Awesome

hmglib - Hierarchical matrices on GPU(s) library

This library provides an entirely GPU-based implementation of the hierarchical (H) matrix approach. H matrices approximate specific dense matrices, e.g., from discretized integral equations or kernel ridge regression, leading to log-linear time complexity in dense matrix-vector products. The library specifically allows to construct and apply an H matrix on GPU(s). To get high performance on GPU(s), techniques like parallel tree traversal, space filling curves and batched operations are applied.

Getting Started

After cloning the repository, go to the source directory and adapt the makefile to your needs.

Prerequisites

The following list of libraries should allow you to compile the example codes.

• Nvidia GPU with CUDA support
• standard development tools (GCC, make, ...)
• recent version of the CUDA Toolkit
• recent version of OpenBLAS
• recent version of MAGMA
• recent version of GSL (GNU Scientific Library)

Compilation of Example Code

After appropriately modifying the makefile, go to the source directory and compile the code by running

make

Example Code

The file hmglib_test.cu is an example code that allows to construct an H matrix for a dense matrix from kernel ridge regression with a Gaussian kernel and random sampling points. After the construction of the H matrix, it runs the H matrix - vector product and compares the solution of this product to a full matrix - vector product. Finally, the error of this produc is reported.

Running the Example Code

Run the example code by entering

./hmglib_test <Nx> <Ny> <k> <c_leaf> <exponent of epsilon> <eta>

A typical (small!) example would be:

./hmglib_test 4096 4096 16 256 -3 1.0

Here, a matrix of 4096 x 4096 is approximated with adaptive cross approximation (ACA) with 16 low-rank terms, a leaf size of 256 and a balancing parameter of 1.0. The exponent of epsilon is currently unused.

How to go further ?

It is first suggested to read reference [1], found in the literature section of this page. To construct H matrices for other matrix types than a Gaussian kernel matrix, an interested user should try to derive a new system assembler class, providing a method to evaluate a given entry of the (dense) matrix. In the example code from above, the implemented class is called gaussian_kernel_system_assembler.

Current State of the Project

The current state of the software project is that it is an experimental code, which was used to create the two manuscripts referenced below. hmglib is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License in LICENSE for more details.

Authors

• Peter Zaspel - Main developer - University of Basel

License

This project is licensed under the LGPL License version 3.0 - see the LICENSE file for details.

Literature

1. P. Zaspel. Algorithmic patterns for H matrices on many-core processors, Preprint 2017-12, Fachbereich Mathematik, Universität Basel, Switzerland, 2017. Also available as arXiv preprint arXiv:1708.09707.
2. H. Harbrecht and P. Zaspel. A scalable H-matrix approach for the solution of boundary integral equations on multi-GPU clusters. Preprint 2018-11, Fachbereich Mathematik, Universität Basel, Switzerland, 2018. Available as arXiv preprint arXiv:1806.11558.