That's cool!
Back when I came up with it (June 11th '08, last day of the DevStation, on the train to the airport
), I had not seen anyone do it. I've not been on DevNet since October, so I wasn't aware of anyone else doing per-pixel lightsource classification.
Would be interesting to compare results. I'm sure if Insomniac is doing it, they must be a good chunk faster at it than me.
...5 minutes later...
OK, I just checked the Insomniac R&D site (I'll have a look at Blur later), where they talk about their deferred system. The way I understand it, they render the light volumes while using the scene depth and then use the stencil test determine if the z-value at a given pixel is in the light volume.
This is not the same. First of all, they are using a prelighting renderer, not a deferred setup. Unless I'm totally off, they will render one light-volume at a time during the prelighting-pass, while we process all lights at once, allowing us to do that in full FP32 on an RGBA8 target.
It's really rather different. Deferred vs. prelighting, stencil-tests vs. SPU.
Anyway, I like their setup. When Wolfgang published the pre-lighting, I was sadly a bit late for us to give it a shot.
Any Insomniacs around?