simtk.org

Hello everyone,

I have been having some trouble with certain simulations in OpenMM (utilizing CustomGBForces) using OpenCL on my MacBook Pro - it is a convenient testing platform. Here are the OpenCL parameters printed out from the context:

OpenCLDeviceIndex: 1
OpenCLDeviceName: HD Graphics 4000
OpenCLPlatformIndex: 0
OpenCLPlatformName: Apple
OpenCLPrecision: single
OpenCLUseCpuPme: false


Has anyone seen any troubles with these boards or any problems within how OpenMM compiles this "script" code into OpenCL in the execution of the simulation programs? If a developer could point me to some of the relevant API routines or source files I could look at to further figure out what is going on in the underlying execution of code that uses the scripting in CustomGBForce (as well as any additional OpenCL debugging information), I'd greatly appreciate it.

The error is that the system sort of disappears (coordinates NaN) and reports energy NaN very early into the simulation, perhaps on the 100th step or so; it minimizes fine but it seems there is a bug somewhere, and these jobs are using some of the routines seen in customgbforces.py w/ some minor changes for experimenting with the model. It seems it does not depend on what is calculated here.

I run these jobs on the Linux machine with CUDA platform and everything proceeds fine and I obtain up to 500ns trajectories, so this may be either an OpenCL error or a driver error with my Intel HD 4000 graphics card which has the OpenCL 1.2 support and installed drivers.

Thanks,
Luke

Hi Luke,

At present, OpenCL is completely broken on OS X. This is caused by bugs in Apple's implementation. They're aware of the bugs. They have test cases to reproduce them. Perhaps some day they will even fix them. Unfortunately, they give us no information at all.

The next update to OpenMM (due out any day now) will simply disable the whole OpenCL platform when running on OS X. Sorry this caused problems for you.

Peter

peastman wrote:Hi Luke,

At present, OpenCL is completely broken on OS X. This is caused by bugs in Apple's implementation. They're aware of the bugs. They have test cases to reproduce them. Perhaps some day they will even fix them. Unfortunately, they give us no information at all.

The next update to OpenMM (due out any day now) will simply disable the whole OpenCL platform when running on OS X. Sorry this caused problems for you.

Peter

Thanks, I was wondering about that and the behavior was always inconsistent. I'll look into a workaround but I suppose you are kind of tied to Apple's OpenCL implementation and can't override it. Perhaps one of these Boot Camp softwares to dual boot Linux would be a good solution for developing on the machine when I'm away from a desktop.

I appreciate your help!