On Thu, 2 Jun 2011 09:08:11 +0200, Bogdan Opanchuk <manti...@gmail.com> wrote: > > I guess there's one thing I hadn't considered when merging the CURAND > > patches, and that is what happens when using shipped boost. As the > > CURAND wrappers end up being a separate binary, they simply don't get > > access to boost, which they need. Not sure what to do. I'm not sure > > setuptools is up to the job of building a separated shared library for > > boost, or if the _curand binary can have access to the boost bits in > > _driver. Any shared library gurus care to comment? > > I am by no means a guru, but can't you just link boost statically to > _curand as you do for _driver? Is the increased size of the module > such a bad thing?
I seem to remember that boost.python gets upset when there are multiple copies of it in a single Python process. > > If worst comes to worst, we'll just shove the _curand wrappers back into > > the main PyCUDA wrapper binary. I've done just that, for lack of better ideas. > In the meantime the second issue should be fixed somehow, I think. > Currently I am using the following workaround: > > - def generate_direction_vectors(count, > direction=direction_vector_set.VECTOR_32): > + def generate_direction_vectors(count, direction=None): > + if direction is None: > + direction = direction_vector_set.VECTOR_32 This is a good idea no matter what, merged. Scott, Bogdan--can you check whether this works for you? Andreas
pgppB3B3ZP2Oa.pgp
Description: PGP signature
_______________________________________________ PyCUDA mailing list PyCUDA@tiker.net http://lists.tiker.net/listinfo/pycuda
