MJPEG acceleration for video capture using VAAPI

content/common/gpu/media/gpu_jpeg_decode_accelerator.cc

Index: content/common/gpu/media/gpu_jpeg_decode_accelerator.cc
diff --git a/content/common/gpu/media/gpu_jpeg_decode_accelerator.cc b/content/common/gpu/media/gpu_jpeg_decode_accelerator.cc
new file mode 100644
index 0000000000000000000000000000000000000000..5354a21a21ee5e3106622a9ddd5a791bb20f4e73
--- /dev/null
+++ b/content/common/gpu/media/gpu_jpeg_decode_accelerator.cc
@@ -0,0 +1,251 @@
+// 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 <vector>
+
+#include "base/bind.h"
+#include "base/command_line.h"
+#include "base/logging.h"
+#include "base/message_loop/message_loop_proxy.h"
+#include "base/stl_util.h"
+#include "content/common/gpu/gpu_channel.h"
+#include "content/common/gpu/gpu_messages.h"
+#include "content/public/common/content_switches.h"
+#include "gpu/command_buffer/common/command_buffer.h"
+#include "ipc/ipc_message_macros.h"
+#include "ipc/ipc_message_utils.h"
+#include "ipc/message_filter.h"
+#include "media/base/limits.h"
+#include "media/filters/jpeg_parser.h"
+#include "ui/gfx/geometry/size.h"
+
+#if defined(OS_CHROMEOS)
+#if defined(ARCH_CPU_X86_FAMILY)
+#include "content/common/gpu/media/vaapi_jpeg_decode_accelerator.h"
+#endif  // defined(ARCH_CPU_X86_FAMILY)
+#endif
+
+namespace base {
+
+void DefaultDeleter<content::GpuJpegDecodeAccelerator>::operator()(
+    void* jpeg_decode_accelerator) const {
+  static_cast<content::GpuJpegDecodeAccelerator*>(jpeg_decode_accelerator)
+      ->Destroy();
+}
+
+}  // namespace base

[wuchengli, 2015/04/15 07:11:58]

add a blank line

[kcwu, 2015/04/16 14:38:28]

Done.

+namespace content {
+
+class GpuJpegDecodeAccelerator::MessageFilter : public IPC::MessageFilter {
+ public:
+  MessageFilter(GpuJpegDecodeAccelerator* owner, int32 host_route_id)
+      : owner_(owner), host_route_id_(host_route_id) {}
+
+  void OnChannelError() override { sender_ = NULL; }
+
+  void OnChannelClosing() override { sender_ = NULL; }
+
+  void OnFilterAdded(IPC::Sender* sender) override { sender_ = sender; }
+
+  void OnFilterRemoved() override {
+    // This will delete |owner_| and |this|.
+    owner_->OnFilterRemoved();
+  }
+
+  bool OnMessageReceived(const IPC::Message& msg) override {
+    if (msg.routing_id() != host_route_id_)
+      return false;
+    DVLOG(3) << __func__;

[wuchengli, 2015/04/15 07:11:58]

The log in OnDecode should be enough. This can be removed.

[kcwu, 2015/04/16 14:38:28]

Done.

+
+    IPC_BEGIN_MESSAGE_MAP(MessageFilter, msg)
+    IPC_MESSAGE_FORWARD(AcceleratedJpegDecoderMsg_Decode, owner_,
+                        GpuJpegDecodeAccelerator::OnDecode)
+    IPC_MESSAGE_UNHANDLED(return false;)
+    IPC_END_MESSAGE_MAP()
+    return true;
+  }
+
+  bool SendOnIOThread(IPC::Message* message) {
+    DCHECK(!message->is_sync());
+    if (!sender_) {
+      delete message;
+      return false;
+    }
+    return sender_->Send(message);
+  }
+
+ protected:
+  virtual ~MessageFilter() {}
+
+ private:
+  GpuJpegDecodeAccelerator* owner_;
+  int32 host_route_id_;
+  // The sender to which this filter was added.
+  IPC::Sender* sender_;
+};
+
+GpuJpegDecodeAccelerator::GpuJpegDecodeAccelerator(
+    GpuChannel* channel,
+    int32 host_route_id,
+    const scoped_refptr<base::MessageLoopProxy>& io_message_loop)
+    : channel_(channel),
+      host_route_id_(host_route_id),
+      filter_removed_(true, false),
+      io_message_loop_(io_message_loop) {
+  child_message_loop_ = base::MessageLoopProxy::current();
+}
+
+GpuJpegDecodeAccelerator::~GpuJpegDecodeAccelerator() {
+}
+
+bool GpuJpegDecodeAccelerator::OnMessageReceived(const IPC::Message& msg) {
+  bool handled = true;
+  IPC_BEGIN_MESSAGE_MAP(GpuJpegDecodeAccelerator, msg)
+    IPC_MESSAGE_HANDLER(AcceleratedJpegDecoderMsg_Destroy, OnDestroy)
+    IPC_MESSAGE_UNHANDLED(handled = false)
+  IPC_END_MESSAGE_MAP()
+  return handled;
+}
+
+bool GpuJpegDecodeAccelerator::Initialize() {
+  DVLOG(3) << __func__;
+  DCHECK(child_message_loop_->BelongsToCurrentThread());
+  DCHECK(!jpeg_decode_accelerator_.get());
+
+  if (!channel_->AddRoute(host_route_id_, this)) {
+    DLOG(ERROR) << "GpuJpegDecodeAccelerator::Initialize(): "
+                   "failed to add route";
+    return false;
+  }
+
+  filter_ = new MessageFilter(this, host_route_id_);
+  channel_->AddFilter(filter_.get());
+
+#if defined(OS_CHROMEOS) && defined(ARCH_CPU_X86_FAMILY)
+  jpeg_decode_accelerator_.reset(new VaapiJpegDecodeAccelerator());
+#else
+  DVLOG(1) << "HW JPEG decode acceleration not available.";
+  return false;
+#endif
+
+  if (!jpeg_decode_accelerator_->Initialize(this)) {
+    return false;
+  }
+  return true;
+}
+
+void GpuJpegDecodeAccelerator::NotifyError(
+    int32_t buffer_id,
+    media::JpegDecodeAccelerator::Error error) {
+  DVLOG(3) << __func__;
+  Send(new AcceleratedJpegDecoderHostMsg_NotifyError(host_route_id_, buffer_id,
+                                                     error));
+}
+
+void GpuJpegDecodeAccelerator::VideoFrameReady(int32_t bitstream_buffer_id) {
+  // This is called from JDA's decode thread.

[wuchengli, 2015/04/15 07:11:58]

This should be called from IO thread to reduce the latency. DCHECK io thread.

[kcwu, 2015/04/20 17:47:58]

I don't think this will reduce latency. The flow is eventually passed from
decode thread to IO thread. It doesn't matter that the caller of VideoFrameReady
PostTask first or ChannelProxy Send() PostTask internally.

+  DVLOG(3) << __func__;
+  Send(new AcceleratedJpegDecoderHostMsg_VideoFrameReady(host_route_id_,
+                                                         bitstream_buffer_id));
+}
+
+void DecodeFinished(scoped_ptr<base::SharedMemory> shm) {
+  // Do nothing. The purpose of this function is just to keep life-cycle of
+  // SharedMemory as VideoFrame

[wuchengli, 2015/04/15 07:11:58]

I don't understand. What do you mean?

[kcwu, 2015/04/20 17:47:58]

Done.
Comment updated.

+  DVLOG(3) << "DecodeFinished";

[wuchengli, 2015/04/15 07:11:58]

__func__

[kcwu, 2015/04/16 14:38:28]

Done.

+}
+
+void GpuJpegDecodeAccelerator::OnDecode(
+    const AcceleratedJpegDecoderMsg_Decode_Params& params) {
+  DVLOG(3) << __func__;
+  DCHECK(io_message_loop_->BelongsToCurrentThread());
+  DCHECK(jpeg_decode_accelerator_.get());
+
+  if (params.input_buffer_id < 0) {
+    DLOG(ERROR) << "BitstreamBuffer id " << params.input_buffer_id

[wuchengli, 2015/04/15 07:11:58]

s/DLOG/LOG/

[kcwu, 2015/04/16 14:38:28]

Done.

+                << " out of range";
+    NotifyError(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)) {
+    DLOG(ERROR) << "Could not map output shared memory for input buffer id "

[wuchengli, 2015/04/15 07:11:58]

s/DLOG/LOG/

[kcwu, 2015/04/16 14:38:28]

Done.

+                << params.input_buffer_id;
+    NotifyError(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,

[wuchengli, 2015/04/15 07:11:58]

We don't really care about the visible size here. Right? This is not passed back
or anything.

[kcwu, 2015/04/16 14:38:28]

Yes

+          shm_memory,
+          params.output_buffer_size,
+          params.output_video_frame_handle,
+          0,
+          base::TimeDelta(),
+          base::Bind(DecodeFinished, base::Passed(&output_shm)));
+
+  if (!frame.get()) {
+    DLOG(ERROR) << "Could not create VideoFrame for input buffer id "

[wuchengli, 2015/04/15 07:11:58]

s/DLOG/LOG/

[kcwu, 2015/04/16 14:38:28]

Done.

+                << params.input_buffer_id;
+    NotifyError(params.input_buffer_id,
+                media::JpegDecodeAccelerator::PLATFORM_FAILURE);
+    return;
+  }
+
+  jpeg_decode_accelerator_->Decode(input_buffer, frame);
+}
+
+void GpuJpegDecodeAccelerator::OnDestroy() {

[wuchengli, 2015/04/15 07:11:58]

DCHECK main thread

[kcwu, 2015/04/20 17:47:58]

Done.

+  DVLOG(3) << __func__;
+  DCHECK(jpeg_decode_accelerator_.get());
+  Destroy();
+}
+
+void GpuJpegDecodeAccelerator::OnFilterRemoved() {
+  // We're destroying; cancel all callbacks.
+  filter_removed_.Signal();
+}
+
+void GpuJpegDecodeAccelerator::Destroy() {

[wuchengli, 2015/04/15 07:11:57]

Do we need this method? Can this be combined with OnDestroy?

[kcwu, 2015/04/20 17:47:58]

There are two flows to destroy GJDA:
1. msg -> OnDestroy -> ... -> remove from gpu_channel
2. something wrong, GpuChannel want to destroy itself and also remove GJDA

Destroy need to be a separate method for case 2 to call.
I may revisit this when adding support for multiple jpeg decoders.

+  // We cannot destroy the JDA before the IO thread message filter is
+  // removed however, since we cannot service incoming messages with JDA gone.
+  // We cannot simply check for existence of JDA on IO thread though, because
+  // we don't want to synchronize the IO thread with the ChildThread.
+  // So we have to wait for the RemoveFilter callback here instead and remove
+  // the JDA after it arrives and before returning.
+  if (filter_.get()) {

[wuchengli, 2015/04/15 07:11:58]

|filter_| should always exist after Initialize. This can be removed.

[kcwu, 2015/04/20 17:47:58]

This is necessary for safety if GpuJpegDecodeAccelerator created and destroyed
without Initialize.

+    channel_->RemoveFilter(filter_.get());
+    filter_removed_.Wait();
+  }
+
+  channel_->RemoveRoute(host_route_id_);

[wuchengli, 2015/04/15 07:11:58]

This can be called in ReleaseJpegDecoder.

[kcwu, 2015/04/20 17:47:58]

AddRoute is in Initialize(). I'd like RemoveRoute is inside this class as well
since they are paired.

+  channel_->ReleaseJpegDecoder();
+  jpeg_decode_accelerator_.reset();
+
+  delete this;
+}
+
+bool GpuJpegDecodeAccelerator::Send(IPC::Message* message) {
+  if (filter_.get() && io_message_loop_->BelongsToCurrentThread())

[wuchengli, 2015/04/15 07:11:58]

Remove filter_.get() check.

[kcwu, 2015/04/20 17:47:58]

Done.

+    return filter_->SendOnIOThread(message);
+  return channel_->Send(message);
+}
+
+}  // namespace content