Files
JustOneCacophony/README.en.md
T
TheM14 fd842db26f
ci / ubuntu-latest (push) Failing after 11s
ci / macos-latest (push) Canceled after 0s
ci / windows-latest (push) Canceled after 0s
ci / release (push) Canceled after 0s
Render SOFA HRIRs directly and place bed channels by their labels
Two changes that belong to the same reproduction model: a bed channel has
one position, defined by its label, and a SOFA set keeps the impulse
responses it was measured with.

CBA bed channels: the position comes from the channel's label in ETSI
TS 103 420 Table B.10, not from the layout the decoder is asked to render
to. The per-layout anchor table displaced bed channels onto the nearest
target speaker, which silently moved content whenever the target layout
changed.

- Drop the anchor table and the geometry lookup; bed_channel_q15 is the
  only source of bed positions, so stream and task no longer need a
  layout to ask for one
- Keep the 17-bit non-standard bed mask in channel order, as Table 13 /
  TS 103 190-2 Table 65 transmit it
- Exclude LFE channels from the object list while keeping them in the
  coverage mask
- Cover the Table B.10 coordinates, the q15 encoding, the 10-bit and
  17-bit masks and the LFE split in tests

SOFA rendering: compiling a SOFA set into a fifth-order directional field
threw away the measured impulse response - the SH projection, the
64-QMF/77-hybrid round trip and the synthetic room replaced it. The full
IRs are now kept and rendered as such, with the reproduction defined by a
configuration that names a personalized model rather than by the renderer
itself.

- Add SphericalMesh: incremental convex hull over the measurement
  directions, barycentric weights inside the containing triangle and
  spherical Voronoi areas for the centre spectrum
- Add FullHrirField: rfft of the selected shell, 200-1500 Hz phase fit for
  the delay, delay-aligned spectra, log-magnitude/phase or MCA
  interpolation, rigid-sphere radial ratio, recording origin and the
  common FFT/2 latency
- Add Uniform/MultiSourcePartitionedConvolver: block 128, numpy FFT
  conventions, zero-fade filter rebuild from the blocks in flight
- Move the Rosella reference arithmetic (profile geometry, direction
  basis, direct gains and room send, table-A room) out of the renderer
  into rosella_reference so both renderers share it
- Add ReferenceEnvironment and SofaReferenceRuntime: JSON configuration
  with model hash and distance-profile mapping, level calibration over
  equal-area directions, the reference room, 1536-sample batching and a
  single latency removal
- Keep .jochrtf an explicit input and the SOFA compile step behind
  --hrtf-cache-policy/--hrtf-cache-dir; add --reference-config and
  --sofa-interpolation, and make --hrtf-radius-m optional for a single
  shell
- Cover the mesh weights and Voronoi areas, the sphere series and the
  partitioned convolver with reference values in tests
- Document both rules in the READMEs and the binaural notes

Rosella renders stay byte-identical (d5877e89...) and so do renders of
the legacy compiled path (6f0a75d2...). Direct SOFA renders match the
Python reference to 1e-4 in level; the residual comes from the
barycentric solve (analytic inverse here, LAPACK there) on near-degenerate
triangles.
2026-11-01 02:54:55 +08:00

12 KiB

JustOneCacophony — C++ Core

中文 · Mathematics · Binaural rendering · SIMD dispatch

The C++ implementation of JustOneCacophony: an execution core for E-AC-3 JOC bitstream parsing, object reconstruction and rendering. It extracts EMDF, ID14 JOC parameters and ID11 OAMD metadata from E-AC-3 syncframes, combines them with the core 5.1 PCM decoded by FFmpeg to rebuild the LFE and 15 object signals, and writes ADM BWF, a WAV for a chosen speaker layout, or a binaural WAV rendered directly from full SOFA HRIRs (or from a compiled .jochrtf directional field).

This is research code, not a complete, standard-conformant or production JOC decoder. It covers the bitstream forms it implements and reports an explicit error on unknown variants instead of pretending everything is in harmony.

Building

cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build
ctest --test-dir build --output-on-failure

Only MSVC (VS 2022, static CRT) is validated locally; Linux and macOS are built and unit-tested by .github/workflows/ci.yml. Floating-point behaviour is part of the byte-exact acceptance, so fast-math is never enabled: /fp:precise on MSVC, -fno-fast-math elsewhere.

The heaviest binaural loops are runtime-dispatched: one binary carries several ISA variants and picks the widest one at startup from CPUID/XGETBV, with the output byte-identical either way (see SIMD dispatch). JOC_ENABLE_AVX2 / JOC_ENABLE_AVX512 (both ON by default) only decide whether those units are linked in: carrying the AVX2 unit buys 12.4% on the SOFA binaural kernel and 1.8% on Rosella, while a CPU without AVX2 still runs the baseline (SSE2) path. To produce a baseline-only binary:

cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DJOC_ENABLE_AVX2=OFF

gives a baseline-ISA (SSE2) build that runs on any x86-64. Non-MSVC builds never receive the flag.

Paths and encoding

Every path inside the library is UTF-8, converted only at the OS boundary (src/foundation/fs_utf8.*): on Windows through std::filesystem::path (UTF-16 inside) into _wfopen/CreateProcessW, and as plain bytes elsewhere. Command line arguments are re-parsed from GetCommandLineW + CommandLineToArgvW on Windows and the console code page is set to UTF-8, so non-ASCII paths (Japanese, Chinese, ...) work for the input, the ffmpeg child process and the output files alike; a non-ASCII path regression case runs in ctest.

Artifacts

Artifact Purpose
joc_core.dll The engine: bitstream parsing, JOC/OAMD, DSP, speaker and binaural rendering, ADM BWF/WAV writing, file task, streaming surface
joc_cli.exe Command line frontend for file tasks
include/joc_core.h Engine, telemetry and file-task interface (pure C)
include/joc_stream.h Embedder-facing streaming push/pull interface (pure C, self-contained)

Command line

The arguments match the reference Python CLI exactly: the input is positional, ADM BWF is the default output, --speaker-layout or --binaural selects the other two modes, and without -o the result lands in output/.

# Default: ADM BWF (inherently 24-bit, so there is no format option)
joc_cli input.m4a
#   -> output/input.adm.wav

joc_cli input.m4a -o out/adm.wav                 # explicit output

# Speaker layout
joc_cli input.m4a --speaker-layout 5.1           # -> output/<name>.5.1.wav
joc_cli input.m4a --speaker-layout 7.1.4 --speaker-output out/714.wav --speaker-format int24

# Binaural (HRTF defaults to <exe>/HRTF/binaural.sofa, then <exe>/HRTF/binaural.personalized_headphone)
joc_cli input.m4a --binaural
joc_cli input.m4a --binaural --sofa-hrtf HRTF/other.sofa   # another SOFA
joc_cli input.m4a --binaural --personalized-headphone      # Rosella personalisation
#   -> output/<name>.binaural.wav
joc_cli input.m4a --binaural --personalized-headphone --rosella-filterbank 2   # captured bank

# Other common switches
joc_cli input.m4a --duration 30 --gain-db -3 --trajectory-mode dense64
joc_cli input.eac3 --metadata-only --print-metadata summary   # parse and print metadata only

--speaker-format / --binaural-format default to float32; an int24 request that would clip follows --clip-action (default ask; a non-interactive terminal must pass continue, float32 or abort). --duration is in seconds, and --object-delay-samples, --speaker-metadata-offset, --binaural-tail-seconds and --binaural-tail-threshold (1e-8, the binaural tail trim) keep the reference defaults. A run always writes <output>.report.json (--report-json overrides it).

--rosella-filterbank 1|2 picks the filterbank of the Rosella .personalized_headphone path: 1 (default) is the public 13-tap / 6-slot bank, 2 is the captured 11-tap / 3-slot bank. Their latencies are 961 and 448 samples, and each keeps its own coefficients, history and flush capacity. The option only applies to the Rosella model; the SOFA and .jochrtf paths reject it. See Rosella filterbank.

Differences from the reference: --backend python and the metadata sidecars (--metadata-dir, --metadata-cache, --metadata-backend sidecar) are unavailable in this build and fail immediately with an explanation instead of being ignored. This build adds --bed (pre-decoded 6-channel float32 PCM, which skips ffmpeg decoding), --kernels, --work-dir, --report-json, --dry-run, --quiet and --rosella-filterbank.

Speaker layouts and channel order

--speaker-layout accepts the eleven layouts below, and the PCM/WAV channel order is the order in the table:

Layout Channel order
2.0 L, R
3.1 L, R, C, LFE
4.0 L, R, Lrs, Rrs
5.1 L, R, C, LFE, Ls, Rs
7.1 L, R, C, LFE, Ls, Rs, Lrs, Rrs
5.1.2 L, R, C, LFE, Ls, Rs, Ltm, Rtm
5.1.4 L, R, C, LFE, Ls, Rs, Ltf, Rtf, Ltr, Rtr
7.1.2 L, R, C, LFE, Ls, Rs, Lrs, Rrs, Ltm, Rtm
7.1.4 L, R, C, LFE, Ls, Rs, Lrs, Rrs, Ltf, Rtf, Ltr, Rtr
9.1.4 L, R, C, LFE, Ls, Rs, Lrs, Rrs, Lw, Rw, Ltf, Rtf, Ltr, Rtr
9.1.6 L, R, C, LFE, Ls, Rs, Lrs, Rrs, Lw, Rw, Ltf, Rtf, Ltm, Rtm, Ltr, Rtr

4.0 is the quad layout this project adds (the traditional FL FR BL BR, here L R Lrs Rrs); it is not part of the Dolby layout set. Every other layout and channel order follows the Dolby specifications.

Multichannel WAV files use WAVE_FORMAT_EXTENSIBLE with dwChannelMask = 0: the file declares no Windows speaker positions, and the PCM channel order is exactly the table above, never rearranged into mask-bit order. Players, ASIO, DAWs and other devices have to map the channel numbers themselves; this project guarantees the content and numbering of each PCM channel, not that a downstream player recognises Dolby speaker positions on its own.

--speaker-layout accepts only the eleven names in the table; anything else fails as an unknown layout.

Library integration

The engine and the file task are exposed by joc_core.h: joc_task_validate / joc_task_execute run a file task and report state events through a callback, and joc_task_result carries frame counts, peak, byte count and SHA-256.

Players and decoder components use joc_stream.h: the caller pushes E-AC-3 bytes and the matching core PCM (or already-rebuilt objects16) at its own pace and pulls rendered PCM. Any chunking is allowed, and the result is byte-identical to the file task.

joc_stream_config config = {0};
config.struct_size = sizeof(config);
config.input = JOC_STREAM_IN_EAC3;      /* or JOC_STREAM_IN_PCM_OBJECTS16 */
config.output = JOC_STREAM_OUT_SPEAKER; /* or BINAURAL / PCM_OBJECTS16 */
config.speaker_layout_name = "5.1";
joc_stream* stream = NULL;
joc_stream_create(&config, &stream);
/* loop: joc_stream_push(...) / joc_stream_pull(...) */
joc_stream_flush(stream);
joc_stream_destroy(stream);

Contract: state is instance-private, so streams coexist; push and pull on one instance must come from the same thread; rendering is stateful, so this version offers no seek - repositioning means decoding from the start of the stream. The kernel latency is 961 samples for speaker/binaural output and joc_stream_flush drains the binaural room tail.

Layout

include/     public C ABI: joc_core.h (engine/file task), joc_stream.h (streaming),
             eac3joc_core.h (upstream ABI)
src/         implementation: eac3_transport, emdf, joc_bitstream, joc_core, oamd,
             timeline, speaker, binaural, hrtf, adm, io, telemetry, task, stream,
             api, cli, simd
tests/       unit tests (CTest, self-contained, no external data)
docs/        mathematics, the binaural rendering flow and SIMD dispatch

Eight files under src/ are byte-identical copies of the upstream JustOneCacophony native library and are never edited (see THIRD_PARTY_NOTICES.md).

Compatibility note

How joc_clipgain is applied

joc_clipgain is folded into the analysis-side scalar: scalar = 1/16 x g (g = 1 + (y/32)·2^(x-4), from the 8-bit field x:3 | y:5). The x16 at the output cancels the 1/16, so the net effect is a multiply by g and the clamp therefore sits after the gain (clamp(16·IQMF·g_interp)). The scalar is kept in two per-frame slots: the first 256 samples of a frame ramp linearly from the previous value, the rest stays constant, and the first frame after a reset takes the new value directly. For g = 1 the interpolation is the identity and clamp(x)·1 == clamp(x·1).

object_delay_samples defaults to 1473, preserving the behaviour of the existing implementation; it is a configurable field and changing it changes the OAMD/ADM time alignment. Upstream investigation suggests the value should be 0; this project keeps the current default to stay byte-identical.

joc_dmx_config_idx 0..4

All five configurations are covered: N=5,7,7,5,7 with the phase switch \sigma(k)=(k\in\{3,4\}) gating the Ls/Rs -j rotation and the band-0 21-tap FIR together. A 7-channel configuration needs an 8-channel bed, mapped by channel code:

k extra pair decoder 8-channel order JOC input (bed index)
1 Lb/Rb FL FR FC LFE BL BR SL SR 0 1 2 6 7 4 5
2 / 4 Tfl/Tfr FL FR FC LFE SL SR TFL TFR 0 1 2 4 5 6 7

ffmpeg gets -ac 6 only for the N=5 configurations; N=7 keeps the decoder's own layout (-ac 6 downmixes 5.1.2/7.1 into 5.1 and -ac 8 remaps 5.1.2 into 7.1; both lose information).

Cross-layer consistency (section 11.2): the substreams' full-band channel count must equal N, a non-zero joc\_ext\_config\_idx is an unmapped variant, and a declared object count above 15 is clamped to 15 and recorded as a diagnostic - those change the DSP input, so a mismatch is refused. The extra pair's identity never reaches the DSP (section 10.1): when the dependent-substream chanmap disagrees with k, the mismatch is recorded as a diagnostic and the frame still renders, with the bed mapping following the channel codes the chanmap declares. The bed layout the OAMD declares may hold more channels than the core substream codes (a 9.1.6 bed with a 5.1 core, for example): only a core channel missing from that layout counts as a mismatch. For bed-only content the 15 JOC object slots are the bed channels, each rendered as a point source at the fixed position its channel label defines (ETSI TS 103 420 table B.10); speaker, ADM and binaural output share those source coordinates and never re-place them per layout, while LFE keeps its separate path. When one audio frame holds several EMDF containers carrying ID11/ID14, they are scored by the association rule of section 8.2 and its Table 56: a container whose OAMD and JOC payloads share a group id and whose ID14 parses completely wins, then one whose ID14 merely parses; only a unique best score is used and the others are recorded as a diagnostic (emdf/extra_container_ignored). The extra pair's 10-slot ADM skeleton placement (k=1\to6/7, k\in\{2,4\}\to8/9) is recorded, but bed channels are not output channels under the current contract.

An audio frame with no JOC EMDF at all (for example when the encoder wrote none for the first frame) emits silent object channels while the QMF analysis and synthesis state keep advancing, and it is counted separately in the statistics.

License

MIT, see LICENSE. Third-party provenance and patent boundaries are in THIRD_PARTY_NOTICES.md.