]> git.sesse.net Git - nageru/blobdiff - nageru/ffmpeg_capture.cpp
Stop leaking SRT sockets on disconnect.
[nageru] / nageru / ffmpeg_capture.cpp
index 2ecdf64c4096d8db46f11d1eb185ed93f79d2ac0..87d1c007e9142a07e21b2c49820c706709b45c26 100644 (file)
@@ -27,13 +27,21 @@ extern "C" {
 #include <utility>
 #include <vector>
 
+#include <Eigen/Core>
+#include <Eigen/LU>
+#include <movit/colorspace_conversion_effect.h>
+
 #include "bmusb/bmusb.h"
 #include "shared/ffmpeg_raii.h"
 #include "ffmpeg_util.h"
 #include "flags.h"
 #include "image_input.h"
 #include "ref_counted_frame.h"
-#include "timebase.h"
+#include "shared/timebase.h"
+
+#ifdef HAVE_SRT
+#include <srt/srt.h>
+#endif
 
 #define FRAME_SIZE (8 << 20)  // 8 MB.
 
@@ -41,6 +49,7 @@ using namespace std;
 using namespace std::chrono;
 using namespace bmusb;
 using namespace movit;
+using namespace Eigen;
 
 namespace {
 
@@ -134,10 +143,10 @@ AVPixelFormat decide_dst_format(AVPixelFormat src_format, bmusb::PixelFormat dst
        return av_get_pix_fmt(best_format);
 }
 
-YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *frame)
+YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *frame, bool is_mjpeg)
 {
        YCbCrFormat format;
-       AVColorSpace colorspace = av_frame_get_colorspace(frame);
+       AVColorSpace colorspace = frame->colorspace;
        switch (colorspace) {
        case AVCOL_SPC_BT709:
                format.luma_coefficients = YCBCR_REC_709;
@@ -191,18 +200,55 @@ YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *f
                format.cb_y_position = 1.0;
                break;
        default:
-               fprintf(stderr, "Unknown chroma location coefficient enum %d from FFmpeg; choosing Rec. 709.\n",
+               fprintf(stderr, "Unknown chroma location coefficient enum %d from FFmpeg; choosing center.\n",
                        frame->chroma_location);
                format.cb_x_position = 0.5;
                format.cb_y_position = 0.5;
                break;
        }
 
+       if (is_mjpeg && !format.full_range) {
+               // Limited-range MJPEG is only detected by FFmpeg whenever a special
+               // JPEG comment is set, which means that in practice, the stream is
+               // almost certainly generated by Futatabi. Override FFmpeg's forced
+               // MJPEG defaults (it disregards the values set in the mux) with what
+               // Futatabi sets.
+               format.luma_coefficients = YCBCR_REC_709;
+               format.cb_x_position = 0.0;
+               format.cb_y_position = 0.5;
+       }
+
        format.cr_x_position = format.cb_x_position;
        format.cr_y_position = format.cb_y_position;
        return format;
 }
 
+RGBTriplet get_neutral_color(AVDictionary *metadata)
+{
+       if (metadata == nullptr) {
+               return RGBTriplet(1.0f, 1.0f, 1.0f);
+       }
+       AVDictionaryEntry *entry = av_dict_get(metadata, "WhitePoint", nullptr, 0);
+       if (entry == nullptr) {
+               return RGBTriplet(1.0f, 1.0f, 1.0f);
+       }
+
+       unsigned x_nom, x_den, y_nom, y_den;
+       if (sscanf(entry->value, " %u:%u , %u:%u", &x_nom, &x_den, &y_nom, &y_den) != 4) {
+               fprintf(stderr, "WARNING: Unable to parse white point '%s', using default white point\n", entry->value);
+               return RGBTriplet(1.0f, 1.0f, 1.0f);
+       }
+
+       double x = double(x_nom) / x_den;
+       double y = double(y_nom) / y_den;
+       double z = 1.0 - x - y;
+
+       Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB);
+       Vector3d rgb = rgb_to_xyz_matrix.inverse() * Vector3d(x, y, z);
+
+       return RGBTriplet(rgb[0], rgb[1], rgb[2]);
+}
+
 }  // namespace
 
 FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned height)
@@ -215,12 +261,36 @@ FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned he
        avformat_network_init();  // In case someone wants this.
 }
 
+#ifdef HAVE_SRT
+FFmpegCapture::FFmpegCapture(int srt_sock, const string &stream_id)
+       : srt_sock(srt_sock),
+         width(0),  // Don't resize; SRT streams typically have stable resolution, and should behave much like regular cards in general.
+         height(0),
+         pixel_format(bmusb::PixelFormat_8BitYCbCrPlanar),
+         video_timebase{1, 1}
+{
+       if (stream_id.empty()) {
+               description = "SRT stream";
+       } else {
+               description = stream_id;
+       }
+       play_as_fast_as_possible = true;
+       play_once = true;
+       last_frame = steady_clock::now();
+}
+#endif
+
 FFmpegCapture::~FFmpegCapture()
 {
        if (has_dequeue_callbacks) {
                dequeue_cleanup_callback();
        }
-       avresample_free(&resampler);
+       swr_free(&resampler);
+#ifdef HAVE_SRT
+       if (srt_sock != -1) {
+               srt_close(srt_sock);
+       }
+#endif
 }
 
 void FFmpegCapture::configure_card()
@@ -262,12 +332,12 @@ std::map<uint32_t, VideoMode> FFmpegCapture::get_available_video_modes() const
        VideoMode mode;
 
        char buf[256];
-       snprintf(buf, sizeof(buf), "%ux%u", width, height);
+       snprintf(buf, sizeof(buf), "%ux%u", sws_last_width, sws_last_height);
        mode.name = buf;
        
        mode.autodetect = false;
-       mode.width = width;
-       mode.height = height;
+       mode.width = sws_last_width;
+       mode.height = sws_last_height;
        mode.frame_rate_num = 60;
        mode.frame_rate_den = 1;
        mode.interlaced = false;
@@ -288,22 +358,38 @@ void FFmpegCapture::producer_thread_func()
                        filename_copy = filename;
                }
 
-               string pathname = search_for_file(filename_copy);
+               string pathname;
+               if (srt_sock == -1) {
+                       pathname = search_for_file(filename_copy);
+               } else {
+                       pathname = description;
+               }
                if (pathname.empty()) {
-                       fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename_copy.c_str());
                        send_disconnected_frame();
+                       if (play_once) {
+                               break;
+                       }
                        producer_thread_should_quit.sleep_for(seconds(1));
+                       fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename_copy.c_str());
                        continue;
                }
                should_interrupt = false;
                if (!play_video(pathname)) {
                        // Error.
-                       fprintf(stderr, "Error when playing %s, sleeping one second and trying again...\n", pathname.c_str());
                        send_disconnected_frame();
+                       if (play_once) {
+                               break;
+                       }
+                       fprintf(stderr, "Error when playing %s, sleeping one second and trying again...\n", pathname.c_str());
                        producer_thread_should_quit.sleep_for(seconds(1));
                        continue;
                }
 
+               if (play_once) {
+                       send_disconnected_frame();
+                       break;
+               }
+
                // Probably just EOF, will exit the loop above on next test.
        }
 
@@ -317,19 +403,21 @@ void FFmpegCapture::send_disconnected_frame()
 {
        // Send an empty frame to signal that we have no signal anymore.
        FrameAllocator::Frame video_frame = video_frame_allocator->alloc_frame();
+       size_t frame_width = width == 0 ? global_flags.width : width;
+       size_t frame_height = height == 0 ? global_flags.height : height;
        if (video_frame.data) {
                VideoFormat video_format;
-               video_format.width = width;
-               video_format.height = height;
+               video_format.width = frame_width;
+               video_format.height = frame_height;
                video_format.frame_rate_nom = 60;
                video_format.frame_rate_den = 1;
                video_format.is_connected = false;
                if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
-                       video_format.stride = width * 4;
-                       video_frame.len = width * height * 4;
+                       video_format.stride = frame_width * 4;
+                       video_frame.len = frame_width * frame_height * 4;
                        memset(video_frame.data, 0, video_frame.len);
                } else {
-                       video_format.stride = width;
+                       video_format.stride = frame_width;
                        current_frame_ycbcr_format.luma_coefficients = YCBCR_REC_709;
                        current_frame_ycbcr_format.full_range = true;
                        current_frame_ycbcr_format.num_levels = 256;
@@ -339,9 +427,9 @@ void FFmpegCapture::send_disconnected_frame()
                        current_frame_ycbcr_format.cb_y_position = 0.0f;
                        current_frame_ycbcr_format.cr_x_position = 0.0f;
                        current_frame_ycbcr_format.cr_y_position = 0.0f;
-                       video_frame.len = width * height * 2;
-                       memset(video_frame.data, 0, width * height);
-                       memset(video_frame.data + width * height, 128, width * height);  // Valid for both NV12 and planar.
+                       video_frame.len = frame_width * frame_height * 2;
+                       memset(video_frame.data, 0, frame_width * frame_height);
+                       memset(video_frame.data + frame_width * frame_height, 128, frame_width * frame_height);  // Valid for both NV12 and planar.
                }
 
                frame_callback(-1, AVRational{1, TIMEBASE}, -1, AVRational{1, TIMEBASE}, timecode++,
@@ -349,6 +437,13 @@ void FFmpegCapture::send_disconnected_frame()
                        FrameAllocator::Frame(), /*audio_offset=*/0, AudioFormat());
                last_frame_was_connected = false;
        }
+
+       if (play_once) {
+               disconnected = true;
+               if (card_disconnected_callback != nullptr) {
+                       card_disconnected_callback();
+               }
+       }
 }
 
 bool FFmpegCapture::play_video(const string &pathname)
@@ -365,10 +460,23 @@ bool FFmpegCapture::play_video(const string &pathname)
                last_modified = buf.st_mtim;
        }
 
-       AVDictionary *opts = nullptr;
-       av_dict_set(&opts, "fflags", "nobuffer", 0);
-
-       auto format_ctx = avformat_open_input_unique(pathname.c_str(), nullptr, &opts, AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
+       AVFormatContextWithCloser format_ctx;
+       if (srt_sock == -1) {
+               // Regular file.
+               format_ctx = avformat_open_input_unique(pathname.c_str(), /*fmt=*/nullptr,
+                       /*options=*/nullptr,
+                       AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
+       } else {
+#ifdef HAVE_SRT
+               // SRT socket, already opened.
+               AVInputFormat *mpegts_fmt = av_find_input_format("mpegts");
+               format_ctx = avformat_open_input_unique(&FFmpegCapture::read_srt_thunk, this,
+                       mpegts_fmt, /*options=*/nullptr,
+                       AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
+#else
+               assert(false);
+#endif
+       }
        if (format_ctx == nullptr) {
                fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
                return false;
@@ -386,6 +494,8 @@ bool FFmpegCapture::play_video(const string &pathname)
        }
 
        int audio_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_AUDIO);
+       int subtitle_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_SUBTITLE);
+       has_last_subtitle = false;
 
        // Open video decoder.
        const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
@@ -407,6 +517,9 @@ bool FFmpegCapture::play_video(const string &pathname)
        unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
                video_codec_ctx.get(), avcodec_close);
 
+       // Used in decode_ycbcr_format().
+       is_mjpeg = video_codecpar->codec_id == AV_CODEC_ID_MJPEG;
+
        // Open audio decoder, if we have audio.
        AVCodecContextWithDeleter audio_codec_ctx;
        if (audio_stream_index != -1) {
@@ -438,18 +551,27 @@ bool FFmpegCapture::play_video(const string &pathname)
                if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) {
                        return true;
                }
+               if (should_interrupt.load()) {
+                       // Check as a failsafe, so that we don't need to rely on avio if we don't have to.
+                       return false;
+               }
                UniqueFrame audio_frame = audio_frame_allocator->alloc_frame();
                AudioFormat audio_format;
 
                int64_t audio_pts;
                bool error;
                AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(), audio_codec_ctx.get(),
-                       pathname, video_stream_index, audio_stream_index, audio_frame.get(), &audio_format, &audio_pts, &error);
+                       pathname, video_stream_index, audio_stream_index, subtitle_stream_index, audio_frame.get(), &audio_format, &audio_pts, &error);
                if (error) {
                        return false;
                }
                if (frame == nullptr) {
                        // EOF. Loop back to the start if we can.
+                       if (format_ctx->pb != nullptr && format_ctx->pb->seekable == 0) {
+                               // Not seekable (but seemingly, sometimes av_seek_frame() would return 0 anyway,
+                               // so don't try).
+                               return true;
+                       }
                        if (av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
                                fprintf(stderr, "%s: Rewind failed, not looping.\n", pathname.c_str());
                                return true;
@@ -472,6 +594,22 @@ bool FFmpegCapture::play_video(const string &pathname)
                }
 
                VideoFormat video_format = construct_video_format(frame.get(), video_timebase);
+               if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
+                       // Invalid frame rate; try constructing it from the previous frame length.
+                       // (This is especially important if we are the master card, for SRT,
+                       // since it affects audio. Not all senders have good timebases
+                       // (e.g., Larix rounds first to timebase 1000 and then multiplies by
+                       // 90 from there, it seems), but it's much better to have an oscillating
+                       // value than just locking at 60.
+                       if (last_pts != 0 && frame->pts > last_pts) {
+                               int64_t pts_diff = frame->pts - last_pts;
+                               video_format.frame_rate_nom = video_timebase.den;
+                               video_format.frame_rate_den = video_timebase.num * pts_diff;
+                       } else {
+                               video_format.frame_rate_nom = 60;
+                               video_format.frame_rate_den = 1;
+                       }
+               }
                UniqueFrame video_frame = make_video_frame(frame.get(), pathname, &error);
                if (error) {
                        return false;
@@ -481,56 +619,67 @@ bool FFmpegCapture::play_video(const string &pathname)
                        if (last_pts == 0 && pts_origin == 0) {
                                pts_origin = frame->pts;        
                        }
-                       next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
-                       if (first_frame && last_frame_was_connected) {
-                               // If reconnect took more than one second, this is probably a live feed,
-                               // and we should reset the resampler. (Or the rate is really, really low,
-                               // in which case a reset on the first frame is fine anyway.)
-                               if (duration<double>(next_frame_start - last_frame).count() >= 1.0) {
-                                       last_frame_was_connected = false;
+                       steady_clock::time_point now = steady_clock::now();
+                       if (play_as_fast_as_possible) {
+                               video_frame->received_timestamp = now;
+                               audio_frame->received_timestamp = now;
+                               next_frame_start = now;
+                       } else {
+                               next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
+                               if (first_frame && last_frame_was_connected) {
+                                       // If reconnect took more than one second, this is probably a live feed,
+                                       // and we should reset the resampler. (Or the rate is really, really low,
+                                       // in which case a reset on the first frame is fine anyway.)
+                                       if (duration<double>(next_frame_start - last_frame).count() >= 1.0) {
+                                               last_frame_was_connected = false;
+                                       }
+                               }
+                               video_frame->received_timestamp = next_frame_start;
+
+                               // The easiest way to get all the rate conversions etc. right is to move the
+                               // audio PTS into the video PTS timebase and go from there. (We'll get some
+                               // rounding issues, but they should not be a big problem.)
+                               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));
                                }
-                       }
-                       video_frame->received_timestamp = next_frame_start;
-
-                       // The easiest way to get all the rate conversions etc. right is to move the
-                       // audio PTS into the video PTS timebase and go from there. (We'll get some
-                       // rounding issues, but they should not be a big problem.)
-                       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;
+                               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;
+                       if (play_as_fast_as_possible) {
+                               finished_wakeup = !producer_thread_should_quit.should_quit();
+                       } else {
+                               finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
                        }
-                       bool finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
                        if (finished_wakeup) {
                                if (audio_frame->len > 0) {
                                        assert(audio_pts != -1);
@@ -542,6 +691,7 @@ bool FFmpegCapture::play_video(const string &pathname)
                                        // audio discontinuity.)
                                        timecode += MAX_FPS * 2 + 1;
                                }
+                               last_neutral_color = get_neutral_color(frame->metadata);
                                frame_callback(frame->pts, video_timebase, audio_pts, audio_timebase, timecode++,
                                        video_frame.get_and_release(), 0, video_format,
                                        audio_frame.get_and_release(), 0, audio_format);
@@ -614,6 +764,7 @@ bool FFmpegCapture::process_queued_commands(AVFormatContext *format_ctx, const s
                        start = compute_frame_start(last_pts, pts_origin, video_timebase, start, rate);
                        pts_origin = last_pts;
                        rate = cmd.new_rate;
+                       play_as_fast_as_possible = (rate >= 10.0);
                        break;
                }
        }
@@ -625,7 +776,7 @@ namespace {
 }  // namespace
 
 AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx, AVCodecContext *audio_codec_ctx,
-       const std::string &pathname, int video_stream_index, int audio_stream_index,
+       const std::string &pathname, int video_stream_index, int audio_stream_index, int subtitle_stream_index,
        FrameAllocator::Frame *audio_frame, AudioFormat *audio_format, int64_t *audio_pts, bool *error)
 {
        *error = false;
@@ -661,6 +812,9 @@ AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCo
                                        *error = true;
                                        return AVFrameWithDeleter(nullptr);
                                }
+                       } else if (pkt.stream_index == subtitle_stream_index) {
+                               last_subtitle = string(reinterpret_cast<const char *>(pkt.data), pkt.size);
+                               has_last_subtitle = true;
                        }
                } else {
                        eof = true;  // Or error, but ignore that for the time being.
@@ -742,41 +896,43 @@ void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator::
            audio_avframe->format != last_src_format ||
            dst_format != last_dst_format ||
            channel_layout != last_channel_layout ||
-           av_frame_get_sample_rate(audio_avframe) != last_sample_rate) {
-               avresample_free(&resampler);
-               resampler = avresample_alloc_context();
+           audio_avframe->sample_rate != last_sample_rate) {
+               swr_free(&resampler);
+               resampler = swr_alloc_set_opts(nullptr,
+                                              /*out_ch_layout=*/AV_CH_LAYOUT_STEREO_DOWNMIX,
+                                              /*out_sample_fmt=*/dst_format,
+                                              /*out_sample_rate=*/OUTPUT_FREQUENCY,
+                                              /*in_ch_layout=*/channel_layout,
+                                              /*in_sample_fmt=*/AVSampleFormat(audio_avframe->format),
+                                              /*in_sample_rate=*/audio_avframe->sample_rate,
+                                              /*log_offset=*/0,
+                                              /*log_ctx=*/nullptr);
+
                if (resampler == nullptr) {
                        fprintf(stderr, "Allocating resampler failed.\n");
-                       exit(1);
+                       abort();
                }
 
-               av_opt_set_int(resampler, "in_channel_layout",  channel_layout,                             0);
-               av_opt_set_int(resampler, "out_channel_layout", AV_CH_LAYOUT_STEREO_DOWNMIX,                0);
-               av_opt_set_int(resampler, "in_sample_rate",     av_frame_get_sample_rate(audio_avframe),    0);
-               av_opt_set_int(resampler, "out_sample_rate",    OUTPUT_FREQUENCY,                           0);
-               av_opt_set_int(resampler, "in_sample_fmt",      audio_avframe->format,                      0);
-               av_opt_set_int(resampler, "out_sample_fmt",     dst_format,                                 0);
-
-               if (avresample_open(resampler) < 0) {
+               if (swr_init(resampler) < 0) {
                        fprintf(stderr, "Could not open resample context.\n");
-                       exit(1);
+                       abort();
                }
 
                last_src_format = AVSampleFormat(audio_avframe->format);
                last_dst_format = dst_format;
                last_channel_layout = channel_layout;
-               last_sample_rate = av_frame_get_sample_rate(audio_avframe);
+               last_sample_rate = audio_avframe->sample_rate;
        }
 
        size_t bytes_per_sample = (audio_format->bits_per_sample / 8) * 2;
        size_t num_samples_room = (audio_frame->size - audio_frame->len) / bytes_per_sample;
 
        uint8_t *data = audio_frame->data + audio_frame->len;
-       int out_samples = avresample_convert(resampler, &data, 0, num_samples_room,
-               const_cast<uint8_t **>(audio_avframe->data), audio_avframe->linesize[0], audio_avframe->nb_samples);
+       int out_samples = swr_convert(resampler, &data, num_samples_room,
+               const_cast<const uint8_t **>(audio_avframe->data), audio_avframe->nb_samples);
        if (out_samples < 0) {
                 fprintf(stderr, "Audio conversion failed.\n");
-                exit(1);
+                abort();
         }
 
        audio_frame->len += out_samples * bytes_per_sample;
@@ -785,23 +941,18 @@ void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator::
 VideoFormat FFmpegCapture::construct_video_format(const AVFrame *frame, AVRational video_timebase)
 {
        VideoFormat video_format;
-       video_format.width = width;
-       video_format.height = height;
+       video_format.width = frame_width(frame);
+       video_format.height = frame_height(frame);
        if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
-               video_format.stride = width * 4;
+               video_format.stride = frame_width(frame) * 4;
        } else if (pixel_format == FFmpegCapture::PixelFormat_NV12) {
-               video_format.stride = width;
+               video_format.stride = frame_width(frame);
        } else {
                assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
-               video_format.stride = width;
+               video_format.stride = frame_width(frame);
        }
        video_format.frame_rate_nom = video_timebase.den;
-       video_format.frame_rate_den = av_frame_get_pkt_duration(frame) * video_timebase.num;
-       if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
-               // Invalid frame rate.
-               video_format.frame_rate_nom = 60;
-               video_format.frame_rate_den = 1;
-       }
+       video_format.frame_rate_den = frame->pkt_duration * video_timebase.num;
        video_format.has_signal = true;
        video_format.is_connected = true;
        return video_format;
@@ -823,7 +974,7 @@ UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &
                sws_dst_format = decide_dst_format(AVPixelFormat(frame->format), pixel_format);
                sws_ctx.reset(
                        sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
-                               width, height, sws_dst_format,
+                               frame_width(frame), frame_height(frame), sws_dst_format,
                                SWS_BICUBIC, nullptr, nullptr, nullptr));
                sws_last_width = frame->width;
                sws_last_height = frame->height;
@@ -839,50 +990,81 @@ UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &
        int linesizes[4] = { 0, 0, 0, 0 };
        if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
                pic_data[0] = video_frame->data;
-               linesizes[0] = width * 4;
-               video_frame->len = (width * 4) * height;
+               linesizes[0] = frame_width(frame) * 4;
+               video_frame->len = (frame_width(frame) * 4) * frame_height(frame);
        } else if (pixel_format == PixelFormat_NV12) {
                pic_data[0] = video_frame->data;
-               linesizes[0] = width;
+               linesizes[0] = frame_width(frame);
 
-               pic_data[1] = pic_data[0] + width * height;
-               linesizes[1] = width;
+               pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
+               linesizes[1] = frame_width(frame);
 
-               video_frame->len = (width * 2) * height;
+               video_frame->len = (frame_width(frame) * 2) * frame_height(frame);
 
                const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
-               current_frame_ycbcr_format = decode_ycbcr_format(desc, frame);
+               current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg);
        } else {
                assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
                const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
 
-               int chroma_width = AV_CEIL_RSHIFT(int(width), desc->log2_chroma_w);
-               int chroma_height = AV_CEIL_RSHIFT(int(height), desc->log2_chroma_h);
+               int chroma_width = AV_CEIL_RSHIFT(int(frame_width(frame)), desc->log2_chroma_w);
+               int chroma_height = AV_CEIL_RSHIFT(int(frame_height(frame)), desc->log2_chroma_h);
 
                pic_data[0] = video_frame->data;
-               linesizes[0] = width;
+               linesizes[0] = frame_width(frame);
 
-               pic_data[1] = pic_data[0] + width * height;
+               pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
                linesizes[1] = chroma_width;
 
                pic_data[2] = pic_data[1] + chroma_width * chroma_height;
                linesizes[2] = chroma_width;
 
-               video_frame->len = width * height + 2 * chroma_width * chroma_height;
+               video_frame->len = frame_width(frame) * frame_height(frame) + 2 * chroma_width * chroma_height;
 
-               current_frame_ycbcr_format = decode_ycbcr_format(desc, frame);
+               current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg);
        }
        sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
 
        return video_frame;
 }
 
-int FFmpegCapture::interrupt_cb_thunk(void *unique)
+int FFmpegCapture::interrupt_cb_thunk(void *opaque)
 {
-       return reinterpret_cast<FFmpegCapture *>(unique)->interrupt_cb();
+       return reinterpret_cast<FFmpegCapture *>(opaque)->interrupt_cb();
 }
 
 int FFmpegCapture::interrupt_cb()
 {
        return should_interrupt.load();
 }
+
+unsigned FFmpegCapture::frame_width(const AVFrame *frame) const
+{
+       if (width == 0) {
+               return frame->width;
+       } else {
+               return width;
+       }
+}
+
+unsigned FFmpegCapture::frame_height(const AVFrame *frame) const
+{
+       if (height == 0) {
+               return frame->height;
+       } else {
+               return width;
+       }
+}
+
+#ifdef HAVE_SRT
+int FFmpegCapture::read_srt_thunk(void *opaque, uint8_t *buf, int buf_size)
+{
+       return reinterpret_cast<FFmpegCapture *>(opaque)->read_srt(buf, buf_size);
+}
+
+int FFmpegCapture::read_srt(uint8_t *buf, int buf_size)
+{
+       SRT_MSGCTRL mc = srt_msgctrl_default;
+       return srt_recvmsg2(srt_sock, reinterpret_cast<char *>(buf), buf_size, &mc);
+}
+#endif