Alpha test: CPU bottleneck reductions

[cabalistic]

Hey everyone,

the last couple of weeks I've been working on an overhaul to TDM's renderer that slightly improves CPU utilization in the game and should thus improve performance in CPU-bottlenecked scenes. This is the next step in our ongoing backend modernization and targeted for 2.09. However, since the changes are quite involved, I'd love to get a headstart on testing, so that we can catch any major problems or regressions early

The 2.08 beta test has absolute priority. But if you happen to be curious and have a bit of extra time to spare, I'd appreciate any feedback you can provide.

Current test build: https://ci.appveyor.com/api/buildjobs/q614b4ojrmat3umc/artifacts/TheDarkMod.7z
You'll need a 2.08 beta installation. Make a copy of your TheDarkModx64.exe and glprogs folder (if present) so that you can go back to 2.08, then extract the files from the archive to your installation.

Instructions for testing:

There are two parts to this. First is a new render backend which significantly reduces CPU draw call overhead by better ordering of draw calls and using modern OpenGL features (if available). You can enable it with cvar r_useNewBackend. Most important thing to test here is that there are no visual artifacts or regressions from the new renderer. So ideally go through some scenes and switch between old and new backend and let me know if you notice any differences
For good measure, you can also try to toggle the cvars r_useMultiDrawIndirect and r_useBindlessTextures, which toggle two particular modern OpenGL features that the new backend can use. Ideally, toggling these should not result in any visual changes, either.

The second part is further parallelization of a particularly compute-heavy step in the renderer. @duzenko spent some time to port BFG threading code to TDM, and this initial work is already in 2.08. However, it wasn't quite stable, so enabling it in 2.08 causes crashes eventually. I spent some more time stabilizing this portion, and combined with the new backend, it should yield some FPS increases in CPU-heavy scenes. To enable this feature, enable the cvar r_useParallelAddModels (caution: disable this when going back to 2.08, or you'll have crashes!). Here, I'm mostly looking for confirmation that it runs without crashes. You can also tweak the level of parallelization with the cvar jobs_numThreads. It defaults to 2, if you have 6 or more cores, I'd advise to increase it to 4. Further increments don't seem to yield much improvement.

Thanks for your time! Oh, and just to give an idea of why I did this:

This is performance with 2.08:

And this is with r_useNewBackend and r_useParallelAddModels enabled:

[HMart]

Holly shit that bump in performance is impressive! Good job!

[Anderson]

+1

[lowenz]

+50% 

Edited June 18, 2020 by lowenz

[cabalistic]

Note: this is obviously a notorious scene, and you need a decently powerful GPU to see improvements from CPU time reduction

[lowenz]

Game crashes walking there (A New Job)

[cabalistic]

Thanks, I can reproduce it. Some kind of memory corruption.

[cabalistic]

Annoying bug. There's a subtle dependency of the order in which certain objects get deleted because they can reference each other, but it's so implicit that you basically can't see it in the code.

I hope I fixed it, but I can't be 100% certain. Would appreciate it if you could test this executable, and let me know if you do find more crashes: https://ci.appveyor.com/api/buildjobs/q614b4ojrmat3umc/artifacts/TheDarkMod.7z

[chakkman]

This is great news for people like me who suffer from a slow and outdated CPU.  Always had performance issues with TDM.

Thanks a lot for your work.

[nbohr1more]

I couldn't find the appveyor for linux so I downloaded the source from here:

https://github.com/fholger/thedarkmod/tree/frontend_parallelization

and built my own executable.

Thus far no crashes but also no significant performance improvements to report. Briarwood just seems to cap at about 30FPS for my system.

I should do the profile thing...

Edit:

OK I seem to be mostly GPU bottlenecked in this scene.

A little more testing and 2.08 beta has drops to 24FPS whereas this build with all options enabled never goes below 30FPS

and is frequently at 35FPS.

[stgatilov]

If you want to detect visual differences between two sets of cvars, the best thing would be to automatically switch them back and forth every second.
Well, if the switching works reliable enough