1 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
5 #include "media/audio/win/audio_low_latency_output_win.h"
7 #include <Functiondiscoverykeys_devpkey.h>
9 #include "base/command_line.h"
10 #include "base/debug/trace_event.h"
11 #include "base/logging.h"
12 #include "base/memory/scoped_ptr.h"
13 #include "base/metrics/histogram.h"
14 #include "base/strings/utf_string_conversions.h"
15 #include "base/win/scoped_propvariant.h"
16 #include "media/audio/win/audio_manager_win.h"
17 #include "media/audio/win/avrt_wrapper_win.h"
18 #include "media/audio/win/core_audio_util_win.h"
19 #include "media/base/limits.h"
20 #include "media/base/media_switches.h"
22 using base::win::ScopedComPtr;
23 using base::win::ScopedCOMInitializer;
24 using base::win::ScopedCoMem;
29 AUDCLNT_SHAREMODE WASAPIAudioOutputStream::GetShareMode() {
30 const CommandLine* cmd_line = CommandLine::ForCurrentProcess();
31 if (cmd_line->HasSwitch(switches::kEnableExclusiveAudio))
32 return AUDCLNT_SHAREMODE_EXCLUSIVE;
33 return AUDCLNT_SHAREMODE_SHARED;
37 int WASAPIAudioOutputStream::HardwareSampleRate(const std::string& device_id) {
38 WAVEFORMATPCMEX format;
39 ScopedComPtr<IAudioClient> client;
40 if (device_id.empty()) {
41 client = CoreAudioUtil::CreateDefaultClient(eRender, eConsole);
43 ScopedComPtr<IMMDevice> device(CoreAudioUtil::CreateDevice(device_id));
46 client = CoreAudioUtil::CreateClient(device);
49 if (!client || FAILED(CoreAudioUtil::GetSharedModeMixFormat(client, &format)))
52 return static_cast<int>(format.Format.nSamplesPerSec);
55 WASAPIAudioOutputStream::WASAPIAudioOutputStream(AudioManagerWin* manager,
56 const std::string& device_id,
57 const AudioParameters& params,
59 : creating_thread_id_(base::PlatformThread::CurrentId()),
64 packet_size_frames_(0),
65 packet_size_bytes_(0),
66 endpoint_buffer_size_frames_(0),
67 device_id_(device_id),
68 device_role_(device_role),
69 share_mode_(GetShareMode()),
70 num_written_frames_(0),
72 audio_bus_(AudioBus::Create(params)) {
75 VLOG(1) << "WASAPIAudioOutputStream::WASAPIAudioOutputStream()";
76 VLOG_IF(1, share_mode_ == AUDCLNT_SHAREMODE_EXCLUSIVE)
77 << "Core Audio (WASAPI) EXCLUSIVE MODE is enabled.";
79 // Load the Avrt DLL if not already loaded. Required to support MMCSS.
80 bool avrt_init = avrt::Initialize();
81 DCHECK(avrt_init) << "Failed to load the avrt.dll";
83 // Set up the desired render format specified by the client. We use the
84 // WAVE_FORMAT_EXTENSIBLE structure to ensure that multiple channel ordering
85 // and high precision data can be supported.
87 // Begin with the WAVEFORMATEX structure that specifies the basic format.
88 WAVEFORMATEX* format = &format_.Format;
89 format->wFormatTag = WAVE_FORMAT_EXTENSIBLE;
90 format->nChannels = params.channels();
91 format->nSamplesPerSec = params.sample_rate();
92 format->wBitsPerSample = params.bits_per_sample();
93 format->nBlockAlign = (format->wBitsPerSample / 8) * format->nChannels;
94 format->nAvgBytesPerSec = format->nSamplesPerSec * format->nBlockAlign;
95 format->cbSize = sizeof(WAVEFORMATEXTENSIBLE) - sizeof(WAVEFORMATEX);
97 // Add the parts which are unique to WAVE_FORMAT_EXTENSIBLE.
98 format_.Samples.wValidBitsPerSample = params.bits_per_sample();
99 format_.dwChannelMask = CoreAudioUtil::GetChannelConfig(device_id, eRender);
100 format_.SubFormat = KSDATAFORMAT_SUBTYPE_PCM;
102 // Store size (in different units) of audio packets which we expect to
103 // get from the audio endpoint device in each render event.
104 packet_size_frames_ = params.frames_per_buffer();
105 packet_size_bytes_ = params.GetBytesPerBuffer();
106 VLOG(1) << "Number of bytes per audio frame : " << format->nBlockAlign;
107 VLOG(1) << "Number of audio frames per packet: " << packet_size_frames_;
108 VLOG(1) << "Number of bytes per packet : " << packet_size_bytes_;
109 VLOG(1) << "Number of milliseconds per packet: "
110 << params.GetBufferDuration().InMillisecondsF();
112 // All events are auto-reset events and non-signaled initially.
114 // Create the event which the audio engine will signal each time
115 // a buffer becomes ready to be processed by the client.
116 audio_samples_render_event_.Set(CreateEvent(NULL, FALSE, FALSE, NULL));
117 DCHECK(audio_samples_render_event_.IsValid());
119 // Create the event which will be set in Stop() when capturing shall stop.
120 stop_render_event_.Set(CreateEvent(NULL, FALSE, FALSE, NULL));
121 DCHECK(stop_render_event_.IsValid());
124 WASAPIAudioOutputStream::~WASAPIAudioOutputStream() {
125 DCHECK_EQ(GetCurrentThreadId(), creating_thread_id_);
128 bool WASAPIAudioOutputStream::Open() {
129 VLOG(1) << "WASAPIAudioOutputStream::Open()";
130 DCHECK_EQ(GetCurrentThreadId(), creating_thread_id_);
134 DCHECK(!audio_client_);
135 DCHECK(!audio_render_client_);
137 // Will be set to true if we ended up opening the default communications
139 bool communications_device = false;
141 // Create an IAudioClient interface for the default rendering IMMDevice.
142 ScopedComPtr<IAudioClient> audio_client;
143 if (device_id_.empty() ||
144 CoreAudioUtil::DeviceIsDefault(eRender, device_role_, device_id_)) {
145 audio_client = CoreAudioUtil::CreateDefaultClient(eRender, device_role_);
146 communications_device = (device_role_ == eCommunications);
148 ScopedComPtr<IMMDevice> device(CoreAudioUtil::CreateDevice(device_id_));
149 DLOG_IF(ERROR, !device) << "Failed to open device: " << device_id_;
151 audio_client = CoreAudioUtil::CreateClient(device);
157 // Extra sanity to ensure that the provided device format is still valid.
158 if (!CoreAudioUtil::IsFormatSupported(audio_client,
161 LOG(ERROR) << "Audio parameters are not supported.";
165 HRESULT hr = S_FALSE;
166 if (share_mode_ == AUDCLNT_SHAREMODE_SHARED) {
167 // Initialize the audio stream between the client and the device in shared
168 // mode and using event-driven buffer handling.
169 hr = CoreAudioUtil::SharedModeInitialize(
170 audio_client, &format_, audio_samples_render_event_.Get(),
171 &endpoint_buffer_size_frames_,
172 communications_device ? &kCommunicationsSessionId : NULL);
176 // We know from experience that the best possible callback sequence is
177 // achieved when the packet size (given by the native device period)
178 // is an even divisor of the endpoint buffer size.
179 // Examples: 48kHz => 960 % 480, 44.1kHz => 896 % 448 or 882 % 441.
180 if (endpoint_buffer_size_frames_ % packet_size_frames_ != 0) {
182 << "Bailing out due to non-perfect timing. Buffer size of "
183 << packet_size_frames_ << " is not an even divisor of "
184 << endpoint_buffer_size_frames_;
188 // TODO(henrika): break out to CoreAudioUtil::ExclusiveModeInitialize()
189 // when removing the enable-exclusive-audio flag.
190 hr = ExclusiveModeInitialization(audio_client,
191 audio_samples_render_event_.Get(),
192 &endpoint_buffer_size_frames_);
196 // The buffer scheme for exclusive mode streams is not designed for max
197 // flexibility. We only allow a "perfect match" between the packet size set
198 // by the user and the actual endpoint buffer size.
199 if (endpoint_buffer_size_frames_ != packet_size_frames_) {
200 LOG(ERROR) << "Bailing out due to non-perfect timing.";
205 // Create an IAudioRenderClient client for an initialized IAudioClient.
206 // The IAudioRenderClient interface enables us to write output data to
207 // a rendering endpoint buffer.
208 ScopedComPtr<IAudioRenderClient> audio_render_client =
209 CoreAudioUtil::CreateRenderClient(audio_client);
210 if (!audio_render_client)
213 // Store valid COM interfaces.
214 audio_client_ = audio_client;
215 audio_render_client_ = audio_render_client;
217 hr = audio_client_->GetService(__uuidof(IAudioClock),
218 audio_clock_.ReceiveVoid());
220 LOG(ERROR) << "Failed to get IAudioClock service.";
228 void WASAPIAudioOutputStream::Start(AudioSourceCallback* callback) {
229 VLOG(1) << "WASAPIAudioOutputStream::Start()";
230 DCHECK_EQ(GetCurrentThreadId(), creating_thread_id_);
234 if (render_thread_) {
235 CHECK_EQ(callback, source_);
241 // Ensure that the endpoint buffer is prepared with silence.
242 if (share_mode_ == AUDCLNT_SHAREMODE_SHARED) {
243 if (!CoreAudioUtil::FillRenderEndpointBufferWithSilence(
244 audio_client_, audio_render_client_)) {
245 LOG(ERROR) << "Failed to prepare endpoint buffers with silence.";
246 callback->OnError(this);
250 num_written_frames_ = endpoint_buffer_size_frames_;
252 // Create and start the thread that will drive the rendering by waiting for
254 render_thread_.reset(
255 new base::DelegateSimpleThread(this, "wasapi_render_thread"));
256 render_thread_->Start();
257 if (!render_thread_->HasBeenStarted()) {
258 LOG(ERROR) << "Failed to start WASAPI render thread.";
260 callback->OnError(this);
264 // Start streaming data between the endpoint buffer and the audio engine.
265 HRESULT hr = audio_client_->Start();
267 PLOG(ERROR) << "Failed to start output streaming: " << std::hex << hr;
269 callback->OnError(this);
273 void WASAPIAudioOutputStream::Stop() {
274 VLOG(1) << "WASAPIAudioOutputStream::Stop()";
275 DCHECK_EQ(GetCurrentThreadId(), creating_thread_id_);
279 // Stop output audio streaming.
280 HRESULT hr = audio_client_->Stop();
282 PLOG(ERROR) << "Failed to stop output streaming: " << std::hex << hr;
283 source_->OnError(this);
286 // Make a local copy of |source_| since StopThread() will clear it.
287 AudioSourceCallback* callback = source_;
290 // Flush all pending data and reset the audio clock stream position to 0.
291 hr = audio_client_->Reset();
293 PLOG(ERROR) << "Failed to reset streaming: " << std::hex << hr;
294 callback->OnError(this);
297 // Extra safety check to ensure that the buffers are cleared.
298 // If the buffers are not cleared correctly, the next call to Start()
299 // would fail with AUDCLNT_E_BUFFER_ERROR at IAudioRenderClient::GetBuffer().
300 // This check is is only needed for shared-mode streams.
301 if (share_mode_ == AUDCLNT_SHAREMODE_SHARED) {
302 UINT32 num_queued_frames = 0;
303 audio_client_->GetCurrentPadding(&num_queued_frames);
304 DCHECK_EQ(0u, num_queued_frames);
308 void WASAPIAudioOutputStream::Close() {
309 VLOG(1) << "WASAPIAudioOutputStream::Close()";
310 DCHECK_EQ(GetCurrentThreadId(), creating_thread_id_);
312 // It is valid to call Close() before calling open or Start().
313 // It is also valid to call Close() after Start() has been called.
316 // Inform the audio manager that we have been closed. This will cause our
318 manager_->ReleaseOutputStream(this);
321 void WASAPIAudioOutputStream::SetVolume(double volume) {
322 VLOG(1) << "SetVolume(volume=" << volume << ")";
323 float volume_float = static_cast<float>(volume);
324 if (volume_float < 0.0f || volume_float > 1.0f) {
327 volume_ = volume_float;
330 void WASAPIAudioOutputStream::GetVolume(double* volume) {
331 VLOG(1) << "GetVolume()";
332 *volume = static_cast<double>(volume_);
335 void WASAPIAudioOutputStream::Run() {
336 ScopedCOMInitializer com_init(ScopedCOMInitializer::kMTA);
338 // Increase the thread priority.
339 render_thread_->SetThreadPriority(base::kThreadPriority_RealtimeAudio);
341 // Enable MMCSS to ensure that this thread receives prioritized access to
343 DWORD task_index = 0;
344 HANDLE mm_task = avrt::AvSetMmThreadCharacteristics(L"Pro Audio",
347 (mm_task && avrt::AvSetMmThreadPriority(mm_task, AVRT_PRIORITY_CRITICAL));
349 // Failed to enable MMCSS on this thread. It is not fatal but can lead
350 // to reduced QoS at high load.
351 DWORD err = GetLastError();
352 LOG(WARNING) << "Failed to enable MMCSS (error code=" << err << ").";
355 HRESULT hr = S_FALSE;
359 HANDLE wait_array[] = { stop_render_event_.Get(),
360 audio_samples_render_event_.Get() };
361 UINT64 device_frequency = 0;
363 // The device frequency is the frequency generated by the hardware clock in
364 // the audio device. The GetFrequency() method reports a constant frequency.
365 hr = audio_clock_->GetFrequency(&device_frequency);
367 PLOG_IF(ERROR, error) << "Failed to acquire IAudioClock interface: "
370 // Keep rendering audio until the stop event or the stream-switch event
371 // is signaled. An error event can also break the main thread loop.
372 while (playing && !error) {
373 // Wait for a close-down event, stream-switch event or a new render event.
374 DWORD wait_result = WaitForMultipleObjects(arraysize(wait_array),
379 switch (wait_result) {
380 case WAIT_OBJECT_0 + 0:
381 // |stop_render_event_| has been set.
384 case WAIT_OBJECT_0 + 1:
385 // |audio_samples_render_event_| has been set.
386 error = !RenderAudioFromSource(device_frequency);
394 if (playing && error) {
395 // Stop audio rendering since something has gone wrong in our main thread
396 // loop. Note that, we are still in a "started" state, hence a Stop() call
397 // is required to join the thread properly.
398 audio_client_->Stop();
399 PLOG(ERROR) << "WASAPI rendering failed.";
403 if (mm_task && !avrt::AvRevertMmThreadCharacteristics(mm_task)) {
404 PLOG(WARNING) << "Failed to disable MMCSS";
408 bool WASAPIAudioOutputStream::RenderAudioFromSource(UINT64 device_frequency) {
409 TRACE_EVENT0("audio", "RenderAudioFromSource");
411 HRESULT hr = S_FALSE;
412 UINT32 num_queued_frames = 0;
413 uint8* audio_data = NULL;
415 // Contains how much new data we can write to the buffer without
416 // the risk of overwriting previously written data that the audio
417 // engine has not yet read from the buffer.
418 size_t num_available_frames = 0;
420 if (share_mode_ == AUDCLNT_SHAREMODE_SHARED) {
421 // Get the padding value which represents the amount of rendering
422 // data that is queued up to play in the endpoint buffer.
423 hr = audio_client_->GetCurrentPadding(&num_queued_frames);
424 num_available_frames =
425 endpoint_buffer_size_frames_ - num_queued_frames;
427 DLOG(ERROR) << "Failed to retrieve amount of available space: "
432 // While the stream is running, the system alternately sends one
433 // buffer or the other to the client. This form of double buffering
434 // is referred to as "ping-ponging". Each time the client receives
435 // a buffer from the system (triggers this event) the client must
436 // process the entire buffer. Calls to the GetCurrentPadding method
437 // are unnecessary because the packet size must always equal the
438 // buffer size. In contrast to the shared mode buffering scheme,
439 // the latency for an event-driven, exclusive-mode stream depends
440 // directly on the buffer size.
441 num_available_frames = endpoint_buffer_size_frames_;
444 // Check if there is enough available space to fit the packet size
445 // specified by the client.
446 if (num_available_frames < packet_size_frames_)
449 DLOG_IF(ERROR, num_available_frames % packet_size_frames_ != 0)
450 << "Non-perfect timing detected (num_available_frames="
451 << num_available_frames << ", packet_size_frames="
452 << packet_size_frames_ << ")";
454 // Derive the number of packets we need to get from the client to
455 // fill up the available area in the endpoint buffer.
456 // |num_packets| will always be one for exclusive-mode streams and
457 // will be one in most cases for shared mode streams as well.
458 // However, we have found that two packets can sometimes be
460 size_t num_packets = (num_available_frames / packet_size_frames_);
462 for (size_t n = 0; n < num_packets; ++n) {
463 // Grab all available space in the rendering endpoint buffer
464 // into which the client can write a data packet.
465 hr = audio_render_client_->GetBuffer(packet_size_frames_,
468 DLOG(ERROR) << "Failed to use rendering audio buffer: "
473 // Derive the audio delay which corresponds to the delay between
474 // a render event and the time when the first audio sample in a
475 // packet is played out through the speaker. This delay value
476 // can typically be utilized by an acoustic echo-control (AEC)
477 // unit at the render side.
479 uint32 audio_delay_bytes = 0;
480 hr = audio_clock_->GetPosition(&position, NULL);
482 // Stream position of the sample that is currently playing
483 // through the speaker.
484 double pos_sample_playing_frames = format_.Format.nSamplesPerSec *
485 (static_cast<double>(position) / device_frequency);
487 // Stream position of the last sample written to the endpoint
488 // buffer. Note that, the packet we are about to receive in
489 // the upcoming callback is also included.
490 size_t pos_last_sample_written_frames =
491 num_written_frames_ + packet_size_frames_;
493 // Derive the actual delay value which will be fed to the
494 // render client using the OnMoreData() callback.
495 audio_delay_bytes = (pos_last_sample_written_frames -
496 pos_sample_playing_frames) * format_.Format.nBlockAlign;
499 // Read a data packet from the registered client source and
500 // deliver a delay estimate in the same callback to the client.
502 int frames_filled = source_->OnMoreData(
503 audio_bus_.get(), audio_delay_bytes);
504 uint32 num_filled_bytes = frames_filled * format_.Format.nBlockAlign;
505 DCHECK_LE(num_filled_bytes, packet_size_bytes_);
507 // Note: If this ever changes to output raw float the data must be
508 // clipped and sanitized since it may come from an untrusted
509 // source such as NaCl.
510 const int bytes_per_sample = format_.Format.wBitsPerSample >> 3;
511 audio_bus_->Scale(volume_);
512 audio_bus_->ToInterleaved(
513 frames_filled, bytes_per_sample, audio_data);
516 // Release the buffer space acquired in the GetBuffer() call.
517 // Render silence if we were not able to fill up the buffer totally.
518 DWORD flags = (num_filled_bytes < packet_size_bytes_) ?
519 AUDCLNT_BUFFERFLAGS_SILENT : 0;
520 audio_render_client_->ReleaseBuffer(packet_size_frames_, flags);
522 num_written_frames_ += packet_size_frames_;
528 HRESULT WASAPIAudioOutputStream::ExclusiveModeInitialization(
529 IAudioClient* client, HANDLE event_handle, uint32* endpoint_buffer_size) {
530 DCHECK_EQ(share_mode_, AUDCLNT_SHAREMODE_EXCLUSIVE);
532 float f = (1000.0 * packet_size_frames_) / format_.Format.nSamplesPerSec;
533 REFERENCE_TIME requested_buffer_duration =
534 static_cast<REFERENCE_TIME>(f * 10000.0 + 0.5);
536 DWORD stream_flags = AUDCLNT_STREAMFLAGS_NOPERSIST;
537 bool use_event = (event_handle != NULL &&
538 event_handle != INVALID_HANDLE_VALUE);
540 stream_flags |= AUDCLNT_STREAMFLAGS_EVENTCALLBACK;
541 VLOG(2) << "stream_flags: 0x" << std::hex << stream_flags;
543 // Initialize the audio stream between the client and the device.
544 // For an exclusive-mode stream that uses event-driven buffering, the
545 // caller must specify nonzero values for hnsPeriodicity and
546 // hnsBufferDuration, and the values of these two parameters must be equal.
547 // The Initialize method allocates two buffers for the stream. Each buffer
548 // is equal in duration to the value of the hnsBufferDuration parameter.
549 // Following the Initialize call for a rendering stream, the caller should
550 // fill the first of the two buffers before starting the stream.
551 HRESULT hr = S_FALSE;
552 hr = client->Initialize(AUDCLNT_SHAREMODE_EXCLUSIVE,
554 requested_buffer_duration,
555 requested_buffer_duration,
556 reinterpret_cast<WAVEFORMATEX*>(&format_),
559 if (hr == AUDCLNT_E_BUFFER_SIZE_NOT_ALIGNED) {
560 LOG(ERROR) << "AUDCLNT_E_BUFFER_SIZE_NOT_ALIGNED";
562 UINT32 aligned_buffer_size = 0;
563 client->GetBufferSize(&aligned_buffer_size);
564 VLOG(1) << "Use aligned buffer size instead: " << aligned_buffer_size;
566 // Calculate new aligned periodicity. Each unit of reference time
567 // is 100 nanoseconds.
568 REFERENCE_TIME aligned_buffer_duration = static_cast<REFERENCE_TIME>(
569 (10000000.0 * aligned_buffer_size / format_.Format.nSamplesPerSec)
572 // It is possible to re-activate and re-initialize the audio client
573 // at this stage but we bail out with an error code instead and
574 // combine it with a log message which informs about the suggested
575 // aligned buffer size which should be used instead.
576 VLOG(1) << "aligned_buffer_duration: "
577 << static_cast<double>(aligned_buffer_duration / 10000.0)
579 } else if (hr == AUDCLNT_E_INVALID_DEVICE_PERIOD) {
580 // We will get this error if we try to use a smaller buffer size than
581 // the minimum supported size (usually ~3ms on Windows 7).
582 LOG(ERROR) << "AUDCLNT_E_INVALID_DEVICE_PERIOD";
588 hr = client->SetEventHandle(event_handle);
590 VLOG(1) << "IAudioClient::SetEventHandle: " << std::hex << hr;
595 UINT32 buffer_size_in_frames = 0;
596 hr = client->GetBufferSize(&buffer_size_in_frames);
598 VLOG(1) << "IAudioClient::GetBufferSize: " << std::hex << hr;
602 *endpoint_buffer_size = buffer_size_in_frames;
603 VLOG(2) << "endpoint buffer size: " << buffer_size_in_frames;
607 void WASAPIAudioOutputStream::StopThread() {
608 if (render_thread_ ) {
609 if (render_thread_->HasBeenStarted()) {
610 // Wait until the thread completes and perform cleanup.
611 SetEvent(stop_render_event_.Get());
612 render_thread_->Join();
615 render_thread_.reset();
617 // Ensure that we don't quit the main thread loop immediately next
618 // time Start() is called.
619 ResetEvent(stop_render_event_.Get());