Drawing massive amounts of textured cubes in XNA 4.0

Question

I am trying to write a custom Minecraft Classic multiplayer client in XNA 4.0, but I am completely stumped when it comes to actually drawing the world in the game. Each block is a cube in 3D space, and it is possible for it to have different textures on each side. I have been reading around the Internet, and found out that for a cube to have a different texture on each side, each face needs its own set of vertices. That makes a total of 24 vertices for each cube, and if you have a world that consists of 64*64*64 cubes (or possibly even more!), that makes a lot of vertices.

In my original code, I split up the texture map I had into separate textures, and applied these before drawing each side of every cube. I was told that this is a very expensive approach, and that I should keep the textures in the same file, and simply use the UV coordinates to map certain subtextures onto the cube. This didn't do much for performance though, since the sheer amount of vertices is simply too much. I was also told to collect the vertices in a VertexBuffer and draw them all at once, but this didn't help much either, and occasionally causes an exception when the number of vertices exceeds the maximum size of the buffer. Any attempt I've tried to make cubes share vertices has also failed, resulting in massive slowdown and glitchy cubes.

I have no idea what to do with this. I am pretty good at programming in general, but any kind of 3D programming or game development completely escapes me.

Here is the method I use to draw the cubes. I have two global lists List<VertexPositionTexture> and List<int>, one for vertices and one for indices. When drawing, I iterate through all of the cubes in the world and do RenderShape on the ones that aren't empty (like Air). The shape class that I have is pasted below. The commented code in the AddVertices method is the attempt to make cubes share vertices. When all of the cubes' vertices have been added to the list, the data is pasted into a VertexBuffer and IndexBuffer, and DrawIndexedPrimitives is called.

To be honest, I am probably doing it completely wrong, but I really have no idea how to do it, and there are no tutorials that actually describe how to draw lots of objects, only extremely simple ones. I had to figure out how to redo the BasicShape to have several textures myself.

The shape: http://pastebin.com/zNUFPygP

Answer 1

You can get a copy of the code I wrote with a few other devs called TechCraft:

http://techcraft.codeplex.com

Its free and open source. It should show you how to create an engine similar to Minecrafts.

Answer 2

There are a lot of things you can do to speed this up:

What you want to do is bake a region of cubes into a vertex buffer. What I mean by this is to take all of the cubes in a small area, and put them all into one vertex buffer. Only update this buffer when a cube changes.

In a world like minecraft's, LOTS of faces are occluding each other. The biggest thing you can do is to hide faces that are shared between two cubes. Imagine two cubes sitting right next to each other, you don't really need to draw the face in between, since it can never be seen anyway. In our engine, this resulted in 20 times less vertices.

 _ _      _ _
|_|_| == |_ _|

As for your textures, it is a good idea, like you said, to use a texture atlas. This greatly reduces your draw calls.

Good luck! And if you feel like cheating, look at Infiniminer. Infiniminer is the game minecraft was based off. It's written in XNA and is open-source!

Answer 3

You need to think about reducing the size of the problem. How can you produce the same image by doing less work?

If your cubes are spaced at regular intervals and are all the same size, you may not need to store the vertices at all - your shader may be able to calculate the vertex positions as it runs. If they are different sizes and not spaced at regular intervals, then you may still be able to use some for onf instancing (where you supply the position and size of a cube to a shader and it works out where to render the vertices to make a cube appear at that location)

If your cubes obscure anything behnd them, then you only need to draw the front-most cubes - anything behind them is just not visible. A natural approach for this would be to use an octree data structure, which divides 3D space into voxels (cubes). Using an octree you could quickly deternine which cubes are visible, and just draw those cubes - so rather than drawing 64x64x64 cubes, you may find you nly have to draw a few hundred per frame. You will also find that as the camera moves, the set of visible cubes will not change much, so you may be able to use this "temporal coherence" to update your data structures to minimise the work that needs to be done to decide which cubes are visible.