Base device_memory_resource on cuda::stream_ordered_memory_resource#883
Base device_memory_resource on cuda::stream_ordered_memory_resource#883harrism wants to merge 33 commits intorapidsai:branch-21.12from
Conversation
…source<cuda::memory_kind::device>`
robertmaynard
left a comment
There was a problem hiding this comment.
libcudacxx headers will need to be marked as non-system for users of rmm.
This is necessary so that the version that rmm installs is used instead of the older version providied by the CUDA Toolkit.
We can re-use cudf's logic for this: https://github.com/rapidsai/cudf/blob/branch-21.12/cpp/
robertmaynard
left a comment
There was a problem hiding this comment.
The big thing I am noticing as far as build-system impacts it the usage of 1.5 instead of 1.4.
This means projects that use cuDF or cuCollections ( or any other consumer of libcudacxx ) and RMM will have two versions of libcudacxx on the include path. This occurs because instead of having a find_package(libcudacxx) each project side-steps and injects the internally packaged cudacxx headers on the include line.
Long term we need libcudacxx to provide proper CMake module support so projects don't install a copy in a custom 'internal' location.
Co-authored-by: Robert Maynard <robertjmaynard@gmail.com>
Co-authored-by: Robert Maynard <robertjmaynard@gmail.com>
@robertmaynard is that something you would be able / willing to contribute to libcudacxx? |
|
rerun tests |
[gpuCI] Forward-merge branch-21.12 to branch-22.02 [skip gpuci]
[gpuCI] Forward-merge branch-21.12 to branch-22.02 [skip gpuci]
This PR adds a script to find the cmake-format-rapids-cmake.json file in a standard location and run the cmake-format or cmake-lint programs with that config file. The script fails gracefully when the file cannot be found and is therefore suitable for use as a pre-commit hook in scenarios where no build directory (containing the config file) exists yet. A corresponding pre-commit configuration is added here as well, replacing the old cmake-format hook which did not use the rapids-cmake config file. Resolves rapidsai#903. Authors: - Vyas Ramasubramani (https://github.com/vyasr) Approvers: - Mark Harris (https://github.com/harrism) - Robert Maynard (https://github.com/robertmaynard) - Rong Ou (https://github.com/rongou) URL: rapidsai#913
[gpuCI] Forward-merge branch-21.12 to branch-22.02 [skip gpuci]
…m_ordered_resource`
|
This PR has been labeled |
|
This PR has been labeled |
|
Superseded by #1095 |
This PR is a stepping stone toward basing RMM memory resources on libcu++
cuda::memory_resource. This PR is intended to make it possible to transition dependent libraries (e.g. libcudf) to usingcuda::resource_viewinstead ofdevice_memory_resourcepointers. Doing that will make it easier to change RMM over with less disruption.device_memory_resourceinheritcuda::stream_ordered_memory_resource<cuda::memory_kind::device>Note that
cuda::memory_resourcecurrently only exists in a branch of libcu++ that has not been merged/released yet, so this should not be merged yet.