Cucheb - CUDA accelerated large sparse eigensolvers

Jared L. Aurentz, Vassilis Kalantzis and Yousef Saad, 2017

GitHub

This README file is written in mark down. For the best experience please view this file, along with the rest of the Cucheb library on GitHub.

Introduction

Cucheb is a collection of C++ subroutines for accurately and efficiently solving large sparse matrix eigenvalue problems using NVIDIA brand GPUs. These methods are well suited for computing eigenvalues and eigenvectors of matrices arising in quantum physics, structural engineering and network analysis.

Current features

cucheb-v0.1.2 has the following features:

• double precision eigensolvers for real symmetric matrices

Dependencies

Cucheb depends on the NVIDIA CUDA Toolkit (v7.0 or greater).

User-level programs

There are many files in the Cucheb library but only a few will be necessary for most users. The first set of files are the objects used to store computed quantities, such as eigenvalues and eigenvectors. Below we give a brief description of each file and a link for further information:

• cuchebmatrix - object for storing sparse matrices
• cucheblanczos - object for storing computed eigenvalues and eigenvectors

The next set of files are programs used to initialize and delete objects and compute eigenvalues and eigenvectors:

• cuchebmatrix_init - initializes a cuchebmatrix object using a sparse matrix stored in Matrix Market Format
• cuchebmatrix_destroy - frees all memory associated with an instance of a cuchebmatrix object
• cuchebmatrix_print - prints basic propertied of an instance of a cuchebmatrix object
• cuchebmatrix_lanczos - computes all eigenvalues and eigenvectors in a user-defined interval using the Lanczos method and stores the output in a cucheblanczos object
• cuchebmatrix_filteredlanczos
  • computes all eigenvalues and eigenvectors in a user-defined interval using the filtered Lanczos procedure and stores the output in a cucheblanczos object
• cucheblanczos_init - initializes a cucheblanczos object
• cucheblanczos_destroy - frees all memory associated with an instance of a cucheblanczos object
• cucheblanczos_print - prints basic propertied of an instance of a cucheblanczos object

Installation

Cucheb is built on top of the NVIDIA CUDA Toolkit and a small number of C++ standard libraries. You must have the toolkit installed before the library can be built.

Linux

To install on a Linux machine simply move into the Cucheb root directory, edit the file make.inc to suit your system and type:

make install

This creates a shared object library libcucheb.so.version and copies it into the user specified installation directory. The installation does not create any symbolic links or export any library paths. To add cucheb to the library path type:

export LD_LIBRARY_PATH=/path/to/cucheb/lib:$LD_LIBRARY_PATH

Tests

You can find several examples for using Cucheb in the tests subdirectory. Once Cucheb has been installed the tests can be compiled by typing:

make tests

The test matrices are included as part of Cucheb but more can be found at the Sparse Matrix Collection.

First example

When using Cucheb for the first time we suggest that you start with testparsec.cu. This tests computes a subset of the ground states of a molecular Hamiltonian, which we consider a standard problem for Cucheb. To execute testparsec.cu once compiled type:

./tests/parsec/testparsec

CAUTION: When working on a cluster the relative path to the matrices in the test source files may not be exported properly. To fix this modify the specific test source file to use the absolute path to the matrix.

Removing Cucheb

If the source directory has not been removed simply move into the Cucheb root directory and type:

make uninstall

If the source directory has been removed the install directory will have to be removed explicitly by the user.

Questions and issues

If you have any questions or encounter any issues while using Cucheb please file an issue on the Cucheb issues page of Github.