In #1645 we implemented rendering layers, but there's some optimization still possible.
my idea there was (and we should discuss/continue from there):
The renderer has two "levels", the high level variant is level 2, the lower level (closer to opengl instructions) is level 1.
We should have something layer-like (ordered multi-pass rendering) in both renderer levels.
There's basically two things to do multiple passes in renderer "level 1" to draw their layers:
- objects (town center e.g.) -> draw shadows, their building parts
- terrain -> overlay and blending
Layer 2 render passes could use layer/pass features in layer1 then, if it makes sense.
To prevent binding the framebuffer when drawing in layer 1 too often, the resulting layer1 instructions somehow have to be batched (so the buffer can remain to be bound in the correct order) - this was our long-term idea for the multiple layers anyway.
Concrete idea to implement this:
- renderer users submit all render instructions to the renderer as they need them
- and have it all in memory, yes
- but later we could optimize maybe by processing one map area after the other? especially when zoomed out)
- re-order and optimize (minimize binding switches of textures, shaders, ...) the draw instructions
- from this optimized list, perform the draw calls (to the correct framebuffer) with minimal binding switches (we ordered for that just before)
In #1645 we implemented rendering layers, but there's some optimization still possible.
my idea there was (and we should discuss/continue from there):
The renderer has two "levels", the high level variant is level 2, the lower level (closer to opengl instructions) is level 1.
We should have something layer-like (ordered multi-pass rendering) in both renderer levels.
There's basically two things to do multiple passes in renderer "level 1" to draw their layers:
Layer 2 render passes could use layer/pass features in layer1 then, if it makes sense.
To prevent binding the framebuffer when drawing in layer 1 too often, the resulting layer1 instructions somehow have to be batched (so the buffer can remain to be bound in the correct order) - this was our long-term idea for the multiple layers anyway.
Concrete idea to implement this: