daisytuner / docc / 24102763944

push

github

Separate can_be_applied and apply for GPUTilling during Loop Scheduling (#632)

* restructure stride minimization pass

* remove iterative bottleneck and add/fix tests

* loop distribute rework

* Loop distributionfix and unit tests

* Refactor loop scheduler to split the can be applied from the apply to save quadratic analysis recalculation cost.

* enforce can be applied for polly scheduler

* add branching

* guard rpc scheduler

* remove redundant analysis

* remove multi invalidation

* add resnet subset tests

* remove redundant invalidates

* light weight invalidation of scope analysis

* revert stride minimization and perfect loop distribution to main

* update tenstorrent scheduler

* Drop submodels from PR

* extend candidate selection to skip impossible candidates fast

* Stream line kls criterions to reuse them for the scheduler.
Saperating can be applied and apply calls

* remove polly header

* Use new RTX label

53 of 67 new or added lines in 2 files covered. (79.1%)

28868 of 44540 relevant lines covered (64.81%)

605.56 hits per line