RADV Gets A Performance Improving Workaround For Vega, Helps Dota 2

[phoronix]

Phoronix: RADV Gets A Performance Improving Workaround For Vega, Helps Dota 2

In my recent 12-way Linux GPU tests with the very newest Radeon/NVIDIA drivers the RX 580 in particular was performing great on the open-source RADV driver against the NVIDIA competition while the latest Vega GPUs were a bit lagging still. Fortunately, RADV co-founder Bas Nieuwenhuizen has landed another performance optimization...

http://www.phoronix.com/scan.php?page=news_item&px=RADV-Dota2-GFX9-Scissors

[GunpowaderGuy]

One of the points of vulkan was ending this ad hoc hackery , if the inefiency is the game's fault then the driver shouldn't bend over backwards to fix it , just expose a good way ( performance and closeness to metal /how much it diffcults cross hardware support ) to deal with it

[Masush5]

One of the points of vulkan was ending this ad hoc hackery , if the inefiency is the game's fault then the driver shouldn't bend over backwards to fix it , just expose a good way ( performance and closeness to metal /how much it diffcultks cross hardware support ) to deal with it

This doesn't work around a game bug, but rather a hardware bug present in Vega/Raven in a more performant way than previously. Dota 2 was just used as an example benchmark.

[ermo]

This doesn't work around a game bug, but rather a hardware bug present in Vega/Raven in a more performant way than previously. Dota 2 was just used as an example benchmark.

To the best of your knowledge, does Vega have many hardware bugs?

[kiffmet]

Does having to work around this bug cost a lot of performance compared to a hypothetical bug-free Vega?

[Masush5]

To the best of your knowledge, does Vega have many hardware bugs?

I just follow the mailing list; I really wouldn't know.

[VikingGe]

Dota 2 was just used as an example benchmark.

This. I don't get why people think that every mention of a specific title is "ad hoc hackery". It's not. It's an improvement that applies to everything and just happens to improve Dota 2 more than other games.

It's almost as if people think that getting better performance in a game is somehow a bad thing.

[GunpowaderGuy]

I got the impression because as i understood it , the article says that dota is cpu limited because of using the api irresponsibly , or at least not optimally , which happened to impact Vega's performance particularly bad ; apparently that is not the case ( should source have source 2 have been better optimized anyways ? ) but either way this is an ad hoc "solution"

[BNieuwenhuizen]

One of the points of vulkan was ending this ad hoc hackery , if the inefiency is the game's fault then the driver shouldn't bend over backwards to fix it , just expose a good way ( performance and closeness to metal /how much it diffcults cross hardware support ) to deal with it

We did not. Vega has a hardware bug with 2 known workarounds:

1) simple but can have a performance cost.
2) less simple but almost no performance cost.

We previously thought the cost by workaround 1 was negligible and hence it was not worth implementing 2, but I noticed that rendering the GUI in dota2 there was worst case behavior for workaround 1 which caused some slowdown.

Part of the difficulty finding this was that my CPU is worse than what Michael uses for his benchmarks, so I'm CPU limited on 4k with my Vega, and hence did not see the performance difference until I forced the GPU to low power.

[smitty3268]

To the best of your knowledge, does Vega have many hardware bugs?

Every GPU has hardware bugs. They get designed and sent out to production much faster than something like a traditional CPU, with much more frequent major architectural changes. Plus they only typically expose higher level APIs such as GL, DX, etc. that can much more easily work around any problems in the driver whereas a CPU only has firmware available to fix things.