[AstroPy] Image combine
Wed Mar 7 10:08:44 CST 2012
On Wed, Mar 7, 2012 at 9:02 AM, Tiago Ribeiro de Souza <email@example.com> wrote:
> Certainly, there are a lot of things to consider. There is a rater interesting paper about this on astro-ph (http://arxiv.org/abs/1008.4623), if you are interested.
Ted Kisner's talk at the Future of Astrocomputing conference from Dec
2010 addresses OpenCL in astronomy as well. In the video file he
discusses many of the shortcomings and challenges more than comes
across in the slides.
http://hipacc.ucsc.edu/FOA_Program.html (with links to both PDF and MP4 files)
> Em 07/03/2012, às 12:52, Derek Homeier escreveu:
>> On 7 Mar 2012, at 16:35, Tiago Ribeiro de Souza wrote:
>>> As far as I concern openCL is an attempt to provide a hardware-free sdk for GPU programming. I agree that it is still far from being user-friendly, though. I am also not aware of any astronomical code widely distributed but, the tutorial in macresearch shows an application used in chemistry (I guess) that uses openCL for scientific computation. At the same time, it seems to me that, once you write parallelized code in openCL it works both with CPU and GPUs in the same way.
>> That's highly unlikely, since even within openCL I expect you have to consider the
>> type of hardware available to you, e.g. standard graphics cards will only provide
>> significant speed-up for single-precision data, while the newest Tesla models
>> (or successors) have double-precision vector registers. You might be able to code
>> routines for both cases and multi-threading over several CPU cores, but again, this
>> is probably best done at a lower level, as I'd expect a lot of these tasks are quite
>> generic for image data or any kind of large arrays in 2 or more D.
>> For rather simple tasks like addition/subtraction/division one would probably also
>> have to carefully weigh the gain in CPU/GPU power against the overhead for moving
>> these large amounts of data around.
>> AstroPy mailing list
> AstroPy mailing list
More information about the AstroPy