reflapack3 for Debian ---------------------- The lapack code itself has not been modified from the lapack source package, version 3.0.20000531a-25. However the blas library, against which lapack is linked, has had backwards binary incompatible corrections applied. Support for the old and new API in both source and binary form are provided to aid in the transition. This requires a change in package name for the lapack packages. Please see the documentation for the refblas3 package for details on the modifications to the cblas API. ============================================================================= original lapack README.Debian ============================================================================= lapack for Debian ---------------------- These libraries were built from the 20000531 Lapack release, as amended by applying all of the unincorporated updates listed in the release notes (e.g. files in the patch subdirectory of the lapack directory at netlib.) The new build process runs the tester binaries after building the libraries, and collects the results in the files /usr/share/doc/lapack/test_results. One can search for any failures with 'grep -i fail /usr/share/doc/lapack/test_results'. The failures in the ?GV and ?XV tester routines are known and expected. The following explanation can be found at http://www.netlib.org/lapack/html/installation.hints: The only known testing failures are in condition number estimation routines in the generalized nonsymmetric eigenproblem testing. Specifically in sgd.out, dgd.out, cgd.out and zgd.out. The cause for the failures of some test cases is that the mathematical algorithm used for estimating the condition numbers could over- or under-estimate the true values in a certain factor in some rare cases. Further details can be found in LAPACK Working Note 87. This version of lapack also introduces a few testing errors in the single precision real symmetric and complex Hermitian eigenvalue routines, at least on ix86 platforms. Correspondence regarding this issue with the lapack upstream authors follows at the end of this file. Suffice it to say that there is no known solution at present, but that the errors should be confined to loss of at most 2-3 decimal places of precision in the routine results. For this reason, Debian also supplies lapack libraries based on the earlier release in the lapack99 package, which, though having errors of its own, does not exhibit this particular failing. ============================================================================= From: Camm Maguire Subject: Re: New Debian maintainer To: Susan Blackford cc: camm@enhanced.com Date: 15 Nov 2001 15:33:11 -0500 Received: from intech19.enhanced.com [192.168.2.19] (mail) by intech5.enhanced.com with esmtp (Exim 3.12 #1 (Debian)) id 164TCL-0005qT-00; Thu, 15 Nov 2001 15:33:21 -0500 Received: from camm by intech19.enhanced.com with local (Exim 3.12 #1 (Debian)) id 164TCB-0002gL-00; Thu, 15 Nov 2001 15:33:11 -0500 References: <01111312052804.00517@idefix> In-Reply-To: Susan Blackford's message of "Tue, 13 Nov 2001 12:05:28 -0500" Message-ID: <547ksrdbaw.fsf@intech19.enhanced.com> X-Mailer: Gnus v5.7/Emacs 20.7 X-UIDL: 9358c79a10286a739f7134b4a7627355 Lines: 92 Xref: intech19.enhanced.com mail.camm:27060 Greetings, and thanks for your reply! Susan Blackford writes: > Hi, > > If the LIN timers (or any of the testers/timers) dies when loading, > then you should reduce the size of NMAX in the=20 > LAPACK/TESTING/LIN/*chkaa.f > LAPACK/TESTING/EIG/*chkee.f > LAPACK/TIMING/LIN/*timaa.f > LAPACK/TIMING/EIG/*timee.f > Thanks for this. I read a usenet post indicating that MAXVAL could be reduced to 7 or 8 in the ?timaa.f, and this seems to work too. Is either solution acceptable? Secondarily, I'm upgrading the package from the 1999 3.0 to the latest, plus the unincorporated patches listed on the release note pages. With the old version, I just get the expected _gd.out errors on test, but the new version has a few errors in the symmetric eigenroutine single and complex precision as well (_sep.out): 1999: ============================================================================= cgd.out: CGV drivers: 4 out of 1092 tests failed to pass the threshold cgd.out: 9 = 1/ULP if DIFEST <> 0 or DIFTRU > ULP*norm(A,B) when reordering fails cgd.out: CGX drivers: 6 out of 1500 tests failed to pass the threshold dgd.out: 9 = 1/ULP if DIFEST <> 0 or DIFTRU > ULP*norm(A,B) when reordering fails dgd.out: SGX drivers: 6 out of 1500 tests failed to pass the threshold dgd.out: DXV drivers: 200 out of 5000 tests failed to pass the threshold dtest.out: DLS drivers: 2250 out of 65268 tests failed to pass the threshold sgd.out: 9 = 1/ULP if DIFEST <> 0 or DIFTRU > ULP*norm(A,B) when reordering fails sgd.out: SGX drivers: 8 out of 1500 tests failed to pass the threshold sgd.out: SXV drivers: 37 out of 5000 tests failed to pass the threshold zgd.out: ZGV drivers: 4 out of 1092 tests failed to pass the threshold zgd.out: 9 = 1/ULP if DIFEST <> 0 or DIFTRU > ULP*norm(A,B) when reordering fails zgd.out: CGX drivers: 6 out of 1500 tests failed to pass the threshold zgd.out: ZXV drivers: 24 out of 5000 tests failed to pass the threshold latest: ============================================================================= cgd.out: CGV drivers: 4 out of 1092 tests failed to pass the threshold csep.out: CST: 1 out of 4662 tests failed to pass the threshold csep.out: CST drivers: 1 out of 11664 tests failed to pass the threshold csep.out: CST: 1 out of 4662 tests failed to pass the threshold dgd.out: DXV drivers: 200 out of 5000 tests failed to pass the threshold sgd.out: SXV drivers: 37 out of 5000 tests failed to pass the threshold ssep.out: SST: 1 out of 4662 tests failed to pass the threshold ssep.out: SST drivers: 1 out of 14256 tests failed to pass the threshold zgd.out: ZGV drivers: 4 out of 1092 tests failed to pass the threshold zgd.out: ZXV drivers: 24 out of 5000 tests failed to pass the threshold ============================================================================= I am somewhat hesitant to introduce these new errors into the official lib of the Debian distribution. I've seen these errors reported in the release_notes in similar setups to mine (g77 2.95, Linux, PIII). The latest lib was compiled with the suggested flags in the Linux make.inc in INSTALL (-funroll-all-loops -O3), whereas the older version was compiled with (-funroll-all-loops -O3 -fomit-frame-pointer -ff90). I can't imagine that the omission of optimizing flags could introduce new errors, but am willing to recompile with the old settings if you think this would help. Any advice appreciated! > > 3) Is this also the place to submit patches/suggestions? > > Yes, absolutely... :) > > Susan > > > --=20 > --------------------------------------------- > Susan Blackford > Member of Technical Staff > Myricom, Inc. > --------------------------------------------- > > -- Camm Maguire camm@enhanced.com ========================================================================== "The earth is but one country, and mankind its citizens." -- Baha'u'llah ============================================================================= From: Edward Anderson Subject: Re: New Debian maintainer To: camm@enhanced.com CC: Susan Blackford Date: Mon, 26 Nov 2001 09:51:48 -0500 Organization: Lockheed Martin (contractor to the U.S. EPA) Received: from localhost [127.0.0.1] (root) by intech5.enhanced.com with esmtp (Exim 3.12 #1 (Debian)) id 168N8I-0001aM-00; Mon, 26 Nov 2001 09:53:18 -0500 Received: from 204.213.252.11 by localhost with POP3 (fetchmail-5.3.3) for camm@enhanced.com (multi-drop); Mon, 26 Nov 2001 09:53:18 -0500 (EST) Received: from mail.eticomm.net [204.213.252.17] by ux1.cyberenet.net with esmtp (Exim 1.61 #1) id 168N7M-0002gE-00; Mon, 26 Nov 2001 09:52:20 -0500 Received: from [134.67.64.25] (helo=nesc.epa.gov) by relay.eticomm.net with esmtp (Exim 3.33 #2) id 168N7M-000CuP-00 for camm@enhanced.com; Mon, 26 Nov 2001 09:52:20 -0500 Received: from epa.gov (yew.nesc.epa.gov [134.67.64.48]) by nesc.epa.gov (8.9.3+Sun/8.9.1) with ESMTP id JAA17204; Mon, 26 Nov 2001 09:51:49 -0500 (EST) Sender: dle@nesc.epa.gov Message-ID: <3C025704.86683917@epa.gov> X-Mailer: Mozilla 4.75C-SGI [en] (X11; I; IRIX 6.5 IP32) X-Accept-Language: en MIME-Version: 1.0 References: <004701c17372$2a3580e0$2fee80ac@compaq> Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit X-UIDL: 11a1ca53c0909d8b833a35f7beea5f16 Lines: 167 Xref: intech19.enhanced.com mail.camm:27339 Camm: I have my own version of LAPACK which incorporates some changes not yet in the netlib version. It's hard to say which if any of these might fix the errors you are seeing. But if you could send me the full text of the output files containing tests that are failing, I may have a few suggestions for you. --Ed Anderson Ed Anderson wrote: > > ----- Original Message ----- > From: Camm Maguire > To: > Sent: Monday, November 19, 2001 5:05 PM > Subject: [Camm Maguire ] Re: New Debian maintainer > > > Susan-- > > > > Resending this to the lapack address, as I haven't heard back > > from my direct reply to you. Any advice most appreciated! > > > > Take care, > > -- > > Camm Maguire camm@enhanced.com > > ========================================================================== > > "The earth is but one country, and mankind its citizens." -- Baha'u'llah > > ------- Start of forwarded message ------- > > To: Susan Blackford > > cc: camm@enhanced.com > > Subject: Re: New Debian maintainer > > References: > <01111312052804.00517@idefix> > > From: Camm Maguire > > Date: 15 Nov 2001 15:33:11 -0500 > > Message-ID: <547ksrdbaw.fsf@intech19.enhanced.com> > > > > Greetings, and thanks for your reply! > > > > Susan Blackford writes: > > > > > Hi, > > > > > > If the LIN timers (or any of the testers/timers) dies when loading, > > > then you should reduce the size of NMAX in the=20 > > > LAPACK/TESTING/LIN/*chkaa.f > > > LAPACK/TESTING/EIG/*chkee.f > > > LAPACK/TIMING/LIN/*timaa.f > > > LAPACK/TIMING/EIG/*timee.f > > > > > > > Thanks for this. I read a usenet post indicating that MAXVAL could be > > reduced to 7 or 8 in the ?timaa.f, and this seems to work too. Is > > either solution acceptable? > > > > Secondarily, I'm upgrading the package from the 1999 3.0 to the > > latest, plus the unincorporated patches listed on the release note > > pages. > > > > With the old version, I just get the expected _gd.out errors on test, > > but the new version has a few errors in the symmetric eigenroutine > > single and complex precision as well (_sep.out): > > > > 1999: > > > ============================================================================ > = > > cgd.out: CGV drivers: 4 out of 1092 tests failed to pass the > threshold > > cgd.out: 9 = 1/ULP if DIFEST <> 0 or DIFTRU > ULP*norm(A,B) when > reordering fails > > cgd.out: CGX drivers: 6 out of 1500 tests failed to pass the > threshold > > dgd.out: 9 = 1/ULP if DIFEST <> 0 or DIFTRU > ULP*norm(A,B) when > reordering fails > > dgd.out: SGX drivers: 6 out of 1500 tests failed to pass the > threshold > > dgd.out: DXV drivers: 200 out of 5000 tests failed to pass the > threshold > > dtest.out: DLS drivers: 2250 out of 65268 tests failed to pass the > threshold > > sgd.out: 9 = 1/ULP if DIFEST <> 0 or DIFTRU > ULP*norm(A,B) when > reordering fails > > sgd.out: SGX drivers: 8 out of 1500 tests failed to pass the > threshold > > sgd.out: SXV drivers: 37 out of 5000 tests failed to pass the > threshold > > zgd.out: ZGV drivers: 4 out of 1092 tests failed to pass the > threshold > > zgd.out: 9 = 1/ULP if DIFEST <> 0 or DIFTRU > ULP*norm(A,B) when > reordering fails > > zgd.out: CGX drivers: 6 out of 1500 tests failed to pass the > threshold > > zgd.out: ZXV drivers: 24 out of 5000 tests failed to pass the > threshold > > > > latest: > > > ============================================================================ > = > > cgd.out: CGV drivers: 4 out of 1092 tests failed to pass the > threshold > > csep.out: CST: 1 out of 4662 tests failed to pass the threshold > > csep.out: CST drivers: 1 out of 11664 tests failed to pass the > threshold > > csep.out: CST: 1 out of 4662 tests failed to pass the threshold > > dgd.out: DXV drivers: 200 out of 5000 tests failed to pass the > threshold > > sgd.out: SXV drivers: 37 out of 5000 tests failed to pass the > threshold > > ssep.out: SST: 1 out of 4662 tests failed to pass the threshold > > ssep.out: SST drivers: 1 out of 14256 tests failed to pass the > threshold > > zgd.out: ZGV drivers: 4 out of 1092 tests failed to pass the > threshold > > zgd.out: ZXV drivers: 24 out of 5000 tests failed to pass the > threshold > > > ============================================================================ > = > > > > I am somewhat hesitant to introduce these new errors into the official > > lib of the Debian distribution. I've seen these errors reported in > > the release_notes in similar setups to mine (g77 2.95, Linux, PIII). > > The latest lib was compiled with the suggested flags in the Linux > > make.inc in INSTALL (-funroll-all-loops -O3), whereas the older > > version was compiled with (-funroll-all-loops -O3 -fomit-frame-pointer > > -ff90). I can't imagine that the omission of optimizing flags could > > introduce new errors, but am willing to recompile with the old > > settings if you think this would help. > > > > Any advice appreciated! > > > > > > > > > > 3) Is this also the place to submit patches/suggestions? > > > > > > Yes, absolutely... :) > > > > > > Susan > > > > > > > > > --=20 > > > --------------------------------------------- > > > Susan Blackford > > > Member of Technical Staff > > > Myricom, Inc. > > > --------------------------------------------- > > > > > > > > > > -- > > Camm Maguire camm@enhanced.com > > ========================================================================== > > "The earth is but one country, and mankind its citizens." -- Baha'u'llah > > > > ------- End of forwarded message ------- ---------------------------------------------------------------- Ed Anderson Scientific Applications Consultant Lockheed Martin Technology Services (contractor to the U.S. EPA) Phone: 919-541-0299 mailto:Anderson.Edward@epa.gov ============================================================================= From: Edward Anderson Subject: Re: New Debian maintainer To: Camm Maguire CC: Edward Anderson , Susan Blackford Date: Fri, 30 Nov 2001 10:30:36 -0500 Organization: Lockheed Martin (contractor to the U.S. EPA) Received: from localhost [127.0.0.1] (root) by intech5.enhanced.com with esmtp (Exim 3.12 #1 (Debian)) id 169peX-0001br-00; Fri, 30 Nov 2001 10:32:37 -0500 Received: from 204.213.252.11 by localhost with POP3 (fetchmail-5.3.3) for camm@enhanced.com (multi-drop); Fri, 30 Nov 2001 10:32:37 -0500 (EST) Received: from mail.eticomm.net [204.213.252.17] by ux1.cyberenet.net with esmtp (Exim 1.61 #1) id 169pd6-0002EM-00; Fri, 30 Nov 2001 10:31:08 -0500 Received: from linden.nesc.epa.gov ([134.67.64.25] helo=nesc.epa.gov) by relay.eticomm.net with esmtp (Exim 3.33 #2) id 169pd5-0003vM-00 for camm@enhanced.com; Fri, 30 Nov 2001 10:31:07 -0500 Received: from epa.gov (yew.nesc.epa.gov [134.67.64.48]) by nesc.epa.gov (8.9.3+Sun/8.9.1) with ESMTP id KAA14465; Fri, 30 Nov 2001 10:30:36 -0500 (EST) Sender: dle@nesc.epa.gov Message-ID: <3C07A61C.57A40172@epa.gov> X-Mailer: Mozilla 4.75C-SGI [en] (X11; I; IRIX 6.5 IP32) X-Accept-Language: en MIME-Version: 1.0 References: <004701c17372$2a3580e0$2fee80ac@compaq> <3C025704.86683917@epa.gov> <544rndhvce.fsf@intech19.enhanced.com> Content-Type: multipart/mixed; boundary="------------6C386FC925298A0C59F101F0" X-UIDL: 7119754484cd74abeaed169b5c841fe4 Lines: 770 Xref: intech19.enhanced.com mail.camm:27462 This is a multi-part message in MIME format. --------------6C386FC925298A0C59F101F0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Camm Maguire wrote: > > Greetings! Sorry to bother you again -- just wondering if you got my > last reply regarding the netlib blas being used, and might have any > suggestions for things I could try. > > Take care, and thanks again, > -- > Camm Maguire camm@enhanced.com > ========================================================================== > "The earth is but one country, and mankind its citizens." -- Baha'u'llah Camm, Sorry for the slow response. I've looked at your failing tests, and the test ratios are all relatively small. This usually means a loss of orthogonality or improper scaling that causes underflow, which most often occurs in a situation where a sum of squares is required. Besides the obvious cases where you are taking a 2-norm of a vector or the Frobenius norm of a matrix, it can also occur any time you divide by a complex number. LAPACK has auxiliary routines for doing a scaled complex divide (CLADIV) and for taking the square root of a sum of squares of two or three numbers (SLAPY2 and SLAPY3), but scaling errors keep cropping up when new software is introduced, and even these auxiliary routines aren't as safe as they should be. The biggest offenders are the 2-norm routines SNRM2, SCNRM2, DNRM2, and DZNRM2. The original netlib routine had some difficult to follow code (known as "Blue's algorithm"); unfortunately it was so hard to understand that most vendors implemented their own version, and got the scaling wrong. A few years ago, the netlib version was replaced with one based on LAPACK's SLASSQ, and they got the scaling wrong too. You can find some more description of these problems in my online technical report, "Installing LAPACK 3 on CRAY machines", http://www.cs.utk.edu/~eanderso/lapack3.html. I am working on an updated version more suitable to IEEE arithmetic machines with gradual underflow, using the IBM SP at my site as a test bed. To see if the 2-norms may be at fault, I will attach prototypes of my latest xLASSQ versions and associated BLAS that I have tested on an IBM SP. I would like to hard code all the scaling constants in xLASSQ for speed but compute them in this version. If this doesn't help, then you may need a larger set of changes, either from a future LAPACK release or my IBM SP patch when it is ready. --Ed Anderson ---------------------------------------------------------------- Ed Anderson Scientific Applications Consultant Lockheed Martin Technology Services (contractor to the U.S. EPA) Phone: 919-541-0299 mailto:Anderson.Edward@epa.gov --------------6C386FC925298A0C59F101F0 Content-Type: text/plain; charset=us-ascii; name="tocamm" Content-Transfer-Encoding: 7bit Content-Disposition: inline; filename="tocamm" SUBROUTINE SLASSQ( N, X, INCX, SCL, SUMSQ ) ! ! -- LAPACK auxiliary routine (version 2.0) -- ! Univ. of Tennessee, Univ. of California Berkeley, NAG Ltd., ! Courant Institute, Argonne National Lab, and Rice University ! October 31, 1992 ! 10-15-99: New version based on LAWN 126 (eca) ! 08-31-01: Use smaller constants for 32-bit IEEE (eca) ! 09-25-01: Don't scale by < SAFMIN or more than SAFMAX (eca) ! ! .. Scalar Arguments .. INTEGER INCX, N REAL SCL, SUMSQ ! .. ! .. Array Arguments .. REAL X( * ) ! .. ! ! Purpose ! ======= ! ! SLASSQ returns the values scl and smsq such that ! ! ( scl**2 )*smsq = x( 1 )**2 +...+ x( n )**2 + ( scale**2 )*sumsq, ! ! where x( i ) = X( 1 + ( i - 1 )*INCX ). The value of sumsq is ! assumed to be non-negative and scl returns the value ! ! scl = max( scale, abs( x( i ) ) ). ! ! scale and sumsq must be supplied in SCL and SUMSQ and ! scl and smsq are overwritten on SCL and SUMSQ respectively. ! ! Arguments ! ========= ! ! N (input) INTEGER ! The number of elements to be used from the vector X. ! ! X (input) REAL array, dimension (N) ! The vector for which a scaled sum of squares is computed. ! x( i ) = X( 1 + ( i - 1 )*INCX ), 1 <= i <= n. ! ! INCX (input) INTEGER ! The increment between successive values of the vector X. ! INCX > 0. ! ! SCL (input/output) REAL ! On entry, the value scale in the equation above. ! On exit, SCL is overwritten with scl , the scaling factor ! for the sum of squares. ! ! SUMSQ (input/output) REAL ! On entry, the value sumsq in the equation above. ! On exit, SUMSQ is overwritten with smsq , the basic sum of ! squares from which scl has been factored out. ! ! ===================================================================== ! REAL ONE, ZERO PARAMETER ( ONE = 1.0E0, ZERO = 0.0E0 ) INTEGER I, IX, IX2 REAL CUTHI, CUTLO, HITEST, SAFMAX, SAFMIN, SMAX, SQMAX INTRINSIC ABS, EPSILON, MAX, REAL, TINY DATA CUTLO / 1.0E-15 / DATA CUTHI / 1.0E+19 / IF( N.LE.0 ) RETURN HITEST = CUTHI / REAL( N+1 ) !eca SAFMIN = SLAMCH( 'Safe minimum' ) / SLAMCH( 'Epsilon' ) SAFMIN = TINY( ONE ) / EPSILON( ONE ) SAFMAX = ONE / SAFMIN IF( SUMSQ.EQ.ZERO ) SCL = ONE IF( INCX.EQ.1 ) THEN ! ! Pass through once to find the maximum value in X. ! SMAX = ZERO DO I = 1, N SMAX = MAX( SMAX, ABS(X(I)) ) END DO SQMAX = MAX( SCL*SQRT( SUMSQ ), SMAX ) ! IF( SCL.EQ.ONE .AND. SQMAX.GT.CUTLO .AND. SQMAX.LT.HITEST ) & THEN ! ! If SCL = ONE and SQMAX is greater than CUTLO and less than ! HITEST, no scaling should be needed. ! DO I = 1, N SUMSQ = SUMSQ + X(I)**2 END DO ELSE IF( SMAX.GT.ZERO ) THEN ! ! Scale by SQMAX if SCL = ONE, otherwise scale by ! max( SQMAX, SCL ). ! SQMAX = MIN( MAX( SQMAX, SAFMIN ), SAFMAX ) IF( SCL.EQ.ONE .OR. SCL.LT.SQMAX ) THEN SUMSQ = ( SUMSQ*( SCL / SQMAX ) )*( SCL / SQMAX ) SCL = SQMAX END IF ! ! Add the sum of squares of values of X scaled by SCL. ! DO I = 1, N SUMSQ = SUMSQ + ( X(I) / SCL )**2 END DO END IF ELSE ! ! Pass through once to find the maximum value in X. ! SMAX = ZERO IX = 1 IF( INCX.LT.0 ) IX = 1 - (N-1)*INCX DO I = 1, N SMAX = MAX( SMAX, ABS(X(IX)) ) IX = IX + INCX END DO SQMAX = MAX( SCL*SQRT( SUMSQ ), SMAX ) ! IF( SCL.EQ.ONE .AND. SQMAX.GT.CUTLO .AND. SQMAX.LT.HITEST ) & THEN ! ! If SCL = ONE and SQMAX is greater than CUTLO and less than ! HITEST, no scaling should be needed. ! IX = 1 IF( INCX.LT.0 ) IX = 1 - (N-1)*INCX DO I = 1, N SUMSQ = SUMSQ + X(IX)**2 IX = IX + INCX END DO ELSE IF( SMAX.GT.ZERO ) THEN ! ! Scale by SQMAX if SCL = ONE, otherwise scale by ! max( SQMAX, SCL ). ! SQMAX = MIN( MAX( SQMAX, SAFMIN ), SAFMAX ) IF( SCL.EQ.ONE .OR. SCL.LT.SQMAX ) THEN SUMSQ = ( SUMSQ*( SCL / SQMAX ) )*( SCL / SQMAX ) SCL = SQMAX END IF ! ! Add the sum of squares of values of X scaled by SCL. ! IX = 1 IF( INCX.LT.0 ) IX = 1 - (N-1)*INCX DO I = 1, N SUMSQ = SUMSQ + ( X(IX) / SCL )**2 IX = IX + INCX END DO END IF END IF RETURN END SUBROUTINE CLASSQ( N, X, INCX, SCL, SUMSQ ) ! ! -- LAPACK auxiliary routine (version 2.0) -- ! Univ. of Tennessee, Univ. of California Berkeley, NAG Ltd., ! Courant Institute, Argonne National Lab, and Rice University ! October 31, 1992 ! 10-15-99: New version based on LAWN 126 (eca) ! 08-31-01: Use smaller constants for 32-bit IEEE (eca) ! 09-25-01: Don't scale by < SAFMIN or more than SAFMAX (eca) ! ! .. Scalar Arguments .. INTEGER INCX, N REAL SCL, SUMSQ ! .. ! .. Array Arguments .. COMPLEX X( * ) ! .. ! ! Purpose ! ======= ! ! CLASSQ returns the values scl and ssq such that ! ! ( scl**2 )*ssq = x( 1 )**2 +...+ x( n )**2 + ( scale**2 )*sumsq, ! ! where x( i ) = abs( X( 1 + ( i - 1 )*INCX ) ). The value of sumsq is ! assumed to be non-negative and scl returns the value ! ! scl = max( scale, abs( real( x( i ) ) ), abs( aimag( x( i ) ) ) ), ! i ! ! scale and sumsq must be supplied in SCL and SUMSQ respectively. ! SCL and SUMSQ are overwritten by scl and ssq respectively. ! ! Arguments ! ========= ! ! N (input) INTEGER ! The number of elements to be used from the vector X. ! ! X (input) COMPLEX array, dimension (N) ! The vector x as described above. ! x( i ) = X( 1 + ( i - 1 )*INCX ), 1 <= i <= n. ! ! INCX (input) INTEGER ! The increment between successive values of the vector X. ! INCX > 0. ! ! SCL (input/output) REAL ! On entry, the value scale in the equation above. ! On exit, SCALE is overwritten with the value scl . ! ! SUMSQ (input/output) REAL ! On entry, the value sumsq in the equation above. ! On exit, SUMSQ is overwritten with the value ssq . ! ! ===================================================================== ! REAL ONE, ZERO PARAMETER ( ONE = 1.0E0, ZERO = 0.0E0 ) INTEGER I, IX, IX2 REAL CUTHI, CUTLO, HITEST, SAFMAX, SAFMIN, SMAX, SQMAX INTRINSIC ABS, EPSILON, MAX, REAL, TINY DATA CUTLO / 1.0E-15 / DATA CUTHI / 1.0E+19 / IF( N.LE.0 ) RETURN HITEST = CUTHI / REAL( 2*N+1 ) !eca SAFMIN = SLAMCH( 'Safe minimum' ) / SLAMCH( 'Epsilon' ) SAFMIN = TINY( ONE ) / EPSILON( ONE ) SAFMAX = ONE / SAFMIN IF( SUMSQ.EQ.ZERO ) SCL = ONE IF( INCX.EQ.1 ) THEN ! ! Pass through once to find the maximum value in X. ! SMAX = ZERO DO I = 1, N SMAX = MAX( SMAX, ABS(REAL(X(I))), ABS(AIMAG(X(I))) ) END DO SQMAX = MAX( SCL*SQRT( SUMSQ ), SMAX ) ! IF( SCL.EQ.ONE .AND. SQMAX.GT.CUTLO .AND. SQMAX.LT.HITEST ) & THEN ! ! If SCL = ONE and SQMAX is greater than CUTLO and less than ! HITEST, no scaling should be needed. ! DO I = 1, N SUMSQ = SUMSQ + REAL(X(I))**2 + AIMAG(X(I))**2 END DO ELSE IF( SMAX.GT.ZERO ) THEN ! ! Scale by SQMAX if SCL = ONE, otherwise scale by ! max( SQMAX, SCL ). ! SQMAX = MIN( MAX( SQMAX, SAFMIN ), SAFMAX ) IF( SCL.EQ.ONE .OR. SCL.LT.SQMAX ) THEN SUMSQ = ( SUMSQ*( SCL / SQMAX ) )*( SCL / SQMAX ) SCL = SQMAX END IF ! ! Add the sum of squares of values of X scaled by SCL. ! DO I = 1, N SUMSQ = SUMSQ + ( REAL(X(I)) / SCL )**2 + & ( AIMAG(X(I)) / SCL )**2 END DO END IF ELSE ! ! Pass through once to find the maximum value in X. ! SMAX = ZERO IX = 1 IF( INCX.LT.0 ) IX = 1 - (N-1)*INCX DO I = 1, N SMAX = MAX( SMAX, ABS(REAL(X(IX))), ABS(AIMAG(X(IX))) ) IX = IX + INCX END DO SQMAX = MAX( SCL*SQRT( SUMSQ ), SMAX ) ! IF( SCL.EQ.ONE .AND. SQMAX.GT.CUTLO .AND. SQMAX.LT.HITEST ) & THEN ! ! If SCL = ONE and SQMAX is greater than CUTLO and less than ! HITEST, no scaling should be needed. ! IX = 1 IF( INCX.LT.0 ) IX = 1 - (N-1)*INCX DO I = 1, N SUMSQ = SUMSQ + REAL(X(IX))**2 + AIMAG(X(IX))**2 IX = IX + INCX END DO ELSE IF( SMAX.GT.ZERO ) THEN ! ! Scale by SQMAX if SCL = ONE, otherwise scale by ! max( SQMAX, SCL ). ! SQMAX = MIN( MAX( SQMAX, SAFMIN ), SAFMAX ) IF( SCL.EQ.ONE .OR. SCL.LT.SQMAX ) THEN SUMSQ = ( SUMSQ*( SCL / SQMAX ) )*( SCL / SQMAX ) SCL = SQMAX END IF ! ! Add the sum of squares of values of X scaled by SCL. ! IX = 1 IF( INCX.LT.0 ) IX = 1 - (N-1)*INCX DO I = 1, N SUMSQ = SUMSQ + ( REAL(X(IX)) / SCL )**2 + & ( AIMAG(X(IX)) / SCL )**2 IX = IX + INCX END DO END IF END IF RETURN END SUBROUTINE DLASSQ( N, X, INCX, SCL, SUMSQ ) ! ! -- LAPACK auxiliary routine (version 2.0) -- ! Univ. of Tennessee, Univ. of California Berkeley, NAG Ltd., ! Courant Institute, Argonne National Lab, and Rice University ! October 31, 1992 ! 10-15-99: New version based on LAWN 126 (eca) ! 09-25-01: Don't scale by < SAFMIN or more than SAFMAX (eca) ! ! .. Scalar Arguments .. INTEGER INCX, N DOUBLE PRECISION SCL, SUMSQ ! .. ! .. Array Arguments .. DOUBLE PRECISION X( * ) ! .. ! ! Purpose ! ======= ! ! DLASSQ returns the values scl and smsq such that ! ! ( scl**2 )*smsq = x( 1 )**2 +...+ x( n )**2 + ( scale**2 )*sumsq, ! ! where x( i ) = X( 1 + ( i - 1 )*INCX ). The value of sumsq is ! assumed to be non-negative and scl returns the value ! ! scl = max( scale, abs( x( i ) ) ). ! ! scale and sumsq must be supplied in SCL and SUMSQ and ! scl and smsq are overwritten on SCL and SUMSQ respectively. ! ! Arguments ! ========= ! ! N (input) INTEGER ! The number of elements to be used from the vector X. ! ! X (input) DOUBLE PRECISION array, dimension (N) ! The vector for which a scaled sum of squares is computed. ! x( i ) = X( 1 + ( i - 1 )*INCX ), 1 <= i <= n. ! ! INCX (input) INTEGER ! The increment between successive values of the vector X. ! INCX > 0. ! ! SCL (input/output) DOUBLE PRECISION ! On entry, the value scale in the equation above. ! On exit, SCL is overwritten with scl , the scaling factor ! for the sum of squares. ! ! SUMSQ (input/output) DOUBLE PRECISION ! On entry, the value sumsq in the equation above. ! On exit, SUMSQ is overwritten with smsq , the basic sum of ! squares from which scl has been factored out. ! ! ===================================================================== ! DOUBLE PRECISION ONE, ZERO PARAMETER ( ONE = 1.0D0, ZERO = 0.0D0 ) INTEGER I, IX, IX2 DOUBLE PRECISION CUTHI, CUTLO, HITEST, SAFMAX, SAFMIN, SMAX, SQMAX INTRINSIC ABS, MAX, DBLE, EPSILON, SQRT, TINY DATA CUTLO / 1.00104154759155046D-146 / DATA CUTHI / 9.48075190810917589D+153 / IF( N.LE.0 ) RETURN HITEST = CUTHI / DBLE( N+1 ) !eca SAFMIN = DLAMCH( 'Safe minimum' ) / DLAMCH( 'Epsilon' ) SAFMIN = TINY( ONE ) / EPSILON( ONE ) SAFMAX = ONE / SAFMIN IF( SUMSQ.EQ.ZERO ) SCL = ONE IF( INCX.EQ.1 ) THEN ! ! Pass through once to find the maximum value in X. ! SMAX = ZERO DO I = 1, N SMAX = MAX( SMAX, ABS(X(I)) ) END DO SQMAX = MAX( SCL*SQRT( SUMSQ ), SMAX ) ! IF( SCL.EQ.ONE .AND. SQMAX.GT.CUTLO .AND. SQMAX.LT.HITEST ) & THEN ! ! If SCL = ONE and SQMAX is greater than CUTLO and less than ! HITEST, no scaling should be needed. ! DO I = 1, N SUMSQ = SUMSQ + X(I)**2 END DO ELSE IF( SMAX.GT.ZERO ) THEN ! ! Scale by SQMAX if SCL = ONE, otherwise scale by ! max( SQMAX, SCL ). ! SQMAX = MIN( MAX( SQMAX, SAFMIN ), SAFMAX ) IF( SCL.EQ.ONE .OR. SCL.LT.SQMAX ) THEN SUMSQ = ( SUMSQ*( SCL / SQMAX ) )*( SCL / SQMAX ) SCL = SQMAX END IF ! ! Add the sum of squares of values of X scaled by SCL. ! DO I = 1, N SUMSQ = SUMSQ + ( X(I) / SCL )**2 END DO END IF ELSE ! ! Pass through once to find the maximum value in X. ! SMAX = ZERO IX = 1 IF( INCX.LT.0 ) IX = 1 - (N-1)*INCX DO I = 1, N SMAX = MAX( SMAX, ABS(X(IX)) ) IX = IX + INCX END DO SQMAX = MAX( SCL*SQRT( SUMSQ ), SMAX ) ! IF( SCL.EQ.ONE .AND. SQMAX.GT.CUTLO .AND. SQMAX.LT.HITEST ) & THEN ! ! If SCL = ONE and SQMAX is greater than CUTLO and less than ! HITEST, no scaling should be needed. ! IX = 1 IF( INCX.LT.0 ) IX = 1 - (N-1)*INCX DO I = 1, N SUMSQ = SUMSQ + X(IX)**2 IX = IX + INCX END DO ELSE IF( SMAX.GT.ZERO ) THEN ! ! Scale by SQMAX if SCL = ONE, otherwise scale by ! max( SQMAX, SCL ). ! SQMAX = MIN( MAX( SQMAX, SAFMIN ), SAFMAX ) IF( SCL.EQ.ONE .OR. SCL.LT.SQMAX ) THEN SUMSQ = ( SUMSQ*( SCL / SQMAX ) )*( SCL / SQMAX ) SCL = SQMAX END IF ! ! Add the sum of squares of values of X scaled by SCL. ! IX = 1 IF( INCX.LT.0 ) IX = 1 - (N-1)*INCX DO I = 1, N SUMSQ = SUMSQ + ( X(IX) / SCL )**2 IX = IX + INCX END DO END IF END IF RETURN END SUBROUTINE ZLASSQ( N, X, INCX, SCL, SUMSQ ) ! ! -- LAPACK auxiliary routine (version 2.0) -- ! Univ. of Tennessee, Univ. of California Berkeley, NAG Ltd., ! Courant Institute, Argonne National Lab, and Rice University ! October 31, 1992 ! 10-15-99: New version based on LAWN 126 (eca) ! 09-25-01: Don't scale by < SAFMIN or more than SAFMAX (eca) ! ! .. Scalar Arguments .. INTEGER INCX, N DOUBLE PRECISION SCL, SUMSQ ! .. ! .. Array Arguments .. COMPLEX*16 X( * ) ! .. ! ! Purpose ! ======= ! ! ZLASSQ returns the values scl and ssq such that ! ! ( scl**2 )*ssq = x( 1 )**2 +...+ x( n )**2 + ( scale**2 )*sumsq, ! ! where x( i ) = abs( X( 1 + ( i - 1 )*INCX ) ). The value of sumsq is ! assumed to be non-negative and scl returns the value ! ! scl = max( scale, abs( real( x( i ) ) ), abs( aimag( x( i ) ) ) ), ! i ! ! scale and sumsq must be supplied in SCL and SUMSQ respectively. ! SCL and SUMSQ are overwritten by scl and ssq respectively. ! ! Arguments ! ========= ! ! N (input) INTEGER ! The number of elements to be used from the vector X. ! ! X (input) COMPLEX*16 array, dimension (N) ! The vector x as described above. ! x( i ) = X( 1 + ( i - 1 )*INCX ), 1 <= i <= n. ! ! INCX (input) INTEGER ! The increment between successive values of the vector X. ! INCX > 0. ! ! SCL (input/output) DOUBLE PRECSION ! On entry, the value scale in the equation above. ! On exit, SCALE is overwritten with the value scl . ! ! SUMSQ (input/output) DOUBLE PRECSION ! On entry, the value sumsq in the equation above. ! On exit, SUMSQ is overwritten with the value ssq . ! ! ===================================================================== ! DOUBLE PRECISION ONE, ZERO PARAMETER ( ONE = 1.0D0, ZERO = 0.0D0 ) INTEGER I, IX, IX2 DOUBLE PRECISION CUTHI, CUTLO, HITEST, SAFMAX, SAFMIN, SMAX, SQMAX INTRINSIC ABS, MAX, DBLE, EPSILON, SQRT, TINY DATA CUTLO / 1.00104154759155046D-146 / DATA CUTHI / 9.48075190810917589D+153 / IF( N.LE.0 ) RETURN HITEST = CUTHI / DBLE( 2*N+1 ) !eca SAFMIN = DLAMCH( 'Safe minimum' ) / DLAMCH( 'Epsilon' ) SAFMIN = TINY( ONE ) / EPSILON( ONE ) SAFMAX = ONE / SAFMIN IF( SUMSQ.EQ.ZERO ) SCL = ONE IF( INCX.EQ.1 ) THEN ! ! Pass through once to find the maximum value in X. ! SMAX = ZERO DO I = 1, N SMAX = MAX( SMAX, ABS(DBLE(X(I))), ABS(DIMAG(X(I))) ) END DO SQMAX = MAX( SCL*SQRT( SUMSQ ), SMAX ) ! IF( SCL.EQ.ONE .AND. SQMAX.GT.CUTLO .AND. SQMAX.LT.HITEST ) & THEN ! ! If SCL = ONE and SQMAX is greater than CUTLO and less than ! HITEST, no scaling should be needed. ! DO I = 1, N SUMSQ = SUMSQ + DBLE(X(I))**2 + DIMAG(X(I))**2 END DO ELSE IF( SMAX.GT.ZERO ) THEN ! ! Scale by SQMAX if SCL = ONE, otherwise scale by ! max( SQMAX, SCL ). ! SQMAX = MIN( MAX( SQMAX, SAFMIN ), SAFMAX ) IF( SCL.EQ.ONE .OR. SCL.LT.SQMAX ) THEN SUMSQ = ( SUMSQ*( SCL / SQMAX ) )*( SCL / SQMAX ) SCL = SQMAX END IF ! ! Add the sum of squares of values of X scaled by SCL. ! DO I = 1, N SUMSQ = SUMSQ + ( DBLE(X(I)) / SCL )**2 + & ( DIMAG(X(I)) / SCL )**2 END DO END IF ELSE ! ! Pass through once to find the maximum value in X. ! SMAX = ZERO IX = 1 IF( INCX.LT.0 ) IX = 1 - (N-1)*INCX DO I = 1, N SMAX = MAX( SMAX, ABS(DBLE(X(IX))), ABS(DIMAG(X(IX))) ) IX = IX + INCX END DO SQMAX = MAX( SCL*SQRT( SUMSQ ), SMAX ) ! IF( SCL.EQ.ONE .AND. SQMAX.GT.CUTLO .AND. SQMAX.LT.HITEST ) & THEN ! ! If SCL = ONE and SQMAX is greater than CUTLO and less than ! HITEST, no scaling should be needed. ! IX = 1 IF( INCX.LT.0 ) IX = 1 - (N-1)*INCX DO I = 1, N SUMSQ = SUMSQ + DBLE(X(IX))**2 + DIMAG(X(IX))**2 IX = IX + INCX END DO ELSE IF( SMAX.GT.ZERO ) THEN ! ! Scale by SQMAX if SCL = ONE, otherwise scale by ! max( SQMAX, SCL ). ! SQMAX = MIN( MAX( SQMAX, SAFMIN ), SAFMAX ) IF( SCL.EQ.ONE .OR. SCL.LT.SQMAX ) THEN SUMSQ = ( SUMSQ*( SCL / SQMAX ) )*( SCL / SQMAX ) SCL = SQMAX END IF ! ! Add the sum of squares of values of X scaled by SCL. ! IX = 1 IF( INCX.LT.0 ) IX = 1 - (N-1)*INCX DO I = 1, N SUMSQ = SUMSQ + ( DBLE(X(IX)) / SCL )**2 + & ( DIMAG(X(IX)) / SCL )**2 IX = IX + INCX END DO END IF END IF RETURN END real function SNRM2( n, x, incx ) ! ! -- Level 1 BLAS -- ! E. Anderson, NESC ! November 17, 1999 ! integer n, incx real x(*) real scl, sumsq external SLASSQ intrinsic SQRT scl = 1.0 sumsq = 0.0 call SLASSQ( n, x, incx, scl, sumsq ) SNRM2 = scl*SQRT( sumsq ) return end real function SCNRM2( n, x, incx ) ! ! -- Level 1 BLAS -- ! E. Anderson, NESC ! November 17, 1999 ! integer n, incx complex x(*) real scl, sumsq external CLASSQ intrinsic SQRT scl = 1.0 sumsq = 0.0 call CLASSQ( n, x, incx, scl, sumsq ) SCNRM2 = scl*SQRT( sumsq ) return end double precision function DNRM2( n, x, incx ) ! ! -- Level 1 BLAS -- ! E. Anderson, NESC ! November 17, 1999 ! integer n, incx double precision x(*) double precision scl, sumsq external DLASSQ intrinsic DSQRT scl = 1.0D0 sumsq = 0.0D0 call DLASSQ( n, x, incx, scl, sumsq ) DNRM2 = scl*DSQRT( sumsq ) return end double precision function DZNRM2( n, x, incx ) ! ! -- Level 1 BLAS -- ! E. Anderson, NESC ! November 17, 1999 ! integer n, incx complex*16 x(*) double precision scl, sumsq external ZLASSQ intrinsic DSQRT scl = 1.0D0 sumsq = 0.0D0 call ZLASSQ( n, x, incx, scl, sumsq ) DZNRM2 = scl*DSQRT( sumsq ) return end --------------6C386FC925298A0C59F101F0-- ============================================================================= From: Edward Anderson Subject: Re: New Debian maintainer To: Camm Maguire Date: Mon, 03 Dec 2001 08:42:42 -0500 Organization: Lockheed Martin (contractor to the U.S. EPA) Received: from localhost [127.0.0.1] (root) by intech5.enhanced.com with esmtp (Exim 3.12 #1 (Debian)) id 16AtPH-0000Jg-00; Mon, 03 Dec 2001 08:45:15 -0500 Received: from 204.213.252.11 by localhost with POP3 (fetchmail-5.3.3) for camm@enhanced.com (multi-drop); Mon, 03 Dec 2001 08:45:15 -0500 (EST) Received: from mail.eticomm.net [204.213.252.17] by ux1.cyberenet.net with esmtp (Exim 1.61 #1) id 16AtNS-0005d8-00; Mon, 3 Dec 2001 08:43:22 -0500 Received: from linden.nesc.epa.gov ([134.67.64.25] helo=nesc.epa.gov) by relay.eticomm.net with esmtp (Exim 3.33 #2) id 16AtNR-0002PS-00 for camm@enhanced.com; Mon, 03 Dec 2001 08:43:21 -0500 Received: from epa.gov (yew.nesc.epa.gov [134.67.64.48]) by nesc.epa.gov (8.9.3+Sun/8.9.1) with ESMTP id IAA01538 for ; Mon, 3 Dec 2001 08:42:42 -0500 (EST) Sender: dle@nesc.epa.gov Message-ID: <3C0B8152.6CDAF8BD@epa.gov> X-Mailer: Mozilla 4.75C-SGI [en] (X11; I; IRIX 6.5 IP32) X-Accept-Language: en MIME-Version: 1.0 References: <004701c17372$2a3580e0$2fee80ac@compaq> <3C025704.86683917@epa.gov> <544rndhvce.fsf@intech19.enhanced.com> <3C07A61C.57A40172@epa.gov> <547ks7kft4.fsf@intech19.enhanced.com> Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit X-UIDL: b458a850136eaa1c731faa8144f7d567 Lines: 57 Xref: intech19.enhanced.com mail.camm:27531 Camm: If you only have a Fortran 77 compiler then you won't be able to use the Fortran 90 intrinsic functions EPSILON and TINY. You will need to use the calls to SLAMCH/DLAMCH which appear in the comments of the code I sent, and remove the declarations of EPSILON and TINY from the INTRINSIC line. Should you decide to use these versions, you would want to put the LAPACK auxiliary routines xLASSQ in LAPACK/SRC and the BLAS routines xNRM2 in LAPACK/BLAS/SRC. --Ed Camm Maguire wrote: > > Sorry to bother you again -- one other item. I'm having a bit of > trouble compiling the lassq with g77: > > f77 -pipe -c -O6 -fomit-frame-pointer -fPIC classq.f > classq.f: In subroutine `classq': > classq.f:63: warning: > INTRINSIC ABS, EPSILON, MAX, REAL, TINY > ^ > Reference to unimplemented intrinsic `EPSILON' at (^) (assumed EXTERNAL) > classq.f:63: > INTRINSIC ABS, EPSILON, MAX, REAL, TINY > ^ > Invalid declaration of or reference to symbol `epsilon' at (^) [initially seen at (^)] > classq.f:63: warning: > INTRINSIC ABS, EPSILON, MAX, REAL, TINY > ^ > Reference to unimplemented intrinsic `TINY' at (^) (assumed EXTERNAL) > classq.f:63: > INTRINSIC ABS, EPSILON, MAX, REAL, TINY > ^ > Invalid declaration of or reference to symbol `tiny' at (^) [initially seen at (^)] > make: *** [classq.o] Error 1 > > Have I merely put these files in the wrong place? Or does your > fortran compiler define more intrinsics than g77? Can I hardcode > these values from the lapack lamch output? > > Take care, > -- > Camm Maguire camm@enhanced.com > ========================================================================== > "The earth is but one country, and mankind its citizens." -- Baha'u'llah ---------------------------------------------------------------- Ed Anderson Scientific Applications Consultant Lockheed Martin Technology Services (contractor to the U.S. EPA) Phone: 919-541-0299 mailto:Anderson.Edward@epa.gov ============================================================================= From: Edward Anderson Subject: Re: New Debian maintainer To: Camm Maguire Date: Tue, 04 Dec 2001 09:00:44 -0500 Organization: Lockheed Martin (contractor to the U.S. EPA) Received: from localhost [127.0.0.1] (root) by intech5.enhanced.com with esmtp (Exim 3.12 #1 (Debian)) id 16BGBM-0000Ti-00; Tue, 04 Dec 2001 09:04:24 -0500 Received: from 204.213.252.11 by localhost with POP3 (fetchmail-5.3.3) for camm@enhanced.com (multi-drop); Tue, 04 Dec 2001 09:04:24 -0500 (EST) Received: from mail.eticomm.net [204.213.252.17] by ux1.cyberenet.net with esmtp (Exim 1.61 #1) id 16BG8K-0001XT-00; Tue, 4 Dec 2001 09:01:16 -0500 Received: from linden.nesc.epa.gov ([134.67.64.25] helo=nesc.epa.gov) by relay.eticomm.net with esmtp (Exim 3.33 #2) id 16BG8K-000FX3-00 for camm@enhanced.com; Tue, 04 Dec 2001 09:01:16 -0500 Received: from epa.gov (yew.nesc.epa.gov [134.67.64.48]) by nesc.epa.gov (8.9.3+Sun/8.9.1) with ESMTP id JAA07649 for ; Tue, 4 Dec 2001 09:00:44 -0500 (EST) Sender: dle@nesc.epa.gov Message-ID: <3C0CD70C.EA958B8C@epa.gov> X-Mailer: Mozilla 4.75C-SGI [en] (X11; I; IRIX 6.5 IP32) X-Accept-Language: en MIME-Version: 1.0 References: <004701c17372$2a3580e0$2fee80ac@compaq> <3C025704.86683917@epa.gov> <544rndhvce.fsf@intech19.enhanced.com> <3C07A61C.57A40172@epa.gov> <547ks7kft4.fsf@intech19.enhanced.com> <3C0B8152.6CDAF8BD@epa.gov> <54oflgyq59.fsf@intech19.enhanced.com> Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit X-UIDL: adfa4279f91ebf2369b8aaef02ca151f Lines: 23 Xref: intech19.enhanced.com mail.camm:27561 Camm, There's some danger in continuing in this vein because you'll end up with your own version of LAPACK that differs from every other one out there. But that's probably not so unusual for a software installation. You may be able to fix some bad test ratios in the symmetric eigenvalue driver routines by adjusting the value of "FUDGE" in sstebz/dstebz. As the name implies, this was a "fudge factor" and it is adjustable. I increased it from 2.0 to 2.1 for the CRAY T3E, but I heard from someone else (can't find the mail now) that they had set it even higher, like 2.5. Give it a whirl and see what happens. --Ed ---------------------------------------------------------------- Ed Anderson Scientific Applications Consultant Lockheed Martin Technology Services (contractor to the U.S. EPA) Phone: 919-541-0299 mailto:Anderson.Edward@epa.gov ============================================================================= From: Edward Anderson Subject: Re: New Debian maintainer To: Camm Maguire CC: Susan Blackford Date: Wed, 05 Dec 2001 16:27:32 -0500 Organization: Lockheed Martin (contractor to the U.S. EPA) Received: from localhost [127.0.0.1] (root) by intech5.enhanced.com with esmtp (Exim 3.12 #1 (Debian)) id 16Bjcx-0002Uh-00; Wed, 05 Dec 2001 16:30:51 -0500 Received: from 204.213.252.11 by localhost with POP3 (fetchmail-5.3.3) for camm@enhanced.com (multi-drop); Wed, 05 Dec 2001 16:30:51 -0500 (EST) Received: from mail.eticomm.net [204.213.252.17] by ux1.cyberenet.net with esmtp (Exim 1.61 #1) id 16BjaG-0005He-00; Wed, 5 Dec 2001 16:28:04 -0500 Received: from linden.nesc.epa.gov ([134.67.64.25] helo=nesc.epa.gov) by relay.eticomm.net with esmtp (Exim 3.33 #2) id 16BjaF-000IIY-00 for camm@enhanced.com; Wed, 05 Dec 2001 16:28:03 -0500 Received: from epa.gov (yew.nesc.epa.gov [134.67.64.48]) by nesc.epa.gov (8.9.3+Sun/8.9.1) with ESMTP id QAA18500; Wed, 5 Dec 2001 16:27:32 -0500 (EST) Sender: dle@nesc.epa.gov Message-ID: <3C0E9144.3035D152@epa.gov> X-Mailer: Mozilla 4.75C-SGI [en] (X11; I; IRIX 6.5 IP32) X-Accept-Language: en MIME-Version: 1.0 References: <004701c17372$2a3580e0$2fee80ac@compaq> <3C025704.86683917@epa.gov> <544rndhvce.fsf@intech19.enhanced.com> <3C07A61C.57A40172@epa.gov> <547ks7kft4.fsf@intech19.enhanced.com> <3C0B8152.6CDAF8BD@epa.gov> <54oflgyq59.fsf@intech19.enhanced.com> <3C0CD70C.EA958B8C@epa.gov> <54vgflms1a.fsf@intech19.enhanced.com> Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit X-UIDL: 7c86af4c1b5cd2c960fef3c9e2448126 Lines: 68 Xref: intech19.enhanced.com mail.camm:27607 Camm Maguire wrote: > > Greetings, and thank you again for the suggestion. I saw the > adjustment to 2.1, but changing it a bit all the way up to 3.5 made no > difference at all in the test results. I don't know what this means. > > I'd also like to ask you about the significance of the errors, and > which set is 'preferable'. Are routine errors worse than driver > errors? Why did the complex driver errors get so much larger in > magnitude with your blas changes? (everything else seems to have > decreased.) What effect in practice will the magnitudes of the errors > seen in both sets have on typical library use? And lastly, is it known > how these errors got introduced into the code in the last year? (1999 > 3.0 doesn't have them.) > Camm: With a few exceptions, all the mathematics of LAPACK is contained in the computational routines. If you get a failure in a driver routine when all the computational routines passed their tests, it is usually because one path through the driver routines didn't agree with another path through the driver routines, and the writer of the test code thought that they should. I regard all the errors as equal in importance, but many of the eigensystem tests are not very well documented. There are a lot of calculations in LAPACK that are sensitive to the scaling and even the order of operations. Also, the LAPACK test code deliberately generates some numerically sensitive test problems to exercise parts of the code. Sometimes a change in the scaling will make a particular test ratio disappear, while creating a bad test ratio somewhere else. It's hard to say which method is "better" without more analysis. I've managed to get all except the xgd tests to pass on a CRAY PVP, CRAY MPP, and IBM SP, but my set of changes is rather large and not limited to numerical accuracy (some are to improve performance). In my experience, test ratios under 1000 or so are usually not that serious. Most of the test ratios are scaled so that an order of magnitude is a lost decimal place of accuracy, and the thresholds are set to about 50. So, if your test ratio were 1000, you would have lost about 3 places of accuracy -- a lot in single precision, but probably in the noise in double precision. If you were concerned about the accuracy of a particular solution in the context of an application, LAPACK has error bounds and condition number information built into the "expert" driver routines that may offer some guidance about what to expect. Your best bet as the installer of LAPACK would be to document the failures you currently see, ask that they be posted in the LAPACK release notes, and wait for something better to come along. It will also give you a baseline for comparison if and when there is a new LAPACK release. Just by looking at the error messages that were generated, you've already done more than some library developers I know! I'll put you on my list of people to notify when my (totally unsupported and unauthorized) patch for the IBM SP is available. It's been useful for me to see that some set of users still needs Fortran 77 compatibility. --Ed ---------------------------------------------------------------- Ed Anderson Scientific Applications Consultant Lockheed Martin Technology Services (contractor to the U.S. EPA) Phone: 919-541-0299 mailto:Anderson.Edward@epa.gov ============================================================================= -- Camm Maguire , Tue, 29 Jan 2002 11:42:34 -0500