g77 optimization fails with current versions

From: "Olofsen, E. (ANST)" E.Olofsen@lumc.nl Subject: [NMusers] g77 optimization fails with current versions Date: 12/11/2003 9:28 AM Dear NONMEM users, NONMEM + ADVAN3 compiled using g77 versions 3.2.2 and 3.2.3 with -O3 (and possibly with -O2 and other versions) may produce error messages such as PK PARAMETER FOR K12 IS NON-POSITIVE or PRED RETURNS DIFFERENT OUTPUTS WITH THE SAME INPUTS When I compiled PREDI with -O3 -fno-gcse, these error messages disappeared. This is probably related to a bug in these, rather current, versions of gcc, and other NONMEM subroutines may be compiled incorrectly as well. I'm currently checking the gcc bug archives and testing the latest version, 3.3.2, but it seems the bug hasn't been solved yet. Erik Olofsen Leiden University Medical Center

From: "Bachman, William (MYD)" bachmanw@iconus.com Subject: [NMusers] reporting "errors" or "bugs" to nmusers Date: 12/12/2003 7:57 AM In my opinion, unless you can provide the control streams and data to reproduce the issues and give the details of your installation and error messages, it is not useful to report alleged "bugs" or "errors". Particularly if you have used non-standard installations or procedures. We cannot assess the value of these statements. nmconsult@globomaxnm.com GloboMax The Strategic Pharmaceutical Development Division of ICON plc 7250 Parkway Drive, Suite 430 Hanover, MD 21076 Voice: (410) 782-2205 FAX: (410) 712-0737

From: "Olofsen, E. (ANST)" E.Olofsen@lumc.nl Subject: [NMusers] g77 optimization fails with current versions Date: 12/12/2003 8:14 AM Dear NONMEM users, Yesterday I reported about the g77 optimization bug. I've now recompiled version 3.2.3 and compiled the latest version, 3.3.2, and did some more runs with ADVAN3 and two different data sets. With 3.2.3 and -O3, it is possible to get a variety of error messages, or possibly no error message at all, so I don't think the results with -O3 (and possibly -O2) can be trusted. I've tested this under two different Linux distributions and kernels. With 3.2.3 and -O3 -fno-gcse, runs may or may not be different from runs without optimization, but this may just be because with optimized code rounding errors accumulate differently. Runs with the latest version, 3.3.2, were consistent, so it appears the bug has been removed. Interestingly, runs were also more consistent than those of using version 2.95.3 with the model/data combination with low stability. Erik Olofsen Leiden University Medical Center

From: Erik Olofsen [mailto:e.olofsen@xs4all.nl] Subject: RE: reporting "errors" or "bugs" to nmusers Date: Friday, December 12, 2003 1:28 PM Bill, I installed NONMEM from the floppies using SETUP on another computer and this gives the same problem. >> 1. the control stream and data file that reproduces the issue. Proprietary >> information should be masked (e.g product names, physioligical marker >> identification, etc.), data may be a subset of the whole sufficient to >> reproduce the issue. Files are attached (the same as yesterday + the report file). >> 2. complete details of system: >> - operating system and version, e.g. Windows 2000 or Red Hat Linux 9 Gentoo Linux version 1.4, with kernel 2.4.21 >> - compiler and version, e.g. Compaq Visual Fortran 6.6 or GNU g77 3.2 GNU Fortran (GCC 3.2.2) 3.2.2 20030205 (release) >> - NONMEM version and level, e.g. NONMEM V Level 1.1 NONMEM V Level 1.1 >> - method used to install NONMEM, e.g setup.bat or SETUP C-shell script >> supplied with NONMEM distribution, user-defined custom installation (Note: >> We cannot trouble-shoot custom installs.) The SETUP C-shell script with compiler option -O3 So now I've tested different computers, different Linux distributions (Slackware too), my NONMEM source directory and the original files (NSIZES etc.) and my and the original installation procedures. The problem appears to lie with compilers 3.2.2 and 3.2.3. I might try to find the time to look at MinGW. Erik

From: Erik Olofsen [mailto:e.olofsen@xs4all.nl] Subject: [NMusers] RE: reporting "errors" or "bugs" to nmusers -> -O3 optimization r esults corroborated Date: 12/15/2003 9:50 AM Bill, >> My results with gcc 3.2.2 on Red Hat Linux 9 are as follows: >> >> -O and '-O3 -fno-gcse': >> MINIMIZATION TERMINATED >> DUE TO PROXIMITY OF LAST ITERATION EST. TO A VALUE >> AT WHICH THE OBJ. FUNC. IS INFINITE (ERROR=136) >> >> obj. fcn for -O: 785.8 >> obj. fcn for '-O3 -fno-gcse': 498.5 It could be that the latter result is also correctly computed, but that with this unstable model/data combination, rounding errors accumulate in a different way in the (correctly) optimized code causing NONMEM to abort at quite another point. >> -O3: PK PARAMETER FOR K12 IS NON-POSITIVE >> >> These results corroborate your results. Additionally, results with runs of >> CONTROL5 and THEOPP (distributed with NONMEM) indicate there indeed is a >> problem with -O3 optimization in general: >> >> -O and '-O3 -fno-gcse': minimization successful >> -O3: PK PARAMETERS FOR KA AND K ARE EQUAL And (see above) I guess with -O and '-O3 -fno-gcse' the obj. func. values are equal. >> My recommendation for g77 would be to use -O optimization for versions 3.2.2 >> and 3.2.3 and test optimization options for future versions. >> Have you been able to get any additional information on the source of the >> problem? I haven't found a description of exactly this problem on the gnu bug lists, but some other reports of problems with optimization. Indeed it is best to be careful with strong optimization with any compiler. Erik

From: "Bachman, William (MYD)" bachmanw@iconus.com Subject: RE: reporting "errors" or "bugs" to nmusers -> -O3 optimization r esults corroborated Date: 12/16/2003 9:56 AM Erik, My results with gcc 3.2.2 on Red Hat Linux 9 are as follows: -O and '-O3 -fno-gcse': MINIMIZATION TERMINATED DUE TO PROXIMITY OF LAST ITERATION EST. TO A VALUE AT WHICH THE OBJ. FUNC. IS INFINITE (ERROR=136) obj. fcn for -O: 785.8 obj. fcn for '-O3 -fno-gcse': 498.5 -O3: PK PARAMETER FOR K12 IS NON-POSITIVE These results corroborate your results. Additionally, results with runs of CONTROL5 and THEOPP (distributed with NONMEM) indicate there indeed is a problem with -O3 optimization in general: -O and '-O3 -fno-gcse': minimization successful -O3: PK PARAMETERS FOR KA AND K ARE EQUAL Thank you for providing the additional information to evaluate the issue. My recommendation for g77 would be to use -O optimization for versions 3.2.2 and 3.2.3 and test optimization options for future versions. Have you been able to get any additional information on the source of the problem? Bill nmconsult@globomaxnm.com GloboMax The Strategic Pharmaceutical Development Division of ICON plc 7250 Parkway Drive, Suite 430 Hanover, MD 21076 Voice: (410) 782-2205 FAX: (410) 712-0737 _______________________________________________________