react-native-audio-api 0.11.0-alpha.0 → 0.11.0-nightly-bfab178-20251107

package/common/cpp/audioapi/core/sources/StreamerNode.cpp

@@ -25,14 +25,12 @@ StreamerNode::StreamerNode(BaseAudioContext *context)
       codecpar_(nullptr),
       pkt_(nullptr),
       frame_(nullptr),
-      pendingFrame_(nullptr),
-      bufferedBus_(nullptr),
-      bufferedBusIndex_(0),
-      maxBufferSize_(0),
-      audio_stream_index_(-1),
       swrCtx_(nullptr),
       resampledData_(nullptr),
-      maxResampledSamples_(0) {}
+      bufferedBus_(nullptr),
+      audio_stream_index_(-1),
+      maxResampledSamples_(0),
+      processedSamples_(0) {}
 
 StreamerNode::~StreamerNode() {
   cleanup();
@@ -66,26 +64,22 @@ bool StreamerNode::initialize(const std::string &input_url) {
     return false;
   }
 
-  maxBufferSize_ = BUFFER_LENGTH_SECONDS * codecCtx_->sample_rate;
-  // If decoding is faster than playing, we buffer few seconds of audio
-  bufferedBus_ = std::make_shared<AudioBus>(
-      maxBufferSize_, codecpar_->ch_layout.nb_channels, codecCtx_->sample_rate);
-
   channelCount_ = codecpar_->ch_layout.nb_channels;
   audioBus_ = std::make_shared<AudioBus>(
       RENDER_QUANTUM_SIZE, channelCount_, context_->getSampleRate());
 
+  auto [sender, receiver] = channels::spsc::channel<
+      StreamingData,
+      channels::spsc::OverflowStrategy::WAIT_ON_FULL,
+      channels::spsc::WaitStrategy::ATOMIC_WAIT>(CHANNEL_CAPACITY);
+  sender_ = std::move(sender);
+  receiver_ = std::move(receiver);
+
   streamingThread_ = std::thread(&StreamerNode::streamAudio, this);
-  streamFlag.store(true);
   isInitialized_ = true;
   return true;
 }
 
-void StreamerNode::stop(double when) {
-  AudioScheduledSourceNode::stop(when);
-  streamFlag.store(false);
-}
-
 bool StreamerNode::setupResampler() {
   // Allocate resampler context
   swrCtx_ = swr_alloc();
@@ -122,30 +116,27 @@ bool StreamerNode::setupResampler() {
 }
 
 void StreamerNode::streamAudio() {
-  while (streamFlag.load()) {
-    if (pendingFrame_ != nullptr) {
-      if (!processFrameWithResampler(pendingFrame_)) {
+  while (!isNodeFinished_.load(std::memory_order_acquire)) {
+    if (av_read_frame(fmtCtx_, pkt_) < 0) {
+      return;
+    }
+    if (pkt_->stream_index == audio_stream_index_) {
+      if (avcodec_send_packet(codecCtx_, pkt_) != 0) {
         return;
       }
-    } else {
-      if (av_read_frame(fmtCtx_, pkt_) < 0) {
+      if (avcodec_receive_frame(codecCtx_, frame_) != 0) {
         return;
       }
-      if (pkt_->stream_index == audio_stream_index_) {
-        if (avcodec_send_packet(codecCtx_, pkt_) != 0) {
-          return;
-        }
-        if (avcodec_receive_frame(codecCtx_, frame_) != 0) {
-          return;
-        }
-        if (!processFrameWithResampler(frame_)) {
-          return;
-        }
+      if (!processFrameWithResampler(frame_)) {
+        return;
       }
-      av_packet_unref(pkt_);
     }
-    std::this_thread::sleep_for(std::chrono::milliseconds(10));
+    av_packet_unref(pkt_);
   }
+  StreamingData dummy;
+  while (receiver_.try_receive(dummy) ==
+         channels::spsc::ResponseStatus::SUCCESS)
+    ; // clear the receiver
 }
 
 std::shared_ptr<AudioBus> StreamerNode::processNode(
@@ -154,34 +145,44 @@ std::shared_ptr<AudioBus> StreamerNode::processNode(
   size_t startOffset = 0;
   size_t offsetLength = 0;
   updatePlaybackInfo(processingBus, framesToProcess, startOffset, offsetLength);
+  isNodeFinished_.store(isFinished(), std::memory_order_release);
 
   if (!isPlaying() && !isStopScheduled()) {
     processingBus->zero();
     return processingBus;
   }
 
-  // If we have enough buffered data, copy to output bus
-  if (bufferedBusIndex_ >= framesToProcess) {
-    Locker locker(mutex_);
+  int bufferRemaining = bufferedBusSize_ - processedSamples_;
+  int alreadyProcessed = 0;
+  if (bufferRemaining < framesToProcess) {
+    if (bufferedBus_ != nullptr) {
+      for (int ch = 0; ch < processingBus->getNumberOfChannels(); ch++) {
+        memcpy(
+            processingBus->getChannel(ch)->getData(),
+            bufferedBus_->getChannel(ch)->getData() + processedSamples_,
+            bufferRemaining * sizeof(float));
+      }
+      framesToProcess -= bufferRemaining;
+      alreadyProcessed += bufferRemaining;
+    }
+    StreamingData data;
+    auto res = receiver_.try_receive(data);
+    if (res == channels::spsc::ResponseStatus::SUCCESS) {
+      bufferedBus_ = std::make_shared<AudioBus>(std::move(data.bus));
+      bufferedBusSize_ = data.size;
+      processedSamples_ = 0;
+    } else {
+      bufferedBus_ = nullptr;
+    }
+  }
+  if (bufferedBus_ != nullptr) {
     for (int ch = 0; ch < processingBus->getNumberOfChannels(); ch++) {
       memcpy(
-          processingBus->getChannel(ch)->getData(),
-          bufferedBus_->getChannel(ch)->getData(),
-          offsetLength * sizeof(float));
-
-      memmove(
-          bufferedBus_->getChannel(ch)->getData(),
-          bufferedBus_->getChannel(ch)->getData() + offsetLength,
-          (maxBufferSize_ - offsetLength) * sizeof(float));
+          processingBus->getChannel(ch)->getData() + alreadyProcessed,
+          bufferedBus_->getChannel(ch)->getData() + processedSamples_,
+          framesToProcess * sizeof(float));
     }
-    bufferedBusIndex_ -= offsetLength;
-  } else {
-    if (VERBOSE)
-      printf(
-          "Buffer underrun: have %zu, need %zu\n",
-          bufferedBusIndex_,
-          (size_t)framesToProcess);
-    processingBus->zero();
+    processedSamples_ += framesToProcess;
   }
 
   return processingBus;
@@ -220,22 +221,21 @@ bool StreamerNode::processFrameWithResampler(AVFrame *frame) {
     return false;
   }
 
-  // Check if converted data fits in buffer
-  if (bufferedBusIndex_ + converted_samples > maxBufferSize_) {
-    pendingFrame_ = frame;
+  // if we would like to finish dont copy anything
+  if (this->isFinished()) {
     return true;
-  } else {
-    pendingFrame_ = nullptr;
   }
-
-  // Copy converted data to our buffer
-  Locker locker(mutex_);
+  auto bus = AudioBus(
+      static_cast<size_t>(converted_samples),
+      codecCtx_->ch_layout.nb_channels,
+      context_->getSampleRate());
   for (int ch = 0; ch < codecCtx_->ch_layout.nb_channels; ch++) {
     auto *src = reinterpret_cast<float *>(resampledData_[ch]);
-    float *dst = bufferedBus_->getChannel(ch)->getData() + bufferedBusIndex_;
+    float *dst = bus.getChannel(ch)->getData();
     memcpy(dst, src, converted_samples * sizeof(float));
   }
-  bufferedBusIndex_ += converted_samples;
+  StreamingData data{std::move(bus), static_cast<size_t>(converted_samples)};
+  sender_.send(std::move(data));
   return true;
 }
 
@@ -280,7 +280,7 @@ bool StreamerNode::setupDecoder() {
 }
 
 void StreamerNode::cleanup() {
-  streamFlag.store(false);
+  this->playbackState_ = PlaybackState::FINISHED;
   // cleanup cannot be called from the streaming thread so there is no need to
   // check if we are in the same thread
   streamingThread_.join();

package/common/cpp/audioapi/core/sources/StreamerNode.h

@@ -11,6 +11,7 @@
 #pragma once
 
 #include <audioapi/core/sources/AudioScheduledSourceNode.h>
+#include <audioapi/utils/AudioBus.h>
 
 #ifndef AUDIO_API_TEST_SUITE
 extern "C" {
@@ -27,8 +28,35 @@ extern "C" {
 #include <memory>
 #include <string>
 #include <atomic>
+#include <utility>
+#ifndef AUDIO_API_TEST_SUITE
+#include <audioapi/utils/SpscChannel.hpp>
 
-static bool constexpr VERBOSE = false;
+static constexpr audioapi::channels::spsc::OverflowStrategy STREAMER_NODE_SPSC_OVERFLOW_STRATEGY =
+    audioapi::channels::spsc::OverflowStrategy::WAIT_ON_FULL;
+static constexpr audioapi::channels::spsc::WaitStrategy STREAMER_NODE_SPSC_WAIT_STRATEGY =
+    audioapi::channels::spsc::WaitStrategy::ATOMIC_WAIT;
+#endif
+
+static constexpr bool VERBOSE = false;
+static constexpr int CHANNEL_CAPACITY = 32;
+
+struct StreamingData{
+  audioapi::AudioBus bus;
+  size_t size;
+  StreamingData() = default;
+  StreamingData(audioapi::AudioBus b, size_t s) : bus(b), size(s) {}
+  StreamingData(const StreamingData& data) : bus(data.bus), size(data.size) {}
+  StreamingData(StreamingData&& data) noexcept : bus(std::move(data.bus)), size(data.size) {}
+  StreamingData& operator=(const StreamingData& data) {
+    if (this == &data) {
+      return *this;
+    }
+    bus = data.bus;
+    size = data.size;
+    return *this;
+  }
+};
 
 namespace audioapi {
 
@@ -43,7 +71,6 @@ class StreamerNode : public AudioScheduledSourceNode {
    * @brief Initialize all necessary ffmpeg components for streaming audio
   */
   bool initialize(const std::string& inputUrl);
-  void stop(double when) override;
 
  protected:
   std::shared_ptr<AudioBus> processNode(const std::shared_ptr<AudioBus>& processingBus, int framesToProcess) override;
@@ -56,19 +83,20 @@ class StreamerNode : public AudioScheduledSourceNode {
   AVCodecParameters* codecpar_;
   AVPacket* pkt_;
   AVFrame* frame_; // Frame that is currently being processed
-  AVFrame* pendingFrame_; // Frame that is saved if bufferedBus is full
-  std::shared_ptr<AudioBus> bufferedBus_; // audio bus for buffering hls frames
-  size_t bufferedBusIndex_; // index in the buffered bus where we write the next frame
-  size_t maxBufferSize_; // maximum size of the buffered bus
-  int audio_stream_index_; // index of the audio stream channel in the input
   SwrContext* swrCtx_;
   uint8_t** resampledData_; // weird ffmpeg way of using raw byte pointers for resampled data
+
+  std::shared_ptr<AudioBus> bufferedBus_; // audio bus for buffering hls frames
+  size_t bufferedBusSize_; // size of currently buffered bus
+  int audio_stream_index_; // index of the audio stream channel in the input
   int maxResampledSamples_;
-  std::mutex mutex_;
+  size_t processedSamples_;
+
   std::thread streamingThread_;
-  std::atomic<bool> streamFlag; // Flag to control the streaming thread
-  static constexpr float BUFFER_LENGTH_SECONDS = 5.0f; // Length of the buffer in seconds
+  std::atomic<bool> isNodeFinished_; // Flag to control the streaming thread
   static constexpr int INITIAL_MAX_RESAMPLED_SAMPLES = 8192; // Initial size for resampled data
+  channels::spsc::Sender<StreamingData, STREAMER_NODE_SPSC_OVERFLOW_STRATEGY, STREAMER_NODE_SPSC_WAIT_STRATEGY> sender_;
+  channels::spsc::Receiver<StreamingData, STREAMER_NODE_SPSC_OVERFLOW_STRATEGY, STREAMER_NODE_SPSC_WAIT_STRATEGY> receiver_;
 
   /**
    * @brief Setting up the resampler

package/common/cpp/audioapi/core/utils/AudioNodeManager.cpp

@@ -1,5 +1,6 @@
 #include <audioapi/core/AudioNode.h>
 #include <audioapi/core/AudioParam.h>
+#include <audioapi/core/effects/ConvolverNode.h>
 #include <audioapi/core/sources/AudioScheduledSourceNode.h>
 #include <audioapi/core/utils/AudioNodeManager.h>
 #include <audioapi/core/utils/Locker.h>
@@ -221,6 +222,10 @@ inline bool AudioNodeManager::nodeCanBeDestructed(
   if constexpr (std::is_base_of_v<AudioScheduledSourceNode, U>) {
     return node.use_count() == 1 &&
         (node->isUnscheduled() || node->isFinished());
+  } else if constexpr (std::is_base_of_v<
+                           ConvolverNode,
+                           U>) { // convolver overrides disabling behavior
+    return node.use_count() == 1 && !node->isEnabled();
   }
   return node.use_count() == 1;
 }

package/common/cpp/audioapi/core/utils/Constants.h

@@ -1,5 +1,6 @@
 #pragma once
 
+#include <numbers>
 #include <cmath>
 #include <limits>
 
@@ -19,7 +20,7 @@ static constexpr float MOST_POSITIVE_SINGLE_FLOAT = static_cast<float>(std::nume
 static constexpr float MOST_NEGATIVE_SINGLE_FLOAT = static_cast<float>(std::numeric_limits<float>::lowest());
 static float LOG2_MOST_POSITIVE_SINGLE_FLOAT = std::log2(MOST_POSITIVE_SINGLE_FLOAT);
 static float LOG10_MOST_POSITIVE_SINGLE_FLOAT = std::log10(MOST_POSITIVE_SINGLE_FLOAT);
-static constexpr float PI = static_cast<float>(M_PI);
+static constexpr float PI = std::numbers::pi_v<float>;
 
 // buffer sizes
 static constexpr size_t PROMISE_VENDOR_THREAD_POOL_WORKER_COUNT = 4;

package/common/cpp/audioapi/core/utils/worklets/SafeIncludes.h

@@ -5,14 +5,8 @@
 #include <string>
 #include <memory>
 
-#ifdef __APPLE__
-  /// We cannot make any conditional logic inside podspec but it should automatically compile those files
-  /// they should be accessible if someone has react-native-worklets in node_modules
-  #if __has_include(<worklets/WorkletRuntime/WorkletRuntime.h>)
-    #define RN_AUDIO_API_ENABLE_WORKLETS 1
-  #else
-    #define RN_AUDIO_API_ENABLE_WORKLETS 0
-  #endif
+#if ANDROID
+  #include <fbjni/detail/Environment.h>
 #endif
 
 #ifndef RN_AUDIO_API_TEST
@@ -78,4 +72,14 @@ class SerializableWorklet {
 struct RuntimeRegistry {
   std::weak_ptr<worklets::WorkletRuntime> uiRuntime;
   std::shared_ptr<worklets::WorkletRuntime> audioRuntime;
+
+#if ANDROID
+    ~RuntimeRegistry() {
+        facebook::jni::ThreadScope::WithClassLoader(
+            [this]() {
+                uiRuntime.reset();
+                audioRuntime.reset();
+            });
+    }
+#endif
 };

package/common/cpp/audioapi/core/utils/worklets/WorkletsRunner.cpp

@@ -4,7 +4,7 @@ namespace audioapi {
 
 WorkletsRunner::WorkletsRunner(
     std::weak_ptr<worklets::WorkletRuntime> weakRuntime,
-    std::shared_ptr<worklets::SerializableWorklet> shareableWorklet,
+    const std::shared_ptr<worklets::SerializableWorklet> &shareableWorklet,
     bool shouldLockRuntime)
     : weakRuntime_(std::move(weakRuntime)),
       shouldLockRuntime(shouldLockRuntime) {
@@ -33,8 +33,8 @@ WorkletsRunner::WorkletsRunner(
 WorkletsRunner::WorkletsRunner(WorkletsRunner &&other)
     : weakRuntime_(std::move(other.weakRuntime_)),
       unsafeRuntimePtr(other.unsafeRuntimePtr),
-      shouldLockRuntime(other.shouldLockRuntime),
-      workletInitialized(other.workletInitialized) {
+      workletInitialized(other.workletInitialized),
+      shouldLockRuntime(other.shouldLockRuntime) {
   if (workletInitialized) {
     std::memcpy(&unsafeWorklet, &other.unsafeWorklet, sizeof(unsafeWorklet));
     other.workletInitialized = false;

package/common/cpp/audioapi/core/utils/worklets/WorkletsRunner.h

@@ -28,7 +28,7 @@ class WorkletsRunner {
  public:
   explicit WorkletsRunner(
     std::weak_ptr<worklets::WorkletRuntime> weakRuntime,
-    std::shared_ptr<worklets::SerializableWorklet> shareableWorklet,
+    const std::shared_ptr<worklets::SerializableWorklet>& shareableWorklet,
     bool shouldLockRuntime = true);
   WorkletsRunner(WorkletsRunner&&);
   ~WorkletsRunner();

package/common/cpp/audioapi/dsp/AudioUtils.cpp

@@ -30,4 +30,4 @@ float linearToDecibels(float value) {
 float decibelsToLinear(float value) {
   return pow(10, value / 20);
 }
-} // namespace audioapi::dsp
+} // namespace audioapi::dsp

package/common/cpp/audioapi/dsp/Convolver.cpp

@@ -0,0 +1,213 @@
+// implementation of linear convolution algorithm described in this paper:
+// https://publications.rwth-aachen.de/record/466561/files/466561.pdf page 110
+
+#if defined(__ARM_NEON)
+#include <arm_neon.h>
+#endif
+
+#include <audioapi/core/sources/AudioBuffer.h>
+#include <audioapi/dsp/Convolver.h>
+#include <audioapi/dsp/VectorMath.h>
+#include <audioapi/utils/AudioArray.h>
+#include <chrono>
+#include <iostream>
+
+namespace audioapi {
+
+Convolver::Convolver()
+    : _blockSize(0),
+      _segSize(0),
+      _segCount(0),
+      _fftComplexSize(0),
+      _segments(),
+      _segmentsIR(),
+      _fftBuffer(0),
+      _fft(nullptr),
+      _preMultiplied(),
+      _current(0),
+      _inputBuffer(0) {}
+
+void Convolver::reset() {
+  _blockSize = 0;
+  _segSize = 0;
+  _segCount = 0;
+  _fftComplexSize = 0;
+  _current = 0;
+  _fft = nullptr;
+  _segments.clear();
+  _segmentsIR.clear();
+  _preMultiplied.clear();
+  _fftBuffer.zero();
+  _inputBuffer.zero();
+}
+
+bool Convolver::init(
+    size_t blockSize,
+    const audioapi::AudioArray &ir,
+    size_t irLen) {
+  reset();
+  // blockSize must be a power of two
+  if ((blockSize & (blockSize - 1))) {
+    return false;
+  }
+
+  // Ignore zeros at the end of the impulse response because they only waste
+  // computation time
+  _blockSize = blockSize;
+  _trueSegmentCount = (size_t)(std::ceil((float)irLen / (float)_blockSize));
+  while (irLen > 0 && ::fabs(ir[irLen - 1]) < 10e-3) {
+    --irLen;
+  }
+
+  if (irLen == 0) {
+    return true;
+  }
+
+  // The length-N is split into P = N/B length-B sub filters
+  _segCount = (size_t)(std::ceil((float)irLen / (float)_blockSize));
+  _segSize = 2 * _blockSize;
+  // size of the FFT is 2B, so the complex size is B+1, due to the
+  // complex-conjugate symmetricity
+  _fftComplexSize = _segSize / 2 + 1;
+  _fft = std::make_shared<dsp::FFT>((int)_segSize);
+  _fftBuffer.resize(_segSize);
+
+  // segments preparation
+  for (int i = 0; i < _segCount; ++i) {
+    aligned_vec_complex vec(_fftComplexSize, std::complex<float>(0.0f, 0.0f));
+    _segments.push_back(vec);
+  }
+
+  // ir preparation
+  for (int i = 0; i < _segCount; ++i) {
+    aligned_vec_complex segment(_fftComplexSize);
+    const size_t remainingSamples = irLen - (i * _blockSize);
+    const size_t samplesToCopy = std::min(_blockSize, remainingSamples);
+
+    if (samplesToCopy > 0) {
+      memcpy(
+          _fftBuffer.getData(),
+          ir.getData() + i * _blockSize,
+          samplesToCopy * sizeof(float));
+    }
+    // Each sub filter is zero-padded to length 2B and transformed using a
+    // 2B-point real-to-complex FFT.
+    memset(_fftBuffer.getData() + _blockSize, 0, _blockSize * sizeof(float));
+    _fft->doFFT(_fftBuffer.getData(), segment);
+    segment.at(0).imag(0.0f); // ensure DC component is real
+    _segmentsIR.push_back(segment);
+  }
+
+  _preMultiplied = aligned_vec_complex(_fftComplexSize);
+  _inputBuffer.resize(_segSize);
+  _current = 0;
+
+  return true;
+}
+
+/// @brief Fast pairwise complex multiplication using ARM NEON intrinsics
+/// @param ir Impulse response
+/// @param audio Input audio signal
+/// @param pre Output buffer for pre-multiplied results
+/// @note IMPORTANT: ir, audio, and pre must be the same size and should be
+/// aligned to 16 bytes for optimal performance
+void pairwise_complex_multiply_fast(
+    const Convolver::aligned_vec_complex &ir,
+    const Convolver::aligned_vec_complex &audio,
+    Convolver::aligned_vec_complex &pre) {
+  size_t n = ir.size();
+
+/// @note Using ARM NEON intrinsics for SIMD optimization
+/// This implementation is on average 2x faster than the scalar version on ARM
+/// architectures With 16-byte alignment it can be even faster up to 2.5x
+#ifdef __ARM_NEON
+  size_t j = 0;
+
+  // Main vector loop: process 4 complex samples (8 floats) per iteration using
+  // vld2q/vst2q deinterleave
+  for (; j <= n - 4; j += 4) {
+    // load de-interleaved real/imag for 4 complex values
+    float32x4x2_t ir_de = vld2q_f32(reinterpret_cast<const float *>(&ir[j]));
+    float32x4x2_t a_de = vld2q_f32(reinterpret_cast<const float *>(&audio[j]));
+    float32x4x2_t pre_de = vld2q_f32(reinterpret_cast<float *>(&pre[j]));
+
+    float32x4_t ir_re = ir_de.val[0];
+    float32x4_t ir_im = ir_de.val[1];
+    float32x4_t a_re = a_de.val[0];
+    float32x4_t a_im = a_de.val[1];
+
+    // real = ir_re * a_re - ir_im * a_im
+    float32x4_t real = vmulq_f32(ir_re, a_re);
+    real = vmlsq_f32(real, ir_im, a_im);
+    // imag = ir_re * a_im + ir_im * a_re
+    float32x4_t imag = vmulq_f32(ir_re, a_im);
+    imag = vmlaq_f32(imag, ir_im, a_re);
+
+    // accumulate into pre
+    float32x4_t new_re = vaddq_f32(pre_de.val[0], real);
+    float32x4_t new_im = vaddq_f32(pre_de.val[1], imag);
+
+    float32x4x2_t out_de;
+    out_de.val[0] = new_re;
+    out_de.val[1] = new_im;
+
+    vst2q_f32(reinterpret_cast<float *>(&pre[j]), out_de);
+  }
+
+  // Tail
+  for (; j < n; ++j) {
+    pre[j] += ir[j] * audio[j];
+  }
+
+#else
+  // Fallback scalar implementation
+  for (size_t i = 0; i < n; ++i) {
+    pre[i] += ir[i] * audio[i];
+  }
+#endif
+}
+
+void Convolver::process(float *data, float *outputData) {
+  // The input buffer acts as a 2B-point sliding window of the input signal.
+  // With each new input block, the right half of the input buffer is shifted
+  // to the left and the new block is stored in the right half.
+  memmove(
+      _inputBuffer.getData(),
+      _inputBuffer.getData() + _blockSize,
+      _blockSize * sizeof(float));
+  memcpy(_inputBuffer.getData() + _blockSize, data, _blockSize * sizeof(float));
+
+  // All contents (DFT spectra) in the FDL are shifted up by one slot.
+  _current = (_current > 0) ? _current - 1 : _segCount - 1;
+  // A 2B-point real-to-complex FFT is computed from the input buffer,
+  // resulting in B+1 complex-conjugate symmetric DFT coefficients. The
+  // result is stored in the first FDL slot.
+  // _current marks first FDL slot, which is the current input block.
+  _fft->doFFT(_inputBuffer.getData(), _segments[_current]);
+  _segments[_current][0].imag(0.0f); // ensure DC component is real
+
+  // The P sub filter spectra are pairwisely multiplied with the input spectra
+  // in the FDL. The results are accumulated in the frequency-domain.
+  memset(
+      _preMultiplied.data(),
+      0,
+      _preMultiplied.size() * sizeof(std::complex<float>));
+  // this is a bottleneck of the algorithm
+  for (int i = 0; i < _segCount; ++i) {
+    const int indexAudio = (_current + i) % _segCount;
+    const auto &impulseResponseSegment = _segmentsIR[i];
+    const auto &audioSegment = _segments[indexAudio];
+    pairwise_complex_multiply_fast(
+        impulseResponseSegment, audioSegment, _preMultiplied);
+  }
+  // Of the accumulated spectral convolutions, an 2B-point complex-to-real
+  // IFFT is computed. From the resulting 2B samples, the left half is
+  // discarded and the right half is returned as the next output block.
+  _fft->doInverseFFT(_preMultiplied, _fftBuffer.getData());
+
+  memcpy(
+      outputData,
+      _fftBuffer.getData() + _blockSize,
+      _blockSize * sizeof(float));
+}
+} // namespace audioapi

package/common/cpp/audioapi/dsp/Convolver.h

@@ -0,0 +1,45 @@
+#pragma once
+
+#include <audioapi/utils/AudioArray.h>
+#include <audioapi/dsp/FFT.h>
+#include <vector>
+#include <cstring>
+#include <complex>
+#include <memory>
+#include <audioapi/utils/AlignedAllocator.hpp>
+
+namespace audioapi {
+
+class AudioBuffer;
+
+class Convolver {
+  using aligned_vec_complex =
+      std::vector<std::complex<float>, AlignedAllocator<std::complex<float>, 16>>;
+
+ public:
+    Convolver();
+    bool init(size_t blockSize, const AudioArray &ir, size_t irLen);
+    void process(float* inputData, float* outputData);
+    void reset();
+    inline size_t getSegCount() const { return _trueSegmentCount; }
+
+ private:
+    size_t _trueSegmentCount;
+    size_t _blockSize;
+    size_t _segSize;
+    size_t _segCount;
+    size_t _fftComplexSize;
+    std::vector<aligned_vec_complex> _segments;
+    std::vector<aligned_vec_complex> _segmentsIR;
+    AudioArray _fftBuffer;
+    std::shared_ptr<dsp::FFT> _fft;
+    aligned_vec_complex _preMultiplied;
+    size_t _current;
+    AudioArray _inputBuffer;
+
+  friend void pairwise_complex_multiply_fast(
+      const aligned_vec_complex& ir,
+      const aligned_vec_complex& audio,
+      aligned_vec_complex& pre);
+};
+} // namespace audioapi

package/common/cpp/audioapi/dsp/FFT.cpp

@@ -12,30 +12,4 @@ FFT::~FFT() {
   pffft_aligned_free(work_);
 }
 
-void FFT::doFFT(float *in, std::vector<std::complex<float>> &out) {
-  pffft_transform_ordered(
-      pffftSetup_,
-      in,
-      reinterpret_cast<float *>(&out[0]),
-      work_,
-      PFFFT_FORWARD);
-
-  dsp::multiplyByScalar(
-      reinterpret_cast<float *>(&out[0]),
-      0.5f,
-      reinterpret_cast<float *>(&out[0]),
-      size_ * 2);
-}
-
-void FFT::doInverseFFT(std::vector<std::complex<float>> &in, float *out) {
-  pffft_transform_ordered(
-      pffftSetup_,
-      reinterpret_cast<float *>(&in[0]),
-      out,
-      work_,
-      PFFFT_BACKWARD);
-
-  dsp::multiplyByScalar(out, 1.0f / static_cast<float>(size_), out, size_);
-}
-
 } // namespace audioapi::dsp