Intel® Fortran Compiler
Build applications that can scale for the future with optimized code designed for Intel® Xeon® and compatible processors.
29277 Discussions

What does #6541 mean?

JVanB
Valued Contributor II
‎02-05-2016 06:04 PM
1,094 Views

There was a question in stack overflow about implementing Matlab's diff function in Fortran. First attempt:

! diff.f90
module M
   use ISO_FORTRAN_ENV
   implicit none
   private
   public diff
   interface diff
      module procedure diff2
   end interface diff
   contains
recursive function diff2(A,n,dim) result(B)
   real(REAL64)  A(:,:)
   real(REAL64), allocatable :: B(:,:)
   integer, optional :: n, dim
   integer Ndim
   integer shB(rank(A))
   integer shA(rank(A))
   integer peB(rank(A))
   integer i

   if(present(n)) then
      if(n <= 0) then
!         allocate(B,source=A)
         B = A
         return
      else if(n > 1) then
!         allocate(B,source=diff(diff(A,n-1,dim),1,dim))
         B = diff(diff(A,n-1,dim),1,dim)
         return
      end if
   end if
   if(present(dim)) then
      if(dim <= 0 .OR. dim > rank(A)) then
!         allocate(B,source=A)
         B = A
         return
      else
         Ndim = dim
      end if
   else
      Ndim = 1
   end if
   shB = shape(A)
   shB(Ndim) = shB(Ndim)-1
   if(any(shB <= 0)) then
      allocate(B(shB(1),shB(2)))
      return
   end if
   peB = [(i,i=1,rank(A))]
   if(Ndim /= rank(A)) then
      peB([Ndim,rank(A)]) = peB([rank(A),Ndim])
   end if
   shA = shape(A)
!   allocate(B,source=reshape(reshape(eoshift(A,1,dim=Ndim)-A,shB1,order=peB),shB,order=peB))
   B = reshape(reshape(eoshift(A,1,dim=Ndim)-A,shA(peB),order=peB),shB,order=peB)
end function diff2
end module M

! main.f90
program P
   use M
   use ISO_FORTRAN_ENV
   implicit none
   real(REAL64), allocatable :: A(:,:), B(:,:)
   integer M1, N
   character(20) fmt
   integer i,j
   integer dim

   do dim = 1, rank(A)
      M1 = 3
      N = 4
      write(*,'(a)') 'A ='
      A = reshape([((10*i+j,i=1,M1),j=1,N)],[M1,N])
      write(fmt,'(*(g0))') '(',size(A,2),'(g0:1x))'
      write(*,fmt) transpose(A)
      B = diff(A,dim=dim)
      write(fmt,'(*(g0))') '(',size(B,2),'(g0:1x))'
      write(*,'(a)') 'B ='
      write(*,fmt) transpose(B)
   end do
end program P

Worked OK with gfortran, but with ifort I got:

diff.f90(55): error #6541: This element is not yet supported in this context.
[SHA]
   B = reshape(reshape(eoshift(A,1,dim=Ndim)-A,shA(peB),order=peB),shB,order=peB
)
-----------------------------------------------^
diff.f90(55): error #6361: An array-valued argument is required in this context.
   [RESHAPE]
   B = reshape(reshape(eoshift(A,1,dim=Ndim)-A,shA(peB),order=peB),shB,order=peB
)
---------------^
compilation aborted for diff.f90 (code 1)

And there is that error #6541, so what does it mean?

I figured the way to go was to create a rank-invariant expression for diff(A,1,dim=Ndim) so that it could be used in a function or subroutine with an assumed-rank input. Is this possible? Will it be possible when SELECT RANK is available? Actually a test program that incrementally wrote out a rank-invariant expression seemed to work OK, both on gfortran and ifort:

program Q
   use ISO_FORTRAN_ENV
   implicit none
   real(REAL64), allocatable :: A(:,:,:)
   integer M, N, P
   integer i, j, k
   character(30) fmt
   integer Ndim

   M = 3
   N = 4
   P = 5
do Ndim = 1, rank(A)
   allocate(A(M,N,P))
   write(fmt,'(a,i0,a)') '(',size(A,2),'(g0:1x))'
   A = reshape([(((100*i**2+10*j**2+k**2,i=1,M),j=1,N),k=1,P)],shape(A))
   write(*,'(a)') 'A ='
do i = 1, P
   write(*,fmt) transpose(A(:,:,i))
   write(*,'()')
end do
   write(*,'(*(g0:1x))') 'shape(A) =',shape(A)
   write(*,'(*(g0:1x))') '[(i,i=1,rank(A))] =',[(i,i=1,rank(A))]
   write(*,'(*(g0:1x))') 'eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) =', &
      eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim)
   write(*,'(*(g0:1x))') 'eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)) =', &
      eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A))
   write(*,'(*(g0:1x))') '[(i,i=1,rank(A))] + ' // &
      'eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) + ' // &
      'eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)) =', &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) + &
      eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A))
   write(fmt,'(a,i0,a)') '(',rank(A),'(g0:1x))'
   write(*,'(a)') 'spread([1,(0,i=2,rank(A))],2,rank(A)) ='
   write(*,fmt) transpose(spread([1,(0,i=2,rank(A))],2,rank(A)))
   write(*,'(a)') 'eoshift(' // &
      'spread([1,(0,i=2,rank(A))],2,rank(A)), 1-(' // &
      '[(i,i=1,rank(A))] +' // &
      'eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) +' // &
      'eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)))) ='
   write(*,fmt) eoshift( &
      spread([1,(0,i=2,rank(A))],2,rank(A)), 1-( &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) + &
      eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A))))
   write(*,'(a)') 'matmul(eoshift(' // &
      'spread([1,(0,i=2,rank(A))],2,rank(A)), 1-(' // &
      '[(i,i=1,rank(A))] +' // &
      'eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) +' // &
      'eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)))),' // &
      'shape(A)) ='
   write(*,'(*(g0:1x))') matmul(eoshift( &
      spread([1,(0,i=2,rank(A))],2,rank(A)), 1-( &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) + &
      eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)))), &
      shape(A))
   write(fmt,'(*(g0))') '(',M,'(',N,'(g0:1x)/))'
   write(*,'(a)') 'eoshift(A,1,dim=Ndim)-A ='
   write(*,fmt) reshape(eoshift(A,1,dim=Ndim)-A,[N,M,P],order=[2,1,3])
   write(fmt,'("(",i0,"(",i0,"(g0:1x)/))")') matmul(reshape([1,0,0, &
                                                            0,1,0], &
      [2,3],order=[2,1]), &
      matmul(eoshift( &
      spread([1,(0,i=2,rank(A))],2,rank(A)), 1-( &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) + &
      eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)))), &
      shape(A)))
   write(*,'(a)') 'reshape(' // &
      'eoshift(A,1,dim=Ndim)-A,' // &
      'matmul(eoshift(' // &
      'spread([1,(0,i=2,rank(A))],2,rank(A)), 1-(' // &
      '[(i,i=1,rank(A))] +' // &
      'eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) +' // &
      'eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)))),' // &
      'shape(A)),' // &
      'order =' // &
      '[(i,i=1,rank(A))] +' // &
      'eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) +' // &
      'eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)))'
   write(*,fmt) reshape(reshape( &
      eoshift(A,1,dim=Ndim)-A, &
      matmul(eoshift( &
      spread([1,(0,i=2,rank(A))],2,rank(A)), 1-( &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) + &
      eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)))), &
      shape(A)), &
      order = &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) + &
      eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A))), &
      matmul(eoshift( &
      spread([1,(0,i=2,rank(A))],2,rank(A)), 1-( &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) + &
      eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)))), &
      shape(A)), &
      order=[2,1,3])
   write(*,'(a)') 'shape(A)-eoshift([1,(0,i=2,rank(A))],1-Ndim) ='
   write(*,'(*(g0:1x))') shape(A)-eoshift([1,(0,i=2,rank(A))],1-Ndim)
   write(fmt,'(*(g0))') '(',M-merge(1,0,Ndim==1),'(',N-merge(1,0,Ndim==2),'(g0:1x)/))'
   write(*,'(a)') 'reshape(' // &
      'reshape(' // &
      'eoshift(A,1,dim=Ndim)-A,' // &
      'matmul(eoshift(' // &
      'spread([1,(0,i=2,rank(A))],2,rank(A)), 1-(' // &
      '[(i,i=1,rank(A))] +' // &
      'eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) +' // &
      'eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)))),' // &
      'shape(A)),' // &
      'order =' // &
      '[(i,i=1,rank(A))] +' // &
      'eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) +' // &
      'eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A))),' // &
      'shape(A)-eoshift([1,(0,i=2,rank(A))],1-Ndim),' // &
      'order =' // &
      '[(i,i=1,rank(A))] +' // &
      'eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) +' // &
      'eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)))'
   write(*,fmt) reshape(reshape( &
      reshape( &
      eoshift(A,1,dim=Ndim)-A, &
      matmul(eoshift( &
      spread([1,(0,i=2,rank(A))],2,rank(A)), 1-( &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) + &
      eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A)))), &
      shape(A)), &
      order = &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) + &
      eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A))), &
      shape(A)-eoshift([1,(0,i=2,rank(A))],1-Ndim), &
      order = &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim,(0,i=2,rank(A))],1-Ndim) + &
      eoshift([Ndim-rank(A),(0,i=2,rank(A))],1-rank(A))), &
      [N-merge(1,0,Ndim==2),M-merge(1,0,Ndim==1),P-merge(1,0,Ndim==3)],order=[2,1,3])
   deallocate(A)
end do
end program Q

(ifort needs /assume:realloc_lhs for the above.) Encouraged by this result, I started on a version with a constant rank-invariant expression:

program P
   use ISO_FORTRAN_ENV
   implicit none
   integer, parameter :: M = 3
   integer, parameter :: N = 4
   real(REAL64) :: A(M,N)
   integer i, j
   parameter(A = reshape([((10*i**2+j**2,i=1,M),j=1,N)],shape(A)))
   integer, parameter :: Ndim1 = 1
   real(REAL64) :: B1(M-merge(1,0,Ndim1==1),N-merge(1,0,Ndim1==2))
   parameter(B1 = reshape( &
      reshape( &
      eoshift(A,1,dim=Ndim1)-A, &
      matmul(eoshift( &
      spread([1,(0,i=2,rank(A))],2,rank(A)), 1-( &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim1,(0,i=2,rank(A))],1-Ndim1) + &
      eoshift([Ndim1-rank(A),(0,i=2,rank(A))],1-rank(A)))), &
      shape(A)), &
      order = &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim1,(0,i=2,rank(A))],1-Ndim1) + &
      eoshift([Ndim1-rank(A),(0,i=2,rank(A))],1-rank(A))), &
      shape(A)-eoshift([1,(0,i=2,rank(A))],1-Ndim1), &
      order = &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim1,(0,i=2,rank(A))],1-Ndim1) + &
      eoshift([Ndim1-rank(A),(0,i=2,rank(A))],1-rank(A))))
   integer, parameter :: Ndim2 = 2
   real(REAL64) :: B2(M-merge(1,0,Ndim2==1),N-merge(1,0,Ndim2==2))
   parameter(B2 = reshape( &
      reshape( &
      eoshift(A,1,dim=Ndim2)-A, &
      matmul(eoshift( &
      spread([1,(0,i=2,rank(A))],2,rank(A)), 1-( &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim2,(0,i=2,rank(A))],1-Ndim2) + &
      eoshift([Ndim2-rank(A),(0,i=2,rank(A))],1-rank(A)))), &
      shape(A)), &
      order = &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim2,(0,i=2,rank(A))],1-Ndim2) + &
      eoshift([Ndim2-rank(A),(0,i=2,rank(A))],1-rank(A))), &
      shape(A)-eoshift([1,(0,i=2,rank(A))],1-Ndim2), &
      order = &
      [(i,i=1,rank(A))] + &
      eoshift([rank(A)-Ndim2,(0,i=2,rank(A))],1-Ndim2) + &
      eoshift([Ndim2-rank(A),(0,i=2,rank(A))],1-rank(A))))
   character(30) fmt

   write(fmt,'(*(g0))') '(',size(A,2),'(g0:1x))'
   write(*,'(a)') 'A ='
   write(*,fmt) transpose(A)
end program P

But that one got a C0000005 with ifort.

 

0 Kudos

Link Copied

6 Replies
Steven_L_Intel1
Employee
‎02-06-2016 07:37 AM
1,094 Views

I have no idea what that means. If this were a specification or initialization expression, there might be some missing functionality but this is just an assignment. I will ask the developers.

0 Kudos
Copy link
JVanB
Valued Contributor II
‎02-06-2016 07:45 AM
1,094 Views

Well it's sort of an initialization expression in that

20 SHAPE shall be a rank-one integer array. SIZE (x), where x is the actual argument corresponding to

21 SHAPE, shall be a constant expression whose value is positive and less than 16. It shall not have

22 an element whose value is negative.

0 Kudos
Copy link
Steven_L_Intel1
Employee
‎02-06-2016 08:00 AM
1,094 Views

Hmm, I see. SHA(Peb) is not a constant expression but it is required to be. Maybe that's what the compiler tried to say. Still a rather unhelpful message.

0 Kudos
Copy link
JVanB
Valued Contributor II
‎02-06-2016 09:01 AM
1,094 Views

Looking over the standard, shA(peB) is a variable in that shA is a variable and this is an array section of it with a vector subscript. Then the question is whether SIZE(shA(peB)) is a constant expression, and that flies because SIZE is an array inquiry function, so it's a specification inquiry and the result comes ultimately from SIZE(peB), which is given by RANK(A), and that is OK as a constant expression via the specification inquiry clause, all the way from F95 at least (if F95 had a RANK intrinsic).

We can confirm that SIZE(shA(peB)) is a constant expression by running it through our handy-dandy constant expression tester:

module M
   use ISO_FORTRAN_ENV
   implicit none
   contains
      recursive subroutine S(A)
         real(REAL64) A(:,:)
         integer peB(rank(A))
         integer shA(rank(A))
         type T
            integer :: x = 0
         end type T
         type(T) x(SIZE(shA(peB)))
         save
         print '(i0)', x(1)%x
         x(1)%x = 1
      end subroutine S
end module M

program P
   use M
   implicit none
   real(REAL64) A(2,2)
   A = 7
   call S(A)
   call S(A)
end program P

And both ifort and gfortran print out:

0
1

So it is a constant expression.

 

0 Kudos
Copy link
Steven_L_Intel1
Employee
‎02-07-2016 04:22 PM
1,094 Views

The first case escalated as DPD200381581. The ICE is DPD200381582.

0 Kudos
Copy link
Steven_L_Intel1
Employee
‎07-15-2016 06:31 AM
1,094 Views

First case (DPD200381581) is fixed for the 17.0 product release. The ICE is still in progress.

0 Kudos
Copy link
Reply

Community support is provided Monday to Friday. Other contact methods are available here.

Intel does not verify all solutions, including but not limited to any file transfers that may appear in this community. Accordingly, Intel disclaims all express and implied warranties, including without limitation, the implied warranties of merchantability, fitness for a particular purpose, and non-infringement, as well as any warranty arising from course of performance, course of dealing, or usage in trade.

For more complete information about compiler optimizations, see our Optimization Notice.