Skip to content

Benchmark kernel launch overhead and partitioning #774

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

Merged
vchuravy merged 2 commits into from
Sep 12, 2026
+70 −1
Merged

Benchmark kernel launch overhead and partitioning #774

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c270eaf
Benchmark kernel launch overhead and partitioning
giordano Sep 11, 2026
00c47a3
Report allocations in the PR benchmark comment
giordano Sep 11, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion .github/workflows/benchmark_pr.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,7 @@ jobs:
path: plots
- name: Create markdown table from benchmarks
run: |
benchpkgtable ${{ steps.extract-package-name.outputs.package_name }} --rev="${{github.event.repository.default_branch}},${{github.event.pull_request.head.sha}}" --input-dir=results/ --ratio > table.md
benchpkgtable ${{ steps.extract-package-name.outputs.package_name }} --rev="${{github.event.repository.default_branch}},${{github.event.pull_request.head.sha}}" --input-dir=results/ --ratio --mode=time,memory > table.md
echo '### Benchmark Results' > body.md
echo '' >> body.md
echo '<details>' >> body.md
Expand Down
69 changes: 69 additions & 0 deletions benchmark/benchmarks.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,3 +65,72 @@ let default = BenchmarkGroup()
end
SUITE["saxpy"]["default"] = default
end

# Launch overhead: a problem of a single workgroup, so that the time is dominated by the
# host-side work of a launch. The kernel is constructed in the setup and the backend
# synchronized in the teardown, so only the launch itself is measured.
@kernel function scale_kernel!(A, @Const(B))
I = @index(Global)
@inbounds A[I] = 2 * B[I]
end

@kernel function scale_kernel_3d!(A, @Const(B))
i, j, k = @index(Global, NTuple)
@inbounds A[i, j, k] = 2 * B[i, j, k]
end

SUITE["launch"] = BenchmarkGroup()

let launch = BenchmarkGroup(), n = 16, dims = (4, 4, 4)
launch["dynamic workgroup, dynamic ndrange"] = @benchmarkable kernel(A, B, ndrange = $n) setup = (
kernel = scale_kernel!($BACKEND);
A = KernelAbstractions.zeros($BACKEND, Float64, $n);
B = rand!(KernelAbstractions.zeros($BACKEND, Float64, $n))
) teardown = KernelAbstractions.synchronize($BACKEND)

launch["dynamic workgroup, dynamic ndrange, workgroupsize given"] = @benchmarkable kernel(A, B, ndrange = $n, workgroupsize = $n) setup = (
kernel = scale_kernel!($BACKEND);
A = KernelAbstractions.zeros($BACKEND, Float64, $n);
B = rand!(KernelAbstractions.zeros($BACKEND, Float64, $n))
) teardown = KernelAbstractions.synchronize($BACKEND)

launch["static workgroup, dynamic ndrange"] = @benchmarkable kernel(A, B, ndrange = $n) setup = (
kernel = scale_kernel!($BACKEND, $n);
A = KernelAbstractions.zeros($BACKEND, Float64, $n);
B = rand!(KernelAbstractions.zeros($BACKEND, Float64, $n))
) teardown = KernelAbstractions.synchronize($BACKEND)

launch["static workgroup, static ndrange"] = @benchmarkable kernel(A, B) setup = (
kernel = scale_kernel!($BACKEND, $n, $n);
A = KernelAbstractions.zeros($BACKEND, Float64, $n);
B = rand!(KernelAbstractions.zeros($BACKEND, Float64, $n))
) teardown = KernelAbstractions.synchronize($BACKEND)

launch["3D static workgroup, dynamic ndrange"] = @benchmarkable kernel(A, B, ndrange = $dims) setup = (
kernel = scale_kernel_3d!($BACKEND, $dims);
A = KernelAbstractions.zeros($BACKEND, Float64, $dims...);
B = rand!(KernelAbstractions.zeros($BACKEND, Float64, $dims...))
) teardown = KernelAbstractions.synchronize($BACKEND)

launch["3D static workgroup, static ndrange"] = @benchmarkable kernel(A, B) setup = (
kernel = scale_kernel_3d!($BACKEND, $dims, $dims);
A = KernelAbstractions.zeros($BACKEND, Float64, $dims...);
B = rand!(KernelAbstractions.zeros($BACKEND, Float64, $dims...))
) teardown = KernelAbstractions.synchronize($BACKEND)

SUITE["launch"] = launch
end

# Host-side partitioning of the iteration space, independent of the backend.
let partition = BenchmarkGroup(), ndrange = Ref((1024,)), workgroupsize = Ref((16,))
partition["dynamic workgroup, dynamic ndrange"] = @benchmarkable KernelAbstractions.partition(kernel, $ndrange[], $workgroupsize[]) setup = (
kernel = scale_kernel!($BACKEND)
)
partition["static workgroup, dynamic ndrange"] = @benchmarkable KernelAbstractions.partition(kernel, $ndrange[], nothing) setup = (
kernel = scale_kernel!($BACKEND, 16)
)
partition["static workgroup, static ndrange"] = @benchmarkable KernelAbstractions.partition(kernel, nothing, nothing) setup = (
kernel = scale_kernel!($BACKEND, 16, 1024)
)
SUITE["partition"] = partition
end
Loading