I'm currently trying to implement bump mapping which requires to have a "tangent space". I read through some tutorials, specifically the following two:
Both tutorials avoid expensive matrix computation in the fragment shader which would be required if the shading computation would happen in the camera space as usual (as I'm used to, at least).
They introduce the tangent space which might be different per vertex (or even per fragment if the surface is smoothed). If I understand it correctly, for efficient bump mapping (i.e. to minimize computations in the fragment shader), they convert everything needed for light computation into this tangent space using the vertex shader. But I wonder if model space is a good alternative to compute light shading in.
My questions regarding this topic are:
For shading computation in tangent space, what exactly do I pass between vertex and fragment shaders? Do I really need to convert light positions in tangent space, requiring O(number of lights) of varying variables? This will for example not work for deferred shading or if the light positions aren't known for some other reason in the vertex shader. There has to be a (still efficient) alternative, which I guess is shading computation in model space.
If I pass model space varyings, is it a good idea to still perform shading computations in tangent space, i.e. convert light positions in fragment shader? Or is it better to perform shading computations in model space? Which will be faster? (In both cases I need a TBN matrix, but one case requires a model-to-tangent transform, the other a tangent-to-model transform.)
I currently pass per-vertex normal, tangent and bitangent (orhtonormal) to the vertex shader. If I understand it correctly, the orthonormalization is only required if I want to quickly build a model-to-tangent space matrix which requires inversion of a matrix containing the TBN vectors. If they are orthogonal, this is simply a transposition. But if I don't need vectors in the tangent space, I don't need an inversion but simply the original TBN vectors in a matrix which is then the tangent-to-model matrix. Wouldn't this simplify everything?
Normal mapping is usually done in tangent space because the normal maps are given in this space. So if you pre-transform the (relatively little) input data to tangent space in the vertex shader, you don't need extra computation in the fragment shader. That requires that all input data is available, of course. I haven't done bump mapping with deferred shading, but using the model space seems to be a good idea. World space would probably be even better, because you'll need world space vectors in the end to render to the G-buffers.
If you pass model space vectors, I would recommend to perform the calculations in this space. Then the fragment shader would have to transform one normal from tangent space to model space. In the other case it would have to transform n light attributes from model space to tangent space, which should take n times longer.
If you don't need the inverse TBN matrix, a non-orthonormal coordinate system should be fine. At least I don't see any reason, why it should not.