#include typedef struct { PetscReal lambda; /* The default step length for the update */ Vec * dX; /* The change in X */ Vec * dF; /* The change in F */ PetscInt m; /* the number of kept previous steps */ PetscScalar * alpha; PetscScalar * beta; PetscScalar * rho; } QNContext; #undef __FUNCT__ #define __FUNCT__ "LBGFSApplyJinv_Private" PetscErrorCode LBGFSApplyJinv_Private(SNES snes, PetscInt it, Vec g, Vec z) { PetscErrorCode ierr; QNContext * qn = (QNContext *)snes->data; Vec * dX = qn->dX; Vec * dF = qn->dF; PetscScalar * alpha = qn->alpha; PetscScalar * beta = qn->beta; PetscScalar * rho = qn->rho; PetscInt k, i; PetscInt m = qn->m; PetscScalar t; PetscInt l = m; PetscFunctionBegin; if (it < m) l = it; ierr = VecCopy(g, z);CHKERRQ(ierr); /* outward recursion starting at iteration k's update and working back */ for (i = 0; i < l; i++) { k = (it - i - 1) % m; /* k = (it + i - l) % m; */ ierr = VecDot(dX[k], z, &t);CHKERRQ(ierr); alpha[k] = t*rho[k]; ierr = VecAXPY(z, -alpha[k], dF[k]);CHKERRQ(ierr); } /* inner application of the initial inverse jacobian approximation */ /* right now it's just the identity. Nothing needs to go here. */ /* inward recursion starting at the first update and working forward*/ for (i = 0; i < l; i++) { /* k = (it - i - 1) % m; */ k = (it + i - l) % m; ierr = VecDot(dF[k], z, &t);CHKERRQ(ierr); beta[k] = rho[k]*t; ierr = VecAXPY(z, (alpha[k] - beta[k]), dX[k]); } ierr = VecScale(z, 1.0);CHKERRQ(ierr); PetscFunctionReturn(0); } #undef __FUNCT__ #define __FUNCT__ "QNLineSearchQuadratic" PetscErrorCode QNLineSearchQuadratic(SNES snes,void *lsctx,Vec X,Vec F,Vec Y,PetscReal fnorm,PetscReal dummyXnorm,Vec G,Vec W,PetscReal *dummyYnorm,PetscReal *gnorm,PetscBool *flag) { PetscInt i; PetscReal alphas[3] = {0.0, 0.5, 1.0}; PetscReal norms[3]; PetscReal alpha,a,b; PetscErrorCode ierr; PetscFunctionBegin; norms[0] = fnorm; /* Calculate trial solutions */ for(i=1; i < 3; ++i) { /* Calculate X^{n+1} = (1 - \alpha) X^n + \alpha Y */ ierr = VecCopy(X, W);CHKERRQ(ierr); ierr = VecAXPBY(W, alphas[i], 1 - alphas[i], Y);CHKERRQ(ierr); ierr = SNESComputeFunction(snes, W, F);CHKERRQ(ierr); ierr = VecNorm(F, NORM_2, &norms[i]);CHKERRQ(ierr); } for(i = 0; i < 3; ++i) { norms[i] = PetscSqr(norms[i]); } /* Fit a quadratic: If we have x_{0,1,2} = 0, x_1, x_2 which generate norms y_{0,1,2} a = (x_1 y_2 - x_2 y_1 + (x_2 - x_1) y_0)/(x^2_2 x_1 - x_2 x^2_1) b = (x^2_1 y_2 - x^2_2 y_1 + (x^2_2 - x^2_1) y_0)/(x_2 x^2_1 - x^2_2 x_1) c = y_0 x_min = -b/2a If we let x_{0,1,2} = 0, 0.5, 1.0 a = 2 y_2 - 4 y_1 + 2 y_0 b = -y_2 + 4 y_1 - 3 y_0 c = y_0 */ a = (alphas[1]*norms[2] - alphas[2]*norms[1] + (alphas[2] - alphas[1])*norms[0])/(PetscSqr(alphas[2])*alphas[1] - alphas[2]*PetscSqr(alphas[1])); b = (PetscSqr(alphas[1])*norms[2] - PetscSqr(alphas[2])*norms[1] + (PetscSqr(alphas[2]) - PetscSqr(alphas[1]))*norms[0])/(alphas[2]*PetscSqr(alphas[1]) - PetscSqr(alphas[2])*alphas[1]); /* Check for positive a (concave up) */ if (a >= 0.0) { alpha = -b/(2.0*a); alpha = PetscMin(alpha, alphas[2]); alpha = PetscMax(alpha, alphas[0]); } else { alpha = 1.0; } ierr = VecAXPBY(X, alpha, 1 - alpha, Y);CHKERRQ(ierr); ierr = SNESComputeFunction(snes, X, F);CHKERRQ(ierr); if (alpha != 1.0) { ierr = VecNorm(F, NORM_2, gnorm);CHKERRQ(ierr); } else { *gnorm = PetscSqrtReal(norms[2]); } *flag = PETSC_TRUE; PetscFunctionReturn(0); } #undef __FUNCT__ #define __FUNCT__ "SNESSolve_QN" static PetscErrorCode SNESSolve_QN(SNES snes) { PetscErrorCode ierr; QNContext * qn = (QNContext*) snes->data; Vec x, xold; Vec f, fold; Vec w, y; PetscInt i, k; PetscReal fnorm, xnorm; PetscInt m = qn->m; PetscBool ls_OK; PetscScalar rhosc; Vec * dX = qn->dX; Vec * dF = qn->dF; PetscScalar * rho = qn->rho; /* basically just a regular newton's method except for the application of the jacobian */ PetscFunctionBegin; x = snes->vec_sol; xold = snes->vec_sol_update; /* dX_k */ w = snes->work[1]; f = snes->vec_func; fold = snes->work[0]; y = snes->work[2]; snes->reason = SNES_CONVERGED_ITERATING; ierr = PetscObjectTakeAccess(snes);CHKERRQ(ierr); snes->iter = 0; snes->norm = 0.; ierr = PetscObjectGrantAccess(snes);CHKERRQ(ierr); ierr = SNESComputeFunction(snes,x,f);CHKERRQ(ierr); if (snes->domainerror) { snes->reason = SNES_DIVERGED_FUNCTION_DOMAIN; PetscFunctionReturn(0); } ierr = VecNorm(f, NORM_2, &fnorm);CHKERRQ(ierr); /* fnorm <- ||F|| */ if (PetscIsInfOrNanReal(fnorm)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FP,"Infinite or not-a-number generated in norm"); ierr = PetscObjectTakeAccess(snes);CHKERRQ(ierr); snes->norm = fnorm; ierr = PetscObjectGrantAccess(snes);CHKERRQ(ierr); SNESLogConvHistory(snes,fnorm,0); ierr = SNESMonitor(snes,0,fnorm);CHKERRQ(ierr); /* set parameter for default relative tolerance convergence test */ snes->ttol = fnorm*snes->rtol; /* test convergence */ ierr = (*snes->ops->converged)(snes,0,0.0,0.0,fnorm,&snes->reason,snes->cnvP);CHKERRQ(ierr); if (snes->reason) PetscFunctionReturn(0); ierr = VecCopy(f, fold);CHKERRQ(ierr); ierr = VecCopy(x, xold);CHKERRQ(ierr); for(i = 0; i < snes->max_its; i++) { /* general purpose update */ if (snes->ops->update) { ierr = (*snes->ops->update)(snes, snes->iter);CHKERRQ(ierr); } /* apply the current iteration of the approximate jacobian */ ierr = LBGFSApplyJinv_Private(snes, i, f, y);CHKERRQ(ierr); /* line search for lambda */ ierr = VecAYPX(y,-1.0,x);CHKERRQ(ierr); ierr = QNLineSearchQuadratic(snes, PETSC_NULL, x, f, y, fnorm, xnorm, 0, w,&xnorm, &fnorm, &ls_OK);CHKERRQ(ierr); ierr = SNESComputeFunction(snes, x, f);CHKERRQ(ierr); ierr = VecNorm(f, NORM_2, &fnorm);CHKERRQ(ierr); if (PetscIsInfOrNanReal(fnorm)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FP,"Infinite or not-a-number generated in norm"); ierr = PetscObjectTakeAccess(snes);CHKERRQ(ierr); snes->norm = fnorm; ierr = PetscObjectGrantAccess(snes);CHKERRQ(ierr); SNESLogConvHistory(snes,fnorm,i+1); ierr = SNESMonitor(snes,i+1,fnorm);CHKERRQ(ierr); /* set parameter for default relative tolerance convergence test */ ierr = (*snes->ops->converged)(snes,i+1,0.0,0.0,fnorm,&snes->reason,snes->cnvP);CHKERRQ(ierr); if (snes->reason) PetscFunctionReturn(0); /* set the differences */ k = i % m; ierr = VecCopy(f, dF[k]);CHKERRQ(ierr); ierr = VecAXPY(dF[k], -1.0, fold);CHKERRQ(ierr); ierr = VecCopy(x, dX[k]);CHKERRQ(ierr); ierr = VecAXPY(dX[k], -1.0, xold);CHKERRQ(ierr); ierr = VecDot(dX[k], dF[k], &rhosc);CHKERRQ(ierr); rho[k] = 1. / rhosc; ierr = VecCopy(f, fold);CHKERRQ(ierr); ierr = VecCopy(x, xold);CHKERRQ(ierr); } if (i == snes->max_its) { ierr = PetscInfo1(snes, "Maximum number of iterations has been reached: %D\n", snes->max_its);CHKERRQ(ierr); if (!snes->reason) snes->reason = SNES_DIVERGED_MAX_IT; } PetscFunctionReturn(0); } #undef __FUNCT__ #define __FUNCT__ "SNESSetUp_QN" static PetscErrorCode SNESSetUp_QN(SNES snes) { QNContext * qn = (QNContext *)snes->data; PetscErrorCode ierr; PetscFunctionBegin; ierr = VecDuplicateVecs(snes->vec_sol, qn->m, &qn->dX);CHKERRQ(ierr); ierr = VecDuplicateVecs(snes->vec_sol, qn->m, &qn->dF);CHKERRQ(ierr); ierr = PetscMalloc3(qn->m, PetscScalar, &qn->alpha, qn->m, PetscScalar, &qn->beta, qn->m, PetscScalar, &qn->rho);CHKERRQ(ierr); ierr = SNESDefaultGetWork(snes,3);CHKERRQ(ierr); PetscFunctionReturn(0); } #undef __FUNCT__ #define __FUNCT__ "SNESReset_QN" static PetscErrorCode SNESReset_QN(SNES snes) { PetscErrorCode ierr; QNContext * qn; PetscFunctionBegin; if (snes->data) { qn = (QNContext *)snes->data; if (qn->dX) { ierr = VecDestroyVecs(qn->m, &qn->dX);CHKERRQ(ierr); } if (qn->dF) { ierr = VecDestroyVecs(qn->m, &qn->dF);CHKERRQ(ierr); } ierr = PetscFree3(qn->alpha, qn->beta, qn->rho);CHKERRQ(ierr); } if (snes->work) {ierr = VecDestroyVecs(snes->nwork,&snes->work);CHKERRQ(ierr);} PetscFunctionReturn(0); } #undef __FUNCT__ #define __FUNCT__ "SNESDestroy_QN" static PetscErrorCode SNESDestroy_QN(SNES snes) { PetscErrorCode ierr; PetscFunctionBegin; ierr = SNESReset_QN(snes);CHKERRQ(ierr); ierr = PetscFree(snes->data);CHKERRQ(ierr); PetscFunctionReturn(0); } #undef __FUNCT__ #define __FUNCT__ "SNESSetFromOptions_QN" static PetscErrorCode SNESSetFromOptions_QN(SNES snes) { PetscErrorCode ierr; QNContext * qn; PetscFunctionBegin; qn = (QNContext *)snes->data; ierr = PetscOptionsHead("SNES QN options");CHKERRQ(ierr); ierr = PetscOptionsReal("-snes_ls_damping", "Damping parameter", "SNES", qn->lambda, &qn->lambda, PETSC_NULL);CHKERRQ(ierr); ierr = PetscOptionsInt("-snes_qn_m", "Number of past states saved for L-Broyden methods", "SNES", qn->m, &qn->m, PETSC_NULL);CHKERRQ(ierr); ierr = PetscOptionsTail();CHKERRQ(ierr); PetscFunctionReturn(0); } /* -------------------------------------------------------------------------- */ /*MC SNESQN - Limited-Memory Quasi-Newton methods for the solution of nonlinear systems. Options Database: + -snes_qn_m - Number of past states saved for the L-Broyden methods. + -snes_ls_damping - The damping parameter on the update to x. Notes: This implements the L-BFGS algorithm for the solution of F(x) = 0 using previous change in F(x) and x to form the approximate inverse Jacobian using a series of multiplicative rank-one updates. This will eventually be generalized to implement several limited-memory Broyden methods. References: L-Broyden Methods: a generalization of the L-BFGS method to the limited memory Broyden family, M. B. Reed, International Journal of Computer Mathematics, vol. 86, 2009. Level: beginner .seealso: SNESCreate(), SNES, SNESSetType(), SNESLS, SNESTR M*/ EXTERN_C_BEGIN #undef __FUNCT__ #define __FUNCT__ "SNESCreate_QN" PetscErrorCode SNESCreate_QN(SNES snes) { PetscErrorCode ierr; QNContext * qn; PetscFunctionBegin; snes->ops->setup = SNESSetUp_QN; snes->ops->solve = SNESSolve_QN; snes->ops->destroy = SNESDestroy_QN; snes->ops->setfromoptions = SNESSetFromOptions_QN; snes->ops->view = 0; snes->ops->reset = SNESReset_QN; snes->usespc = PETSC_TRUE; snes->usesksp = PETSC_FALSE; ierr = PetscNewLog(snes, QNContext, &qn);CHKERRQ(ierr); snes->data = (void *) qn; qn->m = 100; qn->lambda = 1.; qn->dX = PETSC_NULL; qn->dF = PETSC_NULL; PetscFunctionReturn(0); } EXTERN_C_END