Graphite phone call from 2009_04_08
Attendees: Christophe, Tobias, Sebastian, Konrad, Harsha, David, Li, Albert
Status:
- Reductions: Reductions switched out of SSA. Per phi node introduce an array of dimension zero. Everything is just working.
- Benefits:
- No need to handle all the phi nodes.
- Problems with cloog to hoisting out conditions
- Was way faster to implement.
- Try to go back to SSA, by reintroducing phi nodes.
- No idea. Ask Diego Novillo and Kenny.
- Richi was against going out of SSA. But going all out of SSA.
- Benefits:
- Patch to handle NE_EXPR and EQ_EXPR. Bootstrapped and committed.
- Necessary for fortran.
- Automatic parallelization - Li started:
- Added flag -fgraphite-force-parallel
Got CompileFarm account
- Started to get git working
- Git repositories on the compile farm.
- Automatic test for git repositories.
- Data dependencies:
- Konrad wants to write pdf file how the polyhedron look like.
- As we want handle general scattering functions and access
- polyhedrons we need a more general data dependency analysis.
- Should be done as soon as possible.
- Konrad wants to get it done before Sebastian comes to INRIA
- Switch from polyhedron to union of polyhedron in pbb domain.
- Necessary for NE_EXPR and EQ_EXPR.
- Switch to union of polyhedron for access functions.
- Union of access functions and domains are independent. They do
- not need to have the same number of disjoint polyhedrons.
- Agreed.
- Union of access functions and domains are independent. They do
- Razya in France
- Sebastian at INRIA from 20/04/2009.
- Jan starts on working on a full working implementation on PCP
- Jan spends the next couple of month dedicated to PCP
- Harsha
- Starts to work with Jan to get started with PCP
- Wants to work on Interchange heuristics, fusion/fission
- Vectorizer
- Code is not handled by Vectorizer
- Loops where not in the form that can be handled by the vectoriser.
- Hot loops information is not yet updated (edges and frequencies)