Expand description
The #[kernel] attribute — write a hydroplane kernel as a plain generic function.
From a function whose first parameter is the SIMD context, #[kernel] generates the kernel
struct and Kernel impl (MatrixKernel with #[kernel(matrix)], adding the .tiles()
surface), a callable wrapper that runs dispatch for you, and a <name>_on(ctx, …) companion
that runs the body on a caller-supplied context — call it from another kernel to skip dispatch.
#[hydroplane::kernel]
fn any_overlap<'a, T: Scalar>(ctx: Gang, soa: &'a Soa<T>, q: [T; 4]) -> bool { /* body */ }
let hit = any_overlap(&soa, q); // call site — no struct, no impl, no `dispatch`The context is written bare (ctx: Gang) or concrete (ctx: Gang<f32>); every later parameter
becomes a struct field carried across the dispatch boundary, so borrows need explicit
lifetimes. The dispatch element is the generic bound by Scalar/FloatScalar/IntScalar (or
the one named T), inferred for concrete kernels; attribute overrides: scalar = U, extra
element names (OR’d into the dispatch type-combo), and unroll = N, which caps the ILP unroll
factor for this kernel. tiny and noalias are still accepted but no
longer change the generated code: the body always runs in an inlined _on.
Inside the body, a * b + c and c + a * b on varyings compile to a fused multiply-add, as
.fma would; on anything else they stay the two operators.
Attribute Macros§
- kernel
- Generate a
Kernel(or, with#[kernel(matrix)], aMatrixKernel) and a dispatching wrapper from one annotated function. See the crate docs for the full shape.