Refactor the cuda_memcpy functions to make them more usable
#16945
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
vuule
added
improvement
3 tasks
bdice
reviewed
Sep 27, 2024
| impl::cuda_memcpy_async( | ||
| dst.data(), | ||
| src.data(), | ||
| std::min(dst.size_bytes(), src.size_bytes()), |
There was a problem hiding this comment.
This might cause surprises to users if they unintentionally pass spans of different sizes. We could instead throw in this case.
Yes, this should be a runtime check and it should throw. If the caller wants to copy subspans, the caller can create subspans. Spans, as view types, are meant to make this easy.
There was a problem hiding this comment.
Sure.
Just one thing to keep in mind for this use case:
host_span{my_hv.data(), subsize} is not the same as host_span{my_hv}.subspan(0, subsize) because the first one will not know if it's pointing to pinned memory.
3 tasks
pmattione-nvidia
approved these changes
Sep 30, 2024
bdice
approved these changes
Oct 1, 2024
mythrocks
reviewed
Oct 1, 2024
| source_data.size() * sizeof(T), | ||
| is_pinned ? host_memory_kind::PINNED : host_memory_kind::PAGEABLE, | ||
| stream); | ||
| cuda_memcpy_async<T>(ret, source_data, stream); |
There was a problem hiding this comment.
Oh, yes. This is much better.
|
/merge |
rapids-bot bot
pushed a commit
that referenced
this pull request
Oct 18, 2024
Depends on #16945 Added `cudf::detail::device_scalar`, derived from `rmm::device_scalar`. The new class overrides function members that perform copies between host and device. New implementation uses a `cudf::detail::host_vector` as a bounce buffer to avoid performing a pageable copy. Replaced `rmm::device_scalar` with `cudf::detail::device_scalar` across libcudf. Authors: - Vukasin Milovanovic (https://github.com/vuule) - Vyas Ramasubramani (https://github.com/vyasr) Approvers: - Basit Ayantunde (https://github.com/lamarrr) - Vyas Ramasubramani (https://github.com/vyasr) - David Wendt (https://github.com/davidwendt) URL: #16947
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
As we expanded the use of the
cuda_memcpyfunctions, we realized that they are not very ergonomic, as they require caller to queryis_device_accessibleand pass the correctPAGEABLE/PINNEDenum based on this.This PR aims to make the
cuda_memcpyfunctions easier to use, and the call site changes hopefully showcase this. The new implementation takes spans as parameters and relies on thehost_span::is_device_accessibleto enable copy strategies for pinned memory. Host spans set this flag during construction; creating a host span from acudf::detail::host_vectorwill correctly propagateis_device_accessible. Thus, call can simply* call thecuda_memcpyfunctions with their containers as parameters and rely on implicit conversion tohost_span/device_span. Bonus - there's no way to mix up host and device memory pointers 👍Sharp edges:
The API copies themin(input.size(), output.size())bytes, as this is what we can do safely. This might cause surprises to users if they unintentionally pass spans of different sizes. We could instead throw in this case.Checklist