[Unity] GPU Instancing Tutorial

If you see anything wrong with this video please point it out with examples or references, I'm just starting to get into this section of the Unity documentation.
Timestamps:
0:00 - Intro
0:25 - GPU Comparison Chart
4:35 - Programming
Need critique, motivation, or help? Join our Discord server: / discord
=Credits=
Created by Kyle Andrews
• Website: gamesbykyle.com/
• Artwork: andrewk9.artstation.com/
• Follow me: / @gamesbykyle
=MUSIC=
Song: Culture Code feat. Karra - Make Me Move (James Roche Remix) [NCS Release]
Music provided by NoCopyrightSounds.
Video: • Culture Code feat. Kar...
Download: ncs.io/MakeMeMoveRemix

[@purpl3grape]

As much as the gpu instancing optimization is talked about in the community, it's surprising to find not that much coverage on it. This is great.

[@underpowerjet]

Thank for this tutorial!!! I am going to be trying this out to generate grass on dynamic LOD planet terrain. Recently, I decided to use Light Weight Render Pipeline for the Shager Graph. My custom Gerometry Shaders for generating grass are now completely borken due to LWRP. I am going to stick with LWRP because I hate writing shader code. This looks like it will work for the solution I'm looking for. This is awsome!

[@guir0]

Nice job, to the point!! We need a follow up video ;) Suggestion/Challenge: NavMesh with boids on a plane to see how much it can handle! Cheers!!

[@andreasnull]

Nice tutorial. Some comments regarding the performance: The main problem with your performance is that you are using Linq in the Update method. This is overkill. You actually invoke a Select over a collection which is bad and then ToList which again is not the best for performance. You probably need to create a solution with dictionaries. Also, the ObjData class can be a struct, which I don't think will matter for this example, but if you want in the future to create ObjData during gameplay this is the way to go.

[@ThePsycho211]

Then throw Foreach out of the window and it will be even better.

[@NumianVonKami]

thank you! what i did for my project is to initialize the batch.Select(...) part on the Start() method, as my objects wont change position.

[@personmuc943]

What if i want to use multiple Meshes and Colors? Should i use multiple matrix4x4 lists?

[@danieles6684]

I know this is old as hell, but the built in GPU instancing wasn't working for me, so I wanted to try this. This ended up working a lot better. That being said, the selection via linq of the matrices generates a lot of garbage, so I found my FPS tanking every few frames from garbage collection. Just storing the matrices as the batches directly fix this so I wasn't generating new garbage every frame.

[@YangSunWoo]

The liberal uses of 'fuck' really helped me keep my attention on the tutorial, thanks!

[@Waffle4569]

The GPU performance in editor might be explained by VRAM usage. Once your GPU runs out of VRAM, performance drastically drops.

[@courageouscuber9278]

I did this to render grass in my game. I had to do a few changes to make the grass stick to the terrain, and I had to make the grass disappear when the player gets further away so that the tris count doesn't get too high, but it works and my game is running smoothly! Thank you so much!

[@nukeythewise9424]

Could you please tell me what methods you used?

[@varunkushwah451]

Pls tell us

[@ivanfilatov5482]

How to instance box with colliders? Trees in terrain support colliders, i dont think drawMesh can do the same

[@jessicasan9643]

Can i use in Android device? The objects may have collider, rigidbody and can i use with Unet?

[@BlendcraftCreations]

This should answer your questions: bfy.tw/JdFR

[@Waffle4569]

I've gotten much better performance with 100k cubes in editor (GTX 570) by combining the cubes into one mesh and using a vertex shader. If a combined mesh was then instanced, you'd probably be able to go even higher with little performance impact. Instancing the individual cubes adds up too much, when you could be doing a lot more in a draw call.

[@tokyowarfare6729]

GPU instancing is not here to overcome issues bith static batches BUT with dynamic batching. You should keep doing static batching for your static stuff and consider using GPU agains Dynamic for your dynamic stuff

[@kirienitram]

Hi Kyle. I was wondering why this is more efficiently than ticking of the draw instanced checkbox? I'm new to Unity, but I know a fair amount about graphics programming.

[@BlendcraftCreations]

I haven't touched Unity since I made this video so I'm not sure what's the most efficient way to do anytime anymore. Sorry I couldn't be more helpful.

[@kirienitram]

@@BlendcraftCreations No worries! I didn't look deeper into it since I asked, so I don't know anything more either. Thanks for the answering nonetheless!

[@boulangermarguiertheo8937]

Could we add physical interactions with those cubes ? Similar way to a rigidbody ? Gravity & collisions.

[@comradekenobi8146]

Boulanger Marguier Théo You'd have to write a custom physics system to do gravity -- Unity's in-built system will choke on that many objects.

[@jesper9622]

@@comradekenobi8146 Thats unneccesary, just remove the collider component of the cubes if you just want to do gravity

[@comradekenobi8146]

@@jesper9622 The original comment also mentioned collisions. I meant to say "to do gravity and collisions."

[@jesper9622]

@@comradekenobi8146 Ok hm, with collisions i think that will be hard to beat unitys physics system and make it support meshcolliders, friction, raycasts

[@comradekenobi8146]

Jesper Unity's physics system is more versatile, but there are simpler ways to handle basic collision detection that could run better with that many objects that wouldn't support mesh colliders, raycasts, or friction.

[@caburnicus2285]

I haven't even finished watching the tutorial, but thumbs up already for convenient curse words...how the fuck do we run this thing? Lets find out!

[@WreckedRevival]

I tested this with deferred vs forward rendering for 100k instances.. (cam set to perspective) Forward Render Batches | saved by batching | FPS | Render Thread | Tris | Verts | Set Pass 1202 | 598,800 | 34.5 - 38.9 | 25.9 - 26.8ms | 7.2m | 14.4m | 8 Deferred Batches | saved by batching | FPS | Render Thread | Tris | Verts | Set Pass 1007 | 498,000 | 41.7 - 46.8 | 20.9 - 21.8ms | 6m | 12m | 12

[@BlendcraftCreations]

I've never heard of the deferred vs forward rendering, and after a few quick google searches it's something I think I'll have to experiment with. Thanks for sharing!

[@salemazzam6502]

You should update your visual studio to 2017 and unity to 2018

[@BlendcraftCreations]

I believe unity 2018 was released (just) after I made this. Also I use VS Code instead of studio now, it's much nicer.

[@Kelewann]

I was following your tutorial to familiarise myself with GPU instancing. Linq has pretty poor performance and by directly storing the matrix4x4 in list for the batches, I got a solid 60fps with 100k cubes in editor (I have similar pc specs to yours). Also I do notice a weird thing and have no idea what's ups. If I manually put 10k cubes in the editor scene and then run it, I get 10k "saved by batching" as expected. But when doing it as you do through code I get 50k "saved by batching"... If you try this with linqs you will see it has pretty heavy performance impact, as if it was actually instantiating 50k.

[@BlendcraftCreations]

I really couldn't find a ton of info on GPU Instancing when I was researching this during the week before filming so I made it work with what I knew/had available. If you have any references to how it could be done better I'd love to read/watch through them.

[@Kelewann]

Sorry I don't have any other reference. Your video is actually the best material on the subject that I could find. I just wanted to give feedback on things I found while following your tutorial. I'll let you know if I find anything better.

[@BlendcraftCreations]

I'm looking forward to it. I'm just glad I didn't miss anything major while researching this topic. After I hit the 6th page of Google I tend to lose hope in finding something useful.

[@Gmania999]

Hi, very good result, but i dont understand how to write a script "the matrix4x4 in list for the batches", could you share the script pls)

[@Kelewann]

I'll put up the full project on github soon. Not that it's much. If you take BlendCraft code in the video, I just added a list of Matrix4x4 and when creating the object I directly put the Matrix in this new list. Then instead of getting the matrix from the list of object with a Linq querry, I just give directly the list of matrix. Here is a pastebin copy-paste: pastebin.com/r63PiSa1

[@simonfarre4907]

Also seems like a unity problem even in build. 10000 cubes should not be that taxing to draw instanced on the GPU if it was handcoded. Which is a strong argument for building your own engine.

[@MattRix]

It's not a Unity problem, he's just not using it correctly.

[@RiversJ]

Old video, just about everything that could go wrong with the implementation went wrong. Other comments contain his reply and pointers by others how to do it right. Do it right with indirect version and you can pump out a million of those cubes even on the hardware of 3 years back.

[@tokyowarfare6729]

having higher FPS in editor than build makes no sense. I've never seen this behaviour.

[@BlendcraftCreations]

At what point was the FPS higher in the editor than the build?

[@tokyowarfare6729]

@@BlendcraftCreations 2:55

[@colonthree]

4:35

[@User33423]

Meshes that have a low number of vertices can’t be processed efficiently using GPU instancing because the GPU can’t distribute the work in a way that fully uses the GPU’s resources. This processing inefficiency can have a detrimental effect on performance. The threshold at which inefficiencies begin depends on the GPU, but as a general rule, don’t use GPU instancing for meshes that have fewer than 256 vertices. If you want to render a mesh with a low number of vertices many times, best practice is to create a single buffer that contains all the mesh information and use that to draw the meshes.

[@YAS-vm8ko]

8:31

[@chromosome24]

My dog food keeps producing dog shit.

[@shloh532]

Hi :D

[@BlendcraftCreations]

Hello

[@Gasimo]

I WANT 16BG RAM

[@BlendcraftCreations]

downloadmoreram.com

[@Hello-qg4yk]

@@BlendcraftCreations hey!

[@scenes_dot]

using System; using System.Collections; using System.Collections.Generic; using UnityEngine; using System.Linq; public class Spawner : MonoBehaviour { public class ObjData { public Vector3 pos; public Quaternion rot; public Vector3 scale; public Matrix4x4 matrix { get { return Matrix4x4.TRS(pos, rot, scale); } } public ObjData(Vector3 _pos, Vector3 _scale, Quaternion _rot) { pos = _pos; scale = _scale; rot = _rot; } } public int instanceNum; public Vector3 maxPos; public Mesh objMesh; public Material objMat; private List batches = new List(); void Start() { int batchIndexNum = 0; List currBatch = new List(); for (int i = 0; i < instanceNum; i++) { AddObj(currBatch, i); batchIndexNum++; if (batchIndexNum >= 1000) { batches.Add(currBatch); currBatch = BuildNewBatch(); batchIndexNum = 0; } } } void AddObj(List _currBatch, int _i) { Vector3 position = new Vector3(UnityEngine.Random.Range(-maxPos.x, maxPos.x), UnityEngine.Random.Range(-maxPos.y, maxPos.y), UnityEngine.Random.Range(-maxPos.z, maxPos.z) ); _currBatch.Add(new ObjData(position, new Vector3(2, 2, 2), Quaternion.identity)); } private List BuildNewBatch() { return new List(); } void Update() { RenderBatch(); } private void RenderBatch() { foreach (var batch in batches) { Graphics.DrawMeshInstanced(objMesh, 0, objMat, batch.Select((a) => a.matrix).ToList()); } } }

[@BlendcraftCreations]

This is why god made pastebin.com/ use that and try again.