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 bps [PETSC_DIR=</path/to/petsc>] [CEED_DIR=</path/to/libceed>] 28 // 29 // Sample runs: 30 // 31 // bps -problem bp1 32 // bps -problem bp2 -ceed /cpu/self 33 // bps -problem bp3 -ceed /gpu/occa 34 // bps -problem bp4 -ceed /cpu/occa 35 // bps -problem bp5 -ceed /omp/occa 36 // bps -problem bp6 -ceed /ocl/occa 37 // 38 //TESTARGS -ceed {ceed_resource} -test -problem bp2 39 40 /// @file 41 /// CEED BPs example using PETSc 42 /// See bpsdmplex.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 "common.h" 50 #include "bp1.h" 51 #include "bp2.h" 52 #include "bp3.h" 53 #include "bp4.h" 54 55 static void Split3(PetscInt size, PetscInt m[3], bool reverse) { 56 for (PetscInt d=0,sizeleft=size; d<3; d++) { 57 PetscInt try = (PetscInt)PetscCeilReal(PetscPowReal(sizeleft, 1./(3 - d))); 58 while (try * (sizeleft / try) != sizeleft) try++; 59 m[reverse ? 2-d : d] = try; 60 sizeleft /= try; 61 } 62 } 63 64 static PetscInt Max3(const PetscInt a[3]) { 65 return PetscMax(a[0], PetscMax(a[1], a[2])); 66 } 67 static PetscInt Min3(const PetscInt a[3]) { 68 return PetscMin(a[0], PetscMin(a[1], a[2])); 69 } 70 static void GlobalNodes(const PetscInt p[3], const PetscInt irank[3], 71 PetscInt degree, const PetscInt melem[3], 72 PetscInt mnodes[3]) { 73 for (int d=0; d<3; d++) 74 mnodes[d] = degree*melem[d] + (irank[d] == p[d]-1); 75 } 76 static PetscInt GlobalStart(const PetscInt p[3], const PetscInt irank[3], 77 PetscInt degree, const PetscInt melem[3]) { 78 PetscInt start = 0; 79 // Dumb brute-force is easier to read 80 for (PetscInt i=0; i<p[0]; i++) { 81 for (PetscInt j=0; j<p[1]; j++) { 82 for (PetscInt k=0; k<p[2]; k++) { 83 PetscInt mnodes[3], ijkrank[] = {i,j,k}; 84 if (i == irank[0] && j == irank[1] && k == irank[2]) return start; 85 GlobalNodes(p, ijkrank, degree, melem, mnodes); 86 start += mnodes[0] * mnodes[1] * mnodes[2]; 87 } 88 } 89 } 90 return -1; 91 } 92 static int CreateRestriction(Ceed ceed, const CeedInt melem[3], 93 CeedInt P, CeedInt ncomp, 94 CeedElemRestriction *Erestrict) { 95 const PetscInt nelem = melem[0]*melem[1]*melem[2]; 96 PetscInt mnodes[3], *idx, *idxp; 97 98 // Get indicies 99 for (int d=0; d<3; d++) mnodes[d] = melem[d]*(P-1) + 1; 100 idxp = idx = malloc(nelem*P*P*P*sizeof idx[0]); 101 for (CeedInt i=0; i<melem[0]; i++) { 102 for (CeedInt j=0; j<melem[1]; j++) { 103 for (CeedInt k=0; k<melem[2]; k++,idxp += P*P*P) { 104 for (CeedInt ii=0; ii<P; ii++) { 105 for (CeedInt jj=0; jj<P; jj++) { 106 for (CeedInt kk=0; kk<P; kk++) { 107 if (0) { // This is the C-style (i,j,k) ordering that I prefer 108 idxp[(ii*P+jj)*P+kk] = (((i*(P-1)+ii)*mnodes[1] 109 + (j*(P-1)+jj))*mnodes[2] 110 + (k*(P-1)+kk)); 111 } else { // (k,j,i) ordering for consistency with MFEM example 112 idxp[ii+P*(jj+P*kk)] = (((i*(P-1)+ii)*mnodes[1] 113 + (j*(P-1)+jj))*mnodes[2] 114 + (k*(P-1)+kk)); 115 } 116 } 117 } 118 } 119 } 120 } 121 } 122 123 // Setup CEED restriction 124 CeedElemRestrictionCreate(ceed, nelem, P*P*P, mnodes[0]*mnodes[1]*mnodes[2], ncomp, 125 CEED_MEM_HOST, CEED_OWN_POINTER, idx, Erestrict); 126 127 PetscFunctionReturn(0); 128 } 129 130 // Data for PETSc 131 typedef struct User_ *User; 132 struct User_ { 133 MPI_Comm comm; 134 VecScatter ltog; // Scatter for all entries 135 VecScatter ltog0; // Skip Dirichlet values 136 VecScatter gtogD; // global-to-global; only Dirichlet values 137 Vec Xloc, Yloc; 138 CeedVector xceed, yceed; 139 CeedOperator op; 140 CeedVector rho; 141 Ceed ceed; 142 }; 143 144 // BP Options 145 typedef enum { 146 CEED_BP1 = 0, CEED_BP2 = 1, CEED_BP3 = 2, 147 CEED_BP4 = 3, CEED_BP5 = 4, CEED_BP6 = 5 148 } bpType; 149 static const char *const bpTypes[] = {"bp1","bp2","bp3","bp4","bp5","bp6", 150 "bpType","CEED_BP",0 151 }; 152 153 // BP specific data 154 typedef struct { 155 CeedInt ncompu, qdatasize, qextra; 156 CeedQFunctionUser setup, apply, error; 157 const char *setupfname, *applyfname, *errorfname; 158 CeedEvalMode inmode, outmode; 159 CeedQuadMode qmode; 160 } bpData; 161 162 bpData bpOptions[6] = { 163 [CEED_BP1] = { 164 .ncompu = 1, 165 .qdatasize = 1, 166 .qextra = 1, 167 .setup = SetupMass, 168 .apply = Mass, 169 .error = Error, 170 .setupfname = SetupMass_loc, 171 .applyfname = Mass_loc, 172 .errorfname = Error_loc, 173 .inmode = CEED_EVAL_INTERP, 174 .outmode = CEED_EVAL_INTERP, 175 .qmode = CEED_GAUSS 176 }, 177 [CEED_BP2] = { 178 .ncompu = 3, 179 .qdatasize = 1, 180 .qextra = 1, 181 .setup = SetupMass3, 182 .apply = Mass3, 183 .error = Error3, 184 .setupfname = SetupMass3_loc, 185 .applyfname = Mass3_loc, 186 .errorfname = Error3_loc, 187 .inmode = CEED_EVAL_INTERP, 188 .outmode = CEED_EVAL_INTERP, 189 .qmode = CEED_GAUSS 190 }, 191 [CEED_BP3] = { 192 .ncompu = 1, 193 .qdatasize = 6, 194 .qextra = 1, 195 .setup = SetupDiff, 196 .apply = Diff, 197 .error = Error, 198 .setupfname = SetupDiff_loc, 199 .applyfname = Diff_loc, 200 .errorfname = Error_loc, 201 .inmode = CEED_EVAL_GRAD, 202 .outmode = CEED_EVAL_GRAD, 203 .qmode = CEED_GAUSS 204 }, 205 [CEED_BP4] = { 206 .ncompu = 3, 207 .qdatasize = 6, 208 .qextra = 1, 209 .setup = SetupDiff3, 210 .apply = Diff3, 211 .error = Error3, 212 .setupfname = SetupDiff3_loc, 213 .applyfname = Diff3_loc, 214 .errorfname = Error3_loc, 215 .inmode = CEED_EVAL_GRAD, 216 .outmode = CEED_EVAL_GRAD, 217 .qmode = CEED_GAUSS 218 }, 219 [CEED_BP5] = { 220 .ncompu = 1, 221 .qdatasize = 6, 222 .qextra = 0, 223 .setup = SetupDiff, 224 .apply = Diff, 225 .error = Error, 226 .setupfname = SetupDiff_loc, 227 .applyfname = Diff_loc, 228 .errorfname = Error_loc, 229 .inmode = CEED_EVAL_GRAD, 230 .outmode = CEED_EVAL_GRAD, 231 .qmode = CEED_GAUSS_LOBATTO 232 }, 233 [CEED_BP6] = { 234 .ncompu = 3, 235 .qdatasize = 6, 236 .qextra = 0, 237 .setup = SetupDiff3, 238 .apply = Diff3, 239 .error = Error3, 240 .setupfname = SetupDiff3_loc, 241 .applyfname = Diff3_loc, 242 .errorfname = Error3_loc, 243 .inmode = CEED_EVAL_GRAD, 244 .outmode = CEED_EVAL_GRAD, 245 .qmode = CEED_GAUSS_LOBATTO 246 } 247 }; 248 249 // This function uses libCEED to compute the action of the mass matrix 250 static PetscErrorCode MatMult_Mass(Mat A, Vec X, Vec Y) { 251 PetscErrorCode ierr; 252 User user; 253 PetscScalar *x, *y; 254 255 PetscFunctionBeginUser; 256 ierr = MatShellGetContext(A, &user); CHKERRQ(ierr); 257 ierr = VecScatterBegin(user->ltog, X, user->Xloc, INSERT_VALUES, 258 SCATTER_REVERSE); CHKERRQ(ierr); 259 ierr = VecScatterEnd(user->ltog, X, user->Xloc, INSERT_VALUES, SCATTER_REVERSE); 260 CHKERRQ(ierr); 261 ierr = VecZeroEntries(user->Yloc); CHKERRQ(ierr); 262 263 ierr = VecGetArrayRead(user->Xloc, (const PetscScalar **)&x); CHKERRQ(ierr); 264 ierr = VecGetArray(user->Yloc, &y); CHKERRQ(ierr); 265 CeedVectorSetArray(user->xceed, CEED_MEM_HOST, CEED_USE_POINTER, x); 266 CeedVectorSetArray(user->yceed, CEED_MEM_HOST, CEED_USE_POINTER, y); 267 268 CeedOperatorApply(user->op, user->xceed, user->yceed, 269 CEED_REQUEST_IMMEDIATE); 270 ierr = CeedVectorSyncArray(user->yceed, CEED_MEM_HOST); CHKERRQ(ierr); 271 272 ierr = VecRestoreArrayRead(user->Xloc, (const PetscScalar **)&x); CHKERRQ(ierr); 273 ierr = VecRestoreArray(user->Yloc, &y); CHKERRQ(ierr); 274 275 if (Y) { 276 ierr = VecZeroEntries(Y); CHKERRQ(ierr); 277 ierr = VecScatterBegin(user->ltog, user->Yloc, Y, ADD_VALUES, SCATTER_FORWARD); 278 CHKERRQ(ierr); 279 ierr = VecScatterEnd(user->ltog, user->Yloc, Y, ADD_VALUES, SCATTER_FORWARD); 280 CHKERRQ(ierr); 281 } 282 PetscFunctionReturn(0); 283 } 284 285 // This function uses libCEED to compute the action of the Laplacian with 286 // Dirichlet boundary conditions 287 static PetscErrorCode MatMult_Diff(Mat A, Vec X, Vec Y) { 288 PetscErrorCode ierr; 289 User user; 290 PetscScalar *x, *y; 291 292 PetscFunctionBeginUser; 293 ierr = MatShellGetContext(A, &user); CHKERRQ(ierr); 294 295 // Global-to-local 296 ierr = VecScatterBegin(user->ltog0, X, user->Xloc, INSERT_VALUES, 297 SCATTER_REVERSE); CHKERRQ(ierr); 298 ierr = VecScatterEnd(user->ltog0, X, user->Xloc, INSERT_VALUES, 299 SCATTER_REVERSE); 300 CHKERRQ(ierr); 301 ierr = VecZeroEntries(user->Yloc); CHKERRQ(ierr); 302 303 // Setup CEED vectors 304 ierr = VecGetArrayRead(user->Xloc, (const PetscScalar **)&x); CHKERRQ(ierr); 305 ierr = VecGetArray(user->Yloc, &y); CHKERRQ(ierr); 306 CeedVectorSetArray(user->xceed, CEED_MEM_HOST, CEED_USE_POINTER, x); 307 CeedVectorSetArray(user->yceed, CEED_MEM_HOST, CEED_USE_POINTER, y); 308 309 // Apply CEED operator 310 CeedOperatorApply(user->op, user->xceed, user->yceed, 311 CEED_REQUEST_IMMEDIATE); 312 ierr = CeedVectorSyncArray(user->yceed, CEED_MEM_HOST); CHKERRQ(ierr); 313 314 // Restore PETSc vectors 315 ierr = VecRestoreArrayRead(user->Xloc, (const PetscScalar **)&x); CHKERRQ(ierr); 316 ierr = VecRestoreArray(user->Yloc, &y); CHKERRQ(ierr); 317 318 // Local-to-global 319 ierr = VecZeroEntries(Y); CHKERRQ(ierr); 320 ierr = VecScatterBegin(user->gtogD, X, Y, INSERT_VALUES, SCATTER_FORWARD); 321 CHKERRQ(ierr); 322 ierr = VecScatterEnd(user->gtogD, X, Y, INSERT_VALUES, SCATTER_FORWARD); 323 CHKERRQ(ierr); 324 ierr = VecScatterBegin(user->ltog0, user->Yloc, Y, ADD_VALUES, SCATTER_FORWARD); 325 CHKERRQ(ierr); 326 ierr = VecScatterEnd(user->ltog0, user->Yloc, Y, ADD_VALUES, SCATTER_FORWARD); 327 CHKERRQ(ierr); 328 329 PetscFunctionReturn(0); 330 } 331 332 // This function calculates the error in the final solution 333 static PetscErrorCode ComputeErrorMax(User user, CeedOperator op_error, Vec X, 334 CeedVector target, PetscReal *maxerror) { 335 PetscErrorCode ierr; 336 PetscScalar *x; 337 CeedVector collocated_error; 338 CeedInt length; 339 340 PetscFunctionBeginUser; 341 CeedVectorGetLength(target, &length); 342 CeedVectorCreate(user->ceed, length, &collocated_error); 343 344 // Global-to-local 345 ierr = VecScatterBegin(user->ltog, X, user->Xloc, INSERT_VALUES, 346 SCATTER_REVERSE); CHKERRQ(ierr); 347 ierr = VecScatterEnd(user->ltog, X, user->Xloc, INSERT_VALUES, SCATTER_REVERSE); 348 CHKERRQ(ierr); 349 350 // Setup CEED vector 351 ierr = VecGetArrayRead(user->Xloc, (const PetscScalar **)&x); CHKERRQ(ierr); 352 CeedVectorSetArray(user->xceed, CEED_MEM_HOST, CEED_USE_POINTER, x); 353 354 // Apply CEED operator 355 CeedOperatorApply(op_error, user->xceed, collocated_error, 356 CEED_REQUEST_IMMEDIATE); 357 358 // Restore PETSc vector 359 VecRestoreArrayRead(user->Xloc, (const PetscScalar **)&x); CHKERRQ(ierr); 360 361 // Reduce max error 362 *maxerror = 0; 363 const CeedScalar *e; 364 CeedVectorGetArrayRead(collocated_error, CEED_MEM_HOST, &e); 365 for (CeedInt i=0; i<length; i++) { 366 *maxerror = PetscMax(*maxerror, PetscAbsScalar(e[i])); 367 } 368 CeedVectorRestoreArrayRead(collocated_error, &e); 369 ierr = MPI_Allreduce(MPI_IN_PLACE, maxerror, 370 1, MPIU_REAL, MPIU_MAX, user->comm); CHKERRQ(ierr); 371 372 // Cleanup 373 CeedVectorDestroy(&collocated_error); 374 375 PetscFunctionReturn(0); 376 } 377 378 int main(int argc, char **argv) { 379 PetscInt ierr; 380 MPI_Comm comm; 381 char ceedresource[PETSC_MAX_PATH_LEN] = "/cpu/self"; 382 double my_rt_start, my_rt, rt_min, rt_max; 383 PetscInt degree, qextra, localnodes, localelem, melem[3], mnodes[3], p[3], 384 irank[3], lnodes[3], lsize, ncompu = 1; 385 PetscScalar *r; 386 PetscBool test_mode, benchmark_mode, write_solution; 387 PetscMPIInt size, rank; 388 Vec X, Xloc, rhs, rhsloc; 389 Mat mat; 390 KSP ksp; 391 VecScatter ltog, ltog0, gtogD; 392 User user; 393 Ceed ceed; 394 CeedBasis basisx, basisu; 395 CeedElemRestriction Erestrictx, Erestrictu, Erestrictxi, Erestrictui, 396 Erestrictqdi; 397 CeedQFunction qf_setup, qf_apply, qf_error; 398 CeedOperator op_setup, op_apply, op_error; 399 CeedVector xcoord, rho, rhsceed, target; 400 CeedInt P, Q; 401 const CeedInt dim = 3, ncompx = 3; 402 bpType bpChoice; 403 404 ierr = PetscInitialize(&argc, &argv, NULL, help); 405 if (ierr) return ierr; 406 comm = PETSC_COMM_WORLD; 407 ierr = PetscOptionsBegin(comm, NULL, "CEED BPs in PETSc", NULL); CHKERRQ(ierr); 408 bpChoice = CEED_BP1; 409 ierr = PetscOptionsEnum("-problem", 410 "CEED benchmark problem to solve", NULL, 411 bpTypes, (PetscEnum)bpChoice, (PetscEnum *)&bpChoice, 412 NULL); CHKERRQ(ierr); 413 ncompu = bpOptions[bpChoice].ncompu; 414 test_mode = PETSC_FALSE; 415 ierr = PetscOptionsBool("-test", 416 "Testing mode (do not print unless error is large)", 417 NULL, test_mode, &test_mode, NULL); CHKERRQ(ierr); 418 benchmark_mode = PETSC_FALSE; 419 ierr = PetscOptionsBool("-benchmark", 420 "Benchmarking mode (prints benchmark statistics)", 421 NULL, benchmark_mode, &benchmark_mode, NULL); 422 CHKERRQ(ierr); 423 write_solution = PETSC_FALSE; 424 ierr = PetscOptionsBool("-write_solution", 425 "Write solution for visualization", 426 NULL, write_solution, &write_solution, NULL); 427 CHKERRQ(ierr); 428 degree = test_mode ? 3 : 1; 429 ierr = PetscOptionsInt("-degree", "Polynomial degree of tensor product basis", 430 NULL, degree, °ree, NULL); CHKERRQ(ierr); 431 qextra = bpOptions[bpChoice].qextra; 432 ierr = PetscOptionsInt("-qextra", "Number of extra quadrature points", 433 NULL, qextra, &qextra, NULL); CHKERRQ(ierr); 434 ierr = PetscOptionsString("-ceed", "CEED resource specifier", 435 NULL, ceedresource, ceedresource, 436 sizeof(ceedresource), NULL); CHKERRQ(ierr); 437 localnodes = 1000; 438 ierr = PetscOptionsInt("-local", 439 "Target number of locally owned nodes per process", 440 NULL, localnodes, &localnodes, NULL); CHKERRQ(ierr); 441 ierr = PetscOptionsEnd(); CHKERRQ(ierr); 442 P = degree + 1; 443 Q = P + qextra; 444 445 // Determine size of process grid 446 ierr = MPI_Comm_size(comm, &size); CHKERRQ(ierr); 447 Split3(size, p, false); 448 449 // Find a nicely composite number of elements no less than localnodes 450 for (localelem = PetscMax(1, localnodes / (degree*degree*degree)); ; 451 localelem++) { 452 Split3(localelem, melem, true); 453 if (Max3(melem) / Min3(melem) <= 2) break; 454 } 455 456 // Find my location in the process grid 457 ierr = MPI_Comm_rank(comm, &rank); CHKERRQ(ierr); 458 for (int d=0,rankleft=rank; d<dim; d++) { 459 const int pstride[3] = {p[1] *p[2], p[2], 1}; 460 irank[d] = rankleft / pstride[d]; 461 rankleft -= irank[d] * pstride[d]; 462 } 463 464 GlobalNodes(p, irank, degree, melem, mnodes); 465 466 // Setup global vector 467 ierr = VecCreate(comm, &X); CHKERRQ(ierr); 468 ierr = VecSetSizes(X, mnodes[0]*mnodes[1]*mnodes[2]*ncompu, PETSC_DECIDE); 469 CHKERRQ(ierr); 470 ierr = VecSetUp(X); CHKERRQ(ierr); 471 472 // Print summary 473 if (!test_mode) { 474 CeedInt gsize; 475 ierr = VecGetSize(X, &gsize); CHKERRQ(ierr); 476 ierr = PetscPrintf(comm, 477 "\n-- CEED Benchmark Problem %d -- libCEED + PETSc --\n" 478 " libCEED:\n" 479 " libCEED Backend : %s\n" 480 " Mesh:\n" 481 " Number of 1D Basis Nodes (p) : %d\n" 482 " Number of 1D Quadrature Points (q) : %d\n" 483 " Global nodes : %D\n" 484 " Process Decomposition : %D %D %D\n" 485 " Local Elements : %D = %D %D %D\n" 486 " Owned nodes : %D = %D %D %D\n", 487 bpChoice+1, ceedresource, P, Q, gsize/ncompu, p[0], 488 p[1], p[2], localelem, melem[0], melem[1], melem[2], 489 mnodes[0]*mnodes[1]*mnodes[2], mnodes[0], mnodes[1], mnodes[2]); 490 CHKERRQ(ierr); 491 } 492 493 { 494 lsize = 1; 495 for (int d=0; d<dim; d++) { 496 lnodes[d] = melem[d]*degree + 1; 497 lsize *= lnodes[d]; 498 } 499 ierr = VecCreate(PETSC_COMM_SELF, &Xloc); CHKERRQ(ierr); 500 ierr = VecSetSizes(Xloc, lsize*ncompu, PETSC_DECIDE); CHKERRQ(ierr); 501 ierr = VecSetUp(Xloc); CHKERRQ(ierr); 502 503 // Create local-to-global scatter 504 PetscInt *ltogind, *ltogind0, *locind, l0count; 505 IS ltogis, ltogis0, locis; 506 PetscInt gstart[2][2][2], gmnodes[2][2][2][dim]; 507 508 for (int i=0; i<2; i++) { 509 for (int j=0; j<2; j++) { 510 for (int k=0; k<2; k++) { 511 PetscInt ijkrank[3] = {irank[0]+i, irank[1]+j, irank[2]+k}; 512 gstart[i][j][k] = GlobalStart(p, ijkrank, degree, melem); 513 GlobalNodes(p, ijkrank, degree, melem, gmnodes[i][j][k]); 514 } 515 } 516 } 517 518 ierr = PetscMalloc1(lsize, <ogind); CHKERRQ(ierr); 519 ierr = PetscMalloc1(lsize, <ogind0); CHKERRQ(ierr); 520 ierr = PetscMalloc1(lsize, &locind); CHKERRQ(ierr); 521 l0count = 0; 522 for (PetscInt i=0,ir,ii; ir=i>=mnodes[0], ii=i-ir*mnodes[0], i<lnodes[0]; i++) { 523 for (PetscInt j=0,jr,jj; jr=j>=mnodes[1], jj=j-jr*mnodes[1], j<lnodes[1]; j++) { 524 for (PetscInt k=0,kr,kk; kr=k>=mnodes[2], kk=k-kr*mnodes[2], k<lnodes[2]; k++) { 525 PetscInt here = (i*lnodes[1]+j)*lnodes[2]+k; 526 ltogind[here] = 527 gstart[ir][jr][kr] + (ii*gmnodes[ir][jr][kr][1]+jj)*gmnodes[ir][jr][kr][2]+kk; 528 if ((irank[0] == 0 && i == 0) 529 || (irank[1] == 0 && j == 0) 530 || (irank[2] == 0 && k == 0) 531 || (irank[0]+1 == p[0] && i+1 == lnodes[0]) 532 || (irank[1]+1 == p[1] && j+1 == lnodes[1]) 533 || (irank[2]+1 == p[2] && k+1 == lnodes[2])) 534 continue; 535 ltogind0[l0count] = ltogind[here]; 536 locind[l0count++] = here; 537 } 538 } 539 } 540 ierr = ISCreateBlock(comm, ncompu, lsize, ltogind, PETSC_OWN_POINTER, 541 <ogis); CHKERRQ(ierr); 542 ierr = VecScatterCreate(Xloc, NULL, X, ltogis, <og); CHKERRQ(ierr); 543 CHKERRQ(ierr); 544 ierr = ISCreateBlock(comm, ncompu, l0count, ltogind0, PETSC_OWN_POINTER, 545 <ogis0); CHKERRQ(ierr); 546 ierr = ISCreateBlock(comm, ncompu, l0count, locind, PETSC_OWN_POINTER, 547 &locis); CHKERRQ(ierr); 548 ierr = VecScatterCreate(Xloc, locis, X, ltogis0, <og0); CHKERRQ(ierr); 549 { 550 // Create global-to-global scatter for Dirichlet values (everything not in 551 // ltogis0, which is the range of ltog0) 552 PetscInt xstart, xend, *indD, countD = 0; 553 IS isD; 554 const PetscScalar *x; 555 ierr = VecZeroEntries(Xloc); CHKERRQ(ierr); 556 ierr = VecSet(X, 1.0); CHKERRQ(ierr); 557 ierr = VecScatterBegin(ltog0, Xloc, X, INSERT_VALUES, SCATTER_FORWARD); 558 CHKERRQ(ierr); 559 ierr = VecScatterEnd(ltog0, Xloc, X, INSERT_VALUES, SCATTER_FORWARD); 560 CHKERRQ(ierr); 561 ierr = VecGetOwnershipRange(X, &xstart, &xend); CHKERRQ(ierr); 562 ierr = PetscMalloc1(xend-xstart, &indD); CHKERRQ(ierr); 563 ierr = VecGetArrayRead(X, &x); CHKERRQ(ierr); 564 for (PetscInt i=0; i<xend-xstart; i++) { 565 if (x[i] == 1.) indD[countD++] = xstart + i; 566 } 567 ierr = VecRestoreArrayRead(X, &x); CHKERRQ(ierr); 568 ierr = ISCreateGeneral(comm, countD, indD, PETSC_COPY_VALUES, &isD); 569 CHKERRQ(ierr); 570 ierr = PetscFree(indD); CHKERRQ(ierr); 571 ierr = VecScatterCreate(X, isD, X, isD, >ogD); CHKERRQ(ierr); 572 ierr = ISDestroy(&isD); CHKERRQ(ierr); 573 } 574 ierr = ISDestroy(<ogis); CHKERRQ(ierr); 575 ierr = ISDestroy(<ogis0); CHKERRQ(ierr); 576 ierr = ISDestroy(&locis); CHKERRQ(ierr); 577 } 578 579 // Set up libCEED 580 CeedInit(ceedresource, &ceed); 581 582 // CEED bases 583 CeedBasisCreateTensorH1Lagrange(ceed, dim, ncompu, P, Q, 584 bpOptions[bpChoice].qmode, &basisu); 585 CeedBasisCreateTensorH1Lagrange(ceed, dim, ncompx, 2, Q, 586 bpOptions[bpChoice].qmode, &basisx); 587 588 // CEED restrictions 589 CreateRestriction(ceed, melem, P, ncompu, &Erestrictu); 590 CreateRestriction(ceed, melem, 2, dim, &Erestrictx); 591 CeedInt nelem = melem[0]*melem[1]*melem[2]; 592 CeedElemRestrictionCreateIdentity(ceed, nelem, Q*Q*Q, nelem*Q*Q*Q, ncompu, 593 &Erestrictui); 594 CeedElemRestrictionCreateIdentity(ceed, nelem, 595 Q*Q*Q, 596 nelem*Q*Q*Q, 597 bpOptions[bpChoice].qdatasize, &Erestrictqdi); 598 CeedElemRestrictionCreateIdentity(ceed, nelem, Q*Q*Q, nelem*Q*Q*Q, 1, 599 &Erestrictxi); 600 { 601 CeedScalar *xloc; 602 CeedInt shape[3] = {melem[0]+1, melem[1]+1, melem[2]+1}, len = 603 shape[0]*shape[1]*shape[2]; 604 xloc = malloc(len*ncompx*sizeof xloc[0]); 605 for (CeedInt i=0; i<shape[0]; i++) { 606 for (CeedInt j=0; j<shape[1]; j++) { 607 for (CeedInt k=0; k<shape[2]; k++) { 608 xloc[((i*shape[1]+j)*shape[2]+k) + 0*len] = 1.*(irank[0]*melem[0]+i) / 609 (p[0]*melem[0]); 610 xloc[((i*shape[1]+j)*shape[2]+k) + 1*len] = 1.*(irank[1]*melem[1]+j) / 611 (p[1]*melem[1]); 612 xloc[((i*shape[1]+j)*shape[2]+k) + 2*len] = 1.*(irank[2]*melem[2]+k) / 613 (p[2]*melem[2]); 614 } 615 } 616 } 617 CeedVectorCreate(ceed, len*ncompx, &xcoord); 618 CeedVectorSetArray(xcoord, CEED_MEM_HOST, CEED_OWN_POINTER, xloc); 619 } 620 621 // Create the Q-function that builds the operator (i.e. computes its 622 // quadrature data) and set its context data 623 CeedQFunctionCreateInterior(ceed, 1, bpOptions[bpChoice].setup, 624 bpOptions[bpChoice].setupfname, &qf_setup); 625 CeedQFunctionAddInput(qf_setup, "x", ncompx, CEED_EVAL_INTERP); 626 CeedQFunctionAddInput(qf_setup, "dx", ncompx*dim, CEED_EVAL_GRAD); 627 CeedQFunctionAddInput(qf_setup, "weight", 1, CEED_EVAL_WEIGHT); 628 CeedQFunctionAddOutput(qf_setup, "rho", bpOptions[bpChoice].qdatasize, 629 CEED_EVAL_NONE); 630 CeedQFunctionAddOutput(qf_setup, "true_soln", ncompu, CEED_EVAL_NONE); 631 CeedQFunctionAddOutput(qf_setup, "rhs", ncompu, CEED_EVAL_INTERP); 632 633 // Set up PDE operator 634 CeedQFunctionCreateInterior(ceed, 1, bpOptions[bpChoice].apply, 635 bpOptions[bpChoice].applyfname, &qf_apply); 636 // Add inputs and outputs 637 CeedInt gradInScale = bpOptions[bpChoice].inmode==CEED_EVAL_GRAD ? 3 : 1; 638 CeedInt gradOutScale = bpOptions[bpChoice].outmode==CEED_EVAL_GRAD ? 3 : 1; 639 CeedQFunctionAddInput(qf_apply, "u", ncompu*gradInScale, 640 bpOptions[bpChoice].inmode); 641 CeedQFunctionAddInput(qf_apply, "rho", bpOptions[bpChoice].qdatasize, 642 CEED_EVAL_NONE); 643 CeedQFunctionAddOutput(qf_apply, "v", ncompu*gradOutScale, 644 bpOptions[bpChoice].outmode); 645 646 // Create the error qfunction 647 CeedQFunctionCreateInterior(ceed, 1, bpOptions[bpChoice].error, 648 bpOptions[bpChoice].errorfname, &qf_error); 649 CeedQFunctionAddInput(qf_error, "u", ncompu, CEED_EVAL_INTERP); 650 CeedQFunctionAddInput(qf_error, "true_soln", ncompu, CEED_EVAL_NONE); 651 CeedQFunctionAddOutput(qf_error, "error", ncompu, CEED_EVAL_NONE); 652 653 // Create the persistent vectors that will be needed in setup 654 CeedInt nqpts; 655 CeedBasisGetNumQuadraturePoints(basisu, &nqpts); 656 CeedVectorCreate(ceed, bpOptions[bpChoice].qdatasize*nelem*nqpts, &rho); 657 CeedVectorCreate(ceed, nelem*nqpts*ncompu, &target); 658 CeedVectorCreate(ceed, lsize*ncompu, &rhsceed); 659 660 // Create the operator that builds the quadrature data for the ceed operator 661 CeedOperatorCreate(ceed, qf_setup, NULL, NULL, &op_setup); 662 CeedOperatorSetField(op_setup, "x", Erestrictx, CEED_NOTRANSPOSE, 663 basisx, CEED_VECTOR_ACTIVE); 664 CeedOperatorSetField(op_setup, "dx", Erestrictx, CEED_NOTRANSPOSE, 665 basisx, CEED_VECTOR_ACTIVE); 666 CeedOperatorSetField(op_setup, "weight", Erestrictxi, CEED_NOTRANSPOSE, 667 basisx, CEED_VECTOR_NONE); 668 CeedOperatorSetField(op_setup, "rho", Erestrictqdi, CEED_NOTRANSPOSE, 669 CEED_BASIS_COLLOCATED, CEED_VECTOR_ACTIVE); 670 CeedOperatorSetField(op_setup, "true_soln", Erestrictui, CEED_NOTRANSPOSE, 671 CEED_BASIS_COLLOCATED, target); 672 CeedOperatorSetField(op_setup, "rhs", Erestrictu, CEED_TRANSPOSE, 673 basisu, rhsceed); 674 675 // Create the mass or diff operator 676 CeedOperatorCreate(ceed, qf_apply, NULL, NULL, &op_apply); 677 CeedOperatorSetField(op_apply, "u", Erestrictu, CEED_TRANSPOSE, 678 basisu, CEED_VECTOR_ACTIVE); 679 CeedOperatorSetField(op_apply, "rho", Erestrictqdi, CEED_NOTRANSPOSE, 680 CEED_BASIS_COLLOCATED, rho); 681 CeedOperatorSetField(op_apply, "v", Erestrictu, CEED_TRANSPOSE, 682 basisu, CEED_VECTOR_ACTIVE); 683 684 // Create the error operator 685 CeedOperatorCreate(ceed, qf_error, NULL, NULL, &op_error); 686 CeedOperatorSetField(op_error, "u", Erestrictu, CEED_TRANSPOSE, 687 basisu, CEED_VECTOR_ACTIVE); 688 CeedOperatorSetField(op_error, "true_soln", Erestrictui, CEED_NOTRANSPOSE, 689 CEED_BASIS_COLLOCATED, target); 690 CeedOperatorSetField(op_error, "error", Erestrictui, CEED_NOTRANSPOSE, 691 CEED_BASIS_COLLOCATED, CEED_VECTOR_ACTIVE); 692 693 694 // Set up Mat 695 ierr = PetscMalloc1(1, &user); CHKERRQ(ierr); 696 user->comm = comm; 697 user->ltog = ltog; 698 if (bpChoice != CEED_BP1 && bpChoice != CEED_BP2) { 699 user->ltog0 = ltog0; 700 user->gtogD = gtogD; 701 } 702 user->Xloc = Xloc; 703 ierr = VecDuplicate(Xloc, &user->Yloc); CHKERRQ(ierr); 704 CeedVectorCreate(ceed, lsize*ncompu, &user->xceed); 705 CeedVectorCreate(ceed, lsize*ncompu, &user->yceed); 706 user->op = op_apply; 707 user->rho = rho; 708 user->ceed = ceed; 709 710 ierr = MatCreateShell(comm, mnodes[0]*mnodes[1]*mnodes[2]*ncompu, 711 mnodes[0]*mnodes[1]*mnodes[2]*ncompu, 712 PETSC_DECIDE, PETSC_DECIDE, user, &mat); CHKERRQ(ierr); 713 if (bpChoice == CEED_BP1 || bpChoice == CEED_BP2) { 714 ierr = MatShellSetOperation(mat, MATOP_MULT, (void(*)(void))MatMult_Mass); 715 CHKERRQ(ierr); 716 } else { 717 ierr = MatShellSetOperation(mat, MATOP_MULT, (void(*)(void))MatMult_Diff); 718 CHKERRQ(ierr); 719 } 720 ierr = MatCreateVecs(mat, &rhs, NULL); CHKERRQ(ierr); 721 722 // Get RHS vector 723 ierr = VecDuplicate(Xloc, &rhsloc); CHKERRQ(ierr); 724 ierr = VecZeroEntries(rhsloc); CHKERRQ(ierr); 725 ierr = VecGetArray(rhsloc, &r); CHKERRQ(ierr); 726 CeedVectorSetArray(rhsceed, CEED_MEM_HOST, CEED_USE_POINTER, r); 727 728 // Setup rho, rhs, and target 729 CeedOperatorApply(op_setup, xcoord, rho, CEED_REQUEST_IMMEDIATE); 730 ierr = CeedVectorSyncArray(rhsceed, CEED_MEM_HOST); CHKERRQ(ierr); 731 CeedVectorDestroy(&xcoord); 732 733 // Gather RHS 734 ierr = VecRestoreArray(rhsloc, &r); CHKERRQ(ierr); 735 ierr = VecZeroEntries(rhs); CHKERRQ(ierr); 736 ierr = VecScatterBegin(ltog, rhsloc, rhs, ADD_VALUES, SCATTER_FORWARD); 737 CHKERRQ(ierr); 738 ierr = VecScatterEnd(ltog, rhsloc, rhs, ADD_VALUES, SCATTER_FORWARD); 739 CHKERRQ(ierr); 740 CeedVectorDestroy(&rhsceed); 741 742 ierr = KSPCreate(comm, &ksp); CHKERRQ(ierr); 743 { 744 PC pc; 745 ierr = KSPGetPC(ksp, &pc); CHKERRQ(ierr); 746 if (bpChoice == CEED_BP1 || bpChoice == CEED_BP2) { 747 ierr = PCSetType(pc, PCJACOBI); CHKERRQ(ierr); 748 ierr = PCJacobiSetType(pc, PC_JACOBI_ROWSUM); CHKERRQ(ierr); 749 } else { 750 ierr = PCSetType(pc, PCNONE); CHKERRQ(ierr); 751 } 752 ierr = KSPSetType(ksp, KSPCG); CHKERRQ(ierr); 753 ierr = KSPSetNormType(ksp, KSP_NORM_NATURAL); CHKERRQ(ierr); 754 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 755 PETSC_DEFAULT); CHKERRQ(ierr); 756 } 757 ierr = KSPSetFromOptions(ksp); CHKERRQ(ierr); 758 ierr = KSPSetOperators(ksp, mat, mat); CHKERRQ(ierr); 759 // First run, if benchmarking 760 if (benchmark_mode) { 761 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 1); 762 CHKERRQ(ierr); 763 my_rt_start = MPI_Wtime(); 764 ierr = KSPSolve(ksp, rhs, X); CHKERRQ(ierr); 765 my_rt = MPI_Wtime() - my_rt_start; 766 // Set maxits based on first iteration timing 767 if (my_rt > 0.02) { 768 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 5); 769 CHKERRQ(ierr); 770 } else { 771 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 20); 772 CHKERRQ(ierr); 773 } 774 } 775 // Timed solve 776 my_rt_start = MPI_Wtime(); 777 ierr = KSPSolve(ksp, rhs, X); CHKERRQ(ierr); 778 my_rt = MPI_Wtime() - my_rt_start; 779 { 780 KSPType ksptype; 781 KSPConvergedReason reason; 782 PetscReal rnorm; 783 PetscInt its; 784 ierr = KSPGetType(ksp, &ksptype); CHKERRQ(ierr); 785 ierr = KSPGetConvergedReason(ksp, &reason); CHKERRQ(ierr); 786 ierr = KSPGetIterationNumber(ksp, &its); CHKERRQ(ierr); 787 ierr = KSPGetResidualNorm(ksp, &rnorm); CHKERRQ(ierr); 788 if (!test_mode || reason < 0 || rnorm > 1e-8) { 789 ierr = PetscPrintf(comm, 790 " KSP:\n" 791 " KSP Type : %s\n" 792 " KSP Convergence : %s\n" 793 " Total KSP Iterations : %D\n" 794 " Final rnorm : %e\n", 795 ksptype, KSPConvergedReasons[reason], its, 796 (double)rnorm); CHKERRQ(ierr); 797 } 798 if (benchmark_mode && (!test_mode)) { 799 CeedInt gsize; 800 ierr = VecGetSize(X, &gsize); CHKERRQ(ierr); 801 MPI_Reduce(&my_rt, &rt_min, 1, MPI_DOUBLE, MPI_MIN, 0, comm); 802 MPI_Reduce(&my_rt, &rt_max, 1, MPI_DOUBLE, MPI_MAX, 0, comm); 803 ierr = PetscPrintf(comm, 804 " Performance:\n" 805 " CG Solve Time : %g (%g) sec\n" 806 " DoFs/Sec in CG : %g (%g) million\n", 807 rt_max, rt_min, 1e-6*gsize*its/rt_max, 808 1e-6*gsize*its/rt_min); CHKERRQ(ierr); 809 } 810 } 811 812 { 813 PetscReal maxerror; 814 ierr = ComputeErrorMax(user, op_error, X, target, &maxerror); CHKERRQ(ierr); 815 PetscReal tol = (bpChoice == CEED_BP1 || bpChoice == CEED_BP2) ? 5e-3 : 5e-2; 816 if (!test_mode || maxerror > tol) { 817 ierr = PetscPrintf(comm, 818 " Pointwise Error (max) : %e\n", 819 (double)maxerror); CHKERRQ(ierr); 820 } 821 } 822 823 if (write_solution) { 824 PetscViewer vtkviewersoln; 825 826 ierr = PetscViewerCreate(comm, &vtkviewersoln); CHKERRQ(ierr); 827 ierr = PetscViewerSetType(vtkviewersoln, PETSCVIEWERVTK); CHKERRQ(ierr); 828 ierr = PetscViewerFileSetName(vtkviewersoln, "solution.vtk"); CHKERRQ(ierr); 829 ierr = VecView(X, vtkviewersoln); CHKERRQ(ierr); 830 ierr = PetscViewerDestroy(&vtkviewersoln); CHKERRQ(ierr); 831 } 832 833 ierr = VecDestroy(&rhs); CHKERRQ(ierr); 834 ierr = VecDestroy(&rhsloc); CHKERRQ(ierr); 835 ierr = VecDestroy(&X); CHKERRQ(ierr); 836 ierr = VecDestroy(&user->Xloc); CHKERRQ(ierr); 837 ierr = VecDestroy(&user->Yloc); CHKERRQ(ierr); 838 ierr = VecScatterDestroy(<og); CHKERRQ(ierr); 839 ierr = VecScatterDestroy(<og0); CHKERRQ(ierr); 840 ierr = VecScatterDestroy(>ogD); CHKERRQ(ierr); 841 ierr = MatDestroy(&mat); CHKERRQ(ierr); 842 ierr = KSPDestroy(&ksp); CHKERRQ(ierr); 843 844 CeedVectorDestroy(&user->xceed); 845 CeedVectorDestroy(&user->yceed); 846 CeedVectorDestroy(&user->rho); 847 CeedVectorDestroy(&target); 848 CeedOperatorDestroy(&op_setup); 849 CeedOperatorDestroy(&op_apply); 850 CeedOperatorDestroy(&op_error); 851 CeedElemRestrictionDestroy(&Erestrictu); 852 CeedElemRestrictionDestroy(&Erestrictx); 853 CeedElemRestrictionDestroy(&Erestrictui); 854 CeedElemRestrictionDestroy(&Erestrictxi); 855 CeedElemRestrictionDestroy(&Erestrictqdi); 856 CeedQFunctionDestroy(&qf_setup); 857 CeedQFunctionDestroy(&qf_apply); 858 CeedQFunctionDestroy(&qf_error); 859 CeedBasisDestroy(&basisu); 860 CeedBasisDestroy(&basisx); 861 CeedDestroy(&ceed); 862 ierr = PetscFree(user); CHKERRQ(ierr); 863 return PetscFinalize(); 864 } 865