codec-video-prep-legacy-exact 0.2.5.post4

Legacy-exact codec video preprocessing wheel

codec-video-prep-legacy-exact

This repository is a clean legacy-exact source tree for reproducing old codec preprocessing results. It is intentionally separate from the newer codec-video-prep package.

The legacy pipeline uses the old cv_reader.api.read_video_cb backend. It does not use cv_reader_fast, and this repository does not build or package cv_reader_fast.

Contents

codec_selector/              # legacy selector pipeline
src/cv_reader/               # old cv_reader Python package and C++ extension source
tool/                        # legacy scripts used by the CLI
ffmpeg/ffmpeg_patch/         # patched FFmpeg source files
ffmpeg/install_ffmpeg.sh     # downloads FFmpeg 5.1, applies patches, builds static libs
build_scripts/               # wheel build and verification helpers
codec_video_prep_legacy_exact_cli.py
setup.py
pyproject.toml

Generated files are intentionally excluded from the clean tree:

*.so
build/
dist/
*.egg-info/
ffmpeg/ffmpeg_install/
ffmpeg/ffmpeg_source/
__pycache__/

Build patched FFmpeg

Run once per build machine/container:

cd /video_vit/yunyaoyan/code/codec-video-prep-legacy-exact
bash build_scripts/build_patched_ffmpeg.sh

This downloads FFmpeg 5.1, copies files from ffmpeg/ffmpeg_patch/ into the FFmpeg source tree, and installs static FFmpeg libraries under:

ffmpeg/ffmpeg_install/

setup.py reads FFmpeg flags from:

ffmpeg/ffmpeg_install/lib/pkgconfig

Build wheels

Build for available local Python versions:

cd /video_vit/yunyaoyan/code/codec-video-prep-legacy-exact
bash build_scripts/build_wheels.sh

Choose Python versions:

PY_LIST="3.10 3.11 3.12" bash build_scripts/build_wheels.sh

Set version:

LEGACY_EXACT_VERSION=0.2.5.post4 bash build_scripts/build_wheels.sh

Set a platform tag when building inside a target platform/container:

PLAT_NAME=manylinux_2_35_x86_64 bash build_scripts/build_wheels.sh
PLAT_NAME=manylinux_2_39_aarch64 bash build_scripts/build_wheels.sh

For aarch64, run the same scripts inside an aarch64 container or machine after building patched FFmpeg there.

Verification

The wheel verifier runs automatically from build_wheels.sh. You can also run it manually:

python3 build_scripts/verify_legacy_wheel.py dist/*.whl

It checks:

• exactly one cv_reader/api*.so is present;
• no cv_reader_fast files are present;
• no unexpected cv_reader/*.so files are present.

Manual inspection:

unzip -l dist/codec_video_prep_legacy_exact-*.whl | grep -E 'cv_reader.*\.so|cv_reader_fast'

Expected output should include cv_reader/api...so and should not include cv_reader_fast.

Smoke test

python3.10 -m pip install --force-reinstall dist/codec_video_prep_legacy_exact-*-cp310-*.whl
cd /tmp
python3.10 - <<'PY'
from cv_reader import api
print(api.__file__)
PY
codec-video-prep-legacy-exact --help

Legacy behavior note

Do not replace cv_reader.api with the newer cv_reader_fast path if you need old src_patch_position.npy reproduction. The old exact behavior depends on cv_reader.api.read_video_cb and its historical frame/bitcost alignment behavior.