[Mitgcm-support] [Fwd: RE: [Fwd: Re: GM and tamc (and not taf)]]

[mitgcm-support at dev.mitgcm.org]

FYI:
Adding to my last e-mail to ecco-all
regarding TAF.

-Patrick



-----Original Message-----
From: Patrick Heimbach [mailto:heimbach at MIT.EDU]
Sent: Friday, January 03, 2003 12:29 PM
To: cwunsch; John Marshall; Chris Hill; Detlef B. Stammer
Subject: [Fwd: Re: GM and tamc (and not taf)]


FYI:
Ralf found the problem with TAF and will
install a new TAF version.

-Patrick

-------- Original Message --------
Subject: Re: GM and tamc (and not taf)
Date: Fri, 3 Jan 2003 18:08:50 +0100
From: Ralf Giering <Ralf.Giering at fastopt.de>
Organization: FastOpt
To: heimbach at MIT.EDU
CC: Thomas Kaminski <Thomas.Kaminski at fastopt.de>
References: <3E146EA6.3070602 at mit.edu>

Hi Patrick,

It turns out that the current TAF version (1.4.19) also generates
correct code. Do you remember that a used taf-1.4.17 as you did because
I found a difference to the 1.4.19 generated code. That difference
actually matters! It is in the adjoint of the implicit
solver.
The fix was from version 1.4.17 to 1.4.18 because of incorrect adjoint
code of
a totally different model.

We missed the bug because the GM schema is not in the MIT model code we
use to
quality control a new TAF version. We will add the c47 code to our
quality control code set.

So I will finish TAF version 1.4.20 and install it at MIT.

Ralf


On Thursday 02 January 2003 05:53 pm, you wrote:
  > Hi Ralf,
  >
  > here's some news re. the GM problem that should
  > greatly help you in finding the problem.
  >
  > I've generated the adjoint in the GM setup using
  > TAMC instead of TAF, and - siehe da - it works OK,
  > finite difference gradient checks look reliable,
  > and tangent linear vs. adjoint gradients are identical.
  >
  > I had never thought about doing this because I was under
  > the impression that TAF is more reliable these days,
  > but apparently TAMC still seems to be very useful now
  > for performing consistency checks.
  >
  > I ran these tests on slough.mit.edu, see
  > /data37/heimbach/ecco/c47_for_ralf_2/
  > Both adjoint codes are in adjoint/
  > tamc_code_ad.f_taf
  > tamc_code_ad.f_tamc
  > The outputs are in exe/
  > run.out_taf
  > run.out_tamc
  >
  > (you can "grep ph-grd run.out_ta*")
  >
  > This, of course, raises the issue (again), of whether
  > it would be possible to make the tamc vs. taf "output",
  > i.e. code, more similar to be able to detect possible
  > problems more easily.
  > Right now, minor changes in the output format make it
  > very hard to diff TAF- and TAMC-generated code.
  >
  > Cheers
  > -Patrick



-- 
_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/
Patrick Heimbach     Massachusetts Institute of Technology
FON: +1/617/253-5259                    EAPS, Room 54-1518
FAX: +1/617/253-4464               77 Massachusetts Avenue
mailto:heimbach at mit.edu                 Cambridge MA 02139
http://www.mit.edu/~heimbach/                       U.S.A.