ANN_Force biasing force plugin

This is the ANN_Force biasing force plugin for OpenMM (https://github.com/pandegroup/openmm) and accelerated sampling with data-augmented autoencoders framework (https://github.com/weiHelloWorld/accelerated_sampling_with_autoencoder).

Dependency

OpenMM simulation pacakge: https://github.com/pandegroup/openmm

SWIG C/C++ wrapper: http://www.swig.org/

CMake: https://cmake.org/

Installation

mkdir build
cd build
ccmake ..

Modify CMake installation settings as needed. Turn on CUDA option if you would like to run simulation on CUDA. Then run

make install
make PythonInstall

Root permission may be needed.

Quick start

This package should be used together with OpenMM simulation package and training results of neural networks. A good starting point would be tests in this framework: https://github.com/weiHelloWorld/accelerated_sampling_with_autoencoder.

TODO: examples

Testing

mkdir build
cd build
ccmake ..
make install
echo "running tests for numerical calculations..."
make test

cd ../openmmapi/tests
echo "running test for Python wrapper..."
python test_Python_wrapper.py

Why don't we use OpenMM + Plumed with custom expressions for umbrella sampling?

Because it is slow to use Plumed with OpenMM on CUDA. The way it works is that at every time step OpenMM copies current state (positions/velocities/etc) to plumed and then receives calculated results from plumed and do post-computation (https://github.com/peastman/openmm-plumed/blob/master/platforms/cuda/src/CudaPlumedKernels.cpp#L198). Since plumed does not run on CUDA, there would be significant communication overhead.

Implementation details for CUDA platform

Here are some implementation details for CUDA platform if you are interested:

• Parallelization is realized by duplicating a single ANN_Force to get multiple copies for many threads. By doing this, it can be well integrated into implementation of CudaBondedUtilities class (a force can only be assigned to one thread at most by default, that is why we need to get multiple copies for the force).

• We use two-step code generation to improve performance:

  • use C++ to dynamically generate cuda framework code. The purpose of this additional step is to
    • create code with minimal computation (remove unnecessary conditionals and loops), where C++ works for pre-preprocessing optimization
    • reduce memory access by replacing variable in shared memory with number literals
  • do preprocessing to get runnable cuda code.

• Inter-block synchronization is not done (I use all threads in one block to avoid this issue), but based on current benchmark, I do not expect it would accelerate computation too much.

• There is still room for optimization, in additional to synchronization issue, how to store variables in proper memory space and how to achieve workload balance among different threads would possibly make a difference.

TODO

• serialization?

• better synchronization for CUDA implementation?

Contact

For any questions, feel free to contact weichen9@illinois.edu or open a github issue.