Graphite phone call from 2009_04_01
Attendees: Razya, Li, Konrad, Jan, Tobi, David, Sebastian, Christophe
Discussed topics:
- Data dependencies: Tobias committed a patch filling the access polyhedrons
- Not yet enabled as there are some bootstrap bugs
- All arrays access the same alias set
- Reductions are not yet included in the access polyhedrons (Same as last week, but now really committed)
- Bug in building of scattering matrices fixed.
- Bug in condition detection on its way to be fixed.
- Automatic Parallelization:
- Li will start integrating graphite and code generation (If his summer of code proposal is accepted)
- Mark all loops parallel if new flag "-fgraphite-force-parallel" is set.
- How to connect auto par?
- Directly in graphite code generation (clast to gimple): We call the autopar backend after generating a serial gimple loop. This is the easiest approach. Also we still access to the polyhedral information so we can use it to give autopar more information.
- By annotating loops: In the graphite pass we annotate gimple loops as parallel . Later the autopar pass is scheduled, which bypasses its analysis if it sees a loop annotated parallel. This allows to schedule scalar optimizations in between graphite and autopar, but we loose the polyhedral information. As we do not need the additional information, but we definitely need some scalar optimization like PRE the "annotating loops" approach seems to be our first try.