media/filters/ffmpeg_audio_decoder.cc

   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.
   4
   5 #include "media/filters/ffmpeg_audio_decoder.h"
   6
   7 #include "base/callback_helpers.h"
   8 #include "base/single_thread_task_runner.h"
   9 #include "media/base/audio_buffer.h"
  10 #include "media/base/audio_bus.h"
  11 #include "media/base/audio_decoder_config.h"
  12 #include "media/base/audio_discard_helper.h"
  13 #include "media/base/bind_to_current_loop.h"
  14 #include "media/base/decoder_buffer.h"
  15 #include "media/base/limits.h"
  16 #include "media/base/timestamp_constants.h"
  17 #include "media/ffmpeg/ffmpeg_common.h"
  18 #include "media/filters/ffmpeg_glue.h"
  19
  20 namespace media {
  21
  22 // Returns true if the decode result was end of stream.
  23 static inline bool IsEndOfStream(int result,
  24                                  int decoded_size,
  25                                  const scoped_refptr<DecoderBuffer>& input) {
  26   // Three conditions to meet to declare end of stream for this decoder:
  27   // 1. FFmpeg didn't read anything.
  28   // 2. FFmpeg didn't output anything.
  29   // 3. An end of stream buffer is received.
  30   return result == 0 && decoded_size == 0 && input->end_of_stream();
  31 }
  32
  33 // Return the number of channels from the data in |frame|.
  34 static inline int DetermineChannels(AVFrame* frame) {
  35 #if defined(CHROMIUM_NO_AVFRAME_CHANNELS)
  36   // When use_system_ffmpeg==1, libav's AVFrame doesn't have channels field.
  37   return av_get_channel_layout_nb_channels(frame->channel_layout);
  38 #else
  39   return frame->channels;
  40 #endif
  41 }
  42
  43 // Called by FFmpeg's allocation routine to free a buffer. |opaque| is the
  44 // AudioBuffer allocated, so unref it.
  45 static void ReleaseAudioBufferImpl(void* opaque, uint8* data) {
  46   scoped_refptr<AudioBuffer> buffer;
  47   buffer.swap(reinterpret_cast<AudioBuffer**>(&opaque));
  48 }
  49
  50 // Called by FFmpeg's allocation routine to allocate a buffer. Uses
  51 // AVCodecContext.opaque to get the object reference in order to call
  52 // GetAudioBuffer() to do the actual allocation.
  53 static int GetAudioBuffer(struct AVCodecContext* s, AVFrame* frame, int flags) {
  54   DCHECK(s->codec->capabilities & CODEC_CAP_DR1);
  55   DCHECK_EQ(s->codec_type, AVMEDIA_TYPE_AUDIO);
  56
  57   // Since this routine is called by FFmpeg when a buffer is required for audio
  58   // data, use the values supplied by FFmpeg (ignoring the current settings).
  59   // FFmpegDecode() gets to determine if the buffer is useable or not.
  60   AVSampleFormat format = static_cast<AVSampleFormat>(frame->format);
  61   SampleFormat sample_format = AVSampleFormatToSampleFormat(format);
  62   int channels = DetermineChannels(frame);
  63   if (channels <= 0 || channels >= limits::kMaxChannels) {
  64     DLOG(ERROR) << "Requested number of channels (" << channels
  65                 << ") exceeds limit.";
  66     return AVERROR(EINVAL);
  67   }
  68
  69   int bytes_per_channel = SampleFormatToBytesPerChannel(sample_format);
  70   if (frame->nb_samples <= 0)
  71     return AVERROR(EINVAL);
  72
  73   if (s->channels != channels) {
  74     DLOG(ERROR) << "AVCodecContext and AVFrame disagree on channel count.";
  75     return AVERROR(EINVAL);
  76   }
  77
  78   // Determine how big the buffer should be and allocate it. FFmpeg may adjust
  79   // how big each channel data is in order to meet the alignment policy, so
  80   // we need to take this into consideration.
  81   int buffer_size_in_bytes =
  82       av_samples_get_buffer_size(&frame->linesize[0],
  83                                  channels,
  84                                  frame->nb_samples,
  85                                  format,
  86                                  AudioBuffer::kChannelAlignment);
  87   // Check for errors from av_samples_get_buffer_size().
  88   if (buffer_size_in_bytes < 0)
  89     return buffer_size_in_bytes;
  90   int frames_required = buffer_size_in_bytes / bytes_per_channel / channels;
  91   DCHECK_GE(frames_required, frame->nb_samples);
  92   scoped_refptr<AudioBuffer> buffer = AudioBuffer::CreateBuffer(
  93       sample_format,
  94       ChannelLayoutToChromeChannelLayout(s->channel_layout, s->channels),
  95       channels,
  96       s->sample_rate,
  97       frames_required);
  98
  99   // Initialize the data[] and extended_data[] fields to point into the memory
 100   // allocated for AudioBuffer. |number_of_planes| will be 1 for interleaved
 101   // audio and equal to |channels| for planar audio.
 102   int number_of_planes = buffer->channel_data().size();
 103   if (number_of_planes <= AV_NUM_DATA_POINTERS) {
 104     DCHECK_EQ(frame->extended_data, frame->data);
 105     for (int i = 0; i < number_of_planes; ++i)
 106       frame->data[i] = buffer->channel_data()[i];
 107   } else {
 108     // There are more channels than can fit into data[], so allocate
 109     // extended_data[] and fill appropriately.
 110     frame->extended_data = static_cast<uint8**>(
 111         av_malloc(number_of_planes * sizeof(*frame->extended_data)));
 112     int i = 0;
 113     for (; i < AV_NUM_DATA_POINTERS; ++i)
 114       frame->extended_data[i] = frame->data[i] = buffer->channel_data()[i];
 115     for (; i < number_of_planes; ++i)
 116       frame->extended_data[i] = buffer->channel_data()[i];
 117   }
 118
 119   // Now create an AVBufferRef for the data just allocated. It will own the
 120   // reference to the AudioBuffer object.
 121   void* opaque = NULL;
 122   buffer.swap(reinterpret_cast<AudioBuffer**>(&opaque));
 123   frame->buf[0] = av_buffer_create(
 124       frame->data[0], buffer_size_in_bytes, ReleaseAudioBufferImpl, opaque, 0);
 125   return 0;
 126 }
 127
 128 FFmpegAudioDecoder::FFmpegAudioDecoder(
 129     const scoped_refptr<base::SingleThreadTaskRunner>& task_runner,
 130     const scoped_refptr<MediaLog>& media_log)
 131     : task_runner_(task_runner),
 132       state_(kUninitialized),
 133       av_sample_format_(0),
 134       media_log_(media_log) {
 135 }
 136
 137 FFmpegAudioDecoder::~FFmpegAudioDecoder() {
 138   DCHECK(task_runner_->BelongsToCurrentThread());
 139
 140   if (state_ != kUninitialized)
 141     ReleaseFFmpegResources();
 142 }
 143
 144 std::string FFmpegAudioDecoder::GetDisplayName() const {
 145   return "FFmpegAudioDecoder";
 146 }
 147
 148 void FFmpegAudioDecoder::Initialize(const AudioDecoderConfig& config,
 149                                     const InitCB& init_cb,
 150                                     const OutputCB& output_cb) {
 151   DCHECK(task_runner_->BelongsToCurrentThread());
 152   DCHECK(!config.is_encrypted());
 153
 154   FFmpegGlue::InitializeFFmpeg();
 155
 156   config_ = config;
 157   InitCB bound_init_cb = BindToCurrentLoop(init_cb);
 158
 159   if (!config.IsValidConfig() || !ConfigureDecoder()) {
 160     bound_init_cb.Run(false);
 161     return;
 162   }
 163
 164   // Success!
 165   output_cb_ = BindToCurrentLoop(output_cb);
 166   state_ = kNormal;
 167   bound_init_cb.Run(true);
 168 }
 169
 170 void FFmpegAudioDecoder::Decode(const scoped_refptr<DecoderBuffer>& buffer,
 171                                 const DecodeCB& decode_cb) {
 172   DCHECK(task_runner_->BelongsToCurrentThread());
 173   DCHECK(!decode_cb.is_null());
 174   CHECK_NE(state_, kUninitialized);
 175   DecodeCB decode_cb_bound = BindToCurrentLoop(decode_cb);
 176
 177   if (state_ == kError) {
 178     decode_cb_bound.Run(kDecodeError);
 179     return;
 180   }
 181
 182   // Do nothing if decoding has finished.
 183   if (state_ == kDecodeFinished) {
 184     decode_cb_bound.Run(kOk);
 185     return;
 186   }
 187
 188   DecodeBuffer(buffer, decode_cb_bound);
 189 }
 190
 191 void FFmpegAudioDecoder::Reset(const base::Closure& closure) {
 192   DCHECK(task_runner_->BelongsToCurrentThread());
 193
 194   avcodec_flush_buffers(codec_context_.get());
 195   state_ = kNormal;
 196   ResetTimestampState();
 197   task_runner_->PostTask(FROM_HERE, closure);
 198 }
 199
 200 void FFmpegAudioDecoder::DecodeBuffer(
 201     const scoped_refptr<DecoderBuffer>& buffer,
 202     const DecodeCB& decode_cb) {
 203   DCHECK(task_runner_->BelongsToCurrentThread());
 204   DCHECK_NE(state_, kUninitialized);
 205   DCHECK_NE(state_, kDecodeFinished);
 206   DCHECK_NE(state_, kError);
 207   DCHECK(buffer.get());
 208
 209   // Make sure we are notified if http://crbug.com/49709 returns.  Issue also
 210   // occurs with some damaged files.
 211   if (!buffer->end_of_stream() && buffer->timestamp() == kNoTimestamp()) {
 212     DVLOG(1) << "Received a buffer without timestamps!";
 213     decode_cb.Run(kDecodeError);
 214     return;
 215   }
 216
 217   bool has_produced_frame;
 218   do {
 219     has_produced_frame = false;
 220     if (!FFmpegDecode(buffer, &has_produced_frame)) {
 221       state_ = kError;
 222       decode_cb.Run(kDecodeError);
 223       return;
 224     }
 225     // Repeat to flush the decoder after receiving EOS buffer.
 226   } while (buffer->end_of_stream() && has_produced_frame);
 227
 228   if (buffer->end_of_stream())
 229     state_ = kDecodeFinished;
 230
 231   decode_cb.Run(kOk);
 232 }
 233
 234 bool FFmpegAudioDecoder::FFmpegDecode(
 235     const scoped_refptr<DecoderBuffer>& buffer,
 236     bool* has_produced_frame) {
 237   DCHECK(!*has_produced_frame);
 238
 239   AVPacket packet;
 240   av_init_packet(&packet);
 241   if (buffer->end_of_stream()) {
 242     packet.data = NULL;
 243     packet.size = 0;
 244   } else {
 245     packet.data = const_cast<uint8*>(buffer->data());
 246     packet.size = buffer->data_size();
 247   }
 248
 249   // Each audio packet may contain several frames, so we must call the decoder
 250   // until we've exhausted the packet.  Regardless of the packet size we always
 251   // want to hand it to the decoder at least once, otherwise we would end up
 252   // skipping end of stream packets since they have a size of zero.
 253   do {
 254     int frame_decoded = 0;
 255     const int result = avcodec_decode_audio4(
 256         codec_context_.get(), av_frame_.get(), &frame_decoded, &packet);
 257
 258     if (result < 0) {
 259       DCHECK(!buffer->end_of_stream())
 260           << "End of stream buffer produced an error! "
 261           << "This is quite possibly a bug in the audio decoder not handling "
 262           << "end of stream AVPackets correctly.";
 263
 264       MEDIA_LOG(DEBUG, media_log_)
 265           << "Dropping audio frame which failed decode with timestamp: "
 266           << buffer->timestamp().InMicroseconds()
 267           << " us, duration: " << buffer->duration().InMicroseconds()
 268           << " us, packet size: " << buffer->data_size() << " bytes";
 269
 270       break;
 271     }
 272
 273     // Update packet size and data pointer in case we need to call the decoder
 274     // with the remaining bytes from this packet.
 275     packet.size -= result;
 276     packet.data += result;
 277
 278     scoped_refptr<AudioBuffer> output;
 279     const int channels = DetermineChannels(av_frame_.get());
 280     if (frame_decoded) {
 281       if (av_frame_->sample_rate != config_.samples_per_second() ||
 282           channels != ChannelLayoutToChannelCount(config_.channel_layout()) ||
 283           av_frame_->format != av_sample_format_) {
 284         DLOG(ERROR) << "Unsupported midstream configuration change!"
 285                     << " Sample Rate: " << av_frame_->sample_rate << " vs "
 286                     << config_.samples_per_second()
 287                     << ", Channels: " << channels << " vs "
 288                     << ChannelLayoutToChannelCount(config_.channel_layout())
 289                     << ", Sample Format: " << av_frame_->format << " vs "
 290                     << av_sample_format_;
 291
 292         if (config_.codec() == kCodecAAC &&
 293             av_frame_->sample_rate == 2 * config_.samples_per_second()) {
 294           MEDIA_LOG(DEBUG, media_log_)
 295               << "Implicit HE-AAC signalling is being"
 296               << " used. Please use mp4a.40.5 instead of"
 297               << " mp4a.40.2 in the mimetype.";
 298         }
 299         // This is an unrecoverable error, so bail out.
 300         av_frame_unref(av_frame_.get());
 301         return false;
 302       }
 303
 304       // Get the AudioBuffer that the data was decoded into. Adjust the number
 305       // of frames, in case fewer than requested were actually decoded.
 306       output = reinterpret_cast<AudioBuffer*>(
 307           av_buffer_get_opaque(av_frame_->buf[0]));
 308
 309       DCHECK_EQ(ChannelLayoutToChannelCount(config_.channel_layout()),
 310                 output->channel_count());
 311       const int unread_frames = output->frame_count() - av_frame_->nb_samples;
 312       DCHECK_GE(unread_frames, 0);
 313       if (unread_frames > 0)
 314         output->TrimEnd(unread_frames);
 315       av_frame_unref(av_frame_.get());
 316     }
 317
 318     // WARNING: |av_frame_| no longer has valid data at this point.
 319     const int decoded_frames = frame_decoded ? output->frame_count() : 0;
 320     if (IsEndOfStream(result, decoded_frames, buffer)) {
 321       DCHECK_EQ(packet.size, 0);
 322     } else if (discard_helper_->ProcessBuffers(buffer, output)) {
 323       *has_produced_frame = true;
 324       output_cb_.Run(output);
 325     }
 326   } while (packet.size > 0);
 327
 328   return true;
 329 }
 330
 331 void FFmpegAudioDecoder::ReleaseFFmpegResources() {
 332   codec_context_.reset();
 333   av_frame_.reset();
 334 }
 335
 336 bool FFmpegAudioDecoder::ConfigureDecoder() {
 337   if (!config_.IsValidConfig()) {
 338     DLOG(ERROR) << "Invalid audio stream -"
 339                 << " codec: " << config_.codec()
 340                 << " channel layout: " << config_.channel_layout()
 341                 << " bits per channel: " << config_.bits_per_channel()
 342                 << " samples per second: " << config_.samples_per_second();
 343     return false;
 344   }
 345
 346   if (config_.is_encrypted()) {
 347     DLOG(ERROR) << "Encrypted audio stream not supported";
 348     return false;
 349   }
 350
 351   // Release existing decoder resources if necessary.
 352   ReleaseFFmpegResources();
 353
 354   // Initialize AVCodecContext structure.
 355   codec_context_.reset(avcodec_alloc_context3(NULL));
 356   AudioDecoderConfigToAVCodecContext(config_, codec_context_.get());
 357
 358   codec_context_->opaque = this;
 359   codec_context_->get_buffer2 = GetAudioBuffer;
 360   codec_context_->refcounted_frames = 1;
 361
 362   AVCodec* codec = avcodec_find_decoder(codec_context_->codec_id);
 363   if (!codec || avcodec_open2(codec_context_.get(), codec, NULL) < 0) {
 364     DLOG(ERROR) << "Could not initialize audio decoder: "
 365                 << codec_context_->codec_id;
 366     ReleaseFFmpegResources();
 367     state_ = kUninitialized;
 368     return false;
 369   }
 370
 371   // Success!
 372   av_frame_.reset(av_frame_alloc());
 373   av_sample_format_ = codec_context_->sample_fmt;
 374
 375   if (codec_context_->channels !=
 376       ChannelLayoutToChannelCount(config_.channel_layout())) {
 377     DLOG(ERROR) << "Audio configuration specified "
 378                 << ChannelLayoutToChannelCount(config_.channel_layout())
 379                 << " channels, but FFmpeg thinks the file contains "
 380                 << codec_context_->channels << " channels";
 381     ReleaseFFmpegResources();
 382     state_ = kUninitialized;
 383     return false;
 384   }
 385
 386   ResetTimestampState();
 387   return true;
 388 }
 389
 390 void FFmpegAudioDecoder::ResetTimestampState() {
 391   discard_helper_.reset(new AudioDiscardHelper(config_.samples_per_second(),
 392                                                config_.codec_delay()));
 393   discard_helper_->Reset(config_.codec_delay());
 394 }
 395
 396 }  // namespace media