GotoBLAS on Intel Core Duo + SS12

GotoBLAS is renowned as the fastest BLAS library. It has been used by many supercomputing sites to run HPL, the infamous benchmark that determines their position in the Top500.

I've had a chance to test the combo of Goto+HPL using Sun's compilers (Sun Studio 12 aka. SS12) on OpenSUSE 11.0 x86 (i.e. 32-bit version). The machine is a laptop: IBM Thinkpad R60 with Intel Core Duo T2400 1.83GHz. [In the past, I've had relied on ACML when building HPL for AMD platforms.]

Here're the steps to build GotoBLAS on the abovementioned environment, assuming that the source archive has been extracted and you're in the source-code directory:

1. Open detect in a text editor, then change FCOMPILER=NULL to FCOMPILER=SUN
2. If your SS12 is installed in a non-standard directory (e.g. /tmp/sunstudio12), then open Makefile.rule in a text editor. Then, apply the following two changes:
   
   • ifeq ($(F_COMPILER), SUN)
     COMPILER_F77 = f90
     into
     ifeq ($(F_COMPILER), SUN)
     COMPILER_F77 = /tmp/sunstudio12/bin/f90
     
   • ifneq ($(OSNAME), SunOS)
     SUNPATH = /opt/sunstudio12
     into
     ifneq ($(OSNAME), SunOS)
     SUNPATH = /tmp/sunstudio12
     

And that's it! Should you wonder that I forgot about the C compiler: no, I don't. Leave it to use gcc.

Note that on this particular environment, -fast must not be used as it results in a floating-point exception on one of GotoBLAS's self-tests. YMMV (your mileage may vary) though.

Note2 that -xtarget=native is safe (again, it has to be qualified as “only on this particular CPU architecture”).

In addition to the built-in self-tests, another test was to build HPL with libgoto.a (and CT8), and then run the xhpl. And, not surprisingly, everything went smoothly; no further hiccups or whatsoever.