Page 1 of 1

[posila] [0.17.46] Major performance regression compared to .45

Posted: Fri Jun 07, 2019 3:24 pm
by eradicator
What?

On .45 i can run an empty map (just as an example) at 6000UPS/76FPS just fine. On .46 everything above speed = 2 has heavy micro (more like macro...) stutters, the UPS/FPS display fluctuates +/-90% and moving feels like 5FPS.

From the stats it looks like render suddenly takes 10ms instead of 1ms.


45.png
45.png (2.58 MiB) Viewed 3766 times
46.png
46.png (3.18 MiB) Viewed 3766 times

Re: [0.17.46] Major performance regression compared to .45

Posted: Fri Jun 07, 2019 4:54 pm
by posila
Hello darkness my old friend ...

I have 3 big questions for you:

- Do you still have flickering issues?

- Does the performance issue persist if you turn on v-sync?

- If not, does it return back to 0.17.45 performance if you edit config.ini and uncomment and set dxgi-presentation-mode=copy?

Re: [0.17.46] Major performance regression compared to .45

Posted: Fri Jun 07, 2019 7:25 pm
by eradicator
(All tested on 0.17.47)
posila wrote:
Fri Jun 07, 2019 4:54 pm
Hello darkness my old friend ...
yami.png
yami.png (15.22 KiB) Viewed 3671 times
posila wrote:
Fri Jun 07, 2019 4:54 pm
- Do you still have flickering issues?
It's difficult to tell now that flickerglitch4 has several seconds of input lag even on speed=1. I was about to say "maybe not"....
posila wrote:
Fri Jun 07, 2019 4:54 pm
- Does the performance issue persist if you turn on v-sync?
But then i turned Vsync on. Performance went up. And a huge fullscreen atlas jumped right into my face.
posila wrote:
Fri Jun 07, 2019 4:54 pm
- If not, does it return back to 0.17.45 performance if you edit config.ini and uncomment and set dxgi-presentation-mode=copy?
From having done faar too many trials now, dxgi=copy has almost the same performance as before (from the logs it looks like this was default before?). Not editing the config and turning Vsync on also has normal performance. From a "gut feeling-comparison" (yea..bad measurement) i'd say either option is about 95% performance of .45.

Also after even more trials i can say that Vsync definetly has *some* effect on the flicker. Flicker can start both with vsync on and off. When the flicker occurs then *toggling* vsync (i.e. vsync = not vsync) makes the flicker disappear, toggling it back makes it reappear...most of the time.

Btw i noticed that with dxgi=copy the log says "discard", not sure if that is correct.
(Log: D3D Feature Level: 11.0, DXGI 1.2+, SwapChain: 1,discard,-,DNW*,-,none)

Also also...i want on the backer names list for all this work :p.

Re: [0.17.46] Major performance regression compared to .45

Posted: Fri Jun 07, 2019 7:47 pm
by Rseding91
I'm still of the opinion that we should just ignore any issues when people are using intel integrated GPUs. None of this seems worth the effort with how many problems there are around the GPUs.

Re: [0.17.46] Major performance regression compared to .45

Posted: Sat Jun 08, 2019 6:55 am
by eradicator
Rseding91 wrote:
Fri Jun 07, 2019 7:47 pm
I'm still of the opinion that we should just ignore any issues when people are using intel integrated GPUs. None of this seems worth the effort with how many problems there are around the GPUs.
I just report stuff i find, hopefully being helpful. If whoever-gets-to-decide-these-things decides to ignore garbage IGPs (wink wink) please do tell me early... trying to reliably reproduce this magic shit isn't free for me either. And if i had to decide between this and pre-composed multi-layer icons...

Also because someone linked it earlier: https://en.wikipedia.org/wiki/Escalation_of_commitment

Re: [0.17.46] Major performance regression compared to .45

Posted: Sat Jun 08, 2019 7:13 am
by posila
Well, what has come out of this is that there is not silver bullet setting for presentation model, and it just needs to be available in graphics settings, so people can change it if they have problems with the default one.

Re: [0.17.46] Major performance regression compared to .45

Posted: Sat Jun 08, 2019 7:32 pm
by gamestefan
Same issue here, I made a little screen recoding for you including the debug options enabled. When I am on zoomed in zoom levels in a forest I get a stuttering effect. I noticed that in the middle section behind "Sprites ... Calls: G xxx T xxx" the values for G and T behind calls can spike up to 3700 normal looks like G ~70 T ~120. Not sure what that means, but I guess that's a lot ;)
Even though I get the stuttering and these large values my FPS/USP still remains at 60/60.

I am including the log file for the specs and graphics settings.

Re: [0.17.46] Major performance regression compared to .45

Posted: Sun Jun 09, 2019 9:29 pm
by posila
Thanks for the reports.

Special thanks for the video. It made me immediately realize what did I do wrong and that we're flooding GPU with commands to draw frames due to CPU being able to queue the commands faster than GPU process them (in your case).
Fixed for 0.17.48