GPU instancing

[nowakpe]

Is there a method similar to Unity "https://docs.unity3d.com/ScriptReference/Graphics.DrawMeshInstanced.html"
I could not find a similar simple method directly in UE4, so I am expecting that it is not, but if someone has any ideas? Thanks.

Update:
I found it.
UInstancedStaticMeshComponent::BatchUpdateInstancesTransforms
But it is not yet exposed.
Right now only for single instance InstancedStaticMeshComponent.UpdateInstanceTransform

[nxrighthere]

UInstancedStaticMeshComponent::BatchUpdateInstancesTransforms() is essentially doing the same thing under the hood. Just use markRenderStateDirty for the last element in a batch with InstancedStaticMeshComponent.UpdateInstanceTransform(), as stated in the API reference. See this test.

[nxrighthere]

Maybe it can be also achieved with Vector3 and explicit layout? It is probably more performant to use 16 bytes, because of the vectorization, SIMD stuff...

I have no control over .NET vector types from System.Numerics, so intermediate vector types are used to map them into Unreal vector types.

By "4 float vectors" I meant that the FTransform is Quaternion (4 floats), Location (4 floats), and Scale (4 floats).

Ah, that makes sense yes.

Yes, but if you used directly c++ in Unreal you do not have to copy anything also, it is just because that method accepts the type of TArray.

Try to use this constructor:

TArray(const ElementType* Ptr, SizeType Count)

[nowakpe]

Try to use this constructor:

Checked, it is also using memcpy internally. But thanks for the tip.

Edit: Hm ok, I also tried it and for 150K instances, it gives another 10 fps :D maybe I missed some branching.

[nowakpe]

The result https://www.linkedin.com/posts/petr-nowak-513198156_unrealengine-ue4-unrealclr-activity-6723503591308464128-kZyp
(Didn't know that linked in will decrease the quality of the video that much :-D)

[nxrighthere]

Pretty cool! May you share it here in the Showcase category, please?

Yes, all those websites are screwing video quality.

[nxrighthere]

Transforms are now used directly where appropriate 8aa5c6e. Thanks for participation.