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 <atomic>
  11 #include <cstdint>
  12 #include <mutex>
  13
  14 #include "defs.h"
  15 #include "flags.h"
  16 #include "metrics.h"
  17 #include "mux.h"
  18 #include "print_latency.h"
  19 #include "timebase.h"
  20 #include "x264_dynamic.h"
  21 #include "x264_speed_control.h"
  22
  23 extern "C" {
  24 #include <libavcodec/avcodec.h>
  25 #include <libavformat/avformat.h>
  26 }
  27
  28 using namespace movit;
  29 using namespace std;
  30 using namespace std::chrono;
  31
  32 namespace {
  33
  34 // X264Encoder can be restarted if --record-x264-video is set, so make these
  35 // metrics global.
  36 atomic<int64_t> metric_x264_queued_frames{0};
  37 atomic<int64_t> metric_x264_max_queued_frames{X264_QUEUE_LENGTH};
  38 atomic<int64_t> metric_x264_dropped_frames{0};
  39 atomic<int64_t> metric_x264_output_frames_i{0};
  40 atomic<int64_t> metric_x264_output_frames_p{0};
  41 atomic<int64_t> metric_x264_output_frames_b{0};
  42 Histogram metric_x264_crf;
  43 LatencyHistogram x264_latency_histogram;
  44 once_flag x264_metrics_inited;
  45
  46 void update_vbv_settings(x264_param_t *param)
  47 {
  48         if (global_flags.x264_bitrate == -1) {
  49                 return;
  50         }
  51         if (global_flags.x264_vbv_buffer_size < 0) {
  52                 param->rc.i_vbv_buffer_size = param->rc.i_bitrate;  // One-second VBV.
  53         } else {
  54                 param->rc.i_vbv_buffer_size = global_flags.x264_vbv_buffer_size;
  55         }
  56         if (global_flags.x264_vbv_max_bitrate < 0) {
  57                 param->rc.i_vbv_max_bitrate = param->rc.i_bitrate;  // CBR.
  58         } else {
  59                 param->rc.i_vbv_max_bitrate = global_flags.x264_vbv_max_bitrate;
  60         }
  61 }
  62
  63 }  // namespace
  64
  65 X264Encoder::X264Encoder(AVOutputFormat *oformat)
  66         : wants_global_headers(oformat->flags & AVFMT_GLOBALHEADER),
  67           dyn(load_x264_for_bit_depth(global_flags.x264_bit_depth))
  68 {
  69         call_once(x264_metrics_inited, [](){
  70                 global_metrics.add("x264_queued_frames", &metric_x264_queued_frames, Metrics::TYPE_GAUGE);
  71                 global_metrics.add("x264_max_queued_frames", &metric_x264_max_queued_frames, Metrics::TYPE_GAUGE);
  72                 global_metrics.add("x264_dropped_frames", &metric_x264_dropped_frames);
  73                 global_metrics.add("x264_output_frames", {{ "type", "i" }}, &metric_x264_output_frames_i);
  74                 global_metrics.add("x264_output_frames", {{ "type", "p" }}, &metric_x264_output_frames_p);
  75                 global_metrics.add("x264_output_frames", {{ "type", "b" }}, &metric_x264_output_frames_b);
  76
  77                 metric_x264_crf.init_uniform(50);
  78                 global_metrics.add("x264_crf", &metric_x264_crf);
  79                 x264_latency_histogram.init("x264");
  80         });
  81
  82         size_t bytes_per_pixel = global_flags.x264_bit_depth > 8 ? 2 : 1;
  83         frame_pool.reset(new uint8_t[global_flags.width * global_flags.height * 2 * bytes_per_pixel * X264_QUEUE_LENGTH]);
  84         for (unsigned i = 0; i < X264_QUEUE_LENGTH; ++i) {
  85                 free_frames.push(frame_pool.get() + i * (global_flags.width * global_flags.height * 2 * bytes_per_pixel));
  86         }
  87         encoder_thread = thread(&X264Encoder::encoder_thread_func, this);
  88 }
  89
  90 X264Encoder::~X264Encoder()
  91 {
  92         should_quit = true;
  93         queued_frames_nonempty.notify_all();
  94         encoder_thread.join();
  95         if (dyn.handle) {
  96                 dlclose(dyn.handle);
  97         }
  98 }
  99
 100 void X264Encoder::add_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const uint8_t *data, const ReceivedTimestamps &received_ts)
 101 {
 102         assert(!should_quit);
 103
 104         QueuedFrame qf;
 105         qf.pts = pts;
 106         qf.duration = duration;
 107         qf.ycbcr_coefficients = ycbcr_coefficients;
 108         qf.received_ts = received_ts;
 109
 110         {
 111                 lock_guard<mutex> lock(mu);
 112                 if (free_frames.empty()) {
 113                         fprintf(stderr, "WARNING: x264 queue full, dropping frame with pts %ld\n", pts);
 114                         ++metric_x264_dropped_frames;
 115                         return;
 116                 }
 117
 118                 qf.data = free_frames.front();
 119                 free_frames.pop();
 120         }
 121
 122         size_t bytes_per_pixel = global_flags.x264_bit_depth > 8 ? 2 : 1;
 123         memcpy(qf.data, data, global_flags.width * global_flags.height * 2 * bytes_per_pixel);
 124
 125         {
 126                 lock_guard<mutex> lock(mu);
 127                 queued_frames.push(qf);
 128                 queued_frames_nonempty.notify_all();
 129                 metric_x264_queued_frames = queued_frames.size();
 130         }
 131 }
 132
 133 void X264Encoder::init_x264()
 134 {
 135         x264_param_t param;
 136         dyn.x264_param_default_preset(&param, global_flags.x264_preset.c_str(), global_flags.x264_tune.c_str());
 137
 138         param.i_width = global_flags.width;
 139         param.i_height = global_flags.height;
 140         param.i_csp = X264_CSP_NV12;
 141         if (global_flags.x264_bit_depth > 8) {
 142                 param.i_csp |= X264_CSP_HIGH_DEPTH;
 143         }
 144         param.b_vfr_input = 1;
 145         param.i_timebase_num = 1;
 146         param.i_timebase_den = TIMEBASE;
 147         param.i_keyint_max = 50; // About one second.
 148         if (global_flags.x264_speedcontrol) {
 149                 param.i_frame_reference = 16;  // Because speedcontrol is never allowed to change this above what we set at start.
 150         }
 151
 152         // NOTE: These should be in sync with the ones in quicksync_encoder.cpp (sps_rbsp()).
 153         param.vui.i_vidformat = 5;  // Unspecified.
 154         param.vui.b_fullrange = 0;
 155         param.vui.i_colorprim = 1;  // BT.709.
 156         param.vui.i_transfer = 13;  // sRGB.
 157         if (global_flags.ycbcr_rec709_coefficients) {
 158                 param.vui.i_colmatrix = 1;  // BT.709.
 159         } else {
 160                 param.vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 161         }
 162
 163         if (!isinf(global_flags.x264_crf)) {
 164                 param.rc.i_rc_method = X264_RC_CRF;
 165                 param.rc.f_rf_constant = global_flags.x264_crf;
 166         } else {
 167                 param.rc.i_rc_method = X264_RC_ABR;
 168                 param.rc.i_bitrate = global_flags.x264_bitrate;
 169         }
 170         update_vbv_settings(&param);
 171         if (param.rc.i_vbv_max_bitrate > 0) {
 172                 // If the user wants VBV control to cap the max rate, it is
 173                 // also reasonable to assume that they are fine with the stream
 174                 // constantly being around that rate even for very low-complexity
 175                 // content; the obvious and extreme example being a static
 176                 // black picture.
 177                 //
 178                 // One would think it's fine to have low-complexity content use
 179                 // less bitrate, but it seems to cause problems in practice;
 180                 // e.g. VLC seems to often drop the stream (similar to a buffer
 181                 // underrun) in such cases, but only when streaming from Nageru,
 182                 // not when reading a dump of the same stream from disk.
 183                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 184                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 185                 // but it's a typical case of problems that can arise. Similarly,
 186                 // TCP's congestion control is not always fond of the rate staying
 187                 // low for a while and then rising quickly -- a variation on the same
 188                 // problem.
 189                 //
 190                 // We solve this by simply asking x264 to fill in dummy bits
 191                 // in these cases, so that the bitrate stays reasonable constant.
 192                 // It's a waste of bandwidth, but it makes things go much more
 193                 // smoothly in these cases. (We don't do it if VBV control is off
 194                 // in general, not the least because it makes no sense and x264
 195                 // thus ignores the parameter.)
 196                 param.rc.b_filler = 1;
 197         }
 198
 199         // Occasionally players have problem with extremely low quantizers;
 200         // be on the safe side. Shouldn't affect quality in any meaningful way.
 201         param.rc.i_qp_min = 5;
 202
 203         for (const string &str : global_flags.x264_extra_param) {
 204                 const size_t pos = str.find(',');
 205                 if (pos == string::npos) {
 206                         if (dyn.x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 207                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 208                         }
 209                 } else {
 210                         const string key = str.substr(0, pos);
 211                         const string value = str.substr(pos + 1);
 212                         if (dyn.x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 213                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 214                                         key.c_str(), value.c_str());
 215                         }
 216                 }
 217         }
 218
 219         if (global_flags.x264_bit_depth > 8) {
 220                 dyn.x264_param_apply_profile(&param, "high10");
 221         } else {
 222                 dyn.x264_param_apply_profile(&param, "high");
 223         }
 224
 225         param.b_repeat_headers = !wants_global_headers;
 226
 227         x264 = dyn.x264_encoder_open(&param);
 228         if (x264 == nullptr) {
 229                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 230                 exit(1);
 231         }
 232
 233         if (global_flags.x264_speedcontrol) {
 234                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 235         }
 236
 237         if (wants_global_headers) {
 238                 x264_nal_t *nal;
 239                 int num_nal;
 240
 241                 dyn.x264_encoder_headers(x264, &nal, &num_nal);
 242
 243                 for (int i = 0; i < num_nal; ++i) {
 244                         if (nal[i].i_type == NAL_SEI) {
 245                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 246                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 247                         } else {
 248                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 249                         }
 250                 }
 251         }
 252 }
 253
 254 void X264Encoder::encoder_thread_func()
 255 {
 256         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 257                 perror("nice()");
 258                 // No exit; it's not fatal.
 259         }
 260         pthread_setname_np(pthread_self(), "x264_encode");
 261         init_x264();
 262         x264_init_done = true;
 263
 264         bool frames_left;
 265
 266         do {
 267                 QueuedFrame qf;
 268
 269                 // Wait for a queued frame, then dequeue it.
 270                 {
 271                         unique_lock<mutex> lock(mu);
 272                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 273                         if (!queued_frames.empty()) {
 274                                 qf = queued_frames.front();
 275                                 queued_frames.pop();
 276                         } else {
 277                                 qf.pts = -1;
 278                                 qf.duration = -1;
 279                                 qf.data = nullptr;
 280                         }
 281
 282                         metric_x264_queued_frames = queued_frames.size();
 283                         frames_left = !queued_frames.empty();
 284                 }
 285
 286                 encode_frame(qf);
 287
 288                 {
 289                         lock_guard<mutex> lock(mu);
 290                         free_frames.push(qf.data);
 291                 }
 292
 293                 // We should quit only if the should_quit flag is set _and_ we have nothing
 294                 // in either queue.
 295         } while (!should_quit || frames_left || dyn.x264_encoder_delayed_frames(x264) > 0);
 296
 297         dyn.x264_encoder_close(x264);
 298 }
 299
 300 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 301 {
 302         x264_nal_t *nal = nullptr;
 303         int num_nal = 0;
 304         x264_picture_t pic;
 305         x264_picture_t *input_pic = nullptr;
 306
 307         if (qf.data) {
 308                 dyn.x264_picture_init(&pic);
 309
 310                 pic.i_pts = qf.pts;
 311                 if (global_flags.x264_bit_depth > 8) {
 312                         pic.img.i_csp = X264_CSP_NV12 | X264_CSP_HIGH_DEPTH;
 313                         pic.img.i_plane = 2;
 314                         pic.img.plane[0] = qf.data;
 315                         pic.img.i_stride[0] = global_flags.width * sizeof(uint16_t);
 316                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height * sizeof(uint16_t);
 317                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint32_t);
 318                 } else {
 319                         pic.img.i_csp = X264_CSP_NV12;
 320                         pic.img.i_plane = 2;
 321                         pic.img.plane[0] = qf.data;
 322                         pic.img.i_stride[0] = global_flags.width;
 323                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height;
 324                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint16_t);
 325                 }
 326                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 327
 328                 input_pic = &pic;
 329
 330                 frames_being_encoded[qf.pts] = qf.received_ts;
 331         }
 332
 333         // See if we have a new bitrate to change to.
 334         unsigned new_rate = new_bitrate_kbit.exchange(0);  // Read and clear.
 335         if (new_rate != 0) {
 336                 bitrate_override_func = [new_rate](x264_param_t *param) {
 337                         param->rc.i_bitrate = new_rate;
 338                         update_vbv_settings(param);
 339                 };
 340         }
 341
 342         auto ycbcr_coefficients_override_func = [qf](x264_param_t *param) {
 343                 if (qf.ycbcr_coefficients == YCBCR_REC_709) {
 344                         param->vui.i_colmatrix = 1;  // BT.709.
 345                 } else {
 346                         assert(qf.ycbcr_coefficients == YCBCR_REC_601);
 347                         param->vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 348                 }
 349         };
 350
 351         if (speed_control) {
 352                 speed_control->set_config_override_function([this, ycbcr_coefficients_override_func](x264_param_t *param) {
 353                         if (bitrate_override_func) {
 354                                 bitrate_override_func(param);
 355                         }
 356                         ycbcr_coefficients_override_func(param);
 357                 });
 358         } else {
 359                 x264_param_t param;
 360                 dyn.x264_encoder_parameters(x264, &param);
 361                 if (bitrate_override_func) {
 362                         bitrate_override_func(&param);
 363                 }
 364                 ycbcr_coefficients_override_func(&param);
 365                 dyn.x264_encoder_reconfig(x264, &param);
 366         }
 367
 368         if (speed_control) {
 369                 speed_control->before_frame(float(free_frames.size()) / X264_QUEUE_LENGTH, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE);
 370         }
 371         dyn.x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic);
 372         if (speed_control) {
 373                 speed_control->after_frame();
 374         }
 375
 376         if (num_nal == 0) return;
 377
 378         if (IS_X264_TYPE_I(pic.i_type)) {
 379                 ++metric_x264_output_frames_i;
 380         } else if (IS_X264_TYPE_B(pic.i_type)) {
 381                 ++metric_x264_output_frames_b;
 382         } else {
 383                 ++metric_x264_output_frames_p;
 384         }
 385
 386         metric_x264_crf.count_event(pic.prop.f_crf_avg);
 387
 388         if (frames_being_encoded.count(pic.i_pts)) {
 389                 ReceivedTimestamps received_ts = frames_being_encoded[pic.i_pts];
 390                 frames_being_encoded.erase(pic.i_pts);
 391
 392                 static int frameno = 0;
 393                 print_latency("Current x264 latency (video inputs → network mux):",
 394                         received_ts, (pic.i_type == X264_TYPE_B || pic.i_type == X264_TYPE_BREF),
 395                         &frameno, &x264_latency_histogram);
 396         } else {
 397                 assert(false);
 398         }
 399
 400         // We really need one AVPacket for the entire frame, it seems,
 401         // so combine it all.
 402         size_t num_bytes = buffered_sei.size();
 403         for (int i = 0; i < num_nal; ++i) {
 404                 num_bytes += nal[i].i_payload;
 405         }
 406
 407         unique_ptr<uint8_t[]> data(new uint8_t[num_bytes]);
 408         uint8_t *ptr = data.get();
 409
 410         if (!buffered_sei.empty()) {
 411                 memcpy(ptr, buffered_sei.data(), buffered_sei.size());
 412                 ptr += buffered_sei.size();
 413                 buffered_sei.clear();
 414         }
 415         for (int i = 0; i < num_nal; ++i) {
 416                 memcpy(ptr, nal[i].p_payload, nal[i].i_payload);
 417                 ptr += nal[i].i_payload;
 418         }
 419
 420         AVPacket pkt;
 421         memset(&pkt, 0, sizeof(pkt));
 422         pkt.buf = nullptr;
 423         pkt.data = data.get();
 424         pkt.size = num_bytes;
 425         pkt.stream_index = 0;
 426         if (pic.b_keyframe) {
 427                 pkt.flags = AV_PKT_FLAG_KEY;
 428         } else {
 429                 pkt.flags = 0;
 430         }
 431         pkt.duration = reinterpret_cast<intptr_t>(pic.opaque);
 432
 433         for (Mux *mux : muxes) {
 434                 mux->add_packet(pkt, pic.i_pts, pic.i_dts);
 435         }
 436 }