/*
 * Copyright © 2016 Mozilla Foundation
 *
 * This program is made available under an ISC-style license.  See the
 * accompanying file LICENSE for details.
 */
#ifndef NOMINMAX
#define NOMINMAX
#endif // NOMINMAX
#include "cubeb/cubeb.h"
#include "cubeb_audio_dump.h"
#include "cubeb_log.h"
#include "cubeb_resampler.h"
// #define ENABLE_NORMAL_LOG
// #define ENABLE_VERBOSE_LOG
#include "common.h"
#include "cubeb_resampler_internal.h"
#include "gtest/gtest.h"
#include <algorithm>
#include <cmath>
#include <iostream>
#include <queue>
#include <stdio.h>
#include <thread>

/* Windows cmath USE_MATH_DEFINE thing... */
const float PI = 3.14159265359f;

/* Testing all sample rates is very long, so if THOROUGH_TESTING is not defined,
 * only part of the test suite is ran. */
#ifdef THOROUGH_TESTING
/* Some standard sample rates we're testing with. */
const uint32_t sample_rates[] = {8000,  16000, 32000, 44100,
                                 48000, 88200, 96000, 192000};
/* The maximum number of channels we're resampling. */
const uint32_t max_channels = 2;
/* The minimum an maximum number of milliseconds we're resampling for. This is
 * used to simulate the fact that the audio stream is resampled in chunks,
 * because audio is delivered using callbacks. */
const uint32_t min_chunks = 10; /* ms */
const uint32_t max_chunks = 30; /* ms */
const uint32_t chunk_increment = 1;

#else

const uint32_t sample_rates[] = {
    8000,
    44100,
    48000,
};
const uint32_t max_channels = 2;
const uint32_t min_chunks = 10; /* ms */
const uint32_t max_chunks = 30; /* ms */
const uint32_t chunk_increment = 10;
#endif

// #define DUMP_ARRAYS
#ifdef DUMP_ARRAYS
/**
 * Files produced by dump(...) can be converted to .wave files using:
 *
 * sox -c <channel_count> -r <rate> -e float -b 32  file.raw file.wav
 *
 * for floating-point audio, or:
 *
 * sox -c <channel_count> -r <rate> -e unsigned -b 16  file.raw file.wav
 *
 * for 16bit integer audio.
 */

/* Use the correct implementation of fopen, depending on the platform. */
void
fopen_portable(FILE ** f, const char * name, const char * mode)
{
#ifdef WIN32
  fopen_s(f, name, mode);
#else
  *f = fopen(name, mode);
#endif
}

template <typename T>
void
dump(const char * name, T * frames, size_t count)
{
  FILE * file;
  fopen_portable(&file, name, "wb");

  if (!file) {
    fprintf(stderr, "error opening %s\n", name);
    return;
  }

  if (count != fwrite(frames, sizeof(T), count, file)) {
    fprintf(stderr, "error writing to %s\n", name);
  }
  fclose(file);
}
#else
template <typename T>
void
dump(const char * name, T * frames, size_t count)
{
}
#endif

// The more the ratio is far from 1, the more we accept a big error.
float
epsilon_tweak_ratio(float ratio)
{
  return ratio >= 1 ? ratio : 1 / ratio;
}

// Epsilon values for comparing resampled data to expected data.
// The bigger the resampling ratio is, the more lax we are about errors.
template <typename T>
T
epsilon(float ratio);

template <>
float
epsilon(float ratio)
{
  return 0.08f * epsilon_tweak_ratio(ratio);
}

template <>
int16_t
epsilon(float ratio)
{
  return static_cast<int16_t>(10 * epsilon_tweak_ratio(ratio));
}

void
test_delay_lines(uint32_t delay_frames, uint32_t channels, uint32_t chunk_ms)
{
  const size_t length_s = 2;
  const size_t rate = 44100;
  const size_t length_frames = rate * length_s;
  delay_line<float> delay(delay_frames, channels, rate);
  auto_array<float> input;
  auto_array<float> output;
  uint32_t chunk_length = channels * chunk_ms * rate / 1000;
  uint32_t output_offset = 0;
  uint32_t channel = 0;

  /** Generate diracs every 100 frames, and check they are delayed. */
  input.push_silence(length_frames * channels);
  for (uint32_t i = 0; i < input.length() - 1; i += 100) {
    input.data()[i + channel] = 0.5;
    channel = (channel + 1) % channels;
  }
  dump("input.raw", input.data(), input.length());
  while (input.length()) {
    uint32_t to_pop =
        std::min<uint32_t>(input.length(), chunk_length * channels);
    float * in = delay.input_buffer(to_pop / channels);
    input.pop(in, to_pop);
    delay.written(to_pop / channels);
    output.push_silence(to_pop);
    delay.output(output.data() + output_offset, to_pop / channels);
    output_offset += to_pop;
  }

  // Check the diracs have been shifted by `delay_frames` frames.
  for (uint32_t i = 0; i < output.length() - delay_frames * channels + 1;
       i += 100) {
    ASSERT_EQ(output.data()[i + channel + delay_frames * channels], 0.5);
    channel = (channel + 1) % channels;
  }

  dump("output.raw", output.data(), output.length());
}
/**
 * This takes sine waves with a certain `channels` count, `source_rate`, and
 * resample them, by chunk of `chunk_duration` milliseconds, to `target_rate`.
 * Then a sample-wise comparison is performed against a sine wave generated at
 * the correct rate.
 */
template <typename T>
void
test_resampler_one_way(uint32_t channels, uint32_t source_rate,
                       uint32_t target_rate, float chunk_duration)
{
  size_t chunk_duration_in_source_frames =
      static_cast<uint32_t>(ceil(chunk_duration * source_rate / 1000.));
  float resampling_ratio = static_cast<float>(source_rate) / target_rate;
  cubeb_resampler_speex_one_way<T> resampler(channels, source_rate, target_rate,
                                             3);
  auto_array<T> source(channels * source_rate * 10);
  auto_array<T> destination(channels * target_rate * 10);
  auto_array<T> expected(channels * target_rate * 10);
  uint32_t phase_index = 0;
  uint32_t offset = 0;
  const uint32_t buf_len = 2; /* seconds */

  // generate a sine wave in each channel, at the source sample rate
  source.push_silence(channels * source_rate * buf_len);
  while (offset != source.length()) {
    float p = phase_index++ / static_cast<float>(source_rate);
    for (uint32_t j = 0; j < channels; j++) {
      source.data()[offset++] = 0.5 * sin(440. * 2 * PI * p);
    }
  }

  dump("input.raw", source.data(), source.length());

  expected.push_silence(channels * target_rate * buf_len);
  // generate a sine wave in each channel, at the target sample rate.
  // Insert silent samples at the beginning to account for the resampler
  // latency.
  offset = resampler.latency() * channels;
  for (uint32_t i = 0; i < offset; i++) {
    expected.data()[i] = 0.0f;
  }
  phase_index = 0;
  while (offset != expected.length()) {
    float p = phase_index++ / static_cast<float>(target_rate);
    for (uint32_t j = 0; j < channels; j++) {
      expected.data()[offset++] = 0.5 * sin(440. * 2 * PI * p);
    }
  }

  dump("expected.raw", expected.data(), expected.length());

  // resample by chunk
  uint32_t write_offset = 0;
  destination.push_silence(channels * target_rate * buf_len);
  while (write_offset < destination.length()) {
    size_t output_frames = static_cast<uint32_t>(
        floor(chunk_duration_in_source_frames / resampling_ratio));
    uint32_t input_frames = resampler.input_needed_for_output(output_frames);
    resampler.input(source.data(), input_frames);
    source.pop(nullptr, input_frames * channels);
    resampler.output(
        destination.data() + write_offset,
        std::min(output_frames,
                 (destination.length() - write_offset) / channels));
    write_offset += output_frames * channels;
  }

  dump("output.raw", destination.data(), expected.length());

  // compare, taking the latency into account
  bool fuzzy_equal = true;
  for (uint32_t i = resampler.latency() + 1; i < expected.length(); i++) {
    float diff = fabs(expected.data()[i] - destination.data()[i]);
    if (diff > epsilon<T>(resampling_ratio)) {
      fprintf(stderr, "divergence at %d: %f %f (delta %f)\n", i,
              expected.data()[i], destination.data()[i], diff);
      fuzzy_equal = false;
    }
  }
  ASSERT_TRUE(fuzzy_equal);
}

template <typename T>
cubeb_sample_format
cubeb_format();

template <>
cubeb_sample_format
cubeb_format<float>()
{
  return CUBEB_SAMPLE_FLOAT32NE;
}

template <>
cubeb_sample_format
cubeb_format<short>()
{
  return CUBEB_SAMPLE_S16NE;
}

struct osc_state {
  osc_state()
      : input_phase_index(0), output_phase_index(0), output_offset(0),
        input_channels(0), output_channels(0)
  {
  }
  uint32_t input_phase_index;
  uint32_t max_output_phase_index;
  uint32_t output_phase_index;
  uint32_t output_offset;
  uint32_t input_channels;
  uint32_t output_channels;
  uint32_t output_rate;
  uint32_t target_rate;
  auto_array<float> input;
  auto_array<float> output;
};

uint32_t
fill_with_sine(float * buf, uint32_t rate, uint32_t channels, uint32_t frames,
               uint32_t initial_phase)
{
  uint32_t offset = 0;
  for (uint32_t i = 0; i < frames; i++) {
    float p = initial_phase++ / static_cast<float>(rate);
    for (uint32_t j = 0; j < channels; j++) {
      buf[offset++] = 0.5 * sin(440. * 2 * PI * p);
    }
  }
  return initial_phase;
}

long
data_cb_resampler(cubeb_stream * /*stm*/, void * user_ptr,
                  const void * input_buffer, void * output_buffer,
                  long frame_count)
{
  osc_state * state = reinterpret_cast<osc_state *>(user_ptr);
  const float * in = reinterpret_cast<const float *>(input_buffer);
  float * out = reinterpret_cast<float *>(output_buffer);

  state->input.push(in, frame_count * state->input_channels);

  /* Check how much output frames we need to write */
  uint32_t remaining =
      state->max_output_phase_index - state->output_phase_index;
  uint32_t to_write = std::min<uint32_t>(remaining, frame_count);
  state->output_phase_index =
      fill_with_sine(out, state->target_rate, state->output_channels, to_write,
                     state->output_phase_index);

  return to_write;
}

template <typename T>
bool
array_fuzzy_equal(const auto_array<T> & lhs, const auto_array<T> & rhs, T epsi)
{
  uint32_t len = std::min(lhs.length(), rhs.length());

  for (uint32_t i = 0; i < len; i++) {
    if (fabs(lhs.at(i) - rhs.at(i)) > epsi) {
      std::cout << "not fuzzy equal at index: " << i << " lhs: " << lhs.at(i)
                << " rhs: " << rhs.at(i)
                << " delta: " << fabs(lhs.at(i) - rhs.at(i))
                << " epsilon: " << epsi << std::endl;
      return false;
    }
  }
  return true;
}

template <typename T>
void
test_resampler_duplex(uint32_t input_channels, uint32_t output_channels,
                      uint32_t input_rate, uint32_t output_rate,
                      uint32_t target_rate, float chunk_duration)
{
  cubeb_stream_params input_params;
  cubeb_stream_params output_params;
  osc_state state;

  input_params.format = output_params.format = cubeb_format<T>();
  state.input_channels = input_params.channels = input_channels;
  state.output_channels = output_params.channels = output_channels;
  input_params.rate = input_rate;
  state.output_rate = output_params.rate = output_rate;
  state.target_rate = target_rate;
  input_params.prefs = output_params.prefs = CUBEB_STREAM_PREF_NONE;
  long got;

  cubeb_resampler * resampler = cubeb_resampler_create(
      (cubeb_stream *)nullptr, &input_params, &output_params, target_rate,
      data_cb_resampler, (void *)&state, CUBEB_RESAMPLER_QUALITY_VOIP,
      CUBEB_RESAMPLER_RECLOCK_NONE);

  long latency = cubeb_resampler_latency(resampler);

  const uint32_t duration_s = 2;
  int32_t duration_frames = duration_s * target_rate;
  uint32_t input_array_frame_count =
      ceil(chunk_duration * input_rate / 1000) +
      ceilf(static_cast<float>(input_rate) / target_rate) * 2;
  uint32_t output_array_frame_count = chunk_duration * output_rate / 1000;
  auto_array<float> input_buffer(input_channels * input_array_frame_count);
  auto_array<float> output_buffer(output_channels * output_array_frame_count);
  auto_array<float> expected_resampled_input(input_channels * duration_frames);
  auto_array<float> expected_resampled_output(output_channels * output_rate *
                                              duration_s);

  state.max_output_phase_index = duration_s * target_rate;

  expected_resampled_input.push_silence(input_channels * duration_frames);
  expected_resampled_output.push_silence(output_channels * output_rate *
                                         duration_s);

  /* expected output is a 440Hz sine wave at 16kHz */
  fill_with_sine(expected_resampled_input.data() + latency, target_rate,
                 input_channels, duration_frames - latency, 0);
  /* expected output is a 440Hz sine wave at 32kHz */
  fill_with_sine(expected_resampled_output.data() + latency, output_rate,
                 output_channels, output_rate * duration_s - latency, 0);

  while (state.output_phase_index != state.max_output_phase_index) {
    uint32_t leftover_samples = input_buffer.length() * input_channels;
    input_buffer.reserve(input_array_frame_count);
    state.input_phase_index = fill_with_sine(
        input_buffer.data() + leftover_samples, input_rate, input_channels,
        input_array_frame_count - leftover_samples, state.input_phase_index);
    long input_consumed = input_array_frame_count;
    input_buffer.set_length(input_array_frame_count);

    got = cubeb_resampler_fill(resampler, input_buffer.data(), &input_consumed,
                               output_buffer.data(), output_array_frame_count);

    /* handle leftover input */
    if (input_array_frame_count != static_cast<uint32_t>(input_consumed)) {
      input_buffer.pop(nullptr, input_consumed * input_channels);
    } else {
      input_buffer.clear();
    }

    state.output.push(output_buffer.data(), got * state.output_channels);
  }

  dump("input_expected.raw", expected_resampled_input.data(),
       expected_resampled_input.length());
  dump("output_expected.raw", expected_resampled_output.data(),
       expected_resampled_output.length());
  dump("input.raw", state.input.data(), state.input.length());
  dump("output.raw", state.output.data(), state.output.length());

  // This is disabled because the latency estimation in the resampler code is
  // slightly off so we can generate expected vectors.
  // See https://github.com/kinetiknz/cubeb/issues/93
  // ASSERT_TRUE(array_fuzzy_equal(state.input, expected_resampled_input,
  // epsilon<T>(input_rate/target_rate)));
  // ASSERT_TRUE(array_fuzzy_equal(state.output, expected_resampled_output,
  // epsilon<T>(output_rate/target_rate)));

  cubeb_resampler_destroy(resampler);
}

#define array_size(x) (sizeof(x) / sizeof(x[0]))

TEST(cubeb, resampler_one_way)
{
  /* Test one way resamplers */
  for (uint32_t channels = 1; channels <= max_channels; channels++) {
    for (uint32_t source_rate = 0; source_rate < array_size(sample_rates);
         source_rate++) {
      for (uint32_t dest_rate = 0; dest_rate < array_size(sample_rates);
           dest_rate++) {
        for (uint32_t chunk_duration = min_chunks; chunk_duration < max_chunks;
             chunk_duration += chunk_increment) {
          fprintf(stderr,
                  "one_way: channels: %d, source_rate: %d, dest_rate: %d, "
                  "chunk_duration: %d\n",
                  channels, sample_rates[source_rate], sample_rates[dest_rate],
                  chunk_duration);
          test_resampler_one_way<float>(channels, sample_rates[source_rate],
                                        sample_rates[dest_rate],
                                        chunk_duration);
        }
      }
    }
  }
}

TEST(cubeb, DISABLED_resampler_duplex)
{
  for (uint32_t input_channels = 1; input_channels <= max_channels;
       input_channels++) {
    for (uint32_t output_channels = 1; output_channels <= max_channels;
         output_channels++) {
      for (uint32_t source_rate_input = 0;
           source_rate_input < array_size(sample_rates); source_rate_input++) {
        for (uint32_t source_rate_output = 0;
             source_rate_output < array_size(sample_rates);
             source_rate_output++) {
          for (uint32_t dest_rate = 0; dest_rate < array_size(sample_rates);
               dest_rate++) {
            for (uint32_t chunk_duration = min_chunks;
                 chunk_duration < max_chunks;
                 chunk_duration += chunk_increment) {
              fprintf(stderr,
                      "input channels:%d output_channels:%d input_rate:%d "
                      "output_rate:%d target_rate:%d chunk_ms:%d\n",
                      input_channels, output_channels,
                      sample_rates[source_rate_input],
                      sample_rates[source_rate_output], sample_rates[dest_rate],
                      chunk_duration);
              test_resampler_duplex<float>(input_channels, output_channels,
                                           sample_rates[source_rate_input],
                                           sample_rates[source_rate_output],
                                           sample_rates[dest_rate],
                                           chunk_duration);
            }
          }
        }
      }
    }
  }
}

TEST(cubeb, resampler_delay_line)
{
  for (uint32_t channel = 1; channel <= 2; channel++) {
    for (uint32_t delay_frames = 4; delay_frames <= 40;
         delay_frames += chunk_increment) {
      for (uint32_t chunk_size = 10; chunk_size <= 30; chunk_size++) {
        fprintf(stderr, "channel: %d, delay_frames: %d, chunk_size: %d\n",
                channel, delay_frames, chunk_size);
        test_delay_lines(delay_frames, channel, chunk_size);
      }
    }
  }
}

long
test_output_only_noop_data_cb(cubeb_stream * /*stm*/, void * /*user_ptr*/,
                              const void * input_buffer, void * output_buffer,
                              long frame_count)
{
  EXPECT_TRUE(output_buffer);
  EXPECT_TRUE(!input_buffer);
  return frame_count;
}

TEST(cubeb, resampler_output_only_noop)
{
  cubeb_stream_params output_params;
  int target_rate;

  output_params.rate = 44100;
  output_params.channels = 1;
  output_params.format = CUBEB_SAMPLE_FLOAT32NE;
  target_rate = output_params.rate;

  cubeb_resampler * resampler = cubeb_resampler_create(
      (cubeb_stream *)nullptr, nullptr, &output_params, target_rate,
      test_output_only_noop_data_cb, nullptr, CUBEB_RESAMPLER_QUALITY_VOIP,
      CUBEB_RESAMPLER_RECLOCK_NONE);
  const long out_frames = 128;
  float out_buffer[out_frames];
  long got;

  got =
      cubeb_resampler_fill(resampler, nullptr, nullptr, out_buffer, out_frames);

  ASSERT_EQ(got, out_frames);

  cubeb_resampler_destroy(resampler);
}

long
test_drain_data_cb(cubeb_stream * /*stm*/, void * user_ptr,
                   const void * input_buffer, void * output_buffer,
                   long frame_count)
{
  EXPECT_TRUE(output_buffer);
  EXPECT_TRUE(!input_buffer);
  auto cb_count = static_cast<int *>(user_ptr);
  (*cb_count)++;
  return frame_count - 1;
}

TEST(cubeb, resampler_drain)
{
  cubeb_stream_params output_params;
  int target_rate;

  output_params.rate = 44100;
  output_params.channels = 1;
  output_params.format = CUBEB_SAMPLE_FLOAT32NE;
  target_rate = 48000;
  int cb_count = 0;

  cubeb_resampler * resampler = cubeb_resampler_create(
      (cubeb_stream *)nullptr, nullptr, &output_params, target_rate,
      test_drain_data_cb, &cb_count, CUBEB_RESAMPLER_QUALITY_VOIP,
      CUBEB_RESAMPLER_RECLOCK_NONE);

  const long out_frames = 128;
  float out_buffer[out_frames];
  long got;

  do {
    got = cubeb_resampler_fill(resampler, nullptr, nullptr, out_buffer,
                               out_frames);
  } while (got == out_frames);

  /* The callback should be called once but not again after returning <
   * frame_count. */
  ASSERT_EQ(cb_count, 1);

  cubeb_resampler_destroy(resampler);
}

// gtest does not support using ASSERT_EQ and friend in a function that returns
// a value.
void
check_output(const void * input_buffer, void * output_buffer, long frame_count)
{
  ASSERT_EQ(input_buffer, nullptr);
  ASSERT_EQ(frame_count, 256);
  ASSERT_TRUE(!!output_buffer);
}

long
cb_passthrough_resampler_output(cubeb_stream * /*stm*/, void * /*user_ptr*/,
                                const void * input_buffer, void * output_buffer,
                                long frame_count)
{
  check_output(input_buffer, output_buffer, frame_count);
  return frame_count;
}

TEST(cubeb, resampler_passthrough_output_only)
{
  // Test that the passthrough resampler works when there is only an output
  // stream.
  cubeb_stream_params output_params;

  const size_t output_channels = 2;
  output_params.channels = output_channels;
  output_params.rate = 44100;
  output_params.format = CUBEB_SAMPLE_FLOAT32NE;
  int target_rate = output_params.rate;

  cubeb_resampler * resampler = cubeb_resampler_create(
      (cubeb_stream *)nullptr, nullptr, &output_params, target_rate,
      cb_passthrough_resampler_output, nullptr, CUBEB_RESAMPLER_QUALITY_VOIP,
      CUBEB_RESAMPLER_RECLOCK_NONE);

  float output_buffer[output_channels * 256];

  long got;
  for (uint32_t i = 0; i < 30; i++) {
    got = cubeb_resampler_fill(resampler, nullptr, nullptr, output_buffer, 256);
    ASSERT_EQ(got, 256);
  }

  cubeb_resampler_destroy(resampler);
}

// gtest does not support using ASSERT_EQ and friend in a function that returns
// a value.
void
check_input(const void * input_buffer, void * output_buffer, long frame_count)
{
  ASSERT_EQ(output_buffer, nullptr);
  ASSERT_EQ(frame_count, 256);
  ASSERT_TRUE(!!input_buffer);
}

long
cb_passthrough_resampler_input(cubeb_stream * /*stm*/, void * /*user_ptr*/,
                               const void * input_buffer, void * output_buffer,
                               long frame_count)
{
  check_input(input_buffer, output_buffer, frame_count);
  return frame_count;
}

TEST(cubeb, resampler_passthrough_input_only)
{
  // Test that the passthrough resampler works when there is only an output
  // stream.
  cubeb_stream_params input_params;

  const size_t input_channels = 2;
  input_params.channels = input_channels;
  input_params.rate = 44100;
  input_params.format = CUBEB_SAMPLE_FLOAT32NE;
  int target_rate = input_params.rate;

  cubeb_resampler * resampler = cubeb_resampler_create(
      (cubeb_stream *)nullptr, &input_params, nullptr, target_rate,
      cb_passthrough_resampler_input, nullptr, CUBEB_RESAMPLER_QUALITY_VOIP,
      CUBEB_RESAMPLER_RECLOCK_NONE);

  float input_buffer[input_channels * 256];

  long got;
  for (uint32_t i = 0; i < 30; i++) {
    long int frames = 256;
    got = cubeb_resampler_fill(resampler, input_buffer, &frames, nullptr, 0);
    ASSERT_EQ(got, 256);
  }

  cubeb_resampler_destroy(resampler);
}

template <typename T>
long
seq(T * array, int stride, long start, long count)
{
  uint32_t output_idx = 0;
  for (int i = 0; i < count; i++) {
    for (int j = 0; j < stride; j++) {
      array[output_idx + j] = static_cast<T>(start + i);
    }
    output_idx += stride;
  }
  return start + count;
}

template <typename T>
void
is_seq(T * array, int stride, long count, long expected_start)
{
  uint32_t output_index = 0;
  for (long i = 0; i < count; i++) {
    for (int j = 0; j < stride; j++) {
      ASSERT_EQ(array[output_index + j], expected_start + i);
    }
    output_index += stride;
  }
}

template <typename T>
void
is_not_seq(T * array, int stride, long count, long expected_start)
{
  uint32_t output_index = 0;
  for (long i = 0; i < count; i++) {
    for (int j = 0; j < stride; j++) {
      ASSERT_NE(array[output_index + j], expected_start + i);
    }
    output_index += stride;
  }
}

struct closure {
  int input_channel_count;
};

// gtest does not support using ASSERT_EQ and friend in a function that returns
// a value.
template <typename T>
void
check_duplex(const T * input_buffer, T * output_buffer, long frame_count,
             int input_channel_count)
{
  ASSERT_EQ(frame_count, 256);
  // Silence scan-build warning.
  ASSERT_TRUE(!!output_buffer);
  assert(output_buffer);
  ASSERT_TRUE(!!input_buffer);
  assert(input_buffer);

  int output_index = 0;
  int input_index = 0;
  for (int i = 0; i < frame_count; i++) {
    // output is two channels, input one or two channels.
    if (input_channel_count == 1) {
      output_buffer[output_index] = output_buffer[output_index + 1] =
          input_buffer[i];
    } else if (input_channel_count == 2) {
      output_buffer[output_index] = input_buffer[input_index];
      output_buffer[output_index + 1] = input_buffer[input_index + 1];
    }
    output_index += 2;
    input_index += input_channel_count;
  }
}

long
cb_passthrough_resampler_duplex(cubeb_stream * /*stm*/, void * user_ptr,
                                const void * input_buffer, void * output_buffer,
                                long frame_count)
{
  closure * c = reinterpret_cast<closure *>(user_ptr);
  check_duplex<float>(static_cast<const float *>(input_buffer),
                      static_cast<float *>(output_buffer), frame_count,
                      c->input_channel_count);
  return frame_count;
}

TEST(cubeb, resampler_passthrough_duplex_callback_reordering)
{
  // Test that when pre-buffering on resampler creation, we can survive an input
  // callback being delayed.

  cubeb_stream_params input_params;
  cubeb_stream_params output_params;

  const int input_channels = 1;
  const int output_channels = 2;

  input_params.channels = input_channels;
  input_params.rate = 44100;
  input_params.format = CUBEB_SAMPLE_FLOAT32NE;

  output_params.channels = output_channels;
  output_params.rate = input_params.rate;
  output_params.format = CUBEB_SAMPLE_FLOAT32NE;

  int target_rate = input_params.rate;

  closure c;
  c.input_channel_count = input_channels;

  cubeb_resampler * resampler = cubeb_resampler_create(
      (cubeb_stream *)nullptr, &input_params, &output_params, target_rate,
      cb_passthrough_resampler_duplex, &c, CUBEB_RESAMPLER_QUALITY_VOIP,
      CUBEB_RESAMPLER_RECLOCK_NONE);

  const long BUF_BASE_SIZE = 256;
  float input_buffer_prebuffer[input_channels * BUF_BASE_SIZE * 2];
  float input_buffer_glitch[input_channels * BUF_BASE_SIZE * 2];
  float input_buffer_normal[input_channels * BUF_BASE_SIZE];
  float output_buffer[output_channels * BUF_BASE_SIZE];

  long seq_idx = 0;
  long output_seq_idx = 0;

  long prebuffer_frames =
      ARRAY_LENGTH(input_buffer_prebuffer) / input_params.channels;
  seq_idx =
      seq(input_buffer_prebuffer, input_channels, seq_idx, prebuffer_frames);

  long got =
      cubeb_resampler_fill(resampler, input_buffer_prebuffer, &prebuffer_frames,
                           output_buffer, BUF_BASE_SIZE);

  output_seq_idx += BUF_BASE_SIZE;

  // prebuffer_frames will hold the frames used by the resampler.
  ASSERT_EQ(prebuffer_frames, BUF_BASE_SIZE);
  ASSERT_EQ(got, BUF_BASE_SIZE);

  for (uint32_t i = 0; i < 300; i++) {
    long int frames = BUF_BASE_SIZE;
    // Simulate that sometimes, we don't have the input callback on time
    if (i != 0 && (i % 100) == 0) {
      long zero = 0;
      got =
          cubeb_resampler_fill(resampler, input_buffer_normal /* unused here */,
                               &zero, output_buffer, BUF_BASE_SIZE);
      is_seq(output_buffer, 2, BUF_BASE_SIZE, output_seq_idx);
      output_seq_idx += BUF_BASE_SIZE;
    } else if (i != 0 && (i % 100) == 1) {
      // if this is the case, the on the next iteration, we'll have twice the
      // amount of input frames
      seq_idx =
          seq(input_buffer_glitch, input_channels, seq_idx, BUF_BASE_SIZE * 2);
      frames = 2 * BUF_BASE_SIZE;
      got = cubeb_resampler_fill(resampler, input_buffer_glitch, &frames,
                                 output_buffer, BUF_BASE_SIZE);
      is_seq(output_buffer, 2, BUF_BASE_SIZE, output_seq_idx);
      output_seq_idx += BUF_BASE_SIZE;
    } else {
      // normal case
      seq_idx =
          seq(input_buffer_normal, input_channels, seq_idx, BUF_BASE_SIZE);
      long normal_input_frame_count = 256;
      got = cubeb_resampler_fill(resampler, input_buffer_normal,
                                 &normal_input_frame_count, output_buffer,
                                 BUF_BASE_SIZE);
      is_seq(output_buffer, 2, BUF_BASE_SIZE, output_seq_idx);
      output_seq_idx += BUF_BASE_SIZE;
    }
    ASSERT_EQ(got, BUF_BASE_SIZE);
  }

  cubeb_resampler_destroy(resampler);
}

// Artificially simulate output thread underruns,
// by building up artificial delay in the input.
// Check that the frame drop logic kicks in.
TEST(cubeb, resampler_drift_drop_data)
{
  for (uint32_t input_channels = 1; input_channels < 3; input_channels++) {
    cubeb_stream_params input_params;
    cubeb_stream_params output_params;

    const int output_channels = 2;
    const int sample_rate = 44100;

    input_params.channels = input_channels;
    input_params.rate = sample_rate;
    input_params.format = CUBEB_SAMPLE_FLOAT32NE;

    output_params.channels = output_channels;
    output_params.rate = sample_rate;
    output_params.format = CUBEB_SAMPLE_FLOAT32NE;

    int target_rate = input_params.rate;

    closure c;
    c.input_channel_count = input_channels;

    cubeb_resampler * resampler = cubeb_resampler_create(
        (cubeb_stream *)nullptr, &input_params, &output_params, target_rate,
        cb_passthrough_resampler_duplex, &c, CUBEB_RESAMPLER_QUALITY_VOIP,
        CUBEB_RESAMPLER_RECLOCK_NONE);

    const long BUF_BASE_SIZE = 256;

    // The factor by which the deadline is missed. This is intentionally
    // kind of large to trigger the frame drop quickly. In real life, multiple
    // smaller under-runs would accumulate.
    const long UNDERRUN_FACTOR = 10;
    // Number buffer used for pre-buffering, that some backends do.
    const long PREBUFFER_FACTOR = 2;

    std::vector<float> input_buffer_prebuffer(input_channels * BUF_BASE_SIZE *
                                              PREBUFFER_FACTOR);
    std::vector<float> input_buffer_glitch(input_channels * BUF_BASE_SIZE *
                                           UNDERRUN_FACTOR);
    std::vector<float> input_buffer_normal(input_channels * BUF_BASE_SIZE);
    std::vector<float> output_buffer(output_channels * BUF_BASE_SIZE);

    long seq_idx = 0;
    long output_seq_idx = 0;

    long prebuffer_frames =
        input_buffer_prebuffer.size() / input_params.channels;
    seq_idx = seq(input_buffer_prebuffer.data(), input_channels, seq_idx,
                  prebuffer_frames);

    long got = cubeb_resampler_fill(resampler, input_buffer_prebuffer.data(),
                                    &prebuffer_frames, output_buffer.data(),
                                    BUF_BASE_SIZE);

    output_seq_idx += BUF_BASE_SIZE;

    // prebuffer_frames will hold the frames used by the resampler.
    ASSERT_EQ(prebuffer_frames, BUF_BASE_SIZE);
    ASSERT_EQ(got, BUF_BASE_SIZE);

    for (uint32_t i = 0; i < 300; i++) {
      long int frames = BUF_BASE_SIZE;
      if (i != 0 && (i % 100) == 1) {
        // Once in a while, the output thread misses its deadline.
        // The input thread still produces data, so it ends up accumulating.
        // Simulate this by providing a much bigger input buffer. Check that the
        // sequence is now unaligned, meaning we've dropped data to keep
        // everything in sync.
        seq_idx = seq(input_buffer_glitch.data(), input_channels, seq_idx,
                      BUF_BASE_SIZE * UNDERRUN_FACTOR);
        frames = BUF_BASE_SIZE * UNDERRUN_FACTOR;
        got =
            cubeb_resampler_fill(resampler, input_buffer_glitch.data(), &frames,
                                 output_buffer.data(), BUF_BASE_SIZE);
        is_seq(output_buffer.data(), 2, BUF_BASE_SIZE, output_seq_idx);
        output_seq_idx += BUF_BASE_SIZE;
      } else if (i != 0 && (i % 100) == 2) {
        // On the next iteration, the sequence should be broken
        seq_idx = seq(input_buffer_normal.data(), input_channels, seq_idx,
                      BUF_BASE_SIZE);
        long normal_input_frame_count = 256;
        got = cubeb_resampler_fill(resampler, input_buffer_normal.data(),
                                   &normal_input_frame_count,
                                   output_buffer.data(), BUF_BASE_SIZE);
        is_not_seq(output_buffer.data(), output_channels, BUF_BASE_SIZE,
                   output_seq_idx);
        // Reclock so that we can use is_seq again.
        output_seq_idx = output_buffer[BUF_BASE_SIZE * output_channels - 1] + 1;
      } else {
        // normal case
        seq_idx = seq(input_buffer_normal.data(), input_channels, seq_idx,
                      BUF_BASE_SIZE);
        long normal_input_frame_count = 256;
        got = cubeb_resampler_fill(resampler, input_buffer_normal.data(),
                                   &normal_input_frame_count,
                                   output_buffer.data(), BUF_BASE_SIZE);
        is_seq(output_buffer.data(), output_channels, BUF_BASE_SIZE,
               output_seq_idx);
        output_seq_idx += BUF_BASE_SIZE;
      }
      ASSERT_EQ(got, BUF_BASE_SIZE);
    }

    cubeb_resampler_destroy(resampler);
  }
}

static long
passthrough_resampler_fill_eq_input(cubeb_stream * stream, void * user_ptr,
                                    void const * input_buffer,
                                    void * output_buffer, long nframes)
{
  // gtest does not support using ASSERT_EQ and friends in a
  // function that returns a value.
  [nframes, input_buffer]() {
    ASSERT_EQ(nframes, 32);
    const float * input = static_cast<const float *>(input_buffer);
    for (int i = 0; i < 64; ++i) {
      ASSERT_FLOAT_EQ(input[i], 0.01 * i);
    }
  }();
  return nframes;
}

TEST(cubeb, passthrough_resampler_fill_eq_input)
{
  uint32_t channels = 2;
  uint32_t sample_rate = 44100;
  passthrough_resampler<float> resampler =
      passthrough_resampler<float>(nullptr, passthrough_resampler_fill_eq_input,
                                   nullptr, channels, sample_rate);

  long input_frame_count = 32;
  long output_frame_count = 32;
  float input[64] = {};
  float output[64] = {};
  for (uint32_t i = 0; i < input_frame_count * channels; ++i) {
    input[i] = 0.01 * i;
  }
  long got =
      resampler.fill(input, &input_frame_count, output, output_frame_count);
  ASSERT_EQ(got, output_frame_count);
  // Input frames used must be equal to output frames.
  ASSERT_EQ(input_frame_count, output_frame_count);
}

static long
passthrough_resampler_fill_short_input(cubeb_stream * stream, void * user_ptr,
                                       void const * input_buffer,
                                       void * output_buffer, long nframes)
{
  // gtest does not support using ASSERT_EQ and friends in a
  // function that returns a value.
  [nframes, input_buffer]() {
    ASSERT_EQ(nframes, 32);
    const float * input = static_cast<const float *>(input_buffer);
    // First part contains the input
    for (int i = 0; i < 32; ++i) {
      ASSERT_FLOAT_EQ(input[i], 0.01 * i);
    }
    // missing part contains silence
    for (int i = 32; i < 64; ++i) {
      ASSERT_FLOAT_EQ(input[i], 0.0);
    }
  }();
  return nframes;
}

TEST(cubeb, passthrough_resampler_fill_short_input)
{
  uint32_t channels = 2;
  uint32_t sample_rate = 44100;
  passthrough_resampler<float> resampler = passthrough_resampler<float>(
      nullptr, passthrough_resampler_fill_short_input, nullptr, channels,
      sample_rate);

  long input_frame_count = 16;
  long output_frame_count = 32;
  float input[64] = {};
  float output[64] = {};
  for (uint32_t i = 0; i < input_frame_count * channels; ++i) {
    input[i] = 0.01 * i;
  }
  long got =
      resampler.fill(input, &input_frame_count, output, output_frame_count);
  ASSERT_EQ(got, output_frame_count);
  // Input frames used are less than the output frames due to glitch.
  ASSERT_EQ(input_frame_count, output_frame_count - 16);
}

static long
passthrough_resampler_fill_input_left(cubeb_stream * stream, void * user_ptr,
                                      void const * input_buffer,
                                      void * output_buffer, long nframes)
{
  // gtest does not support using ASSERT_EQ and friends in a
  // function that returns a value.
  int iteration = *static_cast<int *>(user_ptr);
  if (iteration == 1) {
    [nframes, input_buffer]() {
      ASSERT_EQ(nframes, 32);
      const float * input = static_cast<const float *>(input_buffer);
      for (int i = 0; i < 64; ++i) {
        ASSERT_FLOAT_EQ(input[i], 0.01 * i);
      }
    }();
  } else if (iteration == 2) {
    [nframes, input_buffer]() {
      ASSERT_EQ(nframes, 32);
      const float * input = static_cast<const float *>(input_buffer);
      for (int i = 0; i < 32; ++i) {
        // First part contains the reamaining input samples from previous
        // iteration (since they were more).
        ASSERT_FLOAT_EQ(input[i], 0.01 * (i + 64));
        // next part contains the new buffer
        ASSERT_FLOAT_EQ(input[i + 32], 0.01 * i);
      }
    }();
  } else if (iteration == 3) {
    [nframes, input_buffer]() {
      ASSERT_EQ(nframes, 32);
      const float * input = static_cast<const float *>(input_buffer);
      for (int i = 0; i < 32; ++i) {
        // First part (16 frames) contains the reamaining input samples
        // from previous iteration (since they were more).
        ASSERT_FLOAT_EQ(input[i], 0.01 * (i + 32));
      }
      for (int i = 0; i < 16; ++i) {
        // next part (8 frames) contains the new input buffer.
        ASSERT_FLOAT_EQ(input[i + 32], 0.01 * i);
        // last part (8 frames) contains silence.
        ASSERT_FLOAT_EQ(input[i + 32 + 16], 0.0);
      }
    }();
  }
  return nframes;
}

TEST(cubeb, passthrough_resampler_fill_input_left)
{
  const uint32_t channels = 2;
  const uint32_t sample_rate = 44100;
  int iteration = 0;
  passthrough_resampler<float> resampler = passthrough_resampler<float>(
      nullptr, passthrough_resampler_fill_input_left, &iteration, channels,
      sample_rate);

  long input_frame_count = 48; // 32 + 16
  const long output_frame_count = 32;
  float input[96] = {};
  float output[64] = {};
  for (uint32_t i = 0; i < input_frame_count * channels; ++i) {
    input[i] = 0.01 * i;
  }

  // 1st iteration, add the extra input.
  iteration = 1;
  long got =
      resampler.fill(input, &input_frame_count, output, output_frame_count);
  ASSERT_EQ(got, output_frame_count);
  // Input frames used must be equal to output frames.
  ASSERT_EQ(input_frame_count, output_frame_count);

  // 2st iteration, use the extra input from previous iteration,
  // 16 frames are remaining in the input buffer.
  input_frame_count = 32; // we need 16 input frames but we get more;
  iteration = 2;
  got = resampler.fill(input, &input_frame_count, output, output_frame_count);
  ASSERT_EQ(got, output_frame_count);
  // Input frames used must be equal to output frames.
  ASSERT_EQ(input_frame_count, output_frame_count);

  // 3rd iteration, use the extra input from previous iteration.
  // 16 frames are remaining in the input buffer.
  input_frame_count = 16 - 8; // We need 16 more input frames but we only get 8.
  iteration = 3;
  got = resampler.fill(input, &input_frame_count, output, output_frame_count);
  ASSERT_EQ(got, output_frame_count);
  // Input frames used are less than the output frames due to glitch.
  ASSERT_EQ(input_frame_count, output_frame_count - 8);
}

TEST(cubeb, individual_methods)
{
  const uint32_t channels = 2;
  const uint32_t sample_rate = 44100;
  const uint32_t frames = 256;

  delay_line<float> dl(10, channels, sample_rate);
  uint32_t frames_needed1 = dl.input_needed_for_output(0);
  ASSERT_EQ(frames_needed1, 0u);

  cubeb_resampler_speex_one_way<float> one_way(
      channels, sample_rate, sample_rate, CUBEB_RESAMPLER_QUALITY_DEFAULT);
  float buffer[channels * frames] = {0.0};
  // Add all frames in the resampler's internal buffer.
  one_way.input(buffer, frames);
  // Ask for less than the existing frames, this would create a uint overlflow
  // without the fix.
  uint32_t frames_needed2 = one_way.input_needed_for_output(0);
  ASSERT_EQ(frames_needed2, 0u);
}

struct sine_wave_state {
  float frequency;
  int sample_rate;
  size_t count = 0;
  sine_wave_state(float freq, int rate) : frequency(freq), sample_rate(rate) {}
};

long
data_cb(cubeb_stream * stream, void * user_ptr, void const * input_buffer,
        void * output_buffer, long nframes)
{
  sine_wave_state * state = static_cast<sine_wave_state *>(user_ptr);
  float * out = static_cast<float *>(output_buffer);
  double phase_increment = 2.0f * M_PI * state->frequency / state->sample_rate;

  for (int i = 0; i < nframes; i++) {
    float sample = sin(phase_increment * state->count);
    state->count++;
    out[i] = sample * 0.8;
  }
  return nframes;
}

// This implements 4.6.2 from "Standard for Digitizing Waveform Recorders"
// (in particular Annex A), then returns the estimated amplitude, phase, and the
// sum of squared error relative to a sine wave sampled at `sample_rate` and of
// frequency `frequency`. This is also described in "Numerical methods for
// engineers" chapter 19.1, and explained at
// https://www.youtube.com/watch?v=afQszl_OwKo and videos of the same series.
// In practice here we're sending a perfect 1khz sine wave into a good
// resampler, and despite the resampling ratio being quite extreme sometimes,
// we're expecting a very good fit.
float
fit_sine(const std::vector<float> & signal, float sample_rate, float frequency,
         float & out_amplitude, float & out_phase)
{
  // The formulation below is exact for samples spanning an integer number of
  // periods. It can be important for `signal` to be trimmed to an integer
  // number of periods if it doesn't contain a lot of periods.
  double phase_incr = 2.0 * M_PI * frequency / sample_rate;

  double sum_cos = 0.0;
  double sum_sin = 0.0;
  for (size_t i = 0; i < signal.size(); ++i) {
    double c = std::cos(phase_incr * static_cast<double>(i));
    double s = std::sin(phase_incr * static_cast<double>(i));
    sum_cos += signal[i] * c;
    sum_sin += signal[i] * s;
  }

  double amplitude = 2.0f * std::sqrt(sum_cos * sum_cos + sum_sin * sum_sin) /
                     static_cast<double>(signal.size());
  double phi = std::atan2(sum_cos, sum_sin);

  out_amplitude = amplitude;
  out_phase = phi;

  // Compute sum of squared errors relative to the fitted sine wave
  double sse = 0.0;
  for (size_t i = 0; i < signal.size(); ++i) {
    // Use known amplitude here instead instead of the from the fitted function.
    double fit = 0.8 * std::sin(phase_incr * i + phi);
    double diff = signal[i] - fit;
    sse += diff * diff;
  }

  return sse;
}

// Finds the offset of the start of an input_freq sine wave sampled at
// target_rate in data. Remove the leading silence from data.
size_t
find_sine_start(const std::vector<float> & data, float input_freq,
                float target_rate)
{
  const size_t POINTS = 10;
  size_t skipped = 0;

  while (skipped + POINTS < data.size()) {
    double phase = 0;
    double phase_increment = 2.0f * M_PI * input_freq / target_rate;
    bool fits_sine = true;

    for (size_t i = 0; i < POINTS; i++) {
      float expected = sin(phase) * 0.8;
      float actual = data[skipped + i];
      if (fabs(expected - actual) > 0.1) {
        // doesn't fit a sine, skip to next start point
        fits_sine = false;
        break;
      }
      phase += phase_increment;
      if (phase > 2.0f * M_PI) {
        phase -= 2.0f * M_PI;
      }
    }

    if (!fits_sine) {
      skipped++;
      continue;
    }

    // Found the start of the sine wave
    size_t sine_start = skipped;
    return sine_start;
  }

  return skipped;
}

// This class tracks the monotonicity of a certain value, and reports if it
// increases too much monotonically.
struct monotonic_state {
  explicit monotonic_state(const char * what, int source_rate, int target_rate,
                           int block_size)
      : what(what), source_rate(source_rate), target_rate(target_rate),
        block_size(block_size)
  {
  }
  ~monotonic_state()
  {
    float ratio =
        static_cast<float>(source_rate) / static_cast<float>(target_rate);
    // Only report if there has been a meaningful increase in buffering. Do
    // not warn if the buffering was constant and small.
    if (monotonic && max_value && max_value != max_step) {
      printf("%s is monotonically increasing, max: %zu, max_step: %zu, "
             "in: %dHz, out: "
             "%dHz, block_size: %d, ratio: %lf\n",
             what, max_value, max_step, source_rate, target_rate, block_size,
             ratio);
    }
    // Arbitrary limit: if more than this number of frames has been buffered,
    // print a message.
    constexpr int BUFFER_SIZE_THRESHOLD = 20;
    if (max_value > BUFFER_SIZE_THRESHOLD) {
      printf("%s, unexpected large max buffering value, max: %zu, max_step: "
             "%zu, in: %dHz, out: %dHz, block_size: %d, ratio: %lf\n",
             what, max_value, max_step, source_rate, target_rate, block_size,
             ratio);
    }
  }
  void set_new_value(size_t new_value)
  {
    if (new_value < value) {
      monotonic = false;
    } else {
      max_step = std::max(max_step, new_value - value);
    }
    value = new_value;
    max_value = std::max(value, max_value);
  }
  // Textual representation of this measurement
  const char * what;
  // Resampler parameters for this test case
  int source_rate = 0;
  int target_rate = 0;
  int block_size = 0;
  // Current buffering value
  size_t value = 0;
  // Max buffering value increment
  size_t max_step = 0;
  // Max buffering value observerd
  size_t max_value = 0;
  // Whether the value has only increased or not
  bool monotonic = true;
};

// Setting this to 1 dumps a bunch of wave file to the local directory for
// manual inspection of the resampled output
constexpr int DUMP_OUTPUT = 0;

// Source and target sample-rates in Hz, typical values.
const int rates[] = {16000, 32000, 44100, 48000, 96000, 192000, 384000};
// Block size in frames, except the first element, that is in millisecond
// Power of two are typical on Windows WASAPI IAudioClient3, macOS,
// Linux Pipewire and Jack. 10ms is typical on Windows IAudioClient and
// IAudioClient2. 96, 192 are not uncommon on some Android devices.
constexpr int WASAPI_MS_BLOCK = 10;
const int block_sizes[] = {WASAPI_MS_BLOCK, 96, 128, 192, 256, 512, 1024, 2048};
// Enough iterations to catch rounding/drift issues, but not too many to avoid
// having a test that is too long to run.
constexpr int ITERATION_COUNT = 1000;
// 1 kHz input sine wave
const float input_freq = 1000.0f;

struct ThreadPool {
  std::vector<std::thread> workers;
  std::queue<std::function<void()>> tasks;
  std::mutex queue_mutex;
  std::condition_variable condition;
  bool stop;

  ThreadPool(size_t threads) : stop(false)
  {
    for (size_t i = 0; i < threads; ++i) {
      workers.emplace_back([this] {
        while (true) {
          std::function<void()> task;
          {
            std::unique_lock<std::mutex> lock(queue_mutex);
            condition.wait(lock, [this] { return stop || !tasks.empty(); });
            if (stop && tasks.empty())
              return;
            task = std::move(tasks.front());
            tasks.pop();
          }
          task();
        }
      });
    }
  }

  void enqueue(std::function<void()> task)
  {
    {
      std::unique_lock<std::mutex> lock(queue_mutex);
      tasks.push(std::move(task));
    }
    condition.notify_one();
  }

  ~ThreadPool()
  {
    {
      std::unique_lock<std::mutex> lock(queue_mutex);
      stop = true;
    }
    condition.notify_all();
    for (std::thread & worker : workers) {
      worker.join();
    }
  }
};

static void
run_test(int source_rate, int target_rate, int block_size)
{
  int effective_block_size = block_size;
  // special case: Windows/WASAPI works in blocks of 10ms regardless of
  // the rate.
  if (effective_block_size == WASAPI_MS_BLOCK) {
    effective_block_size = target_rate / 100; // 10ms
  }
  sine_wave_state state(input_freq, source_rate);
  cubeb_stream_params out_params = {};
  out_params.channels = 1;
  out_params.rate = target_rate;
  out_params.format = CUBEB_SAMPLE_FLOAT32NE;

  cubeb_audio_dump_session_t session = nullptr;
  cubeb_audio_dump_stream_t dump_stream = nullptr;
  if constexpr (DUMP_OUTPUT) {
    cubeb_audio_dump_init(&session);
    char buf[256];
    snprintf(buf, 256, "test-%dHz-to-%dhz-%d-block.wav", source_rate,
             target_rate, effective_block_size);
    cubeb_audio_dump_stream_init(session, &dump_stream, out_params, buf);
    cubeb_audio_dump_start(session);
  }
  cubeb_resampler * resampler = cubeb_resampler_create(
      nullptr, nullptr, &out_params, source_rate, data_cb, &state,
      CUBEB_RESAMPLER_QUALITY_DEFAULT, CUBEB_RESAMPLER_RECLOCK_NONE);
  ASSERT_NE(resampler, nullptr);

  std::vector<float> data(effective_block_size * out_params.channels);
  int i = ITERATION_COUNT;
  // For now this only tests the output side (out_... measurements).
  //  We could expect the resampler to be symmetrical, but we could
  //  test both sides at once.
  // - ..._in is the input buffer of the resampler, containing
  // unresampled  frames
  // - ..._out is the output buffer, containing resampled frames.
  monotonic_state in_in_max("in_in", source_rate, target_rate,
                            effective_block_size);
  monotonic_state in_out_max("in_out", source_rate, target_rate,
                             effective_block_size);
  monotonic_state out_in_max("out_in", source_rate, target_rate,
                             effective_block_size);
  monotonic_state out_out_max("out_out", source_rate, target_rate,
                              effective_block_size);

  std::vector<float> resampled;
  resampled.reserve(ITERATION_COUNT * effective_block_size *
                    out_params.channels);
  while (i--) {
    int64_t got = cubeb_resampler_fill(resampler, nullptr, nullptr, data.data(),
                                       effective_block_size);
    ASSERT_EQ(got, effective_block_size);
    cubeb_resampler_stats stats = cubeb_resampler_stats_get(resampler);

    resampled.insert(resampled.end(), data.begin(), data.end());

    in_in_max.set_new_value(stats.input_input_buffer_size);
    in_out_max.set_new_value(stats.input_output_buffer_size);
    out_in_max.set_new_value(stats.output_input_buffer_size);
    out_out_max.set_new_value(stats.output_output_buffer_size);
  }

  cubeb_resampler_destroy(resampler);

  // Example of an error, off by one every block or so, resulting in a
  // silent sample. This is enough to make all the tests fail.
  //
  // for (uint32_t i = 0; i < resampled.size(); i++) {
  //   if (!(i % (effective_block_size))) {
  //     resampled[i] = 0.0;
  //   }
  // }

  // This roughly finds the start of the sine wave and strips it from
  // data.
  size_t skipped = 0;
  skipped = find_sine_start(resampled, input_freq, target_rate);

  resampled.erase(resampled.begin(), resampled.begin() + skipped);

  if constexpr (DUMP_OUTPUT) {
    cubeb_audio_dump_write(dump_stream, resampled.data(), resampled.size());
  }

  float amplitude = 0;
  float phase = 0;

  // Fit our resampled sine wave, get an MSE value
  double sse = fit_sine(resampled, target_rate, input_freq, amplitude, phase);
  double mse = sse / resampled.size();

  // Code to print JSON to plot externally
  // printf("\t[%d,%d,%d,%.10e,%lf,%lf],\n", source_rate, target_rate,
  //        effective_block_size, mse, amplitude, phase);

  // Value found after running the tests on Linux x64
  ASSERT_LT(mse, 3.22e-07);

  if constexpr (DUMP_OUTPUT) {
    cubeb_audio_dump_stop(session);
    cubeb_audio_dump_stream_shutdown(session, dump_stream);
    cubeb_audio_dump_shutdown(session);
  }
}

// This tests checks three things:
// - Whenever resampling from a source rate to a target rate with a certain
//  block size, the correct number of frames is provided back from the
//  resampler, to the backend.
// - While resampling, internal buffers are kept under control and aren't
// growing unbounded.
// - The output signal is a 1khz sine (as is the input)
TEST(cubeb, resampler_typical_uses)
{
  cubeb * ctx;
  common_init(&ctx, "Cubeb resampler test");

  size_t concurrency = std::max(1u, std::thread::hardware_concurrency());
  std::condition_variable cv;
  std::mutex mutex;
  size_t task_count = 0;
  ThreadPool pool(concurrency);

  for (int source_rate : rates) {
    for (int target_rate : rates) {
      for (int block_size : block_sizes) {
        {
          std::unique_lock<std::mutex> lock(mutex);
          ++task_count;
        }
        pool.enqueue([&, source_rate, target_rate, block_size] {
          run_test(source_rate, target_rate, block_size);
          {
            std::unique_lock<std::mutex> lock(mutex);
            --task_count;
          }
          cv.notify_one();
        });
      }
    }
  }

  std::unique_lock<std::mutex> lock(mutex);
  cv.wait(lock, [&] { return task_count == 0; });
  cubeb_destroy(ctx);
}
#undef NOMINMAX
#undef DUMP_ARRAYS