MJPEG acceleration for video capture using VAAPI, the GPU and IPC part

content/common/gpu/media/gpu_jpeg_decode_accelerator.cc

+// Copyright 2015 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/common/gpu/media/gpu_jpeg_decode_accelerator.h"
+
+#include <stdint.h>
+
+#include <map>
+
+#include "base/bind.h"
+#include "base/logging.h"
+#include "base/memory/shared_memory.h"
+#include "base/single_thread_task_runner.h"
+#include "base/stl_util.h"
+#include "base/trace_event/trace_event.h"
+#include "content/common/gpu/gpu_channel.h"
+#include "content/common/gpu/gpu_messages.h"
+#include "ipc/ipc_message_macros.h"
+#include "ipc/message_filter.h"
+#include "media/filters/jpeg_parser.h"
+#include "ui/gfx/geometry/size.h"
+
+#if defined(OS_CHROMEOS) && defined(ARCH_CPU_X86_FAMILY)
+#include "content/common/gpu/media/vaapi_jpeg_decode_accelerator.h"
+#endif
+
+namespace {
+
+void DecodeFinished(scoped_ptr<base::SharedMemory> shm) {
+  // Do nothing. Because VideoFrame is backed by |shm|, the purpose of this
+  // function is to just keep reference of |shm| to make sure it lives util
+  // decode finishes.
+}
+
+}  // namespace
+
+namespace content {
+
+class GpuJpegDecodeAccelerator::Client
+    : public media::JpegDecodeAccelerator::Client {
+ public:
+  Client(content::GpuJpegDecodeAccelerator* owner, int32 route_id)
+      : owner_(owner->AsWeakPtr()), route_id_(route_id) {}
+
+  // media::JpegDecodeAccelerator::Client implementation.
+  void VideoFrameReady(int32_t bitstream_buffer_id) override {
+    if (owner_)
+      owner_->NotifyDecodeStatus(route_id_, bitstream_buffer_id,
+                                 media::JpegDecodeAccelerator::NO_ERROR);
+  }
+
+  void NotifyError(int32_t bitstream_buffer_id,
+                   media::JpegDecodeAccelerator::Error error) override {
+    if (owner_)
+      owner_->NotifyDecodeStatus(route_id_, bitstream_buffer_id, error);
+  }
+
+  void Decode(const media::BitstreamBuffer& bitstream_buffer,
+              const scoped_refptr<media::VideoFrame>& video_frame) {
+    DCHECK(accelerator_);
+    accelerator_->Decode(bitstream_buffer, video_frame);
+  }
+
+  void set_accelerator(scoped_ptr<media::JpegDecodeAccelerator> accelerator) {
+    accelerator_ = accelerator.Pass();
+  }
+
+ private:
+  base::WeakPtr<content::GpuJpegDecodeAccelerator> owner_;
+  int32 route_id_;
+  scoped_ptr<media::JpegDecodeAccelerator> accelerator_;
+};
+
+// Create, destroy, and RemoveClient run on child thread. All other methods run
+// on IO thread.
+class GpuJpegDecodeAccelerator::MessageFilter : public IPC::MessageFilter {
+ public:
+  explicit MessageFilter(GpuJpegDecodeAccelerator* owner)
+      : owner_(owner->AsWeakPtr()),
+        child_task_runner_(owner_->child_task_runner_),
+        io_task_runner_(owner_->io_task_runner_) {}
+
+  void OnChannelError() override { sender_ = NULL; }

[Pawel Osciak, 2015/05/28 09:13:17]

s/NULL/nullptr/ here and in other places.

[kcwu, 2015/05/28 12:10:27]

Done.

+
+  void OnChannelClosing() override { sender_ = NULL; }
+
+  void OnFilterAdded(IPC::Sender* sender) override { sender_ = sender; }
+
+  bool OnMessageReceived(const IPC::Message& msg) override {
+    const int32 route_id = msg.routing_id();
+    if (client_map_.count(route_id) == 0)
+      return false;
+
+    bool handled = true;
+    IPC_BEGIN_MESSAGE_MAP_WITH_PARAM(MessageFilter, msg, &route_id)
+      IPC_MESSAGE_HANDLER(AcceleratedJpegDecoderMsg_Decode, OnDecodeOnIOThread)
+      IPC_MESSAGE_HANDLER(AcceleratedJpegDecoderMsg_Destroy,
+                          OnDestroyOnIOThread)
+      IPC_MESSAGE_UNHANDLED(handled = false)
+    IPC_END_MESSAGE_MAP()
+    return handled;
+  }
+
+  bool SendOnIOThread(IPC::Message* message) {
+    DCHECK(!message->is_sync());
+    if (!sender_) {
+      delete message;
+      return false;
+    }
+    return sender_->Send(message);
+  }
+
+  void AddClientOnIOThread(int32 route_id,
+                           Client* client,
+                           IPC::Message* reply_msg) {
+    DCHECK(io_task_runner_->BelongsToCurrentThread());
+    DCHECK(client_map_.count(route_id) == 0);
+
+    client_map_[route_id] = client;
+    GpuMsg_CreateJpegDecoder::WriteReplyParams(reply_msg, true);
+    SendOnIOThread(reply_msg);
+  }
+
+  void OnDestroyOnIOThread(const int32* route_id) {
+    DCHECK(io_task_runner_->BelongsToCurrentThread());
+    const auto& it = client_map_.find(*route_id);
+    DCHECK(it != client_map_.end());

[Pawel Osciak, 2015/05/28 09:13:17]

DCHECK_NE

[kcwu, 2015/05/28 12:10:27]

DCHECK_NE is not for iterator. The arguments need to be printable.

+    Client* client = it->second;
+    DCHECK(client);
+    client_map_.erase(it);
+
+    child_task_runner_->PostTask(
+        FROM_HERE, base::Bind(&MessageFilter::DestroyClient, this, client));
+  }
+
+  void DestroyClient(Client* client) {
+    DCHECK(child_task_runner_->BelongsToCurrentThread());
+    delete client;
+    if (owner_)
+      owner_->ClientRemoved();
+  }
+
+  void NotifyDecodeStatusOnIOThread(int32 route_id,
+                                    int32_t buffer_id,
+                                    media::JpegDecodeAccelerator::Error error) {
+    DCHECK(io_task_runner_->BelongsToCurrentThread());
+    SendOnIOThread(new AcceleratedJpegDecoderHostMsg_DecodeAck(
+        route_id, buffer_id, error));
+  }
+
+  void OnDecodeOnIOThread(
+      const int32* route_id,
+      const AcceleratedJpegDecoderMsg_Decode_Params& params) {
+    DCHECK(io_task_runner_->BelongsToCurrentThread());
+    DCHECK(route_id);
+    TRACE_EVENT0("jpeg", "GpuJpegDecodeAccelerator::MessageFilter::OnDecode");
+
+    if (params.input_buffer_id < 0) {
+      LOG(ERROR) << "BitstreamBuffer id " << params.input_buffer_id
+                 << " out of range";
+      NotifyDecodeStatusOnIOThread(
+          *route_id, params.input_buffer_id,
+          media::JpegDecodeAccelerator::INVALID_ARGUMENT);
+      return;
+    }
+
+    media::BitstreamBuffer input_buffer(params.input_buffer_id,
+                                        params.input_buffer_handle,
+                                        params.input_buffer_size);
+
+    scoped_ptr<base::SharedMemory> output_shm(
+        new base::SharedMemory(params.output_video_frame_handle, false));
+    if (!output_shm->Map(params.output_buffer_size)) {
+      LOG(ERROR) << "Could not map output shared memory for input buffer id "
+                 << params.input_buffer_id;
+      NotifyDecodeStatusOnIOThread(
+          *route_id, params.input_buffer_id,
+          media::JpegDecodeAccelerator::PLATFORM_FAILURE);
+      return;
+    }
+
+    uint8* shm_memory = reinterpret_cast<uint8*>(output_shm->memory());
+    scoped_refptr<media::VideoFrame> frame =
+        media::VideoFrame::WrapExternalPackedMemory(
+            media::VideoFrame::I420,
+            params.coded_size,
+            gfx::Rect(params.coded_size),
+            params.coded_size,
+            shm_memory,
+            params.output_buffer_size,
+            params.output_video_frame_handle,
+            0,
+            base::TimeDelta(),
+            base::Bind(DecodeFinished, base::Passed(&output_shm)));
+
+    if (!frame.get()) {
+      LOG(ERROR) << "Could not create VideoFrame for input buffer id "
+                 << params.input_buffer_id;
+      NotifyDecodeStatusOnIOThread(
+          *route_id, params.input_buffer_id,
+          media::JpegDecodeAccelerator::PLATFORM_FAILURE);
+      return;
+    }
+
+    DCHECK(client_map_.count(*route_id) > 0);

[Pawel Osciak, 2015/05/28 09:13:17]

DCHECK_GT

[kcwu, 2015/05/28 12:10:27]

Done.

+    Client* client = client_map_[*route_id];
+    client->Decode(input_buffer, frame);
+  }
+
+ protected:
+  ~MessageFilter() override {
+    // Make sure |client_map_| is deleted on child thread.
+    DCHECK(child_task_runner_->BelongsToCurrentThread());

[kcwu, 2015/05/27 14:13:23]

The ownership of |Client| is maintained by the filter. However, I am not sure
the filter is always deleted on child thread.

IIUC, if ~GpuJpegDecodeAccelerator doesn't wait for the filter, the filter may
be still live due to line 286 and ~MessageFilter is invoked on IO thread.

To solve this, I guess I have to PostTask a pure function to STLDeleteValues on
child thread.

[piman, 2015/05/27 21:08:40]

Correct, it could be deleted on either the child thread or the IO thread.
Correct.
Yes, that would be a reasonable option.

+    STLDeleteValues(&client_map_);
+  }
+
+ private:
+  base::WeakPtr<GpuJpegDecodeAccelerator> owner_;
+
+  // GPU child task runner.
+  scoped_refptr<base::SingleThreadTaskRunner> child_task_runner_;
+
+  // GPU IO task runner.
+  scoped_refptr<base::SingleThreadTaskRunner> io_task_runner_;
+
+  // The sender to which this filter was added.
+  IPC::Sender* sender_;
+
+  // A map from route id to JpegDecodeAccelerator.
+  // Unless in destructor (on child thread), |client_map_| should only be
+  // accessed on IO thread.
+  std::map<int32, Client*> client_map_;

[Pawel Osciak, 2015/05/28 09:13:17]

Could we use a scoper?

[kcwu, 2015/05/28 12:10:27]

Scoper implys it will automatically delete Client once the map is being deleted.
But the filter doesn't guarantee it is deleted on the right thread. So I prefer
to handle map's life cycle manually.

+};
+
+GpuJpegDecodeAccelerator::GpuJpegDecodeAccelerator(
+    GpuChannel* channel,
+    const scoped_refptr<base::SingleThreadTaskRunner>& io_task_runner)
+    : channel_(channel),
+      child_task_runner_(base::ThreadTaskRunnerHandle::Get()),
+      io_task_runner_(io_task_runner),
+      client_number_(0) {
+}
+
+GpuJpegDecodeAccelerator::~GpuJpegDecodeAccelerator() {
+  DCHECK(CalledOnValidThread());
+  if (client_number_ > 0)
+    channel_->RemoveFilter(filter_.get());
+}
+
+bool GpuJpegDecodeAccelerator::OnMessageReceived(const IPC::Message& msg) {
+  // Messages are actually handled in filter on IO thread.
+  return false;
+}
+
+void GpuJpegDecodeAccelerator::AddClient(int32 route_id,
+                                         IPC::Message* reply_msg) {
+  DCHECK(CalledOnValidThread());
+  scoped_ptr<media::JpegDecodeAccelerator> accelerator;
+
+// When adding more platforms, GpuJpegDecoder::Supported need
+// update as well.
+#if defined(OS_CHROMEOS) && defined(ARCH_CPU_X86_FAMILY)
+  accelerator.reset(new VaapiJpegDecodeAccelerator(io_task_runner_));
+#else
+  DVLOG(1) << "HW JPEG decode acceleration not available.";
+#endif
+
+  scoped_ptr<Client> client(new Client(this, route_id));
+  if (!accelerator.get() || !accelerator->Initialize(client.get())) {
+    DLOG(ERROR) << "JPEG accelerator Initialize failed";
+    GpuMsg_CreateJpegDecoder::WriteReplyParams(reply_msg, false);
+    Send(reply_msg);
+    return;
+  }
+  client->set_accelerator(accelerator.Pass());
+
+  if (client_number_ == 0) {
+    filter_ = new MessageFilter(this);
+    // This should be before AddClientOnIOThread.
+    channel_->AddFilter(filter_.get());
+  }
+  client_number_++;
+
+  io_task_runner_->PostTask(
+      FROM_HERE, base::Bind(&MessageFilter::AddClientOnIOThread, filter_,
+                            route_id, client.release(), reply_msg));
+}
+
+void GpuJpegDecodeAccelerator::NotifyDecodeStatus(
+    int32 route_id,
+    int32_t buffer_id,
+    media::JpegDecodeAccelerator::Error error) {
+  DCHECK(CalledOnValidThread());
+  Send(new AcceleratedJpegDecoderHostMsg_DecodeAck(route_id, buffer_id, error));
+}
+
+void GpuJpegDecodeAccelerator::ClientRemoved() {
+  DCHECK(CalledOnValidThread());
+  DCHECK_GT(client_number_, 0);
+  client_number_--;
+  if (client_number_ == 0) {
+    channel_->RemoveFilter(filter_.get());
+    filter_ = nullptr;
+  }
+}
+
+bool GpuJpegDecodeAccelerator::Send(IPC::Message* message) {
+  DCHECK(CalledOnValidThread());
+  return channel_->Send(message);
+}
+
+}  // namespace content