Hi PGI,
Are the Mathematical Standard Library Functions Maximum ULPs in CUDA FORTRAN/OpenACC the same as those published in the CUDA C Programming Guide Programming Guide :: CUDA Toolkit Documentation?
Cheers,
Kyle
Hi PGI,
Are the Mathematical Standard Library Functions Maximum ULPs in CUDA FORTRAN/OpenACC the same as those published in the CUDA C Programming Guide Programming Guide :: CUDA Toolkit Documentation?
Cheers,
Kyle
Hi Kyle,
If the compiler generates CUDA C that uses these routines (and you’re not using -ta/-Mcuda=fastmath), then yes you can use this as a guide. However, there other factors may effect the numerical accuracy of your program, such a FMA, order of operations, and other optimizations.
Hey Mat,
You said
If the compiler generates CUDA C that uses these routines
Does your response still hold if I use “-Mcuda=llvm”?
Cheer,
Kyle
Does your response still hold if I use “-Mcuda=llvm”?
Sorry for being a bit non-committal with the following response. I would say that in the cases where the same routines are called, then yes. What I don’t is how many are called.
A better question might be how can I compile for bit-exactness? This is something we’re working on but no information as to when it would be available. For now, we can’t guarantee bit exactness, but you can lower optimization and add “-Mcuda=nofma” to help.