%% how one DispatchGraph runs on Metal (no GPU work graphs there), with the scheduler ported from vkd3d-proton.
flowchart TD
    A["CreateStateObject<br/>(EXECUTABLE, WORK_GRAPH)"] --> B["Program build on the CPU:<br/>sort nodes by ID and array index,<br/>place each on every level it can run at,<br/>unroll recursion"]
    B --> C["Each node translated once<br/>DXIL → SPIR-V → MSL;<br/>up to 4 Metal pipelines via<br/>function constants"]
    D["DispatchGraph<br/>(CPU or GPU records)"] --> E["Level 0: dispatch the entry<br/>node on the uploaded records"]
    C --> E
    E --> F["Nodes write records to a payload<br/>ring and count them per node<br/>(atomics in the backing memory)"]
    F --> G["wg_distribute: counts →<br/>threadgroups per node"]
    G --> H["wg_offsets: each record's<br/>payload offset, grouped by node"]
    H --> I["Level n: one indirect<br/>dispatch per node,<br/>reading the other ring"]
    I --> J{"more levels?"}
    J -- "yes: swap the rings" --> F
    J -- no --> K["Done: 3 dispatches<br/>and barriers per level"]