I probably should have added, it's not just the learnings from DEC Gem
folks, but also the old "Kuck and Associates" team formerly in Champaign
(Intel moved them all to Austin).
On Mon, Feb 24, 2020 at 11:15 AM Clem Cole <clemc(a)ccc.com> wrote:
First please continue this discussion on COFF (which has been CC'ed).
While Fortran is interesting to many, it not a UNIX topic per say.
Also, as I have noted in other places, I work for Intel - these comments
are my own and I'm not trying to sell you anything. Just passing on 45+
years of programming experience.
On Mon, Feb 24, 2020 at 10:34 AM Adam Thornton <athornton(a)gmail.com>
I would think that FORTRAN is likelier to be
passed around as folk wisdom
and ancient PIs (uh, Primary Investigators, not the detective kind)
thrusting a dog-eared FORTRAN IV manual at their new grad students and
snarling "RTFM!" than as actual college courses.
FWIW: I was at CMU last week recruiting. Fortran, even at a leading CS
place like CMU, is hardly "folk wisdom". All the science PhD's (Chem, Mat
Sci, Bio, Physics) that I interviewed all knew and used Fortran (nad listed
on their CV's) as their primary language for their science.
As I've quipped before, Fortran pays my own (and a lot of other people's
salaries in the industry). Check out:
Fortran is about 90% of the codes
running (FWIW: I have seen similar statistics from other large HPC sites
- you'll need to poke).
While I do not write in it, I believe there are three reasons why these
statistics are true and* going to be true for a very long time*:
1. The math being used has not changed. Just open up the codes and
look at what they are doing. You will find that they all are all solving
systems of partial differential equations using linear algebra (-- see the
movie: "Hidden Figures").
2. 50-75 years of data sets with know qualities and programs to work
with them. If you were able to replace the codes magically with something
'better' - (from MathLab to Julia or Python to Java) all their data would
have to be requalified (it is like the QWERTY keyboard - that shipped
sailed years ago).
3. The *scientists want to do their science* for their work to get
their degree or prize. The computer and its programs *are a tool*
for them look at data *to do their science*. They don't care as long
as they get their work done.
Besides Adam's mention of flang, there is, of course, gfortran; but there
are also commerical compilers available for use: Qualify for Free
Software | Intel® Software
believe PGI offers something similar, but I have not checked in a while.
Most 'production' codes use a real compiler like Intel, PGI or Cray's.
FWIW: the largest number of LLVM developers are at Intel now. IMO,
while flang is cute, it will be a toy for a while, as the LLVM IL really
can not handle Fortran easily. There is a huge project to put a number of
the learnings from the DEC Gem compilers into LLVM and one piece is gutting
the internal IL and making work for parallel architectures. The >>hope<<
by many of my peeps, (still unproven) is that at some point the FOSS world
will produce a compiler as good a Gem or the current Intel icc/ifort set.
(Hence, Intel is forced to support 3 different compiler technologies
internally in the technical languages group).