AlgorithmicReverb.h File Reference

World-class 16-line FDN reverb with Jot absorption and Hadamard mixing. More...

#include "../Core/RingBuffer.h"
#include "../Core/DryWetMixer.h"
#include "../Core/DspMath.h"
#include "../Core/AudioSpec.h"
#include "../Core/AudioBuffer.h"
#include "../Core/DenormalGuard.h"
#include "../Core/Biquad.h"
#include "../Core/StateBlob.h"
#include <algorithm>
#include <array>
#include <atomic>
#include <cmath>
#include <cstddef>
#include <cstdint>
#include <utility>
#include <vector>

Include dependency graph for AlgorithmicReverb.h:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Classes
class	dspark::AlgorithmicReverb< T >
	16-line FDN reverb with Jot absorption and 6 presets. More...
struct	dspark::AlgorithmicReverb< T >::SmoothRandomLFO
	Hermite-interpolated random noise generator for organic modulation. More...
struct	dspark::AlgorithmicReverb< T >::CachedParams
struct	dspark::AlgorithmicReverb< T >::PresetValues

Namespaces
namespace	dspark
	Main namespace for the DSPark framework.

Detailed Description

World-class 16-line FDN reverb with Jot absorption and Hadamard mixing.

Architecture:

Input (mono sum)
  |
  v
[Pre-delay: 0-200ms]
  |
  v
[Input Diffusion: 8 cascaded allpass, 1.0-9.5ms]
  |
  +--> [Early Reflections: 40 taps with progressive HF absorption, L/R decorrelated]
  |
  +--> [ER-to-Late gap]
  |       |
  |       v
  |    [Parallel Allpass Diffuser: 16 parallel AP + Hadamard -> 16 delay + Hadamard]
  |    (2-step, 16^2 = 256 echo paths, each FDN line gets unique dense input)
  |       |
  |       v
  |    [FDN Core: 16 delay lines]
  |      +- Read with dual smooth-random-LFO modulated delay
  |      +- Householder 16x16 reflection (I - (2/N)J, lossless) feedback mix
  |      +- Jot absorption filter (1st-order shelving) per line
  |      +- Bass shelf (1-pole) per line
  |      +- 2 feedback allpass per line
  |      +- DC blocker + soft limiter
  |      +- Write back + input injection
  |       |
  |       v
  |    [Output: sign-weighted + Dattorro multi-tap]
  |       |
  |       v
  |    [Output Diffusion: 2 allpass/channel, L/R decorrelated]
  |
  v
[Combine early + late] -> [Tone EQ: Biquad LP + HP (12 dB/oct)] -> DryWetMixer -> Output

Key features:

• Jot absorption filter (Jot 1991): 1st-order pole-zero shelving IIR per delay line for smooth frequency-dependent decay, the #1 factor for natural sound. DC and Nyquist gains anchor the mid/HF T60 exactly; the transition sits at the high crossover. Separate bass shelf for independent LF control.
• Householder 16x16 FDN feedback (I - (2/N)J): unitary reflection, eigenvalues {-1 once, +1 x15}, |lambda|=1 so it is lossless and zero-coloring. (The Hadamard 16x16 below is the input diffuser's mix, not the feedback matrix.)
• Parallel allpass diffuser (Signalsmith-inspired): 16 parallel allpass
  • 2-step Hadamard mixing (all eigenvalues +/-1) -> 256 unique echo paths per input sample. Each FDN line receives a different, densely-mixed version of the input.
• Feedback allpass: 2 regular allpass per delay line for in-loop density
• Output diffusion: 2 allpass per channel with decorrelated delays, smears residual temporal patterns
• Smooth random modulation (Lexicon-style): Hermite-interpolated noise replaces periodic sine LFOs for organic character
• Multi-tap output (Dattorro-style): 7 taps/channel from different delay line positions for true temporal decorrelation
• Tone correction EQ: Biquad high/low cut (12 dB/oct) for tonal shaping
• Early reflections: 40-tap with progressive frequency absorption simulating wall absorption; late taps are naturally darker
• Soft saturation: fast branchless rational soft-clip (transparent below +/-1), more musical than a hard clamp, prevents pipeline stalls.
• Allpass interpolation: in modulated FDN reads, preserves HF over hundreds of feedback iterations (linear/cubic causes cumulative dulling)
• Stereo width: M/S width control on late reverb tail
• Eco quality mode (setQuality): opt-in reduced engine (8 FDN lines, control-rate modulation, linear allpass interpolation, no extra output taps, single-stage input scatter, 12 early taps) at a fraction of the CPU cost, for embedded and other constrained targets. The default Full quality path is bit-identical to previous releases.

Threading: prepare() belongs to the setup thread (allocates). processBlock(), processSample() and reset() belong to the audio thread. All setters are lock-free atomic publications, safe from any thread; topology changes (type, quality) and coefficient refreshes are drained at the start of the next processBlock()/processSample() on the audio thread. Non-finite setter arguments are ignored.

Four levels of API complexity:

• Level 1: reverb.setType(Hall);
• Level 2: reverb.setSize(0.7f); reverb.setDamping(0.3f);
• Level 3: reverb.setHighDecayMultiplier(0.4f);
• Level 4: Inherit and override protected members.

References:

• Jot & Chaigne (1991): FDN with frequency-dependent decay (shelving absorption)
• Dattorro (1997, JAES): multi-tap output, plate topology
• Griesinger / Lexicon 480L: random modulation, Spin/Wander
• Sean Costello / Valhalla DSP: practical FDN design, absorbent allpass
• Valimaki et al. (2012, IEEE): "50 Years of Artificial Reverberation"
• Smith (CCRMA): Hadamard matrices, prime power delay selection

Dependencies: RingBuffer.h, DryWetMixer.h, DspMath.h, Biquad.h, AudioSpec.h, AudioBuffer.h, DenormalGuard.h, StateBlob.h.

dspark::AlgorithmicReverb<float> reverb;
reverb.prepare(spec);
reverb.setType(dspark::AlgorithmicReverb<float>::Type::Hall);
reverb.setDecay(2.0f);
reverb.setMix(0.3f);
reverb.processBlock(buffer);
 
// Advanced: tune frequency-dependent decay
reverb.setHighDecayMultiplier(0.4f);  // HF decays 2.5x faster
reverb.setBassDecayMultiplier(1.3f);  // bass lingers 1.3x longer

Definition in file AlgorithmicReverb.h.