Official Hybrid Renderer V2 (0.4.0)

Is there also a possibility of a help class / method to create your own prefabs / archetype based on in the future? Like:

Code (CSharp):

1. RenderFactory.CreateArchetype(params Type[] customComponent);
2. RenderFactory.CreatePrefab(params IDataComponent[] customComponentData);

That might be helpful if the game have procedural content, so we would not have to constantly adapt your own code and would always be up to date.

 

is there any ETA for when lightmaps are going to be supported by the hybrid renderer v2?
also hybrid render v2 is really laggy

 

If you create good abstractions for conversion and keep them public, this won't really be an issue. Or at least it will minimize the pain. I've hit some built in conversion that was easy to work with, others where it was way too monolithic, it seems to be hit and miss. That and there is no good pattern for overriding built in conversion. Often you have to wholesale replace the conversion system, and then hope that the author didn't make the abstractions it does use (if any) internal.

 

I thought I was wrong regarding my thread from yesterday and HDRP performance, but turns out that was half right. Problem is, it's hard to tell if this is a bug. @SebastianAaltonen , there are some screencaps attached to demonstrate the issue.

A procedurally-generated quad is set as the mesh in a RenderMesh entity component, and (later) that entity is mass instantiated. The instances are put in another chunk (different components). With 2500 instances (plus the original), SRP batcher shows 2 quads (for the two chunks), so all is well. At 5000, I'd expect 2 quads. I see 20 vertices and 30 indices (five quads). At 6500, I see 24 vertices and 36 indices (6 quads). At 7000, the editor crashes with an error that the OS restarted the GPU (too much uploaded).

There are two issues:

1. The number of quads is unexpected. That's explained by your prefab advice. BUT!
2. I wouldn't expect a high end GPU to hang with fewer than a dozen quads, but it does here.

So, the question: Is this related to procedurally generating rather than using a prefab as you describe, or is this a bug I should report? Just want to be sure before I add to anyone's workload.

2500 Instances (Frame Debugger and Entity Debugger): https://gyazo.com/bd6b2ceb2e8682cf593da7193d673cb8
5000 Instances (Same debuggers) https://gyazo.com/cc1aa95b0c0ea079dfacca61c9944646
6500 Instances (Likewise) https://gyazo.com/027f902770ce296fd6d71166c8a66f16

Tomorrow I'll try this starting with a prefab, but in the meantime, I'm still learning. So, I'm a bit shy about calling something a bug without confirmation.

Finally, suggestion:

If issues are avoided just by using a prefab, why not have a method that just generates a prefab and puts it through the necessary paces behind the scenes, to automate all the same? It could take a mesh and material as input, and return a reference to an entity with the necessary components. It's a simple fix to use a prefab, but for developers who come along later, it could save some headache to have it fixed behind the scenes. Heck, we could write that ourselves, but it seems too simple while solving problems too complex for it to not be done already.

 

In an early version of ECS with buildin-renderer, i even displayed 300,000 instances of a Mesh on the screen at the same time at over 30fps. So 7000 should not be an problem.

 

Curious how the dev's think V2 culling would stack up against a partition based gpu culling scheme? Like Vegetation Studio uses for example.

The one drawback of having large partitions like that for our use case is vegetation is dynamic. Rebuilding a partition is a noticeable framerate hit, and worst case we can have those hit once every few seconds. So a more granular system would be appealing. But it's a trade off, we can't take a significant constant perf hit or it's not worth it.

 

If they're instantiated in batches of 2500, I can beat that.

https://gyazo.com/e213f1fc43376187dc67535d5deb50d3


The renderer checks for changes, and only uploads changed data to the GPU. That would explain it and be the end of it, except that I don't have to batch transform and rotation changes like that. And just going by the idea that too much is being sent in one draw call (still odd with so few quads), it should do the same thing when the quads are moved or rotated. On top of that, if I understand batch rendering, it shouldn't matter how many entities are being rendered (GPU side). Though it may slow things down CPU side (preparing the batch).

Mind, that's just 500k + 1 quads all the same color, all pasted to the camera like a worldspace UI canvas. Then 500k of them are rotated behind the camera. AND in an actual UI, there would never be a need to do that. But this is still only 1 million tris. Plus, the 500k behind the camera should be culled. A whole bag of concern, but I'm trying to focus in on what seems to be closer to the root of it all.

 

Hi @SebastianAaltonen! Your previous post here stated that we should wait until Hybrid Renderer 0.4.2. That version was just recently released, but according to the change log, only the dependencies were updated. Was that really the version you were talking about? Or did the expected version numbers change?
https://docs.unity3d.com/Packages/com.unity.rendering.hybrid@0.4/changelog/CHANGELOG.html

 

I'm also confused about the state of hybrid renderer but it seems to be working.
I have latest entities 0.9, URP 9 and HR 0.4.2. - no errors and I can build - so I think it's fine.

 

using HR 0.4.2, turning the V2 flag on, everything is still rendered on 0/0/0 ...

 

Just some random results I did testing on a large scene if anyone was interested (all safety checks off)



Machine is a bit wonky, 3900x but a 1600 GTX, usually pretty GPU capped.

Not really sure why tris/verts are so much higher in HRv2 but performance is good

 

in editor but all safeties debugging off
~8800 render meshes. it's a small low poly city simulation.
no lods or any optimizations.
was really just a curious test of an asset i had on hand.

should definitely do it in a build, but i was mostly just doing some quick testing.

 

-edit-

correction, trying to figure out why default was slower than hdrp, turns out for some reason default scene had a second directional light. results updated

-endedit-

Same scene with HDRP because I was curious



original to compare





Empty project except scene with a bunch of renderers.
Everything using defaults.
No optimization.
All safety turned off.
Results from editor not runtime so take it with a grain of salt

 

I thought this was already in version 0.4.2

The problem i have is still the old one: My "terrain" is completly black and my camera goes black the moment my terrain goes into camera view. I need to enable "Stop Nans" on camera to see anything, but them the terrain is just black and everything else is displayed differently. If i dont create the terrain, then everything is displayed correctly so it looks like the error is inside this shader. I attached a screenshot of the shader, maybe you can help me here

PS: And it looks like Amplify shader doesnt work with the new V2 renderer (still everything is displayed on 0/0/0), but this is not your problem i guess.

 

I also had the issue that procedurally generated RenderMeshes where black, until I manually added all the additional components that Hybrid Renderer V2 MeshRendererConversion.cs would automatically add to Prefabs. Do your black entities have those new components?

• WorldToLocal
  
• PerInstanceCullingTag
  
• BuiltinMaterialPropertyUnity_RenderingLayer
  
• BuiltinMaterialPropertyUnity_WorldTransformParams
  
• BuiltinMaterialPropertyUnity_LightData

There are more, conditionally setup and initialized. See MeshRendererConversion.cs for details.

Also, I think the automatic re-importing of ShaderGraph upon selecting Hybrid Renderer V2 is a fix in the ShaderGraph package (> 9.0.0 preview14?), and not in the Hybrid Renderer package.

 

I create my prefabs "correctly" according to 0.4.2, but only when enabling V2 via flag it gets dark. So what am i missing for components?

 

I just wanted to voice my hope for using URP Lit Sprites with the hybrid renderer as well. Is adding the ability to render Lit 2D Sprite Entities to an existing URP 2D project something on the roadmap for hybrid rendering?

 

Here is the code how i create the planes:

Code (csharp):

1. var archePlane = EntityManager.CreateArchetype(typeof(Translation),
2.     typeof(Rotation), typeof(LocalToWorld), typeof(RenderBounds), typeof(PerInstanceCullingTag));
3.  
4. //filling material dictionary
5. var mesh = Utils.CreateMesh(planeDivisions, planeSize, 80);
6.  
7. for (int i = -(gidSize / 2); i <= (gidSize / 2); i++)
8. {
9.     for (int ii = -(gidSize / 2); ii <= (gidSize / 2); ii++)
10.     {
11.         var id = new int2(i, ii);
12.         var rMesh = Utils.CreateRenderMesh(mesh, new Material(WorldManager.Instance.PlaneMaterial));
13.  
14.         rMesh.material.SetFloat(Constants.CurrentMeshscaleId, meshScale);
15.         rMesh.material.SetFloat(Constants.NewMeshscaleId, meshScale);
16.         _renderMeshDictionary.Add(id, rMesh);
17.  
18.         var entity = EntityManager.CreateEntity(archePlane);
19.  
20.         EntityManager.SetName(entity, "Plane");
21.         EntityManager.AddSharedComponentData(entity, _renderMeshDictionary[id]);
22.  
23.  
24.         EntityManager.SetComponentData(entity, new RenderBounds() {Value = rMesh.mesh.bounds.ToAABB()});
25.         EntityManager.AddComponentData(entity, new BuiltinMaterialPropertyUnity_RenderingLayer
26.         {
27.             Value = new uint4((uint) _renderMeshDictionary[id].layer, 0, 0, 0)
28.         });
29.  
30.         EntityManager.SetComponentData(entity, new Rotation());
31.  
32.         EntityManager.AddComponentData(entity, new BuiltinMaterialPropertyUnity_WorldTransformParams
33.         {
34.             Value = new float4(0, 0, 0, 1)
35.         });
36.  
37.         EntityManager.SetComponentData(entity,
38.             new Translation {Value = IdMath.GetPlanePositionFromId(id, planeSize)});
39.     }
40. }
41.  

EDIT: i found that when i disable the light in SCENE! tab i can see my terrain, but this doesnt work ingame when i disable the light

 

You are not setting BuiltinMaterialPropertyUnity_LightData and WorldToLocal

 

This is for URP, iam using HDRP so i thought i dont need it

 

Code (CSharp):

1. if ENABLE_HYBRID_RENDERER_V2 && UNITY_2020_1_OR_NEWER && (HDRP_9_0_0_OR_NEWER || URP_9_0_0_OR_NEWER)
2.             // Hybrid V2 uploads WorldToLocal from C# to rendering. Need to refresh it.
3.             // TODO: Do this on GPU side in the copy compute shader. Would be free in BW bound shader!
4.             dstEntityManager.AddComponent(entity, ComponentType.ReadWrite<WorldToLocal>());

 

1239332

No project folder but thorough repo steps as it was produced from an empty project using only an asset store asset (which I'm sure you can acquire.)

 

I had it, but i forgot to turn the sun intensity up it works, thank you

 

Since you guys seem to be moving full steam ahead with this idea that we should always start from converting a prefab then add what we need, it would be really helpful if we had a way to copy all components from one entity to another, or to merge entities basically.

As it is now we if we have an entity created at runtime that already has a bunch of components and we wanted to make it render, then we have to manually copy the relevant components one by one to or from the instantiated prefab. I don't know if it's possible while still being performant/without using reflection but a CopyAllComponents(Entity a, Entity b) method would be nice.

 

An entity command buffer can be used to instantiate entities within a query, so we can do that.

Code (CSharp):

1. EndSimulationEntityCommandBufferSystem endSimECBSys
2.      = World.GetOrCreateSystem<EndSimulationEntityCommandBufferSystem>();
3.  
4. var comBuff = endSimECBSys.CreateCommandBuffer().ToConcurrent();
5.  
6. Entities
7.      .WithAny<YourComponentTag>()
8.      .ForEach((int entityInQueryIndex, Entity entity) =>
9.      {
10.           comBuff.Instantiate(entityInQueryIndex, entity);
11.      }).ScheduleParallel();
12.  
13. endSimECBSys.AddJobHandleForProducer(this.Dependency);
14.  
15.  

Lack of typos not guaranteed, but this shows the gist of it. Declare your command buffer system to set a barrier. That determines when the buffer will play back. Your options are beginning of initialization (phase 1), end of initialization, beginning of simulation (phase 2), end of simulation, and beginning of presentation (phase 3). Retrieve that system through World. Then, declare the buffer. Leave off the concurrent and ToConcurrent if you want it on the main thread. Isntantiate clones your entity, at the job index set by (auto-incrementing) Int entityInQueryIndex. Select your entity with your query coming into the ForEach. Schedule parallel, schedule, or run (compiler should prompt you for necessary additional methods). Finally, tell the producer about the job, to make sure it's done.

If you think you're going to free up memory in the job too early, call Complete() on it to have the main thread wait right then. You can also use this to defer scheduling to a point of your choosing that frame (at least, within that system).

I'm a bit sleepy, but this should give you plenty to go by searching the documentation.

 

That's not the problem. The problem is this: Copying arbitrary component data between 2 entities

 

I'm also seeing issues with specific shaders rendering at origin. Is this supposed to be fixed or still incoming?

 

Hmm well the shader say's not SRP batcher compatible, so that's probably it.

Edit: Take that back shader say's it is batcher compatible.

 

Instantiating the same entity prefab over and over results in batch counts climbing up consistently. Is that expected?

 

So I've been evaluating V2 for handling vegetation. Overall very impressive. Especially the main thread overhead.

Culling is definitely a weak link. Not sure if it's just not doing an early out where it should or something is requiring touching every instance every time. It performs like it's iterating every instance.

 

I'm sorry but my repo steps boiled down to

1. Download pre built scene from asset store, hit play __ benchmark 1
2. Download URP package, upgrade materials, hit play __ benchmark 2
3. Download Hybrid render package, add subscene, hit play __ benchmark 3
4. Add ENABLE_HYBRID_RENDERER_V2, hit play __ benchmark 4

If one of these steps is too complicated, I think you guys have a bigger problem on the usability of your package system and asset-store because i really can't see what is complicated about this.

If I was to upload the project it would literally just be the naked assets from the store as there is literally nothing else in the project. I no longer have this setup. It was just a quick test for our company to compare the performance of different pipelines for a future project.

 

Culling by chunk isn't really, it's still doing per instance work just less of it. It's obvious if you look at how the job cpu time scales with per instance and without.

It seems to be roughly 25% or so less expensive at best. Which for vegetation isn't enough. Not sure if this is even fixable with how batch render group works?

 

The Hybrid renderer V2 takes up 1.5G of my memory space, why is that when there are only two materials in the sence?

 

Which pre built scene is that?

 

VERSUS

...and apologies for my useless post

 

Ya I sort of pieced this together. I was using a third party asset from the asset store since I didn't have any vegetation with SRP shaders, and it was on sale. And they had the basic instancing setup but not the DOTS specific stuff. It seems for complex shaders authors are using SG but then modifying the generated source to get around what's not supported by SG. That was the case here and actually every case I've seen so far.

 

Getting very inconsistent results between the hybrid renderer and the standard one. Identical models & materials being lit differently when side by side in the same scene, only difference between them is one has convert to entity and the other doesn't. Is there a way to fix this?