Memaccess iscached

This shows how to implement iscached (needed only for algorithms with WAR dependencies) on top of memaccess

Ok let's wait for !27 (merged) to be merged before reviewing this one.

Sure, it's just a way for me to get work done ahead of the review :)

added 4 commits

• cb1f95b1...7d740d57 - 2 commits from branch solverstack:master
• f2771553 - Allow runtimes to have their own memory access macros
• f70278b0 - Reimplement taking iscached into account on top of memaccess

Compare with previous version

I have rebased on top of !31 (merged)

added 2 commits

• 0adc69b4 - Add begin/end memory access hooks for runtimes
• 06a83a3b - Merge branch 'memaccess-runtime' into memaccess-iscached

Compare with previous version

added 3 commits

• 46082f53 - typo
• 15527d6b - Add execution rank parameter to MORSE_RANK_CHANGED
• f1e0480d - Merge branch 'memaccess-runtime' into memaccess-iscached

Compare with previous version

added 3 commits

• fb248312 - Also add RANK_CHANGED callback for runtimes
• 1c047268 - Add do/while to fix usage in if blocks
• ff80eb94 - Merge branch 'memaccess-runtime' into memaccess-iscached

Compare with previous version

added 10 commits

• ff80eb94...bed75474 - 3 commits from branch solverstack:master
• 6745049c - Allow runtimes to have their own memory access macros
• afa472de - Add begin/end memory access hooks for runtimes
• 9cfedae6 - typo
• e5d8288c - Add execution rank parameter to MORSE_RANK_CHANGED
• d1f7a467 - Also add RANK_CHANGED callback for runtimes
• afbd9e49 - Add do/while to fix usage in if blocks
• 72b58f4c - Reimplement taking iscached into account on top of memaccess

Compare with previous version

We have noticed this afternoon that algorithms with pivoting typically have RAW dependencies, dropping the pruning did fix execution.

Really ? In which case ? You have an example, just to know, because I can't see when.

@tcojean ?

It's a WAW on the pivot vector handle in the tile LU with partial pivoting. I rose my thumb (again).

well, it's actually RWAW :) But the point is that while we do have RWAW in algorithms already (e.g. the gemms in cholesky), it's always on the same node. In the pivot vector handle case, the handle moves between nodes, and thus nodes needs to be notified of the invalidation.

Yeah but we also have this in QR with the upper tile that is forwarded through all nodes then. Ok, I'm not raising my thunb yet, let's discuss about it on tuesday, I think I got it and I agree with it, but just want to be sure first :)

added 4 commits

• 72b58f4c...a2680333 - 3 commits from branch solverstack:master
• 9b0f73a9 - Reimplement taking iscached into account on top of memaccess

Compare with previous version

added 1 commit

• f7bf5d35 - Remove MORSE_WAR_DEPENDENCIES

Compare with previous version

resolved all discussions

enabled an automatic merge when the pipeline for f7bf5d35 succeeds