When would it ever be appropriate to select 'uniformly sample hemisphere'? I feel that the computational cost of doing 'importance sample x' is not sufficiently large for most workloads to warrant avoiding it in favor of doing uniform sampling. Maybe I am missing something?
There is probably a tradeoff between the number of objects in the scene and the runtime for sampling the lights for example. If we have a ton of lights in the scene, sampling from each light may be pretty inefficient, moreso than sampling from the hemisphere.