| 1 | // Copyright (C) 2021 The Qt Company Ltd. |
| 2 | // SPDX-License-Identifier: LicenseRef-Qt-Commercial OR LGPL-3.0-only OR GPL-2.0-only OR GPL-3.0-only |
| 3 | |
| 4 | #include "playbackengine/qffmpegaudiorenderer_p.h" |
| 5 | #include "qaudiosink.h" |
| 6 | #include "qaudiooutput.h" |
| 7 | #include "qaudiobufferoutput.h" |
| 8 | #include "private/qplatformaudiooutput_p.h" |
| 9 | #include <QtCore/qloggingcategory.h> |
| 10 | |
| 11 | #include "qffmpegresampler_p.h" |
| 12 | #include "qffmpegmediaformatinfo_p.h" |
| 13 | |
| 14 | QT_BEGIN_NAMESPACE |
| 15 | |
| 16 | static Q_LOGGING_CATEGORY(qLcAudioRenderer, "qt.multimedia.ffmpeg.audiorenderer" ); |
| 17 | |
| 18 | namespace QFFmpeg { |
| 19 | |
| 20 | using namespace std::chrono_literals; |
| 21 | using namespace std::chrono; |
| 22 | |
| 23 | namespace { |
| 24 | constexpr auto DesiredBufferTime = 110000us; |
| 25 | constexpr auto MinDesiredBufferTime = 22000us; |
| 26 | constexpr auto MaxDesiredBufferTime = 64000us; |
| 27 | constexpr auto MinDesiredFreeBufferTime = 10000us; |
| 28 | |
| 29 | // It might be changed with #ifdef, as on Linux, QPulseAudioSink has quite unstable timings, |
| 30 | // and it needs much more time to make sure that the buffer is overloaded. |
| 31 | constexpr auto BufferLoadingMeasureTime = 400ms; |
| 32 | |
| 33 | constexpr auto DurationBias = 2ms; // avoids extra timer events |
| 34 | |
| 35 | qreal sampleRateFactor() { |
| 36 | // Test purposes: |
| 37 | // |
| 38 | // The env var describes a factor for the sample rate of |
| 39 | // audio data that we feed to the audio sink. |
| 40 | // |
| 41 | // In some cases audio sink might consume data slightly slower or faster than expected; |
| 42 | // even though the synchronization in the audio renderer is supposed to handle it, |
| 43 | // it makes sense to experiment with QT_MEDIA_PLAYER_AUDIO_SAMPLE_RATE_FACTOR != 1. |
| 44 | // |
| 45 | // Set QT_MEDIA_PLAYER_AUDIO_SAMPLE_RATE_FACTOR > 1 (e.g. 1.01 - 1.1) to test high buffer loading |
| 46 | // or try compensating too fast data consumption by the audio sink. |
| 47 | // Set QT_MEDIA_PLAYER_AUDIO_SAMPLE_RATE_FACTOR < 1 to test low buffer loading |
| 48 | // or try compensating too slow data consumption by the audio sink. |
| 49 | |
| 50 | |
| 51 | static const qreal result = []() { |
| 52 | const auto sampleRateFactorStr = qEnvironmentVariable(varName: "QT_MEDIA_PLAYER_AUDIO_SAMPLE_RATE_FACTOR" ); |
| 53 | bool ok = false; |
| 54 | const auto result = sampleRateFactorStr.toDouble(ok: &ok); |
| 55 | return ok ? result : 1.; |
| 56 | }(); |
| 57 | |
| 58 | return result; |
| 59 | } |
| 60 | |
| 61 | QAudioFormat audioFormatFromFrame(const Frame &frame) |
| 62 | { |
| 63 | return QFFmpegMediaFormatInfo::audioFormatFromCodecParameters( |
| 64 | codecPar: *frame.codecContext()->stream()->codecpar); |
| 65 | } |
| 66 | |
| 67 | std::unique_ptr<QFFmpegResampler> createResampler(const Frame &frame, |
| 68 | const QAudioFormat &outputFormat) |
| 69 | { |
| 70 | return std::make_unique<QFFmpegResampler>(args: frame.codecContext(), args: outputFormat, |
| 71 | args: frame.startTime().get()); |
| 72 | } |
| 73 | |
| 74 | } // namespace |
| 75 | |
| 76 | AudioRenderer::AudioRenderer(const TimeController &tc, QAudioOutput *output, |
| 77 | QAudioBufferOutput *bufferOutput) |
| 78 | : Renderer(tc), m_output(output), m_bufferOutput(bufferOutput) |
| 79 | { |
| 80 | if (output) { |
| 81 | // TODO: implement the signals in QPlatformAudioOutput and connect to them, QTBUG-112294 |
| 82 | connect(sender: output, signal: &QAudioOutput::deviceChanged, context: this, slot: &AudioRenderer::onDeviceChanged); |
| 83 | connect(sender: output, signal: &QAudioOutput::volumeChanged, context: this, slot: &AudioRenderer::updateVolume); |
| 84 | connect(sender: output, signal: &QAudioOutput::mutedChanged, context: this, slot: &AudioRenderer::updateVolume); |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | void AudioRenderer::setOutput(QAudioOutput *output) |
| 89 | { |
| 90 | setOutputInternal(actual&: m_output, desired: output, changeHandler: [this](QAudioOutput *) { onDeviceChanged(); }); |
| 91 | } |
| 92 | |
| 93 | void AudioRenderer::setOutput(QAudioBufferOutput *bufferOutput) |
| 94 | { |
| 95 | setOutputInternal(actual&: m_bufferOutput, desired: bufferOutput, |
| 96 | changeHandler: [this](QAudioBufferOutput *) { m_bufferOutputChanged = true; }); |
| 97 | } |
| 98 | |
| 99 | AudioRenderer::~AudioRenderer() |
| 100 | { |
| 101 | freeOutput(); |
| 102 | } |
| 103 | |
| 104 | void AudioRenderer::updateVolume() |
| 105 | { |
| 106 | if (m_sink) |
| 107 | m_sink->setVolume(m_output->isMuted() ? 0.f : m_output->volume()); |
| 108 | } |
| 109 | |
| 110 | void AudioRenderer::onDeviceChanged() |
| 111 | { |
| 112 | m_deviceChanged = true; |
| 113 | } |
| 114 | |
| 115 | Renderer::RenderingResult AudioRenderer::renderInternal(Frame frame) |
| 116 | { |
| 117 | if (frame.isValid()) |
| 118 | updateOutputs(frame); |
| 119 | |
| 120 | // push to sink first in order not to waste time on resampling |
| 121 | // for QAudioBufferOutput |
| 122 | const RenderingResult result = pushFrameToOutput(frame); |
| 123 | |
| 124 | if (m_lastFramePushDone) |
| 125 | pushFrameToBufferOutput(frame); |
| 126 | // else // skip pushing the same data to QAudioBufferOutput |
| 127 | |
| 128 | m_lastFramePushDone = result.done; |
| 129 | |
| 130 | return result; |
| 131 | } |
| 132 | |
| 133 | AudioRenderer::RenderingResult AudioRenderer::pushFrameToOutput(const Frame &frame) |
| 134 | { |
| 135 | if (!m_ioDevice || !m_resampler) |
| 136 | return {}; |
| 137 | |
| 138 | Q_ASSERT(m_sink); |
| 139 | |
| 140 | auto firstFrameFlagGuard = qScopeGuard(f: [&]() { m_firstFrameToSink = false; }); |
| 141 | |
| 142 | const SynchronizationStamp syncStamp{ .audioSinkState: m_sink->state(), .audioSinkBytesFree: m_sink->bytesFree(), |
| 143 | .bufferBytesWritten: m_bufferedData.offset, .timePoint: RealClock::now() }; |
| 144 | |
| 145 | if (!m_bufferedData.isValid()) { |
| 146 | if (!frame.isValid()) { |
| 147 | if (std::exchange(obj&: m_drained, new_val: true)) |
| 148 | return {}; |
| 149 | |
| 150 | const auto time = bufferLoadingTime(syncStamp); |
| 151 | |
| 152 | qCDebug(qLcAudioRenderer) << "Draining AudioRenderer, time:" << time; |
| 153 | |
| 154 | return { .done: time.count() == 0, .recheckInterval: time }; |
| 155 | } |
| 156 | |
| 157 | m_bufferedData = { .buffer: m_resampler->resample(frame: frame.avFrame()) }; |
| 158 | } |
| 159 | |
| 160 | if (m_bufferedData.isValid()) { |
| 161 | // synchronize after "QIODevice::write" to deliver audio data to the sink ASAP. |
| 162 | auto syncGuard = qScopeGuard(f: [&]() { updateSynchronization(stamp: syncStamp, frame); }); |
| 163 | |
| 164 | const auto bytesWritten = m_ioDevice->write(data: m_bufferedData.data(), len: m_bufferedData.size()); |
| 165 | |
| 166 | m_bufferedData.offset += bytesWritten; |
| 167 | |
| 168 | if (m_bufferedData.size() <= 0) { |
| 169 | m_bufferedData = {}; |
| 170 | |
| 171 | return {}; |
| 172 | } |
| 173 | |
| 174 | const auto remainingDuration = durationForBytes(bytes: m_bufferedData.size()); |
| 175 | |
| 176 | return { .done: false, |
| 177 | .recheckInterval: std::min(a: remainingDuration + DurationBias, b: m_timings.actualBufferDuration / 2) }; |
| 178 | } |
| 179 | |
| 180 | return {}; |
| 181 | } |
| 182 | |
| 183 | void AudioRenderer::pushFrameToBufferOutput(const Frame &frame) |
| 184 | { |
| 185 | if (!m_bufferOutput) |
| 186 | return; |
| 187 | |
| 188 | if (frame.isValid()) { |
| 189 | Q_ASSERT(m_bufferOutputResampler); |
| 190 | |
| 191 | // TODO: get buffer from m_bufferedData if resample formats are equal |
| 192 | QAudioBuffer buffer = m_bufferOutputResampler->resample(frame: frame.avFrame()); |
| 193 | emit m_bufferOutput->audioBufferReceived(buffer); |
| 194 | } else { |
| 195 | emit m_bufferOutput->audioBufferReceived(buffer: {}); |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | void AudioRenderer::onPlaybackRateChanged() |
| 200 | { |
| 201 | m_resampler.reset(); |
| 202 | } |
| 203 | |
| 204 | std::chrono::milliseconds AudioRenderer::timerInterval() const |
| 205 | { |
| 206 | constexpr auto MaxFixableInterval = 50ms; |
| 207 | |
| 208 | const auto interval = Renderer::timerInterval(); |
| 209 | |
| 210 | if (m_firstFrameToSink || !m_sink || m_sink->state() != QAudio::IdleState |
| 211 | || interval > MaxFixableInterval) |
| 212 | return interval; |
| 213 | |
| 214 | return 0ms; |
| 215 | } |
| 216 | |
| 217 | void AudioRenderer::onPauseChanged() |
| 218 | { |
| 219 | m_firstFrameToSink = true; |
| 220 | Renderer::onPauseChanged(); |
| 221 | } |
| 222 | |
| 223 | void AudioRenderer::initResampler(const Frame &frame) |
| 224 | { |
| 225 | // We recreate resampler whenever format is changed |
| 226 | |
| 227 | auto resamplerFormat = m_sinkFormat; |
| 228 | resamplerFormat.setSampleRate( |
| 229 | qRound(d: m_sinkFormat.sampleRate() / playbackRate() * sampleRateFactor())); |
| 230 | m_resampler = createResampler(frame, outputFormat: resamplerFormat); |
| 231 | } |
| 232 | |
| 233 | void AudioRenderer::freeOutput() |
| 234 | { |
| 235 | qCDebug(qLcAudioRenderer) << "Free audio output" ; |
| 236 | if (m_sink) { |
| 237 | m_sink->reset(); |
| 238 | |
| 239 | // TODO: inestigate if it's enough to reset the sink without deleting |
| 240 | m_sink.reset(); |
| 241 | } |
| 242 | |
| 243 | m_ioDevice = nullptr; |
| 244 | |
| 245 | m_bufferedData = {}; |
| 246 | m_deviceChanged = false; |
| 247 | m_sinkFormat = {}; |
| 248 | m_timings = {}; |
| 249 | m_bufferLoadingInfo = {}; |
| 250 | } |
| 251 | |
| 252 | void AudioRenderer::updateOutputs(const Frame &frame) |
| 253 | { |
| 254 | if (m_deviceChanged) { |
| 255 | freeOutput(); |
| 256 | m_resampler.reset(); |
| 257 | } |
| 258 | |
| 259 | if (m_bufferOutput) { |
| 260 | if (m_bufferOutputChanged) { |
| 261 | m_bufferOutputChanged = false; |
| 262 | m_bufferOutputResampler.reset(); |
| 263 | } |
| 264 | |
| 265 | if (!m_bufferOutputResampler) { |
| 266 | QAudioFormat outputFormat = m_bufferOutput->format(); |
| 267 | if (!outputFormat.isValid()) |
| 268 | outputFormat = audioFormatFromFrame(frame); |
| 269 | m_bufferOutputResampler = createResampler(frame, outputFormat); |
| 270 | } |
| 271 | } |
| 272 | |
| 273 | if (!m_output) |
| 274 | return; |
| 275 | |
| 276 | if (!m_sinkFormat.isValid()) { |
| 277 | m_sinkFormat = audioFormatFromFrame(frame); |
| 278 | m_sinkFormat.setChannelConfig(m_output->device().channelConfiguration()); |
| 279 | } |
| 280 | |
| 281 | if (!m_sink) { |
| 282 | // Insert a delay here to test time offset synchronization, e.g. QThread::sleep(1) |
| 283 | m_sink = std::make_unique<QAudioSink>(args: m_output->device(), args&: m_sinkFormat); |
| 284 | updateVolume(); |
| 285 | m_sink->setBufferSize(m_sinkFormat.bytesForDuration(microseconds: DesiredBufferTime.count())); |
| 286 | m_ioDevice = m_sink->start(); |
| 287 | m_firstFrameToSink = true; |
| 288 | |
| 289 | connect(sender: m_sink.get(), signal: &QAudioSink::stateChanged, context: this, |
| 290 | slot: &AudioRenderer::onAudioSinkStateChanged); |
| 291 | |
| 292 | m_timings.actualBufferDuration = durationForBytes(bytes: m_sink->bufferSize()); |
| 293 | m_timings.maxSoundDelay = qMin(a: MaxDesiredBufferTime, |
| 294 | b: m_timings.actualBufferDuration - MinDesiredFreeBufferTime); |
| 295 | m_timings.minSoundDelay = MinDesiredBufferTime; |
| 296 | |
| 297 | Q_ASSERT(DurationBias < m_timings.minSoundDelay |
| 298 | && m_timings.maxSoundDelay < m_timings.actualBufferDuration); |
| 299 | } |
| 300 | |
| 301 | if (!m_resampler) |
| 302 | initResampler(frame); |
| 303 | } |
| 304 | |
| 305 | void AudioRenderer::updateSynchronization(const SynchronizationStamp &stamp, const Frame &frame) |
| 306 | { |
| 307 | if (!frame.isValid()) |
| 308 | return; |
| 309 | |
| 310 | Q_ASSERT(m_sink); |
| 311 | |
| 312 | const auto bufferLoadingTime = this->bufferLoadingTime(syncStamp: stamp); |
| 313 | const auto currentFrameDelay = frameDelay(frame, timePoint: stamp.timePoint); |
| 314 | const auto writtenTime = durationForBytes(bytes: stamp.bufferBytesWritten); |
| 315 | const auto soundDelay = currentFrameDelay + bufferLoadingTime - writtenTime; |
| 316 | |
| 317 | auto synchronize = [&](microseconds fixedDelay, microseconds targetSoundDelay) { |
| 318 | // TODO: investigate if we need sample compensation here |
| 319 | |
| 320 | changeRendererTime(offset: fixedDelay - targetSoundDelay); |
| 321 | if (qLcAudioRenderer().isDebugEnabled()) { |
| 322 | // clang-format off |
| 323 | qCDebug(qLcAudioRenderer) |
| 324 | << "Change rendering time:" |
| 325 | << "\n First frame:" << m_firstFrameToSink |
| 326 | << "\n Delay (frame+buffer-written):" << currentFrameDelay << "+" |
| 327 | << bufferLoadingTime << "-" |
| 328 | << writtenTime << "=" |
| 329 | << soundDelay |
| 330 | << "\n Fixed delay:" << fixedDelay |
| 331 | << "\n Target delay:" << targetSoundDelay |
| 332 | << "\n Buffer durations (min/max/limit):" << m_timings.minSoundDelay |
| 333 | << m_timings.maxSoundDelay |
| 334 | << m_timings.actualBufferDuration |
| 335 | << "\n Audio sink state:" << stamp.audioSinkState; |
| 336 | // clang-format on |
| 337 | } |
| 338 | }; |
| 339 | |
| 340 | const auto loadingType = soundDelay > m_timings.maxSoundDelay ? BufferLoadingInfo::High |
| 341 | : soundDelay < m_timings.minSoundDelay ? BufferLoadingInfo::Low |
| 342 | : BufferLoadingInfo::Moderate; |
| 343 | |
| 344 | if (loadingType != m_bufferLoadingInfo.type) { |
| 345 | // qCDebug(qLcAudioRenderer) << "Change buffer loading type:" << |
| 346 | // m_bufferLoadingInfo.type |
| 347 | // << "->" << loadingType << "soundDelay:" << soundDelay; |
| 348 | m_bufferLoadingInfo = { .type: loadingType, .timePoint: stamp.timePoint, .delay: soundDelay }; |
| 349 | } |
| 350 | |
| 351 | if (loadingType != BufferLoadingInfo::Moderate) { |
| 352 | const auto isHigh = loadingType == BufferLoadingInfo::High; |
| 353 | const auto shouldHandleIdle = stamp.audioSinkState == QAudio::IdleState && !isHigh; |
| 354 | |
| 355 | auto &fixedDelay = m_bufferLoadingInfo.delay; |
| 356 | |
| 357 | fixedDelay = shouldHandleIdle ? soundDelay |
| 358 | : isHigh ? qMin(a: soundDelay, b: fixedDelay) |
| 359 | : qMax(a: soundDelay, b: fixedDelay); |
| 360 | |
| 361 | if (stamp.timePoint - m_bufferLoadingInfo.timePoint > BufferLoadingMeasureTime |
| 362 | || (m_firstFrameToSink && isHigh) || shouldHandleIdle) { |
| 363 | const auto targetDelay = isHigh |
| 364 | ? (m_timings.maxSoundDelay + m_timings.minSoundDelay) / 2 |
| 365 | : m_timings.minSoundDelay + DurationBias; |
| 366 | |
| 367 | synchronize(fixedDelay, targetDelay); |
| 368 | m_bufferLoadingInfo = { .type: BufferLoadingInfo::Moderate, .timePoint: stamp.timePoint, .delay: targetDelay }; |
| 369 | } |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | microseconds AudioRenderer::bufferLoadingTime(const SynchronizationStamp &syncStamp) const |
| 374 | { |
| 375 | Q_ASSERT(m_sink); |
| 376 | |
| 377 | if (syncStamp.audioSinkState == QAudio::IdleState) |
| 378 | return microseconds(0); |
| 379 | |
| 380 | const auto bytes = qMax(a: m_sink->bufferSize() - syncStamp.audioSinkBytesFree, b: 0); |
| 381 | |
| 382 | #ifdef Q_OS_ANDROID |
| 383 | // The hack has been added due to QAndroidAudioSink issues (QTBUG-118609). |
| 384 | // The method QAndroidAudioSink::bytesFree returns 0 or bufferSize, intermediate values are not |
| 385 | // available now; to be fixed. |
| 386 | if (bytes == 0) |
| 387 | return m_timings.minSoundDelay + MinDesiredBufferTime; |
| 388 | #endif |
| 389 | |
| 390 | return durationForBytes(bytes); |
| 391 | } |
| 392 | |
| 393 | void AudioRenderer::onAudioSinkStateChanged(QAudio::State state) |
| 394 | { |
| 395 | if (state == QAudio::IdleState && !m_firstFrameToSink && !m_deviceChanged) |
| 396 | scheduleNextStep(); |
| 397 | } |
| 398 | |
| 399 | microseconds AudioRenderer::durationForBytes(qsizetype bytes) const |
| 400 | { |
| 401 | return microseconds(m_sinkFormat.durationForBytes(byteCount: static_cast<qint32>(bytes))); |
| 402 | } |
| 403 | |
| 404 | } // namespace QFFmpeg |
| 405 | |
| 406 | QT_END_NAMESPACE |
| 407 | |
| 408 | #include "moc_qffmpegaudiorenderer_p.cpp" |
| 409 | |