Adding fast CUDA bindings to Standard ML
For more information, check out our website at https://rohany.github.io/StandardML-GPU
To see SML-GPU in action, clone the repository with git clone https://github.com/rohany/StandardML-GPU/
.
Install the dependencies
sudo apt-get install mlton
Follow the Nvidia CUDA setup guide to install CUDA. Then head over to the tests directory, and make sure that cuda path in the Makefiles is the same as the path to your installation. Also ensure that your CUDA compute compatability in the lib makefile is the highest that that of your card supports.
Then just run make, ./<executable> <input size>
, and see the speedups!