I am currently parallelising a code using Coarray Fortran. I have a communications failure and I cannot explain this mistake. Intel is awfully slow to give an answer to this problem (> 2 months).
Do you know any online courses, books, training sessions, forums,... in order to learn more about Coarrays?
There is a nice summary of coarrays in a paper by John Reid on the Fortran standards web site (www.nag.co.uk/sc22wg5)
ftp://ftp.nag.co.uk/sc22wg5/N1801-N1850/N1824.pdf
You might also check http://www.fortranplus.co.uk/ , particularly the "resources" document. It has a lot of information about training and books.
What is the nature of the "failure"? Is the program hanging (communications deadlock), or values just not getting to the place you expect?
Do you have access to a Cray system at one of the PRACE sites in Europe? If so, you could try the Cray compiler - an alternate set of diagnostic messages might be helpful.
Thank you for your answer !
I've already read this paper. Actually, I am looking for something more detailed which deals also with object-oriented programming and coarrays.
Resources from fortranplus look great. It will be very helpful.
Concerning the failure: the program is written in an object-oriented way. When an object on an image i asks for an integer located in the "same" object on the image j/=i, it is not able to get it correctly. The value that it gets is wrong (no compile time error nor execution error). What I write looks like:
subroutine sub1(this)
class(main_obj) , codimension[*] :: this
this%obj2%nbx = this[1]%obj2%nbx
end subroutine
sub 1 is only called from this_image() > 1. If this[1]%obj2%nbx = 10, I get something like this[3]%obj2%nbx = 95842 for example.
I have no access to Cray computers. How can I have one or where can I ask for one ?
I tried a simple example that includes your subroutine sub1:
> cat test.f90
module modu
implicit none
type :: sub_obj
integer :: nbx
end type sub_obj
type :: main_obj
type(sub_obj) :: obj2
contains
procedure :: sub1
end type main_obj
contains
subroutine sub1(this)
class(main_obj) , codimension[*] :: this
this%obj2%nbx = this[1]%obj2%nbx
end subroutine
end module modu
program test
use modu
implicit none
type(main_obj) :: x[*]
integer :: mype
mype = this_image()
x%obj2%nbx = mype
sync all
if (mype /= 1) then
call x%sub1()
print *, "value = ",x%obj2%nbx, " on image ",mype
end if
end program test
> ftn test.f90
> aprun -n8 ./a.out
value = 1 on image 3
value = 1 on image 7
value = 1 on image 5
value = 1 on image 6
value = 1 on image 4
value = 1 on image 2
value = 1 on image 8
Application 12795879 resources: utime ~0s, stime ~0s,
which seems to work as expected (using the Cray compiler). If this is failing for your environment, perhaps a bug report is in order.
For access to a Cray system for testing, the place to start in Europe is probably http://www.prace-ri.eu of which Belgium is a member. Both the UK and Finland provide Cray systems as part of this consortium.
I've modified your code to be closer to what I really wrote and I was able to reproduce the bug. The project is in attachment. Using the Intel compiler, I get the following output:
value = 1 on image 1
value = 1 on image 2
value = 1 on image 6
value = 1 on image 8
value = 1 on image 3
value = 1 on image 5
value = 1 on image 4
value = 1 on image 7
When I switch lines 14 and 15 in ComMod.f90, I have the following output:
value = 1 on image 1
value = 0 on image 2
value = 0 on image 8
value = 0 on image 6
value = 0 on image 5
value = 0 on image 3
value = 0 on image 4
value = 0 on image 7
which is not correct any more Are you able to reproduce this with the Cray compiler or am I doing something wrong ?
Just for clarification in your example, I added comments:
type :: main_obj
class(other_obj) , pointer :: ptr ! line 15
type(sub_obj) :: obj2 ! line 14
contains
procedure :: sub1
end type main_obj
With the Cray compiler I get the same output with lines 14 and 15 in either order. The order of these components should not matter.
I suspect the (Intel compiler) problem is in the execution of your version of sub1:
subroutine sub1(this)
class(main_obj) , codimension[*] :: this
allocate(other_obj :: this%ptr)
this%obj2%nbx = this[1]%obj2%nbx
end subroutine
If you remove the ALLOCATE statement, is the output correct for the original order of lines 14 and 15? If so, I would focus on the allocate in the bug report.
After some more tests, it seems that the allocate statement is not the problem. Only the order in which lines 14 and 15 are written plays a role.
I will report the bug at Intel.
Thank you for your help !!
The following article may be of interest:
Article Fortran 2008 coarrays
- Badges
- Science topic
- Similar topics
- Filing