AudioCommon: extract AudioStretcher class for time-stretching

This commit is contained in:
Michael Maltese
2017-04-23 19:08:33 -07:00
parent a4508e85e8
commit 87a467fe42
7 changed files with 123 additions and 75 deletions

View File

@ -14,19 +14,10 @@
#include "Common/Swap.h"
#include "Core/ConfigManager.h"
CMixer::CMixer(unsigned int BackendSampleRate) : m_sampleRate(BackendSampleRate)
CMixer::CMixer(unsigned int BackendSampleRate)
: m_sampleRate(BackendSampleRate), m_stretcher(BackendSampleRate)
{
INFO_LOG(AUDIO_INTERFACE, "Mixer is initialized");
m_sound_touch.setChannels(2);
m_sound_touch.setSampleRate(BackendSampleRate);
m_sound_touch.setPitch(1.0);
m_sound_touch.setTempo(1.0);
m_sound_touch.setSetting(SETTING_USE_QUICKSEEK, 0);
m_sound_touch.setSetting(SETTING_SEQUENCE_MS, 62);
m_sound_touch.setSetting(SETTING_SEEKWINDOW_MS, 28);
m_sound_touch.setSetting(SETTING_OVERLAP_MS, 8);
DPL2Reset();
}
@ -146,10 +137,10 @@ unsigned int CMixer::Mix(short* samples, unsigned int num_samples)
if (!m_is_stretching)
{
m_sound_touch.clear();
m_stretcher.Clear();
m_is_stretching = true;
}
StretchAudio(m_stretch_buffer.data(), available_samples, samples, num_samples);
m_stretcher.StretchAudio(m_stretch_buffer.data(), available_samples, samples, num_samples);
}
else
{
@ -181,60 +172,6 @@ unsigned int CMixer::MixSurround(float* samples, unsigned int num_samples)
return available_samples;
}
void CMixer::StretchAudio(const short* in, unsigned int num_in, short* out, unsigned int num_out)
{
const double time_delta = static_cast<double>(num_out) / m_sampleRate; // seconds
// We were given actual_samples number of samples, and num_samples were requested from us.
double current_ratio = static_cast<double>(num_in) / static_cast<double>(num_out);
const double max_latency = SConfig::GetInstance().m_audio_stretch_max_latency;
const double max_backlog = m_sampleRate * max_latency / 1000.0 / m_stretch_ratio;
const double backlog_fullness = m_sound_touch.numSamples() / max_backlog;
if (backlog_fullness > 5.0)
{
// Too many samples in backlog: Don't push anymore on
num_in = 0;
}
// We ideally want the backlog to be about 50% full.
// This gives some headroom both ways to prevent underflow and overflow.
// We tweak current_ratio to encourage this.
constexpr double tweak_time_scale = 0.5; // seconds
current_ratio *= 1.0 + 2.0 * (backlog_fullness - 0.5) * (time_delta / tweak_time_scale);
// This low-pass filter smoothes out variance in the calculated stretch ratio.
// The time-scale determines how responsive this filter is.
constexpr double lpf_time_scale = 1.0; // seconds
const double m_lpf_gain = 1.0 - std::exp(-time_delta / lpf_time_scale);
m_stretch_ratio += m_lpf_gain * (current_ratio - m_stretch_ratio);
// Place a lower limit of 10% speed. When a game boots up, there will be
// many silence samples. These do not need to be timestretched.
m_stretch_ratio = std::max(m_stretch_ratio, 0.1);
m_sound_touch.setTempo(m_stretch_ratio);
DEBUG_LOG(AUDIO, "Audio stretching: samples:%u/%u ratio:%f backlog:%f gain: %f", num_in, num_out,
m_stretch_ratio, backlog_fullness, m_lpf_gain);
m_sound_touch.putSamples(in, num_in);
const size_t samples_received = m_sound_touch.receiveSamples(out, num_out);
if (samples_received != 0)
{
m_last_stretched_sample[0] = out[samples_received * 2 - 2];
m_last_stretched_sample[1] = out[samples_received * 2 - 1];
}
// Preform padding if we've run out of samples.
for (size_t i = samples_received; i < num_out; i++)
{
out[i * 2 + 0] = m_last_stretched_sample[0];
out[i * 2 + 1] = m_last_stretched_sample[1];
}
}
void CMixer::MixerFifo::PushSamples(const short* samples, unsigned int num_samples)
{
// Cache access in non-volatile variable