138774f0aSPeter Brune #include <../src/snes/impls/ngmres/snesngmres.h> /*I "petscsnes.h" I*/ 238774f0aSPeter Brune #include <petscblaslapack.h> 338774f0aSPeter Brune 438774f0aSPeter Brune PetscErrorCode SNESNGMRESUpdateSubspace_Private(SNES snes,PetscInt ivec,PetscInt l,Vec F,PetscReal fnorm,Vec X) 538774f0aSPeter Brune { 638774f0aSPeter Brune SNES_NGMRES *ngmres = (SNES_NGMRES*) snes->data; 738774f0aSPeter Brune Vec *Fdot = ngmres->Fdot; 838774f0aSPeter Brune Vec *Xdot = ngmres->Xdot; 938774f0aSPeter Brune 1038774f0aSPeter Brune PetscFunctionBegin; 11*08401ef6SPierre Jolivet PetscCheck(ivec <= l,PetscObjectComm((PetscObject)snes),PETSC_ERR_ARG_WRONGSTATE,"Cannot update vector %D with space size %D!",ivec,l); 129566063dSJacob Faibussowitsch PetscCall(VecCopy(F,Fdot[ivec])); 139566063dSJacob Faibussowitsch PetscCall(VecCopy(X,Xdot[ivec])); 141aa26658SKarl Rupp 1538774f0aSPeter Brune ngmres->fnorms[ivec] = fnorm; 1638774f0aSPeter Brune PetscFunctionReturn(0); 1738774f0aSPeter Brune } 1838774f0aSPeter Brune 19b3c6a99cSPeter Brune PetscErrorCode SNESNGMRESFormCombinedSolution_Private(SNES snes,PetscInt ivec,PetscInt l,Vec XM,Vec FM,PetscReal fMnorm,Vec X,Vec XA,Vec FA) 2038774f0aSPeter Brune { 2138774f0aSPeter Brune SNES_NGMRES *ngmres = (SNES_NGMRES*) snes->data; 2238774f0aSPeter Brune PetscInt i,j; 2338774f0aSPeter Brune Vec *Fdot = ngmres->Fdot; 2438774f0aSPeter Brune Vec *Xdot = ngmres->Xdot; 2538774f0aSPeter Brune PetscScalar *beta = ngmres->beta; 2638774f0aSPeter Brune PetscScalar *xi = ngmres->xi; 2738774f0aSPeter Brune PetscScalar alph_total = 0.; 2838774f0aSPeter Brune PetscReal nu; 2938774f0aSPeter Brune Vec Y = snes->work[2]; 3038774f0aSPeter Brune PetscBool changed_y,changed_w; 3138774f0aSPeter Brune 3238774f0aSPeter Brune PetscFunctionBegin; 3338774f0aSPeter Brune nu = fMnorm*fMnorm; 3438774f0aSPeter Brune 3538774f0aSPeter Brune /* construct the right hand side and xi factors */ 36b3c6a99cSPeter Brune if (l > 0) { 379566063dSJacob Faibussowitsch PetscCall(VecMDotBegin(FM,l,Fdot,xi)); 389566063dSJacob Faibussowitsch PetscCall(VecMDotBegin(Fdot[ivec],l,Fdot,beta)); 399566063dSJacob Faibussowitsch PetscCall(VecMDotEnd(FM,l,Fdot,xi)); 409566063dSJacob Faibussowitsch PetscCall(VecMDotEnd(Fdot[ivec],l,Fdot,beta)); 41b3c6a99cSPeter Brune for (i = 0; i < l; i++) { 42b3c6a99cSPeter Brune Q(i,ivec) = beta[i]; 43b3c6a99cSPeter Brune Q(ivec,i) = beta[i]; 44b3c6a99cSPeter Brune } 45b3c6a99cSPeter Brune } else { 46b3c6a99cSPeter Brune Q(0,0) = ngmres->fnorms[ivec]*ngmres->fnorms[ivec]; 47b3c6a99cSPeter Brune } 48b3c6a99cSPeter Brune 491aa26658SKarl Rupp for (i = 0; i < l; i++) beta[i] = nu - xi[i]; 501aa26658SKarl Rupp 5138774f0aSPeter Brune /* construct h */ 5238774f0aSPeter Brune for (j = 0; j < l; j++) { 5338774f0aSPeter Brune for (i = 0; i < l; i++) { 5438774f0aSPeter Brune H(i,j) = Q(i,j)-xi[i]-xi[j]+nu; 5538774f0aSPeter Brune } 5638774f0aSPeter Brune } 5738774f0aSPeter Brune if (l == 1) { 5838774f0aSPeter Brune /* simply set alpha[0] = beta[0] / H[0, 0] */ 591aa26658SKarl Rupp if (H(0,0) != 0.) beta[0] = beta[0]/H(0,0); 601aa26658SKarl Rupp else beta[0] = 0.; 6138774f0aSPeter Brune } else { 629566063dSJacob Faibussowitsch PetscCall(PetscBLASIntCast(l,&ngmres->m)); 639566063dSJacob Faibussowitsch PetscCall(PetscBLASIntCast(l,&ngmres->n)); 6438774f0aSPeter Brune ngmres->info = 0; 6538774f0aSPeter Brune ngmres->rcond = -1.; 669566063dSJacob Faibussowitsch PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF)); 6738774f0aSPeter Brune #if defined(PETSC_USE_COMPLEX) 688b83055fSJed Brown PetscStackCallBLAS("LAPACKgelss",LAPACKgelss_(&ngmres->m,&ngmres->n,&ngmres->nrhs,ngmres->h,&ngmres->lda,ngmres->beta,&ngmres->ldb,ngmres->s,&ngmres->rcond,&ngmres->rank,ngmres->work,&ngmres->lwork,ngmres->rwork,&ngmres->info)); 6938774f0aSPeter Brune #else 708b83055fSJed Brown PetscStackCallBLAS("LAPACKgelss",LAPACKgelss_(&ngmres->m,&ngmres->n,&ngmres->nrhs,ngmres->h,&ngmres->lda,ngmres->beta,&ngmres->ldb,ngmres->s,&ngmres->rcond,&ngmres->rank,ngmres->work,&ngmres->lwork,&ngmres->info)); 7138774f0aSPeter Brune #endif 729566063dSJacob Faibussowitsch PetscCall(PetscFPTrapPop()); 73*08401ef6SPierre Jolivet PetscCheck(ngmres->info >= 0,PetscObjectComm((PetscObject)snes),PETSC_ERR_LIB,"Bad argument to GELSS"); 74*08401ef6SPierre Jolivet PetscCheck(ngmres->info <= 0,PetscObjectComm((PetscObject)snes),PETSC_ERR_LIB,"SVD failed to converge"); 7538774f0aSPeter Brune } 7638774f0aSPeter Brune for (i=0; i<l; i++) { 772c71b3e2SJacob Faibussowitsch PetscCheckFalse(PetscIsInfOrNanScalar(beta[i]),PetscObjectComm((PetscObject)snes),PETSC_ERR_LIB,"SVD generated inconsistent output"); 7838774f0aSPeter Brune } 7938774f0aSPeter Brune alph_total = 0.; 801aa26658SKarl Rupp for (i = 0; i < l; i++) alph_total += beta[i]; 811aa26658SKarl Rupp 829566063dSJacob Faibussowitsch PetscCall(VecCopy(XM,XA)); 839566063dSJacob Faibussowitsch PetscCall(VecScale(XA,1.-alph_total)); 849566063dSJacob Faibussowitsch PetscCall(VecMAXPY(XA,l,beta,Xdot)); 8538774f0aSPeter Brune /* check the validity of the step */ 869566063dSJacob Faibussowitsch PetscCall(VecCopy(XA,Y)); 879566063dSJacob Faibussowitsch PetscCall(VecAXPY(Y,-1.0,X)); 889566063dSJacob Faibussowitsch PetscCall(SNESLineSearchPostCheck(snes->linesearch,X,Y,XA,&changed_y,&changed_w)); 8946159c86SPeter Brune if (!ngmres->approxfunc) { 90efd4aadfSBarry Smith if (snes->npc && snes->npcside== PC_LEFT) { 919566063dSJacob Faibussowitsch PetscCall(SNESApplyNPC(snes,XA,NULL,FA)); 9246159c86SPeter Brune } else { 939566063dSJacob Faibussowitsch PetscCall(SNESComputeFunction(snes,XA,FA)); 9446159c86SPeter Brune } 9535f895b4SBarry Smith } else { 969566063dSJacob Faibussowitsch PetscCall(VecCopy(FM,FA)); 979566063dSJacob Faibussowitsch PetscCall(VecScale(FA,1.-alph_total)); 989566063dSJacob Faibussowitsch PetscCall(VecMAXPY(FA,l,beta,Fdot)); 99077c4231SPeter Brune } 10038774f0aSPeter Brune PetscFunctionReturn(0); 10138774f0aSPeter Brune } 10238774f0aSPeter Brune 103b3c6a99cSPeter Brune PetscErrorCode SNESNGMRESNorms_Private(SNES snes,PetscInt l,Vec X,Vec F,Vec XM,Vec FM,Vec XA,Vec FA,Vec D,PetscReal *dnorm,PetscReal *dminnorm,PetscReal *xMnorm,PetscReal *fMnorm,PetscReal *yMnorm, PetscReal *xAnorm,PetscReal *fAnorm,PetscReal *yAnorm) 10438774f0aSPeter Brune { 10538774f0aSPeter Brune SNES_NGMRES *ngmres = (SNES_NGMRES*) snes->data; 106b3c6a99cSPeter Brune PetscReal dcurnorm,dmin = -1.0; 10738774f0aSPeter Brune Vec *Xdot = ngmres->Xdot; 10838774f0aSPeter Brune PetscInt i; 10938774f0aSPeter Brune 11038774f0aSPeter Brune PetscFunctionBegin; 111b3c6a99cSPeter Brune if (xMnorm) { 1129566063dSJacob Faibussowitsch PetscCall(VecNormBegin(XM,NORM_2,xMnorm)); 113b3c6a99cSPeter Brune } 114b3c6a99cSPeter Brune if (fMnorm) { 1159566063dSJacob Faibussowitsch PetscCall(VecNormBegin(FM,NORM_2,fMnorm)); 116b3c6a99cSPeter Brune } 117b3c6a99cSPeter Brune if (yMnorm) { 1189566063dSJacob Faibussowitsch PetscCall(VecCopy(X,D)); 1199566063dSJacob Faibussowitsch PetscCall(VecAXPY(D,-1.0,XM)); 1209566063dSJacob Faibussowitsch PetscCall(VecNormBegin(D,NORM_2,yMnorm)); 121b3c6a99cSPeter Brune } 122b3c6a99cSPeter Brune if (xAnorm) { 1239566063dSJacob Faibussowitsch PetscCall(VecNormBegin(XA,NORM_2,xAnorm)); 124b3c6a99cSPeter Brune } 12538774f0aSPeter Brune if (fAnorm) { 1269566063dSJacob Faibussowitsch PetscCall(VecNormBegin(FA,NORM_2,fAnorm)); 12738774f0aSPeter Brune } 128b3c6a99cSPeter Brune if (yAnorm) { 1299566063dSJacob Faibussowitsch PetscCall(VecCopy(X,D)); 1309566063dSJacob Faibussowitsch PetscCall(VecAXPY(D,-1.0,XA)); 1319566063dSJacob Faibussowitsch PetscCall(VecNormBegin(D,NORM_2,yAnorm)); 132b3c6a99cSPeter Brune } 13338774f0aSPeter Brune if (dnorm) { 1349566063dSJacob Faibussowitsch PetscCall(VecCopy(XA,D)); 1359566063dSJacob Faibussowitsch PetscCall(VecAXPY(D,-1.0,XM)); 1369566063dSJacob Faibussowitsch PetscCall(VecNormBegin(D,NORM_2,dnorm)); 13738774f0aSPeter Brune } 13838774f0aSPeter Brune if (dminnorm) { 13938774f0aSPeter Brune for (i=0; i<l; i++) { 1409566063dSJacob Faibussowitsch PetscCall(VecCopy(Xdot[i],D)); 1419566063dSJacob Faibussowitsch PetscCall(VecAXPY(D,-1.0,XA)); 1429566063dSJacob Faibussowitsch PetscCall(VecNormBegin(D,NORM_2,&ngmres->xnorms[i])); 14338774f0aSPeter Brune } 14438774f0aSPeter Brune } 1459566063dSJacob Faibussowitsch if (xMnorm) PetscCall(VecNormEnd(XM,NORM_2,xMnorm)); 1469566063dSJacob Faibussowitsch if (fMnorm) PetscCall(VecNormEnd(FM,NORM_2,fMnorm)); 1479566063dSJacob Faibussowitsch if (yMnorm) PetscCall(VecNormEnd(D,NORM_2,yMnorm)); 1489566063dSJacob Faibussowitsch if (xAnorm) PetscCall(VecNormEnd(XA,NORM_2,xAnorm)); 1499566063dSJacob Faibussowitsch if (fAnorm) PetscCall(VecNormEnd(FA,NORM_2,fAnorm)); 1509566063dSJacob Faibussowitsch if (yAnorm) PetscCall(VecNormEnd(D,NORM_2,yAnorm)); 1519566063dSJacob Faibussowitsch if (dnorm) PetscCall(VecNormEnd(D,NORM_2,dnorm)); 15238774f0aSPeter Brune if (dminnorm) { 15338774f0aSPeter Brune for (i=0; i<l; i++) { 1549566063dSJacob Faibussowitsch PetscCall(VecNormEnd(D,NORM_2,&ngmres->xnorms[i])); 15538774f0aSPeter Brune dcurnorm = ngmres->xnorms[i]; 156b3c6a99cSPeter Brune if ((dcurnorm < dmin) || (dmin < 0.0)) dmin = dcurnorm; 15738774f0aSPeter Brune } 158b3c6a99cSPeter Brune *dminnorm = dmin; 15938774f0aSPeter Brune } 16038774f0aSPeter Brune PetscFunctionReturn(0); 16138774f0aSPeter Brune } 16238774f0aSPeter Brune 163b3c6a99cSPeter Brune PetscErrorCode SNESNGMRESSelect_Private(SNES snes,PetscInt k_restart,Vec XM,Vec FM,PetscReal xMnorm,PetscReal fMnorm,PetscReal yMnorm,Vec XA,Vec FA,PetscReal xAnorm,PetscReal fAnorm,PetscReal yAnorm,PetscReal dnorm,PetscReal fminnorm,PetscReal dminnorm,Vec X,Vec F,Vec Y,PetscReal *xnorm,PetscReal *fnorm,PetscReal *ynorm) 16438774f0aSPeter Brune { 16538774f0aSPeter Brune SNES_NGMRES *ngmres = (SNES_NGMRES*) snes->data; 166422a814eSBarry Smith SNESLineSearchReason lssucceed; 167422a814eSBarry Smith PetscBool selectA; 16838774f0aSPeter Brune 16938774f0aSPeter Brune PetscFunctionBegin; 17038774f0aSPeter Brune if (ngmres->select_type == SNES_NGMRES_SELECT_LINESEARCH) { 17138774f0aSPeter Brune /* X = X + \lambda(XA - X) */ 17238774f0aSPeter Brune if (ngmres->monitor) { 1739566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPrintf(ngmres->monitor,"||F_A||_2 = %e, ||F_M||_2 = %e\n",fAnorm,fMnorm)); 17438774f0aSPeter Brune } 1759566063dSJacob Faibussowitsch PetscCall(VecCopy(FM,F)); 1769566063dSJacob Faibussowitsch PetscCall(VecCopy(XM,X)); 1779566063dSJacob Faibussowitsch PetscCall(VecCopy(XA,Y)); 1789566063dSJacob Faibussowitsch PetscCall(VecAYPX(Y,-1.0,X)); 17938774f0aSPeter Brune *fnorm = fMnorm; 1809566063dSJacob Faibussowitsch PetscCall(SNESLineSearchApply(ngmres->additive_linesearch,X,F,fnorm,Y)); 1819566063dSJacob Faibussowitsch PetscCall(SNESLineSearchGetReason(ngmres->additive_linesearch,&lssucceed)); 1829566063dSJacob Faibussowitsch PetscCall(SNESLineSearchGetNorms(ngmres->additive_linesearch,xnorm,fnorm,ynorm)); 183422a814eSBarry Smith if (lssucceed) { 18438774f0aSPeter Brune if (++snes->numFailures >= snes->maxFailures) { 18538774f0aSPeter Brune snes->reason = SNES_DIVERGED_LINE_SEARCH; 18638774f0aSPeter Brune PetscFunctionReturn(0); 18738774f0aSPeter Brune } 18838774f0aSPeter Brune } 18938774f0aSPeter Brune if (ngmres->monitor) { 1909566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPrintf(ngmres->monitor,"Additive solution: ||F||_2 = %e\n",*fnorm)); 19138774f0aSPeter Brune } 19238774f0aSPeter Brune } else if (ngmres->select_type == SNES_NGMRES_SELECT_DIFFERENCE) { 19338774f0aSPeter Brune selectA = PETSC_TRUE; 19438774f0aSPeter Brune /* Conditions for choosing the accelerated answer */ 19538774f0aSPeter Brune /* Criterion A -- the norm of the function isn't increased above the minimum by too much */ 1961aa26658SKarl Rupp if (fAnorm >= ngmres->gammaA*fminnorm) selectA = PETSC_FALSE; 1971aa26658SKarl Rupp 19838774f0aSPeter Brune /* Criterion B -- the choice of x^A isn't too close to some other choice */ 19938774f0aSPeter Brune if (ngmres->epsilonB*dnorm<dminnorm || PetscSqrtReal(*fnorm)<ngmres->deltaB*PetscSqrtReal(fminnorm)) { 2001aa26658SKarl Rupp } else selectA=PETSC_FALSE; 2011aa26658SKarl Rupp 20238774f0aSPeter Brune if (selectA) { 20338774f0aSPeter Brune if (ngmres->monitor) { 2049566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPrintf(ngmres->monitor,"picked X_A, ||F_A||_2 = %e, ||F_M||_2 = %e\n",fAnorm,fMnorm)); 20538774f0aSPeter Brune } 20638774f0aSPeter Brune /* copy it over */ 207b3c6a99cSPeter Brune *xnorm = xAnorm; 20838774f0aSPeter Brune *fnorm = fAnorm; 209b3c6a99cSPeter Brune *ynorm = yAnorm; 2109566063dSJacob Faibussowitsch PetscCall(VecCopy(FA,F)); 2119566063dSJacob Faibussowitsch PetscCall(VecCopy(XA,X)); 21238774f0aSPeter Brune } else { 21338774f0aSPeter Brune if (ngmres->monitor) { 2149566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPrintf(ngmres->monitor,"picked X_M, ||F_A||_2 = %e, ||F_M||_2 = %e\n",fAnorm,fMnorm)); 21538774f0aSPeter Brune } 216b3c6a99cSPeter Brune *xnorm = xMnorm; 21738774f0aSPeter Brune *fnorm = fMnorm; 218b3c6a99cSPeter Brune *ynorm = yMnorm; 2199566063dSJacob Faibussowitsch PetscCall(VecCopy(XM,Y)); 2209566063dSJacob Faibussowitsch PetscCall(VecAXPY(Y,-1.0,X)); 2219566063dSJacob Faibussowitsch PetscCall(VecCopy(FM,F)); 2229566063dSJacob Faibussowitsch PetscCall(VecCopy(XM,X)); 22338774f0aSPeter Brune } 22438774f0aSPeter Brune } else { /* none */ 225b3c6a99cSPeter Brune *xnorm = xAnorm; 22638774f0aSPeter Brune *fnorm = fAnorm; 227b3c6a99cSPeter Brune *ynorm = yAnorm; 2289566063dSJacob Faibussowitsch PetscCall(VecCopy(FA,F)); 2299566063dSJacob Faibussowitsch PetscCall(VecCopy(XA,X)); 23038774f0aSPeter Brune } 23138774f0aSPeter Brune PetscFunctionReturn(0); 23238774f0aSPeter Brune } 23338774f0aSPeter Brune 23423b3e82cSAsbjørn Nilsen Riseth PetscErrorCode SNESNGMRESSelectRestart_Private(SNES snes,PetscInt l,PetscReal fMnorm, PetscReal fAnorm,PetscReal dnorm,PetscReal fminnorm,PetscReal dminnorm,PetscBool *selectRestart) 23538774f0aSPeter Brune { 23638774f0aSPeter Brune SNES_NGMRES *ngmres = (SNES_NGMRES*)snes->data; 23738774f0aSPeter Brune 23838774f0aSPeter Brune PetscFunctionBegin; 23938774f0aSPeter Brune *selectRestart = PETSC_FALSE; 24038774f0aSPeter Brune /* difference stagnation restart */ 24121687c63SPeter Brune if ((ngmres->epsilonB*dnorm > dminnorm) && (PetscSqrtReal(fAnorm) > ngmres->deltaB*PetscSqrtReal(fminnorm)) && l > 0) { 24238774f0aSPeter Brune if (ngmres->monitor) { 2439566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPrintf(ngmres->monitor,"difference restart: %e > %e\n",ngmres->epsilonB*dnorm,dminnorm)); 24438774f0aSPeter Brune } 24538774f0aSPeter Brune *selectRestart = PETSC_TRUE; 24638774f0aSPeter Brune } 24738774f0aSPeter Brune /* residual stagnation restart */ 24838774f0aSPeter Brune if (PetscSqrtReal(fAnorm) > ngmres->gammaC*PetscSqrtReal(fminnorm)) { 24938774f0aSPeter Brune if (ngmres->monitor) { 2509566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPrintf(ngmres->monitor,"residual restart: %e > %e\n",PetscSqrtReal(fAnorm),ngmres->gammaC*PetscSqrtReal(fminnorm))); 25138774f0aSPeter Brune } 25238774f0aSPeter Brune *selectRestart = PETSC_TRUE; 25338774f0aSPeter Brune } 25423b3e82cSAsbjørn Nilsen Riseth 25523b3e82cSAsbjørn Nilsen Riseth /* F_M stagnation restart */ 25623b3e82cSAsbjørn Nilsen Riseth if (ngmres->restart_fm_rise && fMnorm > snes->norm) { 25723b3e82cSAsbjørn Nilsen Riseth if (ngmres->monitor) { 2589566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPrintf(ngmres->monitor,"F_M rise restart: %e > %e\n",fMnorm,snes->norm)); 25923b3e82cSAsbjørn Nilsen Riseth } 26023b3e82cSAsbjørn Nilsen Riseth *selectRestart = PETSC_TRUE; 26123b3e82cSAsbjørn Nilsen Riseth } 26223b3e82cSAsbjørn Nilsen Riseth 26338774f0aSPeter Brune PetscFunctionReturn(0); 26438774f0aSPeter Brune } 265