1 2 #include <../src/snes/impls/vi/ss/vissimpl.h> /*I "petscsnes.h" I*/ 3 4 /*@ 5 SNESVIComputeMeritFunction - Evaluates the merit function for the mixed complementarity problem. 6 7 Input Parameter: 8 . phi - the `Vec` holding the evaluation of the semismooth function 9 10 Output Parameters: 11 + merit - the merit function 1/2 ||phi||^2 12 - phinorm - the two-norm of the vector, ||phi|| 13 14 Level: developer 15 16 .seealso: `SNESVINEWTONSSLS`, `SNESVIComputeFunction()` 17 @*/ 18 PetscErrorCode SNESVIComputeMeritFunction(Vec phi, PetscReal *merit, PetscReal *phinorm) 19 { 20 PetscFunctionBegin; 21 PetscCall(VecNormBegin(phi, NORM_2, phinorm)); 22 PetscCall(VecNormEnd(phi, NORM_2, phinorm)); 23 *merit = 0.5 * (*phinorm) * (*phinorm); 24 PetscFunctionReturn(PETSC_SUCCESS); 25 } 26 27 static inline PetscScalar Phi(PetscScalar a, PetscScalar b) 28 { 29 return a + b - PetscSqrtScalar(a * a + b * b); 30 } 31 32 static inline PetscScalar DPhi(PetscScalar a, PetscScalar b) 33 { 34 if ((PetscAbsScalar(a) >= 1.e-6) || (PetscAbsScalar(b) >= 1.e-6)) return 1.0 - a / PetscSqrtScalar(a * a + b * b); 35 else return .5; 36 } 37 38 /*@ 39 SNESVIComputeFunction - Provides the function that reformulates a system of nonlinear equations in mixed complementarity form to a system of nonlinear 40 equations in semismooth form. 41 42 Input Parameters: 43 + snes - the SNES context 44 . X - current iterate 45 - functx - user defined function context 46 47 Output Parameter: 48 . phi - the evaluation of Semismooth function at X 49 50 Level: developer 51 52 .seealso: `SNESVINEWTONSSLS`, `SNESVIComputeMeritFunction()` 53 @*/ 54 PetscErrorCode SNESVIComputeFunction(SNES snes, Vec X, Vec phi, void *functx) 55 { 56 SNES_VINEWTONSSLS *vi = (SNES_VINEWTONSSLS *)snes->data; 57 Vec Xl = snes->xl, Xu = snes->xu, F = snes->vec_func; 58 PetscScalar *phi_arr, *f_arr, *l, *u; 59 const PetscScalar *x_arr; 60 PetscInt i, nlocal; 61 62 PetscFunctionBegin; 63 PetscCall((*vi->computeuserfunction)(snes, X, F, functx)); 64 PetscCall(VecGetLocalSize(X, &nlocal)); 65 PetscCall(VecGetArrayRead(X, &x_arr)); 66 PetscCall(VecGetArray(F, &f_arr)); 67 PetscCall(VecGetArray(Xl, &l)); 68 PetscCall(VecGetArray(Xu, &u)); 69 PetscCall(VecGetArray(phi, &phi_arr)); 70 71 for (i = 0; i < nlocal; i++) { 72 if ((PetscRealPart(l[i]) <= PETSC_NINFINITY) && (PetscRealPart(u[i]) >= PETSC_INFINITY)) { /* no constraints on variable */ 73 phi_arr[i] = f_arr[i]; 74 } else if (PetscRealPart(l[i]) <= PETSC_NINFINITY) { /* upper bound on variable only */ 75 phi_arr[i] = -Phi(u[i] - x_arr[i], -f_arr[i]); 76 } else if (PetscRealPart(u[i]) >= PETSC_INFINITY) { /* lower bound on variable only */ 77 phi_arr[i] = Phi(x_arr[i] - l[i], f_arr[i]); 78 } else if (l[i] == u[i]) { 79 phi_arr[i] = l[i] - x_arr[i]; 80 } else { /* both bounds on variable */ 81 phi_arr[i] = Phi(x_arr[i] - l[i], -Phi(u[i] - x_arr[i], -f_arr[i])); 82 } 83 } 84 85 PetscCall(VecRestoreArrayRead(X, &x_arr)); 86 PetscCall(VecRestoreArray(F, &f_arr)); 87 PetscCall(VecRestoreArray(Xl, &l)); 88 PetscCall(VecRestoreArray(Xu, &u)); 89 PetscCall(VecRestoreArray(phi, &phi_arr)); 90 PetscFunctionReturn(PETSC_SUCCESS); 91 } 92 93 /* 94 SNESVIComputeBsubdifferentialVectors - Computes the diagonal shift (Da) and row scaling (Db) vectors needed for the 95 the semismooth jacobian. 96 */ 97 PetscErrorCode SNESVIComputeBsubdifferentialVectors(SNES snes, Vec X, Vec F, Mat jac, Vec Da, Vec Db) 98 { 99 PetscScalar *l, *u, *x, *f, *da, *db, da1, da2, db1, db2; 100 PetscInt i, nlocal; 101 102 PetscFunctionBegin; 103 PetscCall(VecGetArray(X, &x)); 104 PetscCall(VecGetArray(F, &f)); 105 PetscCall(VecGetArray(snes->xl, &l)); 106 PetscCall(VecGetArray(snes->xu, &u)); 107 PetscCall(VecGetArray(Da, &da)); 108 PetscCall(VecGetArray(Db, &db)); 109 PetscCall(VecGetLocalSize(X, &nlocal)); 110 111 for (i = 0; i < nlocal; i++) { 112 if ((PetscRealPart(l[i]) <= PETSC_NINFINITY) && (PetscRealPart(u[i]) >= PETSC_INFINITY)) { /* no constraints on variable */ 113 da[i] = 0; 114 db[i] = 1; 115 } else if (PetscRealPart(l[i]) <= PETSC_NINFINITY) { /* upper bound on variable only */ 116 da[i] = DPhi(u[i] - x[i], -f[i]); 117 db[i] = DPhi(-f[i], u[i] - x[i]); 118 } else if (PetscRealPart(u[i]) >= PETSC_INFINITY) { /* lower bound on variable only */ 119 da[i] = DPhi(x[i] - l[i], f[i]); 120 db[i] = DPhi(f[i], x[i] - l[i]); 121 } else if (l[i] == u[i]) { /* fixed variable */ 122 da[i] = 1; 123 db[i] = 0; 124 } else { /* upper and lower bounds on variable */ 125 da1 = DPhi(x[i] - l[i], -Phi(u[i] - x[i], -f[i])); 126 db1 = DPhi(-Phi(u[i] - x[i], -f[i]), x[i] - l[i]); 127 da2 = DPhi(u[i] - x[i], -f[i]); 128 db2 = DPhi(-f[i], u[i] - x[i]); 129 da[i] = da1 + db1 * da2; 130 db[i] = db1 * db2; 131 } 132 } 133 134 PetscCall(VecRestoreArray(X, &x)); 135 PetscCall(VecRestoreArray(F, &f)); 136 PetscCall(VecRestoreArray(snes->xl, &l)); 137 PetscCall(VecRestoreArray(snes->xu, &u)); 138 PetscCall(VecRestoreArray(Da, &da)); 139 PetscCall(VecRestoreArray(Db, &db)); 140 PetscFunctionReturn(PETSC_SUCCESS); 141 } 142 143 /* 144 SNESVIComputeJacobian - Computes the jacobian of the semismooth function.The Jacobian for the semismooth function is an element of the B-subdifferential of the Fischer-Burmeister function for complementarity problems. 145 146 Input Parameters: 147 . Da - Diagonal shift vector for the semismooth jacobian. 148 . Db - Row scaling vector for the semismooth jacobian. 149 150 Output Parameters: 151 . jac - semismooth jacobian 152 . jac_pre - optional preconditioning matrix 153 154 Note: 155 The semismooth jacobian matrix is given by 156 jac = Da + Db*jacfun 157 where Db is the row scaling matrix stored as a vector, 158 Da is the diagonal perturbation matrix stored as a vector 159 and jacfun is the jacobian of the original nonlinear function. 160 */ 161 PetscErrorCode SNESVIComputeJacobian(Mat jac, Mat jac_pre, Vec Da, Vec Db) 162 { 163 /* Do row scaling and add diagonal perturbation */ 164 PetscFunctionBegin; 165 PetscCall(MatDiagonalScale(jac, Db, NULL)); 166 PetscCall(MatDiagonalSet(jac, Da, ADD_VALUES)); 167 if (jac != jac_pre) { /* If jac and jac_pre are different */ 168 PetscCall(MatDiagonalScale(jac_pre, Db, NULL)); 169 PetscCall(MatDiagonalSet(jac_pre, Da, ADD_VALUES)); 170 } 171 PetscFunctionReturn(PETSC_SUCCESS); 172 } 173 174 /* 175 SNESVIComputeMeritFunctionGradient - Computes the gradient of the merit function psi. 176 177 Input Parameters: 178 phi - semismooth function. 179 H - semismooth jacobian 180 181 Output Parameter: 182 dpsi - merit function gradient 183 184 Note: 185 The merit function gradient is computed as follows 186 dpsi = H^T*phi 187 */ 188 PetscErrorCode SNESVIComputeMeritFunctionGradient(Mat H, Vec phi, Vec dpsi) 189 { 190 PetscFunctionBegin; 191 PetscCall(MatMultTranspose(H, phi, dpsi)); 192 PetscFunctionReturn(PETSC_SUCCESS); 193 } 194 195 /* 196 SNESSolve_VINEWTONSSLS - Solves the complementarity problem with a semismooth Newton 197 method using a line search. 198 199 Input Parameter: 200 . snes - the SNES context 201 202 Application Interface Routine: SNESSolve() 203 204 Note: 205 This implements essentially a semismooth Newton method with a 206 line search. The default line search does not do any line search 207 but rather takes a full Newton step. 208 209 Developer Note: the code in this file should be slightly modified so that this routine need not exist and the SNESSolve_NEWTONLS() routine is called directly with the appropriate wrapped function and Jacobian evaluations 210 211 */ 212 PetscErrorCode SNESSolve_VINEWTONSSLS(SNES snes) 213 { 214 SNES_VINEWTONSSLS *vi = (SNES_VINEWTONSSLS *)snes->data; 215 PetscInt maxits, i, lits; 216 SNESLineSearchReason lssucceed; 217 PetscReal gnorm, xnorm = 0, ynorm; 218 Vec Y, X, F; 219 KSPConvergedReason kspreason; 220 DM dm; 221 DMSNES sdm; 222 223 PetscFunctionBegin; 224 PetscCall(SNESGetDM(snes, &dm)); 225 PetscCall(DMGetDMSNES(dm, &sdm)); 226 227 vi->computeuserfunction = sdm->ops->computefunction; 228 sdm->ops->computefunction = SNESVIComputeFunction; 229 230 snes->numFailures = 0; 231 snes->numLinearSolveFailures = 0; 232 snes->reason = SNES_CONVERGED_ITERATING; 233 234 maxits = snes->max_its; /* maximum number of iterations */ 235 X = snes->vec_sol; /* solution vector */ 236 F = snes->vec_func; /* residual vector */ 237 Y = snes->work[0]; /* work vectors */ 238 239 PetscCall(PetscObjectSAWsTakeAccess((PetscObject)snes)); 240 snes->iter = 0; 241 snes->norm = 0.0; 242 PetscCall(PetscObjectSAWsGrantAccess((PetscObject)snes)); 243 244 PetscCall(SNESVIProjectOntoBounds(snes, X)); 245 PetscCall(SNESComputeFunction(snes, X, vi->phi)); 246 if (snes->domainerror) { 247 snes->reason = SNES_DIVERGED_FUNCTION_DOMAIN; 248 sdm->ops->computefunction = vi->computeuserfunction; 249 PetscFunctionReturn(PETSC_SUCCESS); 250 } 251 /* Compute Merit function */ 252 PetscCall(SNESVIComputeMeritFunction(vi->phi, &vi->merit, &vi->phinorm)); 253 254 PetscCall(VecNormBegin(X, NORM_2, &xnorm)); /* xnorm <- ||x|| */ 255 PetscCall(VecNormEnd(X, NORM_2, &xnorm)); 256 SNESCheckFunctionNorm(snes, vi->merit); 257 258 PetscCall(PetscObjectSAWsTakeAccess((PetscObject)snes)); 259 snes->norm = vi->phinorm; 260 PetscCall(PetscObjectSAWsGrantAccess((PetscObject)snes)); 261 PetscCall(SNESLogConvergenceHistory(snes, vi->phinorm, 0)); 262 263 /* test convergence */ 264 PetscCall(SNESConverged(snes, 0, 0.0, 0.0, vi->phinorm)); 265 PetscCall(SNESMonitor(snes, 0, vi->phinorm)); 266 if (snes->reason) { 267 sdm->ops->computefunction = vi->computeuserfunction; 268 PetscFunctionReturn(PETSC_SUCCESS); 269 } 270 271 for (i = 0; i < maxits; i++) { 272 /* Call general purpose update function */ 273 PetscTryTypeMethod(snes, update, snes->iter); 274 275 /* Solve J Y = Phi, where J is the semismooth jacobian */ 276 277 /* Get the jacobian -- note that the function must be the original function for snes_fd and snes_fd_color to work for this*/ 278 sdm->ops->computefunction = vi->computeuserfunction; 279 PetscCall(SNESComputeJacobian(snes, X, snes->jacobian, snes->jacobian_pre)); 280 SNESCheckJacobianDomainerror(snes); 281 sdm->ops->computefunction = SNESVIComputeFunction; 282 283 /* Get the diagonal shift and row scaling vectors */ 284 PetscCall(SNESVIComputeBsubdifferentialVectors(snes, X, F, snes->jacobian, vi->Da, vi->Db)); 285 /* Compute the semismooth jacobian */ 286 PetscCall(SNESVIComputeJacobian(snes->jacobian, snes->jacobian_pre, vi->Da, vi->Db)); 287 /* Compute the merit function gradient */ 288 PetscCall(SNESVIComputeMeritFunctionGradient(snes->jacobian, vi->phi, vi->dpsi)); 289 PetscCall(KSPSetOperators(snes->ksp, snes->jacobian, snes->jacobian_pre)); 290 PetscCall(KSPSolve(snes->ksp, vi->phi, Y)); 291 PetscCall(KSPGetConvergedReason(snes->ksp, &kspreason)); 292 293 if (kspreason < 0) { 294 if (++snes->numLinearSolveFailures >= snes->maxLinearSolveFailures) { 295 PetscCall(PetscInfo(snes, "iter=%" PetscInt_FMT ", number linear solve failures %" PetscInt_FMT " greater than current SNES allowed, stopping solve\n", snes->iter, snes->numLinearSolveFailures)); 296 snes->reason = SNES_DIVERGED_LINEAR_SOLVE; 297 break; 298 } 299 } 300 PetscCall(KSPGetIterationNumber(snes->ksp, &lits)); 301 snes->linear_its += lits; 302 PetscCall(PetscInfo(snes, "iter=%" PetscInt_FMT ", linear solve iterations=%" PetscInt_FMT "\n", snes->iter, lits)); 303 /* 304 if (snes->ops->precheck) { 305 PetscBool changed_y = PETSC_FALSE; 306 PetscUseTypeMethod(snes,precheck ,X,Y,snes->precheck,&changed_y); 307 } 308 309 if (PetscLogPrintInfo) PetscCall(SNESVICheckResidual_Private(snes,snes->jacobian,F,Y,G,W)); 310 */ 311 /* Compute a (scaled) negative update in the line search routine: 312 Y <- X - lambda*Y 313 and evaluate G = function(Y) (depends on the line search). 314 */ 315 PetscCall(VecCopy(Y, snes->vec_sol_update)); 316 ynorm = 1; 317 gnorm = vi->phinorm; 318 PetscCall(SNESLineSearchApply(snes->linesearch, X, vi->phi, &gnorm, Y)); 319 PetscCall(SNESLineSearchGetReason(snes->linesearch, &lssucceed)); 320 PetscCall(SNESLineSearchGetNorms(snes->linesearch, &xnorm, &gnorm, &ynorm)); 321 PetscCall(PetscInfo(snes, "fnorm=%18.16e, gnorm=%18.16e, ynorm=%18.16e, lssucceed=%d\n", (double)vi->phinorm, (double)gnorm, (double)ynorm, (int)lssucceed)); 322 if (snes->reason == SNES_DIVERGED_FUNCTION_COUNT) break; 323 if (snes->domainerror) { 324 snes->reason = SNES_DIVERGED_FUNCTION_DOMAIN; 325 sdm->ops->computefunction = vi->computeuserfunction; 326 PetscFunctionReturn(PETSC_SUCCESS); 327 } 328 if (lssucceed) { 329 if (++snes->numFailures >= snes->maxFailures) { 330 PetscBool ismin; 331 snes->reason = SNES_DIVERGED_LINE_SEARCH; 332 PetscCall(SNESVICheckLocalMin_Private(snes, snes->jacobian, vi->phi, X, gnorm, &ismin)); 333 if (ismin) snes->reason = SNES_DIVERGED_LOCAL_MIN; 334 break; 335 } 336 } 337 /* Update function and solution vectors */ 338 vi->phinorm = gnorm; 339 vi->merit = 0.5 * vi->phinorm * vi->phinorm; 340 /* Monitor convergence */ 341 PetscCall(PetscObjectSAWsTakeAccess((PetscObject)snes)); 342 snes->iter = i + 1; 343 snes->norm = vi->phinorm; 344 snes->xnorm = xnorm; 345 snes->ynorm = ynorm; 346 PetscCall(PetscObjectSAWsGrantAccess((PetscObject)snes)); 347 PetscCall(SNESLogConvergenceHistory(snes, snes->norm, lits)); 348 /* Test for convergence, xnorm = || X || */ 349 if (snes->ops->converged != SNESConvergedSkip) PetscCall(VecNorm(X, NORM_2, &xnorm)); 350 PetscCall(SNESConverged(snes, snes->iter, xnorm, ynorm, vi->phinorm)); 351 PetscCall(SNESMonitor(snes, snes->iter, snes->norm)); 352 if (snes->reason) break; 353 } 354 sdm->ops->computefunction = vi->computeuserfunction; 355 PetscFunctionReturn(PETSC_SUCCESS); 356 } 357 358 /* 359 SNESSetUp_VINEWTONSSLS - Sets up the internal data structures for the later use 360 of the SNES nonlinear solver. 361 362 Input Parameter: 363 . snes - the SNES context 364 365 Application Interface Routine: SNESSetUp() 366 367 Note: 368 For basic use of the SNES solvers, the user need not explicitly call 369 SNESSetUp(), since these actions will automatically occur during 370 the call to SNESSolve(). 371 */ 372 PetscErrorCode SNESSetUp_VINEWTONSSLS(SNES snes) 373 { 374 SNES_VINEWTONSSLS *vi = (SNES_VINEWTONSSLS *)snes->data; 375 376 PetscFunctionBegin; 377 PetscCall(SNESSetUp_VI(snes)); 378 PetscCall(VecDuplicate(snes->vec_sol, &vi->dpsi)); 379 PetscCall(VecDuplicate(snes->vec_sol, &vi->phi)); 380 PetscCall(VecDuplicate(snes->vec_sol, &vi->Da)); 381 PetscCall(VecDuplicate(snes->vec_sol, &vi->Db)); 382 PetscCall(VecDuplicate(snes->vec_sol, &vi->z)); 383 PetscCall(VecDuplicate(snes->vec_sol, &vi->t)); 384 PetscFunctionReturn(PETSC_SUCCESS); 385 } 386 387 PetscErrorCode SNESReset_VINEWTONSSLS(SNES snes) 388 { 389 SNES_VINEWTONSSLS *vi = (SNES_VINEWTONSSLS *)snes->data; 390 391 PetscFunctionBegin; 392 PetscCall(SNESReset_VI(snes)); 393 PetscCall(VecDestroy(&vi->dpsi)); 394 PetscCall(VecDestroy(&vi->phi)); 395 PetscCall(VecDestroy(&vi->Da)); 396 PetscCall(VecDestroy(&vi->Db)); 397 PetscCall(VecDestroy(&vi->z)); 398 PetscCall(VecDestroy(&vi->t)); 399 PetscFunctionReturn(PETSC_SUCCESS); 400 } 401 402 /* 403 SNESSetFromOptions_VINEWTONSSLS - Sets various parameters for the SNESVI method. 404 405 Input Parameter: 406 . snes - the SNES context 407 408 Application Interface Routine: SNESSetFromOptions() 409 */ 410 static PetscErrorCode SNESSetFromOptions_VINEWTONSSLS(SNES snes, PetscOptionItems *PetscOptionsObject) 411 { 412 PetscFunctionBegin; 413 PetscCall(SNESSetFromOptions_VI(snes, PetscOptionsObject)); 414 PetscOptionsHeadBegin(PetscOptionsObject, "SNES semismooth method options"); 415 PetscOptionsHeadEnd(); 416 PetscFunctionReturn(PETSC_SUCCESS); 417 } 418 419 /*MC 420 SNESVINEWTONSSLS - Semi-smooth solver for variational inequalities based on Newton's method 421 422 Options Database Keys: 423 + -snes_type <vinewtonssls,vinewtonrsls> a semi-smooth solver, a reduced space active set method 424 - -snes_vi_monitor - prints the number of active constraints at each iteration. 425 426 Level: beginner 427 428 References: 429 + * - T. S. Munson, F. Facchinei, M. C. Ferris, A. Fischer, and C. Kanzow. The semismooth 430 algorithm for large scale complementarity problems. INFORMS Journal on Computing, 13 (2001). 431 - * - T. S. Munson, and S. Benson. Flexible Complementarity Solvers for Large Scale 432 Applications, Optimization Methods and Software, 21 (2006). 433 434 Notes: 435 This family of algorithm is much like an interior point method. 436 437 The reduced space active set solvers `SNESVINEWTONRSLS` provide an alternative approach that does not result in extremely ill-conditioned linear systems 438 439 .seealso: `SNESVINEWTONRSLS`, `SNESVISetVariableBounds()`, `SNESVISetComputeVariableBounds()`, `SNESCreate()`, `SNES`, `SNESSetType()`, `SNESVINEWTONRSLS`, `SNESNEWTONTR`, `SNESLineSearchSetType()`, `SNESLineSearchSetPostCheck()`, `SNESLineSearchSetPreCheck()` 440 M*/ 441 PETSC_EXTERN PetscErrorCode SNESCreate_VINEWTONSSLS(SNES snes) 442 { 443 SNES_VINEWTONSSLS *vi; 444 SNESLineSearch linesearch; 445 446 PetscFunctionBegin; 447 snes->ops->reset = SNESReset_VINEWTONSSLS; 448 snes->ops->setup = SNESSetUp_VINEWTONSSLS; 449 snes->ops->solve = SNESSolve_VINEWTONSSLS; 450 snes->ops->destroy = SNESDestroy_VI; 451 snes->ops->setfromoptions = SNESSetFromOptions_VINEWTONSSLS; 452 snes->ops->view = NULL; 453 454 snes->usesksp = PETSC_TRUE; 455 snes->usesnpc = PETSC_FALSE; 456 457 PetscCall(SNESGetLineSearch(snes, &linesearch)); 458 if (!((PetscObject)linesearch)->type_name) { 459 PetscCall(SNESLineSearchSetType(linesearch, SNESLINESEARCHBT)); 460 PetscCall(SNESLineSearchBTSetAlpha(linesearch, 0.0)); 461 } 462 463 snes->alwayscomputesfinalresidual = PETSC_FALSE; 464 465 PetscCall(PetscNew(&vi)); 466 snes->data = (void *)vi; 467 468 PetscCall(PetscObjectComposeFunction((PetscObject)snes, "SNESVISetVariableBounds_C", SNESVISetVariableBounds_VI)); 469 PetscCall(PetscObjectComposeFunction((PetscObject)snes, "SNESVISetComputeVariableBounds_C", SNESVISetComputeVariableBounds_VI)); 470 PetscFunctionReturn(PETSC_SUCCESS); 471 } 472