https://gcc.gnu.org/bugzilla/show_bug.cgi?id=51119
--- Comment #33 from Jerry DeLisle <jvdelisle at gcc dot gnu.org> ---
With #pragma GCC optimize ( "-O3" )
$ gfc -static -O2 -finline-matmul-limit=0 compare.f90
$ ./a.out
=========================================================
================ MEASURED GIGAFLOPS =
=========================================================
Matmul Matmul
fixed Matmul variable
Size Loops explicit refMatmul assumed explicit
=========================================================
2 2000 0.055 0.051 0.038 0.056
4 2000 0.408 0.274 0.318 0.408
8 2000 0.644 0.711 1.287 1.831
16 2000 2.507 2.591 2.521 2.579
32 2000 3.573 2.300 3.506 3.573
64 2000 4.628 2.196 4.462 4.629
128 2000 5.030 2.393 5.304 5.054
256 477 4.802 2.367 5.573 4.854
512 59 3.907 1.856 5.234 4.035
1024 7 3.891 1.178 5.222 4.022
2048 1 3.901 1.500 5.238 4.033
and with no #pragma it is better than the -O3 version
$ gfc -static -O2 -finline-matmul-limit=0 compare.f90
$ ./a.out
=========================================================
================ MEASURED GIGAFLOPS =
=========================================================
Matmul Matmul
fixed Matmul variable
Size Loops explicit refMatmul assumed explicit
=========================================================
2 2000 0.054 0.052 0.043 0.057
4 2000 0.397 0.281 0.316 0.414
8 2000 0.691 0.773 1.831 1.995
16 2000 2.493 2.691 2.521 2.512
32 2000 3.629 2.301 3.623 3.572
64 2000 4.557 2.072 4.568 4.468
128 2000 5.282 2.387 5.291 5.284
256 477 5.629 2.369 5.620 5.605
512 59 5.215 1.874 5.240 5.216
1024 7 5.212 1.174 5.217 5.217
2048 1 5.230 1.499 5.234 5.229
Still a good improvement over gfortran6 on the larger matrices.
