[Rtk-users] CUFFT error

[ghostcz]

Hi
If I use spacing =0.4623  , size=[800 2000 1000]
------------- Device memory usage ------------
Total memory = 5368512512 bytes
Free memory = 16805888 bytes
Used memory = 5351706624 bytes
Used memory = 99.687%
----------------------------------------------then it crashed.

If I use spacing =0.462   , size=[800 2000 1000]
------------- Device memory usage ------------
Total memory = 5368512512 bytes
Free memory = 146837504 bytes
Used memory = 5221675008 bytes
Used memory = 97.2648%
----------------------------------------------didn’t crash, but a near miss.

Yes, it looks like a memory problem. But why a small change in spacing will increase the memory consumption?

Best regards,
Louie



From: Simon Rit 
Sent: Tuesday, August 04, 2015 1:27 PM
To: ghostcz 
Cc: rtk-users at public.kitware.com 
Subject: Re: [Rtk-users] CUFFT error

Then I don't see what else could be the problem. If it's a GPU memory issue, you'll see it by adding the following command
   {                                                                                                                                                                                                                                                                              
    size_t free, used, total;
    cudaMemGetInfo(&free, &total);
    used = total - free;

    std::cout << std::endl;
    std::cout << "------------- Device memory usage ------------" << std::endl;
    std::cout << "Total memory = " << total << " bytes" << std::endl;
    std::cout << "Free memory = " << free << " bytes" << std::endl;
    std::cout << "Used memory = " << used << " bytes" << std::endl;
    std::cout << "Used memory = " << (used * 100.0 / total) << "%" << std::endl;
    std::cout << "----------------------------------------------" << std::endl;
    std::cout << std::endl;
  }

before the crashing command. If the memory is the issue, then I'm not sure I understand why either, I quite agree with your comments. If it's not a memory issue, then I don't know what's the problem...

Simon


On Tue, Aug 4, 2015 at 1:05 PM, ghostcz <ghostcz at hotmail.com> wrote:

  Thank you.
  I kept the dimension to the same value. I think reducing the spacing will need less rows in the projections, because the physical length will reduce with the spacing.
  Interestingly, it also works if I use 0.0462 instead of 0.04623 as spacing.
  Yes, I use –subsetsize=1 and –division=2 while –lowmem enabled. The origin is always updated with –0.5*spacing*(size-1);

  Best regards,
  Louie

  From: Simon Rit 
  Sent: Tuesday, August 04, 2015 12:09 PM
  To: ghostcz 
  Cc: rtk-users at public.kitware.com 
  Subject: Re: [Rtk-users] CUFFT error

  Hi,

  From this page, you have

CUFFT_ALLOC_FAILED   = 2,  //  cuFFT failed to allocate GPU or CPU memoryThis is not suprising that you have a GPU memory issue when decreasing the spacing because a finer volume spacing might require  more rows of the projections for the reconstruction and therefore you can end up with being out of memory. Try processing one projection at a time with --subsetsize 1. If it's still too large, you'll have to change your --dimension to process a smaller volume in the y direction.

  Simon


  On Tue, Aug 4, 2015 at 10:49 AM, ghostcz <ghostcz at hotmail.com> wrote:

    Dear rtk users/developers,

    I was testing the rtkfdk.exe application.
    The program ran into an error and caught an exception saying:
    “
    itk::ExceptionObject (00000000002EEA60)
    Location: "unknown"
    File: C:/Users/liy/root/RTK/code/rtkCudaFFTRampImageFilter.cu
    Line: 97
    Description: itk::ERROR: CUFFT ERROR #2”
    To save you a trip to the documentation, Line: 97 is here:
    “
    if(fftDimension.z==1)
        result = cufftPlan2d(&fftFwd, inputDimension.y, inputDimension.x, CUFFT_R2C);
      else
        result = cufftPlan3d(&fftFwd, inputDimension.z, inputDimension.y, inputDimension.x, CUFFT_R2C);
      CUFFT_CHECK_ERROR(result);
    “

    This occurred when --spacing=0.04623, while the program works just fine if I use --spacing=0.05.
    Is there anyone can help me with this?

    Best regards,
    Louie

    _______________________________________________
    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/20150804/c16a7c01/attachment-0010.html>