git.sesse.net Git - nageru/blob - x264_encoder.cpp

   1 #include "x264_encoder.h"
   2
   3 #include <assert.h>
   4 #include <stdio.h>
   5 #include <stdlib.h>
   6 #include <string.h>
   7 #include <unistd.h>
   8 #include <x264.h>
   9 #include <cstdint>
  10
  11 #include "defs.h"
  12 #include "flags.h"
  13 #include "mux.h"
  14 #include "print_latency.h"
  15 #include "timebase.h"
  16 #include "x264_speed_control.h"
  17
  18 extern "C" {
  19 #include <libavcodec/avcodec.h>
  20 #include <libavformat/avformat.h>
  21 }
  22
  23 using namespace movit;
  24 using namespace std;
  25 using namespace std::chrono;
  26
  27 namespace {
  28
  29 void update_vbv_settings(x264_param_t *param)
  30 {
  31         if (global_flags.x264_vbv_buffer_size < 0) {
  32                 param->rc.i_vbv_buffer_size = param->rc.i_bitrate;  // One-second VBV.
  33         } else {
  34                 param->rc.i_vbv_buffer_size = global_flags.x264_vbv_buffer_size;
  35         }
  36         if (global_flags.x264_vbv_max_bitrate < 0) {
  37                 param->rc.i_vbv_max_bitrate = param->rc.i_bitrate;  // CBR.
  38         } else {
  39                 param->rc.i_vbv_max_bitrate = global_flags.x264_vbv_max_bitrate;
  40         }
  41 }
  42
  43 }  // namespace
  44
  45 X264Encoder::X264Encoder(AVOutputFormat *oformat)
  46         : wants_global_headers(oformat->flags & AVFMT_GLOBALHEADER)
  47 {
  48         frame_pool.reset(new uint8_t[global_flags.width * global_flags.height * 2 * X264_QUEUE_LENGTH]);
  49         for (unsigned i = 0; i < X264_QUEUE_LENGTH; ++i) {
  50                 free_frames.push(frame_pool.get() + i * (global_flags.width * global_flags.height * 2));
  51         }
  52         encoder_thread = thread(&X264Encoder::encoder_thread_func, this);
  53 }
  54
  55 X264Encoder::~X264Encoder()
  56 {
  57         should_quit = true;
  58         queued_frames_nonempty.notify_all();
  59         encoder_thread.join();
  60 }
  61
  62 void X264Encoder::add_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const uint8_t *data, const ReceivedTimestamps &received_ts)
  63 {
  64         assert(!should_quit);
  65
  66         QueuedFrame qf;
  67         qf.pts = pts;
  68         qf.duration = duration;
  69         qf.ycbcr_coefficients = ycbcr_coefficients;
  70         qf.received_ts = received_ts;
  71
  72         {
  73                 lock_guard<mutex> lock(mu);
  74                 if (free_frames.empty()) {
  75                         fprintf(stderr, "WARNING: x264 queue full, dropping frame with pts %ld\n", pts);
  76                         return;
  77                 }
  78
  79                 qf.data = free_frames.front();
  80                 free_frames.pop();
  81         }
  82
  83         memcpy(qf.data, data, global_flags.width * global_flags.height * 2);
  84
  85         {
  86                 lock_guard<mutex> lock(mu);
  87                 queued_frames.push(qf);
  88                 queued_frames_nonempty.notify_all();
  89         }
  90 }
  91
  92 void X264Encoder::init_x264()
  93 {
  94         x264_param_t param;
  95         x264_param_default_preset(&param, global_flags.x264_preset.c_str(), global_flags.x264_tune.c_str());
  96
  97         param.i_width = global_flags.width;
  98         param.i_height = global_flags.height;
  99         param.i_csp = X264_CSP_NV12;
 100         param.b_vfr_input = 1;
 101         param.i_timebase_num = 1;
 102         param.i_timebase_den = TIMEBASE;
 103         param.i_keyint_max = 50; // About one second.
 104         if (global_flags.x264_speedcontrol) {
 105                 param.i_frame_reference = 16;  // Because speedcontrol is never allowed to change this above what we set at start.
 106         }
 107
 108         // NOTE: These should be in sync with the ones in quicksync_encoder.cpp (sps_rbsp()).
 109         param.vui.i_vidformat = 5;  // Unspecified.
 110         param.vui.b_fullrange = 0;
 111         param.vui.i_colorprim = 1;  // BT.709.
 112         param.vui.i_transfer = 2;  // Unspecified (since we use sRGB).
 113         if (global_flags.ycbcr_rec709_coefficients) {
 114                 param.vui.i_colmatrix = 1;  // BT.709.
 115         } else {
 116                 param.vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 117         }
 118
 119         param.rc.i_rc_method = X264_RC_ABR;
 120         param.rc.i_bitrate = global_flags.x264_bitrate;
 121         update_vbv_settings(&param);
 122         if (param.rc.i_vbv_max_bitrate > 0) {
 123                 // If the user wants VBV control to cap the max rate, it is
 124                 // also reasonable to assume that they are fine with the stream
 125                 // constantly being around that rate even for very low-complexity
 126                 // content; the obvious and extreme example being a static
 127                 // black picture.
 128                 //
 129                 // One would think it's fine to have low-complexity content use
 130                 // less bitrate, but it seems to cause problems in practice;
 131                 // e.g. VLC seems to often drop the stream (similar to a buffer
 132                 // underrun) in such cases, but only when streaming from Nageru,
 133                 // not when reading a dump of the same stream from disk.
 134                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 135                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 136                 // but it's a typical case of problems that can arise. Similarly,
 137                 // TCP's congestion control is not always fond of the rate staying
 138                 // low for a while and then rising quickly -- a variation on the same
 139                 // problem.
 140                 //
 141                 // We solve this by simply asking x264 to fill in dummy bits
 142                 // in these cases, so that the bitrate stays reasonable constant.
 143                 // It's a waste of bandwidth, but it makes things go much more
 144                 // smoothly in these cases. (We don't do it if VBV control is off
 145                 // in general, not the least because it makes no sense and x264
 146                 // thus ignores the parameter.)
 147                 param.rc.b_filler = 1;
 148         }
 149
 150         // Occasionally players have problem with extremely low quantizers;
 151         // be on the safe side. Shouldn't affect quality in any meaningful way.
 152         param.rc.i_qp_min = 5;
 153
 154         for (const string &str : global_flags.x264_extra_param) {
 155                 const size_t pos = str.find(',');
 156                 if (pos == string::npos) {
 157                         if (x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 158                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 159                         }
 160                 } else {
 161                         const string key = str.substr(0, pos);
 162                         const string value = str.substr(pos + 1);
 163                         if (x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 164                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 165                                         key.c_str(), value.c_str());
 166                         }
 167                 }
 168         }
 169
 170         x264_param_apply_profile(&param, "high");
 171
 172         param.b_repeat_headers = !wants_global_headers;
 173
 174         x264 = x264_encoder_open(&param);
 175         if (x264 == nullptr) {
 176                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 177                 exit(1);
 178         }
 179
 180         if (global_flags.x264_speedcontrol) {
 181                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 182         }
 183
 184         if (wants_global_headers) {
 185                 x264_nal_t *nal;
 186                 int num_nal;
 187
 188                 x264_encoder_headers(x264, &nal, &num_nal);
 189
 190                 for (int i = 0; i < num_nal; ++i) {
 191                         if (nal[i].i_type == NAL_SEI) {
 192                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 193                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 194                         } else {
 195                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 196                         }
 197                 }
 198         }
 199 }
 200
 201 void X264Encoder::encoder_thread_func()
 202 {
 203         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 204                 perror("nice()");
 205                 // No exit; it's not fatal.
 206         }
 207         init_x264();
 208         x264_init_done = true;
 209
 210         bool frames_left;
 211
 212         do {
 213                 QueuedFrame qf;
 214
 215                 // Wait for a queued frame, then dequeue it.
 216                 {
 217                         unique_lock<mutex> lock(mu);
 218                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 219                         if (!queued_frames.empty()) {
 220                                 qf = queued_frames.front();
 221                                 queued_frames.pop();
 222                         } else {
 223                                 qf.pts = -1;
 224                                 qf.duration = -1;
 225                                 qf.data = nullptr;
 226                         }
 227
 228                         frames_left = !queued_frames.empty();
 229                 }
 230
 231                 encode_frame(qf);
 232
 233                 {
 234                         lock_guard<mutex> lock(mu);
 235                         free_frames.push(qf.data);
 236                 }
 237
 238                 // We should quit only if the should_quit flag is set _and_ we have nothing
 239                 // in either queue.
 240         } while (!should_quit || frames_left || x264_encoder_delayed_frames(x264) > 0);
 241
 242         x264_encoder_close(x264);
 243 }
 244
 245 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 246 {
 247         x264_nal_t *nal = nullptr;
 248         int num_nal = 0;
 249         x264_picture_t pic;
 250         x264_picture_t *input_pic = nullptr;
 251
 252         if (qf.data) {
 253                 x264_picture_init(&pic);
 254
 255                 pic.i_pts = qf.pts;
 256                 pic.img.i_csp = X264_CSP_NV12;
 257                 pic.img.i_plane = 2;
 258                 pic.img.plane[0] = qf.data;
 259                 pic.img.i_stride[0] = global_flags.width;
 260                 pic.img.plane[1] = qf.data + global_flags.width * global_flags.height;
 261                 pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint16_t);
 262                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 263
 264                 input_pic = &pic;
 265
 266                 frames_being_encoded[qf.pts] = qf.received_ts;
 267         }
 268
 269         // See if we have a new bitrate to change to.
 270         unsigned new_rate = new_bitrate_kbit.exchange(0);  // Read and clear.
 271         if (new_rate != 0) {
 272                 bitrate_override_func = [new_rate](x264_param_t *param) {
 273                         param->rc.i_bitrate = new_rate;
 274                         update_vbv_settings(param);
 275                 };
 276         }
 277
 278         auto ycbcr_coefficients_override_func = [qf](x264_param_t *param) {
 279                 if (qf.ycbcr_coefficients == YCBCR_REC_709) {
 280                         param->vui.i_colmatrix = 1;  // BT.709.
 281                 } else {
 282                         assert(qf.ycbcr_coefficients == YCBCR_REC_601);
 283                         param->vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 284                 }
 285         };
 286
 287         if (speed_control) {
 288                 speed_control->set_config_override_function([this, ycbcr_coefficients_override_func](x264_param_t *param) {
 289                         if (bitrate_override_func) {
 290                                 bitrate_override_func(param);
 291                         }
 292                         ycbcr_coefficients_override_func(param);
 293                 });
 294         } else {
 295                 x264_param_t param;
 296                 x264_encoder_parameters(x264, &param);
 297                 if (bitrate_override_func) {
 298                         bitrate_override_func(&param);
 299                 }
 300                 ycbcr_coefficients_override_func(&param);
 301                 x264_encoder_reconfig(x264, &param);
 302         }
 303
 304         if (speed_control) {
 305                 speed_control->before_frame(float(free_frames.size()) / X264_QUEUE_LENGTH, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE);
 306         }
 307         x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic);
 308         if (speed_control) {
 309                 speed_control->after_frame();
 310         }
 311
 312         if (num_nal == 0) return;
 313
 314         if (frames_being_encoded.count(pic.i_pts)) {
 315                 ReceivedTimestamps received_ts = frames_being_encoded[pic.i_pts];
 316                 frames_being_encoded.erase(pic.i_pts);
 317
 318                 static int frameno = 0;
 319                 print_latency("Current x264 latency (video inputs → network mux):",
 320                         received_ts, (pic.i_type == X264_TYPE_B || pic.i_type == X264_TYPE_BREF),
 321                         &frameno);
 322         } else {
 323                 assert(false);
 324         }
 325
 326         // We really need one AVPacket for the entire frame, it seems,
 327         // so combine it all.
 328         size_t num_bytes = buffered_sei.size();
 329         for (int i = 0; i < num_nal; ++i) {
 330                 num_bytes += nal[i].i_payload;
 331         }
 332
 333         unique_ptr<uint8_t[]> data(new uint8_t[num_bytes]);
 334         uint8_t *ptr = data.get();
 335
 336         if (!buffered_sei.empty()) {
 337                 memcpy(ptr, buffered_sei.data(), buffered_sei.size());
 338                 ptr += buffered_sei.size();
 339                 buffered_sei.clear();
 340         }
 341         for (int i = 0; i < num_nal; ++i) {
 342                 memcpy(ptr, nal[i].p_payload, nal[i].i_payload);
 343                 ptr += nal[i].i_payload;
 344         }
 345
 346         AVPacket pkt;
 347         memset(&pkt, 0, sizeof(pkt));
 348         pkt.buf = nullptr;
 349         pkt.data = data.get();
 350         pkt.size = num_bytes;
 351         pkt.stream_index = 0;
 352         if (pic.b_keyframe) {
 353                 pkt.flags = AV_PKT_FLAG_KEY;
 354         } else {
 355                 pkt.flags = 0;
 356         }
 357         pkt.duration = reinterpret_cast<intptr_t>(pic.opaque);
 358
 359         for (Mux *mux : muxes) {
 360                 mux->add_packet(pkt, pic.i_pts, pic.i_dts);
 361         }
 362 }