David Airlie Tackling RADV Vulkan Conformance

[airlied]

How? RADV currently isn't completely Vulkan compliant, so I wouldn't consider that a better option. The closed drivers have "adequate" performance. I don't see why the closed driver is bad if it runs everything you need it to (assuming you follow its guidelines).

For the record, I'm not saying the closed drivers are good, just simply adequate. I'm not favoring them either. I have no doubt RADV can eventually be better.

You do understand, this set of results means radv is Vulkan compliant, it just hasn't been submitted to the procedure for getting it stamped.

[pal666]

Why are you recommending people use anv which only supports incapable hardware, and not radv which supports a wide variety of capable hardware?

because for playing with api anv is more than enough

[airlied]

we are commenting on article which says you are keeping up 54.7% of time. and in previous articles you were few times slower than radeonsi

Vulkan compliance doesn't work like you think. The driver is compliant in all but name now. No driver runs 100% of the tests. Someone should run vulkan 1.0.2 against amdgpu-pro 16.60 and see what the results are, you might be surprised.

I expect tessellation shaders would bring the number up quite a bit, number of vulkan apps in existance using tessellation? 0.

Otherwise I don't think we are missing anything to really make the conformance figures that much different, as-is the driver in that branch passes conforomance.

Dave.

[pal666]

it is more like: every day the users of GCN1.0-1.1 will go down... because hardware broke or just by natural-obsolescence and users of 1,2 and newer will go up.

there are still users of pre-gcn hardware. and there is no full-fledged amd driver for it, so they really could make use of mesa improvements without duplication

[duby229]

because for playing with api anv is more than enough

Unless you have an amdgpu, in which case radv is more than enough. I can see if you are only running an Intel APU, then fine anv is your only option. But if you have an AMD card in your system then radv is obviously the better option. Anv really only is an option when it's the only option. That's a perfectly fine reason for it to exist, but lets recognize it for what it is.

[pal666]

Vulkan compliance doesn't work like you think. The driver is compliant in all but name now. No driver runs 100% of the tests. Someone should run vulkan 1.0.2 against amdgpu-pro 16.60 and see what the results are, you might be surprised.

the nice thing about amdgpu-pro is that it will be improved even if you retire

I expect tessellation shaders would bring the number up quite a bit, number of vulkan apps in existance using tessellation? 0.

it does not make radv working solution

Otherwise I don't think we are missing anything to really make the conformance figures that much different, as-is the driver in that branch passes conforomance.

what is your percentage on vega?

[indepe]

what features? We are keeping up pretty well, especially with features with compliance tests.

Dave.

I didn't quite understand (either): You seem to be talking about your WIP branch, and the must-pass tests, when you say you pass 54%. So that sounds like you are only half-conformant (unless 100% refers to something else than the must-pass list).

[pal666]

radv is more than enough.

radv does not come for free. amd will not stop developing amdgpu-pro because of radv, intel will not stop developing anv because of radv. but airlied stopped developing mesa because of radv. so you have duplicate driver and lost man-year of improvements of something which people really need

[indepe]

radv does not come for free. amd will not stop developing amdgpu-pro because of radv, intel will not stop developing anv because of radv. but airlied stopped developing mesa because of radv. so you have duplicate driver and lost man-year of improvements of something which people really need

One may hope that RADV will eventually be faster, as Mesa OpenGL is significantly faster.

[schmidtbag]

it is more like: every day the users of GCN1.0-1.1 will go down... because hardware broke or just by natural-obsolescence and users of 1,2 and newer will go up.

so in fact it is more important to get the features on new hardware instead of over-engineering support of old hardware.

Every day is a bit of a stretch, but yes, older GCN hardware will remain on a decline. But that doesn't mean older hardware should be ignored. Keep in mind, I'm not just arguing for GCN 1.0-1.1 - like I said before, there are others in need of Vulkan attention too. Anyway, older GCN hardware isn't obsolete. It may be inefficient, but still very capable.

the users can just buy new hardware: problem solved.

Not everyone can just throw away hundreds of dollars when their perfectly functional and par hardware should be capable. If money was never an issue, Nvidia would have a lot more sales (maybe not for Linux specifically, but in general).

@arlied
Pal666's comments seem to address your questions toward me. Considering he and I disagree with each other a LOT more often than we agree, that should stress the significance of my perspective. Though, personally I have a more optimistic outlook. I do see a good future for RADV, and steady improvements.