[MetaSchedule] Performance Alignment - NRM and SFM (CUDA) #559
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
|
@zxybazh Hi Xiyou, could you help go through the changed file list to check whether some changed file has been upstreamed? If so, we need to send PRs to mainline to update these files in the short future. Thanks a lot! |
|
Of course. Most of the changes were introduced in #9761, #9780, #9789(not merged yet) and #9799(not merged yet). And in these PRs I followed the file list here and upstreamed all files related to my part, i.e., Mutators, Schedule Rules and Postprocs are included but only header file and the main implementation, no concrete classes included. You may refer to this commit as the latest one on my side. Let me know if I missed anything, thanks! |
|
@zxybazh Got it, thanks! I went through your PRs and found no conflicts 😉 . As your parts haven't introduce concrete classes so far, there should be no problem 😆 . |
spectrometerHBH
added a commit
that referenced
this pull request
Dec 30, 2021
[Meta Schedule][M3c] Schedule Rules, Mutator & Postprocs (#485) [Meta Schedule][M3c] PostOrderApply (#486) Fix Post Order Apply (#490) [MetaSchedule] Relay Integration (#489) [M3c][Meta Schedule] Add Trace Correctness Test for PostOrderApply (#492) Fix replay trace. (#493) [M3c][Meta Schedule] Implement the Replay Func class. (#495) [PR] Test script for meta-schedule task extraction. Interface to load… (#494) [Meta Schedule Refactor] Get child blocks (#500) Read-at && Write-at (#497) [M3c][Meta Schedule] Measure Callbacks (#498) [Bug] Fix Infinite Loop Caused When Calling Methods Not Overrided In PyClass (#496) [MetaSchedule] Sample-Perfect-Tile (#501) [MetaSchedule] TE Workloads (#502) [TensorIR] GetProducer, GetConsumer (#506) [MetaScheduleRefactor] Annotate&Unannotate (#505) [MetaSchedule] Multi-Level-Tiling & Auto-Inline (#503) [Tests] Add unittests for auto-inline and multi-level-tiling (#508) [Meta Schedule] Minor Fixes (#507) [MetaSchedule] Rewrite Cooperative-Fetching / Unbound-Block / Reduction-Block (#509) [MetaSchedule] Rewrite Parallel-Vectorize-Unroll / Verify-GPU / Disallow-Dynamic-Loops (#499) [Meta Schedule] Add Helper Function & Minor Modification (#512) [MetaSchedule] Test for Rewrite Parallel-Vectorize-Unroll (#513) [Meta Schedule] Feature Extractor & Cost Model (#510) Blockize & Tensorize (#514) Layout Rewriting: Suggest-Index-Map (#520) [MetaSchedule] Parallel-Vectorize-Unroll & Random-Compute-Location (#516) [Meta Schedule] Per-Store-Feature (#521) Add traced schedule for blockize & tensorize (#526) [Meta Schedule] Add XGBoost Model & Random Model (#519) User-Interface: Tune-TIR (#525) User-Interface: Tune-TE (#527) [Minor] More logging on python (#528) Get CUDA tuning working (#529) [MetaSchedule] TensorRT BYOC (#518) [BugFix] LocalBuilder API (#531) [Meta Schedule] Add Cost Model Update Measure Callback (#530) [Bugfix] BuilderInput with default params (#532) [MetaSchedule] Mutator-Tile-Size, Mutate-Parallel, Mutate-Unroll (#534) [Meta Schedule] Evolutionary Search (#522) [BugFix] Remove duplicated definition of MakeMultinomialSampler (#535) [Meta Schedule] Fix some bugs (#537) Initiate Experiments for CPU Performance Alignment with Ansor (#538) [Meta Schedule] Tweak experiment scripts (#539) [Meta Schedule] Initiate experiments on CUDA (#540) [TIR][Schedule] Buffer transform (#523) Auto Tensor Core (#524) Working on Evo Search (#542) [Meta Schedule] Add Replay Tuning Interface (#543) Evolutionary Search on CPU (#544) Misc improvement over the error message (#545) [TIR][Schedule] Software pipelining (#533) [Meta Schedule Refactor] fixing unit tests (#547) [MetaSchedule] Mutator-Compute-Location (#548) Misc Improvement of Evolutionary Search (#549) Hotfix for software pipeline (#552) Misc Improvement (#550) [Cherry-Pick][TensorIR] Primitive "SetScope" (#9738) (#555) Rule RFactor (#551) [MemHammer] Rewrite Rules (#554) [MetaSchedule] Schedule Rule: Cross-Thread Reduction (#556) [MetaSchedule] Performance Alignment - NRM and SFM (CUDA) (#559) [MetaSchedule] Perf Alignment - NRM on CUDA (#560) [TIR] Reorder the block iters of the blocks generated by RFactor (#561) Co-authored-by: Siyuan Feng <Hzfengsy@sjtu.edu.cn> Co-authored-by: Bohan Hou <32121147+spectrometerHBH@users.noreply.github.com> Co-authored-by: Hongyi Jin <3231950289@qq.com> Co-authored-by: Ruihang Lai <lairuihangdongdong@qq.com> Co-authored-by: Junru Shao <junrushao1994@gmail.com> Co-authored-by: Wuwei Lin <wuwei@apache.org> Co-authored-by: Sunghyun Park <49998730+sunggg@users.noreply.github.com> Co-authored-by: Xiyou Zhou <xiyou@octoml.ai>
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.
The performance of batch normalization and softmax on CUDA is almost aligned:
To be clear, the gap isn’t caused by rule cross-thread reduction, for sure. The detailed reason is unknown yet, partly because I cannot reproduce Ansor’s history best GFLOPS number except when tuning with Ansor. Thus let's dig out the issues later in the future.