[ITK-users] GPU Implementation

[Jim Miller]

A simple solution is to downsample the images. Instead of 512x512x350 images, downsample to 256x256x175 images. 

With the regularization in the registration algorithm, the results will be very similar on the two datasets. 

If you wanted, after running the registration on the downsampled versions, you can upsample the deformation field and run a few iterations of the algorithm on the cpu at full resolution. 

Jim

> On Dec 30, 2014, at 7:57 AM, Xiaopeng Yang <yxp233 at postech.ac.kr> wrote:
> 
> Hi Jim,
>  
> How are you doing?
>  
> I have a new question about dataset size. I did an experiment to see how large of a dataset I can handle and the result was 512 × 512 × 140 (140 CT slices). Then I upgraded my graphic card which has 12 GB memory (previous card has 4 GB memory) and did the same experiment. However, only up to 218 slices, the new graphic card worked properly.  In my case, sometimes I need to process more than 350 slices, what would be a solution for processing this large dataset? I know there are 16 GB and 24 GB graphic cards, but I am not sure whether they can handle up to 350 slices or not and they are very expensive as you know. Maybe there should be any solution by coding? Thank you.
>  
> Best regards,
> Xiaopeng
>  
> From: Jim Miller [mailto:millerjv at gmail.com] 
> Sent: Thursday, December 11, 2014 12:49 AM
> To: Xiaopeng Yang
> Cc: John Drescher; Insight-Users at Itk. Org
> Subject: Re: [ITK-users] GPU Implementation
>  
> Great. You can experiment further to see how large of a dataset you can handle. Then decide if you need to upgrade your graphic card to one with more memory to support the dataset size you want to process. 
> 
> Jim
> 
> On Dec 10, 2014, at 9:04 AM, Xiaopeng Yang <yxp233 at postech.ac.kr> wrote:
> 
> Hi Jim,
>  
> Thanks for your suggestion. I tested with a smaller dataset: 512 × 512 × 75 and it worked. No crash occurred.
>  
> Best regards,
> Xiaopeng
>  
> From: Jim Miller [mailto:millerjv at gmail.com] 
> Sent: Saturday, December 06, 2014 12:50 PM
> To: Xiaopeng Yang
> Cc: John Drescher; Insight-Users at Itk. Org
> Subject: Re: [ITK-users] GPU Implementation
>  
> I'd start by using a smaller dataset. Perhaps downsample your dataset to 256x256x150 to see if that works on your card. 
> 
> Jim
> 
> On Dec 5, 2014, at 8:39 PM, Xiaopeng Yang <yxp233 at postech.ac.kr> wrote:
> 
> Hi Jim,
>  
> So you mean I need to change graphics card or set some parameters? How to allocate the memory of graphics card?
>  
> Thanks,
> Xiaopeng
>  
> From: Jim Miller [mailto:millerjv at gmail.com] 
> Sent: Saturday, December 06, 2014 10:11 AM
> To: Xiaopeng Yang
> Cc: John Drescher; Insight-Users at Itk. Org
> Subject: Re: [ITK-users] GPU Implementation
>  
> I estimate that you would need over 2GB on your graphics card to run the demons algorithm on that size of a dataset. 
> 
> Jim
> 
> On Dec 5, 2014, at 6:45 PM, Xiaopeng Yang <yxp233 at postech.ac.kr> wrote:
> 
> Dear Jim,
>  
> The images are quite large: 512 * 512 *300.
>  
> I did not build test when I compile ITK.
>  
> Thanks,
> Xiaopeng
>  
> From: Jim Miller [mailto:millerjv at gmail.com] 
> Sent: Friday, December 05, 2014 11:15 PM
> To: Xiaopeng Yang
> Cc: John Drescher; Insight-Users at Itk. Org
> Subject: Re: [ITK-users] GPU Implementation
>  
> How large are the images that you are using?
> 
> Jim
> 
> On Dec 5, 2014, at 2:43 AM, Xiaopeng Yang <yxp233 at postech.ac.kr> wrote:
> 
> Hi John,
>  
> Sorry I mean the program got crashed after showing #define T float.
>  
> Thanks,
> Xiaopeng
>  
> From: John Drescher [mailto:drescherjm at gmail.com] 
> Sent: Friday, December 05, 2014 4:24 PM
> To: Xiaopeng Yang
> Cc: Jim Miller; Insight-Users at Itk. Org
> Subject: Re: [ITK-users] GPU Implementation
>  
>  
>  
> On Fri, Dec 5, 2014 at 12:41 AM, Xiaopeng Yang <yxp233 at postech.ac.kr> wrote:
> Dear Jim,
>  
> Thanks for your help. I built the test successfully, but I met an error when I run the program. Please see the attached image file.
>  
> Best,
> Xiaopeng
>  
> <image001.png>
>  
>  
> Is that an error?
> 
> John
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/insight-users/attachments/20141230/e88d7476/attachment.html>