[gmx-developers] Branches w/working OpenCL support

[Mirco Wahab]

On 25.06.2015 17:02, Mark Abraham wrote:
> I got local access to a machine with two AMD GPUs, and ran into the same
> initialization issues I'd seen with Nvidia runtimes, so I have updated
> https://gerrit.gromacs.org/4314 so that in 5.1 GROMACS will never use
> more than one OpenCL device per node. We can probably fix that in the
> longer term, but it seems like issues of host-side concurrency are not
> treated in the OpenCL standard (or at least the implementations we've
> targeted), and we didn't do a good enough job of anticipating and
> managing that.

In this moment I completed compilation and run some first tests on
my AMD box (AMD Phenom II X6 1090t + R9-270X Pitcairn GCN 1.1).

Compilation/Installation:
  I fetched refs/changes/14/4314/38 against the current git master,
  compilation (VS 2013/MSVC 18.00.31101, AMD App-SDK 3.0.0-beta)
  went smooth, not a single error (built against a local fftw 3.3.3),
  so that looks really promising ...

First run tests (for posting the status here):
  - single water box (2916 SPC)
  - NVT + RF  ok.
  - NVT + PME ok.
  - NPT + RF  ok.
  - NPT + PME ok.

All runs did complete without any problem!
Run times are promising too on this
very small system (ns/day):

cpu only - rf/npt: 32.482, pme/npt: 20.971
cpu/gpu  - rf/npt: 63.530, pme/npt: 51.429
            rf/nvt: 64.001, pme/nvt: 51.125

The gpu kernels do get compiled in this version
at the start of every run and are located in the
%TEMP% directory - and deleted after loading, So they
are not longer cached/stored. The time for gpu kernel
compilation is >30 sec on this box (maybe this rewards
an extra message). The structure of this file is
(looks like) an ELF-binary?

As it looks, the OpenCL extension could be a
very useable attribute for Gromacs users. I'm
impressed.


Thanks & Regards

M.