Merging GPU autoscheduler

[BachiLi]

We should at some point port the GPU autoscheduler (in standalone_autoscheduler_gpu) to the main branch. @aekul

[neoskypig]

What is the plan of this issue?

[aekul]

Hopefully it will be merged in the near future but there are no concrete plans at the moment.

[ryanstout]

@aekul Is there concern of bitrot on the standalone_autoscheduler_gpu branch? I would offer to help if I was a bit more familiar with halide internals. I would hate to see that branch not get merged in.

[TH3CHARLie]

I am trying to push this forward, here's my rough plan:

1. Submit an initial PR to merge gpu autoscheduler code from the branch into src/autoschedulers/anderson2021. This PR aims to fix incompatibility between the old code base and master. Build pass and minimal tests are what we need at this moment. (1 giant PR)
2. Since adams2019 and anderson2021 shares some code(e.g. Weights.cpp), it makes sense to me that we extract the common parts and make them reusable across schedulers. We need to make sure this will not damage the existing plugin system and will not break existing tests (if any). (1 or several refactoring PRs)
3. Adding tests, CMake, documentations to anderson2021 (several PRs)

cc @BachiLi @abadams @aekul

[steven-johnson]

Is this still active?

[ryanstout]

@steven-johnson I would love to see it land at some point, not holding my breath at this point though :-)

[steven-johnson]

@steven-johnson I would love to see it land at some point, not holding my breath at this point though :-)

What are the obstacles to moving forward with it, as far as you know?

[ryanstout]

@steven-johnson I think it's just really stale at this point. I wasn't involved in it though, so thats an outsider's look at it. I'm sure there's a lot of details to get it brought up before it could be merged. Thanks!

[TH3CHARLie]

yeah the biggest obstacle is that Luke or myself could not find time to push this forward before, I will meet with Andrew and Derek later today and I will bring this up to them and see if we can have a swift plan to merge this.

[aekul]

I haven't had a lot of time to work on this but have been preparing a basic PR lately that should be ready sometime this week. It doesn't attempt to merge/de-duplicate the code that's shared with adams2019 so if that is a blocker, then more time/work will be required.

[ashishUthama]

Been meaning to try this out too but was unable to hack out something from the branch to use. I would welcome a quick "v.1" version as an end user.

[antonysigma]

Hi I noticed the PR #6856 has been merged around 4 days ago. What is the release schedule? v16.0 ?

Asking because I am compiling the Halide binary from source, which I found is not scalable in a long run.

[abadams]

Yes, this should go out in Halide 16