habemus-papadum-vtenc 0.3.0

Zero-PyAV macOS H.264 encode (Apple VideoToolbox) for pdum.rfb

habemus-papadum-vtenc (import pdum.vtenc)

macOS host NV12 → H.264 Annex B via Apple's VideoToolbox (VTCompressionSession), with no PyAV and no ffmpeg. The companion encoder for pdum.rfb (PyPI: habemus-papadum-rfb) on Apple Silicon — the counterpart of habemus-papadum-nvenc on NVIDIA. A uv workspace member of this repo. Design notes: docs/mlx_metal_videotoolbox_encoder_design.md.

Why it exists:

1. Hardware H.264 on macOS without PyAV. VideoToolbox is the Apple-Silicon hardware encoder; this binds it directly, so the GPU path needs no ffmpeg layer.
2. MLX-friendly. Its encode() takes any Python buffer-protocol object, so an evaluated MLX mx.array (Apple-Silicon unified memory) feeds it directly.

What's ours

Everything is ours — there is no vendored SDK (VideoToolbox/CoreVideo/CoreMedia are macOS system frameworks):

src/cpp/vtenc_ext.mm        OURS — the only native code; a thin pybind11 binding over
                            VTCompressionSession (Objective-C++).
src/pdum/vtenc/__init__.py  OURS — Python surface + single-extension loader.
CMakeLists.txt              OURS — pybind11 3.0.4; -framework links; one _vtenc module.
build-wheel.sh              OURS — self-contained wheel build (delocate).

Behaviour (matches the pdum.rfb invariants)

• NV12 in → H.264 Annex B out (start codes, in-band SPS/PPS on every IDR — what the browser's WebCodecs VideoDecoder wants).
• Low-latency, no frame reordering (no B-frames ⇒ output order == input order) and synchronous 1-in-1-out: each encode() returns its own frame's access unit (CompleteFrames after each submit) — required for correct seq attribution.
• BT.601 limited range VUI (matches pdum.rfb's gpu.rgb_to_nv12 kernel), so a browser decodes the color correctly.
• Fixed-resolution, even dimensions; one VTCompressionSession per instance.

Usage

import numpy as np
from pdum.vtenc import VtEncoder

enc = VtEncoder(1920, 1080, fps=30, bitrate=12_000_000)
nv12 = np.zeros((1080 * 3 // 2, 1920), dtype=np.uint8)   # contiguous NV12 (Y then UV)
# ... fill nv12 (e.g. from an evaluated MLX array) ...
annexb = enc.encode(nv12, force_idr=True)                # bytes; H.264 Annex B
annexb += enc.flush()
print(enc.codec_string)                                  # e.g. "avc1.420028" (from the SPS)
enc.close()

encode() accepts any contiguous (H*3//2, W) uint8 buffer-protocol object — numpy or an evaluated MLX mx.array (call mx.eval(frame) first; MLX is lazy).

VtEncoder.codec_string is the avc1.PPCCLL string derived from the actual emitted SPS (VideoToolbox picks the level from the resolution, so it is not a constant — 1080p Baseline is avc1.420028, not avc1.42E01F). Empty until the first keyframe.

Build & test (local, CMake)

cmake -S . -B build -G Ninja
cmake --build build -j

Build wheels (maintainer)

./build-wheel.sh                                 # cp314 -> dist/habemus_papadum_vtenc-*.whl
PYTHON_VERSIONS="3.12 3.13 3.14" ./build-wheel.sh

Requires only Xcode Command Line Tools (clang + the macOS SDK frameworks); the full Metal toolchain is not needed for v1. The wheel bundles nothing beyond the extension — the frameworks come from macOS, as they must. Publishing to PyPI is done by scripts/publish.sh, not from CI.

Scope / caveats

• Fixed-resolution NV12 in, Annex B out, one encoder per instance. H.264 only (HEVC is a follow-up). No EncoderBackend/serve() wiring yet — that's the pdum.rfb integration.
• Input is a host-visible (CPU / unified-memory) NV12 buffer, memcpy'd into an encoder-owned CVPixelBuffer. Wrapping an MLX unified-memory buffer as the CVPixelBuffer backing directly (true zero-copy) is a follow-up.