Open ttnghia opened 2 months ago
Merged PR 2011 does much of this work; this conversion has now been performed at all of the kernel interfaces. However the resource allocator in SparkResourceAdaptor.jni still uses device_memory_resource and needs to be updated.
Further work is blocked until RMM updates its memory resource classes. The cuDF Java initialize() function ultimately creates the RMM C++ memory resources, but they all still inherit from device_memory_resource. The RMM team says that they still need to figure some things out before they can update them to use the new memory concepts.
In anticipation of the upcoming changes, the following draft PRs have been created, containing all of the work that can be done until then:
cuDF Java Code Draft PR spark-rapids-jni Draft PR
The remaining work that needs to be done before we can update our Java code is being tracked here:
In order to move along with
rmm
code refactoring, similar to https://github.com/rapidsai/cudf/pull/15507, we need to perform a repository-wide conversion fromrmm::mr::device_memory_resource *
intormm::device_async_resource_ref
.This needs to be done ASAP before
rmm::mr::device_memory_resource *
is removed fromrmm
.