(see http://icl.cs.utk.edu/lapack-forum/viewtopic.php?f=13&t=2893)
This is related to the LAPACK-XBLAS routine: zherfsx.f
Here is what I did:
- Introduce IINFO so that INFO is not overwritten
- Use IGNORE_CWISE as suggested to prevent use of unitialize variable PARAMS
But this did not fix the problem reported.
INFO has new value in ZHERFSX (see description of INFO between ZHESVX and ZHESVXX)
This is set on line 634 (IF ( INFO .LE. N ) INFO = N + J) of zherfsx.f
And this is not handled by the testing LIN/zdrvhex.f
I just add .AND. INFO.LE.N at line 638 to avoid raising an error when INFO = N + J
Please send feedback as I am not sure this is the best way to fix the issue.
I will commit other precision once fix approved.
Thanks
Julie
dimension 2*NMAX to accomodate the calls to CHESVXX in CDRVHEX, ZHESVXX in
ZDRVHEX, CSYSVXX in CDRVSYX, and ZSYSVXX in ZDRVSYX. (The dimension of RWORK in
the caller subroutines --, e.g., ZCHKAA -- is way large enough.)
Use \par instead of \details for section.
add a Contributors Section and a Reference Section.
Remove (some) verbatim section when not needed.
Those changes have been done by hand so I am not sure I manage to catch them all.
The following variables will control the precision to be built:
BUILD_SINGLE
BUILD_DOUBLE
BUILD_COMPLEX
BUILD_COMPLEX16
For mixed precision SINGLE/DOUBLE routines, both BUILD_SINGLE and BUILD_DOUBLE needs to be on.
(same for COMPLEX/COMPLEX16)
Patch was based on 3.3.0, so required a little bit of merging.
Thank you Sven for the corrections.
Julie
JL
TESTING/EIG
zerrgg.f
Declare ILO, IHI
SRC
---
I have corrected all the following warnings and errors:
c/d/s/zsysv.f, Unused external reference ILAENV
>JL OK
c/d/s/zsysv.f, Unused local variable NB
>JL ALREADY CORRECTED
sgsvj0.f, Unused intrinsic AMIN1
> JL OK
d/sorbdb.f, Unused intrinsic MIN
> JL OK
sorcsd.f, Unused intrinsic SIN
sorcsd.f, Unused intrinsic COS
> JL OK MERGE
cpoequb.f, Unused intrinsic AIMAG
cpoequb.f, Unused intrinsic REAL
cpoequb.f, Unused local variable ZDUM
zgeequb.f, REAL --> DBLE
> JL OK
sorcsd.f, line 330: Inconsistent data type REAL (previously INTEGER) for argument 14 in reference to SORBDB
sorcsd.f, line 340: Inconsistent data type REAL (previously INTEGER) for argument 4 in reference to SORGQR
sorcsd.f, line 345: Inconsistent data type REAL (previously INTEGER) for argument 4 in reference to SORGQR
sorcsd.f, line 356: Inconsistent data type REAL (previously INTEGER) for argument 4 in reference to SORGLQ
sorcsd.f, line 363: Inconsistent data type REAL (previously INTEGER) for argument 4 in reference to SORGLQ
sorcsd.f, line 369: Inconsistent data type REAL (previously INTEGER) for argument 4 in reference to SORGLQ
sorcsd.f, line 374: Inconsistent data type REAL (previously INTEGER) for argument 4 in reference to SORGLQ
sorcsd.f, line 385: Inconsistent data type REAL (previously INTEGER) for argument 4 in reference to SORGQR
sorcsd.f, line 392: Inconsistent data type REAL (previously INTEGER) for argument 4 in reference to SORGQR
sorcsd.f, line 399: Inconsistent data type REAL (previously INTEGER) for argument 9 in reference to SBBCSD
> JL OK MERGE
ila(s/d/c/z)lr.f
In the WHILE loop, at about line 59 I inserted
IF (I.EQ.0) THEN
EXIT
END IF
since, otherwise when I = 0, A(0,J) is referenced.
> JL OK
TESTING/MATGEN
--------------
s/dlatm7.f
line 187: I replaced
IF( N.GT.1 ) THEN
by
IF( N.GT.1 .AND. RANK.GT.1 ) THEN
Otherwise I get a division by zero.
> JL OK
s/dlaror.f
I moved:
INFO = 0
from line 125 to line 121 in front of
IF( N.EQ.0 .OR. M.EQ.0 )
$ RETURN
otherwise INFO is not defined on return when N or M are zero.
Similar change for
c/zlaror.f
> JL OK
TESTING/LIN
-----------
s/derrsy.f
In calls to S/DSYTRI2: I replaced IW by IW(1) since the dummy argument
(LWORK) is a scalar.
> JL OK
c/zerrrfp.f
I made ALPHA and BETA REAL/DOUBLE and introduced COMPLEX/COMPLEX*16
CALPHA since calls to C/ZHFRK need real alpha and beta. In calls to
CTFSM, I then replaced ALPHA with CALPHA.
In cerrrfp.f, I replaced 1.0D0 with 1.0E0 (not strictly necessary, but
cleaner).
> JL OK
dpot06.f
Removed the unused declaration of LSAME
> JL OK
d/zdrvac.f
Removed the unused declaration of LSAME
> JL OK MERGE
s/d/c/z/ckcsd.f
Removed the unused declarations of S/D/C/ZLANGE, SIN and COS
> JL OK
c/zgennd.f
Removed the unused variable OUT
> JL already DONE
Corrected the following warnings and errors:
dchksy.f, Unused local variable MYWORK
> JL already DONE
ddrvgbx.f, line 792: Different number of arguments from the first call of DGBT02
> JL OK
cdrvgbx.f, line 726: Inconsistent data type REAL (previously COMPLEX) for argument 4 in reference to CLASET
cdrvgbx.f, line 728: Inconsistent data type REAL (previously COMPLEX) for argument 4 in reference to CLASET
cdrvgbx.f, line 780: Inconsistent data type REAL (previously COMPLEX) for argument 10 in reference to CGBT01
cdrvgbx.f, line 794: Different number of arguments from the first call of CGBT02
> JL OK
zdrvgbx.f, line 726: Inconsistent data type DOUBLE PRECISION (previously COMPLEX(KIND(0d0))) for argument 4 in reference to ZLASET
zdrvgbx.f, line 728: Inconsistent data type DOUBLE PRECISION (previously COMPLEX(KIND(0d0))) for argument 4 in reference to ZLASET
zdrvgbx.f, line 780: Inconsistent data type DOUBLE PRECISION (previously COMPLEX(KIND(0d0))) for argument 10 in reference to ZGBT01
zdrvgbx.f, line 794: Different number of arguments from the first call of ZGBT02
> JL OK
TESTING/EIG
-----------
s/d/c/zchkee.f
I inserted
CALL XLAENV( 12, 1 )
at line 1208, since IPARMS(ISPEC) with ISPEC = 12 needs to be set.
(I think I had to do this with them all, certainly with zchkee.f)
> JL OK zchkee.f cchkee.f
c/zchkee.f
In line 2316, the 15th argument of C/ZCKCSD should be real, not complex,
so I replaced A(1,7) by RWORK.
> JL OK
c/zcsdts.f
In calls to C/ZHERK I have replaced ONE by REALONE as arguments 5 and 8
should be real.
> JL OK
s/d/c/zcsdts.f
Division by zero occurs at line 155 in computing EPS when M = 0, so I
have modified
EPS2 = MAX( ULP, CLANGE( '1', M, M, WORK, LDX, RWORK ) / REAL( M ) )
to
IF (M.GT.0) THEN
EPS2 = MAX( ULP,
$ CLANGE( '1', M, M, WORK, LDX, RWORK ) / REAL( M ) )
ELSE
EPS2 = ULP
END IF
> JL OK
SRC/VARIANTS/lu/CR
------------------
s/d/c/zgetrf
Removed the unused intrinsic MOD
> JL OK
"We notice that errorcheckers tests aren’t correct in some cases.
errorcheckers of ?ggevx, ?pstrf, ?pstf2
Some output parameters are read only in the errorcheckers tests."
reported by omitrofa (olga.mitrofanova@intel.com)
See Forum 2644
Test of CHERFSX function doesn't work because there is an error
in ../TESTING/LIN/cerrhex.f ( lines 284-330 ).
This function uses the diagonal pivoting factorization of a Hermitian
indefinite matrix instead of the diagonal pivoting factorization of
a Hermitian indefinite packed matrix. When CHERFSX is included in
conditional "IF( LSAMEN( 2, C2, 'HE' ) ) THEN " the test works.
Below is the help that goes with it.
Please send feedback. I tired to include several levels of output and several kind of errors
So far, I just tested it on my machine, that has no major problems.
I just tested "numerical errors", not "illegal error" neither "info error".
The goal is to have our nightly testing use it and also users.
TO DO:
- Test other kinds of errors on machines
- See to include it in ctest and buildbot (need to contact Don and Brad/Chuck)
- See if a type of error option is required.
./lapack_testing.py [-h|--help] [-s |--short] [-f |--file] [-e |--error] [-p p |--prec p] [-t test |--test test] [-n | --number]
- h is to print this message
- f is to use directly the output of the LAPACK testing (.out files). By default, the script will run all the LAPACK tests
LEVEL OF OUTPUT
- x is to print a detailed summary
- e is to print only the error summary
- s is to print a short summary
- n is to print the numbers of failing tests (turn on summary mode)
SECLECTION OF TESTS:
- p [s/c/d/z] is to indicate the PRECISION to run:
s=single
d=double
c=complex
z=double complex
x=all [DEFAULT]
- t [lin/eig/mixed/rfp/all] is to indicate which TEST FAMILY to run:
lin=Linear Equation
eig=Eigen Problems
mixed=mixed-precision
rfp=rfp format
all=all tests [DEFAULT]
EXAMPLES:
./lapack_testing.py -n -f
Will return the numbers of failed tests by analyzing the LAPACK output
./lapack_testing.py -n -f -p s
Will return the numbers of failed tests in REAL precision by analyzing the LAPACK output
./lapack_testing.py -n -f -p s -t eig
Will return the numbers of failed tests in REAL precision by analyzing only the LAPACK output of EIGEN testings
Written by Julie Langou (June 2011)
Test the case where N=M (Workspace should be 1), so put LWORK at 0 to make it failed
Test the case where N!=M (Workspace should be more than MAX(1,M)), so put M=2, N=3 ans LWORK at 1 to make it failed
Correct some typos in [dz]drgsx testings.
igor175