1 // Kaeru (換える), a simple transcoder intended for use with Nageru.
2 // This is experimental code, not yet supported.
4 #include "audio_encoder.h"
5 #include "basic_stats.h"
8 #include "ffmpeg_capture.h"
11 #include "quittable_sleeper.h"
13 #include "x264_encoder.h"
22 using namespace bmusb;
23 using namespace movit;
25 using namespace std::chrono;
26 using namespace std::placeholders;
28 Mixer *global_mixer = nullptr;
29 X264Encoder *global_x264_encoder = nullptr;
31 BasicStats *global_basic_stats = nullptr;
32 QuittableSleeper should_quit;
33 MuxMetrics stream_mux_metrics;
35 int write_packet(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
37 static bool seen_sync_markers = false;
38 static string stream_mux_header;
39 HTTPD *httpd = (HTTPD *)opaque;
41 if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
42 seen_sync_markers = true;
43 } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
44 // We don't know if this is a keyframe or not (the muxer could
45 // avoid marking it), so we just have to make the best of it.
46 type = AVIO_DATA_MARKER_SYNC_POINT;
49 if (type == AVIO_DATA_MARKER_HEADER) {
50 stream_mux_header.append((char *)buf, buf_size);
51 httpd->set_header(stream_mux_header);
53 httpd->add_data((char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
58 unique_ptr<Mux> create_mux(HTTPD *httpd, AVOutputFormat *oformat, X264Encoder *x264_encoder, AudioEncoder *audio_encoder)
60 AVFormatContext *avctx = avformat_alloc_context();
61 avctx->oformat = oformat;
63 uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
64 avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, httpd, nullptr, nullptr, nullptr);
65 avctx->pb->write_data_type = &write_packet;
66 avctx->pb->ignore_boundary_point = 1;
67 avctx->flags = AVFMT_FLAG_CUSTOM_IO;
69 string video_extradata = x264_encoder->get_global_headers();
72 mux.reset(new Mux(avctx, global_flags.width, global_flags.height, Mux::CODEC_H264, video_extradata, audio_encoder->get_codec_parameters().get(), COARSE_TIMEBASE,
73 /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, { &stream_mux_metrics }));
74 stream_mux_metrics.init({{ "destination", "http" }});
78 void video_frame_callback(FFmpegCapture *video, X264Encoder *x264_encoder, AudioEncoder *audio_encoder,
79 int64_t video_pts, AVRational video_timebase,
80 int64_t audio_pts, AVRational audio_timebase,
82 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
83 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
85 if (video_pts >= 0 && video_frame.len > 0) {
86 ReceivedTimestamps ts;
87 ts.ts.push_back(steady_clock::now());
89 video_pts = av_rescale_q(video_pts, video_timebase, AVRational{ 1, TIMEBASE });
90 int64_t frame_duration = TIMEBASE * video_format.frame_rate_den / video_format.frame_rate_nom;
91 x264_encoder->add_frame(video_pts, frame_duration, video->get_current_frame_ycbcr_format().luma_coefficients, video_frame.data + video_offset, ts);
92 global_basic_stats->update(frame_num++, /*dropped_frames=*/0);
94 if (audio_frame.len > 0) {
95 // FFmpegCapture takes care of this for us.
96 assert(audio_format.num_channels == 2);
97 assert(audio_format.sample_rate == OUTPUT_FREQUENCY);
99 // TODO: Reduce some duplication against AudioMixer here.
100 size_t num_samples = audio_frame.len / (audio_format.bits_per_sample / 8);
101 vector<float> float_samples;
102 float_samples.resize(num_samples);
103 if (audio_format.bits_per_sample == 16) {
104 const int16_t *src = (const int16_t *)audio_frame.data;
105 float *dst = &float_samples[0];
106 for (size_t i = 0; i < num_samples; ++i) {
107 *dst++ = le16toh(*src++) * (1.0f / 32768.0f);
109 } else if (audio_format.bits_per_sample == 32) {
110 const int32_t *src = (const int32_t *)audio_frame.data;
111 float *dst = &float_samples[0];
112 for (size_t i = 0; i < num_samples; ++i) {
113 *dst++ = le32toh(*src++) * (1.0f / 2147483648.0f);
118 audio_pts = av_rescale_q(audio_pts, audio_timebase, AVRational{ 1, TIMEBASE });
119 audio_encoder->encode_audio(float_samples, audio_pts);
122 if (video_frame.owner) {
123 video_frame.owner->release_frame(video_frame);
125 if (audio_frame.owner) {
126 audio_frame.owner->release_frame(audio_frame);
130 void audio_frame_callback(Mux *mux, const AVPacket *pkt, AVRational timebase)
132 mux->add_packet(*pkt, pkt->pts, pkt->dts == AV_NOPTS_VALUE ? pkt->pts : pkt->dts, timebase, /*stream_index=*/1);
135 void adjust_bitrate(int signal)
137 int new_bitrate = global_flags.x264_bitrate;
138 if (signal == SIGUSR1) {
140 if (new_bitrate > 100000) {
141 fprintf(stderr, "Ignoring SIGUSR1, can't increase bitrate below 100000 kbit/sec (currently at %d kbit/sec)\n",
142 global_flags.x264_bitrate);
144 fprintf(stderr, "Increasing bitrate to %d kbit/sec due to SIGUSR1.\n", new_bitrate);
145 global_flags.x264_bitrate = new_bitrate;
146 global_x264_encoder->change_bitrate(new_bitrate);
148 } else if (signal == SIGUSR2) {
150 if (new_bitrate < 100) {
151 fprintf(stderr, "Ignoring SIGUSR2, can't decrease bitrate below 100 kbit/sec (currently at %d kbit/sec)\n",
152 global_flags.x264_bitrate);
154 fprintf(stderr, "Decreasing bitrate to %d kbit/sec due to SIGUSR2.\n", new_bitrate);
155 global_flags.x264_bitrate = new_bitrate;
156 global_x264_encoder->change_bitrate(new_bitrate);
161 void request_quit(int signal)
166 int main(int argc, char *argv[])
168 parse_flags(PROGRAM_KAERU, argc, argv);
169 if (optind + 1 != argc) {
170 usage(PROGRAM_KAERU);
173 global_flags.num_cards = 1; // For latency metrics.
176 avformat_network_init();
180 AVOutputFormat *oformat = av_guess_format(global_flags.stream_mux_name.c_str(), nullptr, nullptr);
181 assert(oformat != nullptr);
183 unique_ptr<AudioEncoder> audio_encoder;
184 if (global_flags.stream_audio_codec_name.empty()) {
185 audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
187 audio_encoder.reset(new AudioEncoder(global_flags.stream_audio_codec_name, global_flags.stream_audio_codec_bitrate, oformat));
190 unique_ptr<X264Encoder> x264_encoder(new X264Encoder(oformat));
191 unique_ptr<Mux> http_mux = create_mux(&httpd, oformat, x264_encoder.get(), audio_encoder.get());
192 if (global_flags.transcode_audio) {
193 audio_encoder->add_mux(http_mux.get());
195 x264_encoder->add_mux(http_mux.get());
196 global_x264_encoder = x264_encoder.get();
198 FFmpegCapture video(argv[optind], global_flags.width, global_flags.height);
199 video.set_pixel_format(FFmpegCapture::PixelFormat_NV12);
200 video.set_frame_callback(bind(video_frame_callback, &video, x264_encoder.get(), audio_encoder.get(), _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11));
201 if (!global_flags.transcode_audio) {
202 video.set_audio_callback(bind(audio_frame_callback, http_mux.get(), _1, _2));
204 video.configure_card();
205 video.start_bm_capture();
206 video.change_rate(2.0); // Be sure never to really fall behind, but also don't dump huge amounts of stuff onto x264.
208 BasicStats basic_stats(/*verbose=*/false, /*use_opengl=*/false);
209 global_basic_stats = &basic_stats;
210 httpd.start(global_flags.http_port);
212 signal(SIGUSR1, adjust_bitrate);
213 signal(SIGUSR2, adjust_bitrate);
214 signal(SIGINT, request_quit);
216 while (!should_quit.should_quit()) {
217 should_quit.sleep_for(hours(1000));
220 video.stop_dequeue_thread();
221 // Stop the x264 encoder before killing the mux it's writing to.
222 x264_encoder.reset();