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 <functional>
  13 #include <mutex>
  14
  15 #include "defs.h"
  16 #include "flags.h"
  17 #include "metrics.h"
  18 #include "mux.h"
  19 #include "print_latency.h"
  20 #include "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 once_flag x264_metrics_inited;
  47
  48 void update_vbv_settings(x264_param_t *param)
  49 {
  50         if (global_flags.x264_bitrate == -1) {
  51                 return;
  52         }
  53         if (global_flags.x264_vbv_buffer_size < 0) {
  54                 param->rc.i_vbv_buffer_size = param->rc.i_bitrate;  // One-second VBV.
  55         } else {
  56                 param->rc.i_vbv_buffer_size = global_flags.x264_vbv_buffer_size;
  57         }
  58         if (global_flags.x264_vbv_max_bitrate < 0) {
  59                 param->rc.i_vbv_max_bitrate = param->rc.i_bitrate;  // CBR.
  60         } else {
  61                 param->rc.i_vbv_max_bitrate = global_flags.x264_vbv_max_bitrate;
  62         }
  63 }
  64
  65 }  // namespace
  66
  67 X264Encoder::X264Encoder(AVOutputFormat *oformat)
  68         : wants_global_headers(oformat->flags & AVFMT_GLOBALHEADER),
  69           dyn(load_x264_for_bit_depth(global_flags.x264_bit_depth))
  70 {
  71         call_once(x264_metrics_inited, [](){
  72                 global_metrics.add("x264_queued_frames", &metric_x264_queued_frames, Metrics::TYPE_GAUGE);
  73                 global_metrics.add("x264_max_queued_frames", &metric_x264_max_queued_frames, Metrics::TYPE_GAUGE);
  74                 global_metrics.add("x264_dropped_frames", &metric_x264_dropped_frames);
  75                 global_metrics.add("x264_output_frames", {{ "type", "i" }}, &metric_x264_output_frames_i);
  76                 global_metrics.add("x264_output_frames", {{ "type", "p" }}, &metric_x264_output_frames_p);
  77                 global_metrics.add("x264_output_frames", {{ "type", "b" }}, &metric_x264_output_frames_b);
  78
  79                 metric_x264_crf.init_uniform(50);
  80                 global_metrics.add("x264_crf", &metric_x264_crf);
  81                 x264_latency_histogram.init("x264");
  82         });
  83
  84         size_t bytes_per_pixel = global_flags.x264_bit_depth > 8 ? 2 : 1;
  85         frame_pool.reset(new uint8_t[global_flags.width * global_flags.height * 2 * bytes_per_pixel * X264_QUEUE_LENGTH]);
  86         for (unsigned i = 0; i < X264_QUEUE_LENGTH; ++i) {
  87                 free_frames.push(frame_pool.get() + i * (global_flags.width * global_flags.height * 2 * bytes_per_pixel));
  88         }
  89         encoder_thread = thread(&X264Encoder::encoder_thread_func, this);
  90 }
  91
  92 X264Encoder::~X264Encoder()
  93 {
  94         should_quit = true;
  95         queued_frames_nonempty.notify_all();
  96         encoder_thread.join();
  97         if (dyn.handle) {
  98                 dlclose(dyn.handle);
  99         }
 100 }
 101
 102 void X264Encoder::add_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const uint8_t *data, const ReceivedTimestamps &received_ts)
 103 {
 104         assert(!should_quit);
 105
 106         QueuedFrame qf;
 107         qf.pts = pts;
 108         qf.duration = duration;
 109         qf.ycbcr_coefficients = ycbcr_coefficients;
 110         qf.received_ts = received_ts;
 111
 112         {
 113                 lock_guard<mutex> lock(mu);
 114                 if (free_frames.empty()) {
 115                         fprintf(stderr, "WARNING: x264 queue full, dropping frame with pts %ld\n", pts);
 116                         ++metric_x264_dropped_frames;
 117                         return;
 118                 }
 119
 120                 qf.data = free_frames.front();
 121                 free_frames.pop();
 122         }
 123
 124         size_t bytes_per_pixel = global_flags.x264_bit_depth > 8 ? 2 : 1;
 125         memcpy(qf.data, data, global_flags.width * global_flags.height * 2 * bytes_per_pixel);
 126
 127         {
 128                 lock_guard<mutex> lock(mu);
 129                 queued_frames.push(qf);
 130                 queued_frames_nonempty.notify_all();
 131                 metric_x264_queued_frames = queued_frames.size();
 132         }
 133 }
 134
 135 void X264Encoder::init_x264()
 136 {
 137         x264_param_t param;
 138         dyn.x264_param_default_preset(&param, global_flags.x264_preset.c_str(), global_flags.x264_tune.c_str());
 139
 140         param.i_width = global_flags.width;
 141         param.i_height = global_flags.height;
 142         param.i_csp = X264_CSP_NV12;
 143         if (global_flags.x264_bit_depth > 8) {
 144                 param.i_csp |= X264_CSP_HIGH_DEPTH;
 145         }
 146         param.b_vfr_input = 1;
 147         param.i_timebase_num = 1;
 148         param.i_timebase_den = TIMEBASE;
 149         param.i_keyint_max = 50; // About one second.
 150         if (global_flags.x264_speedcontrol) {
 151                 param.i_frame_reference = 16;  // Because speedcontrol is never allowed to change this above what we set at start.
 152         }
 153 #if X264_BUILD >= 153
 154         param.i_bitdepth = global_flags.x264_bit_depth;
 155 #endif
 156
 157         // NOTE: These should be in sync with the ones in quicksync_encoder.cpp (sps_rbsp()).
 158         param.vui.i_vidformat = 5;  // Unspecified.
 159         param.vui.b_fullrange = 0;
 160         param.vui.i_colorprim = 1;  // BT.709.
 161         param.vui.i_transfer = 13;  // sRGB.
 162         if (global_flags.ycbcr_rec709_coefficients) {
 163                 param.vui.i_colmatrix = 1;  // BT.709.
 164         } else {
 165                 param.vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 166         }
 167
 168         if (!isinf(global_flags.x264_crf)) {
 169                 param.rc.i_rc_method = X264_RC_CRF;
 170                 param.rc.f_rf_constant = global_flags.x264_crf;
 171         } else {
 172                 param.rc.i_rc_method = X264_RC_ABR;
 173                 param.rc.i_bitrate = global_flags.x264_bitrate;
 174         }
 175         update_vbv_settings(&param);
 176         if (param.rc.i_vbv_max_bitrate > 0) {
 177                 // If the user wants VBV control to cap the max rate, it is
 178                 // also reasonable to assume that they are fine with the stream
 179                 // constantly being around that rate even for very low-complexity
 180                 // content; the obvious and extreme example being a static
 181                 // black picture.
 182                 //
 183                 // One would think it's fine to have low-complexity content use
 184                 // less bitrate, but it seems to cause problems in practice;
 185                 // e.g. VLC seems to often drop the stream (similar to a buffer
 186                 // underrun) in such cases, but only when streaming from Nageru,
 187                 // not when reading a dump of the same stream from disk.
 188                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 189                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 190                 // but it's a typical case of problems that can arise. Similarly,
 191                 // TCP's congestion control is not always fond of the rate staying
 192                 // low for a while and then rising quickly -- a variation on the same
 193                 // problem.
 194                 //
 195                 // We solve this by simply asking x264 to fill in dummy bits
 196                 // in these cases, so that the bitrate stays reasonable constant.
 197                 // It's a waste of bandwidth, but it makes things go much more
 198                 // smoothly in these cases. (We don't do it if VBV control is off
 199                 // in general, not the least because it makes no sense and x264
 200                 // thus ignores the parameter.)
 201                 param.rc.b_filler = 1;
 202         }
 203
 204         // Occasionally players have problem with extremely low quantizers;
 205         // be on the safe side. Shouldn't affect quality in any meaningful way.
 206         param.rc.i_qp_min = 5;
 207
 208         for (const string &str : global_flags.x264_extra_param) {
 209                 const size_t pos = str.find(',');
 210                 if (pos == string::npos) {
 211                         if (dyn.x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 212                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 213                         }
 214                 } else {
 215                         const string key = str.substr(0, pos);
 216                         const string value = str.substr(pos + 1);
 217                         if (dyn.x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 218                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 219                                         key.c_str(), value.c_str());
 220                         }
 221                 }
 222         }
 223
 224         if (global_flags.x264_bit_depth > 8) {
 225                 dyn.x264_param_apply_profile(&param, "high10");
 226         } else {
 227                 dyn.x264_param_apply_profile(&param, "high");
 228         }
 229
 230         param.b_repeat_headers = !wants_global_headers;
 231
 232         x264 = dyn.x264_encoder_open(&param);
 233         if (x264 == nullptr) {
 234                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 235                 exit(1);
 236         }
 237
 238         if (global_flags.x264_speedcontrol) {
 239                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 240         }
 241
 242         if (wants_global_headers) {
 243                 x264_nal_t *nal;
 244                 int num_nal;
 245
 246                 dyn.x264_encoder_headers(x264, &nal, &num_nal);
 247
 248                 for (int i = 0; i < num_nal; ++i) {
 249                         if (nal[i].i_type == NAL_SEI) {
 250                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 251                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 252                         } else {
 253                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 254                         }
 255                 }
 256         }
 257 }
 258
 259 void X264Encoder::encoder_thread_func()
 260 {
 261         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 262                 perror("nice()");
 263                 // No exit; it's not fatal.
 264         }
 265         pthread_setname_np(pthread_self(), "x264_encode");
 266         init_x264();
 267         x264_init_done = true;
 268
 269         bool frames_left;
 270
 271         do {
 272                 QueuedFrame qf;
 273
 274                 // Wait for a queued frame, then dequeue it.
 275                 {
 276                         unique_lock<mutex> lock(mu);
 277                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 278                         if (!queued_frames.empty()) {
 279                                 qf = queued_frames.front();
 280                                 queued_frames.pop();
 281                         } else {
 282                                 qf.pts = -1;
 283                                 qf.duration = -1;
 284                                 qf.data = nullptr;
 285                         }
 286
 287                         metric_x264_queued_frames = queued_frames.size();
 288                         frames_left = !queued_frames.empty();
 289                 }
 290
 291                 encode_frame(qf);
 292
 293                 {
 294                         lock_guard<mutex> lock(mu);
 295                         free_frames.push(qf.data);
 296                 }
 297
 298                 // We should quit only if the should_quit flag is set _and_ we have nothing
 299                 // in either queue.
 300         } while (!should_quit || frames_left || dyn.x264_encoder_delayed_frames(x264) > 0);
 301
 302         dyn.x264_encoder_close(x264);
 303 }
 304
 305 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 306 {
 307         x264_nal_t *nal = nullptr;
 308         int num_nal = 0;
 309         x264_picture_t pic;
 310         x264_picture_t *input_pic = nullptr;
 311
 312         if (qf.data) {
 313                 dyn.x264_picture_init(&pic);
 314
 315                 pic.i_pts = qf.pts;
 316                 if (global_flags.x264_bit_depth > 8) {
 317                         pic.img.i_csp = X264_CSP_NV12 | X264_CSP_HIGH_DEPTH;
 318                         pic.img.i_plane = 2;
 319                         pic.img.plane[0] = qf.data;
 320                         pic.img.i_stride[0] = global_flags.width * sizeof(uint16_t);
 321                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height * sizeof(uint16_t);
 322                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint32_t);
 323                 } else {
 324                         pic.img.i_csp = X264_CSP_NV12;
 325                         pic.img.i_plane = 2;
 326                         pic.img.plane[0] = qf.data;
 327                         pic.img.i_stride[0] = global_flags.width;
 328                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height;
 329                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint16_t);
 330                 }
 331                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 332
 333                 input_pic = &pic;
 334
 335                 frames_being_encoded[qf.pts] = qf.received_ts;
 336         }
 337
 338         unsigned new_rate = new_bitrate_kbit.load();  // Can be 0 for no change.
 339         if (speed_control) {
 340                 speed_control->set_config_override_function(bind(&speed_control_override_func, new_rate, qf.ycbcr_coefficients, _1));
 341         } else {
 342                 x264_param_t param;
 343                 dyn.x264_encoder_parameters(x264, &param);
 344                 speed_control_override_func(new_rate, qf.ycbcr_coefficients, &param);
 345                 dyn.x264_encoder_reconfig(x264, &param);
 346         }
 347
 348         if (speed_control) {
 349                 float queue_fill_ratio;
 350                 {
 351                         lock_guard<mutex> lock(mu);
 352                         queue_fill_ratio = float(free_frames.size()) / X264_QUEUE_LENGTH;
 353                 }
 354                 speed_control->before_frame(queue_fill_ratio, 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, &x264_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 }
 422
 423 void X264Encoder::speed_control_override_func(unsigned bitrate_kbit, movit::YCbCrLumaCoefficients ycbcr_coefficients, x264_param_t *param)
 424 {
 425         if (bitrate_kbit != 0) {
 426                 param->rc.i_bitrate = bitrate_kbit;
 427                 update_vbv_settings(param);
 428         }
 429
 430         if (ycbcr_coefficients == YCBCR_REC_709) {
 431                 param->vui.i_colmatrix = 1;  // BT.709.
 432         } else {
 433                 assert(ycbcr_coefficients == YCBCR_REC_601);
 434                 param->vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 435         }
 436 }