[albert-de-montserrat]

Since v0.16.1, @parallel <kernelcall> reads the kernel's metadata (nb_parallel_indices, memopt) at launch. The name of the metadata function was derived from the callee expression, so the lookup only works when the kernel is called by its own name. Calling a kernel through a struct field or a function argument fails:

struct Foo{F}; f::F; end
bar(s::Foo, A) = @parallel s.f(A)  # FieldError: type Foo has no field `##META, ParallelStencil#…, f`
g(f, A) = @parallel f(A)           # UndefVarError: `##META, ParallelStencil#…, f` not defined

Change

The metadata is now stored as an additional method of the kernel function itself,

@inline kernel(::ParallelStencil.KernelMetadata, args...) = <metadata module>

and the call site emits callee(ParallelStencil.KernelMetadata(), args...). Any callee expression that evaluates to the kernel (f, Mod.f, s.f, a function argument) thus finds the metadata by dispatch, in whatever scope the kernel is defined (including let blocks and function bodies). This also covers module-qualified calls (#224) and replaces the special case added for them in #225. The lookup still returns a constant module, so with const metadata memopt and nb_parallel_indices remain constant-folded.

The metadata method is defined after the kernel, as KernelAbstractions' @kernel defines the kernel constructors only if the kernel name is not yet bound.

test/runtests.jl: the stderr filter for metadata method-overwrite warnings is updated to the new method signature.

Tests

• New testset in test/test_parallel.jl ("6. metadata"): @metadata and @parallel called through a NamedTuple field and through a function argument.
• The MWE from Issue with functions inside structures #234 (with the backend swapped to Threads, and to Metal with Float32) errors on main and gives the correct result with this change.
• Pkg.test() passes locally (Julia 1.13, macOS aarch64) with Threads, Polyester, KernelAbstractions (CPU) and Metal. CUDA and AMDGPU were not tested locally.

Fixes #234

🤖 Generated with Claude Code

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.00%. Comparing base (6cf80ee) to head (532d2e9).
⚠️ Report is 8 commits behind head on main.
❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

☔ 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.

[omlins]

@luraess Why does it not trigger now the GPU CI?

[luraess]

@luraess Why does it not trigger now the GPU CI?

Ah, I forgot, CSCS CI cannot run from forks. So to have it running one would need to rebase on a branch from PS's main.

[omlins]

@albert-de-montserrat : thanks for the PR! There are unit tests that are looking for the old metadata functions (e.g. with ###META.) - this needs to be adapted in order to verify that the new metadata methods were created instead

[omlins]

There are unit tests that are looking for the old metadata functions (e.g. with ###META.) - this needs to be adapted in order to verify that the new metadata methods were created instead

[albert-de-montserrat]

Unit tests are passing now :)