On newer versions of openmmtools, GPU can't be used if using mixed precision due to mixed precision check

https://github.com/choderalab/openmmtools/blob/3f2cfc0d9ea5ee69b270b1f7705867548f8ca037/openmmtools/utils.py#L513

This function always fails due to this part:

    if platform.getName() in ['CUDA', 'OpenCL']:
        from simtk import openmm
        properties = { 'Precision' : precision }
        system = openmm.System()
        integrator = openmm.VerletIntegrator(0.001)
        try:
            context = openmm.Context(system, integrator, properties)
            del context, integrator
            return True
        except Exception as e:
            return False

It always returns

TypeError: Wrong number or type of arguments for overloaded function 'new_Context'.
  Possible C/C++ prototypes are:
    OpenMM::Context::Context(OpenMM::System const &,OpenMM::Integrator &)
    OpenMM::Context::Context(OpenMM::System const &,OpenMM::Integrator &,OpenMM::Platform &)
    OpenMM::Context::Context(OpenMM::System const &,OpenMM::Integrator &,OpenMM::Platform &,std::map< std::string,std::string,std::less< std::string >,std::allocator< std::pair< std::string const,std::string > > > const &)
    OpenMM::Context::Context(OpenMM::Context const &)

Try it:

precision = "mixed"
from simtk import openmm
properties = { 'Precision' : precision }
system = openmm.System()
integrator = openmm.VerletIntegrator(0.001)
context = openmm.Context(system, integrator, properties)

Thus, whenever configuring the platform, if trying to check available platforms that allow mixed precision, the CUDA and OpenCL ones will always fail.

[jchodera]

Thanks for the fantastic bug report and minimal example to reproduce!

The syntax for line 551 should be changed from

context = openmm.Context(system, integrator, properties)

to

context = openmm.Context(system, integrator, platform, properties)

I’ll open a PR.

[aced125]

Thanks @jchodera and @peastman for the very quick fix indeed!