KI Sub-groups by christiangnrd · Pull Request #668 · JuliaGPU/KernelAbstractions.jl

christiangnrd · 2025-12-23T01:44:16Z

No description provided.

github-actions · 2025-12-23T01:45:20Z

Your PR no longer requires formatting changes. Thank you for your contribution!

codecov · 2026-01-07T17:55:10Z

Codecov Report

❌ Patch coverage is 75.55556% with 11 lines in your changes missing coverage. Please review.
✅ Project coverage is 64.19%. Comparing base (9e79dd0) to head (be3a2e3).

Files with missing lines	Patch %	Lines
src/pocl/backend.jl	73.07%	7 Missing ⚠️
src/interface.jl	0.00%	3 Missing ⚠️
src/pocl/compiler/compilation.jl	88.88%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #668      +/-   ##
==========================================
+ Coverage   63.98%   64.19%   +0.20%     
==========================================
  Files          23       23              
  Lines        1966     2008      +42     
==========================================
+ Hits         1258     1289      +31     
- Misses        708      719      +11

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

christiangnrd · 2026-06-16T21:57:00Z

@maleadt @vchuravy This is ready for final review. If/once accepted I’ll merge the backend subgroup branches into their respective ka 0.10 branch and undo the new branch redirection before merging this one

maleadt · 2026-06-17T17:11:13Z

I think the ideal solution would be to have get_num_sub_groups always return at least 1 so we can force a sub-group size.

FWIW, I fixed that, and pocl_jll should include the patch merged upstream to make get_num_sub_groups behave sensibly.

christiangnrd · 2026-06-17T17:34:57Z

FWIW, I fixed that, and pocl_jll should include the patch merged upstream to make get_num_sub_groups behave sensibly.

I never got around to removing setting the subgroup size by default so no changes necessary.

Also, I bumped the required pocl_standalone_jll to 7.1.3, could/should we yank 7.1.3+0 to guarantee no segfaults from libpocl mismatches?

…overrides Adds KI.sub_group_ballot(pred::Bool) → UInt32: returns a bitmask with bit (lane-1) set for every lane where pred is true. CPU fallback returns UInt32(pred). SPIR-V/OpenCL has no ballot intrinsic so no POCL override is added. Also adds ext/CUDAExt.jl, the KA extension for CUDA, providing @device_override implementations of all sub-group intrinsics from JuliaGPU#668: - sub_group_size(::CUDABackend), shfl_down_types(::CUDABackend) [host-side] - get_sub_group_{size,max_size,local_id,id,num}() [device-side] - sub_group_barrier(), shfl_down(), sub_group_ballot() [device-side] Builds on top of JuliaGPU#668 (KI Sub-groups). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

vchuravy

Looks good from my side, van all current backends implement this?

christiangnrd · 2026-07-02T10:30:19Z

Looks good from my side, van all current backends implement this?

I have a branch that passes tests for every backend. I had to implement shuffle_down manually with shuffle for a couple of them.

I guess the newly supported cuda opencl is failing I’ll have to take a look.

vchuravy · 2026-07-02T13:49:41Z

    return config
 end
-@noinline function _compiler_config(dev; kernel = true, name = nothing, always_inline = false, kwargs...)
+@noinline function _compiler_config(dev; kernel = true, name = nothing, always_inline = false, sub_group_size::Union{Nothing, Int} = 32, kwargs...)


Did you try different defaults here?

I didn't. I just went with what is most common on the GPU side

Co-Authored-By: Anton Smirnov <tonysmn97@gmail.com>

christiangnrd · 2026-07-02T14:44:09Z

@vchuravy I added a supports_subgroup function to KA that defaults to true since it only really needs to be defined in OpenCL.jl. If you agree with that approach I'll revert CI pointing to my intrinsicsnew branches and merge this!

vchuravy · 2026-07-03T06:57:15Z

    return Int(device().max_work_group_size)
 end
+function KI.sub_group_size(::POCLBackend)::Int
+    sg_sizes = cl.device().sub_group_sizes


Can you add a comment explaining the logic here?

Suggested change

sg_sizes = cl.device().sub_group_sizes

# POCL can technically support any sub_group size.

# Check for common values used on GPUs then

# return 1 otherwise

sg_sizes = cl.device().sub_group_sizes

christiangnrd mentioned this pull request Jan 2, 2026

shfl_down intrinsic #661

Closed

christiangnrd force-pushed the subgroups branch from 6852410 to 84730d2 Compare January 2, 2026 22:38

christiangnrd marked this pull request as draft January 3, 2026 19:47

Hamiltonian-Action reviewed Jan 4, 2026

View reviewed changes

Comment thread test/interface.jl

christiangnrd force-pushed the subgroups branch 7 times, most recently from daea025 to 6343fd2 Compare January 7, 2026 16:55

christiangnrd force-pushed the subgroups branch from 6343fd2 to a86cc8c Compare January 14, 2026 01:31

christiangnrd force-pushed the subgroups branch from a86cc8c to 6f4a517 Compare February 2, 2026 17:44

christiangnrd force-pushed the subgroups branch 5 times, most recently from f0a545c to 8858825 Compare February 19, 2026 23:54

christiangnrd marked this pull request as ready for review February 20, 2026 00:13

christiangnrd requested a review from vchuravy February 20, 2026 00:13

christiangnrd force-pushed the subgroups branch 3 times, most recently from 7ddba7e to 8de2c42 Compare March 24, 2026 10:29

christiangnrd mentioned this pull request Mar 24, 2026

Skip benchmarking CI on forks #688

Merged

JuliaGPU deleted a comment from github-actions Bot Mar 24, 2026

christiangnrd force-pushed the subgroups branch from 8de2c42 to 82d32db Compare March 24, 2026 13:16

christiangnrd force-pushed the subgroups branch 2 times, most recently from 409a1e7 to 08a8130 Compare May 28, 2026 22:53

christiangnrd mentioned this pull request Jun 16, 2026

Restore a fixed subgroup size in kernels JuliaGPU/OpenCL.jl#442

Merged

christiangnrd force-pushed the subgroups branch from edfa0bb to aa5ddcf Compare June 16, 2026 19:36

christiangnrd added this to the 0.10.0 milestone Jun 18, 2026

christiangnrd force-pushed the subgroups branch from 56c597e to 8abe14f Compare June 29, 2026 17:52

christiangnrd force-pushed the subgroups branch from 8abe14f to 89f02da Compare July 1, 2026 14:06

vchuravy approved these changes Jul 2, 2026

View reviewed changes

christiangnrd force-pushed the subgroups branch 3 times, most recently from b08bc66 to 2a9d804 Compare July 2, 2026 13:15

vchuravy reviewed Jul 2, 2026

View reviewed changes

christiangnrd and others added 11 commits July 2, 2026 11:17

Initial subgroups support

acd5028

shfl_down intrinsics

1493056

Co-Authored-By: Anton Smirnov <tonysmn97@gmail.com>

All backends are synchronizing

4277409

Local backend support

36b852f

[Temp] Use new intrinsics feature branches in CI

195b160

Fixes to pocl backend in line with OpenCL fixes

dd4935d

Temp?

82e89f0

Fix

23dbfc0

Correct branch

8d8b726

Add supports_subgroups

ebbd840

Fix

14a2d61

christiangnrd force-pushed the subgroups branch from 9246927 to 14a2d61 Compare July 2, 2026 14:17

Format

be3a2e3

vchuravy reviewed Jul 3, 2026

View reviewed changes

-    sg_sizes = cl.device().sub_group_sizes
+    # POCL can technically support any sub_group size.
+    #  Check for common values used on GPUs then
+    #  return 1 otherwise
+    sg_sizes = cl.device().sub_group_sizes

Uh oh!

Conversation

christiangnrd commented Dec 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Dec 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

codecov Bot commented Jan 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

christiangnrd commented Jun 16, 2026

Uh oh!

maleadt commented Jun 17, 2026

Uh oh!

christiangnrd commented Jun 17, 2026

Uh oh!

vchuravy left a comment

Choose a reason for hiding this comment

Uh oh!

christiangnrd commented Jul 2, 2026

Uh oh!

vchuravy Jul 2, 2026

Choose a reason for hiding this comment

Uh oh!

christiangnrd Jul 2, 2026

Choose a reason for hiding this comment

Uh oh!

christiangnrd commented Jul 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

vchuravy Jul 3, 2026

Choose a reason for hiding this comment

Uh oh!

christiangnrd Jul 3, 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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

christiangnrd commented Dec 23, 2025 •

edited

Loading

github-actions Bot commented Dec 23, 2025 •

edited

Loading

codecov Bot commented Jan 7, 2026 •

edited

Loading

christiangnrd commented Jul 2, 2026 •

edited

Loading

christiangnrd Jul 3, 2026 •

edited

Loading