Benchmarking sparse linear algebra libraries on CPU, GPU and FPGA.

Real-life sparse matrices from this collection are used.

You should have installed:

1. For CPU benchmarking:

• armadillo and its dependencies
• PetSC
• trilinos
• Note! For maximum performance these libraries usually require vendor specific libraries (e.g. Intel MKL) or optimized linear algebra packs (see their documentation for details)

2. For GPU benchmarking:

• CUDA and its dependecies
• cusp
• Note! Make sure you use a version of cusp that is compatible with CUDA (e.g. cusp 0.4.0 with CUDA 5.5)
• Note! cuda external libraries (such as cusp) are assumed to be installed in the user's home directory (~/cuda); you can specify a different directory using the CUDA_PATH Makefile variable

3. For FPGA benchmarking:

• coming soon...

4. Other:

• python2.7 (including the wget package)

After installing dependencies run:

git clone https://github.com/paul-g/sparse-bench.git`
cd sparse-bench/benchmark/<specific benchmark>
make
./<binary name>

That's it!

TODO: add a script to run all the benchmarks and produce some nice results. Coming soon...