Make compute_strides function a public function

[yasahi-hpc]

This PR aims at moving compute_strides under KokkosFFT::Impl namespace. This function can be used not only for rocFFT but also for cuFFTMp with small differences.

[PaulGannay]

I would rename IntType as IndexType

[yasahi-hpc]

OK

[yasahi-hpc]

May be called this extents_type?
Following mdspan convention

[PaulGannay]

I would say no, as I understand it, in MDSpan, Extends is a composite type grouping three different informations:

• Extents::index_type
• Extents::size_type
• Extents::rank_type

Here, Int_Type corresponds only to the index_type. But I may be mistaken?

[yasahi-hpc]

That is correct
What I intent was indeed index_type
https://en.cppreference.com/w/cpp/container/mdspan/extent.html

[PaulGannay]

Ok, yes, I agree with you, index_type would be an even better name.

[PaulGannay]

I understood the old

  // (n0, n1, n2) -> (1, n0, n0*n1)
  // (n0, n1) -> (1, n0)
  // (n0) -> (1)

but I can't make sense of this example.
Could you explain what kind of stride you are computing?

[yasahi-hpc]

You may imagine Views v in LayoutRight.
If we compute the strides from Right (the contiguous dimension) to Left, it would be

• v(n0) -> (1)
• v(n0, n1) -> (1, n1)
• v(n0, n1, n2) -> (1, n2, n2 * n1)

I do not know why it is given in this order though. (As you can see, in cuFFTMp the order is reversed. This makes more sense)

• v(n0) -> (1)
• v(n0, n1) -> (n1, 1)
• v(n0, n1, n2) -> (n2 * n1, n2, 1)

[PaulGannay]

Ok, this is clear, thank you.
Maybe it deserve an explanation in the comment too?

Otherwise it looks good to me.

[yasahi-hpc]

@PaulGannay

Thank you for approval. I will merge this after CI passes