Hi Brian,
This sort of thing can occur when the CUDA runtime library that gets used at runtime does not match the CUDA version used to compile the code. This is inline with what you suggest below regarding CUDA 7 etc. I would suggest starting from a brand new Amber untar and build it again. Make sure that your CUDA_HOME variable (and anything in your LD_LIBRARY_PATH) match the cuda compiler version that 'nvcc -V' returns.
If the problem still persists let me know.
All the best
Ross
> On Mar 25, 2015, at 10:30 AM, Brian Radak <brian.radak.accts.gmail.com> wrote:
>
> I've been able to successfully compile pmemd.cuda on a new desktop running
> Ubuntu 14.04 with a Quadro K2200 (I know it probably won't be fast, but it
> has Kepler architecture so it should be good for small tests).
>
>
> This was my sequence:
> configure -cuda gnu
> make -j8
> cd test && make test.cuda
>
> However all of the tests error out and most or all give the msg:
>
> cudaMemcpyToSymbol: SetSim copy to cSim failed invalid device symbol
>
> Suggestions?
>
> Possibly relevant - Without thinking I originally installed cuda 7.0. I
> then had to uninstall that as well as the updated driver (346 or
> something?) and then install the "old" 6.5/340 toolkit/driver. I don't
> *think* that was a problem, bc I can compile and run the cuda-samples.
>
> Thanks,
> Brian
>
> --
> Brian Radak
> Postdoctoral Scholar
> University of Chicago
> Department of Biochemistry & Molecular Biology
> Gordon Center for Integrative Science, W323A
> 929 E. 57th St.
> Chicago, IL 60637-1454
> Tel: 773/834-2812
> e-mail: radak.uchicago.edu
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Mar 25 2015 - 11:00:03 PDT
