https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81108
Bug ID: 81108
Summary: OpenMP doacross (omp do/for ordered) performance
Product: gcc
Version: 7.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libgomp
Assignee: unassigned at gcc dot gnu.org
Reporter: jeff.science at gmail dot com
CC: jakub at gcc dot gnu.org
Target Milestone: ---
Created attachment 41560
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=41560&action=edit
sequential code
When I use "doacross" OpenMP parallelism as described in
https://developers.redhat.com/blog/2016/03/22/what-is-new-in-openmp-4-5-3/, the
performance is ~30x worse than sequential execution.
This code can be effectively parallelized by blocking the wavefront, which I
have implemented as well. I see the same behavior in Fortran and C++, so I
believe it is relatively agnostic to the front-end. Clang has the same issue,
which we determined is because the dependency analysis is applied to the
collapsed loop, inhibiting all parallelism, although I don't understand why
this is so much worse than a serial implementation.
If it helps, the project is https://github.com/ParRes/Kernels, although I will
attach all of the code here.
# Sequential
do j=2,n
do i=2,m
grid(i,j) = grid(i-1,j) + grid(i,j-1) - grid(i-1,j-1)
enddo
enddo
$ ./p2p 40 1000 1000
Parallel Research Kernels
Fortran Serial pipeline execution on 2D
Number of iterations = 40
Grid sizes = 1000 1000
> traverse in the m dimension
Solution validates
Rate (MFlop/s): 866.508739 Avg time (s): 0.002303
# OpenMP "doacross"
!$omp do ordered(2) collapse(2)
do j=2,n
do i=2,m
!$omp ordered depend(sink:j,i-1) depend(sink:j-1,i)
depend(sink:j-1,i-1)
grid(i,j) = grid(i-1,j) + grid(i,j-1) - grid(i-1,j-1)
!$omp ordered depend(source)
enddo
enddo
!$omp end do
$ ./p2p-openmp-doacross 40 1000 1000
Parallel Research Kernels
Fortran OpenMP pipeline execution on 2D
Number of threads = 4
Number of iterations = 40
Grid sizes = 1000 1000
Solution validates
Rate (MFlop/s): 17.855468 Avg time (s): 0.111787