CFD Online Discussion Forums

CFD Online Discussion Forums (
-   OpenFOAM (
-   -   OpenFOAM and OpenCL (

Arrow October 21, 2009 14:22

OpenFOAM and OpenCL
Hi all...

I thought we needed a new thread to discuss the possibility of accelerating OpenFOAM on heterogenious systems with OpenCL.

I think OpenCL is the way to go as it is more general and open than a port to CUDA( which is proprietary to nVidia).

OpenCL has an active working group with respect to C++ and tools and libraries similar to BLAST are coming along.

Best to start discussing the issues.

andrewryan October 21, 2009 15:31

Did u read this thread?

Arrow October 23, 2009 20:18

My bad I guess
Stupid newbee tricks...

Just started the thread because the other one was started in 2006.

Porting OpenFOAM to a GPGPU environment is a pressing technological imperative.

Both AMD and nVidia have gpu's out or coming real soon now that promise to boost HPC 100x.

Just loudly making the point that the spirit of OpenFOAM as OSS should demand a OSS heterogenious computing solution like OpenCL.

The good people at Apple and AMD have a C++ binding for OpenCL available.

CL::xx name space

Tools are available now and will get better. ( see MAGMA project).

Just saying

olesen October 26, 2009 04:25


Originally Posted by Arrow (Post 233574)
Best to start discussing the issues.

A few issues come to mind: How much time is required to move the matrices from main memory to GPU and back? Are the GPUs capable of double precision?

Since you are the one promoting the idea, it would be nice if you provided a bit more background about your own feasibility assessment and perhaps about what coding you already attempted in this direction.

andrewryan October 26, 2009 05:59

> Are the GPUs capable of double precision?

As far as I read the Nvidia GPUs have very limited double support atm. However the next gen Nvidia Fermi should fully support double precision, see Nvidias whitepaper [1]


eugene October 26, 2009 08:17

My 2 cents.

Intel will be rolling out a x86 compliant Larabee GPU architecture in the next year. These processor cores will eventually be integrated on-chip with the cpu or as drop in replacements in multi-socket systems with CSI interconnect removing all memory transfer latencies. This combination in my opinion strongly disincentives investment in an OpenCL specific branch since it is highly likely this approach will be superseded within the next 2-3 years by the x86 compliant "co-processor" model.

Of course this is just conjecture made up from a patchwork of information, the reality of Intel's architecture may be quite different.

elvis October 26, 2009 11:25

You mean, people should wait to hear an anouncement on the Icon Open Source CFD International Conference ;-)

All times are GMT -4. The time now is 13:15.