Precomputed BSP Trees and Static Meshes

Entry posted by Josh · August 23, 2019

1,853 views

The GMF2 data format gives us fine control to enable fast load times. Vertex and indice data is stored in the exact same format the GPU uses so everything is read straight into video memory. There are some other optimizations I have wanted to make for a long time, and our use of big CAD models makes this a perfect time to implement these improvements.

Precomputed BSP Trees

For raycast (pick) operations, a BSP structure needs to be constructed from the mesh data. In Leadwerks Engine this structure is computed after a model is loaded. The "Bober Station: model from The Zone is a 46,000 poly model. Here are the build times for its BSP structure:

Release: 500 milliseconds
Debug: 6143 milliseconds

This is not our biggest mesh. Let's assume a two million poly mesh, which is actually something we come across with CAD data. With this size model our load times increase 40x:

Release: 20 seconds
Debug: 4 minutes

Those numbers are not good. To reduce load times I have embedded precomputed BSP structures into the GMF2 format so that they can be loaded from the file. I don't have any numbers yet because there is currently no GMF1 to GMF2 pipeline, but I believe this will significantly reduce load times, especially for large models.

I am interested in the idea of replacing the BSP float vertex values with shorts to reduce memory, but I'm not going to worry about it right now.

Static Meshes

I also am adding making the engine dump vertex and indice data from memory by default when a mesh is sent to the GPU. This will reduce memory usage, since you normally don't need a copy of the vertex data in memory for any reason.

void Mesh::Finalize(const bool makestatic)
{
	GameEngine::Get()->renderingthreadmanager->AddInstruction(std::bind(&RenderMesh::Modify, rendermesh, vertices, indices));
	if (makestatic)
	{
		if (collider == nullptr) UpdateCollider();
		vertices.clear();
		vertices.shrink_to_fit();
		indices.clear();
		indices.shrink_to_fit();
	}
}

This change reduced our system memory usage in "The Zone" by 40 mb.

2 Comments

Recommended Comments

Josh 15,479

Posted August 23, 2019

I tested a 1.5 million poly mesh and the BSP generation time was about three seconds. Not terrible given the size, but I would like to completely eliminate that annoying delay.