daisytuner / docc / 30614148740

65%

push

github

Adding ProducerIntoConsumer MapFusion into the new, recursive MapFusionPass (#873)

Refactored the code of MapFusion Transform into MapFusionByAccessWorker.

The old transform still exists for its use as standalone transformation for now and also to run the old pass to cover everything not yet implemented in the new ByAccessWorker.

Refactored MapFusionByDomain into LoopFusionPass, the sequential & recursive loop-fusion logic into a base that handles the existing by-domain fusion and the by-access fusion, separately disablable.

For this, the existing per-loop caches of LoopFusionPass have been extended to contain the data needed by by-access fusion (assumptions, non-map loops and more details about subsets and locations of accesses).

LoopFusionPass still maintains these caches as fusions happen, so it never needs to clear any analysis.

For now, the MapFusionByAccessWorker has not been verified and skips any ConsumerIntoProducer fusions, as that is a greater change to how the cached data needs to be updated and its also a smaller portion of compile-time.

The MapFusionByAccessWorker was refactored to use the FusionLoopCandidate structure to access any cached data. It can be used as a basis for a new Transformation, that simply creates these per-loop structures on the fly, so that Transform and efficient pass can share code again.
The ByAccessWorker was also optimized to support per "fusion container" selection of replacing memory accesses with local transients and to consider the domain-equality checks already done by the NewMapFusionPass.

If the domains match, we can leave in the indirect writes in, which makes the fused code is a valid substitute for the original loop, avoiding the need for any loop-duplication.

Elliding same-subset indirect write-read patterns in the same block or successive blocks is a far easier problem to solve, then leaving the original loop lying around and hoping that it can be proven completely dead and removed by later ... (continued)

1388 of 1894 new or added lines in 13 files covered. (73.28%)

45088 of 69697 relevant lines covered (64.69%)

725.27 hits per line