skills.rest
NVIDIA
NVIDIAOFFICIAL
Explore all skills by this author →
WebsiteGitHub

doca-compress

Guides DOCA Compress programming for hardware-accelerated DEFLATE and LZ4 decompression on BlueField DPUs.

3.2k|370|Updated Feb 25, 2026
Search Tags:#compression#Debugging#deflate#doca#bluefield#hardware offload#lz4
One-click install
npx skills add https://github.com/NVIDIA/skills --skill doca-compress
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: doca-compress
Source: https://github.com/NVIDIA/skills/tree/main/skills/doca-compress
Command: npx skills add https://github.com/NVIDIA/skills --skill doca-compress

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Developers offloading compression to NVIDIA BlueField DPUs or ConnectX NICs face a complex API surface: per-task capability queries, mmap permission matrices, buffer sizing against device limits, and layered DOCA_ERROR_* diagnostics. This Skill teaches AI agents the correct DOCA Compress workflows so they produce grounded, version-aware guidance instead of hallucinated API calls.

Core Features & Use Cases

  • Task configuration guidance: Walks through enabling compress-deflate, decompress-deflate, decompress-lz4-stream, and decompress-lz4-block tasks on a doca_compress context, including decompress-only as a valid standalone shape.
  • Capability discovery and sizing: Enforces runtime doca_compress_cap_* queries for task support and per-task maximum buffer sizes before any submission, plus the source/destination mmap permission matrix.
  • Offload path selection: Provides the size-threshold rule for deciding between accelerator offload and CPU zlib/zstd, and routes pure mmap-to-mmap copies to doca-dma.
  • Debug and validation workflows: Supplies a round-trip smoke test methodology, an iterative test loop, a five-phase debug loop, and a bounded rollback procedure.
  • Use Case: A developer asks their agent to "decompress incoming network data on the BlueField" — the agent loads this Skill, queries device capabilities, configures a decompress-only context with correct mmap permissions, and validates with a known DEFLATE fixture before bulk traffic.

Quick Start

Ask your agent to help you configure a DOCA Compress context to offload gzip-style DEFLATE compression of a large buffer on your BlueField DPU, including capability checks and a round-trip validation.

Frequently Asked Questions about doca-compress

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I offload gzip compression to a BlueField DPU with DOCA?▼

Create a doca_compress context, enable the doca_compress_task_compress_deflate task before doca_ctx_start, and size buffers against the cap query. DOCA Compress supports DEFLATE, the same algorithm behind gzip and zlib, on the encode side.

When should I use DOCA Compress instead of CPU zlib or zstd?▼

Use DOCA Compress for bulk inputs of a few KiB or larger, repeated operations, or data already pinned in doca_mmap memory. For tiny one-shot buffers, CPU compression with zlib or zstd is faster than the DMA-to-accelerator round-trip.

Does DOCA Compress support LZ4 compression?▼

DOCA Compress supports LZ4 only on the decompress side, via the lz4_stream and lz4_block task families, each independently capability-gated. LZ4 encode is not supported; use a CPU LZ4 library for outbound LZ4 encoding.

Can I use DOCA Compress only for decompression?▼

Yes, a decompress-only consumer is a fully valid standalone shape. Enable only doca_compress_task_decompress_deflate (or an LZ4 decode task) without enabling the compress task, and validate against a CPU-produced fixture.

Why does doca_compress task submission return DOCA_ERROR_INVALID_VALUE?▼

This usually means the source buffer exceeds the per-task max buffer size from the capability query, or the destination buffer is too small for worst-case output. Re-run the matching doca_compress_cap_*_get_max_buf_size query and resize.

What mmap permissions does DOCA Compress require?▼

The source mmap needs at least DOCA_ACCESS_FLAG_LOCAL_READ_ONLY and the destination mmap needs DOCA_ACCESS_FLAG_LOCAL_READ_WRITE, both set before the first task submission. Missing permissions surface as DOCA_ERROR_NOT_PERMITTED at submit time.

Related Skills

Explore More Software Engineering Skills→
NVIDIANVIDIA
14

k8s-network-engineer

Configure and troubleshoot NVIDIA networking deployments on Kubernetes.

Official
Advanced
coolsocketcoolsocket
1

cloud-run-benchmark

...

Community
Advanced
FernanSuozaFernanSuoza

embedded-code-development

Implement deterministic embedded firmware changes with explicit resource handling and unit tests.

Community
Advanced
pepegarpepegar

mermaid

Encode Mermaid diagram code into live.mermaid.com URLs and decode URLs back to code.

Community
Basic