Quick question about performance

Zeozen

Couldn't find this anywhere so please link me there if this question is already answered something.

What gives best performance in a real-time situation (video game) between
a) a high triangle count model with vertex-splits but no texture, and
b) a very lowpoly model with few vertex splits but a high resolution texture?

In both cases I'm assuming a fairly simple pixel and vertex shading situation.

I would think that the first option at least saves more space on disk.

Thanks!:)

Edit: I did more snooping around and found the answer, sorry for my earlier lack of detective skills v.v

musashidan

It's generally good practice to log your self-solution results for others in future who stumble upon your thread with a similar question.

ZacD

There isn't really a solid answer, but I have some information to add.

GPUs render in 2x2 squares, if there's 2 triangles in a square, it has to render that square an extra time. If there's three, 3 times, etc. UE4 has a view mode called quad overdraw that shows this. Generally, if you have a lot of triangles, it's not an issue if most of them are larger than the 2x2 pixels. 

http://blog.selfshadow.com/2012/11/12/counting-quads/

This is why artists are told to avoid long skinny tris. 

So I guess, if your tri density is causing a lot of quad overdraw, it's an issue. 

Memory is also a different matter. It's a case of, use as much available memory as you want, without going over the limit. So at the most basic level, textures are "free" until you run out of memory. 

Zeozen

musashidan said:

It's generally good practice to log your self-solution results for others in future who stumble upon your thread with a similar question.

Yes you're completely right, my bad. I got most of the info I needed from this: http://polycount.com/discussion/50588/technical-talk-faq-game-art-optimisation-do-polygon-counts-really-matter/p1

A summary of what I needed: There's a range between 1-300 or something tris that has no difference in cost.
Tris are in general very cheap nowadays, textures are a bigger bottleneck. Bottom line is of course to use as much is needed to make it look how it's supposed to.

ZacD said:

Memory is also a different matter. It's a case of, use as much available memory as you want, without going over the limit. So at the most basic level, textures are "free" until you run out of memory. 

That's one way of looking at it, however another way of looking at it is what you save one place you can use another, so there's never anything for free.

But that's a very interesting read ZacD, thanks! For the project I'm currently working on, props will generally be a bit away from the camera, so that's definitely something to think about.

RN

There's some information from game engines:

- http://docs.cryengine.com/display/SDKDOC2/Getting+Started+Modeling#GettingStartedModeling-TexturingandUVWs
- http://wiki.unity3d.com/index.php?title=General_Performance_Tips
- https://docs.unity3d.com/Manual/OptimizingGraphicsPerformance.html
- https://docs.unrealengine.com/latest/INT/Engine/Performance/Scalability/ScalabilityAndYou/index.html

Zeozen

Those are good reads RN! Can't believe I missed that one from ue4. An especially interesting tidbit was this:

"Mix different components in similar ratios (some hardware might be only slow in some feature) bad: many objects simple material, few objects many lights with complex materials. "

I don't really get this, fewer lights and simpler materials seem to me like they would give the most performance anyway..

RN

It's worded in a confusing way. The way it talks about ratio and hardware that may be slow with only one feature, I think they mean "use the same amount of simple-material objects and complex-material objects", as that's the safest choice when you don't know the size and speed of the memory, CPU and GPU.

Zeozen

Sure, but what I don't get is that some hardware can do complex stuff better than simple stuff. I guess simple isn't always simpler?