Originally posted by: khanna

Hello 

I'm attempting to experiment with Power9's native "quadruple precision" 128-bit floating-point hardware using XLC/XLF. The data type I'm using is "long double" in C and "real*16" in Fortran. 

It is difficult to tell whether or not the compiler is using the native 128-bit fp hardware or just doing some software emulation. The performance I get from basic test codes using "long double" data type is 10x slower than "double". So, I suspect that is just through software implementation and not native hardware.

After some digging .. I noted that I should have a special library "libC128core.a" and special compiler commands like xlc++core128, xlC128core_r, etc. But my install of XLC doesn't come with any of that. (I'm on Linux .. CentOS7). 

Anyone with feedback? Does anyone have any reference / sample code on how to take advantage of the Power9's native 128-bit fp hardware? 

Thanks!

Originally posted by: Rafik_Zurob

Hello.  The XL compilers currently support the "double double" long double format only.  They do not support the IEEE 128-bit long double format that the POWER 9 hardware supports.

Originally posted by: khanna

OK. Do you happen do know if GCC supports it? 

Originally posted by: Rafik_Zurob

It does, but it depends on the GCC version you have.  This page has some information:  https://gcc.gnu.org/gcc-6/changes.html  (e.g. You need -mfloat128 to turn it on).  If you have a newer GCC than that, please check the GCC manual to see if they use different options.