As a start, you must decide how to compute that distance. Of course at the end it will be a subtraction between two values: the light position and the position of the fragment being shaded.

For this operation to give meaningful results, those two values must be in the same coordinate space. There are a couple I want to mention:

• World-space is pretty easy to mangle as light entities are usually managed in world-coords. It requires some extra work in VS and PS. I have been told the value ranges might be a problem although I never observed this issue;
• Eye-space is "native" in the processing and "denser" in terms of values produced. I think it's historically the most widely used coordinate space. Needs some mangling CPU-side but GPU has it easy;
• Tangent-space (or "surface-local"): the light position is transformed so each fragment getting shaded gets a specific vector to the light WRT itself. Standard for initial DOT3-based bump-mapping implementations.

Are you familiar with those notions? Those are the hard part. Once you get the two points in the correct coordinate space, it will just be a matter of evaluating an attenuation function scaled by the inverse of distance (squared).

I'm sure you'll be somewhat confused right now. Where do you want to start?

I'm afraid I don't know know anything more about the lights used in Doom 3 than a few phrases said by John Carmack in one of his presentations, but if you are looking for volumetric lights this might help: http://www.humus.name/index.php?page=3D&ID=36

I'd like to implement something like this one day for a few atmospheric lights, but my BRDF research will take up quite some time, so if you manage to learn something about that demo, please share :).