this reads confused, but i found it interesting precisely because of that...
the whole discourse on the over operator and compositing trees makes a real meal out of something very simple. i'm not sure that "order of compositing doesn't matter" (vs. order in layers) is a particularly deep insight... its the sort of thing we just intuitively know from real-life experience afaik. not worthy of a discussion with trees and equations...
i'd be curious why the author got stuck with compositing frame buffers together though - this is usually very avoidable, even in the cases described, and even on old hardware. i'm guessing there are some /really/ hairy edge case things that are not mentioned...
also, all this talk as if everything is built with dynamic, upload-every-frame, vertex buffers as well, instead of (given that its 2d) some well organised quad renderer... i doubt changing shader/material parameters was the dominant cost somehow.
there are certainly somethings you like to do in 2d (post-processing filters) that are hard in 3d, but drawing your basic stuff isn't one of them - there is no mystical difference between 2d and 3d, one is just a subset of the other - if you ignore the hacky pixel space stuff we do in post (which we can also do over 3d renders) which is a trick of the representation (pixels) and not the geometric space (2d).
I don't think you read it very carefully. Some of the operators he's supporting are not OVERs. These "filter effects" and blends are non-associative and that's where the tricky stuff comes in.
the whole discourse on the over operator and compositing trees makes a real meal out of something very simple. i'm not sure that "order of compositing doesn't matter" (vs. order in layers) is a particularly deep insight... its the sort of thing we just intuitively know from real-life experience afaik. not worthy of a discussion with trees and equations...
i'd be curious why the author got stuck with compositing frame buffers together though - this is usually very avoidable, even in the cases described, and even on old hardware. i'm guessing there are some /really/ hairy edge case things that are not mentioned...
also, all this talk as if everything is built with dynamic, upload-every-frame, vertex buffers as well, instead of (given that its 2d) some well organised quad renderer... i doubt changing shader/material parameters was the dominant cost somehow.
there are certainly somethings you like to do in 2d (post-processing filters) that are hard in 3d, but drawing your basic stuff isn't one of them - there is no mystical difference between 2d and 3d, one is just a subset of the other - if you ignore the hacky pixel space stuff we do in post (which we can also do over 3d renders) which is a trick of the representation (pixels) and not the geometric space (2d).