[DOC]: Document alignment requirements for temporary storage when using cub::ScanTileState directly

[akeley98NV]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Is this for new documentation, or an update to existing docs?

New

Describe the incorrect/future/missing documentation

The high-level APIs for decoupled look-back don't have any alignment requirements AFAIK but this seems not the case if you use cub::ScanTileState directly. I had to use compute-sanitizer to figure out why my code crashed and it seems that d_temp_storage needs to be 8-byte aligned in this case.

I am making this a documentation request, not a bug request, as this behavior seems correct to me (if I'm working at this low level then I don't want cub padding or aligning pointers behind my back). I just suggest that this be documented.

If this is a correction, please provide a link to the incorrect documentation. If this is a new documentation request, please link to where you have looked.

I was just copying the example code here: https://github.com/NVIDIA/cccl/blob/main/cub/examples/device/example_device_decoupled_look_back.cu

I actually didn't see any documentation for ScanTileState in the HTML docs, I just looked around in the header files and didn't find anything about the alignment issue. In fact I was not really sure this is part of the public API except that Jake Hemstad told me it was (NVIDIA-internal link): https://nvidia.slack.com/archives/C07DK328L/p1697046387099159?thread_ts=1696996019.051079&cid=C07DK328L

This seems to be distinct from this issue: #910 (I'm specifically not expecting cub to align pointers for me)