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

Re: Towards mature AMD GPU compute support



Hi Xuanteng,

I'm dropping Debian Science from this thread because this level of detail is not relevant to that audience.

On 2025-02-16 23:00, Xuanteng Huang wrote:
I’m drafting a release plan for ROCm 6.3.2 [1], please let me know anything missing or should be revised.

Thanks. One new package in ROCm 6.3 is hipblas-common [2], which is essentially just a header file that defines a few enums used by both hipBLAS and hipBLASLt [3]. An important change for rocBLAS in ROCm 6.3 is that it depends on hipBLASLt. The purpose of that dependency relationship is essentially to centralize the database of assembly kernels.

I'm not sure we can get both hipblas-common and hipblaslt into unstable in time for Trixie, so I suspect that we will probably end up shipping rocBLAS 6.2. That's fine. We should still try to get everything else onto ROCm 6.3.

So should we skip the 6.1.2 release and directly work on 6.3.2, in consideration of the Trixie freeze?

Yes and no. We almost have the HIP Runtime and below on ROCm 6.1. We should finish that by uploading rocm-hipamd 6.1.2. In parallel with that, we can begin updating all the mathlibs directly to ROCm 6.3.2. Then, when rocm-hipamd 6.1.2 clears NEW, we can update to rocm-hipamd 6.3.2.

tl;dr: the runtime should not be skipping ROCm 6.1.2, but it's fine for the mathlibs to go directly to ROCm 6.3.2.

[1]: https://salsa.debian.org/rocm-team/community/team-project/-/merge_requests/1
[2]: https://github.com/ROCm/hipBLAS-common
[3]: I suggested wrapping the enum definitions in preprocessor guards and just defining them in both projects (like was done when hipsolver needed a few hipblas enums), but alas my motion did not carry.

Reply to: