<html><head><meta http-equiv="content-type" content="text/html; charset=us-ascii"></head><body style="overflow-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;"><div><br></div> Ah, there is <a href="https://petsc.org/release/manualpages/Mat/MATSOLVERSPQR/#matsolverspqr">https://petsc.org/release/manualpages/Mat/MATSOLVERSPQR/#matsolverspqr</a> See also <a href="https://petsc.org/release/manualpages/Mat/MatGetFactor/#matgetfactor">https://petsc.org/release/manualpages/Mat/MatGetFactor/#matgetfactor</a> and <a href="https://petsc.org/release/manualpages/Mat/MatQRFactorSymbolic/">https://petsc.org/release/manualpages/Mat/MatQRFactorSymbolic/</a> <div><br></div><div><br><div><br><blockquote type="cite"><div>On Aug 29, 2023, at 1:17 PM, Jed Brown <jed@jedbrown.org> wrote:</div><br class="Apple-interchange-newline"><div><div>Suitesparse includes a sparse QR algorithm. The main issue is that (even with pivoting) the R factor has the same nonzero structure as a Cholesky factor of A^T A, which is generally much denser than a factor of A, and this degraded sparsity impacts Q as well.<br><br>I wonder if someone would like to contribute a sparse QR to PETSc. It could have a default implementation via Cholesky QR and the ability to call SPQR from Suitesparse.<br><br>Barry Smith <bsmith@petsc.dev> writes:<br><br><blockquote type="cite"> Are the nonzero structures of all the rows related? If they are, one could devise a routine to take advantage of this relationship, but if the nonzero structures of each row are "randomly" different from all the other rows, then it is difficult to see how one can take advantage of the sparsity.<br><br><br><br><blockquote type="cite">On Aug 29, 2023, at 12:50 PM, Thanasis Boutsikakis <thanasis.boutsikakis@corintis.com> wrote:<br><br>Hi all, I have the following code that orthogonalizes a PETSc matrix. The problem is that this implementation requires that the PETSc matrix is dense, otherwise, it fails at bv.SetFromOptions(). Hence the assert in orthogonality().<br><br>What could I do in order to be able to orthogonalize sparse matrices as well? Could I convert it efficiently? (I tried to no avail)<br><br>Thanks!<br><br>"""Experimenting with matrix orthogonalization"""<br><br>import contextlib<br>import sys<br>import time<br>import numpy as np<br>from firedrake import COMM_WORLD<br>from firedrake.petsc import PETSc<br><br>import slepc4py<br><br>slepc4py.init(sys.argv)<br>from slepc4py import SLEPc<br><br>from numpy.testing import assert_array_almost_equal<br><br>EPSILON_USER = 1e-4<br>EPS = sys.float_info.epsilon<br><br><br>def Print(message: str):<br> """Print function that prints only on rank 0 with color<br><br> Args:<br> message (str): message to be printed<br> """<br> PETSc.Sys.Print(message)<br><br><br>def create_petsc_matrix(input_array, sparse=True):<br> """Create a PETSc matrix from an input_array<br><br> Args:<br> input_array (np array): Input array<br> partition_like (PETSc mat, optional): Petsc matrix. Defaults to None.<br> sparse (bool, optional): Toggle for sparese or dense. Defaults to True.<br><br> Returns:<br> PETSc mat: PETSc matrix<br> """<br> # Check if input_array is 1D and reshape if necessary<br> assert len(input_array.shape) == 2, "Input array should be 2-dimensional"<br> global_rows, global_cols = input_array.shape<br><br> size = ((None, global_rows), (global_cols, global_cols))<br><br> # Create a sparse or dense matrix based on the 'sparse' argument<br> if sparse:<br> matrix = PETSc.Mat().createAIJ(size=size, comm=COMM_WORLD)<br> else:<br> matrix = PETSc.Mat().createDense(size=size, comm=COMM_WORLD)<br> matrix.setUp()<br><br> local_rows_start, local_rows_end = matrix.getOwnershipRange()<br><br> for counter, i in enumerate(range(local_rows_start, local_rows_end)):<br> # Calculate the correct row in the array for the current process<br> row_in_array = counter + local_rows_start<br> matrix.setValues(<br> i, range(global_cols), input_array[row_in_array, :], addv=False<br> )<br><br> # Assembly the matrix to compute the final structure<br> matrix.assemblyBegin()<br> matrix.assemblyEnd()<br><br> return matrix<br><br><br>def orthogonality(A): # sourcery skip: avoid-builtin-shadow<br> """Checking and correcting orthogonality<br><br> Args:<br> A (PETSc.Mat): Matrix of size [m x k].<br><br> Returns:<br> PETSc.Mat: Matrix of size [m x k].<br> """<br> # Check if the matrix is dense<br> mat_type = A.getType()<br> assert mat_type in (<br> "seqdense",<br> "mpidense",<br> ), "A must be a dense matrix. SLEPc.BV().createFromMat() requires a dense matrix."<br><br> m, k = A.getSize()<br><br> Phi1 = A.getColumnVector(0)<br> Phi2 = A.getColumnVector(k - 1)<br><br> # Compute dot product using PETSc function<br> dot_product = Phi1.dot(Phi2)<br><br> if abs(dot_product) > min(EPSILON_USER, EPS * m):<br> Print(" Matrix is not orthogonal")<br><br> # Type can be CHOL, GS, mro(), SVQB, TSQR, TSQRCHOL<br> _type = SLEPc.BV().OrthogBlockType.GS<br><br> bv = SLEPc.BV().createFromMat(A)<br> bv.setFromOptions()<br> bv.setOrthogonalization(_type)<br> bv.orthogonalize()<br><br> A = bv.createMat()<br><br> Print(" Matrix successfully orthogonalized")<br><br> # # Assembly the matrix to compute the final structure<br> if not A.assembled:<br> A.assemblyBegin()<br> A.assemblyEnd()<br> else:<br> Print(" Matrix is orthogonal")<br><br> return A<br><br><br># --------------------------------------------<br># EXP: Orthogonalization of an mpi PETSc matrix<br># --------------------------------------------<br><br>m, k = 11, 7<br># Generate the random numpy matrices<br>np.random.seed(0) # sets the seed to 0<br>A_np = np.random.randint(low=0, high=6, size=(m, k))<br><br>A = create_petsc_matrix(A_np, sparse=False)<br><br>A_orthogonal = orthogonality(A)<br><br># --------------------------------------------<br># TEST: Orthogonalization of a numpy matrix<br># --------------------------------------------<br># Generate A_np_orthogonal<br>A_np_orthogonal, _ = np.linalg.qr(A_np)<br><br># Get the local values from A_orthogonal<br>local_rows_start, local_rows_end = A_orthogonal.getOwnershipRange()<br>A_orthogonal_local = A_orthogonal.getValues(<br> range(local_rows_start, local_rows_end), range(k)<br>)<br><br># Assert the correctness of the multiplication for the local subset<br>assert_array_almost_equal(<br> np.abs(A_orthogonal_local),<br> np.abs(A_np_orthogonal[local_rows_start:local_rows_end, :]),<br> decimal=5,<br>)<br></blockquote></blockquote></div></div></blockquote></div><br></div></body></html>