[Arith] Support integer BufferLoad in IntervalSetEvaluator #10327
Conversation
This file contains 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
MasterJH5574
requested review from
tqchen,
junrushao,
vinx13,
comaniac,
jroesch,
areusch,
yzhliu,
merrymercy and
icemelon
as code owners
Hzfengsy
requested changes
Feb 20, 2022
tests/python/unittest/test_tir_transform_compact_buffer_region.py
Outdated
Show resolved
Hide resolved
| T.reads( | ||
| A_indptr[i], | ||
| A_indptr[i + 1], | ||
| A_data[A_indptr[i] + 0 : A_indptr[i] + (A_indptr[i + 1] - A_indptr[i])], |
There was a problem hiding this comment.
Suggested change
| A_data[A_indptr[i] + 0 : A_indptr[i] + (A_indptr[i + 1] - A_indptr[i])], | |
| A_data[A_indptr[i] : A_indptr[i + 1]], |
There was a problem hiding this comment.
Due to some parser issue we're not allowed to simplify the expression for now:
error: Unsupported index type, expected int or tvm.tir.PrimExpr, but got <class 'tvm.script.tir.node.BufferSlice'>
--> tensorir.py:431:44
|
431 | T.reads(A_indptr[vi : vi + 1], A_data[A_indptr[vi]: A_indptr[vi + 1]])
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
note: run with `TVM_BACKTRACE=1` environment variable to display a backtrace.
I suppose this is some bug that can be easily fixed, though I didn't get enough time to take a deeper look into the parser.
There was a problem hiding this comment.
One problem with this analysis is that we cannot assume A_indptr[i+1] >= A_indptr[i].
There was a problem hiding this comment.
Good catch! I never thought of this before. Is it okay by specifying assume_in_range(A_indptr[i + 1] - A_indptr[i], [0, 127]) in the future? (not pretty sure)
There was a problem hiding this comment.
Yes I think so, but we merge this one first.
MasterJH5574
force-pushed
the
feature/2022-02-20-evaluator-buffferload
branch
from
97c8d56
to
4fe1bb4
Compare
Hzfengsy
approved these changes
Feb 21, 2022
spectrometerHBH
approved these changes
Feb 21, 2022
pfk-beta
pushed a commit
to pfk-beta/tvm
that referenced
this pull request
Apr 11, 2022
…nge/loop start and end (apache#10370) A quick fix of the parser issue mentioned in apache#10327 . Ranges and loops require `start` and `stop` to be PrimExpr, however, `BufferSlice` is not always scalar so it's not a `PrimExpr`. This PR performs the transformation.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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 PR introduces the support of integer BufferLoadNodes in
arith::IntervalSetEvaluatoras well as a unittest.Prior to this PR, the evaluator always return
IntervalSet::Everything()(that is[-∞, +∞]) once it encounters a BufferLoad. However, as long as 1) the buffer data is of type integer and 2) the indices of the BufferLoad contain no variable that needs to be relaxed during evaluation, we can keep the BufferLoad as it is.Besides, we're considering introducing
assume_in_rangefor buffers for the use of range/region analysis. Withassume_in_range, even when the indices of BufferLoad contain some variables to be relaxed, we can just return the asserted range, which is more compact, instead of the universe.cc @yzh119 @Hzfengsy @spectrometerHBH @junrushao1994 @tqchen