Floating Point Texture Support?

Anyone happen to know how widely supported FP32 textures are? Can I safely assume that anything which does Shader Model 2.0 will support them?

Also, if there's a way to detect support in DX9, that might be useful too.

Full HDR is Geforce 7+ and X1000+
below that, full FP32 without fakes and restrictions is not present.

Especially ATI did not have full 32bit precision

Are you sure? My Geforce old 6600GT creates FP32 textures without complaint.

Might be that it is already GeForce 6

There is an article on how Valve integrated HDR into HL2 on what obstacles they need to get out of the way.

In that article they mention that one of the GeForce generations (not sure if it was GF5 or 6 now that I think more about it) needed to use 16Bit as 32bit was not fully hardware supported. In the end thought it would make sense if the GF6 were capable of it, as NVidia Tech is always a generation in front compared to ATI

At simplest you create a little test app that you offer to people to see where it works full hardware accelerated and where through fakes / lower precision that especially generates profiling data. So you know which generations actually were capable of doing it at a usefull performance. Because support only is not of much use unless you plan to write a still renderer ;-)

Holy crap, you weren't kidding about the performance of 32 bit floating point textures, were you? Even on an 8600GTS, switching to 16 bit textures from 32 bit caused an FPS jump of 50%. From 600 FPS to almost 1000 FPS. Ok, I'm convinced. Assuming it proves to have enough accuracy, I'm sticking with 16 bit textures.

HDR textures really cripple performance on older cards, even my 7800gt takes a massive speed hit.

Really? Gee, I wonder if I would get a further performance boost if I rewrote the shader to encode the depth values into a standard ARGB texture and ditched floating point textures altogether? Then I wouldn't have to worry about compatibility at all.

That has to be worth a test when I have some spare time.