[MetaSchedule] Fix thread bindings of MultiLevelTilingTensorCore #13243
Conversation
|
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.
Generated by tvm-bot |
|
BTW, i was curious why it's not caught by the unittests of search space generation? |
|
the unit test doesn't use default config here |
|
we might want to add a unittest to cover the rule, but i understand it's kind of urgent. please set a reminder to add a regression test :-) |
|
@tvm-bot rerun |
| if (tile_binds.defined()) { | ||
| for (const String& tile_bind : tile_binds.value()) { | ||
| CHECK_NE(tile_bind, "threadIdx.x") << "Cannot bind to threadIdx.x when using tensor core."; | ||
| } | ||
| } |
There was a problem hiding this comment.
Does this check (along with existing tests) now cover the regression this fixes? If not adding a test to protect against the regression in the future will surely save you and others time!
There was a problem hiding this comment.
Oops looks like @junrushao already asked my question, I trust your judgements :)
There was a problem hiding this comment.
Yes this check covered the broken case that incorrectly using threadIdx.x. As a follow up, I'm thinking also adding some e2e search space generation test for the default config for each target.
vinx13
force-pushed
the
fix/ms-tensorcore
branch
from
6c07b5f to
4d12929
Compare
7 participants
Fixed #13204.
cc @junrushao