Graph Replay
graph_replay captures a compact GPU kernel sequence once, instantiates it as
an execution graph, then repeatedly replays that graph. It tests graph capture,
instantiation, launch, synchronization, and deterministic replay output.
Pantheon uses the portable HIP graph API. On NVIDIA, the CUDA build maps this to CUDA Graphs. On AMD ROCm, it uses HIP Graphs. If the installed driver or runtime cannot capture the sequence, the test exits with the underlying runtime error instead of reporting a proxy result.
What it exercises
- Stream capture of the kernel sequence.
- Execution graph instantiation and repeated graph launches.
- Launch and synchronization overhead under sustained replay.
- Input integrity and deterministic replay output with
--verify.
The reported graph-steps/s value is the rate of the captured sequence, not a
model inference or training throughput measurement.