Import SincResampler from Chromium.

Committing the originals to make further reviews cleaner.

TBR=bjornv
BUG=webrtc:1395

Review URL: https://webrtc-codereview.appspot.com/1096010

git-svn-id: http://webrtc.googlecode.com/svn/trunk/webrtc@3508 4adac7df-926f-26a2-2b94-8c16560cd09d

3 files changed, 852 insertions, 0 deletions

diff --git a/common_audio/resampler/sinc_resampler.cc b/common_audio/resampler/sinc_resampler.cc
new file mode 100644
index 00000000..d836fc7c
--- /dev/null
+++ b/common_audio/resampler/sinc_resampler.cc
@@ -0,0 +1,347 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+//
+// Input buffer layout, dividing the total buffer into regions (r0_ - r5_):
+//
+// |----------------|-----------------------------------------|----------------|
+//
+//                                   kBlockSize + kKernelSize / 2
+//                   <--------------------------------------------------------->
+//                                              r0_
+//
+//  kKernelSize / 2   kKernelSize / 2         kKernelSize / 2   kKernelSize / 2
+// <---------------> <--------------->       <---------------> <--------------->
+//        r1_               r2_                     r3_               r4_
+//
+//                                                     kBlockSize
+//                                     <--------------------------------------->
+//                                                        r5_
+//
+// The algorithm:
+//
+// 1) Consume input frames into r0_ (r1_ is zero-initialized).
+// 2) Position kernel centered at start of r0_ (r2_) and generate output frames
+//    until kernel is centered at start of r4_ or we've finished generating all
+//    the output frames.
+// 3) Copy r3_ to r1_ and r4_ to r2_.
+// 4) Consume input frames into r5_ (zero-pad if we run out of input).
+// 5) Goto (2) until all of input is consumed.
+//
+// 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 "media/base/sinc_resampler.h"
+
+#include <cmath>
+
+#include "base/cpu.h"
+#include "base/logging.h"
+#include "build/build_config.h"
+
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+#include <xmmintrin.h>
+#endif
+
+#if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+#include <arm_neon.h>
+#endif
+
+namespace media {
+
+namespace {
+
+enum {
+  // 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+.
+  kKernelSize = 32,
+
+  // The number of destination frames generated per processing pass.  Affects
+  // how often and for how much SincResampler calls back for input.  Must be
+  // greater than kKernelSize.
+  kBlockSize = 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.
+  kKernelOffsetCount = 32,
+  kKernelStorageSize = kKernelSize * (kKernelOffsetCount + 1),
+
+  // The size (in samples) of the internal buffer used by the resampler.
+  kBufferSize = kBlockSize + kKernelSize
+};
+
+}  // namespace
+
+const int SincResampler::kMaximumLookAheadSize = kBufferSize;
+
+SincResampler::SincResampler(double io_sample_rate_ratio, const ReadCB& read_cb)
+    : io_sample_rate_ratio_(io_sample_rate_ratio),
+      virtual_source_idx_(0),
+      buffer_primed_(false),
+      read_cb_(read_cb),
+      // Create input buffers with a 16-byte alignment for SSE optimizations.
+      kernel_storage_(static_cast<float*>(
+          base::AlignedAlloc(sizeof(float) * kKernelStorageSize, 16))),
+      input_buffer_(static_cast<float*>(
+          base::AlignedAlloc(sizeof(float) * kBufferSize, 16))),
+      // Setup various region pointers in the buffer (see diagram above).
+      r0_(input_buffer_.get() + kKernelSize / 2),
+      r1_(input_buffer_.get()),
+      r2_(r0_),
+      r3_(r0_ + kBlockSize - kKernelSize / 2),
+      r4_(r0_ + kBlockSize),
+      r5_(r0_ + kKernelSize / 2) {
+  // Ensure kKernelSize is a multiple of 32 for easy SSE optimizations; causes
+  // r0_ and r5_ (used for input) to always be 16-byte aligned by virtue of
+  // input_buffer_ being 16-byte aligned.
+  DCHECK_EQ(kKernelSize % 32, 0) << "kKernelSize must be a multiple of 32!";
+  DCHECK_GT(kBlockSize, kKernelSize)
+      << "kBlockSize must be greater than kKernelSize!";
+  // Basic sanity checks to ensure buffer regions are laid out correctly:
+  // r0_ and r2_ should always be the same position.
+  DCHECK_EQ(r0_, r2_);
+  // r1_ at the beginning of the buffer.
+  DCHECK_EQ(r1_, input_buffer_.get());
+  // r1_ left of r2_, r2_ left of r5_ and r1_, r2_ size correct.
+  DCHECK_EQ(r2_ - r1_, r5_ - r2_);
+  // r3_ left of r4_, r5_ left of r0_ and r3_ size correct.
+  DCHECK_EQ(r4_ - r3_, r5_ - r0_);
+  // r3_, r4_ size correct and r4_ at the end of the buffer.
+  DCHECK_EQ(r4_ + (r4_ - r3_), r1_ + kBufferSize);
+  // r5_ size correct and at the end of the buffer.
+  DCHECK_EQ(r5_ + kBlockSize, r1_ + kBufferSize);
+
+  memset(kernel_storage_.get(), 0,
+         sizeof(*kernel_storage_.get()) * kKernelStorageSize);
+  memset(input_buffer_.get(), 0, sizeof(*input_buffer_.get()) * kBufferSize);
+
+  InitializeKernel();
+}
+
+SincResampler::~SincResampler() {}
+
+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;
+
+  // |sinc_scale_factor| is basically the normalized cutoff frequency of the
+  // low-pass filter.
+  double sinc_scale_factor =
+      io_sample_rate_ratio_ > 1.0 ? 1.0 / io_sample_rate_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;
+
+  // Generates a set of windowed sinc() kernels.
+  // We generate a range of sub-sample offsets from 0.0 to 1.0.
+  for (int offset_idx = 0; offset_idx <= kKernelOffsetCount; ++offset_idx) {
+    double subsample_offset =
+        static_cast<double>(offset_idx) / kKernelOffsetCount;
+
+    for (int i = 0; i < kKernelSize; ++i) {
+      // Compute the sinc with offset.
+      double s =
+          sinc_scale_factor * M_PI * (i - kKernelSize / 2 - subsample_offset);
+      double sinc = (!s ? 1.0 : sin(s) / s) * sinc_scale_factor;
+
+      // Compute Blackman window, matching the offset of the sinc().
+      double x = (i - subsample_offset) / kKernelSize;
+      double window = kA0 - kA1 * cos(2.0 * M_PI * x) + kA2
+          * cos(4.0 * M_PI * x);
+
+      // Window the sinc() function and store at the correct offset.
+      kernel_storage_.get()[i + offset_idx * kKernelSize] = sinc * window;
+    }
+  }
+}
+
+void SincResampler::Resample(float* destination, int frames) {
+  int remaining_frames = frames;
+
+  // Step (1) -- Prime the input buffer at the start of the input stream.
+  if (!buffer_primed_) {
+    read_cb_.Run(r0_, kBlockSize + kKernelSize / 2);
+    buffer_primed_ = true;
+  }
+
+  // Step (2) -- Resample!
+  while (remaining_frames) {
+    while (virtual_source_idx_ < kBlockSize) {
+      // |virtual_source_idx_| lies in between two kernel offsets so figure out
+      // what they are.
+      int source_idx = static_cast<int>(virtual_source_idx_);
+      double subsample_remainder = virtual_source_idx_ - source_idx;
+
+      double virtual_offset_idx = subsample_remainder * kKernelOffsetCount;
+      int offset_idx = static_cast<int>(virtual_offset_idx);
+
+      // We'll compute "convolutions" for the two kernels which straddle
+      // |virtual_source_idx_|.
+      float* k1 = kernel_storage_.get() + offset_idx * kKernelSize;
+      float* k2 = k1 + kKernelSize;
+
+      // Initialize input pointer based on quantized |virtual_source_idx_|.
+      float* input_ptr = r1_ + source_idx;
+
+      // Figure out how much to weight each kernel's "convolution".
+      double kernel_interpolation_factor = virtual_offset_idx - offset_idx;
+      *destination++ = Convolve(
+          input_ptr, k1, k2, kernel_interpolation_factor);
+
+      // Advance the virtual index.
+      virtual_source_idx_ += io_sample_rate_ratio_;
+
+      if (!--remaining_frames)
+        return;
+    }
+
+    // Wrap back around to the start.
+    virtual_source_idx_ -= kBlockSize;
+
+    // Step (3) Copy r3_ to r1_ and r4_ to r2_.
+    // This wraps the last input frames back to the start of the buffer.
+    memcpy(r1_, r3_, sizeof(*input_buffer_.get()) * (kKernelSize / 2));
+    memcpy(r2_, r4_, sizeof(*input_buffer_.get()) * (kKernelSize / 2));
+
+    // Step (4)
+    // Refresh the buffer with more input.
+    read_cb_.Run(r5_, kBlockSize);
+  }
+}
+
+int SincResampler::ChunkSize() {
+  return kBlockSize / io_sample_rate_ratio_;
+}
+
+void SincResampler::Flush() {
+  virtual_source_idx_ = 0;
+  buffer_primed_ = false;
+  memset(input_buffer_.get(), 0, sizeof(*input_buffer_.get()) * kBufferSize);
+}
+
+float SincResampler::Convolve(const float* input_ptr, const float* k1,
+                              const float* k2,
+                              double kernel_interpolation_factor) {
+  // Rely on function level static initialization to keep ConvolveProc selection
+  // thread safe.
+  typedef float (*ConvolveProc)(const float* src, const float* k1,
+                                const float* k2,
+                                double kernel_interpolation_factor);
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+  static const ConvolveProc kConvolveProc =
+      base::CPU().has_sse() ? Convolve_SSE : Convolve_C;
+#elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+  static const ConvolveProc kConvolveProc = Convolve_NEON;
+#else
+  static const ConvolveProc kConvolveProc = Convolve_C;
+#endif
+
+  return kConvolveProc(input_ptr, k1, k2, kernel_interpolation_factor);
+}
+
+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.
+  int n = kKernelSize;
+  while (n--) {
+    sum1 += *input_ptr * *k1++;
+    sum2 += *input_ptr++ * *k2++;
+  }
+
+  // Linearly interpolate the two "convolutions".
+  return (1.0 - kernel_interpolation_factor) * sum1
+      + kernel_interpolation_factor * sum2;
+}
+
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+float SincResampler::Convolve_SSE(const float* input_ptr, const float* k1,
+                                  const float* k2,
+                                  double kernel_interpolation_factor) {
+  // Ensure |k1|, |k2| are 16-byte aligned for SSE usage.  Should always be true
+  // so long as kKernelSize is a multiple of 16.
+  DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(k1) & 0x0F);
+  DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(k2) & 0x0F);
+
+  __m128 m_input;
+  __m128 m_sums1 = _mm_setzero_ps();
+  __m128 m_sums2 = _mm_setzero_ps();
+
+  // Based on |input_ptr| alignment, we need to use loadu or load.  Unrolling
+  // these loops hurt performance in local testing.
+  if (reinterpret_cast<uintptr_t>(input_ptr) & 0x0F) {
+    for (int i = 0; i < kKernelSize; i += 4) {
+      m_input = _mm_loadu_ps(input_ptr + i);
+      m_sums1 = _mm_add_ps(m_sums1, _mm_mul_ps(m_input, _mm_load_ps(k1 + i)));
+      m_sums2 = _mm_add_ps(m_sums2, _mm_mul_ps(m_input, _mm_load_ps(k2 + i)));
+    }
+  } else {
+    for (int i = 0; i < kKernelSize; i += 4) {
+      m_input = _mm_load_ps(input_ptr + i);
+      m_sums1 = _mm_add_ps(m_sums1, _mm_mul_ps(m_input, _mm_load_ps(k1 + i)));
+      m_sums2 = _mm_add_ps(m_sums2, _mm_mul_ps(m_input, _mm_load_ps(k2 + i)));
+    }
+  }
+
+  // Linearly interpolate the two "convolutions".
+  m_sums1 = _mm_mul_ps(m_sums1, _mm_set_ps1(1.0 - kernel_interpolation_factor));
+  m_sums2 = _mm_mul_ps(m_sums2, _mm_set_ps1(kernel_interpolation_factor));
+  m_sums1 = _mm_add_ps(m_sums1, m_sums2);
+
+  // Sum components together.
+  float result;
+  m_sums2 = _mm_add_ps(_mm_movehl_ps(m_sums1, m_sums1), m_sums1);
+  _mm_store_ss(&result, _mm_add_ss(m_sums2, _mm_shuffle_ps(
+      m_sums2, m_sums2, 1)));
+
+  return result;
+}
+#endif
+
+#if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+float SincResampler::Convolve_NEON(const float* input_ptr, const float* k1,
+                                   const float* k2,
+                                   double kernel_interpolation_factor) {
+  float32x4_t m_input;
+  float32x4_t m_sums1 = vmovq_n_f32(0);
+  float32x4_t m_sums2 = vmovq_n_f32(0);
+
+  const float* upper = input_ptr + kKernelSize;
+  for (; input_ptr < upper; ) {
+    m_input = vld1q_f32(input_ptr);
+    input_ptr += 4;
+    m_sums1 = vmlaq_f32(m_sums1, m_input, vld1q_f32(k1));
+    k1 += 4;
+    m_sums2 = vmlaq_f32(m_sums2, m_input, vld1q_f32(k2));
+    k2 += 4;
+  }
+
+  // Linearly interpolate the two "convolutions".
+  m_sums1 = vmlaq_f32(
+      vmulq_f32(m_sums1, vmovq_n_f32(1.0 - kernel_interpolation_factor)),
+      m_sums2, vmovq_n_f32(kernel_interpolation_factor));
+
+  // Sum components together.
+  float32x2_t m_half = vadd_f32(vget_high_f32(m_sums1), vget_low_f32(m_sums1));
+  return vget_lane_f32(vpadd_f32(m_half, m_half), 0);
+}
+#endif
+
+}  // namespace media
diff --git a/common_audio/resampler/sinc_resampler.h b/common_audio/resampler/sinc_resampler.h
new file mode 100644
index 00000000..a1d3cf71
--- /dev/null
+++ b/common_audio/resampler/sinc_resampler.h
@@ -0,0 +1,100 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#ifndef MEDIA_BASE_SINC_RESAMPLER_H_
+#define MEDIA_BASE_SINC_RESAMPLER_H_
+
+#include "base/callback.h"
+#include "base/gtest_prod_util.h"
+#include "base/memory/aligned_memory.h"
+#include "base/memory/scoped_ptr.h"
+#include "media/base/media_export.h"
+
+namespace media {
+
+// SincResampler is a high-quality single-channel sample-rate converter.
+class MEDIA_EXPORT SincResampler {
+ public:
+  // The maximum number of samples that may be requested from the callback ahead
+  // of the current position in the stream.
+  static const int kMaximumLookAheadSize;
+
+  // Callback type 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.
+  typedef base::Callback<void(float* destination, int frames)> ReadCB;
+
+  // 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.
+  SincResampler(double io_sample_rate_ratio, const ReadCB& read_cb);
+  virtual ~SincResampler();
+
+  // Resample |frames| of data from |read_cb_| into |destination|.
+  void Resample(float* destination, int frames);
+
+  // The maximum size in frames that guarantees Resample() will only make a
+  // single call to |read_cb_| for more data.
+  int ChunkSize();
+
+  // Flush all buffered data and reset internal indices.
+  void Flush();
+
+ private:
+  FRIEND_TEST_ALL_PREFIXES(SincResamplerTest, Convolve);
+  FRIEND_TEST_ALL_PREFIXES(SincResamplerTest, ConvolveBenchmark);
+
+  void InitializeKernel();
+
+  // Compute convolution of |k1| and |k2| over |input_ptr|, resultant sums are
+  // linearly interpolated using |kernel_interpolation_factor|.  On x86, the
+  // underlying implementation is chosen at run time based on SSE support.  On
+  // ARM, NEON support is chosen at compile time based on compilation flags.
+  static float Convolve(const float* input_ptr, const float* k1,
+                        const float* k2, double kernel_interpolation_factor);
+  static float Convolve_C(const float* input_ptr, const float* k1,
+                          const float* k2, double kernel_interpolation_factor);
+  static float Convolve_SSE(const float* input_ptr, const float* k1,
+                            const float* k2,
+                            double kernel_interpolation_factor);
+  static float Convolve_NEON(const float* input_ptr, const float* k1,
+                             const float* k2,
+                             double kernel_interpolation_factor);
+
+  // 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.
+  ReadCB read_cb_;
+
+  // 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.
+  scoped_ptr_malloc<float, base::ScopedPtrAlignedFree> kernel_storage_;
+
+  // Data from the source is copied into this buffer for each processing pass.
+  scoped_ptr_malloc<float, base::ScopedPtrAlignedFree> input_buffer_;
+
+  // Pointers to the various regions inside |input_buffer_|.  See the diagram at
+  // the top of the .cc file for more information.
+  float* const r0_;
+  float* const r1_;
+  float* const r2_;
+  float* const r3_;
+  float* const r4_;
+  float* const r5_;
+
+  DISALLOW_COPY_AND_ASSIGN(SincResampler);
+};
+
+}  // namespace media
+
+#endif  // MEDIA_BASE_SINC_RESAMPLER_H_
diff --git a/common_audio/resampler/sinc_resampler_unittest.cc b/common_audio/resampler/sinc_resampler_unittest.cc
new file mode 100644
index 00000000..0f718f23
--- /dev/null
+++ b/common_audio/resampler/sinc_resampler_unittest.cc
@@ -0,0 +1,405 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// MSVC++ requires this to be set before any other includes to get M_PI.
+#define _USE_MATH_DEFINES
+
+#include <cmath>
+
+#include "base/bind.h"
+#include "base/bind_helpers.h"
+#include "base/command_line.h"
+#include "base/logging.h"
+#include "base/string_number_conversions.h"
+#include "base/strings/stringize_macros.h"
+#include "base/time.h"
+#include "build/build_config.h"
+#include "media/base/sinc_resampler.h"
+#include "testing/gmock/include/gmock/gmock.h"
+#include "testing/gtest/include/gtest/gtest.h"
+
+using testing::_;
+
+namespace media {
+
+static const double kSampleRateRatio = 192000.0 / 44100.0;
+static const double kKernelInterpolationFactor = 0.5;
+
+// Command line switch for runtime adjustment of ConvolveBenchmark iterations.
+static const char kConvolveIterations[] = "convolve-iterations";
+
+// Helper class to ensure ChunkedResample() functions properly.
+class MockSource {
+ public:
+  MOCK_METHOD2(ProvideInput, void(float* destination, int frames));
+};
+
+ACTION(ClearBuffer) {
+  memset(arg0, 0, arg1 * sizeof(float));
+}
+
+ACTION(FillBuffer) {
+  // Value chosen arbitrarily such that SincResampler resamples it to something
+  // easily representable on all platforms; e.g., using kSampleRateRatio this
+  // becomes 1.81219.
+  memset(arg0, 64, arg1 * sizeof(float));
+}
+
+// Test requesting multiples of ChunkSize() frames results in the proper number
+// of callbacks.
+TEST(SincResamplerTest, ChunkedResample) {
+  MockSource mock_source;
+
+  // Choose a high ratio of input to output samples which will result in quick
+  // exhaustion of SincResampler's internal buffers.
+  SincResampler resampler(
+      kSampleRateRatio,
+      base::Bind(&MockSource::ProvideInput, base::Unretained(&mock_source)));
+
+  static const int kChunks = 2;
+  int max_chunk_size = resampler.ChunkSize() * kChunks;
+  scoped_array<float> resampled_destination(new float[max_chunk_size]);
+
+  // Verify requesting ChunkSize() frames causes a single callback.
+  EXPECT_CALL(mock_source, ProvideInput(_, _))
+      .Times(1).WillOnce(ClearBuffer());
+  resampler.Resample(resampled_destination.get(), resampler.ChunkSize());
+
+  // Verify requesting kChunks * ChunkSize() frames causes kChunks callbacks.
+  testing::Mock::VerifyAndClear(&mock_source);
+  EXPECT_CALL(mock_source, ProvideInput(_, _))
+      .Times(kChunks).WillRepeatedly(ClearBuffer());
+  resampler.Resample(resampled_destination.get(), max_chunk_size);
+}
+
+// Test flush resets the internal state properly.
+TEST(SincResamplerTest, Flush) {
+  MockSource mock_source;
+  SincResampler resampler(
+      kSampleRateRatio,
+      base::Bind(&MockSource::ProvideInput, base::Unretained(&mock_source)));
+  scoped_array<float> resampled_destination(new float[resampler.ChunkSize()]);
+
+  // Fill the resampler with junk data.
+  EXPECT_CALL(mock_source, ProvideInput(_, _))
+      .Times(1).WillOnce(FillBuffer());
+  resampler.Resample(resampled_destination.get(), resampler.ChunkSize() / 2);
+  ASSERT_NE(resampled_destination[0], 0);
+
+  // Flush and request more data, which should all be zeros now.
+  resampler.Flush();
+  testing::Mock::VerifyAndClear(&mock_source);
+  EXPECT_CALL(mock_source, ProvideInput(_, _))
+      .Times(1).WillOnce(ClearBuffer());
+  resampler.Resample(resampled_destination.get(), resampler.ChunkSize() / 2);
+  for (int i = 0; i < resampler.ChunkSize() / 2; ++i)
+    ASSERT_FLOAT_EQ(resampled_destination[i], 0);
+}
+
+// Define platform independent function name for Convolve* tests.
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+#define CONVOLVE_FUNC Convolve_SSE
+#elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+#define CONVOLVE_FUNC Convolve_NEON
+#endif
+
+// Ensure various optimized Convolve() methods return the same value.  Only run
+// this test if other optimized methods exist, otherwise the default Convolve()
+// will be tested by the parameterized SincResampler tests below.
+#if defined(CONVOLVE_FUNC)
+TEST(SincResamplerTest, Convolve) {
+  // Initialize a dummy resampler.
+  MockSource mock_source;
+  SincResampler resampler(
+      kSampleRateRatio,
+      base::Bind(&MockSource::ProvideInput, base::Unretained(&mock_source)));
+
+  // The optimized Convolve methods are slightly more precise than Convolve_C(),
+  // so comparison must be done using an epsilon.
+  static const double kEpsilon = 0.00000005;
+
+  // Use a kernel from SincResampler as input and kernel data, this has the
+  // benefit of already being properly sized and aligned for Convolve_SSE().
+  double result = resampler.Convolve_C(
+      resampler.kernel_storage_.get(), resampler.kernel_storage_.get(),
+      resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+  double result2 = resampler.CONVOLVE_FUNC(
+      resampler.kernel_storage_.get(), resampler.kernel_storage_.get(),
+      resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+  EXPECT_NEAR(result2, result, kEpsilon);
+
+  // Test Convolve() w/ unaligned input pointer.
+  result = resampler.Convolve_C(
+      resampler.kernel_storage_.get() + 1, resampler.kernel_storage_.get(),
+      resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+  result2 = resampler.CONVOLVE_FUNC(
+      resampler.kernel_storage_.get() + 1, resampler.kernel_storage_.get(),
+      resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+  EXPECT_NEAR(result2, result, kEpsilon);
+}
+#endif
+
+// Benchmark for the various Convolve() methods.  Make sure to build with
+// branding=Chrome so that DCHECKs are compiled out when benchmarking.  Original
+// benchmarks were run with --convolve-iterations=50000000.
+TEST(SincResamplerTest, ConvolveBenchmark) {
+  // Initialize a dummy resampler.
+  MockSource mock_source;
+  SincResampler resampler(
+      kSampleRateRatio,
+      base::Bind(&MockSource::ProvideInput, base::Unretained(&mock_source)));
+
+  // Retrieve benchmark iterations from command line.
+  int convolve_iterations = 10;
+  std::string iterations(CommandLine::ForCurrentProcess()->GetSwitchValueASCII(
+      kConvolveIterations));
+  if (!iterations.empty())
+    base::StringToInt(iterations, &convolve_iterations);
+
+  printf("Benchmarking %d iterations:\n", convolve_iterations);
+
+  // Benchmark Convolve_C().
+  base::TimeTicks start = base::TimeTicks::HighResNow();
+  for (int i = 0; i < convolve_iterations; ++i) {
+    resampler.Convolve_C(
+        resampler.kernel_storage_.get(), resampler.kernel_storage_.get(),
+        resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+  }
+  double total_time_c_ms =
+      (base::TimeTicks::HighResNow() - start).InMillisecondsF();
+  printf("Convolve_C took %.2fms.\n", total_time_c_ms);
+
+#if defined(CONVOLVE_FUNC)
+  // Benchmark with unaligned input pointer.
+  start = base::TimeTicks::HighResNow();
+  for (int j = 0; j < convolve_iterations; ++j) {
+    resampler.CONVOLVE_FUNC(
+        resampler.kernel_storage_.get() + 1, resampler.kernel_storage_.get(),
+        resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+  }
+  double total_time_optimized_unaligned_ms =
+      (base::TimeTicks::HighResNow() - start).InMillisecondsF();
+  printf(STRINGIZE(CONVOLVE_FUNC) "(unaligned) took %.2fms; which is %.2fx "
+         "faster than Convolve_C.\n", total_time_optimized_unaligned_ms,
+         total_time_c_ms / total_time_optimized_unaligned_ms);
+
+  // Benchmark with aligned input pointer.
+  start = base::TimeTicks::HighResNow();
+  for (int j = 0; j < convolve_iterations; ++j) {
+    resampler.CONVOLVE_FUNC(
+        resampler.kernel_storage_.get(), resampler.kernel_storage_.get(),
+        resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+  }
+  double total_time_optimized_aligned_ms =
+      (base::TimeTicks::HighResNow() - start).InMillisecondsF();
+  printf(STRINGIZE(CONVOLVE_FUNC) " (aligned) took %.2fms; which is %.2fx "
+         "faster than Convolve_C and %.2fx faster than "
+         STRINGIZE(CONVOLVE_FUNC) " (unaligned).\n",
+         total_time_optimized_aligned_ms,
+         total_time_c_ms / total_time_optimized_aligned_ms,
+         total_time_optimized_unaligned_ms / total_time_optimized_aligned_ms);
+#endif
+}
+
+#undef CONVOLVE_FUNC
+
+// Fake audio source for testing the resampler.  Generates a sinusoidal linear
+// chirp (http://en.wikipedia.org/wiki/Chirp) which can be tuned to stress the
+// resampler for the specific sample rate conversion being used.
+class SinusoidalLinearChirpSource {
+ public:
+  SinusoidalLinearChirpSource(int sample_rate, int samples,
+                              double max_frequency)
+      : sample_rate_(sample_rate),
+        total_samples_(samples),
+        max_frequency_(max_frequency),
+        current_index_(0) {
+    // Chirp rate.
+    double duration = static_cast<double>(total_samples_) / sample_rate_;
+    k_ = (max_frequency_ - kMinFrequency) / duration;
+  }
+
+  virtual ~SinusoidalLinearChirpSource() {}
+
+  void ProvideInput(float* destination, int frames) {
+    for (int i = 0; i < frames; ++i, ++current_index_) {
+      // Filter out frequencies higher than Nyquist.
+      if (Frequency(current_index_) > 0.5 * sample_rate_) {
+        destination[i] = 0;
+      } else {
+        // Calculate time in seconds.
+        double t = static_cast<double>(current_index_) / sample_rate_;
+
+        // Sinusoidal linear chirp.
+        destination[i] = sin(2 * M_PI * (kMinFrequency * t + (k_ / 2) * t * t));
+      }
+    }
+  }
+
+  double Frequency(int position) {
+    return kMinFrequency + position * (max_frequency_ - kMinFrequency)
+        / total_samples_;
+  }
+
+ private:
+  enum {
+    kMinFrequency = 5
+  };
+
+  double sample_rate_;
+  int total_samples_;
+  double max_frequency_;
+  double k_;
+  int current_index_;
+
+  DISALLOW_COPY_AND_ASSIGN(SinusoidalLinearChirpSource);
+};
+
+typedef std::tr1::tuple<int, int, double, double> SincResamplerTestData;
+class SincResamplerTest
+    : public testing::TestWithParam<SincResamplerTestData> {
+ public:
+  SincResamplerTest()
+      : input_rate_(std::tr1::get<0>(GetParam())),
+        output_rate_(std::tr1::get<1>(GetParam())),
+        rms_error_(std::tr1::get<2>(GetParam())),
+        low_freq_error_(std::tr1::get<3>(GetParam())) {
+  }
+
+  virtual ~SincResamplerTest() {}
+
+ protected:
+  int input_rate_;
+  int output_rate_;
+  double rms_error_;
+  double low_freq_error_;
+};
+
+// Tests resampling using a given input and output sample rate.
+TEST_P(SincResamplerTest, Resample) {
+  // Make comparisons using one second of data.
+  static const double kTestDurationSecs = 1;
+  int input_samples = kTestDurationSecs * input_rate_;
+  int output_samples = kTestDurationSecs * output_rate_;
+
+  // Nyquist frequency for the input sampling rate.
+  double input_nyquist_freq = 0.5 * input_rate_;
+
+  // Source for data to be resampled.
+  SinusoidalLinearChirpSource resampler_source(
+      input_rate_, input_samples, input_nyquist_freq);
+
+  SincResampler resampler(
+      input_rate_ / static_cast<double>(output_rate_),
+      base::Bind(&SinusoidalLinearChirpSource::ProvideInput,
+                 base::Unretained(&resampler_source)));
+
+  // TODO(dalecurtis): If we switch to AVX/SSE optimization, we'll need to
+  // allocate these on 32-byte boundaries and ensure they're sized % 32 bytes.
+  scoped_array<float> resampled_destination(new float[output_samples]);
+  scoped_array<float> pure_destination(new float[output_samples]);
+
+  // Generate resampled signal.
+  resampler.Resample(resampled_destination.get(), output_samples);
+
+  // Generate pure signal.
+  SinusoidalLinearChirpSource pure_source(
+      output_rate_, output_samples, input_nyquist_freq);
+  pure_source.ProvideInput(pure_destination.get(), output_samples);
+
+  // Range of the Nyquist frequency (0.5 * min(input rate, output_rate)) which
+  // we refer to as low and high.
+  static const double kLowFrequencyNyquistRange = 0.7;
+  static const double kHighFrequencyNyquistRange = 0.9;
+
+  // Calculate Root-Mean-Square-Error and maximum error for the resampling.
+  double sum_of_squares = 0;
+  double low_freq_max_error = 0;
+  double high_freq_max_error = 0;
+  int minimum_rate = std::min(input_rate_, output_rate_);
+  double low_frequency_range = kLowFrequencyNyquistRange * 0.5 * minimum_rate;
+  double high_frequency_range = kHighFrequencyNyquistRange * 0.5 * minimum_rate;
+  for (int i = 0; i < output_samples; ++i) {
+    double error = fabs(resampled_destination[i] - pure_destination[i]);
+
+    if (pure_source.Frequency(i) < low_frequency_range) {
+      if (error > low_freq_max_error)
+        low_freq_max_error = error;
+    } else if (pure_source.Frequency(i) < high_frequency_range) {
+      if (error > high_freq_max_error)
+        high_freq_max_error = error;
+    }
+    // TODO(dalecurtis): Sanity check frequencies > kHighFrequencyNyquistRange.
+
+    sum_of_squares += error * error;
+  }
+
+  double rms_error = sqrt(sum_of_squares / output_samples);
+
+  // Convert each error to dbFS.
+  #define DBFS(x) 20 * log10(x)
+  rms_error = DBFS(rms_error);
+  low_freq_max_error = DBFS(low_freq_max_error);
+  high_freq_max_error = DBFS(high_freq_max_error);
+
+  EXPECT_LE(rms_error, rms_error_);
+  EXPECT_LE(low_freq_max_error, low_freq_error_);
+
+  // All conversions currently have a high frequency error around -6 dbFS.
+  static const double kHighFrequencyMaxError = -6.02;
+  EXPECT_LE(high_freq_max_error, kHighFrequencyMaxError);
+}
+
+// Almost all conversions have an RMS error of around -14 dbFS.
+static const double kResamplingRMSError = -14.58;
+
+// Thresholds chosen arbitrarily based on what each resampling reported during
+// testing.  All thresholds are in dbFS, http://en.wikipedia.org/wiki/DBFS.
+INSTANTIATE_TEST_CASE_P(
+    SincResamplerTest, SincResamplerTest, testing::Values(
+        // To 44.1kHz
+        std::tr1::make_tuple(8000, 44100, kResamplingRMSError, -62.73),
+        std::tr1::make_tuple(11025, 44100, kResamplingRMSError, -72.19),
+        std::tr1::make_tuple(16000, 44100, kResamplingRMSError, -62.54),
+        std::tr1::make_tuple(22050, 44100, kResamplingRMSError, -73.53),
+        std::tr1::make_tuple(32000, 44100, kResamplingRMSError, -63.32),
+        std::tr1::make_tuple(44100, 44100, kResamplingRMSError, -73.53),
+        std::tr1::make_tuple(48000, 44100, -15.01, -64.04),
+        std::tr1::make_tuple(96000, 44100, -18.49, -25.51),
+        std::tr1::make_tuple(192000, 44100, -20.50, -13.31),
+
+        // To 48kHz
+        std::tr1::make_tuple(8000, 48000, kResamplingRMSError, -63.43),
+        std::tr1::make_tuple(11025, 48000, kResamplingRMSError, -62.61),
+        std::tr1::make_tuple(16000, 48000, kResamplingRMSError, -63.96),
+        std::tr1::make_tuple(22050, 48000, kResamplingRMSError, -62.42),
+        std::tr1::make_tuple(32000, 48000, kResamplingRMSError, -64.04),
+        std::tr1::make_tuple(44100, 48000, kResamplingRMSError, -62.63),
+        std::tr1::make_tuple(48000, 48000, kResamplingRMSError, -73.52),
+        std::tr1::make_tuple(96000, 48000, -18.40, -28.44),
+        std::tr1::make_tuple(192000, 48000, -20.43, -14.11),
+
+        // To 96kHz
+        std::tr1::make_tuple(8000, 96000, kResamplingRMSError, -63.19),
+        std::tr1::make_tuple(11025, 96000, kResamplingRMSError, -62.61),
+        std::tr1::make_tuple(16000, 96000, kResamplingRMSError, -63.39),
+        std::tr1::make_tuple(22050, 96000, kResamplingRMSError, -62.42),
+        std::tr1::make_tuple(32000, 96000, kResamplingRMSError, -63.95),
+        std::tr1::make_tuple(44100, 96000, kResamplingRMSError, -62.63),
+        std::tr1::make_tuple(48000, 96000, kResamplingRMSError, -73.52),
+        std::tr1::make_tuple(96000, 96000, kResamplingRMSError, -73.52),
+        std::tr1::make_tuple(192000, 96000, kResamplingRMSError, -28.41),
+
+        // To 192kHz
+        std::tr1::make_tuple(8000, 192000, kResamplingRMSError, -63.10),
+        std::tr1::make_tuple(11025, 192000, kResamplingRMSError, -62.61),
+        std::tr1::make_tuple(16000, 192000, kResamplingRMSError, -63.14),
+        std::tr1::make_tuple(22050, 192000, kResamplingRMSError, -62.42),
+        std::tr1::make_tuple(32000, 192000, kResamplingRMSError, -63.38),
+        std::tr1::make_tuple(44100, 192000, kResamplingRMSError, -62.63),
+        std::tr1::make_tuple(48000, 192000, kResamplingRMSError, -73.44),
+        std::tr1::make_tuple(96000, 192000, kResamplingRMSError, -73.52),
+        std::tr1::make_tuple(192000, 192000, kResamplingRMSError, -73.52)));
+
+}  // namespace media