Add VaapiVideoEncodeAccelerator for HW-accelerated video encode.

content/common/gpu/media/vaapi_video_encode_accelerator.h

Index: content/common/gpu/media/vaapi_video_encode_accelerator.h
diff --git a/content/common/gpu/media/vaapi_video_encode_accelerator.h b/content/common/gpu/media/vaapi_video_encode_accelerator.h
new file mode 100644
index 0000000000000000000000000000000000000000..23dd58cce1682ea2cb3291265b63629575c53e63
--- /dev/null
+++ b/content/common/gpu/media/vaapi_video_encode_accelerator.h
@@ -0,0 +1,308 @@
+// Copyright (c) 2014 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.
+
+#ifndef CONTENT_COMMON_GPU_MEDIA_VAAPI_VIDEO_ENCODE_ACCELERATOR_H_
+#define CONTENT_COMMON_GPU_MEDIA_VAAPI_VIDEO_ENCODE_ACCELERATOR_H_
+
+#include <list>
+#include <queue>
+
+#include "base/memory/linked_ptr.h"
+#include "base/threading/thread.h"
+#include "content/common/content_export.h"
+#include "content/common/gpu/media/h264_dpb.h"
+#include "content/common/gpu/media/va_surface.h"
+#include "content/common/gpu/media/vaapi_wrapper.h"
+#include "media/filters/h264_bitstream_buffer.h"
+#include "media/video/video_encode_accelerator.h"
+
+namespace content {
+
+// A VideoEncodeAccelerator implementation that uses VA-API
+// (http://www.freedesktop.org/wiki/Software/vaapi) for HW-accelerated
+// video encode.
+class CONTENT_EXPORT VaapiVideoEncodeAccelerator
+    : public media::VideoEncodeAccelerator {
+ public:
+  explicit VaapiVideoEncodeAccelerator(Display* x_display);
+  virtual ~VaapiVideoEncodeAccelerator();
+
+  // media::VideoEncodeAccelerator implementation.
+  virtual bool Initialize(media::VideoFrame::Format format,
+                          const gfx::Size& input_visible_size,
+                          media::VideoCodecProfile output_profile,
+                          uint32 initial_bitrate,
+                          Client* client) OVERRIDE;
+  virtual void Encode(const scoped_refptr<media::VideoFrame>& frame,
+                      bool force_keyframe) OVERRIDE;
+  virtual void UseOutputBitstreamBuffer(
+      const media::BitstreamBuffer& buffer) OVERRIDE;
+  virtual void RequestEncodingParametersChange(uint32 bitrate,
+                                               uint32 framerate) OVERRIDE;
+  virtual void Destroy() OVERRIDE;
+  static std::vector<media::VideoEncodeAccelerator::SupportedProfile>
+      GetSupportedProfiles();
+
+  // UMA errors this class reports.
+  enum VAVEAEncoderFailure {
+    VAAPI_ERROR = 0,
+    VAVEA_ENCODER_FAILURES_MAX,
+  };
+
+ private:
+  enum {
+    // Need 2 surfaces for each frame: one for input data and one for
+    // reconstructed picture, which is later used for reference.
+    kMinSurfacesToEncode = 2,
+
+    // Subjectively chosen.
+    kNumInputBuffers = 4,
+    kMaxNumReferenceFrames = 4,
+
+    // We need up to kMaxNumReferenceFrames surfaces for reference, plus one
+    // for input and one for encode (which will be added to the set of reference
+    // frames for subsequent frames). Actual execution of HW encode is done
+    // in parallel, and we want to process more frames in the meantime.
+    // To have kNumInputBuffers in flight, we need a full set of reference +
+    // encode surfaces (i.e. kMaxNumReferenceFrames + kMinSurfacesToEncode), and
+    // (kNumInputBuffers - 1) of kMinSurfacesToEncode for the remaining frames
+    // in flight.
+    kNumSurfaces = kMaxNumReferenceFrames + kMinSurfacesToEncode +
+                   kMinSurfacesToEncode * (kNumInputBuffers - 1),

[wuchengli, 2014/06/19 07:31:48]

Is it simpler to use "kNumSurfaces = kMaxNumReferenceFrames +
kMinSurfacesToEncode * kNumInputBuffers"? It looks like we need
kMaxNumReferenceFrames and every input buffer needs kMinSurfacesToEncode.

[Pawel Osciak, 2014/06/19 10:59:59]

I felt this was easier to understand, because this way it matches the
description above exactly.

+
+    // An IDR every 128 frames, an I frame every 30 and no B frames.
+    kIDRPeriod = 128,
+    kIPeriod = 30,
+    kIPPeriod = 1,
+
+    kDefaultFramerate = 30,
+
+    // HRD parameters (ch. E.2.2 in spec).
+    kBitRateScale = 0,  // bit_rate_scale for SPS HRD parameters.
+    kCPBSizeScale = 0,  // cpb_size_scale for SPS HRD parameters.
+
+    kDefaultQP = 26,
+    // All Intel codecs can do at least 4.1.
+    kDefaultLevelIDC = 41,
+    kChromaFormatIDC = 1,  // 4:2:0
+
+    // Arbitrarily chosen bitrate window size for rate control, in ms.
+    kCPBWindowSizeMs = 1500,
+  };
+
+  // Reference picture list.
+  typedef std::list<scoped_refptr<VASurface> > RefPicList;
+
+  // Encode job for one frame. Created when an input frame is awaiting and
+  // enough resources are available to proceed. Once the job is prepared and
+  // submitted to the hardware, it awaits on the submitted_encode_jobs_ queue
+  // for an output bitstream buffer to become available. Once one is ready,
+  // the encoded bytes are downloaded to it and job resources are released
+  // and become available for reuse.
+  struct EncodeJob {
+    // Input surface for video frame data.
+    scoped_refptr<VASurface> input_surface;
+    // Surface for a reconstructed picture, which is used for reference
+    // for subsequent frames.
+    scoped_refptr<VASurface> recon_surface;
+    // Buffer that will contain output bitstream for this frame.
+    VABufferID coded_buffer;
+    // Reference surfaces required to encode this picture. We keep references
+    // to them here, because we may discard some of them from ref_pic_list*

[wuchengli, 2014/06/19 07:31:48]

You mean |ref_pic_list0_|?

[Pawel Osciak, 2014/06/19 10:59:59]

No, this is intentional. I want to suggest that when/if we add ref_pic_list1, it
should also be included here.

+    // before the HW job is done.
+    RefPicList reference_surfaces;
+    // True if this job will produce a keyframe. Used to report
+    // to BitstreamBufferReady().
+    bool keyframe;
+
+    EncodeJob();
+  };
+
+  // Encoder state.
+  enum State {
+    kUninitialized,
+    kEncoding,
+    kError,
+  };
+
+  // Holds input frames coming from the client ready to be encoded.
+  struct InputFrameRef;
+  // Holds output buffers coming from the client ready to be filled.
+  struct BitstreamBufferRef;
+
+  // Tasks for each of the VEA interface calls to be executed on the
+  // encoder thread.
+  void InitializeTask();
+  void EncodeTask(const scoped_refptr<media::VideoFrame>& frame,
+                  bool force_keyframe);
+  void UseOutputBitstreamBufferTask(scoped_ptr<BitstreamBufferRef> buffer_ref);
+  void RequestEncodingParametersChangeTask(uint32 bitrate, uint32 framerate);
+  void DestroyTask();
+
+  // Prepare and schedule an encode job if we have an input to encode
+  // and enough resources to proceed.
+  void EncodeFrameTask();
+
+  // Fill current_sps_/current_pps_ with current values.
+  void UpdateSPS();
+  void UpdatePPS();
+  void UpdateRates(uint32 bitrate, uint32 framerate);
+
+  // Generate packed SPS and PPS in packed_sps_/packed_pps_, using
+  // values in current_sps_/current_pps_.
+  void GeneratePackedSPS();
+  void GeneratePackedPPS();
+
+  // Check if we have sufficient resources for a new encode job, claim them and
+  // fill current_encode_job_ with them.
+  // Return false if we cannot start a new job yet, true otherwise.
+  bool PrepareNextJob();
+
+  // Begin a new frame, making it a keyframe if |force_keyframe| is true,
+  // updating current_pic_.
+  void BeginFrame(bool force_keyframe);
+
+  // End current frame, updating reference picture lists and storing current
+  // job in the jobs awaiting completion on submitted_encode_jobs_.
+  void EndFrame();
+
+  // Submit parameters for the current frame to the hardware.
+  bool SubmitFrameParameters();
+  // Submit keyframe headers to the hardware if the current frame is a keyframe.
+  bool SubmitHeadersIfNeeded();
+
+  // Upload image data from |frame| to the input surface for current job.
+  bool UploadFrame(const scoped_refptr<media::VideoFrame>& frame);
+
+  // Execute encode in hardware. This does not block and will return before
+  // the job is finished.
+  bool ExecuteEncode();
+
+  // Callback that returns a no longer used VASurfaceID to
+  // available_va_surface_ids_ for reuse.
+  void RecycleVASurfaceID(VASurfaceID va_surface_id);
+
+  // Tries to return a bitstream buffer if both a submitted job awaits to
+  // be completed and we have bitstream buffers from the client available
+  // to download the encoded data to.
+  void TryToReturnBitstreamBuffer();
+
+  // Puts the encoder into en error state and notifies client about the error.
+  void NotifyError(Error error);
+
+  // Sets the encoder state on the correct thread.
+  void SetState(State state);
+
+  // VaapiWrapper is the owner of all HW resources (surfaces and buffers)
+  // and will free them on destruction.
+  scoped_ptr<VaapiWrapper> vaapi_wrapper_;
+
+  // Input profile and sizes.
+  media::VideoCodecProfile profile_;
+  gfx::Size visible_size_;
+  gfx::Size coded_size_;  // Macroblock-aligned.
+  // Width/height in macroblocks.
+  unsigned int mb_width_;
+  unsigned int mb_height_;
+
+  // Maximum size of the reference list 0.
+  unsigned int max_ref_idx_l0_size_;
+
+  // Initial QP.
+  unsigned int qp_;
+
+  // IDR frame period.
+  unsigned int idr_period_;
+  // I frame period.
+  unsigned int i_period_;
+  // IP period, i.e. how often do we need to have either an I or a P frame in
+  // the stream. Period of 1 means we can have no B frames.
+  unsigned int ip_period_;
+
+  // Size in bytes required for input bitstream buffers.
+  size_t output_buffer_byte_size_;
+
+  Display* x_display_;
+
+  // All of the members below musty be accessed on the encoder_thread_,
+  // while it is running.
+
+  // Encoder state. Encode tasks will only run in kEncoding state.
+  State state_;
+
+  // frame_num to be used for the next frame.
+  unsigned int frame_num_;
+  // frame_num of the previous IDR.
+  unsigned int last_idr_frame_num_;
+
+  // Current bitrate in bps.
+  unsigned int bitrate_;
+  // Current bitrate in fps.
+  unsigned int framerate_;
+  // CPB size in bits, i.e. bitrate in kbps * window size in ms/1000.
+  unsigned int cpb_size_;
+  // True if the parameters have changed and we need to submit a keyframe
+  // with updated parameters.
+  bool encoding_parameters_changed_;
+
+  // Job currently being prepared for encode.
+  scoped_ptr<EncodeJob> current_encode_job_;
+
+  // Current SPS, PPS and their packed versions. Packed versions are their NALUs
+  // in AnnexB format *without* emulation prevention three-byte sequences
+  // (those will be added by the driver).
+  media::H264SPS current_sps_;
+  media::H264BitstreamBuffer packed_sps_;
+  media::H264PPS current_pps_;
+  media::H264BitstreamBuffer packed_pps_;
+
+  // Picture currently being prepared for encode.
+  H264Picture current_pic_;
+
+  // VA surfaces available for reuse.
+  std::vector<VASurfaceID> available_va_surface_ids_;
+
+  // VA buffers for coded frames.
+  std::vector<VABufferID> available_va_buffer_ids_;
+
+  // Currently active reference surfaces.
+  RefPicList ref_pic_list0_;
+
+  // Callback via which finished VA surfaces are returned to us.
+  VASurface::ReleaseCB va_surface_release_cb_;
+
+  // VideoFrames passed from the client, waiting to be encoded.
+  std::queue<linked_ptr<InputFrameRef> > encoder_input_queue_;
+
+  // BitstreamBuffers mapped, ready to be filled.
+  std::queue<linked_ptr<BitstreamBufferRef> > available_bitstream_buffers_;
+
+  // Jobs submitted for encode, awaiting bitstream buffers to become available.
+  std::queue<linked_ptr<EncodeJob> > submitted_encode_jobs_;
+
+  // Encoder thread. All tasks are executed on it.
+  base::Thread encoder_thread_;
+  scoped_refptr<base::MessageLoopProxy> encoder_thread_proxy_;
+
+  const scoped_refptr<base::MessageLoopProxy> child_message_loop_proxy_;
+
+  // To expose client callbacks from VideoEncodeAccelerator.
+  // NOTE: all calls to these objects *MUST* be executed on
+  // child_message_loop_proxy_.
+  scoped_ptr<base::WeakPtrFactory<Client> > client_ptr_factory_;
+  base::WeakPtr<Client> client_;
+
+  // WeakPtr to post from the encoder thread back to the ChildThread, as it may
+  // outlive this. Posting from the ChildThread using base::Unretained(this)
+  // to the encoder thread is safe, because this always outlives the encoder

[wuchengli, 2014/06/18 15:57:44]

s/this/|this|/

[Pawel Osciak, 2014/06/19 01:31:11]

Done.

+  // thread (it's a member of this class).
+  base::WeakPtr<VaapiVideoEncodeAccelerator> weak_this_;
+  base::WeakPtrFactory<VaapiVideoEncodeAccelerator> weak_this_ptr_factory_;
+
+  DISALLOW_COPY_AND_ASSIGN(VaapiVideoEncodeAccelerator);
+};
+
+}  // namespace content
+
+#endif  // CONTENT_COMMON_GPU_MEDIA_VAAPI_VIDEO_ENCODE_ACCELERATOR_H_