Measure the VBR target against the nominal rate

[thomas-vilte]

Description

There were two places where the VBR path was using the frame budget while the reference uses the requested bitrate.

1. equiv_rate. This feeds the intensity band, trim, and temporal VBR factor. libopus caps it with the nominal bitrate (celt_encoder.c:1926):

equiv_rate = ((opus_int32)nbCompressedBytes*8*50 << (3-LM)) - (40*C+20)*((400>>LM) - 50);
if (st->bitrate!=OPUS_BITRATE_MAX)
   equiv_rate = IMIN(equiv_rate, st->bitrate - (40*C+20)*((400>>LM) - 50));

In VBR, nbCompressedBytes is the output buffer size, so without that IMIN every rate-dependent decision would see a bitrate corresponding to hundreds of kb/s. pion was always deriving equiv_rate from the frame budget.

Instrumenting the reference at 24 kb/s:

VBR: equiv=24000 with nbComp=1275
CBR: equiv=23600 with nbComp=59

In CBR, st->bitrate is OPUS_BITRATE_MAX, so the cap does not apply.

pion was returning 23600 in both cases. That put VBR 24 kb/s into the previous intensity-table entry: band 9 instead of the reference's band 10.

2. vbr_rate. The base used to calculate the target comes directly from the bitrate (celt_encoder.c:1904):

vbr_rate = bitrate_to_bits(st->bitrate, mode->Fs, frame_size)<<BITRES;
effectiveBytes = vbr_rate>>(3+BITRES);

This is the integer portion of the frame — 60 bytes at 24 kb/s — including the header byte that the CBR payload budget excludes. The reservoir absorbs the difference.

pion was using the 59-byte payload budget instead, one byte less per frame.

For this, the CELT layer needs to know the nominal bitrate, so I added SetBitrate alongside the existing setters.

Measurement

Intensity-band agreement against opus_demo, over 3000 frames:

mode	before	after
VBR 24 kb/s	0.0%	100%
VBR 96 kb/s	100%	100%
CBR 24 and 96	100%	100%

The 24 kb/s VBR case was a complete mismatch: the reference selected band 10 for all 3000 frames while pion selected band 9.

This also affects the bands that are actually coded:

Packet size and SNR in VBR:

bitrate	before	after	libopus
24000	51.44 B / 6.1840 dB	52.44 B / 6.2953 dB	53.50 B / 6.3764 dB
48000	111.55 B / 9.8859 dB	112.56 B / 9.9074 dB	113.71 B / 10.0016 dB
96000	241.81 B / 14.9433 dB	242.81 B / 14.9785 dB	244.03 B / 15.0810 dB

mode	before	after
VBR 24 kb/s	90.7%	95.9%
VBR 96 kb/s	99.9%	100%

CBR remains unchanged, verified at 24, 48, and 96 kb/s: 7.0676 / 10.1179 / 15.0721 dB, identical digit for digit.

What remains

In VBR, pion is now within roughly 0.5–2% of libopus in bytes and 0.08–0.10 dB in SNR.

In CBR, it remains essentially matched: +0.014 / +0.015 / −0.002 dB at 24, 48, and 96 kb/s.

The only metric still below 99% is the number of coded bands at 24 kb/s: 95.9% in VBR and 96.9% in CBR. Since it happens in both modes, this is not a VBR-path issue and should be addressed separately.

Reference issue

Part of #34.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.11%. Comparing base (00b48f4) to head (0e676e5).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #206      +/-   ##
==========================================
+ Coverage   93.09%   93.11%   +0.01%     
==========================================
  Files          58       58              
  Lines       10389    10407      +18     
==========================================
+ Hits         9672     9690      +18     
  Misses        505      505              
  Partials      212      212              

Flag	Coverage Δ
go	93.11% <100.00%> (+0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.