[AutoTVM] Added Droplet algorithm in TVM

[canesche]

This PR adds the Droplet Search algorithm to AutoTVM. Droplet search
is a variation of the Coordinate Descent Algorithm. It relies on the
empirical observation that the search space formed by typical kernels
contains, within the same convex region, origin (the kernel without
optimizations) and the best kernel configuration.

Droplet search is deployed in the same setting as any other search
algorithm that is implemented in AutoTVM. Droplet search is expected
to run faster than gridSearch, randomSearch, GA, and XGBoost. It tends
to find kernels that are as good as those found by these other
techniques.

A detailed account of the algorithm can be found in this paper:

https://homepages.dcc.ufmg.br/~fernando/publications/papers/DropletSearch.pdf

[tvm-bot]

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.

• cc @elvin-n, @Icemist _{See #10317 for details}

_{Generated by tvm-bot}

[tqchen]

Thank you for your contribution, more recently we are moving most of our automation effort toward MetaSchedule, which helps to unify different approaches and can also work as AutoTVM tuning and works for TensorIR.

As a result, it would be great to think about how to make it work for metaschedule. In the meantime, the interface is still compatible so such tuner can be plugged into the existing infra

[leandron]

Hey @canesche, thanks for the PR.

Just noticed this is blocked on an error related to the commit messages guideline (https://tvm.apache.org/docs/contribute/pull_request.html#commit-message-guideline), specifically on the commit titles ending with ., so when you fix that, then we'll have a full CI round and will be able to review it properly.

It would be also good to have some unit testing added in this PR, so that we validate and prevent it breaking in future.

[Icemist]

Please make sure the different types of indexing work correctly (including multi_filter).
For this make tests similar to test_autotvm_ga_tuner.py, test_autotvm_index_tuner.py, test_autotvm_space.py from tests/python/unittest/ folder.
👍🏻

[leandron]

LGTM, thanks @canesche @Icemist. Happy for this PR to be merged once CI is green.

[canesche]

Hi @leandron, thanks! I hope to contribute more often to TVM in the future.

[leandron]

This is now merged. Thanks @canesche @Icemist @tqchen!