paranumal / hipBone

hipBone

This is the hipBone repository. hipBone is a GPU port of the original proxy application called Nekbone.

It solves a screened Poisson equation in a box using a conjugate gradient method.

How to compile hipBone

There are a couple of prerequisites for building hipBone;

• An MPI stack. Any will work;
• OpenBlas.

Installing MPI and OpenBlas can be done using whatever package manager your operating system provides.

To build and run hipBone, there is an included run.sh script which will build the third party OCCA, then build hipBone, and run several problem sizes and output figures of merit.

To build hipBone manually:

$ git clone --recursive <hipBone repo>
$ cd /path/to/hipBone
$ export OPENBLAS_DIR=/path/to/openblas
$ make -j `nproc`

How to run hipBone

Here is an example CORAL-2 problem size that you can run on one GPU:

$ mpirun -np 1 ./hipBone -m HIP -nx 24 -ny 24 -nz 24 -p 14

Here is the meaning of each of the command line options

• nx: the number of spectral elements in the x-direction per MPI rank
• ny: the number of spectral elements in the y-direction per MPI rank
• nz: the number of spectral elements in the z-direction per MPI rank
• p: the order of the polynomial used to approximate the solution
• m: the mode to run OCCA in, HIP is for AMD GPUs but CUDA and Serial are also supported

Running on multiple GPUs can by done by passing a larger argument to np and specifying the number of MPI ranks in each coordinate direction:

$ mpirun -np 2 ./hipBone -m HIP -nx 24 -ny 24 -nz 24 -px 2 -py 1 -pz 1 -p 14

You must specify either:

1. All of px, py, pz, or
2. None of px, py, or pz.

If all of px, py and pz are specified then the product px*py*pz must equal the argument passed to np. If none of px, py or pz are specified then the np must be a cube and hipBone will use an equal number of MPI ranks in each coordinate direction.

Verifying correctness

To verify that the computation is correct, add the -v option to the command line. Example output towards the end of the run may look like this:

CG: it 96, r norm 1.328996666475e-19, alpha = 5.291357e-01
CG: it 97, r norm 2.552900554560e-19, alpha = 1.990951e+00
CG: it 98, r norm 3.836827649728e-19, alpha = 3.269689e+00
CG: it 99, r norm 2.629545869383e-19, alpha = 1.509263e+00
CG: it 100, r norm 2.045530932453e-19, alpha = 8.445030e-01
hipBone: 3, 2744, 0.0249, 100, 9.08e-06,  3.7,  2.3, 1.10e+07; N, DOFs, elapsed, iterations, time per DOF, avg BW (GB/s), avg GFLOPs, DOFs*iterations/ranks*time
hipBone: NekBone FOM =  2.6 GFLOPs.

The printed value of r norm at the end of 100 CG iterations should be small.

As per the Nekbone CORAL-2 Benchmark summary:

Benchmark results are considered correct if the reported r norm is small, generally less than 1e-8, after 100 conjugate gradient iterations.

How to clean build objects

To clean the hipBone build objects:

$ cd /path/to/hipBone/repo
$ make realclean

Please invoke make help for more supported options.

References

HipBone: A performance-portable GPU-accelerated C++ version of the NekBone benchmark: arXiv version: Chalmers N., Mishra A., McDougall D., Warburton T., 2022. HipBone: A performance-portable GPU-accelerated C++ version of the NekBone benchmark.

To cite this repo directly:

@MISC{ChalmersMishraMcDougallWarburtonHipBone2022,
      author = "Chalmers, N. and Mishra, A. and McDougall, D. and Warburton, T.",
      title = "{HipBone}: a performance-portable GPU-accelerated C++ version of the NekBone benchmark",
      year = "2022",
      url = "https://github.com/paranumal/hipBone",
      doi = "10.5281/zenodo.6362839",
      note = "Release 1.1.0"
      }