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