Swizziling in cute

[N1GHTR0]

Hi everyone,
Apologies in advance if this is the wrong place or if my question seems dumb.

I have a float shared memory region of size 19x128, and its layout is defined as:

Layout lsmem = make_layout(make_shape(Int<19>{}, Int<128>{}),
make_stride(Int<128>{}, Int<1>{}));

I want to swizzle it using Swizzle<3, 2, 5>{}, inspired by Lei Mao's blog:
https://leimao.github.io/blog/CuTe-Swizzle/

However, after applying this swizzle, I run into issues — I cannot partition the layout using either make_tiled_copy or local_partition. As far as I understand, both of these use zipped_divide internally, and it seems to break due to the swizzled layout.

I suspect the root cause is the 19x128 size: since 19 is not divisible by 8 (the repeat factor introduced by this swizzle), the resulting swizzled layout doesn't conform nicely. That is, the swizzle repeats every 8 rows, and 19 being not a multiple of 8 causes the pattern to become irregular or incompatible.

If anyone has encountered a similar situation or has suggestions on how to handle this (either a workaround or an alternative approach), I’d really appreciate any guidance.

Thanks in advance!

[thakkarV]

#1397 (comment)

[N1GHTR0]

Thank you 🙏🙏