[FEA] Refactor RMM in terms of cuda::mr::memory_resource

[harrism]

The <cuda/memory_resource> header provides a standard C++ interface for heterogeneous, stream-ordered memory allocation tailored to the needs of CUDA C++ developers. This design builds off of the success ofRMM and evolves the design based on lessons learned. <cuda/memory_resource> is not intended to replace RMM, but instead moves the definition of the memory allocation interface to a more centralized home in CCCL. RMM will remain as a collection of implementations of the cuda::mr interfaces.

At a high level, the header provides:

• the cuda::get_property infrastructure to tag a user defined type with a given property;
• the cuda::mr::{async}_resource and cuda::mr::{async}_resource_with concepts that provide proper constraints for arbitrary memory resources;
• the cuda::mr::{async}_resource_ref is a type-erased memory resource wrapper that enables consumers to specify properties of resources that they expect.

See the CCCL documentation for <cuda/memory_resource> for more details.

RMM began refactoring on top of cuda/memory_resource in #1095. A complete transition involves replacing the use of device_memory_resource in all RAPIDS repos (and other dependent libraries) and refactoring all RMM memory resources to implement the cuda::[async_]memory_resource properties rather than deriving from rmm::mr::device_memory_resource and rmm::mr::host_memory_resource.

Tasks

Beta Give feedback

1. Introduce cuda::mr::{async_}resource_ref to RMM and use in place of device_memory_resource pointers. Introduce conversions between rmm::cuda_stream_view and cuda::stream_ref. Done in Use cuda::mr::memory_resource instead of raw device_memory_resource #1095 (23.12)
2. Require explicit initial pool size in pool_memory_resource. This is because it is difficult to automatically size the pool when the memory is not device memory. Done in Require explicit pool size in pool_memory_resource and move some things out of detail namespace #1417
3. Deprecate and remove get_mem_info methods from memory resources. #1388
   4 of 4
   cpp feature request tech debt +3
   
4. Deprecate and remove supports_streams() from memory resources #1389
   3 of 3
   cpp feature request +2
   
5. [FEA] Change semantics of RMM memory resource equality comparison #1402
   feature request +1
   
6. [FEA] Add a convenience alias in RMM for cuda::mr::async_resource_ref<cuda::mr::device_accessible> #1442
   feature request +1
   
7. Refactor RMM tests to use cuda::resource_ref instead of device_memory_resource pointers. #1444
   feature request +1
8. Replace device_memory_resource* with device_async_resource_ref cudf#15498
   improvement +1
9. Replace device_memory_resource* with device_async_resource_ref raft#2261
   improvement +1
10. Replace device_memory_resource* with device_async_resource_ref cugraph#4333
    improvement +1
11. Replace device_memory_resource* with device_async_resource_ref cuml#5839
    improvement +1
12. Replace device_memory_resource* with device_async_resource_ref cuspatial#1371
    improvement +1
13. Convert any other RAPIDS repos to use rmm::mr::device_async_resource_ref instead of device_memory_resource pointers
14. Refactor RMM resources to implement cuda::memory_resource concepts and deprecate and remove device_memory_resource and host_memory_resource base classes. #1445
    9 of 14
    Epic feature request +2
    

Tasks

Beta Give feedback

No tasks being tracked yet.