[AutoDiff] Fix over-consume when differentiating tuple instruction.
#28257
+8
−4
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Instruction visitors in `PullbackEmitter` should not consume adjoint values preemptively, becuase adjoint values are managed "globally" in a basic block by the block temporary mechanism. The crasher in TF-962 was caused by the previously unexercised logic in `PullbackEmitter::visitTupleInst`, where a `destructure_tuple` instruction is emitted with an adjoint value being its operand. This casues an over-consume. This patch fixes this by creating a copy of the adjoint value before destructuring it, and recording all destructured elements as block temporaries. TODO: The differentiation transform rarely visits a `tuple` instruction. More tests should be added, for example, cases where the tuple type's tangent type is not a tuple (`([Float], Int)`). Resolves TF-962.
dan-zheng
approved these changes
Nov 14, 2019
Contributor
dan-zheng
left a comment
dan-zheng
left a comment
There was a problem hiding this comment.
Great!
Instruction visitors in
PullbackEmittershould not consume adjoint values preemptively, becuase adjoint values are managed "globally" in a basic block by the block temporary mechanism.
I'll keep this principle in mind!
Contributor
|
@swift-ci Please test tensorflow |
Contributor
Thanks for pointing this out! |
dan-zheng
reviewed
Nov 15, 2019
bgogul
pushed a commit
that referenced
this pull request
Nov 19, 2019
…#28257) Instruction visitors in `PullbackEmitter` should not consume adjoint values preemptively, because adjoint values are managed "globally" in a basic block by the block temporary mechanism. The crasher in TF-962 was caused by the previously unexercised logic in `PullbackEmitter::visitTupleInst`, where a `destructure_tuple` instruction is emitted with an adjoint value being its operand. This causes an over-consume. This patch fixes this by creating a copy of the adjoint value before destructuring it, and recording all destructured elements as block temporaries. TODO: The differentiation transform rarely visits `tuple` instructions. More tests should be added, for example, cases where the tuple type's tangent type is not a tuple (`([Float], Int)`). Resolves TF-962.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Instruction visitors in
PullbackEmittershould not consume adjoint values preemptively, becuase adjoint values are managed "globally" in a basic block by the block temporary mechanism. The crasher in TF-962 was caused by the previously unexercised logic inPullbackEmitter::visitTupleInst, where adestructure_tupleinstruction is emitted with an adjoint value being its operand. This causes an over-consume.This patch fixes this by creating a copy of the adjoint value before destructuring it, and recording all destructured elements as block temporaries.
TODO: The differentiation transform rarely visits a
tupleinstruction. More tests should be added, for example, cases where the tuple type's tangent type is not a tuple (([Float], Int)).Resolves TF-962.