cuda_mpi_autotools_proj_template

A template project for CUDA+MPI with autotools build system

Build for the first time

Run autogen.sh script.

./autogen.sh

It will create all autotools related files.

Run configure

./configure --with-cuda=$CUDA_ROOT --enable-mpi

where CUDA_ROOT is an environment variable set to the location of the CUDA toolkit (e.g. /usr/local/cuda-6.5)

Run make

This will build an executable src/testHelloMpiCuda.

Start application

You can run it on a multi-GPU cluster with

mpirun -np N_PROC_MPI ./testHelloMpiCuda

Reference

For a detailled discussion about coupling CUDA and MPI (initialize CUDA context before/after MPI_Init), see the following references:

• https://github.com/parallel-forall/code-samples/tree/master/posts/cuda-aware-mpi-example
• http://www.open-mpi.org/faq/?category=running#mpi-cuda-support