Use a wait-free algorithm for audio queueing to minimize audio hiccups. This closes #8

This commit is contained in:
Cameron Gutman 2016-02-18 12:27:12 -05:00
parent cf5b2ccbb4
commit 3ce2d68250

View File

@ -1,81 +1,42 @@
#include "moonlight.hpp" #include "moonlight.hpp"
#define MAX_CHANNEL_COUNT 6 #define MAX_CHANNEL_COUNT 2
#define FRAME_SIZE 240 #define FRAME_SIZE 240
typedef struct decoded_sample_entry { #define CIRCULAR_BUFFER_SIZE 32
struct decoded_sample_entry *next;
int sampleLength;
short sampleBuffer[1];
} decoded_sample_entry_t;
#define MAX_QUEUE_LENGTH 14 // This code uses volatiles for synchronization between the producer and consumer side. This is
#define QUEUE_PRUNING_LENGTH 7 // only safe because this code executes under very specific conditions, namely that the framework
// ensures AudioPlayerSampleCallback and AudDecDecodeAndPlaySample are each only active on one thread
// at a time.
static int s_OpusChannelCount; static short s_CircularBuffer[CIRCULAR_BUFFER_SIZE][FRAME_SIZE * MAX_CHANNEL_COUNT];
static decoded_sample_entry_t* s_SampleQueueHead; static int s_ReadIndex = 0;
static decoded_sample_entry_t* s_SampleQueueTail; static int s_WriteIndex = 0;
static int s_SampleQueueLength;
static pthread_mutex_t s_SampleQueueLock;
static void ReapSampleQueue() {
decoded_sample_entry_t *entry;
while (s_SampleQueueHead) {
entry = s_SampleQueueHead->next;
free(s_SampleQueueHead);
s_SampleQueueHead = entry;
}
s_SampleQueueTail = NULL;
s_SampleQueueLength = 0;
}
static void AudioPlayerSampleCallback(void* samples, uint32_t buffer_size, void* data) { static void AudioPlayerSampleCallback(void* samples, uint32_t buffer_size, void* data) {
unsigned char* buffer = (unsigned char *)samples; // It should only ask us for complete buffers
int offset = 0; assert(buffer_size == FRAME_SIZE * MAX_CHANNEL_COUNT * sizeof(short));
pthread_mutex_lock(&s_SampleQueueLock); // If the indexes aren't equal, we have a sample
if (s_WriteIndex != s_ReadIndex) {
memcpy(samples, s_CircularBuffer[s_ReadIndex], buffer_size);
while (s_SampleQueueHead && s_SampleQueueHead->sampleLength <= buffer_size - offset) { // Use a full memory barrier to ensure the circular buffer is read before incrementing the index
decoded_sample_entry_t* lastEnt; __sync_synchronize();
memcpy(&buffer[offset], s_SampleQueueHead->sampleBuffer, s_SampleQueueHead->sampleLength); // This can race with the reader in the AudDecDecodeAndPlaySample function. This is
offset += s_SampleQueueHead->sampleLength; // not a problem because at worst, it just won't see that we've consumed this sample yet.
s_ReadIndex = (s_ReadIndex + 1) % CIRCULAR_BUFFER_SIZE;
lastEnt = s_SampleQueueHead;
s_SampleQueueHead = s_SampleQueueHead->next;
free(lastEnt);
s_SampleQueueLength--;
// Remove another sample if we're in pruning mode
if (s_SampleQueueLength > QUEUE_PRUNING_LENGTH) {
lastEnt = s_SampleQueueHead;
s_SampleQueueHead = s_SampleQueueHead->next;
free(lastEnt);
s_SampleQueueLength--;
} }
} else {
memset(samples, 0, buffer_size);
if (!s_SampleQueueHead) {
s_SampleQueueTail = NULL;
}
pthread_mutex_unlock(&s_SampleQueueLock);
// Zero the remaining portion of the sample buffer to reduce noise when underflowing
if (buffer_size != offset) {
memset(&buffer[offset], 0, buffer_size - offset);
} }
} }
void MoonlightInstance::AudDecInit(int audioConfiguration, POPUS_MULTISTREAM_CONFIGURATION opusConfig) { void MoonlightInstance::AudDecInit(int audioConfiguration, POPUS_MULTISTREAM_CONFIGURATION opusConfig) {
int rc; int rc;
pthread_mutex_init(&s_SampleQueueLock, NULL);
s_OpusChannelCount = opusConfig->channelCount;
g_Instance->m_OpusDecoder = opus_multistream_decoder_create(opusConfig->sampleRate, g_Instance->m_OpusDecoder = opus_multistream_decoder_create(opusConfig->sampleRate,
opusConfig->channelCount, opusConfig->channelCount,
opusConfig->streams, opusConfig->streams,
@ -83,56 +44,41 @@ void MoonlightInstance::AudDecInit(int audioConfiguration, POPUS_MULTISTREAM_CON
opusConfig->mapping, opusConfig->mapping,
&rc); &rc);
pp::AudioConfig audioConfig = pp::AudioConfig(g_Instance, PP_AUDIOSAMPLERATE_48000, FRAME_SIZE * 3); g_Instance->m_AudioPlayer = pp::Audio(g_Instance, pp::AudioConfig(g_Instance, PP_AUDIOSAMPLERATE_48000, FRAME_SIZE),
AudioPlayerSampleCallback, NULL);
g_Instance->m_AudioPlayer = pp::Audio(g_Instance, audioConfig, AudioPlayerSampleCallback, NULL);
// Start playback now // Start playback now
g_Instance->m_AudioPlayer.StartPlayback(); g_Instance->m_AudioPlayer.StartPlayback();
} }
void MoonlightInstance::AudDecCleanup(void) { void MoonlightInstance::AudDecCleanup(void) {
pthread_mutex_destroy(&s_SampleQueueLock); // Stop playback
g_Instance->m_AudioPlayer.StopPlayback();
if (g_Instance->m_OpusDecoder) { if (g_Instance->m_OpusDecoder) {
opus_multistream_decoder_destroy(g_Instance->m_OpusDecoder); opus_multistream_decoder_destroy(g_Instance->m_OpusDecoder);
} }
ReapSampleQueue();
} }
void MoonlightInstance::AudDecDecodeAndPlaySample(char* sampleData, int sampleLength) { void MoonlightInstance::AudDecDecodeAndPlaySample(char* sampleData, int sampleLength) {
decoded_sample_entry_t* entry = (decoded_sample_entry_t*)malloc(sizeof(decoded_sample_entry_t) + int decodeLen;
(s_OpusChannelCount * FRAME_SIZE * sizeof(short)));
if (entry) { // Check if there is space for this sample in the buffer. Again, this can race
int decodeLen = opus_multistream_decode(g_Instance->m_OpusDecoder, (unsigned char *)sampleData, sampleLength, // but in the worst case, we'll not see the sample callback having consumed a sample.
entry->sampleBuffer, FRAME_SIZE, 0); if (((s_WriteIndex + 1) % CIRCULAR_BUFFER_SIZE) == s_ReadIndex) {
return;
}
decodeLen = opus_multistream_decode(g_Instance->m_OpusDecoder, (unsigned char *)sampleData, sampleLength,
s_CircularBuffer[s_WriteIndex], FRAME_SIZE, 0);
if (decodeLen > 0) { if (decodeLen > 0) {
entry->sampleLength = decodeLen * s_OpusChannelCount * sizeof(short); // Use a full memory barrier to ensure the circular buffer is written before incrementing the index
entry->next = NULL; __sync_synchronize();
pthread_mutex_lock(&s_SampleQueueLock); // This can race with the reader in the sample callback, however this is a benign
// race since we'll either read the original value of s_WriteIndex (which is safe,
if (s_SampleQueueLength == MAX_QUEUE_LENGTH) { // we just won't consider this sample) or the new value of s_WriteIndex
printf("Reaped sample queue\n"); s_WriteIndex = (s_WriteIndex + 1) % CIRCULAR_BUFFER_SIZE;
ReapSampleQueue();
}
if (!s_SampleQueueTail) {
s_SampleQueueHead = s_SampleQueueTail = entry;
}
else {
s_SampleQueueTail->next = entry;
s_SampleQueueTail = entry;
}
s_SampleQueueLength++;
pthread_mutex_unlock(&s_SampleQueueLock);
}
else {
free(entry);
}
} }
} }