[Bahaatbb]

Proposed changes

np.tile function in mlx as a op mlx.core.tile

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[angeloskath]

Thanks for adding this. However, tile should be implemented using broadcast_to and reshape rather than repeat to avoid doing multiple unnecessary copies.

See #318 for an improved repeat implementation that could be used as a guide for a tile.

[Bahaatbb]

@angeloskath Hey, Happy new year 🎉.
sorry if my first implementation is not up for the standards, tho I must say that it is almost the same implementation as numpy.
tho, I went back and tried this implementation and hacked together this code using your idea, the below code is just a prototype, to take your approval to proceed.

def tile(arr, repeats):
    d = len(repeats)
    
    if d < arr.ndim:
        repeats = [1] * (arr.ndim - d) + repeats
    else:
        arr = np.expand_dims(arr, axis=tuple(range(d - arr.ndim)))

    for i, rep in enumerate(repeats):
        if rep != 1:
            arr = np.expand_dims(arr, axis=i)
            arr = np.broadcast_to(arr, arr.shape[:i] + (rep,) + arr.shape[i+1:])
            arr = np.reshape(arr, arr.shape[:i] + (-1,) + arr.shape[i+2:])
            
    return arr

Note: I change the implementation code to match python code.

[angeloskath]

Sorry for being late to review it. I left some comments, I think after these it should be mergeable. Thanks!

[angeloskath]

Copy is not needed here, you can simply do array arr = a;.

[angeloskath]

This is already much better than before, however again each of these reshapes will be an actual copy and there is no need to launch many. Basically this is like calling repeat() many times.

The better way to do it is to keep 3 vectors.

• a reshape vector which will do all the expands at once (start with a.shape() and add 1s where needed)
• a broadcast to vector which will have the number of repeats where 1s were added in the above
• final reshape vector where the repeats are merged to the axis to their left

After that the implementation is as simple as the following:

auto out = reshape(a, expand_shape, s);
out = broadcast_to(out, repeat_shape, s);
out = reshape(out, final_shape, s);

[angeloskath]

Extremely nitpicky, feel free to ignore. Using an ostringstream would be better as it would allow you to also provide the whole repeats vector. Something like:

std::ostringstream msg;
msg << "[tile] Repeats argument cannot contain negative values but "
    << repeats << " was provided.";
throw std::invalid_argument(msg.str());

[dc-dc-dc]

is this pr stale ?

[Bahaatbb]

Sorry, I haven't had the chance to get back to this, I think did what is required @dc-dc-dc .
I will close this and look for more in the future.