Need optimization Ideas

[brittle harness]

Hello there so ... I was following the strawman videos to make my own version fo the strawman project sample.

https://github.com/Touhma/ECS-Strawman

Everything seems to run pretty fine. but I'm wondering what's next on the optimisation side ? Like everything work well at low entities count ( 5k ) but as soon as the count go 100x that it's struggling.

What can be done to improve performances at that level ?

From what i'm seeing the issue is not coming from the Systems code that are bursted but more on the rendering side.

I'm using in this example ECS : 1.0.16

So i'm wondering :

What's next ?
What is going on ?
What can I do sot reduce the rendering time issue ? ( if the issue actually come from here )

Any resources will be appreciated 🙂

[visual geyser]

hierarchy profiler is useless for multithreaded code

use timeline and figure out what's the candidate for optimizations

[brittle harness]

Sure i've done that but that where i'm not sure what is causing that much overhead

A lot of idling

[visual geyser]

looks like you are gpu bottlenecked?

not sure

check your GPU load

[brittle harness]

2080TI graphic cards , it should be able to show a few hundred thousands basic meshes x)

[visual geyser]

how do you render stuff?

[brittle harness]

Very Basic Mesh too

[visual geyser]

but how do you render it

[brittle harness]

Basic Baker with a prefab reference
With a basic Spawner Script :

 [UpdateInGroup(typeof(InitializationSystemGroup))]
    public partial struct SpawnSystem : ISystem
    {
        public void OnCreate(ref SystemState state)
            => state.RequireForUpdate<ConfigComponent>();

        public void OnUpdate(ref SystemState state)
        {
            ConfigComponent config = SystemAPI.GetSingleton<ConfigComponent>();

            NativeArray<Entity> instances = state.EntityManager.Instantiate
                (config.Prefab, config.SpawnCount, Allocator.Temp);

            Random rand = new(config.RandomSeed);

            foreach (Entity entity in instances)
            {
                RefRW<LocalTransform> transform = SystemAPI.GetComponentRW<LocalTransform>(entity);
                RefRW<DancerComponent> dancer = SystemAPI.GetComponentRW<DancerComponent>(entity);
                RefRW<WalkerComponent> walker = SystemAPI.GetComponentRW<WalkerComponent>(entity);

                transform.ValueRW = LocalTransform.FromPositionRotation(rand.NextOnDisk() * config.SpawnRadius, rand.NextYRotation());

                dancer.ValueRW = DancerComponent.Random(rand.NextUInt());
                walker.ValueRW = WalkerComponent.Random(rand.NextUInt());
            }

            state.Enabled = false;
        }
    }

• Very simple prefab

The project use URP , Nothing weird done on it

Basic setup

[visual geyser]

so you just instantiate baked prefab?

[brittle harness]

Yup

It's a "Very Basic" example 🙂

No gpu instancing , no nothing.

I'm just taking a basic mesh , a basic UPR material, putting that in a prefab, then instanciating it a few thousand time after the prefab is baked.

[visual geyser]

If you look away with camera, is it still bad?

[quasi vault]

100x 5k?

[brittle harness]

500 000k @quasi vault

[quasi vault]

Yeah...

These are probably all setup as dynamics

[brittle harness]

in the batch rendered you mean ?

[quasi vault]

As in you haven't marked them static

Well that's definitely true as you're changing transform

[brittle harness]

"haven't named them static" -> Not sure to follow you here

[quasi vault]

Marked sorry

Typing on phone

[visual geyser]

I think half a mil meshes is too much even for 2080ti

[quasi vault]

Anyway just trying to point you're trying to push a gigabit of transform data to the gpu each second

[visual geyser]

bus capped?

[brittle harness]

Lowl

[quasi vault]

My quick 5am maths says they shouldn't be capped

But it's hard to know what is stalling exactly

[brittle harness]

let me check if that's really what's happening

Just turned off the system updating the transform and i'll mark them as static 😄

[visual geyser]

so bandwidth is 616gb/s

[quasi vault]

Where are you getting that?

[visual geyser]

some tech site

[quasi vault]

That's the bandwidth of gpu reading gpu memory

Not the bandwidth of pci4

[visual geyser]

ah

[quasi vault]

Which is 32gbs

On a 16x

Hence they should be fine throughput wise it's still very high

But that doesn't mean it's not causing latent

[visual geyser]

all right

[brittle harness]

Just so you know in the baker of the prefab : TransformUsageFlags.None -> Don't seem to change stuff

[visual geyser]

you need to click static

on game object

transform usage flags are additive

mesh renderer baker adds Dynamic

your None adds nothing

in the end you have Dynamic - thus nothing changes

[quasi vault]

Hmm thought mesh renderer adds renderable?
Anyway all prefabs are forced dynamic and something else probably does it anyway. Need to hit that static.

[brittle harness]

Technically saving 3ms

[quasi vault]

Gpu still high I assume?

(this 1 hand texting not going well)

[brittle harness]

Yup

100% used that's definitely a gpu bottleneck

lol

How the hell do people show 2millions cubes on the screen without burning their cards ?

[visual geyser]

afaik, BRG only batches in small ones

[quasi vault]

Just for curiosity, I don't think it has an effect in editor but can you go to project settings and turn off graphics jobs

[visual geyser]

like 100 per batch

so for such massive instancing you might want custom solution

[quasi vault]

Who shows 2 million cubes?

[brittle harness]

welp Nothing change indeed 🙂

[quasi vault]

Wait

You have 761 million tris

A normal game might have 5-20 mill

[brittle harness]

wait

that don't add up ...

the mesh is 684 triangles

sooooo time 500k should be half that

[quasi vault]

Yeah then a light?

But even half that is really high

So yes 2 million cubes would be 11x less tris 😅

[brittle harness]

yeah but not 11x more performances

[quasi vault]

Wouldn't expect it to be since its a lot more batches, just pointing out the comparison you're making

CPU time is halved though

[brittle harness]

Yep 🙂

But the gpu still getting raped

[quasi vault]

Side note, 2 million objects is the hard limit in unreal engine for nanite

As it's not expected you're getting close to this

[brittle harness]

So I guess at that point it's just me trying to do stupid things 😄

What's weird IMO tho is why is the CPU halfed but not the GPU time

I mean I disabled Vsync so obviously the gpu would run full power all the time

[quasi vault]

Render thread != gpu time

[brittle harness]

Just the calcs on the CPU to send to the GPU ?

[quasi vault]

Render thread is a dedicated thread on CPU for doing gpu work

[brittle harness]

Make sense so nothing to see here

[quasi vault]

Check this gpu line on profiler instead

[brittle harness]

yeah with the

• strawman mesh : 48ms on "other" & 22.25 ms on shadows
• Cube Mesh : 21ms & 7.4ms

Definitely have a huuuge impact

With a Quad Mesh : 4.65 ms & 1ms

I would expect to have a bigger fps gain tho