https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83064
Bug ID: 83064
Summary: DO CONCURRENT inconsistent results
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: fortran
Assignee: unassigned at gcc dot gnu.org
Reporter: cfztol at hotmail dot com
Target Milestone: ---
This bug is related to 83017. I'm trying to parallelize a pure function with
the do concurrent construct. The following table shows compile flags and
results from the test program at the end of this message. GCC revision r254890:
Unrolled do-loop
Options Parallel Correct
-Og -ftree-parallelize-loops=2 N Y
-O1 -ftree-parallelize-loops=2 Y N - arbitrary
-O2 -ftree-parallelize-loops=2 Y N - always zero
-O3 -ftree-parallelize-loops=2 Y Y
-Ofast -ftree-parallelize-loops=2 Y Y
Modulo inside do-loop, or Indexed via host associated array
Options Parallel Correct
-Og -ftree-parallelize-loops=2 N Y
-O1 -ftree-parallelize-loops=2 Y N - arbitrary
-O2 -ftree-parallelize-loops=2 Y N - arbitrary
-O3 -ftree-parallelize-loops=2 Y N - arbitrary
-Ofast -ftree-parallelize-loops=2 Y N - arbitrary
So the loop is parallelized always, unless -Og optimization level is used.
However, the computed value of PI is only correct in a few cases, depending of
optimization level and details of the pure function. I think that is a bug -
especially since consecutive runs give different results (marked with
"arbitrary" in the table above).
Here's my test program:
program main
use, intrinsic :: iso_fortran_env
implicit none
integer, parameter :: nsplit = 4
integer(int64), parameter :: ne = 200000000
integer(int64) :: stride, low(nsplit), high(nsplit), edof(ne), i
real(real64), dimension(nsplit) :: pi
edof(1::4) = 1
edof(2::4) = 2
edof(3::4) = 3
edof(4::4) = 4
stride = ceiling(real(ne)/nsplit)
do i = 1, nsplit
high(i) = stride*i
end do
do i = 2, nsplit
low(i) = high(i-1) + 1
end do
low(1) = 1
high(nsplit) = ne
pi = 0
do concurrent (i = 1:nsplit)
pi(i) = sum(compute( low(i), high(i) ))
end do
print *, "PI", 4*sum(pi)
print *, "PI", 4*atan(1.0)
contains
pure function compute( low, high ) result( tmp )
integer(int64), intent(in) :: low, high
real(real64), dimension(nsplit) :: tmp
integer(int64) :: j, k
tmp = 0
! Unrolled loop
! do j = low, high, 4
! k = 1
! tmp(k) = tmp(k) + (-1)**(j+1) / real( 2*j-1 )
! k = 2
! tmp(k) = tmp(k) + (-1)**(j+2) / real( 2*j+1 )
! k = 3
! tmp(k) = tmp(k) + (-1)**(j+3) / real( 2*j+3 )
! k = 4
! tmp(k) = tmp(k) + (-1)**(j+4) / real( 2*j+5 )
! end do
! Loop with modulo operation
! do j = low, high
! k = mod( j, nsplit ) + 1
! tmp(k) = tmp(k) + (-1)**(j+1) / real( 2*j-1 )
! end do
! Loop with subscripting via host association
do j = low, high
k = edof(j)
tmp(k) = tmp(k) + (-1.0_real64)**(j+1) / real( 2*j-1 )
end do
end function
end program main
