First time posting in the forums, so apologies (and please educate me) if I make any faux-pas.
Also, something of a newcomer to GPU computing, so again the same caveats! :)
Setup: I’m running on a Macbook Pro under Snow Leopard (OS X 10.6.8)
Problem & Current Deductions:
I have some python code that relies on the old CUBLAS API (i.e. from cublas.h rather than the newer cublas_v2.h).
This code is trying to get some of the functions it needs from libcublas.dylib.
However, running: nm libcublas.dylib | grep cublas
shows me that symbol table is missing entries from the legacy API (i.e. the functions present are the newer *_v2 ones.)
I know that the “right” solution is to update the python bindings, but unfortunately I don’t have the budget for that at present. Also, rolling back the CUDA version isn’t an option because of other dependencies on newer CUDA functionality.
As a fast interim solution, I thought maybe I could roll my own dylib with the necessary functionality built in, and then just use this in the python bindings instead. Unfortunately, I don’t know how to do this!
Does this sound like a reasonable solution? (And if not, an explanation of why would be very much appreciated.)
Could anyone give me any pointers as to the best way to go about building such a library?
Many thanks in advance!