EqualizerNBands.h

Go to the documentation of this file.

#pragma once
#include <math.h>
#include <string.h>
 
#include <limits>
 
#include "AudioTools/CoreAudio/AudioBasic/Collections/Vector.h"
#include "AudioTools/CoreAudio/AudioOutput.h"
#include "AudioTools/CoreAudio/AudioStreams.h"
#include "AudioToolsConfig.h"
 
namespace audio_tools {
 
template <typename SampleT = int16_t, typename AccT = int64_t,
          int NUM_TAPS = 128, int NUM_BANDS = 12>
class EqualizerNBands : public ModifyingStream {
 public:
  EqualizerNBands() { setBandGains(0.0f); };
  EqualizerNBands(Print& out) : EqualizerNBands() { setOutput(out); }
 
  EqualizerNBands(Stream& in) : EqualizerNBands() { setStream(in); }
 
  EqualizerNBands(AudioOutput& out) : EqualizerNBands() {
    setOutput(out);
    out.addNotifyAudioChange(*this);
  }
 
  EqualizerNBands(AudioStream& stream) : EqualizerNBands() {
    setStream(stream);
    stream.addNotifyAudioChange(*this);
  }
 
  ~EqualizerNBands() { end(); }
 
  void setStream(Stream& io) override {
    p_out = &io;
    p_io = &io;
  };
 
  void setOutput(Print& out) override { p_out = &out; }
 
  bool begin(AudioInfo info) {
    setAudioInfo(info);
    return begin();
  }
 
  bool begin() override {
    currentSampleRate = audioInfo().sample_rate;
    if (currentSampleRate <= 0) {
      LOGE("Invalid sample rate: %d", currentSampleRate);
      return false;
    }
    setupFrequencies(currentSampleRate);
    preCalculateWindow();
 
    // Initialize double-buffering pointers
    activeKernel = kernelA;
    updateKernel = kernelB;
 
    // Initialize both kernels with pass-through (identity) filter
    initializeKernel(kernelA);
    initializeKernel(kernelB);
 
    // assign output or source
    if (p_io) filtered.setStream(*p_io);
    if (p_out) filtered.setOutput(*p_out);
    filtered.begin(audioInfo());
 
    // set filters for all channels
    fir_vector.resize(audioInfo().channels);
    for (int ch = 0; ch < audioInfo().channels; ch++) {
      fir_vector[ch].setKernel(activeKernel);
      filtered.setFilter(ch, &fir_vector[ch]);
    }
 
    // calculate the kernel
    bool rc = updateFIRKernel();
 
    // Log the initial band settings for visibility
    for (int band = 0; band < NUM_BANDS; band++) {
      LOGI("Band %d: Freq=%.2fHz, Gain=%.2fdB", band, getBandFrequency(band),
           getBandDB(band));
    }
    return rc;
  }
 
  void end() {
    fir_vector.clear();
    currentSampleRate = 0;
  }
 
  bool setBandGain(int band, float volume) {
    // Map -1.0 to 1.0 to -90DB to +12dB
    float vol_db = volume < 0 ? map<float>(volume, -1.0f, 0.0f, -90.0f, 0.0f) : map<float>(volume, 0.0f, 1.0f, 0.0f, 12.0f);
    return setBandDB(band, vol_db);
  }
 
  bool setBandDB(int band, float gainDb) {
    if (band < 0 || band >= NUM_BANDS) return false;
    float db = min(gainDb, 12.0f);
    db = max(db, -90.0f);
    pendingGains[band] = db;
    gainsDirty = true;
    return true;
  }
 
  bool setBandGains(float volume) {
    for (int band = 0; band < NUM_BANDS; band++) {
      setBandGain(band, volume);
    }
    return true;
  }
 
  float getBandGain(int band) {
    if (band < 0 || band >= NUM_BANDS) return 0.0f;
    return map<float>(pendingGains[band], -12.0f, 12.0f, -1.0f, 1.0f);
  }
 
  float getBandDB(int band) const {
    if (band < 0 || band >= NUM_BANDS) return 0.0f;
    return pendingGains[band];
  }
 
  float getBandFrequency(int band) const {
    if (band < 0 || band >= NUM_BANDS) return 0.0f;
    return centerFreqs[band];
  }
 
  int getBandCount() const { return NUM_BANDS; }
 
  void setAutoUpdate(bool enabled) { autoUpdate = enabled; }
 
  // update the FIR kernel after changing gains
  bool update() { return updateFIRKernel(); }
 
  size_t write(const uint8_t* data, size_t len) override {
    maybeUpdateKernel();
    return filtered.write(data, len);
  }
 
  size_t readBytes(uint8_t* data, size_t len) override {
    maybeUpdateKernel();
    return filtered.readBytes(data, len);
  }
 
  void flush() override { filtered.flush(); }
 
 protected:
  // Simple re-entrancy guard: prevents concurrent kernel updates from
  // corrupting scratch buffers / updateKernel.
  volatile bool isUpdating = false;
  // Indicates new gains are pending and kernel should be refreshed.
  volatile bool gainsDirty = false;
  // Auto-update kernel during streaming. Default is false to preserve
  // explicit update() behavior.
  bool autoUpdate = false;
 
  // Custom FIR Filter Class
  class EQFIRFilter : public Filter<SampleT> {
   public:
    EQFIRFilter() : activeKernel(nullptr) {}
 
    void setKernel(volatile int16_t* kernel) { activeKernel = kernel; }
 
    SampleT process(SampleT sample) override {
      if (activeKernel == nullptr) {
        LOGE("Kernel not set!");
        return sample;  // Pass-through if no kernel set
      }
 
      xHistory[idxHist] = sample;
 
      // Use AccT to prevent overflow and/or allow float accumulation
      AccT acc = (AccT)0;
      int idx = idxHist;
 
      for (int n = 0; n < NUM_TAPS; n++) {
        // Coefficients are Q15 int16_t
        acc += (AccT)xHistory[idx] * (AccT)activeKernel[n];
        if (--idx < 0) idx = NUM_TAPS - 1;
      }
 
      if (++idxHist >= NUM_TAPS) idxHist = 0;
 
      // Convert back from Q15 and saturate
      return fromQ15(acc);
    }
 
   protected:
    SampleT xHistory[NUM_TAPS] = {(SampleT)0};
    int idxHist = 0;
    volatile int16_t* activeKernel = nullptr;  // Pointer to active kernel
 
    static inline SampleT fromQ15(AccT acc) {
      // Default path: integer SampleT (current behavior)
      if constexpr (std::numeric_limits<SampleT>::is_integer) {
        // Shift back from Q15
        if constexpr (std::numeric_limits<AccT>::is_integer) {
          acc >>= 15;
        } else {
          acc = acc / (AccT)(1 << 15);
        }
 
        // Saturation to SampleT range
        const AccT hi = (AccT)std::numeric_limits<SampleT>::max();
        const AccT lo = (AccT)std::numeric_limits<SampleT>::min();
        if (acc > hi) acc = hi;
        if (acc < lo) acc = lo;
        return (SampleT)acc;
      } else {
        // Floating output sample: scale Q15 back to approximately [-1..1]
        return (SampleT)(acc / (AccT)(1 << 15));
      }
    }
  };
 
  // Q15 range is [-32768, 32767]. Use 32767 for +1.0 to avoid overflow/wrap.
  static constexpr float Q15_SCALE = 32767.0f;
  float centerFreqs[NUM_BANDS];
  // Gains in dB used by the kernel design.
  float gains[NUM_BANDS] = {0};
  // Gains written by user-facing setters. Copied to gains transactionally.
  float pendingGains[NUM_BANDS] = {0};
 
  // Scratch buffer used during kernel design. Kept as member to avoid static
  // storage (shared across instances and non-reentrant).
  float tempFloat[NUM_TAPS] = {0};
 
  // Double-buffering: two kernels for thread-safe updates
  int16_t kernelA[NUM_TAPS];
  int16_t kernelB[NUM_TAPS];
  volatile int16_t* activeKernel;  // Pointer to currently used kernel
  volatile int16_t* updateKernel;  // Pointer to kernel being updated
 
  float windowCoeffs[NUM_TAPS];  // Pre-calculated Blackman window
  int currentSampleRate = 0;
 
  Print* p_out = nullptr;        
  Stream* p_io = nullptr;      
  Vector<EQFIRFilter> fir_vector;  
  FilteredStream<SampleT, SampleT> filtered;
 
  // Centralized interrupt guards to avoid repeated #if defined(ARDUINO) blocks.
  inline void enterCritical() {
#if defined(ARDUINO)
    noInterrupts();
#endif
  }
 
  inline void exitCritical() {
#if defined(ARDUINO)
    interrupts();
#endif
  }
 
  // Update kernel if any gain changes are pending.
  inline void maybeUpdateKernel() {
    if (!autoUpdate) return;
    if (gainsDirty) {
      if (updateFIRKernel()) {
        gainsDirty = false;
      }
    }
  }
 
  template <typename T>
  float map(T x, T in_min, T in_max, T out_min, T out_max) {
    return (x - in_min) * (out_max - out_min) / (in_max - in_min) + out_min;
  }
 
  // Helper Sinc
  float sinc(float x) {
    if (fabsf(x) < 1e-8f) return 1.0f;
    return sinf(PI * x) / (PI * x);
  }
 
  // Setup Center Frequencies Logarithmically spaced between 20Hz and Nyquist
  void setupFrequencies(int sampleRate) {
    if (NUM_BANDS <= 0) return;
    float fMin = log10f(20.0f);
    float fMax = log10f(sampleRate / 2.0f);  // Nyquist frequency
    if (NUM_BANDS == 1) {
      centerFreqs[0] = powf(10.0f, (fMin + fMax) * 0.5f);
      LOGD("Only one band: center frequency set to %.2f Hz", centerFreqs[0]);
      return;
    }
    float step = (fMax - fMin) / (float)(NUM_BANDS - 1);
    for (int i = 0; i < NUM_BANDS; i++) {
      centerFreqs[i] = powf(10.0f, fMin + step * (float)i);
      LOGD("Band %d: center frequency = %.2f Hz", i, centerFreqs[i]);
    }
  }
 
  // Pre-calculate Blackman window coefficients (performance optimization)
  void preCalculateWindow() {
    const float N_minus_1 = (float)(NUM_TAPS - 1);
    for (int n = 0; n < NUM_TAPS; n++) {
      windowCoeffs[n] = 0.42f - 0.5f * cosf(2.0f * PI * n / N_minus_1) +
                        0.08f * cosf(4.0f * PI * n / N_minus_1);
    }
  }
 
  // Initialize a kernel to pass-through (identity filter)
  void initializeKernel(volatile int16_t* kernel) {
    const int M = (NUM_TAPS - 1) / 2;
    for (int i = 0; i < NUM_TAPS; i++) {
      if (i == M) {
        kernel[i] = (int16_t)Q15_SCALE;  // Unity gain at center tap
      } else {
        kernel[i] = 0;
      }
    }
  }
 
  bool updateFIRKernel() {
    if (currentSampleRate <= 0) {
      LOGE("Invalid sample rate: %d", currentSampleRate);
      return false;  // Not initialized yet
    }
 
    // Prevent concurrent updates (e.g., from multiple tasks/threads).
    // We keep the critical section tiny: only the check/set of the flag.
    enterCritical();
    if (isUpdating) {
      exitCritical();
      return false;
    }
    isUpdating = true;
    exitCritical();
 
    // Transactional gain update: copy user-updated pendingGains into gains.
    // Keep this critical section short to avoid blocking audio processing.
    enterCritical();
    memcpy(gains, pendingGains, sizeof(gains));
    exitCritical();
 
    memset(tempFloat, 0, sizeof(tempFloat));
 
    const int M = (NUM_TAPS - 1) / 2;
    const float sampleRateFloat = (float)currentSampleRate;
 
    // Base impulse (Pass-through)
    tempFloat[M] = 1.0f;
 
    for (int i = 0; i < NUM_BANDS; i++) {
      // Skip bands with 0dB gain to save precision
      if (fabs(gains[i]) < 0.1f) continue;
 
      // Calculate Linear Gain Delta
      // If gain is +6dB (2.0x), we add (2.0 - 1.0) = +1.0 to the impulse
      float linGain = powf(10.0f, gains[i] / 20.0f) - 1.0f;
 
      // Use actual sample rate
      float fL_hz = centerFreqs[i] * 0.707f;  // Lower Edge (-3dB point)
      float fH_hz = centerFreqs[i] * 1.414f;  // Upper Edge (+3dB point)
 
      // Enforce minimum bandwidth so the windowed-sinc FIR can resolve
      // this band.  The Blackman window main-lobe width is ~4/N in
      // normalised frequency, i.e. 4*Fs/N Hz.  Bands narrower than that
      // are effectively invisible to the filter and produce a near-flat
      // (no-effect) response.
      float minBwHz = 4.0f * sampleRateFloat / (float)NUM_TAPS;
      float actualBwHz = fH_hz - fL_hz;
      if (actualBwHz < minBwHz) {
        float expand = (minBwHz - actualBwHz) * 0.5f;
        fL_hz = fL_hz - expand;
        fH_hz = fH_hz + expand;
        if (fL_hz < 1.0f) fL_hz = 1.0f;
      }
 
      float fL = fL_hz / sampleRateFloat;
      float fH = fH_hz / sampleRateFloat;
 
      // Clamp to valid normalized frequency range [0, 0.5] (Nyquist)
      if (fL < 0.0f) fL = 0.0f;
      if (fH < 0.0f) fH = 0.0f;
      if (fL > 0.5f) fL = 0.5f;
      if (fH > 0.5f) fH = 0.5f;
      if (fH <= fL) continue;
 
      // Evaluate the windowed bandpass magnitude at the center frequency
      // so we can normalise to unity.  The Blackman window reduces the
      // passband peak below 1.0; without compensation the actual
      // boost/cut is weaker than requested.
      float wCenter = 2.0f * PI * centerFreqs[i] / sampleRateFloat;
      float hReal = 0.0f;
      float hImag = 0.0f;
      for (int n = 0; n < NUM_TAPS; n++) {
        float nM = (float)(n - M);
        float bpW = ((2.0f * fH * sinc(2.0f * fH * nM)) -
                     (2.0f * fL * sinc(2.0f * fL * nM))) *
                    windowCoeffs[n];
        hReal += bpW * cosf(wCenter * n);
        hImag -= bpW * sinf(wCenter * n);
      }
      float bpMag = sqrtf(hReal * hReal + hImag * hImag);
      float normFactor = (bpMag > 1e-6f) ? (1.0f / bpMag) : 1.0f;
 
      for (int n = 0; n < NUM_TAPS; n++) {
        float nM = (float)(n - M);
 
        // Use pre-calculated window coefficients
        float window = windowCoeffs[n];
 
        // Bandpass filter: highpass - lowpass
        float bp = (2.0f * fH * sinc(2.0f * fH * nM)) -
                   (2.0f * fL * sinc(2.0f * fL * nM));
 
        // Add the normalised, weighted bandpass to the master kernel
        tempFloat[n] += bp * window * normFactor * linGain;
      }
    }
 
    // Update the inactive kernel (no interruption needed for writes)
    for (int i = 0; i < NUM_TAPS; i++) {
      // DIRECT CONVERSION (No Auto-Normalization)
      // This ensures +6dB actually outputs louder voltage
      int32_t q = (int32_t)(tempFloat[i] * Q15_SCALE);
 
      // Hard Clip the Kernel Coefficients to prevent wrap-around
      if (q > 32767) q = 32767;
      if (q < -32768) q = -32768;
      updateKernel[i] = (int16_t)q;
    }
 
    // Atomically swap the kernel pointers
    // This is the only operation that needs to be atomic
    enterCritical();
    volatile int16_t* temp = activeKernel;
    activeKernel = updateKernel;
    updateKernel = temp;
 
    // Update filter references to new active kernel
    for (auto& fir : fir_vector) {
      fir.setKernel(activeKernel);
    }
    exitCritical();
 
    // Release re-entrancy guard
    enterCritical();
    isUpdating = false;
    gainsDirty = false;
    exitCritical();
 
    LOGI("FIR kernel updated with new gains for %d bands /%d taps.", NUM_BANDS,
         NUM_TAPS);
    return true;
  }
};
 
}  // namespace audio_tools