Development Blog

What happened to 

where you claimed 200 fps for 10000 characters? Was that all shared skeletons?

That's pretty neat so far.  I wonder how that same test will compare with dynamic shadows.

42 minutes ago, Ma-Shell said:

What happened to 

where you claimed 200 fps for 10000 characters? Was that all shared skeletons?

I believe the bottleneck in this situation is actually polygons rendered. I previously chose the goblin model because it was low poly, so I should do some more tests using that model instead of the 4000 poly crawler model.

I wish there was a standard for this but it really depends on your game.  I wonder how many triangles and bones these models are in Shadows of Mordor:

Or these in Arkham Knight:

I ran the test using the goblin model. On this machine 1024 uniquely animated goblins onscreen runs at 200 FPS. With the old renderer it runs at 380 FPS, but it's not exactly an even comparison. The Vulkan renderer is doing a depth-prepass, so all skinned vertices are actually being rendered twice, once for the prepass and once for the final pass.

If all things were the same, I am sure the VK renderer would be going faster. So we are within the same range as my original tests, and now we have full lighting, normal mapping, and other stuff being processed.

The engine will now allocate threads for animation based on how many skeletons there are and the total processing time for the last frame. It will attempt to allocate as many threads as are needed to keep the total processing time down within 60 FPS.

 

If I use shared animations so that less data is being uploaded to the GPU each frame, I get 400 FPS in the Vulkan renderer.

Actually I have found the bottleneck is in the data sent to the GPU. I added some tricks and now 1024 goblines are running at 450 FPS on my laptop, a bit faster than the original OpenGL version, even though everything is being drawn twice in the Vulkan renderer for the depth pre-pass.

I am satisifed.