archived-dots

[rotund token]

it is then async loaded when you open a subscene

so yes, it can take multiple frames to load

[timber ivy]

so a setcomponentdata call on a section entity could then fail if the section isn't fully loaded when its called correct?

[rotund token]

when you call load subscene

it gives you an async handle

you need to wait till that is done before doing anything

[timber ivy]

yeah im in the process of doing that now but issue mentioned it didn't work like that so I just wanted to clarify before I continued

[rotund token]

alternatively you shouldn't even do this at all

you should just be doing this with jobs/queries

when the entity exists

the job will run

if not

it won't run

[timber ivy]

Ah I guess I could make a spawnvehiclesystem that handles this you're right

thank you for the suggestion

[rustic rain]

Are entities that are part of loading subscene present in a world?

or are they loaded only on last frame of loading?

as in, all at the same time?

[rotund token]

They are all loaded at once

[rustic rain]

I wonder if you can load subscene from build in editor

not as authoring, but as just entities

[timber ivy]

So one of my auto converted entities has children entities, is there a way to take the root entity and copy it and it will copy all children as well? cause EntityManager.Instantiate just copies the root entity.

[rotund token]

Hmm I thought it should work

With linked entity

But tbh I avoid hierarchies so I'm not that familiar with it

[timber ivy]

Its not a linked entity according to the inspector

there aren't any linked entity components being added when its converted

[rustic rain]

LinkedEntityGroup buffer

is created during normal conversion for all children in Transform hierarchy

during instantiation of prefab

all entities that belong to that group are copied

[timber ivy]

ill read into it

but in this case I was able to solve my problem with a mesh combine tool xd

[robust scaffold]

Imagine: IBlobComponent...

[robust scaffold]

Well, a blob ref is just a pointer... Reinterpret as a BlobType** ugh. Disgusting. But it works.

Burst does not like generic double pointers sadly.

[viral sonnet]

why would you ever do that?

[robust scaffold]

do what?

[viral sonnet]

reinterpret as blobassetreference<T> is the next best thing

using a pointer pointer for blobs

[robust scaffold]

A BAR is a pointer itself. By reinterpreting as a pointer pointer, you remove Unity's annoying message of requiring use of a ref to access blob data.

[viral sonnet]

hehe, i see. you can also get the ptr to the blob and save just the ptr. double ptr is performance hell

[robust scaffold]

Shape<Circle>** comes from a ICD containing a single BAR of a Shape<Circle>.

[robust scaffold]

The first pointer is the NA component data. The second pointer is the blob.

[viral sonnet]

yeah funky setup

[robust scaffold]

But it works

[viral sonnet]

in that case all good

[robust scaffold]

Burst throws an error if it's a generic double pointer sadly

[viral sonnet]

i didn't mean it like that. just haven't seen nativearrays with it

[robust scaffold]

A NativeArray<BlobAssetReference<BlobType>> => BlobType**.

The redirection of the second pointer breaks vectorization sadly. It is random access.

What it does is reduce memory usage and increase chunk capacity. I've shaved off nearly 88 bytes (22 floats) from every entity.

[viral sonnet]

you moved those bytes to the blob?

[robust scaffold]

yep

[viral sonnet]

yeah blobs are great for that

[robust scaffold]

Hrm. No difference in performance though. The actual chunk iteration jobs are sub 0.1ms and singlethreaded.

[viral sonnet]

oh nice, my blobs are taking quite the hit. but they are muuuch more than just 22 floats. i'd have to check how much size they take

hm, seems i need internal access to find that out

around 350-500 bytes 😅

[robust scaffold]

You storing a book per entity?

[viral sonnet]

good lord, not per entity 😄 it's my spell definition database

and yeah, it's a shit ton of data

[viral sonnet]

anyone remember this? unity said the want to make some cap on chunk capacity or smth

[rotund token]

yes

[viral sonnet]

and still going with 16k chunks? how does this work then when the capacity would be very high? empty chunk space?

[rotund token]

it's 128
and imo it's actually somewhat of a problem

i'm hoping that it's only for archetypes with enable bits... but i feel like this won't be the case

[robust scaffold]

128 components per entity?

[viral sonnet]

i remembered the mention about this while looking at my capacities and I'm just not getting it - why? enabled bits might be the best answer. yeah, i also hope this is just for enabled bits archetypes. i'd hate the idea of such an arbitrary cap

[rotund token]

per chunk

128 entities per chunk

[robust scaffold]

Wait, what? Why

[rotund token]

One related change we did make for 1.0 is putting an artificial cap of 128 entities per chunk, to simplify and optimize the enabled-bits processing code (128 bits = one 16-byte SIMD register). Previously you could have a chunk capacity as high as 2048 under pathological conditions, but in practice almost every archetype we've seen is already comfortably below the new capacity limit, so we don't expect this to result in too many additional chunks.

[robust scaffold]

where is this?

[rotund token]

https://forum.unity.com/threads/new-enabled-disabled-state-filtering.902531/#post-8098883

[robust scaffold]

I guess... I can see the reasoning

[safe lintel]

dont suppose anyone encountered a physics bug when you destroy physics colliders+joints entities, the physics system does like 10x more work(!) and theres an immediate death spiral?

[robust scaffold]

Death spiral is just fixed step. As for destruction, how are you doing so? Just delete entity?

[safe lintel]

yeah just destroy entity via entitymanager.destroy. theres a minor step of removing some system state component but I dont understand this at all

[robust scaffold]

Do you know what job is taking so long? I really only know the BVH builder but I'm 99% sure there's no spiral-able issue in their BVH world reconstruction

Possibly if the destroyed entity's AABB somehow persists but 0'ed out. That'll cause the midpoint split to end up in a neverending loop.

[safe lintel]

it definitely feels like something like that is happening

I swear there are less physics entities in the below profiler than the top 🥲

maybe the only wrinkle is that Im relying on LinkedEntityGroup to destroy the ragdoll rather than doing it manually? I dont ever recall getting this before in previous ragdoll experiments but now that Im thinking about it, Im not sure if I ever actually destoyed entity ragdolls before

[robust scaffold]

Possibly. The create contacts job is the one killing your simulation.

My hypothesis is that when you destroy an entity with a LEG, the target entity is destroyed but the linked children persist until the next frame, or end of frame.

During that time, physics attempts to reconstruct a joint but one of the entities are invalid. And it moves forward with the invalid data, resulting in a broken while loop (maybe)

[safe lintel]

yeah it doesnt last forever but its a good few seconds of terrible performance after the fact before it settles down. definitely gonna need to do some experiments. Im a bit worried it seems the havok guys havent really been active in a while if this is really a bug and not some dumb implementation thing on my part

[robust scaffold]

When you program your own physics engine, all the bug are both your implementation and your fault.

Join me in making it crystal clear who to blame when the physics engine explodes. Roll your own engine.

[safe lintel]

this is specifically why I didnt write my own engine 😟

my code already explodes enough when using off the shelf parts!

[robust scaffold]

Ha, I think I broke live conversion personally. Thankfully pressing the clear cache is simple enough

[viral sonnet]

this is getting nerdy AF 😄

[robust scaffold]

Using ref, disgusting

You can halve the width of that function by using direct pointer access

You can probably make that generic...

[viral sonnet]

open for input. i looked at generics and datatypes yesterday. wasn't that great, so, wdym?

[robust scaffold]

Just looks like repeated code. Replace the <short>, <half>, <float>, etc with a <T> and you can just copy paste the method name instead of the code.

But for this short of actual code, it's on the edge of worth generifying

[viral sonnet]

T doesn't work for simple data types

and yeah, i'm changing the unsafeUtility methods. figured as I'm using byte ptr and byte index it doesn't work anyway

[robust scaffold]

ref *(float*)((byte*)statChangeValuePtr[statChangeByteIndex]) I guess that isnt much better...

[viral sonnet]

changed it to this

[robust scaffold]

That first line, the short value, wont work.

Adding to a pointer changes the target of the pointer, you still need to pull the actual data out

*((short*)Ptr + Index) (dont need the parenthesis around the second.

[rotund token]

(original was better)

[viral sonnet]

oh right, *(short*) should work

[robust scaffold]

Yep, I think you need parenthesis around the Ptr+Idx or else it'll be adding the number index to the short pulled from ptr

[viral sonnet]

yeah, i liked the verbose nature of it but the index gets multiplied by the size of the data type so sadly not an option

[robust scaffold]

Soon, my secret goal of making everyone write c++ will be achieved.

[rotund token]

if another c# dev jumps in, or just enzi in 6 months time what version is going to make sense faster to them

[viral sonnet]

yeah, i'm aware. already have that.

[rotund token]

because i'm already just looking at it and thinking, wtf you doing mate

[viral sonnet]

i looked at some unsafeUtility method but couldn't find one which is not using a sizeof multiplier

and i'll probably inline method this because you are right. this gets hard to figure out and any slight change will make it explode 😄

[robust scaffold]

I just make everything either a float or int. Easier for everyone that way. Value type switches are just more headaches than it's worth unless you generify everything

[viral sonnet]

eh, most programming task are too easy anyway. need some challenge 🤣

and i'm certainly not mad enough to write my own physics engine

that's for sure!

[robust scaffold]

As you said, it's a challenge. And really makes me think in SIMD. Simple mathematics is too slow. Must do 4 at once.

[viral sonnet]

(i've already done that in flash AS3 😩 )

[robust scaffold]

Pointer madness.

[viral sonnet]

yeah physics and simd go well together

having pointers makes c# fun again

[robust scaffold]

Having to write unsafe everywhere is not fun

[viral sonnet]

make the top class/struct unsafe. done 😄

[robust scaffold]

So big.

[viral sonnet]

this whole thing turned out to be quite the nice case for codegen. i have to create so many index lookups. doing that by hand. ooff

yep, that's a big one 😄

[errant hawk]

is this series still relevant with 0.51+ ? https://www.youtube.com/playlist?list=PL-05SQhI5rIYuRbANfxjt7G2C9oFch0al

[rotund token]

probably 90%

[robust scaffold]

And the 10% will make you very confused.

[errant hawk]

Trying to get an idea of the workflow, starting point, etc of dots. I already know how to use burst/jobs quite well so that is not an issue

[robust scaffold]

That's great. Basically, entities are your automatically managed NAs and the IJobs replaced with IJobEntityBatch to get NSlices of the bigger NA.

[robust scaffold]

https://dots-tutorial.moetsi.com/ Since you already have your feet wet, I recommend this up to date tutorial.

[rotund token]

+1

only tutorial i've seen i like

(though just read the manual)

[robust scaffold]

It touches basically all aspects of DOTS ecosystem currently. From the basic entity all the way to UI and VR integration.

[errant hawk]

Alright cool, thankyou. This along with a few other sources is likely what I'll need to get started on porting my multiplayer project to ECS. It's not necessary to use DOTS, but it's going on servers, and reducing the pref cost of each instance running saves money and resources in the long run.

[rotund token]

my strong advice is don't port an existing project to entities

if you want performance on an existing project, just stick with jobs + burst

[errant hawk]

It's getting completely rewritten, but the assets are being reused/modified to work with DOTS

[rotund token]

oh if you're writing from scratch then sure ok go ahead

[viral sonnet]

hm, does it make any difference if I add 2 short or 2 half values as long as they are 16 bit? i'm not sure if i can generalize this on bit level or if i should stick to data types

[robust scaffold]

2 shorts are fine. Integer addition doesnt care about bit count. 2 halves definitely not.

[viral sonnet]

that's what i thought. hm, well okay then 🙂

[robust scaffold]

You must re-expand the halfs to floats before operating any math on it. Floats have 2 components, the exponent and mantissa, and can not be added like an int.

Unless you want to implement half arithmetic. It's just like floats but halve the bitshifts.

[viral sonnet]

i was expecting var statChangeHalfValue = *(half*) changeValuePtr; ref var tmpHalfValue = ref *(half*) valuePtr; tmpHalfValue += statChangeHalfValue; to fail as I can't find the + operator for it

[robust scaffold]

There's an implicit cast to float.

[viral sonnet]

ok then

[robust scaffold]

Check. Implicit casting is sneaky.

[robust scaffold]

God, I forgot how amazing swizzling is.

It's been too long since I had to hammer out any shaders but burst is keeping me sharp

[errant hawk]

quick question regarding the dots workflow: is it better than prior to 0.51? I am not a fan of all the boilerplate or custom tools that need to be made first just to do something simple in earlier versions

[rotund token]

depends how fast you want your code to be

[robust scaffold]

^

It's easier to use with far less boilerplate if you use the code-generation version but you're not gonna see the maximum burst performance using them

[rotund token]

you can still get great performance boosts from burst and easy multi threading from jobs

and will be potentially a lot faster than gamoebjects

but yeah, you won't get max performance without writing it yourself

[robust scaffold]

Rendering honestly is the largest bottleneck. Using the code gen Entities.ForEach or IJobEntity will give you objectively far better performance on a larger scale than GOs.

You just wont get performance to support thousands to tens of thousands simultaneously

[errant hawk]

So what is the alternative to using the Entities.ForEach/IJobEntity that isn't codegen?

[rotund token]

writing the jobs yourself with IJobEntityBatch

that's what all the code gen turns into

[errant hawk]

oh, so just native Jobs/Burst

[robust scaffold]

Basically

[errant hawk]

Im kind of bothered that IConvertGameObjectToEntity exists. DOTS workflow should not incorporate the concept of GameObjects

The only link between the two should be the ability to share data, ie; cache value from GameObject, queue job, await computation, apply results (just like normal jobs currently not using entites)

and really the only use as of right now that GOs serve in a DOTS-based project is for UI/ease of use/setup. If unity can cut down on that reliance, DOTS will be much better overall

[robust scaffold]

GOs are the human facing interface.

They're not evil. I've grown to like them quite a lot as an authoring tool despite going completely DOTS with no GOs during play.

[robust scaffold]

GOs in DOTS are just a skin. They're a launching platform for 100s to 1,000s of entities.

This single monobehavior is 24 different components when authored.

Dont think of GO's as good or bad. They're just a well fleshed out inspector tool in DOTS.

[errant hawk]

I don't hate GOs, I like their ease-of-use and wide range of uses. The bothersome part is that you can take an GameObject, use it as a "template" and create an entity from it. Instead, an entity should be created from some sort of custom entity editor much like the current GO inspector for components etc.

[robust scaffold]

The GO inspector is the custom entity editor.

These GOs wont exist during gameplay. They only exist inside the editor. You ship the entities found in the DOTS Hierarchy.

[rotund token]

So if they invented EntityObjects and they existed in scene and could be prefabs and used the exact same inspector that'd be fine?!

Why spend years re-inventing the wheel

[errant hawk]

No, instead it would be a list/collection of prefab entities you can use that are created using the editor to instantiate during runtime. Since Entities are data-driven, it makes more sense for them to be "prefabs" in a storage sense in the editor project folder. When you need to create a new entity, referencing the entity prefab would be the go-to.

For entites/GameObjects that exist in the scene, Im fine with them technically being GameObjects, but presented to the user they should come with obvious entity restrictions and not this hacky ConvertToEntity script that gets added when the checkbox is selected.

[robust scaffold]

That's called a subscene.

[rotund token]

^

[robust scaffold]

Subscenes are DOTS's prefab and authoring tool all in one. You can nest subscenes. You can edit subscenes. You can selectively activate (load) or deactivate them.

[errant hawk]

DOTS honestly should have it's own editor

[rotund token]

you need to watch the 1.0 talk

[robust scaffold]

https://www.youtube.com/watch?v=p4ct4vHWYt0&list=PLX2vGYjWbI0TMJrYFmLi-inr7DiIkGO4W

Watch that. I've gone from a subscene hater to holy shit I cant live without them

[errant hawk]

unlisted... go figure that is the reason I couldn't find any context about that video

[rotund token]

https://www.youtube.com/playlist?list=PLX2vGYjWbI0TMJrYFmLi-inr7DiIkGO4W

part of their gdc 2022 playlist

[robust scaffold]

I dont know why unity unlists that talk. It's possibly the most solid demonstration of DOTS 1.0 so far.

[rotund token]

but yeah it's weird it's unlisted

[robust scaffold]

It singlehandedly re-kindled my hope in DOTS. Being a salty old vet in these coding trenches.

[errant hawk]

I generally follow development in the forums, so seeing this video is a huge refreshment if it turns out to be what I hope it covers

[errant hawk]

3mins in and they're using a foreach Query which cleans up the workflow quite a lot IMO

[robust scaffold]

That's the preliminary on-main-thread entities foreach using the IAspect<> Interface coming with 1.0.

If they somehow can extend that to scheduled jobs, that will be fantastic but the foreach IAspect on the mainthread is still nice to see.

[rustic rain]

GOs are custom editor

Conversion workflow with gos is great

Subscenes make it so no gos make it to the build

[errant hawk]

Im liking the light humor and the way examples are presented to solving issues. My favorite so far is the ability to manipulate entites and have the changes reflect it edit mode as well (dirty.) I also like that there is now less abstraction between authoring mode, and runtime mode

[errant hawk]

cool, watched it all and Im liking what Im seeing so far

[robust scaffold]

Great. 1.0 is probably months away so now is a good time to get familiar with the basics of DOTS. It wont change much with 1.0.

[rustic rain]

Months away?

They gave any tips about eta?

[robust scaffold]

Im being optimistic for the new guy

probably years

[rustic rain]

Ah yes

That's more like it

[robust scaffold]

(although 2022 LTS is their goal according to the very small blurb at the end of the 2022.2 beta features list)

Got physics working. Only discrete collisions though.

[rustic rain]

Can't wait for jornal

[robust scaffold]

Journal?

[rustic rain]

Ability to track structural changes history

[robust scaffold]

Is that not already in entities? I havent used it before

[rustic rain]

I wasn't able to find anything about using it

[robust scaffold]

https://docs.unity3d.com/Packages/com.unity.entities@0.51/manual/entities-journaling.html

Only through an attached debugger currently though

[rotund token]

yeah you can see a lot of info through debugger now

its great

look at entity field, can look it up in every world

all it's components etc

was such a great addition

[rustic rain]

bruh

I barely have systembase to attach breakpoint xD

hmmm

Does Rider support journal?

[timber ivy]

So I have an entity that I am converting from a gameobject. It has one of unities physic authoring scripts. Physics Shape attached. Later after it is converted to an entity and spawned with entitymanager I add a PhysicsVelocity component to it. The moment I do the entities Translation and LW all get NaN as the values. What am I missing?

[rustic rain]

I suggest you to create entity through authoring and see it yourself in inspector

then create it manually in code

[rotund token]

yes

[timber ivy]

oh good idea

thank you

[rustic rain]

how do you access it?

[rotund token]

a good way to see how to use it is turn on break points on exception

[rustic rain]

I can't find that journal anywhere in debug options

[timber ivy]

hmm so I am not missing any components. This problem appears to be from copying an entity. If I do not copy the entity and just create a new one the problem doesn't persist. Maybe the colliders are being reused and its causing this not sure, but I know how to fix it now at least.

[rotund token]

Maybe the colliders are being reused and its causing this not sure, but I know how to fix it now at least.
collider is definitely re-used

if you want to make something that's static -> dynamic

convert it as a dynamic

and remove the component

it needs to setup a bunch of other stuff

[errant hawk]

they said sometime during the 2022.2 beta

likely this year imo

[rustic rain]

have you ensured your translation/rotation (scale?)are correct?

if you'll have invalid rotation

you'll certainly have NaN errors

[timber ivy]

Yes they are just 0

[rotund token]

what do you mean by debug options?

[rotund token]

journaling records are automatically attached to the stack of any exception

[rustic rain]

0,0,0,0 rotation is invalid

[timber ivy]

its 0 0 0 1

[rustic rain]

what about translation?

try to ensure they are all 0

[rotund token]

there's no window - you have to break point

[timber ivy]

They are

[rustic rain]

yeah, normal breakpoints didn't show

and I had no idea about exceptions

[timber ivy]

Its cause the colliders being copied so the rigiddbody doesn't know how to behave

[rotund token]

but any exception you have from entities should attach journaling info

(turn journaling off and compare)

[rotund token]

you can inspect entities in normal break points

[rustic rain]

oh

[rotund token]

that's not really journaling info, but just another very useful feature

[rustic rain]

well, no

I need journal

[timber ivy]

Is there a helper method cause I am not finding any info how to do this anywhere

[rustic rain]

I have this weird exception triggered on entity from subscene (I assume) which is not meant to be destroyed at all
How can I breakpoint this code with journal?

[rotund token]

you can see it should be debugging the journal info out for you already

[rustic rain]

well, I don't see it

[rotund token]

in your log?

what type of info are you looking for?

[rustic rain]

what destroyed this entity

ok, I figured it out without journal

xD

[rotund token]

mine literally shows this info in the exception

ArgumentException: The entity does not exist. Entity (190:1) was previously destroyed by Beyond.Environment.Corruption.CorruptionSystem.

or if done via ECB

ArgumentException: The entity does not exist. Entity (190:1) was previously destroyed by Unity.Entities.BeginPresentationEntityCommandBufferSystem. This command was requested from Beyond.Environment.Corruption.CorruptionSystem.

this is from journaling

[rustic rain]

yeah and in my case it said ConvertToEntitySystem

which did not really help

[timber ivy]

Does the new input system work in jobs?

[rotund token]

just have an input system

that writes your input to a component

[rustic rain]

yeah, I created generic components for each action

Started<>
Performed<>
Canceled<>

and everytime action is pressed - there's singleton component

with it

[rustic rain]

bruuuuh

how can I solve this problem?

I have trigger events

which sometimes need to be disabled only for query

which is basically based on write group

        private struct JumpInteractionJob : ITriggerEventsJob
        {
            public EntityCommandBuffer buffer;
            [NativeDisableUnsafePtrRestriction] public EntityQuery kek;

            public void Execute(TriggerEvent triggerEvent)
            {
                if (kek.Matches(triggerEvent.EntityA) && kek.Matches(triggerEvent.EntityB))
                {
                    Interact(triggerEvent.EntityA);
                    Interact(triggerEvent.EntityB);
                }
            }

            private void Interact(in Entity e) { buffer.AddComponent<InteractedTag>(e); }
        }

Current implementation doesn't work

because EntityQuery methods aren't allowed in jobs

[viral sonnet]

cast entityquery to entity array and build a hashset

[rustic rain]

hmm

I wonder if I can do opposite

get all filtered entities

so this way I will have very small hashset

[errant hawk]

Im getting memory leaks with a super simple test SystemBase that modifies a cube GameObject in a subscene

It seems it has something to do with subscenes

there's a second memory leak error that points to AsyncLoadSceneOperation.cs

nvm, I forgot the subscene had the ConvertToEntity enabled... 😅 weird how that checkbox exists in the first place.

[rustic rain]

it's ignored afaik

in subscene

[errant hawk]

Ignored as in it shouldn't be used?

[rustic rain]

in subscene - yes

everything inside subscene gets converted

[errant hawk]

yeah it doesnt make much sense to convert the subscene itself

[rustic rain]

oooh, you literally converted subscene

kek

[errant hawk]

internally the subscene is a GameObject, so that's why it's showing up in the first place (and also why it can actually convert a subscene, but not without errors)

[rustic rain]

they use some kind of hooks

to inject subscenes into world

they are not meant to be converted through normal interface

[errant hawk]

Yeah I was messing around last night and left that checked... forgot to uncheck it and I remembered the error I had last night was because of it 😅

is EntityManager usable inside of a burstified job? or at least a normal job? The docs claim you cant, but source code says otherwise

[rustic rain]

no

[whole gyro]

I was thinking about a similar idea the other day. How's it working for you? I assume you are relying heavily on RequireSingletonForUpdate which is why you went the singleton route.

[rustic rain]

yeah, I do rely on singleton for updates

works out fine, since it's literally same thing as

normal calls

from new input system

the only thing that bothers me is that it can be codegened

into a very nice wrapper

which generates all components manually, creates system that registers for all callbacks

and etc

[whole gyro]

Yeah, I guess I'm wondering if its worth splitting input into one component per action like that. I use a single big component for all my input right now. I can definitely see the benefits of the more granular approach, but wasn't sure if it was overkill.

[whole gyro]

codegened from an ActionMap?

[rustic rain]

yeah

you have list of actions

you generate tags for buttons

special components for values

so for example if you want to generate that for moving mouse, which gives Vector2 value

you simply make such component

literally same as OOP new input system

and fully burstable

potentially

it can be even improved more

by creating single system to handle all of them

[whole gyro]

yeah sounds like a nice setup. Wonder if unity will provide a better way to work with input in dots

Do you mean a single system for populating the tags and components, also codegened?

[rustic rain]

I already have single system

which is hand written atm

and looks horrible to manage

it simply resets archetype of input entity

[whole gyro]

yeah same for me, but all editing the same component instead of one per input event

[rustic rain]

so all input components are gone

meanwhile adding components to that entity is done by new input system callbacks

[whole gyro]

By new input system callbacks do you mean the codegened c# interface that looks like OnMyInput(InputAction.CallbackContext context)?

[rustic rain]

yeah

I haven't really dug into how new input system works outside of that

so potentially can be made even faster

[whole gyro]

yeah that's what I'm using too. Thought it has been around for a while so making sure I didn't miss some new thing there

[robust scaffold]

I'm still using the old input system. I have that control system wired from years back in my first unity project and it works even today

[rustic rain]

old input system is not that cool with custom bindings to centralized actions

[robust scaffold]

WASD, the only bindings you ever need

[rustic rain]

what if user wants ESDF?

or uses gamepad

[robust scaffold]

sucks to be them

[viral sonnet]

there's really nothing more that i hate more and that's html and css. say hello to ui toolkit

[robust scaffold]

they're deliberately making it HTML and CSS because that's what game UI designers are familiar with

[viral sonnet]

i can't even make this dropdown and textfield in 2 columns and i made the froggy flex tutorial a few times. ...ugh

webdev hacks are everywhere

flex is at least better thand oldschool css. doesn't help when theui builder is just weird

(I'm fat fingering on mobile again.)

i'm sure if i get around it, it'll be great. much better than the old way of building editor ui

anyone has experience in ui toolkit for editor ui? there's some controls but i don't know how to style them. everything's greyed out

like the textfield. can't make the label smaller

[robust scaffold]

Might be better off asking in #🧰┃ui-toolkit

[viral sonnet]

yeah

found some thread. maybe that will help me

[robust scaffold]

I only played with it to render text and such, not actually build a UI

[rustic rain]

can probably just give it some classes

[viral sonnet]

yep, custom selector .my-text-field > .unity-base-field__label - this UI builder is very unintuitive. i had to find out now that min-width of this label is set to 150px. doesn't really show me. shame, hope this gets better. it also says when hovering over some class. double click to create selector but nothing is happening. you have to type it in 😄

[eager pawn]

is it still RenderMesh we use to draw 2d sprites?

[rustic rain]

nah, 2d is drawn through hybrid go

[eager pawn]

oh okay i was about to add the hybrid v2 thingy and universal RP, is that needed

and ig this would be able to fix my problem with mesh and material being a managed type cuz hybrid go is non-burstable anyway

[rustic rain]

unless you implement your own renderer yeah

[errant hawk]

Best way to get a specific entity? Such as the local player that is in the same chunk of NPCs/other players.

or should the local player(s) be tagged to separate it

[solemn hollow]

var packedEnablerCounts = enablerCounts.Reinterpret<int4>();
                    for (int j = 0; j < enablerCounts.Length; j++)
                    {
                        bool valid = packedEnablerCounts[j] > new int4(0,0,0,0);
                    }```

how could i do something like this? 
i want that bool4 to be turned into a bool "valid" that is true if all evaluations of the bool4 are true.

[errant hawk]

math.all

[solemn hollow]

if you handle that local player seperately in a lot of ways id say it makes sense to have a tag.

[solemn hollow]

thanks! im not very familiar with element wise operations on vectors. that saved me a ton of time

[errant hawk]

yeah np. math has a LOT of cool features that I wish the normal Mathf library had

[viral sonnet]

this starts to look like something

tutorials are all over the place for ui toolkit. some make it harder than it actually is

[errant hawk]

Is my only option to use a job to evaluate the values of the entity(s) before passing it off to the HUD? Example: health value for each local player (split screen)

Entities.ForEach doesn't have an ability to await it like .Complete() for jobs, so parsing data is not possible.

[solemn hollow]

Entities.ForEach returns a jobhandle. you can .Complete() that. It is the same as with jobs

[solemn hollow]

i dont know how other people are getting their data to UI but i use unbursted Systems to write to monobehaviour world. im not using UI Toolkit for runtime yet

[errant hawk]

No it doesn't, the only type that is returned is a ForEachLambdaJobDescription

[solemn hollow]

ah well it only does if you use .Schedule() or ScheduleParallel()

else there is no reason to await since its synchronous with .Run

[errant hawk]

That's what I'm using. Can you provide a snippet of it returning the JobHandle ?

[solemn hollow]

             {
                 
             })
            .Schedule(Dependency);```

Where Dependency is the Systems Dependency jobhandle.

[rotund token]

^ if you don't pass in a handle it's handled for you via code gen

Is you pass in a handle it'll return a handle and you can manage it yourself

[errant hawk]

Ok thanks I think that will work, but I think im going to stick to using jobs instead of using ForEach. Too many restrictions that shouldn't exist atm

[solemn hollow]

yep

var packedEnablerCounts = enablerCounts.AsNativeArray().Reinterpret<int4>();
                    for (int j = 0; j < enablerCounts.Length; j++)
                    {
                        var isValid = math.all(packedEnablerCounts[j] > new int4(0, 0, 0, 0));
                        var didChange = possibilities[j].valid != isValid;
                        
                        possibilities[j] = new Possibility
                        {
                            valid = isValid
                        };

                        if (didChange)
                        {
                            //do some expensive stuff
                        }
                    }```

how should i go about getting rid of that branch at the bottom to make that loop SIMD compatible?

write to a didChangeArray and do an extra loop over that outside of this faster loop?

[errant hawk]

That would be faster yes. The isValid is converted from a 4-wide bool4, so it already breaks SIMD vector intrinsics

[safe lintel]

ok my physics performance woes were due to an extra PhysicsStep(didnt realize had two). two will annihilate performance in certain scenarios(but not all apparently)

[solemn hollow]

so id have to first create an bool4 isValid in a SIMDed Loop and then in another do the reduction to bool isValid? that second loop must be able to use SIMD too somehow right?

[solemn hollow]

thats the physics deathspiral Tertle talks about alot. with standard settings going below 60fps guarantees a bad time

[safe lintel]

imo it should error having more than one PhysicsStep to protect dummies like me, but it only manifested under certain obscure circumstances, enabling a ragdoll one way but not another. 🤷‍♂️

[viral sonnet]

you can only really make the top vectorized. the bottom would require another loop. something that could work, would be to use left shift (math.compress) and tzcnt to count the number that can be skipped. having said that, i don't think this will be worth it much if this runs every frame anyway. just let it branch. the expensive part will probably still be scalar so you can make only ns improvements with the intro

@robust scaffold might have some better ideas

[solemn hollow]

its not a performance bottleneck atm. i just thought the datalayout fits perfectly here to try to learn some easy SIMD. gotta start somewhere

i do have some other places where SIMD should actually do alot for performance but those im not touching right now 😄

[viral sonnet]

try it, i also think skipping is a nice way to utilize some simd

[solemn hollow]

first time i hear about skipping. so give me some time to google it 😄

[robust scaffold]

Huh, depends on what you're trying to do

Early returns and code skips are great but it really depends on what you're trying to do.

[solemn hollow]

i need to message neighborcells of this cell running the code about changes to the possibility array. Changes to possibilities can only occur if in all 4 directions (the int4) the enabler count is positive.

[robust scaffold]

So you want to run // do some expensive stuff only if packedEnablerCounts[j] has values other than 0?

[solemn hollow]

yes AND if possibilities[j].valid was also false before.

[robust scaffold]

Does the expensive stuff depend on the specific value in PEC[j] and return a int4?

[solemn hollow]

no. its something like Get the neighbor cell from Grid and a struct containing the changed PossibilityIndex ( j ) to a dynamic buffer

[robust scaffold]

If the expensive stuff computes only a single float or int, you're gonna be hard pressed to SIMD it

This looks like a simple bit check

Instead of an if (int4 > int4.zero), try instead just a byte and then toggle on bit flags.

Not everything needs to be SIMD with SOA format.

The code blurb I just posted above calculates the maximum distance out of 4 points simultaneously using SIMD. That's a easy situation to SIMD. If you're just using a flag check, you dont need a math.all()

Use a byte and save chunk memory.

[solemn hollow]

hmm sry i dont think i understood it. i just dont know the tools i have to my disposal at all. its so hard to find good learning resources on it

[errant hawk]

Im pretty sure the top still wont work:

    public void Execute () {
        for (int i = 0; i < 100; i++) {
            Unity.Burst.CompilerServices.Loop.ExpectVectorized();

            bool changed = math.all(vals[i] > int4.zero);

            if (changed) results[i] = vals[i];
        }
    }

[viral sonnet]

that's expecting loop unrolling and something very different. automatic loop unrolling will never work with scalar branches

[robust scaffold]

What values do you expect packedEnablerCounts[j] to be?

[robust scaffold]

Expect vectorized is very basic. Look at the burst inspector to tell if it's really vectorized. Look for VP operations.

[errant hawk]

Is there an alternative to the automatic loop unrolling? I've been really wanting more control over how a loop gets vectorzied... but never looked into it

[robust scaffold]

There's manual unrolling. But it is painful and I do not recommend.

[solemn hollow]

that depends on the TileRuleset of the WaveFunctionCollapse. id practically expect not more than 10k

[robust scaffold]

You can probably help burst by instead doing: results[i] = math.select(results[i], vals[i], math.all(vals[i] > int4.zero))

[robust scaffold]

I would imagine a WFC algo can be SIMD'ed since I've seen compute shaders generate them as well. But short of giving me a code snippet, all I can say is that doing so is not for the faint of heart.

[solemn hollow]

as far as i understood it the compute shader versions have to do alot more work due to the parallelization. They calculate every tile every frame. Im doing a version where i do have parallel propagation without the need to update everything

its not good for my heart thats true

[errant hawk]

well, in a performance-critical environment it can matter (especially since I have a library that aims for the most absolute top performance possible using burst.) Got any sources on how to force Burst to acknowledge the manual unrolling?

[robust scaffold]

I dont know the specifics of the algo but it should be possible.

[robust scaffold]

Manual unrolling is called writing the assembly by hand. Reinterpret your NAs to v256 and write out all the VP operations manually.

[robust scaffold]

https://blog.unity.com/games/updated-guide-for-using-neon-intrinsics-in-unity-burst

Let me tell you, I took a shot at doing this by hand for a sweep and clear algo on a tree and, while it worked, manual vectorization means having to manually support all instruction sets.

That article describes manual vectorization for Neon, which is AMD chips. The code wont work on Intel for example. You'll need to write out the intel intrinsics as well. And if you want to support mobile, those as well.

[errant hawk]

Oh screw that im not writing manual assembly. I was hoping it was just some unity limbo you had to go through :/

[robust scaffold]

Work with burst, not against. Vectorization means working with float4s or straight floats.

[solemn hollow]

the "old" naive brutefore way of wfc would benefit alot from SIMD im sure.
the newer more performant way generates lookup datastructures before the actual algo starts. you dont have to (or cant) loop linerally over arrays alot actually.

[robust scaffold]

reading up on WFC, from my initial impression, you can probably calculate the entropy on the four neighboring tiles simultaneously using SIMD then collapse the wavefunction using math.cmin() to select the minimum and left pack / compress to conduct logic using the lowest entropy.

The lowest entropy probably cant be done in SIMD unless you're gonna use SIMD first to obtain lowest neighboring entropy and write back to the source array and a second pass that operates on 4 tiles at once to calculate the resulting tile.

[solemn hollow]

you are mixing the collapse with the propagation step though. for collapse you take the global lowest entropy cell. you never need to calculate the lowest entropy of neighbors

[robust scaffold]

This is just from some random article on the topic. I imagine you know the specifics.

[solemn hollow]

but you are not wrong in that the entropy can be done in parallel. there is actually only one really good article i found on WFC that explains the overlapping model combined with the performance improvements. the other articles simplify quite alot. if you are curious : https://www.gridbugs.org/wave-function-collapse/

[robust scaffold]

Hrm yea. Seems right.

[solemn hollow]

im still intruiged about the math.compress and tzcnt @viral sonnet mentioned

what does the compress do with an int4?

[robust scaffold]

Compress is left pack.

Tzcnt is the standard count low zeros

[solemn hollow]

so it packs every int4 that is 0 to the right of the array?

[robust scaffold]

This is my usage of compress to obtain overlapping AABBs when querying a BVH tree for broadphase physics.

The first compress uses a bool mask of 4 possible leaves containing entities and left packs a compIdx and compDat (name of an arrays of ints) of the two halves of an entity struct.

[solemn hollow]

ah so the bool4 defines which ones are packed left i see

[robust scaffold]

Yep

it's just a really fancy shuffle.

[solemn hollow]

and that enables me to avoid branching because i can just sort everything i want to skip to the end?

[robust scaffold]

Yes. But make sure this is appropriate for the math and logic you want to use.

I'm left packing / sorting in order to obtain all possible overlaps from a quadtree like node array that I then pass into a narrowphase system.

It might be more efficient, depending on your math, to just integrate the mathematics into this system and skip the use of compress

you have the bool4 mask. You can just zero out portions that are not masked.

[solemn hollow]

in the comments it says its useful for building intersections. atm i am using hashsets to build intersections between dynamic buffers in my AI. is there a much more efficient way with compress?

ah nvm. they would need to be in the same order right?

[robust scaffold]

Hashsets obtains uniques. Compress obtains equals. So they're two complete different systems

[solemn hollow]

i learned a lot again. thank you! ill head off to bed for now and ponder over how i will never get to use all those fancy tools 😄

[devout prairie]

is there any difference between doing:
RequireForUpdate(GetEntityQuery(ComponentType.ReadOnly<MyComponent>()));
and:

        {
            All = new ComponentType[] { typeof(MyComponent) }
        }));

[robust scaffold]

nope, no difference.

wait

All = new ComponentType[] { typeof(MyComponent) } translates to RequireForUpdate(GetEntityQuery(typeof(MyComponent)));

All does not imply readonly

you must also ComponentType.ReadOnly<MyComponent>() inside the all array

typeof is read write

[devout prairie]

right yeah

i was wondering why do the longer version really

[robust scaffold]

Any and none querys.

[devout prairie]

unless you need Any or none ofc

yeah

[robust scaffold]

I rarely use the inline ones since I have so many any queries.

[devout prairie]

right yeah

[robust scaffold]

Although I hate to say it but I think my physics is hot garbage, damn. Well that was a great learning session. I'm gonna go back to old monobehaviors.

It works, kinda, it's more efficient than unity's by miles due to SIMD and burst everywhere but it just doesnt have the same stability as box2d and I really dont know how

[devout prairie]

ahh that's painful

no ideas of why you are having stability problems?

[robust scaffold]

I need continual collision detection for substep collisions.

or else it just penetrates into objects and bounces off violently

ugh, it works fine enough. and it's deterministic because it's all in burst (or at least as deterministic as DOTS physics is)

onto other things, i spent enough weeks on this already

[devout prairie]

i guess you could always go back to it with a fresh approach at a later point

[robust scaffold]

Box2D is open source, just way beyond my current ability to replicate in SIMD though

why did unity stop 2d dots. I mean, I know why, money, but i am very sad

[devout prairie]

yeah i do wonder if they'll pick that up later, it seems a no brainer to leverage dots in that area

[errant hawk]

is the order of entities in a chunk deterministic? If it's not, is there any way I can define the order? I need the entity index to be consistent on both the server and client because I do not want to make the operation an O(n^2)

[robust scaffold]

If you do not use any shared components, I think so. Dont quote me on this but nothing in the entity manager uses floats and integer operations are cross platform deterministic.

Even without burst. Integer math is deterministic. The second you add floats into it, not anymore.

[errant hawk]

Not too concerned with determinism of floats, just that the entities on the server and client have the same memory layout so I can do something like: clientPlayer[0].someValue = serverPlayer[0].someValue and not worry about modifying the wrong player/entity/etc

[robust scaffold]

That might be an issue with network latency though. If a structural change occurs and the client doesnt replicate it in time, this will not be valid.

[errant hawk]

I think you're miss-understanding my question

[robust scaffold]

Memory layout is not an issue though. It's 16kb of linearly allocated memory blocks per chunk.

[rotund token]

entities itself is deterministic as far as i'm aware

[robust scaffold]

if there is no structural changes, that section of code will be valid.

[errant hawk]

same with components or nah?

[rotund token]

chunks is entities

but your client might not be the same as your server

you can have different entities on each

and the time things spawn might differ due to network conditions

but if you create things in the same order in different sessions on the server

the chunk layout will be identical

[errant hawk]

alright thanks

[viral sonnet]

aw damn! but yeah, i know what you're talking about. as mentioned i wrote a physics engine in AS3 because we needed a very specific web demo and concave colliding. most was straight forward but then stability issues hit and i spent weeks full of tweaking on it. that's honestly where writing a physics engine stops being fun. so yeah, if you have a fallback model, go for that, for the sake of productivity. but revisit it some time later. i think it's worth it what you started.

[robust scaffold]

Yea. And I think my BVH builder and broad/narrowphase is superior to unity's personally. But beyond that and the actual collision interactions, not so much.

I'm keeping my BVH though. Pinnacle of my coding right now.

[viral sonnet]

that shit takes a lot of patience and time to get right

box2d is the default for unity 2d physics?

[robust scaffold]

Yea

[viral sonnet]

and dots physics is not an option?

[robust scaffold]

3d locked Z axis has terrible top down stability

[viral sonnet]

anyone knows how i can trigger source generatos before i google? 😅 i actually need the source gen to read a json file

[robust scaffold]

Source gens run every time the code recompiles. Which is why having too many gets really chunky

[viral sonnet]

in case the json doesn't work out i need to go in a roundabout way :/

[robust scaffold]

It should work... should

[viral sonnet]

hm, if not, i can just write the code file myself from a template. it's not that complicated, just a bunch of indices that are a pain to write

[viral sonnet]

yep, no need for ISourceGenerator

[timber ivy]

are you able to read values from a referene type stored in ISharedComponentData in a parallel job?

[rotund token]

not in a burst job

but technically yes

i wrote a SharedComponentDataFromIndex<T> that let's you do this

[timber ivy]

hmm, I found like 5 different ways to deal "use" animation curves in dots. All of them are way too memory heavy. So I thought about using ISharedComponentData since most of the entities will be sharing this data

[rotund token]

why are they anymore heavy than a shared component?

if the animation curve is shared between multiple assets

just use a blob to store it

[timber ivy]

interesting ok, that will make sure the memory isn't duplicated ig? ima go read up on this

if I make a blob asset then pass it to components doesn't it just make a copy of the it since its a valuetype?

nvm this is what blobassetreference is ig

[timber ivy]

is it possible to create an authoring component for blobassetreference?

i tried but ig it doesnt' work cause the pointers shift when you enter play mode idk?

[safe lintel]

if you're feeling mildly adventurous @timber ivy you can just copypaste the Animation.Curve blobasset from the Unity.Animation package, it should have conversion in there as well

[timber ivy]

I thought about doing the same from unity.timeline.dots

they have animation curves in there too

that are dots compatible

[safe lintel]

It’s most likely the same one. Wonder what’s going on with that package

[rustic rain]

You can

Blob assets are saved in subscenes

And references are remapped correctly by Unity

During loading

[timber ivy]

Ok yeah it works if you do it in one of the conversion systems

but if you try to do it from an editor script it doesn't seem to work

[rustic rain]

Ahem

Subscenes conversion is done in editor only

[timber ivy]

I litterally copied my code from a custom editor window to a GameObjectConversionSystem and the blob reference worked

wasn't working in my custom editor window

[rustic rain]

I don't really understand what you're doing

[timber ivy]

oh i see what the problem was. Its is because blobassetreference isn't serializable

[rustic rain]

We'll, not really

It is serializable, just like entity

But

It requires remapping

Because otherwise during load it's invalid

They use some kind of memory offset approach for blobs that makes it possible

[hollow jolt]

how do i convert Matrix4x4 to float4x4 ?

vertices[i] = vertexMatrix.MultiplyPoint3x4(v);
normals[i] = vertexMatrix.MultiplyVector(n);

where v and n are Vector3

[rotund token]

                float4x4 f = m;```

[rustic rain]

(float4x4)yourmatrix;

[rotund token]

it has an implicit cast

[hollow jolt]

tried math.mul but there is nothing for float4x4 and float3

[timber ivy]

will entities foreach work on gameobjects and monobehaviors?

[rustic rain]

no, but they work on entities with monobehaviours

[timber ivy]

Entities with monobehaviors?

[rustic rain]

yeah, you can add managed objects

[timber ivy]

oh so I can attach a camera to an entity for example?

[rustic rain]

yeah, that's pretty much how camera conversion works

[timber ivy]

so how do you do this?

[rustic rain]

if you want to go with this approach, I suggest you to make conversion system

dstManager.AddComponentObject(entity, object);

[timber ivy]

oh awesome thanks!

[rustic rain]

keep in mind, it'll keep actual game object around

[timber ivy]

and the downsides are ofc can only be ran without burst and on the main thread

any other?

[hollow jolt]

no overloads

[rustic rain]

what does error say?

[rustic rain]

it's slow, it's managed, it might not work very well if companion link system doesn't catch up

for built in types

like SpriteRenderer

it automatically creates gameobject clone

with same spriterenderer settings

for other monobehaviours

[timber ivy]

"wont catch up" Hmm does this mean I will have trouble making a camera smoothly follow an entity then?

[rustic rain]

I'm not sure if it will work the same every time

[rustic rain]

no it means, that potentially your camera attached to entity might be literally new Camera()

[hollow jolt]

i can however do the following , so i could cast vec3 to flaot4 , any ideas which one is the correct way ( would the matrix be the 1st or the 2nd argument ) ?

[timber ivy]

ah yikes

[rustic rain]

looks like your matrix multiplication is simply incorrect

[hollow jolt]

the first line is what im trying to replicate

or rather im trying to figure out how to write it using Unity.Mathematics

[hollow jolt]

yep

hence the float4

any ideas which one is the correct way ( would the matrix be the 1st or the 2nd argument ) ?

[rustic rain]

look at example

A*B=C

[hollow jolt]

so if u put float4 on the LHS the result would also have a single row

[rustic rain]

well, you can literally see the return type

of your math.mul

after you assigned parameters

[hollow jolt]

both return float4 lol

so im guessing the vector is on the LHS

var result = math.mul( ( float4 ) vec3 , (float4x4) matrix ).xyz ;

gonna try something like that

i wonder would it be worth it

is Matrix4x4 is burst compatible ?

well this is the closest i got
from :

vertices[i] = vertexMatrix.MultiplyPoint3x4(v);
normals[i] = vertexMatrix.MultiplyVector(n).normalized;

to :

vertices[i] = math.mul( ( float4x4 ) vertexMatrix, new float4( v, 0 ) ).xyz;
normals[i] = math.normalize( math.mul( new float4( n, 0 ), ( float4x4 ) vertexMatrix ).xyz );

for the vert , the input is M4x4 x Vec , for the norm the input Vec x M4x4

however the math . mul doesn't not preserve position

[timber ivy]

Does the position of the game object automatically get synced to the position of the entity?

[rustic rain]

yes, it's built in

companion link

system

[timber ivy]

alright cool thank you!

[rustic rain]

hmmm, anyone remembers how to make command buffer for ISystem?

that can be played back multiple times

ah, nvm

        /// <summary>
        /// The <see cref="EntityCommandBuffer"/> can be played back more than once.
        /// </summary>
        /// <remarks>Even though the EntityCommandBuffer can be played back more than once, no commands can be added after the first playback.</remarks>
        MultiPlayback

[rotund token]

ooo i'm super interested what your use case is

i've never really found a good use for it

[rustic rain]

delayed component adding

but I was able to resolve it with other tools

like system groups and ordering

but tbh

I could use ECB in many other systems

which are for no reason managed rn

after all you can schedule jobs in ISystem

btw, I figured an interseting option

[timber ivy]

This isn't happening for me.

i attached a CameraComponentAuthoring script to the main camera in the scene. The authoring component is just autogenerated from this

[GenerateAuthoringComponent]
public struct CameraComponent : IComponentData
{
    public float smoothTime;
    public Entity observable;
    public float3 offset;
    public Vector3 velocity;
}

when the entity is created it looks like the pic below. However when I update the values of the translation component the cameras gameobject doesn't follow

[rustic rain]

I want to make helper method to get ref of actual ComponentData

so instead of doing double em call for GetComponentData and then SetComponentData

I could just do GetComponentDataReference

with ref value as return

[rustic rain]

doesn't look like you have companion link

[timber ivy]

I have been thinking about this as well but wasn't sure it would work

[rustic rain]

I think you need special compiler declaration

[errant hawk]

Im trying to get a subscene's entities in a SystemBase but having trouble finding info on how to do that. Thoughts?

[rustic rain]

HYBRID_ENTITIES_CAMERA_CONVERSION

[timber ivy]

Ah that explains why I can't manually add it

[rustic rain]

add this to Player options

[timber ivy]

awesome thank you!

[rotund token]

btw doing it on main camera causes, oddities

which is why it's disabled

[rustic rain]

that's not really meant to be done, what exactly are you trying to do?

[timber ivy]

can I just remove the main camera tag from the camera or do I need 2 cameras in the scene

[errant hawk]

Trying to get a chunk but it requires an entity

[rustic rain]

ahem

what kind of chunk

for what

[rustic rain]

just try it with main

and see how it works first

maybe you won't need any special treatment at all

[rotund token]

yeah if you remove main camera tag it wont have issues

but i think srp needs one

it still works btw on main camera

you might just get oddities in editor

like camera starting in wrong position when entering the game

as far as i'm aware these are editor only issues

and if you aren't lerping camera you probably won't notice anyway

[errant hawk]

EntityManager.GetChunk(Entity) doesn't specify any other parameters other-than Entity so idk what you mean by "what kind of chunk"

[rustic rain]

well yeah, because it gives you chunk where this exact entity is stored

why do you need it?

[rotund token]

just query for the subscene?

[errant hawk]

trying to get info from a specific entity

specifically for UI purposes

but using GOs

[rustic rain]

well, that's not really supported so you'd need hacky ways to make it work

if it's UI related

why not just store data in GO

this way you can get reference to it with Object.FindObjectOfType<>()

[errant hawk]

you're telling me I can write to an entity, but I cannot read for it to be used elsewhere?

[rustic rain]

when you store entity in subscene

that means you store it in large chunk of serialized data which is not loaded during world creation

so in order to get data from subscene in System you either will have to do it on every frame

or in OnStartRunning

which will require you to implement some algorithm to ensure it only loads it once (assuming it's important)

so that's why it's all hacky

and rather be done differently

[rotund token]

that didn't sound like their problem

[rotund token]

what's stopping you querying this entity?

[rustic rain]

trying to get info from a specific entity
I assume he just stored settings for smth in here

query sounds like an overkill for that kind of purpose

allthough, completely ok

if it's some kind of true ECS UI interface

[errant hawk]

these are networked entites so storing the context of their UI on a GO isn't really ideal. Would rather just read the entity and copy the data to the UI

[rustic rain]

well then

RequireSingletonForUpdate<YourEntityComp>();

and then in OnStartRunning
GetSingleton<YourEntityComp>();

and do all your work in on start running

but that would be called every single time there's such entity added to world when it didn't exist previously

[rotund token]

correct me if i'm wrong, but the problem sounded like extevious wants to get a specific entity and write it's data to the UI

not query a UI entity

[rustic rain]

var ptr = access->EntityComponentStore->GetComponentDataRawRW(entity, typeIndex);
Hmm

can I somehow cast ptr to ref T?

        public static unsafe ref T GetComponentReference<T>(this EntityManager em, in Entity entity)
            where T : struct, IComponentData
        {
            var access = em.GetCheckedEntityDataAccess();
            var typeIndex = TypeManager.GetTypeIndex<T>();

            access->EntityComponentStore->AssertEntityHasComponent(entity, typeIndex);
            access->EntityComponentStore->AssertNotZeroSizedComponent(typeIndex);
            if (!access->IsInExclusiveTransaction)
                access->DependencyManager->CompleteWriteDependency(typeIndex);

            var ptr = access->EntityComponentStore->GetComponentDataRawRW(entity, typeIndex);
           
        }

        [BurstCompatible(GenericTypeArguments = new[] { typeof(BurstCompatibleComponentData) })]
        public static unsafe ref T GetComponentReference<T>(this EntityManager em, in Entity entity)
            where T : unmanaged, IComponentData
        {
            var access = em.GetCheckedEntityDataAccess();
            var typeIndex = TypeManager.GetTypeIndex<T>();

            access->EntityComponentStore->AssertEntityHasComponent(entity, typeIndex);
            access->EntityComponentStore->AssertNotZeroSizedComponent(typeIndex);
            if (!access->IsInExclusiveTransaction)
                access->DependencyManager->CompleteWriteDependency(typeIndex);

            var ptr = access->EntityComponentStore->GetComponentDataRawRW(entity, typeIndex);
            var structure = (T*)ptr;
            return ref *structure;
        }

hmmm

let's see if it'll work

oh well

it's not even crashing

😅

[errant hawk]

you can shorten that last line to: return ref *(T*)ptr;

[rustic rain]

yeah, that's really interesting thing

just gotta remember

that structural changes invalidate it

[solemn hollow]

oof how to debug without burst when burst is actually required to get any kind of performance? if i disable burst i need to wait minutes until i run into the error i wanna debug lol. with burst its seconds...

[rustic rain]

NativeDebugger in VS

haven't ever used it though

[solemn hollow]

hmm any idea if its possible in rider?

[rustic rain]

not that I know of

[rotund token]

I haven't been able to get it to work

I can hook riders native debugger up to native code run by unity but not burst

For some reason

[rustic rain]

how do you hook VS?

[solemn hollow]

:/

[rustic rain]

I can't even attach VS to Unity

hmm

hmmm

[timber ivy]

im using vecctor3.smoothdamp so basically same thing

Quaternion.LookRotation

is there something similar to this in the math/collections library?

[timber ivy]

yeah works fine with main camera for now thanks

[rustic rain]

quaternion.LookRotation

[timber ivy]

Do I have to do something speicial to set the rotation of a rigidbody in unity.physics? in vanilla physics it just works

in unity physics with dots they are fighting each other

[rustic rain]

rigidbody?

there are no rigidbodies in dots

[rotund token]

They call them different things like physics body and rigidtransform

But it's the same thing

[rustic rain]

btw

is it possible to get triggerevents list

manually from physics world?

[robust scaffold]

Wait. Since I uninstalled entities, I can move to 2022.2... IM FREE

[rustic rain]

🤔

uninstalled entities?

herecy

[robust scaffold]

Yea. Went back to trusty pile of NA.

[pastel field]

Do you guys separate Data, Authoring & System in separate files ?
I would have thought this was a good practice, but my code base is growing and I'm starting to find it annoying. I hesitate to put everything in one file.

[robust scaffold]

I'm handing out raw pointers to everything and I want them to still be valid if I structural change

[robust scaffold]

Three folders: Authoring conversion, components, and system files.

[rustic rain]

Files yes, folders no

usually I keep mechanics together

[pastel field]

That's what I'm doing right now, "convention" oop, but I don't see much benefit in it.

[rustic rain]

unless it's some very general components

that are used by multiple mechanics

[robust scaffold]

My components are often a dozen lines or less but I have a lot of them. My authoring is quite large but one off and conglomerates (all the components for entities get added in one authoring system). My systems one per job

[pastel field]

This is perhaps the only thing that motivates me to keep doing it, as some Data structures are very small.

[robust scaffold]

I mean, you structure your project however you like. There is no set rules for this

[rustic rain]

"Just drop everything in /Scripts folder and it's fine"

[robust scaffold]

That's what I do for in progress mechanics. Then I rename scripts to whatever it actually is, add an assembly def file, and create a new scripts folder

[rustic rain]

it's fine

[pastel field]

Speaking of assembly def, Is there a faster way to add dependencies? It's a bit messy at the moment.

[robust scaffold]

Manually adding them one by one until the scripts stop throwing errors about missing namespace is basically the way to do kt

[pastel field]

It's a pain 🙂

[robust scaffold]

Keep your dependencies low and use the least amount of packages as possible and it's less amount of pain

[pastel field]

Especially since sometimes the addiction is not very explicit, I've been known to look for a long time for what the addiction was. The experience could be more "automated".

[robust scaffold]

That's just regular implicit assembly definition (using the scripts folder)

[rustic rain]

https://assetstore.unity.com/packages/tools/utilities/auto-asmdef-156502

[pastel field]

I would like to, but it's Unity that makes it difficult with its duplicates (Unity.Physics,Unity.Physics.Editor, Unity.Physics.Hybrid, Havok.Physics, Havok.Physics.Editor, Havok.Physics.Hybrid,...)

[rustic rain]

really don't understand why Rider doesn't do

[pastel field]

That sounds interesting 🙂

[robust scaffold]

Or you can just buy that asset.

Ugh. Migrating to 2022 is gonna be pain

Or 2023... choices. FREEDOM in versions...

[rustic rain]

one thing Rider does tho

is lets you easily add references through code

but still

so annoying it doesn't add reference automatically

[pastel field]

true

[rustic rain]

and now you have to reimplement everything yourself

must be heaven for you 😅

[pastel field]

The official editor could be better at this

[robust scaffold]

Pointers everywhere. Structural changes off main thread and parallel. It's so liberating

[robust scaffold]

Although I am doing this because I moved back to built in physics 2d. So that's one thing I don't have to reimplement. And a functional sprite renderer

[rustic rain]

wha

hybrid?

herecy

[robust scaffold]

I chopped my entire project back down to my custom lighting engine. Archived my physics. Deleted my Sprite renderer

[robust scaffold]

Not hybrid. Pure GOs.

No entities

[rustic rain]

oh god

[robust scaffold]

Freedom tastes amazing

[rustic rain]

I feel like I'm in cage when using GOs

[robust scaffold]

What's stopping you from just rolling a mountain of native arrays?

Component data replaced with direct pointers to the NA element. Although resizing means invalidating all the pointers and reissuing them to all the GOs which is very painful

[rustic rain]

no chunks

no poggie woggie memory layout

[robust scaffold]

You now have 1 chunk. One giant universal chunk.

[rustic rain]

yeah...

not much help here

for CPU cache

besides

Unity entities worked with pointers a lot too

just under a safe hood

and tons of checks for you

[robust scaffold]

If you iterate in sequence, the cpu knows how to properly manage the cache

You dont need to divide up arrays in 128 sized chunks to get performance. In fact, that reduces performance

[rustic rain]

and how would it work

with fragmentation

[robust scaffold]

Fragmentation?

[rustic rain]

well yeah

if for example

there are 50 blue entities and 50 red entities

and they are all

go through 0-1-0-1

kind of thing in your array

and you only need one color

[robust scaffold]

You can have multiple native arrays

[rustic rain]

so how that would handle multiple archetypes

[robust scaffold]

1 per entity archetype and then string them together. Or use multiple sequential or parallel jobs

[rustic rain]

so array for each archetype

aaand how would that work

if you need multiple data from 2 different arrays

CPU cache will be lost with it

[robust scaffold]

Yes. Or you can merge archetype similar components like one array for all colors and the two separate arrays for different components

[robust scaffold]

The CPU cache doesnt need to be handheld. It's quite intelligent in knowing what information is required so long as you include it. What helps performance is following along its assumptions. That the next data to be iterated on is the best element when incrementing the pointer.

[rustic rain]

yeah, not really

the fastest cache is very small

if you load 1 large memory chunk and then the other

first one you load will be unloaded to less faster cache layer

[robust scaffold]

Yes. Extremely small. 64 bytes. That's 1 struct.

[rustic rain]

so each time you grab your component data

for one iteration per entity

you basically reload cache

[robust scaffold]

My entities? A pile of NAs is just that. Arrays of ints or structs of ints. It's the best cache friendly format

[rustic rain]

if you only work with it

and only it

[robust scaffold]

You can not get any better than NA<int>

[rustic rain]

but if you work multiple data at the same time

where you need data from several arrays

you end losing cache value with every call

to different array

[robust scaffold]

What, is the cache reloading every time two arrays are called? What do you think is happening when two native arrays are added together and written into a third? The classic example of vectorization is pointed to as good cache utilization because its linear access. Despite multiple arrays

[rustic rain]

it is realoding when you need to access 2 parts of memory

and even your arrays will be multiple parts of memory, because cache can't load it all at once

so it all be chunked for it anyway

[robust scaffold]

It can load from two sections of memory simultaneously. Divide up the cache into 2 parts for data from two sections of memory. It's not one continuous block

[rustic rain]

yeah and then you end up loading memory x N times per array used

[robust scaffold]

It can load from parts of an array as well. Load 64 elements from one, 64 elements from a second, and 64 elements from a third and so on

What do you think a chunk is? It's a 16KB section of memory with multiple NAs

[rustic rain]

what I'm saying is that in case of chunks - it loads chunk and boom. YOu can access any data from entitiy - all already loaded in CPU cache

meanwhile in your case

for each component

it'll have to access dfiferent parts of memory

and what if some of it's memory will be in RAM and not L3 or smth?

yeah, it's good when you want to just iterate over 1 array only

[robust scaffold]

So? It's just one initial access to declare that these memory locations will be required for the execution of this job

[rustic rain]

but when it comes to multiple, it seems slower

[rustic rain]

I highly doubt CPU is THAT smart with memory management

[robust scaffold]

This isnt vacuum tube processors. The CPU and memory access has grown to be incredibly smart.

And plus, what is there to be intelligent about? It knows it has it's set cache size. It knows which arrays are required

[rustic rain]

it's still less fast though

[robust scaffold]

It's a simple division operation to calculate the amount of elements it can load into the cache at once from every array

[rustic rain]

in case of chunks you load 1 chunk per lots of entities

let's say per 32

so per every 32 entities it'll load slow memory 1 time

and in case of multiple arrays, it'll have to access slow memory x N component type times

[robust scaffold]

Not all 32 entities will be able to fit though. That cache is shared between all compute applications using whatever memory they want as eell

[rustic rain]

in best case scenario for not less than 32

[rustic rain]

they are all fit

[robust scaffold]

In a completely isolated cpu running only your program, yes chunks can be read in sequence and fit in perfectly. But when other applications are also running, nothing will be in sequence

[rustic rain]

chunks usually contain even more

I mean, sometimes it's literally 139 xD

[robust scaffold]

I'm just saying, this entire discussion can be concluded with a simple performance test. Generate a random number from one value and write that into an output.

[rustic rain]

but it's not about reading chunks in sequence

it's about reading contents of chunk

[robust scaffold]

One would assume reading and writing from one array is better than doing so from a set of linked arrays

That's what a chunk is. A linked array set

[rustic rain]

a bit different:

1. have chunks with 10 different values. All must be multiplief/added/divided and etc between each other.
2. have 10 arrays do the same thing

[robust scaffold]

Okay. Have 10 arrays of size 10,000 vs 10 structs containing 10 arrays of size 100

In fact, I dont need to do that, ArowX has that benchmark on the forums

[rustic rain]

as long as chunk size is ~ size of chunk in entities

[robust scaffold]

Have the arrays be of 4x4 TRS matrices and read and write to a corresponding float3 position value.

[rustic rain]

no no sir, I specifically mean what I mentioned above

this is what I think is worst case scenario for multiple array approach

[robust scaffold]

I'm standing by ArowX's benchmark post as my proof that a size 10,000 NA performs better than an equivalent chunk implementation

[rustic rain]

1 array? or multiple arrays?

[robust scaffold]

Just have 10 position values multiplied against a 4x4 TRS then

[viral sonnet]

you have to ask yourself, is it practical?

isolated tests can be made faster. doesn't mean they are worth anything in practice. like how often do you use a float3 alone? if a single value is important why use float3. if you add/multiply/etc... a bunch of float3s they are as fast

to add something to the discussion. chunks will be slower as they are not linear in memory. but when i say slower. it'll be marginal. like extremely small. a quite good test is reading nativelist/array against nativestream. nativestream also uses small blocks of memory (4k)

[robust scaffold]

The smaller the block, the higher the cost vs linearly arrayed elements. Which is why chunk size should be maximized

[viral sonnet]

totally with you, the chunk size of 16k is not good. with their enabled comp feature they seem to be doubling down though

[viral sonnet]

entities that are rarely added/destroyed and are big should have at least 32/64k

[rustic rain]

I'll do test

here's array approach

where ```cs
private static float4 DoWork(float a0, float2 a1, float3 a2, int a3, int2 a4, int3 a5, float4 a6, int4 a7,
float4 a8, int4 a9)
{
var kek = a0 * a1;
var lul = a2 * a3;
kek /= a4;
lul /= a5;

        var adds = a6 + a7;
        var subs = a8 - a9;

        adds = new float4(adds.x * kek.x, adds.y * lul.x, adds.z * kek.y, adds.w * lul.y) / lul.z;
        adds *= subs;
        return adds;
    }

[viral sonnet]

do yourself the favor and write a performancetest

will save a lot of time

[rustic rain]

now gotta do the same with Entities

[viral sonnet]

    "name": "Tests",
    "rootNamespace": "",
    "references": [
        "UnityEngine.TestRunner",
        "UnityEditor.TestRunner",
        "Unity.Burst",
        "Unity.Jobs",
        "Unity.Jobs.Editor",
        "Unity.Collections",
        "Unity.Collections.BurstCompatibilityGen",
        "Unity.Entities",
        "Unity.Entities.Tests",
        "Unity.Entities.Editor",
        "Unity.Entities.Editor.Tests",
        "Unity.PerformanceTesting",
        "Unity.PerformanceTesting.Editor",
        "Unity.Entities.Hybrid",
        "Unity.Transforms",
        "Unity.Mathematics"
    ],
    "includePlatforms": [
        "Editor"
    ],
    "excludePlatforms": [],
    "allowUnsafeCode": true,
    "overrideReferences": true,
    "precompiledReferences": [
        "nunit.framework.dll"
    ],
    "autoReferenced": false,
    "defineConstraints": [
        "UNITY_INCLUDE_TESTS"
    ],
    "versionDefines": [],
    "noEngineReferences": false
}``` here's an asmdef for it

and some example to get you going

the annoying part of performancetests is setting them up. asmdef, boilerplate code. once done. excellent. saves so much time and timings are accurate.

i'll post a class to test ecs also. courtesy of some forum member with an anime avatar whose name i can't remember

[rustic rain]

let's see

oh wait, forgot to assign values to entities

also ReadOnly attributes

[robust scaffold]

https://forum.unity.com/threads/hybrid-rendering-package-becomes-entities-graphics.1323792/

More prep for 1.0

[viral sonnet]

posted 2 mins ago. nice!

[robust scaffold]

gotta continuously refresh the forum page

[viral sonnet]

F5 dat sheet

prepare for disappointment with your post 😄

[robust scaffold]

Hahahahahaha. I am 99% sure it's gonna be a no.

If it's a yes. I'm uninstalling 2023.1 and going back to entities.

[rustic rain]

what are these errors?

[robust scaffold]

use ref?

[rustic rain]

        public partial struct TestJob : IJobEntity
        {
            public void Execute(ref ResultComp result, in FloatComp a0, in Float2Comp a1, in Float3Comp a2,
                in IntComp a3, in Int2Comp a4,
                in Int3Comp a5, in Float4Comp a6, in Int4Comp a7, in OtherFloat4Comp a8, in OtherInt4Comp a9
            )
            {
                result.value = Test.DoWork(a0.value,
                    a1.value,
                    a2.value,
                    a3.value,
                    a4.value,
                    a5.value,
                    a6.value,
                    a7.value,
                    a8.value,
                    a9.value);
            }
        }

I mean

why?

[robust scaffold]

IJobEntity

[rustic rain]

I want RO access

[robust scaffold]

cursed

IJobEntityBatch

[devout prairie]

ohhh

[rustic rain]

ok let's go

huh

maybe, I should test in build

[robust scaffold]

multiply the number of arrays by 10 as well

the length of them

not the actual number of arrays

[rustic rain]

10k not enough?

[robust scaffold]

100k. You're measuring in the nanoseconds

[rustic rain]

all right

I also need separate project for this

so I can actually test it in build

ok, here's editor

here's build

welp

array wins by 2

@robust scaffold what would be fair IJobFor settings for that kind of array?

I want to test parallel too

job.Schedule(sampleSize, 16).Complete();

array batchSize

sampleSize = 10 mil

Editor

I'll play a bit with values

smaller sampleSize yields same result

huh

I mean

batchSize

for parallel in arrays

let's leave it at 16 then

interesting

pretty much same result

I'd say

chunks win, because in real life you won't have such nice synthetic 10 mil array xD

[robust scaffold]

You say that

[rustic rain]

yeah, but if you'd need to have a query over multiple archetypes

you'll run into a same problem, why chunks lose here

they do way more job, than just iterating over 1 large array

combining multiple arrays

per archetype

and etc

I have a feeling, if pure array equals to perfomance with chunks, that means chunks compensate somehow

for such losses

[robust scaffold]

Chunks compensate with flexability. If you know beforehand the archetype of the entities you're creating and the way you will be using them, hardcoding the array is faster

Not faster in coding time, but faster in performance

[rustic rain]

ofc hardcoding it is faster

but then you will have a very hard time creating mechanics along that hardcoded stuff

I'v been there, kek

this project... this is my 5th full redo

from scratch

kind of how I learn ECS, kek

everytime I recreate it better than it used to be

1. dots ecs, 2. full OOP, 3 another dots attempt, 4 managed dots ecs attempt (works ok btw), 5 current fully unmanaged attempt

so far only 2, 4 and 5 worked out

[pliant pike]

I've done that I have tons of projects where I've started with dots and got better over time