C[omp]ute

Welcome to my blog, which was once a mailing list of the same name and is still generated by mail. Please reply via the "comment" links.

Always interested in offers/projects/new ideas. Eclectic experience in fields like: numerical computing; Python web; Java enterprise; functional languages; GPGPU; SQL databases; etc. Based in Santiago, Chile; telecommute worldwide. CV; email.

Personal Projects

Choochoo Training Diary

Last 100 entries

Image Processing with CUDA / Python (Dynamic Pipelines)

From: "andrew cooke" <andrew@...>

Date: Sun, 3 Aug 2008 13:20:21 -0400 (CLT)

I've been thinking about how best to apply CUDA to image processing
(particularly in astronomy, which is what I know).

Many image processing tasks are *very* suitable for parallelisation, since
the processing is "per-pixel" across several arrays (for example,
subtracting oen image from another).

So my idea is to write a parser (in Python) that accepts an expression in
a "language" that describes image processing and compiles that (literally)
to a CUDA process.

The motivation for this approach is to reduce the memory transfer overhead
compared to the alternative: a library of primitive operations.  If add,
multiple, etc are each separate primitives then data must be loaded for
each.  But in many cases it should be possible to combine these so that
all data are loaded just once, and the "per pixel" operation does several
steps.  Thie lowers the data transfer cost while at the same time grouping
the arithmetic processing into a single "chunk" (useful because it can run
while other data are loading).

It may not be clear that there is sufficient complexity, but my idea is to
support "error" and "uality" data too.

Andrew

Re: CUDA and astronomical image processing

From: andrew cooke <andrew@...>

Date: Sun, 8 Jul 2012 08:35:38 -0400

Hi,

I need to update my blog - it doesn't handle multipart mime messages like 
yours (which is why it wasn't displayed).

Anyway, no - I never got any further.  One reason was that I thought Theano
http://deeplearning.net/software/theano/ did a lot of what I was planning.  So
if you're looking for something like I described, you might want to consider
that.

Cheers,
Andrew

On Sun, Jul 08, 2012 at 11:47:42AM +0930, Andrew Cool wrote:
> Hi Andrew,
> 
> It's now 2012. Did you get anywhere with your CUDA processing?
> 
> Regards,
> 
> Andrew Cool
> 
> www.skippysky.com.au

CUDA and astronomical image processing

From: "Andrew Cool" <andrew@...>

Date: Sun, 8 Jul 2012 11:47:42 +0930

Hi Andrew,

It's now 2012. Did you get anywhere with your CUDA processing?

Regards,

Andrew Cool

www.skippysky.com.au

 

From: "andrew cooke" <andrew@...> 

Date: Sun, 3 Aug 2008 13:20:21 -0400 (CLT) 

I've been thinking about how best to apply CUDA to image processing

(particularly in astronomy, which is what I know).

 

Many image processing tasks are *very* suitable for parallelisation, since

the processing is "per-pixel" across several arrays (for example,

subtracting oen image from another).

 

So my idea is to write a parser (in Python) that accepts an expression in

a "language" that describes image processing and compiles that (literally)

to a CUDA process.

 

The motivation for this approach is to reduce the memory transfer overhead

compared to the alternative: a library of primitive operations.  If add,

multiple, etc are each separate primitives then data must be loaded for

each.  But in many cases it should be possible to combine these so that

all data are loaded just once, and the "per pixel" operation does several

steps.  Thie lowers the data transfer cost while at the same time grouping

the arithmetic processing into a single "chunk" (useful because it can run

while other data are loading).

 

It may not be clear that there is sufficient complexity, but my idea is to

support "error" and "uality" data too.

 

Andrew

 

 

All of us could take a lesson from the weather........ It pays no attention
to Criticism.

Comment on this post