6 #include <movit/image_format.h>
13 #include <sys/socket.h>
20 #include <libavutil/mem.h>
21 #include <libavformat/avformat.h>
22 #include <libavformat/avio.h>
23 #include <libavutil/avutil.h>
24 #include <libavutil/rational.h>
27 #include "audio_encoder.h"
29 #include "av1_encoder.h"
32 #include "shared/ffmpeg_raii.h"
34 #include "shared/httpd.h"
35 #include "shared/mux.h"
36 #include "quicksync_encoder.h"
37 #include "shared/timebase.h"
38 #include "x264_encoder.h"
39 #include "video_encoder.h"
40 #include "shared/metrics.h"
41 #include "shared/ref_counted_gl_sync.h"
42 #include "shared/shared_defs.h"
44 class RefCountedFrame;
47 using namespace std::chrono;
48 using namespace movit;
52 string generate_local_dump_filename(int frame)
54 time_t now = time(NULL);
56 localtime_r(&now, &now_tm);
59 strftime(timestamp, sizeof(timestamp), "%F-%H%M%S%z", &now_tm);
61 // Use the frame number to disambiguate between two cuts starting
62 // on the same second.
64 snprintf(filename, sizeof(filename), "%s/%s%s-f%02d%s",
65 global_flags.recording_dir.c_str(),
66 LOCAL_DUMP_PREFIX, timestamp, frame % 100, LOCAL_DUMP_SUFFIX);
72 VideoEncoder::VideoEncoder(ResourcePool *resource_pool, QSurface *surface, const std::string &va_display, int width, int height, HTTPD *httpd, DiskSpaceEstimator *disk_space_estimator)
73 : resource_pool(resource_pool), surface(surface), va_display(va_display), width(width), height(height), httpd(httpd), disk_space_estimator(disk_space_estimator)
75 // TODO: If we're outputting AV1, we can't use MPEG-TS currently.
76 srt_oformat = av_guess_format("mpegts", nullptr, nullptr);
77 assert(srt_oformat != nullptr);
79 oformat = av_guess_format(global_flags.stream_mux_name.c_str(), nullptr, nullptr);
80 assert(oformat != nullptr);
81 if (global_flags.stream_audio_codec_name.empty()) {
82 stream_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
84 stream_audio_encoder.reset(new AudioEncoder(global_flags.stream_audio_codec_name, global_flags.stream_audio_codec_bitrate, oformat));
86 if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
87 x264_encoder.reset(new X264Encoder(oformat, /*use_separate_disk_params=*/false));
89 VideoCodecInterface *http_encoder = x264_encoder.get();
90 VideoCodecInterface *disk_encoder = x264_encoder.get();
92 if (global_flags.av1_video_to_http) {
93 av1_encoder.reset(new AV1Encoder(oformat));
94 http_encoder = av1_encoder.get();
97 if (global_flags.x264_separate_disk_encode) {
98 x264_disk_encoder.reset(new X264Encoder(oformat, /*use_separate_disk_params=*/true));
99 disk_encoder = x264_disk_encoder.get();
102 string filename = generate_local_dump_filename(/*frame=*/0);
103 quicksync_encoder.reset(new QuickSyncEncoder(filename, resource_pool, surface, va_display, width, height, oformat, http_encoder, disk_encoder, disk_space_estimator));
105 open_output_streams();
106 stream_audio_encoder->add_mux(http_mux.get());
107 if (srt_mux != nullptr) {
108 stream_audio_encoder->add_mux(srt_mux.get());
110 quicksync_encoder->set_http_mux(http_mux.get());
111 if (srt_mux != nullptr) {
112 quicksync_encoder->set_srt_mux(srt_mux.get());
114 if (global_flags.x264_video_to_http) {
115 x264_encoder->add_mux(http_mux.get());
116 if (srt_mux != nullptr) {
117 x264_encoder->add_mux(srt_mux.get());
121 if (global_flags.av1_video_to_http) {
122 av1_encoder->add_mux(http_mux.get());
123 if (srt_mux != nullptr) {
124 av1_encoder->add_mux(srt_mux.get());
130 VideoEncoder::~VideoEncoder()
133 quicksync_encoder->shutdown();
134 x264_encoder.reset(nullptr);
135 x264_disk_encoder.reset(nullptr);
136 quicksync_encoder->close_file();
137 quicksync_encoder.reset(nullptr);
138 while (quicksync_encoders_in_shutdown.load() > 0) {
143 void VideoEncoder::do_cut(int frame)
145 string filename = generate_local_dump_filename(frame);
146 printf("Starting new recording: %s\n", filename.c_str());
148 // Do the shutdown of the old encoder in a separate thread, since it can
149 // take some time (it needs to wait for all the frames in the queue to be
150 // done encoding, for one) and we are running on the main mixer thread.
151 // However, since this means both encoders could be sending packets at
152 // the same time, it means pts could come out of order to the stream mux,
153 // and we need to plug it until the shutdown is complete.
155 lock(qs_mu, qs_audio_mu);
156 lock_guard<mutex> lock1(qs_mu, adopt_lock), lock2(qs_audio_mu, adopt_lock);
157 QuickSyncEncoder *old_encoder = quicksync_encoder.release(); // When we go C++14, we can use move capture instead.
158 X264Encoder *old_x264_encoder = nullptr;
159 X264Encoder *old_x264_disk_encoder = nullptr;
160 if (global_flags.x264_video_to_disk) {
161 old_x264_encoder = x264_encoder.release();
163 if (global_flags.x264_separate_disk_encode) {
164 old_x264_disk_encoder = x264_disk_encoder.release();
166 thread([old_encoder, old_x264_encoder, old_x264_disk_encoder, this]{
167 old_encoder->shutdown();
168 delete old_x264_encoder;
169 delete old_x264_disk_encoder;
170 old_encoder->close_file();
173 // We cannot delete the encoder here, as this thread has no OpenGL context.
174 // We'll deal with it in begin_frame().
175 lock_guard<mutex> lock(qs_mu);
176 qs_needing_cleanup.emplace_back(old_encoder);
179 if (global_flags.x264_video_to_disk) {
180 x264_encoder.reset(new X264Encoder(oformat, /*use_separate_disk_params=*/false));
181 assert(global_flags.x264_video_to_http);
182 if (global_flags.x264_video_to_http) {
183 x264_encoder->add_mux(http_mux.get());
185 if (overriding_bitrate != 0) {
186 x264_encoder->change_bitrate(overriding_bitrate);
189 X264Encoder *http_encoder = x264_encoder.get();
190 X264Encoder *disk_encoder = x264_encoder.get();
191 if (global_flags.x264_separate_disk_encode) {
192 x264_disk_encoder.reset(new X264Encoder(oformat, /*use_separate_disk_params=*/true));
193 disk_encoder = x264_disk_encoder.get();
196 quicksync_encoder.reset(new QuickSyncEncoder(filename, resource_pool, surface, va_display, width, height, oformat, http_encoder, disk_encoder, disk_space_estimator));
197 quicksync_encoder->set_http_mux(http_mux.get());
200 void VideoEncoder::change_x264_bitrate(unsigned rate_kbit)
202 overriding_bitrate = rate_kbit;
203 x264_encoder->change_bitrate(rate_kbit);
206 void VideoEncoder::add_audio(int64_t pts, std::vector<float> audio)
208 // Take only qs_audio_mu, since add_audio() is thread safe
209 // (we can only conflict with do_cut(), which takes qs_audio_mu)
210 // and we don't want to contend with begin_frame().
212 lock_guard<mutex> lock(qs_audio_mu);
213 quicksync_encoder->add_audio(pts, audio);
215 stream_audio_encoder->encode_audio(audio, pts + quicksync_encoder->global_delay());
218 bool VideoEncoder::is_zerocopy() const
220 // Explicitly do _not_ take qs_mu; this is called from the mixer,
221 // and qs_mu might be contended. is_zerocopy() is thread safe
222 // and never called in parallel with do_cut() (both happen only
223 // from the mixer thread).
224 return quicksync_encoder->is_zerocopy();
227 bool VideoEncoder::begin_frame(int64_t pts, int64_t duration, movit::YCbCrLumaCoefficients ycbcr_coefficients, const std::vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
229 lock_guard<mutex> lock(qs_mu);
230 qs_needing_cleanup.clear(); // Since we have an OpenGL context here, and are called regularly.
231 return quicksync_encoder->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);
234 RefCountedGLsync VideoEncoder::end_frame()
236 want_srt_metric_update = true;
237 lock_guard<mutex> lock(qs_mu);
238 return quicksync_encoder->end_frame();
241 void VideoEncoder::open_output_streams()
243 for (bool is_srt : {false, true}) {
244 if (is_srt && global_flags.srt_destination_host.empty()) {
248 AVFormatContext *avctx = avformat_alloc_context();
249 avctx->oformat = is_srt ? srt_oformat : oformat;
251 uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
252 avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
254 avctx->pb->write_packet = &VideoEncoder::write_srt_packet_thunk;
256 avctx->pb->write_data_type = &VideoEncoder::write_packet2_thunk;
257 avctx->pb->ignore_boundary_point = 1;
260 Mux::Codec video_codec;
261 if (global_flags.av1_video_to_http) {
262 video_codec = Mux::CODEC_AV1;
264 video_codec = Mux::CODEC_H264;
267 avctx->flags = AVFMT_FLAG_CUSTOM_IO;
269 string video_extradata;
270 if (global_flags.x264_video_to_http) {
271 video_extradata = x264_encoder->get_global_headers();
273 } else if (global_flags.av1_video_to_http) {
274 video_extradata = av1_encoder->get_global_headers();
278 Mux *mux = new Mux(avctx, width, height, video_codec, video_extradata, stream_audio_encoder->get_codec_parameters().get(),
279 get_color_space(global_flags.ycbcr_rec709_coefficients), COARSE_TIMEBASE,
280 /*write_callback=*/nullptr, is_srt ? Mux::WRITE_BACKGROUND : Mux::WRITE_FOREGROUND, { is_srt ? &srt_mux_metrics : &http_mux_metrics });
283 srt_mux_metrics.init({{ "destination", "srt" }});
284 srt_metrics.init({{ "cardtype", "output" }});
285 global_metrics.add("srt_num_connection_attempts", {{ "cardtype", "output" }}, &metric_srt_num_connection_attempts);
288 http_mux_metrics.init({{ "destination", "http" }});
293 int VideoEncoder::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
295 VideoEncoder *video_encoder = (VideoEncoder *)opaque;
296 return video_encoder->write_packet2(buf, buf_size, type, time);
299 int VideoEncoder::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
301 if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
302 seen_sync_markers = true;
303 } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
304 // We don't know if this is a keyframe or not (the muxer could
305 // avoid marking it), so we just have to make the best of it.
306 type = AVIO_DATA_MARKER_SYNC_POINT;
309 if (type == AVIO_DATA_MARKER_HEADER) {
310 http_mux_header.append((char *)buf, buf_size);
311 httpd->set_header(HTTPD::StreamID{ HTTPD::MAIN_STREAM, 0 }, http_mux_header);
313 httpd->add_data(HTTPD::StreamID{ HTTPD::MAIN_STREAM, 0 }, (char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
318 int VideoEncoder::write_srt_packet_thunk(void *opaque, uint8_t *buf, int buf_size)
320 VideoEncoder *video_encoder = (VideoEncoder *)opaque;
321 return video_encoder->write_srt_packet(buf, buf_size);
324 static string print_addrinfo(const addrinfo *ai)
326 char hoststr[NI_MAXHOST], portstr[NI_MAXSERV];
327 if (getnameinfo(ai->ai_addr, ai->ai_addrlen, hoststr, sizeof(hoststr), portstr, sizeof(portstr), NI_DGRAM | NI_NUMERICHOST | NI_NUMERICSERV) != 0) {
328 return "<unknown address>"; // Should basically never happen, since we're not doing DNS lookups.
331 if (ai->ai_family == AF_INET6) {
332 return string("[") + hoststr + "]:" + portstr;
334 return string(hoststr) + ":" + portstr;
338 int VideoEncoder::open_srt_socket()
340 int sock = srt_create_socket();
342 fprintf(stderr, "srt_create_socket(): %s\n", srt_getlasterror_str());
346 SRT_TRANSTYPE live = SRTT_LIVE;
347 if (srt_setsockopt(sock, 0, SRTO_TRANSTYPE, &live, sizeof(live)) < 0) {
348 fprintf(stderr, "srt_setsockopt(SRTO_TRANSTYPE): %s\n", srt_getlasterror_str());
353 if (srt_setsockopt(sock, 0, SRTO_LATENCY, &global_flags.srt_output_latency_ms, sizeof(global_flags.srt_output_latency_ms)) < 0) {
354 fprintf(stderr, "srt_setsockopt(SRTO_LATENCY): %s\n", srt_getlasterror_str());
359 if (!global_flags.srt_streamid.empty()) {
360 if (srt_setsockopt(sock, 0, SRTO_STREAMID, global_flags.srt_streamid.data(), global_flags.srt_streamid.size()) < 0) {
361 fprintf(stderr, "srt_setsockopt(SRTO_STREAMID): %s\n", srt_getlasterror_str());
367 if (!global_flags.srt_passphrase.empty()) {
368 if (srt_setsockopt(sock, 0, SRTO_PASSPHRASE, global_flags.srt_passphrase.data(), global_flags.srt_passphrase.size()) < 0) {
369 fprintf(stderr, "srt_setsockopt(SRTO_PASSPHRASE): %s\n", srt_getlasterror_str());
378 int VideoEncoder::connect_to_srt()
380 // We need to specify SOCK_DGRAM as a hint, or we'll get all addresses
381 // three times (for each of TCP, UDP, raw).
383 memset(&hints, 0, sizeof(hints));
384 hints.ai_flags = AI_ADDRCONFIG;
385 hints.ai_socktype = SOCK_DGRAM;
388 int ret = getaddrinfo(global_flags.srt_destination_host.c_str(), global_flags.srt_destination_port.c_str(), &hints, &ai);
390 fprintf(stderr, "getaddrinfo(%s:%s): %s\n", global_flags.srt_destination_host.c_str(), global_flags.srt_destination_port.c_str(), gai_strerror(ret));
394 unique_ptr<addrinfo, decltype(freeaddrinfo) *> ai_cleanup(ai, &freeaddrinfo);
396 for (const addrinfo *cur = ai; cur != nullptr; cur = cur->ai_next) {
397 // Seemingly, srt_create_socket() isn't universal; once we try to connect,
398 // it gets locked to either IPv4 or IPv6. So we need to create a new one
399 // for every address we try.
400 int sock = open_srt_socket();
405 ++metric_srt_num_connection_attempts;
407 // We do a non-blocking connect, so that we can check should_quit
408 // every now and then.
410 if (srt_setsockopt(sock, 0, SRTO_RCVSYN, &blocking, sizeof(blocking)) < 0) {
411 fprintf(stderr, "srt_setsockopt(SRTO_SNDSYN=0): %s\n", srt_getlasterror_str());
415 if (srt_connect(sock, cur->ai_addr, cur->ai_addrlen) < 0) {
416 fprintf(stderr, "srt_connect(%s): %s\n", print_addrinfo(cur).c_str(), srt_getlasterror_str());
420 int eid = srt_epoll_create();
422 fprintf(stderr, "srt_epoll_create(): %s\n", srt_getlasterror_str());
426 int modes = SRT_EPOLL_ERR | SRT_EPOLL_OUT;
427 if (srt_epoll_add_usock(eid, sock, &modes) < 0) {
428 fprintf(stderr, "srt_epoll_usock(): %s\n", srt_getlasterror_str());
430 srt_epoll_release(eid);
434 while (!should_quit.load()) {
435 SRTSOCKET errfds[1], writefds[1];
436 int num_errfds = 1, num_writefds = 1;
437 int poll_time_ms = 100;
438 int ret = srt_epoll_wait(eid, errfds, &num_errfds, writefds, &num_writefds, poll_time_ms, 0, 0, 0, 0);
440 if (srt_getlasterror(nullptr) == SRT_ETIMEOUT) {
443 fprintf(stderr, "srt_epoll_wait(): %s\n", srt_getlasterror_str());
445 srt_epoll_release(eid);
448 } else if (ret > 0) {
449 // The SRT epoll framework is pretty odd, but seemingly,
450 // this is the way. Getting the same error code as srt_connect()
451 // would normally return seems to be impossible, though.
452 ok = (num_errfds == 0);
454 fprintf(stderr, "num_errfds=%d num_writefds=%d last_err=%s\n", num_errfds, num_writefds, srt_getlasterror_str());
458 srt_epoll_release(eid);
459 if (should_quit.load()) {
464 fprintf(stderr, "Connected to destination SRT endpoint at %s.\n", print_addrinfo(cur).c_str());
467 fprintf(stderr, "srt_connect(%s): %s\n", print_addrinfo(cur).c_str(), srt_getlasterror_str());
472 // Out of candidates, so give up.
476 int VideoEncoder::write_srt_packet(uint8_t *buf, int buf_size)
478 if (want_srt_metric_update.exchange(false) && srt_sock != -1) {
479 srt_metrics.update_srt_stats(srt_sock);
482 bool has_drained = false;
483 bool trying_reconnect = false;
484 steady_clock::time_point first_connect_start;
486 while (buf_size > 0 && !should_quit.load()) {
487 if (srt_sock == -1) {
488 if (!trying_reconnect) {
489 first_connect_start = steady_clock::now();
490 trying_reconnect = true;
492 srt_sock = connect_to_srt();
493 if (srt_sock == -1) {
495 if (!has_drained && duration<double>(steady_clock::now() - first_connect_start).count() >= global_flags.srt_output_latency_ms * 1e-3) {
496 // The entire concept for SRT is to have fixed, low latency.
497 // If we've been out for more than a latency period, we shouldn't
498 // try to send the entire backlog. (But we should be tolerant
499 // of a quick disconnect and reconnect.) Maybe it would be better
500 // to have a sliding window of how much we remove, but it quickly
501 // starts getting esoteric, so juts drop it all.
502 fprintf(stderr, "WARNING: No SRT connection for more than %d ms, dropping data.\n",
503 global_flags.srt_output_latency_ms);
509 srt_metrics.update_srt_stats(srt_sock);
512 // Now that we're reconnected, we can start accepting data again,
513 // but discard the rest of this write (it is very old by now).
517 int to_send = min(buf_size, SRT_LIVE_DEF_PLSIZE);
518 int ret = srt_send(srt_sock, (char *)buf, to_send);
520 fprintf(stderr, "srt_send(): %s\n", srt_getlasterror_str());
522 srt_metrics.metric_srt_uptime_seconds = 0.0 / 0.0;
523 if (!trying_reconnect) {
524 first_connect_start = steady_clock::now();
525 trying_reconnect = true;
527 srt_sock = connect_to_srt();