 |
Done via rev 9023 to 9028 |
 |
Err, nope, that was something else. This is still open - it basically means to remove the r_useNewBackend switch entirely and everything that the old backend replicates from the new one. What the mentioned revisions removed was even older code :D |
 |
Wait a few months first. Most of the players have just started using TDM 2.09. As of now, I recall only these common issues: 1) new AMD bug with bindless textures, and 2) old AMD bug with 64-bit FBO. None of which are about the new backend, actually =) |
|
|
The investigation in 0005598 showed that new backend is sometimes a bit slower than the old one: (stgatilov) For me, it is: 190 FPS: default settings 210 FPS: +set r_fboColorBits 32 215 FPS: +set r_tonemap 0 245 FPS: +set r_useNewBackend 0 (cabalistic) As for the new backend, a good portion of the difference at least appears to lie within the depth stage. There were some concessions I made to support the multidraw approach (even when it's disabled) that might be at fault here. In any case, the drop with the new backend definitely needs to be fixed. It is not supposed to be a step backwards. I'll take care of it... somehow :) So perhaps old backend should live a bit longer... Here is an important discussion about the future of the new backend: https://forums.thedarkmod.com/index.php?/topic/20949-on-bindless-textures-multidraws-and-improvements-to-the-new-render-backend/ Basically, Cabalistic says bindless textures and multidrawindirect should be removed. Doing so will probably fix the performance issue... cite: "I sacrificed some scissoring in certain stages to be able to batch more draws - and while that makes no difference on fast cards, it might be hurting slower cards in certain situations." Also it will simplify our code a lot. I'd say removing bindless textures and multidrawindirect are the most actual tasks. |
|
|
Agreed; I'll find some time for it next year :) |
|
|
Closing this since we plan to merge two backends and return to simpler approaches. Like uniforms, and later maybe per-surface, per-object, per-view, per-XXX UBOs . |
- Anonymous
