[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: ROCM host architectures



Hi Étienne,

On 2/27/23 14:46, Étienne Mollier wrote:
Hi all,

Cordell Bloor, on 2023-02-22:
On 2/21/23 14:06, Christian Kastner wrote:
I propose that unless we expect this builds to pass and the packages to
also be usable, we should initially limit to:

      Architecture: amd64 arm64 ppc64el
That seems reasonable.
I initially hoped to bring some support at least to the riscv64
architecture, as it seemed a promising platform in the future,
but resolving the current build issue remained out of my reach
for now.  The selection indicated by Christian looks fair enough
to me as well.

Finding test hardware is perhaps the most difficult part of this. Do you know if the AMDGPU kernel module works on riscv64? The only non-x86 hardware that I know of that uses AMD GPUs is the Raptor Talos II [1].

I think the reason HIP doesn't work on pcc64el is because hipcc tries to use `-lclang_rt.builtins-$(uname -m)`, but that library doesn't exist on that architecture. My understanding is that the purpose of the rt builtins is to enable fp16 support [2], so perhaps it could just be omitted on platforms where it doesn't exist.

On a related note, there's a pair of upstream patches that (A) fix clang version parsing on Debian [3], and (B) remove direct references to clang_rt.builtins from hipcc [4]. We still do have one more upload of HIP coming to allow the versions with all the bugfixes to migrate to Bookworm, but it's pretty late in the cycle for big patches. Nevertheless, they're interesting and relevant so I wanted to mention them.

Sincerely,
Cory Bloor

[1]: https://www.phoronix.com/news/AMDKFD-Compute-POWERPC
[2]: https://github.com/ROCm-Developer-Tools/HIP/pull/2219
[3]: https://github.com/ROCm-Developer-Tools/HIP/commit/0c443d12011da16a036057e0472ae59c68bc901f [4]: https://github.com/ROCm-Developer-Tools/HIP/commit/a388eb1e6ffa19184a10078353bd8e9dbd739a60


Reply to: