[Rtk-users] rtkfdk doesn't work correctly after a recent update

Fri Nov 20 11:28:22 EST 2015

Hi Yang,
Thanks a lot for sharing the data. I have good news. I could reproduce the
bug and I have solutions. First, about your dataset: it's a huge dataset
and the hannY requires zero padding in both x and y. Because of this, cufft
has to take the 2D FFT of projection subsets that are 4095x2048x16. Quite
big!
When I ran it on my cuda 7.0, I immediately got:

itk::ExceptionObject (0x8b9cb30)
Location: "void CUDA_fft_convolution(const int3&, const int2&, float*,
cufftComplex*)"
File: /home/srit/src/rtk/rtk/code/rtkCudaFFTConvolutionImageFilter.cu
Line: 97
Description: itk::ERROR: CUFFT ERROR #2

i.e., a GPU alloc issue, even if I have the largest possible memory on a
nvidia desktop computer. Then, I ran it with cuda 6.5 and I got no error
message but a result similar to yours.
Conclusions:
- if you add the option --subsetsize 1 to your rtkfdk command line, that
should reconstruct correctly
- if you upgrade cuda, you should get an understandable error message,
- I don't think I'm going to look for a solution because the pb is probably
in cufft but if someone is willing to, be my guest!
Simon

On Mon, Nov 16, 2015 at 4:47 PM, Yang-Kyun Park <theday79 at gmail.com> wrote:

> Hi Simon,
>
> For better understanding, I'm attaching the problematic case including
> projection files, geometry file(xml) and reconstructed images.(please find
> the GDrive links. The file size is ~800 MB)
> This is phantom data and there is no patient information in it.
>
> Regarding the fftw library, I'm using [fftw-3.3.4-dll64].
>
> By the way, I found another clue. When I've tried it with a head phantom
> image (full-fan, ~ 190 projections), there was no problem. Is there any
> possibility that DDF and parker short filter play some roles on this issue?
>
> Feel free to let me know if further data is needed.
>
> Thanks!
>
> Yang
>
> 
>  FFTW_issue.zip
> <https://drive.google.com/file/d/0BxuP3mON9V-qbVdQd0luc1lIb2M/view?usp=drive_web>
> 
>
> On Mon, Nov 16, 2015 at 10:07 AM, Simon Rit <
> simon.rit at creatis.insa-lyon.fr> wrote:
>
>> Hi,
>> Thanks. It's really weird... When you run the tests with FFTW (compiling
>> the "test" target), do they fail?
>> Can you provide your geometry file? It might be specific to this geometry.
>> Thanks again for looking into this issue, we'll find what's the problem.
>> Simon
>>
>> On Mon, Nov 16, 2015 at 3:54 PM, Yang K Park <theday79 at gmail.com> wrote:
>>
>>> Hi Simon,
>>>
>>>
>>>
>>> Sorry for the late response.
>>>
>>>
>>>
>>> As you suggested, I’ve tried rtkramp+CUDA with only one projection and
>>> it was successfully done for both w/ FFTW and w/o FFTW.
>>>
>>> (Two results were exactly same.)
>>>
>>>
>>>
>>> Regarding  the CUDA version, I’m using v6.5.
>>>
>>>
>>>
>>> Thanks a lot for the help!
>>>
>>>
>>>
>>> Yang
>>>
>>>
>>>
>>>
>>>
>>> *From:* simon.rit at gmail.com [mailto:simon.rit at gmail.com] *On Behalf Of *Simon
>>> Rit
>>> *Sent:* Thursday, November 12, 2015 5:59 PM
>>>
>>> *To:* Yang K Park <theday79 at gmail.com>
>>> *Cc:* rtk-users at openrtk.org
>>> *Subject:* Re: [Rtk-users] rtkfdk doesn't work correctly after a recent
>>> update
>>>
>>>
>>>
>>> Hi,
>>>
>>> Ok. Can you try rtkramp with only one projection? Then you shouldn't
>>> encounter this regions issue (that should be solved one day but that's a
>>> different story).
>>>
>>> Can you also indicate which version of CUDA you're using?
>>>
>>> FYI, we use ITK for the FFT of the kernel of the ramp filter and cufft
>>> for the FFT of the projections. So if they are not in the same format,
>>> you're in trouble and that might be the issue here... The weird part is
>>> that I didn't reproduce this problem on linux (with and without FFTW) but
>>> it might be a CUDA version difference. I'll give it another try tomorrow
>>> when you provide the CUDA version.
>>>
>>> Thanks again for reporting the issue,
>>> Simon
>>>
>>>
>>>
>>> On Thu, Nov 12, 2015 at 11:39 PM, Yang K Park <theday79 at gmail.com>
>>> wrote:
>>>
>>> Hi Simon,
>>>
>>>
>>>
>>> Thanks for the advice. The current RTK was compiled with FFTW to use the
>>> phase extraction feature from Amsterdam shroud.
>>>
>>>
>>>
>>> Regarding your suggestions,
>>>
>>>
>>>
>>> 1)      rtkramp w/ or w/o CUDA:
>>>
>>> (command line: “ rtkramp --path projPath --regexp .*.his --hardware cuda
>>> --output Ramp_wCUDA.mha” )
>>>
>>>
>>>
>>> Then, I have encountered following error.
>>>
>>> “CudaFFTRampImageFilter<..> CudaFFTConvolutionImageFilter assumes that
>>> input requested and buffered regions are equal.”
>>>
>>>
>>>
>>>
>>>
>>> Without CUDA, it was successfully done.
>>>
>>>
>>>
>>> 2)      I’ve re-compiled *ITK w/o FFTW* and built RTK again. r*tkfdk
>>> worked fine.*
>>>
>>> -          I’ve also repeated “rtkramp+CUDA” with this rtk version and
>>> the error was same as above.
>>>
>>>
>>>
>>> In conclusion, my *FFTW seems to affect CUDA-based fdk reconstruction
>>> only*.
>>>
>>> But I do need FFTW for the phase extraction. What should I do?
>>>
>>>
>>>
>>> Thanks.
>>>
>>>
>>>
>>> Yang
>>>
>>>
>>>
>>>
>>>
>>> -----Original Message-----
>>> From: simon.rit at gmail.com [mailto:simon.rit at gmail.com] On Behalf Of
>>> Simon Rit
>>> Sent: Wednesday, November 11, 2015 4:16 PM
>>> To: Yang K Park <theday79 at gmail.com>
>>> Cc: rtk-users at openrtk.org
>>> Subject: Re: [Rtk-users] rtkfdk doesn't work correctly after a recent
>>> update
>>>
>>>
>>>
>>> Hi,
>>>
>>> I didn't manage to reproduce the two problems on linux with the same ITK
>>> version.
>>>
>>> For point 1, did you copy itkMacro.h by hand? Because I don't see how
>>> replacing <itkMacro.h> by "itkMacro.h" can help if itkMacro.h is not in the
>>> same folder as rtkCudaUtilities.hcu For point 2, no clue. Do you use FFTW?
>>> It looks like a ramp filter issue. Can you try rtkramp --path projPath
>>> --regexp .*.his, with and without --hardware cuda? Maybe you can provide
>>> your geometry file in case it's a specific geometric issue.
>>>
>>> Simon
>>>
>>>
>>>
>>> On Wed, Nov 11, 2015 at 7:50 PM, Yang K Park <theday79 at gmail.com> wrote:
>>>
>>> > Hi all,
>>>
>>> >
>>>
>>> >
>>>
>>> >
>>>
>>> > I’ve updated rtk with the up-to-date revision and have encountered a
>>>
>>> > couple of problems.
>>>
>>> >
>>>
>>> >
>>>
>>> >
>>>
>>> > 1.       Compile error: RTK couldn’t find itkMacro.h while it compiled
>>>
>>> > rtkCudaUtilities.hcu. It was solved by a minor fetching:
>>>
>>> >
>>>
>>> > #include <itkMacro.h>   à #include "itkMacro.h" in rtkCudaUtilities.hcu
>>>
>>> >
>>>
>>> >
>>>
>>> >
>>>
>>> > 2.       Output of rtkfdk with CUDA was not the same as before. It was
>>> very
>>>
>>> > blurry as seen in the attached image.
>>>
>>> >
>>>
>>> > a.       With “FDK with CPU” or “SART with CUDA”, there was no problem.
>>>
>>> >
>>>
>>> >
>>>
>>> >
>>>
>>> >
>>>
>>> >
>>>
>>> > I’m using VS2013 with win7, itk4.8.1.
>>>
>>> >
>>>
>>> >
>>>
>>> >
>>>
>>> > The command line I used was :
>>>
>>> >
>>>
>>> > rtkfdk --geometry 1111.xml --path projPath --verbos --regexp .*.his
>>>
>>> > --output output.mha --spacing 1,1,1 --dimension 400,200,400 --hardware
>>>
>>> > cuda --pad 1 --hann 5 --hannY 5
>>>
>>> >
>>>
>>> >
>>>
>>> >
>>>
>>> > Does anyone have any idea about this?
>>>
>>> >
>>>
>>> >
>>>
>>> >
>>>
>>> > Thanks.
>>>
>>> >
>>>
>>> >
>>>
>>> >
>>>
>>> > Yang
>>>
>>> >
>>>
>>> >
>>>
>>> >
>>>
>>> >
>>>
>>> > _______________________________________________
>>>
>>> > Rtk-users mailing list
>>>
>>> > Rtk-users at public.kitware.com
>>>
>>> > http://public.kitware.com/mailman/listinfo/rtk-users
>>>
>>> >
>>>
>>>
>>>
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/rtk-users/attachments/20151120/4e6c0e1e/attachment-0010.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image001.png
Type: image/png
Size: 5752 bytes
Desc: not available
URL: <http://public.kitware.com/pipermail/rtk-users/attachments/20151120/4e6c0e1e/attachment-0010.png>