FORALL inside SINGLE OpenMP Directive Error

[Tim_Stitt]

Hi all,

I'm trying to compile a code with ifort v11.0 on a linux system and receive the following compiler error:

error #7636: This statement or directive is not permitted within the body of an OpenMP SINGLE directive

FORALL (t = 1:nrows)

----^

/afs/crc.nd.edu/user/t/tstitt/packages/cp2k/intel/cp2k/makefiles/../src/lib/dbcsr_methods.F(1183): error #7636: This statement or directive is not permitted within the body of an OpenMP SINGLE directive

END FORALL

----^

compilation aborted

...

Can anyone suggest a workaround?

Thanks.

[pbkenned1]

Interesting, I don't know why SINGLE is not allowed. I can't find anything in the OpenMP 3.0 spec that forbids it.

As a workaround, try using the WORKSHARE directive; FORALL statements/constructs are allowed under WORKSHARE. Note that Intel's WORKSHARE is only executed by a single thread, so the semantics are equivalent to using SINGLE.

I'll find out from the OpenMP developers why we're disallowing FORALL under SINGLE. GNU accepts it:

> cat U77885-OpenMP-single.f90

program U77885

implicit none

integer, parameter :: n=4

real x(n), y(n,n)

integer i

!$omp parallel

!$omp single

forall(i = 1:n) x(i) = i

!$omp end single

!$omp workshare

forall(i = 1:n) y(i, i) = 2.0 * x(i)

!$omp end workshare

!$omp end parallel

print *, 'y(i, i) =',(y(i, i), i=1,n)

end program U77885

> gfortran --version

GNU Fortran (GCC) 4.5.0 20090924 (experimental) [trunk revision 152147]

> gfortran -fopenmp U77885-OpenMP-single.f90

> ./a.out

y(i, i) = 2.0000000 4.0000000 6.0000000 8.0000000

>

Patrick Kennedy
Intel Developer Support

[TimP]

Due to the implementation, as Patrick mentioned, the reported problem involves unexpected nested SINGLE regions. Patrick's suggestion removes the hidden duplication.
The next ifort release implements do concurrent, which fits some requirements better than forall.

[jimdempseyatthecove]

Tim,

>>problem involves unexpected nested SINGLE

Does this imply the FORALL is implicitly

!$OMP SINGLE
FORALL(...)
!$OMP END SINGLE

IOW the equivilent to the above !$OMP... being injected into the code stream?

Jim

[TimP]

Compile the example above with the SINGLE directives removed, for example, with -openmp -S set. You would still see calls to __kmp_single and __kmp_end_single in the expansion of FORALL. I don't know if you consider it equivalent to what you wrote.

[jimdempseyatthecove]

Yes, that is what I would consider as equivilent.

The OP could use the workshare as mentioned or possibly

INTEGER :: SINGLETHREAD

SINGLETHREAD = -1
!$OMP PARALLEL SHARED(SINGLETHREAD)
...
!$OMP SINGLE
SINGLETHREAD = OMP_GET_THREAD_NUM()
!$OMP END SINGLE
IF(OMP_GET_THREAD_NUM() .EQ. SINGLETHREAD) THEN
... (original single code here up to but not including FORALL)
FORALL(...)
... (remainder of single code here)
ENDIF ! end single section
...
!$OMP END PARALLEL

Jim Dempsey

[pbkenned1]

FORALL will be accepted withina SINGLE directive in an update to the 12.1 compiler. It was a straight bug, pure and simple. I'll post an update here when a compiler with the fix is available.

Tracking number for this defect: DPD200161616

Patrick Kennedy
Intel Developer Support

[pbkenned1]

This issue isresolved in update #7 (package l_fcompxe_2011.7.256).

Patrick