Add FFmpeg audio decoder for the clear key CDM.

webkit/media/crypto/ppapi/ffmpeg_cdm_audio_decoder.cc

Index: webkit/media/crypto/ppapi/ffmpeg_cdm_audio_decoder.cc
diff --git a/webkit/media/crypto/ppapi/ffmpeg_cdm_audio_decoder.cc b/webkit/media/crypto/ppapi/ffmpeg_cdm_audio_decoder.cc
new file mode 100644
index 0000000000000000000000000000000000000000..d33861145dbede40157e9761508e355d7fddff89
--- /dev/null
+++ b/webkit/media/crypto/ppapi/ffmpeg_cdm_audio_decoder.cc
@@ -0,0 +1,354 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "webkit/media/crypto/ppapi/ffmpeg_cdm_audio_decoder.h"
+

[xhwang, 2012/10/24 08:18:15]

add #include <algorithm> for std::min

[Tom Finegan, 2012/10/24 21:16:44]

Done.

+#include "base/logging.h"
+#include "media/base/buffers.h"
+#include "media/base/limits.h"
+
+// Include FFmpeg header files.
+extern "C" {
+// Temporarily disable possible loss of data warning.
+MSVC_PUSH_DISABLE_WARNING(4244);
+#include <libavcodec/avcodec.h>
+MSVC_POP_WARNING();
+}  // extern "C"
+
+namespace webkit_media {
+
+// Maximum number of channels with defined order in the Vorbis specification.
+// http://www.xiph.org/vorbis/doc/Vorbis_I_spec.html
+static const int kMaxVorbisChannels = 8;
+
+static CodecID CdmAudioCodecToCodecID(
+    cdm::AudioDecoderConfig::AudioCodec audio_codec) {
+  switch (audio_codec) {
+    case cdm::AudioDecoderConfig::kCodecVorbis:
+      return CODEC_ID_VORBIS;
+    default:
+      NOTREACHED() << "Unsupported cdm::AudioCodec: " << audio_codec;
+  }
+
+  return CODEC_ID_NONE;
+}
+
+static void CdmAudioDecoderConfigToAVCodecContext(
+    const cdm::AudioDecoderConfig& config,
+    AVCodecContext* codec_context) {
+  codec_context->codec_type = AVMEDIA_TYPE_AUDIO;
+  codec_context->codec_id = CdmAudioCodecToCodecID(config.codec);
+
+  switch (config.bits_per_channel) {
+    case 8:
+      codec_context->sample_fmt = AV_SAMPLE_FMT_U8;
+      break;
+    case 16:
+      codec_context->sample_fmt = AV_SAMPLE_FMT_S16;
+      break;
+    case 32:
+      codec_context->sample_fmt = AV_SAMPLE_FMT_S32;
+      break;
+    default:
+      DVLOG(1) << "CdmAudioDecoderConfigToAVCodecContext() Unsupported bits "
+                  "per channel: " << config.bits_per_channel;
+      codec_context->sample_fmt = AV_SAMPLE_FMT_NONE;
+  }
+
+  codec_context->channels = config.channel_count;
+  codec_context->sample_rate = config.samples_per_second;
+
+  if (config.extra_data) {
+    codec_context->extradata_size = config.extra_data_size;
+    codec_context->extradata = reinterpret_cast<uint8_t*>(
+        av_malloc(config.extra_data_size + FF_INPUT_BUFFER_PADDING_SIZE));
+    memcpy(codec_context->extradata, config.extra_data,
+           config.extra_data_size);
+    memset(codec_context->extradata + config.extra_data_size, '\0',
+           FF_INPUT_BUFFER_PADDING_SIZE);
+  } else {
+    codec_context->extradata = NULL;
+    codec_context->extradata_size = 0;
+  }
+}
+
+// Returns true when the decode result was end of stream.
+static inline bool IsEndOfOutputStream(int result,
+                                       int decoded_size,
+                                       bool is_end_of_input_stream) {
+  // Three conditions to meet to declare end of stream for this decoder:
+  // 1. FFmpeg didn't read anything.
+  // 2. FFmpeg didn't output anything.
+  // 3. An end of stream buffer is received.
+  return result == 0 && decoded_size == 0 && is_end_of_input_stream;
+}
+
+FFmpegCdmAudioDecoder::FFmpegCdmAudioDecoder(cdm::Allocator* allocator)
+    : allocator_(allocator),

[xhwang, 2012/10/24 08:18:15]

initialization order should match declaration order.

[Tom Finegan, 2012/10/24 21:16:44]

Done.

+      codec_context_(NULL),
+      av_frame_(NULL),
+      is_initialized_(false),
+      bits_per_channel_(0),
+      samples_per_second_(0),
+      bytes_per_frame_(0),
+      output_timestamp_base_(media::kNoTimestamp()),
+      total_frames_decoded_(0),

[xhwang, 2012/10/24 08:18:15]

If we really need to use double for total_frames_decoded_, we should use 0.0
here.

[Tom Finegan, 2012/10/24 21:16:44]

It's int64_t now.

+      last_input_timestamp_(media::kNoTimestamp()),
+      output_bytes_to_drop_(0) {
+}
+
+FFmpegCdmAudioDecoder::~FFmpegCdmAudioDecoder() {
+  ReleaseFFmpegResources();
+}
+
+bool FFmpegCdmAudioDecoder::Initialize(const cdm::AudioDecoderConfig& config) {
+  DVLOG(1) << "Initialize()";
+
+  if (!IsValidConfig(config)) {
+    LOG(ERROR) << "Initialize(): invalid audio decoder configuration.";
+    return false;
+  }
+
+  if (is_initialized_) {
+    LOG(ERROR) << "Initialize(): Already initialized.";
+    return false;
+  }
+
+  // Release existing resources if necessary.
+  ReleaseFFmpegResources();

[xhwang, 2012/10/24 08:18:15]

In which case do we need this? Are we worrying that Initialize() is called
without de-init? Shouldn't the check on line 113 cover this already?

[Tom Finegan, 2012/10/24 21:16:44]

Done, here and in the video decoder. Copied from media version.

+
+  // Initialize AVCodecContext structure.
+  codec_context_ = avcodec_alloc_context3(NULL);
+  CdmAudioDecoderConfigToAVCodecContext(config, codec_context_);
+  DCHECK_EQ(CODEC_ID_VORBIS, codec_context_->codec_id);
+
+  AVCodec* codec = avcodec_find_decoder(codec_context_->codec_id);
+  if (!codec) {
+    LOG(ERROR) << "Initialize(): avcodec_find_decoder failed.";
+    return false;
+  }
+
+  int status;
+  if ((status = avcodec_open2(codec_context_, codec, NULL)) < 0) {
+    LOG(ERROR) << "Initialize(): avcodec_open2 failed: " << status;
+    return false;
+  }
+
+  av_frame_ = avcodec_alloc_frame();
+  is_initialized_ = true;

[xhwang, 2012/10/24 08:18:15]

Move this down to line 144.

[Tom Finegan, 2012/10/24 21:16:44]

Done.

+  bits_per_channel_ = config.bits_per_channel;
+  samples_per_second_ = config.samples_per_second;
+  bytes_per_frame_ = codec_context_->channels * bits_per_channel_ / 8;
+  serialized_audio_frames_.reserve(bytes_per_frame_ * samples_per_second_);
+
+  return true;
+}
+
+void FFmpegCdmAudioDecoder::Deinitialize() {
+  DVLOG(1) << "Deinitialize()";
+  ReleaseFFmpegResources();
+  is_initialized_ = false;
+  ResetOutputTime();
+}
+
+void FFmpegCdmAudioDecoder::Reset() {
+  DVLOG(1) << "Reset()";
+  avcodec_flush_buffers(codec_context_);
+  ResetOutputTime();
+}
+
+// static
+bool FFmpegCdmAudioDecoder::IsValidConfig(
+    const cdm::AudioDecoderConfig& config) {
+  return config.codec == cdm::AudioDecoderConfig::kCodecVorbis &&
+      config.channel_count > 0 && config.channel_count <= kMaxVorbisChannels &&

[xhwang, 2012/10/24 08:18:15]

split this to two lines to be consistent with the rest. also indent to align
with "config.codec == ... "?

[Tom Finegan, 2012/10/24 21:16:44]

Done.

+      config.bits_per_channel > 0 &&
+      config.bits_per_channel <= media::limits::kMaxBitsPerSample &&
+      config.samples_per_second > 0 &&
+      config.samples_per_second <= media::limits::kMaxSampleRate;
+}
+
+cdm::Status FFmpegCdmAudioDecoder::DecodeBuffer(
+    const uint8_t* compressed_buffer,
+    int32_t compressed_buffer_size,
+    int64_t input_timestamp,
+    cdm::AudioFrames* decoded_frames) {
+  const bool is_end_of_stream = compressed_buffer_size == 0;
+  base::TimeDelta timestamp =
+      base::TimeDelta::FromMicroseconds(input_timestamp);
+  if (!is_end_of_stream) {
+    if (last_input_timestamp_ == media::kNoTimestamp()) {
+      if (codec_context_->codec_id == CODEC_ID_VORBIS &&
+          timestamp < base::TimeDelta()) {
+        // Dropping frames for negative timestamps as outlined in section A.2
+        // in the Vorbis spec. http://xiph.org/vorbis/doc/Vorbis_I_spec.html
+        int frames_to_drop = floor(
+            0.5 + -timestamp.InSecondsF() * samples_per_second_);
+        output_bytes_to_drop_ = bytes_per_frame_ * frames_to_drop;
+      } else {
+        last_input_timestamp_ = timestamp;
+      }
+    } else if (timestamp != media::kNoTimestamp()) {
+      if (timestamp < last_input_timestamp_) {
+        base::TimeDelta diff = timestamp - last_input_timestamp_;
+        DVLOG(1) << "Input timestamps are not monotonically increasing! "
+                 << " ts " << timestamp.InMicroseconds() << " us"
+                 << " diff " << diff.InMicroseconds() << " us";
+        return cdm::kDecodeError;
+      }
+
+      last_input_timestamp_ = timestamp;
+    }
+  }
+
+  AVPacket packet;
+  av_init_packet(&packet);
+  packet.data = const_cast<uint8_t*>(compressed_buffer);
+  packet.size = compressed_buffer_size;
+
+  // Each audio packet may contain several frames, so we must call the decoder
+  // until we've exhausted the packet.  Regardless of the packet size we always
+  // want to hand it to the decoder at least once, otherwise we would end up
+  // skipping end of stream packets since they have a size of zero.
+  do {
+    // Reset frame to default values.
+    avcodec_get_frame_defaults(av_frame_);
+
+    int frame_decoded = 0;
+    int result = avcodec_decode_audio4(
+        codec_context_, av_frame_, &frame_decoded, &packet);
+
+    if (result < 0) {
+      DCHECK(!is_end_of_stream)
+          << "End of stream buffer produced an error! "
+          << "This is quite possibly a bug in the audio decoder not handling "
+          << "end of stream AVPackets correctly.";
+
+      DLOG(ERROR)
+          << "Error decoding an audio frame with timestamp: "
+          << timestamp.InMicroseconds() << " us, duration: "
+          << timestamp.InMicroseconds() << " us, packet size: "
+          << compressed_buffer_size << " bytes";
+
+      // TODO(tomfinegan): Return cdm::kDecodeError here when
+      // http://crbug.com/145276 is fixed.
+      break;

[xhwang, 2012/10/24 08:18:15]

The break here will hit line 315. In that case we are returning kSuccess with a
null buffer. I think that'll hit a DCHECK on the way to media code. How about
just return kDecodeError here? I don't really know the impact of this bug
though.

[Tom Finegan, 2012/10/24 21:16:44]

Done.

+    }
+
+    // Update packet size and data pointer in case we need to call the decoder
+    // with the remaining bytes from this packet.
+    packet.size -= result;
+    packet.data += result;
+
+    if (output_timestamp_base_ == media::kNoTimestamp() && !is_end_of_stream) {
+      DCHECK(timestamp != media::kNoTimestamp());
+      if (output_bytes_to_drop_ > 0) {
+        // If we have to drop samples it always means the timeline starts at 0.
+        output_timestamp_base_ = base::TimeDelta();
+      } else {
+        output_timestamp_base_ = timestamp;
+      }
+    }
+
+    const uint8* decoded_audio_data = NULL;

[xhwang, 2012/10/24 08:18:15]

we are usng uint8_t in other places in this file.

[Tom Finegan, 2012/10/24 21:16:44]

_t added; missed this one after the copy from media.

+    int decoded_audio_size = 0;
+    if (frame_decoded) {
+      int output_sample_rate = av_frame_->sample_rate;
+      if (output_sample_rate != samples_per_second_) {
+        DLOG(ERROR) << "Output sample rate (" << output_sample_rate
+                    << ") doesn't match expected rate " << samples_per_second_;
+        return cdm::kDecodeError;
+      }
+
+      decoded_audio_data = av_frame_->data[0];
+      decoded_audio_size = av_samples_get_buffer_size(
+          NULL, codec_context_->channels, av_frame_->nb_samples,
+          codec_context_->sample_fmt, 1);
+    }
+
+    if (decoded_audio_size > 0 && output_bytes_to_drop_ > 0) {
+      int dropped_size = std::min(decoded_audio_size, output_bytes_to_drop_);
+      decoded_audio_data += dropped_size;
+      decoded_audio_size -= dropped_size;
+      output_bytes_to_drop_ -= dropped_size;
+    }
+
+    if (decoded_audio_size > 0) {
+      DCHECK_EQ(decoded_audio_size % bytes_per_frame_, 0)
+          << "Decoder didn't output full frames";
+
+      base::TimeDelta output_timestamp = GetNextOutputTimestamp();

[xhwang, 2012/10/24 08:18:15]

hmm, ISTM that this timestamp is the time of the last decoded frame, not the
first one. I know this is from media code. I'll double check tomorrow morning.

[xhwang, 2012/10/24 16:33:21]

Okay, now I understand how it works ;) Could you please move line 291 to 282?
These two lines are closely related.

[Tom Finegan, 2012/10/24 21:16:44]

Done.

+
+      // Serialize the audio samples into |serialized_audio_frames_|.
+      SerializeInt64(output_timestamp.InMicroseconds());
+      SerializeInt64(decoded_audio_size);
+      serialized_audio_frames_.insert(
+          serialized_audio_frames_.end(),
+          decoded_audio_data,
+          decoded_audio_data + decoded_audio_size);
+
+      total_frames_decoded_ += decoded_audio_size / bytes_per_frame_;
+    } else if (IsEndOfOutputStream(result,
+                                   decoded_audio_size,
+                                   is_end_of_stream)) {

[xhwang, 2012/10/24 08:18:15]

I am overwhelmed by the logic in this function (I know it's from media code, so
I am not complaining :)) 

Let's look at the logic here. IsEndOfOutputStream() can be true only when
is_end_of_stream is true. But in this case, packet.size must be zero. So this
do/while loop can only run once. Since we have no data, "result" must be zero
here (negative result has triggered break on line 236).  So what really matters
is the decoded_audio_size. It must be zero here since we are at the "else if"
for "if (decoded_audio_size > 0)". So the condition here is really equivalent to
"else if (is_end_of_stream)".

Anyway, in this case, we should return kNeedMoreData instead of serialize an end
of stream buffer. (see below comment).

[Tom Finegan, 2012/10/24 21:16:44]

Done (returning kNeedMoreData). Changed nothing else. Should I change the
conditions?

+      DCHECK_EQ(packet.size, 0);
+      // Serialize an end of stream buffer.
+      SerializeInt64(0);
+      SerializeInt64(0);

[xhwang, 2012/10/24 08:18:15]

The new contract is: suppose when the input is EOS, we can still produce N
frames. Then we return N frames with kSuccess as normal. Then in the next call
(the input must be EOS again), we return kNeedMoreData.

[Tom Finegan, 2012/10/24 21:16:44]

Done.

+    }
+  } while (packet.size > 0);
+
+  if (serialized_audio_frames_.size() > 0) {
+    decoded_frames->set_buffer(
+        allocator_->Allocate(serialized_audio_frames_.size()));
+    if (!decoded_frames->buffer()) {
+      LOG(ERROR) << "DecodeBuffer() cdm::Allocator::Allocate failed.";
+      return cdm::kDecodeError;
+    }
+    memcpy(decoded_frames->buffer()->data(),
+           &serialized_audio_frames_[0],
+           serialized_audio_frames_.size());
+    serialized_audio_frames_.clear();
+  }
+
+  return cdm::kSuccess;
+}
+
+void FFmpegCdmAudioDecoder::ResetOutputTime() {

[xhwang, 2012/10/24 08:18:15]

This function name is a little misleading as we are also setting
last_input_timestamp_.

[Tom Finegan, 2012/10/24 21:16:44]

Renamed to ResetAudioTimingData().

+  output_timestamp_base_ = media::kNoTimestamp();
+  total_frames_decoded_ = 0;
+  last_input_timestamp_ = media::kNoTimestamp();
+  output_bytes_to_drop_ = 0;
+}
+
+void FFmpegCdmAudioDecoder::ReleaseFFmpegResources() {
+  DVLOG(1) << "ReleaseFFmpegResources()";
+
+  if (codec_context_) {
+    av_free(codec_context_->extradata);
+    avcodec_close(codec_context_);
+    av_free(codec_context_);
+    codec_context_ = NULL;
+  }
+  if (av_frame_) {
+    av_free(av_frame_);
+    av_frame_ = NULL;
+  }
+}
+
+base::TimeDelta FFmpegCdmAudioDecoder::GetNextOutputTimestamp() const {
+  DCHECK(output_timestamp_base_ != media::kNoTimestamp());
+  double decoded_us = (total_frames_decoded_ / samples_per_second_) *
+      base::Time::kMicrosecondsPerSecond;
+  return output_timestamp_base_ +
+      base::TimeDelta::FromMicroseconds(decoded_us);
+}
+
+void FFmpegCdmAudioDecoder::SerializeInt64(int64 value) {
+  const uint8_t* ptr = reinterpret_cast<uint8_t*>(&value);
+  serialized_audio_frames_.insert(serialized_audio_frames_.end(),
+                                  ptr, ptr + sizeof(value));
+}
+
+}  // namespace webkit_media