この文書はまだ翻訳されていません。本文は英語のまま表示されています。

M2 report: RGB encode, the comb rungs, WASM, and the measured budget

Run stamp: 2026-09-01, third commit of this repository, rustc 1.97.1, cargo test --workspace 45 tests green, clippy clean, MUTATE=1 reddens 27 tests. Throughput measured on an AMD Ryzen 5 5600X (6 cores, one thread used), node v24.0.1 for the wasm rows. Every number below is pasted from the runs described; M5 puts them under a check-self-counts scan.

The primary source, finally in hand

SMPTE's own repository serves ST 170M-2004 (stabilized 2010): https://pub.smpte.org/latest/st170/st170-20041130-pub.zip, zip sha256 dc9a1f0487385b697de0caca0a8371d61d84bffe7352bcf707818cc9bd8df2bd. The M1 obligation is closed: clause 6.1 states the base matrix digit for digit ("precise values; i.e., 0.587 G = 587/1000 G"), and clause 6.2's I/Q formulas re-derive from the transcribed reduction factors and the 33-degree rotation to all four printed decimals. data/yuv-matrix.toml records the confirmation; data/broadcast-timing.toml is a fresh transcription of Table 1 (levels), Table 2 (horizontal timing) and clauses 5.1, 7, 8, 10 and 11. Clause 11's own rationals land exactly on M0's: f_sc = 5 MHz x 63/88 = 315/88 MHz, f_V = 59.94005994 Hz = 60,000/1,001.

Two findings that reshaped the plan, both from the primary:

Equiband U/V is the standard, not a compromise. Clause 7.2's primary spec bandlimits B-Y and R-Y identically (less than 2 dB down at 1.3 MHz, at least 20 dB down at 3.6 MHz); the split I/Q bandwidths are its NTSC-1953 continuation note. The spec's section 4.2 step 2 ("Y to 4.2 MHz, I to 1.3, Q to 0.4") should cite this at v0.3: Y is explicitly unrestricted (7.1) and the I/Q split is optional history.
Clause 10 is self-normalizing for the decoder. The base equation scales luma and chroma by the same 0.925 and adds 7.5 IRE of setup, so a decoder referenced to black 7.5 / white 100 IRE recovers exactly 0.492 (B-Y) and 0.877 (R-Y): the reduction-factor U and V the transcribed inverse matrix expects, with no extra constants anywhere.

The RGB encoder, and the bars gate

ntsc-source-rgb: video-level G'B'R' (or sRGB bytes through the sRGB EOTF and clause 5.1's camera OETF), the clause 6.1 matrix, linear resample onto the 2271-sample active region, the colour-difference lowpass (Blackman-windowed sinc, cutoff scanned to the 7.2 template with most margin: 2.60 MHz, measured -0.00 dB at 1.3 MHz and -51.8 dB at 3.6, recorded in data/filters/rgb-encoder.toml), clause 10's base equation, and Table 1/2 sync, blanking and burst with the nine burst-free lines per field honoured. Simplifications (rectangular envelopes, line-granularity vertical sync, untuned SC-H) are recorded in the module header and invisible to M2's oracles.

The known-answer gate (tests/bars.rs), all inside Table 1's +/-1 IRE:

Each 75% bar's waveform mean lands on both the clause-10 derivation and the published waveform-monitor column (76.9 / 69.0 / 56.1 / 48.2 / 36.1 / 28.2 / 15.4 IRE), which are compared against each other, so a wrong matrix or a wrong setup breaks the agreement.
Burst peak-to-peak at 40 IRE (tolerance justified from the 12-sample grid's worst peak-sampling offset, cos 15 degrees), sync at -40, blanking at 0.
Each bar's chroma amplitude within 3% of clause 10's own scales.
The decode roundtrip returns every bar centre within 0.02 of the input video level.

The comb rungs

Rung B is refused by name on the NES profile at construction, per the spec: the 120-degree line residue means two adjacent lines cannot cancel; the attenuation a two-line comb would deliver there is 0.866 in its chroma path with half the chroma left in luma (the phasor arithmetic the temporal-comb tests measure directly). On broadcast it works: the comb luma's chroma-band energy on a bars line collapses to under 2% of the raw line's.
Rung C on the NES profile is the native comb, and the M2 gate is shown: on the stripe frame the three-line mean's subcarrier-locked chroma is under 1% of the raw line's, while the chroma path still demodulates the stripes' artifact colour at the notch rung's own saturation. On a solid frame the comb removes the fundamental to under 1e-3 where the notch leaves its (1 - 0.907) residue.
Rung D, measured against the spec's claim. Spec v0.2 section 5 says rendering-enabled alternate frames are 180 degrees apart and a two-frame average cancels chroma. Its own section 3.2 residues say 4 and 8 samples: 120 and 240 degrees, never 180. Measured (tests/combs.rs): the two-frame comb on the Even/OddShort pattern attenuates its chroma path to 0.866 of a single frame (|1 - e^(-i120)| / 2) and leaves half the fundamental in luma; it cannot cancel. Three full frames (the rendering-disabled pattern, residue 4 each: the Battletoads case) cancel exactly: luma fundamental under 2e-3, saturation preserved to 1.000. For v0.3: Rung D as specced does not exist; the honest temporal comb is three-frame on the full-frame pattern. The OddShort path is exercised through decode either way, which was the spec's stated reason for the rung.

Three more instrument lessons, recorded at the fix

Band energy is not chroma. The stripe frame's dot-rate luma square (2.685 MHz) lives inside the chroma band; a comb rightly keeps it, and a band-RMS instrument convicted a comb that was cancelling perfectly (the solid-frame probe measured 5e-4 of the raw). The instruments are now synchronous demodulation for subcarrier-locked claims, band RMS only where the fundamental is the only band occupant.
A borrowed decoder carries its own gain. The temporal comb run through a notch decoder inherited its 0.907 chroma-gain divisor and measured 1.103 saturation: exactly 1/0.907. The comb now overrides the gain to unity and says why.
A truncated tick manufactured a slow display. The broadcast pacing test's dt of 1,001e9/30,000 ns rounds down; the fix is a 3-tick pattern summing to exactly three periods.

Throughput, measured (the budget's replacement)

Ryzen 5 5600X, one thread, --release, best of 3, full pipeline = encode

decode at 2048 x 240 output:

Path	frames/s	Note
NES notch, native	7.54	3.7 Msamples/s decoded
NES comb3, native	8.49
NES temporal-3 decode, native	12.41	demod tail only, encode shared
Broadcast notch decode, native	8.21	240 lines of 2271
Broadcast comb2 decode, native	9.97
NES notch, wasm (node v24)	4.87
NES comb3, wasm	5.20
NES notch, wasm +simd128	4.97	inside noise
NES comb3, wasm +simd128	5.18	inside noise

The spec's authored budget said Rung A "should fit comfortably in a single thread with SIMD" at 42.95 Msamples/s; the measurement says otherwise: the naive per-sample FIRs cost about 253 multiply-adds per sample (51-tap chroma bandpass plus two 101-tap U/V lowpasses), an order of magnitude off real time, and the simd128 flag alone changes nothing because nothing is written to vectorize. The lever, named and not yet built: chroma is 0.6 MHz wide after demodulation, so decimating before the U/V lowpass cuts the dominant cost by the decimation factor, and the line convolutions are the natural place for explicit SIMD. M3 can run on these numbers (the CRT stages are per-pixel, not per-sample); real-time display is an optimization milestone, not a correctness one.

The pacing policy (spec section 7) is implemented and tested with derived, not retyped, expectations: a 60 Hz display against the NES rendering-enabled rate drops exactly the beat frequency (about one frame per ten seconds), a 120 Hz display duplicates every other callback, and a display at exactly 30,000/1,001 Hz never drifts.

Carried forward

M3: the CRT stages, analytic tests, and the shell gate.
Real-time decode (decimation, explicit SIMD) when a display consumer exists; the wasm bridge and stats are ready for it.
Spec v0.3 notes from this milestone: the Rung D correction, the clause 7.2 equiband citation, and M1's field-rate correction.

Pulled at build time from ntsc-crt/docs/m2-report.md; the repository is the one copy.