update code.

VocieProcess/common_audio/resampler/push_sinc_resampler.cc

@@ -0,0 +1,102 @@
+/*
+ *  Copyright (c) 2013 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "common_audio/resampler/push_sinc_resampler.h"
+
+#include <cstring>
+
+#include "common_audio/include/audio_util.h"
+#include "rtc_base/checks.h"
+
+namespace webrtc {
+
+PushSincResampler::PushSincResampler(size_t source_frames,
+                                     size_t destination_frames)
+    : resampler_(new SincResampler(source_frames * 1.0 / destination_frames,
+                                   source_frames,
+                                   this)),
+      source_ptr_(nullptr),
+      source_ptr_int_(nullptr),
+      destination_frames_(destination_frames),
+      first_pass_(true),
+      source_available_(0) {}
+
+PushSincResampler::~PushSincResampler() {}
+
+size_t PushSincResampler::Resample(const int16_t* source,
+                                   size_t source_length,
+                                   int16_t* destination,
+                                   size_t destination_capacity) {
+  if (!float_buffer_.get())
+    float_buffer_.reset(new float[destination_frames_]);
+
+  source_ptr_int_ = source;
+  // Pass nullptr as the float source to have Run() read from the int16 source.
+  Resample(nullptr, source_length, float_buffer_.get(), destination_frames_);
+  FloatS16ToS16(float_buffer_.get(), destination_frames_, destination);
+  source_ptr_int_ = nullptr;
+  return destination_frames_;
+}
+
+size_t PushSincResampler::Resample(const float* source,
+                                   size_t source_length,
+                                   float* destination,
+                                   size_t destination_capacity) {
+  RTC_CHECK_EQ(source_length, resampler_->request_frames());
+  RTC_CHECK_GE(destination_capacity, destination_frames_);
+  // Cache the source pointer. Calling Resample() will immediately trigger
+  // the Run() callback whereupon we provide the cached value.
+  source_ptr_ = source;
+  source_available_ = source_length;
+
+  // On the first pass, we call Resample() twice. During the first call, we
+  // provide dummy input and discard the output. This is done to prime the
+  // SincResampler buffer with the correct delay (half the kernel size), thereby
+  // ensuring that all later Resample() calls will only result in one input
+  // request through Run().
+  //
+  // If this wasn't done, SincResampler would call Run() twice on the first
+  // pass, and we'd have to introduce an entire `source_frames` of delay, rather
+  // than the minimum half kernel.
+  //
+  // It works out that ChunkSize() is exactly the amount of output we need to
+  // request in order to prime the buffer with a single Run() request for
+  // `source_frames`.
+  if (first_pass_)
+    resampler_->Resample(resampler_->ChunkSize(), destination);
+
+  resampler_->Resample(destination_frames_, destination);
+  source_ptr_ = nullptr;
+  return destination_frames_;
+}
+
+void PushSincResampler::Run(size_t frames, float* destination) {
+  // Ensure we are only asked for the available samples. This would fail if
+  // Run() was triggered more than once per Resample() call.
+  RTC_CHECK_EQ(source_available_, frames);
+
+  if (first_pass_) {
+    // Provide dummy input on the first pass, the output of which will be
+    // discarded, as described in Resample().
+    std::memset(destination, 0, frames * sizeof(*destination));
+    first_pass_ = false;
+    return;
+  }
+
+  if (source_ptr_) {
+    std::memcpy(destination, source_ptr_, frames * sizeof(*destination));
+  } else {
+    for (size_t i = 0; i < frames; ++i)
+      destination[i] = static_cast<float>(source_ptr_int_[i]);
+  }
+  source_available_ -= frames;
+}
+
+}  // namespace webrtc

VocieProcess/common_audio/resampler/push_sinc_resampler.h

@@ -0,0 +1,88 @@
+/*
+ *  Copyright (c) 2013 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef COMMON_AUDIO_RESAMPLER_PUSH_SINC_RESAMPLER_H_
+#define COMMON_AUDIO_RESAMPLER_PUSH_SINC_RESAMPLER_H_
+
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+
+#include "api/audio/audio_view.h"
+#include "common_audio/resampler/sinc_resampler.h"
+
+namespace webrtc {
+
+// A thin wrapper over SincResampler to provide a push-based interface as
+// required by WebRTC. SincResampler uses a pull-based interface, and will
+// use SincResamplerCallback::Run() to request data upon a call to Resample().
+// These Run() calls will happen on the same thread Resample() is called on.
+class PushSincResampler : public SincResamplerCallback {
+ public:
+  // Provide the size of the source and destination blocks in samples. These
+  // must correspond to the same time duration (typically 10 ms) as the sample
+  // ratio is inferred from them.
+  PushSincResampler(size_t source_frames, size_t destination_frames);
+  ~PushSincResampler() override;
+
+  PushSincResampler(const PushSincResampler&) = delete;
+  PushSincResampler& operator=(const PushSincResampler&) = delete;
+
+  // Perform the resampling. `source_frames` must always equal the
+  // `source_frames` provided at construction. `destination_capacity` must be
+  // at least as large as `destination_frames`. Returns the number of samples
+  // provided in destination (for convenience, since this will always be equal
+  // to `destination_frames`).
+  template <typename S, typename D>
+  size_t Resample(const MonoView<S>& source, const MonoView<D>& destination) {
+    return Resample(&source[0], SamplesPerChannel(source), &destination[0],
+                    SamplesPerChannel(destination));
+  }
+
+  size_t Resample(const int16_t* source,
+                  size_t source_frames,
+                  int16_t* destination,
+                  size_t destination_capacity);
+  size_t Resample(const float* source,
+                  size_t source_frames,
+                  float* destination,
+                  size_t destination_capacity);
+
+  // Delay due to the filter kernel. Essentially, the time after which an input
+  // sample will appear in the resampled output.
+  static float AlgorithmicDelaySeconds(int source_rate_hz) {
+    return 1.f / source_rate_hz * SincResampler::kKernelSize / 2;
+  }
+
+ protected:
+  // Implements SincResamplerCallback.
+  void Run(size_t frames, float* destination) override;
+
+ private:
+  friend class PushSincResamplerTest;
+  SincResampler* get_resampler_for_testing() { return resampler_.get(); }
+
+  std::unique_ptr<SincResampler> resampler_;
+  std::unique_ptr<float[]> float_buffer_;
+  const float* source_ptr_;
+  const int16_t* source_ptr_int_;
+  const size_t destination_frames_;
+
+  // True on the first call to Resample(), to prime the SincResampler buffer.
+  bool first_pass_;
+
+  // Used to assert we are only requested for as much data as is available.
+  size_t source_available_;
+};
+
+}  // namespace webrtc
+
+#endif  // COMMON_AUDIO_RESAMPLER_PUSH_SINC_RESAMPLER_H_

VocieProcess/common_audio/resampler/sinc_resampler.cc

@@ -0,0 +1,366 @@
+/*
+ *  Copyright (c) 2013 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Modified from the Chromium original:
+// src/media/base/sinc_resampler.cc
+
+// Initial input buffer layout, dividing into regions r0_ to r4_ (note: r0_, r3_
+// and r4_ will move after the first load):
+//
+// |----------------|-----------------------------------------|----------------|
+//
+//                                        request_frames_
+//                   <--------------------------------------------------------->
+//                                    r0_ (during first load)
+//
+//  kKernelSize / 2   kKernelSize / 2         kKernelSize / 2   kKernelSize / 2
+// <---------------> <--------------->       <---------------> <--------------->
+//        r1_               r2_                     r3_               r4_
+//
+//                             block_size_ == r4_ - r2_
+//                   <--------------------------------------->
+//
+//                                                  request_frames_
+//                                    <------------------ ... ----------------->
+//                                               r0_ (during second load)
+//
+// On the second request r0_ slides to the right by kKernelSize / 2 and r3_, r4_
+// and block_size_ are reinitialized via step (3) in the algorithm below.
+//
+// These new regions remain constant until a Flush() occurs.  While complicated,
+// this allows us to reduce jitter by always requesting the same amount from the
+// provided callback.
+//
+// The algorithm:
+//
+// 1) Allocate input_buffer of size: request_frames_ + kKernelSize; this ensures
+//    there's enough room to read request_frames_ from the callback into region
+//    r0_ (which will move between the first and subsequent passes).
+//
+// 2) Let r1_, r2_ each represent half the kernel centered around r0_:
+//
+//        r0_ = input_buffer_ + kKernelSize / 2
+//        r1_ = input_buffer_
+//        r2_ = r0_
+//
+//    r0_ is always request_frames_ in size.  r1_, r2_ are kKernelSize / 2 in
+//    size.  r1_ must be zero initialized to avoid convolution with garbage (see
+//    step (5) for why).
+//
+// 3) Let r3_, r4_ each represent half the kernel right aligned with the end of
+//    r0_ and choose block_size_ as the distance in frames between r4_ and r2_:
+//
+//        r3_ = r0_ + request_frames_ - kKernelSize
+//        r4_ = r0_ + request_frames_ - kKernelSize / 2
+//        block_size_ = r4_ - r2_ = request_frames_ - kKernelSize / 2
+//
+// 4) Consume request_frames_ frames into r0_.
+//
+// 5) Position kernel centered at start of r2_ and generate output frames until
+//    the kernel is centered at the start of r4_ or we've finished generating
+//    all the output frames.
+//
+// 6) Wrap left over data from the r3_ to r1_ and r4_ to r2_.
+//
+// 7) If we're on the second load, in order to avoid overwriting the frames we
+//    just wrapped from r4_ we need to slide r0_ to the right by the size of
+//    r4_, which is kKernelSize / 2:
+//
+//        r0_ = r0_ + kKernelSize / 2 = input_buffer_ + kKernelSize
+//
+//    r3_, r4_, and block_size_ then need to be reinitialized, so goto (3).
+//
+// 8) Else, if we're not on the second load, goto (4).
+//
+// Note: we're glossing over how the sub-sample handling works with
+// `virtual_source_idx_`, etc.
+
+// MSVC++ requires this to be set before any other includes to get M_PI.
+#define _USE_MATH_DEFINES
+
+#include "common_audio/resampler/sinc_resampler.h"
+
+#include <math.h>
+#include <stdint.h>
+#include <string.h>
+
+#include <limits>
+
+#include "rtc_base/checks.h"
+#include "rtc_base/system/arch.h"
+#include "system_wrappers/include/cpu_features_wrapper.h"  // kSSE2, WebRtc_G...
+
+namespace webrtc {
+
+namespace {
+
+double SincScaleFactor(double io_ratio) {
+  // `sinc_scale_factor` is basically the normalized cutoff frequency of the
+  // low-pass filter.
+  double sinc_scale_factor = io_ratio > 1.0 ? 1.0 / io_ratio : 1.0;
+
+  // The sinc function is an idealized brick-wall filter, but since we're
+  // windowing it the transition from pass to stop does not happen right away.
+  // So we should adjust the low pass filter cutoff slightly downward to avoid
+  // some aliasing at the very high-end.
+  // TODO(crogers): this value is empirical and to be more exact should vary
+  // depending on kKernelSize.
+  sinc_scale_factor *= 0.9;
+
+  return sinc_scale_factor;
+}
+
+}  // namespace
+
+const size_t SincResampler::kKernelSize;
+
+// If we know the minimum architecture at compile time, avoid CPU detection.
+void SincResampler::InitializeCPUSpecificFeatures() {
+#if defined(WEBRTC_HAS_NEON)
+  convolve_proc_ = Convolve_NEON;
+#elif defined(WEBRTC_ARCH_X86_FAMILY)
+  // Using AVX2 instead of SSE2 when AVX2/FMA3 supported.
+  if (GetCPUInfo(kAVX2) && GetCPUInfo(kFMA3))
+    convolve_proc_ = Convolve_AVX2;
+  else if (GetCPUInfo(kSSE2))
+    convolve_proc_ = Convolve_SSE;
+  else
+    convolve_proc_ = Convolve_C;
+#else
+  // Unknown architecture.
+  convolve_proc_ = Convolve_C;
+#endif
+}
+
+SincResampler::SincResampler(double io_sample_rate_ratio,
+                             size_t request_frames,
+                             SincResamplerCallback* read_cb)
+    : io_sample_rate_ratio_(io_sample_rate_ratio),
+      read_cb_(read_cb),
+      request_frames_(request_frames),
+      input_buffer_size_(request_frames_ + kKernelSize),
+      // Create input buffers with a 32-byte alignment for SIMD optimizations.
+      kernel_storage_(static_cast<float*>(
+          AlignedMalloc(sizeof(float) * kKernelStorageSize, 32))),
+      kernel_pre_sinc_storage_(static_cast<float*>(
+          AlignedMalloc(sizeof(float) * kKernelStorageSize, 32))),
+      kernel_window_storage_(static_cast<float*>(
+          AlignedMalloc(sizeof(float) * kKernelStorageSize, 32))),
+      input_buffer_(static_cast<float*>(
+          AlignedMalloc(sizeof(float) * input_buffer_size_, 32))),
+      convolve_proc_(nullptr),
+      r1_(input_buffer_.get()),
+      r2_(input_buffer_.get() + kKernelSize / 2) {
+  InitializeCPUSpecificFeatures();
+  RTC_DCHECK(convolve_proc_);
+  RTC_DCHECK_GT(request_frames_, 0);
+  Flush();
+  RTC_DCHECK_GT(block_size_, kKernelSize);
+
+  memset(kernel_storage_.get(), 0,
+         sizeof(*kernel_storage_.get()) * kKernelStorageSize);
+  memset(kernel_pre_sinc_storage_.get(), 0,
+         sizeof(*kernel_pre_sinc_storage_.get()) * kKernelStorageSize);
+  memset(kernel_window_storage_.get(), 0,
+         sizeof(*kernel_window_storage_.get()) * kKernelStorageSize);
+
+  InitializeKernel();
+}
+
+SincResampler::~SincResampler() {}
+
+void SincResampler::UpdateRegions(bool second_load) {
+  // Setup various region pointers in the buffer (see diagram above).  If we're
+  // on the second load we need to slide r0_ to the right by kKernelSize / 2.
+  r0_ = input_buffer_.get() + (second_load ? kKernelSize : kKernelSize / 2);
+  r3_ = r0_ + request_frames_ - kKernelSize;
+  r4_ = r0_ + request_frames_ - kKernelSize / 2;
+  block_size_ = r4_ - r2_;
+
+  // r1_ at the beginning of the buffer.
+  RTC_DCHECK_EQ(r1_, input_buffer_.get());
+  // r1_ left of r2_, r4_ left of r3_ and size correct.
+  RTC_DCHECK_EQ(r2_ - r1_, r4_ - r3_);
+  // r2_ left of r3.
+  RTC_DCHECK_LT(r2_, r3_);
+}
+
+void SincResampler::InitializeKernel() {
+  // Blackman window parameters.
+  static const double kAlpha = 0.16;
+  static const double kA0 = 0.5 * (1.0 - kAlpha);
+  static const double kA1 = 0.5;
+  static const double kA2 = 0.5 * kAlpha;
+
+  // Generates a set of windowed sinc() kernels.
+  // We generate a range of sub-sample offsets from 0.0 to 1.0.
+  const double sinc_scale_factor = SincScaleFactor(io_sample_rate_ratio_);
+  for (size_t offset_idx = 0; offset_idx <= kKernelOffsetCount; ++offset_idx) {
+    const float subsample_offset =
+        static_cast<float>(offset_idx) / kKernelOffsetCount;
+
+    for (size_t i = 0; i < kKernelSize; ++i) {
+      const size_t idx = i + offset_idx * kKernelSize;
+      const float pre_sinc = static_cast<float>(
+          M_PI * (static_cast<int>(i) - static_cast<int>(kKernelSize / 2) -
+                  subsample_offset));
+      kernel_pre_sinc_storage_[idx] = pre_sinc;
+
+      // Compute Blackman window, matching the offset of the sinc().
+      const float x = (i - subsample_offset) / kKernelSize;
+      const float window = static_cast<float>(kA0 - kA1 * cos(2.0 * M_PI * x) +
+                                              kA2 * cos(4.0 * M_PI * x));
+      kernel_window_storage_[idx] = window;
+
+      // Compute the sinc with offset, then window the sinc() function and store
+      // at the correct offset.
+      kernel_storage_[idx] = static_cast<float>(
+          window * ((pre_sinc == 0)
+                        ? sinc_scale_factor
+                        : (sin(sinc_scale_factor * pre_sinc) / pre_sinc)));
+    }
+  }
+}
+
+void SincResampler::SetRatio(double io_sample_rate_ratio) {
+  if (fabs(io_sample_rate_ratio_ - io_sample_rate_ratio) <
+      std::numeric_limits<double>::epsilon()) {
+    return;
+  }
+
+  io_sample_rate_ratio_ = io_sample_rate_ratio;
+
+  // Optimize reinitialization by reusing values which are independent of
+  // `sinc_scale_factor`.  Provides a 3x speedup.
+  const double sinc_scale_factor = SincScaleFactor(io_sample_rate_ratio_);
+  for (size_t offset_idx = 0; offset_idx <= kKernelOffsetCount; ++offset_idx) {
+    for (size_t i = 0; i < kKernelSize; ++i) {
+      const size_t idx = i + offset_idx * kKernelSize;
+      const float window = kernel_window_storage_[idx];
+      const float pre_sinc = kernel_pre_sinc_storage_[idx];
+
+      kernel_storage_[idx] = static_cast<float>(
+          window * ((pre_sinc == 0)
+                        ? sinc_scale_factor
+                        : (sin(sinc_scale_factor * pre_sinc) / pre_sinc)));
+    }
+  }
+}
+
+void SincResampler::Resample(size_t frames, float* destination) {
+  size_t remaining_frames = frames;
+
+  // Step (1) -- Prime the input buffer at the start of the input stream.
+  if (!buffer_primed_ && remaining_frames) {
+    read_cb_->Run(request_frames_, r0_);
+    buffer_primed_ = true;
+  }
+
+  // Step (2) -- Resample!  const what we can outside of the loop for speed.  It
+  // actually has an impact on ARM performance.  See inner loop comment below.
+  const double current_io_ratio = io_sample_rate_ratio_;
+  const float* const kernel_ptr = kernel_storage_.get();
+  while (remaining_frames) {
+    // `i` may be negative if the last Resample() call ended on an iteration
+    // that put `virtual_source_idx_` over the limit.
+    //
+    // Note: The loop construct here can severely impact performance on ARM
+    // or when built with clang.  See https://codereview.chromium.org/18566009/
+    for (int i = static_cast<int>(
+             ceil((block_size_ - virtual_source_idx_) / current_io_ratio));
+         i > 0; --i) {
+      RTC_DCHECK_LT(virtual_source_idx_, block_size_);
+
+      // `virtual_source_idx_` lies in between two kernel offsets so figure out
+      // what they are.
+      const int source_idx = static_cast<int>(virtual_source_idx_);
+      const double subsample_remainder = virtual_source_idx_ - source_idx;
+
+      const double virtual_offset_idx =
+          subsample_remainder * kKernelOffsetCount;
+      const int offset_idx = static_cast<int>(virtual_offset_idx);
+
+      // We'll compute "convolutions" for the two kernels which straddle
+      // `virtual_source_idx_`.
+      const float* const k1 = kernel_ptr + offset_idx * kKernelSize;
+      const float* const k2 = k1 + kKernelSize;
+
+      // Ensure `k1`, `k2` are 32-byte aligned for SIMD usage.  Should always be
+      // true so long as kKernelSize is a multiple of 32.
+      RTC_DCHECK_EQ(0, reinterpret_cast<uintptr_t>(k1) % 32);
+      RTC_DCHECK_EQ(0, reinterpret_cast<uintptr_t>(k2) % 32);
+
+      // Initialize input pointer based on quantized `virtual_source_idx_`.
+      const float* const input_ptr = r1_ + source_idx;
+
+      // Figure out how much to weight each kernel's "convolution".
+      const double kernel_interpolation_factor =
+          virtual_offset_idx - offset_idx;
+      *destination++ =
+          convolve_proc_(input_ptr, k1, k2, kernel_interpolation_factor);
+
+      // Advance the virtual index.
+      virtual_source_idx_ += current_io_ratio;
+
+      if (!--remaining_frames)
+        return;
+    }
+
+    // Wrap back around to the start.
+    virtual_source_idx_ -= block_size_;
+
+    // Step (3) -- Copy r3_, r4_ to r1_, r2_.
+    // This wraps the last input frames back to the start of the buffer.
+    memcpy(r1_, r3_, sizeof(*input_buffer_.get()) * kKernelSize);
+
+    // Step (4) -- Reinitialize regions if necessary.
+    if (r0_ == r2_)
+      UpdateRegions(true);
+
+    // Step (5) -- Refresh the buffer with more input.
+    read_cb_->Run(request_frames_, r0_);
+  }
+}
+
+#undef CONVOLVE_FUNC
+
+size_t SincResampler::ChunkSize() const {
+  return static_cast<size_t>(block_size_ / io_sample_rate_ratio_);
+}
+
+void SincResampler::Flush() {
+  virtual_source_idx_ = 0;
+  buffer_primed_ = false;
+  memset(input_buffer_.get(), 0,
+         sizeof(*input_buffer_.get()) * input_buffer_size_);
+  UpdateRegions(false);
+}
+
+float SincResampler::Convolve_C(const float* input_ptr,
+                                const float* k1,
+                                const float* k2,
+                                double kernel_interpolation_factor) {
+  float sum1 = 0;
+  float sum2 = 0;
+
+  // Generate a single output sample.  Unrolling this loop hurt performance in
+  // local testing.
+  size_t n = kKernelSize;
+  while (n--) {
+    sum1 += *input_ptr * *k1++;
+    sum2 += *input_ptr++ * *k2++;
+  }
+
+  // Linearly interpolate the two "convolutions".
+  return static_cast<float>((1.0 - kernel_interpolation_factor) * sum1 +
+                            kernel_interpolation_factor * sum2);
+}
+
+}  // namespace webrtc

VocieProcess/common_audio/resampler/sinc_resampler.h

@@ -0,0 +1,181 @@
+/*
+ *  Copyright (c) 2013 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Modified from the Chromium original here:
+// src/media/base/sinc_resampler.h
+
+#ifndef COMMON_AUDIO_RESAMPLER_SINC_RESAMPLER_H_
+#define COMMON_AUDIO_RESAMPLER_SINC_RESAMPLER_H_
+
+#include <stddef.h>
+
+#include <memory>
+
+#include "rtc_base/gtest_prod_util.h"
+#include "rtc_base/memory/aligned_malloc.h"
+#include "rtc_base/system/arch.h"
+
+namespace webrtc {
+
+// Callback class for providing more data into the resampler.  Expects `frames`
+// of data to be rendered into `destination`; zero padded if not enough frames
+// are available to satisfy the request.
+class SincResamplerCallback {
+ public:
+  virtual ~SincResamplerCallback() {}
+  virtual void Run(size_t frames, float* destination) = 0;
+};
+
+// SincResampler is a high-quality single-channel sample-rate converter.
+class SincResampler {
+ public:
+  // The kernel size can be adjusted for quality (higher is better) at the
+  // expense of performance.  Must be a multiple of 32.
+  // TODO(dalecurtis): Test performance to see if we can jack this up to 64+.
+  static const size_t kKernelSize = 32;
+
+  // Default request size.  Affects how often and for how much SincResampler
+  // calls back for input.  Must be greater than kKernelSize.
+  static const size_t kDefaultRequestSize = 512;
+
+  // The kernel offset count is used for interpolation and is the number of
+  // sub-sample kernel shifts.  Can be adjusted for quality (higher is better)
+  // at the expense of allocating more memory.
+  static const size_t kKernelOffsetCount = 32;
+  static const size_t kKernelStorageSize =
+      kKernelSize * (kKernelOffsetCount + 1);
+
+  // Constructs a SincResampler with the specified `read_cb`, which is used to
+  // acquire audio data for resampling.  `io_sample_rate_ratio` is the ratio
+  // of input / output sample rates.  `request_frames` controls the size in
+  // frames of the buffer requested by each `read_cb` call.  The value must be
+  // greater than kKernelSize.  Specify kDefaultRequestSize if there are no
+  // request size constraints.
+  SincResampler(double io_sample_rate_ratio,
+                size_t request_frames,
+                SincResamplerCallback* read_cb);
+  virtual ~SincResampler();
+
+  SincResampler(const SincResampler&) = delete;
+  SincResampler& operator=(const SincResampler&) = delete;
+
+  // Resample `frames` of data from `read_cb_` into `destination`.
+  void Resample(size_t frames, float* destination);
+
+  // The maximum size in frames that guarantees Resample() will only make a
+  // single call to `read_cb_` for more data.
+  size_t ChunkSize() const;
+
+  size_t request_frames() const { return request_frames_; }
+
+  // Flush all buffered data and reset internal indices.  Not thread safe, do
+  // not call while Resample() is in progress.
+  void Flush();
+
+  // Update `io_sample_rate_ratio_`.  SetRatio() will cause a reconstruction of
+  // the kernels used for resampling.  Not thread safe, do not call while
+  // Resample() is in progress.
+  //
+  // TODO(ajm): Use this in PushSincResampler rather than reconstructing
+  // SincResampler.  We would also need a way to update `request_frames_`.
+  void SetRatio(double io_sample_rate_ratio);
+
+  float* get_kernel_for_testing() { return kernel_storage_.get(); }
+
+ private:
+  FRIEND_TEST_ALL_PREFIXES(SincResamplerTest, Convolve);
+  FRIEND_TEST_ALL_PREFIXES(SincResamplerTest, ConvolveBenchmark);
+
+  void InitializeKernel();
+  void UpdateRegions(bool second_load);
+
+  // Selects runtime specific CPU features like SSE.  Must be called before
+  // using SincResampler.
+  // TODO(ajm): Currently managed by the class internally. See the note with
+  // `convolve_proc_` below.
+  void InitializeCPUSpecificFeatures();
+
+  // Compute convolution of `k1` and `k2` over `input_ptr`, resultant sums are
+  // linearly interpolated using `kernel_interpolation_factor`.  On x86 and ARM
+  // the underlying implementation is chosen at run time.
+  static float Convolve_C(const float* input_ptr,
+                          const float* k1,
+                          const float* k2,
+                          double kernel_interpolation_factor);
+#if defined(WEBRTC_ARCH_X86_FAMILY)
+  static float Convolve_SSE(const float* input_ptr,
+                            const float* k1,
+                            const float* k2,
+                            double kernel_interpolation_factor);
+  static float Convolve_AVX2(const float* input_ptr,
+                             const float* k1,
+                             const float* k2,
+                             double kernel_interpolation_factor);
+#elif defined(WEBRTC_HAS_NEON)
+  static float Convolve_NEON(const float* input_ptr,
+                             const float* k1,
+                             const float* k2,
+                             double kernel_interpolation_factor);
+#endif
+
+  // The ratio of input / output sample rates.
+  double io_sample_rate_ratio_;
+
+  // An index on the source input buffer with sub-sample precision.  It must be
+  // double precision to avoid drift.
+  double virtual_source_idx_;
+
+  // The buffer is primed once at the very beginning of processing.
+  bool buffer_primed_;
+
+  // Source of data for resampling.
+  SincResamplerCallback* read_cb_;
+
+  // The size (in samples) to request from each `read_cb_` execution.
+  const size_t request_frames_;
+
+  // The number of source frames processed per pass.
+  size_t block_size_;
+
+  // The size (in samples) of the internal buffer used by the resampler.
+  const size_t input_buffer_size_;
+
+  // Contains kKernelOffsetCount kernels back-to-back, each of size kKernelSize.
+  // The kernel offsets are sub-sample shifts of a windowed sinc shifted from
+  // 0.0 to 1.0 sample.
+  std::unique_ptr<float[], AlignedFreeDeleter> kernel_storage_;
+  std::unique_ptr<float[], AlignedFreeDeleter> kernel_pre_sinc_storage_;
+  std::unique_ptr<float[], AlignedFreeDeleter> kernel_window_storage_;
+
+  // Data from the source is copied into this buffer for each processing pass.
+  std::unique_ptr<float[], AlignedFreeDeleter> input_buffer_;
+
+  // Stores the runtime selection of which Convolve function to use.
+  // TODO(ajm): Move to using a global static which must only be initialized
+  // once by the user. We're not doing this initially, because we don't have
+  // e.g. a LazyInstance helper in webrtc.
+  typedef float (*ConvolveProc)(const float*,
+                                const float*,
+                                const float*,
+                                double);
+  ConvolveProc convolve_proc_;
+
+  // Pointers to the various regions inside `input_buffer_`.  See the diagram at
+  // the top of the .cc file for more information.
+  float* r0_;
+  float* const r1_;
+  float* const r2_;
+  float* r3_;
+  float* r4_;
+};
+
+}  // namespace webrtc
+
+#endif  // COMMON_AUDIO_RESAMPLER_SINC_RESAMPLER_H_