Re: PyCUDA and arch flags

classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view
|

Re: PyCUDA and arch flags

Andreas Kloeckner
Aleksandar Donev <[hidden email]> writes:

> We have a machine now that has both a Titan X (Maxwell) and a Titan V
> (Volta) card, which have different architectures. My students/postdocs
> have been running PyCUDA codes but I am not sure if we need to do
> anything different in this case -- does PyCUDA determine the
> architecture automatically?
>
> Based on this:
>
> https://docs.nvidia.com/cuda/volta-compatibility-guide/index.html
>
> the right thing to do is to compile with nvcc and the flags
>
> -gencode=arch=compute_61,code=sm_61
> -gencode=arch=compute_70,code=sm_70
>
> Do we need to do anything or will it just work?

Cc'ing the mailing list for archival.

You will only have one context per GPU. PyCUDA's SourceModule (assuming
that that's what you're using) will automatically pass the right arch
flags for the GPU in that context.

Andreas

_______________________________________________
PyCUDA mailing list
[hidden email]
https://lists.tiker.net/listinfo/pycuda