I've seen some shadows optimisations from leadwerks engine, however they don't want to explain me precisly what they've done, so I'm going to try to do this alone.
My goal is to use shadows faster with no rendering difference with my actual system. I think a great thing would be that instead of using 1 shadow map per light, we use 3 render target per light : 1 used for static shadow map (not moving objects), another for dynamic shadow map (moving objects), and the last one would be a Z-buffer from the static shadow map render to produce the dynamic one with Z-sorting (by ouself in the shader). The trick is to render the static shadow map / Z-buffer 1 time, and then only update the dynamic shadow map with moving objects. At the end we merge both static & dynamic maps to produce our shadows.
Here is a little illustration I did :
I would like to know what you think of this system.
Is it better to use idle/moving object lists or just define directly static/dynamic nodes forever so we don't have to check if a node moved from the last frame?
What do you think in term of performance ? Using 3 render target instead of 1 should run slower on the GPU, however if you have a scene of 300 000 polygons and only like 1000 polygones moving you can go up to 300 times faster.
[ cube map :
normal render : 6 * 300 000 = 1 800 000 polys
only moving objects : 6000 polys ]
Did anyone on this forum attempted such a system ?
I think it's called a cache system and it can be also used with reflexion maps. But here I'm only after shadows... for now
Thanks all
[EDIT :](edit2 :Following remark is wrong) I didn't thought, but how can we don't update the static buffer if the camera moves? shadow maps are camera dependent, so this would be an optimisation for non-moving camera only ? That sucks
