Radeon RADV+ACO Vulkan Performance Is In Great Shape For Mesa 20.2

[cl333r]

Mesa 20.2 stable with this new default should debut around the end of August and make it into the autumn Linux distribution releases.

I thought everybody in the US says Fall instead of Autumn.

[anth]

I wonder how these compilers would compare with a slower CPU. The main advantage of ACO is compile time, so we might see a bigger difference in minimum frame rate.

[zboszor]

Michael I've been playing a lot of Mad Max with the Vulkan rendered and had huge issues with micro-stuttering, even on my new 3950X with 32 GB of RAM. Still with my R290 though.

The thing is that when I tried the same game with the same settings in Windows, I got a steady 60 fps consistently.

So I was hoping that perhaps ACO will solve those micro-stutters. You would be able to see those on your graph of frame-times, right?

It would be interesting to see you test that, even though you probably wouldn't be testing on a R290. But if you would test on that (perhaps together with a slew of older cards that still support Vulkan and ACO) it would be great! I'd give you a $10 tip, since I'm already a lifetime member

I think you might be interested in this patch: https://gitlab.freedesktop.org/mesa/..._requests/5865
According to the MR description, cache misses have caused stuttering which is greatly smoothed by some extra caching in a hash.

[TemplarGR]

Michael I've been playing a lot of Mad Max with the Vulkan rendered and had huge issues with micro-stuttering, even on my new 3950X with 32 GB of RAM. Still with my R290 though.

The thing is that when I tried the same game with the same settings in Windows, I got a steady 60 fps consistently.

So I was hoping that perhaps ACO will solve those micro-stutters. You would be able to see those on your graph of frame-times, right?

It would be interesting to see you test that, even though you probably wouldn't be testing on a R290. But if you would test on that (perhaps together with a slew of older cards that still support Vulkan and ACO) it would be great! I'd give you a $10 tip, since I'm already a lifetime member

Perhaps there is some configuration needed. I don't own Mad Max so i can't test it, but for example in Civ VI i had a lot of issues, it couldn't even load properly it took tens of minutes to allow me to start a map. I then found out that the issue was a low ulimit. For some odd reason the port didn't ask for a higher one and stayed at the default 1024 and that caused a lot of issues and freezes. I had to put every single setting at the lowest just so i can start a game properly. But then i went into systemd configuration and changed the default ulimit and then Civ VI ran properly, even faster than on Windows. I heard that this solves issues in a lot of games, and it is even recommended by Lutris to do if you are using esync with WINE.

Also, perhaps the issue could be with the number of threads. Some older ports have issues with more than a certain number of cores. Maybe that's it? I don't know, just throwing some ideas here. Sometimes even Vsync could be the culprit. Or even just a bad port.

[Mez']

I wonder how these compilers would compare with a slower CPU. The main advantage of ACO is compile time, so we might see a bigger difference in minimum frame rate.

As an absolute non geek, I'm also wondering how a compiler can affect the performance of the end result driver. My initial non geek thought would be that it's just a matter of time of compilation, not performance.
Some vulgarization from specialists would be welcome.

[ntropy]

Vega 64 runs amazing with aco ever since it got released. The issues I had with Linux Gaming got purged more and more by the good and eager devs of the community. Regarding issues with mesa-aco I can report none for half a year now.

Btw, Apex Legends runs with 144 fps on 1080p and very high settings

As an absolute non geek, I'm also wondering how a compiler can affect the performance of the end result driver. My initial non geek thought would be that it's just a matter of time of compilation, not performance.
Some vulgarization from specialists would be welcome.

If you render a frame with the GPU, the frame will go through a pipeline of elements in which the final picture gets composed. The most basic pipeline has a set of vertices at the beginning, so a set of coordinates and then it goes into a vertex shader, a tesselation shader, a fragment shader and much more. This is the so called render pipeline. All these shaders get compiled real-time and it depends on the game how its done. Some games pre-compile a lot of shaders on first load, others are more gpu-intensive and calculate everything anew with real lightning and so on. These little shader compilation, need a compiler and llvm wasnt made for this originally. So with aco there was a big boost

[Azpegath]

Azpegath are you sure that you are using vulkan? I can't test it right now, but iirc you should select vulkan on the launcher. With a lower end than yours, I got 60fps easily

Yes, I'm certain. The FPS counter in the game actually said it was a fairly stable 60 FPS. I was running at highest quality, with 60 fps but stuttering.

Of course it could be related to power settings on the card or something.

[Mez']

If you render a frame with the GPU, the frame will go through a pipeline of elements in which the final picture gets composed. The most basic pipeline has a set of vertices at the beginning, so a set of coordinates and then it goes into a vertex shader, a tesselation shader, a fragment shader and much more. This is the so called render pipeline. All these shaders get compiled real-time and it depends on the game how its done. Some games pre-compile a lot of shaders on first load, others are more gpu-intensive and calculate everything anew with real lightning and so on. These little shader compilation, need a compiler and llvm wasnt made for this originally. So with aco there was a big boost

It's still quite abstract to me...
But if it compiles complex stuff in real-time during the games (I thought it was the radeon driver package compilation ), I understand why the use of one or the other can lead to differences in performance. Your explanation served its purpose. Thanks for taking the time.

[BNieuwenhuizen]

I think you might be interested in this patch: https://gitlab.freedesktop.org/mesa/..._requests/5865
According to the MR description, cache misses have caused stuttering which is greatly smoothed by some extra caching in a hash.

That is unrelated to shader cache misses, and is mostly around stuttering around window resizes on APUs.

[ntropy]

It's still quite abstract to me...
But if it compiles complex stuff in real-time during the games (I thought it was the radeon driver package compilation ), I understand why the use of one or the other can lead to differences in performance. Your explanation served its purpose. Thanks for taking the time.

Its quite a complex process and there is way more to it. Here is a nice explanation: https://www.youtube.com/watch?v=qHpKfrkpt4c
Before he shows the rendering pipeline from vertex to fragment shader, he explains the command buffer, so dont get confused there in the beginning.

I have a very basic OpenGL Project here: https://github.com/ntropy83/pandora
If you go to the folder bin/shaders you will find the programm code for the fragment and vertex shader. In my program there are only 2 of those programming snippets, in AAA games you have several thousand. Those little programm snippets are used to calculate the vertices or the mesh of an object then in game and need to be compiled.

Sometimes you see that steam downloads in Linux, Vulkan shader definitions for a game. Those are precompiled shaders, it downloads to minimize stuttering for you in game.

The process is the same in windows, but you have much more precompiled shaders there and often games already ship them. So the stuttering is never an issue there. It will be gone for Linux too eventually, with more people playing and compiling shaders. Some stuttering tho will persist with heavy graphics intense titles, but that is for windows too. And if your GPU is fast enough, you wont even notice.