Re: RFC: Register Build-Profile for ROCm and CUDA?
On 2025-05-06 01:07, Cordell Bloor wrote:
> However, the biggest reason why ROCm-enabled packages take so long to
> build is that they're being built a dozen times in order to cover all
> GPU architectures. When working on HIP-enabled packages, I often trim
> the GPU targets list down to a single GPU until I'm ready for my final
> test.
That's what I tend to do as well, but it's still a few extra steps. With
a build profile, simply passing --profiles=norocm to sbuild when needed
would suffice.
> I wonder if perhaps we could solve this problem by providing a
> centralized way to specify the HIP build targets and create a standard
> way to override them. A few developers have expressed the need for
> something like that anyway, most recently with regards to spfft [1].
We should (I recall mentioning this once, ages ago, but I don't know why
I never followed up on that).
More specifically, there should be a single utility, maintained by the
Debian ROCm Team, which for current release emits the list of
architectures supported by the ROCm stack in that release.
Optional features could be to print this info for other releases, or to
print the list of architectures installed on the current host.
This information should also be included in the source package metadata,
something like an X-AMDGPU-Targets field, so that reverse dependencies
can be binNMUed whenever the team changes the supported GPU list for a
release.
However, this is slightly complicated by the fact that the list of
supported GPUs is not universal AFAIUI. I think it was MIOPen, or one of
its dependencies, where we first ran into the issue that fewer GPUs are
supported than other ROCm packages.
Best,
Christian
Reply to: