[AstroPy] Image combine

Derek Homeier derek@astro.physik.uni-goettingen...
Wed Mar 7 09:40:32 CST 2012


On 7 Mar 2012, at 16:17, Perry Greenfield wrote:

> Do you have any good example of code like this is successfully  
> distributed to users in the wild? So far, the dependencies on the  
> specific GPU hardware seem to make this impractical right now. But  
> maybe I'm missing something.
> 
> And I'm also a bit confused by the first statement. Are you arguing  
> that we should be porting SPP to GPUs?

I'd agree that optimisation is an important issue; enabling GPU is one part of that, 
classical parallelisation (multi-core/shared memory first) another part. 
But these are rather two sides of the medal; if the basic tasks are not available 
people will not be willing or even able to switch from IRAF or other suites. 
The best perspectives for speed-ups OTOH lie most likely in the underlying 
routines, i.e. for the most part the numpy core routines or even below (like the 
general-purpose numeric libraries numpy or scipy are calling), and that is 
where any optimisation effort is best spent - just my 2p. 

Cheers,
						Derek

> On Mar 7, 2012, at 10:11 AM, Tiago Ribeiro de Souza wrote:
> 
>> Hi guys,
>> 
>> Even though I understand that people want to be free from using  
>> IRAF, I don't agree that only  translating its basic functionality  
>> to other programming language is actually useful. If we ought to  
>> look to the future I would say that we should start an effort to  
>> provide a GPU-enabled suite of image reduction facility ratter than  
>> replicating existing code in different languages. Given the level of  
>> paralalization of image reduction (say image combine, subtraction,  
>> division, source extraction, spectral extraction and more!) GPU  
>> programming will definitively be much more important and useful.
>> 
>> For those who are interested, check out this tutorial at http://www.macresearch.org/opencl 
>> and http://developer.nvidia.com/content/september-2009-opencl-public-downloads
> 


More information about the AstroPy mailing list