CUDA Toolkit and Driver 2.3b for OS X released IFFT_011209 nbody, madelbrot NOT compiled by errors &

i seen that Nvidia did offer new Mac OS X 2.3b SDK (CUDA C 10.5 + OpenCL 10.6) now.
I used before the 2.3a SDK.
Make common worked :)
Make C CUDA demos problems:
Now 3 C CUDA demos cant build anymore (compiler errors):
IFFT_011209, nbody, madelbrot - all other (> 15) build with success / Mac OSX 10.5.8
Would be nice if that could be looked at and fixed.

I only looked at the C CUDA things for now (10.5), not OpenCL (i will do that later after booting into 10.6 and install 2.3b SDK there also).

EDIT: The 2.3b Version is only for the SDK, the drivers are not updated also (still 2.3x). Sorry for the little wrong headline. (only SDK is updated, not also drivers)

I only had issues with the Mandelbrot and nbody demos not compiling. The Mondelbrot one had issues with the cu file, and the nbody had a long list of issues that I didn’t really investigate.

Incidentally the OpenCL demos continue to not compile for me in both the a and b revisions of the 2.3 SDK. I get errors about libGLEW.a not being of the required architecture. The error pops up for a few demos such as oclBoxFilter and oclMedianFilter but I guess those codes don’t actually USE libGLEW. nbody does though and I get a lot of undefined symbol errors all related to libGLEW. I’d like to see those fixed so I can do a proper comparison of OpenCL vs CUDA on the mac.

Messing around with the OpenCL demos in the latest SDK (and I assume this would work equally well on the 2.3a version), specifically the nbody demo, I found that the libGLEW.a library included in common/lib is 64bit only but the compiler is doing everything in 32 bits. This seems to be a function of the terminal session I’m running in any case the file does not care whether you’re using 32 or 64 bits, it always references the libGLEW.a archive which is 64 bits only.

Another quick mention. In the 2.3b version of the nbody demo they apparently are changing to a template design and it doesn’t work. If you back up to the version from 2.3a it DOES work though. That’s what I did and it’s SLOW. I’m not sure if that is a fault of their code (and that’s why they’re changing it in 2.3b?) or not. It seems that most stuff still runs 5x or more slower on OpenCL than it does in CUDA. Hope that helps someone else trying to make this work and couldn’t find anything.