[Rtk-users] New itk::CudaImage

[Simon Rit]

Dear users,
Yesterday, I have merged the branch ITK-Cuda in the master branch. This
development introduces a new image format, itk::CudaImage. This format has
been developed by Kitware for us (and, eventually, ITK). It is based on
itk::GPUImage which is for OpenCL. The goal of this class is to avoid
unnecessary memory transfers between the CPU and the GPU memories.
They have also developed a mechanism to generate kernels for any pixel type
but this is not yet used in RTK. If you want to use it, you can look at this
commit<https://github.com/SimonRit/RTK/commit/bd31fddcd4672645624a3da8e777a2393104f409>but
be aware that the mechanism requires the use of the same compiler for
both the normal and the CUDA code.
I am very pleased with the result and I'll let you discover the details of
this mechanism in the code. The one and only drawback (to my knowledge) is
that we lose the compability with CUDA 3.2 so you must upgrade to a newer
CUDA version if you still use this one. For the rest, we have managed to
preserve the ITK 3.20 compatibility.
These observations are based on our regression
tests<http://my.cdash.org/index.php?project=RTK>and we are looking
forward to your experience. Let us know via this mailing
list if you encounter any other problem.
Simon
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/rtk-users/attachments/20130906/cbe4df37/attachment-0008.html>