Provide a Chrome-owned buffer to FFmpeg for video decoding, instead of

media/filters/ffmpeg_video_decoder.cc

 // Copyright (c) 2012 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 "media/filters/ffmpeg_video_decoder.h"
 
 #include "base/bind.h"
 #include "base/callback_helpers.h"
 #include "base/command_line.h"
 #include "base/message_loop.h"
@@ skipping 43 matching lines @@
     : message_loop_factory_cb_(message_loop_cb),
       message_loop_(NULL),
       state_(kUninitialized),
       codec_context_(NULL),
       av_frame_(NULL),
       frame_rate_numerator_(0),
       frame_rate_denominator_(0),
       decryptor_(NULL) {
 }
 
+int FFmpegVideoDecoder::GetVideoBuffer(AVFrame* frame) {
+  VideoFrame::Format format = PixelFormatToVideoFormat(codec_context_->pix_fmt);
+  if (format == VideoFrame::INVALID)
+    return AVERROR(EINVAL);
+  DCHECK(format == VideoFrame::YV12 || format == VideoFrame::YV16);
+
+  int width = codec_context_->width;
+  int height = codec_context_->height;
+  int ret;
+  if ((ret = av_image_check_size(width, height, 0, NULL)) < 0)
+    return ret;
+
+  scoped_refptr<VideoFrame> video_frame =
+      VideoFrame::CreateFrame(format, width, height, 16u,

[DaleCurtis, 2012/06/14 23:35:37]

Will this need to change to 32-bytes with AVX instructions? 

[rbultje1, 2012/06/14 23:42:07]

Yes and no. So the alignment here serves two purposes: consider something like
VP8 or H264, where the video is encoded in 16x16 macroblocks. If the video has a
non-16x16 size, we need to round it up to 16x16 sizes for video coding purposes,
and ignore the data outside the visible area. More concretely, if the video has
a size of 1x1 pixels, then we still decode 1x1 macroblocks (16x16 pixels) and
simply don't display everything outside the initial 1x1 pixels. But the extra
data needs to be allocated and is written to.

In addition to that, there's alignment and so on, which is only in horizontal
dimensions. That's currently also 16 (for SSE2), which is very convenient. For
AVX2, that will change to 32 which will make all of this a massive pain, since
then vertical alignment needs to be 16 but horizontal 32. Let's ignore that
until it comes up.

+                              kNoTimestamp(), kNoTimestamp());
+
+  for (int i = 0; i < 3; i++) {

[DaleCurtis, 2012/06/14 23:35:37]

s/3/VideoFrame::kMaxPlanes/ ?

+    frame->base[i] = video_frame->data(i);
+    frame->data[i] = video_frame->data(i);
+    frame->linesize[i] = video_frame->stride(i);
+  }
+
+  frame->opaque = video_frame.release();
+  frame->type = FF_BUFFER_TYPE_USER;
+  frame->pkt_pts = codec_context_->pkt ? codec_context_->pkt->pts :

[DaleCurtis, 2012/06/14 23:35:37]

Alignment seems funky, consider just using a 4-sp indent on the next line. Up to
you.

+                                         AV_NOPTS_VALUE;
+  frame->width = codec_context_->width;
+  frame->height = codec_context_->height;
+  frame->format = codec_context_->pix_fmt;
+
+  return 0;
+}
+
+static int GetVideoBufferImpl(AVCodecContext* s, AVFrame* frame) {
+  FFmpegVideoDecoder* vd = static_cast<FFmpegVideoDecoder*>(s->opaque);
+  return vd->GetVideoBuffer(frame);
+}
+
+static void ReleaseVideoBufferImpl(AVCodecContext* s, AVFrame* frame) {
+  // We're releasing the refenence to the buffer allocated in

[Ami GONE FROM CHROMIUM, 2012/06/15 06:53:30]

typo

+  // GetVideoBuffer() here, so the explicit Release() here is
+  // intentional.
+  scoped_refptr<VideoFrame> video_frame =
+      static_cast<VideoFrame*>(frame->opaque);
+  video_frame->Release();
+  memset(frame->data, 0, sizeof(frame->data));

[Ami GONE FROM CHROMIUM, 2012/06/15 06:53:30]

unnecessary?  (only do #if !NDEBUG ?)

[rbultje1, 2012/06/15 19:10:27]

We discussed this earlier, it's a FFmpeg API requirement in the callback. I'll
add a comment in the next revision.

+  frame->opaque = NULL;
+}
+
 void FFmpegVideoDecoder::Initialize(const scoped_refptr<DemuxerStream>& stream,
                                     const PipelineStatusCB& status_cb,
                                     const StatisticsCB& statistics_cb) {
   if (!message_loop_) {
     message_loop_ = message_loop_factory_cb_.Run();
     message_loop_factory_cb_.Reset();
 
     message_loop_->PostTask(FROM_HERE, base::Bind(
         &FFmpegVideoDecoder::Initialize, this,
         stream, status_cb, statistics_cb));
@@ skipping 23 matching lines @@
 
   // Initialize AVCodecContext structure.
   codec_context_ = avcodec_alloc_context();
   VideoDecoderConfigToAVCodecContext(config, codec_context_);
 
   // Enable motion vector search (potentially slow), strong deblocking filter
   // for damaged macroblocks, and set our error detection sensitivity.
   codec_context_->error_concealment = FF_EC_GUESS_MVS | FF_EC_DEBLOCK;
   codec_context_->err_recognition = AV_EF_CAREFUL;
   codec_context_->thread_count = GetThreadCount(codec_context_->codec_id);
+  codec_context_->opaque = this;
+  codec_context_->flags |= CODEC_FLAG_EMU_EDGE;
+  codec_context_->get_buffer = GetVideoBufferImpl;
+  codec_context_->release_buffer = ReleaseVideoBufferImpl;
 
   AVCodec* codec = avcodec_find_decoder(codec_context_->codec_id);
   if (!codec) {
     status_cb.Run(PIPELINE_ERROR_DECODE);
     return;
   }
 
   if (avcodec_open2(codec_context_, codec, NULL) < 0) {
     status_cb.Run(PIPELINE_ERROR_DECODE);
     return;
@@ skipping 238 matching lines @@
   // The decoder is in a bad state and not decoding correctly.
   // Checking for NULL avoids a crash in CopyPlane().
   if (!av_frame_->data[VideoFrame::kYPlane] ||
       !av_frame_->data[VideoFrame::kUPlane] ||
       !av_frame_->data[VideoFrame::kVPlane]) {
     LOG(ERROR) << "Video frame was produced yet has invalid frame data.";
     *video_frame = NULL;
     return false;
   }
 
-  // We've got a frame! Make sure we have a place to store it.
-  *video_frame = AllocateVideoFrame();
-  if (!(*video_frame)) {
-    LOG(ERROR) << "Failed to allocate video frame";
+  if (!av_frame_->opaque) {
+    LOG(ERROR) << "VideoFrame object associated with frame data not set.";
     return false;
   }
+  *video_frame = static_cast<VideoFrame *>(av_frame_->opaque);
 
   // Determine timestamp and calculate the duration based on the repeat picture
   // count.  According to FFmpeg docs, the total duration can be calculated as
   // follows:
   //   fps = 1 / time_base
   //
   //   duration = (1 / fps) + (repeat_pict) / (2 * fps)
   //            = (2 + repeat_pict) / (2 * fps)
   //            = (2 + repeat_pict) / (2 * (1 / time_base))
   DCHECK_LE(av_frame_->repeat_pict, 2);  // Sanity check.
   AVRational doubled_time_base;
   doubled_time_base.num = frame_rate_denominator_;
   doubled_time_base.den = frame_rate_numerator_ * 2;
 
   (*video_frame)->SetTimestamp(
       base::TimeDelta::FromMicroseconds(av_frame_->reordered_opaque));
   (*video_frame)->SetDuration(
       ConvertFromTimeBase(doubled_time_base, 2 + av_frame_->repeat_pict));
 
-  // Copy the frame data since FFmpeg reuses internal buffers for AVFrame
-  // output, meaning the data is only valid until the next
-  // avcodec_decode_video() call.
-  int y_rows = codec_context_->height;
-  int uv_rows = codec_context_->height;
-  if (codec_context_->pix_fmt == PIX_FMT_YUV420P) {
-    uv_rows /= 2;
-  }
-
-  CopyYPlane(av_frame_->data[0], av_frame_->linesize[0], y_rows, *video_frame);
-  CopyUPlane(av_frame_->data[1], av_frame_->linesize[1], uv_rows, *video_frame);
-  CopyVPlane(av_frame_->data[2], av_frame_->linesize[2], uv_rows, *video_frame);
-
   return true;
 }
 
 void FFmpegVideoDecoder::DeliverFrame(
     const scoped_refptr<VideoFrame>& video_frame) {
   // Reset the callback before running to protect against reentrancy.
   base::ResetAndReturn(&read_cb_).Run(kOk, video_frame);
 }
 
 void FFmpegVideoDecoder::ReleaseFFmpegResources() {
   if (codec_context_) {
     av_free(codec_context_->extradata);
     avcodec_close(codec_context_);
     av_free(codec_context_);
     codec_context_ = NULL;
   }
   if (av_frame_) {
     av_free(av_frame_);
     av_frame_ = NULL;
   }
 }
 
-scoped_refptr<VideoFrame> FFmpegVideoDecoder::AllocateVideoFrame() {
-  VideoFrame::Format format = PixelFormatToVideoFormat(codec_context_->pix_fmt);
-  size_t width = codec_context_->width;
-  size_t height = codec_context_->height;
-
-  return VideoFrame::CreateFrame(format, width, height,
-                                 kNoTimestamp(), kNoTimestamp());
-}
-
 }  // namespace media