%% where the perf harness measures a frame. bench.exe times its own calls; MTL12_PERF times the layer's unix side in any program.
flowchart TB
    A["Windows program<br/>(bench.exe or an app)"] -- "record:<br/>Reset..Close" --> B["d3d12core.dll<br/>(PE side)<br/>records commands"]
    B -- "ExecuteCommandLists<br/>= submit" --> C["mtl12.so<br/>(unix side)<br/>replays into Metal"]
    C --> D["Metal 4<br/>command buffers"]
    D --> E["GPU"]
    E -. "timestamps around<br/>the list = gpu_us" .-> A
    C -. "MTL12_PERF: thread CPU time<br/>per call, by kind" .-> F["one JSON line<br/>per Present"]
    E -. "commit feedback:<br/>GPU start/end" .-> F
    A -. "QueryPerformanceCounter:<br/>record, submit, frame" .-> G["result: line"]
    F --> H["run script:<br/>median, p95, p99<br/>+ history file"]
    G --> H