On Sun, Jan 23, 2022, Michael Coleman wrote:
>
>The documented Amber build procedure isn't working for us
It would be great (if you have the time) to see what is going wrong with
the regular build procedure. Is it this problem?
>The 'cuda.10.2.89' works around an exasperating undocumented behavior of
>the Amber build, which is that GPU code is generated or not generated for
>different cards depending on the CUDA version number. In particular, K80
>code is not generated for CUDA 11, even though CUDA 11 supports K80s. The
>Amber code doesn't fail obviously for lack of this code, but rather one
>gets an obscure diagnostic.
For Amber developers: is there a reason that support for SM3.0 is turned
off for Cuda 11 (in CudaConfig.cmake)? Note that (officially?) the K80
card is supposed to use SM3.7, but there is a comment (outdated?) in
CudaConfig.cmake that SM3.0 "may be better". But then we turn off SM3.0 for
Cuda 11, and never use SM3.
For Mike: if the above is the problem, you might try just editing
CudaConfig.cmake for your installation to allow either SM3.0 or SM3.7 code
to be generated. And if other things are problematic, it would good to know
about them. In any event, thanks for your work on this, including on spack.
....dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon Jan 24 2022 - 06:30:02 PST
