I am new to Parallel computing and I've been trying to parallelize a sequential Fortran program which I have worked on before, using OpenMP, My main code is similar to this example :
program main
use omp_lib
implicit none
integer :: i
double precision :: y , test , a(5)
common/y/y
external :: test
a=[1,2,3,4,5]
do i=1,5
y=a(i)
print*,test(2.d0)
end do
end program main
double precision function test(x)
implicit none
double precision :: y , x
common/y/y
test=y*x
return
end function
I am trying to use !$OMP PARALLEL DO , but I have problem with the variable "y" which is declared global and used in the function "test" but it is changed within each thread. How can I parallelize such a code without declaring "y" as the argument of the function "test"?
Also, is there any way that each thread can carry the value of iteration counter "i" privately, throughout the whole program (including through the functions)?
OMP uses a DEFAULT() type for all variables that exist prior to the start of the OMP construct. The default value of DEFAULT() is shared. That means your common block variables are shared variables by default. Functions inside OMP constructs however use their local scope to create private versions of function local variables. The y variable that you're passing to the function through a COMMON block isn't function local though. That means it retains the shared type, and is common across all threads.
If this was just another variable, then you could resolve this by declaring y as a PRIVATE() variable in the OMP construct. Instead, because this is inside a COMMON block you need to declare y as THREADPRIVATE in every location that references the common block. If you miss it in even one place it'll default to being a shared variable across all of your threads. Assuming it even compiles.
Related
I'm trying to teach myself Fortran, and have been messing around with linking multiple files together. Following examples, I have been writing programs in one file, functions in another, and using interface blocks in my main program to refer to the external function.
I was testing how much information was needed in the interface block and realised that I can remove it entirely.
My program:
program test
implicit none
real :: x, y, func
x = 3
y = func(x)
print *, y
end program test
And the function file:
function func(x)
implicit none
real :: x, func
func = x**3
end function func
I then compile it using gfortran -o test test.f90 func.f90 and the code works as expected. My question is, why do I not need to include an interface block in my program file? Is it simply a matter of good practice, or does defining func as a real variable serve as shorthand? I am on Windows, having installed gfortran through minGW.
As an aside/related question, if I instead use a subroutine:
subroutine func(x,y)
implicit none
real :: x,y
y = x**3
end subroutine func
And change the line y = func(x) to call func(x,y) then the code will work fine without any interface block or declaration. Why is this?
The declaration real :: func in the main program here declares func to be a function with (default) real result. This function has an interface in the main program as a result, so it is legitimate to reference that function with y = func(x).
The interface in this case is implicit. In this way, the main program knows exactly three things about func:
it is a function (with that name);
it has the external attribute;
it has real result.
The reference to the function is compatible with that knowledge. Further, how you reference the function matches precisely the properties of the function itself.
Equally, in the case of the subroutine, a call func(x,y) tells the main program exactly three things, again with the implicit interface:
it is a subroutine (with that name);
it has the external attribute;
it takes two real arguments.
Those three things again match the subroutine's definition, so things are fine.
Loosely, then, you don't need an interface block in these cases because the implicit interfaces are good enough.
There are times when an explicit interface is required and in most (nearly all) cases an explicit interface is better. As you can see in other questions and answers, there are usually better ways to provide an explicit interface than using an interface block.
Why do I not need to include an interface block in my program file?
Since Fortran 90, the recommended way to define reusable functions and subroutines is to use modules.
module func_m
contains
function func(x)
implicit none
real :: x, func
func = x**3
end function func
end module func_m
Then write use func_m in the main program, before implicit none: gfortran -c func_m.f90, gfortran -c test.f90 and gfortran -o test test.o func_m.o.
When you use modules, the compiler will check the type of the arguments of the functions. You also do not need to declare real :: func as the declarations are taken from the module.
When you compile as a "simple" object file, the compiler will simply call whatever function is named func without verification, as long as such a function is given in an object file.
The interface block is a kind of "in between". In your case, you could add one in the program file. This would force you to follow that declaration in the program. But it will not prevent linking to wrong functions at link time: there is no guarantee that the interface is right. It is mostly useful if you need to call C or FORTRAN 77 code from someone else, for which you couldn't use modules.
And change the line y = func(x) to call func(x,y) then the code will
work fine without any interface block or declaration. Why is this?
The interface issue is orthogonal to the function vs subroutine issue.
There are some cases where interface blocks are needed. For example if the called subroutine uses a pointer, allocatable or assumed shape arrays (this list is not complete, see the standard for more):
integer, pointer :: x(:,:) ! pointer attribute
integer, allocatable :: x(:,:) ! pointer attribute
integer :: a(:,:) ! assumed shape array
The pointer/allocatable have the advantage that the called subroutine can allocate it for the calling function. Assumed shape arrays automatically transfer the dimensions to the called subroutine.
If you use any of this your program will crash without explicit interface. Easiest solution is like said in the other answer: use modules to have the interface automtically correct. We use a perl script automatically extracting interfaces to have interface check without rewriting the code to modules (and avoid long compile times until all compilers reliably support Fortran 2008 submodules...)
I have a problem understanding why a variable (i) declared in a subroutine is seen in a contained subroutine, but that this is not true for a function (fie) which results in a compilation error. I searched for an answer and also tried to see if I could find something in the Fortran 95 standard but in vain.
I wrote a small example program:
program pgm
call a
end
subroutine a
implicit none
integer :: i
double precision :: fie
i = 7
call b
!write(*,*) fie(9)
contains
subroutine b
double precision :: x
!double precision :: fie
x = i
x = x + fie(i)
write(*,*) x
end subroutine
end subroutine
double precision function fie(ii)
implicit none
integer, intent(in) :: ii
fie = ii
end function
When compiling this with gfortran under cygwin (gfortran 5.4.0) I get the following error message:
$ gfortran aa.f90
aa.f90:20:15:
x = x + fie(i)
1
Error: ‘fie’ at (1) is not a function
When enabling either of the commented lines the program compiles and runs correctly.
I saw a similar error message when using the Intel compiler (Intel Fortran 12.1.7.367, indeed quite old).
It looks like fie has to be made available either in the contained routine or has to be used in the encompassing subroutine, but as said I could not find an answer on the net or in the Fortran 95 standard (or maybe I didn't look for the right words).
Any explanation?
The simplest fix is to use
double precision, external :: fie
the external attribute (also can be specified by the external statement) says: this is a procedure, I am not declaring a local variable.
For the declaration without the external to be interpreted as a function declaration the function reference must be present within the function body. Internal functions don't count. And therefore the compiler created a local double precision variable called fie.
Thank's to IanH for the relevant standard rule (from Fortran 2008 (16.5.1.4p5), but Fortran 95 will have an equivalent):
If an external or dummy procedure with an implicit interface is
accessed via host association, then it shall have the EXTERNAL
attribute in the host scoping unit; if it is invoked as a function in
the inner scoping unit, its type and type parameters shall be
established in the host scoping unit. The type and type parameters of
a function with the EXTERNAL attribute are established in a scoping
unit if that scoping unit explicitly declares them, invokes the
function, accesses the function from a module, or accesses the
function from its host where its type and type parameters are
established.
Of course explicit interfaces (best using modules) are much better than external functions.
Currently in my program I have several common blocks spread across several subprograms and functions. I sometimes forget to change all instances of a common block as I add variables to it. I want to make these common blocks into modules so I can add and remove variables to the module in one place without worrying about updating all instances of the module across my subprograms.
Do I need to include 'use' statements in the program that initializes the variables in the module or do I include the program in the module? I normally would use common blocks for this but am trying to implement modules because I think they will help my code remain readable as complexity increases.
NOTE: Some values of the variables in the modules need to be able to change as they are passed from one program to another.
I tried to write a simplified test program to become acquainted with modules but could not get it to work. I am familiar with fortran 77 but have never used modules before. I appreciate any help or advice.
I am using gfortran 4.6.1
Main.f
program main
use Words
use Vals
double precision x,y,z
character*5 Greet
integer i
Greet = 'Hello'
x = 4.1
y = 5.2
z = 10.0
i = 3
call foo ()
end program main
subroutine foo ()
use Words
use Vals
print *, Greet
z = x + y
print *, z
print *, i
end subroutine
module Words
character*5 Greet
save
end module
module Vals
double precision x,y
integer int
save
end module
You only need one instance of a module. You make it known to any main program or procedure (subroutine or function) that uses it with a use statement. If you have a subroutine that sets values then, like any other, it has to have a use statement. If are are setting initial values you can do so in the declaration. If the module is used by the main program then it will always be in scope and the values of the variables will persist throughout the run of the program. If the module is only used by procedure, in principle the module will go out of scope when none of those procedures are in the call chain and the compiler is allowed to forget the values of the module variables. (It is doubtful that any Fortran compiler actually does this.) This can be prevented by declaring each variable with SAVE. SAVE is implicit if you declare the variable with an initial value.
Normally you have to compile the modules first before they are used so that the compiler "knows" about them when it encounters the use statement. This is done either by putting them first in the file or compiling their files first. Here is your example reordered:
module Words
character*5 Greet
save
end module
module Vals
double precision x,y
integer i
save
end module
module my_subs
contains
subroutine foo ()
use Words
use Vals
double precision :: z
print *, Greet
z = x + y
print *, z
print *, i
end subroutine
end module my_subs
program main
use Words
use Vals
use my_subs
Greet = 'Hello'
x = 4.1
y = 5.2
i = 3
call foo ()
end program main
There are a couple of reasons why your code will not compile:
You have your modules positioned after your main program, which means they will not have been compiled by the time you use them there. Either put these in separate files and compile them before the main program, or put them before the main program.
You re-declare variables from your module in the main program, which the compiler will interpret as a name conflict. All module variables with the public attribute (which is the default) will become available in the scope where you use the module; this is called "use association". In other words, use vals is enough to make x, y and int available.
Additionally, modules are more like singleton objects rather than just data containers. They can also contain procedures, listed after a contains statement, which aids grouping variables and related procedures together. An example would be grouping your two modules into one, along with subroutine foo:
module vals
implicit none
double precision :: x = 4.1, y = 5.2
integer :: i = 3
character(5) :: greet = 'Hello'
contains
subroutine foo()
double precision :: z ! New local variable
print *, Greet
z = x + y
print *, z
print *, i
end subroutine
end module
Above, I have used the "new" :: double colon operator, which allows declaring and initialising multiple variables at once. Since module variables are already save'd implicitly, this is just fine.
Alternatively, if these modules are intended to be separate, you can also have contains sections in your main program (or any subprogram) and place the subroutine there. The advantage is that procedures declared in such a way always have explicit interfaces, which greatly benefits error diagnosis by the compiler and is even required in certain newer cases. This is one of the main improvements of F90, since F77 only dealt with external subprograms and implicit interfaces.
The simplest way to do common blocks is to have one include file per common block and make all your declarations within the include file. That way, the common block is only ever declared in one place. The problem of having to convert the code to use modules then just magically disappears.
Also, if you are starting new code, prefixing the name of the common block variables with the same name as the named common block will make coding a lot easier. It is a pain initially and the prima donnas who have been coding for years will refuse to conform. The people who maintain the code will just find it so easy: no greps or groks. Just by looking at the name, you know which common block it comes from.
Keeping the same convention with modules also helps. If all routine names and variable names begin with the module name, then just by looking at the name, you know which module it comes from.
I'm working on this F77 code with extensive use of COMMON BLOCKS, and those COMMON BLOCKS are also used in a C program (after name mangling). If foo(a, b, ...) is the main routine of the fortran code, is there anyway to make the COMMON BLOCKS local to foo() so I can do something like the following?
program new_control_routine()
...
call foo(a1,b1,...)
...
call foo(a2,b2,...)
...
end
The point is to have multiple calls of "foo" with each of them having independent COMMON BLOCKS, even though they those blocks have the same names. The only way I know is to rename the COMMON BLOCKS and related subroutines in the second call of "foo", so it becomes another routine "foo1", so I can do:
programe new_control_routine()
...
call foo(a1,b1,...)
...
call foo1(a2,b2,...)
But this is certainly mundane and error-prone.
The point of common blocks is that they are global, accessible to any routine that uses their name. Could you make the variables in the common blocks arrays, and pass an index, so that your different calls accessed different elements?
Why limit yourself to FORTRAN 77? With Fortran 95/2003 you could create a user-defined type and create a array-variable of that type. Then you pass the index of the element that you want to work on. The user-defined type can be a complex group of variables. The variable can either be an argument to the subroutines or a module variable.
I am updating legacy code and I need to use a simple mathematical function inside a subroutine. I cannot figure out how to do this. I have a function that works when called from a test program. What do I need to do differently for a subroutine?
example:
subroutine foo(i,j,k)
i = bar(j,k)
stuff = otherstuff
return
end
other info:
bar is an erf approximation.
I am using the PGF90 compiler.
I am new to FORTRAN from C.
thanks!
Basically, calling from a program or a subroutine shouldn't differ. Does the code really look like this, without any declarations? This means all variables will have implicit types: variables with names starting with the letters i-n will be integer, all others real; this also holds for function return values. The code you show, tries to assign a real (bar()) to an integer (i).
If you're writing new code, always start programs and procedures with IMPLICIT NONE. This forces you to explicitly include type declarations for all variables and function return values, greatly reducing errors.