#include /*I "petscmat.h" I*/ PETSC_EXTERN PetscErrorCode VecGetRootType_Private(Vec, VecType *); typedef struct { Mat A; /* sparse matrix */ Mat U, V; /* dense tall-skinny matrices */ Vec c; /* sequential vector containing the diagonal of C */ Vec work1, work2; /* sequential vectors that hold partial products */ Vec xl, yl; /* auxiliary sequential vectors for matmult operation */ } Mat_LRC; static PetscErrorCode MatMult_LRC_kernel(Mat N, Vec x, Vec y, PetscBool transpose) { Mat_LRC *Na = (Mat_LRC *)N->data; PetscMPIInt size; Mat U, V; PetscFunctionBegin; U = transpose ? Na->V : Na->U; V = transpose ? Na->U : Na->V; PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)N), &size)); if (size == 1) { PetscCall(MatMultHermitianTranspose(V, x, Na->work1)); if (Na->c) PetscCall(VecPointwiseMult(Na->work1, Na->c, Na->work1)); if (Na->A) { if (transpose) { PetscCall(MatMultTranspose(Na->A, x, y)); } else { PetscCall(MatMult(Na->A, x, y)); } PetscCall(MatMultAdd(U, Na->work1, y, y)); } else { PetscCall(MatMult(U, Na->work1, y)); } } else { Mat Uloc, Vloc; Vec yl, xl; const PetscScalar *w1; PetscScalar *w2; PetscInt nwork; xl = transpose ? Na->yl : Na->xl; yl = transpose ? Na->xl : Na->yl; PetscCall(VecGetLocalVector(y, yl)); PetscCall(MatDenseGetLocalMatrix(U, &Uloc)); PetscCall(MatDenseGetLocalMatrix(V, &Vloc)); /* multiply the local part of V with the local part of x */ PetscCall(VecGetLocalVectorRead(x, xl)); PetscCall(MatMultHermitianTranspose(Vloc, xl, Na->work1)); PetscCall(VecRestoreLocalVectorRead(x, xl)); /* form the sum of all the local multiplies: this is work2 = V'*x = sum_{all processors} work1 */ PetscCall(VecGetArrayRead(Na->work1, &w1)); PetscCall(VecGetArrayWrite(Na->work2, &w2)); PetscCall(VecGetLocalSize(Na->work1, &nwork)); PetscCallMPI(MPIU_Allreduce(w1, w2, nwork, MPIU_SCALAR, MPIU_SUM, PetscObjectComm((PetscObject)N))); PetscCall(VecRestoreArrayRead(Na->work1, &w1)); PetscCall(VecRestoreArrayWrite(Na->work2, &w2)); if (Na->c) { /* work2 = C*work2 */ PetscCall(VecPointwiseMult(Na->work2, Na->c, Na->work2)); } if (Na->A) { /* form y = A*x or A^t*x */ if (transpose) { PetscCall(MatMultTranspose(Na->A, x, y)); } else { PetscCall(MatMult(Na->A, x, y)); } /* multiply-add y = y + U*work2 */ PetscCall(MatMultAdd(Uloc, Na->work2, yl, yl)); } else { /* multiply y = U*work2 */ PetscCall(MatMult(Uloc, Na->work2, yl)); } PetscCall(VecRestoreLocalVector(y, yl)); } PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode MatMult_LRC(Mat N, Vec x, Vec y) { PetscFunctionBegin; PetscCall(MatMult_LRC_kernel(N, x, y, PETSC_FALSE)); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode MatMultTranspose_LRC(Mat N, Vec x, Vec y) { PetscFunctionBegin; PetscCall(MatMult_LRC_kernel(N, x, y, PETSC_TRUE)); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode MatDestroy_LRC(Mat N) { Mat_LRC *Na = (Mat_LRC *)N->data; PetscFunctionBegin; PetscCall(MatDestroy(&Na->A)); PetscCall(MatDestroy(&Na->U)); PetscCall(MatDestroy(&Na->V)); PetscCall(VecDestroy(&Na->c)); PetscCall(VecDestroy(&Na->work1)); PetscCall(VecDestroy(&Na->work2)); PetscCall(VecDestroy(&Na->xl)); PetscCall(VecDestroy(&Na->yl)); PetscCall(PetscFree(N->data)); PetscCall(PetscObjectComposeFunction((PetscObject)N, "MatLRCGetMats_C", NULL)); PetscCall(PetscObjectComposeFunction((PetscObject)N, "MatLRCSetMats_C", NULL)); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode MatLRCGetMats_LRC(Mat N, Mat *A, Mat *U, Vec *c, Mat *V) { Mat_LRC *Na = (Mat_LRC *)N->data; PetscFunctionBegin; if (A) *A = Na->A; if (U) *U = Na->U; if (c) *c = Na->c; if (V) *V = Na->V; PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode MatLRCSetMats_LRC(Mat N, Mat A, Mat U, Vec c, Mat V) { Mat_LRC *Na = (Mat_LRC *)N->data; PetscFunctionBegin; PetscCall(PetscObjectReference((PetscObject)A)); PetscCall(PetscObjectReference((PetscObject)U)); PetscCall(PetscObjectReference((PetscObject)V)); PetscCall(PetscObjectReference((PetscObject)c)); PetscCall(MatDestroy(&Na->A)); PetscCall(MatDestroy(&Na->U)); PetscCall(MatDestroy(&Na->V)); PetscCall(VecDestroy(&Na->c)); Na->A = A; Na->U = U; Na->c = c; Na->V = V; PetscFunctionReturn(PETSC_SUCCESS); } /*@ MatLRCGetMats - Returns the constituents of an LRC matrix Not collective Input Parameter: . N - matrix of type `MATLRC` Output Parameters: + A - the (sparse) matrix . U - first dense rectangular (tall and skinny) matrix . c - a sequential vector containing the diagonal of C - V - second dense rectangular (tall and skinny) matrix Level: intermediate Notes: The returned matrices should not be destroyed by the caller. `U`, `c`, `V` may be `NULL` if not needed .seealso: [](ch_matrices), `MatLRCSetMats()`, `Mat`, `MATLRC`, `MatCreateLRC()` @*/ PetscErrorCode MatLRCGetMats(Mat N, Mat *A, Mat *U, Vec *c, Mat *V) { PetscFunctionBegin; PetscUseMethod(N, "MatLRCGetMats_C", (Mat, Mat *, Mat *, Vec *, Mat *), (N, A, U, c, V)); PetscFunctionReturn(PETSC_SUCCESS); } /*@ MatLRCSetMats - Sets the constituents of an LRC matrix Logically collective Input Parameters: + N - matrix of type `MATLRC` . A - the (sparse) matrix . U - first dense rectangular (tall and skinny) matrix . c - a sequential vector containing the diagonal of C - V - second dense rectangular (tall and skinny) matrix Level: intermediate Note: If `V` is `NULL`, then it is assumed to be identical to `U`. .seealso: [](ch_matrices), `MatLRCGetMats()`, `Mat`, `MATLRC`, `MatCreateLRC()` @*/ PetscErrorCode MatLRCSetMats(Mat N, Mat A, Mat U, Vec c, Mat V) { PetscInt k, k1, m, n, m1, n1; PetscBool match; PetscFunctionBegin; if (A) PetscValidHeaderSpecific(A, MAT_CLASSID, 2); PetscValidHeaderSpecific(U, MAT_CLASSID, 3); if (c) PetscValidHeaderSpecific(c, VEC_CLASSID, 4); if (V) { PetscValidHeaderSpecific(V, MAT_CLASSID, 5); PetscCheckSameComm(U, 3, V, 5); } if (A) PetscCheckSameComm(A, 2, U, 3); if (!V) V = U; PetscCall(PetscObjectBaseTypeCompareAny((PetscObject)U, &match, MATSEQDENSE, MATMPIDENSE, "")); PetscCheck(match, PetscObjectComm((PetscObject)U), PETSC_ERR_SUP, "Matrix U must be of type dense, found %s", ((PetscObject)U)->type_name); PetscCall(PetscObjectBaseTypeCompareAny((PetscObject)V, &match, MATSEQDENSE, MATMPIDENSE, "")); PetscCheck(match, PetscObjectComm((PetscObject)U), PETSC_ERR_SUP, "Matrix V must be of type dense, found %s", ((PetscObject)V)->type_name); PetscCall(PetscStrcmp(U->defaultvectype, V->defaultvectype, &match)); PetscCheck(match, PetscObjectComm((PetscObject)U), PETSC_ERR_ARG_WRONG, "Matrix U and V must have the same VecType %s != %s", U->defaultvectype, V->defaultvectype); if (A) { PetscCall(PetscStrcmp(A->defaultvectype, U->defaultvectype, &match)); PetscCheck(match, PetscObjectComm((PetscObject)U), PETSC_ERR_ARG_WRONG, "Matrix A and U must have the same VecType %s != %s", A->defaultvectype, U->defaultvectype); } PetscCall(MatGetSize(U, NULL, &k)); PetscCall(MatGetSize(V, NULL, &k1)); PetscCheck(k == k1, PetscObjectComm((PetscObject)U), PETSC_ERR_ARG_INCOMP, "U and V have different number of columns (%" PetscInt_FMT " vs %" PetscInt_FMT ")", k, k1); PetscCall(MatGetLocalSize(U, &m, NULL)); PetscCall(MatGetLocalSize(V, &n, NULL)); if (A) { PetscCall(MatGetLocalSize(A, &m1, &n1)); PetscCheck(m == m1, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Local dimensions of U %" PetscInt_FMT " and A %" PetscInt_FMT " do not match", m, m1); PetscCheck(n == n1, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Local dimensions of V %" PetscInt_FMT " and A %" PetscInt_FMT " do not match", n, n1); } if (c) { PetscMPIInt size, csize; PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)U), &size)); PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)c), &csize)); PetscCall(VecGetSize(c, &k1)); PetscCheck(k == k1, PetscObjectComm((PetscObject)c), PETSC_ERR_ARG_INCOMP, "The length of c %" PetscInt_FMT " does not match the number of columns of U and V (%" PetscInt_FMT ")", k1, k); PetscCheck(csize == 1 || csize == size, PetscObjectComm((PetscObject)c), PETSC_ERR_ARG_INCOMP, "U and c must have the same communicator size %d != %d", size, csize); } PetscCall(MatSetSizes(N, m, n, PETSC_DECIDE, PETSC_DECIDE)); PetscUseMethod(N, "MatLRCSetMats_C", (Mat, Mat, Mat, Vec, Mat), (N, A, U, c, V)); PetscFunctionReturn(PETSC_SUCCESS); } static PetscErrorCode MatSetUp_LRC(Mat N) { Mat_LRC *Na = (Mat_LRC *)N->data; Mat A = Na->A; Mat U = Na->U; Mat V = Na->V; Vec c = Na->c; Mat Uloc; PetscMPIInt size, csize = 0; PetscBool sym = (PetscBool)(U == V), dummy; PetscFunctionBegin; PetscCall(MatSetVecType(N, U->defaultvectype)); // Flag matrix as symmetric if A is symmetric and U == V if (A && sym) PetscCall(MatIsSymmetricKnown(A, &dummy, &sym)); PetscCall(MatSetOption(N, MAT_SYMMETRIC, sym)); PetscCall(MatDenseGetLocalMatrix(Na->U, &Uloc)); PetscCall(MatCreateVecs(Uloc, &Na->work1, NULL)); PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)U), &size)); if (c) PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)c), &csize)); if (size != 1) { Mat Vloc; if (Na->c && csize != 1) { /* scatter parallel vector to sequential */ VecScatter sct; PetscCall(VecScatterCreateToAll(Na->c, &sct, &c)); PetscCall(VecScatterBegin(sct, Na->c, c, INSERT_VALUES, SCATTER_FORWARD)); PetscCall(VecScatterEnd(sct, Na->c, c, INSERT_VALUES, SCATTER_FORWARD)); PetscCall(VecScatterDestroy(&sct)); PetscCall(VecDestroy(&Na->c)); Na->c = c; } PetscCall(MatDenseGetLocalMatrix(Na->V, &Vloc)); PetscCall(VecDuplicate(Na->work1, &Na->work2)); PetscCall(MatCreateVecs(Vloc, NULL, &Na->xl)); PetscCall(MatCreateVecs(Uloc, NULL, &Na->yl)); } // Internally create a scaling vector if roottypes do not match if (Na->c) { VecType rt1, rt2; PetscBool match; PetscCall(VecGetRootType_Private(Na->work1, &rt1)); PetscCall(VecGetRootType_Private(Na->c, &rt2)); PetscCall(PetscStrcmp(rt1, rt2, &match)); if (!match) { PetscCall(VecDuplicate(Na->c, &c)); PetscCall(VecCopy(Na->c, c)); PetscCall(VecDestroy(&Na->c)); Na->c = c; } } N->assembled = PETSC_TRUE; N->preallocated = PETSC_TRUE; PetscFunctionReturn(PETSC_SUCCESS); } PETSC_EXTERN PetscErrorCode MatCreate_LRC(Mat N) { Mat_LRC *Na; PetscFunctionBegin; PetscCall(PetscObjectChangeTypeName((PetscObject)N, MATLRC)); PetscCall(PetscNew(&Na)); N->data = (void *)Na; N->ops->destroy = MatDestroy_LRC; N->ops->setup = MatSetUp_LRC; N->ops->mult = MatMult_LRC; N->ops->multtranspose = MatMultTranspose_LRC; PetscCall(PetscObjectComposeFunction((PetscObject)N, "MatLRCGetMats_C", MatLRCGetMats_LRC)); PetscCall(PetscObjectComposeFunction((PetscObject)N, "MatLRCSetMats_C", MatLRCSetMats_LRC)); PetscFunctionReturn(PETSC_SUCCESS); } /*MC MATLRC - "lrc" - a matrix object that behaves like A + U*C*V' Note: The matrix A + U*C*V' is not formed! Rather the matrix object performs the matrix-vector product `MatMult()`, by first multiplying by A and then adding the other term. Level: advanced .seealso: [](ch_matrices), `Mat`, `MatCreateLRC()`, `MatMult()`, `MatLRCGetMats()`, `MatLRCSetMats()` M*/ /*@ MatCreateLRC - Creates a new matrix object that behaves like A + U*C*V' of type `MATLRC` Collective Input Parameters: + A - the (sparse) matrix (can be `NULL`) . U - dense rectangular (tall and skinny) matrix . V - dense rectangular (tall and skinny) matrix - c - a vector containing the diagonal of C (can be `NULL`) Output Parameter: . N - the matrix that represents A + U*C*V' Level: intermediate Notes: The matrix A + U*C*V' is not formed! Rather the new matrix object performs the matrix-vector product `MatMult()`, by first multiplying by A and then adding the other term. `C` is a diagonal matrix (represented as a vector) of order k, where k is the number of columns of both `U` and `V`. If `A` is `NULL` then the new object behaves like a low-rank matrix U*C*V'. Use `V`=`U` (or `V`=`NULL`) for a symmetric low-rank correction, A + U*C*U'. If `c` is `NULL` then the low-rank correction is just U*V'. If a sequential `c` vector is used for a parallel matrix, PETSc assumes that the values of the vector are consistently set across processors. .seealso: [](ch_matrices), `Mat`, `MATLRC`, `MatLRCGetMats()` @*/ PetscErrorCode MatCreateLRC(Mat A, Mat U, Vec c, Mat V, Mat *N) { PetscFunctionBegin; PetscCall(MatCreate(PetscObjectComm((PetscObject)U), N)); PetscCall(MatSetType(*N, MATLRC)); PetscCall(MatLRCSetMats(*N, A, U, c, V)); PetscCall(MatSetUp(*N)); PetscFunctionReturn(PETSC_SUCCESS); }