Tensordot implementation #1545
Open
Tensordot implementation #1545
+444
−0
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
Adds a full implementation of `tensordot` for n-dimensional arrays, supporting both numeric and paired axis specifications via the new `AxisSpec` enum. The design mirrors NumPy’s `tensordot` behaviour while integrating cleanly with ndarray’s trait-based approach for the ``dot`` product and existing shape and stride logic. All internal reshaping and permutation operations use `unwrap` and `expect` with explicit safety reasoning: each call is guarded by dimension and axis validation, ensuring panics can only occur under invalid `AxisSpec` input. Documentation and inline comments describe these invariants and the exact failure conditions. Includes tests verifying correct contraction for both paired and integer axis modes.
nilgoyette
reviewed
Nov 12, 2025
| /// # use ndarray::linalg::AxisSpec; | ||
| /// let axes = AxisSpec::Pair(vec![1, -1], vec![0, 2]); | ||
| /// ``` | ||
| Pair(Vec<isize>, Vec<isize>), |
Collaborator
There was a problem hiding this comment.
I know both vectors must be of equal length, but must the number of elements be equal to the dimension?
I always try to reduce the number of memory allocations. I wonder if there's a better way than to accept 2 vectors. If the maximum number of elements = the dimension, we might be able to use constant size arrays. Although there are 2 dimensions here. Is one of the dimension always >= to the other?
src/linalg/impl_linalg.rs
Outdated
| /// - The computed product of reshaped dimensions does not equal the | ||
| /// array’s total element count (which would indicate internal logic error). | ||
| #[track_caller] | ||
| fn tensordot(&self, rhs: &Rhs, axes: AxisSpec) -> Self::Output; |
Collaborator
There was a problem hiding this comment.
I have never used tensordot. Is this a kind of function that gets called often with the same arguments? If so, wa might want to use &AxisSpec instead.
src/linalg/impl_linalg.rs
Outdated
|
|
||
| let c2 = a2.dot(&b2); | ||
|
|
||
| let mut out_shape: Vec<usize> = notin_a.iter().map(|&ax| a.shape()[ax]).collect(); |
Collaborator
There was a problem hiding this comment.
You're allocating 3 vectors in this function. Can't you re-use one of them for out_shape?
Updates ``tensordot_impl`` by removing one vector allocation, eliminating repeated axis-membership scans, and reducing shape-index lookups. Also switches axes_a/axes_b to borrowed inputs for potential reuse upstream. Replace ``notin_a`` + ``clone`` with direct construction of ``out_shape``, removing 1 allocation and 1 clone. Allocation count is now: - ``is_contracted_a`` → ``O(ndim(a))`` - is_contracted_b → ``O(ndim(b))`` - newaxes_a → ``O(ndim(a))`` - newaxes_b → ``O(ndim(b))`` - out_shape → ``O(ndim(a) + ndim(b) - contracted)`` All sizes are exactly determined by the axis mask and do not depend on runtime data beyond shape rank. Precompute boolean membership arrays for contracted axes, replacing multiple ``iter().any()`` scans with O(1) lookups. Cache ``a.shape()`` and ``b.shape()`` slices to avoid repeated indexing. Update signature to accept borrowed axis lists, allowing the caller to reuse them without moving.
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.
Tensordot implementation Issue #1517
Adds a full implementation of
tensordotfor n-dimensional arrays,supporting both numeric and paired axis specifications via the new
AxisSpecenum. The design mirrors NumPy’stensordotbehaviour whileintegrating cleanly with ndarray’s trait-based approach for the
dotproduct and existing shape and stride logic.
All internal reshaping and permutation operations use
unwrapandexpectwith explicit safety reasoning: each call is guarded bydimension and axis validation, ensuring panics can only occur under
invalid
AxisSpecinput. Documentation and inline commentsdescribe these invariants and the exact failure conditions.
Includes tests verifying correct contraction for
both paired and integer axis modes.