BUG: <PERF: Using np.reshape(tensor) is slow>

[HarryCollins2]

Describe the issue:

I made a test on this problem and the time difference on tensor is large (10x). Seems that it needs a fix for the data structure that can be used as an array.

Reproduce the code example:

import numpy as np
import torch
import time

a_np = np.random.rand(1000, 1000)
b_np = np.random.rand(1000, 1000)
a_torch = torch.from_numpy(a_np)
b_torch = torch.from_numpy(b_np)


%timeit reshaped_np = np.reshape(a_torch, (2000, 500))
%timeit reshaped_torch = np.reshape(a_torch.numpy(), (2000, 500))

### Error message:

```shell
17.5 µs ± 155 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)
1.43 µs ± 11.9 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

Python and NumPy Versions:

1.26.0
3.9.18 (main, Sep 11 2023, 08:25:10)
[Clang 14.0.6 ]

Runtime Environment:

No response

Context for the issue:

No response

[seberg]

That is for torch to know, maybe torch converts to ndarray and that is just slow. There is nothing about NumPy involvement though, NumPy would allow torch to make it fast in principle through protocols.

[HarryCollins2]

Thank you for your response. But pandas and matplotlib both support the data structure having array interface to be handled like array, maybe it only small modification in numpy to convert it and it can boost performance. pandas-dev/pandas#44616, matplotlib/matplotlib#25882

[seberg]

I cannot say about whether torch wants or can do it. But as I said: NumPy provides the protocols (not just __array__). If anyone can do something about it, it's torch and not NumPy.

[HarryCollins2]

Ok, thanks for your patience and clarification