Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!swrinde!zaphod.mps.ohio-state.edu!maverick.ksu.ksu.edu!unmvax!uokmax!alliant.backbone.uoknor.edu!jay
From: jay@alliant.backbone.uoknor.edu (Jay Liang)
Newsgroups: comp.sys.alliant
Subject: FX/FORTRAN MATMUL substitution
Keywords: FX/FORTRAN, MATMUL, fortran0, fortran1
Message-ID: <1990Aug6.203133.28803@uokmax.uucp>
Date: 6 Aug 90 20:31:33 GMT
Sender: news@uokmax.uucp (Nets)
Reply-To: jay@alliant.backbone.uoknor.edu (Jay Liang)
Organization: Geosciences Computing Network, University of Oklahoma, Norman, OK
Lines: 24


Does anybody know of a way to turn off the FX/FORTRAN compiler routine
substitution feature?  Specifically in fortran0, it substitute matrix
multiplication portion of the source code by MUTMUL routine which is
done in the fortran0 stage of the compilation.  I have tried skipping the
fortran0 part of the compilation and start with fortran1 and ld but the
results are about 50 times SLOWER than compile with 

	% fortran -O -DAS file.f

I have also tried to use the subroutine calls to trick the compiler
to not substituting MATMUL but the best results I have got so far was
about 4 times slower than the compiler optimized results.  One related
question: does anyone have access to the method Alliant does its
fortran1 conversion.  My conjecture is that by using the subroutine
calls, it was done in vector mode only, hence the 4 time difference
and by skipping fortran0, it is done in concurrent mode only hence the
~50 times slowdown.  Any help or comments are welcome.  Thanks in
advance.
--
Jay Liang (jay@alliant.backbone.uoknor.edu)
Geosciences Computing Network
University of Oklahoma
Norman, OK 73019