Fix for evaluation containing array and its inverse. #3671
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
A bug was found in the jit calculation tree where under certain circumstances the incorrect result would be found when both an array and a transpose of the same array are used in the same evaluation. When the transpose method is used on an array the treatment depends on the structure of the jit tree of that array. In the particular case that the array is linear (e.g. not a sub-array) and is not simply a buffer (i.e. a combination of buffers and/or scalars with some operators) a moddim node will be added to the root of the tree with the new dimensions. When the tree is evaluated the dimensions of the child buffer(s) of a moddim node are changed and the moddim node is deleted. If an evaluation happens to include both an array and the transpose of that same array then when the moddim node is applied it changes the dimensions of the children for the remaining lifetime of the evaluation including any subsequent use of these nodes. As a result if the transpose of an array is used in an expression followed by the array itself the second instance will also be transposed. In this fix instead of creating a moddims node a deep copy of the calculation tree is made at that point with the transpose applied to the child buffer nodes. These buffer node copies will have the transposed dimensions and strides but will still point to the same memory location for the data so no copy of the underlying data needs to be made.
umar456
reviewed
Jun 30, 2025
|
|
||
| return out; | ||
| } | ||
|
|
||
| template<typename T> | ||
| Array<T> moddimOp(const Array<T> &in, af::dim4 outDim) { |
There was a problem hiding this comment.
It looks like moddims is not a jittable operation anymore with this change. what are the performance implications of this?
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.
A bug was found in the jit calculation tree where under certain circumstances the incorrect result would be found when both an array and a transpose of the same array are used in the same evaluation. When the transpose method is used on an array the treatment depends on the structure of the jit tree of that array. In the particular case that the array is linear (e.g. not a sub-array) and is not simply a buffer (i.e. a combination of buffers and/or scalars with some operators) a moddim node will be added to the root of the tree with the new dimensions. When the tree is evaluated the dimensions of the child buffer(s) of a moddim node are changed and the moddim node is deleted. If an evaluation happens to include both an array and the transpose of that same array then when the moddim node is applied it changes the dimensions of the children for the remaining lifetime of the evaluation including any subsequent use of these nodes. As a result if the transpose of an array is used in an expression followed by the array itself the second instance will also be transposed.
In this fix instead of creating a moddims node a deep copy of the calculation tree is made at that point with the transpose applied to the child buffer nodes. These buffer node copies will have the transposed dimensions and strides but will still point to the same memory location for the data so no copy of the underlying data needs to be made.
Checklist