Compaq Visual Fortran compared to Gnu g77 compiler

From:Nick Holford Subject:[NMusers] Compaq Visual Fortran compared to Gnu g77 compiler Date:Wed, 07 Aug 2002 08:29:43 +1200 Bill, I list below the results I got on Peter Bonate's valfit2 problem. The runs indicate the various compilers and versions I tested with NONMEM V 1.1. g77 versions used -fno-backslash -O and df versions used /fltconsistency /optimize:4 /fast. This problem is numerically unstable and does not give a satisfactory solution with any compiler. However, lower objective function values are produced by the df compiler versions compared with the g77 compiler versions. I have found the same behaviour on other examples and typically the Compaq Visual Fortran (df65 and later) will get lower objective function values than g77 on problems that do not converge readily. If the problem converges then both compilers will converge at the same value and the parameter estimates are essentially identical. It is my interpretation of these results that g77 is unable to continue the search for a lower objective because it loses precision. The compiler version is a factor that also needs to be considered when reporting the results. Note that the Compaq version (df 6.6) is better than the earlier Digital version (6.0A) and the g77 3.1 version is better than the earlier 0.5.25 version using the same criterion. I had reported this version specific issue earlier to nmusers with a different problem (Re: [NMusers] Digital Visual Fortran vers 5.0 vs 6.6 March 20 2002). #Run Sig Obj Errors df66 2.5 8280.513 ROUNDING_ERRORS_(ERROR=134) g7731 2.7 8313.35 ROUNDING_ERRORS_(ERROR=134) g770525 2.5 8439.316 ROUNDING_ERRORS_(ERROR=134) df60a . 8612.169 PROXIMITY_OF_LAST_ITERATION_EST._TO_A_VALUE_AT_WHICH_THE_OBJ._FUNC._IS_INFINITE_(ERROR=136)