Experiments for the paper Compositional Deep Learning in Futhark

There are two Futhark test programs: conv.fut and mlp.fut. To run then, first construct their input data files by running make. Then, assuming you have installed the Futhark compiler and have a working OpenCL setup, you can run the benchmarks:

futhark bench --backend=opencl conv.fut mlp.fut

Alternatively, to use the CUDA backend:

futhark bench --backend=cuda conv.fut mlp.fut

The two backends should perform roughly the same. We used the OpenCL backend for the paper.

These benchmarks were developed for Futhark 0.11.2. If you read this in the distant future, it is possible that the then-current version of the compiler may not be able to compile them anymore. If this happens, please tell us.

The library

The actual deeplearning library is properly its own GitHub repository, but embedded here in its entirety. Run futhark pkg upgrade && futhark pkg sync if you want to bump it to the newest version.