Right now when culling the models to render for the directional light shadow map pass I just do a view frustrum culling using the main camera. At some angles, the objects will be outside the view but still obviously should cast shadows.
My question is, how do you cull/gather all the relevant shadow casters for a directional light?
There is an advanced technology based on compute shaders that will analyze the depth buffer of the main camera, and return the min depth, and max depth, with 1 pixel precision (reduce operation). This information can be used to create a very precise frustum for the shadow camera. Elements outside the view that casts shadows are only in the frustum that goes from the viewed zone to the light. (oblique in the direction of the lighting).
One issue with this is that you don't know the min/max before rendering the world, and you need the shadow map to render the world. So to palliate to that, you can simply use a min/max pair that is one frame late. And to mitigate potential little artefacts if the player moves too fast, you can inflate the volume slightly by some epsilon.
I don't do a frustum cull with the main camera, but the shadow camera alone.
I calculate the AABB (axis aligned bounding box) of the shadow camera, then do the culling with that. A directional shadow camera is a box in itself (orthographic projection matrix projects to a side of the box). After the box is rotated (for different light direction), the AABB can still be computed easily (but it will be a little bigger than originally, so it is a bit wasteful).
I compute it like this: Rotate every corner of the bounding box, then find the new min and max corner from the bunch of new coordinates. An AABB is essentially just a pair of min and max points.
You can check AABB-AABB intersections after that which should be quite fast.
