Is it faster for all threads in a warp to compute the same value, or for one thread to compute it and broadcast it to the other threads?

[maxj]

@Curefab

Then you can calculate 32 locations and distribute one after the other to all 32 threads for common processing of each location.

Ah, thank you, this idea hadn’t occurred to me (I’m quite new to CUDA programming). Very helpful :)
I suppose it would mean launching 32x fewer threads, so might not always be ideal in terms of occupancy if the inputs aren’t sufficiently large, though I guess I could tweak this to use only a portion of the 32 values in such cases.

@Curefab

Very basic integer computations can be done with the uniform instructions, which is like a separate lane and execution unit common for all 32 threads

Is there any way of knowing what the compiler will be able to recognise as uniform without somehow inspecting the SASS? E.g. is it likely to know that dividing the thread index by 64 will give the same result for all threads in the warp?

@Curefab

There is no advantage (for basic arithmetic computations) to just use one lane or one thread.

I guess this leaves open the possibility that in some cases (non-“basic” arithmetic) it could be faster?