[microTVM] tuning on micro targets with meta-schedule by mkatanbaf · Pull Request #13514 · apache/tvm

mkatanbaf · 2022-11-29T19:56:03Z

adds support for tuning microTVM models using meta-schedule.

Summary of the changes:

adds "c" to the targets supported by meta-schedule
implements a builder and runner for micro devices
runs a simple tuning job for verification

tvm-bot · 2022-11-29T19:56:07Z

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

Built docs for commit 21b25d5 can be found here.
cc @alanmacd, @gromero, @leandron, @mehrdadh _{See #10317 for details}

_{Generated by tvm-bot}

mkatanbaf · 2022-12-07T18:40:58Z

cc @areusch @guberti @zxybazh

zxybazh

Thanks for send the PR for micro targets MS tuning! Generally looks ok to me except a few spots needs clarification and minor changes. Would be great if someone else could take a look at the relay executor and tvm rpc changes.

guberti

I'm not the right person to do an in-depth review here, but I'm really excited for these changes. Looking at this was helpful for understanding meta-schedule - I left a few nits too.

mkatanbaf · 2022-12-10T01:06:13Z

@tvm-bot rerun

mehrdadh · 2022-12-12T17:13:45Z

@mkatanbaf you might need to rebase with main.

areusch

thanks @mkatanbaf , added a few more comments/questions

areusch · 2023-01-03T17:30:02Z

+        mod = IRModule({"main": prim_func})
+        runtime = Runtime("crt", {"system-lib": True})
+        mod = RemoveWeightLayoutRewriteBlock(skip_ndarray_rewrite=True)(mod)
+        rt_mod = tvm_build(mod, target=target, runtime=runtime)


just to confirm--are the relay.build changes needed in this PR? if not, could we remove them until we figure out how to wrap a TIR function in a Relay function?

No, they are not. We continue to use tvm_build in tuning. The changes in relay_integration.py are needed for compiling the relay program using the MetaSchedule tuning database.

mkatanbaf · 2023-01-09T19:03:25Z

@areusch @mehrdadh @zxybazh Could you please take another look?

mehrdadh · 2023-01-09T19:56:41Z

LGTM, I'll wait for the others and then merge it.
cc @guberti @zxybazh

guberti · 2023-01-09T20:44:52Z

LGTM, I think we can merge once tests pass.

zxybazh

Generally looking good, just one question about the schedule rule set up.

zxybazh · 2023-01-09T23:24:54Z


+Array<ScheduleRule> ScheduleRule::DefaultMicro() {
+  return {
+      ScheduleRule::AutoInline(


Is there any specific reason that ApplyCustomRule is not included as a rule here?

No, there was not. I copied the rules from the Hexagon rules a while ago, and the ApplyCustomRule was not added then. I added the rule.

Hi, after you add the ApplyCustomRule you may also consider to add other rules in the current Hexagon rules like InlineConstantScalars.

@zxybazh I think this could be in a follow up PR since the PR has been open for a while. thanks!

Thanks @zxybazh, I already added the rest of Hexagon rules except the ParallelizeVectorizeUnroll which is not needed on the micro targets.

mehrdadh · 2023-01-10T01:11:10Z

@tvm-bot rerun

adds support for tuning microTVM models using meta-schedule. Summary of the changes: adds "c" to the targets supported by meta-schedule implements a builder and runner for micro devices runs a simple tuning job for verification Co-authored-by: Mohamad <mkatanbaf@users.noreply.github.com>

mkatanbaf force-pushed the micro-with-meta-schedule branch 3 times, most recently from 341cecf to bc41156 Compare December 7, 2022 18:15

mkatanbaf marked this pull request as ready for review December 7, 2022 18:40

mkatanbaf changed the title ~~[microTVM] micro tuning with meta-schedule~~ [microTVM] tuning on micro targets with meta-schedule Dec 7, 2022

zxybazh reviewed Dec 8, 2022

View reviewed changes

Comment thread python/tvm/meta_schedule/relay_integration.py

Comment thread python/tvm/meta_schedule/relay_integration.py

Comment thread python/tvm/contrib/micro/meta_schedule/test_autotune_ms.py Outdated

Comment thread python/tvm/contrib/micro/meta_schedule/rpc_runner_micro.py Outdated

masahi reviewed Dec 8, 2022

View reviewed changes

Comment thread src/meta_schedule/schedule_rule/schedule_rule.cc Outdated

guberti reviewed Dec 9, 2022

View reviewed changes

mkatanbaf force-pushed the micro-with-meta-schedule branch from c0396fa to fc04678 Compare December 12, 2022 18:36

areusch reviewed Jan 3, 2023

View reviewed changes

Comment thread src/meta_schedule/schedule_rule/schedule_rule.cc Outdated

mkatanbaf force-pushed the micro-with-meta-schedule branch from bf80e6d to afda2c1 Compare January 3, 2023 23:49

mehrdadh requested changes Jan 4, 2023

View reviewed changes

Comment thread python/tvm/contrib/micro/meta_schedule/test_autotune_ms.py Outdated

Comment thread python/tvm/contrib/micro/meta_schedule/test_autotune_ms.py

mehrdadh reviewed Jan 9, 2023

View reviewed changes

Comment thread tests/python/unittest/test_micro_ms_tuning.py Outdated

mehrdadh requested review from guberti and zxybazh January 9, 2023 19:56

guberti approved these changes Jan 9, 2023

View reviewed changes

mkatanbaf added 5 commits January 9, 2023 13:29

added support for tuning microTVM models using meta_schedule.

a5e8c1f

addressing comments

2acf767

addressing comments

94b9a58

set stack and heap size for the test

eef2c3e

nit

0c09eb7

mkatanbaf force-pushed the micro-with-meta-schedule branch from 8322e09 to 0c09eb7 Compare January 9, 2023 21:30

zxybazh reviewed Jan 9, 2023

View reviewed changes

updating schedule rules

908b679

fixing the test on fvp

9b224e7

mehrdadh approved these changes Jan 11, 2023

View reviewed changes

mehrdadh merged commit c2bc1ec into apache:main Jan 11, 2023

ysh329 mentioned this pull request Apr 17, 2023

[Release] v0.12.0 Release Candidate Notes #14645

Closed

Conversation

mkatanbaf commented Nov 29, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

tvm-bot commented Nov 29, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mkatanbaf commented Dec 7, 2022

Uh oh!

zxybazh left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

guberti left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

mkatanbaf commented Dec 10, 2022

Uh oh!

mehrdadh commented Dec 12, 2022

Uh oh!

areusch left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

areusch Jan 3, 2023

Choose a reason for hiding this comment

Uh oh!

mkatanbaf Jan 3, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

mkatanbaf commented Jan 9, 2023

Uh oh!

Uh oh!

mehrdadh commented Jan 9, 2023

Uh oh!

guberti commented Jan 9, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

zxybazh left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zxybazh Jan 9, 2023

Choose a reason for hiding this comment

Uh oh!

mkatanbaf Jan 9, 2023

Choose a reason for hiding this comment

Uh oh!

zxybazh Jan 10, 2023

Choose a reason for hiding this comment

Uh oh!

mehrdadh Jan 11, 2023

Choose a reason for hiding this comment

Uh oh!

mkatanbaf Jan 11, 2023

Choose a reason for hiding this comment

Uh oh!

mehrdadh commented Jan 10, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

mkatanbaf commented Nov 29, 2022 •

edited

Loading

tvm-bot commented Nov 29, 2022 •

edited

Loading

mkatanbaf Jan 3, 2023 •

edited

Loading

guberti commented Jan 9, 2023 •

edited

Loading

zxybazh left a comment •

edited

Loading