Fix doubled Rosella gain and bed DRC decoding
build / windows (push) Has been cancelled
build / release (push) Has been cancelled

This commit is contained in:
TheM14 committed 2026-09-25 21:23:13 +08:00
1 parent 188cb9ebfb
commit 259688ef1b
3 files changed
+157 -142

No files matched your search

+144 -134
View File
@@ -1,134 +1,144 @@
# Verification # Verification␍␍
␍␍
Measured results only. Each entry names the command or the log line it came from, so it can Measured results only. Each entry names the command or the log line it came from, so it can␍␍
be reproduced. Environment: Windows x64 host, official **foobar2000 1.6.19 x86** portable be reproduced. Environment: Windows x64 host, official **foobar2000 1.6.19 x86** portable␍␍
installation, official SDK 2026-09-17 pinned to `FOOBAR2000_TARGET_VERSION 80`, MSVC 14.44, installation, official SDK 2026-09-17 pinned to `FOOBAR2000_TARGET_VERSION 80`, MSVC 14.44,␍␍
ffmpeg 8.0. ffmpeg 8.0.␍␍
␍␍
Test material is supplied locally and is **not** part of this repository: the `testdata/` and Test material is supplied locally and is **not** part of this repository: the `testdata/` and␍␍
`vectors/` files of the upstream renderer project, and — for the binaural measurements — an `vectors/` files of the upstream renderer project, and — for the binaural measurements — an␍␍
HRTF file. Everything except binaural rendering runs without any HRTF; binaural runs take the HRTF file. Everything except binaural rendering runs without any HRTF; binaural runs take the␍␍
file as an argument (`tests/render_harness.cpp --hrtf …`) or use the default location beside file as an argument (`tests/render_harness.cpp --hrtf …`) or use the default location beside␍␍
the DLL. the DLL.␍␍
␍␍
## Component and renderer ## Component and renderer␍␍
␍␍
| Check | Result | | Check | Result |␍␍
|---|---| |---|---|␍␍
| Sources compiled in, nothing loaded at run time | log: `core: in-process renderer 0.1.0-m1 (abi 3), component built against abi 3` | | Sources compiled in, nothing loaded at run time | log: `core: in-process renderer 0.1.0-m1 (abi 3), component built against abi 3` |␍␍
| One artefact, no companion DLL | `dist\*.fb2k-component` holds `foo_input_joc.dll` and `README.md` only | | One artefact, no companion DLL | `dist\*.fb2k-component` holds `foo_input_joc.dll` and `README.md` only |␍␍
| Kernel sources untouched | the upstream working tree's file timestamps are unchanged; it is only ever read | | Kernel sources untouched | the upstream working tree's file timestamps are unchanged; it is only ever read |␍␍
| Both architectures build | `build\Win32\foo_input_joc.dll`, `build\x64\foo_input_joc.dll` | | Both architectures build | `build\Win32\foo_input_joc.dll`, `build\x64\foo_input_joc.dll` |␍␍
␍␍
## Playback in foobar2000 1.6.19 x86 ## Playback in foobar2000 1.6.19 x86␍␍
␍␍
| Case | Log evidence | | Case | Log evidence |␍␍
|---|---| |---|---|␍␍
| Speaker 7.1, 5 s file | `stream created, 8 output channel(s), layout=7.1` … `frames_in=157 frames_out=157 samples_out=241152` — 157 × 1536 exactly | | Speaker 7.1, 5 s file | `stream created, 8 output channel(s), layout=7.1` … `frames_in=157 frames_out=157 samples_out=241152` — 157 × 1536 exactly |␍␍
| Binaural, SOFA | `stream created, 2 output channel(s)` … `end of stream after 481215 frames` (241152 source + 240063 tail) | | Binaural, SOFA | `stream created, 2 output channel(s)` … `end of stream after 481215 frames` (241152 source + 240063 tail) |␍␍
| Binaural, Rosella model | `stream created, 2 output channel(s)` … `end of stream after 481855 frames` | | Binaural, Rosella model | `stream created, 2 output channel(s)` … `end of stream after 481855 frames` |␍␍
| Binaural, HRTF path left empty | resolves to `<component directory>\HRTF\binaural.sofa` and produces the same 481215 frames as naming that file explicitly | | Binaural, HRTF path left empty | resolves to `<component directory>\HRTF\binaural.sofa` and produces the same 481215 frames as naming that file explicitly |␍␍
| Full 238 s file, binaural SOFA | `eac3 frames queued=7436, bed frames pushed=7436` … `samples_out=11661759`; no ffmpeg process left behind | | Full 238 s file, binaural SOFA | `eac3 frames queued=7436, bed frames pushed=7436` … `samples_out=11661759`; no ffmpeg process left behind |␍␍
| Installed from the `.fb2k-component` package | unpacked into `user-components\foo_input_joc\`, plays with the default-folder HRTF | | Installed from the `.fb2k-component` package | unpacked into `user-components\foo_input_joc\`, plays with the default-folder HRTF |␍␍
␍␍
## Bitstream recognition ## Bitstream recognition␍␍
␍␍
`tests/scan_crosscheck.cpp` compares, frame by frame, the syncframe lengths and the JOC `tests/scan_crosscheck.cpp` compares, frame by frame, the syncframe lengths and the JOC␍␍
verdict this component computes against the renderer's own `joc_eac3_frame_bytes()` / verdict this component computes against the renderer's own `joc_eac3_frame_bytes()` /␍␍
`joc_parse_eac3_frame()`: `joc_parse_eac3_frame()`:␍␍
␍␍
``` ```␍␍
testdata\gold_forever.eac3 frames=64 plugin_joc=64 kernel_joc=64 len_mismatch=0 verdict_mismatch=0 testdata\gold_forever.eac3 frames=64 plugin_joc=64 kernel_joc=64 len_mismatch=0 verdict_mismatch=0␍␍
vectors\valid.eac3 frames=7 plugin_joc=7 kernel_joc=7 len_mismatch=0 verdict_mismatch=0 vectors\valid.eac3 frames=7 plugin_joc=7 kernel_joc=7 len_mismatch=0 verdict_mismatch=0␍␍
build\plain_eac3.eac3 frames=64 plugin_joc=0 kernel_joc=0 len_mismatch=0 verdict_mismatch=0 build\plain_eac3.eac3 frames=64 plugin_joc=0 kernel_joc=0 len_mismatch=0 verdict_mismatch=0␍␍
crosscheck: 3 file(s), AGREES WITH CORE crosscheck: 3 file(s), AGREES WITH CORE␍␍
``` ```␍␍
␍␍
Ten corrupt vectors were compared as well: no frame-length disagreement, verdicts agreed on Ten corrupt vectors were compared as well: no frame-length disagreement, verdicts agreed on␍␍
9 of 10. The one difference is `corrupt_truncated_huffman.eac3`: this component only tests 9 of 10. The one difference is `corrupt_truncated_huffman.eac3`: this component only tests␍␍
for the container, while the renderer also parses the payload and reports a truncated for the container, while the renderer also parses the payload and reports a truncated␍␍
bitstream later. bitstream later.␍␍
␍␍
## Plain E-AC-3 is handed back ## Plain E-AC-3 is handed back␍␍
␍␍
Playing a file the component's own encoder produced without JOC: Playing a file the component's own encoder produced without JOC:␍␍
␍␍
``` ```␍␍
decoder: open "...plain_eac3.eac3" reason=1 bytes=144384 frames=8 with_joc=0 decoder: open "...plain_eac3.eac3" reason=1 bytes=144384 frames=8 with_joc=0␍␍
decoder: yielding to the built-in decoder (at least one examined syncframe has no JOC EMDF container) decoder: yielding to the built-in decoder (at least one examined syncframe has no JOC EMDF container)␍␍
``` ```␍␍
␍␍
The file then plays through the built-in decoder; no decode log appears for it. Verified both The file then plays through the built-in decoder; no decode log appears for it. Verified both␍␍
before and after the renderer was compiled in. before and after the renderer was compiled in.␍␍
␍␍
## Output identical to the reference renderer ## Output identical to the reference renderer␍␍
␍␍
`tests/render_harness.cpp` drives the component's own engine and writes a WAV in the same `tests/render_harness.cpp` drives the component's own engine and writes a WAV in the same␍␍
format the reference command-line renderer writes, so the two files can be compared byte for format the reference command-line renderer writes, so the two files can be compared byte for␍␍
byte. 30 s of the reference file, speaker 5.1: byte. 30 s of the reference file, speaker 5.1:␍␍
␍␍
| Product | Whole-file SHA-256 | | Product | Whole-file SHA-256 |␍␍
|---|---| |---|---|␍␍
| reference renderer (`--speaker-layout 5.1 --bed … --duration 30`) | `99a8e3edbd1c047a3c0f547eaf85e56941f882af9e65468fbde0a9f18b6c5b6e` | | reference renderer (`--speaker-layout 5.1 --bed … --duration 30`) | `99a8e3edbd1c047a3c0f547eaf85e56941f882af9e65468fbde0a9f18b6c5b6e` |␍␍
| this component's engine, x64 | `99a8e3edbd1c047a3c0f547eaf85e56941f882af9e65468fbde0a9f18b6c5b6e` | | this component's engine, x64 | `99a8e3edbd1c047a3c0f547eaf85e56941f882af9e65468fbde0a9f18b6c5b6e` |␍␍
| this component's engine, x86 | `99a8e3edbd1c047a3c0f547eaf85e56941f882af9e65468fbde0a9f18b6c5b6e` | | this component's engine, x86 | `99a8e3edbd1c047a3c0f547eaf85e56941f882af9e65468fbde0a9f18b6c5b6e` |␍␍
␍␍
34,578,500 bytes each. Binaural with a real SOFA file: with the same input frame count the 34,578,500 bytes each. Binaural with a real SOFA file: with the same input frame count the␍␍
whole file is identical too (`588a15ce5526977f…baa88`, 12,158,780 bytes), and over the whole whole file is identical too (`588a15ce5526977f…baa88`, 12,158,780 bytes), and over the whole␍␍
238 s file the total sample count matches exactly (11,661,759 = 11,420,735 program + 238 s file the total sample count matches exactly (11,661,759 = 11,420,735 program +␍␍
241,024 tail) with identical peak (1.144561172) and an identical SHA-256 over the reference 241,024 tail) with identical peak (1.144561172) and an identical SHA-256 over the reference␍␍
renderer's entire payload. renderer's entire payload.␍␍
␍␍
The one structural difference is the tail: the reference renderer trims trailing samples The one structural difference is the tail: the reference renderer trims trailing samples␍␍
below 1e-8 and this component returns the tail in full. below 1e-8 and this component returns the tail in full.␍␍
␍␍
## Gain ## Gain␍␍
␍␍
Same file, speaker 5.1, rendered with and without attenuation: Same file, speaker 5.1, rendered with and without attenuation:␍␍
␍␍
| Check | Result | | Check | Result |␍␍
|---|---| |---|---|␍␍
| Peak | 0.183568597 → 0.092002235, i.e. exactly −6.000000 dB | | Peak | 0.183568597 → 0.092002235, i.e. exactly −6.000000 dB |␍␍
| RMS over the whole signal | exactly −6.000000 dB | | RMS over the whole signal | exactly −6.000000 dB |␍␍
| Per-sample, 1,446,912 floats | largest deviation from the ideal scaling is 2.1e-06 relative; the gain is applied in double precision before the DSP, so the remaining difference is float32 rounding | | Per-sample, 1,446,912 floats | largest deviation from the ideal scaling is 2.1e-06 relative; the gain is applied in double precision before the DSP, so the remaining difference is float32 rounding |␍␍
| +6 dB | same check passes | | +6 dB | same check passes |␍␍
| The switch | switch on and −6 dB → log `gain=-6.00 dB`, delivered peak 0.000063; switch off with −6 dB still stored → `gain=0.00 dB`, peak 0.000126; switch on and 0 dB → peak 0.000126 | | The switch | switch on and −6 dB → log `gain=-6.00 dB`, delivered peak 0.000063; switch off with −6 dB still stored → `gain=0.00 dB`, peak 0.000126; switch on and 0 dB → peak 0.000126 |␍␍
␍␍
## Preferences page ## Preferences page␍␍
␍␍
`tests/prefs_layout_check.cpp` builds the page from the component's own dialog resource and `tests/prefs_layout_check.cpp` builds the page from the component's own dialog resource and␍␍
asserts, per control, that it is enabled, lies inside the client area, and is not covered by asserts, per control, that it is enabled, lies inside the client area, and is not covered by␍␍
another interactive control (static text and group boxes are transparent to the mouse, as another interactive control (static text and group boxes are transparent to the mouse, as␍␍
they are for real clicks): they are for real clicks):␍␍
␍␍
``` ```␍␍
dialog client=495x367 non-client=0x0 dialog client=495x367 non-client=0x0␍␍
WS_CAPTION=no WS_BORDER=no WS_CHILD=yes WS_VISIBLE=yes WS_CAPTION=no WS_BORDER=no WS_CHILD=yes WS_VISIBLE=yes␍␍
content extent=485x354 client=495x367 (everything fits) content extent=485x354 client=495x367 (everything fits)␍␍
controls=26 problems=0 controls=26 problems=0␍␍
``` ```␍␍
␍␍
The page has no caption of its own — the host draws the frame — and no control is ever The page has no caption of its own — the host draws the frame — and no control is ever␍␍
disabled, which is what "the option is there but cannot be clicked" otherwise looks like. disabled, which is what "the option is there but cannot be clicked" otherwise looks like.␍␍
␍␍
Not verified here: how the page and the `%joc_*%` fields look on screen; that needs a human Not verified here: how the page and the `%joc_*%` fields look on screen; that needs a human␍␍
in front of the window. in front of the window.␍␍
␍␍
## Container support (mp4 / m4a / mov / mkv / mka / webm) ## Container support (mp4 / m4a / mov / mkv / mka / webm)␍␍
␍␍
Verified with the Win32 build, a 5.1 speaker layout, and the component installed in a Verified with the Win32 build, a 5.1 speaker layout, and the component installed in a␍␍
portable foobar2000 1.6.19 profile: portable foobar2000 1.6.19 profile:␍␍
␍␍
- `ffmpeg -i <bare .eac3> -c:a copy` muxed into MP4 and Matroska, then `-map 0:a:0 -c:a copy - `ffmpeg -i <bare .eac3> -c:a copy` muxed into MP4 and Matroska, then `-map 0:a:0 -c:a copy␍␍
-f eac3` extracted again, is byte-identical to a direct `-t 30 -c:a copy` of the source -f eac3` extracted again, is byte-identical to a direct `-t 30 -c:a copy` of the source␍␍
(same SHA-256) — the renderer therefore sees the stored syncframes, not a re-encode. (same SHA-256) — the renderer therefore sees the stored syncframes, not a re-encode.␍␍
- The header probe (`tests/container_scan_test.cpp`, no foobar2000 involved) reports, for the - The header probe (`tests/container_scan_test.cpp`, no foobar2000 involved) reports, for the␍␍
same material: `joc.mp4 -> mp4 eac3=1 audio#0 codec=ec-3 30.016 s`, same material: `joc.mp4 -> mp4 eac3=1 audio#0 codec=ec-3 30.016 s`,␍␍
`joc.mkv -> matroska eac3=1 audio#0 codec=A_EAC3 30.016 s`, `plain_eac3.mp4 -> ec-3`, `joc.mkv -> matroska eac3=1 audio#0 codec=A_EAC3 30.016 s`, `plain_eac3.mp4 -> ec-3`,␍␍
`ac3.mp4 -> ac-3 (declined)`, `aac.mp4 -> mp4a (declined)`. `ac3.mp4 -> ac-3 (declined)`, `aac.mp4 -> mp4a (declined)`.␍␍
- End to end, with the component ordered ahead of the container reader in - End to end, with the component ordered ahead of the container reader in␍␍
Preferences -> Decoding: `open()` is called, the track is found, the JOC verdict is positive, Preferences -> Decoding: `open()` is called, the track is found, the JOC verdict is positive,␍␍
the file is claimed, and playback reaches `end of stream`. the file is claimed, and playback reaches `end of stream`.␍␍
- Negative case end to end: an MP4 holding E-AC-3 without JOC yields with - Negative case end to end: an MP4 holding E-AC-3 without JOC yields with␍␍
`E-AC-3 track 0 carries no JOC`, and the built-in decoder plays it. `E-AC-3 track 0 carries no JOC`, and the built-in decoder plays it.␍␍
- Bare `.eac3` / `.ec3` handling is unchanged. - Bare `.eac3` / `.ec3` handling is unchanged.
## Bed decode and gain
* The 5.1 bed is decoded with -drc_scale 0 -target_level 0, so it is taken as stored and the
decoder does not apply the stream's dynrng or target-level metadata. The resulting bed is
byte-identical to the one the reference renderer uses (same SHA-256).
* The master gain is applied once, by the JOC kernel, on every output path. Measured at -6 dB
against 0 dB: speaker 5.1 and SOFA binaural give 0.50119 in peak and in per-channel RMS,
Rosella binaural gives 0.50119 as well, and a 0 dB render is unaffected by the renderer's
output gain.
+1 -1
View File
@@ -118,7 +118,7 @@ Status Stream::create(const Config& config) {
} }
render_options.object_delay_samples = config_.object_delay_samples; render_options.object_delay_samples = config_.object_delay_samples;
render_options.tail_seconds = config_.tail_seconds; render_options.tail_seconds = config_.tail_seconds;
render_options.output_gain = std::pow(10.0, config_.gain_db / 20.0); render_options.output_gain = 1.0;
if (status.ok()) { if (status.ok()) {
status = rosella_.open(model, render_options); status = rosella_.open(model, render_options);
} }
+12 -7
View File
@@ -94,10 +94,11 @@ public:
// ffmpeg_path, the input file and whatever should follow "-i <input>" are // ffmpeg_path, the input file and whatever should follow "-i <input>" are
// separate: the 5.1 bed and the E-AC-3 metadata stream of a container file are // separate: the 5.1 bed and the E-AC-3 metadata stream of a container file are
// both ffmpeg output, they only differ in those arguments. // both ffmpeg output, they only differ in those arguments. input_arguments come
// before -i and carry the decoder options the bed needs.
bool start(const std::string& ffmpeg_path, const std::string& input_path, bool start(const std::string& ffmpeg_path, const std::string& input_path,
const std::wstring& output_arguments, const char* label, const std::wstring& input_arguments, const std::wstring& output_arguments,
const std::wstring& stderr_path, std::string* error, const char* label, const std::wstring& stderr_path, std::string* error,
std::size_t pipe_bytes = kBedPipeBytes) { std::size_t pipe_bytes = kBedPipeBytes) {
SECURITY_ATTRIBUTES attributes{}; SECURITY_ATTRIBUTES attributes{};
attributes.nLength = sizeof(attributes); attributes.nLength = sizeof(attributes);
@@ -118,7 +119,9 @@ public:
CREATE_ALWAYS, 0, nullptr); CREATE_ALWAYS, 0, nullptr);
std::wstring command = L"\"" + utf8_to_wide(ffmpeg_path) + L"\""; std::wstring command = L"\"" + utf8_to_wide(ffmpeg_path) + L"\"";
command += L" -hide_banner -loglevel error -nostdin -y -i \""; command += L" -hide_banner -loglevel error -nostdin -y ";
command += input_arguments; // input options must precede -i
command += L" -i \"";
command += utf8_to_wide(input_path); command += utf8_to_wide(input_path);
command += L"\" "; command += L"\" ";
command += output_arguments; command += output_arguments;
@@ -395,7 +398,7 @@ bool probe_container_joc(const std::string& ffmpeg_path, const std::string& path
return length == 0 ? std::wstring(L"NUL") return length == 0 ? std::wstring(L"NUL")
: std::wstring(temp) + L"joc_container_probe.log"; : std::wstring(temp) + L"joc_container_probe.log";
}(); }();
if (!pipe.start(ffmpeg_path, path, arguments, "probe", stderr_path, &error, 1u << 20)) { if (!pipe.start(ffmpeg_path, path, L"", arguments, "probe", stderr_path, &error, 1u << 20)) {
if (detail != nullptr) *detail = error; if (detail != nullptr) *detail = error;
return false; return false;
} }
@@ -604,10 +607,12 @@ bool Engine::start(const std::string& input_path, const Settings& settings, std:
// The 5.1 core PCM, exactly as the reference renderer's own core decode does it: // The 5.1 core PCM, exactly as the reference renderer's own core decode does it:
// 5.1 interleaved float32 at 48 kHz, the layout the renderer expects // 5.1 interleaved float32 at 48 kHz, the layout the renderer expects
// (L R C LFE Ls Rs). // (L R C LFE Ls Rs).
// -drc_scale 0 -target_level 0: the bed is taken as stored, without the stream's
// dynrng or target-level metadata being applied by the decoder.
std::wstring bed_arguments = L"-map 0:a:"; std::wstring bed_arguments = L"-map 0:a:";
bed_arguments += std::to_wstring(settings.audio_index); bed_arguments += std::to_wstring(settings.audio_index);
bed_arguments += L" -vn -ac 6 -ar 48000 -c:a pcm_f32le -f f32le -"; bed_arguments += L" -vn -ac 6 -ar 48000 -c:a pcm_f32le -f f32le -";
if (!impl.bed.start(settings.ffmpeg_path, input_path, bed_arguments, "bed", if (!impl.bed.start(settings.ffmpeg_path, input_path, L"-drc_scale 0 -target_level 0", bed_arguments, "bed",
stderr_path_for(L"joc_ffmpeg_bed.log"), error)) { stderr_path_for(L"joc_ffmpeg_bed.log"), error)) {
return false; return false;
} }
@@ -618,7 +623,7 @@ bool Engine::start(const std::string& input_path, const Settings& settings, std:
std::wstring stream_arguments = L"-map 0:a:"; std::wstring stream_arguments = L"-map 0:a:";
stream_arguments += std::to_wstring(settings.audio_index); stream_arguments += std::to_wstring(settings.audio_index);
stream_arguments += L" -vn -c:a copy -f eac3 -"; stream_arguments += L" -vn -c:a copy -f eac3 -";
if (!impl.eac3_pipe.start(settings.ffmpeg_path, input_path, stream_arguments, "metadata", if (!impl.eac3_pipe.start(settings.ffmpeg_path, input_path, L"", stream_arguments, "metadata",
stderr_path_for(L"joc_ffmpeg_stream.log"), error, stderr_path_for(L"joc_ffmpeg_stream.log"), error,
1u << 20)) { 1u << 20)) {
return false; return false;