Mutations for metaheuristics

[cosunae]

This is an issue to discuss what mutations,how to perform them on the IIR and what functionality we need to apply them:

• <i,j,k> block sizes: We need to add these parameters to the IIR level
• stencil function <-> tmp precomputation replacement. We would need to configure the Pass so that it makes a replacement in each of both direction of a single (random?) tmp computation.
• Elemental EPU transformation: In a DAG of EPUs, take a single EPU inside a MS into another MS. We need a new pass that does that. And we will need to re-run some passes inside the modified MSs, like the fusion of stages, identification of caches, etc.
• we need to add EPUs label as a derived info of the StatementAccessPair level. We need to use the ms splitting and stage with maxSplit option in such a way that does not reorganize the stms but rather extracts EPUs labels. We also need to build the DAG of EPUs
• Cache mutation pass: The IIR will contain a subset of caches that is generated by the Caches Pass. We need to either remove one cache from the IIR or re-run the Caches pass and from the
• Level Fusion of Stages: Sometimes fusion all the statements possible due to dependencies will create artificially large redundant computations, that could be avoided with a more fine grained split of stages. We are not sure how we can experiment with multiple permutations here, probably it can be left for a second stage.

More???

[twicki]

roadmap.pdf

[cosunae]

I think we can archive this and close it , @twicki ?

[twicki]

agreed