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