Macro offload_kernel

The offload_kernel macro is applied to a function to generate two separate definitions: a host-side wrapper for dispatch and a device-side kernel.

The macro does not perform the offload itself. It generates the necessary code required by the compiler's offloading infrastructure.

Usage example:

#[offload_kernel]
fn foo(a: &[f32], b: &[f32], c: *mut f32) {
    *c = a[0] + b[0];
}

This expands to the host-side function:

#[unsafe(no_mangle)]
#[inline(never)]
fn foo(_: &[f32], _: &[f32], _: *mut f32) {
    ::core::panicking::panic("not implemented")
}

And the device-side kernel:

#[rustc_offload_kernel]
#[unsafe(no_mangle)]
unsafe extern "gpu-kernel" fn foo(a: &[f32], b: &[f32], c: *mut f32) {
    *c = a[0] + b[0];
}