Flush and reset internal state after seek

This commit adds the following behavior to `seek` so that `seek`
works after a frame is decoded.

1. Flush the decoder buffer.
2. Recreate filter graphs (so that internal state is re-initialized)
3. Discard the buffered tensor. (decoded chunks)

Also it disallows negative values for seek timestamp.

torchaudio/csrc/ffmpeg/buffer.cpp

@@ -285,5 +285,9 @@ torch::Tensor Buffer::pop_all() {
   return torch::cat(ret, 0);
 }
 
+void Buffer::flush() {
+  chunks.clear();
+}
+
 } // namespace ffmpeg
 } // namespace torchaudio

torchaudio/csrc/ffmpeg/buffer.h

@@ -40,6 +40,8 @@ class Buffer {
 
   c10::optional<torch::Tensor> pop_chunk();
 
+  void flush();
+
  private:
   virtual torch::Tensor pop_one_chunk() = 0;
   torch::Tensor pop_all();

torchaudio/csrc/ffmpeg/decoder.cpp

@@ -16,5 +16,9 @@ int Decoder::get_frame(AVFrame* pFrame) {
   return avcodec_receive_frame(pCodecContext, pFrame);
 }
 
+void Decoder::flush_buffer() {
+  avcodec_flush_buffers(pCodecContext);
+}
+
 } // namespace ffmpeg
 } // namespace torchaudio

torchaudio/csrc/ffmpeg/decoder.h

@@ -24,6 +24,8 @@ class Decoder {
   int process_packet(AVPacket* pPacket);
   // Fetch a decoded frame
   int get_frame(AVFrame* pFrame);
+  // Flush buffer (for seek)
+  void flush_buffer();
 };
 
 } // namespace ffmpeg

torchaudio/csrc/ffmpeg/ffmpeg.cpp

@@ -156,5 +156,9 @@ AVFilterGraph* get_filter_graph() {
 } // namespace
 AVFilterGraphPtr::AVFilterGraphPtr()
     : Wrapper<AVFilterGraph, AVFilterGraphDeleter>(get_filter_graph()) {}
+
+void AVFilterGraphPtr::reset() {
+  ptr.reset(get_filter_graph());
+}
 } // namespace ffmpeg
 } // namespace torchaudio

torchaudio/csrc/ffmpeg/ffmpeg.h

@@ -129,6 +129,7 @@ struct AVFilterGraphDeleter {
 };
 struct AVFilterGraphPtr : public Wrapper<AVFilterGraph, AVFilterGraphDeleter> {
   AVFilterGraphPtr();
+  void reset();
 };
 } // namespace ffmpeg
 } // namespace torchaudio

torchaudio/csrc/ffmpeg/filter_graph.cpp

@@ -8,11 +8,11 @@ FilterGraph::FilterGraph(
     AVRational time_base,
     AVCodecParameters* codecpar,
     std::string filter_description)
-    : filter_description(filter_description) {
-  add_src(time_base, codecpar);
-  add_sink();
-  add_process();
-  create_filter();
+    : input_time_base(time_base),
+      codecpar(codecpar),
+      filter_description(filter_description),
+      media_type(codecpar->codec_type) {
+  init();
 }
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -62,18 +62,29 @@ std::string get_video_src_args(
 
 } // namespace
 
-void FilterGraph::add_src(AVRational time_base, AVCodecParameters* codecpar) {
-  if (media_type != AVMEDIA_TYPE_UNKNOWN) {
-    throw std::runtime_error("Source buffer is already allocated.");
-  }
-  media_type = codecpar->codec_type;
+void FilterGraph::init() {
+  add_src();
+  add_sink();
+  add_process();
+  create_filter();
+}
+
+void FilterGraph::reset() {
+  pFilterGraph.reset();
+  buffersrc_ctx = nullptr;
+  buffersink_ctx = nullptr;
+
+  init();
+}
+
+void FilterGraph::add_src() {
   std::string args;
   switch (media_type) {
     case AVMEDIA_TYPE_AUDIO:
-      args = get_audio_src_args(time_base, codecpar);
+      args = get_audio_src_args(input_time_base, codecpar);
       break;
     case AVMEDIA_TYPE_VIDEO:
-      args = get_video_src_args(time_base, codecpar);
+      args = get_video_src_args(input_time_base, codecpar);
       break;
     default:
       throw std::runtime_error("Only audio/video are supported.");

torchaudio/csrc/ffmpeg/filter_graph.h

@@ -5,13 +5,19 @@ namespace torchaudio {
 namespace ffmpeg {
 
 class FilterGraph {
-  AVMediaType media_type = AVMEDIA_TYPE_UNKNOWN;
+  // Parameters are kept for reset (recreating the underlying filter_graph struct)
+  AVRational input_time_base;
+  AVCodecParameters* codecpar;
+  std::string filter_description;
+
+  // constant.
+  AVMediaType media_type;
+
   AVFilterGraphPtr pFilterGraph;
   // AVFilterContext is freed as a part of AVFilterGraph
   // so we do not manage the resource.
   AVFilterContext* buffersrc_ctx = nullptr;
   AVFilterContext* buffersink_ctx = nullptr;
-  const std::string filter_description;
 
  public:
   FilterGraph(
@@ -35,8 +41,12 @@ class FilterGraph {
   //////////////////////////////////////////////////////////////////////////////
   // Configuration methods
   //////////////////////////////////////////////////////////////////////////////
+  void init();
+
+  void reset();
+
  private:
-  void add_src(AVRational time_base, AVCodecParameters* codecpar);
+  void add_src();
 
   void add_sink();
 

torchaudio/csrc/ffmpeg/sink.cpp

@@ -53,5 +53,11 @@ int Sink::process_frame(AVFrame* pFrame) {
 bool Sink::is_buffer_ready() const {
   return buffer->is_ready();
 }
+
+void Sink::flush() {
+  filter.reset();
+  buffer->flush();
+}
+
 } // namespace ffmpeg
 } // namespace torchaudio

torchaudio/csrc/ffmpeg/sink.h

@@ -22,6 +22,8 @@ class Sink {
 
   int process_frame(AVFrame* frame);
   bool is_buffer_ready() const;
+
+  void flush();
 };
 
 } // namespace ffmpeg

torchaudio/csrc/ffmpeg/stream_processor.cpp

@@ -82,6 +82,13 @@ int StreamProcessor::process_packet(AVPacket* packet) {
   return ret;
 }
 
+void StreamProcessor::flush() {
+  decoder.flush_buffer();
+  for (auto& ite : sinks) {
+    ite.second.flush();
+  }
+}
+
 // 0: some kind of success
 // <0: Some error happened
 int StreamProcessor::send_frame(AVFrame* pFrame) {

torchaudio/csrc/ffmpeg/stream_processor.h

@@ -68,6 +68,10 @@ class StreamProcessor {
   // - Sending NULL will drain (flush) the internal
   int process_packet(AVPacket* packet);
 
+  // flush the internal buffer of decoder.
+  // To be use when seeking
+  void flush();
+
  private:
   int send_frame(AVFrame* pFrame);
 

torchaudio/csrc/ffmpeg/streamer.cpp

@@ -136,11 +136,20 @@ bool Streamer::is_buffer_ready() const {
 // Configure methods
 ////////////////////////////////////////////////////////////////////////////////
 void Streamer::seek(double timestamp) {
+  if (timestamp < 0) {
+    throw std::invalid_argument("timestamp must be non-negative.");
+  }
+
   int64_t ts = static_cast<int64_t>(timestamp * AV_TIME_BASE);
   int ret = avformat_seek_file(pFormatContext, -1, INT64_MIN, ts, INT64_MAX, 0);
   if (ret < 0) {
     throw std::runtime_error("Failed to seek. (" + av_err2string(ret) + ".)");
   }
+  for (const auto& it : processors) {
+    if (it) {
+      it->flush();
+    }
+  }
 }
 
 void Streamer::add_audio_stream(