Hi guys!

This is my current issue, I'm trying to create 3d billboards, now this is how I think I should do it:

(Sample: Render thousands of grass instances)

Create Instance Buffer ( -> Position, Rotation)

Send Mesh Stuff + Instance Buffer

Shader:

{
  float4x4 rMatrix = generateRotationMatrix(instanceRot);
  position = mul(position, rMatrix);
  position += instancepos;

  ...apply World, View, Proj matrices here.
  ...
}

But is this the best way? Because I don't think passing a constant buffer with thousands of rotation matrices would be so good, am I wrong?

What does you're experience tell you to do here?

So the real question is: How can I individually rotate thousands of instances to the cameras view?

Thank You, like always!

You don't even need to store a matrix in your per-instance buffer (which I'd use a dynamic vertex buffer for, with D3D11_INPUT_PER_INSTANCE_DATA in the corresponding input layout so that you won't run into cbuffer space limits); you can just store a combined MVP in your per-frame cbuffer, then the position of each instance in your per-instance buffer.

In your per-frame cbuffer you also store a couple of vectors which you can extract from your view matrix, and that gives you everything you need to do billboarding, at a significantly reduced data cost and much higher ceiling on how many instances you can have per draw call.

A discussion of the technique (implemented in software, but which you should be easily able to convert to shader code) is available here: http://www.mvps.org/directx/articles/view_oriented_billboards.htm

Thank you mhagain and menohack!

mhagain, I followed your link and it is actually working now, and now I'm translating it into shader code, but here is my worry:

How can I translate the vertices individually, as each vertex has it's own location.

I guess I could use the SV_VertexID and then use some if's, but would that be slow?

I'm not asking for you to write the shader code, just guide me.

The easiest way would be to use a geometry shader. Input one point, output a 4-vert tristrip, and just lift it from the software version. You don't need to use instancing with this method, although having the GS stage active will introduce some (small) additional overhead. If that's acceptable, then go do it.

Another method would be to use an additional per-vertex buffer, containing 4 verts. This buffer can be static, and each vertex is 2 floats: {{-1, -1}, {1, -1}, {-1, 1}, {1, 1}} works well for one tristrip ordering (other tristrip orderings will be different, of course). Let's assume that this is called "offset" in your VS input struct, that "position" is a float3, and you get the following:

vs_out.position = mul (float4 (vs_in.position + right * offset.x + up * offset.y, 1.0f), globalMVP);

If you want each billboard to have a variable scale ("scale" in your input struct) modify like so:

vs_out.position = mul (float4 (vs_in.position + (right * offset.x + up * offset.y) * vs_in.scale, 1.0f), globalMVP)

This isn't exactly the same as the example in the link I posted; that one produces a diamond shape, this one is a square.

These "offsets" can alternatively be indexed via SV_VertexID and also be reused for texcoords, by the way.

Ohh but luckily I am going after a square

And if needed, I found this one: http://www.evolution3d.net/my3dstuff/papers/papers.php?key=billboards

It's basically a copy, but with the squares!

... Still Working on Billboards...

Thank you all for improving my knowledge on Billboards!

See you next time!