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

   1 #include <string.h>
   2 #include <unistd.h>
   3
   4 #include "defs.h"
   5 #include "flags.h"
   6 #include "mux.h"
   7 #include "timebase.h"
   8 #include "x264_encoder.h"
   9 #include "x264_speed_control.h"
  10
  11 extern "C" {
  12 #include <libavformat/avformat.h>
  13 }
  14
  15 using namespace std;
  16
  17 namespace {
  18
  19 void update_vbv_settings(x264_param_t *param)
  20 {
  21         if (global_flags.x264_vbv_buffer_size < 0) {
  22                 param->rc.i_vbv_buffer_size = param->rc.i_bitrate;  // One-second VBV.
  23         } else {
  24                 param->rc.i_vbv_buffer_size = global_flags.x264_vbv_buffer_size;
  25         }
  26         if (global_flags.x264_vbv_max_bitrate < 0) {
  27                 param->rc.i_vbv_max_bitrate = param->rc.i_bitrate;  // CBR.
  28         } else {
  29                 param->rc.i_vbv_max_bitrate = global_flags.x264_vbv_max_bitrate;
  30         }
  31 }
  32
  33 }  // namespace
  34
  35 X264Encoder::X264Encoder(AVOutputFormat *oformat)
  36         : wants_global_headers(oformat->flags & AVFMT_GLOBALHEADER)
  37 {
  38         frame_pool.reset(new uint8_t[WIDTH * HEIGHT * 2 * X264_QUEUE_LENGTH]);
  39         for (unsigned i = 0; i < X264_QUEUE_LENGTH; ++i) {
  40                 free_frames.push(frame_pool.get() + i * (WIDTH * HEIGHT * 2));
  41         }
  42         encoder_thread = thread(&X264Encoder::encoder_thread_func, this);
  43 }
  44
  45 X264Encoder::~X264Encoder()
  46 {
  47         should_quit = true;
  48         queued_frames_nonempty.notify_all();
  49         encoder_thread.join();
  50 }
  51
  52 void X264Encoder::add_frame(int64_t pts, int64_t duration, const uint8_t *data)
  53 {
  54         QueuedFrame qf;
  55         qf.pts = pts;
  56         qf.duration = duration;
  57
  58         {
  59                 lock_guard<mutex> lock(mu);
  60                 if (free_frames.empty()) {
  61                         fprintf(stderr, "WARNING: x264 queue full, dropping frame with pts %ld\n", pts);
  62                         return;
  63                 }
  64
  65                 qf.data = free_frames.front();
  66                 free_frames.pop();
  67         }
  68
  69         memcpy(qf.data, data, WIDTH * HEIGHT * 2);
  70
  71         {
  72                 lock_guard<mutex> lock(mu);
  73                 queued_frames.push(qf);
  74                 queued_frames_nonempty.notify_all();
  75         }
  76 }
  77
  78 void X264Encoder::init_x264()
  79 {
  80         x264_param_t param;
  81         x264_param_default_preset(&param, global_flags.x264_preset.c_str(), global_flags.x264_tune.c_str());
  82
  83         param.i_width = WIDTH;
  84         param.i_height = HEIGHT;
  85         param.i_csp = X264_CSP_NV12;
  86         param.b_vfr_input = 1;
  87         param.i_timebase_num = 1;
  88         param.i_timebase_den = TIMEBASE;
  89         param.i_keyint_max = 50; // About one second.
  90         if (global_flags.x264_speedcontrol) {
  91                 param.i_frame_reference = 16;  // Because speedcontrol is never allowed to change this above what we set at start.
  92         }
  93
  94         // NOTE: These should be in sync with the ones in h264encode.cpp (sbs_rbsp()).
  95         param.vui.i_vidformat = 5;  // Unspecified.
  96         param.vui.b_fullrange = 0;
  97         param.vui.i_colorprim = 1;  // BT.709.
  98         param.vui.i_transfer = 2;  // Unspecified (since we use sRGB).
  99         param.vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 100
 101
 102         param.rc.i_rc_method = X264_RC_ABR;
 103         param.rc.i_bitrate = global_flags.x264_bitrate;
 104         update_vbv_settings(&param);
 105         if (param.rc.i_vbv_max_bitrate > 0) {
 106                 // If the user wants VBV control to cap the max rate, it is
 107                 // also reasonable to assume that they are fine with the stream
 108                 // constantly being around that rate even for very low-complexity
 109                 // content; the obvious and extreme example being a static
 110                 // black picture.
 111                 //
 112                 // One would think it's fine to have low-complexity content use
 113                 // less bitrate, but it seems to cause problems in practice;
 114                 // e.g. VLC seems to often drop the stream (similar to a buffer
 115                 // underrun) in such cases, but only when streaming from Nageru,
 116                 // not when reading a dump of the same stream from disk.
 117                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 118                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 119                 // but it's a typical case of problems that can arise. Similarly,
 120                 // TCP's congestion control is not always fond of the rate staying
 121                 // low for a while and then rising quickly -- a variation on the same
 122                 // problem.
 123                 //
 124                 // We solve this by simply asking x264 to fill in dummy bits
 125                 // in these cases, so that the bitrate stays reasonable constant.
 126                 // It's a waste of bandwidth, but it makes things go much more
 127                 // smoothly in these cases. (We don't do it if VBV control is off
 128                 // in general, not the least because it makes no sense and x264
 129                 // thus ignores the parameter.)
 130                 param.rc.b_filler = 1;
 131         }
 132
 133         // Occasionally players have problem with extremely low quantizers;
 134         // be on the safe side. Shouldn't affect quality in any meaningful way.
 135         param.rc.i_qp_min = 5;
 136
 137         for (const string &str : global_flags.x264_extra_param) {
 138                 const size_t pos = str.find(',');
 139                 if (pos == string::npos) {
 140                         if (x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 141                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 142                         }
 143                 } else {
 144                         const string key = str.substr(0, pos);
 145                         const string value = str.substr(pos + 1);
 146                         if (x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 147                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 148                                         key.c_str(), value.c_str());
 149                         }
 150                 }
 151         }
 152
 153         x264_param_apply_profile(&param, "high");
 154
 155         param.b_repeat_headers = !wants_global_headers;
 156
 157         x264 = x264_encoder_open(&param);
 158         if (x264 == nullptr) {
 159                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 160                 exit(1);
 161         }
 162
 163         if (global_flags.x264_speedcontrol) {
 164                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 165         }
 166
 167         if (wants_global_headers) {
 168                 x264_nal_t *nal;
 169                 int num_nal;
 170
 171                 x264_encoder_headers(x264, &nal, &num_nal);
 172
 173                 for (int i = 0; i < num_nal; ++i) {
 174                         if (nal[i].i_type == NAL_SEI) {
 175                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 176                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 177                         } else {
 178                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 179                         }
 180                 }
 181         }
 182 }
 183
 184 void X264Encoder::encoder_thread_func()
 185 {
 186         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 187                 perror("nice()");
 188                 // No exit; it's not fatal.
 189         }
 190         init_x264();
 191
 192         bool frames_left;
 193
 194         do {
 195                 QueuedFrame qf;
 196
 197                 // Wait for a queued frame, then dequeue it.
 198                 {
 199                         unique_lock<mutex> lock(mu);
 200                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 201                         if (!queued_frames.empty()) {
 202                                 qf = queued_frames.front();
 203                                 queued_frames.pop();
 204                         } else {
 205                                 qf.pts = -1;
 206                                 qf.duration = -1;
 207                                 qf.data = nullptr;
 208                         }
 209
 210                         frames_left = !queued_frames.empty();
 211                 }
 212
 213                 encode_frame(qf);
 214
 215                 {
 216                         lock_guard<mutex> lock(mu);
 217                         free_frames.push(qf.data);
 218                 }
 219
 220                 // We should quit only if the should_quit flag is set _and_ we have nothing
 221                 // in either queue.
 222         } while (!should_quit || frames_left || x264_encoder_delayed_frames(x264) > 0);
 223
 224         x264_encoder_close(x264);
 225 }
 226
 227 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 228 {
 229         x264_nal_t *nal = nullptr;
 230         int num_nal = 0;
 231         x264_picture_t pic;
 232         x264_picture_t *input_pic = nullptr;
 233
 234         if (qf.data) {
 235                 x264_picture_init(&pic);
 236
 237                 pic.i_pts = qf.pts;
 238                 pic.img.i_csp = X264_CSP_NV12;
 239                 pic.img.i_plane = 2;
 240                 pic.img.plane[0] = qf.data;
 241                 pic.img.i_stride[0] = WIDTH;
 242                 pic.img.plane[1] = qf.data + WIDTH * HEIGHT;
 243                 pic.img.i_stride[1] = WIDTH / 2 * sizeof(uint16_t);
 244                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 245
 246                 input_pic = &pic;
 247         }
 248
 249         // See if we have a new bitrate to change to.
 250         unsigned new_rate = new_bitrate_kbit.exchange(0);  // Read and clear.
 251         if (new_rate != 0) {
 252                 if (speed_control) {
 253                         speed_control->set_config_override_function([new_rate](x264_param_t *param) {
 254                                 param->rc.i_bitrate = new_rate;
 255                                 update_vbv_settings(param);
 256                         });
 257                 } else {
 258                         x264_param_t param;
 259                         x264_encoder_parameters(x264, &param);
 260                         param.rc.i_bitrate = new_rate;
 261                         update_vbv_settings(&param);
 262                         x264_encoder_reconfig(x264, &param);
 263                 }
 264         }
 265
 266         if (speed_control) {
 267                 speed_control->before_frame(float(free_frames.size()) / X264_QUEUE_LENGTH, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE);
 268         }
 269         x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic);
 270         if (speed_control) {
 271                 speed_control->after_frame();
 272         }
 273
 274         // We really need one AVPacket for the entire frame, it seems,
 275         // so combine it all.
 276         size_t num_bytes = buffered_sei.size();
 277         for (int i = 0; i < num_nal; ++i) {
 278                 num_bytes += nal[i].i_payload;
 279         }
 280
 281         unique_ptr<uint8_t[]> data(new uint8_t[num_bytes]);
 282         uint8_t *ptr = data.get();
 283
 284         if (!buffered_sei.empty()) {
 285                 memcpy(ptr, buffered_sei.data(), buffered_sei.size());
 286                 ptr += buffered_sei.size();
 287                 buffered_sei.clear();
 288         }
 289         for (int i = 0; i < num_nal; ++i) {
 290                 memcpy(ptr, nal[i].p_payload, nal[i].i_payload);
 291                 ptr += nal[i].i_payload;
 292         }
 293
 294         AVPacket pkt;
 295         memset(&pkt, 0, sizeof(pkt));
 296         pkt.buf = nullptr;
 297         pkt.data = data.get();
 298         pkt.size = num_bytes;
 299         pkt.stream_index = 0;
 300         if (pic.b_keyframe) {
 301                 pkt.flags = AV_PKT_FLAG_KEY;
 302         } else {
 303                 pkt.flags = 0;
 304         }
 305         pkt.duration = reinterpret_cast<intptr_t>(pic.opaque);
 306
 307         mux->add_packet(pkt, pic.i_pts, pic.i_dts);
 308 }