On Sun, 20 Jul 2008 09:54:39 -0700, mabshoff <[EMAIL PROTECTED]> wrote:
> On Jul 20, 9:21 am, Simon Beaumont <[EMAIL PROTECTED]> wrote: > > Hi Simon, > >> I am just about to embark on integrating come CUDA libraries into >> sage. I was not sure of the best route to go - I am considering the >> pycuda libraries as a starting point - this a pure kernel approach - I >> but would also like to get the CUDA blas and fft libraries integrated. > > Various people have started looking into this, but so far no one has > produced code. One big issue (at least for me) with pycuda is the > requirement for boost, but I am not sure how that could be overcome. > Since I am personally only interested in the CUDA BLAS the suggested > way to hook it into Sage was directly via Cython, but recreating > pycuda via Cython seems like a large waste of effort and should be > avoided at all costs. This will boil down to what you're trying to accomplish. Exposing the full CUDA api to python might be a large task. OTOH, building a capability using C/C++ CUDA and exposing a simplified API to SAGE through Cython might be straightforward and avoid some of the issues below. Given the problems CUDA targets and the performance characteristics of the GPU/CPU/memory integration, its almost certain that a great deal of that interface needn't and shouldn't be pushed up through Cython into Python. > > The main issue with boost I see is that PolyBoRi ships with a subset > of boost and installs it into $SAGE_LOCAL/include/boost. I assume that > it will not be enough of boost, i.e. boost python is not part of it. > Since PolyBoRi also has an interface to Python using boost Python we > might be able to add the bits needed to the polybori.spkg, otherwise I > see potentially huge trouble by colliding boost versions in the tree. > And shipping boost itself is not really an option due to its rather > large size. > >> (I think cuda-python) can do this. I'm sure I'm not the first down >> this road and wondered which would be the most useful. I'd also >> appreciate some tips and pointers into integration of sage arrays and >> matrices to make this as native as possible. > > I think using numpy arrays here for now might be the way to go, but > that limits you to numpy data types. Since we are talking about > numerical computations it seems that we will not lose functionality > here at all. There's a GSOC activity looking at Python's buffer interface which is coordinated with the Cython activity http://wiki.cython.org/enhancements/buffer and is related to PEP 3118: http://www.python.org/dev/peps/pep-3118/ Travis Oliphant of Numpy fame is the author. Putting a buffer api on top of CUDA might be a good place to start and given its broad use integrating C/C++ features throughout Python, you might get a lot of capability for free. For example, mmap files expose a buffer interface. The new buffer interface will likely provide those characteristics you seek from Numpy and Cython looks to be very interested in how to make it all work. > >> Of course this work would >> be shared with the community. I have plans to make some CUDA hardware >> available over the web using sage and have also have some longer term >> plans for a modeling environment based on it. > > Cool. There are various people waiting to but the next generation of > Tesla hardware, i.e. the one that actually provides IEEE doubles. I am > basically waiting for it to become available since the things I am > interested in do require IEEE precision and close to IEEE is not > enough. > >> Any advice and pointers most welcome. > > Please keep us up to date how things are going and let us know about > any problems you have. This is an area we should definitely make some > progress this year. > >> Regards, >> >> Simon > > Cheers, > > Michael > > -- Glenn H. Tarbox, PhD || 206-494-0819 || [EMAIL PROTECTED] "Don't worry about people stealing your ideas. If your ideas are any good you'll have to ram them down peoples throats" -- Howard Aiken --~--~---------~--~----~------------~-------~--~----~ To post to this group, send email to sage-devel@googlegroups.com To unsubscribe from this group, send email to [EMAIL PROTECTED] For more options, visit this group at http://groups.google.com/group/sage-devel URLs: http://www.sagemath.org -~----------~----~----~----~------~----~------~--~---
