Files
pytorch/torch/testing
nikitaved edc2d539d1 torch.tensordot: performance improvements when contracting to a scalar. (#145936)
As per title.
Fixes https://github.com/pytorch/pytorch/issues/145731

Touches only compute. The CPU overhead can potentially be further reduced.

Before:
```python
In [3]: n = 512

In [4]: A = torch.rand(n, n)

In [5]: B = torch.rand(n, n)

In [6]: %timeit torch.tensordot(A, B, [[0, 1], [0, 1]])
2.04 ms ± 70 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [7]: %timeit torch.tensordot(A, B, [[0, 1], [1, 0]])
2.85 ms ± 191 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [8]: %timeit torch.tensordot(A, B, [[1, 0], [0, 1]])
2.9 ms ± 133 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [9]: %timeit torch.tensordot(A, B, [[1, 0], [1, 0]])
4.07 ms ± 262 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
```

After
```python
In [2]: n = 512

In [3]: A = torch.rand(n, n)

In [4]: B = torch.rand(n, n)

In [5]: %timeit torch.tensordot(A, B, [[0, 1], [0, 1]])
30.7 µs ± 2.51 µs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

In [6]: %timeit torch.tensordot(A, B, [[0, 1], [1, 0]])
141 µs ± 6.52 µs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

In [7]: %timeit torch.tensordot(A, B, [[1, 0], [0, 1]])
142 µs ± 4.03 µs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

In [8]: %timeit torch.tensordot(A, B, [[1, 0], [1, 0]])
62.8 µs ± 4.31 µs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

```

Pull Request resolved: https://github.com/pytorch/pytorch/pull/145936
Approved by: https://github.com/albanD, https://github.com/ngimel
2025-05-13 10:57:30 +00:00
..