X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=ffmpeg_capture.cpp;h=b7db50c5513beea1376b70031d608cab6059c154;hb=4a300e3cab7b1b1ef5a32e1f4a7ec319c48e95e5;hp=4f6e1685f282e358339a1e95dbf700525da719b1;hpb=7be71674af58d35d61928d2af6750efc5536bab6;p=nageru

diff --git a/ffmpeg_capture.cpp b/ffmpeg_capture.cpp
index 4f6e168..b7db50c 100644
--- a/ffmpeg_capture.cpp
+++ b/ffmpeg_capture.cpp
@@ -208,7 +208,6 @@ YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *f
 FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned height)
 	: filename(filename), width(width), height(height), video_timebase{1, 1}
 {
-	// Not really used for anything.
 	description = "Video: " + filename;
 
 	last_frame = steady_clock::now();
@@ -283,9 +282,15 @@ void FFmpegCapture::producer_thread_func()
 	pthread_setname_np(pthread_self(), thread_name);
 
 	while (!producer_thread_should_quit.should_quit()) {
-		string pathname = search_for_file(filename);
-		if (filename.empty()) {
-			fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename.c_str());
+		string filename_copy;
+		{
+			lock_guard<mutex> lock(filename_mu);
+			filename_copy = filename;
+		}
+
+		string pathname = search_for_file(filename_copy);
+		if (pathname.empty()) {
+			fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename_copy.c_str());
 			send_disconnected_frame();
 			producer_thread_should_quit.sleep_for(seconds(1));
 			continue;
@@ -491,6 +496,38 @@ bool FFmpegCapture::play_video(const string &pathname)
 			int64_t audio_pts_as_video_pts = av_rescale_q(audio_pts, audio_timebase, video_timebase);
 			audio_frame->received_timestamp = compute_frame_start(audio_pts_as_video_pts, pts_origin, video_timebase, start, rate);
 
+			if (audio_frame->len != 0) {
+				// The received timestamps in Nageru are measured after we've just received the frame.
+				// However, pts (especially audio pts) is at the _beginning_ of the frame.
+				// If we have locked audio, the distinction doesn't really matter, as pts is
+				// on a relative scale and a fixed offset is fine. But if we don't, we will have
+				// a different number of samples each time, which will cause huge audio jitter
+				// and throw off the resampler.
+				//
+				// In a sense, we should have compensated by adding the frame and audio lengths
+				// to video_frame->received_timestamp and audio_frame->received_timestamp respectively,
+				// but that would mean extra waiting in sleep_until(). All we need is that they
+				// are correct relative to each other, though (and to the other frames we send),
+				// so just align the end of the audio frame, and we're fine.
+				size_t num_samples = (audio_frame->len * 8) / audio_format.bits_per_sample / audio_format.num_channels;
+				double offset = double(num_samples) / OUTPUT_FREQUENCY -
+					double(video_format.frame_rate_den) / video_format.frame_rate_nom;
+				audio_frame->received_timestamp += duration_cast<steady_clock::duration>(duration<double>(offset));
+			}
+
+			steady_clock::time_point now = steady_clock::now();
+			if (duration<double>(now - next_frame_start).count() >= 0.1) {
+				// If we don't have enough CPU to keep up, or if we have a live stream
+				// where the initial origin was somehow wrong, we could be behind indefinitely.
+				// In particular, this will give the audio resampler problems as it tries
+				// to speed up to reduce the delay, hitting the low end of the buffer every time.
+				fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n",
+					pathname.c_str(),
+					1e3 * duration<double>(now - next_frame_start).count());
+				pts_origin = frame->pts;
+				start = next_frame_start = now;
+				timecode += MAX_FPS * 2 + 1;
+			}
 			bool finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
 			if (finished_wakeup) {
 				if (audio_frame->len > 0) {