Jetson Video Capability
Keep three authorities separate:
- API query: raw fields, never operation or product-support proof.
- Authenticated NVIDIA sample: report or exact operation evidence.
- Applicable NVIDIA documentation: product-support verdict.
Scope gates
- For a request solely for PSNR, SSIM, or other objective quality metrics, say
this skill does not provide them and a separately authorized quality workflow
is required; do nothing else.
- For Netflix, Widevine, PlayReady, or clearly content-protected streaming,
state only that this skill covers hardware encode/decode of user-supplied
non-DRM bitstreams, not content-DRM playback. Do not claim whether the
service works, describe Jetson certification, CDM, or secure-playback
requirements, or recommend a browser, DRM module, workaround, or bypass;
then stop. An unqualified “DRM” may mean Linux DRM/KMS; ask which meaning if
context does not resolve it.
- A bare “video SDK” is ambiguous: ask native Video Codec SDK,
PyNvVideoCodec, or both before probing. An otherwise unqualified capability
request uses the native-preferred fallback in
surface-selection-contract.md.
Read-only discovery
Capability work depends on jetson-video-setup for a fresh, read-only
installation check. Invoke that skill through public dispatch and consume its
reported exact native package/Samples root or exact PyNvVideoCodec interpreter,
version, and loaded module path. Do not locate or import setup's files. A
missing or mismatched surface is unknown/not_ready, never codec unsupported;
route repair to setup without mutating anything here.
Engine capability queries belong here, not in setup. For PyNvVideoCodec, use
the exact selected interpreter to call the public GetEncoderCaps and
GetDecoderCaps APIs as described in
capability-queries.md. A broad encoder
catalog covers H.264, HEVC, and AV1; a broad decoder catalog covers all ten
families across four chroma formats and three bit depths (120 exact tuples).
A bounded request queries only named members of the applicable catalog set.
Preserve every scoped record, error, and GPU ordinal. Nonzero-GPU helper results
remain unknown when the public helper selects only GPU 0.
For native reports, reuse authenticated package-owned binaries or build only
the required report target in a fresh user-owned tree, then run
AppEncCuda -ec and/or AppDec -dc using the build, identity, and grammar
rules in the capability reference. A query-only request does not authorize
package installation or an encode/decode operation. If the package, source,
tool, interpreter, or runtime-library identity cannot be established, report
the result unknown and name the missing setup prerequisite.
Decision flow
- Classify scope and selected surface before any probe.
- Query only the selected surface. Preserve raw records, errors, exact GPU,
release, interpreter/binary identities, argv, and evidence classification.
- Resolve the most exact live product identity using the device-tree paths and
NUL handling in
capability-queries.md,
then cross-check it against NVIDIA's current
Video Encode and Decode Support Matrix
and the release-matched SDK 13.0
NVENC
or NVDEC
application note. Never infer SKU from memory, engine count, capability
fields, or operation behavior. Preserve source URL, retrieval date, table
title, exact row/column labels, and cell value, using that reference's manual
capture record.
Generic family identity stays non-exact; use candidate-row consensus only
when every authenticated candidate agrees.
- Report documentation
No as the final unsupported product verdict even if
an API or diagnostic operation is positive; this ends the normal
availability check. Documentation Yes establishes documented support;
live availability additionally needs the matching authenticated operation.
Missing, unretrievable, or conflicting documentation remains unknown.
When documentation is unknown, do not present positive capability fields as
available options: label each affected codec unknown beside them and state
the retrieval failure with the verdict.
- For a PyNvVideoCodec encode-availability operation, request setup's
full-samples profile and carry its exact interpreter into the recipe and
pipeline stages. Do not select the smaller decode-performance/smoke profile.
- Run an exact operation only when the user requests live availability and
documentation is positive or unknown, or explicitly requests a diagnostic
despite a negative verdict. A diagnostic under documentation
No reports
only operation_verified or operation_failed for that exact tuple and
never changes the unsupported product verdict. A tuple is the exact surface,
GPU, codec, profile, chroma, bit depth, dimensions, input format, and control
set tested. Resolve one recipe with jetson-video-recipe, then use
jetson-video-pipeline for encode followed by independent decode of the
exact output identity. A failed tuple never generalizes to the product.
Evidence rules
GetEncoderCaps success is capability_reported, supported=null,
operation_status=not_tested.
GetDecoderCaps bIsSupported=1/0 records raw API true/false and whether
returned limits apply; neither value is the documentation verdict.
- Missing enums, calls, fields, prerequisites, or nonzero-GPU authority are
unknown, not unsupported.
- Native
-ec/-dc text is official_sample_report; preserve raw values and
never rewrite them as API or product claims.
- When reporting Main10 support, note that NVENC can convert verified 8-bit
input internally and that P010 is the exact no-input-bit-depth-conversion
path; a live claim still requires an authenticated 10-bit operation.
- AV1 output is operation evidence only after the pipeline's package-owned
decoder consumes the exact artifact and reports the expected frame count.
- For a VP9 encode question, report only that VP9 remains a decoder family and
no released NVENC route exists. Do not add a generic list of other encoders
or offer a diagnostic encode operation.
- Presets, tuning, package presence, throughput, and successful concurrent
streams do not establish codec support or NVENC/NVDEC engine count.
- Codec/API capability, NVENC/NVDEC availability, and successful bounded
operations do not establish DMA-BUF, NvSciBuf, CUDA-memory sharing, zero
copy, or cross-stage synchronization compatibility.
For a Python API query, create a short task-local program from this Markdown
and the installed public SDK, run it with the exact selected interpreter, and
keep its raw output with the result. Installation belongs to setup and
operation validation belongs to pipeline. This skill owns engine queries,
classification, documentation reconciliation, and the compact direct report
above.
1---2name: jetson-video-capability3license: "Apache-2.0"4description: >-5 Use when Jetson codec, profile, chroma, bit-depth, dimension, engine-count,6 or operational support must be reconciled from live APIs, authenticated7 NVIDIA samples, and NVIDIA documentation; also applies the content-DRM scope.8metadata:9 source: https://skillmd.com/skills/nvidia/jetson-video-capability10 author: "Vinit Bansal <vinitkumarb@nvidia.com>"11 tags: [jetson, video-codec-sdk, pynvvideocodec, nvenc, nvdec, capability]12 languages: [markdown]13 data-classification: public14---1516# Jetson Video Capability1718Keep three authorities separate:1920- API query: raw fields, never operation or product-support proof.21- Authenticated NVIDIA sample: report or exact operation evidence.22- Applicable NVIDIA documentation: product-support verdict.2324## Scope gates2526- For a request solely for PSNR, SSIM, or other objective quality metrics, say27 this skill does not provide them and a separately authorized quality workflow28 is required; do nothing else.29- For Netflix, Widevine, PlayReady, or clearly content-protected streaming,30 state only that this skill covers hardware encode/decode of user-supplied31 non-DRM bitstreams, not content-DRM playback. Do not claim whether the32 service works, describe Jetson certification, CDM, or secure-playback33 requirements, or recommend a browser, DRM module, workaround, or bypass;34 then stop. An unqualified “DRM” may mean Linux DRM/KMS; ask which meaning if35 context does not resolve it.36- A bare “video SDK” is ambiguous: ask native Video Codec SDK,37 PyNvVideoCodec, or both before probing. An otherwise unqualified capability38 request uses the native-preferred fallback in39 [surface-selection-contract.md](references/surface-selection-contract.md).4041## Read-only discovery4243Capability work depends on `jetson-video-setup` for a fresh, read-only44installation check. Invoke that skill through public dispatch and consume its45reported exact native package/Samples root or exact PyNvVideoCodec interpreter,46version, and loaded module path. Do not locate or import setup's files. A47missing or mismatched surface is `unknown`/`not_ready`, never codec unsupported;48route repair to setup without mutating anything here.4950Engine capability queries belong here, not in setup. For PyNvVideoCodec, use51the exact selected interpreter to call the public `GetEncoderCaps` and52`GetDecoderCaps` APIs as described in53[capability-queries.md](references/capability-queries.md). A broad encoder54catalog covers H.264, HEVC, and AV1; a broad decoder catalog covers all ten55families across four chroma formats and three bit depths (120 exact tuples).56A bounded request queries only named members of the applicable catalog set.57Preserve every scoped record, error, and GPU ordinal. Nonzero-GPU helper results58remain unknown when the public helper selects only GPU 0.5960For native reports, reuse authenticated package-owned binaries or build only61the required report target in a fresh user-owned tree, then run62`AppEncCuda -ec` and/or `AppDec -dc` using the build, identity, and grammar63rules in the capability reference. A query-only request does not authorize64package installation or an encode/decode operation. If the package, source,65tool, interpreter, or runtime-library identity cannot be established, report66the result `unknown` and name the missing setup prerequisite.6768## Decision flow69701. Classify scope and selected surface before any probe.712. Query only the selected surface. Preserve raw records, errors, exact GPU,72 release, interpreter/binary identities, argv, and evidence classification.733. Resolve the most exact live product identity using the device-tree paths and74 NUL handling in75 [capability-queries.md](references/capability-queries.md#documentation-reconciliation),76 then cross-check it against NVIDIA's current77 [Video Encode and Decode Support Matrix](https://developer.nvidia.com/video-encode-decode-support-matrix)78 and the release-matched SDK 13.079 [NVENC](https://docs.nvidia.com/video-technologies/video-codec-sdk/13.0/nvenc-application-note/index.html)80 or [NVDEC](https://docs.nvidia.com/video-technologies/video-codec-sdk/13.0/nvdec-application-note/index.html)81 application note. Never infer SKU from memory, engine count, capability82 fields, or operation behavior. Preserve source URL, retrieval date, table83 title, exact row/column labels, and cell value, using that reference's manual84 capture record.85 Generic family identity stays non-exact; use candidate-row consensus only86 when every authenticated candidate agrees.874. Report documentation `No` as the final unsupported product verdict even if88 an API or diagnostic operation is positive; this ends the normal89 availability check. Documentation `Yes` establishes documented support;90 live availability additionally needs the matching authenticated operation.91 Missing, unretrievable, or conflicting documentation remains unknown.92 When documentation is unknown, do not present positive capability fields as93 available options: label each affected codec `unknown` beside them and state94 the retrieval failure with the verdict.955. For a PyNvVideoCodec encode-availability operation, request setup's96 `full-samples` profile and carry its exact interpreter into the recipe and97 pipeline stages. Do not select the smaller decode-performance/smoke profile.986. Run an exact operation only when the user requests live availability and99 documentation is positive or unknown, or explicitly requests a diagnostic100 despite a negative verdict. A diagnostic under documentation `No` reports101 only `operation_verified` or `operation_failed` for that exact tuple and102 never changes the unsupported product verdict. A tuple is the exact surface,103 GPU, codec, profile, chroma, bit depth, dimensions, input format, and control104 set tested. Resolve one recipe with `jetson-video-recipe`, then use105 `jetson-video-pipeline` for encode followed by independent decode of the106 exact output identity. A failed tuple never generalizes to the product.107108## Evidence rules109110- `GetEncoderCaps` success is `capability_reported`, `supported=null`,111 `operation_status=not_tested`.112- `GetDecoderCaps` `bIsSupported=1/0` records raw API true/false and whether113 returned limits apply; neither value is the documentation verdict.114- Missing enums, calls, fields, prerequisites, or nonzero-GPU authority are115 unknown, not unsupported.116- Native `-ec/-dc` text is `official_sample_report`; preserve raw values and117 never rewrite them as API or product claims.118- When reporting Main10 support, note that NVENC can convert verified 8-bit119 input internally and that P010 is the exact no-input-bit-depth-conversion120 path; a live claim still requires an authenticated 10-bit operation.121- AV1 output is operation evidence only after the pipeline's package-owned122 decoder consumes the exact artifact and reports the expected frame count.123- For a VP9 encode question, report only that VP9 remains a decoder family and124 no released NVENC route exists. Do not add a generic list of other encoders125 or offer a diagnostic encode operation.126- Presets, tuning, package presence, throughput, and successful concurrent127 streams do not establish codec support or NVENC/NVDEC engine count.128- Codec/API capability, NVENC/NVDEC availability, and successful bounded129 operations do not establish DMA-BUF, NvSciBuf, CUDA-memory sharing, zero130 copy, or cross-stage synchronization compatibility.131132For a Python API query, create a short task-local program from this Markdown133and the installed public SDK, run it with the exact selected interpreter, and134keep its raw output with the result. Installation belongs to setup and135operation validation belongs to pipeline. This skill owns engine queries,136classification, documentation reconciliation, and the compact direct report137above.