announcement_2011_05_18

To users of GPU devices for scientific computation,

Currently there are many implementations of numpy-inspired gpu ndarrays (n dimensional array). We are trying to regroup all of the effort into one project. We would like a base type to facilitate argument-passing in python, python compiled modules, and in other language too, so we are proposing one in C with Python bindings.

We created a mailing list [1] to discuss this. There is a start of an implementation at [2].

In our plan, the compyte project will become a dependency of Theano[3], PyOpenCL[4] and PyCuda[5]. These projects already implement large subsets of a gpu ndarray functionality (bindings for CUDA libraries, code-generation for elemwise and reduction computations). The plan is to refactor and combine some of that work and port it into compyte.

We encourage everybody interested in this work to subscribe to the mailing list [1] and keep an eye on the project wiki [2]. If you are aware of other efforts in this direction, please feel free to share them.

Thanks

• Frédéric Bastien
• Andreas Klochner (?)
• James Bergstra

[1] http://lists.tiker.net/listinfo/gpundarray

[2] https://github.com/inducer/compyte

[3] http://deeplearning.net/software/theano/

[4] http://documen.tician.de/pyopencl

[5] http://documen.tician.de/pycuda