#include <../src/tao/constrained/impls/ipm/pdipm.h> /* TaoPDIPMEvaluateFunctionsAndJacobians - Evaluate the objective function f, gradient fx, constraints, and all the Jacobians at current vector Collective Input Parameter: + tao - solver context - x - vector at which all objects to be evaluated Level: beginner .seealso: `TAOPDIPM`, `TaoPDIPMUpdateConstraints()`, `TaoPDIPMSetUpBounds()` */ static PetscErrorCode TaoPDIPMEvaluateFunctionsAndJacobians(Tao tao, Vec x) { TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; PetscFunctionBegin; /* Compute user objective function and gradient */ PetscCall(TaoComputeObjectiveAndGradient(tao, x, &pdipm->obj, tao->gradient)); /* Equality constraints and Jacobian */ if (pdipm->Ng) { PetscCall(TaoComputeEqualityConstraints(tao, x, tao->constraints_equality)); PetscCall(TaoComputeJacobianEquality(tao, x, tao->jacobian_equality, tao->jacobian_equality_pre)); } /* Inequality constraints and Jacobian */ if (pdipm->Nh) { PetscCall(TaoComputeInequalityConstraints(tao, x, tao->constraints_inequality)); PetscCall(TaoComputeJacobianInequality(tao, x, tao->jacobian_inequality, tao->jacobian_inequality_pre)); } PetscFunctionReturn(PETSC_SUCCESS); } /* TaoPDIPMUpdateConstraints - Update the vectors ce and ci at x Collective Input Parameter: + tao - Tao context - x - vector at which constraints to be evaluated Level: beginner .seealso: `TAOPDIPM`, `TaoPDIPMEvaluateFunctionsAndJacobians()` */ static PetscErrorCode TaoPDIPMUpdateConstraints(Tao tao, Vec x) { TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; PetscInt i, offset, offset1, k, xstart; PetscScalar *carr; const PetscInt *ubptr, *lbptr, *bxptr, *fxptr; const PetscScalar *xarr, *xuarr, *xlarr, *garr, *harr; PetscFunctionBegin; PetscCall(VecGetOwnershipRange(x, &xstart, NULL)); PetscCall(VecGetArrayRead(x, &xarr)); PetscCall(VecGetArrayRead(tao->XU, &xuarr)); PetscCall(VecGetArrayRead(tao->XL, &xlarr)); /* (1) Update ce vector */ PetscCall(VecGetArrayWrite(pdipm->ce, &carr)); if (pdipm->Ng) { /* (1.a) Inserting updated g(x) */ PetscCall(VecGetArrayRead(tao->constraints_equality, &garr)); PetscCall(PetscArraycpy(carr, garr, pdipm->ng)); PetscCall(VecRestoreArrayRead(tao->constraints_equality, &garr)); } /* (1.b) Update xfixed */ if (pdipm->Nxfixed) { offset = pdipm->ng; PetscCall(ISGetIndices(pdipm->isxfixed, &fxptr)); /* global indices in x */ for (k = 0; k < pdipm->nxfixed; k++) { i = fxptr[k] - xstart; carr[offset + k] = xarr[i] - xuarr[i]; } } PetscCall(VecRestoreArrayWrite(pdipm->ce, &carr)); /* (2) Update ci vector */ PetscCall(VecGetArrayWrite(pdipm->ci, &carr)); if (pdipm->Nh) { /* (2.a) Inserting updated h(x) */ PetscCall(VecGetArrayRead(tao->constraints_inequality, &harr)); PetscCall(PetscArraycpy(carr, harr, pdipm->nh)); PetscCall(VecRestoreArrayRead(tao->constraints_inequality, &harr)); } /* (2.b) Update xub */ offset = pdipm->nh; if (pdipm->Nxub) { PetscCall(ISGetIndices(pdipm->isxub, &ubptr)); for (k = 0; k < pdipm->nxub; k++) { i = ubptr[k] - xstart; carr[offset + k] = xuarr[i] - xarr[i]; } } if (pdipm->Nxlb) { /* (2.c) Update xlb */ offset += pdipm->nxub; PetscCall(ISGetIndices(pdipm->isxlb, &lbptr)); /* global indices in x */ for (k = 0; k < pdipm->nxlb; k++) { i = lbptr[k] - xstart; carr[offset + k] = xarr[i] - xlarr[i]; } } if (pdipm->Nxbox) { /* (2.d) Update xbox */ offset += pdipm->nxlb; offset1 = offset + pdipm->nxbox; PetscCall(ISGetIndices(pdipm->isxbox, &bxptr)); /* global indices in x */ for (k = 0; k < pdipm->nxbox; k++) { i = bxptr[k] - xstart; /* local indices in x */ carr[offset + k] = xuarr[i] - xarr[i]; carr[offset1 + k] = xarr[i] - xlarr[i]; } } PetscCall(VecRestoreArrayWrite(pdipm->ci, &carr)); /* Restoring Vectors */ PetscCall(VecRestoreArrayRead(x, &xarr)); PetscCall(VecRestoreArrayRead(tao->XU, &xuarr)); PetscCall(VecRestoreArrayRead(tao->XL, &xlarr)); PetscFunctionReturn(PETSC_SUCCESS); } /* TaoPDIPMSetUpBounds - Create upper and lower bound vectors of x Collective Input Parameter: . tao - holds pdipm and XL & XU Level: beginner .seealso: `TAOPDIPM`, `TaoPDIPMUpdateConstraints` */ static PetscErrorCode TaoPDIPMSetUpBounds(Tao tao) { TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; const PetscScalar *xl, *xu; PetscInt n, *ixlb, *ixub, *ixfixed, *ixfree, *ixbox, i, low, high, idx; MPI_Comm comm; PetscInt sendbuf[5], recvbuf[5]; PetscFunctionBegin; /* Creates upper and lower bounds vectors on x, if not created already */ PetscCall(TaoComputeVariableBounds(tao)); PetscCall(VecGetLocalSize(tao->XL, &n)); PetscCall(PetscMalloc5(n, &ixlb, n, &ixub, n, &ixfree, n, &ixfixed, n, &ixbox)); PetscCall(VecGetOwnershipRange(tao->XL, &low, &high)); PetscCall(VecGetArrayRead(tao->XL, &xl)); PetscCall(VecGetArrayRead(tao->XU, &xu)); for (i = 0; i < n; i++) { idx = low + i; if ((PetscRealPart(xl[i]) > PETSC_NINFINITY) && (PetscRealPart(xu[i]) < PETSC_INFINITY)) { if (PetscRealPart(xl[i]) == PetscRealPart(xu[i])) { ixfixed[pdipm->nxfixed++] = idx; } else ixbox[pdipm->nxbox++] = idx; } else { if ((PetscRealPart(xl[i]) > PETSC_NINFINITY) && (PetscRealPart(xu[i]) >= PETSC_INFINITY)) { ixlb[pdipm->nxlb++] = idx; } else if ((PetscRealPart(xl[i]) <= PETSC_NINFINITY) && (PetscRealPart(xu[i]) < PETSC_INFINITY)) { ixub[pdipm->nxlb++] = idx; } else ixfree[pdipm->nxfree++] = idx; } } PetscCall(VecRestoreArrayRead(tao->XL, &xl)); PetscCall(VecRestoreArrayRead(tao->XU, &xu)); PetscCall(PetscObjectGetComm((PetscObject)tao, &comm)); sendbuf[0] = pdipm->nxlb; sendbuf[1] = pdipm->nxub; sendbuf[2] = pdipm->nxfixed; sendbuf[3] = pdipm->nxbox; sendbuf[4] = pdipm->nxfree; PetscCallMPI(MPIU_Allreduce(sendbuf, recvbuf, 5, MPIU_INT, MPI_SUM, comm)); pdipm->Nxlb = recvbuf[0]; pdipm->Nxub = recvbuf[1]; pdipm->Nxfixed = recvbuf[2]; pdipm->Nxbox = recvbuf[3]; pdipm->Nxfree = recvbuf[4]; if (pdipm->Nxlb) PetscCall(ISCreateGeneral(comm, pdipm->nxlb, ixlb, PETSC_COPY_VALUES, &pdipm->isxlb)); if (pdipm->Nxub) PetscCall(ISCreateGeneral(comm, pdipm->nxub, ixub, PETSC_COPY_VALUES, &pdipm->isxub)); if (pdipm->Nxfixed) PetscCall(ISCreateGeneral(comm, pdipm->nxfixed, ixfixed, PETSC_COPY_VALUES, &pdipm->isxfixed)); if (pdipm->Nxbox) PetscCall(ISCreateGeneral(comm, pdipm->nxbox, ixbox, PETSC_COPY_VALUES, &pdipm->isxbox)); if (pdipm->Nxfree) PetscCall(ISCreateGeneral(comm, pdipm->nxfree, ixfree, PETSC_COPY_VALUES, &pdipm->isxfree)); PetscCall(PetscFree5(ixlb, ixub, ixfixed, ixbox, ixfree)); PetscFunctionReturn(PETSC_SUCCESS); } /* TaoPDIPMInitializeSolution - Initialize `TAOPDIPM` solution X = [x; lambdae; lambdai; z]. X consists of four subvectors in the order [x; lambdae; lambdai; z]. These four subvectors need to be initialized and its values copied over to X. Instead of copying, we use `VecPlaceArray()`/`VecResetArray()` functions to share the memory locations for X and the subvectors Collective Input Parameter: . tao - Tao context Level: beginner */ static PetscErrorCode TaoPDIPMInitializeSolution(Tao tao) { TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; PetscScalar *Xarr, *z, *lambdai; PetscInt i; const PetscScalar *xarr, *h; PetscFunctionBegin; PetscCall(VecGetArrayWrite(pdipm->X, &Xarr)); /* Set Initialize X.x = tao->solution */ PetscCall(VecGetArrayRead(tao->solution, &xarr)); PetscCall(PetscArraycpy(Xarr, xarr, pdipm->nx)); PetscCall(VecRestoreArrayRead(tao->solution, &xarr)); /* Initialize X.lambdae = 0.0 */ if (pdipm->lambdae) PetscCall(VecSet(pdipm->lambdae, 0.0)); /* Initialize X.lambdai = push_init_lambdai, X.z = push_init_slack */ if (pdipm->Nci) { PetscCall(VecSet(pdipm->lambdai, pdipm->push_init_lambdai)); PetscCall(VecSet(pdipm->z, pdipm->push_init_slack)); /* Additional modification for X.lambdai and X.z */ PetscCall(VecGetArrayWrite(pdipm->lambdai, &lambdai)); PetscCall(VecGetArrayWrite(pdipm->z, &z)); if (pdipm->Nh) { PetscCall(VecGetArrayRead(tao->constraints_inequality, &h)); for (i = 0; i < pdipm->nh; i++) { if (h[i] < -pdipm->push_init_slack) z[i] = -h[i]; if (pdipm->mu / z[i] > pdipm->push_init_lambdai) lambdai[i] = pdipm->mu / z[i]; } PetscCall(VecRestoreArrayRead(tao->constraints_inequality, &h)); } PetscCall(VecRestoreArrayWrite(pdipm->lambdai, &lambdai)); PetscCall(VecRestoreArrayWrite(pdipm->z, &z)); } PetscCall(VecRestoreArrayWrite(pdipm->X, &Xarr)); PetscFunctionReturn(PETSC_SUCCESS); } /* TaoSNESJacobian_PDIPM - Evaluate the Hessian matrix at X Input Parameter: snes - SNES context X - KKT Vector *ctx - pdipm context Output Parameter: J - Hessian matrix Jpre - matrix to build the preconditioner from */ static PetscErrorCode TaoSNESJacobian_PDIPM(SNES snes, Vec X, Mat J, Mat Jpre, PetscCtx ctx) { Tao tao = (Tao)ctx; TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; PetscInt i, row, cols[2], Jrstart, rjstart, nc, j; const PetscInt *aj, *ranges, *Jranges, *rranges, *cranges; const PetscScalar *Xarr, *aa; PetscScalar vals[2]; PetscInt proc, nx_all, *nce_all = pdipm->nce_all; PetscFunctionBegin; PetscCall(MatGetOwnershipRanges(Jpre, &Jranges)); PetscCall(MatGetOwnershipRange(Jpre, &Jrstart, NULL)); PetscCall(MatGetOwnershipRangesColumn(tao->hessian, &rranges)); PetscCall(MatGetOwnershipRangesColumn(tao->hessian, &cranges)); PetscCall(VecGetArrayRead(X, &Xarr)); /* (1) insert Z and Ci to the 4th block of Jpre -- overwrite existing values */ if (pdipm->solve_symmetric_kkt) { /* 1 for eq 17 revised pdipm doc 0 for eq 18 (symmetric KKT) */ vals[0] = 1.0; for (i = 0; i < pdipm->nci; i++) { row = Jrstart + pdipm->off_z + i; cols[0] = Jrstart + pdipm->off_lambdai + i; cols[1] = row; vals[1] = Xarr[pdipm->off_lambdai + i] / Xarr[pdipm->off_z + i]; PetscCall(MatSetValues(Jpre, 1, &row, 2, cols, vals, INSERT_VALUES)); } } else { for (i = 0; i < pdipm->nci; i++) { row = Jrstart + pdipm->off_z + i; cols[0] = Jrstart + pdipm->off_lambdai + i; cols[1] = row; vals[0] = Xarr[pdipm->off_z + i]; vals[1] = Xarr[pdipm->off_lambdai + i]; PetscCall(MatSetValues(Jpre, 1, &row, 2, cols, vals, INSERT_VALUES)); } } /* (2) insert 2nd row block of Jpre: [ grad g, 0, 0, 0] */ if (pdipm->Ng) { PetscCall(MatGetOwnershipRange(tao->jacobian_equality, &rjstart, NULL)); for (i = 0; i < pdipm->ng; i++) { row = Jrstart + pdipm->off_lambdae + i; PetscCall(MatGetRow(tao->jacobian_equality, i + rjstart, &nc, &aj, &aa)); proc = 0; for (j = 0; j < nc; j++) { while (aj[j] >= cranges[proc + 1]) proc++; cols[0] = aj[j] - cranges[proc] + Jranges[proc]; PetscCall(MatSetValue(Jpre, row, cols[0], aa[j], INSERT_VALUES)); } PetscCall(MatRestoreRow(tao->jacobian_equality, i + rjstart, &nc, &aj, &aa)); if (pdipm->kkt_pd) { /* add shift \delta_c */ PetscCall(MatSetValue(Jpre, row, row, -pdipm->deltac, INSERT_VALUES)); } } } /* (3) insert 3rd row block of Jpre: [ -grad h, 0, deltac, I] */ if (pdipm->Nh) { PetscCall(MatGetOwnershipRange(tao->jacobian_inequality, &rjstart, NULL)); for (i = 0; i < pdipm->nh; i++) { row = Jrstart + pdipm->off_lambdai + i; PetscCall(MatGetRow(tao->jacobian_inequality, i + rjstart, &nc, &aj, &aa)); proc = 0; for (j = 0; j < nc; j++) { while (aj[j] >= cranges[proc + 1]) proc++; cols[0] = aj[j] - cranges[proc] + Jranges[proc]; PetscCall(MatSetValue(Jpre, row, cols[0], -aa[j], INSERT_VALUES)); } PetscCall(MatRestoreRow(tao->jacobian_inequality, i + rjstart, &nc, &aj, &aa)); if (pdipm->kkt_pd) { /* add shift \delta_c */ PetscCall(MatSetValue(Jpre, row, row, -pdipm->deltac, INSERT_VALUES)); } } } /* (4) insert 1st row block of Jpre: [Wxx, grad g', -grad h', 0] */ if (pdipm->Ng) { /* grad g' */ PetscCall(MatTranspose(tao->jacobian_equality, MAT_REUSE_MATRIX, &pdipm->jac_equality_trans)); } if (pdipm->Nh) { /* grad h' */ PetscCall(MatTranspose(tao->jacobian_inequality, MAT_REUSE_MATRIX, &pdipm->jac_inequality_trans)); } PetscCall(VecPlaceArray(pdipm->x, Xarr)); PetscCall(TaoComputeHessian(tao, pdipm->x, tao->hessian, tao->hessian_pre)); PetscCall(VecResetArray(pdipm->x)); PetscCall(MatGetOwnershipRange(tao->hessian, &rjstart, NULL)); for (i = 0; i < pdipm->nx; i++) { row = Jrstart + i; /* insert Wxx = fxx + ... -- provided by user */ PetscCall(MatGetRow(tao->hessian, i + rjstart, &nc, &aj, &aa)); proc = 0; for (j = 0; j < nc; j++) { while (aj[j] >= cranges[proc + 1]) proc++; cols[0] = aj[j] - cranges[proc] + Jranges[proc]; if (row == cols[0] && pdipm->kkt_pd) { /* add shift deltaw to Wxx component */ PetscCall(MatSetValue(Jpre, row, cols[0], aa[j] + pdipm->deltaw, INSERT_VALUES)); } else { PetscCall(MatSetValue(Jpre, row, cols[0], aa[j], INSERT_VALUES)); } } PetscCall(MatRestoreRow(tao->hessian, i + rjstart, &nc, &aj, &aa)); /* insert grad g' */ if (pdipm->ng) { PetscCall(MatGetRow(pdipm->jac_equality_trans, i + rjstart, &nc, &aj, &aa)); PetscCall(MatGetOwnershipRanges(tao->jacobian_equality, &ranges)); proc = 0; for (j = 0; j < nc; j++) { /* find row ownership of */ while (aj[j] >= ranges[proc + 1]) proc++; nx_all = rranges[proc + 1] - rranges[proc]; cols[0] = aj[j] - ranges[proc] + Jranges[proc] + nx_all; PetscCall(MatSetValue(Jpre, row, cols[0], aa[j], INSERT_VALUES)); } PetscCall(MatRestoreRow(pdipm->jac_equality_trans, i + rjstart, &nc, &aj, &aa)); } /* insert -grad h' */ if (pdipm->nh) { PetscCall(MatGetRow(pdipm->jac_inequality_trans, i + rjstart, &nc, &aj, &aa)); PetscCall(MatGetOwnershipRanges(tao->jacobian_inequality, &ranges)); proc = 0; for (j = 0; j < nc; j++) { /* find row ownership of */ while (aj[j] >= ranges[proc + 1]) proc++; nx_all = rranges[proc + 1] - rranges[proc]; cols[0] = aj[j] - ranges[proc] + Jranges[proc] + nx_all + nce_all[proc]; PetscCall(MatSetValue(Jpre, row, cols[0], -aa[j], INSERT_VALUES)); } PetscCall(MatRestoreRow(pdipm->jac_inequality_trans, i + rjstart, &nc, &aj, &aa)); } } PetscCall(VecRestoreArrayRead(X, &Xarr)); /* (6) assemble Jpre and J */ PetscCall(MatAssemblyBegin(Jpre, MAT_FINAL_ASSEMBLY)); PetscCall(MatAssemblyEnd(Jpre, MAT_FINAL_ASSEMBLY)); if (J != Jpre) { PetscCall(MatAssemblyBegin(J, MAT_FINAL_ASSEMBLY)); PetscCall(MatAssemblyEnd(J, MAT_FINAL_ASSEMBLY)); } PetscFunctionReturn(PETSC_SUCCESS); } /* TaoSnesFunction_PDIPM - Evaluate KKT function at X Input Parameter: snes - SNES context X - KKT Vector *ctx - pdipm Output Parameter: F - Updated Lagrangian vector */ static PetscErrorCode TaoSNESFunction_PDIPM(SNES snes, Vec X, Vec F, PetscCtx ctx) { Tao tao = (Tao)ctx; TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; PetscScalar *Farr; Vec x, L1; PetscInt i; const PetscScalar *Xarr, *carr, *zarr, *larr; PetscFunctionBegin; PetscCall(VecSet(F, 0.0)); PetscCall(VecGetArrayRead(X, &Xarr)); PetscCall(VecGetArrayWrite(F, &Farr)); /* (0) Evaluate f, fx, gradG, gradH at X.x Note: pdipm->x is not changed below */ x = pdipm->x; PetscCall(VecPlaceArray(x, Xarr)); PetscCall(TaoPDIPMEvaluateFunctionsAndJacobians(tao, x)); /* Update ce, ci, and Jci at X.x */ PetscCall(TaoPDIPMUpdateConstraints(tao, x)); PetscCall(VecResetArray(x)); /* (1) L1 = fx + (gradG'*DE + Jce_xfixed'*lambdae_xfixed) - (gradH'*DI + Jci_xb'*lambdai_xb) */ L1 = pdipm->x; PetscCall(VecPlaceArray(L1, Farr)); /* L1 = 0.0 */ if (pdipm->Nci) { if (pdipm->Nh) { /* L1 += gradH'*DI. Note: tao->DI is not changed below */ PetscCall(VecPlaceArray(tao->DI, Xarr + pdipm->off_lambdai)); PetscCall(MatMultTransposeAdd(tao->jacobian_inequality, tao->DI, L1, L1)); PetscCall(VecResetArray(tao->DI)); } /* L1 += Jci_xb'*lambdai_xb */ PetscCall(VecPlaceArray(pdipm->lambdai_xb, Xarr + pdipm->off_lambdai + pdipm->nh)); PetscCall(MatMultTransposeAdd(pdipm->Jci_xb, pdipm->lambdai_xb, L1, L1)); PetscCall(VecResetArray(pdipm->lambdai_xb)); /* L1 = - (gradH'*DI + Jci_xb'*lambdai_xb) */ PetscCall(VecScale(L1, -1.0)); } /* L1 += fx */ PetscCall(VecAXPY(L1, 1.0, tao->gradient)); if (pdipm->Nce) { if (pdipm->Ng) { /* L1 += gradG'*DE. Note: tao->DE is not changed below */ PetscCall(VecPlaceArray(tao->DE, Xarr + pdipm->off_lambdae)); PetscCall(MatMultTransposeAdd(tao->jacobian_equality, tao->DE, L1, L1)); PetscCall(VecResetArray(tao->DE)); } if (pdipm->Nxfixed) { /* L1 += Jce_xfixed'*lambdae_xfixed */ PetscCall(VecPlaceArray(pdipm->lambdae_xfixed, Xarr + pdipm->off_lambdae + pdipm->ng)); PetscCall(MatMultTransposeAdd(pdipm->Jce_xfixed, pdipm->lambdae_xfixed, L1, L1)); PetscCall(VecResetArray(pdipm->lambdae_xfixed)); } } PetscCall(VecResetArray(L1)); /* (2) L2 = ce(x) */ if (pdipm->Nce) { PetscCall(VecGetArrayRead(pdipm->ce, &carr)); for (i = 0; i < pdipm->nce; i++) Farr[pdipm->off_lambdae + i] = carr[i]; PetscCall(VecRestoreArrayRead(pdipm->ce, &carr)); } if (pdipm->Nci) { if (pdipm->solve_symmetric_kkt) { /* (3) L3 = z - ci(x); (4) L4 = Lambdai * e - mu/z *e */ PetscCall(VecGetArrayRead(pdipm->ci, &carr)); larr = Xarr + pdipm->off_lambdai; zarr = Xarr + pdipm->off_z; for (i = 0; i < pdipm->nci; i++) { Farr[pdipm->off_lambdai + i] = zarr[i] - carr[i]; Farr[pdipm->off_z + i] = larr[i] - pdipm->mu / zarr[i]; } PetscCall(VecRestoreArrayRead(pdipm->ci, &carr)); } else { /* (3) L3 = z - ci(x); (4) L4 = Z * Lambdai * e - mu * e */ PetscCall(VecGetArrayRead(pdipm->ci, &carr)); larr = Xarr + pdipm->off_lambdai; zarr = Xarr + pdipm->off_z; for (i = 0; i < pdipm->nci; i++) { Farr[pdipm->off_lambdai + i] = zarr[i] - carr[i]; Farr[pdipm->off_z + i] = zarr[i] * larr[i] - pdipm->mu; } PetscCall(VecRestoreArrayRead(pdipm->ci, &carr)); } } PetscCall(VecRestoreArrayRead(X, &Xarr)); PetscCall(VecRestoreArrayWrite(F, &Farr)); PetscFunctionReturn(PETSC_SUCCESS); } /* Evaluate F(X); then update tao->gnorm0, tao->step = mu, tao->residual = norm2(F_x,F_z) and tao->cnorm = norm2(F_ce,F_ci). */ static PetscErrorCode TaoSNESFunction_PDIPM_residual(SNES snes, Vec X, Vec F, PetscCtx ctx) { Tao tao = (Tao)ctx; TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; PetscScalar *Farr, *tmparr; Vec L1; PetscInt i; PetscReal res[2], cnorm[2]; const PetscScalar *Xarr = NULL; PetscFunctionBegin; PetscCall(TaoSNESFunction_PDIPM(snes, X, F, (void *)tao)); PetscCall(VecGetArrayWrite(F, &Farr)); PetscCall(VecGetArrayRead(X, &Xarr)); /* compute res[0] = norm2(F_x) */ L1 = pdipm->x; PetscCall(VecPlaceArray(L1, Farr)); PetscCall(VecNorm(L1, NORM_2, &res[0])); PetscCall(VecResetArray(L1)); /* compute res[1] = norm2(F_z), cnorm[1] = norm2(F_ci) */ if (pdipm->z) { if (pdipm->solve_symmetric_kkt) { PetscCall(VecPlaceArray(pdipm->z, Farr + pdipm->off_z)); if (pdipm->Nci) { PetscCall(VecGetArrayWrite(pdipm->z, &tmparr)); for (i = 0; i < pdipm->nci; i++) tmparr[i] *= Xarr[pdipm->off_z + i]; PetscCall(VecRestoreArrayWrite(pdipm->z, &tmparr)); } PetscCall(VecNorm(pdipm->z, NORM_2, &res[1])); if (pdipm->Nci) { PetscCall(VecGetArrayWrite(pdipm->z, &tmparr)); for (i = 0; i < pdipm->nci; i++) tmparr[i] /= Xarr[pdipm->off_z + i]; PetscCall(VecRestoreArrayWrite(pdipm->z, &tmparr)); } PetscCall(VecResetArray(pdipm->z)); } else { /* !solve_symmetric_kkt */ PetscCall(VecPlaceArray(pdipm->z, Farr + pdipm->off_z)); PetscCall(VecNorm(pdipm->z, NORM_2, &res[1])); PetscCall(VecResetArray(pdipm->z)); } PetscCall(VecPlaceArray(pdipm->ci, Farr + pdipm->off_lambdai)); PetscCall(VecNorm(pdipm->ci, NORM_2, &cnorm[1])); PetscCall(VecResetArray(pdipm->ci)); } else { res[1] = 0.0; cnorm[1] = 0.0; } /* compute cnorm[0] = norm2(F_ce) */ if (pdipm->Nce) { PetscCall(VecPlaceArray(pdipm->ce, Farr + pdipm->off_lambdae)); PetscCall(VecNorm(pdipm->ce, NORM_2, &cnorm[0])); PetscCall(VecResetArray(pdipm->ce)); } else cnorm[0] = 0.0; PetscCall(VecRestoreArrayWrite(F, &Farr)); PetscCall(VecRestoreArrayRead(X, &Xarr)); tao->gnorm0 = tao->residual; tao->residual = PetscSqrtReal(res[0] * res[0] + res[1] * res[1]); tao->cnorm = PetscSqrtReal(cnorm[0] * cnorm[0] + cnorm[1] * cnorm[1]); tao->step = pdipm->mu; PetscFunctionReturn(PETSC_SUCCESS); } /* PCPostSetup_PDIPM -- called when the KKT matrix is Cholesky factored for the preconditioner. Checks the inertia of Cholesky factor of the KKT matrix. If it does not match the numbers of prime and dual variables, add shifts to the KKT matrix. */ static PetscErrorCode PCPostSetUp_PDIPM(PC pc) { Tao tao; TAO_PDIPM *pdipm; Vec X; SNES snes; KSP ksp; Mat Factor; PetscBool isCHOL; PetscInt nneg, nzero, npos; PetscFunctionBegin; PetscCall(PCGetApplicationContext(pc, &tao)); pdipm = (TAO_PDIPM *)tao->data; X = pdipm->X; snes = pdipm->snes; /* Get the inertia of Cholesky factor */ PetscCall(SNESGetKSP(snes, &ksp)); PetscCall(KSPGetPC(ksp, &pc)); PetscCall(PetscObjectTypeCompare((PetscObject)pc, PCCHOLESKY, &isCHOL)); if (!isCHOL) PetscFunctionReturn(PETSC_SUCCESS); PetscCall(PCFactorGetMatrix(pc, &Factor)); PetscCall(MatGetInertia(Factor, &nneg, &nzero, &npos)); if (npos < pdipm->Nx + pdipm->Nci) { pdipm->deltaw = PetscMax(pdipm->lastdeltaw / 3, 1.e-4 * PETSC_MACHINE_EPSILON); PetscCall(PetscInfo(tao, "Test reduced deltaw=%g; previous MatInertia: nneg %" PetscInt_FMT ", nzero %" PetscInt_FMT ", npos %" PetscInt_FMT "(<%" PetscInt_FMT ")\n", (double)pdipm->deltaw, nneg, nzero, npos, pdipm->Nx + pdipm->Nci)); PetscCall(TaoSNESJacobian_PDIPM(snes, X, pdipm->K, pdipm->K, tao)); PetscCall(PCSetPostSetUp(pc, NULL)); PetscCall(PCSetUp(pc)); PetscCall(MatGetInertia(Factor, &nneg, &nzero, &npos)); if (npos < pdipm->Nx + pdipm->Nci) { pdipm->deltaw = pdipm->lastdeltaw; /* in case reduction update does not help, this prevents that step from impacting increasing update */ while (npos < pdipm->Nx + pdipm->Nci && pdipm->deltaw <= 1. / PETSC_SMALL) { /* increase deltaw */ PetscCall(PetscInfo(tao, " deltaw=%g fails, MatInertia: nneg %" PetscInt_FMT ", nzero %" PetscInt_FMT ", npos %" PetscInt_FMT "(<%" PetscInt_FMT ")\n", (double)pdipm->deltaw, nneg, nzero, npos, pdipm->Nx + pdipm->Nci)); pdipm->deltaw = PetscMin(8 * pdipm->deltaw, PetscPowReal(10, 20)); PetscCall(TaoSNESJacobian_PDIPM(snes, X, pdipm->K, pdipm->K, tao)); PetscCall(PCSetUp(pc)); PetscCall(MatGetInertia(Factor, &nneg, &nzero, &npos)); } PetscCheck(pdipm->deltaw < 1. / PETSC_SMALL, PetscObjectComm((PetscObject)tao), PETSC_ERR_CONV_FAILED, "Reached maximum delta w will not converge, try different initial x0"); PetscCall(PetscInfo(tao, "Updated deltaw %g\n", (double)pdipm->deltaw)); pdipm->lastdeltaw = pdipm->deltaw; pdipm->deltaw = 0.0; } } if (nzero) { /* Jacobian is singular */ if (pdipm->deltac == 0.0) { pdipm->deltac = PETSC_SQRT_MACHINE_EPSILON; } else { pdipm->deltac = pdipm->deltac * PetscPowReal(pdipm->mu, .25); } PetscCall(PetscInfo(tao, "Updated deltac=%g, MatInertia: nneg %" PetscInt_FMT ", nzero %" PetscInt_FMT "(!=0), npos %" PetscInt_FMT "\n", (double)pdipm->deltac, nneg, nzero, npos)); PetscCall(TaoSNESJacobian_PDIPM(snes, X, pdipm->K, pdipm->K, tao)); PetscCall(PCSetPostSetUp(pc, NULL)); PetscCall(PCSetUp(pc)); PetscCall(MatGetInertia(Factor, &nneg, &nzero, &npos)); } PetscCall(PCSetPostSetUp(pc, PCPostSetUp_PDIPM)); PetscFunctionReturn(PETSC_SUCCESS); } /* SNESLineSearch_PDIPM - Custom line search used with PDIPM. Collective Notes: This routine employs a simple backtracking line-search to keep the slack variables (z) and inequality constraints Lagrange multipliers (lambdai) positive, i.e., z,lambdai >=0. It does this by calculating scalars alpha_p and alpha_d to keep z,lambdai non-negative. The decision (x), and the slack variables are updated as X = X - alpha_d*dx. The constraint multipliers are updated as Lambdai = Lambdai + alpha_p*dLambdai. The barrier parameter mu is also updated as mu = mu + z'lambdai/Nci */ static PetscErrorCode SNESLineSearch_PDIPM(SNESLineSearch linesearch, PetscCtx ctx) { Tao tao = (Tao)ctx; TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; SNES snes; Vec X, F, Y; PetscInt i, iter; PetscReal alpha_p = 1.0, alpha_d = 1.0, alpha[4]; PetscScalar *Xarr, *z, *lambdai, dot, *taosolarr; const PetscScalar *dXarr, *dz, *dlambdai; PetscFunctionBegin; PetscCall(SNESLineSearchGetSNES(linesearch, &snes)); PetscCall(SNESGetIterationNumber(snes, &iter)); PetscCall(SNESLineSearchSetReason(linesearch, SNES_LINESEARCH_SUCCEEDED)); PetscCall(SNESLineSearchGetVecs(linesearch, &X, &F, &Y, NULL, NULL)); PetscCall(VecGetArrayWrite(X, &Xarr)); PetscCall(VecGetArrayRead(Y, &dXarr)); z = Xarr + pdipm->off_z; dz = dXarr + pdipm->off_z; for (i = 0; i < pdipm->nci; i++) { if (z[i] - dz[i] < 0.0) alpha_p = PetscMin(alpha_p, 0.9999 * z[i] / dz[i]); } lambdai = Xarr + pdipm->off_lambdai; dlambdai = dXarr + pdipm->off_lambdai; for (i = 0; i < pdipm->nci; i++) { if (lambdai[i] - dlambdai[i] < 0.0) alpha_d = PetscMin(0.9999 * lambdai[i] / dlambdai[i], alpha_d); } alpha[0] = alpha_p; alpha[1] = alpha_d; PetscCall(VecRestoreArrayRead(Y, &dXarr)); PetscCall(VecRestoreArrayWrite(X, &Xarr)); /* alpha = min(alpha) over all processes */ PetscCallMPI(MPIU_Allreduce(alpha, alpha + 2, 2, MPIU_REAL, MPIU_MIN, PetscObjectComm((PetscObject)tao))); alpha_p = alpha[2]; alpha_d = alpha[3]; /* X = X - alpha * Y */ PetscCall(VecGetArrayWrite(X, &Xarr)); PetscCall(VecGetArrayRead(Y, &dXarr)); for (i = 0; i < pdipm->nx; i++) Xarr[i] -= alpha_p * dXarr[i]; for (i = 0; i < pdipm->nce; i++) Xarr[i + pdipm->off_lambdae] -= alpha_d * dXarr[i + pdipm->off_lambdae]; for (i = 0; i < pdipm->nci; i++) { Xarr[i + pdipm->off_lambdai] -= alpha_d * dXarr[i + pdipm->off_lambdai]; Xarr[i + pdipm->off_z] -= alpha_p * dXarr[i + pdipm->off_z]; } PetscCall(VecGetArrayWrite(tao->solution, &taosolarr)); PetscCall(PetscArraycpy(taosolarr, Xarr, pdipm->nx)); PetscCall(VecRestoreArrayWrite(tao->solution, &taosolarr)); PetscCall(VecRestoreArrayWrite(X, &Xarr)); PetscCall(VecRestoreArrayRead(Y, &dXarr)); /* Update mu = mu_update_factor * dot(z,lambdai)/pdipm->nci at updated X */ if (pdipm->z) PetscCall(VecDot(pdipm->z, pdipm->lambdai, &dot)); else dot = 0.0; /* if (PetscAbsReal(pdipm->gradL) < 0.9*pdipm->mu) */ pdipm->mu = pdipm->mu_update_factor * dot / pdipm->Nci; /* Update F; get tao->residual and tao->cnorm */ PetscCall(TaoSNESFunction_PDIPM_residual(snes, X, F, (void *)tao)); tao->niter++; PetscCall(TaoLogConvergenceHistory(tao, pdipm->obj, tao->residual, tao->cnorm, tao->niter)); PetscCall(TaoMonitor(tao, tao->niter, pdipm->obj, tao->residual, tao->cnorm, pdipm->mu)); PetscUseTypeMethod(tao, convergencetest, tao->cnvP); if (tao->reason) PetscCall(SNESSetConvergedReason(snes, SNES_CONVERGED_FNORM_ABS)); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode TaoSolve_PDIPM(Tao tao) { TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; SNESLineSearch linesearch; /* SNESLineSearch context */ Vec dummy; PetscFunctionBegin; PetscCheck(tao->constraints_equality || tao->constraints_inequality, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_NULL, "Equality and inequality constraints are not set. Either set them or switch to a different algorithm"); /* Initialize all variables */ PetscCall(TaoPDIPMInitializeSolution(tao)); /* Set linesearch */ PetscCall(SNESGetLineSearch(pdipm->snes, &linesearch)); PetscCall(SNESLineSearchSetType(linesearch, SNESLINESEARCHSHELL)); PetscCall(SNESLineSearchShellSetApply(linesearch, SNESLineSearch_PDIPM, tao)); PetscCall(SNESLineSearchSetFromOptions(linesearch)); tao->reason = TAO_CONTINUE_ITERATING; /* -tao_monitor for iteration 0 and check convergence */ PetscCall(VecDuplicate(pdipm->X, &dummy)); PetscCall(TaoSNESFunction_PDIPM_residual(pdipm->snes, pdipm->X, dummy, (void *)tao)); PetscCall(TaoLogConvergenceHistory(tao, pdipm->obj, tao->residual, tao->cnorm, tao->niter)); PetscCall(TaoMonitor(tao, tao->niter, pdipm->obj, tao->residual, tao->cnorm, pdipm->mu)); PetscCall(VecDestroy(&dummy)); PetscUseTypeMethod(tao, convergencetest, tao->cnvP); if (tao->reason) PetscCall(SNESSetConvergedReason(pdipm->snes, SNES_CONVERGED_FNORM_ABS)); while (tao->reason == TAO_CONTINUE_ITERATING) { SNESConvergedReason reason; PetscCall(SNESSolve(pdipm->snes, NULL, pdipm->X)); /* Check SNES convergence */ PetscCall(SNESGetConvergedReason(pdipm->snes, &reason)); if (reason < 0) PetscCall(PetscPrintf(PetscObjectComm((PetscObject)pdipm->snes), "SNES solve did not converged due to reason %s\n", SNESConvergedReasons[reason])); /* Check TAO convergence */ PetscCheck(!PetscIsInfOrNanReal(pdipm->obj), PETSC_COMM_SELF, PETSC_ERR_SUP, "User-provided compute function generated infinity or NaN"); } PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode TaoView_PDIPM(Tao tao, PetscViewer viewer) { TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; PetscFunctionBegin; tao->constrained = PETSC_TRUE; PetscCall(PetscViewerASCIIPushTab(viewer)); PetscCall(PetscViewerASCIIPrintf(viewer, "Number of prime=%" PetscInt_FMT ", Number of dual=%" PetscInt_FMT "\n", pdipm->Nx + pdipm->Nci, pdipm->Nce + pdipm->Nci)); if (pdipm->kkt_pd) PetscCall(PetscViewerASCIIPrintf(viewer, "KKT shifts deltaw=%g, deltac=%g\n", (double)pdipm->deltaw, (double)pdipm->deltac)); PetscCall(PetscViewerASCIIPopTab(viewer)); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode TaoSetup_PDIPM(Tao tao) { TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; MPI_Comm comm; PetscMPIInt size; PetscInt row, col, Jcrstart, Jcrend, k, tmp, nc, proc, *nh_all, *ng_all; PetscInt offset, *xa, *xb, i, j, rstart, rend; PetscScalar one = 1.0, neg_one = -1.0; const PetscInt *cols, *rranges, *cranges, *aj, *ranges; const PetscScalar *aa, *Xarr; Mat J; Mat Jce_xfixed_trans, Jci_xb_trans; PetscInt *dnz, *onz, rjstart, nx_all, *nce_all, *Jranges, cols1[2]; PetscFunctionBegin; PetscCall(PetscObjectGetComm((PetscObject)tao, &comm)); PetscCallMPI(MPI_Comm_size(comm, &size)); /* (1) Setup Bounds and create Tao vectors */ PetscCall(TaoPDIPMSetUpBounds(tao)); if (!tao->gradient) { PetscCall(VecDuplicate(tao->solution, &tao->gradient)); PetscCall(VecDuplicate(tao->solution, &tao->stepdirection)); } /* (2) Get sizes */ /* Size of vector x - This is set by TaoSetSolution */ PetscCall(VecGetSize(tao->solution, &pdipm->Nx)); PetscCall(VecGetLocalSize(tao->solution, &pdipm->nx)); /* Size of equality constraints and vectors */ if (tao->constraints_equality) { PetscCall(VecGetSize(tao->constraints_equality, &pdipm->Ng)); PetscCall(VecGetLocalSize(tao->constraints_equality, &pdipm->ng)); } else { pdipm->ng = pdipm->Ng = 0; } pdipm->nce = pdipm->ng + pdipm->nxfixed; pdipm->Nce = pdipm->Ng + pdipm->Nxfixed; /* Size of inequality constraints and vectors */ if (tao->constraints_inequality) { PetscCall(VecGetSize(tao->constraints_inequality, &pdipm->Nh)); PetscCall(VecGetLocalSize(tao->constraints_inequality, &pdipm->nh)); } else { pdipm->nh = pdipm->Nh = 0; } pdipm->nci = pdipm->nh + pdipm->nxlb + pdipm->nxub + 2 * pdipm->nxbox; pdipm->Nci = pdipm->Nh + pdipm->Nxlb + pdipm->Nxub + 2 * pdipm->Nxbox; /* Full size of the KKT system to be solved */ pdipm->n = pdipm->nx + pdipm->nce + 2 * pdipm->nci; pdipm->N = pdipm->Nx + pdipm->Nce + 2 * pdipm->Nci; /* (3) Offsets for subvectors */ pdipm->off_lambdae = pdipm->nx; pdipm->off_lambdai = pdipm->off_lambdae + pdipm->nce; pdipm->off_z = pdipm->off_lambdai + pdipm->nci; /* (4) Create vectors and subvectors */ /* Ce and Ci vectors */ PetscCall(VecCreate(comm, &pdipm->ce)); PetscCall(VecSetSizes(pdipm->ce, pdipm->nce, pdipm->Nce)); PetscCall(VecSetFromOptions(pdipm->ce)); PetscCall(VecCreate(comm, &pdipm->ci)); PetscCall(VecSetSizes(pdipm->ci, pdipm->nci, pdipm->Nci)); PetscCall(VecSetFromOptions(pdipm->ci)); /* X=[x; lambdae; lambdai; z] for the big KKT system */ PetscCall(VecCreate(comm, &pdipm->X)); PetscCall(VecSetSizes(pdipm->X, pdipm->n, pdipm->N)); PetscCall(VecSetFromOptions(pdipm->X)); /* Subvectors; they share local arrays with X */ PetscCall(VecGetArrayRead(pdipm->X, &Xarr)); /* x shares local array with X.x */ if (pdipm->Nx) PetscCall(VecCreateMPIWithArray(comm, 1, pdipm->nx, pdipm->Nx, Xarr, &pdipm->x)); /* lambdae shares local array with X.lambdae */ if (pdipm->Nce) PetscCall(VecCreateMPIWithArray(comm, 1, pdipm->nce, pdipm->Nce, Xarr + pdipm->off_lambdae, &pdipm->lambdae)); /* tao->DE shares local array with X.lambdae_g */ if (pdipm->Ng) { PetscCall(VecCreateMPIWithArray(comm, 1, pdipm->ng, pdipm->Ng, Xarr + pdipm->off_lambdae, &tao->DE)); PetscCall(VecCreate(comm, &pdipm->lambdae_xfixed)); PetscCall(VecSetSizes(pdipm->lambdae_xfixed, pdipm->nxfixed, PETSC_DECIDE)); PetscCall(VecSetFromOptions(pdipm->lambdae_xfixed)); } if (pdipm->Nci) { /* lambdai shares local array with X.lambdai */ PetscCall(VecCreateMPIWithArray(comm, 1, pdipm->nci, pdipm->Nci, Xarr + pdipm->off_lambdai, &pdipm->lambdai)); /* z for slack variables; it shares local array with X.z */ PetscCall(VecCreateMPIWithArray(comm, 1, pdipm->nci, pdipm->Nci, Xarr + pdipm->off_z, &pdipm->z)); } /* tao->DI which shares local array with X.lambdai_h */ if (pdipm->Nh) PetscCall(VecCreateMPIWithArray(comm, 1, pdipm->nh, pdipm->Nh, Xarr + pdipm->off_lambdai, &tao->DI)); PetscCall(VecCreate(comm, &pdipm->lambdai_xb)); PetscCall(VecSetSizes(pdipm->lambdai_xb, pdipm->nci - pdipm->nh, PETSC_DECIDE)); PetscCall(VecSetFromOptions(pdipm->lambdai_xb)); PetscCall(VecRestoreArrayRead(pdipm->X, &Xarr)); /* (5) Create Jacobians Jce_xfixed and Jci */ /* (5.1) PDIPM Jacobian of equality bounds cebound(x) = J_nxfixed */ if (pdipm->Nxfixed) { /* Create Jce_xfixed */ PetscCall(MatCreate(comm, &pdipm->Jce_xfixed)); PetscCall(MatSetSizes(pdipm->Jce_xfixed, pdipm->nxfixed, pdipm->nx, PETSC_DECIDE, pdipm->Nx)); PetscCall(MatSetFromOptions(pdipm->Jce_xfixed)); PetscCall(MatSeqAIJSetPreallocation(pdipm->Jce_xfixed, 1, NULL)); PetscCall(MatMPIAIJSetPreallocation(pdipm->Jce_xfixed, 1, NULL, 1, NULL)); PetscCall(MatGetOwnershipRange(pdipm->Jce_xfixed, &Jcrstart, &Jcrend)); PetscCall(ISGetIndices(pdipm->isxfixed, &cols)); k = 0; for (row = Jcrstart; row < Jcrend; row++) { PetscCall(MatSetValues(pdipm->Jce_xfixed, 1, &row, 1, cols + k, &one, INSERT_VALUES)); k++; } PetscCall(ISRestoreIndices(pdipm->isxfixed, &cols)); PetscCall(MatAssemblyBegin(pdipm->Jce_xfixed, MAT_FINAL_ASSEMBLY)); PetscCall(MatAssemblyEnd(pdipm->Jce_xfixed, MAT_FINAL_ASSEMBLY)); } /* (5.2) PDIPM inequality Jacobian Jci = [tao->jacobian_inequality; ...] */ PetscCall(MatCreate(comm, &pdipm->Jci_xb)); PetscCall(MatSetSizes(pdipm->Jci_xb, pdipm->nci - pdipm->nh, pdipm->nx, PETSC_DECIDE, pdipm->Nx)); PetscCall(MatSetFromOptions(pdipm->Jci_xb)); PetscCall(MatSeqAIJSetPreallocation(pdipm->Jci_xb, 1, NULL)); PetscCall(MatMPIAIJSetPreallocation(pdipm->Jci_xb, 1, NULL, 1, NULL)); PetscCall(MatGetOwnershipRange(pdipm->Jci_xb, &Jcrstart, &Jcrend)); offset = Jcrstart; if (pdipm->Nxub) { /* Add xub to Jci_xb */ PetscCall(ISGetIndices(pdipm->isxub, &cols)); k = 0; for (row = offset; row < offset + pdipm->nxub; row++) { PetscCall(MatSetValues(pdipm->Jci_xb, 1, &row, 1, cols + k, &neg_one, INSERT_VALUES)); k++; } PetscCall(ISRestoreIndices(pdipm->isxub, &cols)); } if (pdipm->Nxlb) { /* Add xlb to Jci_xb */ PetscCall(ISGetIndices(pdipm->isxlb, &cols)); k = 0; offset += pdipm->nxub; for (row = offset; row < offset + pdipm->nxlb; row++) { PetscCall(MatSetValues(pdipm->Jci_xb, 1, &row, 1, cols + k, &one, INSERT_VALUES)); k++; } PetscCall(ISRestoreIndices(pdipm->isxlb, &cols)); } /* Add xbox to Jci_xb */ if (pdipm->Nxbox) { PetscCall(ISGetIndices(pdipm->isxbox, &cols)); k = 0; offset += pdipm->nxlb; for (row = offset; row < offset + pdipm->nxbox; row++) { PetscCall(MatSetValues(pdipm->Jci_xb, 1, &row, 1, cols + k, &neg_one, INSERT_VALUES)); tmp = row + pdipm->nxbox; PetscCall(MatSetValues(pdipm->Jci_xb, 1, &tmp, 1, cols + k, &one, INSERT_VALUES)); k++; } PetscCall(ISRestoreIndices(pdipm->isxbox, &cols)); } PetscCall(MatAssemblyBegin(pdipm->Jci_xb, MAT_FINAL_ASSEMBLY)); PetscCall(MatAssemblyEnd(pdipm->Jci_xb, MAT_FINAL_ASSEMBLY)); /* PetscCall(MatView(pdipm->Jci_xb,PETSC_VIEWER_STDOUT_WORLD)); */ /* (6) Set up ISs for PC Fieldsplit */ if (pdipm->solve_reduced_kkt) { PetscCall(PetscMalloc2(pdipm->nx + pdipm->nce, &xa, 2 * pdipm->nci, &xb)); for (i = 0; i < pdipm->nx + pdipm->nce; i++) xa[i] = i; for (i = 0; i < 2 * pdipm->nci; i++) xb[i] = pdipm->off_lambdai + i; PetscCall(ISCreateGeneral(comm, pdipm->nx + pdipm->nce, xa, PETSC_OWN_POINTER, &pdipm->is1)); PetscCall(ISCreateGeneral(comm, 2 * pdipm->nci, xb, PETSC_OWN_POINTER, &pdipm->is2)); } /* (7) Gather offsets from all processes */ PetscCall(PetscMalloc1(size, &pdipm->nce_all)); /* Get rstart of KKT matrix */ PetscCallMPI(MPI_Scan(&pdipm->n, &rstart, 1, MPIU_INT, MPI_SUM, comm)); rstart -= pdipm->n; PetscCallMPI(MPI_Allgather(&pdipm->nce, 1, MPIU_INT, pdipm->nce_all, 1, MPIU_INT, comm)); PetscCall(PetscMalloc3(size, &ng_all, size, &nh_all, size, &Jranges)); PetscCallMPI(MPI_Allgather(&rstart, 1, MPIU_INT, Jranges, 1, MPIU_INT, comm)); PetscCallMPI(MPI_Allgather(&pdipm->nh, 1, MPIU_INT, nh_all, 1, MPIU_INT, comm)); PetscCallMPI(MPI_Allgather(&pdipm->ng, 1, MPIU_INT, ng_all, 1, MPIU_INT, comm)); PetscCall(MatGetOwnershipRanges(tao->hessian, &rranges)); PetscCall(MatGetOwnershipRangesColumn(tao->hessian, &cranges)); if (pdipm->Ng) { PetscCall(TaoComputeJacobianEquality(tao, tao->solution, tao->jacobian_equality, tao->jacobian_equality_pre)); PetscCall(MatTranspose(tao->jacobian_equality, MAT_INITIAL_MATRIX, &pdipm->jac_equality_trans)); } if (pdipm->Nh) { PetscCall(TaoComputeJacobianInequality(tao, tao->solution, tao->jacobian_inequality, tao->jacobian_inequality_pre)); PetscCall(MatTranspose(tao->jacobian_inequality, MAT_INITIAL_MATRIX, &pdipm->jac_inequality_trans)); } /* Count dnz,onz for preallocation of KKT matrix */ nce_all = pdipm->nce_all; if (pdipm->Nxfixed) PetscCall(MatTranspose(pdipm->Jce_xfixed, MAT_INITIAL_MATRIX, &Jce_xfixed_trans)); PetscCall(MatTranspose(pdipm->Jci_xb, MAT_INITIAL_MATRIX, &Jci_xb_trans)); MatPreallocateBegin(comm, pdipm->n, pdipm->n, dnz, onz); /* 1st row block of KKT matrix: [Wxx; gradCe'; -gradCi'; 0] */ PetscCall(TaoPDIPMEvaluateFunctionsAndJacobians(tao, pdipm->x)); PetscCall(TaoComputeHessian(tao, tao->solution, tao->hessian, tao->hessian_pre)); /* Insert tao->hessian */ PetscCall(MatGetOwnershipRange(tao->hessian, &rjstart, NULL)); for (i = 0; i < pdipm->nx; i++) { row = rstart + i; PetscCall(MatGetRow(tao->hessian, i + rjstart, &nc, &aj, NULL)); proc = 0; for (j = 0; j < nc; j++) { while (aj[j] >= cranges[proc + 1]) proc++; col = aj[j] - cranges[proc] + Jranges[proc]; PetscCall(MatPreallocateSet(row, 1, &col, dnz, onz)); } PetscCall(MatRestoreRow(tao->hessian, i + rjstart, &nc, &aj, NULL)); if (pdipm->ng) { /* Insert grad g' */ PetscCall(MatGetRow(pdipm->jac_equality_trans, i + rjstart, &nc, &aj, NULL)); PetscCall(MatGetOwnershipRanges(tao->jacobian_equality, &ranges)); proc = 0; for (j = 0; j < nc; j++) { /* find row ownership of */ while (aj[j] >= ranges[proc + 1]) proc++; nx_all = rranges[proc + 1] - rranges[proc]; col = aj[j] - ranges[proc] + Jranges[proc] + nx_all; PetscCall(MatPreallocateSet(row, 1, &col, dnz, onz)); } PetscCall(MatRestoreRow(pdipm->jac_equality_trans, i + rjstart, &nc, &aj, NULL)); } /* Insert Jce_xfixed^T' */ if (pdipm->nxfixed) { PetscCall(MatGetRow(Jce_xfixed_trans, i + rjstart, &nc, &aj, NULL)); PetscCall(MatGetOwnershipRanges(pdipm->Jce_xfixed, &ranges)); proc = 0; for (j = 0; j < nc; j++) { /* find row ownership of */ while (aj[j] >= ranges[proc + 1]) proc++; nx_all = rranges[proc + 1] - rranges[proc]; col = aj[j] - ranges[proc] + Jranges[proc] + nx_all + ng_all[proc]; PetscCall(MatPreallocateSet(row, 1, &col, dnz, onz)); } PetscCall(MatRestoreRow(Jce_xfixed_trans, i + rjstart, &nc, &aj, NULL)); } if (pdipm->nh) { /* Insert -grad h' */ PetscCall(MatGetRow(pdipm->jac_inequality_trans, i + rjstart, &nc, &aj, NULL)); PetscCall(MatGetOwnershipRanges(tao->jacobian_inequality, &ranges)); proc = 0; for (j = 0; j < nc; j++) { /* find row ownership of */ while (aj[j] >= ranges[proc + 1]) proc++; nx_all = rranges[proc + 1] - rranges[proc]; col = aj[j] - ranges[proc] + Jranges[proc] + nx_all + nce_all[proc]; PetscCall(MatPreallocateSet(row, 1, &col, dnz, onz)); } PetscCall(MatRestoreRow(pdipm->jac_inequality_trans, i + rjstart, &nc, &aj, NULL)); } /* Insert Jci_xb^T' */ PetscCall(MatGetRow(Jci_xb_trans, i + rjstart, &nc, &aj, NULL)); PetscCall(MatGetOwnershipRanges(pdipm->Jci_xb, &ranges)); proc = 0; for (j = 0; j < nc; j++) { /* find row ownership of */ while (aj[j] >= ranges[proc + 1]) proc++; nx_all = rranges[proc + 1] - rranges[proc]; col = aj[j] - ranges[proc] + Jranges[proc] + nx_all + nce_all[proc] + nh_all[proc]; PetscCall(MatPreallocateSet(row, 1, &col, dnz, onz)); } PetscCall(MatRestoreRow(Jci_xb_trans, i + rjstart, &nc, &aj, NULL)); } /* 2nd Row block of KKT matrix: [grad Ce, deltac*I, 0, 0] */ if (pdipm->Ng) { PetscCall(MatGetOwnershipRange(tao->jacobian_equality, &rjstart, NULL)); for (i = 0; i < pdipm->ng; i++) { row = rstart + pdipm->off_lambdae + i; PetscCall(MatGetRow(tao->jacobian_equality, i + rjstart, &nc, &aj, NULL)); proc = 0; for (j = 0; j < nc; j++) { while (aj[j] >= cranges[proc + 1]) proc++; col = aj[j] - cranges[proc] + Jranges[proc]; PetscCall(MatPreallocateSet(row, 1, &col, dnz, onz)); /* grad g */ } PetscCall(MatRestoreRow(tao->jacobian_equality, i + rjstart, &nc, &aj, NULL)); } } /* Jce_xfixed */ if (pdipm->Nxfixed) { PetscCall(MatGetOwnershipRange(pdipm->Jce_xfixed, &Jcrstart, NULL)); for (i = 0; i < (pdipm->nce - pdipm->ng); i++) { row = rstart + pdipm->off_lambdae + pdipm->ng + i; PetscCall(MatGetRow(pdipm->Jce_xfixed, i + Jcrstart, &nc, &cols, NULL)); PetscCheck(nc == 1, PETSC_COMM_SELF, PETSC_ERR_SUP, "nc != 1"); proc = 0; j = 0; while (cols[j] >= cranges[proc + 1]) proc++; col = cols[j] - cranges[proc] + Jranges[proc]; PetscCall(MatPreallocateSet(row, 1, &col, dnz, onz)); PetscCall(MatRestoreRow(pdipm->Jce_xfixed, i + Jcrstart, &nc, &cols, NULL)); } } /* 3rd Row block of KKT matrix: [ gradCi, 0, deltac*I, -I] */ if (pdipm->Nh) { PetscCall(MatGetOwnershipRange(tao->jacobian_inequality, &rjstart, NULL)); for (i = 0; i < pdipm->nh; i++) { row = rstart + pdipm->off_lambdai + i; PetscCall(MatGetRow(tao->jacobian_inequality, i + rjstart, &nc, &aj, NULL)); proc = 0; for (j = 0; j < nc; j++) { while (aj[j] >= cranges[proc + 1]) proc++; col = aj[j] - cranges[proc] + Jranges[proc]; PetscCall(MatPreallocateSet(row, 1, &col, dnz, onz)); /* grad h */ } PetscCall(MatRestoreRow(tao->jacobian_inequality, i + rjstart, &nc, &aj, NULL)); } /* I */ for (i = 0; i < pdipm->nh; i++) { row = rstart + pdipm->off_lambdai + i; col = rstart + pdipm->off_z + i; PetscCall(MatPreallocateSet(row, 1, &col, dnz, onz)); } } /* Jci_xb */ PetscCall(MatGetOwnershipRange(pdipm->Jci_xb, &Jcrstart, NULL)); for (i = 0; i < (pdipm->nci - pdipm->nh); i++) { row = rstart + pdipm->off_lambdai + pdipm->nh + i; PetscCall(MatGetRow(pdipm->Jci_xb, i + Jcrstart, &nc, &cols, NULL)); PetscCheck(nc == 1, PETSC_COMM_SELF, PETSC_ERR_SUP, "nc != 1"); proc = 0; for (j = 0; j < nc; j++) { while (cols[j] >= cranges[proc + 1]) proc++; col = cols[j] - cranges[proc] + Jranges[proc]; PetscCall(MatPreallocateSet(row, 1, &col, dnz, onz)); } PetscCall(MatRestoreRow(pdipm->Jci_xb, i + Jcrstart, &nc, &cols, NULL)); /* I */ col = rstart + pdipm->off_z + pdipm->nh + i; PetscCall(MatPreallocateSet(row, 1, &col, dnz, onz)); } /* 4-th Row block of KKT matrix: Z and Ci */ for (i = 0; i < pdipm->nci; i++) { row = rstart + pdipm->off_z + i; cols1[0] = rstart + pdipm->off_lambdai + i; cols1[1] = row; PetscCall(MatPreallocateSet(row, 2, cols1, dnz, onz)); } /* diagonal entry */ for (i = 0; i < pdipm->n; i++) dnz[i]++; /* diagonal entry */ /* Create KKT matrix */ PetscCall(MatCreate(comm, &J)); PetscCall(MatSetSizes(J, pdipm->n, pdipm->n, PETSC_DECIDE, PETSC_DECIDE)); PetscCall(MatSetFromOptions(J)); PetscCall(MatSeqAIJSetPreallocation(J, 0, dnz)); PetscCall(MatMPIAIJSetPreallocation(J, 0, dnz, 0, onz)); MatPreallocateEnd(dnz, onz); pdipm->K = J; /* (8) Insert constant entries to K */ /* Set 0.0 to diagonal of K, so that the solver does not complain *about missing diagonal value */ PetscCall(MatGetOwnershipRange(J, &rstart, &rend)); for (i = rstart; i < rend; i++) PetscCall(MatSetValue(J, i, i, 0.0, INSERT_VALUES)); /* In case Wxx has no diagonal entries preset set diagonal to deltaw given */ if (pdipm->kkt_pd) { for (i = 0; i < pdipm->nh; i++) { row = rstart + i; PetscCall(MatSetValue(J, row, row, pdipm->deltaw, INSERT_VALUES)); } } /* Row block of K: [ grad Ce, 0, 0, 0] */ if (pdipm->Nxfixed) { PetscCall(MatGetOwnershipRange(pdipm->Jce_xfixed, &Jcrstart, NULL)); for (i = 0; i < (pdipm->nce - pdipm->ng); i++) { row = rstart + pdipm->off_lambdae + pdipm->ng + i; PetscCall(MatGetRow(pdipm->Jce_xfixed, i + Jcrstart, &nc, &cols, &aa)); proc = 0; for (j = 0; j < nc; j++) { while (cols[j] >= cranges[proc + 1]) proc++; col = cols[j] - cranges[proc] + Jranges[proc]; PetscCall(MatSetValue(J, row, col, aa[j], INSERT_VALUES)); /* grad Ce */ PetscCall(MatSetValue(J, col, row, aa[j], INSERT_VALUES)); /* grad Ce' */ } PetscCall(MatRestoreRow(pdipm->Jce_xfixed, i + Jcrstart, &nc, &cols, &aa)); } } /* Row block of K: [ -grad Ci, 0, 0, I] */ PetscCall(MatGetOwnershipRange(pdipm->Jci_xb, &Jcrstart, NULL)); for (i = 0; i < pdipm->nci - pdipm->nh; i++) { row = rstart + pdipm->off_lambdai + pdipm->nh + i; PetscCall(MatGetRow(pdipm->Jci_xb, i + Jcrstart, &nc, &cols, &aa)); proc = 0; for (j = 0; j < nc; j++) { while (cols[j] >= cranges[proc + 1]) proc++; col = cols[j] - cranges[proc] + Jranges[proc]; PetscCall(MatSetValue(J, col, row, -aa[j], INSERT_VALUES)); PetscCall(MatSetValue(J, row, col, -aa[j], INSERT_VALUES)); } PetscCall(MatRestoreRow(pdipm->Jci_xb, i + Jcrstart, &nc, &cols, &aa)); col = rstart + pdipm->off_z + pdipm->nh + i; PetscCall(MatSetValue(J, row, col, 1, INSERT_VALUES)); } for (i = 0; i < pdipm->nh; i++) { row = rstart + pdipm->off_lambdai + i; col = rstart + pdipm->off_z + i; PetscCall(MatSetValue(J, row, col, 1, INSERT_VALUES)); } /* Row block of K: [ 0, 0, I, ...] */ for (i = 0; i < pdipm->nci; i++) { row = rstart + pdipm->off_z + i; col = rstart + pdipm->off_lambdai + i; PetscCall(MatSetValue(J, row, col, 1, INSERT_VALUES)); } if (pdipm->Nxfixed) PetscCall(MatDestroy(&Jce_xfixed_trans)); PetscCall(MatDestroy(&Jci_xb_trans)); PetscCall(PetscFree3(ng_all, nh_all, Jranges)); /* (9) Set up nonlinear solver SNES */ PetscCall(SNESSetFunction(pdipm->snes, NULL, TaoSNESFunction_PDIPM, (void *)tao)); PetscCall(SNESSetJacobian(pdipm->snes, J, J, TaoSNESJacobian_PDIPM, (void *)tao)); if (pdipm->solve_reduced_kkt) { PC pc; PetscCall(KSPGetPC(tao->ksp, &pc)); PetscCall(PCSetType(pc, PCFIELDSPLIT)); PetscCall(PCFieldSplitSetType(pc, PC_COMPOSITE_SCHUR)); PetscCall(PCFieldSplitSetIS(pc, "2", pdipm->is2)); PetscCall(PCFieldSplitSetIS(pc, "1", pdipm->is1)); } PetscCall(SNESSetFromOptions(pdipm->snes)); /* (10) Setup PCPostSetUp() for pdipm->solve_symmetric_kkt */ if (pdipm->solve_symmetric_kkt) { KSP ksp; PC pc; PetscBool isCHOL; PetscCall(SNESGetKSP(pdipm->snes, &ksp)); PetscCall(KSPGetPC(ksp, &pc)); PetscCall(PCSetPostSetUp(pc, PCPostSetUp_PDIPM)); PetscCall(PetscObjectTypeCompare((PetscObject)pc, PCCHOLESKY, &isCHOL)); if (isCHOL) { Mat Factor; PetscCheck(PetscDefined(HAVE_MUMPS), PetscObjectComm((PetscObject)tao), PETSC_ERR_SUP, "Requires external package MUMPS"); PetscCall(PCFactorGetMatrix(pc, &Factor)); PetscCall(MatMumpsSetIcntl(Factor, 24, 1)); /* detection of null pivot rows */ if (size > 1) PetscCall(MatMumpsSetIcntl(Factor, 13, 1)); /* parallelism of the root node (enable ScaLAPACK) and its splitting */ } } PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode TaoDestroy_PDIPM(Tao tao) { TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; PetscFunctionBegin; /* Freeing Vectors assocaiated with KKT (X) */ PetscCall(VecDestroy(&pdipm->x)); /* Solution x */ PetscCall(VecDestroy(&pdipm->lambdae)); /* Equality constraints lagrangian multiplier*/ PetscCall(VecDestroy(&pdipm->lambdai)); /* Inequality constraints lagrangian multiplier*/ PetscCall(VecDestroy(&pdipm->z)); /* Slack variables */ PetscCall(VecDestroy(&pdipm->X)); /* Big KKT system vector [x; lambdae; lambdai; z] */ /* work vectors */ PetscCall(VecDestroy(&pdipm->lambdae_xfixed)); PetscCall(VecDestroy(&pdipm->lambdai_xb)); /* Legrangian equality and inequality Vec */ PetscCall(VecDestroy(&pdipm->ce)); /* Vec of equality constraints */ PetscCall(VecDestroy(&pdipm->ci)); /* Vec of inequality constraints */ /* Matrices */ PetscCall(MatDestroy(&pdipm->Jce_xfixed)); PetscCall(MatDestroy(&pdipm->Jci_xb)); /* Jacobian of inequality constraints Jci = [tao->jacobian_inequality ; J(nxub); J(nxlb); J(nxbx)] */ PetscCall(MatDestroy(&pdipm->K)); /* Index Sets */ if (pdipm->Nxub) PetscCall(ISDestroy(&pdipm->isxub)); /* Finite upper bound only -inf < x < ub */ if (pdipm->Nxlb) PetscCall(ISDestroy(&pdipm->isxlb)); /* Finite lower bound only lb <= x < inf */ if (pdipm->Nxfixed) PetscCall(ISDestroy(&pdipm->isxfixed)); /* Fixed variables lb = x = ub */ if (pdipm->Nxbox) PetscCall(ISDestroy(&pdipm->isxbox)); /* Boxed variables lb <= x <= ub */ if (pdipm->Nxfree) PetscCall(ISDestroy(&pdipm->isxfree)); /* Free variables -inf <= x <= inf */ if (pdipm->solve_reduced_kkt) { PetscCall(ISDestroy(&pdipm->is1)); PetscCall(ISDestroy(&pdipm->is2)); } /* SNES */ PetscCall(SNESDestroy(&pdipm->snes)); /* Nonlinear solver */ PetscCall(PetscFree(pdipm->nce_all)); PetscCall(MatDestroy(&pdipm->jac_equality_trans)); PetscCall(MatDestroy(&pdipm->jac_inequality_trans)); /* Destroy pdipm */ PetscCall(PetscFree(tao->data)); /* Holding locations of pdipm */ /* Destroy Dual */ PetscCall(VecDestroy(&tao->DE)); /* equality dual */ PetscCall(VecDestroy(&tao->DI)); /* dinequality dual */ PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode TaoSetFromOptions_PDIPM(Tao tao, PetscOptionItems PetscOptionsObject) { TAO_PDIPM *pdipm = (TAO_PDIPM *)tao->data; PetscFunctionBegin; PetscOptionsHeadBegin(PetscOptionsObject, "PDIPM method for constrained optimization"); PetscCall(PetscOptionsReal("-tao_pdipm_push_init_slack", "parameter to push initial slack variables away from bounds", NULL, pdipm->push_init_slack, &pdipm->push_init_slack, NULL)); PetscCall(PetscOptionsReal("-tao_pdipm_push_init_lambdai", "parameter to push initial (inequality) dual variables away from bounds", NULL, pdipm->push_init_lambdai, &pdipm->push_init_lambdai, NULL)); PetscCall(PetscOptionsBool("-tao_pdipm_solve_reduced_kkt", "Solve reduced KKT system using Schur-complement", NULL, pdipm->solve_reduced_kkt, &pdipm->solve_reduced_kkt, NULL)); PetscCall(PetscOptionsReal("-tao_pdipm_mu_update_factor", "Update scalar for barrier parameter (mu) update", NULL, pdipm->mu_update_factor, &pdipm->mu_update_factor, NULL)); PetscCall(PetscOptionsBool("-tao_pdipm_symmetric_kkt", "Solve non reduced symmetric KKT system", NULL, pdipm->solve_symmetric_kkt, &pdipm->solve_symmetric_kkt, NULL)); PetscCall(PetscOptionsBool("-tao_pdipm_kkt_shift_pd", "Add shifts to make KKT matrix positive definite", NULL, pdipm->kkt_pd, &pdipm->kkt_pd, NULL)); PetscOptionsHeadEnd(); PetscFunctionReturn(PETSC_SUCCESS); } /*MC TAOPDIPM - Barrier-based primal-dual interior point algorithm for generally constrained optimization. Options Database Keys: + -tao_pdipm_push_init_lambdai - parameter to push initial dual variables away from bounds (> 0) . -tao_pdipm_push_init_slack - parameter to push initial slack variables away from bounds (> 0) . -tao_pdipm_mu_update_factor - update scalar for barrier parameter (mu) update (> 0) . -tao_pdipm_symmetric_kkt - Solve non-reduced symmetric KKT system - -tao_pdipm_kkt_shift_pd - Add shifts to make KKT matrix positive definite Level: beginner .seealso: `TAOPDIPM`, `Tao`, `TaoType` M*/ PETSC_EXTERN PetscErrorCode TaoCreate_PDIPM(Tao tao) { TAO_PDIPM *pdipm; PC pc; PetscFunctionBegin; tao->ops->setup = TaoSetup_PDIPM; tao->ops->solve = TaoSolve_PDIPM; tao->ops->setfromoptions = TaoSetFromOptions_PDIPM; tao->ops->view = TaoView_PDIPM; tao->ops->destroy = TaoDestroy_PDIPM; PetscCall(PetscNew(&pdipm)); tao->data = (void *)pdipm; pdipm->nx = pdipm->Nx = 0; pdipm->nxfixed = pdipm->Nxfixed = 0; pdipm->nxlb = pdipm->Nxlb = 0; pdipm->nxub = pdipm->Nxub = 0; pdipm->nxbox = pdipm->Nxbox = 0; pdipm->nxfree = pdipm->Nxfree = 0; pdipm->ng = pdipm->Ng = pdipm->nce = pdipm->Nce = 0; pdipm->nh = pdipm->Nh = pdipm->nci = pdipm->Nci = 0; pdipm->n = pdipm->N = 0; pdipm->mu = 1.0; pdipm->mu_update_factor = 0.1; pdipm->deltaw = 0.0; pdipm->lastdeltaw = 3 * 1.e-4; pdipm->deltac = 0.0; pdipm->kkt_pd = PETSC_FALSE; pdipm->push_init_slack = 1.0; pdipm->push_init_lambdai = 1.0; pdipm->solve_reduced_kkt = PETSC_FALSE; pdipm->solve_symmetric_kkt = PETSC_TRUE; /* Override default settings (unless already changed) */ PetscCall(TaoParametersInitialize(tao)); PetscObjectParameterSetDefault(tao, max_it, 200); PetscObjectParameterSetDefault(tao, max_funcs, 500); PetscCall(SNESCreate(((PetscObject)tao)->comm, &pdipm->snes)); PetscCall(SNESSetOptionsPrefix(pdipm->snes, tao->hdr.prefix)); PetscCall(SNESGetKSP(pdipm->snes, &tao->ksp)); PetscCall(PetscObjectReference((PetscObject)tao->ksp)); PetscCall(KSPGetPC(tao->ksp, &pc)); PetscCall(PCSetApplicationContext(pc, (void *)tao)); PetscFunctionReturn(PETSC_SUCCESS); }