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