This is the mail archive of the
gcc-bugs@gcc.gnu.org
mailing list for the GCC project.
[Bug fortran/66189] Block loops for inline matmul
- From: "tkoenig at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>
- To: gcc-bugs at gcc dot gnu dot org
- Date: Sun, 06 Sep 2015 11:32:38 +0000
- Subject: [Bug fortran/66189] Block loops for inline matmul
- Auto-submitted: auto-generated
- References: <bug-66189-4 at http dot gcc dot gnu dot org/bugzilla/>
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66189
--- Comment #2 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
(In reply to Dominique d'Humieres from comment #1)
> IMO the matmul inlining should be restricted to small matrices, thus I am
> not convinced that this worth the work.
For large matrix sizes, an external optimized BLAS is faster. This is why
inline matmul hands over to the external BLAS by default.
Our current library implementation is slower than inline matmul, so if the user
does not use -fexternal-blas, inlining still makes sense, and it also makes
sense to make this fast.