Tonga AMDGPU Performance On Ubuntu 16.04 Has 80~90%+ Performance Of Catalyst

[W.Irrkopf]

Vulkan support is worth it for GCN1.0+. Older cards won't be migrated.

Is this because older cards don't have the necessary hardware support or "just" because AMD does not intend to provide the support (in which case open source developers still might fill the gap)?

[Tomin]

Is this because older cards don't have the necessary hardware support or "just" because AMD does not intend to provide the support (in which case open source developers still might fill the gap)?

It is explained here: https://www.phoronix.com/forums/foru...730#post845730
It's just not worth doing it for those parts, because the gains are small for users, but yes it could be done.

I personally have Radeon HD 6670 card and I could get a HD 5850 (I think that was the model) for free later this year if I wanted, but it's a bit too power hungry for my needs. I think it would be interesting to take part in adding Vulkan support for those parts, but I don't have much experience on graphics programming yet and there is no common base for that in Mesa like there is for OpenGL (another thing is that whether there is any need for that if Vulkan is too simple and low level).

[pal666]

Well, every set of possibilities has a probability distribution. See for example histogram.

The timeframe implicitly assumed by Phoronix readers also has a distribution function.

there was no timeframe in original question

Furthermore, if an event E has low probability of happening it increases the risk of an event Ecu to happen. Ecu contradicts E, and Ecu is currently unknown to us.

what exactly has low probability of happening? people writing code?

[pal666]

That's right, radeon devs can migrate support for older chips into amdgpu over time if they so choose. Question is though, is it worth it?

yes, it is the simplest way to get vulkan

[pal666]

If only one could play Alien: Isolation at all on the free driver stack...

you have no games to play until june mesa release? steam has more games than possible to play by one person

[pal666]

Well, Intel does manage to get code for upcoming chips into their drivers for the most part.

intel has less capable chips and more money

[pal666]

Is this because older cards don't have the necessary hardware support or "just" because AMD does not intend to provide the support (in which case open source developers still might fill the gap)?

technically, everything could be implemented in software. hardware needed to make it fast. if it is not supported by windows driver, it does not make sense

[pal666]

Supporting Vulkan in old GPUs has its limitations.

gcn1.0 is not that old. feel free to post more garbage

[bridgman]

Well, Intel does manage to get code for upcoming chips into their drivers for the most part.

IIRC we have also consistently released code for integrated GPUs prior to launch, from Ontario all the way to Carrizo. The dGPU market is somewhat more competitive and so more concerns about exposing plans too early (and our competition is closed-source-only) but I think we're making progress there as well.

gcn1.0 is not that old. feel free to post more garbage

These threads would be shorter and friendlier if everyone could agree on what "old" means.

Can I propose you stop using the term "old" and be specific about HW generations, at the very least saying "pre-GCN" vs "GCN" ? The Catalyst userspace drivers support all the GCN parts including SI so providing kernel/libdrm support for them makes a lot of sense, but so far I don't think any vendors have done much with Vulkan and older HW architectures, VLIW 4/5 in our case.

[duby229]

IIRC we have also consistently released code for integrated GPUs prior to launch, from Ontario all the way to Carrizo. The dGPU market is somewhat more competitive and so more concerns about exposing plans too early (and our competition is closed-source-only) but I think we're making progress there as well.



These threads would be shorter and friendlier if everyone could agree on what "old" means.

Can I propose you stop using the term "old" and be specific about HW generations, at the very least saying "pre-GCN" vs "GCN" ? The Catalyst userspace drivers support all the GCN parts including SI so providing kernel/libdrm support for them makes a lot of sense, but so far I don't think any vendors have done much with Vulkan and older HW architectures, VLIW 4/5 in our case.

I don't see that as a problem. I have no doubt that VLIW4/5 could run a Vulkan renderer (Evergreen+ I think could), but really it isn't that important. There aren't any recent products released with those architectures... But GCN 1.0 architectures are used in a number of recent products. If you use the libdrm approach on the radeon driver or you port GCN1.0 to amdgpu is your preference. (But from past conversations i got the impression that the ASIC was largely the same as NI, so the best approach probably is the libdrm approach.)