Skip to content

[AsyncAlloc][SYCL][CUDA][Exp] Initial device side implementation for the sycl_ext_oneapi_async_memory_alloc extension #16900

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 45 commits into from
Mar 27, 2025
Merged

[AsyncAlloc][SYCL][CUDA][Exp] Initial device side implementation for the sycl_ext_oneapi_async_memory_alloc extension #16900

merged 45 commits into from
Mar 27, 2025

Conversation

Seanst98
Copy link
Contributor

@Seanst98 Seanst98 commented Feb 6, 2025 •
edited
Loading

Implement the sycl_ext_oneapi_async_memory_alloc extension for asynchronous memory allocation and freeing in CUDA, for device allocated pools only.

SYCL entrypoints which specify host or shared side pools, or pools created by pre-existing allocations will throw.

co-authored-by: Sean Stirling sean.stirling@codeplay.com
co-authored-by: Hugh Delaney hugh.delaney@codeplay.com

@Seanst98 Seanst98 mentioned this pull request Feb 6, 2025
Closed
@EwanC EwanC mentioned this pull request Feb 21, 2025
Merged
@Seanst98 Seanst98 force-pushed the sean/async-alloc branch 2 times, most recently from 30ed8bb to 82a106a Compare February 21, 2025 13:18
@Seanst98 Seanst98 mentioned this pull request Feb 26, 2025
Closed
AerialMantis
AerialMantis approved these changes Mar 26, 2025
View reviewed changes
Copy link
Contributor

@AerialMantis AerialMantis left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for the changes, this LGTM.

@npmiller
Copy link
Contributor

@intel/llvm-gatekeepers I believe this is ready to merge

@martygrant martygrant merged commit faa2365 into intel:sycl Mar 27, 2025
31 checks passed
@npmiller npmiller mentioned this pull request Mar 27, 2025
Merged
@uditagarwal97
Copy link
Contributor

Hi @Seanst98 ,
This PR broke nightly: https://github.com/intel/llvm/actions/runs/14121213165/job/39561695855

/__w/llvm/llvm/src/unified-runtime/source/adapters/cuda/usm.cpp: In constructor 'ur_usm_pool_handle_t_::ur_usm_pool_handle_t_(ur_context_handle_t, ur_device_handle_t, ur_usm_pool_desc_t*)':
/__w/llvm/llvm/src/unified-runtime/source/adapters/cuda/usm.cpp:434:20: error: 'CUmemPoolProps' {aka 'struct CUmemPoolProps_st'} has no member named 'maxSize'
  434 |       MemPoolProps.maxSize =
      |                    ^~~~~~~

This could be due to difference in CUDA version in pre-commit vs. Nightly. The Ubuntu 22 build job in Nightly uses CUDA 12.1, while Ubuntu 24 build job uses CUDA 12.6.3.
Could you please look into this?

@Seanst98
Copy link
Contributor Author

Hi @Seanst98 , This PR broke nightly: https://github.com/intel/llvm/actions/runs/14121213165/job/39561695855

Thanks for bringing this to my attention, I've pushed a PR which addresses this: #17733

sommerlukas pushed a commit that referenced this pull request Mar 31, 2025
@npmiller
[SYCL][AsyncAlloc] Fix minor async alloc issues (#17689)
a8d7f08 
This patch fixes a couple static analysis issues with the recent [async
alloc patch](#16900):

* Use `std::move` for shared pointer in `CGAsyncAlloc` constructor. It
is already passed by-value to the constructor so we can just move it
when assigning it to the member.
* Assert that the queue is available in `AsyncFree`
* Catch any exceptions from the memory pool destructor
* Initialize AsyncAlloc fields in handler
* Add `[[maybe_unused]]` for parameter only used in assert
@npmiller npmiller mentioned this pull request Apr 1, 2025
Open
@Seanst98 Seanst98 deleted the sean/async-alloc branch April 4, 2025 16:39
KornevNikita pushed a commit that referenced this pull request May 27, 2025
@Seanst98 @hdelan
@npmiller @KornevNikita
[AsyncAlloc][SYCL][CUDA][Exp] Initial device side implementation for …
ec11e27 
…the sycl_ext_oneapi_async_memory_alloc extension (#16900)

Implement the
[sycl_ext_oneapi_async_memory_alloc](#14800)
extension for asynchronous memory allocation and freeing in CUDA, for
device allocated pools only.

SYCL entrypoints which specify host or shared side pools, or pools
created by pre-existing allocations will throw.

co-authored-by: Sean Stirling <sean.stirling@codeplay.com>
co-authored-by: Hugh Delaney <hugh.delaney@codeplay.com>

---------

Co-authored-by: Hugh Delaney <hugh.delaney@codeplay.com>
Co-authored-by: Nicolas Miller <nicolas.miller@codeplay.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@EwanC EwanC EwanC approved these changes

@Bensuo Bensuo Bensuo left review comments

@aelovikov-intel aelovikov-intel aelovikov-intel left review comments

@hvdijk hvdijk hvdijk approved these changes

@npmiller npmiller npmiller approved these changes

@AerialMantis AerialMantis AerialMantis approved these changes

@isaacault isaacault isaacault approved these changes

@steffenlarsen steffenlarsen steffenlarsen approved these changes

@maksimsab maksimsab maksimsab approved these changes

@frasercrmck frasercrmck Awaiting requested review from frasercrmck frasercrmck is a code owner automatically assigned from intel/llvm-reviewers-cuda

@uditagarwal97 uditagarwal97 Awaiting requested review from uditagarwal97 uditagarwal97 is a code owner automatically assigned from intel/llvm-reviewers-runtime

@sergey-semenov sergey-semenov Awaiting requested review from sergey-semenov

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.