In DirectX 11, batching primitives for performance, how does this actually work?

Question

I don't seem to be able to understand this. Microsoft says that one of the possible optimizations of for a Direct3D 11 if to batch primitives draw calls. For example in order to draw say 300 triangles calling the Draw method each time, you would instead pack and Draw them all at once.

The thing I don't get is how you would actually draw them in different locations and everything, because for example, I need one different world transformation matrix for each one. The only thing that comes to my mind is to calculate everything on the CPU and set the values for each vertex manually, however this approach doesn't seem like it would be a very good code optimization to me.

I feel like I'm missing something but I can't find the answer anywhere. The question is: how would you be able to pack things in a draw call without needing a render state change, and still get objects in different locations?

Note that I am not asking for pieces of code here, I'm primarily looking for a theorical answer to this.

Answer 1

There is many ways to do Batching. But for DX11, Instancing is the best choise, and it´s realy efficent and easy to use!

there is plenty of resources on DX11 instancing :

Rastertek instancing
MSDN Instancing reference

How it works is that you create an extra Buffer containing all the worldtransform matrixes. that you pass to the shader.

from there, you will get a "perlayout" matrix, that corresponds to each instance.

The old way of doing this is via calculating the world position for all the triangles befor drawing them, and then using the view_projection matrix.

Answer 2

The only thing that comes to my mind is to calculate everything on the CPU and set the values for each vertex manually, however this approach doesn't seem like it would be a very good code optimization to me.

You might be surprised, then.

Obviously, this will depend on the complexity of the geometry you're drawing and the transformations you're applying to it, but you have to remember that the CPU time consumed by Draw calls is not insignificant. It will very quickly add up if you have lots of little batches. So if the geometry and transformations are simple, it can very often make a lot of sense to trade the (relatively expensive) Draw calls for (relatively cheap) CPU-based transformations.

This is how XNA's SpriteBatch works--it accumulates potentially thousands of sprites into a CPU-based buffer by manually transforming vertices, then it sends them all to the GPU at once. Doing a separate Draw call for hundreds or thousands of sprites would bring your framerate to its knees, but calculating the positions of their vertices is easy and fast. It's a huge win.

As Tordin explains in his answer, hardware instancing can let you have your cake and eat it too, and is often a good idea. However, the idea of batching primitives the manual way is a general principle that applies even when hardware instancing isn't available.