You can use either version of the library. ACML 5.1.0 was primarily released to provide support for our newest Family 15h processors, however it includes the same optimizations that 4.4.0 had for our Family 10h processors.
Using the gfortran compiler as an example, if you are building for our newest Bulldozer processors you would use the libraries in gfortran64_fma4, while for the Opteron 6168 you would use the libraries in gfortran64.
Thanks for the answer. However I was wondering whether the library gfortran64 in ACML 5.1.0 contains or not the fast math vector. For example, in ACML 4.4.0 I was using explicitly gfortran64_mv to include math vect, and gfortran64 to don't use math vect. (i.e. CFLAGS = -Ofast -march=native -ftree-vectorize -funsafe-math-optimizations -mveclibabi=acml and linking with gfortran64/lib/libacml_mv.a)
How about ACML 5.1.0? May I use CFLAGS = -Ofast -march=native -ftree-vectorize -funsafe-math-optimizations -mveclibabi=acml and link against gfortran64/lib/libacml.a (or gfortran64 lib or whathever is named)?
As far as I understood I cannot use gfortran64_fma4 of ACML 5.1.0 because I need family 15 CPUs...
From your words I understood that for my needs it should be better to use ACML 4.4.0 (BTW I have a local copy, but the version for gfortran64 is disappeared from the download page).
[Added]: Can you confirm that if I want to use mv library I have to switch to ACML 4.4.0 on fam 10h CPUs?
Ah. Yes, starting with 5.0.0 we removed the fast vector math routines and moved them to their own seperate download.
You can find that at:
For your case it may be better to stay with 4.4.0. 5.1.0 will run on AMD Family 10h CPUs, but you will need to also link in the vector math library.
ACML 4.4.0 (and most previous versions) is still available for download on the ACML Archive page.
Thanks! Now it-s more clear to me!
What a pity that new libraries in some way cannot use all the features of a not-so-old CPU...