[MITgcm-devel] testreport for AD

[Patrick Heimbach]

Hi Ed,

here are some suggestions for AD-related testreport.

1. General
##########
    I would distinguish between testing the adjoint
    and testing the tangent linear; thus have an option
    testreport (-adm|-tlm)
    We could actually have the default to do both
    with the option -ad (for 'automatic differentiation').

2. Build processes:
###################
    (assuming either access to taf or staf is available;
    we may think about whether we'd also want to test
    compatibility w.r.t. tamc)

    >>> -adm <<<
genmake2 -mods=../code_ad
make depend
make adall

    >>> -tlm <<<
genmake2 -mods=../code_ad
make depend
make ftlall

    >>> both <<<
genmake2 -mods=../code_ad
make depend
make adall
make ftlall

3. Running:
###########
    To each directory code_ad(_extension)
    there is an input_ad(_extension)
    and a result_ad(_extension)
    Note that the input_ad(_extension) does not usually
    contains the input fields themselves (they don't change)
    in order not to blow up the size of the download.
    Also, each result_ad(_extension) now contains
    * output.txt_adm
    * output.txt_tlm
    adm and tlm use the exact same inputs

./mitgcmuv_ad >&! output.txt_adm
./mitgcmuv_ftl >&! output.txt_tlm

4. Comparing:
############
    Grep for

grep precision_grdchk_result output.txt_(adm|tlm)

    In each case you'll get a set similar to
(PID.TID 0000.0001)       precision_grdchk_result ADM     1.84044933456284E+14     1.74865283966910E+08
(PID.TID 0000.0001)       precision_grdchk_result ADM     1.84044933456284E+14     1.31884844820653E+08
(PID.TID 0000.0001)       precision_grdchk_result ADM     1.84044933456284E+14     1.96420903150567E+07
(PID.TID 0000.0001)       precision_grdchk_result ADM     1.84044933456284E+14    -9.99869528291716E+06
    or
(PID.TID 0000.0001)       precision_grdchk_result TLM     1.84044933456284E+14     1.74865283966909E+08
(PID.TID 0000.0001)       precision_grdchk_result TLM     1.84044933456284E+14     1.31884844820653E+08
(PID.TID 0000.0001)       precision_grdchk_result TLM     1.84044933456284E+14     1.96420903150604E+07
(PID.TID 0000.0001)       precision_grdchk_result TLM     1.84044933456284E+14    -9.99869528291790E+06

    The precision/format is the same as the one for cg2d_init_res
    The second-last term  in each row is the cost function, the last one is the AD-derived gradient
    (might think about adding the finite differnce gradient as well).
    PASS/FAIL should be based on accuracy of last column.
    In addition info on accuracy of the cost function can be given.

That's it I guess. Le me know if you have any questions.
Cheers
-Patrick

-- 
Patrick Heimbach ........................... M.I.T
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