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

   1 #include "x264_encoder.h"
   2
   3 #include <assert.h>
   4 #include <atomic>
   5 #include <cstdint>
   6 #include <dlfcn.h>
   7 #include <functional>
   8 #include <inttypes.h>
   9 #include <math.h>
  10 #include <memory>
  11 #include <movit/image_format.h>
  12 #include <mutex>
  13 #include <pthread.h>
  14 #include <stdio.h>
  15 #include <stdlib.h>
  16 #include <string.h>
  17 #include <string>
  18 #include <thread>
  19 #include <unistd.h>
  20 #include <vector>
  21 #include <x264.h>
  22
  23 #include "defs.h"
  24 #include "flags.h"
  25 #include "shared/metrics.h"
  26 #include "shared/mux.h"
  27 #include "print_latency.h"
  28 #include "shared/timebase.h"
  29 #include "x264_dynamic.h"
  30 #include "x264_speed_control.h"
  31
  32 extern "C" {
  33 #include <libavcodec/packet.h>
  34 #include <libavformat/avformat.h>
  35 }
  36
  37 using namespace movit;
  38 using namespace std;
  39 using namespace std::chrono;
  40 using namespace std::placeholders;
  41
  42 namespace {
  43
  44 // X264Encoder can be restarted if --record-x264-video is set, so make these
  45 // metrics global.
  46 atomic<int64_t> metric_x264_queued_frames{0};
  47 atomic<int64_t> metric_x264_max_queued_frames{X264_QUEUE_LENGTH};
  48 atomic<int64_t> metric_x264_dropped_frames{0};
  49 atomic<int64_t> metric_x264_output_frames_i{0};
  50 atomic<int64_t> metric_x264_output_frames_p{0};
  51 atomic<int64_t> metric_x264_output_frames_b{0};
  52 Histogram metric_x264_crf;
  53 LatencyHistogram x264_latency_histogram;
  54
  55 atomic<int64_t> metric_x264_disk_queued_frames{0};
  56 atomic<int64_t> metric_x264_disk_max_queued_frames{X264_QUEUE_LENGTH};
  57 atomic<int64_t> metric_x264_disk_dropped_frames{0};
  58 atomic<int64_t> metric_x264_disk_output_frames_i{0};
  59 atomic<int64_t> metric_x264_disk_output_frames_p{0};
  60 atomic<int64_t> metric_x264_disk_output_frames_b{0};
  61 Histogram metric_x264_disk_crf;
  62 LatencyHistogram x264_disk_latency_histogram;
  63
  64 once_flag x264_metrics_inited, x264_disk_metrics_inited;
  65
  66 void update_vbv_settings(x264_param_t *param)
  67 {
  68         if (global_flags.x264_bitrate == -1) {
  69                 return;
  70         }
  71         param->rc.i_vbv_buffer_size = param->rc.i_bitrate;  // One-second VBV.
  72         param->rc.i_vbv_max_bitrate = param->rc.i_bitrate;  // CBR.
  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.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.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.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.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.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 the user wants to cap the max rate, it is also reasonable
 217                 // to assume that they are fine with the stream constantly
 218                 // being around that rate even for very low-complexity content;
 219                 // the obvious and extreme example being a static black picture.
 220                 //
 221                 // One would think it's fine to have low-complexity content use
 222                 // less bitrate, but it seems to cause problems in practice;
 223                 // e.g. VLC seems to often drop the stream (similar to a buffer
 224                 // underrun) in such cases, but only when streaming from Nageru,
 225                 // not when reading a dump of the same stream from disk.
 226                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 227                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 228                 // but it's a typical case of problems that can arise. Similarly,
 229                 // TCP's congestion control is not always fond of the rate staying
 230                 // low for a while and then rising quickly -- a variation on the same
 231                 // problem.
 232                 //
 233                 // We solve this by simply asking x264 to fill in dummy bits
 234                 // in these cases, so that the bitrate stays reasonable constant.
 235                 // It's a waste of bandwidth, but it makes things go much more
 236                 // smoothly in these cases. (We don't do it if VBV control is off
 237                 // in general, not the least because it makes no sense and x264
 238                 // thus ignores the parameter.)
 239                 param.rc.b_filler = 1;
 240         }
 241         if (!use_separate_disk_params) {
 242                 update_vbv_settings(&param);
 243         }
 244
 245         // Occasionally players have problem with extremely low quantizers;
 246         // be on the safe side. Shouldn't affect quality in any meaningful way.
 247         param.rc.i_qp_min = 5;
 248
 249         const vector<string> &extra_param = use_separate_disk_params ? global_flags.x264_separate_disk_extra_param : global_flags.x264_extra_param;
 250         for (const string &str : extra_param) {
 251                 const size_t pos = str.find(',');
 252                 if (pos == string::npos) {
 253                         if (dyn.x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 254                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 255                         }
 256                 } else {
 257                         const string key = str.substr(0, pos);
 258                         const string value = str.substr(pos + 1);
 259                         if (dyn.x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 260                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 261                                         key.c_str(), value.c_str());
 262                         }
 263                 }
 264         }
 265
 266         if (global_flags.bit_depth > 8) {
 267                 dyn.x264_param_apply_profile(&param, "high10");
 268         } else {
 269                 dyn.x264_param_apply_profile(&param, "high");
 270         }
 271
 272         param.b_repeat_headers = !wants_global_headers;
 273
 274         x264 = dyn.x264_encoder_open(&param);
 275         if (x264 == nullptr) {
 276                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 277                 abort();
 278         }
 279
 280         if (!use_separate_disk_params && global_flags.x264_speedcontrol) {
 281                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 282         }
 283
 284         if (wants_global_headers) {
 285                 x264_nal_t *nal;
 286                 int num_nal;
 287
 288                 dyn.x264_encoder_headers(x264, &nal, &num_nal);
 289
 290                 for (int i = 0; i < num_nal; ++i) {
 291                         if (nal[i].i_type == NAL_SEI) {
 292                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 293                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 294                         } else {
 295                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 296                         }
 297                 }
 298         }
 299 }
 300
 301 void X264Encoder::encoder_thread_func()
 302 {
 303         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 304                 perror("nice()");
 305                 // No exit; it's not fatal.
 306         }
 307         pthread_setname_np(pthread_self(), "x264_encode");
 308         init_x264();
 309         x264_init_done = true;
 310
 311         bool frames_left;
 312
 313         do {
 314                 QueuedFrame qf;
 315
 316                 // Wait for a queued frame, then dequeue it.
 317                 {
 318                         unique_lock<mutex> lock(mu);
 319                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 320                         if (!queued_frames.empty()) {
 321                                 qf = queued_frames.front();
 322                                 queued_frames.pop();
 323                         } else {
 324                                 qf.pts = -1;
 325                                 qf.duration = -1;
 326                                 qf.data = nullptr;
 327                         }
 328
 329                         if (use_separate_disk_params) {
 330                                 metric_x264_disk_queued_frames = queued_frames.size();
 331                         } else {
 332                                 metric_x264_queued_frames = queued_frames.size();
 333                         }
 334                         frames_left = !queued_frames.empty();
 335                 }
 336
 337                 encode_frame(qf);
 338
 339                 {
 340                         lock_guard<mutex> lock(mu);
 341                         free_frames.push(qf.data);
 342                 }
 343
 344                 // We should quit only if the should_quit flag is set _and_ we have nothing
 345                 // in either queue.
 346         } while (!should_quit || frames_left || dyn.x264_encoder_delayed_frames(x264) > 0);
 347
 348         dyn.x264_encoder_close(x264);
 349 }
 350
 351 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 352 {
 353         x264_nal_t *nal = nullptr;
 354         int num_nal = 0;
 355         x264_picture_t pic;
 356         x264_picture_t *input_pic = nullptr;
 357
 358         if (qf.data) {
 359                 dyn.x264_picture_init(&pic);
 360
 361                 pic.i_pts = qf.pts;
 362                 if (global_flags.bit_depth > 8) {
 363                         pic.img.i_csp = X264_CSP_NV12 | X264_CSP_HIGH_DEPTH;
 364                         pic.img.i_plane = 2;
 365                         pic.img.plane[0] = qf.data;
 366                         pic.img.i_stride[0] = global_flags.width * sizeof(uint16_t);
 367                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height * sizeof(uint16_t);
 368                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint32_t);
 369                 } else {
 370                         pic.img.i_csp = X264_CSP_NV12;
 371                         pic.img.i_plane = 2;
 372                         pic.img.plane[0] = qf.data;
 373                         pic.img.i_stride[0] = global_flags.width;
 374                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height;
 375                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint16_t);
 376                 }
 377                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 378
 379                 input_pic = &pic;
 380
 381                 frames_being_encoded[qf.pts] = qf.received_ts;
 382         }
 383
 384         unsigned new_rate = new_bitrate_kbit.load();  // Can be 0 for no change.
 385         if (speed_control) {
 386                 speed_control->set_config_override_function(bind(&speed_control_override_func, new_rate, qf.ycbcr_coefficients, _1));
 387         } else {
 388                 x264_param_t param;
 389                 dyn.x264_encoder_parameters(x264, &param);
 390                 speed_control_override_func(new_rate, qf.ycbcr_coefficients, &param);
 391                 dyn.x264_encoder_reconfig(x264, &param);
 392         }
 393
 394         if (speed_control) {
 395                 float queue_fill_ratio;
 396                 {
 397                         lock_guard<mutex> lock(mu);
 398                         queue_fill_ratio = float(free_frames.size()) / X264_QUEUE_LENGTH;
 399                 }
 400                 speed_control->before_frame(queue_fill_ratio, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE);
 401         }
 402         dyn.x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic);
 403         if (speed_control) {
 404                 speed_control->after_frame();
 405         }
 406
 407         if (num_nal == 0) return;
 408
 409         if (use_separate_disk_params) {
 410                 if (IS_X264_TYPE_I(pic.i_type)) {
 411                         ++metric_x264_disk_output_frames_i;
 412                 } else if (IS_X264_TYPE_B(pic.i_type)) {
 413                         ++metric_x264_disk_output_frames_b;
 414                 } else {
 415                         ++metric_x264_disk_output_frames_p;
 416                 }
 417
 418                 metric_x264_disk_crf.count_event(pic.prop.f_crf_avg);
 419         } else {
 420                 if (IS_X264_TYPE_I(pic.i_type)) {
 421                         ++metric_x264_output_frames_i;
 422                 } else if (IS_X264_TYPE_B(pic.i_type)) {
 423                         ++metric_x264_output_frames_b;
 424                 } else {
 425                         ++metric_x264_output_frames_p;
 426                 }
 427
 428                 metric_x264_crf.count_event(pic.prop.f_crf_avg);
 429         }
 430
 431         if (frames_being_encoded.count(pic.i_pts)) {
 432                 ReceivedTimestamps received_ts = frames_being_encoded[pic.i_pts];
 433                 frames_being_encoded.erase(pic.i_pts);
 434
 435                 static int frameno = 0;
 436                 print_latency("Current x264 latency (video inputs → network mux):",
 437                         received_ts, (pic.i_type == X264_TYPE_B || pic.i_type == X264_TYPE_BREF),
 438                         &frameno, &x264_latency_histogram);
 439         } else {
 440                 assert(false);
 441         }
 442
 443         // We really need one AVPacket for the entire frame, it seems,
 444         // so combine it all.
 445         size_t num_bytes = buffered_sei.size();
 446         for (int i = 0; i < num_nal; ++i) {
 447                 num_bytes += nal[i].i_payload;
 448         }
 449
 450         unique_ptr<uint8_t[]> data(new uint8_t[num_bytes]);
 451         uint8_t *ptr = data.get();
 452
 453         if (!buffered_sei.empty()) {
 454                 memcpy(ptr, buffered_sei.data(), buffered_sei.size());
 455                 ptr += buffered_sei.size();
 456                 buffered_sei.clear();
 457         }
 458         for (int i = 0; i < num_nal; ++i) {
 459                 memcpy(ptr, nal[i].p_payload, nal[i].i_payload);
 460                 ptr += nal[i].i_payload;
 461         }
 462
 463         AVPacket pkt;
 464         memset(&pkt, 0, sizeof(pkt));
 465         pkt.buf = nullptr;
 466         pkt.data = data.get();
 467         pkt.size = num_bytes;
 468         pkt.stream_index = 0;
 469         if (pic.b_keyframe) {
 470                 pkt.flags = AV_PKT_FLAG_KEY;
 471         } else {
 472                 pkt.flags = 0;
 473         }
 474         pkt.duration = reinterpret_cast<intptr_t>(pic.opaque);
 475
 476         for (Mux *mux : muxes) {
 477                 mux->add_packet(pkt, pic.i_pts, pic.i_dts);
 478         }
 479 }
 480
 481 void X264Encoder::speed_control_override_func(unsigned bitrate_kbit, movit::YCbCrLumaCoefficients ycbcr_coefficients, x264_param_t *param)
 482 {
 483         if (bitrate_kbit != 0) {
 484                 param->rc.i_bitrate = bitrate_kbit;
 485                 update_vbv_settings(param);
 486         }
 487
 488         if (ycbcr_coefficients == YCBCR_REC_709) {
 489                 param->vui.i_colmatrix = 1;  // BT.709.
 490         } else {
 491                 assert(ycbcr_coefficients == YCBCR_REC_601);
 492                 param->vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 493         }
 494 }