CUDA exp() expf() and __expf()

Question

How to optimize the exp function in cuda? What are the differences between the following in CUDA??

   exp()
   expf()
   __expf()

Answer 1

The differences are explained in the CUDA C Programming Guide, appendix D.

• exp() should be used for double precision, although should be overloaded for single
• expf() should be used for single precision (float)
• __expf() is the fast-math version, the performance is faster with some loss of precision (dependent on the input value, see the guide for more details).

Answer 2

Generally exp() is for doubles, expf() for floats and both are slightly slower than __exp() which is available as a hardware operation. The performance gain usually comes at the cost of accuracy but unless you are really concerned about accuracy it shouldn't be a problem.