Trace a dedicated light shadow ray for scatter events.
Makes it so a volume scene with a distant shadow-linked (excluded)
object renders the same as the scene without shadow linking at all.
Refactored code a bit so that there is less duplication.
The downside is that the camera ray visibility for light shader
is now checked twice (in the intersect_closest and shade_light),
but it is unlikely to cause measurable performance impact.
The goal of this change is to make it so the emitters are not forced
to have MIS disabled when shadow linking in used. On the user level
it means that lights sources sources which are behind excluded shadow
blocker will be visible in sharp glossy reflection.
In order to support this an extra shadow ray is traced in the direction
of the main path, to gather contribution of emitters behind shadow
blockers. This is done in the two new kernels.
First kernel performs light intersection to see if the extra shadow
ray is needed. It is not needed if, for example, there are no light
sources in the direction of the main path.
The second kernel shades the light source and generates the actual
shadow path.
Such separation allows to keep kernels small (so that no intersection
and shading happens in the same kernel). It also helps having good
occupancy on the GPU: the main path will not wait for the shadow
kernels to complete before continuing (the main path is needed by the
extra shadow path generation, to have access to direction and shading
reading state).
Current implementation is limited to lights only: there is no support
of mesh lights yet.
The MIS weight of the new shadow ray needs to be double-checked. It was
verified to give same result as the main path when the same light is
hit.
To avoid contribution from the same light source counted by the shadow
ray and the main path the light source which was chosen to trace the
shadow ray to is excluded from intersection via transparent bounces.
This seems unideal and feels that it could be done via some MIS wights
as well. Doing so is an exercise for later.
The kernel naming could be improved. Suggestions are welcome.
There are also quiet some TODOs in the code. Not sure how much of those
must be resolved prior to merge to the cycles-light-linking branch.
Would be cool to have extra eyes on some of the MIS aspects, which is
easier in the shared branch.
Tested on the CPU and Metal on M2 GPU.
Sergey Sharybin