Quote:Original post by daedalic
Is shader instancing best, or is using two streams the best idea?

As the idea of "shader instancing" is a bit counter-intuitive to me (it is far from being a "true" instancing technique) I want to link this from gamedev and this from MSDN.

Now, let's make the two really short:

• True instancing (in D3D9) reduces the batch count by having a special mechanism that will "somehow" fetch some VS varying in parameters differently for each stream. Each index will be fetched normally and resolve to a set of vertex attribs according to the "stream divider" or "stream frequency".
  This allows to draw a variable amount of batches with limited effort.
  Because stream frequency is a per-stream-port property this requires 1 extra stream.


• Shader instancing is a data trick which really produces a long batch. The vertex attributes are resolved normally.
  This implies everything could be put in a single stream as the "instance" data will really have to be manually replicated. Drawing a variable amount of instances requires more work and more bandwidth but it's still a very good deal as the vertex count is low to start with.

Now, if you think at it, you'll figure out the two things are not really comparable. The longer batch produced by shader instancing is really a standard batch and this means the vertex processing can work more efficiently.
In line of theory, the same could be said for the true instancing but that was not the case. As the complexity of the instanced mesh grows, the benefit of true batching decreases. I think this is the last document talking about D3D9...
You see that long batches will give better performance over the same amount of vertices obtained by "true" instancing.

I personally don't like much the idea of shader instancing (altough I used a variation of it based on uniforms instead of varyings).
Anyway, you see that there's a tradeoff between the two.

Quote:Original post by daedalic
how do you access data from other streams in HLSL?

You don't... need to care. The API fetches the vertex attribs for you depending on the Stream Frequency and VertexDecl.

If we're talking D3D9 here, then "shader instancing" has a few disadvantages:

1. Your batch size is limited by the how much instance data you can stick in the 256 constant registers available to the vertex shader

2. You have to extend both your vertex buffer and your index buffer, and also add an instance index to each vertex

3. Indexing into constant registers causes poor vertex shader performance.

You can mostly get around 1 and 3 by using a texture instead of shader constants, but only for hardware that has decent vertex texturing performance (DX10 class or better). And any hardware that has good vertex texturing performance supports hardware instancing, so you might as well use that instead.