[QST] TMA / CuTe 1D "im2col"

What is your question?

Is there a way to use TMA / CuTe to perform that following load pattern:

Starting with this 1-D tensor,

[1, 2, 3, 4]

I'd like to load the following 2-D tensor in smem without materializing in gmem:

[
 [4, 0, 0, 0]
 [3, 4, 0, 0]
 [2, 3, 4, 0]
 [1, 2, 3, 4]
]

The 2-D tensor will always be square, with dim0==dim1 >= dim of original 1-D tensor.

More generally, there will be N 1-D tensors in gmem:

[
 [1, 2, 3, 4]
 [5, 6, 7, 8]
 ...
 ]

Each of which will be expanded to 2-D tensors (each threadblock indexing one such 2-D tensor)
in smem:

[
[4, 0, 0, 0]
[3, 4, 0, 0]
[2, 3, 4, 0]
[1, 2, 3, 4]
]

[
[8, 0, 0, 0]
[7, 8, 0, 0]
[6, 7, 8, 0]
[5, 6, 7, 8]
]

This is similar to im2col for 4D image tensors, but in this case we're starting with 1-D tensors.

@thakkarV @Ethan-Yan27 @ccecka

No file, test, or entry point is named. Start by identifying the existing TMA/CuTe transfer entry points relevant to the requested expansion; done means a maintainer-confirmed answer about whether this pattern is expressible, with any required implementation scope made explicit.