Accelerate/optimize an open source package under OpenCL (or other).

There is a useful open implementation of an algorithm called clstm ([login to view URL]) it has only appeared recently. It has a sister pure python implementation in the Ocropy package (useful for testing). I am using this code for purely research purposes, but would like to see it accelerated, optimised or parallelised. It is a small package with two binaries, and a python wrapper. The author is still adding useful features, so some degree of forwards compatibility needs to be maintained - including maintaining the python wrapper(s). So in short, I would like a build-able version that exploits GPGPU and/or threads to speed up the training of the algorithm (its a neural network). The package can have defined input and outputs (a training and test set) and so has very definable completion criteria, i.e. the speeded up version gets the same answer only significantly quicker. I test on MacPro (for an example) and have seen up to 30X speed ups on similar software when ported to GPU. The authors style and architecture / command line / wrappers need to be maintained where possible. Conditional compilation, or a command line switch allowing optimisations to be activated/deactivated is required.