If you got things working earlier, the new CUDA installation should not
present a problem. What version is it? We recently made a patch to
support CUDA11, but that was tied together with some Ampere-specific
improvements. Last I knew, compiling with CUDA11 required accepting a
bunch of __syncwarp() operations which caused some slowdowns on older
architectures, but this is the wages of technological advancement. The
code is correct, it just ran noticeably slower on Turing and to a lesser
extent Volta and Pascal. There may have been other changes that kept the
performance up for older architectures while still compiling with the
necessary synchronization for Ampere, but I am not sure on that.
Barring any other changes, you should be able to do this:
(enter the build directory)
make clean
make uninstall
(re-run your customized run_cmake script from the build directory)
(check build report)
make install -j
Alternatively, the older configure script still works. Enter the
installation directory, which could be the same as the source directory but
ideally is established as a separate directory with "lndir ../${AMBERHOME}
. " to create symbolic links to all files. This will allow you to do a
configure-style build while still keeping the original source directory
clean for a cmake build.
make clean
make uninstall
./confiugre -noX11 gnu
make install -j
make clean
./configure -noX11 -cuda gnu
make install -j
Dave
On Mon, Sep 21, 2020 at 5:25 PM Camila Clemente <
camilamaraclemente.gmail.com> wrote:
> Hi!
>
> The cluster where I run has updated a new version of cuda and I must
> compile again. How should I do it?
>
> Thank you
> --
> *Camila*
> _______________________________________________
> 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 Mon Sep 21 2020 - 15:30:02 PDT
