Skip to content

[SYCL][Doc] Simplify non-uniform group design #14604

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
AlexeySachkov merged 18 commits into from
Oct 4, 2024
Merged

[SYCL][Doc] Simplify non-uniform group design #14604

AlexeySachkov merged 18 commits into from
Oct 4, 2024

Conversation

@Pennycook
Copy link
Contributor

@Pennycook Pennycook commented Jul 17, 2024 •
edited
Loading

This commit simplifies the design of the non-uniform group extension as follows:

  • The _group suffix is removed from user-constructed groups to reduce verbosity.

  • get_*_group() functions are replaced with partitioning functions, to clarify that users are not getting an existing group but rather creating a new one.

  • Fixed-size groups are renamed to "chunks". The previous name did not convey that the group is contiguous as well as having a fixed size.

  • Ballot-groups and opportunistic-groups have been combined into a single group type called a "fragment", focusing on what the groups represent rather than how they were constructed.

  • Synchronization behavior is added to all partitioning functions, to reduce the cognitive burden (of remembering which partitioning functions synchronize).

  • An implicit cast from chunks to fragments is introduced. A fragment is a more general representation than a chunk, and so this conversion is always valid.

  • Split tangle into its own extension. The implementation requirements for tangle are different to other non-uniform groups, which may limit their availability to certain device types. Splitting them out into a separate extension will make it easier to set user expectations and to track implementation status across backends.

@Pennycook
[SYCL][Doc] Simplify non-uniform group design
db8fffd 
This commit simplifies the design of the non-uniform group extension as
follows:

- The _group suffix is removed from user-constructed groups to reduce
  verbosity.

- get_*_group() functions are replaced with partitioning functions, to clarify
  that users are not getting an existing group but rather creating a new one.

- Fixed-size groups are renamed to "chunks". The previous name did not convey
  that the group is contiguous as well as having a fixed size.

- Ballot-groups and opportunistic-groups have been combined into a single group
  type called a "fragment", focusing on what the groups represent rather than
  how they were constructed.

- Synchronization behavior is added to all partitioning functions, to reduce the
  cognitive burden (of remembering which partitioning functions synchronize).

- An implicit cast from chunks to fragments is introduced. A fragment is a more
  general representation than a chunk, and so this conversion is always valid.

Signed-off-by: John Pennycook <[email protected]>
@Pennycook Pennycook added the spec extension All issues/PRs related to extensions specifications label Jul 17, 2024
@Pennycook Pennycook requested a review from steffenlarsen July 17, 2024 17:26
Pennycook added 3 commits July 22, 2024 08:53
@Pennycook
[SYCL][Doc] Split tangle into its own extension
d39a740 
The implementation requirements for tangle are different to other non-uniform
groups, which may limit their availability to certain device types. Splitting
them out into a separate extension will make it easier to set user extensions
and to track implementation status across backends.

Signed-off-by: John Pennycook <[email protected]>
@Pennycook
[SYCL][Doc] Add issue regarding tangle convergence
9b521c7 
Signed-off-by: John Pennycook <[email protected]>
@Pennycook
Fix copyright year
ac40eb8 
Signed-off-by: John Pennycook <[email protected]>
@Pennycook Pennycook marked this pull request as ready for review September 19, 2024 08:52
@Pennycook Pennycook requested a review from a team as a code owner September 19, 2024 08:52
gmlueck
gmlueck reviewed Sep 20, 2024
View reviewed changes
@Pennycook
Update header and revision
e28400d 
Signed-off-by: John Pennycook <[email protected]>
@Pennycook
SYCL 2020 => SYCL core
b4461cb 
Signed-off-by: John Pennycook <[email protected]>
AlexeySachkov
AlexeySachkov reviewed Sep 26, 2024
View reviewed changes
Copy link
Contributor

@AlexeySachkov AlexeySachkov left a comment

Choose a reason for hiding this comment

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

Haven't looked at the tangle doc yet

sycl/doc/extensions/experimental/sycl_ext_oneapi_non_uniform_groups.asciidoc
template <typename ParentGroup>
class tangle_group {
template <std::size_t ChunkSize, typename ParentGroup>
Copy link
Contributor

@AlexeySachkov AlexeySachkov Sep 26, 2024

Choose a reason for hiding this comment

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

We should probably update linear_id_type definition here, see the discussion about recorded inconsistency in #13702. Linear id type is smaller than chunk size, meaning that in certain cases we won't be able to enumerate all work-items in a chunk.

Copy link
Contributor Author

@Pennycook Pennycook Sep 27, 2024

Choose a reason for hiding this comment

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

I see a few possible fixes here and I'm not sure what is best:

  1. linear_id_type = std::size_t
    Always matches, no problems.

  2. linear_id_type = ParentGroup::linear_id_type
    Sets uint32_t for sub_group and size_t for group. The situation where somebody provides a ChunkSize larger than the ParentGroup would prevent enumeration of all work-items in the chunk, but that situation is already UB.

  3. linear_id_type = /* smallest unsigned integer with log2(ChunkSize) bits */
    Could use std::uint8_t for ChunkSize < 256. Might lead to future ABI issues if we get something like a std::uint4_t in future (but that seems unlikely, given the number of bits in a byte).

Copy link
Contributor

@AlexeySachkov AlexeySachkov Sep 27, 2024

Choose a reason for hiding this comment

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

I lean towards (2) here, because it makes chunk consistent with its parent group. I.e. if someone had a code for sub-groups, but then introduced some specialization for chunks, types of all IDs they may have in use would still match what APIs return. And this will also be true when we extend chunks to support group.

Copy link
Contributor Author

@Pennycook Pennycook Sep 30, 2024

Choose a reason for hiding this comment

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

Makes sense to me. I made the change in 0f2a471.

@AlexeySachkov AlexeySachkov mentioned this pull request Sep 26, 2024
Closed
AlexeySachkov
AlexeySachkov reviewed Sep 27, 2024
View reviewed changes
Copy link
Contributor

@AlexeySachkov AlexeySachkov left a comment

Choose a reason for hiding this comment

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

tangle part review

@Pennycook
Rename this_kernel => this_work_item
9ec0f2b 
Signed-off-by: John Pennycook <[email protected]>
@Pennycook
Simplify constraints with explicit list of types
100f139 
Signed-off-by: John Pennycook <[email protected]>
@Pennycook
Clarify which work-items join opportunistic groups
c41c4be 
Signed-off-by: John Pennycook <[email protected]>
@Pennycook
Clarify return value of get_group_range()
a4b38be 
Signed-off-by: John Pennycook <[email protected]>
@Pennycook
Apply fixes from intel#13702
cc24795 
Signed-off-by: John Pennycook <[email protected]>
@Pennycook
Replace user-defined => user-constructed
ebe8577 
Signed-off-by: John Pennycook <[email protected]>
@Pennycook
Drop max_local_range() requirement for chunks
6679e81 
Signed-off-by: John Pennycook <[email protected]>
@Pennycook
Allow binary_partition of chunks
3e1b001 
Signed-off-by: John Pennycook <[email protected]>
@Pennycook
Remove unnecessary barrier after entangle
9af5173 
Signed-off-by: John Pennycook <[email protected]>
AlexeySachkov
AlexeySachkov approved these changes Sep 27, 2024
View reviewed changes
Copy link
Contributor

@AlexeySachkov AlexeySachkov left a comment

Choose a reason for hiding this comment

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

All my feedback has been addressed, I don't have any further comments. That issue about liner_id_type of chunk is not a high-priority one to be resolved immediately, because we don't have a device which supports sub-group sizes bigger than uint32_t and potential issue I've highlighted there can't really be encountered right now.

@AlexeySachkov
Copy link
Contributor

AlexeySachkov commented Oct 1, 2024

@intel/dpcpp-specification-reviewers, any further comments before I merge this PR?

@AlexeySachkov
Copy link
Contributor

AlexeySachkov commented Oct 4, 2024

@intel/dpcpp-specification-reviewers, any further comments before I merge this PR?

I treat lack of response as a no, so merging the PR

@AlexeySachkov AlexeySachkov merged commit 761d45d into intel:sycl Oct 4, 2024
@AndreiZibrov AndreiZibrov mentioned this pull request Oct 16, 2024
Closed
@AndreiZibrov AndreiZibrov mentioned this pull request Jul 1, 2025
Merged
steffenlarsen pushed a commit that referenced this pull request Jul 30, 2025
@AndreiZibrov @Pennycook @uditagarwal97
[SYCL] Align Non-Uniform Groups with PR#14604 (#19238)
7e16ea8 
This PR implements the non-uniform groups API redesign proposed in [PR
14604](#14604)

1. Removed redundant `_group` suffix from class names
2. Used more descriptive function names that clearly indicate the
partitioning behaviour
3. United similar functionality to reduce cognitive overhead

  * fixed_size_group → chunk
  * tangle_group → tangle
  * ballot_group → fragment
  * opportunistic_group → merged into fragment
  * get_ballot_group() → binary_partition()
  * get_fixed_size_group<N>() → chunked_partition<N>()
  * get_tangle_group() → entangle()

---------

Signed-off-by: Larsen, Steffen <[email protected]>
Co-authored-by: Andrei Zibrov <[email protected]>
Co-authored-by: John Pennycook <[email protected]>
Co-authored-by: Udit Kumar Agarwal <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@rolandschulz rolandschulz rolandschulz left review comments

@gmlueck gmlueck gmlueck approved these changes

@AlexeySachkov AlexeySachkov AlexeySachkov approved these changes

@steffenlarsen steffenlarsen Awaiting requested review from steffenlarsen

Assignees

No one assigned

Labels

spec extension All issues/PRs related to extensions specifications

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@Pennycook @AlexeySachkov @rolandschulz @gmlueck