Testing The Open-Source "RADV" Radeon Vulkan Driver vs. AMDGPU-PRO

[haagch]

Maybe AMD could simply put some of the people who worked on the closed source vulkan driver to work on radv, even if only part time. The experience could be helpful.

Anyway, for me the important thing to fix right now would be the GPU hangs on polaris. I still haven't bisected and if nobody does it soon, I probably have to do it...

[jonwil]

Also I suspect given the fast progress being made on RADV, the best way forward would be for the AND guys (perhaps with some input from the RADV developers) to figure out which bits of code would be the most useful in terms of making RADV complete and focus on getting the legal reviews (and any changes/work necessary to pass those reviews) happening for those bits of code rather than the other bits that RADV has already covered (or can easily do without AMD help)

[jonwil]

We have opened specifications for consumer PC GPUs, but we also have an active semi-custom business as well.

Last I checked all 3 major game consoles (Wii U, PS4, XB1) are using ATI GPUs of various sorts so I assume that's part of the "custom" business being talked about

[suberimakuri]

Geez, give the guy (bridgeman) a break.
It will be what it will be, he's obviously doing what he can.

[bridgman]

Ah, OK. Yeah, the review is quick... the hard part is dealing with the issues that come out of the review.

In this case we knew most of the issues going in, so dealing with them was more of a priority than any kind of formal review anyways.

[pal666]

if any of those were the plan I would have said that

i was arguing not with you, but with people imagining army of amd linux vulkan developers wasting their time waiting for lawyers doing legal review

[bridgman]

I still don't understand: you have GPU specification opened, and Windows API for writing drivers are open either. So what kind of non-public information the sources could possibly have?

We have opened specifications for consumer PC GPUs, but we also have an active semi-custom business as well. If you look at Windows development kits I think you will find they are not "open as in open source", just downloadable under a "for your use only" license. Even so, in terms of OSes we support more than just Windows.

At the risk of stating the obvious, I can't talk about things that we keep secret.

[artivision]

I can see Vulcan coming everywhere on Linux, even on HD5-6K series and Fermi.

[Hi-Angel]

The driver stack that implements Vulkan was written to leverage code and developer effort across multiple APIs and multiple OSes. In order to open source one piece of that stack (Vulkan on Linux) some pieces need to be rewritten so they no longer share code with those other OSes and APIs, at least not the ones which require non-public information to implement (unfortunately that is pretty much all of them).

I still don't understand: you have GPU specification opened, and Windows API for writing drivers are open either. So what kind of non-public information the sources could possibly have?

[bridgman]

So… Do you know what do they review? Why is it taking so long?

I wish everyone would stop talking about "legal review"; there is no such thing other than legal requiring engineering to confirm that we have rights to publicly distribute all of code and associated IP. The review and mitigation planning is done by architects, not legal staff.

The driver stack that implements Vulkan was written to leverage code and developer effort across multiple APIs and multiple OSes. In order to open source one piece of that stack (Vulkan on Linux) some pieces need to be rewritten so they no longer share code with those other OSes and APIs, at least not the ones which require non-public information to implement (unfortunately that is pretty much all of them).

Why is this happening, given we're talking about a completely new driver which definitely was written with open sourcing it mind? Also, bridgman (no 'e') mentioned some rewrites of the driver: again, what do they rewrite, they just created it!

You seem to be thinking that the driver was written from scratch just for Linux/Vulkan and for open sourcing, which is pretty much the opposite of what we have said. The Vulkan driver is part of a closed-source stack built on existing closed-source code which primarily supports non-open APIs and non-open OSes.

The stack was written with awareness that we would probably want to open source support for one of the dozen-or-so API/OS combinations in the future, and the code was structured to separate open-able from non-open-able to the extent that could be done without significantly slowing the overall project, but other than that it was a closed source project.