Memory Resources

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 of the RAPIDS Memory Manager (RMM) project 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:

cuda::get_property	Infrastructure to tag a user defined type with a given property	stable CCCL 3.1.0 / CUDA 13.1, experimental CCCL 2.2.0 / CUDA 12.3
cuda::mr::{synchronous_}resource and cuda::mr::{synchronous_}resource_with	Concepts that provide proper constraints for arbitrary memory resources.	stable CCCL 3.1.0 / CUDA 13.1, experimental CCCL 2.2.0 / CUDA 12.3
cuda::mr::{synchronous_}resource_ref	A non-owning type-erased memory resource wrapper that enables consumers to specify properties of resources that they expect. resource_ref is still an experimental design, only available if LIBCUDACXX_ENABLE_EXPERIMENTAL_MEMORY_RESOURCE is defined	experimental CCCL 2.2.0 / CUDA 12.3

These features are an evolution of std::pmr::memory_resource that was introduced in C++17. While std::pmr::memory_resource provides a polymorphic memory resource that can be adopted through inheritance, it is not properly suited for heterogeneous systems.

With the current design it ranges from cumbersome to impossible to verify whether a memory resource provides allocations that are e.g. accessible on device, or whether it can utilize other allocation mechanisms.

To better support asynchronous CUDA stream-ordered allocations libcu++ provides cuda::stream_ref as a wrapper around cudaStream_t. The definition of cuda::stream_ref can be found in the <cuda/stream_ref> header.