[Eeglablist] Newtimef CUDA?

[Alejandro]

Hi all,

I have an implementation of time/frequency analysis that uses Matlab's 
cwt (continuous wavelet transform) that could be a good alternative to 
bypass the bottleneck that newtimef is. CWT works in parallel internally 
so you don't have to worry about that, nor to get a fancy video card. If 
anyone would like to help incorporating my code in EEGLAB's study 
machinery I'd be more than happy to help. As an example, computing ERSP 
and ITC for epochs 2 seconds long, 100 trials takes ~ 6 second in a 
decent desktop machine. The only critical input it needs beside a matrix 
with the data (channel or components: timeXtrials ) is a vector of 
scales (something equivalent to 'cycles' in newtimef), which can be set 
in logspace to produce a nice narrow-broad band decomposition (see the 
example in the website below). Right now it lacks from a significant 
test but that is something that can be easily added, for example using 
bootstrap will be something like this:

 >> bootstats = bootstrp( nboot, @myCTWfunction, data, scales);

where bootstrap is Matlab's bootstrap function that also works in parallel.


See "example_erp.mat" in the Downloads section on Mobilabs's google code 
website.
https://code.google.com/p/mobilab/

Regards,
Alejandro

On 10/15/2012 07:39 PM, Makoto Miyakoshi wrote:
> Dear Andrew,
>
> I also had an occasion to process over 100 datasets. Below is a
> practical solution (i.e., without rewriting the code) I found. If you
> want to try it, do it at your own risk.
>
> 1. If you have 100 subjects, then create STUDY first with proper
> STUDY.design. Do not run precompute yet.
> 2. Create 4 25 subjects STUDYs using the same data.
> 3. Run precompute for these 4 in parallel (assuming you have 4 core CPU).
> 4. Load the first 100 subject STUDY and cluster them.
>
> Makoto
>
> 2012/10/11 Andrew Hill <andrewhill at ucla.edu>:
>> Has anyone explored optimizing execution time of newtimef with CUDA or OpenGL?
>>
>> My ERSP runs take a week or so each (120 files, 30-min EEG, 64 channels @ 512 Hz) running on either a 64-bit Mac or Win.
>>
>> Of note is that this keeps CPU & RAM at only ~ 25% load the entire week, so there is some bottleneck that isn't resource limited.
>>
>> Any other suggestions for speeding up ERSP?
>>
>> Thanks,
>> Andrew Hill
>> _______________________________________________
>> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
>> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu
>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20121016/fc4147cb/attachment.html>