Middle-End Arrays

Streamization with graphite

Are we going to generate the parallel codes directly? No. Generating call-backs to the vectorizer and parallelizer from polyhedra to Gimple do we need to have a separate pass for vectorizer after PCP to get the vectorize info form graphite?

IPA optimizations and graphite (by Honza)


memory usage statistics polylib vs ppl.

Next Steps

  1. get identity transformations working. (4.4) - basic block copying should work to get this.(update_ssa)
  2. stabilize code -make loop interchange and blocking working (4.4)
  3. get rid of linear lambda (4.5)
  4. representing the scalar as zero dimensional arrays, reductions (4.5)
  5. start PCP by writing some functions that are fundamentally modular, graphite.h basically having an hypothesis that it does not depend on tree-ssa. (4.5)
  6. distributing cloog, cleanup the code for cloog-ppl , clean the polylib matrices from gcc(4.4)
  7. creating tarballs for cloog and ppl (4.4)
  8. need to build a committee around cloog.
  9. dependence analysis (4.5) - proposing ppl to use array dependence analysis. need to agree on the dependence graph (same as data dependence graph). - Konrad. (Sebastian will talk to him) But this will bind cloog to ppl - so use cloogs's abstraction. need to extend cloog to do this. This can be avoided by just using ppl directly for dependence analysis.
  10. extend ppl to implement parametric solvers. (right now piplib does that) (?)- Roberto how important is this. (4.5)
  11. implement region based scope detection (4.5)
  12. privatization of scalars. (4.5)
  13. cost model - heuristics for interchange - Harsha (4.5)
  14. implementation of PCP - Jan Sjodin. (4.5)
  15. modify cloog to consider only one type of constraint (4.5)
  16. loop distribution (4.5) -Sebastian
  17. loop fusing (4.5) - Dwarak - loop bounds need to identical, data dependence, same context but little by different domains
  18. unroll is a cloog issues, rerolling- Cedric
  19. loop composition can be done after loop transformations are done.

None: Graphite_Workshop_Nov08_Dwarak (last edited 2009-01-18 08:45:07 by cpe-70-116-74-159)