Delay estimation result looks wrong

[RangHo]

Although the main.cc source code follows the logic of WebRTC's delay estimation unit test, the number of estimated delay still looks very off. Below is the source which main.cc is based on.

// Verifies that the delay estimator produces correct delay for artificially
// delayed signals.
TEST(EchoPathDelayEstimator, DelayEstimation) {
  constexpr size_t kNumRenderChannels = 1;
  constexpr size_t kNumCaptureChannels = 1;
  constexpr int kSampleRateHz = 48000;
  constexpr size_t kNumBands = NumBandsForRate(kSampleRateHz);

  Random random_generator(42U);
  std::vector<std::vector<std::vector<float>>> render(
      kNumBands, std::vector<std::vector<float>>(
                     kNumRenderChannels, std::vector<float>(kBlockSize)));
  std::vector<std::vector<float>> capture(kNumCaptureChannels,
                                          std::vector<float>(kBlockSize));
  ApmDataDumper data_dumper(0);
  constexpr size_t kDownSamplingFactors[] = {2, 4, 8};
  for (auto down_sampling_factor : kDownSamplingFactors) {
    EchoCanceller3Config config;
    config.delay.down_sampling_factor = down_sampling_factor;
    config.delay.num_filters = 10;
    for (size_t delay_samples : {30, 64, 150, 200, 800, 4000}) {
      SCOPED_TRACE(ProduceDebugText(delay_samples, down_sampling_factor));
      std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
          RenderDelayBuffer::Create(config, kSampleRateHz, kNumRenderChannels));
      DelayBuffer<float> signal_delay_buffer(delay_samples);
      EchoPathDelayEstimator estimator(&data_dumper, config,
                                       kNumCaptureChannels);

      absl::optional<DelayEstimate> estimated_delay_samples;
      for (size_t k = 0; k < (500 + (delay_samples) / kBlockSize); ++k) {
        RandomizeSampleVector(&random_generator, render[0][0]);
        signal_delay_buffer.Delay(render[0][0], capture[0]);
        render_delay_buffer->Insert(render);

        if (k == 0) {
          render_delay_buffer->Reset();
        }

        render_delay_buffer->PrepareCaptureProcessing();

        auto estimate = estimator.EstimateDelay(
            render_delay_buffer->GetDownsampledRenderBuffer(), capture);

        if (estimate) {
          estimated_delay_samples = estimate;
        }
      }

      if (estimated_delay_samples) {
        // Allow estimated delay to be off by one sample in the down-sampled
        // domain.
        size_t delay_ds = delay_samples / down_sampling_factor;
        size_t estimated_delay_ds =
            estimated_delay_samples->delay / down_sampling_factor;
        EXPECT_NEAR(delay_ds, estimated_delay_ds, 1);
      } else {
        ADD_FAILURE();
      }
    }
  }
}

Maybe reproducing this unit test helps to pinpoint the issue? Not sure if this unit test even passes in the WebRTC repository in the first place.