OSX 10.8 status of OpenMM

[Charles Brooks]

Hi All,

Has anyone had success in running any applications using OpenMM on Mac OSX 10.8? I do not seem to be able to do so. I understand from the earlier message that OpenCL will not work, but I cannot get Cuda to work either.

[Peter Eastman]

CUDA on 10.8 is working for me. What problems are you having?

Peter

[Charles Brooks]

Hi Peter,

I am running CHARMM through our CHARMM/OpenMM interface on a Macbook Pro w/ NVIDIA GeForce GT 650M 1024 MB card. I have the cuda_5.0.36_macos.pkg installed and try with both gfortran (GNU Fortran (GCC) 4.6.2) and ifort (IFORT 11.1 20100401). Both work on reference platform but throw a floating point exception for attempts to run on the Cuda platform. I do not get a traceback from gfortran, but ifort yields the following:

libc++abi.dylib: terminate called throwing an exception
forrtl: error (76): Abort trap signal
Image PC Routine Line Source
libsystem_kernel. 92183A6A Unknown Unknown Unknown
libsystem_c.dylib 9029A4F8 Unknown Unknown Unknown
libc++abi.dylib 95D4280C Unknown Unknown Unknown
libc++abi.dylib 95D40275 Unknown Unknown Unknown
libc++abi.dylib 95D402B5 Unknown Unknown Unknown
libc++abi.dylib 95D4031D Unknown Unknown Unknown
libc++abi.dylib 95D41412 Unknown Unknown Unknown
libOpenMMCuda.dyl 0E3D3B80 Unknown Unknown Unknown
libOpenMMCuda.dyl 0E3AF6C2 Unknown Unknown Unknown
libOpenMMCuda.dyl 0E3B3C3A Unknown Unknown Unknown
libOpenMMCuda.dyl 0E3B4323 Unknown Unknown Unknown
libOpenMMCuda.dyl 0E3A9F55 Unknown Unknown Unknown
libOpenMM.dylib 050AA3F0 Unknown Unknown Unknown
libOpenMM.dylib 050A8C3E Unknown Unknown Unknown
libOpenMM.dylib 05144033 Unknown Unknown Unknown
libOpenMM.dylib 05151589 Unknown Unknown Unknown
charmm 00DAC3A5 Unknown Unknown Unknown
charmm 00DA75AC Unknown Unknown Unknown
Abort

Any ideas would be helpful.

Thanks, Charlie

[Charles Brooks]

P.S. I'm linking against the OpenMM 4.1.1 precompiled libraries installed from the web site.

[Peter Eastman]

Hi Charlie,

CUDA is not binary compatible between releases, so you have to use the exact release that OpenMM was compiled against. OpenMM 4.1.1 was compiled against CUDA 4.1, so it won't work with CUDA 5.0. But if you recompile it from source, then it should work.

I'm hoping the situation will improve in OpenMM 5, since we're switching from the client API to the driver API, but I'm not certain of that yet.

Peter