Re: Project proposal for GSoC.

[Daniel J Sebald]

On 02/25/2012 11:43 PM, Jordi Gutiérrez Hermoso wrote:
> On 25 February 2012 23:26, SAYANTAN DATTA<address@hidden>  wrote:
> > I am looking for opportunities to contribute in the field of parallel
> > processing and GPU computing.As many of the algorithms in mathematics are
> > inherently parallel ,it could save a huge time if we could use the power of
> > GPUs and multicore CPU, especially when the data sets are very large.For
> > example LU decomposition,FFT,matrix inverse ,numerical integration, ODE and
> > PDE solving etc could be computed much faster if they are performed using
> > GPUs and multicore CPUs.
> >
> >   Also I would like to mention that I'm quiet familiar with OpenCL,OpenMP
> > ,MPI and majority of the numerical methods used commonly in engineering.
> >
> >   So is there any scope for implementing the above mentioned idea for octave
> > in this year's GSoC?
>
> Perhaps, but there is a huge problem we keep facing: is there a free
> and GPL compatible implementation of OpenCL? It seems like a big shame
> that the only way to do GPU computing is by giving up control of your
> hardware to some binary blob made by nvidia or ATI. I know last year's
> GSoC had a project to implement OpenCL:
>
>      http://www.phoronix.com/scan.php?page=news_item&px=OTgwOQ
>
> There seems to be more recent progress on nouveau too:
>
>      http://www.phoronix.com/scan.php?page=news_item&px=MTA1Mzk
>
> I personally see getting a working free OpenCL implementation like a
> prerequisite before we can think of how to integrate this into Octave.
> My advice is to see if there is something you can do on that side
> first.

It would be interesting to see how GPUs could be included in Octave. 
Processing video streams would be the obvious choice, but GPUs and 
OpenCL are surely catching on with bigger software companies in fields 
like medical imaging.

I don't think there is anything dissuading writing code for a non-open 
piece of software, just that the Octave portion of things needs to be. 
OpenCL is a standard as well, if I understand correctly.  It would be 
like compiling Octave using a compiler other than gcc.

The issue with multicore, again if I understand correctly, is that 
really isn't a parallel processing technique.  It is more multithreaded. 
 That is, the OS decides how it can schedule the processes or how it 
can halt a process and restart it on another CPU.  To try and get Octave 
to figure out how best to use multicore processors would be a headache. 
 I've just moved to multicore, and it really is nice in some ways. 
It's like having several computers at hand to multitask.  Best to leave 
that to the OS: run multiple Octave processes if need be.  I've done that.

On a CPU, parallel processing takes the form of single instruction, 
multiple operation features of chips, but this is very non-portable. 
That's best for custom hardware/software designs, say a company branding 
a PC for its own application.

So, it seems to me OpenCL would be the direction to move in order to get 
some high power Octave parallel processing.  You'd have to think of the 
application first.  If you think it isn't difficult, some experimenting 
would be OK.

With the rate at which graphics cards are developing, a couple year old 
OpenCL GPU is rather affordable.  Interesting side note is that twenty 
years ago Steve Jobs put an accessible coprocessor on a motherboard and 
it didn't fly.  Today we have coprocessors on graphics cards and it 
seems quite popular.

Dan