topic It's my fault. The problem in Intel® oneAPI Math Kernel Library

gemm problem

Ren_J_ — Thu, 24 Dec 2015 14:30:10 GMT

I calculate matrix matrix product in my program. But I meet some problem.

There are three matrix.

rightv dimension is larger than (svddimbefore,4*Rrealdimbefore).

midmat2 dimension is larger than (rsvddimbefore,4*lsvddimbefore).

coeffguess dimension is larger than (4*Rrealdimbefore,4*lsvddimbefore)

When I use the fortran95 interface

use blas95

call gemm(rightv(1:rsvddimbefore,1:4*Rrealdimbefore),midmat2(1:rsvddimbefore,1:4*lsvddimbefore),&
coeffguess(1:4*Rrealdimbefore,1:4*lsvddimbefore),'T','N',1.0D0,0.0D0)

It failed!

Program received signal SIGSEGV, Segmentation fault.
0: 0x000000000059e35f in dgemm_f95_ ()

But when i use f77 interface

call dgemm('T','N',4*Rrealdimbefore,4*lsvddimbefore,rsvddimbefore,1.0D0,&
rightv(1:rsvddimbefore,1:4*Rrealdimbefore),rsvddimbefore,&
midmat2(1:rsvddimbefore,1:4*lsvddimbefore),rsvddimbefore,0.0D0,coeffguess,4*subMp)

It works.

My question is that why fortran95 interface cannot work in this case.

I used fortran95 interface gemm before, it can work. But this time it doesn't.

Any idea?

Best Regards,

Jiajun

Please show a complete code

mecej4 — Thu, 24 Dec 2015 19:05:00 GMT

Please show a complete code example which demonstrates the problem. In particular, the declarations of the array variables must be known, and the matrices being passed to GEMM must conform to the specifications.

Here is an example code that is similar to what you described in passing the matrix arguments as array sections. It runs correctly (verified using Matlab).

program asec_gem
use blas95
!test calling F95 GEMM with array sections
implicit none
real A(5,4), B(4,5), C(3,3)
integer i,j
!
do i=1,5
   do j=1,4
      A(i,j)=10*i+j
      B(j,i)=100*j+i
   end do
end do
write(*,'(1x,3F8.1)')((A(i,j),j=1,3),i=1,5,2)
write(*,*)
write(*,'(1x,3F8.1)')((B(i,j),j=1,5,2),i=2,4)
write(*,*)
call gemm(A(1:5:2,1:3),B(2:4,1:5:2),C)
write(*,'(1x,3F10.1)')((C(i,j),j=1,3),i=1,3)
end

It's my fault. The problem

Ren_J_ — Tue, 23 Feb 2016 08:41:26 GMT

It's my fault. The problem has solved.

Thank you all.

Jiajun