sham::kernel_call#

sham::kernel_call submits a SYCL kernel over n threads. Inputs and outputs are wrapped in MultiRef; buffer access (get_read_access / get_write_access) and event completion are handled automatically.

Header: shambackends/kernel_call.hpp.

Basic usage#

C++ cannot expand two parameter packs, so inputs and outputs are passed as MultiRef:

sham::kernel_call(
    queue,
    sham::MultiRef{/* inputs */},
    sham::MultiRef{/* outputs (in-out) */},
    n,   // thread count
    [](u32 i, /* input ptrs */, /* output ptrs */) {
        // ...
    });

Minimal example:

sham::kernel_call(
    q,
    sham::MultiRef{buf_in},
    sham::MultiRef{buf_out},
    n,
    [](u32 i, const T *in, T *out) {
        out[i] = in[i];
    });

  • Functor first argument is the index (u32); then pointers in MultiRef order (inputs, then outputs)

  • Read buffers → const T*; write buffers → T*

  • Prefer const sham::DeviceBuffer& for pure inputs

MultiRef#

MultiRef holds references to buffer-like objects passed to kernel_call. Each member is what the kernel accesses — it must provide get_read_access, get_write_access, and complete_event_state (e.g. DeviceBuffer<T> or a custom accessor).

Under the hood#

  • get_read_access / get_write_access on each MultiRef member

  • launch n threads

  • complete_event_state after the kernel