[BUG]: Wrong mask type in cub/cub/block/block_radix_rank.cuh #6106
Description
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
Type of Bug
Something else
Component
CUB
Describe the bug
The implementation uses int for masks that go into intrinsics that expect unsigned types.
So far this has been fine because the intrinsics behave nicely there but we plan on dropping them in favor of internal functions that actually check whether we are using the right type. We should change the types of match_masks and other masks to uint32_t or uint64_t
How to Reproduce
Came up in #6099 when transitioning __clz to cuda::std::countl_zero
Expected behavior
We should use the right type for bitmasks
Reproduction link
No response
Operating System
No response
nvidia-smi output
No response
NVCC version
No response