[Rtk-users] Issues with using ChangeInformationImageFilter on CudaImage

[Chao Wu]

Hi all,

There seems to be a CPU-GPU synchronization issue if an
itk::ChangeInformationImageFilter<itk::CudaImage<...>> is inserted before a
Cuda filter. It works if the ChangeInformationImageFilter is inserted
before a CPU filter.

The issue can be reproduced with the rtkSartCudaTest and the following
patch:

========================================================================
*File: include/rtkSARTConeBeamReconstructionFilter.hxx*
>
>
@@ -18,10 +18,11 @@

 #ifndef rtkSARTConeBeamReconstructionFilter_hxx
 #define rtkSARTConeBeamReconstructionFilter_hxx

 #include "rtkSARTConeBeamReconstructionFilter.h"
+#include <itkChangeInformationImageFilter.h>

 #include <algorithm>

 namespace rtk
 {
@@ -282,10 +283,15 @@ SARTConeBeamReconstructionFilter<TVolumeImage,
TProjectionImage>

   // Declare the image used in the main loop
   typename TVolumeImage::Pointer pimg;
   typename TVolumeImage::Pointer norm;

+  // test: insert itkChangeInformationImageFilter
+  itk::ChangeInformationImageFilter<VolumeType>::Pointer change =
itk::ChangeInformationImageFilter<VolumeType>::New();
+  change->SetInput(this->GetInput(0));
+  m_ForwardProjectionFilter->SetInput(1, change->GetOutput());
+
   // For each iteration, go over each projection
   for(unsigned int iter = 0; iter < m_NumberOfIterations; iter++)
     {
     unsigned int projectionsProcessedInSubset = 0;
     for(unsigned int i = 0; i < nProj; i++)
@@ -326,11 +332,11 @@ SARTConeBeamReconstructionFilter<TVolumeImage,
TProjectionImage>
         else
           pimg = m_AddFilter->GetOutput();
         pimg->Update();
         pimg->DisconnectPipeline();

-        m_ForwardProjectionFilter->SetInput(1, pimg );
+        change->SetInput(pimg);
         m_AddFilter->SetInput2(pimg);
         m_BackProjectionFilter->SetInput(0,
m_ConstantVolumeSource->GetOutput());
         m_BackProjectionNormalizationFilter->SetInput(0,
m_ConstantVolumeSource->GetOutput());

         projectionsProcessedInSubset = 0;
========================================================================

Results from the rtkSartCudaTest:

****** Case 1: Voxel-Based Backprojector ******

Error per Pixel = 0.0206672
MSE = 0.00282352
PSNR = 31.5127dB
QI = 0.989666


Test PASSED!


****** Case 2: Voxel-Based Backprojector, OS-SART with 2 projections per
subset ******

Error per Pixel = 0.0225105
MSE = 0.00346037
PSNR = 30.6294dB
QI = 0.988745


Test PASSED!


****** Case 3: Joseph Backprojector ******

Error per Pixel = 0.0189275
MSE = 0.00259909
PSNR = 31.8724dB
QI = 0.990536


Test PASSED!


****** Case 4: CUDA Voxel-Based Backprojector ******

Error per Pixel = 16.5716
MSE = 525.278
PSNR = -21.1833dB
QI = -7.28582
Test Failed, Error per pixel not valid! 16.5716 instead of 0.032


Best regards,
Chao
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://public.kitware.com/pipermail/rtk-users/attachments/20180720/8a459487/attachment-0001.html>