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/sample_format.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 LogCB& log_cb)
 131     : task_runner_(task_runner),
 132       state_(kUninitialized),
 133       av_sample_format_(0),
 134       log_cb_(log_cb) {
 135 }
 136
 137 FFmpegAudioDecoder::~FFmpegAudioDecoder() {
 138   DCHECK(task_runner_->BelongsToCurrentThread());
 139
 140   if (state_ != kUninitialized) {
 141     ReleaseFFmpegResources();
 142     ResetTimestampState();
 143   }
 144 }
 145
 146 std::string FFmpegAudioDecoder::GetDisplayName() const {
 147   return "FFmpegAudioDecoder";
 148 }
 149
 150 void FFmpegAudioDecoder::Initialize(const AudioDecoderConfig& config,
 151                                     const PipelineStatusCB& status_cb,
 152                                     const OutputCB& output_cb) {
 153   DCHECK(task_runner_->BelongsToCurrentThread());
 154   DCHECK(!config.is_encrypted());
 155
 156   FFmpegGlue::InitializeFFmpeg();
 157
 158   config_ = config;
 159   PipelineStatusCB initialize_cb = BindToCurrentLoop(status_cb);
 160
 161   if (!config.IsValidConfig() || !ConfigureDecoder()) {
 162     initialize_cb.Run(DECODER_ERROR_NOT_SUPPORTED);
 163     return;
 164   }
 165
 166   // Success!
 167   output_cb_ = BindToCurrentLoop(output_cb);
 168   state_ = kNormal;
 169   initialize_cb.Run(PIPELINE_OK);
 170 }
 171
 172 void FFmpegAudioDecoder::Decode(const scoped_refptr<DecoderBuffer>& buffer,
 173                                 const DecodeCB& decode_cb) {
 174   DCHECK(task_runner_->BelongsToCurrentThread());
 175   DCHECK(!decode_cb.is_null());
 176   CHECK_NE(state_, kUninitialized);
 177   DecodeCB decode_cb_bound = BindToCurrentLoop(decode_cb);
 178
 179   if (state_ == kError) {
 180     decode_cb_bound.Run(kDecodeError);
 181     return;
 182   }
 183
 184   // Do nothing if decoding has finished.
 185   if (state_ == kDecodeFinished) {
 186     decode_cb_bound.Run(kOk);
 187     return;
 188   }
 189
 190   DecodeBuffer(buffer, decode_cb_bound);
 191 }
 192
 193 void FFmpegAudioDecoder::Reset(const base::Closure& closure) {
 194   DCHECK(task_runner_->BelongsToCurrentThread());
 195
 196   avcodec_flush_buffers(codec_context_.get());
 197   state_ = kNormal;
 198   ResetTimestampState();
 199   task_runner_->PostTask(FROM_HERE, closure);
 200 }
 201
 202 void FFmpegAudioDecoder::DecodeBuffer(
 203     const scoped_refptr<DecoderBuffer>& buffer,
 204     const DecodeCB& decode_cb) {
 205   DCHECK(task_runner_->BelongsToCurrentThread());
 206   DCHECK_NE(state_, kUninitialized);
 207   DCHECK_NE(state_, kDecodeFinished);
 208   DCHECK_NE(state_, kError);
 209   DCHECK(buffer.get());
 210
 211   // Make sure we are notified if http://crbug.com/49709 returns.  Issue also
 212   // occurs with some damaged files.
 213   if (!buffer->end_of_stream() && buffer->timestamp() == kNoTimestamp()) {
 214     DVLOG(1) << "Received a buffer without timestamps!";
 215     decode_cb.Run(kDecodeError);
 216     return;
 217   }
 218
 219   bool has_produced_frame;
 220   do {
 221     has_produced_frame = false;
 222     if (!FFmpegDecode(buffer, &has_produced_frame)) {
 223       state_ = kError;
 224       decode_cb.Run(kDecodeError);
 225       return;
 226     }
 227     // Repeat to flush the decoder after receiving EOS buffer.
 228   } while (buffer->end_of_stream() && has_produced_frame);
 229
 230   if (buffer->end_of_stream())
 231     state_ = kDecodeFinished;
 232
 233   decode_cb.Run(kOk);
 234 }
 235
 236 bool FFmpegAudioDecoder::FFmpegDecode(
 237     const scoped_refptr<DecoderBuffer>& buffer,
 238     bool* has_produced_frame) {
 239   DCHECK(!*has_produced_frame);
 240
 241   AVPacket packet;
 242   av_init_packet(&packet);
 243   if (buffer->end_of_stream()) {
 244     packet.data = NULL;
 245     packet.size = 0;
 246   } else {
 247     packet.data = const_cast<uint8*>(buffer->data());
 248     packet.size = buffer->data_size();
 249   }
 250
 251   // Each audio packet may contain several frames, so we must call the decoder
 252   // until we've exhausted the packet.  Regardless of the packet size we always
 253   // want to hand it to the decoder at least once, otherwise we would end up
 254   // skipping end of stream packets since they have a size of zero.
 255   do {
 256     int frame_decoded = 0;
 257     const int result = avcodec_decode_audio4(
 258         codec_context_.get(), av_frame_.get(), &frame_decoded, &packet);
 259
 260     if (result < 0) {
 261       DCHECK(!buffer->end_of_stream())
 262           << "End of stream buffer produced an error! "
 263           << "This is quite possibly a bug in the audio decoder not handling "
 264           << "end of stream AVPackets correctly.";
 265
 266       MEDIA_LOG(log_cb_)
 267           << "Dropping audio frame which failed decode with timestamp: "
 268           << buffer->timestamp().InMicroseconds() << " us, duration: "
 269           << buffer->duration().InMicroseconds() << " us, packet size: "
 270           << buffer->data_size() << " bytes";
 271
 272       break;
 273     }
 274
 275     // Update packet size and data pointer in case we need to call the decoder
 276     // with the remaining bytes from this packet.
 277     packet.size -= result;
 278     packet.data += result;
 279
 280     scoped_refptr<AudioBuffer> output;
 281     const int channels = DetermineChannels(av_frame_.get());
 282     if (frame_decoded) {
 283       if (av_frame_->sample_rate != config_.samples_per_second() ||
 284           channels != ChannelLayoutToChannelCount(config_.channel_layout()) ||
 285           av_frame_->format != av_sample_format_) {
 286         DLOG(ERROR) << "Unsupported midstream configuration change!"
 287                     << " Sample Rate: " << av_frame_->sample_rate << " vs "
 288                     << config_.samples_per_second()
 289                     << ", Channels: " << channels << " vs "
 290                     << ChannelLayoutToChannelCount(config_.channel_layout())
 291                     << ", Sample Format: " << av_frame_->format << " vs "
 292                     << av_sample_format_;
 293
 294         if (config_.codec() == kCodecAAC &&
 295             av_frame_->sample_rate == 2 * config_.samples_per_second()) {
 296           MEDIA_LOG(log_cb_) << "Implicit HE-AAC signalling is being used."
 297                              << " Please use mp4a.40.5 instead of mp4a.40.2 in"
 298                              << " the mimetype.";
 299         }
 300         // This is an unrecoverable error, so bail out.
 301         av_frame_unref(av_frame_.get());
 302         return false;
 303       }
 304
 305       // Get the AudioBuffer that the data was decoded into. Adjust the number
 306       // of frames, in case fewer than requested were actually decoded.
 307       output = reinterpret_cast<AudioBuffer*>(
 308           av_buffer_get_opaque(av_frame_->buf[0]));
 309
 310       DCHECK_EQ(ChannelLayoutToChannelCount(config_.channel_layout()),
 311                 output->channel_count());
 312       const int unread_frames = output->frame_count() - av_frame_->nb_samples;
 313       DCHECK_GE(unread_frames, 0);
 314       if (unread_frames > 0)
 315         output->TrimEnd(unread_frames);
 316       av_frame_unref(av_frame_.get());
 317     }
 318
 319     // WARNING: |av_frame_| no longer has valid data at this point.
 320     const int decoded_frames = frame_decoded ? output->frame_count() : 0;
 321     if (IsEndOfStream(result, decoded_frames, buffer)) {
 322       DCHECK_EQ(packet.size, 0);
 323     } else if (discard_helper_->ProcessBuffers(buffer, output)) {
 324       *has_produced_frame = true;
 325       output_cb_.Run(output);
 326     }
 327   } while (packet.size > 0);
 328
 329   return true;
 330 }
 331
 332 void FFmpegAudioDecoder::ReleaseFFmpegResources() {
 333   codec_context_.reset();
 334   av_frame_.reset();
 335 }
 336
 337 bool FFmpegAudioDecoder::ConfigureDecoder() {
 338   if (!config_.IsValidConfig()) {
 339     DLOG(ERROR) << "Invalid audio stream -"
 340                 << " codec: " << config_.codec()
 341                 << " channel layout: " << config_.channel_layout()
 342                 << " bits per channel: " << config_.bits_per_channel()
 343                 << " samples per second: " << config_.samples_per_second();
 344     return false;
 345   }
 346
 347   if (config_.is_encrypted()) {
 348     DLOG(ERROR) << "Encrypted audio stream not supported";
 349     return false;
 350   }
 351
 352   // Release existing decoder resources if necessary.
 353   ReleaseFFmpegResources();
 354
 355   // Initialize AVCodecContext structure.
 356   codec_context_.reset(avcodec_alloc_context3(NULL));
 357   AudioDecoderConfigToAVCodecContext(config_, codec_context_.get());
 358
 359   codec_context_->opaque = this;
 360   codec_context_->get_buffer2 = GetAudioBuffer;
 361   codec_context_->refcounted_frames = 1;
 362
 363   AVCodec* codec = avcodec_find_decoder(codec_context_->codec_id);
 364   if (!codec || avcodec_open2(codec_context_.get(), codec, NULL) < 0) {
 365     DLOG(ERROR) << "Could not initialize audio decoder: "
 366                 << codec_context_->codec_id;
 367     ReleaseFFmpegResources();
 368     state_ = kUninitialized;
 369     return false;
 370   }
 371
 372   // Success!
 373   av_frame_.reset(av_frame_alloc());
 374   discard_helper_.reset(new AudioDiscardHelper(config_.samples_per_second(),
 375                                                config_.codec_delay()));
 376   av_sample_format_ = codec_context_->sample_fmt;
 377
 378   if (codec_context_->channels !=
 379       ChannelLayoutToChannelCount(config_.channel_layout())) {
 380     DLOG(ERROR) << "Audio configuration specified "
 381                 << ChannelLayoutToChannelCount(config_.channel_layout())
 382                 << " channels, but FFmpeg thinks the file contains "
 383                 << codec_context_->channels << " channels";
 384     ReleaseFFmpegResources();
 385     state_ = kUninitialized;
 386     return false;
 387   }
 388
 389   ResetTimestampState();
 390   return true;
 391 }
 392
 393 void FFmpegAudioDecoder::ResetTimestampState() {
 394   discard_helper_->Reset(config_.codec_delay());
 395 }
 396
 397 }  // namespace media