Update to CCCL 2.2.0. #5702
Update to CCCL 2.2.0. #5702
Conversation
bdice
added
5 - DO NOT MERGE
| @@ -76,13 +76,18 @@ void get_probabilities(const raft::handle_t& handle, | |||
| rmm::device_uvector<value_t> deaths(n_clusters, stream); | |||
| thrust::fill(exec_policy, deaths.begin(), deaths.end(), 0.0f); | |||
|
|
|||
| cudaError_t (*reduce_func)(void*, | |||
There was a problem hiding this comment.
Why was the change to a function pointer needed?
There was a problem hiding this comment.
I'm not sure. CUB was failing to compile the previous code because of some API change. This code is borrowed from #5623. @hcho3 Do you have anything to add here?
@divyegala We may have to merge this as-is and revisit the question, or else CI will be blocked once the rest of RAPIDS is updated to CCCL 2.2.0.
There was a problem hiding this comment.
The change was made in b1302bd (#5623) but the commit message only says "Fix build."
There was a problem hiding this comment.
Okay, it's not egregious so I will approve for now.
There was a problem hiding this comment.
I kept getting the build error like
error: no instance of function template "detail::Utils::cub_segmented_reduce" matches the argument list
argument types are: (float *, float *, int, int *, cudaStream_t, <unknown-type>)
because the compiler failed to build the type of cub::DeviceSegmentedReduce::Max. I had to explicitly declare the type of cub::DeviceSegmentedReduce::Max by declaring a function pointer.
bdice
added
improvement
|
/merge |
This PR updates cuml to CCCL 2.2.0. Do not merge until all of RAPIDS is ready to update.
Replaces #5623.