Ffmpeg asm

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

1. C reference first, asm second, byte-exact always -- no "close enough" on pixel output
2. Same-patch checkasm -- a kernel without a checkasm test will be rejected on review
3. INIT_XMM/INIT_YMM/INIT_ZMM is your ISA gate -- never hand-duplicate SSE/AVX/AVX-512 kernels
4. Dispatch registration is part of the patch -- a perfect kernel that nobody calls runs zero times
5. VZEROUPPER on AVX exit, mask-register hygiene on AVX-512 -- Skylake-class transition penalty is real
6. Contribute via ffmpeg's canonical paths, not the GitHub mirror -- code.ffmpeg.org or git send-email to ffmpeg-devel; see references/upstream-workflow.md
7. 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"; see references/upstream-workflow.md "Comment style"

When to use

• Editing .asm under libav*/x86/ or .S under libav*/aarch64/ or libav*/arm/
• Working with x86inc.asm, x86util.asm, aarch64/asm.S macro frameworks
• Writing or modifying *_init.c SIMD dispatch tables
• Adding or fixing tests under tests/checkasm/
• Running FATE (make fate-*) and triaging failures
• Configuring an ffmpeg build (./configure flags, 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-stack or native-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 Elixir
• droo-stack -- general C, Rust, Zig language patterns
• performance-profiler -- higher-level perf workflow, flamegraphs, before reaching for asm
• code-review -- pre-submission self-review against ffmpeg-devel reviewer hot-buttons
• tdd -- red/green/refactor discipline applied to checkasm tests
• focused-fix -- 5-phase methodology for SIMD correctness regressions