A feature request: Auto-configured software rendering threads

[willkuer]

To fix that i made the pr.
Its still questionable if it increases or decreases usability...

@dogen
Can you please make a small benchmark. Fps(uncapped) vs ERT? This could be really helpful.
Please try to use games that are not EE limited in sw mode but it might be hard to find. Every piece of information would be helpful.

[dogen]

To fix that i made the pr.
Its still questionable if it increases or decreases usability...

@dogen
Can you please make a small benchmark. Fps(uncapped) vs ERT? This could be really helpful.
Please try to use games that are not EE limited in sw mode but it might be hard to find. Every piece of information would be helpful.

Sure, and I don't think that many games are EE limited in SW mode actually.

[gregory]

Well if you can post all cases. Limited by ee, limited by gs ...

[dogen]

Alright, I'll edit this post with results. I'm assuming the latest build is fine to test with.

Wow, is it normal for baldur's gate to run faster in software mode than hardware?

Armored Core 2
0->167-175%
1->146-150%
2->128-134%
3->139-142%

Baldur's Gate: Dark Alliance
0->51-52%
1->57-59%
2->60-61%
3->57-58%

Burnout 3
0->47-48%
1->48-50%
2->58-60%
3->58-60%

Dragon Quest VIII
0->71-72%
1->64-66%
2->68-70%
3->63-65%

Final Fantasy XII Demo
0->174-178%
1->168-178%
2->140-144%
3->155-159%

Jak and Daxter
0->74-78%
1->65-69%
2->66-67%
3->68-69%

Jak II
0->49-50%
1->46-47%
2->49-51%
3->51-53%

Metal Gear Solid 2
0->46-49%
1->46-49%
2->61-64%
3->60-67%

So it seems like in most games 2 or 3 threads are as good or better than 1 or 0, unless the game is really light already.

Of course, just because I have a dual core doesn't mean every dual core will perform the same. Haswell added an extra alu per core, so even ivy and sandy bridge CPUs might not be able to handle as many threads. And I have no idea how AMDs handle software mode.

[ssakash]

That seems pretty similar to the tests I've done on my core2duo processor. I only gained speed from 1+ software rendering threads on few games like GOW , Fatal frame , Bakugan: Battle brawlers whereas most of the others only saw a drop in most of the instances.

[dogen]

In case anyone missed my last update.

[willkuer]

Very good. It is probably really as expected. Synchronization overhead seems really a factor for high fps I would not have guessed that.

We asked for haswell dual cores as we said others will hardly perform sufficiently well to even consider sw rendering. Maybe thats not true for the lighter games though.

So we agree on those results that 2/3 is a better solution than zero?



Maybe one should look into the synchronization. Maybe there is a better way to wait for finishing.

[Blyss Sarania]

I agree it's better than the default. Even if it causes lower performance on a handful of weak chips, that's better than the lower performance on almost all modern chips that 0 is causing currently.

[ssakash]

I'd agree with this if it actually uses detection of processor physical cores to determine the default value. if it's going to be set globally for all processors then 0 would be the best choice in my opinion.

[Blyss Sarania]

if it's going to be set globally for all processors then 0 would be the best choice in my opinion.

Why?

If your answer is "Because some people have dual cores or single cores" then that's not good reasoning. You could say the same thing about why 2 is the best choice "Because many people have quad cores or higher"