search

mrdoob / three.js

JavaScript 3D Library.
https://threejs.org/
MIT License
100.37k stars 35.2k forks source link

Raycaster performance drops dramatically #16153

Closed visus100 closed 5 years ago

visus100 commented 5 years ago

Raycaster performance dropdown dramatically in realise 103 but in previous lower works great (checked r102, r99, r98).

There is also strange dependence if mouse hovering on Lee perry mesh fps drops down between 20-30, but if hover mesh and click left mouse button to change camera view (orbit controlls) performance stabile 60fps with raycasting.

I also noticed that when i close chrome console i get 15 fps more, so maybe its not much dramatically but it's not good.

Three.js version
Browser
OS
Usnul commented 5 years ago

I get 60 FPS on your example. There has been a lot of discussion on Raycast performance, I have advocated heavily for a spatial index, which enables logarithmic complexity of raycasting, versus the current linear. Here's an example, assuming a single operation takes 100 nanoseconds, cast times:

method polycount time
spatial index 100 460ns
current 100 10000ns
spatial index 10,000 920ns
current 10,000 1000000ns (1ms)
spatial index 1,000,000 1381ns
current 1,000,000 10000000ns (100ms)

If performance of Raycaster is important for you - I recommend using an external solution with a spatial index.

EDIT:

There used to be an implementation of Octree in three.js, but it was removed. Here's the relevant commit: https://github.com/mrdoob/three.js/commit/7aa204e349d997f957b5f3cf1d7a8c62f3bd6068 You can revert that locally to recover the code if you need it.

My own implementation is closed source, if you really want it - we can talk about the licensing, but you'd probably be served well enough by a plethora of JS spatial index implementations here on github that are totally free.

Here are some of my older posts on the subject, reading through might yield some useful information for you: https://github.com/mrdoob/three.js/issues/13909 https://github.com/mrdoob/three.js/issues/13807 https://github.com/mrdoob/three.js/issues/5571 https://github.com/mrdoob/three.js/issues/12857

arpu commented 5 years ago

@Usnul do you know any implementation of raycaster with spatial index for THREE,js?

visus100 commented 5 years ago

@Usnul What computer specyfication did u test it? Are u try to open console? Are u made the example full screen (for more render space because with smaller render space i also reach 60FPS)?

I tried this on my laptop spec: GTX 960m Intel i7 2.6 GHz 16GB RAM

Above comparing test performance images with R103 vs R102

fps r103

fps r102

Please compare with 102 <script src="https://cdnjs.cloudflare.com/ajax/libs/three.js/102/three.min.js"></script>

Usnul commented 5 years ago

@visus100

my specs are: GTX 1080 i7 4.2GHz 16G Ram (dual channel, 2.6Ghz)

The performance bottleneck is CPU+RAM, mostly the CPU i reckon, having a beefy cache on the CPU helps too.

I did have console open - yes.

@arpu I have edited my answer. Hope that helps.

visus100 commented 5 years ago

@Usnul

You have around 2x or even 3x faster PC Spec then i tested this example (It could be reason u gain 60FPS). So i think we should wait till someone will check this with older pc like mine.

Benchmark comparing Graphic cards

Mugen87 commented 5 years ago

I'm unable to reproduce the issue with your codepen. Performance is stable at 60 FPS. Tested with an 4 GHz Intel Core i7.

BTW: The GPU is not relevant in this issue. Raycasting happens on the CPU.

titansoftime commented 5 years ago

Hmm. getting a solid 60fps on my work pc (Intel Core i3-3240 and a 12 year old cheap video card).

visus100 commented 5 years ago

@titansoftime

Did u also tried to check with open console? Because probably it needs more overload to get sight... And for now i can't check this on other PC's.

However if +1 will get solid 60fps maybe it's something wrong on my side... Then suggest to close...

Mugen87 commented 5 years ago

Yes, I've tested with opened dev tools.

visus100 commented 5 years ago

Ok now i know what is new in r103. Raycaster works on morph targets (in r102 and lower not working after morphTargetInfluences. it's pointing on starting geometry) so math was changed too (that could explain hardware requirements), but still i don't understand why i have this situation: "but if hover mesh and click left mouse button to change camera view (orbit controlls) performance stabile 60fps with raycasting."

Mugen87 commented 5 years ago

Closing, this for now since it seems the issue is specific to your machine. If you can actually proof a bug in the three.js library, you can reopen the issue. For now, please use the forum or stackoverflow for more feedback.

mrdoob commented 5 years ago

@Usnul Wouldn't be opposed if someone wants to bring back the Octree example. It just seemed that no one was willing to maintain it (reimplement it using BufferGeometry).

takahirox commented 5 years ago

I'm glad, too. projectObject() in renderer seems one of the cpu hot spot functions because frustum culling is slow. I hope Octree can speed it up.

Usnul commented 5 years ago

@Usnul Wouldn't be opposed if someone wants to bring back the Octree example. It just seemed that no one was willing to maintain it (reimplement it using BufferGeometry).

Personally, I do not prefer Octree as a general-purpose solution, I lean more towards BVH, and that's what I use primarily myself too.

My mention of the Octree was not meant as a criticism of it going away. I don't mind very much, in fact - I think it's probably a good thing, as it's a fairly major feature and tacking it on the side is very misleading for a lot of people, I think.

In the past, I did advocate for introducing a spatial index (for example Octree) into the core of rendering system for various reasons (see related issues for full discussion), but there is not enough interest to see that become a part of three.js proper, so from my perspective - the discussion is well resolved on this topic: no spatial index in three.js for now.

Current state of things is fine for me, I hope this clarifies my position a bit more. Again - I'm just providing information, not really pushing for anything here :)

mrdoob commented 5 years ago

Personally, I do not prefer Octree as a general-purpose solution, I lean more towards BVH, and that's what I use primarily myself too.

What about introducing an example using BVH then? :)

arpu commented 5 years ago

i think https://github.com/gkjohnson/three-mesh-bvh could be used for an example

Usnul commented 5 years ago

@mrdoob Maybe. The code that I do have is a pretty serious spatial index system that I have been working on for several years now, and it would be too big for an "example" I feel, too many parts.

@arpu yeah, that's a nice example