The polyphase evaluation is PHASES*TAPS_PER_PHASE multiply-adds per channel
(96 for stereo) on every sample, unconditionally, and measured as 42 ns/sample
- 17% of the whole chain, and the largest single item in it. It is also the
only part that is pure throughput rather than a serial dependency chain, which
makes it the one worth attacking.
It is skippable. The interpolator is linear, so over its FIR support
|true peak| <= peak_gain * max|history|, peak_gain = max_phase sum |c|
and when that bound sits under the ceiling the true peak provably cannot reach
it. The limiter's gain target would be 1.0 for that sample either way, so
feeding the sample peak instead cannot change the outcome - a genuinely loud
neighbour still gets caught by its own evaluation.
- oversampler: split max_true_peak into push (cheap, must run every sample or
the FIR history develops holes) and evaluate (the expensive half, skippable).
Expose peak_gain and SUPPORT_LEN so a caller can compute the bound.
- limiter: track the sample peak over the FIR support with the SlidingMax
already used for the look-ahead window, and skip evaluate when bounded under.
Measured, 48 kHz stereo:
limiter, quiet 0.1 (skips) 66 -> 31 ns/sample
limiter, hot 0.95 (always evaluates) 66 -> 72 ns/sample (no real change)
FULL CHAIN 232 -> 195 ns/sample
So this pays off on material that is not pinned to the ceiling and costs
nothing when it is. On a maximizer the hot case is common, so treat the win as
situational rather than a flat 16%.
evaluate_never_exceeds_peak_gain_bound asserts the bound directly against
adversarial input (full-scale square waves, impulses), since the whole
optimisation rests on it, and also asserts it is not vacuously loose.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LU1NRL7fAaZTtFMTTPn2b1
Found by reading the audio-thread code, not by profiling - no profiler is
set up in this project yet. All four are algorithmic or arithmetic wins,
none change the DSP result; the existing tests still pass.
- oversampler: max_true_peak recomputed its circular-history tap index
(including a % TAPS_PER_PHASE) inside the phase x channel loop, though
the index depends only on pos/k - 8x redundant. Hoisted to a per-sample
precomputed index array. This runs unconditionally every sample (it
feeds the limiter's true-peak detector, not gated by metering), so it
was the highest-value fix.
- limiter: the sliding-window max rescanned the entire look-ahead ring
(~73 samples at 48kHz/1.5ms) every sample - O(window) per sample.
Replaced with an O(1)-amortized monotonic-deque sliding max, backed by
preallocated ring arrays and RT-safe. Covered by a new test against a
naive brute-force reference. Also precomputed the constant true-peak
margin gain in prepare() instead of a powf every sample.
- compressor: replaced per-sample log10/powf(10, x/20) with the
equivalent ln/exp forms, mirroring nih-plug's own gain_to_db_fast /
db_to_gain_fast. Called 4x/sample (3 bands + 'All').
- both: replaced % capacity ring-index math (a real integer division,
since capacity is not a power of two) with a conditional-subtract
wrap_add helper. Left the oversampler's % TAPS_PER_PHASE alone - that
divisor is a compile-time constant LLVM already turns into a multiply.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LU1NRL7fAaZTtFMTTPn2b1
Rework Low Curve from a one-sided saturating bend into a bipolar mid-bulge: the
low region is anchored at BOTH the silence floor and the knee, Low Slope tilts
the straight line between them, and Low Curve (-1..1) bows that line in the
middle (4*t*(1-t), peak +/-12 dB) without moving either endpoint. Positive
bulges up (boost the quiet middle), negative down (suppress). Still serial.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a per-channel below-threshold shaper composed in series ahead of the comp:
gain = low_shape(level) + comp(level + low_shape(level)). The compressor's
threshold now sees the shaped level, so a Low Slope boost lifts quiet material
up into compression (and a cut pulls it out). Anchored at the -60 dB silence
floor. Low Curve bends the shaper toward a bounded saturation so the serial
composition doesn't blow up (0 = straight line).
gain_computer split into comp_gain_db + low_gain_db and composed; shared with
the editor gain-curve display. Slider order rearranged to read in signal order
(pre-gain -> low shaper -> compressor -> output). Defaults (slope 1, curve 0)
reproduce the plain compressor; 17 tests pass.
Known: the bipolar behaviour isn't final yet (milestone commit).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add editor/gain_curve.rs: a square panel beside the scrolling plot showing the
channel transfer (out vs in, -60..0 dB) for whichever channel the plot tab
selects. Plots the full wet path — out = (in + pre_gain) + gain_reduction + makeup
— using the shared Compressor::gain_computer (now pub) so it matches the DSP and
the GR meter. Unity-reference diagonal + threshold marker; mix not folded in.
Update README structure/status to reflect the editor/ widget module and the
completed Stage 6 visualisers.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace the per-channel bypass toggle with a smoothed dry/wet `mix` (0..100%,
default 100%). The blend is applied at the compressor output:
out = delayed_input * ((1 - mix) + mix * wet_gain)
Dry and wet share the same delayed input, so it's phase-aligned (parallel
compression, no comb filtering). mix=0 is bit-identical to the old bypass.
The detector now runs even at mix 0, so the GR meter shows the wet gain
reduction regardless of mix, while the level/plot out trace reads the mixed
output. "Bands at 0% = simple full-band comp via All" still holds.
Update README + parameter docs (bypass -> mix).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Stage 6 begins. Add lock-free Meters (atomics shared audio->GUI) with a
peak-with-decay ballistic, published once per block and gated on the editor
being open. Editor draws a per-channel level + gain-reduction meter panel and
a ceiling lamp fed by the output limiter. Compressor/Limiter expose
gain_reduction_db() for this.
Also add a smoothed per-channel pre-gain applied before each compressor (and
before the All compressor), driving the signal into compression and on into
the limiter for a compressed semi-distortion. Pairs with makeup for full
per-channel input/output gain-staging. Compressor DSP untouched; 16 tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Move parameter structs, defaults, and build_settings into src/params.rs;
move the egui editor into src/editor.rs. lib.rs now holds only the plugin
shell, DSP wiring, and process(). No behavior change (16/16 tests pass).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Upgrades the brickwall limiter from sample-peak to true-peak (inter-sample).
- src/dsp/oversampler.rs: 4x polyphase windowed-sinc (4 phases x 12 taps, Blackman,
each phase normalized to unity DC). Detection-only: max_true_peak() returns the
inter-sample max magnitude and discards the upsampled samples; the audio path is
untouched. Built in prepare(), no realtime allocation. Cost ~ one base-rate FIR
per channel; its small group delay is absorbed by the limiter look-ahead, so no
added reported latency.
- src/dsp/limiter.rs: detector peak = max(sample_peak, oversampler.max_true_peak());
targets a 0.3 dB margin under the ceiling to cover the 4x detection residual.
- 16 unit tests (2 new: detects ~3 dB fs/4 inter-sample overshoot; preserves DC
amplitude). README/docs updated: Stage 4 complete.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds the output limiter stage after the 'All' channel. Guarantees the output
never exceeds the ceiling: fixed 1.5 ms look-ahead, stereo-linked sliding-max
peak detection over the look-ahead window -> gain = ceiling/window_max, decoupled
smoothing (fast attack / user release), and a final clamp as the hard guarantee.
- src/dsp/limiter.rs: Limiter (sample-peak; true-peak via oversampling is 4b)
- src/lib.rs: wired as final stage; new globals output_ceiling_db (-24..0) and
limiter_release_ms; latency now the constant three-stage total (bands+All+limiter);
two UI sliders added to the global row
- 14 unit tests (4 new: ceiling guarantee on spikes, loud-sine limiting,
transparency below ceiling, latency)
- README/docs updated (Stage 4 split into 4a done / 4b oversampling)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Investigated the planned global FTZ for the IIR filters and found NIH-plug already
handles it: process_wrapper wraps process()/reset() in a ScopedFtz guard that
enables CPU Flush-To-Zero (x86 MXCSR bit 15 / AArch64 FPCR bit 24, via inline asm,
restored on drop) on the vst3, clap, and standalone paths. SSE is baseline on
x86_64 so FTZ is always active for our build.
So adding our own guard would just duplicate the framework. Instead, removed the
now-redundant flush_denormal() from compressor.rs (the biquads and envelope/RMS
tails already relied on this FTZ) for a single consistent story, and rewrote the
README 'Denormal flushing' note to document that the framework handles it (FTZ,
not DAZ — sufficient for our feed-forward IIR).
No functional change; 10 unit tests still pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Splits the input into low/mid/high with a Linkwitz-Riley 24 dB/oct crossover,
compresses each band, sums them, then runs the sum through a fourth 'All'
compressor. Bypassing the three bands collapses the plugin to a simple full-band
comp driven by 'All' (the crossover sums flat in magnitude).
- src/dsp/biquad.rs: generic RBJ biquad (Transposed Direct Form II), LP/HP/AP
- src/dsp/crossover.rs: 3-band LR4 filterbank; lower band all-pass-compensated at
the higher crossover so the bands sum to flat magnitude (an all-pass, not a
bit-exact null — that only holds for linear-phase FIR). Mirrors nih-plug's
crossover plugin design.
- src/lib.rs: 4 Compressor instances (low/mid/high/all) + Crossover; params
restructured to 4 nested CompressorParams (id_prefix low/mid/high/all) plus
global crossover_low_hz/crossover_high_hz/look_ahead_ms; 4-column lo|mid|hi|all
egui UI; latency = two series stages (bands + all), constant, reported once
- 10 unit tests (adds biquad LP/AP magnitude, crossover flat-magnitude
reconstruction, band-split sanity)
- README: Stage 3 marked done; corrected the 'sum flat' expectation to flat
magnitude (IIR LR sums to an all-pass, not a time-domain null)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Compressor: switchable peak / RMS detection (EnumParam<DetectionMode> in lib.rs
-> use_rms bool in CompressorSettings; DSP stays framework-agnostic). RMS is a
one-pole running mean of the linked squared level with a hardcoded 5 ms window,
updated whenever active so peak<->RMS switching is seamless. New unit test
(RMS compresses a sine less than peak); 6 tests total.
- README: reconciled Implementation Order checklists with actual progress
(Stages 1-2 done; look-ahead/latency + basic UI pulled forward), annotated the
project structure (implemented vs planned), and corrected the Latency and
Denormal-flushing notes to match the code (set_latency_samples once / constant
latency; in-code denormal flush). Overview and goals left unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Implements the single full-band feed-forward compressor (the engine that will be
reused per band + for the 'All' channel). Design follows Giannoulis et al. 2012:
log-domain gain computer with a quadratic soft knee feeding a smooth decoupled
peak detector for attack/release ballistics. Stereo-linked peak detection.
Look-ahead uses a fixed audio delay with a constant reported latency (set once in
initialize); the knob only moves the detector tap within that delay. This avoids
renegotiating latency from process(), which crashed FL Studio when the look-ahead
was adjusted during playback.
- src/dsp/compressor.rs: Compressor + CompressorSettings, RT-safe (no alloc in
process; buffers sized in prepare; envelope denormals flushed in-code)
- src/lib.rs: nested CompressorParams (threshold/ratio/knee/attack/release/makeup/
bypass) + global look-ahead; egui ParamSlider grid; latency reported once
- 5 unit tests (static curve, knee continuity, steady-state convergence, constant
latency); Cargo.toml lib crate-type added so tests link
- README: 'All' channel architecture already documented; look-ahead spec updated
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Mikkeli Matlock
irene