Open
Description
Some CUDA kernels in cuML
and other RAPIDS libraries require memory alignment. This shouldn't be a problem as long as RMM
is used as allocator / memory handler because RMM
allocations are aligned by default. However, the native API do not have any guarantee on the addresses of the pointers provided. It might be necessary to check for memory alignment at multiple places in the codebase. If the condition isn't met (memory isn't aligned correctly), a secondary pathway should be chosen.