This is a repost of my guestpost originally posted on chorasimilarity’s blog.
I designed an algorithm that, given an octree and a location within that octree can compute the cubemap at that location. The algorithm kind of raytraces each face of the cubemap, but does this in such a way that multiple rays are traced simultaneously (aka. mass connected processing).
The faces of the cubemap are axis-aligned, which allows us to use linear interpolation rather than perspective correct interpolation (which includes 2 multiplications and a division). Because of this linear interpolation, the algorithm only needs integer additions, subtractions, and multiplications by a power of two (aka. bitshifts).
For a more thorough explanation of the algorithm please read the Article.
I’ve also written a partial implementation of the algorithm, available at Git-hub, which renders the positive z plane of the cubemap. It currently does so at 2 fps, which, considering I did not yet put effort in optimization and screen space occlusion is not yet implemented, seems promising. The data is still fully loaded into memory at startup. As background I render a fixed cubemap.
Due to space considerations, not all models have been uploaded to git-hub, but a small low-resolution model is available, such that after compiling the code, you can at least see something. The code is rather messy as it is still in an experimental state.