git.sesse.net Git - nageru/blob - 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 <cstdint>
  11
  12 #include "defs.h"
  13 #include "flags.h"
  14 #include "metrics.h"
  15 #include "mux.h"
  16 #include "print_latency.h"
  17 #include "timebase.h"
  18 #include "x264_dynamic.h"
  19 #include "x264_speed_control.h"
  20
  21 extern "C" {
  22 #include <libavcodec/avcodec.h>
  23 #include <libavformat/avformat.h>
  24 }
  25
  26 using namespace movit;
  27 using namespace std;
  28 using namespace std::chrono;
  29
  30 namespace {
  31
  32 void update_vbv_settings(x264_param_t *param)
  33 {
  34         if (global_flags.x264_bitrate == -1) {
  35                 return;
  36         }
  37         if (global_flags.x264_vbv_buffer_size < 0) {
  38                 param->rc.i_vbv_buffer_size = param->rc.i_bitrate;  // One-second VBV.
  39         } else {
  40                 param->rc.i_vbv_buffer_size = global_flags.x264_vbv_buffer_size;
  41         }
  42         if (global_flags.x264_vbv_max_bitrate < 0) {
  43                 param->rc.i_vbv_max_bitrate = param->rc.i_bitrate;  // CBR.
  44         } else {
  45                 param->rc.i_vbv_max_bitrate = global_flags.x264_vbv_max_bitrate;
  46         }
  47 }
  48
  49 }  // namespace
  50
  51 X264Encoder::X264Encoder(AVOutputFormat *oformat)
  52         : wants_global_headers(oformat->flags & AVFMT_GLOBALHEADER),
  53           dyn(load_x264_for_bit_depth(global_flags.x264_bit_depth))
  54 {
  55         size_t bytes_per_pixel = global_flags.x264_bit_depth > 8 ? 2 : 1;
  56         frame_pool.reset(new uint8_t[global_flags.width * global_flags.height * 2 * bytes_per_pixel * X264_QUEUE_LENGTH]);
  57         for (unsigned i = 0; i < X264_QUEUE_LENGTH; ++i) {
  58                 free_frames.push(frame_pool.get() + i * (global_flags.width * global_flags.height * 2 * bytes_per_pixel));
  59         }
  60         encoder_thread = thread(&X264Encoder::encoder_thread_func, this);
  61
  62         global_metrics.add("x264_queued_frames", &metric_x264_queued_frames, Metrics::TYPE_GAUGE);
  63         global_metrics.add("x264_max_queued_frames", &metric_x264_max_queued_frames, Metrics::TYPE_GAUGE);
  64         global_metrics.add("x264_dropped_frames", &metric_x264_dropped_frames);
  65         global_metrics.add("x264_output_frames", {{ "type", "i" }}, &metric_x264_output_frames_i);
  66         global_metrics.add("x264_output_frames", {{ "type", "p" }}, &metric_x264_output_frames_p);
  67         global_metrics.add("x264_output_frames", {{ "type", "b" }}, &metric_x264_output_frames_b);
  68
  69         metric_x264_crf.init_uniform(50);
  70         global_metrics.add("x264_crf", &metric_x264_crf);
  71
  72         latency_histogram.init("x264");
  73 }
  74
  75 X264Encoder::~X264Encoder()
  76 {
  77         should_quit = true;
  78         queued_frames_nonempty.notify_all();
  79         encoder_thread.join();
  80         if (dyn.handle) {
  81                 dlclose(dyn.handle);
  82         }
  83 }
  84
  85 void X264Encoder::add_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const uint8_t *data, const ReceivedTimestamps &received_ts)
  86 {
  87         assert(!should_quit);
  88
  89         QueuedFrame qf;
  90         qf.pts = pts;
  91         qf.duration = duration;
  92         qf.ycbcr_coefficients = ycbcr_coefficients;
  93         qf.received_ts = received_ts;
  94
  95         {
  96                 lock_guard<mutex> lock(mu);
  97                 if (free_frames.empty()) {
  98                         fprintf(stderr, "WARNING: x264 queue full, dropping frame with pts %ld\n", pts);
  99                         ++metric_x264_dropped_frames;
 100                         return;
 101                 }
 102
 103                 qf.data = free_frames.front();
 104                 free_frames.pop();
 105         }
 106
 107         size_t bytes_per_pixel = global_flags.x264_bit_depth > 8 ? 2 : 1;
 108         memcpy(qf.data, data, global_flags.width * global_flags.height * 2 * bytes_per_pixel);
 109
 110         {
 111                 lock_guard<mutex> lock(mu);
 112                 queued_frames.push(qf);
 113                 queued_frames_nonempty.notify_all();
 114                 metric_x264_queued_frames = queued_frames.size();
 115         }
 116 }
 117
 118 void X264Encoder::init_x264()
 119 {
 120         x264_param_t param;
 121         dyn.x264_param_default_preset(&param, global_flags.x264_preset.c_str(), global_flags.x264_tune.c_str());
 122
 123         param.i_width = global_flags.width;
 124         param.i_height = global_flags.height;
 125         param.i_csp = X264_CSP_NV12;
 126         if (global_flags.x264_bit_depth > 8) {
 127                 param.i_csp |= X264_CSP_HIGH_DEPTH;
 128         }
 129         param.b_vfr_input = 1;
 130         param.i_timebase_num = 1;
 131         param.i_timebase_den = TIMEBASE;
 132         param.i_keyint_max = 50; // About one second.
 133         if (global_flags.x264_speedcontrol) {
 134                 param.i_frame_reference = 16;  // Because speedcontrol is never allowed to change this above what we set at start.
 135         }
 136
 137         // NOTE: These should be in sync with the ones in quicksync_encoder.cpp (sps_rbsp()).
 138         param.vui.i_vidformat = 5;  // Unspecified.
 139         param.vui.b_fullrange = 0;
 140         param.vui.i_colorprim = 1;  // BT.709.
 141         param.vui.i_transfer = 2;  // Unspecified (since we use sRGB).
 142         if (global_flags.ycbcr_rec709_coefficients) {
 143                 param.vui.i_colmatrix = 1;  // BT.709.
 144         } else {
 145                 param.vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 146         }
 147
 148         if (!isinf(global_flags.x264_crf)) {
 149                 param.rc.i_rc_method = X264_RC_CRF;
 150                 param.rc.f_rf_constant = global_flags.x264_crf;
 151         } else {
 152                 param.rc.i_rc_method = X264_RC_ABR;
 153                 param.rc.i_bitrate = global_flags.x264_bitrate;
 154         }
 155         update_vbv_settings(&param);
 156         if (param.rc.i_vbv_max_bitrate > 0) {
 157                 // If the user wants VBV control to cap the max rate, it is
 158                 // also reasonable to assume that they are fine with the stream
 159                 // constantly being around that rate even for very low-complexity
 160                 // content; the obvious and extreme example being a static
 161                 // black picture.
 162                 //
 163                 // One would think it's fine to have low-complexity content use
 164                 // less bitrate, but it seems to cause problems in practice;
 165                 // e.g. VLC seems to often drop the stream (similar to a buffer
 166                 // underrun) in such cases, but only when streaming from Nageru,
 167                 // not when reading a dump of the same stream from disk.
 168                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 169                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 170                 // but it's a typical case of problems that can arise. Similarly,
 171                 // TCP's congestion control is not always fond of the rate staying
 172                 // low for a while and then rising quickly -- a variation on the same
 173                 // problem.
 174                 //
 175                 // We solve this by simply asking x264 to fill in dummy bits
 176                 // in these cases, so that the bitrate stays reasonable constant.
 177                 // It's a waste of bandwidth, but it makes things go much more
 178                 // smoothly in these cases. (We don't do it if VBV control is off
 179                 // in general, not the least because it makes no sense and x264
 180                 // thus ignores the parameter.)
 181                 param.rc.b_filler = 1;
 182         }
 183
 184         // Occasionally players have problem with extremely low quantizers;
 185         // be on the safe side. Shouldn't affect quality in any meaningful way.
 186         param.rc.i_qp_min = 5;
 187
 188         for (const string &str : global_flags.x264_extra_param) {
 189                 const size_t pos = str.find(',');
 190                 if (pos == string::npos) {
 191                         if (dyn.x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 192                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 193                         }
 194                 } else {
 195                         const string key = str.substr(0, pos);
 196                         const string value = str.substr(pos + 1);
 197                         if (dyn.x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 198                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 199                                         key.c_str(), value.c_str());
 200                         }
 201                 }
 202         }
 203
 204         if (global_flags.x264_bit_depth > 8) {
 205                 dyn.x264_param_apply_profile(&param, "high10");
 206         } else {
 207                 dyn.x264_param_apply_profile(&param, "high");
 208         }
 209
 210         param.b_repeat_headers = !wants_global_headers;
 211
 212         x264 = dyn.x264_encoder_open(&param);
 213         if (x264 == nullptr) {
 214                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 215                 exit(1);
 216         }
 217
 218         if (global_flags.x264_speedcontrol) {
 219                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 220         }
 221
 222         if (wants_global_headers) {
 223                 x264_nal_t *nal;
 224                 int num_nal;
 225
 226                 dyn.x264_encoder_headers(x264, &nal, &num_nal);
 227
 228                 for (int i = 0; i < num_nal; ++i) {
 229                         if (nal[i].i_type == NAL_SEI) {
 230                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 231                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 232                         } else {
 233                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 234                         }
 235                 }
 236         }
 237 }
 238
 239 void X264Encoder::encoder_thread_func()
 240 {
 241         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 242                 perror("nice()");
 243                 // No exit; it's not fatal.
 244         }
 245         pthread_setname_np(pthread_self(), "x264_encode");
 246         init_x264();
 247         x264_init_done = true;
 248
 249         bool frames_left;
 250
 251         do {
 252                 QueuedFrame qf;
 253
 254                 // Wait for a queued frame, then dequeue it.
 255                 {
 256                         unique_lock<mutex> lock(mu);
 257                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 258                         if (!queued_frames.empty()) {
 259                                 qf = queued_frames.front();
 260                                 queued_frames.pop();
 261                         } else {
 262                                 qf.pts = -1;
 263                                 qf.duration = -1;
 264                                 qf.data = nullptr;
 265                         }
 266
 267                         metric_x264_queued_frames = queued_frames.size();
 268                         frames_left = !queued_frames.empty();
 269                 }
 270
 271                 encode_frame(qf);
 272
 273                 {
 274                         lock_guard<mutex> lock(mu);
 275                         free_frames.push(qf.data);
 276                 }
 277
 278                 // We should quit only if the should_quit flag is set _and_ we have nothing
 279                 // in either queue.
 280         } while (!should_quit || frames_left || dyn.x264_encoder_delayed_frames(x264) > 0);
 281
 282         dyn.x264_encoder_close(x264);
 283 }
 284
 285 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 286 {
 287         x264_nal_t *nal = nullptr;
 288         int num_nal = 0;
 289         x264_picture_t pic;
 290         x264_picture_t *input_pic = nullptr;
 291
 292         if (qf.data) {
 293                 dyn.x264_picture_init(&pic);
 294
 295                 pic.i_pts = qf.pts;
 296                 if (global_flags.x264_bit_depth > 8) {
 297                         pic.img.i_csp = X264_CSP_NV12 | X264_CSP_HIGH_DEPTH;
 298                         pic.img.i_plane = 2;
 299                         pic.img.plane[0] = qf.data;
 300                         pic.img.i_stride[0] = global_flags.width * sizeof(uint16_t);
 301                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height * sizeof(uint16_t);
 302                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint32_t);
 303                 } else {
 304                         pic.img.i_csp = X264_CSP_NV12;
 305                         pic.img.i_plane = 2;
 306                         pic.img.plane[0] = qf.data;
 307                         pic.img.i_stride[0] = global_flags.width;
 308                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height;
 309                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint16_t);
 310                 }
 311                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 312
 313                 input_pic = &pic;
 314
 315                 frames_being_encoded[qf.pts] = qf.received_ts;
 316         }
 317
 318         // See if we have a new bitrate to change to.
 319         unsigned new_rate = new_bitrate_kbit.exchange(0);  // Read and clear.
 320         if (new_rate != 0) {
 321                 bitrate_override_func = [new_rate](x264_param_t *param) {
 322                         param->rc.i_bitrate = new_rate;
 323                         update_vbv_settings(param);
 324                 };
 325         }
 326
 327         auto ycbcr_coefficients_override_func = [qf](x264_param_t *param) {
 328                 if (qf.ycbcr_coefficients == YCBCR_REC_709) {
 329                         param->vui.i_colmatrix = 1;  // BT.709.
 330                 } else {
 331                         assert(qf.ycbcr_coefficients == YCBCR_REC_601);
 332                         param->vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 333                 }
 334         };
 335
 336         if (speed_control) {
 337                 speed_control->set_config_override_function([this, ycbcr_coefficients_override_func](x264_param_t *param) {
 338                         if (bitrate_override_func) {
 339                                 bitrate_override_func(param);
 340                         }
 341                         ycbcr_coefficients_override_func(param);
 342                 });
 343         } else {
 344                 x264_param_t param;
 345                 dyn.x264_encoder_parameters(x264, &param);
 346                 if (bitrate_override_func) {
 347                         bitrate_override_func(&param);
 348                 }
 349                 ycbcr_coefficients_override_func(&param);
 350                 dyn.x264_encoder_reconfig(x264, &param);
 351         }
 352
 353         if (speed_control) {
 354                 speed_control->before_frame(float(free_frames.size()) / X264_QUEUE_LENGTH, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE);
 355         }
 356         dyn.x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic);
 357         if (speed_control) {
 358                 speed_control->after_frame();
 359         }
 360
 361         if (num_nal == 0) return;
 362
 363         if (IS_X264_TYPE_I(pic.i_type)) {
 364                 ++metric_x264_output_frames_i;
 365         } else if (IS_X264_TYPE_B(pic.i_type)) {
 366                 ++metric_x264_output_frames_b;
 367         } else {
 368                 ++metric_x264_output_frames_p;
 369         }
 370
 371         metric_x264_crf.count_event(pic.prop.f_crf_avg);
 372
 373         if (frames_being_encoded.count(pic.i_pts)) {
 374                 ReceivedTimestamps received_ts = frames_being_encoded[pic.i_pts];
 375                 frames_being_encoded.erase(pic.i_pts);
 376
 377                 static int frameno = 0;
 378                 print_latency("Current x264 latency (video inputs → network mux):",
 379                         received_ts, (pic.i_type == X264_TYPE_B || pic.i_type == X264_TYPE_BREF),
 380                         &frameno, &latency_histogram);
 381         } else {
 382                 assert(false);
 383         }
 384
 385         // We really need one AVPacket for the entire frame, it seems,
 386         // so combine it all.
 387         size_t num_bytes = buffered_sei.size();
 388         for (int i = 0; i < num_nal; ++i) {
 389                 num_bytes += nal[i].i_payload;
 390         }
 391
 392         unique_ptr<uint8_t[]> data(new uint8_t[num_bytes]);
 393         uint8_t *ptr = data.get();
 394
 395         if (!buffered_sei.empty()) {
 396                 memcpy(ptr, buffered_sei.data(), buffered_sei.size());
 397                 ptr += buffered_sei.size();
 398                 buffered_sei.clear();
 399         }
 400         for (int i = 0; i < num_nal; ++i) {
 401                 memcpy(ptr, nal[i].p_payload, nal[i].i_payload);
 402                 ptr += nal[i].i_payload;
 403         }
 404
 405         AVPacket pkt;
 406         memset(&pkt, 0, sizeof(pkt));
 407         pkt.buf = nullptr;
 408         pkt.data = data.get();
 409         pkt.size = num_bytes;
 410         pkt.stream_index = 0;
 411         if (pic.b_keyframe) {
 412                 pkt.flags = AV_PKT_FLAG_KEY;
 413         } else {
 414                 pkt.flags = 0;
 415         }
 416         pkt.duration = reinterpret_cast<intptr_t>(pic.opaque);
 417
 418         for (Mux *mux : muxes) {
 419                 mux->add_packet(pkt, pic.i_pts, pic.i_dts);
 420         }
 421 }