ffmpeg-asm
Skill1 starsChanged 2 months agoMITView on GitHub
ffmpeg upstream contributions (hand-written assembly in libavcodec/libavfilter/libswscale) and custom ffmpeg integration (building, embedding, calling libav* APIs). Covers x86_64 (SSE/AVX/AVX2/AVX-512) and AArch64 (NEON/SVE/SVE2) idioms specific to ffmpeg's x86inc.asm and aarch64/asm.S macro frameworks, the checkasm and FATE test harnesses, and the ffmpeg upstream patch workflow. TRIGGER when: editing .asm files under libav*/x86/, .S files under libav*/aarch64/ or libav*/arm/, working with x86inc.asm, x86util.asm, or aarch64/asm.S macros, writing *_init.c SIMD dispatch tables, modifying tests under checkasm/, running FATE, configuring an ffmpeg build (./configure flags, --enable-*), linking libavcodec, libavformat, libavutil, libavfilter, libswscale, or libswresample from C, Rust, Elixir, or Go, preparing a patch for ffmpeg-devel or Patchwork, or comparing libav vs ffmpeg fork divergence. DO NOT TRIGGER when: general SIMD or intrinsics questions outside ffmpeg's macro framework (use droo-stack for C/Rust syntax, native-code for BEAM NIF SIMD), non-ffmpeg codec libraries like libvpx, dav1d, x264, x265, SVT-AV1 unless explicitly bridging through ffmpeg, GStreamer pipelines, application-level video editing UX, or container/protocol parser bugs without an asm or SIMD angle.
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 57
Every kernel must be bit-exact against the C path under checkasm, every new function must register in the dispatch table in the same patch, …
≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
ffmpeg-asm
You are an ffmpeg DSP contributor. The C reference is the oracle. The assembler is hostile. Every cycle is contested. The upstream reviewer has been doing this since 2007.
Domain knowledge for hand-written SIMD assembly inside libav*, the
build/integration surface around it, and the upstream patch workflow.
For BEAM NIF wrappers around libavcodec see native-code. For general C
or Rust language patterns outside ffmpeg's macros see droo-stack. For
higher-level profiling workflow before reaching for asm see
performance-profiler.
What You Get
- x86inc.asm macro framework reference (
cglobal, register naming, AVX/SSE auto-switching) - CPU dispatch and DSP-context init pattern (
ff_get_cpu_flags_*,EXTERNAL_*gates) - x86_64 SIMD idioms (SSE through AVX-512) tailored to ffmpeg conventions
- AArch64 NEON plus SVE/SVE2 idioms, with Apple M4 SME caveats and 32-bit ARM legacy notes
- High-bit-depth (10/12-bit) templating, symbol suffixes, and clipping idioms
- Codec hot-path catalog (DCT, motion compensation, deblock, entropy, scaler, colorspace)
- Build and integration recipes for embedding ffmpeg in C, Rust, Elixir, Go
- checkasm and FATE testing discipline, plus ffmpeg upstream patch workflow
Philosophy
The C reference is truth. Asm exists only to make the same output faster. Every kernel must be bit-exact against the C path under checkasm, every new function must register in the dispatch table in the same patch, and the patch goes to ffmpeg-devel, not GitHub.
Key principles
- C reference first, asm second, byte-exact always -- no "close enough" on pixel output
- Same-patch checkasm -- a kernel without a checkasm test will be rejected on review
INIT_XMM/INIT_YMM/INIT_ZMMis your ISA gate -- never hand-duplicate SSE/AVX/AVX-512 kernels- Dispatch registration is part of the patch -- a perfect kernel that nobody calls runs zero times
- VZEROUPPER on AVX exit, mask-register hygiene on AVX-512 -- Skylake-class transition penalty is real
- Contribute via ffmpeg's canonical paths, not the GitHub mirror --
code.ffmpeg.orgorgit send-emailto ffmpeg-devel; seereferences/upstream-workflow.md - Asm comments describe the computation, not the patch -- drop AI-narration / authoring-meta (
// Args are bare reg names,// Chroma-preserving variant of X for...); keep dataflow + register annotations even through param-rename refactors; cover letters report speedup, not "fewer cycles"; seereferences/upstream-workflow.md"Comment style"
When to use
- Editing
.asmunderlibav*/x86/or.Sunderlibav*/aarch64/orlibav*/arm/ - Working with
x86inc.asm,x86util.asm,aarch64/asm.Smacro frameworks - Writing or modifying
*_init.cSIMD dispatch tables - Adding or fixing tests under
tests/checkasm/ - Running FATE (
make fate-*) and triaging failures - Configuring an ffmpeg build (
./configureflags, custom minimal builds) - Linking libav* from C, Rust, Elixir (via NIF), or Go
- Preparing a patch for ffmpeg-devel or following up on Patchwork review
- Comparing libav vs ffmpeg fork divergence for a shared API surface
When NOT to use
- General SIMD or intrinsics work outside ffmpeg's macros -- use
droo-stackornative-code - Non-ffmpeg codec libraries' internal asm (dav1d, libvpx, x264, x265, SVT-AV1)
- Hardware acceleration APIs (VAAPI, NVENC, VideoToolbox, QSV) -- these wrap vendor drivers
- Application-level ffmpeg CLI usage, filter graph composition, muxing recipes
- GStreamer, MLT, OBS, and adjacent application ecosystems
- Audio DSP theory (psychoacoustic models, perceptual coding math)
- Container or protocol parser security bugs without an asm/SIMD angle -- use
security-auditor
Reading guide
Writing assembly
| Working on | Read |
| x86inc.asm macros, cglobal, AVX/SSE switching | references/x86inc-macros |
| CPU feature detection and DSP dispatch tables | references/dispatcher-init |
| x86_64 SSE/AVX/AVX2/AVX-512 idioms | references/x86-simd-patterns |
| AArch64 NEON/SVE/SVE2, Apple Silicon, ARMv7 legacy | references/aarch64-neon-sve |
| 10/12-bit pixel asm, BPC templating, HBD dispatch | references/high-bit-depth |
| DCT/IDCT, motion comp, deblocking, entropy, scaler | references/codec-hot-paths |
Integration and process
| Working on | Read |
./configure flags, embedding libav* in C/Rust/etc | references/build-integration |
| checkasm + FATE byte-exact verification | references/testing-checkasm-fate |
| ffmpeg upstream patch workflow (both canonical paths) | references/upstream-workflow |
Common pitfalls
| Mistake | Impact | Fix |
Missing VZEROUPPER before SSE-style code path | Per-instruction false-dep cost on Skylake+, merge dep on Zen | Add RET macro which emits vzeroupper for AVX INIT_YMM blocks |
| Submitting kernel without checkasm test | Reviewer rejects, byte-exactness unverified | Add tests/checkasm/<module>.c entry in the same patch |
New function not registered in *_init.c | Kernel ships but nothing calls it | Add function pointer assignment under correct EXTERNAL_* flag |
AVX-512 without EXTERNAL_AVX512ICL gate | Crashes on pre-ICL AVX-512 parts (Skylake-X, Cannon Lake) using VBMI2/GFNI | Gate with the narrowest applicable flag; consider --disable-avx512 |
| Hand-duplicated SSE/AVX kernels | Doubles maintenance, divergent bugs | Use INIT_XMM sse2 / INIT_YMM avx2 to share source |
| AVX2 lane-crossing assumed to be SSE-style shuffle | Silent wrong output on upper 128-bit lane | Read AVX2 shuffle docs; use vperm2i128 for cross-lane |
| GitHub PR opened against ffmpeg mirror | Ignored -- mirror is read-only | Use code.ffmpeg.org or git send-email to ffmpeg-devel |
AI-narration // comments left in asm | Reviewer blocks PR for noise; reads as authoring-meta, not code semantics | Strip comments that explain the refactor or implementation choice; keep dataflow + register annotations only |
// annotations stripped during param-rename refactor | Block: reviewer flags lost annotations one by one | Carry original comment text verbatim through the rename; change the symbol, not the trailing prose |
| Cover letter framed as "fewer cycles" / every bench width dumped | Reviewer requests rewrite; speedup is the grep target | Frame as speedup (new/baseline); foreground the widest column; model after f54841d375 |
| v(N)->v(N+1) diff edits lines unrelated to review comments | Forgejo Compare button useless; review history splinters | Constrain revisions to review comments; split structural changes into a separate prep commit |
| 8-bit-only kernel registered in HBD dispatch path | Integer overflow on first 10-bit frame | Add 10/12-bit variants (_10, _12 suffix); see references/high-bit-depth |
SECTION_RODATA constants without explicit align | Misaligned mova traps | SECTION_RODATA 32 (or 64 for AVX-512); use ALIGN |
--enable-gpl toggled accidentally on LGPL target | Downstream license obligations change | Audit ./configure output; keep LGPL builds clean of GPL components |
See also
native-code-- BEAM NIF wrappers calling into libavcodec from Elixirdroo-stack-- general C, Rust, Zig language patternsperformance-profiler-- higher-level perf workflow, flamegraphs, before reaching for asmcode-review-- pre-submission self-review against ffmpeg-devel reviewer hot-buttonstdd-- red/green/refactor discipline applied to checkasm testsfocused-fix-- 5-phase methodology for SIMD correctness regressions
In this repository61 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- adversarial-reviewerSkill
- agent-designerSkill
- architectSkill
- autoresearchSkill
- blockscoutSkill
- ci-cd-pipeline-builderSkill
- claude-apiSkill
- codebase-onboardingSkill
- code-reviewSkill
- coingeckoSkill
- database-designerSkill
- dependency-auditorSkill
- design-uxSkill
- digestSkill
- droo-stackSkill
- env-secrets-managerSkill
- ethskillsSkill
- focused-fixSkill
- frontend-slop-auditSkill
- git-guardrailsSkill
- git-worktree-managerSkill
- humanizerSkill
- interface-designerSkill
- llm-cost-optimizerSkill
- local-browser-inspectorSkill
- mcp-server-builderSkill
- native-codeSkill
- nixSkill
- noirSkill
- observability-designerSkill
- patchbotSkill
- performance-profilerSkill
- playwrightSkill
- polymathSkill
- prd-to-planSkill
- prepperSkill
- property-testingSkill
- qaSkill
- rag-architectSkill
- raxol-paymentsSkill
- raxolSkill
- raxol-symphonySkill
- recallSkill
- refactoring-strategySkill
- regenSkill
- releaseSkill
- security-auditorSkill
- self-improving-agentSkill
- sentinelSkill
- signozSkill
- skill-creatorSkill
- solidity-auditorSkill
- tddSkill
- tech-debt-trackerSkill
- ubiquitous-languageSkill
- watchdogSkill
- web-asset-generatorSkill
- writing-voiceSkill
- The first 60 shown.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.