE); 746 CeedQFunctionAddOutput(qferror, "error", ncompu, CEED_EVAL_NONE); 747 748 // Create the persistent vectors that will be needed in setup 749 CeedInt nqpts; 750 CeedBasisGetNumQuadraturePoints(basisu, &nqpts); 751 CeedVectorCreate(ceed, bpOptions[bpchoice].qdatasize*nelem*nqpts, &qdata); 752 CeedVectorCreate(ceed, nelem*nqpts*ncompu, &target); 753 CeedVectorCreate(ceed, lsize*ncompu, &rhsceed); 754 755 // Create the operator that builds the quadrature data for the ceed operator 756 CeedOperatorCreate(ceed, qfsetupgeo, CEED_QFUNCTION_NONE, 757 CEED_QFUNCTION_NONE, &opsetupgeo); 758 CeedOperatorSetField(opsetupgeo, "dx", Erestrictx, basisx, 759 CEED_VECTOR_ACTIVE); 760 CeedOperatorSetField(opsetupgeo, "weight", CEED_ELEMRESTRICTION_NONE, basisx, 761 CEED_VECTOR_NONE); 762 CeedOperatorSetField(opsetupgeo, "qdata", Erestrictqdi, 763 CEED_BASIS_COLLOCATED, CEED_VECTOR_ACTIVE); 764 765 // Create the operator that builds the RHS and true solution 766 CeedOperatorCreate(ceed, qfsetuprhs, CEED_QFUNCTION_NONE, 767 CEED_QFUNCTION_NONE, &opsetuprhs); 768 CeedOperatorSetField(opsetuprhs, "x", Erestrictx, basisx, 769 CEED_VECTOR_ACTIVE); 770 CeedOperatorSetField(opsetuprhs, "dx", Erestrictx, basisx, 771 CEED_VECTOR_ACTIVE); 772 CeedOperatorSetField(opsetuprhs, "weight", CEED_ELEMRESTRICTION_NONE, basisx, 773 CEED_VECTOR_NONE); 774 CeedOperatorSetField(opsetuprhs, "true_soln", Erestrictui, 775 CEED_BASIS_COLLOCATED, target); 776 CeedOperatorSetField(opsetuprhs, "rhs", Erestrictu, basisu, 777 CEED_VECTOR_ACTIVE); 778 779 // Create the mass or diff operator 780 CeedOperatorCreate(ceed, qfapply, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, 781 &opapply); 782 CeedOperatorSetField(opapply, "u", Erestrictu, basisu, CEED_VECTOR_ACTIVE); 783 CeedOperatorSetField(opapply, "qdata", Erestrictqdi, CEED_BASIS_COLLOCATED, 784 qdata); 785 CeedOperatorSetField(opapply, "v", Erestrictu, basisu, CEED_VECTOR_ACTIVE); 786 787 // Create the error operator 788 CeedOperatorCreate(ceed, qferror, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, 789 &operror); 790 CeedOperatorSetField(operror, "u", Erestrictu, basisu, CEED_VECTOR_ACTIVE); 791 CeedOperatorSetField(operror, "true_soln", Erestrictui, 792 CEED_BASIS_COLLOCATED, target); 793 CeedOperatorSetField(operror, "error", Erestrictui, CEED_BASIS_COLLOCATED, 794 CEED_VECTOR_ACTIVE); 795 796 // Set up Mat 797 ierr = PetscMalloc1(1, &user); CHKERRQ(ierr); 798 user->comm = comm; 799 user->ltog = ltog; 800 if (bpchoice != CEED_BP1 && bpchoice != CEED_BP2) { 801 user->ltog0 = ltog0; 802 user->gtogD = gtogD; 803 } 804 user->Xloc = Xloc; 805 ierr = VecDuplicate(Xloc, &user->Yloc); CHKERRQ(ierr); 806 CeedVectorCreate(ceed, lsize*ncompu, &user->xceed); 807 CeedVectorCreate(ceed, lsize*ncompu, &user->yceed); 808 user->op = opapply; 809 user->qdata = qdata; 810 user->ceed = ceed; 811 user->memtype = memtyperequested; 812 if (memtyperequested == CEED_MEM_HOST) { 813 user->VecGetArray = VecGetArray; 814 user->VecGetArrayRead = VecGetArrayRead; 815 user->VecRestoreArray = VecRestoreArray; 816 user->VecRestoreArrayRead = VecRestoreArrayRead; 817 } else { 818 user->VecGetArray = VecCUDAGetArray; 819 user->VecGetArrayRead = VecCUDAGetArrayRead; 820 user->VecRestoreArray = VecCUDARestoreArray; 821 user->VecRestoreArrayRead = VecCUDARestoreArrayRead; 822 } 823 824 ierr = MatCreateShell(comm, mnodes[0]*mnodes[1]*mnodes[2]*ncompu, 825 mnodes[0]*mnodes[1]*mnodes[2]*ncompu, 826 PETSC_DECIDE, PETSC_DECIDE, user, &mat); CHKERRQ(ierr); 827 if (bpchoice == CEED_BP1 || bpchoice == CEED_BP2) { 828 ierr = MatShellSetOperation(mat, MATOP_MULT, (void(*)(void))MatMult_Mass); 829 CHKERRQ(ierr); 830 } else { 831 ierr = MatShellSetOperation(mat, MATOP_MULT, (void(*)(void))MatMult_Diff); 832 CHKERRQ(ierr); 833 } 834 if (user->memtype == CEED_MEM_DEVICE) { 835 ierr = MatShellSetVecType(mat, VECCUDA); CHKERRQ(ierr); 836 } 837 838 // Get RHS vector 839 ierr = VecDuplicate(X, &rhs); CHKERRQ(ierr); 840 ierr = VecDuplicate(Xloc, &rhsloc); CHKERRQ(ierr); 841 ierr = VecZeroEntries(rhsloc); CHKERRQ(ierr); 842 ierr = user->VecGetArray(rhsloc, &r); CHKERRQ(ierr); 843 CeedVectorSetArray(rhsceed, user->memtype, CEED_USE_POINTER, r); 844 845 // Setup qdata, rhs, and target 846 CeedOperatorApply(opsetupgeo, xcoord, qdata, CEED_REQUEST_IMMEDIATE); 847 CeedOperatorApply(opsetuprhs, xcoord, rhsceed, CEED_REQUEST_IMMEDIATE); 848 ierr = CeedVectorSyncArray(rhsceed, user->memtype); CHKERRQ(ierr); 849 CeedVectorDestroy(&xcoord); 850 851 // Gather RHS 852 ierr = user->VecRestoreArray(rhsloc, &r); CHKERRQ(ierr); 853 ierr = VecZeroEntries(rhs); CHKERRQ(ierr); 854 ierr = VecScatterBegin(ltog, rhsloc, rhs, ADD_VALUES, SCATTER_FORWARD); 855 CHKERRQ(ierr); 856 ierr = VecScatterEnd(ltog, rhsloc, rhs, ADD_VALUES, SCATTER_FORWARD); 857 CHKERRQ(ierr); 858 CeedVectorDestroy(&rhsceed); 859 860 ierr = KSPCreate(comm, &ksp); CHKERRQ(ierr); 861 { 862 PC pc; 863 ierr = KSPGetPC(ksp, &pc); CHKERRQ(ierr); 864 if (bpchoice == CEED_BP1 || bpchoice == CEED_BP2) { 865 ierr = PCSetType(pc, PCJACOBI); CHKERRQ(ierr); 866 ierr = PCJacobiSetType(pc, PC_JACOBI_ROWSUM); CHKERRQ(ierr); 867 } else { 868 ierr = PCSetType(pc, PCNONE); CHKERRQ(ierr); 869 } 870 ierr = KSPSetType(ksp, KSPCG); CHKERRQ(ierr); 871 ierr = KSPSetNormType(ksp, KSP_NORM_NATURAL); CHKERRQ(ierr); 872 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 873 PETSC_DEFAULT); CHKERRQ(ierr); 874 } 875 ierr = KSPSetFromOptions(ksp); CHKERRQ(ierr); 876 ierr = KSPSetOperators(ksp, mat, mat); CHKERRQ(ierr); 877 // First run, if benchmarking 878 if (benchmark_mode) { 879 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 1); 880 CHKERRQ(ierr); 881 my_rt_start = MPI_Wtime(); 882 ierr = KSPSolve(ksp, rhs, X); CHKERRQ(ierr); 883 my_rt = MPI_Wtime() - my_rt_start; 884 ierr = MPI_Allreduce(MPI_IN_PLACE, &my_rt, 1, MPI_DOUBLE, MPI_MIN, comm); 885 CHKERRQ(ierr); 886 // Set maxits based on first iteration timing 887 if (my_rt > 0.02) { 888 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 5); 889 CHKERRQ(ierr); 890 } else { 891 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 20); 892 CHKERRQ(ierr); 893 } 894 } 895 896 // Timed solve 897 ierr = VecZeroEntries(X); CHKERRQ(ierr); 898 ierr = PetscBarrier((PetscObject)ksp); CHKERRQ(ierr); 899 900 // -- Performance logging 901 ierr = PetscLogStageRegister("Solve Stage", &solvestage); CHKERRQ(ierr); 902 ierr = PetscLogStagePush(solvestage); CHKERRQ(ierr); 903 904 // -- Solve 905 my_rt_start = MPI_Wtime(); 906 ierr = KSPSolve(ksp, rhs, X); CHKERRQ(ierr); 907 my_rt = MPI_Wtime() - my_rt_start; 908 909 // -- Performance logging 910 ierr = PetscLogStagePop(); 911 912 // Output results 913 { 914 KSPType ksptype; 915 KSPConvergedReason reason; 916 PetscReal rnorm; 917 PetscInt its; 918 ierr = KSPGetType(ksp, &ksptype); CHKERRQ(ierr); 919 ierr = KSPGetConvergedReason(ksp, &reason); CHKERRQ(ierr); 920 ierr = KSPGetIterationNumber(ksp, &its); CHKERRQ(ierr); 921 ierr = KSPGetResidualNorm(ksp, &rnorm); CHKERRQ(ierr); 922 if (!test_mode || reason < 0 || rnorm > 1e-8) { 923 ierr = PetscPrintf(comm, 924 " KSP:\n" 925 " KSP Type : %s\n" 926 " KSP Convergence : %s\n" 927 " Total KSP Iterations : %D\n" 928 " Final rnorm : %e\n", 929 ksptype, KSPConvergedReasons[reason], its, 930 (double)rnorm); CHKERRQ(ierr); 931 } 932 if (!test_mode) { 933 ierr = PetscPrintf(comm," Performance:\n"); CHKERRQ(ierr); 934 } 935 { 936 PetscReal maxerror; 937 ierr = ComputeErrorMax(user, operror, X, target, &maxerror); 938 CHKERRQ(ierr); 939 PetscReal tol = 5e-2; 940 if (!test_mode || maxerror > tol) { 941 ierr = MPI_Allreduce(&my_rt, &rt_min, 1, MPI_DOUBLE, MPI_MIN, comm); 942 CHKERRQ(ierr); 943 ierr = MPI_Allreduce(&my_rt, &rt_max, 1, MPI_DOUBLE, MPI_MAX, comm); 944 CHKERRQ(ierr); 945 ierr = PetscPrintf(comm, 946 " Pointwise Error (max) : %e\n" 947 " CG Solve Time : %g (%g) sec\n", 948 (double)maxerror, rt_max, rt_min); CHKERRQ(ierr); 949 } 950 } 951 if (benchmark_mode && (!test_mode)) { 952 ierr = PetscPrintf(comm, 953 " DoFs/Sec in CG : %g (%g) million\n", 954 1e-6*gsize*its/rt_max, 955 1e-6*gsize*its/rt_min); CHKERRQ(ierr); 956 } 957 } 958 959 if (write_solution) { 960 PetscViewer vtkviewersoln; 961 962 ierr = PetscViewerCreate(comm, &vtkviewersoln); CHKERRQ(ierr); 963 ierr = PetscViewerSetType(vtkviewersoln, PETSCVIEWERVTK); CHKERRQ(ierr); 964 ierr = PetscViewerFileSetName(vtkviewersoln, "solution.vtk"); CHKERRQ(ierr); 965 ierr = VecView(X, vtkviewersoln); CHKERRQ(ierr); 966 ierr = PetscViewerDestroy(&vtkviewersoln); CHKERRQ(ierr); 967 } 968 969 ierr = VecDestroy(&rhs); CHKERRQ(ierr); 970 ierr = VecDestroy(&rhsloc); CHKERRQ(ierr); 971 ierr = VecDestroy(&X); CHKERRQ(ierr); 972 ierr = VecDestroy(&user->Xloc); CHKERRQ(ierr); 973 ierr = VecDestroy(&user->Yloc); CHKERRQ(ierr); 974 ierr = VecScatterDestroy(<og); CHKERRQ(ierr); 975 ierr = VecScatterDestroy(<og0); CHKERRQ(ierr); 976 ierr = VecScatterDestroy(>ogD); CHKERRQ(ierr); 977 ierr = MatDestroy(&mat); CHKERRQ(ierr); 978 ierr = KSPDestroy(&ksp); CHKERRQ(ierr); 979 980 CeedVectorDestroy(&user->xceed); 981 CeedVectorDestroy(&user->yceed); 982 CeedVectorDestroy(&user->qdata); 983 CeedVectorDestroy(&target); 984 CeedOperatorDestroy(&opsetupgeo); 985 CeedOperatorDestroy(&opsetuprhs); 986 CeedOperatorDestroy(&opapply); 987 CeedOperatorDestroy(&operror); 988 CeedElemRestrictionDestroy(&Erestrictu); 989 CeedElemRestrictionDestroy(&Erestrictx); 990 CeedElemRestrictionDestroy(&Erestrictui); 991 CeedElemRestrictionDestroy(&Erestrictqdi); 992 CeedQFunctionDestroy(&qfsetupgeo); 993 CeedQFunctionDestroy(&qfsetuprhs); 994 CeedQFunctionDestroy(&qfapply); 995 CeedQFunctionDestroy(&qferror); 996 CeedBasisDestroy(&basisu); 997 CeedBasisDestroy(&basisx); 998 CeedDestroy(&ceed); 999 ierr = PetscFree(user); CHKERRQ(ierr); 1000 return PetscFinalize(); 1001 } 1002