Thank you Jesper, this could indeed be the problem.
I might get a machine with 2.0 CUDA Capabilities in a couple of weeks to
make sure, meanwhile using the OpenMP implementation.
Cheers,
Sami
-----Original Message-----
From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf
Of Jesper Andersson
Sent: Tuesday, 14 June, 2016 11:53 PM
To: [log in to unmask]
Subject: Re: [FSL] eddy_cuda
Dear Sami,
> I'm trying to run eddy_cuda on my (rather old) CUDA enabled Nvidia (Tesla
C1060 with Cuda Driver version 6.5, Cuda Toolbox 6.0 and CUDA Capability
1.3).
>
> The following error from the cuda library appears:
>
> Reading images
> Running Register
>
> ...................Allocated GPU # 0...................
> thrust::system_error thrown in
> CudaVolume::common_assignment_from_newimage_vol after resize() with
message: invalid device function terminate called after throwing an instance
of 'thrust::system::system_error'
> what(): invalid device function
> Aborted (core dumped)
>
> Is it possible that eddy_cuda was compiled with compute_20,sm_20? If so,
would it be possible to come up with a compute_13,sm_13 version?
> If it's not about the CUDA Capability, might anyone have a clue on why is
it happening?
I don't think so. The cuda binary we supply is "fat" and include compute
capabilities 1.0 up to 3.5. My guess is that the thrust library, which we
use in eddy, makes a call to a function that isn't supported by your
hardware. I haven't got access to a CUDA box with capability below 2.0, so I
haven't been able to test it.
Jesper
>
> Thanks in advance!
>
> Cheers,
> Sami
>
> Ps. My OS is Ubuntu 12.04 LTS, 64bit, using the Neurodebian release of FSL
(5.0.9) with the eddy_* binaries taken from the CentOs 6-64bit release.
|