[Insight-developers] CUDA ITK

[Philip Ward]

Dear Developers,

My apologies if my protocol is incorrect in using this list, I am a first time user.

A development team, including myself, from VPAC in Melbourne Australia (www.vpac.org) are currently working on porting some of the ITK to CUDA. One of the reasons for doing this is VPAC manage many clusters, including a Tesla cluster, and some of our clients use ITK. As part of our preliminary research I would like to know if any developers are currently porting to CUDA so that we can avoid any doubling up and wasted effort. We are open to collaboration and will be releasing our code sometime in late February 2010.

Our current tactic is to write a memory management class to wrap around the itkImportImageContainer. This will be used to further increase performance when multiple GPU filters are used in sequence. It will also allow us to give full compatibility with existing CPU filters. The drawback of using GPU's is the overhead of copying memory to and from the device, so gains on particular filters may be lost due to memory transfers to accommodate existing CPU filters. This will be investigated further during our testing. 

I am happy and open to discuss any matters relating to this so do not hesitate to email me. I will be putting up a detailed plan of our initial architecture online soon if anyone would like to view it.

Regards,
Phillip Ward