[Lapack] Vendor libraries, xLA* routines, and SAVE statements.
From: Jason Riedy
Date: Mon, 24 Apr 2006 09:26:28 -0700
And Sven Hammarling writes:
- > We now need to jump through some hoops to build
- > our sgerfsx the condition estimators against MKL.
- No. Just because a vendor makes a silly decision does not mean that we
I over-stated the hoops. I moved our code over to the
new thread-safe versions, which we needed to do anyways.
Besides, you can't blame them for not talking to their
compiler guys. They've only had a widely deployed
solution for problems like this for about two years.
Oh, wait, yes you can.
In many C compilers, the __thread attribute denotes that
a variable is thread-local. Some library and linker tricks
make it work quite efficiently, and it'd be easy enough to
add to their Fortran compilers. Only time it wouldn't
work is if the same xLACON computation is migrated between
- Have you made Intel aware of our new thread safe versions? Or
- would you like me to do that?
I was hoping the guilty party would pipe up, but he may
read only the scalapack list. Someone with more tact than
I have should volunteer the information. ;)
BTW, this problem bites ACML, too. I don't know who to
contact at AMD about it, but someone who hasn't lost the
relevant NERSC seminar announcement does.
 Anyone requesting such a the change will no doubt get
a lesson in how VAX/VMS did it better.