CUDA Warp Divergence

Question

I' m developing with cuda and have an arithmetic problem, which I could implement with or without warp diverengence. With warp divergence it would look like:

float v1;
float v2;
//calculate values of v1 and v2
if(v2 != 0)
  v1 += v2*complicated_math();
//store v1

Without warp divergence the version looks like:

float v1;
float v2;
//calculate values of v1 and v2
v1 += v2*complicated_math();
//store v1

The Question is, which version is faster?

In other words how expensive is a warp disable compared to some extra calculation and addition of 0?

Answer 1

Your question has no single answer. This heavily depends on the amount of extra calculations, divergence frequency, type of hardware, dimensions and many more aspects. The best way is simply to program both and use profiling to determine the best solution in this particular case and situation.