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

Re: ROCm bump to 5.7 - rocm-smi-lib SOVERSIONs bump



On 2023-09-17 17:12, Étienne Mollier wrote:
> M. Zhou, on 2023-09-17:
>> A very straightforward way to confirm this is to grep the pattern
>>   shared_mutex_init
>> within the source code of its reverse dependencies. If there is no
>> package directly calling this API, then it is fine to go ahead
>> without SOVERSION bump.

Generally: no, as this could still break binaries that a user compiled
locally against the old version of the library, so a bump would still
be needed.

Having said that:

> The only oam reverse dependency I identified potentially was the
> rccl (obtained via `build-rdeps`, `apt rdepends` didn't show any
> other reverse dependency than liboam-dev), and it doesn't look
> to make use of such symbol.  Adding slurm-wlm in the pot, as it
> is the reverse dependency making librocm-smi64-1 a key package,
> I don't see it using this symbol either.
> 
> In ROCm SMI lib, shared_mutex_init is only used for the Device
> constructor.  It is probably not part of the API, and probably
> shouldn't be exposed in the first place in the ABI.  I think it
> should be fine to not bump.  I also took some time to recheck
> what was the status with ROCm SMI lib from ROCm 5.2.3, and the
> rocm_smi64 SOVERSION delta was already there.

That (and you previous analysis) sounds reasonable; also, it's not part
of the documentation [1], and the name clashes with the naming scheme
(which is otherwise clean), so it's more than likely that you are right.

Best,
Christian

[1] https://rocm.docs.amd.com/projects/rocm_smi_lib/en/latest/.doxygen/docBin/html/index.html


Reply to: