Implement oversampled peak detection.

diff --git a/mixer.cpp b/mixer.cpp
index 6bf81192fdfd99421398dedc189e23be75344a5d..bc349bef275304a57533281667cc5a892b80feee 100644 (file)
--- a/mixer.cpp
+++ b/mixer.cpp
@@ -153,6 +153,10 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
        r128.integr_start();
 
        locut.init(FILTER_HPF, 2);
+
+       // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
+       // and there's a limit to how important the peak meter is.
+       peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
 }
 
 Mixer::~Mixer()
@@ -182,10 +186,10 @@ int unwrap_timecode(uint16_t current_wrapped, int last)
        }
 }
 
-float find_peak(const vector<float> &samples)
+float find_peak(const float *samples, size_t num_samples)
 {
        float m = fabs(samples[0]);
-       for (size_t i = 1; i < samples.size(); ++i) {
+       for (size_t i = 1; i < num_samples; ++i) {
                m = std::max(m, fabs(samples[i]));
        }
        return m;
@@ -600,8 +604,21 @@ void Mixer::process_audio_one_frame()
 
 //     printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 
-       // Find peak and R128 levels.
-       peak = max<float>(peak, find_peak(samples_out));
+       // Upsample 4x to find interpolated peak.
+       peak_resampler.inp_data = samples_out.data();
+       peak_resampler.inp_count = samples_out.size() / 2;
+
+       vector<float> interpolated_samples_out;
+       interpolated_samples_out.resize(samples_out.size());
+       while (peak_resampler.inp_count > 0) {  // About four iterations.
+               peak_resampler.out_data = &interpolated_samples_out[0];
+               peak_resampler.out_count = interpolated_samples_out.size() / 2;
+               peak_resampler.process();
+               size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
+               peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
+       }
+
+       // Find R128 levels.
        vector<float> left, right;
        deinterleave_samples(samples_out, &left, &right);
        float *ptrs[] = { left.data(), right.data() };
@@ -698,6 +715,7 @@ void Mixer::channel_clicked(int preview_num)
 
 void Mixer::reset_meters()
 {
+       peak_resampler.reset();
        peak = 0.0f;
        r128.reset();
        r128.integr_start();

diff --git a/mixer.h b/mixer.h
index 3a19d3aaef946f6257a27a9ae060c0cf057306ba..0fd07aea05598f82ee0c1a3e2a46aeda821741f0 100644 (file)
--- a/mixer.h
+++ b/mixer.h
@@ -10,6 +10,7 @@
 
 #include <movit/effect_chain.h>
 #include <movit/flat_input.h>
+#include <zita-resampler/resampler.h>
 #include <atomic>
 #include <condition_variable>
 #include <cstddef>
@@ -213,7 +214,7 @@ private:
        audio_level_callback_t audio_level_callback = nullptr;
        Ebu_r128_proc r128;
 
-       // TODO: Implement oversampled peak detection.
+       Resampler peak_resampler;
        std::atomic<float> peak{0.0f};
 
        StereoFilter locut;  // Default cutoff 150 Hz, 24 dB/oct.