thilinarmtb / lsbench

Repository for benchmarking linear solvers on GPU.

Introduction

lsbench is a repository containing benchmark codes for various high performant linear solvers on CPUs and GPUs. It can read a sparse matrix and solve it using cuSparse, AmgX, Hypre or CHOLMOD. An example on how to use the API is in bin/driver.c. A few test matrices are available in tests/ directory. Currently, all the benchmarks are performend in double precision.

Building lsbench

lsbench by defaults download and builds CHOLMOD. You can enable/disable solvers (AmgX, cuSparse, etc.) when configuring the cmake build using ENABLE_<SOLVER>=ON|OFF.

Build requirements

• cmake (>= 3.18)
• CUDAToolkit (>= 11.0) if using AmgX, cuSparse or Hypre with Cuda backend)

Build instructions

1. Clone the repo first using git and cd into the repository:

git clone https://github.com/thilinarmtb/lsbench.git
cd lsbench

2. Then you can use cmake to build the benchmarks:

mkdir build && cd build 
cmake .. -DCMAKE_C_COMPILER=<c-compiler> -DCMAKE_CXX_COMPILER=<cxx-compiler> \
  -DCMAKE_INSTALL_PREFIX=<lsbench-install-dir> \
  -DENABLE_CUSPARSE=ON -DENABLE_AMGX=OFF
make -j8 install
cd -

You can add the bin directory to the PATH variable to access the binary without the full path:

export PATH=${PATH}:<lsbench-install-dir>/bin

Running the benchmarks

Once the benchmarks are built, they can be run using the following command: Do driver --help to see available options (not everything is implemented right now).

driver --solver cholmod --test ./tests/I1_05x05.txt --verbose=1