[TE] Raise error for non-bijective transformation #12926
Merged
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
This is a fix for a bug introduced in apache#12904. Prior to then, an exception was raised when the transformation wouldn't be bijective over the transformed buffer's shape. The PR replaced the bijective check done as part of `DetectIterMap` with a check done on the returned `padding_predicate`. However, this check was not equivalent, and some transformations could erroneously apply, rather than raising an exception as being non-bijective. This commit re-enables the bijectivity check in `DetectIterMap`, and adds a test case for this behavior.
echuraev
reviewed
Sep 28, 2022
Comment on lines
+578
to
+590
| def test_non_divisible_transform_raises_error(): | ||
| A = te.placeholder([1, 3, 8, 8]) | ||
| B = te.compute(A.shape, lambda *indices: A[indices]) | ||
| s = te.create_schedule(B.op) | ||
|
|
||
| transform = lambda n, c, h, w: [n, c // 4, h, w, c % 4] | ||
| # Error occurs here, because the transformation would introduce | ||
| # padding. Padded transforms are supported in TIR-based | ||
| # schedules. | ||
| with pytest.raises(tvm.TVMError): | ||
| s[B].transform_layout(transform) | ||
|
|
||
|
|
Contributor
There was a problem hiding this comment.
nit: Probably we can joint this test case with test_size_one_buffer? Just extend its testing parameters
Contributor
Author
There was a problem hiding this comment.
It could, but I think it would make the test less readable as an example use case, specifically what behavior is being tested, because the desired behavior differs in each case. It would look something like below, but there's nothing to call attention to the fact that is_valid changes the expected behavior, and isn't just a parameter being used in the setup.
shape, transform, is_valid = tvm.testing.parameters(
([1, 8], lambda n, i: [i, n], True),
([1, 1, 8], lambda i, j, k: [j, te.AXIS_SEPARATOR, i, k], True),
([1, 1, 8], lambda i, j, k: [i, te.AXIS_SEPARATOR, j, k], True),
([1, 3, 8, 8], lambda i, j, k: [i, te.AXIS_SEPARATOR, j, k], False),
)
def test_transform_validity(shape, transform, is_valid):
dtype = "int8"
A = te.placeholder(shape, dtype, name="A")
B = te.compute(
shape=A.shape,
fcompute=lambda *indices: A[indices].astype(dtype),
name="B",
)
s = te.create_schedule(B.op)
if is_valid:
s[B].transform_layout(transform)
else:
with pytest.raises(tvm.TVMError):
s[B].transform_layout(transform)
Contributor
There was a problem hiding this comment.
Got it. Thank you. I agree.
vinx13
approved these changes
Oct 14, 2022
xinetzone
pushed a commit
to daobook/tvm
that referenced
this pull request
Nov 10, 2022
This is a fix for a bug introduced in apache#12904. Prior to then, an exception was raised when the transformation wouldn't be bijective over the transformed buffer's shape. The PR replaced the bijective check done as part of `DetectIterMap` with a check done on the returned `padding_predicate`. However, this check was not equivalent, and some transformations could erroneously apply, rather than raising an exception as being non-bijective. This commit re-enables the bijectivity check in `DetectIterMap`, and adds a test case for this behavior.
xinetzone
pushed a commit
to daobook/tvm
that referenced
this pull request
Nov 25, 2022
This is a fix for a bug introduced in apache#12904. Prior to then, an exception was raised when the transformation wouldn't be bijective over the transformed buffer's shape. The PR replaced the bijective check done as part of `DetectIterMap` with a check done on the returned `padding_predicate`. However, this check was not equivalent, and some transformations could erroneously apply, rather than raising an exception as being non-bijective. This commit re-enables the bijectivity check in `DetectIterMap`, and adds a test case for this behavior.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
This is a fix for a bug introduced in #12904. Prior to then, an exception was raised when the transformation wouldn't be bijective over the transformed buffer's shape. The PR replaced the bijective check done as part of
DetectIterMapwith a check done on the returnedpadding_predicate. However, this check was not equivalent, and some transformations could erroneously apply, rather than raising an exception as being non-bijective.This commit re-enables the bijectivity check in
DetectIterMap, and adds a test case for this behavior.