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