cubeb_sink: Perform audio stretching

src/audio_core/cubeb_sink.cpp

@@ -6,8 +6,10 @@
 #include <cstring>
 #include "audio_core/cubeb_sink.h"
 #include "audio_core/stream.h"
+#include "audio_core/time_stretch.h"
 #include "common/logging/log.h"
 #include "common/ring_buffer.h"
+#include "core/settings.h"
 
 namespace AudioCore {
 
@@ -15,14 +17,8 @@ class CubebSinkStream final : public SinkStream {
 public:
     CubebSinkStream(cubeb* ctx, u32 sample_rate, u32 num_channels_, cubeb_devid output_device,
                     const std::string& name)
-        : ctx{ctx}, num_channels{num_channels_} {
-
-        if (num_channels == 6) {
-            // 6-channel audio does not seem to work with cubeb + SDL, so we downsample this to 2
-            // channel for now
-            is_6_channel = true;
-            num_channels = 2;
-        }
+        : ctx{ctx}, is_6_channel{num_channels_ == 6}, num_channels{std::min(num_channels_, 2u)},
+          time_stretch{sample_rate, num_channels} {
 
         cubeb_stream_params params{};
         params.rate = sample_rate;
@@ -89,10 +85,6 @@ public:
         return num_channels;
     }
 
-    u32 GetNumChannelsInQueue() const {
-        return num_channels == 1 ? 1 : 2;
-    }
-
 private:
     std::vector<std::string> device_list;
 
@@ -103,6 +95,7 @@ private:
 
     Common::RingBuffer<s16, 0x10000> queue;
     std::array<s16, 2> last_frame;
+    TimeStretcher time_stretch;
 
     static long DataCallback(cubeb_stream* stream, void* user_data, const void* input_buffer,
                              void* output_buffer, long num_frames);
@@ -161,9 +154,19 @@ long CubebSinkStream::DataCallback(cubeb_stream* stream, void* user_data, const
         return {};
     }
 
-    const size_t num_channels = impl->GetNumChannelsInQueue();
-    const size_t max_samples_to_write = num_channels * num_frames;
-    const size_t samples_written = impl->queue.Pop(buffer, max_samples_to_write);
+    const size_t num_channels = impl->GetNumChannels();
+    const size_t samples_to_write = num_channels * num_frames;
+    size_t samples_written;
+
+    if (Settings::values.enable_audio_stretching) {
+        const std::vector<s16> in{impl->queue.Pop()};
+        const size_t num_in{in.size() / num_channels};
+        s16* const out{reinterpret_cast<s16*>(buffer)};
+        const size_t out_frames = impl->time_stretch.Process(in.data(), num_in, out, num_frames);
+        samples_written = out_frames * num_channels;
+    } else {
+        samples_written = impl->queue.Pop(buffer, samples_to_write);
+    }
 
     if (samples_written >= num_channels) {
         std::memcpy(&impl->last_frame[0], buffer + (samples_written - num_channels) * sizeof(s16),
@@ -171,7 +174,7 @@ long CubebSinkStream::DataCallback(cubeb_stream* stream, void* user_data, const
     }
 
     // Fill the rest of the frames with last_frame
-    for (size_t i = samples_written; i < max_samples_to_write; i += num_channels) {
+    for (size_t i = samples_written; i < samples_to_write; i += num_channels) {
         std::memcpy(buffer + i * sizeof(s16), &impl->last_frame[0], num_channels * sizeof(s16));
     }
 

src/audio_core/time_stretch.cpp

@@ -28,8 +28,8 @@ size_t TimeStretcher::Process(const s16* in, size_t num_in, s16* out, size_t num
     // 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 = 0.3; // seconds
-    const double max_backlog = m_sample_rate * max_latency / 1000.0 / m_stretch_ratio;
+    const double max_latency = 1.0; // seconds
+    const double max_backlog = m_sample_rate * max_latency;
     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
@@ -49,13 +49,13 @@ size_t TimeStretcher::Process(const s16* in, size_t num_in, s16* out, size_t num
     const double lpf_gain = 1.0 - std::exp(-time_delta / lpf_time_scale);
     m_stretch_ratio += lpf_gain * (current_ratio - m_stretch_ratio);
 
-    // Place a lower limit of 10% speed.  When a game boots up, there will be
+    // Place a lower limit of 5% 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_stretch_ratio = std::max(m_stretch_ratio, 0.05);
     m_sound_touch.setTempo(m_stretch_ratio);
 
-    LOG_DEBUG(Audio, "Audio Stretching: samples:{}/{} ratio:{} backlog:{} gain: {}", num_in, num_out,
-              m_stretch_ratio, backlog_fullness, lpf_gain);
+    LOG_DEBUG(Audio, "{:5}/{:5} ratio:{:0.6f} backlog:{:0.6f}", num_in, num_out, m_stretch_ratio,
+              backlog_fullness);
 
     m_sound_touch.putSamples(in, num_in);
     return m_sound_touch.receiveSamples(out, num_out);

src/audio_core/time_stretch.h

@@ -27,7 +27,6 @@ public:
 private:
     u32 m_sample_rate;
     u32 m_channel_count;
-    std::array<s16, 2> m_last_stretched_sample = {};
     soundtouch::SoundTouch m_sound_touch;
     double m_stretch_ratio = 1.0;
 };