diff options
Diffstat (limited to 'services/audioflinger/FastMixer.cpp')
-rw-r--r-- | services/audioflinger/FastMixer.cpp | 246 |
1 files changed, 158 insertions, 88 deletions
diff --git a/services/audioflinger/FastMixer.cpp b/services/audioflinger/FastMixer.cpp index 21df1d7..f27ea17 100644 --- a/services/audioflinger/FastMixer.cpp +++ b/services/audioflinger/FastMixer.cpp @@ -25,6 +25,7 @@ #define ATRACE_TAG ATRACE_TAG_AUDIO +#include "Configuration.h" #include <sys/atomics.h> #include <time.h> #include <utils/Log.h> @@ -44,6 +45,8 @@ #define MIN_WARMUP_CYCLES 2 // minimum number of loop cycles to wait for warmup #define MAX_WARMUP_CYCLES 10 // maximum number of loop cycles to wait for warmup +#define FCC_2 2 // fixed channel count assumption + namespace android { // Fast mixer thread @@ -82,7 +85,7 @@ bool FastMixer::threadLoop() struct timespec oldLoad = {0, 0}; // previous value of clock_gettime(CLOCK_THREAD_CPUTIME_ID) bool oldLoadValid = false; // whether oldLoad is valid uint32_t bounds = 0; - bool full = false; // whether we have collected at least kSamplingN samples + bool full = false; // whether we have collected at least mSamplingN samples #ifdef CPU_FREQUENCY_STATISTICS ThreadCpuUsage tcu; // for reading the current CPU clock frequency in kHz #endif @@ -93,6 +96,12 @@ bool FastMixer::threadLoop() uint32_t warmupCycles = 0; // counter of number of loop cycles required to warmup NBAIO_Sink* teeSink = NULL; // if non-NULL, then duplicate write() to this non-blocking sink NBLog::Writer dummyLogWriter, *logWriter = &dummyLogWriter; + uint32_t totalNativeFramesWritten = 0; // copied to dumpState->mFramesWritten + + // next 2 fields are valid only when timestampStatus == NO_ERROR + AudioTimestamp timestamp; + uint32_t nativeFramesWrittenButNotPresented = 0; // the = 0 is to silence the compiler + status_t timestampStatus = INVALID_OPERATION; for (;;) { @@ -142,7 +151,9 @@ bool FastMixer::threadLoop() preIdle = *current; current = &preIdle; oldTsValid = false; +#ifdef FAST_MIXER_STATISTICS oldLoadValid = false; +#endif ignoreNextOverrun = true; } previous = current; @@ -182,9 +193,12 @@ bool FastMixer::threadLoop() warmupCycles = 0; sleepNs = -1; coldGen = current->mColdGen; +#ifdef FAST_MIXER_STATISTICS bounds = 0; full = false; +#endif oldTsValid = !clock_gettime(CLOCK_MONOTONIC, &oldTs); + timestampStatus = INVALID_OPERATION; } else { sleepNs = FAST_HOT_IDLE_NS; } @@ -220,7 +234,7 @@ bool FastMixer::threadLoop() } else { format = outputSink->format(); sampleRate = Format_sampleRate(format); - ALOG_ASSERT(Format_channelCount(format) == 2); + ALOG_ASSERT(Format_channelCount(format) == FCC_2); } dumpState->mSampleRate = sampleRate; } @@ -236,7 +250,7 @@ bool FastMixer::threadLoop() // implementation; it would be better to have normal mixer allocate for us // to avoid blocking here and to prevent possible priority inversion mixer = new AudioMixer(frameCount, sampleRate, FastMixerState::kMaxFastTracks); - mixBuffer = new short[frameCount * 2]; + mixBuffer = new short[frameCount * FCC_2]; periodNs = (frameCount * 1000000000LL) / sampleRate; // 1.00 underrunNs = (frameCount * 1750000000LL) / sampleRate; // 1.75 overrunNs = (frameCount * 500000000LL) / sampleRate; // 0.50 @@ -375,6 +389,31 @@ bool FastMixer::threadLoop() i = __builtin_ctz(currentTrackMask); currentTrackMask &= ~(1 << i); const FastTrack* fastTrack = ¤t->mFastTracks[i]; + + // Refresh the per-track timestamp + if (timestampStatus == NO_ERROR) { + uint32_t trackFramesWrittenButNotPresented; + uint32_t trackSampleRate = fastTrack->mSampleRate; + // There is currently no sample rate conversion for fast tracks currently + if (trackSampleRate != 0 && trackSampleRate != sampleRate) { + trackFramesWrittenButNotPresented = + ((int64_t) nativeFramesWrittenButNotPresented * trackSampleRate) / + sampleRate; + } else { + trackFramesWrittenButNotPresented = nativeFramesWrittenButNotPresented; + } + uint32_t trackFramesWritten = fastTrack->mBufferProvider->framesReleased(); + // Can't provide an AudioTimestamp before first frame presented, + // or during the brief 32-bit wraparound window + if (trackFramesWritten >= trackFramesWrittenButNotPresented) { + AudioTimestamp perTrackTimestamp; + perTrackTimestamp.mPosition = + trackFramesWritten - trackFramesWrittenButNotPresented; + perTrackTimestamp.mTime = timestamp.mTime; + fastTrack->mBufferProvider->onTimestamp(perTrackTimestamp); + } + } + int name = fastTrackNames[i]; ALOG_ASSERT(name >= 0); if (fastTrack->mVolumeProvider != NULL) { @@ -433,7 +472,7 @@ bool FastMixer::threadLoop() //bool didFullWrite = false; // dumpsys could display a count of partial writes if ((command & FastMixerState::WRITE) && (outputSink != NULL) && (mixBuffer != NULL)) { if (mixBufferState == UNDEFINED) { - memset(mixBuffer, 0, frameCount * 2 * sizeof(short)); + memset(mixBuffer, 0, frameCount * FCC_2 * sizeof(short)); mixBufferState = ZEROED; } if (teeSink != NULL) { @@ -448,7 +487,8 @@ bool FastMixer::threadLoop() dumpState->mWriteSequence++; if (framesWritten >= 0) { ALOG_ASSERT((size_t) framesWritten <= frameCount); - dumpState->mFramesWritten += framesWritten; + totalNativeFramesWritten += framesWritten; + dumpState->mFramesWritten = totalNativeFramesWritten; //if ((size_t) framesWritten == frameCount) { // didFullWrite = true; //} @@ -457,6 +497,18 @@ bool FastMixer::threadLoop() } attemptedWrite = true; // FIXME count # of writes blocked excessively, CPU usage, etc. for dump + + timestampStatus = outputSink->getTimestamp(timestamp); + if (timestampStatus == NO_ERROR) { + uint32_t totalNativeFramesPresented = timestamp.mPosition; + if (totalNativeFramesPresented <= totalNativeFramesWritten) { + nativeFramesWrittenButNotPresented = + totalNativeFramesWritten - totalNativeFramesPresented; + } else { + // HAL reported that more frames were presented than were written + timestampStatus = INVALID_OPERATION; + } + } } // To be exactly periodic, compute the next sleep time based on current time. @@ -498,91 +550,91 @@ bool FastMixer::threadLoop() } } sleepNs = -1; - if (isWarm) { - if (sec > 0 || nsec > underrunNs) { - ATRACE_NAME("underrun"); - // FIXME only log occasionally - ALOGV("underrun: time since last cycle %d.%03ld sec", - (int) sec, nsec / 1000000L); - dumpState->mUnderruns++; - ignoreNextOverrun = true; - } else if (nsec < overrunNs) { - if (ignoreNextOverrun) { - ignoreNextOverrun = false; - } else { + if (isWarm) { + if (sec > 0 || nsec > underrunNs) { + ATRACE_NAME("underrun"); // FIXME only log occasionally - ALOGV("overrun: time since last cycle %d.%03ld sec", + ALOGV("underrun: time since last cycle %d.%03ld sec", (int) sec, nsec / 1000000L); - dumpState->mOverruns++; + dumpState->mUnderruns++; + ignoreNextOverrun = true; + } else if (nsec < overrunNs) { + if (ignoreNextOverrun) { + ignoreNextOverrun = false; + } else { + // FIXME only log occasionally + ALOGV("overrun: time since last cycle %d.%03ld sec", + (int) sec, nsec / 1000000L); + dumpState->mOverruns++; + } + // This forces a minimum cycle time. It: + // - compensates for an audio HAL with jitter due to sample rate conversion + // - works with a variable buffer depth audio HAL that never pulls at a + // rate < than overrunNs per buffer. + // - recovers from overrun immediately after underrun + // It doesn't work with a non-blocking audio HAL. + sleepNs = forceNs - nsec; + } else { + ignoreNextOverrun = false; } - // This forces a minimum cycle time. It: - // - compensates for an audio HAL with jitter due to sample rate conversion - // - works with a variable buffer depth audio HAL that never pulls at a rate - // < than overrunNs per buffer. - // - recovers from overrun immediately after underrun - // It doesn't work with a non-blocking audio HAL. - sleepNs = forceNs - nsec; - } else { - ignoreNextOverrun = false; } - } #ifdef FAST_MIXER_STATISTICS - if (isWarm) { - // advance the FIFO queue bounds - size_t i = bounds & (FastMixerDumpState::kSamplingN - 1); - bounds = (bounds & 0xFFFF0000) | ((bounds + 1) & 0xFFFF); - if (full) { - bounds += 0x10000; - } else if (!(bounds & (FastMixerDumpState::kSamplingN - 1))) { - full = true; - } - // compute the delta value of clock_gettime(CLOCK_MONOTONIC) - uint32_t monotonicNs = nsec; - if (sec > 0 && sec < 4) { - monotonicNs += sec * 1000000000; - } - // compute the raw CPU load = delta value of clock_gettime(CLOCK_THREAD_CPUTIME_ID) - uint32_t loadNs = 0; - struct timespec newLoad; - rc = clock_gettime(CLOCK_THREAD_CPUTIME_ID, &newLoad); - if (rc == 0) { - if (oldLoadValid) { - sec = newLoad.tv_sec - oldLoad.tv_sec; - nsec = newLoad.tv_nsec - oldLoad.tv_nsec; - if (nsec < 0) { - --sec; - nsec += 1000000000; - } - loadNs = nsec; - if (sec > 0 && sec < 4) { - loadNs += sec * 1000000000; + if (isWarm) { + // advance the FIFO queue bounds + size_t i = bounds & (dumpState->mSamplingN - 1); + bounds = (bounds & 0xFFFF0000) | ((bounds + 1) & 0xFFFF); + if (full) { + bounds += 0x10000; + } else if (!(bounds & (dumpState->mSamplingN - 1))) { + full = true; + } + // compute the delta value of clock_gettime(CLOCK_MONOTONIC) + uint32_t monotonicNs = nsec; + if (sec > 0 && sec < 4) { + monotonicNs += sec * 1000000000; + } + // compute raw CPU load = delta value of clock_gettime(CLOCK_THREAD_CPUTIME_ID) + uint32_t loadNs = 0; + struct timespec newLoad; + rc = clock_gettime(CLOCK_THREAD_CPUTIME_ID, &newLoad); + if (rc == 0) { + if (oldLoadValid) { + sec = newLoad.tv_sec - oldLoad.tv_sec; + nsec = newLoad.tv_nsec - oldLoad.tv_nsec; + if (nsec < 0) { + --sec; + nsec += 1000000000; + } + loadNs = nsec; + if (sec > 0 && sec < 4) { + loadNs += sec * 1000000000; + } + } else { + // first time through the loop + oldLoadValid = true; } - } else { - // first time through the loop - oldLoadValid = true; + oldLoad = newLoad; } - oldLoad = newLoad; - } #ifdef CPU_FREQUENCY_STATISTICS - // get the absolute value of CPU clock frequency in kHz - int cpuNum = sched_getcpu(); - uint32_t kHz = tcu.getCpukHz(cpuNum); - kHz = (kHz << 4) | (cpuNum & 0xF); + // get the absolute value of CPU clock frequency in kHz + int cpuNum = sched_getcpu(); + uint32_t kHz = tcu.getCpukHz(cpuNum); + kHz = (kHz << 4) | (cpuNum & 0xF); #endif - // save values in FIFO queues for dumpsys - // these stores #1, #2, #3 are not atomic with respect to each other, - // or with respect to store #4 below - dumpState->mMonotonicNs[i] = monotonicNs; - dumpState->mLoadNs[i] = loadNs; + // save values in FIFO queues for dumpsys + // these stores #1, #2, #3 are not atomic with respect to each other, + // or with respect to store #4 below + dumpState->mMonotonicNs[i] = monotonicNs; + dumpState->mLoadNs[i] = loadNs; #ifdef CPU_FREQUENCY_STATISTICS - dumpState->mCpukHz[i] = kHz; + dumpState->mCpukHz[i] = kHz; #endif - // this store #4 is not atomic with respect to stores #1, #2, #3 above, but - // the newest open and oldest closed halves are atomic with respect to each other - dumpState->mBounds = bounds; - ATRACE_INT("cycle_ms", monotonicNs / 1000000); - ATRACE_INT("load_us", loadNs / 1000); - } + // this store #4 is not atomic with respect to stores #1, #2, #3 above, but + // the newest open & oldest closed halves are atomic with respect to each other + dumpState->mBounds = bounds; + ATRACE_INT("cycle_ms", monotonicNs / 1000000); + ATRACE_INT("load_us", loadNs / 1000); + } #endif } else { // first time through the loop @@ -603,25 +655,43 @@ bool FastMixer::threadLoop() // never return 'true'; Thread::_threadLoop() locks mutex which can result in priority inversion } -FastMixerDumpState::FastMixerDumpState() : +FastMixerDumpState::FastMixerDumpState( +#ifdef FAST_MIXER_STATISTICS + uint32_t samplingN +#endif + ) : mCommand(FastMixerState::INITIAL), mWriteSequence(0), mFramesWritten(0), mNumTracks(0), mWriteErrors(0), mUnderruns(0), mOverruns(0), mSampleRate(0), mFrameCount(0), /* mMeasuredWarmupTs({0, 0}), */ mWarmupCycles(0), mTrackMask(0) #ifdef FAST_MIXER_STATISTICS - , mBounds(0) + , mSamplingN(0), mBounds(0) #endif { mMeasuredWarmupTs.tv_sec = 0; mMeasuredWarmupTs.tv_nsec = 0; +#ifdef FAST_MIXER_STATISTICS + increaseSamplingN(samplingN); +#endif +} + +#ifdef FAST_MIXER_STATISTICS +void FastMixerDumpState::increaseSamplingN(uint32_t samplingN) +{ + if (samplingN <= mSamplingN || samplingN > kSamplingN || roundup(samplingN) != samplingN) { + return; + } + uint32_t additional = samplingN - mSamplingN; // sample arrays aren't accessed atomically with respect to the bounds, // so clearing reduces chance for dumpsys to read random uninitialized samples - memset(&mMonotonicNs, 0, sizeof(mMonotonicNs)); - memset(&mLoadNs, 0, sizeof(mLoadNs)); + memset(&mMonotonicNs[mSamplingN], 0, sizeof(mMonotonicNs[0]) * additional); + memset(&mLoadNs[mSamplingN], 0, sizeof(mLoadNs[0]) * additional); #ifdef CPU_FREQUENCY_STATISTICS - memset(&mCpukHz, 0, sizeof(mCpukHz)); + memset(&mCpukHz[mSamplingN], 0, sizeof(mCpukHz[0]) * additional); #endif + mSamplingN = samplingN; } +#endif FastMixerDumpState::~FastMixerDumpState() { @@ -641,7 +711,7 @@ static int compare_uint32_t(const void *pa, const void *pb) } } -void FastMixerDumpState::dump(int fd) +void FastMixerDumpState::dump(int fd) const { if (mCommand == FastMixerState::INITIAL) { fdprintf(fd, "FastMixer not initialized\n"); @@ -692,9 +762,9 @@ void FastMixerDumpState::dump(int fd) uint32_t newestOpen = bounds & 0xFFFF; uint32_t oldestClosed = bounds >> 16; uint32_t n = (newestOpen - oldestClosed) & 0xFFFF; - if (n > kSamplingN) { + if (n > mSamplingN) { ALOGE("too many samples %u", n); - n = kSamplingN; + n = mSamplingN; } // statistics for monotonic (wall clock) time, thread raw CPU load in time, CPU clock frequency, // and adjusted CPU load in MHz normalized for CPU clock frequency @@ -710,7 +780,7 @@ void FastMixerDumpState::dump(int fd) uint32_t *tail = n >= kTailDenominator ? new uint32_t[n] : NULL; // loop over all the samples for (uint32_t j = 0; j < n; ++j) { - size_t i = oldestClosed++ & (kSamplingN - 1); + size_t i = oldestClosed++ & (mSamplingN - 1); uint32_t wallNs = mMonotonicNs[i]; if (tail != NULL) { tail[j] = wallNs; |