Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Change NdProducer::Dim of axis_windows() to Ix1 #1305

Open
wants to merge 6 commits into
base: master
Choose a base branch
from

Conversation

jonasBoss
Copy link

See #1304

This breaks the public API because the method axis_windows() now returns a new AxisWindows struct. And the dimension of the NdProducer::Dim will always be Ix1.

If the old behavior is needed the method windows() can be used.

I was thinking about adding a method AxisWindows::into_windows(self) -> Windows but I don't think that this will see much use. Let me know if i should implement it.

}


impl<'a, A, D: Dimension> NdProducer for AxisWindows<'a, A, D> {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is the impl_ndproducer! macro unable to handle this impl? (It seems to be used for Windows.)

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

almost, the methods raw_dim, uget_ptr, stride_of and split_at need a different implementation.

let mut window = base.raw_dim();
window[axis_idx] = window_size;

base.slice_each_axis_inplace(|ax_desc| {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am not really happy about the code duplication between this method and Windows::new_with_stride. Is it reasonably possible to have re-use without making this too hard to follow? (Windows::new_with_stride also appears to have additional assertions/consistency checks?)

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There is a way to do this with reusing the Windows struct. I don't know if it is more or less clean. It would look like this:

pub struct Windows<'a, A, D, V>{
    base: ArrayView<'a, A, D>,
    window: D,
    strides: D,
    version: V,
}

pub struct GeneralWindow;
pub struct AxisWindow{index: usize}

impl<'a, A, D: Dimension> NdProducer for Windows<'a, A, D, GeneralWindow> {
   // the default impl 
}

impl<'a, A, D: Dimension> NdProducer for Windows<'a, A, D, AxisWindow> {
   // the impl for the Axis version
}

Now the methods axis_windows and windows would both return Windows but with a different generic V.

The struct GeneralWindow and AxisWindow need to be public, but they can be inaccessible from outside of the crate.


The assertions are not needed for AxisWindows::new because we construct the variables strides and window locally. axsis_windows() asserts that the axis is in range.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The struct GeneralWindow and AxisWindow need to be public, but they can be inaccessible from outside of the crate.

This does sound reasonable to me

I would suggest making Windows and AxisWindows public and have both wrap an WindowsImpl or something like that. (Meaning AxisWindows contains two fields, index: usize and impl: WindowsImpl.)

I would also suggest replacing the term "version" by "variant" to avoid any confusion as "version" often has a temporal connotation.

I don't know if it is more or less clean.

Of course, if this just increases complexity for very little gain in the end, it might not be worth it, you just have to give it a try.

As an alternative, a simple helper function encapsulating the slice-in-place approach using references to the constituent fields of Windows and AxisWindows might work as well.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I just updated to my suggested structure, It actually does reduce code duplication a lot.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am not a fan of exposing these implementation detail generics in the API, hence the suggestion to expose Windows and AxisWindows with the same generic parameters as before which contain a WindowsImpl/Inner/Whatever as a private field. We should not leak this sharing of implementation in the public API.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I thought that This does sound reasonable to [you] 🙁

We need to expose that there is a difference between the return type of windows() and axis_windows() there is no way around that.

This generic is not an implementation detail. It makes a real difference to the public API. Specifically it changes the behavior of <Windows as NdProducer> and the Type of <Windows as NdProducer>::Dim. Otherwise this PR would be pointless.

Now we have this solution which is as minimal as it gets.
Or the solution I had before. Sure I could try and faff around with some helper methods and structs to reduce code duplication for the cost of clarity and more boilerplate.
Having a WindowsInner/Impl/... will break the impl_ndproducer! macro, because base needs to be a field of Windows not of some wrapped type.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We need to expose that there is a difference between the return type of windows() and axis_windows() there is no way around that.

Distinct return types is not the issue. It is return types which only differ in a type parameter which I have a problem with. I want the relation between these two return types to be fully opaque to the caller, so that e.g. we could also not have code reuse in their implementation without that affecting our API.

This generic is not an implementation detail. It makes a real difference to the public API. Specifically it changes the behavior of and the Type of ::Dim. Otherwise this PR would be pointless.

<Windows as NdProducer>::Dim and <AxisWindows as NdProducer>::Dim can be different without the outside world seeing any relation between those two types.

Having a WindowsInner/Impl/... will break the impl_ndproducer! macro, because base needs to be a field of Windows not of some wrapped type.

This is a real issue and maybe manually implementing NdProducer as you did for AxisWindows for both cases is simple enough, but maybe this actually reduces reuse. This is why I also suggested just using an external helper method with completely disjoint struct types.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you for clearing that up. This makes much more sense.

I just extracted the main logic of creating the base into a separate function, and call that from both constructors Windwos::new_with_stride() and AxisWindows::new()

Comment on lines +225 to +232
fn stride_of(&self, axis: Axis) -> isize {
assert_eq!(axis, Axis(0));
self.base.stride_of(Axis(self.axis_idx))
}

fn split_at(self, axis: Axis, index: usize) -> (Self, Self) {
assert_eq!(axis, Axis(0));
let (a, b) = self.base.split_at(Axis(self.axis_idx), index);
Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Here I added assertions to check that this is only called with Axis(0). Do we need this? I could just ignore the axis argument, but that would be confusing when a user tries to call this with a different axis.


In general I am not super confident that this trait is correctly implemented, as there is a severe lack of documentation for a lot of the methods.

Specialize the `Windows` struct with a `variant` field to replace the
`AxisWindows` struct
@jonasBoss
Copy link
Author

Is there anything holding this back that I can do do address?

The CI errors don't seem to be related to this PR.

@adamreichold
Copy link
Collaborator

Is there anything holding this back that I can do do address?

The CI errors don't seem to be related to this PR.

No. Sorry this is waiting on me finding time to look into it again.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

None yet

3 participants