With CUDA 2.0, in emulation mode, sqrt is returning single precision results instead of double precision results. Is this a bug or is there some way to specify whether emulation runs in SP or DP mode?
not sure, but try compiling with -arch sm_13 -deviceemu. not sure if both are required at the same time, but that’s my guess without checking with the compiler team
Many thanks. I added -arch compute_13 in front of the already present -deviceemu and that did the trick.
(NVIDIA: Why is the 2.0 release documentation lagging here?)