Fortran/fr: Difference between revisions

From Alliance Doc
Jump to navigation Jump to search
(Created page with "== Options utiles de compilation ==")
(Created page with "La plupart des compilateurs Fortran modernes offrent des options utiles pour le débogage. * <tt>-fcheck=all</tt> pour le compilateur gfortran et <tt>-check</tt> pour le comp...")
Line 5: Line 5:
== Options utiles de compilation ==
== Options utiles de compilation ==


Most contemporary Fortran compilers have a variety of options that can be very helpful during the debugging phase of code development.
La plupart des compilateurs Fortran modernes offrent des options utiles pour le débogage.  
* <tt>-fcheck=all</tt> for the gfortran compiler and <tt>-check</tt> for the ifort compiler check array bounds and alert for disassociated pointers and uninitialized variables;
* <tt>-fcheck=all</tt> pour le compilateur gfortran et <tt>-check</tt> pour le compilateur ifort vérifient les limites des tableaux et signalent les pointeurs sans cible et les variables non initialisées;
* <tt>-fpe0</tt> (ifort) causes the application to halt for floating point exceptions such as division by zero or the square root of a negative, instead of simply generating a NaN and letting the application run;  
* <tt>-fpe0</tt> (ifort) interrompt l'application dans des cas de virgule flottante (division par zéro ou racine carrée d'un nombre négatif) plutôt que de simplement générer NaN (''not a number'') et laisser l'application se poursuivre;
* during testing, you should use <tt>-O0</tt> to disable optimizations and <tt>-g</tt> to add debugging symbols.     
* pendant les tests, utilisez <tt>-O0</tt> pour désactiver les optimisations et <tt>-g</tt> pour ajouter les symboles de débogage.     


==Numerical Linear Algebra==
==Numerical Linear Algebra==

Revision as of 19:14, 12 June 2017

Other languages:

Fortran est un langage compilé disponible sur les ordinateurs de Calcul Canada où sont installés les compilateurs gfortran et ifort. En général, les langages compilés offrent une meilleure performance; nous vous encourageons donc à écrire vos programmes en Fortran, C ou C++.

Options utiles de compilation

La plupart des compilateurs Fortran modernes offrent des options utiles pour le débogage.

  • -fcheck=all pour le compilateur gfortran et -check pour le compilateur ifort vérifient les limites des tableaux et signalent les pointeurs sans cible et les variables non initialisées;
  • -fpe0 (ifort) interrompt l'application dans des cas de virgule flottante (division par zéro ou racine carrée d'un nombre négatif) plutôt que de simplement générer NaN (not a number) et laisser l'application se poursuivre;
  • pendant les tests, utilisez -O0 pour désactiver les optimisations et -g pour ajouter les symboles de débogage.

Numerical Linear Algebra

Note that modern versions of Fortran, i.e. from Fortran 90 on, include built-in functions to handle basic linear algebra operations like multiplication involving matrices and vectors (matmul and dot_product) and tranposition of matrices (transpose). You should use these or the system-provided BLAS/LAPACK libraries and never attempt to write your own methods for such operations, except as an educational exercise. The BLAS matrix-matrix multiplication routine can be up to 100 times faster than a naive implementation involving three nested loops.

Segmentation Faults

An error that is frequently seen with a Fortran program comes from interface problems. These problems surface if a pointer, a dynamically allocated array or even a function pointer is passed as an argument to a subroutine. There are no compile-time problems, but when the program is ran you see for example the following message:

forrtl: severe (174): SIGSEGV, segmentation fault occurred

To correct this problem, you should ensure that the interface of the subroutine is explicitly defined. This can be done in Fortran using the INTERFACE command. Then the compiler can construct the interface and the segmentation faults are fixed.


When the argument is an allocatable array, you should replace the following code:

File : error_allocate.f90

Program Eigenvalue
implicit none

integer                       :: ierr
integer                       :: ntot
real, dimension(:,:), pointer :: matrix

read(5,*) ntot
ierr = genmat( ntot, matrix )

call Compute_Eigenvalue( ntot, matrix )

deallocate( matrix )
end


by this code:

File : interface_allocate.f90

Program Eigenvalue
implicit none

integer                       :: ierr
integer                       :: ntot
real, dimension(:,:), pointer :: matrix

interface
    function genmat( ntot, matrix )
    implicit none
    integer                       :: genmat
    integer, intent(in)           :: ntot
    real, dimension(:,:), pointer :: matrix
    end function genmat
end interface

read(5,*) ntot
ierr = genmat( ntot, matrix )

call Compute_Eigenvalue( ntot, matrix )

deallocate( matrix )
end


The same principle applies when the argument is a function pointer. Consider, for example, the following code:

File : error_=pointer.f90

Program AreaUnderTheCurve
implicit none

real,parameter :: boundInf = 0.
real,parameter :: boundSup = 1.
real           :: area
real, external :: computeIntegral
real, external :: FunctionToIntegrate

area = computeIntegral( FunctionToIntegrate, boundInf, boundSup )

end

function FunctionToIntegrate( x )
implicit none

real             :: FunctionToIntegrate
real, intent(in) :: x

FunctionToIntegrate = x

end function FunctionToIntegrate

function computeIntegral( func, boundInf, boundSup )
implicit none

real, external   :: func
real, intent(in) :: boundInf, boundSup

...


To avoid segmentation faults you should replace the above code by the following:

File : interface_pointer.f90

Program Eigenvalue
implicit none

real,parameter :: boundInf = 0.
real,parameter :: boundSup = 1.
real           :: area
real, external :: computeIntegral

interface
    function FunctionToIntegrate( x )
    implicit none
    real             :: FunctionToIntegrate
    real, intent(in) :: x
    end function FunctionToIntegrate
end interface

area = computeIntegral( FunctionToIntegrate, boundInf, boundSup )

end


function FunctionToIntegrate( x )
implicit none

real             :: FunctionToIntegrate
real, intent(in) :: x

FunctionToIntegrate = x

end function FunctionToIntegrate


function computeIntegral( func, boundInf, boundSup )
implicit none

real, intent(in) :: boundInf, boundSup

interface
    function func( x )
    implicit none
    real             :: func
    real, intent(in) :: x
    end function func
end interface

...