LEONARDZ@UOGUELPH.BITNET ("Len Zaifman UoGuelph 824-4120 xt 6566", 519) (08/25/89)
I need help on two issues.
1 . I am trying to compile some fortran code with optimisation. At O2 ,
several routines ( this program has several hundred ) hit the error
not optimised beacuase global tables are not large enough, use
Olimit > nnn where nnn is some number larger than 500. The question
is , does it hurt the performance of the program , or the program
size to compile with Olimit nnn,where nnn is greater than the largest
reported size required, for all the routines , whether they need the
space or not ?? The program is large enough as it is , and I am trying
to enhance its runtime execution speed as much as possible.
To that end I would also like someone from SGI to comment on the
advisability of using O3. I was told that at the moment this level of
optimisation is unsafe. Is that true , and what performance enhancement
is expected over O2 optimisation ?? Also, for a very large program such
as this what are the recommended analysis tools on IRIX to help
optimise the performance(hopefully without source code changes ). I am
looking for tools which help restructure the binary(loaded) code to
enhance performance.
2. On the user interface side , I see a need for easier command recall
than using !nnn to recall a command. Recently I worked on an Apollo
where command recall and editting was available through mouse
highlighting, cutting and pasting (ie like a mac). Can this be done
in IRIX now (3.1 x level)? will it be available at 3.2 ?? In
Workspace ??
Thanks
Regards,
Len Zaifmancalvin@dinkum.wpd.sgi.com (Calvin H. Vu) (08/26/89)
In article <89Aug25.093608edt.57688@ugw.utcs.utoronto.ca>, LEONARDZ@UOGUELPH.BITNET ("Len Zaifman UoGuelph 824-4120 xt 6566", 519) writes: > > 1 . I am trying to compile some fortran code with optimisation. At O2 , > several routines ( this program has several hundred ) hit the error > not optimised beacuase global tables are not large enough, use > Olimit > nnn where nnn is some number larger than 500. The question > is , does it hurt the performance of the program , or the program > size to compile with Olimit nnn,where nnn is greater than the largest > reported size required, for all the routines , whether they need the > space or not ?? The program is large enough as it is , and I am trying > to enhance its runtime execution speed as much as possible. The number specified in the -Olimit option is used to set up the internal table inside the optimizer. It will not adversely effect the runtime speed nor the size of your program (other than the fact that your program will be smaller and run faster if optimized). The only bad effect is that the bigger the number, the more memory space will be used by the optimizer, and your program may take longer to compile. > To that end I would also like someone from SGI to comment on the > advisability of using O3. I was told that at the moment this level of > optimisation is unsafe. Is that true , and what performance enhancement > is expected over O2 optimisation ?? The reason we don't recommend -O3 level is because it fails some of our internal QA tests. It provides about 10% improvement over -O2 level, depending on your application. > Also, for a very large program such > as this what are the recommended analysis tools on IRIX to help > optimise the performance(hopefully without source code changes ). I am > looking for tools which help restructure the binary(loaded) code to > enhance performance. > Sorry we don't have any tools that can do optimization on loaded binary code. You can use 'prof' and 'pixie' to figure out all the places in your program that may need to be changed for faster execution. > 2. On the user interface side , I see a need for easier command recall > than using !nnn to recall a command. Recently I worked on an Apollo > where command recall and editting was available through mouse > highlighting, cutting and pasting (ie like a mac). Can this be done > in IRIX now (3.1 x level)? will it be available at 3.2 ?? In > Workspace ?? > We have cut & paste in 3.2. > Regards, > Len Zaifman
davea@quasar.wpd.sgi.com (David B. Anderson) (08/26/89)
In article <41004@sgi.sgi.com> calvin@dinkum.wpd.sgi.com (Calvin H. Vu) writes: >In article <89Aug25.093608edt.57688@ugw.utcs.utoronto.ca>, LEONARDZ@UOGUELPH.BITNET ("Len Zaifman UoGuelph 824-4120 xt 6566", 519) writes: [stuff deleted] >> Also, for a very large program such >> as this what are the recommended analysis tools on IRIX to help >> optimise the performance(hopefully without source code changes ). I am >> looking for tools which help restructure the binary(loaded) code to >> enhance performance. >> > Sorry we don't have any tools that can do optimization on loaded > binary code. [stuff deleted] Calvin is quite correct about release 3.1. However for 3.2 we support cord(1), which restructures a binary. When you get 3.2 see the cord(1), ftoc(1), prof(1), pixie(1), and cc(1) man pages for further information. [ David B. Anderson Silicon Graphics (415)964-1459x3060 davea@sgi.com ]
khb@road.Sun.COM (Keith Bierman - Advanced Languages - Floating Point Group ) (08/26/89)
In article <89Aug25.093608edt.57688@ugw.utcs.utoronto.ca> LEONARDZ@UOGUELPH.BITNET ("Len Zaifman UoGuelph 824-4120 xt 6566", 519) writes: > > I need help on two issues. > > 1 . I am trying to compile some fortran code with optimisation. At O2 , > several routines ( this program has several hundred ) hit .... stuff about table sizes It should be easy enough to try the experiment of compiling a few modules with and without the larger tables. I've never noticed a significant difference, but not working for SGI, I can't promise you that. > To that end I would also like someone from SGI to comment on the > advisability of using O3. I was told that at the moment this level of > optimisation is unsafe. Is that true , and what performance Mileage varies a lot for all higher level optimizers. I strongly suggest you profile your code, and turn the optimizer on high on just the top routines (probably 5 or 10). Optimizers can pessimize code, so pay attention to the profiles as you go. I expect that you have both prof and gprof. prof is less intrusive. For really reducing the intrusiveness (at the cost of lower information content) turn on profiling ONLY on the "link" line, viz. f77 -O your.f -c f77 your.o -p Asking how much benefit O<n> gives over <n-1> is typically silly ... most of the time doesn't matter ... just what goes on in 10% of your code. > > 2. On the user interface side , I see a need for easier command recall At the very least ksh should be available (if only via the ATT toolchest source program, or via bash from RMS). I haven't played around with the SGI machines enough to comment of the cut and paste (sunview, and openlook allow it). Keith H. Bierman |*My thoughts are my own. !! kbierman@sun.com It's Not My Fault | MTS --Only my work belongs to Sun* I Voted for Bill & | Advanced Languages/Floating Point Group Opus | "When the going gets Weird .. the Weird turn PRO"
ijlustig@phoenix.Princeton.EDU (Irvin Lustig) (08/28/89)
Regarding -O3 level optimization, I have a large Fortran source code that works under -O2, but fails miserably under -O3. Who knows what triggers the failure. -Irv Lustig Assistant Professor Dept. of Civil Engineering and Operations Research Princeton University irv%basie@princeton.edu