markus-perl/ffmpeg-build-script v9.0.4

9.0.4

on GitHub

Builds FFmpeg 9.0. Four new libraries, two new options, and one fix that affects every
default build on Linux.

Default builds on Linux had no TLS at all

build_gnutls built the whole gmp/nettle/gnutls chain and then never passed
--enable-gnutls — the line had been commented out. Neither gnutls nor openssl is in
FFmpeg's EXTERNAL_AUTODETECT_LIBRARY_LIST, so nothing filled the gap: an LGPL build could
not open an https:// URL. macOS was unaffected, because FFmpeg autodetects
SecureTransport there. Fixed, and the TLS backend is now a choice
(#178):

./build-ffmpeg --build --tls=gnutls     # default without --enable-gpl-and-non-free
./build-ffmpeg --build --tls=openssl    # default with it

FFmpeg refuses both at once, so exactly one is built. Under --tls=gnutls, libssh and libsrt
are skipped and the script says so before the build starts: libssh has no GnuTLS backend, and
libsrt's uses nettle's struct aes_ctx, removed in nettle 4.0.

GnuTLS also builds on Apple Silicon again — the skip dated from 3.6.16 and the pinned 3.8.13
builds fine.

New libraries

• libvvenc — H.266/VVC encoding
  (#179). FFmpeg's native VVC
  decoder was already there; this adds the encoder. Encoder name libvvenc.

• whisper.cpp — speech recognition through the af_whisper filter
  (#251). Opt-in, because
  exactly one ggml backend is compiled into a static binary and it has to match the machine
  that runs it:

  ./build-ffmpeg --build --whisper=cpu      # portable
  ./build-ffmpeg --build --whisper=metal    # macOS
  ./build-ffmpeg --build --whisper=cuda     # Linux, needs nvcc
  ./build-ffmpeg --build --whisper=vulkan   # Linux, needs glslc and a Vulkan loader

  No models are shipped; af_whisper takes model=/path at runtime. The cuda and
  vulkan backends have not been built or linked by anyone yet — reports welcome.

• libmysofa — the sofalizer filter, for binaural rendering
  (#219).

Leaving libraries out

--disable-lv2 and SKIPRAV1E=yes are gone, replaced by one option that takes any package
name --list-packages prints:

./build-ffmpeg --build --disable=rav1e,x265
./build-ffmpeg --build --disable=lv2 --disable=vulkan

lv2, vulkan, opencl and bluray stand for the whole cluster behind them. Disabling a
package also disables what cannot build without it and says so, so --disable=libogg takes
libvorbis and libtheora with it. Build tools and the TLS stack cannot be disabled; they are
marked always built in --list-packages. Names are checked before anything is built.

CUDA on RTX 50xx

CUDA_COMPUTE_CAPABILITY defaulted to a fixed 75 or 52, below Blackwell's 120, so an RTX 50xx
could not use its AV1 decoder
(#252). The capability is now
read from the installed GPU with nvidia-smi and checked against the CUDA toolkit by
compiling for it; the old defaults remain the fallback. Setting the variable yourself still
wins. Raising the default instead was not an option — FFmpeg cannot produce a
multi-architecture CUDA build, so one fixed value always breaks somebody.

Other changes

• --update replaces the script in place with the latest release. It refuses to run in a
  git checkout, verifies the archive before deleting anything, and leaves packages/ and
  workspace/ alone, so only packages whose version moved are rebuilt.
• --latest is removed. Its only reader was the branch that skipped an outdated package;
  passing it now fails as an unknown option instead of being silently ignored.
• A package whose pinned version changed is always rebuilt. Previously that needed
  --latest, so the default was to keep linking FFmpeg against a library the script no longer
  pinned.
• Every version pin and checksum is back in one table, src/10-versions.sh.
• CI checks that the option list in the README still matches --help.

Upgrading

--disable-lv2 → --disable=lv2, SKIPRAV1E=yes → --disable=rav1e, and --latest should
be dropped. Everything else is compatible.