OpenMM 6.0 binary install OS X 10.9

[Del Lucent]

Hello Everybody,

I have a number of brand new iMacs (nvidia GT750M) running OS X 10.9, the latest X-Code, CUDA 5.5, the Anaconda python dist, and have installed OpenM 6.01. For some reason only the CPU and Reference platforms are detected. Is there something obvious that I'm missing? Thanks for your help!

-Del

[Peter Eastman]

OpenCL doesn't work on Macs. Apple's implement is seriously broken.

If CUDA isn't getting loaded, it's likely to be a library dependency issue. Try typing

otool -L /usr/local/openmm/lib/plugins/libOpenMMCUDA.dylib

That should tell you if it has any unresolved dependencies. Make sure you've installed the CUDA toolkit, not just the driver.

Peter

[Del Lucent]

Ahh, that did it, problems with the wrong CUDA version! Thanks Peter.

[Del Lucent]

Hi Peter,

Unfortunately, I'm still running into problems using the CUDA platform. testInstallation.py seems to fail for the CUDA platform. A simple simulation of villin gives the following error when I try to use the CUDA platform:

Exception: Error launching CUDA compiler: 256
clang: error: unsupported option '-dumpspecs'
clang: error: no input files

I'm not sure how to go about debugging this. I have the CUDA 5.5 (and the sdk) installed and nvcc is in /usr/local/cuda/bin. Any thoughts? The full context of the error is pasted below:

Code: Select all

pdb = PDBFile('/Users/lucent/aaron.cabinaw/WorkshopTutorials/input.pdb') 
forcefield = ForceField('amber99sb.xml', 'tip3p.xml')
platform = Platform.getPlatformByName('CUDA')
system = forcefield.createSystem(pdb.topology, nonbondedMethod=PME, nonbondedCutoff=1.0*nanometer, 
constraints=HBonds) 
integrator = LangevinIntegrator(300*kelvin, 1/picosecond, 0.002*picoseconds)
simulation = Simulation(pdb.topology, system, integrator,platform)
simulation.context.setPositions(pdb.positions) 
---------------------------------------------------------------------------
Exception                                 Traceback (most recent call last)
<ipython-input-13-d2565d5e77a2> in <module>()
      1 # Creating simulation context
----> 2 simulation = Simulation(pdb.topology, system, integrator,platform)
      3 simulation.context.setPositions(pdb.positions)

/Users/lucent/anaconda/lib/python2.7/site-packages/simtk/openmm/app/simulation.pyc in __init__(self, topology, system, integrator, platform, platformProperties)
     75             self.context = mm.Context(system, integrator)
     76         elif platformProperties is None:
---> 77             self.context = mm.Context(system, integrator, platform)
     78         else:
     79             self.context = mm.Context(system, integrator, platform, platformProperties)

/Users/lucent/anaconda/lib/python2.7/site-packages/simtk/openmm/openmm.pyc in __init__(self, *args)
   4861 
   4862 
-> 4863         this = _openmm.new_Context(*args)
   4864         try: self.this.append(this)
   4865         except: self.this = this

Exception: Error launching CUDA compiler: 256
clang: error: unsupported option '-dumpspecs'
clang: error: no input files

[Peter Eastman]

I haven't seen this problem, but it looks like other people have:

http://stackoverflow.com/questions/1964 ... ted-option
http://stackoverflow.com/questions/1935 ... -mavericks

In both of those threads, the answer is to upgrade to a newer version of the CUDA toolkit. What version do you have? For me,

/usr/local/cuda/bin/nvcc --version

gives

Code: Select all

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2013 NVIDIA Corporation
Built on Thu_Sep__5_10:17:14_PDT_2013
Cuda compilation tools, release 5.5, V5.5.0

Peter

[Del Lucent]

I have 5.5:

Code: Select all

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2013 NVIDIA Corporation
Built on Wed_Jul_10_11:16:01_PDT_2013
Cuda compilation tools, release 5.5, V5.5.0

If I upgrade to 6.0, won't that break binary compatibility? I suppose I could just try compiling from source against CUDA 6.0 but I thought this should work with 5.5? Thanks for your help!

[Peter Eastman]

It looks like NVIDIA is being sloppy with version numbers. My version of the toolkit is newer than yours (built in Sept. rather than July), but they both claim to be "release 5.5, V5.5.0". Sigh.

Looking at their page for past CUDA versions, they list 5.5 as being from July 2013. Presumably that's what you downloaded? Yet at some point, there clearly were newer builds than that still listed as 5.5.

Your best bet might be to install CUDA 6.0 and compile from source.

Peter

[Del Lucent]

Compiling from source with cuda 6 seemed to work. As always, thanks for the help Peter!

[Preeti Sahu]

Hi Peter,
We have a Mac OS X 10.8.5 with AMD Radeon. So OpenMM can't be installed in this is it?

[Jason Swails]

OpenMM works fine on macs. It's just the OpenCL platform that doesn't work.