matlab link interface

[cstring]

[kyle]

[cstring]

[cstring]

[kyle]

Make sure you are only using the 64-bit libraries. I see you still have a 32-bit folder listed in your example.

[cstring]

Thanks for the suggestion Kyle. I've tried double checking and I think I'm only including 64-bit libraries. Have you successfully used the matlab link interface on a mac?
Thanks,
Craig

[kyle]

We've successfully tested the link interface with OSX but we don't currently have access to a copy of MATLAB for Mac. I'll see if we find a trial copy to test further.

[chester1248]

cstring -- are you still having problems with cula link in matlab 2011a in OSX? It works for me, but on a Snow Leopard MBP using it's 330M GPU (no PCIe cards). If you are having problems still, what is your latest bash exports, according to "env"?

[rmgarnett]

[chester1248]

Roman, just to double check, maybe you should have your wrapper script (which sets the env vars and then starts matlab) to print out all the variable values (i.e. run "env" before "matlab ..." in your script). The problem you see could, for example, happen if you set the environ variables as you show, but those were not the var settings when matlab is started up.

Alternatively, double check within matlab by calling the matlab getenv() function, for those variables.

If they all report as you showed in your posting, then I would next suspect that CULA isn't actually (fully) installed in the default directories as you might think. So you might want to verify that the CULA library files are in the paths you expected?

Anyway, just in case you didn't already check these things ... good to eliminate the obvious stuff before something more complex that might be going on ...

[rmgarnett]

[john]

You could deactivate or uninstall the Matlab parallel computing toolkit (PCT). CULA's link interface is completely incompatible with PCT, for the reasons you are seeing here - their CUDA version is far behind. Since Matlab can't load both CUDA toolkits, then you need to choose which you prefer.

On Windows, we've seen that you can work around this by ensuring your first CULA call comes before your first PCT call. In that case, the CUDA runtime in your DYLD_LIBRARY_PATH will be loaded rather than the one in PCT. On OSX, your mileage may vary.