https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88713
Thomas Koenig <tkoenig at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |tkoenig at gcc dot gnu.org
Blocks| |36854
Severity|normal |enhancement
--- Comment #4 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
First, a few remarks on the code: It is written in a suboptmial way
regarding the way Fortran lays out its memory.
The code does
real, dimension(N,3), intent(out) :: X
real, dimension(N,10), intent(in) :: BPP
and then
do concurrent (i = 1:N)
X(i,:) = fpdbacksolve(BPP(i,1:3), BPP(i,5:10))
end do
The problem is that BPP(i,1:3) is not contiguous in memory.
Fortran lays out the memory for that array as
BPP(1,1), BPP(2,1), BPP(3,1), BPP(4,1), ..., BPP(1,2)
so you are accessing your memory with a stride of n in the
expressions BPP(i,1:3) and BPP(i,5:10). This is very inefficient
anyway, vectorization would not really help in this case.
So, if you change your code to
real, dimension(3,N), intent(out) :: X
real, dimension(10,N), intent(in) :: BPP
do concurrent (i = 1:N)
X(i,:) = fpdbacksolve(BPP(1:3,i), BPP(5:10,i))
end do
then processBPP will be inlined completely, and memory accesses will
be contiguous.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=36854
[Bug 36854] [meta-bug] fortran front-end optimization
