New ATI Stream OpenCL Platforms

FrgMstr

Just Plain Mean
Staff member
2FA
Joined
May 18, 1997
Messages
58,704
Today, AMD has announced that new ATI Stream development platforms are available from Colfax and Exxact, providing developers with the tools to get started working with OpenCL immediately. These platforms are further proof of the industry’s adoption of open standards for application acceleration, as well as the demand for heterogeneous computing environments. I’ve included the full press release below, which has all of the pertinent information. Feel free to also check out AMD’s latest HPC video below, which discusses ATI Stream. Particularly timely with SC09 this week.
 
good to see they are getting the developer software out there for anyone to use..
 
So the ATi guys left an 5870 in TACC for them to test. Still CUDA is very easy to translate from regular C. Is that the same for OpenCL?
 
OpenCL kernels are written in what is very close to C99. There are a few additions, notably address space qualifiers and support for short vector syntax, and some restrictions around recursion and function pointers. While the host side API might look a little scary at first (kind of like DX vs GL for hello world), that code is largely boilerplate. You can also use the C++ bindings available from Khronos (search for cl.hpp) to help encapsulate much of the setup code.

From a high level, OpenCL, CUDA, DirectCompute are all data parallel programming languages with similar methodology and even syntax on expressing data locality and execution decomposition. Once you've worked with one, the others are pretty straightforward.

If you want to see an example of a pretty large code ported to OpenCL, Los Alamos National Lab has an open source gas dynamics simulation up on sourceforge that can run across all shipping OpenCL implementations (AMD, Apple, Nvidia, IBM) that was being demoed at Supercomputing. See http://sourceforge.net/projects/hypgad/ There are also tutorial slides up on Khronos that walk through simple examples and all of the shipping implementations also provide many samples (nbody, convolution, options pricing, etc), although they are largely for reference and not necessarily for performance.
 
Back
Top