Re: Joining ROCm team, ectrans
Hi Cory, Alastair,
On 2024-10-14 10:51, Cordell Bloor wrote:
> Upon reviewing ectrans, it looks like it is pretty straightforward. You will need to add hipcc, libhipblas-dev, librocblas-dev, libhipfft-dev, and librocfft-dev to the B-D, set HIPCXX=clang++-17 as an environment variable,
I guess this should make it into dpkg eventually, as per [2]? Alternatively:
> and specify which GPU architectures you're building for by passing -DCMAKE_HIP_ARCHITECTURES=<arch> to your CMake configure command.
>
> I'd suggest -DCMAKE_HIP_ARCHITECTURES="gfx803;gfx900;gfx906;gfx908;gfx90a;gfx1010;gfx1030;gfx1100;gfx1101;gfx1102"
For this, we (the Debian ROCm Team) should be absolutely be providing a
Makefile that could be included into d/rules, analogous to
/usr/share/dpkg/*.mk, that exports something like ROCM_SUPPORTED_GPUS
and/or ROCM_<release>_SUPPORTED_GPUS or whatever.
This does not just simplify other maintainer's lives, this allows us to
change this value Archive-wide without having to change N source
packages with hardcoded values; we just update the package with the main
definition, and trigger a binNMU.
This would be a nice to our pkg-rocm-tools or rocm-team-tools (name TBD)
package.
I'm sure that we can provide some other useful functionality, too.
The problem is that not every GPU arch is supported by newer ROCm
releases. That's going to make things complicated. I wonder whether it
wouldn't be best to split these out into -legacy packages, similar to
how the Nvidia packages are structured.
Best,
Christian
> [2]: https://bugs.debian.org/1065439
Reply to: