Ak/skip flush frames

ffmpeg/decoder.c

@@ -1,6 +1,7 @@
 #include "transcoder.h"
 #include "decoder.h"
 #include "logging.h"
+#include "flushing.h"
 
 #include <libavutil/pixfmt.h>
 
@@ -70,7 +71,7 @@ int decode_in(struct input_ctx *ictx, AVPacket *pkt, AVFrame *frame, int *stream
 
   if (!ictx->first_pkt && pkt->flags & AV_PKT_FLAG_KEY && decoder == ictx->vc) {
     ictx->first_pkt = av_packet_clone(pkt);
-    ictx->first_pkt->pts = -1;
+    ictx->first_pkt->pts = FLUSH_FRAME_PTS;
   }
 
   ret = lpms_send_packet(ictx, decoder, pkt);

ffmpeg/encoder.c

@@ -1,5 +1,6 @@
 #include "encoder.h"
 #include "logging.h"
+#include "flushing.h"
 
 #include <libavcodec/avcodec.h>
 #include <libavfilter/buffersrc.h>
@@ -371,6 +372,10 @@ static int encode(AVCodecContext* encoder, AVFrame *frame, struct output_ctx* oc
     ret = avcodec_receive_packet(encoder, pkt);
     if (AVERROR(EAGAIN) == ret || AVERROR_EOF == ret) goto encode_cleanup;
     if (ret < 0) LPMS_ERR(encode_cleanup, "Error receiving packet from encoder");
+    if(pkt->pts == FLUSH_FRAME_PTS) {
+      // This is flush-sentinel frame, do not leak into output
+      return ret;
+    }
     ret = mux(pkt, encoder->time_base, octx, ost);
     if (ret < 0) goto encode_cleanup;
   }

ffmpeg/filter.c

@@ -1,5 +1,6 @@
 #include "filter.h"
 #include "logging.h"
+#include "flushing.h"
 
 #include <libavfilter/buffersrc.h>
 #include <libavfilter/buffersink.h>
@@ -314,7 +315,7 @@ int filtergraph_write(AVFrame *inf, struct input_ctx *ictx, struct output_ctx *o
   } else if (!filter->flushed) { // Flush Frame
     int ts_step;
     inf = (is_video) ? ictx->last_frame_v : ictx->last_frame_a;
-    inf->opaque = (void *) (INT64_MIN); // Store INT64_MIN as pts for flush frames
+    inf->opaque = (void *) (FLUSH_FRAME_PTS);
     filter->flushing = 1;
     if (is_video) {
       ts_step = av_rescale_q(1, av_inv_q(vst->r_frame_rate), vst->time_base);
@@ -344,6 +345,7 @@ int filtergraph_write(AVFrame *inf, struct input_ctx *ictx, struct output_ctx *o
 
 int filtergraph_read(struct input_ctx *ictx, struct output_ctx *octx, struct filter_ctx *filter, int is_video)
 {
+    int64_t original_input_pts;
     AVFrame *frame = filter->frame;
     av_frame_unref(frame);
 
@@ -353,22 +355,24 @@ int filtergraph_read(struct input_ctx *ictx, struct output_ctx *octx, struct fil
     if (AVERROR(EAGAIN) == ret || AVERROR_EOF == ret) return ret;
     else if (ret < 0) LPMS_ERR(fg_read_cleanup, "Error consuming the filtergraph");
 
-    if (frame && ((int64_t) frame->opaque == INT64_MIN)) {
-      // opaque being INT64_MIN means it's a flush packet
-      // don't set flushed flag in case this is a flush from a previous segment
+    original_input_pts = (int64_t)frame->opaque;
+
+    if (frame && ((int64_t) frame->opaque == FLUSH_FRAME_PTS)) {
+      // it's a flush packet. don't set flushed flag in case this is a flush from a previous segment
       if (filter->flushing) filter->flushed = 1;
       ret = lpms_ERR_FILTER_FLUSHED;
     } else if (frame && is_video && octx->fps.den) {
       // We set custom PTS as an input of the filtergraph so we need to
       // re-calculate our output PTS before passing it on to the encoder
       if (filter->pts_diff == INT64_MIN) {
-        int64_t pts = (int64_t)frame->opaque; // original input PTS
-        pts = av_rescale_q_rnd(pts, ictx->ic->streams[ictx->vi]->time_base, av_buffersink_get_time_base(filter->sink_ctx), AV_ROUND_NEAR_INF|AV_ROUND_PASS_MINMAX);
+        int64_t pts = av_rescale_q_rnd(original_input_pts, ictx->ic->streams[ictx->vi]->time_base, av_buffersink_get_time_base(filter->sink_ctx), AV_ROUND_NEAR_INF|AV_ROUND_PASS_MINMAX);
         // difference between rescaled input PTS and the segment's first frame PTS of the filtergraph output
         filter->pts_diff = pts - frame->pts;
       }
       frame->pts += filter->pts_diff; // Re-calculate by adding back this segment's difference calculated at start
     }
+
+    if(original_input_pts == FLUSH_FRAME_PTS) frame->pts = FLUSH_FRAME_PTS;
 fg_read_cleanup:
     return ret;
 }

ffmpeg/flushing.h

@@ -0,0 +1,9 @@
+
+#ifndef _LPMS_FLUSHING_H_
+#define _LPMS_FLUSHING_H_
+
+#include <stdint.h>
+
+#define FLUSH_FRAME_PTS INT64_MIN
+
+#endif // _LPMS_FLUSHING_H_

ffmpeg/nvidia_test.go

@@ -793,3 +793,13 @@ func TestTranscoder_Portrait(t *testing.T) {
 func TestNvidia_DiscontinuityAudioSegment(t *testing.T) {
 	discontinuityAudioSegment(t, Nvidia)
 }
+
+func TestTranscoder_FlushFrameLeak(t *testing.T) {
+	in := &TranscodeOptionsIn{Fname: "../data/stopwatch_8frames.ts", Accel: Nvidia}
+	profile := VideoProfile{Name: "p720", Bitrate: "3000000", Framerate: 5994, FramerateDen: 100, AspectRatio: "16:9", Resolution: "1280x720"}
+	out := []TranscodeOptions{{Profile: profile, Oname: "../data/stopwatch_8frames_720p.ts", Accel: Nvidia}}
+	result, err := Transcode3(in, out)
+	require.NoError(t, err)
+	require.Equal(t, 8, result.Decoded.Frames)
+	require.Equal(t, 8, result.Encoded[0].Frames)
+}

ffmpeg/transcoder.c

@@ -3,6 +3,7 @@
 #include "filter.h"
 #include "encoder.h"
 #include "logging.h"
+#include "flushing.h"
 
 #include <libavcodec/avcodec.h>
 #include <libavformat/avformat.h>
@@ -463,7 +464,7 @@ int handle_video_packet(struct transcode_thread *h, output_results *decoded_resu
     // very first video packet, keep it
     // TODO: this should be called first_video_pkt
     ictx->first_pkt = av_packet_clone(pkt);
-    ictx->first_pkt->pts = -1;
+    ictx->first_pkt->pts = FLUSH_FRAME_PTS;
   }
 
   // TODO: this could probably be done always, because it is a no-op if
@@ -983,7 +984,7 @@ struct transcode_thread* lpms_transcode_new() {
   // keep track of last dts in each stream.
   // used while transmuxing, to skip packets with invalid dts.
   for (int i = 0; i < MAX_OUTPUT_SIZE; i++) {
-    h->ictx.last_dts[i] = -1;
+    h->ictx.last_dts[i] = FLUSH_FRAME_PTS;
   }
   return h;
 }