Trying Out The New OpenGL Threaded Dispatch In Mesa 17.1

[geearf]

But when you are talking about averages you are talking about benchmarking I guess, while I'm talking about realtime numbers while playing the game, or I understood wrong?

What's the difference between those 2? Unless you've looked at every frame's time for the whole time you played...

[cj.wijtmans]

It's probably conflicting with games trying to do the same thing themselves. I guess it's a good thing the way it's supposed to work is with a whitelist of games that benefit from it and not even try to use it on games that aren't on that whitelist.

This.

Game devs know opengl is single threaded so they make the engine work in the the exact same way, one thread as a command buffer the rest of the game logic into other threads. So the difference should be none or even performance worse because you have two command buffers instead of one. Or am i misunderstanding how this is supposed to work now?

[gbil]

What's the difference between those 2? Unless you've looked at every frame's time for the whole time you played...

No but the fps counter in both cases was pretty much steady and if as you wrote the variation was bigger I should have seen in many occasions the fps counter going way down or way up.

When I have again some time I'll use the gallium hud which instead of a single value it presents fps with a graph therefore possibly any up and downs will be way more visible.

[cj.wijtmans]

Vulkan is by design. RADV isn't as of yet. Testing DOTA 2 I get twice the frame rate if I limit it to two cores as opposed to letting it use all 16 logical cores.

It makes much more sense to spawn a worker thread only when your existing worker threads reach 100%. Not only will it allow Turbo to kick in if two worker threads are enough but its also more efficient. And also allows cores to be completely available for other things running. In contrast to server logic where you want requests scattered over your cores as much as possible.

[LinuxID10T]

It makes much more sense to spawn a worker thread only when your existing worker threads reach 100%. Not only will it allow Turbo to kick in if two worker threads are enough but its also more efficient. And also allows cores to be completely available for other things running. In contrast to server logic where you want requests scattered over your cores as much as possible.

My results were without turbo enabled, just a 3.9 GHz all core overclock. There is definitely some other things at play there. I'd like to see if the scaling is the same on my i3, but that will require further testing.

[mitch074]

So, I did some testing on my old Athlon X4, DD3-1333 8 Gb dual channel, RadeonHD 7770 1 Gb. I ran 4 tests, all FullHD low details no v-sync :
* quad core, 1600MHz, GL dispatch off
* quad core, 1600MHz, GL dispatch on
* dual core, 3250MHz, GL dispatch off
* dual core, 3250MHz, GL dispatch on
The segment is a replay of the game's beginning sequence up until DeWitt sees himself in the basin. A lot of it is autoplay, and those segments I look at most.
One can see that the max theoretical performance should be the same in dual and quad core modes as I double frequency when dividing core counts (+some extra because I can't double RAM nor bus speed). The results: max FPS is dual core, 3.2 GHz, GL dispatch off - but some scenes top up a CPU core and fps takes a good dip. Dual core 3.2 GL dispatch on is the best mode: top fps is lower, but average is actually higher and there is much less variation in frame rate. 4-core 1.6 GHz with GL dispatch average performance isn't great (framerate looks choppy) but 4-core 1.6 GHz without GL dispatch is actually jarring in places.
In all cases, enabling GL dispatch causes CPU use to be MUCH MUCH tighter across all cores: most visible in 4-core modes where CPU usage is usually within a 10% bracket across all 4 cores when GL dispatch disabled causes huge disparities (I mean one core with 80% and another with 15%), it can be felt even in 2-core mode when they usually are within 20% of each other - if that. In all cases, GL dispatch provides smoother gameplay.