Is there a recommended technique to use when “eye” coordinates are very large, relative to geometry?
eye @ (10e6,10e6,10e6)
geometry all coordinates in range [1,-1]
Seems like I will run into float precision problems, with such large t-values?
Yes, you will hit precision problems. You can mitigate this for triangle intersection by rebasing your ray near the triangle. See the glass sample for an example of iterative triangle intersection that gets more accurate hits.
But you will also face this issue in BVH traversal. You don’t want your low precision ray to miss some bounding boxes that it should have hit. To address this you will need to pass a closer ray to rtTrace. If your case is as simple as you describe you could manually compute a new ray origin that is close to the unit cube.
Thanks Dave. That is actually the approach I took: compute the rays as doubles, advance the rays close to the target, then finally convert rays to floats.