Difference between revisions of "Modern Fortran"
m (→Hands on 1) |
|||
(11 intermediate revisions by the same user not shown) | |||
Line 13: | Line 13: | ||
* [[#A Brief History of Fortran|Intro, History]] (10 min) | * [[#A Brief History of Fortran|Intro, History]] (10 min) | ||
* [[#New Format, New Syntax|New syntax]] (30 min) | * [[#New Format, New Syntax|New syntax]] (30 min) | ||
− | * [[#Hands | + | * [[#Hands On 1|Hands On 1]](60 min) |
− | * Functions, Modules (45 min) | + | * [[#Procedures and Modules|Functions, Modules]] (45 min) |
− | * Hands | + | * [[#Hands On 2|#Hands On 2]] (30 min) |
* Lunch (1 hr) | * Lunch (1 hr) | ||
* New Array Features (15 min) | * New Array Features (15 min) | ||
Line 128: | Line 128: | ||
==Free Format: some highlights== | ==Free Format: some highlights== | ||
{| | {| | ||
+ | |- valign="top" | ||
| | | | ||
* Columns no longer significant; can start at left margin | * Columns no longer significant; can start at left margin | ||
Line 632: | Line 633: | ||
|} | |} | ||
− | ==Hands | + | ==Hands On 1== |
* In workedexample/f77 is a simplified, F77ized version of a fluid-dynamics code from Ue-Li Pen, CITA, U of Toronto ([http://www.cita.utoronto.ca/~pen/MHD/ http://www.cita.utoronto.ca/~pen/MHD/]) | * In workedexample/f77 is a simplified, F77ized version of a fluid-dynamics code from Ue-Li Pen, CITA, U of Toronto ([http://www.cita.utoronto.ca/~pen/MHD/ http://www.cita.utoronto.ca/~pen/MHD/]) | ||
* For the purposes of this class, we've turned it from a perfectly good f90 code to something that looks more like something your supervisor would dust off and give to you. | * For the purposes of this class, we've turned it from a perfectly good f90 code to something that looks more like something your supervisor would dust off and give to you. | ||
Line 642: | Line 643: | ||
* ~1 hr (for getting logged in and everything working) | * ~1 hr (for getting logged in and everything working) | ||
− | =Procedures and | + | =Procedures and Modules= |
{| | {| | ||
| | | | ||
Line 652: | Line 653: | ||
==Modules== | ==Modules== | ||
{| | {| | ||
+ | |- valign="top" | ||
| | | | ||
* Easiest to show by example | * Easiest to show by example | ||
Line 658: | Line 660: | ||
* “Use” goes before “implicit none” | * “Use” goes before “implicit none” | ||
| | | | ||
− | samples/procedures/simplemod/simplemod.f90 | + | <source lang="fortran"> |
+ | module gravity | ||
+ | implicit none | ||
+ | real, parameter :: G = 6.67e-11 ! MKS units | ||
+ | |||
+ | contains | ||
+ | real function gravforce(x1,x2,m1,m2) | ||
+ | implicit none | ||
+ | real, dimension(3), intent(in) :: x1,x2 | ||
+ | real, intent(in) :: m1, m2 | ||
+ | real :: dist | ||
+ | |||
+ | dist = sqrt(sum((x1-x2)**2)) | ||
+ | gravforce = G * m1 * m2 / dist**2 | ||
+ | end function gravforce | ||
+ | end module gravity | ||
+ | |||
+ | program simplemod | ||
+ | use gravity | ||
+ | implicit none | ||
+ | |||
+ | print *, 'Gravitational constant = ', G | ||
+ | print *, 'Force between 2 1kg masses at [1,0,0] & | ||
+ | &and [0,0,1] is' | ||
+ | |||
+ | print *, gravforce([1.,0.,0.],[0.,0.,1.],1.,1.) | ||
+ | |||
+ | end program simplemod | ||
+ | </source> | ||
+ | (from samples/procedures/simplemod/simplemod.f90) | ||
| | | | ||
|} | |} | ||
Line 664: | Line 695: | ||
==Compiling & Running== | ==Compiling & Running== | ||
{| | {| | ||
+ | |- valign="top" | ||
| | | | ||
* When compiling the code a gravity.mod file is created | * When compiling the code a gravity.mod file is created | ||
Line 669: | Line 701: | ||
* Not compatible between different compilers, versions. | * Not compatible between different compilers, versions. | ||
| | | | ||
+ | <source lang="bash"> | ||
+ | $ ls | ||
+ | simplemod.f90 | ||
+ | |||
+ | $ gfortran -o simplemod simplemod.f90 -Wall | ||
+ | |||
+ | $ ls | ||
+ | gravity.mod simplemod simplemod.f90 | ||
+ | |||
+ | $ ./simplemod | ||
+ | Gravitational constant = 6.6700000E-11 | ||
+ | Force between 2 1kg masses at [1,0,0] and [0,0,1] is | ||
+ | 3.3350003E-11 | ||
+ | </source> | ||
|} | |} | ||
==Modules== | ==Modules== | ||
{| | {| | ||
+ | |- valign="top" | ||
| | | | ||
* function gravforce can “see” the modulewide parameter defined above. | * function gravforce can “see” the modulewide parameter defined above. | ||
* So can main program, through use statement. | * So can main program, through use statement. | ||
| | | | ||
− | + | <source lang="fortran"> | |
− | + | <source lang="fortran"> | |
+ | module gravity | ||
+ | implicit none | ||
+ | real, parameter :: G = 6.67e-11 ! MKS units | ||
− | ==use module, only :== | + | contains |
+ | real function gravforce(x1,x2,m1,m2) | ||
+ | implicit none | ||
+ | real, dimension(3), intent(in) :: x1,x2 | ||
+ | real, intent(in) :: m1, m2 | ||
+ | real :: dist | ||
+ | |||
+ | dist = sqrt(sum((x1-x2)**2)) | ||
+ | gravforce = G * m1 * m2 / dist**2 | ||
+ | end function gravforce | ||
+ | end module gravity | ||
+ | |||
+ | program simplemod | ||
+ | use gravity | ||
+ | implicit none | ||
+ | |||
+ | print *, 'Gravitational constant = ', G | ||
+ | print *, 'Force between 2 1kg masses at [1,0,0] & | ||
+ | &and [0,0,1] is' | ||
+ | |||
+ | print *, gravforce([1.,0.,0.],[0.,0.,1.],1.,1.) | ||
+ | |||
+ | end program simplemod | ||
+ | </source> | ||
+ | (from samples/procedures/simplemod/simplemod.f90) | ||
+ | |} | ||
+ | |||
+ | ==use module, only :== | ||
{| | {| | ||
+ | |- valign="top" | ||
+ | | | ||
* Best practice is to only pull in from the module what you need | * Best practice is to only pull in from the module what you need | ||
* Otherwise, everything. | * Otherwise, everything. | ||
Line 687: | Line 766: | ||
* (Note syntax for continuation of a string...) | * (Note syntax for continuation of a string...) | ||
| | | | ||
+ | <source lang="fortran"> | ||
+ | module gravity | ||
+ | implicit none | ||
+ | real, parameter :: G = 6.67e-11 ! MKS units | ||
+ | |||
+ | contains | ||
+ | real function gravforce(x1,x2,m1,m2) | ||
+ | implicit none | ||
+ | real, dimension(3), intent(in) :: x1,x2 | ||
+ | real, intent(in) :: m1, m2 | ||
+ | real :: dist | ||
+ | |||
+ | dist = sqrt(sum((x1-x2)**2)) | ||
+ | gravforce = G * m1 * m2 / dist**2 | ||
+ | end function gravforce | ||
+ | end module gravity | ||
+ | |||
+ | program simplemod2 | ||
+ | use gravity, only : G, gravforce | ||
+ | implicit none | ||
+ | |||
+ | print *, 'Gravitational constant = ', G | ||
+ | print *, 'Force between 2 1kg masses at [1,0,0] & | ||
+ | &and [0,0,1] is' | ||
+ | |||
+ | print *, gravforce([1.,0.,0.],[0.,0.,1.],1.,1.) | ||
+ | |||
+ | end program simplemod2 | ||
+ | </source> | ||
samples/procedures/simplemod/simplemod2.f90 | samples/procedures/simplemod/simplemod2.f90 | ||
|} | |} | ||
Line 692: | Line 800: | ||
==Modules usually get their own files== | ==Modules usually get their own files== | ||
{| | {| | ||
+ | |- valign="top" | ||
| | | | ||
* For encapsulation | * For encapsulation | ||
Line 699: | Line 808: | ||
* (Main program hasn’t changed much). | * (Main program hasn’t changed much). | ||
| | | | ||
− | samples/procedures/multifilemod/gravity.f90 | + | <source lang="fortran"> |
− | |} | + | module gravity |
+ | implicit none | ||
+ | private | ||
+ | |||
+ | character (len=8), parameter, public :: massunit="kilogram" | ||
+ | character (len=8), parameter, public :: forceunit="Newton" | ||
+ | public :: gravforce | ||
+ | |||
+ | real, parameter :: G = 6.67e-11 ! MKS units | ||
+ | |||
+ | contains | ||
+ | real function distance(x1,x2) | ||
+ | implicit none | ||
+ | real, dimension(3), intent(in) :: x1, x2 | ||
+ | |||
+ | distance = sqrt(sum((x1-x2)**2)) | ||
+ | end function distance | ||
+ | |||
+ | real function gravforce(x1,x2,m1,m2) | ||
+ | implicit none | ||
+ | real, dimension(3), intent(in) :: x1,x2 | ||
+ | real, intent(in) :: m1, m2 | ||
+ | real :: dist | ||
+ | |||
+ | dist = distance(x1,x2) | ||
+ | gravforce = G * m1 * m2 / dist**2 | ||
+ | end function gravforce | ||
+ | end module gravity | ||
+ | </source> | ||
+ | (from samples/procedures/multifilemod/gravity.f90 ) | ||
+ | |} | ||
==Modules usually get their own files== | ==Modules usually get their own files== | ||
{| | {| | ||
+ | |- valign="top" | ||
+ | | | ||
* Compiling gravity.f90 now gives both an .o file (containing the code) and the .mod file as before. | * Compiling gravity.f90 now gives both an .o file (containing the code) and the .mod file as before. | ||
* Compiling the main program (multifilemod.f90) requires the .mod file. | * Compiling the main program (multifilemod.f90) requires the .mod file. | ||
| | | | ||
− | samples/procedures/multifilemod/Makefile | + | <source lang="make"> |
+ | FC=gfortran | ||
+ | FFLAGS=-O3 -Wall | ||
+ | |||
+ | multifilemod: multifilemod.o gravity.o | ||
+ | $(FC) -o $@ multifilemod.o gravity.o | ||
+ | |||
+ | %.mod: %.f90 | ||
+ | $(FC) $(FFLAGS) -c $< | ||
+ | |||
+ | multifilemod.o: multifilemod.f90 gravity.mod | ||
+ | $(FC) $(FFLAGS) -c $< | ||
+ | |||
+ | clean: | ||
+ | rm -f *.o *~ *.mod multifilemod | ||
+ | </source> | ||
+ | (from samples/procedures/multifilemod/Makefile) | ||
|} | |} | ||
==.mod needed for compilation== | ==.mod needed for compilation== | ||
{| | {| | ||
+ | |- valign="top" | ||
| | | | ||
* ...because needs the type information of the constants, | * ...because needs the type information of the constants, | ||
Line 717: | Line 875: | ||
* Can’t compile without these | * Can’t compile without these | ||
| | | | ||
− | samples/procedures/multifilemod/multifilemod.f90 | + | <source lang="fortran"> |
+ | program simplemod2 | ||
+ | use gravity, only : gravforce, massunit, forceunit | ||
+ | implicit none | ||
+ | |||
+ | print *, 'Force between 2 1 ', massunit ,' masses ', & | ||
+ | ' at [1,0,0] and [0,0,1] is' | ||
+ | |||
+ | print *, gravforce([1.,0.,0.],[0.,0.,1.],1.,1.), forceunit | ||
+ | |||
+ | end program simplemod2 | ||
+ | </source> | ||
+ | (from samples/procedures/multifilemod/multifilemod.f90) | ||
|} | |} | ||
==.o needed for linking== | ==.o needed for linking== | ||
{| | {| | ||
+ | |- valign="top" | ||
+ | | | ||
* Linking, however, doesn’t require the .mod file | * Linking, however, doesn’t require the .mod file | ||
* Only requires the .o file from the module code. | * Only requires the .o file from the module code. | ||
* .mod file analogous (but better than) .h files for C code. | * .mod file analogous (but better than) .h files for C code. | ||
| | | | ||
− | samples/procedures/multifilemod/Makefile | + | <source lang="make"> |
+ | FC=gfortran | ||
+ | FFLAGS=-O3 -Wall | ||
+ | |||
+ | multifilemod: multifilemod.o gravity.o | ||
+ | $(FC) -o $@ multifilemod.o gravity.o | ||
+ | |||
+ | %.mod: %.f90 | ||
+ | $(FC) $(FFLAGS) -c $< | ||
+ | |||
+ | multifilemod.o: multifilemod.f90 gravity.mod | ||
+ | $(FC) $(FFLAGS) -c $< | ||
+ | |||
+ | clean: | ||
+ | rm -f *.o *~ *.mod multifilemod | ||
+ | </source> | ||
+ | (from samples/procedures/multifilemod/Makefile) | ||
|} | |} | ||
==Compiling and running== | ==Compiling and running== | ||
{| | {| | ||
+ | |- valign="top" | ||
+ | | | ||
* So compile files with modules first, so those that need them have the .mod files | * So compile files with modules first, so those that need them have the .mod files | ||
* Link the .o files | * Link the .o files | ||
+ | | | ||
+ | <source lang="bash"> | ||
+ | $ make | ||
+ | gfortran -O3 -Wall -c gravity.f90 | ||
+ | gfortran -O3 -Wall -c multifilemod.f90 | ||
+ | gfortran -o multifilemod multifilemod.o gravity.o | ||
+ | reposado-$ ./multifilemod | ||
+ | Force between 2 1 kilogram masses at [1,0,0] and [0,0,1] is | ||
+ | 3.33500033E-11 Newton | ||
+ | </source> | ||
|} | |} | ||
==Private and public== | ==Private and public== | ||
− | * Not all of a module’s | + | {| |
− | content need be public | + | |- valign="top" |
− | * Can give individual | + | | |
− | items public or private | + | * Not all of a module’s content need be public |
− | attribute | + | * Can give individual items public or private attribute |
− | * “private” makes | + | * “private” makes everything private by default |
− | everything private by | + | * Allows hiding implementation specific routines |
− | default | + | | |
− | * Allows hiding | + | <source lang="fortran"> |
− | + | module gravity | |
− | + | implicit none | |
+ | private | ||
− | == | + | character (len=8), parameter, public :: massunit="kilogram" |
− | + | character (len=8), parameter, public :: forceunit="Newton" | |
− | + | public :: gravforce | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | = | + | real, parameter :: G = 6.67e-11 ! MKS units |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | contains | |
− | + | real function distance(x1,x2) | |
− | + | implicit none | |
− | function | + | real, dimension(3), intent(in) :: x1, x2 |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | = | + | distance = sqrt(sum((x1-x2)**2)) |
− | * | + | end function distance |
− | |||
− | * | ||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | real function gravforce(x1,x2,m1,m2) | |
− | + | implicit none | |
− | + | real, dimension(3), intent(in) :: x1,x2 | |
− | + | real, intent(in) :: m1, m2 | |
− | + | real :: dist | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | == | + | dist = distance(x1,x2) |
− | * | + | gravforce = G * m1 * m2 / dist**2 |
− | + | end function gravforce | |
− | * | + | end module gravity |
− | + | </source> | |
− | + | ( from samples/procedures/multifilemod/gravity.f90 ) | |
− | + | |} | |
− | |||
− | |||
− | == | + | ==Procedures== |
− | * | + | {| |
− | + | |- valign="top" | |
− | ( | + | | |
− | * | + | * We’ve already seen procedures defined in new style; let’s look more closely. |
− | + | * Biggest change: intent attribute to “dummy variables” (eg, parameters passed in/out). | |
− | + | * Again, make expectations more explicit, compiler can catch errors, optimize. | |
− | * | + | * Intent(in) - read only. Error to change. |
− | * | + | * Intent(out) - write only. Value undefined before set. |
− | + | * Intent(inout) - read/write. (eg, modify region of an array) | |
− | + | * Also documentation of a sort. | |
+ | | | ||
+ | <source lang="fortran"> | ||
+ | module procedures | ||
− | + | contains | |
− | + | function square(x) result(xsquared) | |
− | + | implicit none | |
− | + | real :: xsquared | |
− | + | real, intent(IN) :: x | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | = | + | xsquared = x*x |
− | + | end function square | |
− | |||
− | |||
− | |||
− | |||
− | * | ||
− | |||
− | |||
− | + | function cube(x) | |
− | + | implicit none | |
− | + | real :: cube | |
− | + | real, intent(IN) :: x | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | = | + | cube = x*x*x |
− | * | + | end function cube |
− | |||
− | |||
− | |||
− | * | ||
− | |||
− | end | ||
− | |||
− | |||
− | + | subroutine squareAndCube(x, squarex, cubex) | |
+ | implicit none | ||
+ | real, intent(in) :: x | ||
+ | real, intent(out) :: squarex | ||
+ | real, intent(out) :: cubex | ||
− | == | + | squarex = square(x) |
− | + | cubex = cube(x) | |
− | + | end subroutine squareAndCube | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | samples/procedures/ | + | end module procedures |
+ | </source> | ||
+ | ( from samples/procedures/funcsub/procedures.f90 ) | ||
+ | |} | ||
− | == | + | ==Functions== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | * Can | + | * Can be typed a couple of ways. |
− | * | + | * Old-style still works (real function square..) |
− | ( | + | * Give a result variable different from function name; set that, type that result (xsquared) |
− | + | * Explicitly type the function name, set that as return value (cube) | |
− | + | * Function return values don’t take intent - always out | |
− | + | | | |
+ | <source lang="fortran"> | ||
+ | function square(x) result(xsquared) | ||
+ | implicit none | ||
+ | real :: xsquared | ||
+ | real, intent(IN) :: x | ||
− | = | + | xsquared = x*x |
− | + | end function square | |
− | |||
− | |||
− | * | ||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | function cube(x) | |
− | + | implicit none | |
− | + | real :: cube | |
− | + | real, intent(IN) :: x | |
− | |||
− | |||
− | |||
− | |||
− | = | + | cube = x*x*x |
− | * | + | end function cube |
− | |||
− | |||
− | |||
− | |||
− | * | ||
− | |||
− | + | subroutine squareAndCube(x, squarex, cubex) | |
− | + | implicit none | |
− | + | real, intent(in) :: x | |
− | + | real, intent(out) :: squarex | |
− | + | real, intent(out) :: cubex | |
− | |||
− | |||
− | |||
− | == | + | squarex = square(x) |
− | + | cubex = cube(x) | |
− | + | end subroutine squareAndCube | |
− | + | </source> | |
− | + | ( from samples/procedures/funcsub/procedures.f90 ) | |
− | + | |} | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | ==Procedure interfaces== | |
− | + | {| | |
− | + | |- valign="top" | |
+ | | | ||
+ | * The interface to a procedure consists of | ||
+ | ** A procedure’s name | ||
+ | ** The arguments, their names, types and all attributes | ||
+ | ** For functions, the return value name and type | ||
+ | * Eg, the procedure, with all the real code stripped out. | ||
+ | * Like a C prototype, but more detailed info | ||
+ | * .mod files contain explicit interfaces to all public module procedures. | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | function square(x) result(xsquared) | ||
+ | implicit none | ||
+ | real :: xsquared | ||
+ | real, intent(IN) :: x | ||
− | = | + | xsquared = x*x |
− | * | + | end function square |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | function cube(x) | |
+ | implicit none | ||
+ | real :: cube | ||
+ | real, intent(IN) :: x | ||
− | = | + | cube = x*x*x |
− | * | + | end function cube |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | subroutine squareAndCube(x, squarex, cubex) | |
+ | implicit none | ||
+ | real, intent(in) :: x | ||
+ | real, intent(out) :: squarex | ||
+ | real, intent(out) :: cubex | ||
− | == | + | squarex = square(x) |
− | + | cubex = cube(x) | |
− | + | end subroutine squareAndCube | |
− | + | </source> | |
− | + | ( from samples/procedures/funcsub/procedures.f90 ) | |
− | + | |} | |
− | |||
− | |||
− | samples/ | ||
− | == | + | ==Procedure interfaces== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | * | + | * To see where interfaces become necessary, consider this sketch of a routine to do trapezoid-rule integration |
− | + | * We want to integrate a passed-in function f, but we don’t know anything about it - type, # of arguments, etc. | |
− | * | + | * Need to “type” f the same way you do with xlo, xhi, n. |
− | + | * You do that for procedures with interfaces | |
− | + | | | |
− | + | <source lang="fortran"> | |
+ | function integratefx(xlo, xhi, f, n) | ||
+ | ! integrate with trapezoid rule | ||
+ | ! .... | ||
− | + | integer :: i | |
+ | real :: dx, xleft, xright | ||
− | == | + | integratefx = 0. |
− | * | + | dx = (xhi-xlo)/n |
− | + | xleft = xlo | |
− | + | do i=0, n-1 | |
− | + | xright = xleft + dx | |
− | + | integratefx = integratefx + dx*(f(xright)+f(xleft))/2. | |
− | + | xleft = xright | |
− | + | enddo | |
− | + | end function integratefx | |
− | + | </source> | |
− | + | (from samples/procedures/interface/integrate.f90 ) | |
− | |||
− | + | [[File:Wiki_trapezoidal.png|300px]] | |
− | + | (from http://en.wikipedia.org/wiki/File:Trapezoidal_rule_illustration_small.svg ) | |
− | + | |} | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | ==Procedure interfaces== | |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Define f as a parameter, give its type via an interface. | ||
+ | * Can then use it, and at compile time compiler ensures function passed in matches thisinterface. | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | function integratefx(xlo, xhi, f, n) | ||
+ | ! integrate with trapezoid rule | ||
+ | implicit none | ||
+ | real, intent(in) :: xlo, xhi | ||
+ | interface | ||
+ | function f(x) | ||
+ | implicit none | ||
+ | real :: f | ||
+ | real, intent(in) :: x | ||
+ | end function f | ||
+ | end interface | ||
+ | integer, intent(in) :: n | ||
+ | real :: integratefx | ||
− | + | integer :: i | |
− | + | real :: dx, xleft, xright | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | integratefx = 0. | |
− | + | dx = (xhi-xlo)/n | |
− | + | xleft = xlo | |
− | + | do i=0, n-1 | |
− | + | xright = xleft + dx | |
− | + | integratefx = integratefx + dx*(f(xright)+f(xleft))/2. | |
− | + | xleft = xright | |
− | + | enddo | |
− | + | end function integratefx | |
− | + | </source> | |
− | + | (from samples/procedures/interface/integrate.f90 ) | |
− | + | |} | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | ==Recursive procedures== | |
− | + | {| | |
− | == | + | |- valign="top" |
− | + | | | |
− | + | * By default, Fortran procedures cannot call themselves (recursion) | |
− | + | * Can be enabled by giving the procedure the recursive attribute | |
− | + | * Subroutines, functions | |
− | * | + | * Recursive functions '''must''' use “result” keyword to return value. |
− | + | | | |
− | + | <source lang="fortran"> | |
− | * | + | recursive function integratefx(xlo, xhi, f, tol) result(integral) |
− | + | ! integrate with trapezoid rule, simpsons rule; | |
− | + | ! if difference between two is larger than | |
− | + | ! relevant tolerance, subdivide region. | |
− | |||
− | |||
− | + | ! ...variable declarations as before... | |
− | |||
− | |||
− | |||
− | |||
− | == | + | dx = xhi-xlo |
− | + | xmid = (xlo+xhi)/2. | |
− | + | trapezoid = dx*(f(xlo)+f(xhi))/2. | |
− | * | + | simpsons = dx/6.*(f(xlo)+4.*f(xmid)+f(xhi)) |
− | + | error = abs(trapezoid-simpsons)/(0.5*(trapezoid+simpsons)) | |
− | * | ||
− | |||
− | |||
− | |||
− | = | + | if (error > tol) then |
− | + | ! too coarse; subdivide | |
− | + | integral = integratefx(xlo,xmid,f,tol) + & | |
− | + | integratefx(xmid,xhi,f,tol) | |
− | + | else | |
− | + | integral = trapezoid | |
+ | endif | ||
+ | end function integratefx | ||
+ | </source> | ||
+ | ( from samples/procedures/recursive/integrate.f90) | ||
+ | |} | ||
− | == | + | ==Pure procedures== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | + | * Procedures are pure or impure depending on whether or not they have “side effects”: | |
− | + | ** Changing things other than their dummy arguments | |
− | + | * Modifying save variables | |
− | + | * Modifying module data | |
− | + | * Printing, etc. | |
− | + | * Optimizations can be made for pure routines which can’t for impure | |
+ | * Label known-pure routines with the pure attribute. | ||
+ | * Almost all the procedures we’ve seen so far are pure. | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | pure subroutine axpy(a, x, y) | ||
+ | ! y = y + a*x | ||
+ | implicit none | ||
+ | real, intent(IN) :: a, x | ||
+ | real, intent(INOUT) :: y | ||
+ | |||
+ | y = y + a*x | ||
+ | end subroutine axpy | ||
− | ( | + | subroutine printaxpy(a, x, y) |
− | + | ! y = y + a*x | |
+ | implicit none | ||
+ | real, intent(IN) :: a, x | ||
+ | real, intent(INOUT) :: y | ||
− | samples/ | + | print *, a, '*', x, ' + ', y, & |
+ | ' = ', a*x+y | ||
+ | y = a*x + y | ||
+ | end subroutine printaxpy | ||
+ | </source> | ||
+ | (from samples/procedures/purity/purity.f90) | ||
+ | |} | ||
− | == | + | ==Optional Arguments== |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Can make arguments optional by using the optional attribute. | ||
+ | * Use present to test. | ||
+ | * Often useful to clarify calling of procedure if many parameters have sensible defaults. | ||
+ | * Avoid code duplication, wrappers of having one version of routine with default values, one with user-supplied | ||
+ | * Can’t use tol if not present; have to use another variable. | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | recursive function integratefx(xlo, xhi, f, tol) result(integral) | ||
+ | ! .... | ||
+ | real, intent(in), optional :: tol | ||
+ | ! .... | ||
− | + | real :: errtol | |
− | + | ! use parameter if passed, | |
− | + | ! else use default | |
− | + | if (present(tol)) then | |
− | ( | + | errtol = tol |
− | + | else | |
− | + | errtol = 1.e-6 | |
− | + | endif | |
− | |||
− | |||
− | |||
− | |||
− | + | ! .... | |
− | + | if (error > errtol) then | |
− | + | ! too coarse; subdivide | |
− | + | integral = integratefx(xlo,xmid,f,errtol) + & | |
− | + | integratefx(xmid,xhi,f,errtol) | |
− | + | else | |
− | + | integral = trapezoid | |
− | + | endif | |
− | + | end function integratefx | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | samples/ | + | </source> |
+ | (from samples/procedures/optional/integrate.f90) | ||
+ | |} | ||
− | == | + | ==Optional Arguments== |
− | + | {| | |
− | + | |- valign="top" | |
− | * | + | | |
− | + | * When calling the procedure, can use the optional argument or not. | |
− | * | + | * Makes sense to leave optional arguments at end - easier to figure out what’s what when it’s omitted. |
− | |||
− | |||
− | at | ||
− | |||
− | |||
− | = | + | | |
− | * | + | <source lang="fortran"> |
− | + | print *, 'Integrating using default tol' | |
− | + | approx = integratefx(0., 2*pi, sinesquared) | |
− | * | + | print *, 'Approximate integral = ', approx |
− | + | print *, 'Exact integral = ', exact | |
− | * | ||
− | |||
− | * | ||
− | |||
− | + | print *, '' | |
− | * | + | print *, 'Integrating using coarser tol' |
− | + | approx = integratefx(0., 2*pi, sinesquared, 0.01) | |
− | * | + | print *, 'Approximate integral = ', approx |
− | |||
− | * | ||
− | |||
− | * | ||
− | |||
− | |||
− | + | ! .... | |
− | + | </source> | |
− | + | (from samples/procedures/optional/optional.f90) | |
− | + | |} | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | samples/ | ||
− | == | + | ==Keyword Arguments== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | + | * To avoid ambiguity with omitted arguments - or really whenever you want - you can specify which value is which explicitly. | |
− | * | + | * Don’t have to be in order. |
− | + | * Can clarify calls of routines with many arguments | |
− | + | | | |
− | * | + | <source lang="fortran"> |
− | + | ! .... | |
− | * | + | print *, 'Integrating using still coarser tol' |
− | + | approx = integratefx(xhi=2*pi, xlo=0., tol=0.5, & | |
− | + | f=sinesquared) | |
+ | print *, 'Approximate integral = ', approx | ||
+ | </source> | ||
+ | (from samples/procedures/optional/optional.f90) | ||
+ | |} | ||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | ==Procedures and Modules Summary== | |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
+ | * Modules let you bundle procedures, constants in useful packages. | ||
+ | * Can have public, private components | ||
+ | * Compiling them generates a .mod file (needed for compiling anything that does a “use modulename”) and an .o file (where the code goes, needed to link together the program). | ||
+ | * New syntax for functions/subroutines: intent (IN/OUT/INOUT) | ||
+ | * New syntax for function return values; result or explicit typing of function in argument list. | ||
+ | * Procedures have interfaces, which are needed for (eg) passing functions | ||
+ | * Optional/keyword arguments | ||
+ | * Pure/recursive procedures | ||
+ | |} | ||
− | + | ==Hands On 2== | |
− | + | {| | |
+ | |- valign="top" | ||
+ | | | ||
+ | * In workedexamples/modules, have have pulled the PBM stuff out into a module. | ||
+ | * Do the same with the hydro routines. What needs to be private? Public? | ||
+ | * The common block (thankfully) only contains constants, can make those module parameters | ||
+ | * ~30 min | ||
+ | |} | ||
− | + | =Fortran arrays= | |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | + | * Fortran made for dealing with scientific data | |
− | * | + | * Arrays built into language |
− | associated, | + | * The type information associated with an array includes rank (# of dimension), size, element type, stride.. |
− | + | * Enables powerful optimizations, programmer-friendly features. | |
− | + | * Can be manipulated like simple scalar variables | |
− | * Can | + | * Elementwise addition, multiplication.. |
− | + | | | |
− | + | <source lang="fortran"> | |
− | + | program basicarrays | |
− | + | implicit none | |
− | + | integer, dimension(5) :: a, b, c | |
+ | integer :: i | ||
− | + | a = [1,2,3,4,5] | |
− | + | b = [(2*i+1, i=1,5)] | |
− | |||
− | |||
− | |||
− | + | print *, 'a = ', a | |
− | + | print *, 'b = ', b | |
− | |||
− | * | ||
− | |||
− | |||
− | |||
− | + | c = a+b | |
− | + | print *, 'c = ', c | |
− | + | c = a*b + 1 | |
+ | print *, 'a*b+1=', c | ||
+ | end program basicarrays | ||
+ | </source> | ||
+ | (from samples/arrays/basic.f90 ) | ||
+ | |} | ||
− | == | + | ==Array constructors== |
− | * | + | {| |
− | + | |- valign="top" | |
− | * | + | | |
− | + | * Can have array constants like numerical constants | |
− | + | * use [] or (/ /), then comma-separated list of values. | |
− | + | * Implied do loops can be used in constructors | |
+ | * (Implied do-loop variables have to be defined) | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | x = [1,2,3,4,5] | ||
+ | x = (/1,2,3,4,5/) | ||
+ | x = [ (i,i=1,5)] | ||
+ | a = [ ((i*j,j=1,3),i=1,5)] | ||
+ | </source> | ||
+ | |} | ||
− | real, | + | ==Elementwise operations== |
− | real, | + | {| |
− | + | |- valign="top" | |
− | + | | | |
+ | * Elementwise operations can be */+-, or application of an elemental function. | ||
+ | * Math intrinsics are all elemental - applied to array, applies to every element. | ||
+ | * Order of execution undefined - allows vectorization, parallelization. | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | program elementwise | ||
+ | implicit none | ||
+ | real, dimension(10) :: x,y,z | ||
+ | integer :: i | ||
+ | real, parameter:: pi = 4.*atan(1.) | ||
− | x | + | x = [(2*pi*(i-1)/9.,i=1,10)] |
− | + | ||
+ | y = sin(x) | ||
+ | z = x*x | ||
− | + | print *, x | |
− | + | print *, y | |
− | + | print *, z | |
− | * | + | end program elementwise |
− | + | </source> | |
− | + | (from samples/arrays/elementwise.f90 ) | |
− | + | |} | |
− | |||
− | + | ==Elemental Functions== | |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
+ | * User can create their own elemental functions | ||
+ | * Label any scalar function with “elemental” should (until recently, must) be pure, so can be applied everywhere at same time. | ||
+ | * Can be faster than loop. | ||
+ | * Can also take multiple arguments: eg z = addsquare(x,y) | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | program elementalfn | ||
+ | implicit none | ||
+ | real, dimension(10) :: x,y,z | ||
+ | integer :: i | ||
+ | real, parameter:: pi = 4.*atan(1.) | ||
− | x | + | x = [(2*pi*(i-1)/9.,i=1,10)] |
− | + | ||
+ | y = sinesquared(x) | ||
+ | z = sin(x)*sin(x) | ||
− | + | print *, x | |
− | * | + | print *, y |
− | + | print *, z | |
− | + | print *,z(::3) | |
− | + | contains | |
− | * | + | elemental function sinesquared(x) |
− | + | implicit none | |
− | + | real :: sinesquared | |
− | + | real, intent(in) :: x | |
− | |||
− | |||
− | + | sinesquared = sin(x)**2 | |
− | + | end function sinesquared | |
− | + | end program elementalfn | |
− | + | </source> | |
− | + | (from samples/arrays/elemental.f90 ) | |
+ | |} | ||
− | == | + | ==Array comparisons== |
− | + | {| | |
− | + | |- valign="top" | |
+ | | | ||
+ | * Array comparisons return an array of logicals of the same size of the arrays. | ||
+ | * Can use any and all to see if any or all of those logicals are true. | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | program comparearrays | ||
+ | implicit none | ||
+ | integer, dimension(5) :: a, b | ||
+ | integer :: i | ||
− | = | + | a = [1,2,3,4,5] |
− | + | b = [(2*i-3, i=1,5)] | |
− | |||
− | |||
− | |||
− | |||
− | * | ||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | print *, 'A = ', a | |
+ | print *, 'B = ', b | ||
− | + | if (any(a > b)) then | |
+ | print *, 'An A is larger than a B' | ||
+ | endif | ||
+ | if (all(a > b)) then | ||
+ | print *, 'All As ares larger than Bs' | ||
+ | else if (all(b > a)) then | ||
+ | print *, 'All Bs are larger than As' | ||
+ | else | ||
+ | print *, 'A, B values overlap' | ||
+ | endif | ||
− | + | end program comparearrays | |
− | + | </source> | |
− | + | (from samples/arrays/compare.f90) | |
− | + | |} | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | ==Array masks== | |
− | + | {| | |
− | + | |- valign="top" | |
+ | | | ||
+ | * These logical arrays can be used to mask several operations | ||
+ | * Only do sums, mins, etc where the mask is true | ||
+ | * Eg, only pick out positive values. | ||
+ | * Many array intrinsics have this mask option | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | program mask | ||
+ | implicit none | ||
+ | integer, dimension(10) :: a | ||
+ | logical, dimension(10) :: pos | ||
+ | integer :: i | ||
− | + | a = [(2*i-7, i=1,10)] | |
− | + | pos = (a > 0) | |
− | |||
− | + | print '(A,10(I4,1X))','A = ', a | |
+ | print *,'# of positive values: ', count(pos) | ||
+ | print *,'Sum of positive values: ', sum(a,pos) | ||
+ | print *,'Minimum positive value: ', minval(a,pos) | ||
− | 3 | + | end program mask |
+ | </source> | ||
+ | ( from samples/arrays/mask.f90 ) | ||
+ | <source lang="bash"> | ||
+ | $ ./mask | ||
+ | A = -5 -3 -1 1 3 5 7 9 11 13 | ||
+ | # of positive values: 7 | ||
+ | Sum of positive values: 49 | ||
+ | Minimum positive value: 1 | ||
+ | </source> | ||
+ | |} | ||
− | + | ==Where construct== | |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * The where construct can be used to easily manipulate sections of array based on arbitrary comparisons. | ||
+ | * Where construct => for whatever indices the comparison is true, set values as follow; otherwise, set other values. | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | program wherearray | ||
+ | implicit none | ||
+ | real, dimension(6) :: a, diva | ||
+ | integer :: i | ||
− | + | a = [(2*i-6, i=1,6)] | |
+ | where (a /= 0) | ||
+ | diva = 1/a | ||
+ | elsewhere | ||
+ | diva = -999 | ||
+ | endwhere | ||
− | 6 | + | print *,'a = ' |
+ | print '(6(F8.3,1X))',a | ||
+ | print *,'1/a = ' | ||
+ | print '(6(F8.3,1X))',diva | ||
− | + | end program wherearray | |
+ | </source> | ||
+ | (from samples/arrays/where.f90) | ||
+ | <source lang="bash"> | ||
+ | $ ./where | ||
+ | a = | ||
+ | -4.000 -2.000 0.000 2.000 4.000 6.000 | ||
+ | 1/a = | ||
+ | -0.250 -0.500 -999.000 0.500 0.250 0.167 | ||
+ | </source> | ||
+ | |} | ||
− | == | + | ==Forall construct== |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Forall is an array assignment statement | ||
+ | * Each line in forall has to be independent. All done “at once” - no guarantees as to order | ||
+ | * If (say) 2 lines in the forall, all of the first line is done, then all of the second. | ||
+ | * Any functions called must be pure | ||
+ | * Can be vectorized or parallelized by compiler | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | program forallarray | ||
+ | implicit none | ||
+ | integer, dimension(6,6) :: a | ||
+ | integer :: i,j | ||
− | + | a = -999 | |
+ | forall (i=1:6, j=1:6, i/=j) | ||
+ | a(i,j) = i-j | ||
+ | endforall | ||
− | == | + | do i=1,6 |
− | + | print '(6(I5,1X))',(a(i,j),j=1,6) | |
− | + | enddo | |
− | + | end program forallarray | |
− | + | </source> | |
− | + | (from samples/arrays/forall.f90) | |
− | + | <source lang="bash"> | |
+ | $ ./forall | ||
+ | -999 -1 -2 -3 -4 -5 | ||
+ | 1 -999 -1 -2 -3 -4 | ||
+ | 2 1 -999 -1 -2 -3 | ||
+ | 3 2 1 -999 -1 -2 | ||
+ | 4 3 2 1 -999 -1 | ||
+ | 5 4 3 2 1 -999 | ||
+ | </source> | ||
+ | |} | ||
− | == | + | ==Array Sections== |
− | * | + | {| |
− | + | |- valign="top" | |
− | + | | | |
− | + | * Generalization of array indexing | |
− | * | + | * Familiar to users of Matlab, IDL, Python.. |
− | + | * Can use “slices” of an array using “index triplet” | |
− | + | * [start]:[end][:step] | |
− | + | * Default start=1, default end=size, default step=1. | |
− | + | * Can be used for each index of multid array | |
− | + | | | |
+ | <source lang="fortran"> | ||
− | + | a([start]:[end][:step]) | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | = | + | a = [1,2,3,4,5,6,7,8,9,10] |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | == | + | a(7:) == [7,8,9,10] |
− | + | a(:3) == [1,2,3] | |
− | + | a(2:4) == [2,3,4] | |
− | + | a(::3) == [1,4,7,10] | |
− | + | a(2:4:2) == [2,4] | |
− | ( | + | a(2) == 2 |
− | + | a(:) == [1,2,3,4,5,6,7,8,9,10] | |
− | + | </source> | |
− | + | |} | |
− | |||
− | |||
− | |||
− | |||
− | == | ||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | == | + | ==Array Sections== |
− | * | + | {| |
− | + | |- valign="top" | |
− | + | | | |
− | + | * This sort of thing is very handy in numerical computation | |
− | + | * Replace do-loops with clearer, shorter, possibly vectorized array operations | |
− | + | * Bigger advantage for multidimensional arrays | |
− | + | | | |
− | + | <source lang="fortran"> | |
− | + | program derivative | |
+ | implicit none | ||
+ | real, dimension(10) :: x | ||
+ | real, dimension(9) :: derivx | ||
+ | integer :: i | ||
+ | real, parameter:: pi = 4.*atan(1.), h=1. | ||
− | + | x = [(2*pi*(i-1)/9.,i=1,10)] | |
+ | |||
+ | derivx = ((x(2:10)-x(1:9))/h) | ||
+ | print *, derivx | ||
− | == | + | do i=1,9 |
− | + | derivx(i) = (x(i+1)-x(i))/h | |
− | + | enddo | |
− | + | print *, derivx | |
− | + | end program derivative | |
− | + | </source> | |
− | * | + | (from samples/arrays/derivative.f90) |
− | + | |} | |
− | |||
− | samples/ | ||
− | == | + | ==Array Sections== |
{| | {| | ||
+ | |- valign="top" | ||
| | | | ||
− | * | + | * The previous sorts of array sections - shifting things leftward and rightward - are so common there are intrinsics for them |
− | * | + | * positive shift shifts elements leftwards (or array bounds rightwards). |
+ | * cshift does circular shift - shifting off the end of the array “wraps around”. | ||
+ | * eoshift fills with zeros, or optional filling. | ||
+ | * Can work on given dimension | ||
| | | | ||
− | + | <source lang="fortran"> | |
+ | |||
+ | |||
+ | a = [1,2,3,4,5] | ||
+ | |||
+ | cshift(a,1) == [2,3,4,5,1] | ||
+ | cshift(a,-1) == [5,1,2,3,4] | ||
+ | eoshift(a,1) ==[2,3,4,5,0] | ||
+ | eoshift(a,-1)==[0,1,2,3,4] | ||
+ | </source> | ||
|} | |} | ||
− | == | + | ==Other important array intrinsics== |
{| | {| | ||
− | + | |- valign="top" | |
− | |||
− | |||
− | |||
| | | | ||
− | + | * minval/maxval - finds min, max element in an array. | |
− | + | * minloc/maxloc - finds location of min/max element | |
+ | * product/sum - returns product/sum of array elements | ||
+ | * reshape - Adjusts shape of array data. Eg: | ||
+ | |||
+ | <source lang="fortran"> | ||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | == | + | 1,4 |
− | + | reshape([1,2,3,4,5,6],[3,2]) == 2,5 | |
− | + | 3,6 | |
− | + | </source> | |
− | * | + | |} |
− | * | + | |
− | to | + | ==Linear algebra in Fortran== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
+ | * Comes built in with transpose, matmul, dot_product for dealing with arrays. | ||
+ | * matmul also does matrix-vector multiplication | ||
+ | * Either use these or system-provided BLAS libraries - never, ever write yourself. | ||
+ | * Consider the following timings for a matrix multiplication - a naive triple loop, a clal to that matmul intrinsic, or an explicit call to a BLAS package. | ||
+ | * factor of 100 difference! | ||
+ | * Note that you can build gfortran to use fast (BLAS) routines for intrinsics... | ||
+ | | | ||
+ | |} | ||
− | == | + | ==Matrix Multiplication Times== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | + | <source lang="fortran"> | |
− | + | !... | |
− | + | ||
+ | print *, 'Experiment with matrix size ', n | ||
+ | print *, 'Times in seconds.' | ||
− | a | + | allocate(a(n,n)) |
− | + | allocate(b(n,n)) | |
− | + | allocate(c(n,n)) | |
− | + | call random_number(a) | |
+ | call random_number(b) | ||
− | == | + | call tick(starttime) |
− | * | + | do j=1,n |
− | + | do i=1,n | |
− | + | c(i,j) = 0. | |
− | + | do k=1,n | |
− | + | c(i,j) = c(i,j)+a(i,k)*b(k,j) | |
− | + | enddo | |
− | + | enddo | |
+ | enddo | ||
+ | looptime = tock(starttime) | ||
− | + | call tick(starttime) | |
− | = | + | c = matmul(a,b) |
− | + | matmultime = tock(starttime) | |
− | + | call tick(starttime) | |
− | + | call sgemm('N','N',n,n,n,1.,a,n,b,n,0.,c,n) | |
− | + | sgemmtime = tock(starttime) | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | print *, 'Triple-loop time: ', looptime | |
− | * | + | print *, 'matmul intrinsic time: ', matmultime |
− | + | print *, 'SGEMM lapack call time:', sgemmtime | |
− | + | !... | |
− | * | + | </source> |
− | + | | | |
− | + | <source lang="bash"> | |
− | + | $ ./matmul 2500 | |
− | + | Experiment with matrix size | |
− | + | 2500 | |
− | + | Triple-loop time: | |
− | samples/ | + | 149.63400 |
+ | matmul intrinsic time: | ||
+ | 10.370000 | ||
+ | SGEMM lapack call time: | ||
+ | 1.4809999 | ||
+ | </source> | ||
+ | (gfortran 4.6, compiled -O3 -march=native using Intel MKL 10.3 for sgemm) | ||
+ | (program from samples/arrays/matmul.f90) | ||
+ | |} | ||
− | == | + | ==Linear algebra in Fortran== |
− | * | + | {| |
− | + | |- valign="top" | |
− | + | | | |
− | + | * Things like transposes work, too: | |
− | + | * see samples/arrays/matrix.f90 | |
− | + | <source lang="fortran"> | |
+ | program matvec | ||
+ | implicit none | ||
+ | integer, dimension(4,5) :: a | ||
+ | integer, dimension(5,4) :: at | ||
+ | integer, dimension(4,4) :: aat | ||
+ | integer :: i | ||
− | + | a = reshape([(i,i=1,4*5)],[4,5]) | |
− | + | at = transpose(a) | |
+ | print *,'A = ' | ||
+ | call printmat(a) | ||
+ | print *,'A^T = ' | ||
+ | call printmat(at) | ||
− | == | + | aat = matmul(a,at) |
− | + | print *,'A . A^T = ' | |
− | + | call printmat(aat) | |
− | + | !... | |
− | + | </source> | |
− | + | | | |
+ | <source lang="bash"> | ||
+ | $ ./matrix | ||
+ | A = | ||
+ | 1 5 9 13 17 | ||
+ | 2 6 10 14 18 | ||
+ | 3 7 11 15 19 | ||
+ | 4 8 12 16 20 | ||
+ | A^T = | ||
+ | 1 2 3 4 | ||
+ | 5 6 7 8 | ||
+ | 9 10 11 12 | ||
+ | 13 14 15 16 | ||
+ | 17 18 19 20 | ||
+ | A . A^T = | ||
+ | 565 610 655 700 | ||
+ | 610 660 710 760 | ||
+ | 655 710 765 820 | ||
+ | 700 760 820 880 | ||
+ | A . A^T subarray = | ||
+ | 610 660 710 | ||
+ | 655 710 765 | ||
+ | 700 760 820 | ||
+ | </source> | ||
+ | |} | ||
− | + | ==Array sizes and Assumed Shape== | |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Printmat routine here is interesting - don’t pass (a,rows,cols), just a. | ||
+ | * Can assume a rank-2 array, and get size at runtime. | ||
+ | * Simplifies call, and eliminates possible inconsistency: what if rows, cols is wrong? | ||
+ | * size(array,dim) gets the size of array in the dim dimension. | ||
+ | * Assumed shape arrays (eg, dimension(:,:)) much better than older ways of passing arrays: | ||
− | + | integer nx, ny | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | integer a(nx,ny) | |
− | |||
− | + | or worse, | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | integer a(*,ny) | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | * | ||
− | |||
− | |||
− | + | * Information is thrown away, possibility of inconsistency. | |
− | * | + | * Here, (:,:) means we know the rank, but don’t know the size yet. |
− | + | | | |
− | * | + | <source lang="fortran"> |
− | + | subroutine printmat(a) | |
− | + | implicit none | |
+ | integer, dimension(:,:) :: a | ||
+ | integer :: nr, nc, i, j | ||
− | == | + | nr = size(a,1) |
− | + | nc = size(a,2) | |
− | + | do i=1,nr | |
− | + | print '(99(I4,1X))', (a(i,j), j=1,nc) | |
− | + | enddo | |
− | + | end subroutine printmat | |
− | + | </source> | |
− | + | (from samples/arrays/matrix.f90) | |
− | + | |} | |
− | samples/ | ||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | == | + | ==Allocatable Arrays== |
− | + | {| | |
− | * | + | |- valign="top" |
− | + | | | |
− | + | * So far, all our programs have had fixed-size arrays, set at compile time. | |
− | + | * To change problem size, have to edit code, recompile. | |
− | * | + | * Has some advantages (optimization, determinism) but very inflexible. |
− | + | * Would like to be able to request memory at run time, make array of desired size. | |
− | + | * Allocatable arrays are arguably most important addition to Fortran. | |
+ | |} | ||
− | == | + | ==Allocate(), Deallocate()== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | + | * Give array a deferred size (eg, dimension(:)) and the attribute allocatable. | |
− | + | * When time to allocate it, use allocate(a(n)). | |
− | * | + | * Deallocate with deallocate(a). |
− | + | * In between, arrays can be used as any other array. | |
− | * | + | * If allocation fails (not enough memory available for request), program will exit. |
− | + | * Can control this by checking for an optional error code, allocate(a(n),stat=ierr) | |
− | * | + | * Can then test if ierr>0 (failure condition) and handle gracefully. |
− | + | * In scientific programming, the default behaviour is often fine, if abrupt - you either have enough memory to run the problem, or you don’t. | |
− | + | | | |
− | + | <source lang="fortran"> | |
− | + | program allocarray | |
− | + | implicit none | |
− | + | integer :: i, n | |
− | + | integer, dimension(:), allocatable :: a | |
− | |||
− | |||
− | |||
− | + | n = 10 | |
− | + | allocate(a(n)) | |
− | + | a = [(i, i=2,20,2)] | |
− | + | print *,'A = ' | |
− | + | print *,a | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | deallocate(a) | |
− | + | end program allocarray | |
− | + | </source> | |
− | + | (from samples/arrays/allocatable.f90 ) | |
− | + | |} | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | ==get_command_argument()== | |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | + | * Previous version still depended on a compiled-in number. | |
− | + | * Can read from file or from console, but Fortran now has standard way to get command-line arguments | |
+ | * Get the count of arguments, and if there’s at least one argument there, get it, read it as integer, and allocate array. | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | program allocarray2 | ||
+ | implicit none | ||
+ | integer :: i, n | ||
+ | integer, dimension(:), allocatable :: a | ||
+ | character(len=30) :: arg | ||
− | + | if (command_argument_count() < 1) then | |
+ | print *,'Use: allocatable N, '//& | ||
+ | ' where N is array size.' | ||
+ | stop | ||
+ | endif | ||
− | + | call get_command_argument(1, arg) | |
− | + | read( arg,'(I30)'), n | |
− | + | ||
− | + | print *,'Allocating array of size ', n | |
− | + | allocate(a(n)) | |
− | * | ||
− | |||
− | |||
− | |||
− | |||
− | samples/ | + | a = [(i,i=1,n)] |
+ | print *, a | ||
+ | |||
+ | deallocate(a) | ||
+ | end program allocarray2 | ||
+ | </source> | ||
+ | (from samples/arrays/allocatable2.f90) | ||
+ | <source lang="bash"> | ||
+ | $ ./allocatable2 | ||
+ | Use: allocatable N, where N is array size. | ||
− | + | $ ./allocatable2 3 | |
− | + | Allocating array of size 3 | |
− | + | 1 2 3 | |
− | |||
− | |||
− | + | $ ./allocatable2 5 | |
+ | Allocating array of size 5 | ||
+ | 1 2 3 4 5 | ||
+ | </source> | ||
+ | |} | ||
− | == | + | ==Hands on #3== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | * | + | * Use array functionality to simplify hydro code -- don't need to pass, array size, and can simplify mathematics using array operations. |
− | + | * In workedexamples/arrays, have modified hydro to allocate u, and pbm to just take array. | |
− | + | * Do the same with the fluid dynamic routines in solver.f90 | |
− | + | * ~30 min | |
− | + | | | |
− | * | + | |} |
− | |||
− | |||
− | + | =Fortran Pointers= | |
− | + | {| | |
− | = | + | |- valign="top" |
− | + | | | |
− | + | * Pointers, or references, refer to another variable. | |
− | + | * Eg, p does not contain a real value, but a reference to another real variable. | |
− | + | * Once associated with another variable, can read/write to it as if it were stored “in” p. | |
− | * | + | | |
− | + | <source lang="fortran"> | |
+ | real, target :: x = 3.2 | ||
+ | real, pointer:: p | ||
+ | p => x | ||
− | + | p | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | x | |
− | + | 3.2 | |
− | + | </source> | |
− | + | samples/pointers/ptr1.f90 | |
+ | |} | ||
− | + | ==Fortran Pointers== | |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Pointers are either associated, null, or undefined; start out life undefined. | ||
+ | * Can associate them to a variable with => , or mark them as not associated with any valid variable by pointing it to null(). | ||
+ | * Reading value from or writing value to a null pointer will cause errors, probably crash. | ||
+ | * Fortran pointers can’t point just anywhere. | ||
+ | * Must reference a variable with the same type, that has the target attribute. | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | real, target :: x = 3.2 | ||
+ | real, pointer:: p | ||
+ | p => null() | ||
+ | p | ||
− | == | + | x |
+ | </source> | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==Fortran Pointers== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Pointers can reference other pointers. | ||
+ | * Actually references what they’re pointing to. | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | real, target :: x = 3.2 | ||
+ | real, pointer:: p1, p2 | ||
+ | p1 => x | ||
+ | p2 => p1 | ||
+ | p1 | ||
+ | p2 | ||
− | + | x | |
+ | 3.2 | ||
+ | </source> | ||
+ | |} | ||
− | == | + | ==Allocating a pointer== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | + | * Pointer doesn’t necessarily have to have another variable to target | |
− | + | * Can allocate memory for p to point to that does not belong to any other pointer. | |
− | + | * Must deallocate it when done | |
− | + | | | |
− | + | <source lang="fortran"> | |
− | + | real, pointer:: p | |
− | ( | + | allocate(p) |
− | + | p = 7.9 | |
− | + | p | |
− | + | 7.9 | |
− | + | | | |
− | + | </source> | |
+ | samples/pointers/ptr2.f90 | ||
+ | |} | ||
− | == | + | ==What are they good for? (1)== |
− | + | {| | |
− | == | + | |- valign="top" |
− | + | | | |
− | + | * Pointers are essential for creating, maintaining dynamic data structures | |
− | + | * Linked lists, trees, heaps.. | |
− | + | * Some of these can be sort-of implemented in arrays, but very awkward | |
− | * | + | * Adaptive meshes, treebased particle solvers need these structures. |
− | + | | | |
− | + | http://en.wikipedia.org/wiki/File:Singly-linked-list.svg | |
− | + | http://en.wikipedia.org/wiki/File:Max-Heap.svg | |
− | + | |} | |
+ | |||
+ | ==What are they good for? (2)== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * A pointer can be of array type, not just scalar | ||
+ | * Fortran pointers + fortran arrays are quite interesting; can create “views” of subarrays | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | real, target, dimension(7) :: x | ||
+ | real, pointer:: p(:) | ||
+ | p => x(2:6) | ||
− | + | p | |
+ | x | ||
+ | 1 | ||
− | + | 2 | |
− | |||
− | |||
− | |||
− | |||
− | + | 3 | |
− | + | 4 | |
− | |||
− | + | 5 | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | 6 | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | + | 7 | |
− | + | </source> | |
− | + | samples/pointers/views.f90 | |
− | + | |} | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | == | + | ==Hands on #4== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | * | + | * Use pointers to provide views into subsets of the arrays in solver.f90 to clarify the functions. |
− | + | * In workedexamples/pointers, have started the process with cfl, hydroflux; try tackling tvd1d, others. | |
− | + | * ~30 min | |
− | + | | | |
− | * | + | |} |
− | |||
− | |||
− | == | + | =Derived Types and Objects= |
− | + | {| | |
+ | |- valign="top" | ||
+ | | | ||
+ | * Often, groups of variables naturally go together to represent a larger structure | ||
+ | * Whenever you find yourself passing the same group of variables to several routines, a good candidate for a derived type. | ||
+ | <source lang="fortran"> | ||
+ | type griddomain | ||
+ | real :: xmin, xmax | ||
+ | real :: ymin, ymax | ||
+ | real :: nx, ny | ||
+ | real, dimension(:,:) :: u | ||
+ | endtype griddomain | ||
− | + | type(griddomain) :: g | |
− | + | g % xmin = -1 | |
+ | g % xmax = +1 | ||
+ | </source> | ||
+ | |} | ||
− | + | ==Derived Types and Objects== | |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Consider interval arithmetic (good for quantification of uncertainties, etc). | ||
+ | * An interval inherently has two values associated with it - the end points. | ||
+ | * Can make this a type. | ||
+ | * Note can access the fields in the type with “%” | ||
+ | * typename (field1val,field2val..) initializes a value of that type. | ||
+ | * Can pass values of this type to functions, etc., just like a built-in type. | ||
+ | | | ||
+ | (from samples/derivedtypes/simple/intervalmath.f90 ) | ||
+ | | | ||
+ | |} | ||
− | + | ==Procedures using types== | |
− | + | {| | |
+ | |- valign="top" | ||
+ | | | ||
+ | * Can start creating library of routines that operate on these new interval types. | ||
+ | * Procedures can take the new type as arguments, functions can return the new type. | ||
+ | | | ||
+ | (from samples/derivedtypes/intervalfunctions/intervalmath.f90 ) | ||
+ | |} | ||
− | + | ==Procedures using types== | |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Would prefer not to have to treat integer and real intervals so differently in main program | ||
+ | * Different types, but adding should be similar. | ||
+ | * Would like to be able to call “addintervals” and have language call the right subroutine, do the right thing. | ||
+ | * Similar to how intrinsics work - sin() works on any kind of real, matmult() works on integer, real, or complex matricies. | ||
+ | | | ||
+ | samples/derivedtypes/genericintervals/interval2.f90 | ||
+ | |} | ||
− | ... | + | ==Generic Interfaces== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
+ | {| | ||
+ | * Generic Interfaces | ||
+ | * addintintervals and addrealintervals share the same interface, (two input parameters, one function return), but different types. | ||
+ | * Put them behind the same interface. | ||
+ | * Now, a call to addintervals is resolved at compile time to one or the other. | ||
+ | | | ||
+ | samples/derivedtypes/genericintervals/intervalmath.f90 | ||
+ | |} | ||
− | + | ==Generic Interfaces== | |
− | + | {| | |
− | + | |- valign="top" | |
+ | | | ||
+ | * Note that everything is private except what is explicitly made public. | ||
+ | * Types are public. | ||
+ | * Generic interfaces are public. | ||
+ | * Type specific routines are not. | ||
+ | * Program using interval math sees only the generic interfaces. | ||
+ | | | ||
+ | samples/derivedtypes/genericintervals/intervalmath.f90 | ||
+ | |} | ||
− | == | + | ==Generic interfaces== |
− | + | {| | |
+ | |- valign="top" | ||
+ | | | ||
+ | * Call to addintervals or subtract intervals goes to the correct typespecific routine. | ||
+ | * As does print interval. | ||
+ | * Could create routines to add real to int interval, etc and add to the same interface. | ||
+ | | | ||
+ | samples/derivedtypes/genericintervals/interval2.f90 | ||
+ | |} | ||
− | + | ==Operator overloading== | |
− | + | {| | |
+ | |- valign="top" | ||
+ | | | ||
+ | * An infix operator is really just “syntactic sugar” for a function which takes two operands and returns a third. | ||
+ | | | ||
+ | <source lang="fortran" | ||
+ | a = b (op) c | ||
+ | => | ||
+ | function op(b,c) | ||
+ | returns a | ||
− | + | a=b | |
+ | => | ||
+ | subroutine assign(a,b) | ||
+ | </source> | ||
+ | |} | ||
− | + | ==Operator overloading== | |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
+ | * Here, we’ve defined two subroutines which set intervals based on an array - 2 ints for an integer interval, or 2 reals for a real interval | ||
+ | | | ||
+ | samples/derivedtypes/intervaloperators/intervalmath.f90 | ||
+ | |} | ||
− | ... | + | ==Generic interfaces== |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Once this is done, can use assignment operator, | ||
+ | * Or add, subtract multiply intervals. | ||
+ | * Can even compose them in complex expressions! Functions automatically composed. | ||
+ | | | ||
+ | samples/derivedtypes/intervaloperators/interval3.f90 | ||
+ | |} | ||
− | + | ==Type bound procedures== | |
− | + | {| | |
+ | |- valign="top" | ||
+ | | | ||
+ | * Types can have not only variables, but procedures. | ||
+ | * Takes us from a type to what is usually called a class. | ||
+ | | | ||
+ | (from samples/derivedtypes/intervaltypebound/intervalmath.f90 ) | ||
+ | |} | ||
− | + | ==Type bound procedures== | |
− | + | {| | |
− | + | |- valign="top" | |
+ | | | ||
+ | * Called like one accesses a field - % | ||
+ | * Operates on data of the particular variable it is invoked from | ||
+ | | | ||
+ | samples/derivedtypes/intervaltypebound/interval3.f90 | ||
+ | |} | ||
− | + | ==Type bound procedures== | |
− | + | {| | |
+ | |- valign="top" | ||
+ | | | ||
+ | * It is implicitly passed as it’s first parameter the variable itself. | ||
+ | * Can take other arguments as well. | ||
+ | | | ||
+ | samples/derivedtypes/intervaltypebound/intervalmath.f90 | ||
+ | |} | ||
− | + | ==Object oriented programming== | |
− | + | {| | |
+ | |- valign="top" | ||
+ | | | ||
+ | * F2003 onwards can do full object oriented programing. | ||
+ | * Types can be derived from other types, inheriting their fields and type-bound procedures, and extending them. | ||
+ | * Goes beyond scope of today, but this is the starting-off point. | ||
+ | | | ||
+ | |} | ||
− | == | + | =Interoperability with other languages= |
− | * | + | {| |
− | + | |- valign="top" | |
− | + | | | |
− | + | * Large scientific software now frequently uses multiple languages, either within a single code or between codes. | |
− | * | + | * Right tool for the job! |
− | + | * Need to know how to make software interact. | |
− | + | * Here we’ll look at C/Fortran code calling each other, and calling Fortran code from python. | |
− | * | + | | |
− | + | |} | |
− | == | + | ==C-interoperability== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | + | * iso_c_binding module contains definitions for interacting with C | |
− | + | * Types, ways to bind procedures to C, etc. | |
− | * | + | * Allows you to call C routines, or bind Fortran routines in a way that they can be called by C. |
− | + | | | |
− | + | |} | |
− | * | ||
− | |||
− | * | ||
− | samples/ | + | ==Calling a C routine from Fortran== |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * As with the case of calling a passed-in function, need an explicit prototype. | ||
+ | * Tells compiler what to do with “sqrtc” function when called. | ||
+ | | | ||
+ | samples/C-interop/call-c-fn/main.f90 | ||
+ | |} | ||
− | == | + | ==Calling a C routine from Fortran== |
− | * | + | {| |
− | + | |- valign="top" | |
− | * | + | | |
− | + | * BIND(C) - tells compiler/ linker will have a C-style, rather than fortran-style name in object file. | |
− | + | * Can optionally supply another name; otherwise, will default to procedure name. | |
− | + | * Here we’re calling it sqrtc to avoid Fortran sqrt() function. | |
− | * | + | * The value the function takes and returns are C doubles; that is, reals of kind(c_double). |
− | + | * Also defined: c_float, integers of kind c_int, c_long, etc. | |
− | + | | | |
+ | |} | ||
− | + | ==Calling a C routine from Fortran== | |
− | + | {| | |
− | == | + | |- valign="top" |
− | * | + | | |
− | * | + | * C function arguments by default are passed by value; Fortran by default are passed by reference. |
− | + | * Passed by value - values copied into function | |
− | + | * Passed by reference pointers to values copied in. | |
− | * | + | * value attribute for C-style arguments. |
− | + | * Compiling - just make sure to link in C library you’re calling | |
− | + | * And that’s it. | |
− | + | | | |
− | + | <source lang="bash"> | |
− | + | $ make | |
+ | gfortran -c main.f90 | ||
+ | gfortran -o main main.o -lm | ||
− | + | $ ./main | |
+ | x = | ||
+ | 2.0000000000000000 | ||
+ | C | ||
+ | sqrt(x) = 1.4142135623730951 | ||
+ | Fortran sqrt(x) = 1.4142135623730951 | ||
+ | </source> | ||
+ | |} | ||
− | == | + | ==C strings== |
− | + | {| | |
− | + | |- valign="top" | |
− | * | + | | |
− | + | * When using C strings, you have to remember that C terminates strings with a special character | |
− | + | * C_NULL_CHAR | |
− | + | * Dealing with functions that return strings is hairier, as they return a pointer, not an array. | |
− | + | | | |
− | = | + | <source lang="bash"> |
− | + | $ make | |
− | + | gfortran-mp-4.4 -c main.f90 | |
− | + | gfortran-mp-4.4 -o main main.o -lc | |
− | + | $ ./main | |
− | + | 1234 = | |
− | + | 1234 | |
− | + | </source> | |
− | + | (from samples/C-interop/c-strings/main.f90 ) | |
− | samples/ | + | |} |
− | == | + | ==Calling Fortran from C== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | + | * To write Fortran routines callable from C, bind the subroutine to a C name | |
− | + | * Can optionally give it a different name, as above | |
+ | * And make sure arguments are of C types. | ||
+ | | | ||
+ | samples/C-interop/c-call-fortran/froutine.f90 | ||
+ | |} | ||
− | == | + | ==Calling Fortran from C== |
− | * | + | {| |
− | + | |- valign="top" | |
− | * | + | | |
− | of | + | * Here, we’ll do something a little trickier and pass C dynamic arrays |
− | ( | + | * In Fortran, we accept the arguments as C pointers |
− | * | + | * We then associate them with fortran pointers to arrays of shape [nx] (1d arrays here) |
− | + | * Then we can do the usual Fortran array operations with them. | |
− | + | * The single scalar argument passed back we just treat as an intent(out) variable | |
− | + | * Of type c_float. | |
− | + | | | |
− | + | (from samples/C-interop/c-call-fortran/froutine.f90 ) | |
− | + | |} | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | * | ||
− | |||
− | |||
− | |||
− | * | ||
− | |||
− | |||
− | samples/ | + | ==More advanced== |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Handling arbitrary C code is possible | ||
+ | * Passing C structs -- create a Fortran derived type with the same fields, use BIND(C) in defining the type. | ||
+ | * C “multidimensional arrays” - have to be careful, they may not be contiguous! Follow pointers. | ||
+ | * Even taking passed function pointers to C | ||
+ | functions is possible | ||
+ | | | ||
+ | (from samples/C-interop/functionptr ) | ||
+ | |} | ||
− | == | + | ==Fortran calling C, which calls Fortran== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | + | samples/C-interop/valueref/driver.f90 | |
− | + | | | |
− | + | |} | |
− | |||
− | |||
− | |||
− | |||
− | |||
− | |||
− | samples/ | ||
− | == | + | ==samples/C-interop/valueref/croutine.c== |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | | | ||
+ | |} | ||
− | + | ==samples/C-interop/valueref/froutine.f90== | |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | | | ||
+ | |} | ||
− | + | ==$ ./main== | |
− | + | {| | |
+ | |- valign="top" | ||
+ | | | ||
+ | <source lang="bash"> | ||
+ | $ ./main | ||
+ | (1) In the FORTRAN routine before call to C | ||
+ | (1) i = | ||
+ | 15 x = | ||
+ | 2.0000000000000000 | ||
+ | (2) In the C routine, got i=15, *x=2.000000 | ||
+ | (2) Changed x | ||
+ | (3) In the FORTRAN routine called by C | ||
+ | (3) Got product p = | ||
+ | 30.000000000000000 | ||
+ | (1) In the FORTRAN routine after call to C | ||
+ | (1) i = 15 x = 3.0000000000000000 prod = 30.000000000000000 | ||
+ | </source> | ||
+ | | | ||
+ | |} | ||
− | + | ==samples/C-interop/valueref/Makefile== | |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | | | ||
+ | |} | ||
− | + | ==F2py== | |
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Comes with scipy, a widely-installed (by scientists) python package. | ||
+ | * Wraps fortran in python, allows you to easily interface. | ||
+ | * fwrap is another option | ||
+ | | | ||
+ | http://www.scipy.org/F2py | ||
+ | |} | ||
− | == | + | ==* Will only use solver module.== |
− | + | {| | |
− | + | |- valign="top" | |
− | * | + | | |
− | + | * Unfortunately, f2py isn’t quite smart enough to understand using parameters to size arrays, so global replace ‘nvars’=4 | |
− | + | * module load use.experimental gcc python/2.7.1 intel/12 | |
− | * | + | * “f2py -m hydro_solver -h hydro_solver.pyf solver.f90” |
+ | | | ||
+ | |} | ||
− | + | ==* generates the following header file (hydro_solver.pyf)== | |
− | + | {| | |
− | == | + | |- valign="top" |
− | + | | | |
− | + | * Comment out stuff we don’t need (lower-level routines) | |
− | + | * f2py -c --fcompiler=gfortran hydro_solver.pyf solver.f90 | |
− | + | | | |
− | + | |} | |
− | |||
− | |||
− | == | + | ==$ ipython== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | + | <source lang="bash"> | |
− | + | In [1]: import hydro_solver | |
− | + | In [2]: hydro_solver.solver. | |
− | + | hydro_solver.solver.cfl | |
− | + | hydro_solver.solver.gamma | |
− | + | hydro_solver.solver.hydroflux | |
− | + | hydro_solver.solver.idens | |
− | + | hydro_solver.solver.iener | |
− | + | hydro_solver.solver.imomx | |
− | |||
− | |||
− | |||
− | + | hydro_solver.solver.imomy | |
− | + | hydro_solver.solver.initialconditions | |
− | + | hydro_solver.solver.timestep | |
− | + | hydro_solver.solver.tvd1d | |
+ | hydro_solver.solver.xsweep | ||
+ | hydro_solver.solver.xytranspose | ||
− | == | + | In [2]: hydro_solver.solver.idens |
− | * | + | Out[2]: array(1, dtype=int32) |
− | * | + | In [3]: import numpy |
− | * | + | In [4]: u = hydro_solver.solver.initialconditions(100,100) |
− | + | In [5]: import pylab | |
− | + | In [6]: pylab.imshow(u[1,:,:]) | |
+ | In [7]: for i in range(100) | ||
+ | ...dt = hydro_solver.solver.timestep(u) | ||
+ | In [8]: pylab.imshow(u[1,:,:]) | ||
+ | </source> | ||
+ | |} | ||
+ | |||
+ | =Coarrays= | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * In F2008, objects can also have a “codimension”. | ||
+ | * An object with a codimension can be “seen” across processes | ||
+ | * Right now, only intel v 12 supports this | ||
+ | | | ||
+ | samples/coarrays/simple.f90 | ||
+ | |} | ||
+ | |||
+ | ==Coarrays== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | 1 | ||
+ | A | ||
B | B | ||
− | |||
− | |||
− | |||
− | + | 2 | |
− | |||
− | |||
− | |||
− | |||
− | + | A | |
+ | B | ||
− | + | N | |
− | |||
− | |||
− | |||
− | + | ... | |
− | + | A | |
− | + | B | |
− | |||
− | |||
− | + | num_images | |
+ | independant | ||
+ | processes | ||
+ | | | ||
+ | |} | ||
− | == | + | ==Coarrays== |
− | + | {| | |
− | + | |- valign="top" | |
− | + | | | |
− | + | 1 | |
− | + | ||
+ | A | ||
+ | B | ||
+ | |||
+ | 2 | ||
+ | |||
+ | A | ||
+ | B | ||
+ | |||
+ | N | ||
+ | |||
+ | ... | ||
− | + | A | |
− | + | B | |
− | |||
− | |||
− | ==Useful Resources== | + | num_images |
− | * http://fortranwiki.org/ | + | independant |
− | ** Reference source; has all standards; Fortran2003/2008 status of major compilers | + | processes |
− | * http://en.wikipedia.org/wiki/Fortran_language_features | + | |
− | ** Succinct summary of new features (spotty past F95) | + | A[1] |
− | * http://stackoverflow.com/questions/tagged/fortran | + | B[N] |
+ | |||
+ | Independent, parallel tasks can | ||
+ | “see” each other’s data as easily as an array index. | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==Sychronization== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * When accessing other processor’s data, must ensure that tasks are synchronized | ||
+ | * Don’t want to read task 1’s data before read is complete! | ||
+ | * sync all | ||
+ | | | ||
+ | samples/coarrays/broadcast.f90 | ||
+ | |} | ||
+ | |||
+ | ==Sychronization== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Other synchronization tools: | ||
+ | * sync images([..]) syncs only images in the list provided | ||
+ | * critical - only one image can enter block at a time | ||
+ | * lock - finer-grained control than critical. | ||
+ | * atomic operations. | ||
+ | | | ||
+ | samples/coarrays/broadcast.f90 | ||
+ | |} | ||
+ | |||
+ | ==1d Diffusion Eqn== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Calculate heat diffusion along a bar | ||
+ | * Finite difference; calculate second derivative using neighbours | ||
+ | * Get new temperature from old temperatures | ||
+ | 1 -2 1 | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==1d Diffusion Eqn== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Initial conditions: | ||
+ | * Use pointers to point to new, old temperatures | ||
+ | * 1..totpoints+2 (pts 1, totpoints+1 “ghost cells) | ||
+ | * Setup x, orig temperature, expected values | ||
+ | | | ||
+ | samples/coarrays/diffusion/diffusion.f90 | ||
+ | |} | ||
+ | |||
+ | ==1d Diffusion Eqn== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Evolution: | ||
+ | * Apply BCs | ||
+ | * Apply finite difference stencil to all real data points | ||
+ | | | ||
+ | samples/coarrays/diffusion/diffusion.f90 | ||
+ | |} | ||
+ | |||
+ | ==1d Diffusion Eqn== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Output calculated values and theory to file output.txt | ||
+ | * Note: Fortran2008, can use “newunit” to find, use an unused IO unit | ||
+ | | | ||
+ | samples/coarrays/diffusion/diffusion.f90 | ||
+ | |} | ||
+ | |||
+ | ==$ ./diffusion== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | [..] | ||
+ | $ gnuplot | ||
+ | [..] | ||
+ | gnuplot> plot 'output.txt' using 1:2 with points title 'Simulated', | ||
+ | 'output.txt' using 1:3 with lines title 'Theory' | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==Coarray Diffusion== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Now we’ll try this in parallel | ||
+ | * Each image runs it’s part of the problem (totpoints/num_images()) | ||
+ | * Communications is like boundary condition handling - except boundary data must be obtained from neighbouring image. | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==Coarray Diffusion== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Everything’s the same except we are only responsible for locpoints points, starting at start. | ||
+ | * Once calculated, we never need totpoints again. All arrays are of size (locpoints), etc. | ||
+ | * For simplicity, we assume here everything divides evenly. | ||
+ | | | ||
+ | samples/coarrays/diffusion/diffusion-coarray.f90 | ||
+ | |} | ||
+ | |||
+ | ==Coarray Diffusion== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Boundary exchange; if we have interior neighbours, get updated data from them so we can calculate our new data | ||
+ | * Note: can’t use pointers here, coarrays can’t be targets. | ||
+ | * Sync all around boundary exchange a little heavy handed; could just sync neighbours. | ||
+ | | | ||
+ | samples/coarrays/diffusion/diffusion-coarray.f90 | ||
+ | |} | ||
+ | |||
+ | ==Coarray Diffusion== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | temperature()[leftneigh] | ||
+ | |||
+ | temperature() | ||
+ | temperature()[rightneigh] | ||
+ | |||
+ | 1 2 3 ... | ||
+ | |||
+ | lnpts+2 | ||
+ | </source> | ||
+ | |} | ||
+ | |||
+ | ==Coarray Diffusion== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Everything else exactly the same. | ||
+ | * For I/O, we each output our own file, prefixed by our image number | ||
+ | * (eg, 1-output.txt, 2output.txt...) | ||
+ | | | ||
+ | samples/coarrays/diffusion/diffusion-coarray.f90 | ||
+ | |} | ||
+ | |||
+ | ==$ export FOR_COARRAY_NUM_IMAGES=3== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | <source lang="fortran"> | ||
+ | # 3 images | ||
+ | $ ./diffusion-coarray | ||
+ | [..] | ||
+ | $ gnuplot | ||
+ | [..] | ||
+ | gnuplot> plot '1-ics.txt' using 1:2, '2-ics.txt' using 1:2, '3-ics.txt' | ||
+ | using 1:2 | ||
+ | </source> | ||
+ | |} | ||
+ | |||
+ | ==gnuplot> plot '1-output.txt' using 1:2, '2-output.txt' using 1:2, '3output.txt' using 1:2, '1-output.txt' using 1:3 with lines title 'Theory',== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | '2-output.txt' using 1:3 with lines title 'Theory', '3-output.txt' using 1:3 | ||
+ | with lines title 'Theory' | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==Parallel Matrix Multiplication== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Consider matrix operations, where matrix is decomposed onto images in blocks | ||
+ | * A given image has corresponding blocks in A, B, responsible for that block in C. | ||
+ | |||
+ | A | ||
+ | |||
+ | B | ||
+ | X | ||
+ | C | ||
+ | = | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==Parallel Matrix Multiplication== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Block matrix multiply exactly like matrix multiply, except each operation is a submatrix operation | ||
+ | * matmul instead of “*”. | ||
+ | * For simplicity, we’ll assume square decomposition. | ||
+ | |||
+ | A | ||
+ | |||
+ | B | ||
+ | X | ||
+ | C | ||
+ | = | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==Parallel Matrix Multiplication== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Each image gets its own block of a, b, c. | ||
+ | * Note [nrows,*]. Coarrays can have any corank(eg, number of codimensions) | ||
+ | * Here, we make decomposition nrows x ncols. | ||
+ | * Can set decomposition array-by-array. | ||
+ | |||
+ | samples/coarrays/blockmatrix.f90 | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==Parallel Matrix Multiplication== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Allocation, deallocation of coarrays is like a sync all | ||
+ | * Forces synchronization across all images | ||
+ | * Any particular coarray must have same size, co-size across all images | ||
+ | |||
+ | samples/coarrays/blockmatrix.f90 | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==Parallel Matrix Multiplication== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * This is the entire parallel multiplication operation. | ||
+ | * a[myrow,k] gets the entire a block from image at myrow,k. | ||
+ | * matmul works with those blocks, updates c with new term. | ||
+ | |||
+ | samples/coarrays/blockmatrix.f90 | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==Coarray Summary== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Coarrays are powerful, simple-to-use addition to parallel programming models available | ||
+ | * Will be widely available soon | ||
+ | * Basic model implemented in Cray fortran for ~10 yrs; well tested and understood. | ||
+ | * Typically implemented with MPI under the hood; can give performance similar to MPI with fewer lines of code. | ||
+ | * Other languages have similar extensions (UPC based on C, Titanium based on Java) but are not part of languages’ standard. | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==Closing Hints== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * Always give the compiler info it needs to help you by being as explicit as possible | ||
+ | * implicit none, end [construct] [name], parameters for constants, intent in/out, use only, etc. | ||
+ | * Always get as much info from compiler as possible - always use -Wall (gfortran) or -warn all (ifort). | ||
+ | | | ||
+ | |} | ||
+ | |||
+ | ==Useful Resources== | ||
+ | {| | ||
+ | |- valign="top" | ||
+ | | | ||
+ | * http://fortranwiki.org/ | ||
+ | ** Reference source; has all standards; Fortran2003/2008 status of major compilers | ||
+ | * http://en.wikipedia.org/wiki/Fortran_language_features | ||
+ | ** Succinct summary of new features (spotty past F95) | ||
+ | * http://stackoverflow.com/questions/tagged/fortran | ||
** ( Programmers Questions & Answers | ** ( Programmers Questions & Answers | ||
+ | |} |
Latest revision as of 14:59, 23 August 2012
Modern Fortran for FORTRAN77 users
Jonathan Dursi
This is the Wiki-fied version of the slides used for the Modern Fortran course at SciNet. To follow along, with both the examples and the hands on, you will want to download the source code and make sure you have a working Fortran 2003 compiler.
Course Overview
|
A Brief History of Fortran
- Only major compiled programming language designed specifically for scientific programming.
- Powerful array operations; many mathematical functions (Bessel functions!) built in; designed to enable compiler optimizations for fast code
- Oldest (54-57 yrs) still-used programming language.
- Most people come to Fortran via being given old code by someone.
- Can’t understand the old code, or quirks of modern language, without understanding it’s history
- 1957 - J.W. Backus et al. In Proceedings Western Joint Computer Conference, Los Angeles, California, February 1957.
- IBM 704
- (Arguably) first modern compiled programming language.
- Idea of compilers at all was controversial at time.
FORTRAN (1957)
|
Incremental changes
|
FORTRAN66
|
FORTRAN77
- The most common to see “in the wild” of old code today
- if/else/endif, better do loops, control of implicit typing
- Character strings, saved variables, IO improvements
- Approved in 1978, beginning long tradition of “optimistic” naming of standards by year.
The interregnum
- Programming languages and techniques were moving quite quickly
- Several attempts were made to make new version, but standardization process very slow, failed repeatedly.
- Absent new real standard, implementations began to grow in many different directions
- Some extensions became quasi-standard, many were peculiar to individual compilers.
Fortran90
- Enormous changes; the basis of modern Fortran (not FORTRAN!)
- Free form, array slices, modules, dynamic memory allocation, derived types...
- Changes so major that took several years for compilers to catch up.
- Modern fortran
And since...
- Fortran95 - modest changes to Fortran90, killed off some deprecated F77 constructs.
- Fortran 2003 - bigger change; object-oriented, C interoperability. Most compilers have pretty good F2003 support.
- Fortran 2008 - mostly minor changes, with one big addition (coarray), other parallel stuff. Compiler-writers getting there.
F90, F95, F2003, F2008..
- We won’t distinguish between versions; we’ll just show you a lot of useful features of modern fortran.
- Will only show widely-implemented features from 2003 and 8, with exception of coarrays; these are being implemented and are very important.
New Format, New Syntax
Free Format: some highlights
|
<source lang="fortran"> program example implicit none integer, parameter :: npts = 10000 real, parameter :: startx=0., endx=1. real, parameter :: dx = (endx-startx)/npts real :: integral, xleft, xright, xmid integer :: i if (npts < 2) then print *,'Too few points!' else integral = 0. ! Simpson's Rule xleft = 0. int: do i=0,npts-1 xright = (i+1)*dx xmid = (xleft+xright)/2. integral = integral + (dx/6.)*(f(xleft) + 4.*f(xmid) + & f(xright)) xleft = xright end do int print *,'Numerical integral is ', integral print *,'Exact soln is ', (endx-startx)/2. - & (sin(2*endx)-sin(2*startx))/4. endif
function f(x) implicit none real :: f real, intent(in) :: x f = sin(x)**2 end function f end program example </source> |
Variable Declarations
<source lang="fortran"> integer i parameter (i=5) </source>
|
<source lang="fortran"> implicit none integer, parameter :: npts = 10000 real, parameter :: startx=0., endx=1. real, parameter :: dx = (endx-startx)/npts real :: integral, xleft, xright, xmid integer :: i </source> |
Variable Initialization
|
<source lang="fortran"> |
...
subroutine testvarinit implicit none integer :: i = 5 print '(A,I3)', 'On entry; i = ', i i = 7 print '(A,I3)', 'Now set; i = ', i end subroutine testvarinit |
...
use inittest |
...
call testvarinit call testvarinit |
...
end program initialization </source> ( From samples/variables/initialization/initialization.f90) <source lang="bash"> $ gfortran initialization.f90 -o initialization $ ./initialization On entry; i = 5 Now set; i = 7 On entry; i = 7 Now set; i = 7 </source> |
---|
Real Kinds
|
<source lang="fortran"> program realkinds use iso_fortran_env implicit none real :: x real(kind=real32) :: x32 real(kind=real64) :: x64 real(kind=real128):: x128 real(kind=selected_real_kind(6)) :: y6 real(kind=selected_real_kind(15)):: y15 print *,'Default:' print *, precision(x), range(x) print *,'Real32:' print *, precision(x32), range(x32) print *,'Real64:' print *, precision(x64), range(x64) print *,'Real128:' print *, precision(x128), range(x128) print *, print *,'Selected Real Kind 6:' print *, precision(y6), range(y6) print *,'Selected Real Kind 15:' print *, precision(y15), range(y15) end program realkinds </source> (from samples/variables/kinds/realkinds.f90) <source lang="bash"> $ ./realkinds Default: 6 37 Real32: 6 37 Real64: 15 307 Real128: 18 4931 Selected Real Kind 6: 6 37 Selected Real Kind 15: 15 307 </source> |
Integer Kinds
|
<source lang="fortran"> program integerkinds use iso_fortran_env implicit none integer :: i integer(kind=int8) :: i8 integer(kind=int16) :: i16 integer(kind=int32) :: i32 integer(kind=int64) :: i64 integer(kind=selected_int_kind(6)) :: j6 integer(kind=selected_int_kind(15)):: j15 print *,'Default:' print *, huge(i) print *,'Int8:' print *, huge(i8) print *,'Int16:' print *, huge(i16) print *,'Int32:' print *, huge(i32) print *,'Int64:' print *, huge(i64) print *, print *,'Selected Integer Kind 6:' print *, huge(j6) print *,'Selected Integer Kind 15:' print *, huge(j15) end program integerkinds </source> (from samples/variables/kinds/intkinds.f90) <source lang="bash"> $ ./intkinds Default: 2147483647 Int8: 127 Int16: 32767 Int32: 2147483647 Int64: 9223372036854775807 Selected Integer Kind 6: 2147483647 Selected Integer Kind 15: 9223372036854775807 </source> |
Strings
|
<source lang="fortran"> program strings implicit none character(len=20) :: hello character(len=20) :: world character(len=30) :: helloworld hello = "Hello" world = "World!" helloworld = trim(hello) // " " // trim(world) print *, helloworld if (hello < world) then print *, '<', hello, '> is smaller.' else print *, '<', world, '> is larger.' endif end program strings </source> (from samples/variables/strings/strings.f90) <source lang="bash"> $ ./strings Hello World! <Hello > is smaller. </source> |
Array declarations
|
<source lang="fortran"> program arrays implicit none real, dimension(3) :: x, y x = [1,2,3] y = 2*x print *, x print *, y end program arrays </source> ( from samples/variables/arrays/arrays.f90) <source lang="bash"> $ ./arrays 1.0000000 2.0000000 3.0000000 2.0000000 4.0000000 6.0000000 </source> |
Do loops
|
<source lang="fortran"> program doi implicit none integer :: i do i=1,10 print *, i, i**2, i**3 enddo end program doi </source> ( from samples/variables/doloops/doi.f90 ) <source lang="bash"> $ ./doi 1 1 1 2 4 8 3 9 27 4 16 64 5 25 125 6 36 216 7 49 343 8 64 512 9 81 729 10 100 1000 </source> |
Named loops
|
<source lang="fortran"> program nameddo implicit none integer :: i, j outer: do i=1,3 inner: do j=1,3 print *, i, j, i*i+j*j enddo inner end do outer end program nameddo </source> ( from samples/variables/doloops/nameddo.f90 ) <source lang="bash"> $ ./nameddo 1 1 2 1 2 5 1 3 10 2 1 5 2 2 8 2 3 13 3 1 10 3 2 13 3 3 18 </source> |
Cycle/exit
|
<source lang="fortran"> program cycleexit implicit none integer :: i do print *, 'Enter a number between 1-13' read *, i if (i>=1 .and. i<=13) exit print *, 'Wrong; try again.' enddo print *, 'Good; you entered ', i end program cycleexit </source> ( from samples/variables/doloops/cycleexit.f90 ) <source lang="bash"> $ more cycleexit-out.txt $ ./cycleexit Enter a number between 1-13 23 Wrong; try again. Enter a number between 1-13 -1 Wrong; try again. Enter a number between 1-13 12 Good; you entered 12 </source> |
Do while
|
<source lang="fortran"> program dowhile implicit none integer :: i i = -1 do while (i < 1 .or. i > 13) print *, 'Enter a number between 1-13' read *, i if (i<1 .or. i>13) print *, 'Wrong; try again.' enddo print *, 'Good; you entered ', i end program dowhile </source> ( from samples/variables/doloops/dowhile.f90 ) |
Hands On 1
- In workedexample/f77 is a simplified, F77ized version of a fluid-dynamics code from Ue-Li Pen, CITA, U of Toronto (http://www.cita.utoronto.ca/~pen/MHD/)
- For the purposes of this class, we've turned it from a perfectly good f90 code to something that looks more like something your supervisor would dust off and give to you.
- Today we’ll be translating this version into a very modern Fortran
- Compile (using make) and run (./hydro)
- Outputs a .pbm file; use “display dens.pbm” to see the result of dense blob of fluid moving through a light medium.
- In workedexamples/freeform, have partly converted the program to new freeform format, with enddos, ending procedures, implicit nones, and new variable declaration syntax.
- Finish doing so - just need to do program hydro, subroutine color, subroutine outputpbm, function cfl. Fix indenting (Don't need to start at col 7 anymore).
- ~1 hr (for getting logged in and everything working)
Procedures and Modules
|
Modules
|
<source lang="fortran"> module gravity implicit none real, parameter :: G = 6.67e-11 ! MKS units contains real function gravforce(x1,x2,m1,m2) implicit none real, dimension(3), intent(in) :: x1,x2 real, intent(in) :: m1, m2 real :: dist dist = sqrt(sum((x1-x2)**2)) gravforce = G * m1 * m2 / dist**2 end function gravforce end module gravity program simplemod use gravity implicit none print *, 'Gravitational constant = ', G print *, 'Force between 2 1kg masses at [1,0,0] & &and [0,0,1] is' print *, gravforce([1.,0.,0.],[0.,0.,1.],1.,1.) end program simplemod </source> (from samples/procedures/simplemod/simplemod.f90) |
Compiling & Running
|
<source lang="bash"> $ ls simplemod.f90 $ gfortran -o simplemod simplemod.f90 -Wall $ ls gravity.mod simplemod simplemod.f90 $ ./simplemod Gravitational constant = 6.6700000E-11 Force between 2 1kg masses at [1,0,0] and [0,0,1] is 3.3350003E-11 </source> |
Modules
|
<source lang="fortran"> <source lang="fortran"> module gravity implicit none real, parameter :: G = 6.67e-11 ! MKS units contains real function gravforce(x1,x2,m1,m2) implicit none real, dimension(3), intent(in) :: x1,x2 real, intent(in) :: m1, m2 real :: dist dist = sqrt(sum((x1-x2)**2)) gravforce = G * m1 * m2 / dist**2 end function gravforce end module gravity program simplemod use gravity implicit none print *, 'Gravitational constant = ', G print *, 'Force between 2 1kg masses at [1,0,0] & &and [0,0,1] is' print *, gravforce([1.,0.,0.],[0.,0.,1.],1.,1.) end program simplemod </source> (from samples/procedures/simplemod/simplemod.f90) |
use module, only :
|
<source lang="fortran"> module gravity implicit none real, parameter :: G = 6.67e-11 ! MKS units contains real function gravforce(x1,x2,m1,m2) implicit none real, dimension(3), intent(in) :: x1,x2 real, intent(in) :: m1, m2 real :: dist dist = sqrt(sum((x1-x2)**2)) gravforce = G * m1 * m2 / dist**2 end function gravforce end module gravity program simplemod2 use gravity, only : G, gravforce implicit none print *, 'Gravitational constant = ', G print *, 'Force between 2 1kg masses at [1,0,0] & &and [0,0,1] is' print *, gravforce([1.,0.,0.],[0.,0.,1.],1.,1.) end program simplemod2 </source> samples/procedures/simplemod/simplemod2.f90 |
Modules usually get their own files
|
<source lang="fortran"> module gravity implicit none private character (len=8), parameter, public :: massunit="kilogram" character (len=8), parameter, public :: forceunit="Newton" public :: gravforce real, parameter :: G = 6.67e-11 ! MKS units contains real function distance(x1,x2) implicit none real, dimension(3), intent(in) :: x1, x2 distance = sqrt(sum((x1-x2)**2)) end function distance real function gravforce(x1,x2,m1,m2) implicit none real, dimension(3), intent(in) :: x1,x2 real, intent(in) :: m1, m2 real :: dist dist = distance(x1,x2) gravforce = G * m1 * m2 / dist**2 end function gravforce end module gravity </source> (from samples/procedures/multifilemod/gravity.f90 ) |
Modules usually get their own files
|
<source lang="make"> FC=gfortran FFLAGS=-O3 -Wall multifilemod: multifilemod.o gravity.o $(FC) -o $@ multifilemod.o gravity.o %.mod: %.f90 $(FC) $(FFLAGS) -c $< multifilemod.o: multifilemod.f90 gravity.mod $(FC) $(FFLAGS) -c $< clean: rm -f *.o *~ *.mod multifilemod </source> (from samples/procedures/multifilemod/Makefile) |
.mod needed for compilation
|
<source lang="fortran"> program simplemod2 use gravity, only : gravforce, massunit, forceunit implicit none print *, 'Force between 2 1 ', massunit ,' masses ', & ' at [1,0,0] and [0,0,1] is' print *, gravforce([1.,0.,0.],[0.,0.,1.],1.,1.), forceunit end program simplemod2 </source> (from samples/procedures/multifilemod/multifilemod.f90) |
.o needed for linking
|
<source lang="make"> FC=gfortran FFLAGS=-O3 -Wall multifilemod: multifilemod.o gravity.o $(FC) -o $@ multifilemod.o gravity.o %.mod: %.f90 $(FC) $(FFLAGS) -c $< multifilemod.o: multifilemod.f90 gravity.mod $(FC) $(FFLAGS) -c $< clean: rm -f *.o *~ *.mod multifilemod </source> (from samples/procedures/multifilemod/Makefile) |
Compiling and running
|
<source lang="bash"> $ make gfortran -O3 -Wall -c gravity.f90 gfortran -O3 -Wall -c multifilemod.f90 gfortran -o multifilemod multifilemod.o gravity.o reposado-$ ./multifilemod Force between 2 1 kilogram masses at [1,0,0] and [0,0,1] is 3.33500033E-11 Newton </source> |
Private and public
|
<source lang="fortran"> module gravity implicit none private character (len=8), parameter, public :: massunit="kilogram" character (len=8), parameter, public :: forceunit="Newton" public :: gravforce real, parameter :: G = 6.67e-11 ! MKS units contains real function distance(x1,x2) implicit none real, dimension(3), intent(in) :: x1, x2 distance = sqrt(sum((x1-x2)**2)) end function distance real function gravforce(x1,x2,m1,m2) implicit none real, dimension(3), intent(in) :: x1,x2 real, intent(in) :: m1, m2 real :: dist dist = distance(x1,x2) gravforce = G * m1 * m2 / dist**2 end function gravforce end module gravity </source> ( from samples/procedures/multifilemod/gravity.f90 ) |
Procedures
|
<source lang="fortran"> module procedures contains function square(x) result(xsquared) implicit none real :: xsquared real, intent(IN) :: x xsquared = x*x end function square function cube(x) implicit none real :: cube real, intent(IN) :: x cube = x*x*x end function cube subroutine squareAndCube(x, squarex, cubex) implicit none real, intent(in) :: x real, intent(out) :: squarex real, intent(out) :: cubex squarex = square(x) cubex = cube(x) end subroutine squareAndCube end module procedures </source> ( from samples/procedures/funcsub/procedures.f90 ) |
Functions
|
<source lang="fortran"> function square(x) result(xsquared) implicit none real :: xsquared real, intent(IN) :: x xsquared = x*x end function square function cube(x) implicit none real :: cube real, intent(IN) :: x cube = x*x*x end function cube subroutine squareAndCube(x, squarex, cubex) implicit none real, intent(in) :: x real, intent(out) :: squarex real, intent(out) :: cubex squarex = square(x) cubex = cube(x) end subroutine squareAndCube </source> ( from samples/procedures/funcsub/procedures.f90 ) |
Procedure interfaces
|
<source lang="fortran"> function square(x) result(xsquared) implicit none real :: xsquared real, intent(IN) :: x xsquared = x*x end function square function cube(x) implicit none real :: cube real, intent(IN) :: x cube = x*x*x end function cube subroutine squareAndCube(x, squarex, cubex) implicit none real, intent(in) :: x real, intent(out) :: squarex real, intent(out) :: cubex squarex = square(x) cubex = cube(x) end subroutine squareAndCube </source> ( from samples/procedures/funcsub/procedures.f90 ) |
Procedure interfaces
|
<source lang="fortran"> function integratefx(xlo, xhi, f, n) |
integrate with trapezoid rule | ....
integer :: i real :: dx, xleft, xright integratefx = 0. dx = (xhi-xlo)/n xleft = xlo do i=0, n-1 xright = xleft + dx integratefx = integratefx + dx*(f(xright)+f(xleft))/2. xleft = xright enddo end function integratefx </source> (from samples/procedures/interface/integrate.f90 ) (from http://en.wikipedia.org/wiki/File:Trapezoidal_rule_illustration_small.svg ) |
---|
Procedure interfaces
|
<source lang="fortran"> function integratefx(xlo, xhi, f, n) |
integrate with trapezoid rule
implicit none real, intent(in) :: xlo, xhi interface function f(x) implicit none real :: f real, intent(in) :: x end function f end interface integer, intent(in) :: n real :: integratefx integer :: i real :: dx, xleft, xright integratefx = 0. dx = (xhi-xlo)/n xleft = xlo do i=0, n-1 xright = xleft + dx integratefx = integratefx + dx*(f(xright)+f(xleft))/2. xleft = xright enddo end function integratefx </source> (from samples/procedures/interface/integrate.f90 ) |
---|
Recursive procedures
|
<source lang="fortran"> recursive function integratefx(xlo, xhi, f, tol) result(integral) |
integrate with trapezoid rule, simpsons rule; | if difference between two is larger than | relevant tolerance, subdivide region. | ...variable declarations as before...
dx = xhi-xlo xmid = (xlo+xhi)/2. trapezoid = dx*(f(xlo)+f(xhi))/2. simpsons = dx/6.*(f(xlo)+4.*f(xmid)+f(xhi)) error = abs(trapezoid-simpsons)/(0.5*(trapezoid+simpsons)) if (error > tol) then |
too coarse; subdivide
integral = integratefx(xlo,xmid,f,tol) + & integratefx(xmid,xhi,f,tol) else integral = trapezoid endif end function integratefx </source> ( from samples/procedures/recursive/integrate.f90) |
---|
Pure procedures
|
<source lang="fortran"> pure subroutine axpy(a, x, y) |
y = y + a*x
implicit none real, intent(IN) :: a, x real, intent(INOUT) :: y y = y + a*x end subroutine axpy subroutine printaxpy(a, x, y) |
y = y + a*x
implicit none real, intent(IN) :: a, x real, intent(INOUT) :: y print *, a, '*', x, ' + ', y, & ' = ', a*x+y y = a*x + y end subroutine printaxpy </source> (from samples/procedures/purity/purity.f90) |
---|
Optional Arguments
|
<source lang="fortran"> recursive function integratefx(xlo, xhi, f, tol) result(integral) |
....
real, intent(in), optional :: tol |
....
real :: errtol |
use parameter if passed, | else use default
if (present(tol)) then errtol = tol else errtol = 1.e-6 endif |
....
if (error > errtol) then |
too coarse; subdivide
integral = integratefx(xlo,xmid,f,errtol) + & integratefx(xmid,xhi,f,errtol) else integral = trapezoid endif end function integratefx </source> (from samples/procedures/optional/integrate.f90) |
---|
Optional Arguments
|
<source lang="fortran"> print *, 'Integrating using default tol' approx = integratefx(0., 2*pi, sinesquared) print *, 'Approximate integral = ', approx print *, 'Exact integral = ', exact print *, print *, 'Integrating using coarser tol' approx = integratefx(0., 2*pi, sinesquared, 0.01) print *, 'Approximate integral = ', approx |
....
</source> (from samples/procedures/optional/optional.f90) |
---|
Keyword Arguments
|
<source lang="fortran"> |
....
print *, 'Integrating using still coarser tol' approx = integratefx(xhi=2*pi, xlo=0., tol=0.5, & f=sinesquared) print *, 'Approximate integral = ', approx </source> (from samples/procedures/optional/optional.f90) |
---|
Procedures and Modules Summary
|
Hands On 2
|
Fortran arrays
|
<source lang="fortran"> program basicarrays implicit none integer, dimension(5) :: a, b, c integer :: i a = [1,2,3,4,5] b = [(2*i+1, i=1,5)]
print *, 'a = ', a print *, 'b = ', b c = a+b print *, 'c = ', c c = a*b + 1 print *, 'a*b+1=', c end program basicarrays </source> (from samples/arrays/basic.f90 ) |
Array constructors
|
<source lang="fortran"> x = [1,2,3,4,5] x = (/1,2,3,4,5/) x = [ (i,i=1,5)] a = [ ((i*j,j=1,3),i=1,5)] </source> |
Elementwise operations
|
<source lang="fortran"> program elementwise implicit none real, dimension(10) :: x,y,z integer :: i real, parameter:: pi = 4.*atan(1.) x = [(2*pi*(i-1)/9.,i=1,10)] y = sin(x) z = x*x print *, x print *, y print *, z end program elementwise </source> (from samples/arrays/elementwise.f90 ) |
Elemental Functions
|
<source lang="fortran"> program elementalfn implicit none real, dimension(10) :: x,y,z integer :: i real, parameter:: pi = 4.*atan(1.) x = [(2*pi*(i-1)/9.,i=1,10)] y = sinesquared(x) z = sin(x)*sin(x) print *, x print *, y print *, z print *,z(::3) contains elemental function sinesquared(x) implicit none real :: sinesquared real, intent(in) :: x sinesquared = sin(x)**2 end function sinesquared end program elementalfn </source> (from samples/arrays/elemental.f90 ) |
Array comparisons
|
<source lang="fortran"> program comparearrays implicit none integer, dimension(5) :: a, b integer :: i a = [1,2,3,4,5] b = [(2*i-3, i=1,5)] print *, 'A = ', a print *, 'B = ', b if (any(a > b)) then print *, 'An A is larger than a B' endif if (all(a > b)) then print *, 'All As ares larger than Bs' else if (all(b > a)) then print *, 'All Bs are larger than As' else print *, 'A, B values overlap' endif end program comparearrays </source> (from samples/arrays/compare.f90) |
Array masks
|
<source lang="fortran"> program mask implicit none integer, dimension(10) :: a logical, dimension(10) :: pos integer :: i a = [(2*i-7, i=1,10)] pos = (a > 0) print '(A,10(I4,1X))','A = ', a print *,'# of positive values: ', count(pos) print *,'Sum of positive values: ', sum(a,pos) print *,'Minimum positive value: ', minval(a,pos) end program mask </source> ( from samples/arrays/mask.f90 ) <source lang="bash"> $ ./mask A = -5 -3 -1 1 3 5 7 9 11 13 # of positive values: 7 Sum of positive values: 49 Minimum positive value: 1 </source> |
Where construct
|
<source lang="fortran"> program wherearray implicit none real, dimension(6) :: a, diva integer :: i a = [(2*i-6, i=1,6)] where (a /= 0) diva = 1/a elsewhere diva = -999 endwhere print *,'a = ' print '(6(F8.3,1X))',a print *,'1/a = ' print '(6(F8.3,1X))',diva end program wherearray </source> (from samples/arrays/where.f90) <source lang="bash"> $ ./where a = -4.000 -2.000 0.000 2.000 4.000 6.000 1/a = -0.250 -0.500 -999.000 0.500 0.250 0.167 </source> |
Forall construct
|
<source lang="fortran"> program forallarray implicit none integer, dimension(6,6) :: a integer :: i,j a = -999 forall (i=1:6, j=1:6, i/=j) a(i,j) = i-j endforall do i=1,6 print '(6(I5,1X))',(a(i,j),j=1,6) enddo end program forallarray </source> (from samples/arrays/forall.f90) <source lang="bash"> $ ./forall -999 -1 -2 -3 -4 -5 1 -999 -1 -2 -3 -4 2 1 -999 -1 -2 -3 3 2 1 -999 -1 -2 4 3 2 1 -999 -1 5 4 3 2 1 -999 </source> |
Array Sections
|
<source lang="fortran"> a([start]:[end][:step]) a = [1,2,3,4,5,6,7,8,9,10] a(7:) == [7,8,9,10] a(:3) == [1,2,3] a(2:4) == [2,3,4] a(::3) == [1,4,7,10] a(2:4:2) == [2,4] a(2) == 2 a(:) == [1,2,3,4,5,6,7,8,9,10] </source> |
Array Sections
|
<source lang="fortran"> program derivative implicit none real, dimension(10) :: x real, dimension(9) :: derivx integer :: i real, parameter:: pi = 4.*atan(1.), h=1. x = [(2*pi*(i-1)/9.,i=1,10)] derivx = ((x(2:10)-x(1:9))/h) print *, derivx do i=1,9 derivx(i) = (x(i+1)-x(i))/h enddo print *, derivx end program derivative </source> (from samples/arrays/derivative.f90) |
Array Sections
|
<source lang="fortran">
cshift(a,1) == [2,3,4,5,1] cshift(a,-1) == [5,1,2,3,4] eoshift(a,1) ==[2,3,4,5,0] eoshift(a,-1)==[0,1,2,3,4] </source> |
Other important array intrinsics
<source lang="fortran">
1,4 reshape([1,2,3,4,5,6],[3,2]) == 2,5 3,6 </source> |
Linear algebra in Fortran
|
Matrix Multiplication Times
<source lang="fortran"> |
...
print *, 'Experiment with matrix size ', n print *, 'Times in seconds.' allocate(a(n,n)) allocate(b(n,n)) allocate(c(n,n)) call random_number(a) call random_number(b) call tick(starttime) do j=1,n do i=1,n c(i,j) = 0. do k=1,n c(i,j) = c(i,j)+a(i,k)*b(k,j) enddo enddo enddo looptime = tock(starttime) call tick(starttime) c = matmul(a,b) matmultime = tock(starttime) call tick(starttime) call sgemm('N','N',n,n,n,1.,a,n,b,n,0.,c,n) sgemmtime = tock(starttime)
print *, 'Triple-loop time: ', looptime print *, 'matmul intrinsic time: ', matmultime print *, 'SGEMM lapack call time:', sgemmtime |
...
</source> |
<source lang="bash"> $ ./matmul 2500 Experiment with matrix size 2500 Triple-loop time: 149.63400 matmul intrinsic time: 10.370000 SGEMM lapack call time: 1.4809999 </source> (gfortran 4.6, compiled -O3 -march=native using Intel MKL 10.3 for sgemm) (program from samples/arrays/matmul.f90) |
---|
Linear algebra in Fortran
<source lang="fortran"> program matvec implicit none integer, dimension(4,5) :: a integer, dimension(5,4) :: at integer, dimension(4,4) :: aat integer :: i a = reshape([(i,i=1,4*5)],[4,5]) at = transpose(a) print *,'A = ' call printmat(a) print *,'A^T = ' call printmat(at) aat = matmul(a,at) print *,'A . A^T = ' call printmat(aat) |
...
</source> |
<source lang="bash"> $ ./matrix A = 1 5 9 13 17 2 6 10 14 18 3 7 11 15 19 4 8 12 16 20 A^T = 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 A . A^T = 565 610 655 700 610 660 710 760 655 710 765 820 700 760 820 880 A . A^T subarray = 610 660 710 655 710 765 700 760 820 </source> |
---|
Array sizes and Assumed Shape
integer nx, ny integer a(nx,ny) or worse, integer a(*,ny)
|
<source lang="fortran"> subroutine printmat(a) implicit none integer, dimension(:,:) :: a integer :: nr, nc, i, j nr = size(a,1) nc = size(a,2) do i=1,nr print '(99(I4,1X))', (a(i,j), j=1,nc) enddo end subroutine printmat </source> (from samples/arrays/matrix.f90) |
Allocatable Arrays
|
Allocate(), Deallocate()
|
<source lang="fortran"> program allocarray implicit none integer :: i, n integer, dimension(:), allocatable :: a n = 10 allocate(a(n)) a = [(i, i=2,20,2)] print *,'A = ' print *,a deallocate(a) end program allocarray </source> (from samples/arrays/allocatable.f90 ) |
get_command_argument()
|
<source lang="fortran"> program allocarray2 implicit none integer :: i, n integer, dimension(:), allocatable :: a character(len=30) :: arg if (command_argument_count() < 1) then print *,'Use: allocatable N, '//& ' where N is array size.' stop endif call get_command_argument(1, arg) read( arg,'(I30)'), n print *,'Allocating array of size ', n allocate(a(n)) a = [(i,i=1,n)] print *, a deallocate(a) end program allocarray2 </source> (from samples/arrays/allocatable2.f90) <source lang="bash"> $ ./allocatable2 Use: allocatable N, where N is array size. $ ./allocatable2 3 Allocating array of size 3 1 2 3 $ ./allocatable2 5 Allocating array of size 5 1 2 3 4 5 </source> |
Hands on #3
|
Fortran Pointers
|
<source lang="fortran"> real, target :: x = 3.2 real, pointer:: p p => x p x 3.2 </source> samples/pointers/ptr1.f90 |
Fortran Pointers
|
<source lang="fortran"> real, target :: x = 3.2 real, pointer:: p p => null() p x </source> |
Fortran Pointers
|
<source lang="fortran"> real, target :: x = 3.2 real, pointer:: p1, p2 p1 => x p2 => p1 p1 p2 x 3.2 </source> |
Allocating a pointer
|
<source lang="fortran"> real, pointer:: p allocate(p) p = 7.9 p 7.9 |
</source> samples/pointers/ptr2.f90 |
What are they good for? (1)
|
http://en.wikipedia.org/wiki/File:Singly-linked-list.svg http://en.wikipedia.org/wiki/File:Max-Heap.svg |
What are they good for? (2)
|
<source lang="fortran"> real, target, dimension(7) :: x real, pointer:: p(:) p => x(2:6) p x 1 2 3 4 5 6 7 </source> samples/pointers/views.f90 |
Hands on #4
|
Derived Types and Objects
<source lang="fortran"> type griddomain real :: xmin, xmax real :: ymin, ymax real :: nx, ny real, dimension(:,:) :: u endtype griddomain type(griddomain) :: g g % xmin = -1 g % xmax = +1 </source> |
Derived Types and Objects
|
(from samples/derivedtypes/simple/intervalmath.f90 ) |
Procedures using types
|
(from samples/derivedtypes/intervalfunctions/intervalmath.f90 ) |
Procedures using types
|
samples/derivedtypes/genericintervals/interval2.f90 |
Generic Interfaces
Generic Interfaces
Generic interfaces
Operator overloading
Operator overloading
Generic interfaces
Type bound procedures
Type bound procedures
Type bound procedures
Object oriented programming
Interoperability with other languages
C-interoperability
Calling a C routine from Fortran
Calling a C routine from Fortran
Calling a C routine from Fortran
C strings
Calling Fortran from C
Calling Fortran from C
More advanced
Fortran calling C, which calls Fortran
samples/C-interop/valueref/croutine.csamples/C-interop/valueref/froutine.f90$ ./main
samples/C-interop/valueref/MakefileF2py
* Will only use solver module.
* generates the following header file (hydro_solver.pyf)
$ ipython
Coarrays
Coarrays
Coarrays
Sychronization
Sychronization
1d Diffusion Eqn
1d Diffusion Eqn
1d Diffusion Eqn
1d Diffusion Eqn
$ ./diffusion
Coarray Diffusion
Coarray Diffusion
Coarray Diffusion
Coarray Diffusion
Coarray Diffusion
$ export FOR_COARRAY_NUM_IMAGES=3
gnuplot> plot '1-output.txt' using 1:2, '2-output.txt' using 1:2, '3output.txt' using 1:2, '1-output.txt' using 1:3 with lines title 'Theory',
Parallel Matrix Multiplication
Parallel Matrix Multiplication
Parallel Matrix Multiplication
Parallel Matrix Multiplication
Parallel Matrix Multiplication
Coarray Summary
Closing Hints
Useful Resources
|