Provides some benchmarks for matrix math libraries.

1. Eigen
2. PyTorch
3. ArrayFire

1. Dense Chained Multiplication - Multiplies 6 dense matrices together in a chain
2. Dense Inversion - Inverts a dense matrix
3. Dense Multiplication - Multiplies 2 dense matrices together

Each benchmark is evaluated on square matrices with increasing matrix edge size

This package relies on several matrix math libraries as well as Google Benchmark. If you are on Ubuntu install these packages as follows

Install with the debian:

sudo apt install libbenchmark-dev

Install with the debian:

sudo apt install libeigen3-dev

1. Download the appropriate zip file from the PyTorch website: https://pytorch.org/get-started/locally/

2. Unzip the file and navigate to it in the terminal

3. Copy it into your install directory: sudo cp -r libtorch/ /usr/local/libtorch/

4. Add libtorch to your PATH and LD_LIBRARY_PATH in your ~/.bashrc file

export PATH="/usr/local/libtorch/share/cmake:$PATH"
export LD_LIBRARY_PATH=/usr/local/libtorch/lib:$LD_LIBRARY_PATH

For best results, follow the instructions on ArrayFire's website. For subpar results, follow my instructions

1. Download the installer from ArrayFire's website: http://arrayfire.org/docs/installing.htm

2. Install into your /opt directory: ./Arrayfire_*_Linux_x86_64.sh --include-subdir --prefix=/opt

3. Add ArrayFire to your PATH and LD_LIBRARY_PATH in your ~/.bashrc file

export PATH="/opt/arrayfire/share/ArrayFire/cmake:$PATH"
export LD_LIBRARY_PATH=/opt/arrayfire/lib64:$LD_LIBRARY_PATH

Absolutely no guarantee of fitness or optimality in these benchmarks. Use at your own risk.

Note: There are a few points of possible error here.

1. GPU calculations - I'm not sure that these benchmarks are accurate for calculations taking place on the GPU. In particular, I know that PyTorch states that GPU calculations happen asynchronously. I have not done any work to solve this
2. Lazy Execution - ArrayFire uses lazy execution. I am using the "benchmark::DoNotOptimize" tag, but I am not certain if that is enough to force it to run. It would appear that it is, since the time tends to increase as the size increases. I would not expect that if it was simply creating a computation graph