/jetson-video-pipeline

Buffer sharing and synchronization contract

The verified sample routes exchange regular filesystem artifacts. A producer finishes before its consumer starts; the consumer reopens the canonical path and reauthenticates its size and SHA-256. The producer owns the artifact until successful process completion, and cleanup waits until the consumer finishes.

For every plan, result, and producer-to-consumer boundary, emit this object directly in the plan or evidence. Do not import or generate it with a runtime helper:

{
  "schema_version": "1.0",
  "transport": "filesystem-artifact",
  "memory_domain": "host-file",
  "copy_required": true,
  "zero_copy_verified": false,
  "allocation_owner": "producer",
  "consumer_access": "reopen-after-producer-completion",
  "synchronization": {
    "mode": "process-completion-and-artifact-reauthentication"
  },
  "verified_sharing_handles": [],
  "verified_sync_primitives": [],
  "not_established": [
    "nvsci-buf", "nvsci-sync", "dma-buf", "nvbufsurface", "egl-image",
    "cuda-external-memory", "cuda-ipc-memory", "cuda-event",
    "cuda-external-semaphore"
  ]
}

not_established is an evidence verdict, not an unsupported-product verdict. CPU-buffer or device-memory sample modes, API availability, header presence, and successful encode/decode do not establish external interoperability. A request for a listed non-file primitive remains unresolved unless an authenticated operation proves its handle, format, ownership, lifetime, signal, wait, and reuse rules. Never silently fall back or claim zero copy.

The traditional codec APIs expose CUDA ordering at specific boundaries:

• CUVIDPROCPARAMS::output_stream orders mapping and decoder-output post-processing performed by cuvidMapVideoFrame(). It is not an input dependency.
• NvEncSetIOCudaStreams() binds CUDA streams used to order encoder input and output processing. This alone does not prove output completion, external sharing, or zero copy.

cuvidParseVideoData() supplies CPU-addressable compressed bytes and cuvidDecodePicture() receives CUVIDPICPARAMS; neither accepts a CUDA stream or event to wait upon before input consumption. Preserve the documented input and metadata lifetimes. Report these APIs as capabilities until the exact in-process operation is authenticated; do not alter the filesystem contract.