MATMUL fails in checking arguments shapes

MATMUL fails in checking arguments shapes

Hi all.

I'm using Intel(R) Fortran Intel(R) 64 Compiler for applications running on Intel(R) 64, Version 18.0.1.163 Build 20171018.

It seems to me that the shapes of MATMUL arguments are not checked for conformability if some of them is an allocatable array.
The consequent result are clearly wrong.

Here is a short example, in which the two arguments are clearly non conformable, being one a 1-by-1 array and the other a 5-by-5 array.
Nevertheless, neither at compile-time (that's understandable) nor at runtime there is a complain when performing MATMUL.

That does not change if the arrays are allocated explicitly.

Note that when compiled using gfortran (I used version 4.9.2) a runtime error is issued.

Did I miss some specific compiling option ?
Is it left to the programmer to ensure arguments conformability ?

Thanks a lot.

GM

 

t_matmul.f90

integer, parameter :: n = 5
real, allocatable :: y(:, :)
real, allocatable :: z(:, :)
integer :: i, j

y = reshape( [ ((i+j/100.0 , i = 1, n), j=1, n) ], [ n, n ])
z = reshape([1.0], [1, 1])

write(*, *) "z: size ", size(z), "   shape ", shape(z)
write(*, *) "y: size ", size(y), "   shape ", shape(y)
write(*, *)

DO i = 1, size(y,1)
  WRITE(*, '(19F10.2)') (y(i, j), j=1, size(y,2))
ENDDO
write(*, *)

write (*, *) "trying y = matmul(z, y) "
write(*, *)
y = matmul(z, y)

DO i = 1, size(y,1)
  WRITE(*, '(19F10.2)') (y(i, j), j=1, size(y,2))
ENDDO
write(*, *)
end
$ ifort -warn All -check All -stdlev=f2008 t_matmul.f90 -o t_matmul
$ ./t_matmul
 z: size            1    shape            1           1
 y: size           25    shape            5           5
 
      1.01      1.02      1.03      1.04      1.05
      2.01      2.02      2.03      2.04      2.05
      3.01      3.02      3.03      3.04      3.05
      4.01      4.02      4.03      4.04      4.05
      5.01      5.02      5.03      5.04      5.05
 
 trying y = matmul(z, y)
 
      1.01      1.02      1.03      1.04      1.05


$ gfortran t_matmul.f90 && ./a.out
 Z: size            1    shape            1           1
 y: size           25    shape            5           5

      1.01      1.02      1.03      1.04      1.05
      2.01      2.02      2.03      2.04      2.05
      3.01      3.02      3.03      3.04      3.05
      4.01      4.02      4.03      4.04      4.05
      5.01      5.02      5.03      5.04      5.05

 trying y = matmul(z, y)

Fortran runtime error: dimension of array B incorrect in MATMUL intrinsic


8 posts / 0 new
Last post
For more complete information about compiler optimizations, see our Optimization Notice.

The compiler is not obligated by the standard to find such a mismatch in shapes. It is a quality of implementation though. Nagfor (which is considered to be a very good analysing/debugging compiler gets it when compiled with -C=all):

Runtime Error: matmul.f90, line 21: Incompatible argument shapes (1x1 and 5x5) to intrinsic MATMUL

gfortran with -fcheck=all gets it as well as you recognised. Indeed, after a quick glance I could also not find a flag for the Intel compiler to diagnose this. 

 

Thanks for your answer, Juergen.

Yes, I understand, and that's why I often use different compilers to catch all possible flaws in my code.

Expecially when you are using automatically allocated allocatable arrays, it would be a valuable help having the compiler checking these issues.

I expected Intel compiler (as well as other commmercial compilers) would be able to do that, given that even gfortan does.

I'll try to live with this.

GM

 

Just wait, maybe one of the Intel folks knows something that is not directly visible in the list of options. 

That would be great !

Best Reply

MATMUL needs to validate array dimensions.

The array shape checking feature was implemented in Intel Fortran Compiler 19.0 Beta and it checks for array shape conformance.

If you write your own MATMUL and it inlines, then the compiler should catch the mismatch in shape at runtime when compiled with –check shape. See https://software.intel.com/en-us/articles/array-shape-check-new-in-intel-fortran-compiler-190

Devorah - Intel® Developer Support

Thanks, Devorah.

I'll be waiting for the Intel Fortran Compiler 19.0 final relase.

GM

Confirmed fixed in upcoming 19.0 final release:

ifort /check:shape matmul001.f90 
d:\matmul001.exe

z: size            1    shape            1           1

y: size           25    shape            5           5



      1.01      1.02      1.03      1.04      1.05

      2.01      2.02      2.03      2.04      2.05

      3.01      3.02      3.03      3.04      3.05

      4.01      4.02      4.03      4.04      4.05

      5.01      5.02      5.03      5.04      5.05



trying y = matmul(z, y)



forrtl: severe (408): fort: (33): Shape mismatch: The extent of dimension 2 of array Z is 1 and the corresponding extent

of array Y is 5



Image              PC                Routine            Line        Source

matmul001.exe      000000013F433911  Unknown               Unknown  Unknown

matmul001.exe      000000013F4218C7  Unknown               Unknown  Unknown

matmul001.exe      000000013F48BD82  Unknown               Unknown  Unknown

matmul001.exe      000000013F47395B  Unknown               Unknown  Unknown

kernel32.dll       0000000076C759CD  Unknown               Unknown  Unknown

ntdll.dll          0000000076ED383D  Unknown               Unknown  Unknown



!!! matmul001.f90

integer, parameter :: n = 5

real, allocatable :: y(:, :)

real, allocatable :: z(:, :)

integer :: i, j



y = reshape( [ ((i+j/100.0 , i = 1, n), j=1, n) ], [ n, n ])

z = reshape([1.0], [1, 1])



write(*, *) "z: size ", size(z), "   shape ", shape(z)

write(*, *) "y: size ", size(y), "   shape ", shape(y)

write(*, *)



DO i = 1, size(y,1)

  WRITE(*, '(19F10.2)') (y(i, j), j=1, size(y,2))

ENDDO

write(*, *)



write (*, *) "trying y = matmul(z, y) "

write(*, *)

y = matmul(z, y)



DO i = 1, size(y,1)

  WRITE(*, '(19F10.2)') (y(i, j), j=1, size(y,2))

ENDDO

write(*, *)

end

 

Devorah - Intel® Developer Support

Leave a Comment

Please sign in to add a comment. Not a member? Join today