Migrate base memory resources to native CCCL resource concept by bdice · Pull Request #2289 · rapidsai/rmm

bdice · 2026-03-11T23:24:33Z

Description

Migrates all 8 base (non-adaptor) memory resources to natively satisfy the CCCL cuda::mr::resource concept, so that concrete types (e.g. cuda_memory_resource&) satisfy the concept without virtual dispatch through device_memory_resource.

Stateless resources get allocate/deallocate/allocate_sync/deallocate_sync accepting cuda::stream_ref directly on the class:

cuda_memory_resource — device_accessible
managed_memory_resource — device_accessible + host_accessible
pinned_host_memory_resource — device_accessible + host_accessible
cuda_async_view_memory_resource — device_accessible
system_memory_resource — device_accessible + host_accessible

Stateful, non-copyable resources use cuda::mr::shared_resource<Impl> with _impl classes extracted to detail/ headers and .cpp source files, matching the adaptor convention from prior PRs (e.g. limiting_resource_adaptor):

cuda_async_memory_resource — device_accessible
cuda_async_managed_memory_resource — device_accessible + host_accessible
sam_headroom_memory_resource — device_accessible + host_accessible

device_memory_resource inheritance is kept for backward compatibility. do_allocate/do_deallocate delegate to the new CCCL methods. Default alignment is rmm::CUDA_ALLOCATION_ALIGNMENT (256 bytes).

Checklist

I am familiar with the Contributing Guidelines.
New or existing tests cover these changes.
The documentation is up to date with these changes.

copy-pr-bot · 2026-03-11T23:24:36Z

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

Migrate all 8 base (non-adaptor) memory resources to natively satisfy the CCCL cuda::mr::resource concept, so that concrete types satisfy the concept without virtual dispatch through device_memory_resource. Stateless resources (cuda_memory_resource, managed_memory_resource, pinned_host_memory_resource, cuda_async_view_memory_resource, system_memory_resource) get allocate/deallocate/allocate_sync/ deallocate_sync with cuda::stream_ref directly on the class. Stateful resources (cuda_async_memory_resource, cuda_async_managed_memory_resource, sam_headroom_memory_resource) use the cuda::mr::shared_resource<Impl> pattern with _impl classes extracted to detail/ headers and .cpp source files, matching the adaptor convention established in prior PRs. device_memory_resource inheritance is kept for backward compatibility, with do_allocate/do_deallocate delegating to the new CCCL methods.

cpp/include/rmm/mr/detail/cuda_async_managed_memory_resource_impl.hpp

cpp/include/rmm/mr/detail/cuda_async_memory_resource_impl.hpp

cpp/include/rmm/mr/detail/sam_headroom_memory_resource_impl.hpp

wence- · 2026-03-13T10:38:18Z

cpp/include/rmm/mr/cuda_async_managed_memory_resource.hpp

  /**
-   * @brief Allocates memory of size at least \p bytes.
-   *
-   * The returned pointer will have at minimum 256 byte alignment.
-   *
-   * @param bytes The size of the allocation
-   * @param stream Stream on which to perform allocation
-   * @return void* Pointer to the newly allocated memory
+   * @brief Enables the `cuda::mr::device_accessible` property
   */
-  void* do_allocate(std::size_t bytes, rmm::cuda_stream_view stream) override
+  RMM_CONSTEXPR_FRIEND void get_property(cuda_async_managed_memory_resource const&,
+                                         cuda::mr::device_accessible) noexcept
  {
-    return pool_.allocate(stream, bytes);
  }



Is it possible to get these properties from the impl via using shared_base::get_property?

get_property is a friend free function (ADL), not a member, so using cannot pull it in. All shared_resource classes duplicate it.

cpp/src/mr/detail/sam_headroom_memory_resource_impl.cpp

cpp/include/rmm/mr/cuda_memory_resource.hpp

cpp/include/rmm/mr/managed_memory_resource.hpp

cpp/include/rmm/mr/pinned_host_memory_resource.hpp

cpp/include/rmm/mr/system_memory_resource.hpp

…or== shortcircuit - Replace (void)param casts with [[maybe_unused]] on parameter declarations in all 5 stateless base resource headers - Add cudaStreamSynchronize in allocate_sync for cuda_memory_resource, managed_memory_resource, pinned_host_memory_resource, system_memory_resource, and sam_headroom_memory_resource_impl - Add this==addressof(other) shortcircuit to operator== in cuda_async_memory_resource_impl and cuda_async_managed_memory_resource_impl - Add <memory> include for std::addressof in the two async impl headers

…ons" This reverts commit 5e4d5de.

wence-

This looks good AFAICT, with some small comments to address/discuss in followup

wence- · 2026-03-16T10:51:34Z

cpp/include/rmm/mr/detail/cuda_async_memory_resource_impl.hpp

@@ -0,0 +1,90 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2021-2026, NVIDIA CORPORATION.


nit: This is a new file, is this the correct copyright line?

I kept an older copyright year because most of the file contents are being moved/copied from another file.

wence- · 2026-03-16T10:56:00Z

cpp/include/rmm/mr/detail/sam_headroom_memory_resource_impl.hpp

+
+#include <cstddef>
+
+namespace RMM_NAMESPACE {


Just a reminder that RMM_NAMESPACE has __attribute((visibility("default"))) IIRC, so all of these impl types are, today, still in the public symbols. I think we do not fix that here, but once the dust has settled perhaps we can go round and more correctly deal with this.

Yes, absolutely. I plan to do a thorough public symbol audit once I get through this migration.

wence- · 2026-03-16T10:58:24Z

cpp/include/rmm/mr/cuda_async_managed_memory_resource.hpp

+  // Begin legacy device_memory_resource compatibility layer
+  using device_memory_resource::allocate;
+  using device_memory_resource::allocate_sync;
+  using device_memory_resource::deallocate;
+  using device_memory_resource::deallocate_sync;


This is going away soon, but remind me, why do we get the concrete implementations from device_memory_resource rather than shared_base?

See previous comments here: #2246 (comment)

This is being fixed in a later phase. I have a draft of it locally. I think it's fine to delay this until that later PR is ready.

cpp/src/mr/detail/cuda_async_memory_resource_impl.cpp

Co-authored-by: Lawrence Mitchell <wence@gmx.li>

github-project-automation bot added this to RMM Project Board Mar 11, 2026

bdice force-pushed the cccl-base-mrs branch from fc6bed8 to 2aaf53c Compare March 12, 2026 20:05

bdice marked this pull request as ready for review March 12, 2026 21:15

bdice requested review from a team as code owners March 12, 2026 21:15

bdice requested review from harrism and miscco and removed request for a team March 12, 2026 21:15

bdice added breaking Breaking change improvement Improvement / enhancement to an existing function labels Mar 12, 2026

bdice self-assigned this Mar 12, 2026

bdice moved this to In Progress in RMM Project Board Mar 12, 2026

bdice mentioned this pull request Mar 12, 2026

[FEA] Support memory resources from CCCL 3.2 #2011

Open

37 tasks

wence- requested changes Mar 13, 2026

View reviewed changes

github-project-automation bot moved this from In Progress to Review in RMM Project Board Mar 13, 2026

bdice added 3 commits March 13, 2026 18:21

Add [[nodiscard]] to all allocate and allocate_sync declarations

5e4d5de

Revert "Add [[nodiscard]] to all allocate and allocate_sync declarati…

b45ab23

…ons" This reverts commit 5e4d5de.

bdice requested a review from wence- March 13, 2026 19:06

bdice mentioned this pull request Mar 14, 2026

Migrate Python/Cython bindings from device_memory_resource to device_async_resource_ref #2300

Draft

wence- approved these changes Mar 16, 2026

View reviewed changes

Update cpp/src/mr/detail/cuda_async_memory_resource_impl.cpp

857d291

Co-authored-by: Lawrence Mitchell <wence@gmx.li>

bdice merged commit c4b5236 into rapidsai:staging Mar 16, 2026
81 of 84 checks passed

github-project-automation bot moved this from Review to Done in RMM Project Board Mar 16, 2026

bdice mentioned this pull request Mar 16, 2026

[FEA] Migrate base memory resources to native CCCL interface #2287

Closed

17 tasks

		@@ -0,0 +1,90 @@
		/*
		* SPDX-FileCopyrightText: Copyright (c) 2021-2026, NVIDIA CORPORATION.


		#include <cstddef>

		namespace RMM_NAMESPACE {

Conversation

bdice commented Mar 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Checklist

Uh oh!

copy-pr-bot bot commented Mar 11, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

wence- Mar 13, 2026

Choose a reason for hiding this comment

Uh oh!

bdice Mar 13, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

wence- left a comment

Choose a reason for hiding this comment

Uh oh!

wence- Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

bdice Mar 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

wence- Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

bdice Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

wence- Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

bdice Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

bdice commented Mar 11, 2026 •

edited

Loading

bdice Mar 16, 2026 •

edited

Loading