Integrate VDA with WebRTC.

content/renderer/media/rtc_video_decoder.cc

Index: content/renderer/media/rtc_video_decoder.cc
diff --git a/content/renderer/media/rtc_video_decoder.cc b/content/renderer/media/rtc_video_decoder.cc
new file mode 100644
index 0000000000000000000000000000000000000000..3200347d1dc78efb4e378b8bd13cb601bf231e25
--- /dev/null
+++ b/content/renderer/media/rtc_video_decoder.cc
@@ -0,0 +1,685 @@
+// Copyright (c) 2013 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 "content/renderer/media/rtc_video_decoder.h"
+
+#include "base/bind.h"
+#include "base/logging.h"
+#include "base/memory/ref_counted.h"
+#include "base/message_loop_proxy.h"
+#include "base/safe_numerics.h"
+#include "base/task_runner_util.h"
+#include "content/child/child_thread.h"
+#include "content/renderer/media/native_handle_impl.h"
+#include "media/base/bind_to_loop.h"
+#include "third_party/webrtc/system_wrappers/interface/ref_count.h"
+
+namespace content {
+
+// A shared memory segment and its allocated size. |shm| is unowned and the
+// users should delete it.
+struct RTCVideoDecoder::SHMBuffer {
+  SHMBuffer(base::SharedMemory* shm, size_t size);
+  ~SHMBuffer();
+  base::SharedMemory* const shm;
+  const size_t size;
+};
+
+RTCVideoDecoder::SHMBuffer::SHMBuffer(base::SharedMemory* shm, size_t size)
+    : shm(shm), size(size) {}
+
+RTCVideoDecoder::SHMBuffer::~SHMBuffer() {}
+
+// Metadata of a bitstream buffer.
+struct RTCVideoDecoder::BufferData {
+  BufferData(int32 bitstream_buffer_id,
+             uint32_t timestamp,
+             int width,
+             int height,
+             size_t sisze);

[Ami GONE FROM CHROMIUM, 2013/06/26 00:11:58]

typo: sisze

[wuchengli, 2013/06/28 15:08:44]

Done.

+  ~BufferData();
+  int32 bitstream_buffer_id;
+  uint32_t timestamp;  // in 90KHz
+  uint32_t width;
+  uint32_t height;
+  size_t size;  // buffer size
+};
+
+RTCVideoDecoder::BufferData::BufferData(int32 bitstream_buffer_id,
+                                        uint32_t timestamp,
+                                        int width,
+                                        int height,
+                                        size_t size)
+    : bitstream_buffer_id(bitstream_buffer_id),
+      timestamp(timestamp),
+      width(width),
+      height(height),
+      size(size) {}
+
+RTCVideoDecoder::BufferData::~BufferData() {}
+
+RTCVideoDecoder::RTCVideoDecoder(
+    const scoped_refptr<media::GpuVideoDecoder::Factories>& factories)
+    : state_(UNINITIALIZED),
+      decode_complete_callback_(NULL),
+      weak_factory_(this),
+      factories_(factories),
+      vda_loop_proxy_(factories_->GetMessageLoop()),
+      decoder_texture_target_(0),
+      next_picture_buffer_id_(0),
+      next_bitstream_buffer_id_(0),
+      reset_bitstream_buffer_id_(0) {
+  // Initialize directly if |vda_loop_proxy_| is the renderer thread.
+  base::WaitableEvent compositor_loop_async_waiter(false, false);
+  if (vda_loop_proxy_->BelongsToCurrentThread()) {
+    Initialize(&compositor_loop_async_waiter);
+    return;
+  }
+  // Post the task if |vda_loop_proxy_| is the compositor thread. Waiting here
+  // is safe because the compositor thread will not be stopped until the
+  // renderer thread shuts down.
+  vda_loop_proxy_->PostTask(FROM_HERE,
+                            base::Bind(&RTCVideoDecoder::Initialize,
+                                       base::Unretained(this),
+                                       &compositor_loop_async_waiter));
+  compositor_loop_async_waiter.Wait();
+}
+
+RTCVideoDecoder::~RTCVideoDecoder() {
+  DVLOG(2) << "~RTCVideoDecoder";
+  // Delete vda and remove |this| from the observer if vda thread is alive.
+  if (vda_loop_proxy_->BelongsToCurrentThread()) {
+    base::MessageLoop::current()->RemoveDestructionObserver(this);
+    DestroyVDA();
+  } else {
+    // VDA should have been destroyed in WillDestroyCurrentMessageLoop.
+    DCHECK(!vda_);
+  }
+
+  // Delete all shared memories.
+  for (size_t i = 0; i < available_shm_segments_.size(); ++i) {
+    available_shm_segments_[i]->shm->Close();
+    delete available_shm_segments_[i];
+  }
+  available_shm_segments_.clear();
+  for (std::map<int32, SHMBuffer*>::iterator it =
+           bitstream_buffers_in_decoder_.begin();
+       it != bitstream_buffers_in_decoder_.end();
+       ++it) {
+    it->second->shm->Close();
+    delete it->second;
+  }
+  bitstream_buffers_in_decoder_.clear();
+  for (std::deque<std::pair<SHMBuffer*, BufferData> >::iterator it =
+           buffers_to_be_decoded_.begin();
+       it != buffers_to_be_decoded_.end();
+       ++it) {
+    it->first->shm->Close();
+    delete it->first;
+  }
+  buffers_to_be_decoded_.clear();
+
+  // Delete WebRTC input buffers.
+  for (std::deque<std::pair<webrtc::EncodedImage, BufferData> >::iterator it =
+           webrtc_buffers_.begin();
+       it != webrtc_buffers_.end();
+       ++it) {
+    delete it->first._buffer;
+  }
+}
+
+scoped_ptr<RTCVideoDecoder> RTCVideoDecoder::Create(
+    const scoped_refptr<media::GpuVideoDecoder::Factories>& factories) {
+  scoped_ptr<RTCVideoDecoder> decoder(new RTCVideoDecoder(factories));
+  decoder->vda_.reset(factories->CreateVideoDecodeAccelerator(
+      media::VP8PROFILE_MAIN, decoder.get()));
+  // vda can be NULL if VP8 is not supported.
+  if (decoder->vda_ != NULL) {
+    decoder->state_ = INITIALIZED;
+  } else {
+    factories->GetMessageLoop()->DeleteSoon(FROM_HERE, decoder.release());
+  }
+  return decoder.Pass();
+}
+
+int32_t RTCVideoDecoder::InitDecode(const webrtc::VideoCodec* codecSettings,
+                                    int32_t /*numberOfCores*/) {
+  DVLOG(2) << "InitDecode";
+  DCHECK_EQ(codecSettings->codecType, webrtc::kVideoCodecVP8);
+  if (codecSettings->codecSpecific.VP8.feedbackModeOn) {
+    LOG(ERROR) << "Feedback mode not supported";
+    return WEBRTC_VIDEO_CODEC_ERROR;
+  }
+
+  base::AutoLock auto_lock(lock_);
+  if (state_ == UNINITIALIZED || state_ == DECODE_ERROR) {
+    LOG(ERROR) << "VDA is not initialized. state=" << state_;
+    return WEBRTC_VIDEO_CODEC_UNINITIALIZED;
+  }
+  return WEBRTC_VIDEO_CODEC_OK;
+}
+
+int32_t RTCVideoDecoder::Decode(
+    const webrtc::EncodedImage& inputImage,
+    bool missingFrames,
+    const webrtc::RTPFragmentationHeader* /*fragmentation*/,
+    const webrtc::CodecSpecificInfo* /*codecSpecificInfo*/,
+    int64_t /*renderTimeMs*/) {
+  DVLOG(3) << "Decode";
+
+  int bitstream_buffer_id;
+  {
+    base::AutoLock auto_lock(lock_);
+    if (state_ == UNINITIALIZED || decode_complete_callback_ == NULL) {
+      LOG(ERROR) << "The decoder has not initialized.";
+      return WEBRTC_VIDEO_CODEC_UNINITIALIZED;
+    }
+    if (state_ == DECODE_ERROR) {
+      LOG(ERROR) << "Decoding error occurred.";
+      return WEBRTC_VIDEO_CODEC_ERROR;
+    }
+    bitstream_buffer_id = next_bitstream_buffer_id_;
+    // Mask against 30 bits, to avoid (undefined) wraparound on signed integer.
+    next_bitstream_buffer_id_ = (next_bitstream_buffer_id_ + 1) & 0x3FFFFFFF;
+  }
+  if (missingFrames || !inputImage._completeFrame) {
+    DLOG(ERROR) << "Missing or incomplete frames.";
+    // Unlike the SW decoder in libvpx, hw decoder cannot handle broken frames.
+    // Return an error to request a key frame.
+    return WEBRTC_VIDEO_CODEC_ERROR;
+  }
+  if (inputImage._frameType == webrtc::kKeyFrame)
+    frame_size_.SetSize(inputImage._encodedWidth, inputImage._encodedHeight);
+
+  // Create buffer metadata.
+  BufferData buffer_data(bitstream_buffer_id,
+                         inputImage._timeStamp,
+                         frame_size_.width(),
+                         frame_size_.height(),
+                         inputImage._length);
+
+  SendPendingBuffersForDecode();

[Ami GONE FROM CHROMIUM, 2013/06/26 00:11:58]

This is a strange place to place this call.  What's the thinking here?

[wuchengli, 2013/06/28 15:08:44]

The buffers should be sent to VDA in order. If any pending buffer is not sent,
the current buffer in |inputImage| should be appended to the pending queue.

+
+  // If the shared memory is available and there are no pending buffers, send
+  // the buffer for decode. If not, save the buffer in the queue for decode
+  // later.
+  SHMBuffer* shm_buffer = NULL;
+  if (webrtc_buffers_.size() == 0)
+    shm_buffer = GetSHM(inputImage._length);
+  if (shm_buffer != NULL)
+    SendBufferForDecode(inputImage, shm_buffer, buffer_data);
+  else
+    SaveToPendingBuffers(inputImage, buffer_data);
+
+  return WEBRTC_VIDEO_CODEC_OK;
+}
+
+int32_t RTCVideoDecoder::RegisterDecodeCompleteCallback(
+    webrtc::DecodedImageCallback* callback) {
+  DVLOG(2) << "RegisterDecodeCompleteCallback";
+  base::AutoLock auto_lock(lock_);
+  decode_complete_callback_ = callback;
+  return WEBRTC_VIDEO_CODEC_OK;
+}
+
+int32_t RTCVideoDecoder::Release() {
+  DVLOG(2) << "Release";
+  // Do not destroy VDA because the decoder will be recycled by
+  // RTCVideoDecoderFactory. Just reset VDA.
+  return Reset();
+}
+
+int32_t RTCVideoDecoder::Reset() {
+  DVLOG(2) << "Reset";
+  base::AutoLock auto_lock(lock_);
+  if (state_ == UNINITIALIZED) {
+    LOG(ERROR) << "Decoder not initialized.";
+    return WEBRTC_VIDEO_CODEC_UNINITIALIZED;
+  }
+  reset_bitstream_buffer_id_ = next_bitstream_buffer_id_;
+  // If VDA is already resetting, no need to request the reset again.
+  if (state_ != RESETTING) {
+    state_ = RESETTING;
+    vda_loop_proxy_->PostTask(
+        FROM_HERE, base::Bind(&RTCVideoDecoder::ResetInternal, weak_this_));
+  }
+  return WEBRTC_VIDEO_CODEC_OK;
+}
+
+void RTCVideoDecoder::NotifyInitializeDone() {
+  DVLOG(2) << "NotifyInitializeDone";
+  NOTREACHED();
+}
+
+void RTCVideoDecoder::ProvidePictureBuffers(uint32 count,
+                                            const gfx::Size& size,
+                                            uint32 texture_target) {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  DVLOG(3) << "ProvidePictureBuffers. texture_target=" << texture_target;
+
+  if (!vda_)
+    return;
+
+  std::vector<uint32> texture_ids;
+  decoder_texture_target_ = texture_target;
+  if (!factories_->CreateTextures(
+          count, size, &texture_ids, decoder_texture_target_)) {
+    NotifyError(media::VideoDecodeAccelerator::PLATFORM_FAILURE);
+    return;
+  }
+  DCHECK_EQ(count, texture_ids.size());
+
+  std::vector<media::PictureBuffer> picture_buffers;
+  for (size_t i = 0; i < texture_ids.size(); ++i) {
+    picture_buffers.push_back(
+        media::PictureBuffer(next_picture_buffer_id_++, size, texture_ids[i]));
+    bool inserted = assigned_picture_buffers_.insert(std::make_pair(
+        picture_buffers.back().id(), picture_buffers.back())).second;
+    DCHECK(inserted);
+  }
+  vda_->AssignPictureBuffers(picture_buffers);
+}
+
+void RTCVideoDecoder::DismissPictureBuffer(int32 id) {
+  DVLOG(3) << "DismissPictureBuffer. id=" << id;
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+
+  std::map<int32, media::PictureBuffer>::iterator it =
+      assigned_picture_buffers_.find(id);
+  if (it == assigned_picture_buffers_.end()) {
+    NOTREACHED() << "Missing picture buffer: " << id;
+    return;
+  }
+
+  media::PictureBuffer buffer_to_dismiss = it->second;
+  assigned_picture_buffers_.erase(it);
+
+  std::set<int32>::iterator at_display_it =
+      picture_buffers_at_display_.find(id);
+
+  if (at_display_it == picture_buffers_at_display_.end()) {
+    // We can delete the texture immediately as it's not being displayed.
+    factories_->DeleteTexture(buffer_to_dismiss.texture_id());
+  } else {
+    // Texture in display. Postpone deletion until after it's returned to us.
+    bool inserted = dismissed_picture_buffers_
+        .insert(std::make_pair(id, buffer_to_dismiss)).second;
+    DCHECK(inserted);
+  }
+}
+
+void RTCVideoDecoder::PictureReady(const media::Picture& picture) {
+  DVLOG(3) << "PictureReady";
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+
+  std::map<int32, media::PictureBuffer>::iterator it =
+      assigned_picture_buffers_.find(picture.picture_buffer_id());
+  if (it == assigned_picture_buffers_.end()) {
+    NOTREACHED() << "Missing picture buffer: " << picture.picture_buffer_id();
+    NotifyError(media::VideoDecodeAccelerator::PLATFORM_FAILURE);
+    return;
+  }
+  const media::PictureBuffer& pb = it->second;
+
+  // Create a media::VideoFrame.
+  uint32_t timestamp = 0, width = 0, height = 0;
+  size_t size = 0;
+  GetBufferData(
+      picture.bitstream_buffer_id(), &timestamp, &width, &height, &size);
+  scoped_refptr<media::VideoFrame> frame =
+      CreateVideoFrame(picture, pb, timestamp, width, height, size);
+  bool inserted =
+      picture_buffers_at_display_.insert(picture.picture_buffer_id()).second;
+  DCHECK(inserted);
+  {
+    // WebRTC expects no frame callback after Release. Drop the frame if VDA is

[Ami GONE FROM CHROMIUM, 2013/06/26 00:11:58]

Given the frame callback is run on a thread other than the one on which the
Release is called, and no lock is held across both invocations, how can this be
guaranteed?

IOW, state_ may not be RESETTING here but then be set to RESETTING between the
release of the lock at l.341 and the call to Decoded() at l.360.

I suspect you need to hold the state lock here while calling the callback.

[wuchengli, 2013/06/28 15:08:44]

You are right. I need to hold the lock while calling the callback. The callback
calls into WebRTC. We cannot return from Release while the callback is running.

+    // resetting.
+    base::AutoLock auto_lock(lock_);
+    if (state_ == RESETTING)
+      return;
+  }
+
+  // Create a webrtc::I420VideoFrame.
+  webrtc::I420VideoFrame decoded_image;
+  // TODO(wuchengli): remove the malloc.
+  decoded_image.CreateEmptyFrame(width, height, width, height / 2, width / 2);
+  webrtc::RefCountImpl<NativeHandleImpl>* handle =
+      new webrtc::RefCountImpl<NativeHandleImpl>();
+  handle->SetHandle(frame.get());
+  decoded_image.set_native_handle(handle);
+  decoded_image.set_timestamp(timestamp);
+
+  // Send to decode callback.
+  webrtc::DecodedImageCallback* callback;
+  {
+    base::AutoLock auto_lock(lock_);
+    callback = decode_complete_callback_;
+  }
+  DCHECK(callback != NULL);
+  callback->Decoded(decoded_image);
+}
+
+scoped_refptr<media::VideoFrame> RTCVideoDecoder::CreateVideoFrame(
+    const media::Picture& picture,
+    const media::PictureBuffer& pb,
+    uint32_t timestamp,
+    uint32_t width,
+    uint32_t height,
+    size_t size) {
+  gfx::Rect visible_rect(width, height);
+  gfx::Size natural_size(width, height);
+  DCHECK(decoder_texture_target_);
+  // Convert timestamp from 90KHz to ms.
+  base::TimeDelta timestamp_ms = base::TimeDelta::FromInternalValue(
+      base::checked_numeric_cast<uint64_t>(timestamp) * 1000 / 90);
+  return media::VideoFrame::WrapNativeTexture(
+      pb.texture_id(),
+      decoder_texture_target_,
+      pb.size(),
+      visible_rect,
+      natural_size,
+      timestamp_ms,
+      base::Bind(&media::GpuVideoDecoder::Factories::ReadPixels,
+                 factories_,
+                 pb.texture_id(),
+                 decoder_texture_target_,
+                 natural_size),
+      media::BindToCurrentLoop(base::Bind(&RTCVideoDecoder::ReusePictureBuffer,
+                                          weak_this_,
+                                          picture.picture_buffer_id())));
+}
+
+void RTCVideoDecoder::NotifyEndOfBitstreamBuffer(int32 id) {
+  DVLOG(3) << "NotifyEndOfBitstreamBuffer. id=" << id;
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+
+  std::map<int32, SHMBuffer*>::iterator it =
+      bitstream_buffers_in_decoder_.find(id);
+  if (it == bitstream_buffers_in_decoder_.end()) {
+    NotifyError(media::VideoDecodeAccelerator::PLATFORM_FAILURE);
+    NOTREACHED() << "Missing bitstream buffer: " << id;
+    return;
+  }
+
+  PutSHM(it->second);
+  bitstream_buffers_in_decoder_.erase(it);
+
+  RequestBufferDecode();
+}
+
+void RTCVideoDecoder::NotifyFlushDone() {
+  DVLOG(3) << "NotifyFlushDone";
+  NOTREACHED() << "Unexpected flush done notification.";
+}
+
+void RTCVideoDecoder::NotifyResetDone() {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  DVLOG(3) << "NotifyResetDone";
+
+  if (!vda_)
+    return;
+
+  input_buffer_data_.clear();
+  {
+    base::AutoLock auto_lock(lock_);
+    state_ = INITIALIZED;
+  }
+  // Send the pending buffers for decoding.
+  RequestBufferDecode();
+}
+
+void RTCVideoDecoder::NotifyError(media::VideoDecodeAccelerator::Error error) {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  if (!vda_)
+    return;
+
+  DLOG(ERROR) << "VDA Error:" << error;
+  DestroyVDA();
+
+  base::AutoLock auto_lock(lock_);
+  state_ = DECODE_ERROR;
+}
+
+void RTCVideoDecoder::WillDestroyCurrentMessageLoop() {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  factories_->Abort();
+  weak_factory_.InvalidateWeakPtrs();
+  DestroyVDA();
+}
+
+void RTCVideoDecoder::Initialize(base::WaitableEvent* waiter) {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  base::MessageLoop::current()->AddDestructionObserver(this);
+  weak_this_ = weak_factory_.GetWeakPtr();
+  waiter->Signal();
+}
+
+void RTCVideoDecoder::RequestBufferDecode() {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  if (!vda_)
+    return;
+
+  while (CanMoreDecodeWorkBeDone()) {
+    // Get a buffer and data from the queue.
+    std::pair<SHMBuffer*, BufferData>* buffer_pair;
+    SHMBuffer* shm_buffer = NULL;
+    BufferData* buffer_data = NULL;
+    {
+      base::AutoLock auto_lock(lock_);
+      // Do not request decode if VDA is resetting.
+      if (buffers_to_be_decoded_.size() == 0 || state_ == RESETTING)
+        return;
+      buffer_pair = &buffers_to_be_decoded_.front();
+      buffers_to_be_decoded_.pop_front();
+      shm_buffer = buffer_pair->first;
+      buffer_data = &buffer_pair->second;
+      // Drop the buffers before Reset.
+      if (buffer_data->bitstream_buffer_id < reset_bitstream_buffer_id_) {

[Ami GONE FROM CHROMIUM, 2013/06/26 00:11:58]

This doesn't handle the wraparound you have at 30bits.
You want something akin to
https://code.google.com/p/chromium/codesearch#chromium/src/third_party/libjin...

[wuchengli, 2013/06/28 15:08:44]

Done. Good catch...

+        available_shm_segments_.push_back(shm_buffer);
+        continue;
+      }
+    }
+
+    // Create a BitstreamBuffer and send to VDA to decode.
+    media::BitstreamBuffer bitstream_buffer(buffer_data->bitstream_buffer_id,
+                                            shm_buffer->shm->handle(),
+                                            buffer_data->size);
+    bool inserted = bitstream_buffers_in_decoder_
+        .insert(std::make_pair(bitstream_buffer.id(), shm_buffer)).second;
+    DCHECK(inserted);
+    RecordBufferData(*buffer_data);
+    vda_->Decode(bitstream_buffer);
+  }
+}
+
+// Maximum number of concurrent VDA::Decode() operations RVD will maintain.
+// Higher values allow better pipelining in the GPU, but also require more
+// resources.
+enum {
+  kMaxInFlightDecodes = 8
+};
+
+bool RTCVideoDecoder::CanMoreDecodeWorkBeDone() {
+  return bitstream_buffers_in_decoder_.size() < kMaxInFlightDecodes;
+}
+
+void RTCVideoDecoder::SendBufferForDecode(
+    const webrtc::EncodedImage& inputImage,
+    SHMBuffer* shm_buffer,
+    const BufferData& buffer_data) {
+  memcpy(shm_buffer->shm->memory(), inputImage._buffer, inputImage._length);
+  std::pair<SHMBuffer*, BufferData> buffer_pair =
+      std::make_pair(shm_buffer, buffer_data);
+
+  // Store the buffer and the metadata to the queue.
+  {
+    base::AutoLock auto_lock(lock_);
+    buffers_to_be_decoded_.push_back(buffer_pair);
+  }
+  vda_loop_proxy_->PostTask(
+      FROM_HERE, base::Bind(&RTCVideoDecoder::RequestBufferDecode, weak_this_));
+}
+
+void RTCVideoDecoder::SendPendingBuffersForDecode() {
+  while (webrtc_buffers_.size() > 0) {
+    // Get a WebRTC buffer from the queue and a shared memory.
+    std::pair<webrtc::EncodedImage, BufferData>* buffer_pair =
+        &webrtc_buffers_.front();

[Ami GONE FROM CHROMIUM, 2013/06/26 00:11:58]

why * instead of const&?

[wuchengli, 2013/06/28 15:08:44]

No special reason. Is there a preference to use const& for local variables in a
function?

[Ami GONE FROM CHROMIUM, 2013/06/28 17:04:00]

Mainly I think what I look for is less ambiguity about mutability.  So it's the
const I look for over the &-vs-*, but once you have a const* that can never be
NULL the usual practice is to use const&.
(because in chromium code the only const*'s you see are things that could have
been const& except they also need to handle NULL :))

+    webrtc::EncodedImage* input_image = &buffer_pair->first;
+    BufferData* buffer_data = &buffer_pair->second;
+    SHMBuffer* shm_buffer = GetSHM(input_image->_length);
+    if (!shm_buffer)
+      return;
+    webrtc_buffers_.pop_front();
+
+    SendBufferForDecode(*input_image, shm_buffer, *buffer_data);
+    delete input_image->_buffer;
+  }
+}
+
+void RTCVideoDecoder::SaveToPendingBuffers(
+    const webrtc::EncodedImage& inputImage,
+    const BufferData& buffer_data) {
+  // Post to the child thread to create shared memory.
+  content::ChildThread::current()->message_loop()->PostTask(
+      FROM_HERE,
+      base::Bind(&RTCVideoDecoder::CreateSHM, weak_this_, inputImage._length));
+
+  // Clone the input image and save it to the queue.

[Ami GONE FROM CHROMIUM, 2013/06/26 00:11:58]

Is it unsafe to simply swap _buffer pointers?

[wuchengli, 2013/06/28 15:08:44]

This should be OK because WebRTC will allocate a new buffer. Unfortunately
inputImage is passed as const reference in Decode().

+  uint8_t* buffer = (uint8_t*)malloc(inputImage._length);
+  memcpy(buffer, inputImage._buffer, inputImage._length);
+  webrtc::EncodedImage encoded_image(
+      buffer, inputImage._length, inputImage._length);
+  std::pair<webrtc::EncodedImage, BufferData> buffer_pair =
+      std::make_pair(encoded_image, buffer_data);
+  webrtc_buffers_.push_back(buffer_pair);
+}
+
+void RTCVideoDecoder::ResetInternal() {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  DVLOG(2) << "ResetInternal";
+  if (vda_)
+    vda_->Reset();
+}
+
+void RTCVideoDecoder::ReusePictureBuffer(int64 picture_buffer_id) {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  DVLOG(3) << "ReusePictureBuffer. id=" << picture_buffer_id;
+
+  if (!vda_)
+    return;
+
+  CHECK(!picture_buffers_at_display_.empty());
+
+  size_t num_erased = picture_buffers_at_display_.erase(picture_buffer_id);
+  DCHECK(num_erased);
+
+  std::map<int32, media::PictureBuffer>::iterator it =
+      assigned_picture_buffers_.find(picture_buffer_id);
+
+  if (it == assigned_picture_buffers_.end()) {
+    // This picture was dismissed while in display, so we postponed deletion.
+    it = dismissed_picture_buffers_.find(picture_buffer_id);
+    DCHECK(it != dismissed_picture_buffers_.end());
+    factories_->DeleteTexture(it->second.texture_id());
+    dismissed_picture_buffers_.erase(it);
+    return;
+  }
+
+  vda_->ReusePictureBuffer(picture_buffer_id);
+}
+
+void RTCVideoDecoder::DestroyTextures() {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  std::map<int32, media::PictureBuffer>::iterator it;
+
+  for (it = assigned_picture_buffers_.begin();
+       it != assigned_picture_buffers_.end();
+       ++it) {
+    factories_->DeleteTexture(it->second.texture_id());
+  }
+  assigned_picture_buffers_.clear();
+
+  for (it = dismissed_picture_buffers_.begin();
+       it != dismissed_picture_buffers_.end();
+       ++it) {
+    factories_->DeleteTexture(it->second.texture_id());
+  }
+  dismissed_picture_buffers_.clear();
+}
+
+void RTCVideoDecoder::DestroyVDA() {
+  DVLOG(2) << "DestroyVDA";
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  if (vda_)
+    vda_.release()->Destroy();
+  DestroyTextures();
+}
+
+// Size of shared-memory segments we allocate.  Since we reuse them we let them
+// be on the beefy side.
+static const size_t kSharedMemorySegmentBytes = 100 << 10;
+
+RTCVideoDecoder::SHMBuffer* RTCVideoDecoder::GetSHM(size_t min_size) {
+  // Reuse a SHM if possible.
+  base::AutoLock auto_lock(lock_);
+  if (!available_shm_segments_.empty() &&
+      available_shm_segments_.back()->size >= min_size) {
+    SHMBuffer* ret = available_shm_segments_.back();
+    available_shm_segments_.pop_back();
+    return ret;
+  }
+  return NULL;
+}
+
+void RTCVideoDecoder::CreateSHM(size_t min_size) {
+  DCHECK(base::MessageLoop::current() ==
+         content::ChildThread::current()->message_loop());
+  DVLOG(2) << "CreateSharedMemory. size=" << min_size;
+  size_t size_to_allocate = std::max(min_size, kSharedMemorySegmentBytes);
+  // Create three shared memory at once so we don't need to trampoline to the

[Ami GONE FROM CHROMIUM, 2013/06/26 00:11:58]

Why 3 and not kMaxInFlightDecodes?
(and method name & doco need to reflect this)

[wuchengli, 2013/06/28 15:08:44]

Because the message can be posted twice or more before CreateSHM is actually
called. It's easy to have kMaxInFlightDecodes * 2 buffers and half of them is
wasted. 

I've changed it to always keeping an extra SHM.

+  // child thread frequently.
+  for (int i = 0; i < 3; i++) {
+    base::SharedMemory* shm = factories_->CreateSharedMemory(size_to_allocate);
+    if (shm != NULL)
+      PutSHM(new SHMBuffer(shm, size_to_allocate));
+  }
+}
+
+void RTCVideoDecoder::PutSHM(SHMBuffer* shm_buffer) {
+  base::AutoLock auto_lock(lock_);
+  available_shm_segments_.push_back(shm_buffer);
+}
+
+void RTCVideoDecoder::RecordBufferData(const BufferData& buffer_data) {
+  input_buffer_data_.push_front(buffer_data);
+  // Why this value?  Because why not.  avformat.h:MAX_REORDER_DELAY is 16, but
+  // that's too small for some pathological B-frame test videos.  The cost of
+  // using too-high a value is low (192 bits per extra slot).
+  static const size_t kMaxInputBufferDataSize = 128;
+  // Pop from the back of the list, because that's the oldest and least likely
+  // to be useful in the future data.
+  if (input_buffer_data_.size() > kMaxInputBufferDataSize)
+    input_buffer_data_.pop_back();
+}
+
+void RTCVideoDecoder::GetBufferData(int32 bitstream_buffer_id,
+                                    uint32_t* timestamp,
+                                    uint32_t* width,
+                                    uint32_t* height,
+                                    size_t* size) {
+  for (std::list<BufferData>::iterator it = input_buffer_data_.begin();
+       it != input_buffer_data_.end();
+       ++it) {
+    if (it->bitstream_buffer_id != bitstream_buffer_id)
+      continue;
+    *timestamp = it->timestamp;
+    *width = it->width;
+    *height = it->height;
+    return;
+  }
+  NOTREACHED() << "Missing bitstream buffer id: " << bitstream_buffer_id;
+}
+
+}  // namespace content