Skip to content

Plugin: cudaCoord

Description Usage
version of coordination number that can run on the GPU with parallelism written in CUDA
Authors: Daniele Rapetti used in 2 tutorialsused in 0 eggs

Details

This is the optimized version of the lesson that I presented in the plumed-school with a step-by-step approach.

CUDACOORDINATION and CUDACOORDINATIONFLOAT depend on CCCL which is automatically fetched by the cuda compiler (if you use nvcc, you have access to the CCCL headers).

The files cudaHelpers.cuh and cudaHelpers.cu contains a few support functions for helping in interfacing PLMD::Vector and PLMD::Tensor with Cuda's thrust, along with the reduction functions baked with Cuda's cub building blocks and their drivers.

[!WARNING] Plumed may refuse to to the calculations if the GPU does not allow for the calculations to be done. We are currently working on a solution for this

A workaround is reducing the size of the cutoff of the neigbor list or to increase the number of threads with THREADS Plumed will present the user wiht an error message with the suggestion.

Compile

With a ready to run plumed in your path: - ./configure - make

you may need to specify the SM of your GPU by modifying the Makefile, for example with:

NVCCCFLAGS = -dc -dlto --gpu-architecture=sm_75 
NVCCLDFLAGS = -shared -dlto --gpu-architecture=sm_75
If you are using a T1000

How to tests

Assuming you are using this plugin from the plugin directory of a plumed source directory: - cd regtest; ln -s ../../../regtest/scripts . - make check

Limitations

CUDACOORDINATION and CUDACOORDINATIONFLOAT work more or less as the standard COORDINATION, except from:

  • work only with orthogonal pbcs or no pbcs at all
  • do not support the SWITCH keyword
  • new THREADS keyword to control the maximum number of threads that can be used by the kernels

Current TODO

  • The GPU device needs to be explicitly selected
  • Integrate the CI