git.sesse.net Git - nageru/blob - nageru/x264_encoder.cpp

   1 #include "x264_encoder.h"
   2
   3 #include <assert.h>
   4 #include <dlfcn.h>
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <string.h>
   8 #include <unistd.h>
   9 #include <x264.h>
  10 #include <atomic>
  11 #include <cstdint>
  12 #include <functional>
  13 #include <mutex>
  14
  15 #include "defs.h"
  16 #include "flags.h"
  17 #include "shared/metrics.h"
  18 #include "shared/mux.h"
  19 #include "print_latency.h"
  20 #include "shared/timebase.h"
  21 #include "x264_dynamic.h"
  22 #include "x264_speed_control.h"
  23
  24 extern "C" {
  25 #include <libavcodec/avcodec.h>
  26 #include <libavformat/avformat.h>
  27 }
  28
  29 using namespace movit;
  30 using namespace std;
  31 using namespace std::chrono;
  32 using namespace std::placeholders;
  33
  34 namespace {
  35
  36 // X264Encoder can be restarted if --record-x264-video is set, so make these
  37 // metrics global.
  38 atomic<int64_t> metric_x264_queued_frames{0};
  39 atomic<int64_t> metric_x264_max_queued_frames{X264_QUEUE_LENGTH};
  40 atomic<int64_t> metric_x264_dropped_frames{0};
  41 atomic<int64_t> metric_x264_output_frames_i{0};
  42 atomic<int64_t> metric_x264_output_frames_p{0};
  43 atomic<int64_t> metric_x264_output_frames_b{0};
  44 Histogram metric_x264_crf;
  45 LatencyHistogram x264_latency_histogram;
  46
  47 atomic<int64_t> metric_x264_disk_queued_frames{0};
  48 atomic<int64_t> metric_x264_disk_max_queued_frames{X264_QUEUE_LENGTH};
  49 atomic<int64_t> metric_x264_disk_dropped_frames{0};
  50 atomic<int64_t> metric_x264_disk_output_frames_i{0};
  51 atomic<int64_t> metric_x264_disk_output_frames_p{0};
  52 atomic<int64_t> metric_x264_disk_output_frames_b{0};
  53 Histogram metric_x264_disk_crf;
  54 LatencyHistogram x264_disk_latency_histogram;
  55
  56 once_flag x264_metrics_inited, x264_disk_metrics_inited;
  57
  58 void update_vbv_settings(x264_param_t *param)
  59 {
  60         if (global_flags.x264_bitrate == -1) {
  61                 return;
  62         }
  63         param->rc.i_vbv_buffer_size = param->rc.i_bitrate;  // One-second VBV.
  64         param->rc.i_vbv_max_bitrate = param->rc.i_bitrate;  // CBR.
  65 }
  66
  67 }  // namespace
  68
  69 X264Encoder::X264Encoder(const AVOutputFormat *oformat, bool use_separate_disk_params)
  70         : wants_global_headers(oformat->flags & AVFMT_GLOBALHEADER),
  71           use_separate_disk_params(use_separate_disk_params),
  72           dyn(load_x264_for_bit_depth(global_flags.bit_depth))
  73 {
  74         if (use_separate_disk_params) {
  75                 call_once(x264_disk_metrics_inited, []{
  76                         global_metrics.add("x264_queued_frames",  {{ "encode", "separate_disk" }}, &metric_x264_disk_queued_frames, Metrics::TYPE_GAUGE);
  77                         global_metrics.add("x264_max_queued_frames", {{ "encode", "separate_disk" }},  &metric_x264_disk_max_queued_frames, Metrics::TYPE_GAUGE);
  78                         global_metrics.add("x264_dropped_frames", {{ "encode", "separate_disk" }},  &metric_x264_disk_dropped_frames);
  79                         global_metrics.add("x264_output_frames", {{ "encode", "separate_disk" }, { "type", "i" }}, &metric_x264_disk_output_frames_i);
  80                         global_metrics.add("x264_output_frames", {{ "encode", "separate_disk" }, { "type", "p" }}, &metric_x264_disk_output_frames_p);
  81                         global_metrics.add("x264_output_frames", {{ "encode", "separate_disk" }, { "type", "b" }}, &metric_x264_disk_output_frames_b);
  82
  83                         metric_x264_disk_crf.init_uniform(50);
  84                         global_metrics.add("x264_crf", {{ "encode", "separate_disk" }}, &metric_x264_disk_crf);
  85                         x264_disk_latency_histogram.init("x264_disk");
  86                 });
  87         } else {
  88                 call_once(use_separate_disk_params ? x264_disk_metrics_inited : x264_metrics_inited, []{
  89                         global_metrics.add("x264_queued_frames",  {{ "encode", "regular" }}, &metric_x264_queued_frames, Metrics::TYPE_GAUGE);
  90                         global_metrics.add("x264_max_queued_frames", {{ "encode", "regular" }},  &metric_x264_max_queued_frames, Metrics::TYPE_GAUGE);
  91                         global_metrics.add("x264_dropped_frames", {{ "encode", "regular" }},  &metric_x264_dropped_frames);
  92                         global_metrics.add("x264_output_frames", {{ "encode", "regular" }, { "type", "i" }}, &metric_x264_output_frames_i);
  93                         global_metrics.add("x264_output_frames", {{ "encode", "regular" }, { "type", "p" }}, &metric_x264_output_frames_p);
  94                         global_metrics.add("x264_output_frames", {{ "encode", "regular" }, { "type", "b" }}, &metric_x264_output_frames_b);
  95
  96                         metric_x264_crf.init_uniform(50);
  97                         global_metrics.add("x264_crf", {{ "encode", "regular" }}, &metric_x264_crf);
  98                         x264_latency_histogram.init("x264");
  99                 });
 100         }
 101
 102         size_t bytes_per_pixel = global_flags.bit_depth > 8 ? 2 : 1;
 103         frame_pool.reset(new uint8_t[global_flags.width * global_flags.height * 2 * bytes_per_pixel * X264_QUEUE_LENGTH]);
 104         for (unsigned i = 0; i < X264_QUEUE_LENGTH; ++i) {
 105                 free_frames.push(frame_pool.get() + i * (global_flags.width * global_flags.height * 2 * bytes_per_pixel));
 106         }
 107         encoder_thread = thread(&X264Encoder::encoder_thread_func, this);
 108 }
 109
 110 X264Encoder::~X264Encoder()
 111 {
 112         should_quit = true;
 113         queued_frames_nonempty.notify_all();
 114         encoder_thread.join();
 115         if (dyn.handle) {
 116                 dlclose(dyn.handle);
 117         }
 118 }
 119
 120 void X264Encoder::add_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const uint8_t *data, const ReceivedTimestamps &received_ts)
 121 {
 122         assert(!should_quit);
 123
 124         QueuedFrame qf;
 125         qf.pts = pts;
 126         qf.duration = duration;
 127         qf.ycbcr_coefficients = ycbcr_coefficients;
 128         qf.received_ts = received_ts;
 129
 130         {
 131                 lock_guard<mutex> lock(mu);
 132                 if (free_frames.empty()) {
 133                         if (use_separate_disk_params) {
 134                                 fprintf(stderr, "WARNING: x264 queue full (disk encoder), dropping frame with pts %" PRId64 "\n", pts);
 135                                 ++metric_x264_disk_dropped_frames;
 136                         } else {
 137                                 fprintf(stderr, "WARNING: x264 queue full, dropping frame with pts %" PRId64 "\n", pts);
 138                                 ++metric_x264_dropped_frames;
 139                         }
 140                         return;
 141                 }
 142
 143                 qf.data = free_frames.front();
 144                 free_frames.pop();
 145         }
 146
 147         size_t bytes_per_pixel = global_flags.bit_depth > 8 ? 2 : 1;
 148         memcpy(qf.data, data, global_flags.width * global_flags.height * 2 * bytes_per_pixel);
 149
 150         {
 151                 lock_guard<mutex> lock(mu);
 152                 queued_frames.push(qf);
 153                 queued_frames_nonempty.notify_all();
 154                 if (use_separate_disk_params) {
 155                         metric_x264_disk_queued_frames = queued_frames.size();
 156                 } else {
 157                         metric_x264_queued_frames = queued_frames.size();
 158                 }
 159         }
 160 }
 161
 162 void X264Encoder::init_x264()
 163 {
 164         x264_param_t param;
 165         if (use_separate_disk_params) {
 166                 dyn.x264_param_default_preset(&param, global_flags.x264_separate_disk_preset.c_str(), global_flags.x264_separate_disk_tune.c_str());
 167         } else {
 168                 dyn.x264_param_default_preset(&param, global_flags.x264_preset.c_str(), global_flags.x264_tune.c_str());
 169         }
 170
 171         param.i_width = global_flags.width;
 172         param.i_height = global_flags.height;
 173         param.i_csp = X264_CSP_NV12;
 174         if (global_flags.bit_depth > 8) {
 175                 param.i_csp |= X264_CSP_HIGH_DEPTH;
 176         }
 177         param.b_vfr_input = 1;
 178         param.i_timebase_num = 1;
 179         param.i_timebase_den = TIMEBASE;
 180         param.i_keyint_max = 50; // About one second.
 181         if (!use_separate_disk_params && global_flags.x264_speedcontrol) {
 182                 param.i_frame_reference = 16;  // Because speedcontrol is never allowed to change this above what we set at start.
 183         }
 184 #if X264_BUILD >= 153
 185         param.i_bitdepth = global_flags.bit_depth;
 186 #endif
 187
 188         // NOTE: These should be in sync with the ones in quicksync_encoder.cpp (sps_rbsp()).
 189         param.vui.i_vidformat = 5;  // Unspecified.
 190         param.vui.b_fullrange = 0;
 191         param.vui.i_colorprim = 1;  // BT.709.
 192         param.vui.i_transfer = 13;  // sRGB.
 193         if (global_flags.ycbcr_rec709_coefficients) {
 194                 param.vui.i_colmatrix = 1;  // BT.709.
 195         } else {
 196                 param.vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 197         }
 198
 199         const double crf = use_separate_disk_params ? global_flags.x264_separate_disk_crf : global_flags.x264_crf;
 200         const int bitrate = use_separate_disk_params ? global_flags.x264_separate_disk_bitrate : global_flags.x264_bitrate;
 201         if (!isinf(crf)) {
 202                 param.rc.i_rc_method = X264_RC_CRF;
 203                 param.rc.f_rf_constant = crf;
 204         } else {
 205                 param.rc.i_rc_method = X264_RC_ABR;
 206                 param.rc.i_bitrate = bitrate;
 207
 208                 // If the user wants to cap the max rate, it is also reasonable
 209                 // to assume that they are fine with the stream constantly
 210                 // being around that rate even for very low-complexity content;
 211                 // the obvious and extreme example being a static black picture.
 212                 //
 213                 // One would think it's fine to have low-complexity content use
 214                 // less bitrate, but it seems to cause problems in practice;
 215                 // e.g. VLC seems to often drop the stream (similar to a buffer
 216                 // underrun) in such cases, but only when streaming from Nageru,
 217                 // not when reading a dump of the same stream from disk.
 218                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 219                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 220                 // but it's a typical case of problems that can arise. Similarly,
 221                 // TCP's congestion control is not always fond of the rate staying
 222                 // low for a while and then rising quickly -- a variation on the same
 223                 // problem.
 224                 //
 225                 // We solve this by simply asking x264 to fill in dummy bits
 226                 // in these cases, so that the bitrate stays reasonable constant.
 227                 // It's a waste of bandwidth, but it makes things go much more
 228                 // smoothly in these cases. (We don't do it if VBV control is off
 229                 // in general, not the least because it makes no sense and x264
 230                 // thus ignores the parameter.)
 231                 param.rc.b_filler = 1;
 232         }
 233         if (!use_separate_disk_params) {
 234                 update_vbv_settings(&param);
 235         }
 236
 237         // Occasionally players have problem with extremely low quantizers;
 238         // be on the safe side. Shouldn't affect quality in any meaningful way.
 239         param.rc.i_qp_min = 5;
 240
 241         const vector<string> &extra_param = use_separate_disk_params ? global_flags.x264_separate_disk_extra_param : global_flags.x264_extra_param;
 242         for (const string &str : extra_param) {
 243                 const size_t pos = str.find(',');
 244                 if (pos == string::npos) {
 245                         if (dyn.x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 246                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 247                         }
 248                 } else {
 249                         const string key = str.substr(0, pos);
 250                         const string value = str.substr(pos + 1);
 251                         if (dyn.x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 252                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 253                                         key.c_str(), value.c_str());
 254                         }
 255                 }
 256         }
 257
 258         if (global_flags.bit_depth > 8) {
 259                 dyn.x264_param_apply_profile(&param, "high10");
 260         } else {
 261                 dyn.x264_param_apply_profile(&param, "high");
 262         }
 263
 264         param.b_repeat_headers = !wants_global_headers;
 265
 266         x264 = dyn.x264_encoder_open(&param);
 267         if (x264 == nullptr) {
 268                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 269                 abort();
 270         }
 271
 272         if (!use_separate_disk_params && global_flags.x264_speedcontrol) {
 273                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 274         }
 275
 276         if (wants_global_headers) {
 277                 x264_nal_t *nal;
 278                 int num_nal;
 279
 280                 dyn.x264_encoder_headers(x264, &nal, &num_nal);
 281
 282                 for (int i = 0; i < num_nal; ++i) {
 283                         if (nal[i].i_type == NAL_SEI) {
 284                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 285                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 286                         } else {
 287                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 288                         }
 289                 }
 290         }
 291 }
 292
 293 void X264Encoder::encoder_thread_func()
 294 {
 295         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 296                 perror("nice()");
 297                 // No exit; it's not fatal.
 298         }
 299         pthread_setname_np(pthread_self(), "x264_encode");
 300         init_x264();
 301         x264_init_done = true;
 302
 303         bool frames_left;
 304
 305         do {
 306                 QueuedFrame qf;
 307
 308                 // Wait for a queued frame, then dequeue it.
 309                 {
 310                         unique_lock<mutex> lock(mu);
 311                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 312                         if (!queued_frames.empty()) {
 313                                 qf = queued_frames.front();
 314                                 queued_frames.pop();
 315                         } else {
 316                                 qf.pts = -1;
 317                                 qf.duration = -1;
 318                                 qf.data = nullptr;
 319                         }
 320
 321                         if (use_separate_disk_params) {
 322                                 metric_x264_disk_queued_frames = queued_frames.size();
 323                         } else {
 324                                 metric_x264_queued_frames = queued_frames.size();
 325                         }
 326                         frames_left = !queued_frames.empty();
 327                 }
 328
 329                 encode_frame(qf);
 330
 331                 {
 332                         lock_guard<mutex> lock(mu);
 333                         free_frames.push(qf.data);
 334                 }
 335
 336                 // We should quit only if the should_quit flag is set _and_ we have nothing
 337                 // in either queue.
 338         } while (!should_quit || frames_left || dyn.x264_encoder_delayed_frames(x264) > 0);
 339
 340         dyn.x264_encoder_close(x264);
 341 }
 342
 343 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 344 {
 345         x264_nal_t *nal = nullptr;
 346         int num_nal = 0;
 347         x264_picture_t pic;
 348         x264_picture_t *input_pic = nullptr;
 349
 350         if (qf.data) {
 351                 dyn.x264_picture_init(&pic);
 352
 353                 pic.i_pts = qf.pts;
 354                 if (global_flags.bit_depth > 8) {
 355                         pic.img.i_csp = X264_CSP_NV12 | X264_CSP_HIGH_DEPTH;
 356                         pic.img.i_plane = 2;
 357                         pic.img.plane[0] = qf.data;
 358                         pic.img.i_stride[0] = global_flags.width * sizeof(uint16_t);
 359                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height * sizeof(uint16_t);
 360                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint32_t);
 361                 } else {
 362                         pic.img.i_csp = X264_CSP_NV12;
 363                         pic.img.i_plane = 2;
 364                         pic.img.plane[0] = qf.data;
 365                         pic.img.i_stride[0] = global_flags.width;
 366                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height;
 367                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint16_t);
 368                 }
 369                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 370
 371                 input_pic = &pic;
 372
 373                 frames_being_encoded[qf.pts] = qf.received_ts;
 374         }
 375
 376         unsigned new_rate = new_bitrate_kbit.load();  // Can be 0 for no change.
 377         if (speed_control) {
 378                 speed_control->set_config_override_function(bind(&speed_control_override_func, new_rate, qf.ycbcr_coefficients, _1));
 379         } else {
 380                 x264_param_t param;
 381                 dyn.x264_encoder_parameters(x264, &param);
 382                 speed_control_override_func(new_rate, qf.ycbcr_coefficients, &param);
 383                 dyn.x264_encoder_reconfig(x264, &param);
 384         }
 385
 386         if (speed_control) {
 387                 float queue_fill_ratio;
 388                 {
 389                         lock_guard<mutex> lock(mu);
 390                         queue_fill_ratio = float(free_frames.size()) / X264_QUEUE_LENGTH;
 391                 }
 392                 speed_control->before_frame(queue_fill_ratio, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE);
 393         }
 394         dyn.x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic);
 395         if (speed_control) {
 396                 speed_control->after_frame();
 397         }
 398
 399         if (num_nal == 0) return;
 400
 401         if (use_separate_disk_params) {
 402                 if (IS_X264_TYPE_I(pic.i_type)) {
 403                         ++metric_x264_disk_output_frames_i;
 404                 } else if (IS_X264_TYPE_B(pic.i_type)) {
 405                         ++metric_x264_disk_output_frames_b;
 406                 } else {
 407                         ++metric_x264_disk_output_frames_p;
 408                 }
 409
 410                 metric_x264_disk_crf.count_event(pic.prop.f_crf_avg);
 411         } else {
 412                 if (IS_X264_TYPE_I(pic.i_type)) {
 413                         ++metric_x264_output_frames_i;
 414                 } else if (IS_X264_TYPE_B(pic.i_type)) {
 415                         ++metric_x264_output_frames_b;
 416                 } else {
 417                         ++metric_x264_output_frames_p;
 418                 }
 419
 420                 metric_x264_crf.count_event(pic.prop.f_crf_avg);
 421         }
 422
 423         if (frames_being_encoded.count(pic.i_pts)) {
 424                 ReceivedTimestamps received_ts = frames_being_encoded[pic.i_pts];
 425                 frames_being_encoded.erase(pic.i_pts);
 426
 427                 static int frameno = 0;
 428                 print_latency("Current x264 latency (video inputs → network mux):",
 429                         received_ts, (pic.i_type == X264_TYPE_B || pic.i_type == X264_TYPE_BREF),
 430                         &frameno, &x264_latency_histogram);
 431         } else {
 432                 assert(false);
 433         }
 434
 435         // We really need one AVPacket for the entire frame, it seems,
 436         // so combine it all.
 437         size_t num_bytes = buffered_sei.size();
 438         for (int i = 0; i < num_nal; ++i) {
 439                 num_bytes += nal[i].i_payload;
 440         }
 441
 442         unique_ptr<uint8_t[]> data(new uint8_t[num_bytes]);
 443         uint8_t *ptr = data.get();
 444
 445         if (!buffered_sei.empty()) {
 446                 memcpy(ptr, buffered_sei.data(), buffered_sei.size());
 447                 ptr += buffered_sei.size();
 448                 buffered_sei.clear();
 449         }
 450         for (int i = 0; i < num_nal; ++i) {
 451                 memcpy(ptr, nal[i].p_payload, nal[i].i_payload);
 452                 ptr += nal[i].i_payload;
 453         }
 454
 455         AVPacket pkt;
 456         memset(&pkt, 0, sizeof(pkt));
 457         pkt.buf = nullptr;
 458         pkt.data = data.get();
 459         pkt.size = num_bytes;
 460         pkt.stream_index = 0;
 461         if (pic.b_keyframe) {
 462                 pkt.flags = AV_PKT_FLAG_KEY;
 463         } else {
 464                 pkt.flags = 0;
 465         }
 466         pkt.duration = reinterpret_cast<intptr_t>(pic.opaque);
 467
 468         for (Mux *mux : muxes) {
 469                 mux->add_packet(pkt, pic.i_pts, pic.i_dts);
 470         }
 471 }
 472
 473 void X264Encoder::speed_control_override_func(unsigned bitrate_kbit, movit::YCbCrLumaCoefficients ycbcr_coefficients, x264_param_t *param)
 474 {
 475         if (bitrate_kbit != 0) {
 476                 param->rc.i_bitrate = bitrate_kbit;
 477                 update_vbv_settings(param);
 478         }
 479
 480         if (ycbcr_coefficients == YCBCR_REC_709) {
 481                 param->vui.i_colmatrix = 1;  // BT.709.
 482         } else {
 483                 assert(ycbcr_coefficients == YCBCR_REC_601);
 484                 param->vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 485         }
 486 }