Periodic Boundary Conditions Distances

High Performance package to calculate various distances with periodic boundary conditions.

Warning

This project is alpha software and not API stable. It serves as an example how to write reusable code with specialized compilation units.

Using pbc_distances

To calculate all pair wise distances without periodic boundaries and double precision use.

from pbc_distances import distance_sse3 as pbc
import numpy as np

pos = np.random.uniform(-1, 1, size=(5000, 3))

dist = pbc.pairwise_distance(pos, pos, box=None, precision='double')

If you have periodic boundary condition just supply the box argument. pbc_distances will try to figure out what box you supplied and use the appropriate algorithm.

The above example uses the SSE3 implementation. Please have a look into the setup.py which SIMD instruction sets are currently compiled.

Besides pairwise distances pbc_distances also supports bond, angle and dihedral.

Building

build only the extension

python setup.py build_ext -i

installation

python setup.py install

Benchmark Results

The repository contains some benchmark results in the results_***.md files. It compares the performance of this module with MDAnalysis and mdtraj. The possible performance gains using this module are up to a factor of 10. Below is the used CPU and compiler.

gcc = 7.2.1 CPU = Intel(R) Core(TM) i7-4600U CPU @ 2.10GHz

Guide for Developers

The library achieves it's performance by using SIMD instructions. To be able to run on many different platforms it uses inastemp_. This allows the library to be compiled against many different SIMD instructions and always choosing the best possible option.

The pbc algorithms are implemented as c++ functors in mic_*.h. Having these functors allows to decouple them from the actual distance calculations. This results in just having to write one distance algorithm and have c++ templates figure out the concrete types. This design let's us reuse code as much as possible and with enabled optimizations -O3 the compiler is also able to remove all boiler plate we created.

To expose the distance to cython it's best to only have pod data types (double, float) as a template parameter. Therefore you might need to create one distance function that takes an template parameter for the SIMD type to use called __distance_...(...) and expose it to cython as

  // vectorized version
  template<typename T>
  _distance_...(...) {
  using VecType = InaVecBestType<T>;
  __distance<VecType, T, mic_func>(...);
  }

  // non vectorized version
  template<typename T>
  _distance_..._scalar(...) {
  using VecType = InaVecSCALAR<T>;
  __distance<VecType, T, mic_func>(...);
  }


As one should note creating a non vectorized version is also simple with
inastemp as the example above shows.

Testing

Due to the code reuse design we only need to test the minimal image convention distances once for every box type. This means there is no need to test all box types for all the different distances (pairwise/bond/angle/dihedral) that pbc_distances provides.

The different distances can then be tested using no pbc conditions.

kain88-de / pbc_distances