Re: (Re)allocation of allocatable arrays on assignment - F2003

[dominiq at lps dot ens dot fr (Dominique Dhumieres)]

Paul,

I realized that it could be faster to explain the origin of the
bound problems. A reported in pr38318 comment 5, the polyhedron
test nf.f90 use automatic arrays in NF2DPrecon and NF3DPrecon.
On my macbook it results a system overhead due to the corresponding
alloc/free of between 7 to 9s (depending on the mood of the macbook)
for a user time of ~20s. I think this test can use automatic allocation
with minimal surgery, hence allowing to compare the costs of automatic
arrays and automatic allocation.

So far my only successful attempt was to use t=x before the loops
in NF2DPrecon. However this results in moving at least 50 time
more data than necessary for a net loss of a factor 4. All my
other attemps led to segmentation faults due to the reported
bound problems.

Cheers,

Dominique