PGI User Forum
 SearchSearch   MemberlistMemberlist     RegisterRegister   ProfileProfile    Log inLog in 

CUDA-x86.

Fortran90: Loop vectorization failed - data dependency

 
Post new topic   Reply to topic    PGI User Forum Forum Index -> Performance and Benchmarking
View previous topic :: View next topic  
Author Message
r2d2085



Joined: 26 Apr 2010
Posts: 2
PostPosted: Mon Apr 26, 2010 11:20 pm    Post subject: Fortran90: Loop vectorization failed - data dependency Reply with quote
Hi,

I'm using PGI-Fortran compilers (10.2.0) on Cray XT5 and having trouble vectorizing part of my code. I would appreciate if someone could shed some light on this issue.

I'm invoking the compiler with the following options:
ftn -fast -Minfo -c test.F90
17, Loop not vectorized: data dependency
Loop unrolled 2 times
22, Loop not vectorized: data dependency
Loop unrolled 2 times

AFAIK there are no dependencies between the loop statements and also across iterations. When I encountered a similar issue in a C program, I used the -Msafeptr flag to resolve it, is there something similar for Fortran90?

Thanks in advance.

Code:


  1 PROGRAM test
  2
  3 IMPLICIT NONE
  4
  5 type  reaction
  6 integer, pointer :: foo(:)
  7 real, pointer :: foo1(:)
  8 real, pointer :: bar(:)
  9 end type reaction
 10
 11 integer num, i, j, k
 12 type(reaction) node
 13 real, allocatable :: temp(:)
 14
 15 allocate(temp(num))
 16
 17     do i = 1, num
 18        k = node%foo(i)
 19        node%bar(k) = node%bar(k) + node%foo1(i)
 20     enddo
 21
 22     do i = 1, num
 23        node%bar(node%foo(i)) = node%bar(node%foo(i)) + node%foo1(i)
 24     enddo
 25
 26 END PROGRAM test

Back to top
View user's profile
mkcolg



Joined: 30 Jun 2004
Posts: 4996
Location: The Portland Group Inc.
PostPosted: Tue Apr 27, 2010 10:49 am    Post subject: Reply with quote
Hi r2d2085,
Code:

 17     do i = 1, num
 18        k = node%foo(i)
 19        node%bar(k) = node%bar(k) + node%foo1(i)
 20     enddo

What happens if some or all values of "k" are the same? The result will depend upon the order in which the vectors are executed and the memory updated. Since the order can vary from run to run, your results will be non-deterministic.

The compiler must be conservative (it's better to get correct answers, then wrong answers fast) and not vectorize this loop.

Hope this helps,
Mat
Back to top
View user's profile
r2d2085



Joined: 26 Apr 2010
Posts: 2
PostPosted: Tue Apr 27, 2010 11:00 pm    Post subject: Reply with quote
Thanks, Mat for the explanation. It was helpful.

I've one more question, if I make sure that "k" always has an unique value, is there some way to tell the compiler to go ahead with its vectorization?
Back to top
View user's profile
mkcolg



Joined: 30 Jun 2004
Posts: 4996
Location: The Portland Group Inc.
PostPosted: Wed Apr 28, 2010 9:10 am    Post subject: Reply with quote
Yes, you can use the "-Mnodepchk" flag. However, use with care since it can cause wrong answers of all value of K are not independent. Also, it will be applied to all loops in the file, not just this one. Because of this, you might want to use the "nodepchk" directive instead since it can be localized to particular loops.

Note that these loops still won't vectorize even with "-Mnodepchk" since there isn't enough work to justify vectorization.

Hope this helps,
Mat
Back to top
View user's profile
Display posts from previous:   
Post new topic   Reply to topic    PGI User Forum Forum Index -> Performance and Benchmarking All times are GMT - 7 Hours
Page 1 of 1

 
Jump to:  
You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum


Powered by phpBB © 2001, 2002 phpBB Group