On Mittwoch 08 April 2009, Nicholas Tung wrote: > It might be more useful if it's a __device__ function / subroutine.
That would be nice--but I don't see how it could be done without recursive launches, which CUDA doesn't have at the moment. > There's > already CUDAPP for some library functions, but it seems unfortunately tied > into C/C++, and your code is much nicer looking :). C/C++ is not a problem, the ridiculous divide between driver and runtime API is. If any one of you get a chance to get on Nvidia's nerves about this, please do. > Also, maybe you should > consider starting a sub-project for developing library functions. I see your point, and I agree partly. I don't intend for PyCUDA to become 'the kitchen sink' for little pieces of CUDA helper code. There's clearly a limit. The GPUArray functionality, for example, could theoretically be sawed off and stuck in a separate package. However, at this point, I think that the cost of that would significantly outweigh any potential benefits, so it's going to stay in. Then, once you accept to have an array package, reductions are a pretty natural part of that. Andreas
signature.asc
Description: This is a digitally signed message part.
_______________________________________________ PyCuda mailing list [email protected] http://tiker.net/mailman/listinfo/pycuda_tiker.net
