| commit | f4d7d23a8bcdbe5bc0eddbbf5bbcfa8669333dc5 | |
| author | Sven Verdoolaege <skimo@kotnet.org> | |
| Fri, 3 Jun 2011 08:39:19 +0000 (3 10:39 +0200) | ||
| committer | Sven Verdoolaege <skimo@kotnet.org> | |
| Sun, 5 Jun 2011 08:55:22 +0000 (5 10:55 +0200) | ||
| tree | 6b25017f56b9f652d90d899e8511bbb4bc44e27d | treesnapshot (tar.gz zip) |
| parent | e3bf06f2245db9624608e1aa6382c66cd3b752c9 | commitdiff |
scheduling: optionally split parallel rows in non-tilable bands
If a scheduling row that is constructed to carry dependences
is of the form, say, { S1[..] -> [2 * g(..)]; S2[..] -> [2 * g(..) + 1] },
then it is usually better to split off the constant term into a separate
band and obtain { S1[..] -> [g(..),0]; S2[..] -> [g(..),1] } instead.
In particular, the code generated from the second schedule is usually
simpler than that from the first schedule.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
If a scheduling row that is constructed to carry dependences
is of the form, say, { S1[..] -> [2 * g(..)]; S2[..] -> [2 * g(..) + 1] },
then it is usually better to split off the constant term into a separate
band and obtain { S1[..] -> [g(..),0]; S2[..] -> [g(..),1] } instead.
In particular, the code generated from the second schedule is usually
simpler than that from the first schedule.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
| include/isl/options.h | diffblobblamehistory | |
| isl_options.c | diffblobblamehistory | |
| isl_schedule.c | diffblobblamehistory |