Attendees: Christophe, David, Konrad, Li, Ramakrishna, Razya, Sebastian, Tobias
- Phone call recording [2:55] (Tobi)
- We will record all phone calls and put them on the wiki.
- No concerns from anybody.
- Moving to CLooG trunk [4:10] (Tobias)
- Move cloog-ppl to the interface in cloog trunk (Patchset available, will be sent to the mailinglist)
- Move graphite to new interface (Patchset available)
- Problem: Interface in cloog trunk is not finished. It is not possible to use it as a library. Some patches necessary to cloog trunk. Plan to work with Sven
- Hacked to make graphite run with cloog polylib. Just one testcase failure in Graphite test cases.
- Tobias became a Graphite reviewer [6:30]
Graphite in 4.5 status page (Tobias)
- Autopar [7:45]
- Basic autopar support for single loops committed (Li)
- Test coverage is too small, work on test cases with conditions and triangular loops (Li)
- outer loop parallelization: . [9:15] (Razya+Li)
- Autopar test cases for reductions fail. (Razya)
- Since splitting autopar
- Razya will test/fix autopar itself and the interaction with graphite.
- loop-interchange/strip-mining [14:25] (Sebastian)
- loop-interchange
- still broken, because of profitablity analysis
- Problem Computation of array sizes (local patch) [16:50]
- We need to observing data reference, but these are never updated by transformations. To analyze poly_dr we have to look into it using the TRANSFORMED SCATTERING.
- strip-mining [14:50]
- enabled in -O2
- passes all SPEC benchmarks, test cases failing in vectorizer (All SLP test cases fail)
- strip mine by a factor of 64. Pranav will work on improvements proposed by David
- loop-interchange
- loop-blocking [21:15]
- Will be implemented by scheduling interchange and strip mining after each other.
- SPeC [22:10] (Sebastian)
- One failure in dealII: we generate a loop, that generate MAX-1 times.
- tried to implement code of upper bound as signed int. (Alberts suggestion)
- next try: add type constrainst on the parameters to make cloog not generate this expression. (We should try with cloog binary). Relations in between parameters will not be added.
- One failure in dealII: we generate a loop, that generate MAX-1 times.
- ivtypes [25:50] (Ramakrishna)
- looking at the casts treatment in vectorizer
- How types work (Sebastian)
- Simple testcases for the types.
- dependency analysis (Konrad) [31:30]
- removing some bugs in data deps
Still two failing bugs in the bootstrap (Problems to reproduce it) Logs
- One problem exposed by interchange in h264 of SPEC.
- Idea of changing the format of poly_drs [38:10]
- Back to old unique representation as ppl maximize() works.
- However problem to get the strides out of array accesses using the PPL helper functions. Sebastian will send a mail explaining the problem.
- Cloog Releases [40:40]
- Tested with C++ infrastucture (David). CLooG headers are not save.
- Need to release a new CLooG ppl release, git is only useful for development
- Contact release managers [42:35] (David)
- start discussing with the release managers about the merge of the graphite branch on trunk
- Point them to Graphite/4.5 not sufficient.
- Need to be proactive
- Microbenchmarks [43:15] (David)
- testsuite for polyhedral transforms
Mentioned here
- Shows good results for IBM team
- No benchmark collection, but available one by one
- We should add them to our tests
- IBM ideas [49:30] (David)
Also mentioned here
- transformations that other implementations plan to implement: conservative transforms on more complex code: symbolic indices, black boxes.
- Unknown access functions might be added easily, however it will be harder to optimize them. The others will be more difficult as they require model changes. (Tobi)
- End [53:30]