MJPEG acceleration for video capture using VAAPI

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 "base/bind.h"
+#include "base/logging.h"
+#include "base/memory/shared_memory.h"
+#include "base/message_loop/message_loop_proxy.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)
+#if defined(ARCH_CPU_X86_FAMILY)

[mcasas, 2015/05/07 00:59:03]

#if defined(OS_CHROMEOS) && defined(ARCH_CPU_X86_FAMILY)

[kcwu, 2015/05/08 14:42:43]

Done.

+#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
+
+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 {
+    owner_->OnFilterRemoved();
+  }
+
+  bool OnMessageReceived(const IPC::Message& msg) override {
+    if (msg.routing_id() != host_route_id_)
+      return false;
+
+    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__;

[mcasas, 2015/05/07 00:59:03]

As we approach something stable you should remove
these __func__ printouts.

[kcwu, 2015/05/08 14:42:43]

Done.

+  DCHECK(child_message_loop_->BelongsToCurrentThread());
+  DCHECK(!jpeg_decode_accelerator_.get());
+
+  if (!channel_->AddRoute(host_route_id_, this)) {
+    LOG(ERROR) << "GpuJpegDecodeAccelerator::Initialize(): "
+                  "failed to add route";
+    return false;
+  }
+
+  filter_ = new MessageFilter(this, host_route_id_);
+  channel_->AddFilter(filter_.get());
+
+// When adding more platforms, GpuJpegDecodeAcceleratorAdapter::Supported need
+// update as well.
+#if defined(OS_CHROMEOS) && defined(ARCH_CPU_X86_FAMILY)
+  jpeg_decode_accelerator_.reset(
+      new VaapiJpegDecodeAccelerator(io_message_loop_));
+#else
+  DVLOG(1) << "HW JPEG decode acceleration not available.";
+  return false;

[mcasas, 2015/05/07 00:59:03]

Move these |return false| further up in the method,
i.e. why AddRoute() and AddFilter() if the platform
does not support jpeg_decode_accelerator_ anyway?

[kcwu, 2015/05/08 14:42:42]

Done.

+#endif
+
+  return jpeg_decode_accelerator_->Initialize(this);
+}
+
+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.
+  DVLOG(3) << __func__;
+  Send(new AcceleratedJpegDecoderHostMsg_VideoFrameReady(host_route_id_,
+                                                         bitstream_buffer_id));
+}
+
+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 live util

[mcasas, 2015/05/07 00:59:03]

s/live/lives/

[kcwu, 2015/05/08 14:42:42]

Done.

+  // decode finished.

[mcasas, 2015/05/07 00:59:03]

s/finished/finishes/

[kcwu, 2015/05/08 14:42:43]

Done.

+  DVLOG(3) << __func__;
+}
+
+void GpuJpegDecodeAccelerator::OnDecode(
+    const AcceleratedJpegDecoderMsg_Decode_Params& params) {
+  DVLOG(3) << __func__;
+  DCHECK(io_message_loop_->BelongsToCurrentThread());
+  DCHECK(jpeg_decode_accelerator_.get());
+  TRACE_EVENT0("jpeg", "GpuJpegDecodeAccelerator::OnDecode");
+
+  if (params.input_buffer_id < 0) {
+    LOG(ERROR) << "BitstreamBuffer id " << params.input_buffer_id
+               << " 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)) {
+    LOG(ERROR) << "Could not map output shared memory for input buffer id "
+               << 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,
+          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;
+    NotifyError(params.input_buffer_id,
+                media::JpegDecodeAccelerator::PLATFORM_FAILURE);
+    return;
+  }
+
+  jpeg_decode_accelerator_->Decode(input_buffer, frame);
+}
+
+void GpuJpegDecodeAccelerator::OnDestroy() {
+  DVLOG(3) << __func__;
+  DCHECK(child_message_loop_->BelongsToCurrentThread());
+  DCHECK(jpeg_decode_accelerator_.get());
+  Destroy();
+}
+
+void GpuJpegDecodeAccelerator::OnFilterRemoved() {
+  // We're destroying; cancel all callbacks.
+  filter_removed_.Signal();
+}
+
+void GpuJpegDecodeAccelerator::Destroy() {
+  DCHECK(child_message_loop_->BelongsToCurrentThread());
+  // 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()) {
+    channel_->RemoveFilter(filter_.get());
+    filter_removed_.Wait();
+  }
+
+  channel_->RemoveRoute(host_route_id_);
+  channel_->ReleaseJpegDecoder(host_route_id_);
+  jpeg_decode_accelerator_.reset();
+
+  delete this;
+}
+
+bool GpuJpegDecodeAccelerator::Send(IPC::Message* message) {
+  if (io_message_loop_->BelongsToCurrentThread())
+    return filter_->SendOnIOThread(message);
+  return channel_->Send(message);
+}
+
+}  // namespace content