Sun, Digital, and Cray use the number of bytes as the KIND-parameter. SunSoft permits the values 1, 4, and 8 for logical variables (with the same result as for the default value 4), the values 1, 2, and 4 for integers, the values 4 and 8 for floating point values, and also the values 4 and 8 for complex values. The names int7, int15, and int31 are mine.
LOGICAL Default byte word double KIND number = 4 1 4 8 INTEGER int7 int15 int31 KIND number = 1 2 4 digits = 7 15 31 radix = 2 2 2 range = 2 4 9 huge = 127 32767 2147483647 bit_size = 8 16 32 REAL single double KIND number = 4 8 digits = 24 53 maxexponent = 128 1024 minexponent = -125 -1021 precision = 6 15 radix = 2 2 range = 37 307 epsilon = 0.11920929E-06 0.22204460E-15 tiny = 0.11754944E-37 0.22250739-307 huge = 0.34028235E+39 0.17976931+309 COMPLEX single double KIND number = 4 8 precision = 6 15 range = 37 307
-ansi: Report non-ANSI extensions -C: Index check -c: Suppress linking, produce .o files -e: Recognize extended (132 character) source lines -fast: Specify common set of performance options -fixed: Interpret all Fortran source files according to fixed form rules -flags: Print summary of compiler options -fnonstd: Non-standard initialization of floating-point hardware -fns: Turn on non-standard floating point mode -free: Interpret all Fortran source files according to free form rules -fround: Select the appropriate IEEE rounding mode -g: Generate debugging information used by dbx -help: Print summary of compiler options -IX: Passed to the linker to add X to the search path for include files -LX: Passed to the linker to add X to the library search path -lX: Read object library (for ld) -MX: Add directory X to module search path -o file: Set name of output file -O: Generate optimized code -onetrip: Perform DO loops at least once -p: Prepare object code to collect data for profiling with prof -pg: Prepare object code to collect data for profiling with gprof -qp: Same as -p Produce file of type (may also be spelled -qproduce) -Qproduce .f90: .f90 (f90 source, Free Form) -Qproduce .f: .f (f90 source, Fixed Form) -Qproduce .lst: .lst (f90 listing) -Qproduce .o: .o (Object file) -Qproduce .s: .s (Assembler source) -r8const: Changes single precision constants into double precision constants -S: Product .s file only (do not assemble or link) -s: Strip the executable -time: Report the execution time for each component -u: Implies IMPLICIT NONE -V: Show version number of each invoked component -v: Verbose mode -w: Suppress compiler warning messages -xtypemap: Choice of precision for integers and floating pointThe following suffices may be used.
Suffix 'a': Object library Suffix 'il': Inline expansion file Suffix 'o': Object file Suffix 'so': Shared object Suffix 's': Assembler source Suffix 'S': Assembler source for cpp Suffix 'for': f90 source (Fixed Form) Suffix 'ftn': f90 source (Fixed Form) Suffix 'f': f90 source (Fixed Form) Suffix 'F': not currently recognized Suffix 'f90': f90 source (Free Form) Suffix 'lst': f90 listing
On Digital there are some simple commands to the compiler to switch precision. On Sun the command is -xtypemap, and the following combinations are permitted:
-xtypemap=real:32,double:64 Default -xtypemap=real:64,double:64 Both the same -xtypemap=real:64,double:128 Both doubledAlso the integers can be chosen as integer:32 or integer:64. Please note that only variables specified with REAL :: VAR are affected, not those specified with REAL (KIND=any) :: VAR, and correspondingly for integers. It is not possible to use real:128 or double:32.
We first look at the program EPSILON or KAHAN. In order to obtain the desired precision we use SELECTED_REAL_KIND.
Script started on Wed Jun 06 12:48:40 2001 on Solaris 8.
water[~]> cat kahan2.f90 PROGRAM KAHAN2 ! This program is intentionally incorrect ! if not compiled with the Sun option "-r8const" IMPLICIT NONE INTEGER, PARAMETER :: DP = SELECTED_REAL_KIND(15,307) REAL (KIND=DP) :: A, B, C, D, E A = 1.0/3.0 B = 4.0*A - 1.0 C = 3.0*B - 1.0 D = 0.5*C E = ABS(D) WRITE(*,*) ' my = ', E END PROGRAM KAHAN2If this program is compiled in the usual way the result will correspond to single precision, if it is compiled with -r8const the result will correspond to double precision.
water[~]> f95 kahan2.f90 water[~]> ./a.out my = 5.960464477539062E-8 ! Single precision result water[~]> f95 -r8const kahan2.f90 water[~]> ./a.out my = 1.1102230246251565E-16 ! Double precision resultWe now switch to quad precision. Here there is nothing corresponding to -r8const or -r16const, so all floating point constants have to be specified explicitly to be of the correct precision. In the example below it is however sufficient to do this with the one third assigned to A, since the other constants are all exact in binary arithmetic.
water[~]> cat kahan4.f90 PROGRAM KAHAN4 IMPLICIT NONE INTEGER, PARAMETER :: QP = SELECTED_REAL_KIND(33,3000) REAL (KIND=QP) :: A, B, C, D, E A = 1.0_QP/3.0_QP B = 4.0_QP*A - 1.0_QP C = 3.0_QP*B - 1.0_QP D = 0.5_QP*C E = ABS(D) WRITE(*,*) ' my = ', E END PROGRAM KAHAN4
water[~]> f95 kahan4.f90 water[~]> ./a.out my = 9.629649721936179265279889712924637E-35 ! Quad precision result water[~]> exitThis result is correct, we may note that 33 is the maximum permitted number of significant digits on this system.
The summation program SUMMATION works with only a a minor difference compared with on the Digital system, it prints two additional digits for the highest precision. The values are correct.
Back to Appendix 6 (of the English version)
Back to Appendix 7 (of the Swedish version)