Collect VTVideoDecodeAccelerator frames into a work queue

content/common/gpu/media/vt_video_decode_accelerator.cc

Index: content/common/gpu/media/vt_video_decode_accelerator.cc
diff --git a/content/common/gpu/media/vt_video_decode_accelerator.cc b/content/common/gpu/media/vt_video_decode_accelerator.cc
index f625fbcbc66728c12d188558dd5c2c6196bbf0f8..cb29a1cd049eadbcdc278ce4af253297414a6b7c 100644
--- a/content/common/gpu/media/vt_video_decode_accelerator.cc
+++ b/content/common/gpu/media/vt_video_decode_accelerator.cc
@@ -7,7 +7,6 @@
 #include <OpenGL/gl.h>
 
 #include "base/bind.h"
-#include "base/callback_helpers.h"
 #include "base/command_line.h"
 #include "base/sys_byteorder.h"
 #include "base/thread_task_runner_handle.h"
@@ -29,12 +28,13 @@ using content_common_gpu_media::StubPathMap;
 
 namespace content {
 
-// Size of NALU length headers in AVCC/MPEG-4 format (can be 1, 2, or 4).
+// Size to use for NALU length headers in AVC format (can be 1, 2, or 4).
 static const int kNALUHeaderLength = 4;
 
-// We only request 5 picture buffers from the client which are used to hold the
-// decoded samples. These buffers are then reused when the client tells us that
-// it is done with the buffer.
+// We request 5 picture buffers from the client, each of which has a texture ID
+// that we can bind decoded frames to. We need enough to satisfy preroll, and
+// enough to avoid unnecessary stalling, but no more than that. The resource
+// requirements are low, as we don't need the textures to be backed by storage.
 static const int kNumPictureBuffers = 5;
 
 // Route decoded frame callbacks back into the VTVideoDecodeAccelerator.
@@ -48,28 +48,20 @@ static void OutputThunk(
     CMTime presentation_duration) {
   VTVideoDecodeAccelerator* vda =
       reinterpret_cast<VTVideoDecodeAccelerator*>(decompression_output_refcon);
-  int32_t bitstream_id = reinterpret_cast<intptr_t>(source_frame_refcon);
-  vda->Output(bitstream_id, status, image_buffer);
+  vda->Output(source_frame_refcon, status, image_buffer);
 }
 
-VTVideoDecodeAccelerator::DecodedFrame::DecodedFrame(
-    int32_t bitstream_id,
-    CVImageBufferRef image_buffer)
-    : bitstream_id(bitstream_id),
-      image_buffer(image_buffer) {
+VTVideoDecodeAccelerator::Task::Task(TaskType type) : type(type) {
 }
 
-VTVideoDecodeAccelerator::DecodedFrame::~DecodedFrame() {
+VTVideoDecodeAccelerator::Task::~Task() {
 }
 
-VTVideoDecodeAccelerator::PendingAction::PendingAction(
-    Action action,
-    int32_t bitstream_id)
-    : action(action),
-      bitstream_id(bitstream_id) {
+VTVideoDecodeAccelerator::Frame::Frame(int32_t bitstream_id)
+    : bitstream_id(bitstream_id) {
 }
 
-VTVideoDecodeAccelerator::PendingAction::~PendingAction() {
+VTVideoDecodeAccelerator::Frame::~Frame() {
 }
 
 VTVideoDecodeAccelerator::VTVideoDecodeAccelerator(
@@ -78,7 +70,7 @@ VTVideoDecodeAccelerator::VTVideoDecodeAccelerator(
     : cgl_context_(cgl_context),
       make_context_current_(make_context_current),
       client_(NULL),
-      has_error_(false),
+      state_(STATE_NORMAL),
       format_(NULL),
       session_(NULL),
       gpu_task_runner_(base::ThreadTaskRunnerHandle::Get()),
@@ -126,10 +118,34 @@ bool VTVideoDecodeAccelerator::Initialize(
   return true;
 }
 
-bool VTVideoDecodeAccelerator::ConfigureDecoder(
-    const std::vector<const uint8_t*>& nalu_data_ptrs,
-    const std::vector<size_t>& nalu_data_sizes) {
+bool VTVideoDecodeAccelerator::FinishDelayedFrames() {
   DCHECK(decoder_thread_.message_loop_proxy()->BelongsToCurrentThread());
+  if (session_) {
+    OSStatus status = VTDecompressionSessionFinishDelayedFrames(session_);
+    if (status) {
+      NOTIFY_STATUS("VTDecompressionSessionFinishDelayedFrames()", status);
+      return false;
+    }
+  }
+  return true;
+}
+
+bool VTVideoDecodeAccelerator::ConfigureDecoder() {
+  DCHECK(decoder_thread_.message_loop_proxy()->BelongsToCurrentThread());
+  DCHECK(!last_sps_.empty());
+  DCHECK(!last_pps_.empty());
+
+  // Build the configuration records.
+  std::vector<const uint8_t*> nalu_data_ptrs;

[DaleCurtis, 2014/11/10 20:48:02]

Yuck, I guess you could const initialize parallel arrays directly, but it's of
questionable cleanliness. I.e.

  const uint8_t* nalu_data_ptrs[] = {
    last_sps_.front(),
    last_spsext_.empty() ? last_pps_.front() : last_spsext_.front(),
    last_spsext_.empty() ? NULL : last_pps_.front()
  }

At least reserve size 3 for each to avoid unnecessary growth?

[sandersd (OOO until July 31), 2014/11/11 18:46:06]

Done. I'll see about a refactor to remove unnecessary heap allocations later.

+  std::vector<size_t> nalu_data_sizes;
+  nalu_data_ptrs.push_back(&last_sps_.front());
+  nalu_data_sizes.push_back(last_sps_.size());
+  if (!last_spsext_.empty()) {
+    nalu_data_ptrs.push_back(&last_spsext_.front());
+    nalu_data_sizes.push_back(last_spsext_.size());
+  }
+  nalu_data_ptrs.push_back(&last_pps_.front());
+  nalu_data_sizes.push_back(last_pps_.size());
 
   // Construct a new format description from the parameter sets.
   // TODO(sandersd): Replace this with custom code to support OS X < 10.9.
@@ -147,11 +163,19 @@ bool VTVideoDecodeAccelerator::ConfigureDecoder(
     return false;
   }
 
-  // If the session is compatible, there's nothing to do.
+  // Flush all frames using the previous configuration to keep things simple.

[DaleCurtis, 2014/11/10 20:48:02]

Shouldn't this happen in the Reset() that precedes a new ConfigureDecoder()
call?

[sandersd (OOO until July 31), 2014/11/11 18:46:07]

The decoder does not need to be reset to get a new configuration. (Although it
also turns out that if I shuffle around when metadata is stored, I don't need
this flush anyway.)

+  if (!FinishDelayedFrames())
+    return false;
+
+  // Store the new configuration data.
+  CMVideoDimensions coded_dimensions =
+      CMVideoFormatDescriptionGetDimensions(format_);
+  coded_size_.SetSize(coded_dimensions.width, coded_dimensions.height);
+
+  // If the session is compatible, there's nothing else to do.
   if (session_ &&
-      VTDecompressionSessionCanAcceptFormatDescription(session_, format_)) {
-      return true;
-  }
+      VTDecompressionSessionCanAcceptFormatDescription(session_, format_))

[DaleCurtis, 2014/11/10 21:10:39]

Multiline if should keep {}

[sandersd (OOO until July 31), 2014/11/11 18:46:07]

Done.

+    return true;
 
   // Prepare VideoToolbox configuration dictionaries.
   base::ScopedCFTypeRef<CFMutableDictionaryRef> decoder_config(
@@ -174,8 +198,6 @@ bool VTVideoDecodeAccelerator::ConfigureDecoder(
           &kCFTypeDictionaryKeyCallBacks,
           &kCFTypeDictionaryValueCallBacks));
 
-  CMVideoDimensions coded_dimensions =
-      CMVideoFormatDescriptionGetDimensions(format_);
 #define CFINT(i) CFNumberCreate(kCFAllocatorDefault, kCFNumberSInt32Type, &i)
   // TODO(sandersd): RGBA option for 4:4:4 video.
   int32_t pixel_format = kCVPixelFormatType_422YpCbCr8;
@@ -207,33 +229,11 @@ bool VTVideoDecodeAccelerator::ConfigureDecoder(
   return true;
 }
 
-void VTVideoDecodeAccelerator::Decode(const media::BitstreamBuffer& bitstream) {
-  DCHECK(CalledOnValidThread());
-  // Not actually a requirement of the VDA API, but we're lazy and use negative
-  // values as flags internally. Revisit that if this actually happens.
-  if (bitstream.id() < 0) {
-    LOG(ERROR) << "Negative bitstream ID";
-    NotifyError(INVALID_ARGUMENT);
-    client_->NotifyEndOfBitstreamBuffer(bitstream.id());
-    return;
-  }
-  pending_bitstream_ids_.push(bitstream.id());
-  decoder_thread_.message_loop_proxy()->PostTask(FROM_HERE, base::Bind(
-      &VTVideoDecodeAccelerator::DecodeTask, base::Unretained(this),
-      bitstream));
-}
-
 void VTVideoDecodeAccelerator::DecodeTask(
-    const media::BitstreamBuffer& bitstream) {
+    const media::BitstreamBuffer& bitstream,
+    Frame* frame) {
   DCHECK(decoder_thread_.message_loop_proxy()->BelongsToCurrentThread());
 
-  // Once we have a bitstream buffer, we must either decode it or drop it.
-  // This construct ensures that the buffer is always dropped unless we call
-  // drop_bitstream.Release().
-  base::ScopedClosureRunner drop_bitstream(base::Bind(
-      &VTVideoDecodeAccelerator::DropBitstream, base::Unretained(this),
-      bitstream.id()));
-
   // Map the bitstream buffer.
   base::SharedMemory memory(bitstream.handle(), true);
   size_t size = bitstream.size();
@@ -245,11 +245,11 @@ void VTVideoDecodeAccelerator::DecodeTask(
   const uint8_t* buf = static_cast<uint8_t*>(memory.memory());
 
   // NALUs are stored with Annex B format in the bitstream buffer (start codes),

[DaleCurtis, 2014/11/10 20:48:02]

Peanut gallery: Seems like this entire NALU conversion process could be
extracted into a helper function?

Is this code necessary because FFmpegDemuxer invokes annex-b bitstream
conversion?  Should we just be disabling that for OSX instead?

[sandersd (OOO until July 31), 2014/11/11 18:46:06]

That's probably true, but there are a lot of outputs. I'll see if I can find a
clean way to do that in a future refactor.
Much of it is, but I would need to extract the SPS/PPS data either way (and for
reordering, the slice headers too), and that is a lot easier using the
H264Parser.

-  // but VideoToolbox expects AVCC/MPEG-4 format (length headers), so we must
-  // rewrite the data.
+  // but VideoToolbox expects AVC-4 format (length headers), so we must rewrite
+  // the data.
   //
-  // 1. Locate relevant NALUs and compute the size of the translated data.
-  //    Also record any parameter sets for VideoToolbox initialization.
+  // Locate relevant NALUs and compute the size of the rewritten data. Also
+  // record any parameter sets for VideoToolbox initialization.
   bool config_changed = false;
   size_t data_size = 0;
   std::vector<media::H264NALU> nalus;
@@ -264,7 +264,6 @@ void VTVideoDecodeAccelerator::DecodeTask(
       NotifyError(PLATFORM_FAILURE);
       return;
     }
-    // TODO(sandersd): Strict ordering rules.
     switch (nalu.nal_unit_type) {
       case media::H264NALU::kSPS:
         last_sps_.assign(nalu.data, nalu.data + nalu.size);
@@ -280,6 +279,15 @@ void VTVideoDecodeAccelerator::DecodeTask(
         last_pps_.assign(nalu.data, nalu.data + nalu.size);
         config_changed = true;
         break;
+      case media::H264NALU::kSliceDataA:
+      case media::H264NALU::kSliceDataB:
+      case media::H264NALU::kSliceDataC:
+        LOG(ERROR) << "Coded slide data partitions not implemented.";
+        NotifyError(PLATFORM_FAILURE);
+        return;
+      case media::H264NALU::kIDRSlice:
+      case media::H264NALU::kNonIDRSlice:
+        // TODO(sandersd): Compute pic_order_count.
       default:
         nalus.push_back(nalu);
         data_size += kNALUHeaderLength + nalu.size;
@@ -287,47 +295,37 @@ void VTVideoDecodeAccelerator::DecodeTask(
     }
   }
 
-  // 2. Initialize VideoToolbox.
-  // TODO(sandersd): Check if the new configuration is identical before
-  // reconfiguring.
+  // Initialize VideoToolbox.
+  // TODO(sandersd): Instead of assuming that the last SPS and PPS units are
+  // always the correct ones, maintain a cache of recent SPS and PPS units and
+  // select from them using the slice header.
   if (config_changed) {

[DaleCurtis, 2014/11/10 20:48:02]

Do you do this here instead of Initialize() since you need the sps/pps stuff?

[sandersd (OOO until July 31), 2014/11/11 18:46:07]

That's part of it, but also the config can change at any frame.

     if (last_sps_.size() == 0 || last_pps_.size() == 0) {
       LOG(ERROR) << "Invalid configuration data";

[DaleCurtis, 2014/11/10 20:48:02]

You should convert all these LOG(ERROR) calls to DLOG(ERROR) once this is
complete. Most users won't see these which ends up just being binary bloat.

[sandersd (OOO until July 31), 2014/11/11 18:46:07]

Ah, you're quite right, the "VDA Error" message printed by GPUVideoDecoder is
also a DLOG(ERROR). This is going to make debugging failures in the wild a lot
harder.

[DaleCurtis, 2014/11/11 21:02:34]

You might file a bug to do this after a default on canary release then so you
can triage quickly.

[sandersd (OOO until July 31), 2014/11/11 22:14:39]

Perhaps it would be better to temporarily add a lot of media log items?

       NotifyError(INVALID_ARGUMENT);
       return;
     }
-    // TODO(sandersd): Check that the SPS and PPS IDs match.
-    std::vector<const uint8_t*> nalu_data_ptrs;
-    std::vector<size_t> nalu_data_sizes;
-    nalu_data_ptrs.push_back(&last_sps_.front());
-    nalu_data_sizes.push_back(last_sps_.size());
-    if (last_spsext_.size() != 0) {
-      nalu_data_ptrs.push_back(&last_spsext_.front());
-      nalu_data_sizes.push_back(last_spsext_.size());
-    }
-    nalu_data_ptrs.push_back(&last_pps_.front());
-    nalu_data_sizes.push_back(last_pps_.size());
-
-    // If ConfigureDecoder() fails, it already called NotifyError().
-    if (!ConfigureDecoder(nalu_data_ptrs, nalu_data_sizes))
+    if (!ConfigureDecoder())
       return;
   }
 
-  // If there are no non-configuration units, immediately return an empty
-  // (ie. dropped) frame. It is an error to create a MemoryBlock with zero
-  // size.
-  if (!data_size)
+  // If there are no non-configuration units, drop the bitstream buffer by
+  // returning an empty frame.
+  if (!data_size) {
+    if (!FinishDelayedFrames())
+      return;
+    DecodeDone(frame);
     return;
+  }
 
-  // If the session is not configured, fail.
+  // If the session is not configured by this point, fail.
   if (!session_) {
-    LOG(ERROR) << "Image slice without configuration data";
+    LOG(ERROR) << "Image slice without configuration";
     NotifyError(INVALID_ARGUMENT);
     return;
   }
 
-  // 3. Allocate a memory-backed CMBlockBuffer for the translated data.
-  // TODO(sandersd): Check that the slice's PPS matches the current PPS.
+  // Create a memory-backed CMBlockBuffer for the translated data.
   base::ScopedCFTypeRef<CMBlockBufferRef> data;
   OSStatus status = CMBlockBufferCreateWithMemoryBlock(
       kCFAllocatorDefault,
@@ -344,7 +342,7 @@ void VTVideoDecodeAccelerator::DecodeTask(
     return;
   }
 
-  // 4. Copy NALU data, inserting length headers.
+  // Copy NALU data into the CMBlockBuffer, inserting length headers.
   size_t offset = 0;
   for (size_t i = 0; i < nalus.size(); i++) {
     media::H264NALU& nalu = nalus[i];
@@ -364,8 +362,8 @@ void VTVideoDecodeAccelerator::DecodeTask(
     offset += nalu.size;
   }
 
-  // 5. Package the data for VideoToolbox and request decoding.
-  base::ScopedCFTypeRef<CMSampleBufferRef> frame;
+  // Package the data in a CMSampleBuffer.
+  base::ScopedCFTypeRef<CMSampleBufferRef> sample;
   status = CMSampleBufferCreate(
       kCFAllocatorDefault,
       data,                 // data_buffer
@@ -378,79 +376,110 @@ void VTVideoDecodeAccelerator::DecodeTask(
       NULL,                 // &sample_timing_array
       0,                    // num_sample_size_entries
       NULL,                 // &sample_size_array
-      frame.InitializeInto());
+      sample.InitializeInto());
   if (status) {
     NOTIFY_STATUS("CMSampleBufferCreate()", status);
     return;
   }
 
+  // Update the frame data.
+  frame->coded_size = coded_size_;
+
+  // Send the frame for decoding.
   // Asynchronous Decompression allows for parallel submission of frames
   // (without it, DecodeFrame() does not return until the frame has been
   // decoded). We don't enable Temporal Processing so that frames are always
   // returned in decode order; this makes it easier to avoid deadlock.
   VTDecodeFrameFlags decode_flags =
       kVTDecodeFrame_EnableAsynchronousDecompression;
-
-  intptr_t bitstream_id = bitstream.id();
   status = VTDecompressionSessionDecodeFrame(
       session_,
-      frame,                                  // sample_buffer
+      sample,                                 // sample_buffer
       decode_flags,                           // decode_flags
-      reinterpret_cast<void*>(bitstream_id),  // source_frame_refcon
+      reinterpret_cast<void*>(frame),         // source_frame_refcon
       NULL);                                  // &info_flags_out
   if (status) {
     NOTIFY_STATUS("VTDecompressionSessionDecodeFrame()", status);
     return;
   }
-
-  // Now that the bitstream is decoding, don't drop it.
-  (void)drop_bitstream.Release();
 }
 
 // This method may be called on any VideoToolbox thread.
 void VTVideoDecodeAccelerator::Output(
-    int32_t bitstream_id,
+    void* source_frame_refcon,
     OSStatus status,
     CVImageBufferRef image_buffer) {
   if (status) {
-    // TODO(sandersd): Handle dropped frames.
     NOTIFY_STATUS("Decoding", status);
-    image_buffer = NULL;
   } else if (CFGetTypeID(image_buffer) != CVPixelBufferGetTypeID()) {
     LOG(ERROR) << "Decoded frame is not a CVPixelBuffer";
     NotifyError(PLATFORM_FAILURE);
-    image_buffer = NULL;
   } else {
-    CFRetain(image_buffer);
+    Frame* frame = reinterpret_cast<Frame*>(source_frame_refcon);
+    frame->image.reset(image_buffer, base::scoped_policy::RETAIN);
+    DecodeDone(frame);
   }
+}
+
+void VTVideoDecodeAccelerator::DecodeDone(Frame* frame) {
+  if (!CalledOnValidThread()) {

[DaleCurtis, 2014/11/10 21:10:39]

If Output() is always going to call on the wrong thread, it seems it'd be better
to always do the post there and then DCHECK here().

[sandersd (OOO until July 31), 2014/11/11 18:46:07]

Done.

+    gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
+        &VTVideoDecodeAccelerator::DecodeDone,
+        weak_this_factory_.GetWeakPtr(), frame));
+  } else {
+    DCHECK_EQ(frame->bitstream_id, pending_frames_.front()->bitstream_id);
+    Task task(TASK_FRAME);
+    task.frame = pending_frames_.front();
+    pending_frames_.pop();
+    pending_tasks_.push(task);
+    ProcessTasks();
+  }
+}
+
+void VTVideoDecodeAccelerator::FlushTask(TaskType type) {
+  DCHECK(decoder_thread_.message_loop_proxy()->BelongsToCurrentThread());
+  FinishDelayedFrames();
+
+  // Always queue a task, even if FinishDelayedFrames() fails, so that
+  // destruction always completes.
   gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
-      &VTVideoDecodeAccelerator::OutputTask,
-      weak_this_factory_.GetWeakPtr(),
-      DecodedFrame(bitstream_id, image_buffer)));
+      &VTVideoDecodeAccelerator::FlushDone,
+      weak_this_factory_.GetWeakPtr(), type));
+}
+
+void VTVideoDecodeAccelerator::FlushDone(TaskType type) {
+  DCHECK(CalledOnValidThread());

[DaleCurtis, 2014/11/10 21:10:39]

CalledOnValidThread isn't very readable when there are multiple threads :) Can
you rename this to to indicate which thread is the valid thread?

[sandersd (OOO until July 31), 2014/11/11 18:46:06]

I could add a wrapper method, but CalledOnValidThread() is a member of
base:NonThreadSafe, and always means the thread the object was constructed on.

[DaleCurtis, 2014/11/11 21:02:34]

Hmm instead of inheriting non-threadsafe you should use a base::ThreadChecker. 
It's recommended to do this when multiple inheritance might occur (as it does
for your class).

[sandersd (OOO until July 31), 2014/11/11 22:14:39]

Done.

+  pending_tasks_.push(Task(type));
+  ProcessTasks();
 }
 
-void VTVideoDecodeAccelerator::OutputTask(DecodedFrame frame) {
+void VTVideoDecodeAccelerator::Decode(const media::BitstreamBuffer& bitstream) {
   DCHECK(CalledOnValidThread());
-  decoded_frames_.push(frame);
-  ProcessDecodedFrames();
+  DCHECK_EQ(assigned_bitstream_ids_.count(bitstream.id()), 0u);
+  assigned_bitstream_ids_.insert(bitstream.id());
+  Frame* frame = new Frame(bitstream.id());

[DaleCurtis, 2014/11/10 21:10:39]

Why does Frame have a bitstream_id instead of owning the bitstream buffer?

[sandersd (OOO until July 31), 2014/11/11 18:46:07]

The ownership and lifetime story was not very clear, as the buffer becomes
invalid once you call NotifyEndOfBitstreamBuffer() but the decoded frame stays
valid (this isn't a big difference right now, but the reorder CL will to that
before the frames enter the reorder buffer).

In general I also wasn't happy with having a reference as a struct member.

(There were stronger technical reasons before but they have been eliminated, so
it's at least feasible to change now.)

[DaleCurtis, 2014/11/11 21:02:34]

I'd spend some time cleaning up the ownership model sooner rather than later if
you have an idea on how.

[sandersd (OOO until July 31), 2014/11/11 22:14:39]

I do not, and this is the best I had on my third rewrite.

The key problem is that the main thread needs to maintain ownership so that it
can clean up properly, but in a logical sense the decoder owns the data while it
is being decoded.

+  pending_frames_.push(make_linked_ptr(frame));
+  decoder_thread_.message_loop_proxy()->PostTask(FROM_HERE, base::Bind(
+      &VTVideoDecodeAccelerator::DecodeTask, base::Unretained(this),
+      bitstream, frame));
 }
 
 void VTVideoDecodeAccelerator::AssignPictureBuffers(
     const std::vector<media::PictureBuffer>& pictures) {
   DCHECK(CalledOnValidThread());
 
-  for (size_t i = 0; i < pictures.size(); i++) {
-    DCHECK(!texture_ids_.count(pictures[i].id()));
-    assigned_picture_ids_.insert(pictures[i].id());
-    available_picture_ids_.push_back(pictures[i].id());
-    texture_ids_[pictures[i].id()] = pictures[i].texture_id();
+  for (const media::PictureBuffer& picture : pictures) {
+    DCHECK(!texture_ids_.count(picture.id()));
+    assigned_picture_ids_.insert(picture.id());
+    available_picture_ids_.push_back(picture.id());
+    texture_ids_[picture.id()] = picture.texture_id();
   }
 
   // Pictures are not marked as uncleared until after this method returns, and
   // they will be broken if they are used before that happens. So, schedule
   // future work after that happens.
   gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
-      &VTVideoDecodeAccelerator::ProcessDecodedFrames,
+      &VTVideoDecodeAccelerator::ProcessTasks,
       weak_this_factory_.GetWeakPtr()));
 }
 
@@ -458,289 +487,179 @@ void VTVideoDecodeAccelerator::ReusePictureBuffer(int32_t picture_id) {
   DCHECK(CalledOnValidThread());
   DCHECK_EQ(CFGetRetainCount(picture_bindings_[picture_id]), 1);
   picture_bindings_.erase(picture_id);
-  // Don't put the picture back in the available list if has been dismissed.
   if (assigned_picture_ids_.count(picture_id) != 0) {
     available_picture_ids_.push_back(picture_id);
-    ProcessDecodedFrames();
-  }
-}
-
-void VTVideoDecodeAccelerator::CompleteAction(Action action) {
-  DCHECK(CalledOnValidThread());
-
-  switch (action) {
-    case ACTION_FLUSH:
-      client_->NotifyFlushDone();
-      break;
-    case ACTION_RESET:
-      client_->NotifyResetDone();
-      break;
-    case ACTION_DESTROY:
-      delete this;
-      break;
-  }
-}
-
-void VTVideoDecodeAccelerator::CompleteActions(int32_t bitstream_id) {
-  DCHECK(CalledOnValidThread());
-  while (!pending_actions_.empty() &&
-         pending_actions_.front().bitstream_id == bitstream_id) {
-    CompleteAction(pending_actions_.front().action);
-    pending_actions_.pop();
+    ProcessTasks();
+  } else {
+    client_->DismissPictureBuffer(picture_id);
   }
 }
 
-void VTVideoDecodeAccelerator::ProcessDecodedFrames() {
+void VTVideoDecodeAccelerator::ProcessTasks() {
   DCHECK(CalledOnValidThread());
 
-  while (!decoded_frames_.empty()) {
-    if (pending_actions_.empty()) {
-      // No pending actions; send frames normally.
-      if (!has_error_)
-        SendPictures(pending_bitstream_ids_.back());
-      return;
-    }
+  while (!pending_tasks_.empty()) {
+    const Task& task = pending_tasks_.front();
 
-    int32_t next_action_bitstream_id = pending_actions_.front().bitstream_id;
-    int32_t last_sent_bitstream_id = -1;
-    switch (pending_actions_.front().action) {
-      case ACTION_FLUSH:
-        // Send frames normally.
-        if (has_error_)
+    switch (state_) {
+      case STATE_NORMAL:
+        if (!ProcessTask(task))
           return;
-        last_sent_bitstream_id = SendPictures(next_action_bitstream_id);
+        pending_tasks_.pop();
         break;
 
-      case ACTION_RESET:
-        // Drop decoded frames.
-        if (has_error_)
-          return;
-        while (!decoded_frames_.empty() &&
-               last_sent_bitstream_id != next_action_bitstream_id) {
-          last_sent_bitstream_id = decoded_frames_.front().bitstream_id;
-          decoded_frames_.pop();
-          DCHECK_EQ(pending_bitstream_ids_.front(), last_sent_bitstream_id);
-          pending_bitstream_ids_.pop();
-          client_->NotifyEndOfBitstreamBuffer(last_sent_bitstream_id);
-        }
+      case STATE_ERROR:
+        // Do nothing until Destroy() is called.
         break;
 
-      case ACTION_DESTROY:
-        // Drop decoded frames, without bookkeeping.
-        while (!decoded_frames_.empty()) {
-          last_sent_bitstream_id = decoded_frames_.front().bitstream_id;
-          decoded_frames_.pop();
+      case STATE_DESTROYING:
+        // Discard tasks until destruction is complete.
+        if (task.type == TASK_DESTROY) {
+          delete this;
+          return;
         }
-
-        // Handle completing the action specially, as it is important not to
-        // access |this| after calling CompleteAction().
-        if (last_sent_bitstream_id == next_action_bitstream_id)
-          CompleteAction(ACTION_DESTROY);
-
-        // Either |this| was deleted or no more progress can be made.
-        return;
+        pending_tasks_.pop();
+        break;
     }
-
-    // If we ran out of buffers (or pictures), no more progress can be made
-    // until more frames are decoded.
-    if (last_sent_bitstream_id != next_action_bitstream_id)
-      return;
-
-    // Complete all actions pending for this |bitstream_id|, then loop to see
-    // if progress can be made on the next action.
-    CompleteActions(next_action_bitstream_id);
   }
 }
 
-int32_t VTVideoDecodeAccelerator::ProcessDroppedFrames(
-    int32_t last_sent_bitstream_id,
-    int32_t up_to_bitstream_id) {
+bool VTVideoDecodeAccelerator::ProcessTask(const Task& task) {
   DCHECK(CalledOnValidThread());
-  // Drop frames as long as there is a frame, we have not reached the next
-  // action, and the next frame has no image.
-  while (!decoded_frames_.empty() &&
-         last_sent_bitstream_id != up_to_bitstream_id &&
-         decoded_frames_.front().image_buffer.get() == NULL) {
-    const DecodedFrame& frame = decoded_frames_.front();
-    DCHECK_EQ(pending_bitstream_ids_.front(), frame.bitstream_id);
-    client_->NotifyEndOfBitstreamBuffer(frame.bitstream_id);
-    last_sent_bitstream_id = frame.bitstream_id;
-    decoded_frames_.pop();
-    pending_bitstream_ids_.pop();
+  DCHECK_EQ(state_, STATE_NORMAL);
+
+  switch (task.type) {
+    case TASK_FRAME:
+      return ProcessFrame(*task.frame);
+
+    case TASK_FLUSH:
+      DCHECK_EQ(task.type, pending_flush_tasks_.front());
+      pending_flush_tasks_.pop();
+      client_->NotifyFlushDone();
+      return true;
+
+    case TASK_RESET:
+      DCHECK_EQ(task.type, pending_flush_tasks_.front());
+      pending_flush_tasks_.pop();
+      client_->NotifyResetDone();
+      return true;
+
+    case TASK_DESTROY:
+      NOTREACHED() << "Can't destroy while in STATE_NORMAL.";
+      NotifyError(ILLEGAL_STATE);
+      return false;
   }
-  return last_sent_bitstream_id;
 }
 
-// TODO(sandersd): If GpuVideoDecoder didn't specifically check the size of
-// textures, this would be unnecessary, as the size is actually a property of
-// the texture binding, not the texture. We rebind every frame, so the size
-// passed to ProvidePictureBuffers() is meaningless.
-void VTVideoDecodeAccelerator::ProcessSizeChange() {
+bool VTVideoDecodeAccelerator::ProcessFrame(const Frame& frame) {
   DCHECK(CalledOnValidThread());
-  DCHECK(!decoded_frames_.empty());
-
-  // Find the size of the next image.
-  const DecodedFrame& frame = decoded_frames_.front();
-  CVImageBufferRef image_buffer = frame.image_buffer.get();
-  size_t width = CVPixelBufferGetWidth(image_buffer);
-  size_t height = CVPixelBufferGetHeight(image_buffer);
-  gfx::Size image_size(width, height);
-
-  if (picture_size_ != image_size) {
-    // Dismiss all assigned picture buffers.
-    for (int32_t picture_id : assigned_picture_ids_)
-      client_->DismissPictureBuffer(picture_id);
-    assigned_picture_ids_.clear();
-    available_picture_ids_.clear();
-
-    // Request new pictures.
-    client_->ProvidePictureBuffers(
-        kNumPictureBuffers, image_size, GL_TEXTURE_RECTANGLE_ARB);
-    picture_size_ = image_size;
+  DCHECK_EQ(state_, STATE_NORMAL);

[DaleCurtis, 2014/11/10 21:10:39]

This needs a comment on exactly what's happening here. The resetting logic seems
backwards at first glance.

[sandersd (OOO until July 31), 2014/11/11 18:46:07]

Done.

+  bool resetting = !pending_flush_tasks_.empty() &&
+                   pending_flush_tasks_.front() == TASK_RESET;
+  if (!resetting && frame.image.get()) {
+    if (picture_size_ != frame.coded_size) {
+      // Dismiss current pictures.
+      for (int32_t picture_id : assigned_picture_ids_)
+        client_->DismissPictureBuffer(picture_id);
+      assigned_picture_ids_.clear();
+      available_picture_ids_.clear();
+
+      // Request new pictures.
+      picture_size_ = frame.coded_size;
+      client_->ProvidePictureBuffers(
+          kNumPictureBuffers, coded_size_, GL_TEXTURE_RECTANGLE_ARB);
+      return false;
+    }
+    if (!SendFrame(frame))
+      return false;
   }
+  assigned_bitstream_ids_.erase(frame.bitstream_id);
+  client_->NotifyEndOfBitstreamBuffer(frame.bitstream_id);
+  return true;
 }
 
-int32_t VTVideoDecodeAccelerator::SendPictures(int32_t up_to_bitstream_id) {
+bool VTVideoDecodeAccelerator::SendFrame(const Frame& frame) {
   DCHECK(CalledOnValidThread());
-  DCHECK(!decoded_frames_.empty());
-
-  // TODO(sandersd): Store the actual last sent bitstream ID?
-  int32_t last_sent_bitstream_id = -1;
+  DCHECK_EQ(state_, STATE_NORMAL);
 
-  last_sent_bitstream_id =
-      ProcessDroppedFrames(last_sent_bitstream_id, up_to_bitstream_id);
-  if (last_sent_bitstream_id == up_to_bitstream_id || decoded_frames_.empty())
-    return last_sent_bitstream_id;
-
-  ProcessSizeChange();
   if (available_picture_ids_.empty())
-    return last_sent_bitstream_id;
+    return false;
+
+  int32_t picture_id = available_picture_ids_.back();
+  IOSurfaceRef surface = CVPixelBufferGetIOSurface(frame.image.get());
 
   if (!make_context_current_.Run()) {
     LOG(ERROR) << "Failed to make GL context current";
     NotifyError(PLATFORM_FAILURE);
-    return last_sent_bitstream_id;
+    return false;
   }
 
   glEnable(GL_TEXTURE_RECTANGLE_ARB);
-  while (!available_picture_ids_.empty() && !has_error_) {
-    DCHECK_NE(last_sent_bitstream_id, up_to_bitstream_id);
-    DCHECK(!decoded_frames_.empty());
-
-    // We don't pop |frame| or |picture_id| until they are consumed, which may
-    // not happen if an error occurs. Conveniently, this also removes some
-    // refcounting.
-    const DecodedFrame& frame = decoded_frames_.front();
-    DCHECK_EQ(pending_bitstream_ids_.front(), frame.bitstream_id);
-    int32_t picture_id = available_picture_ids_.back();
-
-    CVImageBufferRef image_buffer = frame.image_buffer.get();
-    IOSurfaceRef surface = CVPixelBufferGetIOSurface(image_buffer);
-
-    gfx::ScopedTextureBinder
-        texture_binder(GL_TEXTURE_RECTANGLE_ARB, texture_ids_[picture_id]);
-    CGLError status = CGLTexImageIOSurface2D(
-        cgl_context_,                 // ctx
-        GL_TEXTURE_RECTANGLE_ARB,     // target
-        GL_RGB,                       // internal_format
-        picture_size_.width(),        // width
-        picture_size_.height(),       // height
-        GL_YCBCR_422_APPLE,           // format
-        GL_UNSIGNED_SHORT_8_8_APPLE,  // type
-        surface,                      // io_surface
-        0);                           // plane
-    if (status != kCGLNoError) {
-      NOTIFY_STATUS("CGLTexImageIOSurface2D()", status);
-      break;
-    }
-
-    picture_bindings_[picture_id] = frame.image_buffer;
-    client_->PictureReady(media::Picture(
-        picture_id, frame.bitstream_id, gfx::Rect(picture_size_)));
-    available_picture_ids_.pop_back();
-    client_->NotifyEndOfBitstreamBuffer(frame.bitstream_id);
-    last_sent_bitstream_id = frame.bitstream_id;
-    decoded_frames_.pop();
-    pending_bitstream_ids_.pop();
-
-    last_sent_bitstream_id =
-        ProcessDroppedFrames(last_sent_bitstream_id, up_to_bitstream_id);
-    if (last_sent_bitstream_id == up_to_bitstream_id || decoded_frames_.empty())
-      break;
-
-    ProcessSizeChange();
+  gfx::ScopedTextureBinder
+      texture_binder(GL_TEXTURE_RECTANGLE_ARB, texture_ids_[picture_id]);
+  CGLError status = CGLTexImageIOSurface2D(
+      cgl_context_,                 // ctx
+      GL_TEXTURE_RECTANGLE_ARB,     // target
+      GL_RGB,                       // internal_format
+      frame.coded_size.width(),     // width
+      frame.coded_size.height(),    // height
+      GL_YCBCR_422_APPLE,           // format
+      GL_UNSIGNED_SHORT_8_8_APPLE,  // type
+      surface,                      // io_surface
+      0);                           // plane
+  if (status != kCGLNoError) {
+    NOTIFY_STATUS("CGLTexImageIOSurface2D()", status);
+    return false;
   }
   glDisable(GL_TEXTURE_RECTANGLE_ARB);
 
-  return last_sent_bitstream_id;
-}
-
-void VTVideoDecodeAccelerator::FlushTask() {
-  DCHECK(decoder_thread_.message_loop_proxy()->BelongsToCurrentThread());
-  OSStatus status = VTDecompressionSessionFinishDelayedFrames(session_);
-  if (status)
-    NOTIFY_STATUS("VTDecompressionSessionFinishDelayedFrames()", status);
-}
-
-void VTVideoDecodeAccelerator::QueueAction(Action action) {
-  DCHECK(CalledOnValidThread());
-  if (pending_bitstream_ids_.empty()) {
-    // If there are no pending frames, all actions complete immediately.
-    CompleteAction(action);
-  } else {
-    // Otherwise, queue the action.
-    pending_actions_.push(PendingAction(action, pending_bitstream_ids_.back()));
-
-    // Request a flush to make sure the action will eventually complete.
-    decoder_thread_.message_loop_proxy()->PostTask(FROM_HERE, base::Bind(
-        &VTVideoDecodeAccelerator::FlushTask, base::Unretained(this)));
-
-    // See if we can make progress now that there is a new pending action.
-    ProcessDecodedFrames();
-  }
+  available_picture_ids_.pop_back();
+  picture_bindings_[picture_id] = frame.image;
+  client_->PictureReady(media::Picture(
+      picture_id, frame.bitstream_id, gfx::Rect(frame.coded_size)));
+  return true;
 }
 
 void VTVideoDecodeAccelerator::NotifyError(Error error) {
   if (!CalledOnValidThread()) {
     gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
         &VTVideoDecodeAccelerator::NotifyError,
-        weak_this_factory_.GetWeakPtr(),
-        error));
-    return;
+        weak_this_factory_.GetWeakPtr(), error));
+  } else if (state_ == STATE_NORMAL) {
+    state_ = STATE_ERROR;
+    client_->NotifyError(error);
   }
-  has_error_ = true;
-  client_->NotifyError(error);
 }
 
-void VTVideoDecodeAccelerator::DropBitstream(int32_t bitstream_id) {
-  DCHECK(decoder_thread_.message_loop_proxy()->BelongsToCurrentThread());
-  gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
-      &VTVideoDecodeAccelerator::OutputTask,
-      weak_this_factory_.GetWeakPtr(),
-      DecodedFrame(bitstream_id, NULL)));
+void VTVideoDecodeAccelerator::QueueFlush(TaskType type) {
+  DCHECK(CalledOnValidThread());
+  pending_flush_tasks_.push(type);
+  decoder_thread_.message_loop_proxy()->PostTask(FROM_HERE, base::Bind(
+      &VTVideoDecodeAccelerator::FlushTask, base::Unretained(this),
+      type));
+
+  // If this is a new flush reuqest, see if we can make progress.

[DaleCurtis, 2014/11/10 20:48:02]

sp: request

[sandersd (OOO until July 31), 2014/11/11 18:46:06]

Done.

+  if (pending_flush_tasks_.size() == 1)
+    ProcessTasks();
 }
 
 void VTVideoDecodeAccelerator::Flush() {
   DCHECK(CalledOnValidThread());
-  QueueAction(ACTION_FLUSH);
+  QueueFlush(TASK_FLUSH);
 }
 
 void VTVideoDecodeAccelerator::Reset() {
   DCHECK(CalledOnValidThread());
-  QueueAction(ACTION_RESET);
+  QueueFlush(TASK_RESET);
 }
 
 void VTVideoDecodeAccelerator::Destroy() {
   DCHECK(CalledOnValidThread());
-  // Drop any other pending actions.
-  while (!pending_actions_.empty())
-    pending_actions_.pop();
-  // Return all bitstream buffers.
-  while (!pending_bitstream_ids_.empty()) {
-    client_->NotifyEndOfBitstreamBuffer(pending_bitstream_ids_.front());
-    pending_bitstream_ids_.pop();
-  }
-  QueueAction(ACTION_DESTROY);
+  for (int32_t bitstream_id : assigned_bitstream_ids_)
+    client_->NotifyEndOfBitstreamBuffer(bitstream_id);
+  assigned_bitstream_ids_.clear();
+  state_ = STATE_DESTROYING;
+  QueueFlush(TASK_DESTROY);
 }
 
 bool VTVideoDecodeAccelerator::CanDecodeOnIOThread() {