For an algorithm I am currently looking to port to GPU, we require double precision floating point logarithm and exponential calculations. My understanding (based on looking at the ptx isa 2.0) is that these will not be supported yet with compute capability 2.0, (all transcendental functions: sin, cos, exp and log appear to be limited to f32 only).
I just wanted to check that I am reading this correctly, could someone please confirm or refute this?
Thank you,
BW
There have been double precision transcendental functions in CUDA for some time now. They are supported on compute capability 1.3 and greater devices. Appendix C.2 in the CUDA Programming Guide gives the precision of each.
Edit: I should clarify. PTX includes instructions which allow fast approximations to some of the transcendental. If you are not using __sinf(), __cosf(), __logf(), the implementation of the function requires many instructions in both the single and double precision case.
Ah, yes, I misunderstood what I was looking for. Thanks for pointing me in the right direction.
BW