Hi Paul,
In the comparisons on testcases where both gfortran
and amdflang produce working executables, gfortran
holds it own. As a Fortran user, it is nice to have
more than one option in compiling my codes.
However, what I dislike about this type of article is
that an exact list of options is not given. For the
gfortran comparisons, it seems that the only option
used is -O3. As amdflang is doing off-loading
to a GPU, it seems to me that the testing would use
a gfortran with off-loading capabilities. I have not
tried to verify, but suspect amdflang emits code to
detect the CPU and uses optimizations specific to that
CPU (such as fma). I wonder what the results would
look like if gfortran used '-O3 -march=native
-mtune=native' or'-Ofast' instead of '-O3'. I do
not use gfortran with a gpu, so cannot recommend possible
offloading options.
--
steve
On 10/24/26 02:14, Paul Richard Thomas wrote:
That is an interesting article but is not especially helpful for
gfortran development.
I suggest that the authors deploy an LLM, such as Claude, to extract
reproducers for the failing cases. This would be helpful for the
maintainers/developers of all three compilers. I have put the lead
author of the paper in copy to this end.
Paul
On Sat, 24 Oct 2026 at 01:22, Steve Kargl <[email protected]> wrote:
Thought this might be of interest to some.
https://dl.acm.org/doi/10.1145/3837730.3837734
--
steve
