home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.unix.ultrix
- Path: sparky!uunet!charon.amdahl.com!pacbell.com!ames!haven.umd.edu!darwin.sura.net!zaphod.mps.ohio-state.edu!cs.utexas.edu!uwm.edu!ux1.cso.uiuc.edu!cs.uiuc.edu!vela!amaranth
- From: amaranth@vela.acs.oakland.edu (Paul Amaranth)
- Subject: Re: FPU errors on 5000/240 P.Amaranth
- Message-ID: <1992Nov18.222729.24477@vela.acs.oakland.edu>
- Organization: Oakland University, Rochester MI.
- References: <1992Nov17.195628.15041@vela.acs.oakland.edu> <1eeaffINN3tj@nestroy.wu-wien.ac.at> <1eebumINN51i@nestroy.wu-wien.ac.at>
- Date: Wed, 18 Nov 1992 22:27:29 GMT
- Lines: 54
-
- >I am sorry for contradicting. These things are only correct for the right side of the " = " , and only to bracket expressions.
- >Change
- >ainv=1.0/i
- >to
- >ainv=1.0d0/i or ainv=1.0/dble(i)
- >in your program and you will get a different result!
- >namely -86354540.28378928 .
- >on a DEC with FPU - bug AND on a DEC 5000/240 without FPU - bug.
- >
-
- The first expression is evaluated in single precision, then widened to
- double for the assignment, the second is done in double, maintaining
- precision, so of course you will get a different value.
-
- Only 1 of my 4 unfixed 240s was sensitive to this program anyway, so
- getting the same different answer running a different program doesn't
- seem to mean much.
-
- If DEC would send me the tech info on what is wrong, maybe I could look
- at the code generated and figure out why that 10 line program shows a
- sensitivity to the FPU problem. Yeah, right. Like I needed this FPU
- problem in the first place. 'Course, if I knew what was going on, maybe
- I could tell my users they didn't have to rerun that last three months of
- work. Might help defuse the 'Kill the Messenger' syndrome.
-
- OK, so here's the bottom line: that EXACT 10 line program, when run on
- 5000/240s showed a problem on exactly ONE of my four unfixed 240s, all
- of which showed serious problems with the Ponder program. The Smith
- program failed on the same 240, and worked correctly on the other
- three. The Ponder program seems to be the definitive test to show if
- you have a problem. Changing various things in the test program would
- get it to work on the same 240. It was the minimal pathological case
- that caused it to fail modeled on operations ocurring in the Smith
- program. Why would anyone want to fix it?
-
- You think its lousy Fortran? I agree, but that wasn't the object of the
- exercise. The object was to get a short program that failed in the
- same circumstances as the Smith program. I did that, thought it was of
- sufficient interest to post it and have now spent more time explaining
- Fortran semantics that I spent writing the program.
-
- Since I'm taking up bandwith, did anyone else see that program fail? Or
- is it just my one wierdo 240?
-
- I'm getting tired of Fortran semantics, anyone have any idea of exactly
- what's happening in the FPU? What sequence of operations cause the error?
- Something I can look for in the object and determine if my users had
- a problem or not?
-
- --
- Paul Amaranth Manager User Services - office: (313) 370 4541 (also voicemail)
- (internet) amaranth@vela.acs.oakland.edu | This space
- (bitnet) amaranth@oakland | temporarily
- (uucp) ...!uunet!umich!vela!amaranth | empty
-