[The E]

FSO is not bottlenecked by GPU performance in most scenarios

*cough* Integrated Graphics *cough*

Developing with Intel hardware as a target has never and never will be our goal. FSO will always be targeted at real GPUs first and foremost.

[MatthTheGeek]

FSO is not bottlenecked by GPU performance in most scenarios

*cough* Integrated Graphics *cough*

...aren't most of the maths in an integrated graphics card basically made by the CPU anyway ? So we're back to square 1 : CPU performance bottleneck anyway .

[The E]

That's not entirely true for the latest chipsets to come out of Intel, the graphics cores in the i5 and i7 range are pretty capable by themselves. Still, since they're nowhere near as optimized for performance as the corresponding nvidia and AMD offerings, and since the average gamer will have a dedicated GPU, making concessions to intel chipsets is not something that we need to do.

Yes, I know it sucks that you intel people can't use the full range of options available in FSO, but what are we supposed to do about it?

[Nuke]

yea just tick back your detail settings and dont install every effects package available.
hell, just drop in a $50 nvidia or ati card and you will see an improvement.

but frankly, modders should lod their ****, especially their big ****. ive been lazy with my stuff and theirs no excuse. also with lods, detail boxes and well chosen detail distances, you can cram in more detail than would have been possible otherwise. a lot of polies far away comes at the expense of a lot of polies close up.

[jr2]

So, the bottleneck is pretty much all collision detection at this point?  Any way to re-do that part of the code, or is it massively complicated or integrated in such a way that re-doing that would mean re-doing much of the engine?

[The E]

Collision detection is one part of the engine that could be improved. There are a few others as well, and there are areas where we are using suboptimal ways of transferring data from the CPU to the GPU.

BTW, one potentially rather important area that could stand some improvement is the lua integration code, as it seems, it is really easy to bleed performance there.

[General Battuta]

So, the bottleneck is pretty much all collision detection at this point?  Any way to re-do that part of the code, or is it massively complicated or integrated in such a way that re-doing that would mean re-doing much of the engine?

I don't want to come down too hard on you, but if it hasn't been done already, what do you think the answer is? If you're thinking 'not impossible, but a big effort', you're headed in the right direction.

And please don't make some ****ing crack about Valathil. You belittle the great work he's done as well as the great work done by the rest of the SCP coders by suggesting that what he does is easy.

[jr2]

  I know what he's done isn't easy.  He's just pretty damn good at it is all.  Unless I was to the point where I could actually read and understand the code (not even to the point of being able to contribute) I really wouldn't (so therefore, I can't) be able to tell what the deal is with the collision code.  Maybe it was just such a tangled mess that it can't be fixed, and has to be re-done, but they are waiting for some improvement in another area of code that is relate -- heck, I don't know.  Hence why I asked.

[General Battuta]

Swifty has performed some minor collision optimizations already.

[The E]

The problem with the collision code, in a nutshell, is that it was written for retail-level ship and weapon numbers and model detail. In those circumstances, it performs really well. But, as we have been busy adding stuff, making more and more elaborate missions and models, the approach used there just isn't adequate anymore. Sorting out which object collided with what, and where exactly the collision occured, is getting more and more expensive.

Tweaking the algorithms, or switching them out, is definitely possible. But, as this impacts a core functionality of the engine, it is not something where we want to risk breaking it.

Realize that collision detection is a hard problem, and that there are no silver bullets to instantly make it faster. In addition, I cannot stress enough that collision detection is one issue where we have performance problems, but it is not the only one.

[jr2]

GPUs are not well suited for that type of logic, right?  Cause I was just thinking, if there was a way to off-load collision detection on to the GPU...

* jr2 retreats

Stop hitting me with teh broken plasma rifles; it was just a suggestion!

Seriously, though, Folding@Home & SETI @ Home and IIRC Prime95 benefit from GPU acceleration (as do some password-cracking routines), so, why can't the GPU be told to handle it?  I'm sure there's a good answer, I'm just too dense to already be aware of it.

[The E]

GPUs are good at one thing, and one thing only. That one thing being massively parallel vector operations on large datasets. Collision detection, as implemented in FSO, requires an asston of binary split partition tree walking, which is somehting GPUs are very definitely not good at.

Not to mention that the interface between the CPU and the GPU is rather slow (because it involves generating something that very closely resembles a texture on the CPU, transferring it to the GPU memory, doing your operations, and then transferring it back to main memory and reading the results. Which is slow as ****.

[General Battuta]

Well The_E, maybe we could convert our collision detection into vector operations on large datasets! I mean, objects in motion have vectors, right, so we could use the GPU to handle collision detection if we just recoded the

[jr2]

  I have done it!

I have taken over de WORLD!

tHE GREAT gENERAL bATTUTA HAS fallen TO MY eeeevil INFLUENCE!

BWAHAHA!  BWAHA, BWAHAHAHAHA!!!!




Oh.  He wasn't serious?

O i c This changes things.. a bit.. -backs away slowly-

[assasing123]

wasn't this topic about a big ship...?

[jr2]

Topic?  What topic?..

         SQUIRREL!!!!

[Droid803]

This thread is about big ships.
Big ships and many ships.
And many big ships.
And just many ships in general.

[headdie]

wasn't this topic about a big ship...?

this is natural HLP topic progression, if you are lucky it will still be about ships tomorrow, if not it will either be something like how tasty zebra is to eat or locked

[JGZinv]

^_^ BIG SHIPS ^_^

[Legate Damar]

What's with the paint job on that one?