Hacking ffvp9 decoder support for profiling.

media/filters/ffmpeg_video_decoder.cc

Index: media/filters/ffmpeg_video_decoder.cc
diff --git a/media/filters/ffmpeg_video_decoder.cc b/media/filters/ffmpeg_video_decoder.cc
index 92028dce636ab97364e71c1d09651689e3b5dd99..8f842dbd4825e7946178941921a4c1523316f626 100644
--- a/media/filters/ffmpeg_video_decoder.cc
+++ b/media/filters/ffmpeg_video_decoder.cc
@@ -16,6 +16,7 @@
 #include "base/location.h"
 #include "base/single_thread_task_runner.h"
 #include "base/strings/string_number_conversions.h"
+#include "base/sys_info.h"
 #include "media/base/bind_to_current_loop.h"
 #include "media/base/decoder_buffer.h"
 #include "media/base/limits.h"
@@ -25,6 +26,7 @@
 #include "media/base/video_util.h"
 #include "media/ffmpeg/ffmpeg_common.h"
 #include "media/filters/ffmpeg_glue.h"
+#include "third_party/ffmpeg/libavutil/intreadwrite.h"
 
 namespace media {
 
@@ -42,14 +44,27 @@ static const int kMaxDecodeThreads = 16;
 
 // Returns the number of threads given the FFmpeg CodecID. Also inspects the
 // command line for a valid --video-threads flag.
-static int GetThreadCount(AVCodecID codec_id) {
+static int GetThreadCount(const VideoDecoderConfig& config) {
   // Refer to http://crbug.com/93932 for tsan suppressions on decoding.
   int decode_threads = kDecodeThreads;
 
   const base::CommandLine* cmd_line = base::CommandLine::ForCurrentProcess();
   std::string threads(cmd_line->GetSwitchValueASCII(switches::kVideoThreads));
-  if (threads.empty() || !base::StringToInt(threads, &decode_threads))
+  if (threads.empty() || !base::StringToInt(threads, &decode_threads)) {
+    // TODO(chcunningham): Investigate perf with more threads.
+    if (config.codec() == kCodecVP9) {
+      // For VP9 decode when using the default thread count, increase the number
+      // of decode threads to equal the maximum number of tiles possible for
+      // higher resolution streams.
+      if (config.coded_size().width() >= 2048)
+        decode_threads = 8;
+      else if (config.coded_size().width() >= 1024)
+        decode_threads = 4;
+    }
+    decode_threads =
+        std::min(decode_threads, base::SysInfo::NumberOfProcessors());
     return decode_threads;
+  }
 
   decode_threads = std::max(decode_threads, 0);
   decode_threads = std::min(decode_threads, kMaxDecodeThreads);
@@ -75,7 +90,7 @@ bool FFmpegVideoDecoder::IsCodecSupported(VideoCodec codec) {
 }
 
 FFmpegVideoDecoder::FFmpegVideoDecoder()
-    : state_(kUninitialized), decode_nalus_(false) {
+    : state_(kUninitialized), decode_nalus_(false), parser_context_(nullptr) {
   thread_checker_.DetachFromThread();
 }
 
@@ -283,64 +298,88 @@ bool FFmpegVideoDecoder::FFmpegDecode(
     bool* has_produced_frame) {
   DCHECK(!*has_produced_frame);
 
+  size_t remaining_size = buffer->end_of_stream() ? 0 : buffer->data_size();
+
   // Create a packet for input data.
   // Due to FFmpeg API changes we no longer have const read-only pointers.
   AVPacket packet;
-  av_init_packet(&packet);
-  if (buffer->end_of_stream()) {
-    packet.data = NULL;
-    packet.size = 0;
-  } else {
-    packet.data = const_cast<uint8_t*>(buffer->data());
-    packet.size = buffer->data_size();
 
-    // Let FFmpeg handle presentation timestamp reordering.
-    codec_context_->reordered_opaque = buffer->timestamp().InMicroseconds();
-  }
+  do {
+    av_init_packet(&packet);
+    if (buffer->end_of_stream()) {
+      packet.data = NULL;
+      packet.size = 0;
+    } else {
+      // FFvp9 can't handle superframes, which are multiple frames inside a
+      // single packet. Normally its demuxer handles this parsing, but when
+      // ChunkDemuxer is used, the packet isn't split right.
+      if (codec_context_->codec_id == AV_CODEC_ID_VP9) {
+        if (!parser_context_) {
+          parser_context_.reset(av_parser_init(AV_CODEC_ID_VP9));
+        }
+        // TODO(chcunningham): The PTS/DTS stuff is a bit weird, seems mostly
+        // unused in internal parsing code. Setting NOPTS for now
+        // I think superframes have no presentation, so it should be fine to
+        // set the same timestamp (like alt-ref?)
+        int len = av_parser_parse2(
+            parser_context_.get(), codec_context_.get(), &packet.data,
+            &packet.size,
+            buffer->writable_data() + (buffer->data_size() - remaining_size),
+            remaining_size, AV_NOPTS_VALUE, AV_NOPTS_VALUE, -1);
+        remaining_size -= len;
+      } else {
+        packet.data = const_cast<uint8_t*>(buffer->data());
+        packet.size = buffer->data_size();
+        remaining_size = 0;
+      }
+
+      // Let FFmpeg handle presentation timestamp reordering.
+      codec_context_->reordered_opaque = buffer->timestamp().InMicroseconds();
+    }
 
-  int frame_decoded = 0;
-  int result = avcodec_decode_video2(codec_context_.get(),
-                                     av_frame_.get(),
-                                     &frame_decoded,
-                                     &packet);
-  // Log the problem if we can't decode a video frame and exit early.
-  if (result < 0) {
-    LOG(ERROR) << "Error decoding video: " << buffer->AsHumanReadableString();
-    return false;
-  }
+    int got_picture = 0;
+    int result = avcodec_decode_video2(codec_context_.get(), av_frame_.get(),
+                                       &got_picture, &packet);
+    // Log the problem if we can't decode a video frame and exit early.
+    if (result < 0) {
+      LOG(ERROR) << "Error decoding video: " << buffer->AsHumanReadableString();
+      return false;
+    }
 
-  // FFmpeg says some codecs might have multiple frames per packet.  Previous
-  // discussions with rbultje@ indicate this shouldn't be true for the codecs
-  // we use.
-  DCHECK_EQ(result, packet.size);
-
-  // If no frame was produced then signal that more data is required to
-  // produce more frames. This can happen under two circumstances:
-  //   1) Decoder was recently initialized/flushed
-  //   2) End of stream was reached and all internal frames have been output
-  if (frame_decoded == 0) {
-    return true;
-  }
+    // FFmpeg says some codecs might have multiple frames per packet.  Previous
+    // discussions with rbultje@ indicate this shouldn't be true for the codecs
+    // we use.
+    DCHECK_EQ(result, packet.size);
+
+    // If no picture was produced then signal that more data is required to
+    // produce more frames. This can happen under three circumstances:
+    //   1) Decoder was recently initialized/flushed
+    //   2) End of stream was reached and all internal frames have been output
+    //   3) The frame is needed for decoding dependencies, but is not rendered
+    if (got_picture == 0) {
+      continue;
+    }
 
-  // TODO(fbarchard): Work around for FFmpeg http://crbug.com/27675
-  // 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.";
-    av_frame_unref(av_frame_.get());
-    return false;
-  }
+    // TODO(fbarchard): Work around for FFmpeg http://crbug.com/27675
+    // 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.";
+      av_frame_unref(av_frame_.get());
+      return false;
+    }
 
-  scoped_refptr<VideoFrame> frame =
-      reinterpret_cast<VideoFrame*>(av_buffer_get_opaque(av_frame_->buf[0]));
-  frame->set_timestamp(
-      base::TimeDelta::FromMicroseconds(av_frame_->reordered_opaque));
-  *has_produced_frame = true;
-  output_cb_.Run(frame);
+    scoped_refptr<VideoFrame> frame =
+        reinterpret_cast<VideoFrame*>(av_buffer_get_opaque(av_frame_->buf[0]));
+    frame->set_timestamp(
+        base::TimeDelta::FromMicroseconds(av_frame_->reordered_opaque));
+    *has_produced_frame = true;
+    output_cb_.Run(frame);
 
-  av_frame_unref(av_frame_.get());
+    av_frame_unref(av_frame_.get());
+  } while (remaining_size > 0);
   return true;
 }
 
@@ -360,7 +399,7 @@ bool FFmpegVideoDecoder::ConfigureDecoder(bool low_delay) {
   codec_context_.reset(avcodec_alloc_context3(NULL));
   VideoDecoderConfigToAVCodecContext(config_, codec_context_.get());
 
-  codec_context_->thread_count = GetThreadCount(codec_context_->codec_id);
+  codec_context_->thread_count = GetThreadCount(config_);
   codec_context_->thread_type = low_delay ? FF_THREAD_SLICE : FF_THREAD_FRAME;
   codec_context_->opaque = this;
   codec_context_->flags |= CODEC_FLAG_EMU_EDGE;
@@ -376,6 +415,10 @@ bool FFmpegVideoDecoder::ConfigureDecoder(bool low_delay) {
     return false;
   }
 
+  LOG(ERROR) << "Using ffmpeg!";
+  LOG(ERROR) << "Configured FFmpeg decoder with thread count:"
+             << codec_context_->thread_count;
+
   av_frame_.reset(av_frame_alloc());
   return true;
 }