RadeonSI Picks Up Primitive Culling With Async Compute For Performance Wins

[jjmcwill2003]

Are companies now running RadeonSI with the Pro cards?

I work for Altair, a big CAE software company. So far, nothing has been on our development/support radar except the "Pro" drivers.

[Machine]

Many thanks to Marek

Yep, many thanks indeed

[bridgman]

Are companies now running RadeonSI with the Pro cards?

I work for Altair, a big CAE software company. So far, nothing has been on our development/support radar except the "Pro" drivers.

We are still recommending the Pro drivers and that will probably never change... but if it turns out that we can deliver a better Pro solution using the Mesa driver on Linux that is certainly an option.

One of the open questions is how important it is to our customers to have the same driver code running on Linux & Windows though.

[cb88]

Are companies now running RadeonSI with the Pro cards?

I work for Altair, a big CAE software company. So far, nothing has been on our development/support radar except the "Pro" drivers.

I think the main think blocking that in the past was lack of compatibility profiles, of which Mesa does support some at this point and with this change, Geometry processing may actually be faster than the Pro driver.

[nuetzel]

LLVM 8 isn't officially released at least in their official binaries or my distros. Kerberizer llvm-svn repo has arch binaries but they're 2 months old (his pipeline was broken by git migration). I use them and had same prob as guy above. Time to build this monster... https://github.com/llvm/llvm-project/tree/release/8.x

'Guy above' is since 1996 3Dfx/SUN (Yes) with Mesa...;-)
I'll go with LLVM git (9.0), then. - As always.

[whitecat]

What's the difference with HyperZ ?

[ernstp]

So does this make OpenGL faster than RADV Vulkan?

There's not many scenarios where you can compare that...

Edit: DOTA and Talos perhaps

[Brisse]

Guest
Maybe they are already doing it? Who knows...

[smitty3268]

The measurement shows that Nvidia's current algorithm implementing a similar feature is suboptimal. It is unlikely that hardware internals of Nvidia GPUs will prevent them to match Radeon performance.

Does their async compute work as well as AMD's now? I know it didn't when that feature first came out. Of course, it sounds like this can still be done without async, just with slightly worse performance.

Also, there is apparently an AMD patent on something similar, though Marek suggested that his code may be different enough to not fall under it. Still something for NVidia to think about, though.

[smitty3268]

I have a question. Is Vega FE considered a "pro" card in this case?

Looking at the code, it just looks for "PRO" anywhere in the cards marketing name.