#include <../src/snes/impls/tr/trimpl.h> /*I "petscsnes.h" I*/ typedef struct { SNES snes; PetscErrorCode (*convtest)(KSP, PetscInt, PetscReal, KSPConvergedReason *, void *); PetscErrorCode (*convdestroy)(void *); void *convctx; } SNES_TR_KSPConverged_Ctx; const char *const SNESNewtonTRFallbackTypes[] = {"NEWTON", "CAUCHY", "DOGLEG", "SNESNewtonTRFallbackType", "SNES_TR_FALLBACK_", NULL}; const char *const SNESNewtonTRQNTypes[] = {"NONE", "SAME", "DIFFERENT", "SNESNewtonTRQNType", "SNES_TR_QN_", NULL}; static PetscErrorCode SNESComputeJacobian_MATLMVM(SNES snes, Vec X, Mat J, Mat B, void *dummy) { PetscFunctionBegin; // PetscCall(MatLMVMSymBroydenSetDelta(B, _some_delta)); PetscCall(MatLMVMUpdate(B, X, snes->vec_func)); PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY)); PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY)); if (J != B) { // PetscCall(MatLMVMSymBroydenSetDelta(J, _some_delta)); PetscCall(MatLMVMUpdate(J, X, snes->vec_func)); PetscCall(MatAssemblyBegin(J, MAT_FINAL_ASSEMBLY)); PetscCall(MatAssemblyEnd(J, MAT_FINAL_ASSEMBLY)); } PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode SNESTR_KSPConverged_Private(KSP ksp, PetscInt n, PetscReal rnorm, KSPConvergedReason *reason, void *cctx) { SNES_TR_KSPConverged_Ctx *ctx = (SNES_TR_KSPConverged_Ctx *)cctx; SNES snes = ctx->snes; SNES_NEWTONTR *neP = (SNES_NEWTONTR *)snes->data; Vec x; PetscReal nrm; PetscFunctionBegin; /* Determine norm of solution */ PetscCall(KSPBuildSolution(ksp, NULL, &x)); PetscCall(VecNorm(x, neP->norm, &nrm)); if (nrm >= neP->delta) { PetscCall(PetscInfo(snes, "Ending linear iteration early due to exiting trust region, delta=%g, length=%g\n", (double)neP->delta, (double)nrm)); *reason = KSP_CONVERGED_STEP_LENGTH; PetscFunctionReturn(PETSC_SUCCESS); } PetscCall((*ctx->convtest)(ksp, n, rnorm, reason, ctx->convctx)); if (*reason) PetscCall(PetscInfo(snes, "Default or user provided convergence test KSP iterations=%" PetscInt_FMT ", rnorm=%g\n", n, (double)rnorm)); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode SNESTR_KSPConverged_Destroy(void *cctx) { SNES_TR_KSPConverged_Ctx *ctx = (SNES_TR_KSPConverged_Ctx *)cctx; PetscFunctionBegin; PetscCall((*ctx->convdestroy)(ctx->convctx)); PetscCall(PetscFree(ctx)); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode SNESTR_Converged_Private(SNES snes, PetscInt it, PetscReal xnorm, PetscReal pnorm, PetscReal fnorm, SNESConvergedReason *reason, void *dummy) { SNES_NEWTONTR *neP = (SNES_NEWTONTR *)snes->data; PetscFunctionBegin; *reason = SNES_CONVERGED_ITERATING; if (neP->delta < snes->deltatol) { PetscCall(PetscInfo(snes, "Diverged due to too small a trust region %g<%g\n", (double)neP->delta, (double)snes->deltatol)); *reason = SNES_DIVERGED_TR_DELTA; } else if (snes->nfuncs >= snes->max_funcs && snes->max_funcs >= 0) { PetscCall(PetscInfo(snes, "Exceeded maximum number of function evaluations: %" PetscInt_FMT "\n", snes->max_funcs)); *reason = SNES_DIVERGED_FUNCTION_COUNT; } PetscFunctionReturn(PETSC_SUCCESS); } /*@ SNESNewtonTRSetNormType - Specify the type of norm to use for the computation of the trust region. Input Parameters: + snes - the nonlinear solver object - norm - the norm type Level: intermediate .seealso: `SNESNEWTONTR`, `NormType` @*/ PetscErrorCode SNESNewtonTRSetNormType(SNES snes, NormType norm) { PetscBool flg; PetscFunctionBegin; PetscValidHeaderSpecific(snes, SNES_CLASSID, 1); PetscValidLogicalCollectiveEnum(snes, norm, 2); PetscCall(PetscObjectTypeCompare((PetscObject)snes, SNESNEWTONTR, &flg)); if (flg) { SNES_NEWTONTR *tr = (SNES_NEWTONTR *)snes->data; tr->norm = norm; } PetscFunctionReturn(PETSC_SUCCESS); } /*@ SNESNewtonTRSetQNType - Specify to use a quasi-Newton model. Input Parameters: + snes - the nonlinear solver object - use - the type of approximations to be used Level: intermediate Notes: Options for the approximations can be set with the snes_tr_qn_ and snes_tr_qn_pre_ prefixes. .seealso: `SNESNEWTONTR`, `SNESNewtonTRQNType`, `MATLMVM` @*/ PetscErrorCode SNESNewtonTRSetQNType(SNES snes, SNESNewtonTRQNType use) { PetscBool flg; PetscFunctionBegin; PetscValidHeaderSpecific(snes, SNES_CLASSID, 1); PetscValidLogicalCollectiveEnum(snes, use, 2); PetscCall(PetscObjectTypeCompare((PetscObject)snes, SNESNEWTONTR, &flg)); if (flg) { SNES_NEWTONTR *tr = (SNES_NEWTONTR *)snes->data; tr->qn = use; } PetscFunctionReturn(PETSC_SUCCESS); } /*@ SNESNewtonTRSetFallbackType - Set the type of fallback to use if the solution of the trust region subproblem is outside the radius Input Parameters: + snes - the nonlinear solver object - ftype - the fallback type, see `SNESNewtonTRFallbackType` Level: intermediate .seealso: [](ch_snes), `SNESNEWTONTR`, `SNESNewtonTRPreCheck()`, `SNESNewtonTRGetPreCheck()`, `SNESNewtonTRSetPreCheck()`, `SNESNewtonTRSetPostCheck()`, `SNESNewtonTRGetPostCheck()` @*/ PetscErrorCode SNESNewtonTRSetFallbackType(SNES snes, SNESNewtonTRFallbackType ftype) { SNES_NEWTONTR *tr = (SNES_NEWTONTR *)snes->data; PetscBool flg; PetscFunctionBegin; PetscValidHeaderSpecific(snes, SNES_CLASSID, 1); PetscValidLogicalCollectiveEnum(snes, ftype, 2); PetscCall(PetscObjectTypeCompare((PetscObject)snes, SNESNEWTONTR, &flg)); if (flg) tr->fallback = ftype; PetscFunctionReturn(PETSC_SUCCESS); } /*@C SNESNewtonTRSetPreCheck - Sets a user function that is called before the search step has been determined. Allows the user a chance to change or override the trust region decision. Logically Collective Input Parameters: + snes - the nonlinear solver object . func - [optional] function evaluation routine, for the calling sequence see `SNESNewtonTRPreCheck()` - ctx - [optional] user-defined context for private data for the function evaluation routine (may be `NULL`) Level: intermediate Note: This function is called BEFORE the function evaluation within the solver. .seealso: [](ch_snes), `SNESNEWTONTR`, `SNESNewtonTRPreCheck()`, `SNESNewtonTRGetPreCheck()`, `SNESNewtonTRSetPostCheck()`, `SNESNewtonTRGetPostCheck()`, @*/ PetscErrorCode SNESNewtonTRSetPreCheck(SNES snes, PetscErrorCode (*func)(SNES, Vec, Vec, PetscBool *, void *), void *ctx) { SNES_NEWTONTR *tr = (SNES_NEWTONTR *)snes->data; PetscBool flg; PetscFunctionBegin; PetscValidHeaderSpecific(snes, SNES_CLASSID, 1); PetscCall(PetscObjectTypeCompare((PetscObject)snes, SNESNEWTONTR, &flg)); if (flg) { if (func) tr->precheck = func; if (ctx) tr->precheckctx = ctx; } PetscFunctionReturn(PETSC_SUCCESS); } /*@C SNESNewtonTRGetPreCheck - Gets the pre-check function Not Collective Input Parameter: . snes - the nonlinear solver context Output Parameters: + func - [optional] function evaluation routine, for the calling sequence see `SNESNewtonTRPreCheck()` - ctx - [optional] user-defined context for private data for the function evaluation routine (may be `NULL`) Level: intermediate .seealso: [](ch_snes), `SNESNEWTONTR`, `SNESNewtonTRSetPreCheck()`, `SNESNewtonTRPreCheck()` @*/ PetscErrorCode SNESNewtonTRGetPreCheck(SNES snes, PetscErrorCode (**func)(SNES, Vec, Vec, PetscBool *, void *), void **ctx) { SNES_NEWTONTR *tr = (SNES_NEWTONTR *)snes->data; PetscBool flg; PetscFunctionBegin; PetscValidHeaderSpecific(snes, SNES_CLASSID, 1); PetscCall(PetscObjectTypeCompare((PetscObject)snes, SNESNEWTONTR, &flg)); PetscAssert(flg, PetscObjectComm((PetscObject)snes), PETSC_ERR_ARG_WRONG, "Not for type %s", ((PetscObject)snes)->type_name); if (func) *func = tr->precheck; if (ctx) *ctx = tr->precheckctx; PetscFunctionReturn(PETSC_SUCCESS); } /*@C SNESNewtonTRSetPostCheck - Sets a user function that is called after the search step has been determined but before the next function evaluation. Allows the user a chance to change or override the internal decision of the solver Logically Collective Input Parameters: + snes - the nonlinear solver object . func - [optional] function evaluation routine, for the calling sequence see `SNESNewtonTRPostCheck()` - ctx - [optional] user-defined context for private data for the function evaluation routine (may be `NULL`) Level: intermediate Note: This function is called BEFORE the function evaluation within the solver while the function set in `SNESLineSearchSetPostCheck()` is called AFTER the function evaluation. .seealso: [](ch_snes), `SNESNEWTONTR`, `SNESNewtonTRPostCheck()`, `SNESNewtonTRGetPostCheck()`, `SNESNewtonTRSetPreCheck()`, `SNESNewtonTRGetPreCheck()` @*/ PetscErrorCode SNESNewtonTRSetPostCheck(SNES snes, PetscErrorCode (*func)(SNES, Vec, Vec, Vec, PetscBool *, PetscBool *, void *), void *ctx) { SNES_NEWTONTR *tr = (SNES_NEWTONTR *)snes->data; PetscBool flg; PetscFunctionBegin; PetscValidHeaderSpecific(snes, SNES_CLASSID, 1); PetscCall(PetscObjectTypeCompare((PetscObject)snes, SNESNEWTONTR, &flg)); if (flg) { if (func) tr->postcheck = func; if (ctx) tr->postcheckctx = ctx; } PetscFunctionReturn(PETSC_SUCCESS); } /*@C SNESNewtonTRGetPostCheck - Gets the post-check function Not Collective Input Parameter: . snes - the nonlinear solver context Output Parameters: + func - [optional] function evaluation routine, for the calling sequence see `SNESNewtonTRPostCheck()` - ctx - [optional] user-defined context for private data for the function evaluation routine (may be `NULL`) Level: intermediate .seealso: [](ch_snes), `SNESNEWTONTR`, `SNESNewtonTRSetPostCheck()`, `SNESNewtonTRPostCheck()` @*/ PetscErrorCode SNESNewtonTRGetPostCheck(SNES snes, PetscErrorCode (**func)(SNES, Vec, Vec, Vec, PetscBool *, PetscBool *, void *), void **ctx) { SNES_NEWTONTR *tr = (SNES_NEWTONTR *)snes->data; PetscBool flg; PetscFunctionBegin; PetscValidHeaderSpecific(snes, SNES_CLASSID, 1); PetscCall(PetscObjectTypeCompare((PetscObject)snes, SNESNEWTONTR, &flg)); PetscAssert(flg, PetscObjectComm((PetscObject)snes), PETSC_ERR_ARG_WRONG, "Not for type %s", ((PetscObject)snes)->type_name); if (func) *func = tr->postcheck; if (ctx) *ctx = tr->postcheckctx; PetscFunctionReturn(PETSC_SUCCESS); } /*@C SNESNewtonTRPreCheck - Runs the precheck routine Logically Collective Input Parameters: + snes - the solver . X - The last solution - Y - The step direction Output Parameter: . changed_Y - Indicator that the step direction `Y` has been changed. Level: intermediate .seealso: [](ch_snes), `SNESNEWTONTR`, `SNESNewtonTRSetPreCheck()`, `SNESNewtonTRGetPreCheck()`, `SNESNewtonTRPostCheck()` @*/ PetscErrorCode SNESNewtonTRPreCheck(SNES snes, Vec X, Vec Y, PetscBool *changed_Y) { SNES_NEWTONTR *tr = (SNES_NEWTONTR *)snes->data; PetscBool flg; PetscFunctionBegin; PetscValidHeaderSpecific(snes, SNES_CLASSID, 1); PetscCall(PetscObjectTypeCompare((PetscObject)snes, SNESNEWTONTR, &flg)); PetscAssert(flg, PetscObjectComm((PetscObject)snes), PETSC_ERR_ARG_WRONG, "Not for type %s", ((PetscObject)snes)->type_name); *changed_Y = PETSC_FALSE; if (tr->precheck) { PetscCall((*tr->precheck)(snes, X, Y, changed_Y, tr->precheckctx)); PetscValidLogicalCollectiveBool(snes, *changed_Y, 4); } PetscFunctionReturn(PETSC_SUCCESS); } /*@C SNESNewtonTRPostCheck - Runs the postcheck routine Logically Collective Input Parameters: + snes - the solver . X - The last solution . Y - The full step direction - W - The updated solution, W = X - Y Output Parameters: + changed_Y - indicator if step has been changed - changed_W - Indicator if the new candidate solution W has been changed. Note: If Y is changed then W is recomputed as X - Y Level: intermediate .seealso: [](ch_snes), `SNESNEWTONTR`, `SNESNewtonTRSetPostCheck()`, `SNESNewtonTRGetPostCheck()`, `SNESNewtonTRPreCheck()` @*/ PetscErrorCode SNESNewtonTRPostCheck(SNES snes, Vec X, Vec Y, Vec W, PetscBool *changed_Y, PetscBool *changed_W) { SNES_NEWTONTR *tr = (SNES_NEWTONTR *)snes->data; PetscBool flg; PetscFunctionBegin; PetscValidHeaderSpecific(snes, SNES_CLASSID, 1); PetscCall(PetscObjectTypeCompare((PetscObject)snes, SNESNEWTONTR, &flg)); PetscAssert(flg, PetscObjectComm((PetscObject)snes), PETSC_ERR_ARG_WRONG, "Not for type %s", ((PetscObject)snes)->type_name); *changed_Y = PETSC_FALSE; *changed_W = PETSC_FALSE; if (tr->postcheck) { PetscCall((*tr->postcheck)(snes, X, Y, W, changed_Y, changed_W, tr->postcheckctx)); PetscValidLogicalCollectiveBool(snes, *changed_Y, 5); PetscValidLogicalCollectiveBool(snes, *changed_W, 6); } PetscFunctionReturn(PETSC_SUCCESS); } /* stable implementation of roots of a*x^2 + b*x + c = 0 */ static inline void PetscQuadraticRoots(PetscReal a, PetscReal b, PetscReal c, PetscReal *xm, PetscReal *xp) { PetscReal temp = -0.5 * (b + PetscCopysignReal(1.0, b) * PetscSqrtReal(b * b - 4 * a * c)); PetscReal x1 = temp / a; PetscReal x2 = c / temp; *xm = PetscMin(x1, x2); *xp = PetscMax(x1, x2); } /* Computes the quadratic model difference */ static PetscErrorCode SNESNewtonTRQuadraticDelta(SNES snes, Mat J, PetscBool has_objective, Vec Y, Vec GradF, Vec W, PetscReal *yTHy_, PetscReal *gTy_, PetscReal *deltaqm) { PetscReal yTHy, gTy; PetscFunctionBegin; PetscCall(MatMult(J, Y, W)); if (has_objective) PetscCall(VecDotRealPart(Y, W, &yTHy)); else PetscCall(VecDotRealPart(W, W, &yTHy)); /* Gauss-Newton approximation J^t * J */ PetscCall(VecDotRealPart(GradF, Y, &gTy)); *deltaqm = -(-(gTy) + 0.5 * (yTHy)); /* difference in quadratic model, -gTy because SNES solves it this way */ if (yTHy_) *yTHy_ = yTHy; if (gTy_) *gTy_ = gTy; PetscFunctionReturn(PETSC_SUCCESS); } /* Computes the new objective given X = Xk, Y = direction W work vector, on output W = X - Y G work vector, on output G = SNESFunction(W) */ static PetscErrorCode SNESNewtonTRObjective(SNES snes, PetscBool has_objective, Vec X, Vec Y, Vec W, Vec G, PetscReal *gnorm, PetscReal *fkp1) { PetscBool changed_y, changed_w; PetscFunctionBegin; /* TODO: we can add a linesearch here */ PetscCall(SNESNewtonTRPreCheck(snes, X, Y, &changed_y)); PetscCall(VecWAXPY(W, -1.0, Y, X)); /* Xkp1 */ PetscCall(SNESNewtonTRPostCheck(snes, X, Y, W, &changed_y, &changed_w)); if (changed_y && !changed_w) PetscCall(VecWAXPY(W, -1.0, Y, X)); PetscCall(SNESComputeFunction(snes, W, G)); /* F(Xkp1) = G */ PetscCall(VecNorm(G, NORM_2, gnorm)); if (has_objective) PetscCall(SNESComputeObjective(snes, W, fkp1)); else *fkp1 = 0.5 * PetscSqr(*gnorm); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode SNESSetUpQN_NEWTONTR(SNES snes) { SNES_NEWTONTR *tr = (SNES_NEWTONTR *)snes->data; PetscFunctionBegin; PetscCall(MatDestroy(&tr->qnB)); PetscCall(MatDestroy(&tr->qnB_pre)); if (tr->qn) { PetscInt n, N; const char *optionsprefix; Mat B; PetscCall(MatCreate(PetscObjectComm((PetscObject)snes), &B)); PetscCall(SNESGetOptionsPrefix(snes, &optionsprefix)); PetscCall(MatSetOptionsPrefix(B, "snes_tr_qn_")); PetscCall(MatAppendOptionsPrefix(B, optionsprefix)); PetscCall(MatSetType(B, MATLMVMBFGS)); PetscCall(VecGetLocalSize(snes->vec_sol, &n)); PetscCall(VecGetSize(snes->vec_sol, &N)); PetscCall(MatSetSizes(B, n, n, N, N)); PetscCall(MatSetUp(B)); PetscCall(MatSetFromOptions(B)); PetscCall(MatLMVMAllocate(B, snes->vec_sol, snes->vec_func)); tr->qnB = B; if (tr->qn == SNES_TR_QN_DIFFERENT) { PetscCall(MatCreate(PetscObjectComm((PetscObject)snes), &B)); PetscCall(SNESGetOptionsPrefix(snes, &optionsprefix)); PetscCall(MatSetOptionsPrefix(B, "snes_tr_qn_pre_")); PetscCall(MatAppendOptionsPrefix(B, optionsprefix)); PetscCall(MatSetType(B, MATLMVMBFGS)); PetscCall(MatSetSizes(B, n, n, N, N)); PetscCall(MatSetUp(B)); PetscCall(MatSetFromOptions(B)); PetscCall(MatLMVMAllocate(B, snes->vec_sol, snes->vec_func)); tr->qnB_pre = B; } else { PetscCall(PetscObjectReference((PetscObject)tr->qnB)); tr->qnB_pre = tr->qnB; } } PetscFunctionReturn(PETSC_SUCCESS); } /* SNESSolve_NEWTONTR - Implements Newton's Method with trust-region subproblem and adds dogleg Cauchy (Steepest Descent direction) step and direction if the trust region is not satisfied for solving system of nonlinear equations */ static PetscErrorCode SNESSolve_NEWTONTR(SNES snes) { SNES_NEWTONTR *neP = (SNES_NEWTONTR *)snes->data; Vec X, F, Y, G, W, GradF, YU, Yc; PetscInt maxits, lits; PetscReal rho, fnorm, gnorm = 0.0, xnorm = 0.0, delta, ynorm; PetscReal deltaM, fk, fkp1, deltaqm = 0.0, gTy = 0.0, yTHy = 0.0; PetscReal auk, tauk, gfnorm, gfnorm_k, ycnorm, gTBg, objmin = 0.0, beta_k = 1.0; PC pc; Mat J, Jp; PetscBool already_done = PETSC_FALSE, on_boundary; PetscBool clear_converged_test, rho_satisfied, has_objective; SNES_TR_KSPConverged_Ctx *ctx; void *convctx; PetscErrorCode (*convtest)(KSP, PetscInt, PetscReal, KSPConvergedReason *, void *), (*convdestroy)(void *); PetscErrorCode (*objective)(SNES, Vec, PetscReal *, void *); PetscFunctionBegin; PetscCall(SNESGetObjective(snes, &objective, NULL)); has_objective = objective ? PETSC_TRUE : PETSC_FALSE; maxits = snes->max_its; /* maximum number of iterations */ X = snes->vec_sol; /* solution vector */ F = snes->vec_func; /* residual vector */ Y = snes->vec_sol_update; /* update vector */ G = snes->work[0]; /* updated residual */ W = snes->work[1]; /* temporary vector */ GradF = !has_objective ? snes->work[2] : snes->vec_func; /* grad f = J^T F */ YU = snes->work[3]; /* work vector for dogleg method */ Yc = snes->work[4]; /* Cauchy point */ PetscCheck(!snes->xl && !snes->xu && !snes->ops->computevariablebounds, PetscObjectComm((PetscObject)snes), PETSC_ERR_ARG_WRONGSTATE, "SNES solver %s does not support bounds", ((PetscObject)snes)->type_name); PetscCall(PetscObjectSAWsTakeAccess((PetscObject)snes)); snes->iter = 0; PetscCall(PetscObjectSAWsGrantAccess((PetscObject)snes)); /* setup QN matrices if needed */ PetscCall(SNESSetUpQN_NEWTONTR(snes)); /* Set the linear stopping criteria to use the More' trick if needed */ clear_converged_test = PETSC_FALSE; PetscCall(SNESGetKSP(snes, &snes->ksp)); PetscCall(KSPGetConvergenceTest(snes->ksp, &convtest, &convctx, &convdestroy)); if (convtest != SNESTR_KSPConverged_Private) { clear_converged_test = PETSC_TRUE; PetscCall(PetscNew(&ctx)); ctx->snes = snes; PetscCall(KSPGetAndClearConvergenceTest(snes->ksp, &ctx->convtest, &ctx->convctx, &ctx->convdestroy)); PetscCall(KSPSetConvergenceTest(snes->ksp, SNESTR_KSPConverged_Private, ctx, SNESTR_KSPConverged_Destroy)); PetscCall(PetscInfo(snes, "Using Krylov convergence test SNESTR_KSPConverged_Private\n")); } if (!snes->vec_func_init_set) { PetscCall(SNESComputeFunction(snes, X, F)); /* F(X) */ } else snes->vec_func_init_set = PETSC_FALSE; PetscCall(VecNorm(F, NORM_2, &fnorm)); /* fnorm <- || F || */ SNESCheckFunctionNorm(snes, fnorm); PetscCall(VecNorm(X, NORM_2, &xnorm)); /* xnorm <- || X || */ PetscCall(PetscObjectSAWsTakeAccess((PetscObject)snes)); snes->norm = fnorm; PetscCall(PetscObjectSAWsGrantAccess((PetscObject)snes)); delta = neP->delta0; deltaM = neP->deltaM; neP->delta = delta; PetscCall(SNESLogConvergenceHistory(snes, fnorm, 0)); /* test convergence */ rho_satisfied = PETSC_FALSE; PetscCall(SNESConverged(snes, 0, 0.0, 0.0, fnorm)); PetscCall(SNESMonitor(snes, 0, fnorm)); if (snes->reason) PetscFunctionReturn(PETSC_SUCCESS); if (has_objective) PetscCall(SNESComputeObjective(snes, X, &fk)); else fk = 0.5 * PetscSqr(fnorm); /* obj(x) = 0.5 * ||F(x)||^2 */ /* hook state vector to BFGS preconditioner */ PetscCall(KSPGetPC(snes->ksp, &pc)); PetscCall(PCLMVMSetUpdateVec(pc, X)); if (neP->kmdc) PetscCall(KSPSetComputeEigenvalues(snes->ksp, PETSC_TRUE)); while (snes->iter < maxits) { /* calculating Jacobian and GradF of minimization function only once */ if (!already_done) { /* Call general purpose update function */ PetscTryTypeMethod(snes, update, snes->iter); /* apply the nonlinear preconditioner */ if (snes->npc && snes->npcside == PC_RIGHT) { SNESConvergedReason reason; PetscCall(SNESSetInitialFunction(snes->npc, F)); PetscCall(PetscLogEventBegin(SNES_NPCSolve, snes->npc, X, snes->vec_rhs, 0)); PetscCall(SNESSolve(snes->npc, snes->vec_rhs, X)); PetscCall(PetscLogEventEnd(SNES_NPCSolve, snes->npc, X, snes->vec_rhs, 0)); PetscCall(SNESGetConvergedReason(snes->npc, &reason)); if (reason < 0 && reason != SNES_DIVERGED_MAX_IT && reason != SNES_DIVERGED_TR_DELTA) { snes->reason = SNES_DIVERGED_INNER; PetscFunctionReturn(PETSC_SUCCESS); } // XXX PetscCall(SNESGetNPCFunction(snes, F, &fnorm)); } else if (snes->ops->update) { /* if update is present, recompute objective function and function norm */ PetscCall(SNESComputeFunction(snes, X, F)); } /* Jacobian */ J = NULL; Jp = NULL; if (!neP->qnB) { PetscCall(SNESComputeJacobian(snes, X, snes->jacobian, snes->jacobian_pre)); J = snes->jacobian; Jp = snes->jacobian_pre; } else { /* QN model */ PetscCall(SNESComputeJacobian_MATLMVM(snes, X, neP->qnB, neP->qnB_pre, NULL)); J = neP->qnB; Jp = neP->qnB_pre; } SNESCheckJacobianDomainerror(snes); /* objective function */ PetscCall(VecNorm(F, NORM_2, &fnorm)); if (has_objective) PetscCall(SNESComputeObjective(snes, X, &fk)); else fk = 0.5 * PetscSqr(fnorm); /* obj(x) = 0.5 * ||F(x)||^2 */ /* GradF */ if (has_objective) gfnorm = fnorm; else { PetscCall(MatMultTranspose(J, F, GradF)); /* grad f = J^T F */ PetscCall(VecNorm(GradF, NORM_2, &gfnorm)); } PetscCall(VecNorm(GradF, neP->norm, &gfnorm_k)); } already_done = PETSC_TRUE; /* solve trust-region subproblem */ /* first compute Cauchy Point */ PetscCall(MatMult(J, GradF, W)); if (has_objective) PetscCall(VecDotRealPart(GradF, W, &gTBg)); else PetscCall(VecDotRealPart(W, W, &gTBg)); /* B = J^t * J */ /* Eqs 4.11 and 4.12 in Nocedal and Wright 2nd Edition (4.7 and 4.8 in 1st Edition) */ auk = delta / gfnorm_k; if (gTBg < 0.0) tauk = 1.0; else tauk = PetscMin(gfnorm * gfnorm * gfnorm_k / (delta * gTBg), 1); auk *= tauk; ycnorm = auk * gfnorm; PetscCall(VecAXPBY(Yc, auk, 0.0, GradF)); on_boundary = PETSC_FALSE; if (tauk != 1.0) { KSPConvergedReason reason; /* sufficient decrease (see 6.3.27 in Conn, Gould, Toint "Trust Region Methods") beta_k the largest eigenvalue of the Hessian. Here we use the previous estimated value */ objmin = -neP->kmdc * gnorm * PetscMin(gnorm / beta_k, delta); PetscCall(KSPCGSetObjectiveTarget(snes->ksp, objmin)); /* specify radius if looking for Newton step and trust region norm is the l2 norm */ PetscCall(KSPCGSetRadius(snes->ksp, neP->fallback == SNES_TR_FALLBACK_NEWTON && neP->norm == NORM_2 ? delta : 0.0)); PetscCall(KSPSetOperators(snes->ksp, J, Jp)); PetscCall(KSPSolve(snes->ksp, F, Y)); SNESCheckKSPSolve(snes); PetscCall(KSPGetIterationNumber(snes->ksp, &lits)); PetscCall(KSPGetConvergedReason(snes->ksp, &reason)); on_boundary = (PetscBool)(reason == KSP_CONVERGED_STEP_LENGTH); PetscCall(PetscInfo(snes, "iter=%" PetscInt_FMT ", linear solve iterations=%" PetscInt_FMT "\n", snes->iter, lits)); if (neP->kmdc) { /* update estimated Hessian largest eigenvalue */ PetscReal emax, emin; PetscCall(KSPComputeExtremeSingularValues(snes->ksp, &emax, &emin)); if (emax > 0.0) beta_k = emax + 1; } } else { /* Cauchy point is on the boundary, accept it */ on_boundary = PETSC_TRUE; PetscCall(VecCopy(Yc, Y)); PetscCall(PetscInfo(snes, "CP evaluated on boundary. delta: %g, ycnorm: %g, gTBg: %g\n", (double)delta, (double)ycnorm, (double)gTBg)); } PetscCall(VecNorm(Y, neP->norm, &ynorm)); /* decide what to do when the update is outside of trust region */ if (ynorm > delta || ynorm == 0.0) { SNESNewtonTRFallbackType fallback = ynorm > 0.0 ? neP->fallback : SNES_TR_FALLBACK_CAUCHY; PetscCheck(neP->norm == NORM_2 || fallback != SNES_TR_FALLBACK_DOGLEG, PetscObjectComm((PetscObject)snes), PETSC_ERR_SUP, "DOGLEG without l2 norm not implemented"); switch (fallback) { case SNES_TR_FALLBACK_NEWTON: auk = delta / ynorm; PetscCall(VecScale(Y, auk)); PetscCall(PetscInfo(snes, "SN evaluated. delta: %g, ynorm: %g\n", (double)delta, (double)ynorm)); break; case SNES_TR_FALLBACK_CAUCHY: case SNES_TR_FALLBACK_DOGLEG: if (fallback == SNES_TR_FALLBACK_CAUCHY || gTBg <= 0.0) { PetscCall(VecCopy(Yc, Y)); PetscCall(PetscInfo(snes, "CP evaluated. delta: %g, ynorm: %g, ycnorm: %g, gTBg: %g\n", (double)delta, (double)ynorm, (double)ycnorm, (double)gTBg)); } else { /* take linear combination of Cauchy and Newton direction and step */ auk = gfnorm * gfnorm / gTBg; if (gfnorm_k * auk >= delta) { /* first leg: Cauchy point outside of trust region */ PetscCall(VecAXPBY(Y, delta / gfnorm_k, 0.0, GradF)); PetscCall(PetscInfo(snes, "CP evaluated (outside region). delta: %g, ynorm: %g, ycnorm: %g\n", (double)delta, (double)ynorm, (double)ycnorm)); } else { /* second leg */ PetscReal c0, c1, c2, tau = 0.0, tpos, tneg; PetscBool noroots; /* Find solutions of (Eq. 4.16 in Nocedal and Wright) ||p_U + lambda * (p_B - p_U)||^2 - delta^2 = 0, where p_U the Cauchy direction, p_B the Newton direction */ PetscCall(VecAXPBY(YU, auk, 0.0, GradF)); PetscCall(VecAXPY(Y, -1.0, YU)); PetscCall(VecNorm(Y, NORM_2, &c0)); PetscCall(VecDotRealPart(YU, Y, &c1)); c0 = PetscSqr(c0); c2 = PetscSqr(ycnorm) - PetscSqr(delta); PetscQuadraticRoots(c0, 2 * c1, c2, &tneg, &tpos); /* In principle the DL strategy as a unique solution in [0,1] here we check that for some reason we numerically failed to compute it. In that case, we use the Cauchy point */ noroots = PetscIsInfOrNanReal(tneg); if (!noroots) { if (tpos > 1) { if (tneg >= 0 && tneg <= 1) { tau = tneg; } else noroots = PETSC_TRUE; } else if (tpos >= 0) { tau = tpos; } else noroots = PETSC_TRUE; } if (noroots) { /* No roots, select Cauchy point */ PetscCall(VecCopy(Yc, Y)); } else { PetscCall(VecAXPBY(Y, 1.0, tau, YU)); } PetscCall(PetscInfo(snes, "%s evaluated. roots: (%g, %g), tau %g, ynorm: %g, ycnorm: %g, gTBg: %g\n", noroots ? "CP" : "DL", (double)tneg, (double)tpos, (double)tau, (double)ynorm, (double)ycnorm, (double)gTBg)); } } break; default: SETERRQ(PetscObjectComm((PetscObject)snes), PETSC_ERR_SUP, "Unknown fallback mode"); break; } } /* compute the quadratic model difference */ PetscCall(SNESNewtonTRQuadraticDelta(snes, J, has_objective, Y, GradF, W, &yTHy, &gTy, &deltaqm)); /* Compute new objective function */ PetscCall(SNESNewtonTRObjective(snes, has_objective, X, Y, W, G, &gnorm, &fkp1)); if (PetscIsInfOrNanReal(fkp1)) rho = neP->eta1; else { if (deltaqm > 0.0) rho = (fk - fkp1) / deltaqm; /* actual improvement over predicted improvement */ else rho = neP->eta1; /* no reduction in quadratic model, step must be rejected */ } PetscCall(VecNorm(Y, neP->norm, &ynorm)); PetscCall(PetscInfo(snes, "rho=%g, delta=%g, fk=%g, fkp1=%g, deltaqm=%g, gTy=%g, yTHy=%g, ynormk=%g\n", (double)rho, (double)delta, (double)fk, (double)fkp1, (double)deltaqm, (double)gTy, (double)yTHy, (double)ynorm)); /* update the size of the trust region */ if (rho < neP->eta2) delta *= neP->t1; /* shrink the region */ else if (rho > neP->eta3 && on_boundary) delta *= neP->t2; /* expand the region */ delta = PetscMin(delta, deltaM); /* but not greater than deltaM */ /* log 2-norm of update for moniroting routines */ PetscCall(VecNorm(Y, NORM_2, &ynorm)); /* decide on new step */ neP->delta = delta; if (rho > neP->eta1) { rho_satisfied = PETSC_TRUE; } else { rho_satisfied = PETSC_FALSE; PetscCall(PetscInfo(snes, "Trying again in smaller region\n")); /* check to see if progress is hopeless */ PetscCall(SNESTR_Converged_Private(snes, snes->iter, xnorm, ynorm, fnorm, &snes->reason, snes->cnvP)); if (!snes->reason) PetscCall(SNESConverged(snes, snes->iter, xnorm, ynorm, fnorm)); if (snes->reason == SNES_CONVERGED_SNORM_RELATIVE) snes->reason = SNES_DIVERGED_TR_DELTA; snes->numFailures++; /* We're not progressing, so return with the current iterate */ if (snes->reason) break; } if (rho_satisfied) { /* Update function values */ already_done = PETSC_FALSE; fnorm = gnorm; fk = fkp1; /* New residual and linearization point */ PetscCall(VecCopy(G, F)); PetscCall(VecCopy(W, X)); /* Monitor convergence */ PetscCall(PetscObjectSAWsTakeAccess((PetscObject)snes)); snes->iter++; snes->norm = fnorm; snes->xnorm = xnorm; snes->ynorm = ynorm; PetscCall(PetscObjectSAWsGrantAccess((PetscObject)snes)); PetscCall(SNESLogConvergenceHistory(snes, snes->norm, lits)); /* Test for convergence, xnorm = || X || */ PetscCall(VecNorm(X, NORM_2, &xnorm)); PetscCall(SNESConverged(snes, snes->iter, xnorm, ynorm, fnorm)); PetscCall(SNESMonitor(snes, snes->iter, snes->norm)); if (snes->reason) break; } } if (clear_converged_test) { PetscCall(KSPGetAndClearConvergenceTest(snes->ksp, &ctx->convtest, &ctx->convctx, &ctx->convdestroy)); PetscCall(PetscFree(ctx)); PetscCall(KSPSetConvergenceTest(snes->ksp, convtest, convctx, convdestroy)); } PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode SNESSetUp_NEWTONTR(SNES snes) { PetscFunctionBegin; PetscCall(SNESSetWorkVecs(snes, 5)); PetscCall(SNESSetUpMatrices(snes)); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode SNESReset_NEWTONTR(SNES snes) { SNES_NEWTONTR *tr = (SNES_NEWTONTR *)snes->data; PetscFunctionBegin; PetscCall(MatDestroy(&tr->qnB)); PetscCall(MatDestroy(&tr->qnB_pre)); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode SNESDestroy_NEWTONTR(SNES snes) { PetscFunctionBegin; PetscCall(SNESReset_NEWTONTR(snes)); PetscCall(PetscFree(snes->data)); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode SNESSetFromOptions_NEWTONTR(SNES snes, PetscOptionItems *PetscOptionsObject) { SNES_NEWTONTR *ctx = (SNES_NEWTONTR *)snes->data; SNESNewtonTRQNType qn; SNESNewtonTRFallbackType fallback; NormType norm; PetscReal deltatol; PetscBool flg; PetscFunctionBegin; PetscOptionsHeadBegin(PetscOptionsObject, "SNES trust region options for nonlinear equations"); PetscCall(PetscOptionsReal("-snes_tr_eta1", "eta1", "None", ctx->eta1, &ctx->eta1, NULL)); PetscCall(PetscOptionsReal("-snes_tr_eta2", "eta2", "None", ctx->eta2, &ctx->eta2, NULL)); PetscCall(PetscOptionsReal("-snes_tr_eta3", "eta3", "None", ctx->eta3, &ctx->eta3, NULL)); PetscCall(PetscOptionsReal("-snes_tr_t1", "t1", "None", ctx->t1, &ctx->t1, NULL)); PetscCall(PetscOptionsReal("-snes_tr_t2", "t2", "None", ctx->t2, &ctx->t2, NULL)); PetscCall(PetscOptionsReal("-snes_tr_deltaM", "deltaM", "None", ctx->deltaM, &ctx->deltaM, NULL)); PetscCall(PetscOptionsReal("-snes_tr_delta0", "delta0", "None", ctx->delta0, &ctx->delta0, NULL)); PetscCall(PetscOptionsReal("-snes_tr_kmdc", "sufficient decrease parameter", "None", ctx->kmdc, &ctx->kmdc, NULL)); deltatol = snes->deltatol; PetscCall(PetscOptionsReal("-snes_tr_tol", "Trust region tolerance", "SNESSetTrustRegionTolerance", deltatol, &deltatol, &flg)); if (flg) PetscCall(SNESSetTrustRegionTolerance(snes, deltatol)); fallback = ctx->fallback; PetscCall(PetscOptionsEnum("-snes_tr_fallback_type", "Type of fallback if subproblem solution is outside of the trust region", "SNESNewtonTRSetFallbackType", SNESNewtonTRFallbackTypes, (PetscEnum)fallback, (PetscEnum *)&fallback, &flg)); if (flg) PetscCall(SNESNewtonTRSetFallbackType(snes, fallback)); qn = ctx->qn; PetscCall(PetscOptionsEnum("-snes_tr_qn", "Use Quasi-Newton approximations for the model", "SNESNewtonTRSetQNType", SNESNewtonTRQNTypes, (PetscEnum)qn, (PetscEnum *)&qn, &flg)); if (flg) PetscCall(SNESNewtonTRSetQNType(snes, qn)); norm = ctx->norm; PetscCall(PetscOptionsEnum("-snes_tr_norm_type", "Type of norm for trust region bounds", "SNESNewtonTRSetNormType", NormTypes, (PetscEnum)norm, (PetscEnum *)&norm, &flg)); if (flg) PetscCall(SNESNewtonTRSetNormType(snes, norm)); PetscOptionsHeadEnd(); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode SNESView_NEWTONTR(SNES snes, PetscViewer viewer) { SNES_NEWTONTR *tr = (SNES_NEWTONTR *)snes->data; PetscBool iascii; PetscFunctionBegin; PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii)); if (iascii) { PetscCall(PetscViewerASCIIPrintf(viewer, " Trust region tolerance %g\n", (double)snes->deltatol)); PetscCall(PetscViewerASCIIPrintf(viewer, " eta1=%g, eta2=%g, eta3=%g\n", (double)tr->eta1, (double)tr->eta2, (double)tr->eta3)); PetscCall(PetscViewerASCIIPrintf(viewer, " delta0=%g, t1=%g, t2=%g, deltaM=%g\n", (double)tr->delta0, (double)tr->t1, (double)tr->t2, (double)tr->deltaM)); PetscCall(PetscViewerASCIIPrintf(viewer, " kmdc=%g\n", (double)tr->kmdc)); PetscCall(PetscViewerASCIIPrintf(viewer, " fallback=%s\n", SNESNewtonTRFallbackTypes[tr->fallback])); if (tr->qn) PetscCall(PetscViewerASCIIPrintf(viewer, " qn=%s\n", SNESNewtonTRQNTypes[tr->qn])); if (tr->norm != NORM_2) PetscCall(PetscViewerASCIIPrintf(viewer, " norm=%s\n", NormTypes[tr->norm])); } PetscFunctionReturn(PETSC_SUCCESS); } /*MC SNESNEWTONTR - Newton based nonlinear solver that uses trust-region dogleg method with Cauchy direction {cite}`nocedal2006numerical` Options Database Keys: + -snes_tr_tol - trust region tolerance . -snes_tr_eta1 - trust region parameter eta1 <= eta2, rho > eta1 breaks out of the inner iteration (default: eta1=0.001) . -snes_tr_eta2 - trust region parameter, rho <= eta2 shrinks the trust region (default: eta2=0.25) . -snes_tr_eta3 - trust region parameter eta3 > eta2, rho >= eta3 expands the trust region (default: eta3=0.75) . -snes_tr_t1 - trust region parameter, shrinking factor of trust region (default: 0.25) . -snes_tr_t2 - trust region parameter, expanding factor of trust region (default: 2.0) . -snes_tr_deltaM - trust region parameter, max size of trust region (default: MAX_REAL) . -snes_tr_delta0 - trust region parameter, initial size of trust region (default: 0.2) - -snes_tr_fallback_type - Solution strategy to test reduction when step is outside of trust region. Can use scaled Newton direction, Cauchy point (Steepest Descent direction) or dogleg method. Level: beginner .seealso: [](ch_snes), `SNESCreate()`, `SNES`, `SNESSetType()`, `SNESNEWTONLS`, `SNESSetTrustRegionTolerance()`, `SNESNewtonTRPreCheck()`, `SNESNewtonTRGetPreCheck()`, `SNESNewtonTRSetPostCheck()`, `SNESNewtonTRGetPostCheck()`, `SNESNewtonTRSetPreCheck()`, `SNESNewtonTRSetFallbackType()`, `SNESNewtonTRSetQNType()` M*/ PETSC_EXTERN PetscErrorCode SNESCreate_NEWTONTR(SNES snes) { SNES_NEWTONTR *neP; PetscFunctionBegin; snes->ops->setup = SNESSetUp_NEWTONTR; snes->ops->solve = SNESSolve_NEWTONTR; snes->ops->reset = SNESReset_NEWTONTR; snes->ops->destroy = SNESDestroy_NEWTONTR; snes->ops->setfromoptions = SNESSetFromOptions_NEWTONTR; snes->ops->view = SNESView_NEWTONTR; snes->stol = 0.0; snes->usesksp = PETSC_TRUE; snes->npcside = PC_RIGHT; snes->usesnpc = PETSC_TRUE; snes->alwayscomputesfinalresidual = PETSC_TRUE; PetscCall(PetscNew(&neP)); snes->data = (void *)neP; neP->delta = 0.0; neP->delta0 = 0.2; neP->eta1 = 0.001; neP->eta2 = 0.25; neP->eta3 = 0.75; neP->t1 = 0.25; neP->t2 = 2.0; neP->deltaM = 1.e10; neP->norm = NORM_2; neP->fallback = SNES_TR_FALLBACK_NEWTON; neP->kmdc = 0.0; /* by default do not use sufficient decrease */ PetscFunctionReturn(PETSC_SUCCESS); }