跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
视频与音频

jetson-video-benchmark

Use when measuring Jetson Video Codec SDK or PyNvVideoCodec encode/decode throughput, comparing presets or surfaces, testing codec-worker capacity with authenticated samples and user media, or producing a documented clock-scaled or clock-and-resolution-scaled planning estimate when representative content is unavailable. Also use for Jetson video requests asking only for PSNR or SSIM results, to apply this performance skill's scope-only response.

NVIDIA-AI-IOT · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add NVIDIA-AI-IOT/jetson-device-skills --skill jetson-video-benchmark

技能说明


name: jetson-video-benchmark license: "Apache-2.0" description: >- Use when measuring Jetson Video Codec SDK or PyNvVideoCodec encode/decode throughput, comparing presets or surfaces, testing codec-worker capacity with authenticated samples and user media, or producing a documented clock-scaled or clock-and-resolution-scaled planning estimate when representative content is unavailable. Also use for Jetson video requests asking only for PSNR or SSIM results, to apply this performance skill's scope-only response. metadata: author: "Vinit Bansal vinitkumarb@nvidia.com" tags: [jetson, video-codec-sdk, pynvvideocodec, benchmark, nvenc, nvdec] languages: [python] data-classification: public

Jetson Video Benchmark

Purpose

Measure codec-stage FPS and megapixels/second on the current Jetson. Use this skill for encode or decode throughput, P4/P5 comparisons, native-versus-Python comparisons, and increasing-worker capacity tests. When content is unavailable, it can instead produce a clearly labeled SDK-documentation estimate for an exact supported 1080p table row and target maximum video clock. For another requested resolution, it may additionally apply the bounded pixel-area heuristic defined in the estimate reference. Never present either estimate as a target measurement.

Prerequisites

  • For a live measurement:
    • Run on the target Jetson with direct GPU access. A fresh validated nvcodec-environment identity from jetson-video-setup is optional; when supplied, it is authoritative and any invalid or stale identity fails without local fallback. The agent may obtain that identity from setup's public read-only probe; it need not be present in the customer's prompt.
    • Without that identity, native routes inspect only the installed nvidia-video-codec-sdk APT package and its package-owned official sample sources. PyNvVideoCodec routes require an authenticated setup environment or the caller's exact absolute pynvc_interpreter; never scan for a venv. Before asking the customer for that path, invoke setup's public probe when that skill is installed and inspect its typed result. These read-only checks install, repair, register, and smoke-test nothing. Only when neither authority is usable does an explicit pynvc or both request return input_required; local auto records PyNvVideoCodec as not_evaluated and may continue an eligible native branch.
    • A PyNvVideoCodec decode-only performance route can use a validated default pynvc-smoke environment. A Python encode or compare route uses official samples that import Torch and therefore requires a separately provisioned full-samples venv; never upgrade the smoke venv in place.
    • Live encode, compare, and encode-capacity routes require sibling jetson-video-recipe and an exact portable identity for one of its validated schema-2 recipes. If its public validator is absent, preserve dependency_required and its install-and-retry action. Decode routes do not require the recipe skill.
    • When setup is installed, read its shared video content policy before a live measurement. It does not apply to the separate documentation-only estimate path. Setup is not required solely for this policy: without it, require one exact user-selected path or URL, never choose catalog or synthetic media, and preserve source URL, license, attribution, path, size, and SHA-256.
    • Apply that policy's input gate before constructing a live benchmark dry run. Never choose media for the user or use the setup smoke fixture for performance.
  • A documentation answer may be produced off-target when the exact platform, SDK version, table conditions, and configured maximum video-clock facts are supplied with provenance. A clock-scaled estimate also requires a positive configured maximum video clock. It requires no media, recipe, sample authentication, or codec launch. Without that clock, report only the unscaled reference row as target_clock_unavailable, not a platform-scaled estimate.

Compose requested sibling stages

Documentation-only estimates and recipe-free live decode do not require a sibling skill. Live encode, compare, and encode-capacity routes require jetson-video-recipe; SDK installation, repair, a new full-samples Python environment, or one read-only handoff when registered Python authority is required belongs to jetson-video-setup. Check the agent's installed skill catalog before either stage. If the sibling is present, read its SKILL.md and invoke its documented public entry point; pass artifacts as data and never import sibling code. If it is absent, preserve completed input and measurement evidence and say, using the actual names: I can run <stage>, but it requires <skill>, which is not installed. Install <skill> and retry this stage. Never require setup when the caller already supplied authenticated setup evidence or an exact interpreter that passes local authentication, or a recipe for a decode-only or documentation-only request.

Instructions

  1. Apply the scope boundary first. For a request solely for objective quality metrics, including PSNR or SSIM, state only that this performance skill does not provide them and that a separately authorized quality workflow is required, then stop. Do not name or recommend an external tool, and do not offer to configure or run the comparison; do not request media, probe, install anything, or launch an operation.
  2. Classify an in-scope request as a live encode, decode, compare, or camera_capacity measurement, or as a documented_estimate answer. Resolve camera data direction before selecting an encode or decode row. Quality, preset, bitrate, rate-control, recording, or requested codec-output wording is an encode cue; lead with NVENC and mention deco
阅读完整来源 ↗