[AIEX] Improve iterative scheduling convergence strategy #212
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
We are now expecting the schedulers to have improved enough to handle most regressions on their own. (Some of these improvements in next commits)
Ultimately we will have: createTopDownScoreboard checkResourceConflictsTopDown createBottomUpScoreboard checkResourceConflictsBottomUp Note that createTopDownScoreboard is now extra careful and blocks extra cycles for tiny loops. But this is still an NFC change because that scoreboard is only used to determine the number of resrouces that "stick out".
This allows to bias the depth of some SUnits in the hope of moving them up or down in the scoreboard and avoid resource conflicts without necessarily increasing the latency of the whole region.
gbossu
requested review from
abhinay-anubola,
abnikant,
andcarminati,
khallouh,
konstantinschwarz,
martien-de-jong,
SagarMaheshwari99 and
stephenneuendorffer
as code owners
| MachineInstr *ConflictMI = nullptr; | ||
| for (const MachineBundle &B : SuccBundles) { | ||
| for (MachineInstr *MI : B.getInstrs()) { | ||
| if (HR.getHazardType(Scoreboard, MI->getDesc(), HR.getMemoryBanks(MI), |
There was a problem hiding this comment.
Instead of MI->getDesc(), please use *HR.getSelectedAltDescs().getDesc(&MI)
There was a problem hiding this comment.
Oh I didn't realise that part was already merged. Will do
There was a problem hiding this comment.
And that would take away the pre condition of no MSPs?
| for (int C = FirstBlockedCycle; C <= LastBlockedCycle; ++C) { | ||
| Scoreboard[C].blockResources(); | ||
| } | ||
| } |
There was a problem hiding this comment.
I mentioned it in the commit, in the context of how this was used, this is an NFC.
martien-de-jong
previously approved these changes
Oct 15, 2024
This prepares for future work where Multi-slot pseudos won't necessarily be materialized.
martien-de-jong
approved these changes
Oct 15, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This brings a new convergence strategy that can bias the depth of some SUnits. Doing so can shift part of the schedule up or down in the scoreboard and avoid resource conflicts without necessarily increasing the latency of the whole region.
Looking at Add2D, this helps us place the VLD instructions better and avoid bank conflicts with the next iteration of the loop. We can then disable the heuristic in
WAWStickyRegistersEdgesand more aggressively remove WAW edges on status registers. This is what the first commit does.QoR is fine, there are still major regressions coming from
Neg_aie2_1andHardSigmoidTemplated_int8_0(same as already mentioned in #183 (comment)), but other regressions are avoided.And those two benchmarks are expected to be tackled in the new post-pipeliner soon.
Better review commit by commit :)