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         if (global_flags.x264_vbv_buffer_size < 0) {
  64                 param->rc.i_vbv_buffer_size = param->rc.i_bitrate;  // One-second VBV.
  65         } else {
  66                 param->rc.i_vbv_buffer_size = global_flags.x264_vbv_buffer_size;
  67         }
  68         if (global_flags.x264_vbv_max_bitrate < 0) {
  69                 param->rc.i_vbv_max_bitrate = param->rc.i_bitrate;  // CBR.
  70         } else {
  71                 param->rc.i_vbv_max_bitrate = global_flags.x264_vbv_max_bitrate;
  72         }
  73 }
  74
  75 }  // namespace
  76
  77 X264Encoder::X264Encoder(const AVOutputFormat *oformat, bool use_separate_disk_params)
  78         : wants_global_headers(oformat->flags & AVFMT_GLOBALHEADER),
  79           use_separate_disk_params(use_separate_disk_params),
  80           dyn(load_x264_for_bit_depth(global_flags.x264_bit_depth))
  81 {
  82         if (use_separate_disk_params) {
  83                 call_once(x264_disk_metrics_inited, []{
  84                         global_metrics.add("x264_queued_frames",  {{ "encode", "separate_disk" }}, &metric_x264_disk_queued_frames, Metrics::TYPE_GAUGE);
  85                         global_metrics.add("x264_max_queued_frames", {{ "encode", "separate_disk" }},  &metric_x264_disk_max_queued_frames, Metrics::TYPE_GAUGE);
  86                         global_metrics.add("x264_dropped_frames", {{ "encode", "separate_disk" }},  &metric_x264_disk_dropped_frames);
  87                         global_metrics.add("x264_output_frames", {{ "encode", "separate_disk" }, { "type", "i" }}, &metric_x264_disk_output_frames_i);
  88                         global_metrics.add("x264_output_frames", {{ "encode", "separate_disk" }, { "type", "p" }}, &metric_x264_disk_output_frames_p);
  89                         global_metrics.add("x264_output_frames", {{ "encode", "separate_disk" }, { "type", "b" }}, &metric_x264_disk_output_frames_b);
  90
  91                         metric_x264_disk_crf.init_uniform(50);
  92                         global_metrics.add("x264_crf", {{ "encode", "separate_disk" }}, &metric_x264_disk_crf);
  93                         x264_disk_latency_histogram.init("x264_disk");
  94                 });
  95         } else {
  96                 call_once(use_separate_disk_params ? x264_disk_metrics_inited : x264_metrics_inited, []{
  97                         global_metrics.add("x264_queued_frames",  {{ "encode", "regular" }}, &metric_x264_queued_frames, Metrics::TYPE_GAUGE);
  98                         global_metrics.add("x264_max_queued_frames", {{ "encode", "regular" }},  &metric_x264_max_queued_frames, Metrics::TYPE_GAUGE);
  99                         global_metrics.add("x264_dropped_frames", {{ "encode", "regular" }},  &metric_x264_dropped_frames);
 100                         global_metrics.add("x264_output_frames", {{ "encode", "regular" }, { "type", "i" }}, &metric_x264_output_frames_i);
 101                         global_metrics.add("x264_output_frames", {{ "encode", "regular" }, { "type", "p" }}, &metric_x264_output_frames_p);
 102                         global_metrics.add("x264_output_frames", {{ "encode", "regular" }, { "type", "b" }}, &metric_x264_output_frames_b);
 103
 104                         metric_x264_crf.init_uniform(50);
 105                         global_metrics.add("x264_crf", {{ "encode", "regular" }}, &metric_x264_crf);
 106                         x264_latency_histogram.init("x264");
 107                 });
 108         }
 109
 110         size_t bytes_per_pixel = global_flags.x264_bit_depth > 8 ? 2 : 1;
 111         frame_pool.reset(new uint8_t[global_flags.width * global_flags.height * 2 * bytes_per_pixel * X264_QUEUE_LENGTH]);
 112         for (unsigned i = 0; i < X264_QUEUE_LENGTH; ++i) {
 113                 free_frames.push(frame_pool.get() + i * (global_flags.width * global_flags.height * 2 * bytes_per_pixel));
 114         }
 115         encoder_thread = thread(&X264Encoder::encoder_thread_func, this);
 116 }
 117
 118 X264Encoder::~X264Encoder()
 119 {
 120         should_quit = true;
 121         queued_frames_nonempty.notify_all();
 122         encoder_thread.join();
 123         if (dyn.handle) {
 124                 dlclose(dyn.handle);
 125         }
 126 }
 127
 128 void X264Encoder::add_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const uint8_t *data, const ReceivedTimestamps &received_ts)
 129 {
 130         assert(!should_quit);
 131
 132         QueuedFrame qf;
 133         qf.pts = pts;
 134         qf.duration = duration;
 135         qf.ycbcr_coefficients = ycbcr_coefficients;
 136         qf.received_ts = received_ts;
 137
 138         {
 139                 lock_guard<mutex> lock(mu);
 140                 if (free_frames.empty()) {
 141                         if (use_separate_disk_params) {
 142                                 fprintf(stderr, "WARNING: x264 queue full (disk encoder), dropping frame with pts %" PRId64 "\n", pts);
 143                                 ++metric_x264_disk_dropped_frames;
 144                         } else {
 145                                 fprintf(stderr, "WARNING: x264 queue full, dropping frame with pts %" PRId64 "\n", pts);
 146                                 ++metric_x264_dropped_frames;
 147                         }
 148                         return;
 149                 }
 150
 151                 qf.data = free_frames.front();
 152                 free_frames.pop();
 153         }
 154
 155         size_t bytes_per_pixel = global_flags.x264_bit_depth > 8 ? 2 : 1;
 156         memcpy(qf.data, data, global_flags.width * global_flags.height * 2 * bytes_per_pixel);
 157
 158         {
 159                 lock_guard<mutex> lock(mu);
 160                 queued_frames.push(qf);
 161                 queued_frames_nonempty.notify_all();
 162                 if (use_separate_disk_params) {
 163                         metric_x264_disk_queued_frames = queued_frames.size();
 164                 } else {
 165                         metric_x264_queued_frames = queued_frames.size();
 166                 }
 167         }
 168 }
 169
 170 void X264Encoder::init_x264()
 171 {
 172         x264_param_t param;
 173         if (use_separate_disk_params) {
 174                 dyn.x264_param_default_preset(&param, global_flags.x264_separate_disk_preset.c_str(), global_flags.x264_separate_disk_tune.c_str());
 175         } else {
 176                 dyn.x264_param_default_preset(&param, global_flags.x264_preset.c_str(), global_flags.x264_tune.c_str());
 177         }
 178
 179         param.i_width = global_flags.width;
 180         param.i_height = global_flags.height;
 181         param.i_csp = X264_CSP_NV12;
 182         if (global_flags.x264_bit_depth > 8) {
 183                 param.i_csp |= X264_CSP_HIGH_DEPTH;
 184         }
 185         param.b_vfr_input = 1;
 186         param.i_timebase_num = 1;
 187         param.i_timebase_den = TIMEBASE;
 188         param.i_keyint_max = 50; // About one second.
 189         if (!use_separate_disk_params && global_flags.x264_speedcontrol) {
 190                 param.i_frame_reference = 16;  // Because speedcontrol is never allowed to change this above what we set at start.
 191         }
 192 #if X264_BUILD >= 153
 193         param.i_bitdepth = global_flags.x264_bit_depth;
 194 #endif
 195
 196         // NOTE: These should be in sync with the ones in quicksync_encoder.cpp (sps_rbsp()).
 197         param.vui.i_vidformat = 5;  // Unspecified.
 198         param.vui.b_fullrange = 0;
 199         param.vui.i_colorprim = 1;  // BT.709.
 200         param.vui.i_transfer = 13;  // sRGB.
 201         if (global_flags.ycbcr_rec709_coefficients) {
 202                 param.vui.i_colmatrix = 1;  // BT.709.
 203         } else {
 204                 param.vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 205         }
 206
 207         const double crf = use_separate_disk_params ? global_flags.x264_separate_disk_crf : global_flags.x264_crf;
 208         const int bitrate = use_separate_disk_params ? global_flags.x264_separate_disk_bitrate : global_flags.x264_bitrate;
 209         if (!isinf(crf)) {
 210                 param.rc.i_rc_method = X264_RC_CRF;
 211                 param.rc.f_rf_constant = crf;
 212         } else {
 213                 param.rc.i_rc_method = X264_RC_ABR;
 214                 param.rc.i_bitrate = bitrate;
 215         }
 216         if (!use_separate_disk_params) {
 217                 update_vbv_settings(&param);
 218         }
 219         if (param.rc.i_vbv_max_bitrate > 0) {
 220                 // If the user wants VBV control to cap the max rate, it is
 221                 // also reasonable to assume that they are fine with the stream
 222                 // constantly being around that rate even for very low-complexity
 223                 // content; the obvious and extreme example being a static
 224                 // black picture.
 225                 //
 226                 // One would think it's fine to have low-complexity content use
 227                 // less bitrate, but it seems to cause problems in practice;
 228                 // e.g. VLC seems to often drop the stream (similar to a buffer
 229                 // underrun) in such cases, but only when streaming from Nageru,
 230                 // not when reading a dump of the same stream from disk.
 231                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 232                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 233                 // but it's a typical case of problems that can arise. Similarly,
 234                 // TCP's congestion control is not always fond of the rate staying
 235                 // low for a while and then rising quickly -- a variation on the same
 236                 // problem.
 237                 //
 238                 // We solve this by simply asking x264 to fill in dummy bits
 239                 // in these cases, so that the bitrate stays reasonable constant.
 240                 // It's a waste of bandwidth, but it makes things go much more
 241                 // smoothly in these cases. (We don't do it if VBV control is off
 242                 // in general, not the least because it makes no sense and x264
 243                 // thus ignores the parameter.)
 244                 param.rc.b_filler = 1;
 245         }
 246
 247         // Occasionally players have problem with extremely low quantizers;
 248         // be on the safe side. Shouldn't affect quality in any meaningful way.
 249         param.rc.i_qp_min = 5;
 250
 251         const vector<string> &extra_param = use_separate_disk_params ? global_flags.x264_separate_disk_extra_param : global_flags.x264_extra_param;
 252         for (const string &str : extra_param) {
 253                 const size_t pos = str.find(',');
 254                 if (pos == string::npos) {
 255                         if (dyn.x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 256                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 257                         }
 258                 } else {
 259                         const string key = str.substr(0, pos);
 260                         const string value = str.substr(pos + 1);
 261                         if (dyn.x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 262                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 263                                         key.c_str(), value.c_str());
 264                         }
 265                 }
 266         }
 267
 268         if (global_flags.x264_bit_depth > 8) {
 269                 dyn.x264_param_apply_profile(&param, "high10");
 270         } else {
 271                 dyn.x264_param_apply_profile(&param, "high");
 272         }
 273
 274         param.b_repeat_headers = !wants_global_headers;
 275
 276         x264 = dyn.x264_encoder_open(&param);
 277         if (x264 == nullptr) {
 278                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 279                 abort();
 280         }
 281
 282         if (!use_separate_disk_params && global_flags.x264_speedcontrol) {
 283                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 284         }
 285
 286         if (wants_global_headers) {
 287                 x264_nal_t *nal;
 288                 int num_nal;
 289
 290                 dyn.x264_encoder_headers(x264, &nal, &num_nal);
 291
 292                 for (int i = 0; i < num_nal; ++i) {
 293                         if (nal[i].i_type == NAL_SEI) {
 294                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 295                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 296                         } else {
 297                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 298                         }
 299                 }
 300         }
 301 }
 302
 303 void X264Encoder::encoder_thread_func()
 304 {
 305         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 306                 perror("nice()");
 307                 // No exit; it's not fatal.
 308         }
 309         pthread_setname_np(pthread_self(), "x264_encode");
 310         init_x264();
 311         x264_init_done = true;
 312
 313         bool frames_left;
 314
 315         do {
 316                 QueuedFrame qf;
 317
 318                 // Wait for a queued frame, then dequeue it.
 319                 {
 320                         unique_lock<mutex> lock(mu);
 321                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 322                         if (!queued_frames.empty()) {
 323                                 qf = queued_frames.front();
 324                                 queued_frames.pop();
 325                         } else {
 326                                 qf.pts = -1;
 327                                 qf.duration = -1;
 328                                 qf.data = nullptr;
 329                         }
 330
 331                         if (use_separate_disk_params) {
 332                                 metric_x264_disk_queued_frames = queued_frames.size();
 333                         } else {
 334                                 metric_x264_queued_frames = queued_frames.size();
 335                         }
 336                         frames_left = !queued_frames.empty();
 337                 }
 338
 339                 encode_frame(qf);
 340
 341                 {
 342                         lock_guard<mutex> lock(mu);
 343                         free_frames.push(qf.data);
 344                 }
 345
 346                 // We should quit only if the should_quit flag is set _and_ we have nothing
 347                 // in either queue.
 348         } while (!should_quit || frames_left || dyn.x264_encoder_delayed_frames(x264) > 0);
 349
 350         dyn.x264_encoder_close(x264);
 351 }
 352
 353 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 354 {
 355         x264_nal_t *nal = nullptr;
 356         int num_nal = 0;
 357         x264_picture_t pic;
 358         x264_picture_t *input_pic = nullptr;
 359
 360         if (qf.data) {
 361                 dyn.x264_picture_init(&pic);
 362
 363                 pic.i_pts = qf.pts;
 364                 if (global_flags.x264_bit_depth > 8) {
 365                         pic.img.i_csp = X264_CSP_NV12 | X264_CSP_HIGH_DEPTH;
 366                         pic.img.i_plane = 2;
 367                         pic.img.plane[0] = qf.data;
 368                         pic.img.i_stride[0] = global_flags.width * sizeof(uint16_t);
 369                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height * sizeof(uint16_t);
 370                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint32_t);
 371                 } else {
 372                         pic.img.i_csp = X264_CSP_NV12;
 373                         pic.img.i_plane = 2;
 374                         pic.img.plane[0] = qf.data;
 375                         pic.img.i_stride[0] = global_flags.width;
 376                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height;
 377                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint16_t);
 378                 }
 379                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 380
 381                 input_pic = &pic;
 382
 383                 frames_being_encoded[qf.pts] = qf.received_ts;
 384         }
 385
 386         unsigned new_rate = new_bitrate_kbit.load();  // Can be 0 for no change.
 387         if (speed_control) {
 388                 speed_control->set_config_override_function(bind(&speed_control_override_func, new_rate, qf.ycbcr_coefficients, _1));
 389         } else {
 390                 x264_param_t param;
 391                 dyn.x264_encoder_parameters(x264, &param);
 392                 speed_control_override_func(new_rate, qf.ycbcr_coefficients, &param);
 393                 dyn.x264_encoder_reconfig(x264, &param);
 394         }
 395
 396         if (speed_control) {
 397                 float queue_fill_ratio;
 398                 {
 399                         lock_guard<mutex> lock(mu);
 400                         queue_fill_ratio = float(free_frames.size()) / X264_QUEUE_LENGTH;
 401                 }
 402                 speed_control->before_frame(queue_fill_ratio, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE);
 403         }
 404         dyn.x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic);
 405         if (speed_control) {
 406                 speed_control->after_frame();
 407         }
 408
 409         if (num_nal == 0) return;
 410
 411         if (use_separate_disk_params) {
 412                 if (IS_X264_TYPE_I(pic.i_type)) {
 413                         ++metric_x264_disk_output_frames_i;
 414                 } else if (IS_X264_TYPE_B(pic.i_type)) {
 415                         ++metric_x264_disk_output_frames_b;
 416                 } else {
 417                         ++metric_x264_disk_output_frames_p;
 418                 }
 419
 420                 metric_x264_disk_crf.count_event(pic.prop.f_crf_avg);
 421         } else {
 422                 if (IS_X264_TYPE_I(pic.i_type)) {
 423                         ++metric_x264_output_frames_i;
 424                 } else if (IS_X264_TYPE_B(pic.i_type)) {
 425                         ++metric_x264_output_frames_b;
 426                 } else {
 427                         ++metric_x264_output_frames_p;
 428                 }
 429
 430                 metric_x264_crf.count_event(pic.prop.f_crf_avg);
 431         }
 432
 433         if (frames_being_encoded.count(pic.i_pts)) {
 434                 ReceivedTimestamps received_ts = frames_being_encoded[pic.i_pts];
 435                 frames_being_encoded.erase(pic.i_pts);
 436
 437                 static int frameno = 0;
 438                 print_latency("Current x264 latency (video inputs → network mux):",
 439                         received_ts, (pic.i_type == X264_TYPE_B || pic.i_type == X264_TYPE_BREF),
 440                         &frameno, &x264_latency_histogram);
 441         } else {
 442                 assert(false);
 443         }
 444
 445         // We really need one AVPacket for the entire frame, it seems,
 446         // so combine it all.
 447         size_t num_bytes = buffered_sei.size();
 448         for (int i = 0; i < num_nal; ++i) {
 449                 num_bytes += nal[i].i_payload;
 450         }
 451
 452         unique_ptr<uint8_t[]> data(new uint8_t[num_bytes]);
 453         uint8_t *ptr = data.get();
 454
 455         if (!buffered_sei.empty()) {
 456                 memcpy(ptr, buffered_sei.data(), buffered_sei.size());
 457                 ptr += buffered_sei.size();
 458                 buffered_sei.clear();
 459         }
 460         for (int i = 0; i < num_nal; ++i) {
 461                 memcpy(ptr, nal[i].p_payload, nal[i].i_payload);
 462                 ptr += nal[i].i_payload;
 463         }
 464
 465         AVPacket pkt;
 466         memset(&pkt, 0, sizeof(pkt));
 467         pkt.buf = nullptr;
 468         pkt.data = data.get();
 469         pkt.size = num_bytes;
 470         pkt.stream_index = 0;
 471         if (pic.b_keyframe) {
 472                 pkt.flags = AV_PKT_FLAG_KEY;
 473         } else {
 474                 pkt.flags = 0;
 475         }
 476         pkt.duration = reinterpret_cast<intptr_t>(pic.opaque);
 477
 478         for (Mux *mux : muxes) {
 479                 mux->add_packet(pkt, pic.i_pts, pic.i_dts);
 480         }
 481 }
 482
 483 void X264Encoder::speed_control_override_func(unsigned bitrate_kbit, movit::YCbCrLumaCoefficients ycbcr_coefficients, x264_param_t *param)
 484 {
 485         if (bitrate_kbit != 0) {
 486                 param->rc.i_bitrate = bitrate_kbit;
 487                 update_vbv_settings(param);
 488         }
 489
 490         if (ycbcr_coefficients == YCBCR_REC_709) {
 491                 param->vui.i_colmatrix = 1;  // BT.709.
 492         } else {
 493                 assert(ycbcr_coefficients == YCBCR_REC_601);
 494                 param->vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 495         }
 496 }