[ITK-users] ITK Image data pointer and GPU algorithms

[Dženan Zukić]

Hi Gordian,


   1. Using GetBufferPointer is the right way to access raw pixel data for
   your case.
   2. Not using ITK is more error prone way of manipulating image data.
   3. I remember that with more recent versions of CUDA you have unified
   memory architecture, hence you can using the pointer given
   by GetBufferPointer to access pixel data. I don't remember whether this
   implies slower access (you have to account for latency of PCIe bus) or it
   implicitly copies the data to GPU.


Regards,
Dženan

On Thu, Mar 2, 2017 at 11:24 AM, Kabelitz, Gordian <
Gordian.Kabelitz at medma.uni-heidelberg.de> wrote:

> Hello,
>
> i want to copy my image data on the GPU. Therefore I need to retrieve my
> image data from the itk::Image class. The method that seems suitable is
> GetBufferPointer. Hence, I know the size of my image and the pixel type I
> could get a pointer to the image data.
> I cant use itk::GPUImage because my algorithm is not implemented in ITK.
> My idea is to get the data pointer, do my computation and
> My code looks like:
>
> // get float pointer to image data
> ImageType::Pointer image = reader->GetOutput();
> Image->Update();
>
> float* data = image.GetBufferPointer();
> // copy image data to gpu
> gpu_dev->setVoxels(dimension, voxelSize, data);
> [...]
>
> My setVoxels() method is at the end of this mail.
>
> When I try to access the memory on the gpu where the data should be stored
> (at least in my imagination) I get an error.
> The error occurs at the checking line for the data input.
>
> I read [1]  and tried the suggested methods. But I want to work directly
> on the data and don't want to make any copy.
> But couldn't get any closer to a solution.
>
> My questions are:
> 1. Is there a good way to retrieve my image data other than suggested in
> [1] (less copying)?
> 2. Is it error-prone to change the image data without using itk?
> 3. Can I work directly on the itk::image data without making a copy?
>
> [1]: https://itk.org/CourseWare/Training/GettingStarted-V.pdf
>
> Thank you.
> Sincerely,
> Gordian
>
>
> bool CProjection::setVoxels( const int* dimension, const float* voxelSize,
> const float* data)
> {
>         // check input
>         std::cout << "CUDA: Dimension:" << dimension[0] << ", " <<
> dimension[1] << ", " << dimension[2] << "\n";
>         std::cout << "CUDA: Spacing:" << voxelSize[0] << ", " <<
> voxelSize[1] << ", " << voxelSize[2] << "\n";
>         std::cout << "CUDA: Data (3 values):" << data[0] << ", " <<
> data[10] << ", " << data[100] << std::endl;
>
>         cudaExtent volumeSize = make_cudaExtent(dimension[0],
> dimension[1], dimension[2]);
>         volArray_dev = 0;
>
>         // create 3D array
>         const cudaChannelFormatDesc channelDesc =
> cudaCreateChannelDesc<float>();
>         DS_ERROR_CHECK(cudaMalloc3DArray(&volArray_dev, &channelDesc,
> volumeSize));
>
>         // copy data to 3D array
>         cudaMemcpy3DParms copyParams = { 0 };
>         copyParams.srcPtr = make_cudaPitchedPtr((void*)data,
> volumeSize.width * sizeof(float), volumeSize.width, volumeSize.height);
>         copyParams.dstArray = volArray_dev;
>         copyParams.extent = volumeSize;
>         copyParams.kind = cudaMemcpyHostToDevice;
>
>         DS_ERROR_CHECK(cudaMemcpy3D(&copyParams));
>
>         // set texture parameters
>         tex_vox.normalized = false; //do not access with normalized
> texture coordinates
>
>         //use point access
>         tex_vox.filterMode = cudaFilterModePoint;
>         // clamp texture edges for out of boundary access
>         tex_vox.addressMode[0] = cudaAddressModeClamp;
>         tex_vox.addressMode[1] = cudaAddressModeClamp;
>         tex_vox.addressMode[2] = cudaAddressModeClamp;
>
>         // bind array to 3D texture
>         DS_ERROR_CHECK(cudaBindTextureToArray(tex_vox, volArray_dev,
> channelDesc));
>
>         //Set volume information on host and dev
>         for (int i = 0; i < 3; i++) {
>                 volDimension_host[i] = dimension[i];
>                 voxSize_host[i] = voxelSize[i];
>         }
>
>         DS_ERROR_CHECK(cudaMemcpyToSymbol(dc_voxDim, volDimension_host, 3
> * sizeof(int)));
>         DS_ERROR_CHECK(cudaMemcpyToSymbol(dc_voxSize, voxSize_host, 3 *
> sizeof(float)));
>
>         voxelsSet = true;
>         return voxelsSet; //on success
> }
>
>
>
>
>
> _____________________________________
> Powered by www.kitware.com
>
> Visit other Kitware open-source projects at
> http://www.kitware.com/opensource/opensource.html
>
> Kitware offers ITK Training Courses, for more information visit:
> http://www.kitware.com/products/protraining.php
>
> Please keep messages on-topic and check the ITK FAQ at:
> http://www.itk.org/Wiki/ITK_FAQ
>
> Follow this link to subscribe/unsubscribe:
> http://public.kitware.com/mailman/listinfo/insight-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/insight-users/attachments/20170302/a56550e3/attachment.html>