LI77084: -QHOT AT LEVEL=2 MAY BE SLOWER THAN AT LEVEL=1

XL C/C++ for Blue Gene/Q Fix Pack 12 (May 2015 Update) for 12.1
XL C/C++ for Blue Gene/Q Fix Pack 14 (May 2016 Update) for 12.1
XL C/C++ for Blue Gene/Q Fix Pack 15 (October 2017 Update) for 12.1
February 2013 Update for XL C/C++ for Blue Gene/Q, V12.1
May 2013 Update for XL C/C++ for Blue Gene/Q, V12.1
XL C/C++ for Blue Gene/Q Fix Pack 5 (August 2013 Update) for 12.1
XL C/C++ for Blue Gene/Q Fix Pack 6 (November 2013 Update) for 12.1
XL C/C++ for Blue Gene/Q Fix Pack 7 (February 2014 Update) for 12.1
XL C/C++ for Blue Gene/Q Fix Pack 8 (May 2014 Update) for 12.1
XL C/C++ for Blue Gene/Q Fix Pack 9 (August 2014 Update) for 12.1
XL C/C++ for Blue Gene/Q Fix Pack 10 (November 2014 Update) for 12.1
XL C/C++ for Blue Gene/Q Fix Pack 11 (February 2015 Update) for 12.1
XL C/C++ for Blue Gene/Q Fix Pack 13 (August 2015 Update) for 12.1

APAR status

• Closed as suggestion for future release.

Error description

• In general, -qhot=level=2 is available to the user to perform
  more aggressive loop transformations through the polyhedral
  framework, however the performance is not guaranteed to be as
  good or better than -qhot=level=1.
  
  The following example identifies a situation when
  the code compiled with -qhot=level=2 may get slower than the
  code compiled with -qhot=level=1:
  
        ...
        ...
        do j=1,n
        do i=1,n
        dr(i,j)=ar(i,j)+br(i,j)*cr(i,j)-bi(i,j)*ci(i,j)
        di(i,j)=ai(i,j)+br(i,j)*ci(i,j)+bi(i,j)*cr(i,j)
        end do
        end do
  
  Because the access to all of these arrays are stride-1, and
  there are no loop carried dependencies, XL compiler is able to
  easily vectorize the inner loop.
  
  The loop transformations performed under -qhot=level=1 iterates
  through a series of transformations and identifies opportunities
  that are profitable while also taking preventative measures not
  to prevent opportunities from other transformations that are
  deemed to be more profitable. In the case above it is able to
  identify the SIMD opportunity and focus on vectorizing the inner
  loop
  
  Under -qhot=level=2 XL compiler's polyhedral framework is much
  more aggressive and is solely focused on maximizing the
  performance of the entire loop nest without taking into
  consideration further opportunities that can be exploited.
  
  In the case above, -qhot=level=2 decides to perform several
  transformations that are profitable, but also prevents XL
  compiler's auto-simd transformations from occuring.
  
  If the source is built with -qsimd=noauto and -qhot=level=2
  option combination, the performance of the resulting binary may
  be better than the one compiled with -qhot=level=1.
  Currently, the code compiled with -qhot=level=2 option fails to
  recognize the simd opportunity that the compiler was able to
  exploit for -qhot=level=1 compilation which leads to the
  performance regression at -qhot=level=2.
  

Local fix

• For loop intensive High-Performance Computing workloads it is
  recommended to use -O3 or -O3 -qhot at compile and link time.
  By default, -qhot implies -qhot=level=1.
  

Problem summary

Problem conclusion

Temporary fix

Comments

• According to development team, the runtime performance of the
  code provided by client and compiled with -qhot=level=2 will be
  on par or better than the one compiled with -qhot=level=1
  command line options in the future version of XL compilers.
  

APAR Information

• APAR is sysrouted FROM one or more of the following:

• APAR is sysrouted TO one or more of the following:

Fix information

Applicable component levels