[Mesa-dev] Moving amdgpu/addrlib into a git submodule

[Nicolai Hähnle]

On 09.08.2016 17:21, Marek Olšák wrote:
> On Tue, Aug 9, 2016 at 3:47 PM, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
>> Hi everybody,
>>
>> addrlib is the addressing and alignment calculator which is used by
>> radeonsi. It's developed (and also used) internally at AMD, and so far we've
>> had one open source copy living in the Mesa repository at
>> src/gallium/winsys/amdgpu/drm/addrlib.
>>
>> The question of using addrlib in non-Mesa parts of our open-source stack has
>> come up, in particular in relation to compute. We'd obviously like to share
>> the code rather than having multiple copies flying around. Since the
>> interface of addrlib is slow-moving but unstable, shared linking is not an
>> option.
>>
>> I think the best way forward is to create a dedicated repository for addrlib
>> which is then integrated into Mesa as a git submodule.
>>
>> The point of this email is to gather feedback from the Mesa community on
>> this plan, which is explicitly:
>>
>> (0) Create an addrlib repository, say amd/addrlib on fd.o.
>> (1) Add it as a git submodule to the Mesa repository.
>> (2) Fix up whatever aspects of the build system that may be affected
>> (perhaps for building source tarballs).
>> (3) Go back to mostly ignoring addrlib, except for trying to get better at
>> syncing with the internal closed-source version.
>>
>> From initial experiments, the impact on users interested in radeon is that
>> they will have to run `git submodule init` and then occasionally `git
>> submodule update`. Users who do not build radeonsi should be able to ignore
>> the change completely.
>>
>> There are alternatives. For example, ROCm uses Google's repo tool already.
>> But for Mesa, git submodule looks like a lightweight, well supported and
>> overall conservative option that everybody should already have installed. If
>> there are good arguments for something else, let's hear them!
>>
>> Another point: if we proceed with this plan, I think we should consider
>> moving addrlib into src/amd/addrlib. There are two reasons: First,
>> transitioning to a submodule *without* changing the directory is probably
>> more fragile, i.e. what happens when you switch between checkouts before and
>> after the transition. Second, if/when radv ends up being merged into Mesa
>> master, it makes sense to have addrlib there anyway.
>>
>> Thoughts? Complaints? Praise?
>
> I don't know.
>
> How does this ensure that Mesa and ROCm addrlib copies won't diverge?

They won't really be different copies, because both "copies" are really 
checkouts from the same repository. They will occasionally be checkouts 
of _different versions_ from the same repository -- usually that would 
happen after a sync with the internal copy, when one driver updates 
their pointer before the other does. But that's easiy to reconcile. 
Usually it should just mean changing the version pointer in whichever 
driver uses the older version.


> What issues can we expect if Mesa and ROCm addrlib copies diverge?

This is about software maintenance. If we _do_ have separate copies, and 
someone applies a bug fix in one copy, they may forget to apply it to 
the other. When we want to sync with the internal copy, that has to be 
done twice. Basically, all the usual frictions that go with having the 
same (or almost the same) piece of code in more than one place.


> For texture sharing, the buffer metadata is set in a way that doesn't
> leave any room for interpretation. I think it's possible to bypass
> addrlib in this case.

Right, this is orthogonal to interoperability. Multiple drivers with 
different versions of addrlib can coexist in a system.

Nicolai

>
> Marek
>