On a more upbeat note: the only motivation for all this grief was to build pycuda - against boost - against a framework based python. I have not been so impressed with CUDA for basic matrix operations relative to a fast muti-core CPU on relatively small matrices (though I really like the pycuda approach in exposing the full CUDA kernel api in python for application development) it's not that important to me.
I can develop whatever complex kernels I like (in c) and plug them in wherever they are needed (sage or R or whatever). On the other hand ad- hoc I'd like to do basic linear algebra - so numpy works for me and if required (cublas) or better transparent access to a cublas implementation where pragma or a mop might opportunistically implement numpy would makes much more sense to me. How I now see this panning out (in my case) is two distinct use cases: 1. Specialized simulations or sophisticated applications (physics engines, itransformation engines etc.) - build a custom kernel in c and plug it in as a monolith with application specific api. 2. Routine linear algebra on an ad-hoc basis - use pycuda and have blas implement pragmatically or have a pragma/mop provide a numpy specialization. --~--~---------~--~----~------------~-------~--~----~ 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 -~----------~----~----~----~------~----~------~--~---