[REFACTOR][BOYC] Non recursive partitioning #5493
Conversation
|
@mbrookhart please take a look for the mixed mutator pattern. BTW, we will still need to refactor the infertype pass as it is the most frequently used pass. @comaniac @masahi @mbaret @manupa-arm @trevor-m please take a look. |
There was a problem hiding this comment.
We were trying rewriter but it is not applicable to this pass due to https://github.com/apache/incubator-tvm/pull/5493/files#diff-8d2cdf6314f73e4b32892679ad4dc44aR280, which traverses other nodes out of order. As a result, the current solution seems the most suitable.
There was a problem hiding this comment.
Use of the iterative traversals looks great. Since you originally authored the class, I'll believe you about the necessary mutators.
Why so much auto-formatting noise? As a larger conversation, we might want to build a style checker into CI that enforces a particular auto-format implementaiton...
|
Ah I think that's because I manually ran clang-format for the file. We should definitely build style checker in CI. |
| bool found_start_{false}; | ||
| bool found_end_{false}; | ||
| /*! \brief Map from each region output expr node to its output index and TupleGetItem node. */ | ||
| std::unordered_map<Expr, std::pair<int, TupleGetItem>, ObjectHash, ObjectEqual> out_expr_indices; |
There was a problem hiding this comment.
Since both arguments of TupleGetItem (func_call and the index) are in this struct, this TupleGetItem seems redundant to me. Also see my comment at L282.
There was a problem hiding this comment.
If this TupleGetItem is intended to be cached, please come up with a better name than out_expr_indices, since it is not just indices.
There was a problem hiding this comment.
Rename to region_func_out and only cache the function output expressions (Call or TupleGetItem).
comaniac
force-pushed
the
non-recursive-partition
branch
from
3a6f409 to
00be9a9
Compare
comaniac
force-pushed
the
non-recursive-partition
branch
from
00be9a9 to
26404c1
Compare
|
@masahi thanks for valuable suggestions. We've refactored |
|
Thanks @zhiics @comaniac @mbrookhart @manupa-arm |
* non recursive partitioning * refactor maps * rebase upstream * refactor shared output * address comments Co-authored-by: Cody Yu <[email protected]>
* non recursive partitioning * refactor maps * rebase upstream * refactor shared output * address comments Co-authored-by: Cody Yu <[email protected]>
* non recursive partitioning * refactor maps * rebase upstream * refactor shared output * address comments Co-authored-by: Cody Yu <[email protected]>
* non recursive partitioning * refactor maps * rebase upstream * refactor shared output * address comments Co-authored-by: Cody Yu <[email protected]>
This PR refactors the partitioning pass by using non-recursive mutator. It also removes the unnecessary mutators as we only need to look at begin/end annotations which are definitely wrapped in call nodes. In addition, a metadata struct is used to maintain the intermediate data needed for partitioning.