- Fixed Crossfader to adjust for time spent actually adjusting the

output's volume

- Reduced lock contention in WasapiOut and DirectSoundOut by caching
  AddRef()'d copies of important interfaces in the ::Play() methods.

src/contrib/directsoundout/DirectSoundOut.cpp

@@ -187,6 +187,8 @@ void DirectSoundOut::Stop() {
 }
 
 bool DirectSoundOut::Play(IBuffer *buffer, IBufferProvider *provider) {
+    IDirectSoundBuffer8 *outputBuffer = nullptr;
+
     {
         Lock lock(this->stateMutex);
 
@@ -199,64 +201,70 @@ bool DirectSoundOut::Play(IBuffer *buffer, IBufferProvider *provider) {
             return false;
         }
 
-        unsigned char *dst1 = nullptr, *dst2 = nullptr;
-        DWORD size1 = 0, size2 = 0;
-        DWORD availableBytes = 0;
-        DWORD bufferBytes = buffer->Bytes();
-
-        do {
-            availableBytes = getAvailableBytes(
-                this->secondaryBuffer,
-                this->writeOffset,
-                this->bufferSize);
-
-            if (bufferBytes > availableBytes) {
-                int samples = (bufferBytes - availableBytes) / sizeof(float) / channels;
-                int sleepMs = ((long long)(samples * 1000) / rate) + 1;
-                Sleep(sleepMs);
-            }
-        } while (this->state == StatePlaying && availableBytes < bufferBytes);
-
-        if (this->state != StatePlaying) {
-            return false;
-        }
-
-        HRESULT result =
-            this->secondaryBuffer->Lock(
-                writeOffset,
-                bufferBytes,
-                (void **) &dst1, &size1,
-                (void **) &dst2, &size2,
-                0);
-
-        if (result == DSERR_BUFFERLOST) {
-            this->secondaryBuffer->Restore();
-
-            result = this->secondaryBuffer->Lock(
-                writeOffset,
-                bufferBytes,
-                (void **) &dst1, &size1,
-                (void **) &dst2, &size2,
-                0);
-        }
-
-        if (result != DS_OK) {
-            return false;
-        }
-
-        char* bufferPointer = (char *) buffer->BufferPointer();
-
-        memcpy(dst1, bufferPointer, size1);
-        if (size2 > 0) {
-            memcpy(dst2, bufferPointer + size1, size2);
-        }
-
-        writeOffset += bufferBytes;
-        writeOffset %= this->bufferSize;
-
-        this->secondaryBuffer->Unlock((void *)dst1, size1, (void *)dst2, size2);
+        /* reduce lock contention: cache a reference to the buffer here,
+        just in case someone comes along and release it */
+        outputBuffer = this->secondaryBuffer;
+        outputBuffer->AddRef();
     }
 
+    unsigned char *dst1 = nullptr, *dst2 = nullptr;
+    DWORD size1 = 0, size2 = 0;
+    DWORD availableBytes = 0;
+    DWORD bufferBytes = buffer->Bytes();
+
+    do {
+        availableBytes = getAvailableBytes(
+            outputBuffer,
+            this->writeOffset,
+            this->bufferSize);
+
+        if (bufferBytes > availableBytes) {
+            int samples = (bufferBytes - availableBytes) / sizeof(float) / channels;
+            int sleepMs = ((long long)(samples * 1000) / rate) + 1;
+            Sleep(sleepMs);
+        }
+    } while (this->state == StatePlaying && availableBytes < bufferBytes);
+
+    if (this->state != StatePlaying) {
+        return false;
+    }
+
+    HRESULT result =
+        outputBuffer->Lock(
+            writeOffset,
+            bufferBytes,
+            (void **) &dst1, &size1,
+            (void **) &dst2, &size2,
+            0);
+
+    if (result == DSERR_BUFFERLOST) {
+        outputBuffer->Restore();
+
+        result = this->secondaryBuffer->Lock(
+            writeOffset,
+            bufferBytes,
+            (void **) &dst1, &size1,
+            (void **) &dst2, &size2,
+            0);
+    }
+
+    if (result != DS_OK) {
+        return false;
+    }
+
+    char* bufferPointer = (char *) buffer->BufferPointer();
+
+    memcpy(dst1, bufferPointer, size1);
+    if (size2 > 0) {
+        memcpy(dst2, bufferPointer + size1, size2);
+    }
+
+    writeOffset += bufferBytes;
+    writeOffset %= this->bufferSize;
+
+    outputBuffer->Unlock((void *)dst1, size1, (void *)dst2, size2);
+    outputBuffer->Release();
+
     provider->OnBufferProcessed(buffer);
 
     return true;

src/contrib/wasapiout/WasapiOut.cpp

@@ -149,6 +149,12 @@ void WasapiOut::Drain() {
 }
 
 bool WasapiOut::Play(IBuffer *buffer, IBufferProvider *provider) {
+    IAudioRenderClient *renderClient = nullptr;
+    IAudioClient *audioClient = nullptr;
+
+    /* reduce lock contention by snagging the references to the
+    COM interfaces we care about and calling AddRef(), then operating
+    on the local copies. */
     {
         Lock lock(this->stateMutex);
 
@@ -161,27 +167,31 @@ bool WasapiOut::Play(IBuffer *buffer, IBufferProvider *provider) {
             return false;
         }
 
-        UINT32 availableFrames = 0;
-        UINT32 frameOffset = 0;
-        UINT32 samples = (UINT32) buffer->Samples();
-        UINT32 framesToWrite = samples / (UINT32) buffer->Channels();
-        int channels = buffer->Channels();
+        renderClient = this->renderClient;
+        renderClient->AddRef();
 
-        do {
-            this->audioClient->GetCurrentPadding(&frameOffset);
-            availableFrames = (this->outputBufferFrames - frameOffset);
+        audioClient = this->audioClient;
+        audioClient->AddRef();
+    }
 
-            if (availableFrames < framesToWrite) {
-                UINT32 delta = framesToWrite - availableFrames;
-                REFERENCE_TIME sleepTime = (delta * 1000 * 1000 * 10) / buffer->SampleRate();
-                std::this_thread::sleep_for(std::chrono::microseconds(sleepTime));
-            }
-        } while (this->state == StatePlaying && availableFrames < framesToWrite);
+    UINT32 availableFrames = 0;
+    UINT32 frameOffset = 0;
+    UINT32 samples = (UINT32) buffer->Samples();
+    UINT32 framesToWrite = samples / (UINT32) buffer->Channels();
+    int channels = buffer->Channels();
 
-        if (state != StatePlaying) {
-            return false;
+    do {
+        audioClient->GetCurrentPadding(&frameOffset);
+        availableFrames = (this->outputBufferFrames - frameOffset);
+
+        if (availableFrames < framesToWrite) {
+            UINT32 delta = framesToWrite - availableFrames;
+            REFERENCE_TIME sleepTime = (delta * 1000 * 1000 * 10) / buffer->SampleRate();
+            std::this_thread::sleep_for(std::chrono::microseconds(sleepTime));
         }
+    } while (this->state == StatePlaying && availableFrames < framesToWrite);
 
+    if (state == StatePlaying) {
         if (availableFrames >= framesToWrite) {
             BYTE *data = 0;
             HRESULT result = this->renderClient->GetBuffer(framesToWrite, &data);
@@ -191,10 +201,13 @@ bool WasapiOut::Play(IBuffer *buffer, IBufferProvider *provider) {
             }
 
             memcpy(data, buffer->BufferPointer(), sizeof(float) * samples);
-            this->renderClient->ReleaseBuffer(framesToWrite, 0);
+            renderClient->ReleaseBuffer(framesToWrite, 0);
         }
     }
 
+    renderClient->Release();
+    audioClient->Release();
+
     provider->OnBufferProcessed(buffer);
 
     return true;

src/core/audio/Crossfader.cpp

@@ -38,10 +38,12 @@
 #include <core/runtime/Message.h>
 
 #include <algorithm>
+#include <chrono>
 
 using namespace musik::core::audio;
 using namespace musik::core::sdk;
 using namespace musik::core::runtime;
+using namespace std::chrono;
 
 #define TICKS_PER_SECOND 30
 #define TICK_TIME_MILLIS (1000 / TICKS_PER_SECOND)
@@ -51,6 +53,10 @@ using namespace musik::core::runtime;
     this->messageQueue.Post(Message::Create( \
         this, MESSAGE_TICK, 0, 0), TICK_TIME_MILLIS)
 
+#define ENQUEUE_ADJUSTED_TICK(delay) \
+    this->messageQueue.Post(Message::Create( \
+        this, MESSAGE_TICK, 0, 0), delay > 0 ? delay : 0)
+
 #define LOCK(x) \
     std::unique_lock<std::recursive_mutex> lock(x);
 
@@ -214,6 +220,8 @@ void Crossfader::ProcessMessage(IMessage &message) {
         case MESSAGE_TICK: {
             bool emptied = false;
 
+            auto start = system_clock::now().time_since_epoch();
+
             {
                 LOCK(this->contextListLock);
 
@@ -278,12 +286,7 @@ void Crossfader::ProcessMessage(IMessage &message) {
                     }
                 }
 
-                if (this->contextList.size()) {
-                    ENQUEUE_TICK();
-                }
-                else {
-                    emptied = true;
-                }
+                emptied = (this->contextList.size() == 0);
             } /* end critical section */
 
             /* notify outside of the critical section! */
@@ -291,6 +294,11 @@ void Crossfader::ProcessMessage(IMessage &message) {
                 this->Emptied();
                 this->drainCondition.notify_all();
             }
+            else {
+                auto end = system_clock::now().time_since_epoch();
+                int64 duration = duration_cast<milliseconds>(end - start).count();
+                ENQUEUE_ADJUSTED_TICK(TICK_TIME_MILLIS - duration);
+            }
         }
         break;
     }