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

   1 #undef Success
   2
   3 #include "mixer.h"
   4
   5 #include <assert.h>
   6 #include <epoxy/egl.h>
   7 #include <movit/effect_chain.h>
   8 #include <movit/effect_util.h>
   9 #include <movit/flat_input.h>
  10 #include <movit/image_format.h>
  11 #include <movit/init.h>
  12 #include <movit/resource_pool.h>
  13 #include <pthread.h>
  14 #include <stdint.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <algorithm>
  18 #include <chrono>
  19 #include <condition_variable>
  20 #include <cstddef>
  21 #include <cstdint>
  22 #include <memory>
  23 #include <mutex>
  24 #include <ratio>
  25 #include <string>
  26 #include <thread>
  27 #include <utility>
  28 #include <vector>
  29
  30 #include "DeckLinkAPI.h"
  31 #include "LinuxCOM.h"
  32 #include "alsa_output.h"
  33 #include "basic_stats.h"
  34 #include "bmusb/bmusb.h"
  35 #include "bmusb/fake_capture.h"
  36 #ifdef HAVE_CEF
  37 #include "cef_capture.h"
  38 #endif
  39 #include "chroma_subsampler.h"
  40 #include "shared/context.h"
  41 #include "decklink_capture.h"
  42 #include "decklink_output.h"
  43 #include "defs.h"
  44 #include "shared/disk_space_estimator.h"
  45 #include "ffmpeg_capture.h"
  46 #include "flags.h"
  47 #include "input_mapping.h"
  48 #include "shared/metrics.h"
  49 #include "mjpeg_encoder.h"
  50 #include "pbo_frame_allocator.h"
  51 #include "shared/ref_counted_gl_sync.h"
  52 #include "resampling_queue.h"
  53 #include "shared/timebase.h"
  54 #include "timecode_renderer.h"
  55 #include "v210_converter.h"
  56 #include "va_display_with_cleanup.h"
  57 #include "video_encoder.h"
  58
  59 #undef Status
  60 #include <google/protobuf/util/json_util.h>
  61 #include "json.pb.h"
  62
  63 class IDeckLink;
  64 class QOpenGLContext;
  65
  66 using namespace movit;
  67 using namespace std;
  68 using namespace std::chrono;
  69 using namespace std::placeholders;
  70 using namespace bmusb;
  71
  72 Mixer *global_mixer = nullptr;
  73
  74 namespace {
  75
  76 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  77 {
  78         if (interlaced) {
  79                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  80                         input_state->buffered_frames[card_index][frame_num] =
  81                                 input_state->buffered_frames[card_index][frame_num - 1];
  82                 }
  83                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  84         } else {
  85                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  86                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  87                 }
  88         }
  89 }
  90
  91 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned cbcr_width, unsigned cbcr_height, unsigned v210_width)
  92 {
  93         bool first;
  94         switch (userdata->pixel_format) {
  95         case PixelFormat_10BitYCbCr:
  96                 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
  97                 break;
  98         case PixelFormat_8BitYCbCr:
  99                 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
 100                 break;
 101         case PixelFormat_8BitBGRA:
 102                 first = userdata->tex_rgba[field] == 0;
 103                 break;
 104         case PixelFormat_8BitYCbCrPlanar:
 105                 first = userdata->tex_y[field] == 0 || userdata->tex_cb[field] == 0 || userdata->tex_cr[field] == 0;
 106                 break;
 107         default:
 108                 assert(false);
 109         }
 110
 111         if (first ||
 112             width != userdata->last_width[field] ||
 113             height != userdata->last_height[field] ||
 114             cbcr_width != userdata->last_cbcr_width[field] ||
 115             cbcr_height != userdata->last_cbcr_height[field]) {
 116                 // We changed resolution since last use of this texture, so we need to create
 117                 // a new object. Note that this each card has its own PBOFrameAllocator,
 118                 // we don't need to worry about these flip-flopping between resolutions.
 119                 switch (userdata->pixel_format) {
 120                 case PixelFormat_10BitYCbCr:
 121                         glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
 122                         check_error();
 123                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 124                         check_error();
 125                         break;
 126                 case PixelFormat_8BitYCbCr: {
 127                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 128                         check_error();
 129                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 130                         check_error();
 131                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 132                         check_error();
 133                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 134                         check_error();
 135                         break;
 136                 }
 137                 case PixelFormat_8BitYCbCrPlanar: {
 138                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 139                         check_error();
 140                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 141                         check_error();
 142                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cb[field]);
 143                         check_error();
 144                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 145                         check_error();
 146                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cr[field]);
 147                         check_error();
 148                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 149                         check_error();
 150                         break;
 151                 }
 152                 case PixelFormat_8BitBGRA:
 153                         glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 154                         check_error();
 155                         if (global_flags.can_disable_srgb_decoder) {  // See the comments in tweaked_inputs.h.
 156                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 157                         } else {
 158                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 159                         }
 160                         check_error();
 161                         break;
 162                 default:
 163                         assert(false);
 164                 }
 165                 userdata->last_width[field] = width;
 166                 userdata->last_height[field] = height;
 167                 userdata->last_cbcr_width[field] = cbcr_width;
 168                 userdata->last_cbcr_height[field] = cbcr_height;
 169         }
 170         if (global_flags.ten_bit_input &&
 171             (first || v210_width != userdata->last_v210_width[field])) {
 172                 // Same as above; we need to recreate the texture.
 173                 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
 174                 check_error();
 175                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 176                 check_error();
 177                 userdata->last_v210_width[field] = v210_width;
 178         }
 179 }
 180
 181 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
 182 {
 183         if (interlaced_stride) {
 184                 stride *= 2;
 185         }
 186         if (global_flags.flush_pbos) {
 187                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
 188                 check_error();
 189         }
 190
 191         glBindTexture(GL_TEXTURE_2D, tex);
 192         check_error();
 193         if (interlaced_stride) {
 194                 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
 195                 check_error();
 196         } else {
 197                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 198                 check_error();
 199         }
 200
 201         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
 202         check_error();
 203         glBindTexture(GL_TEXTURE_2D, 0);
 204         check_error();
 205         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 206         check_error();
 207 }
 208
 209 }  // namespace
 210
 211 void JitterHistory::register_metrics(const vector<pair<string, string>> &labels)
 212 {
 213         global_metrics.add("input_underestimated_jitter_frames", labels, &metric_input_underestimated_jitter_frames);
 214         global_metrics.add("input_estimated_max_jitter_seconds", labels, &metric_input_estimated_max_jitter_seconds, Metrics::TYPE_GAUGE);
 215 }
 216
 217 void JitterHistory::unregister_metrics(const vector<pair<string, string>> &labels)
 218 {
 219         global_metrics.remove("input_underestimated_jitter_frames", labels);
 220         global_metrics.remove("input_estimated_max_jitter_seconds", labels);
 221 }
 222
 223 void JitterHistory::frame_arrived(steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames)
 224 {
 225         if (expected_timestamp > steady_clock::time_point::min()) {
 226                 expected_timestamp += dropped_frames * nanoseconds(frame_duration * 1000000000 / TIMEBASE);
 227                 double jitter_seconds = fabs(duration<double>(expected_timestamp - now).count());
 228                 history.push_back(orders.insert(jitter_seconds));
 229                 if (jitter_seconds > estimate_max_jitter()) {
 230                         ++metric_input_underestimated_jitter_frames;
 231                 }
 232
 233                 metric_input_estimated_max_jitter_seconds = estimate_max_jitter();
 234
 235                 if (history.size() > history_length) {
 236                         orders.erase(history.front());
 237                         history.pop_front();
 238                 }
 239                 assert(history.size() <= history_length);
 240         }
 241         expected_timestamp = now + nanoseconds(frame_duration * 1000000000 / TIMEBASE);
 242 }
 243
 244 double JitterHistory::estimate_max_jitter() const
 245 {
 246         if (orders.empty()) {
 247                 return 0.0;
 248         }
 249         size_t elem_idx = lrint((orders.size() - 1) * percentile);
 250         if (percentile <= 0.5) {
 251                 return *next(orders.begin(), elem_idx) * multiplier;
 252         } else {
 253                 return *prev(orders.end(), orders.size() - elem_idx) * multiplier;
 254         }
 255 }
 256
 257 void QueueLengthPolicy::register_metrics(const vector<pair<string, string>> &labels)
 258 {
 259         global_metrics.add("input_queue_safe_length_frames", labels, &metric_input_queue_safe_length_frames, Metrics::TYPE_GAUGE);
 260 }
 261
 262 void QueueLengthPolicy::unregister_metrics(const vector<pair<string, string>> &labels)
 263 {
 264         global_metrics.remove("input_queue_safe_length_frames", labels);
 265 }
 266
 267 void QueueLengthPolicy::update_policy(steady_clock::time_point now,
 268                                       steady_clock::time_point expected_next_frame,
 269                                       int64_t input_frame_duration,
 270                                       int64_t master_frame_duration,
 271                                       double max_input_card_jitter_seconds,
 272                                       double max_master_card_jitter_seconds)
 273 {
 274         double input_frame_duration_seconds = input_frame_duration / double(TIMEBASE);
 275         double master_frame_duration_seconds = master_frame_duration / double(TIMEBASE);
 276
 277         // Figure out when we can expect the next frame for this card, assuming
 278         // worst-case jitter (ie., the frame is maximally late).
 279         double seconds_until_next_frame = max(duration<double>(expected_next_frame - now).count() + max_input_card_jitter_seconds, 0.0);
 280
 281         // How many times are the master card expected to tick in that time?
 282         // We assume the master clock has worst-case jitter but not any rate
 283         // discrepancy, ie., it ticks as early as possible every time, but not
 284         // cumulatively.
 285         double frames_needed = (seconds_until_next_frame + max_master_card_jitter_seconds) / master_frame_duration_seconds;
 286
 287         // As a special case, if the master card ticks faster than the input card,
 288         // we expect the queue to drain by itself even without dropping. But if
 289         // the difference is small (e.g. 60 Hz master and 59.94 input), it would
 290         // go slowly enough that the effect wouldn't really be appreciable.
 291         // We account for this by looking at the situation five frames ahead,
 292         // assuming everything else is the same.
 293         double frames_allowed;
 294         if (master_frame_duration < input_frame_duration) {
 295                 frames_allowed = frames_needed + 5 * (input_frame_duration_seconds - master_frame_duration_seconds) / master_frame_duration_seconds;
 296         } else {
 297                 frames_allowed = frames_needed;
 298         }
 299
 300         safe_queue_length = max<int>(floor(frames_allowed), 0);
 301         metric_input_queue_safe_length_frames = safe_queue_length;
 302 }
 303
 304 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 305         : httpd(),
 306           num_cards(num_cards),
 307           mixer_surface(create_surface(format)),
 308           h264_encoder_surface(create_surface(format)),
 309           decklink_output_surface(create_surface(format))
 310 {
 311         memcpy(ycbcr_interpretation, global_flags.ycbcr_interpretation, sizeof(ycbcr_interpretation));
 312         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 313         check_error();
 314
 315         // This nearly always should be true.
 316         global_flags.can_disable_srgb_decoder =
 317                 epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") &&
 318                 epoxy_has_gl_extension("GL_ARB_sampler_objects");
 319
 320         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 321         // will be halved when sampling them, and we need to compensate here.
 322         movit_texel_subpixel_precision /= 2.0;
 323
 324         resource_pool.reset(new ResourcePool);
 325         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 326                 output_channel[i].parent = this;
 327                 output_channel[i].channel = i;
 328         }
 329
 330         ImageFormat inout_format;
 331         inout_format.color_space = COLORSPACE_sRGB;
 332         inout_format.gamma_curve = GAMMA_sRGB;
 333
 334         // Matches the 4:2:0 format created by the main chain.
 335         YCbCrFormat ycbcr_format;
 336         ycbcr_format.chroma_subsampling_x = 2;
 337         ycbcr_format.chroma_subsampling_y = 2;
 338         if (global_flags.ycbcr_rec709_coefficients) {
 339                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 340         } else {
 341                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 342         }
 343         ycbcr_format.full_range = false;
 344         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 345         ycbcr_format.cb_x_position = 0.0f;
 346         ycbcr_format.cr_x_position = 0.0f;
 347         ycbcr_format.cb_y_position = 0.5f;
 348         ycbcr_format.cr_y_position = 0.5f;
 349
 350         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 351         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 352         check_error();
 353         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 354         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 355         display_chain->add_input(display_input);
 356         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 357         display_chain->set_dither_bits(0);  // Don't bother.
 358         display_chain->finalize();
 359
 360         video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, global_flags.width, global_flags.height, &httpd, global_disk_space_estimator));
 361         if (!global_flags.card_to_mjpeg_stream_export.empty()) {
 362                 mjpeg_encoder.reset(new MJPEGEncoder(&httpd, global_flags.va_display));
 363         }
 364
 365         // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
 366         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 367
 368         // Must be instantiated after the theme, as the theme decides the number of FFmpeg inputs.
 369         std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
 370         audio_mixer.reset(new AudioMixer(num_cards, video_inputs.size()));
 371
 372         httpd.add_endpoint("/channels", bind(&Mixer::get_channels_json, this), HTTPD::ALLOW_ALL_ORIGINS);
 373         for (int channel_idx = 2; channel_idx < theme->get_num_channels(); ++channel_idx) {
 374                 char url[256];
 375                 snprintf(url, sizeof(url), "/channels/%d/color", channel_idx);
 376                 httpd.add_endpoint(url, bind(&Mixer::get_channel_color_http, this, unsigned(channel_idx)), HTTPD::ALLOW_ALL_ORIGINS);
 377         }
 378
 379         // Start listening for clients only once VideoEncoder has written its header, if any.
 380         httpd.start(global_flags.http_port);
 381
 382         // First try initializing the then PCI devices, then USB, then
 383         // fill up with fake cards until we have the desired number of cards.
 384         unsigned num_pci_devices = 0;
 385         unsigned card_index = 0;
 386
 387         {
 388                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 389                 if (decklink_iterator != nullptr) {
 390                         for ( ; card_index < num_cards; ++card_index) {
 391                                 IDeckLink *decklink;
 392                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 393                                         break;
 394                                 }
 395
 396                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 397                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 398                                 if (!output->set_device(decklink)) {
 399                                         delete output;
 400                                         output = nullptr;
 401                                 }
 402                                 configure_card(card_index, capture, CardType::LIVE_CARD, output);
 403                                 ++num_pci_devices;
 404                         }
 405                         decklink_iterator->Release();
 406                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 407                 } else {
 408                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 409                 }
 410         }
 411
 412         unsigned num_usb_devices = BMUSBCapture::num_cards();
 413         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 414                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 415                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 416                 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
 417         }
 418         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 419
 420         unsigned num_fake_cards = 0;
 421         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 422                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 423                 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
 424         }
 425
 426         if (num_fake_cards > 0) {
 427                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 428         }
 429
 430         // Initialize all video inputs the theme asked for. Note that these are
 431         // all put _after_ the regular cards, which stop at <num_cards> - 1.
 432         for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
 433                 if (card_index >= MAX_VIDEO_CARDS) {
 434                         fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
 435                         abort();
 436                 }
 437                 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
 438                 video_inputs[video_card_index]->set_card_index(card_index);
 439         }
 440         num_video_inputs = video_inputs.size();
 441
 442 #ifdef HAVE_CEF
 443         // Same, for HTML inputs.
 444         std::vector<CEFCapture *> html_inputs = theme->get_html_inputs();
 445         for (unsigned html_card_index = 0; html_card_index < html_inputs.size(); ++card_index, ++html_card_index) {
 446                 if (card_index >= MAX_VIDEO_CARDS) {
 447                         fprintf(stderr, "ERROR: Not enough card slots available for the HTML inputs the theme requested.\n");
 448                         abort();
 449                 }
 450                 configure_card(card_index, html_inputs[html_card_index], CardType::CEF_INPUT, /*output=*/nullptr);
 451                 html_inputs[html_card_index]->set_card_index(card_index);
 452         }
 453         num_html_inputs = html_inputs.size();
 454 #endif
 455
 456         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 457         BMUSBCapture::start_bm_thread();
 458
 459         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
 460                 cards[card_index].queue_length_policy.reset(card_index);
 461         }
 462
 463         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 464
 465         if (global_flags.ten_bit_input) {
 466                 if (!v210Converter::has_hardware_support()) {
 467                         fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
 468                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 469                         abort();
 470                 }
 471                 v210_converter.reset(new v210Converter());
 472
 473                 // These are all the widths listed in the Blackmagic SDK documentation
 474                 // (section 2.7.3, “Display Modes”).
 475                 v210_converter->precompile_shader(720);
 476                 v210_converter->precompile_shader(1280);
 477                 v210_converter->precompile_shader(1920);
 478                 v210_converter->precompile_shader(2048);
 479                 v210_converter->precompile_shader(3840);
 480                 v210_converter->precompile_shader(4096);
 481         }
 482         if (global_flags.ten_bit_output) {
 483                 if (!v210Converter::has_hardware_support()) {
 484                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 485                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 486                         abort();
 487                 }
 488         }
 489
 490         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 491         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 492         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 493
 494         if (global_flags.enable_alsa_output) {
 495                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 496         }
 497         if (global_flags.output_card != -1) {
 498                 desired_output_card_index = global_flags.output_card;
 499                 set_output_card_internal(global_flags.output_card);
 500         }
 501
 502         output_jitter_history.register_metrics({{ "card", "output" }});
 503 }
 504
 505 Mixer::~Mixer()
 506 {
 507         if (mjpeg_encoder != nullptr) {
 508                 mjpeg_encoder->stop();
 509         }
 510         httpd.stop();
 511         BMUSBCapture::stop_bm_thread();
 512
 513         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
 514                 cards[card_index].capture->stop_dequeue_thread();
 515                 if (cards[card_index].output) {
 516                         cards[card_index].output->end_output();
 517                         cards[card_index].output.reset();
 518                 }
 519         }
 520
 521         video_encoder.reset(nullptr);
 522 }
 523
 524 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
 525 {
 526         printf("Configuring card %d...\n", card_index);
 527
 528         CaptureCard *card = &cards[card_index];
 529         if (card->capture != nullptr) {
 530                 card->capture->stop_dequeue_thread();
 531         }
 532         card->capture.reset(capture);
 533         card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
 534         card->is_cef_capture = (card_type == CardType::CEF_INPUT);
 535         card->may_have_dropped_last_frame = false;
 536         card->type = card_type;
 537         if (card->output.get() != output) {
 538                 card->output.reset(output);
 539         }
 540
 541         PixelFormat pixel_format;
 542         if (card_type == CardType::FFMPEG_INPUT) {
 543                 pixel_format = capture->get_current_pixel_format();
 544         } else if (card_type == CardType::CEF_INPUT) {
 545                 pixel_format = PixelFormat_8BitBGRA;
 546         } else if (global_flags.ten_bit_input) {
 547                 pixel_format = PixelFormat_10BitYCbCr;
 548         } else {
 549                 pixel_format = PixelFormat_8BitYCbCr;
 550         }
 551
 552         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 553         if (card->frame_allocator == nullptr) {
 554                 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height, card_index, mjpeg_encoder.get()));  // 8 MB.
 555         }
 556         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 557         if (card->surface == nullptr) {
 558                 card->surface = create_surface_with_same_format(mixer_surface);
 559         }
 560         while (!card->new_frames.empty()) card->new_frames.pop_front();
 561         card->last_timecode = -1;
 562         card->capture->set_pixel_format(pixel_format);
 563         card->capture->configure_card();
 564
 565         // NOTE: start_bm_capture() happens in thread_func().
 566
 567         DeviceSpec device;
 568         if (card_type == CardType::FFMPEG_INPUT) {
 569                 device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_cards};
 570         } else {
 571                 device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
 572         }
 573         audio_mixer->reset_resampler(device);
 574         audio_mixer->set_display_name(device, card->capture->get_description());
 575         audio_mixer->trigger_state_changed_callback();
 576
 577         // Unregister old metrics, if any.
 578         if (!card->labels.empty()) {
 579                 const vector<pair<string, string>> &labels = card->labels;
 580                 card->jitter_history.unregister_metrics(labels);
 581                 card->queue_length_policy.unregister_metrics(labels);
 582                 global_metrics.remove("input_received_frames", labels);
 583                 global_metrics.remove("input_dropped_frames_jitter", labels);
 584                 global_metrics.remove("input_dropped_frames_error", labels);
 585                 global_metrics.remove("input_dropped_frames_resets", labels);
 586                 global_metrics.remove("input_queue_length_frames", labels);
 587                 global_metrics.remove("input_queue_duped_frames", labels);
 588
 589                 global_metrics.remove("input_has_signal_bool", labels);
 590                 global_metrics.remove("input_is_connected_bool", labels);
 591                 global_metrics.remove("input_interlaced_bool", labels);
 592                 global_metrics.remove("input_width_pixels", labels);
 593                 global_metrics.remove("input_height_pixels", labels);
 594                 global_metrics.remove("input_frame_rate_nom", labels);
 595                 global_metrics.remove("input_frame_rate_den", labels);
 596                 global_metrics.remove("input_sample_rate_hz", labels);
 597         }
 598
 599         // Register metrics.
 600         vector<pair<string, string>> labels;
 601         char card_name[64];
 602         snprintf(card_name, sizeof(card_name), "%d", card_index);
 603         labels.emplace_back("card", card_name);
 604
 605         switch (card_type) {
 606         case CardType::LIVE_CARD:
 607                 labels.emplace_back("cardtype", "live");
 608                 break;
 609         case CardType::FAKE_CAPTURE:
 610                 labels.emplace_back("cardtype", "fake");
 611                 break;
 612         case CardType::FFMPEG_INPUT:
 613                 labels.emplace_back("cardtype", "ffmpeg");
 614                 break;
 615         case CardType::CEF_INPUT:
 616                 labels.emplace_back("cardtype", "cef");
 617                 break;
 618         default:
 619                 assert(false);
 620         }
 621         card->jitter_history.register_metrics(labels);
 622         card->queue_length_policy.register_metrics(labels);
 623         global_metrics.add("input_received_frames", labels, &card->metric_input_received_frames);
 624         global_metrics.add("input_dropped_frames_jitter", labels, &card->metric_input_dropped_frames_jitter);
 625         global_metrics.add("input_dropped_frames_error", labels, &card->metric_input_dropped_frames_error);
 626         global_metrics.add("input_dropped_frames_resets", labels, &card->metric_input_resets);
 627         global_metrics.add("input_queue_length_frames", labels, &card->metric_input_queue_length_frames, Metrics::TYPE_GAUGE);
 628         global_metrics.add("input_queue_duped_frames", labels, &card->metric_input_duped_frames);
 629
 630         global_metrics.add("input_has_signal_bool", labels, &card->metric_input_has_signal_bool, Metrics::TYPE_GAUGE);
 631         global_metrics.add("input_is_connected_bool", labels, &card->metric_input_is_connected_bool, Metrics::TYPE_GAUGE);
 632         global_metrics.add("input_interlaced_bool", labels, &card->metric_input_interlaced_bool, Metrics::TYPE_GAUGE);
 633         global_metrics.add("input_width_pixels", labels, &card->metric_input_width_pixels, Metrics::TYPE_GAUGE);
 634         global_metrics.add("input_height_pixels", labels, &card->metric_input_height_pixels, Metrics::TYPE_GAUGE);
 635         global_metrics.add("input_frame_rate_nom", labels, &card->metric_input_frame_rate_nom, Metrics::TYPE_GAUGE);
 636         global_metrics.add("input_frame_rate_den", labels, &card->metric_input_frame_rate_den, Metrics::TYPE_GAUGE);
 637         global_metrics.add("input_sample_rate_hz", labels, &card->metric_input_sample_rate_hz, Metrics::TYPE_GAUGE);
 638         card->labels = labels;
 639 }
 640
 641 void Mixer::set_output_card_internal(int card_index)
 642 {
 643         // We don't really need to take card_mutex, since we're in the mixer
 644         // thread and don't mess with any queues (which is the only thing that happens
 645         // from other threads), but it's probably the safest in the long run.
 646         unique_lock<mutex> lock(card_mutex);
 647         if (output_card_index != -1) {
 648                 // Switch the old card from output to input.
 649                 CaptureCard *old_card = &cards[output_card_index];
 650                 old_card->output->end_output();
 651
 652                 // Stop the fake card that we put into place.
 653                 // This needs to _not_ happen under the mutex, to avoid deadlock
 654                 // (delivering the last frame needs to take the mutex).
 655                 CaptureInterface *fake_capture = old_card->capture.get();
 656                 lock.unlock();
 657                 fake_capture->stop_dequeue_thread();
 658                 lock.lock();
 659                 old_card->capture = move(old_card->parked_capture);  // TODO: reset the metrics
 660                 old_card->is_fake_capture = false;
 661                 old_card->capture->start_bm_capture();
 662         }
 663         if (card_index != -1) {
 664                 CaptureCard *card = &cards[card_index];
 665                 CaptureInterface *capture = card->capture.get();
 666                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 667                 // several seconds to complete (blocking on DisableVideoInput);
 668                 // see if we can maybe do it asynchronously.
 669                 lock.unlock();
 670                 capture->stop_dequeue_thread();
 671                 lock.lock();
 672                 card->parked_capture = move(card->capture);
 673                 CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 674                 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
 675                 card->queue_length_policy.reset(card_index);
 676                 card->capture->start_bm_capture();
 677                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 678                 card->output->start_output(desired_output_video_mode, pts_int);
 679         }
 680         output_card_index = card_index;
 681         output_jitter_history.clear();
 682 }
 683
 684 namespace {
 685
 686 int unwrap_timecode(uint16_t current_wrapped, int last)
 687 {
 688         uint16_t last_wrapped = last & 0xffff;
 689         if (current_wrapped > last_wrapped) {
 690                 return (last & ~0xffff) | current_wrapped;
 691         } else {
 692                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 693         }
 694 }
 695
 696 DeviceSpec card_index_to_device(unsigned card_index, unsigned num_cards)
 697 {
 698         if (card_index >= num_cards) {
 699                 return DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_cards};
 700         } else {
 701                 return DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
 702         }
 703 }
 704
 705 }  // namespace
 706
 707 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 708                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 709                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 710 {
 711         DeviceSpec device = card_index_to_device(card_index, num_cards);
 712         CaptureCard *card = &cards[card_index];
 713
 714         ++card->metric_input_received_frames;
 715         card->metric_input_has_signal_bool = video_format.has_signal;
 716         card->metric_input_is_connected_bool = video_format.is_connected;
 717         card->metric_input_interlaced_bool = video_format.interlaced;
 718         card->metric_input_width_pixels = video_format.width;
 719         card->metric_input_height_pixels = video_format.height;
 720         card->metric_input_frame_rate_nom = video_format.frame_rate_nom;
 721         card->metric_input_frame_rate_den = video_format.frame_rate_den;
 722         card->metric_input_sample_rate_hz = audio_format.sample_rate;
 723
 724         if (is_mode_scanning[card_index]) {
 725                 if (video_format.has_signal) {
 726                         // Found a stable signal, so stop scanning.
 727                         is_mode_scanning[card_index] = false;
 728                 } else {
 729                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 730                         steady_clock::time_point now = steady_clock::now();
 731                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 732                         if (sec_since_last_switch > switch_time_s) {
 733                                 // It isn't this mode; try the next one.
 734                                 mode_scanlist_index[card_index]++;
 735                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 736                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 737                                 last_mode_scan_change[card_index] = now;
 738                         }
 739                 }
 740         }
 741
 742         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 743         assert(frame_length > 0);
 744
 745         size_t num_samples = (audio_frame.len > audio_offset) ? (audio_frame.len - audio_offset) / audio_format.num_channels / (audio_format.bits_per_sample / 8) : 0;
 746         if (num_samples > OUTPUT_FREQUENCY / 10 && card->type != CardType::FFMPEG_INPUT) {
 747                 printf("%s: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n",
 748                         spec_to_string(device).c_str(), int(audio_frame.len), int(audio_offset),
 749                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 750                 if (video_frame.owner) {
 751                         video_frame.owner->release_frame(video_frame);
 752                 }
 753                 if (audio_frame.owner) {
 754                         audio_frame.owner->release_frame(audio_frame);
 755                 }
 756                 return;
 757         }
 758
 759         int dropped_frames = 0;
 760         if (card->last_timecode != -1) {
 761                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 762         }
 763
 764         // Number of samples per frame if we need to insert silence.
 765         // (Could be nonintegral, but resampling will save us then.)
 766         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 767
 768         if (dropped_frames > MAX_FPS * 2) {
 769                 fprintf(stderr, "%s lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 770                         spec_to_string(device).c_str(), card->last_timecode, timecode);
 771                 audio_mixer->reset_resampler(device);
 772                 dropped_frames = 0;
 773                 ++card->metric_input_resets;
 774         } else if (dropped_frames > 0) {
 775                 // Insert silence as needed.
 776                 fprintf(stderr, "%s dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 777                         spec_to_string(device).c_str(), dropped_frames, timecode);
 778                 card->metric_input_dropped_frames_error += dropped_frames;
 779
 780                 bool success;
 781                 do {
 782                         success = audio_mixer->add_silence(device, silence_samples, dropped_frames);
 783                 } while (!success);
 784         }
 785
 786         if (num_samples > 0) {
 787                 audio_mixer->add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, audio_frame.received_timestamp);
 788
 789                 // Audio for the MJPEG stream. We don't resample; audio that's not in 48 kHz
 790                 // just gets dropped for now.
 791                 //
 792                 // Only bother doing MJPEG encoding if there are any connected clients
 793                 // that want the stream.
 794                 if (httpd.get_num_connected_multicam_clients() > 0) {
 795                         vector<int32_t> converted_samples = convert_audio_to_fixed32(audio_frame.data + audio_offset, num_samples, audio_format, 2);
 796                         lock_guard<mutex> lock(card_mutex);
 797                         if (card->new_raw_audio.empty()) {
 798                                 card->new_raw_audio = move(converted_samples);
 799                         } else {
 800                                 // For raw audio, we don't really synchronize audio and video;
 801                                 // we just put the audio in frame by frame, and if a video frame is
 802                                 // dropped, we still keep the audio, which means it will be added
 803                                 // to the beginning of the next frame. It would probably be better
 804                                 // to move the audio pts earlier to show this, but most players can
 805                                 // live with some jitter, and in a lot of ways, it's much nicer for
 806                                 // Futatabi to have all audio locked to a video frame.
 807                                 card->new_raw_audio.insert(card->new_raw_audio.end(), converted_samples.begin(), converted_samples.end());
 808
 809                                 // Truncate to one second, just to be sure we don't have infinite buildup in case of weirdness.
 810                                 if (card->new_raw_audio.size() > OUTPUT_FREQUENCY * 2) {
 811                                         size_t excess_samples = card->new_raw_audio.size() - OUTPUT_FREQUENCY * 2;
 812                                         card->new_raw_audio.erase(card->new_raw_audio.begin(), card->new_raw_audio.begin() + excess_samples);
 813                                 }
 814                         }
 815                 }
 816         }
 817
 818         // Done with the audio, so release it.
 819         if (audio_frame.owner) {
 820                 audio_frame.owner->release_frame(audio_frame);
 821         }
 822
 823         card->last_timecode = timecode;
 824
 825         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 826         if (card->type == CardType::FFMPEG_INPUT && userdata != nullptr) {
 827                 FFmpegCapture *ffmpeg_capture = static_cast<FFmpegCapture *>(card->capture.get());
 828                 userdata->has_last_subtitle = ffmpeg_capture->get_has_last_subtitle();
 829                 userdata->last_subtitle = ffmpeg_capture->get_last_subtitle();
 830         }
 831
 832         size_t cbcr_width, cbcr_height, cbcr_offset, y_offset;
 833         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 834         if (userdata != nullptr && userdata->pixel_format == PixelFormat_8BitYCbCrPlanar) {
 835                 // The calculation above is wrong for planar Y'CbCr, so just override it.
 836                 assert(card->type == CardType::FFMPEG_INPUT);
 837                 assert(video_offset == 0);
 838                 expected_length = video_frame.len;
 839
 840                 userdata->ycbcr_format = (static_cast<FFmpegCapture *>(card->capture.get()))->get_current_frame_ycbcr_format();
 841                 cbcr_width = video_format.width / userdata->ycbcr_format.chroma_subsampling_x;
 842                 cbcr_height = video_format.height / userdata->ycbcr_format.chroma_subsampling_y;
 843                 cbcr_offset = video_format.width * video_format.height;
 844                 y_offset = 0;
 845         } else {
 846                 // All the other Y'CbCr formats are 4:2:2.
 847                 cbcr_width = video_format.width / 2;
 848                 cbcr_height = video_format.height;
 849                 cbcr_offset = video_offset / 2;
 850                 y_offset = video_frame.size / 2 + video_offset / 2;
 851         }
 852         if (video_frame.len - video_offset == 0 ||
 853             video_frame.len - video_offset != expected_length) {
 854                 if (video_frame.len != 0) {
 855                         printf("%s: Dropping video frame with wrong length (%zu; expected %zu)\n",
 856                                 spec_to_string(device).c_str(), video_frame.len - video_offset, expected_length);
 857                 }
 858                 if (video_frame.owner) {
 859                         video_frame.owner->release_frame(video_frame);
 860                 }
 861
 862                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 863                 // so that pts can go up accordingly.
 864                 {
 865                         lock_guard<mutex> lock(card_mutex);
 866                         CaptureCard::NewFrame new_frame;
 867                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 868                         new_frame.length = frame_length;
 869                         new_frame.interlaced = false;
 870                         new_frame.dropped_frames = dropped_frames;
 871                         new_frame.received_timestamp = video_frame.received_timestamp;
 872                         card->new_frames.push_back(move(new_frame));
 873                         card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
 874                 }
 875                 card->new_frames_changed.notify_all();
 876                 return;
 877         }
 878
 879         unsigned num_fields = video_format.interlaced ? 2 : 1;
 880         steady_clock::time_point frame_upload_start;
 881         bool interlaced_stride = false;
 882         if (video_format.interlaced) {
 883                 // Send the two fields along as separate frames; the other side will need to add
 884                 // a deinterlacer to actually get this right.
 885                 assert(video_format.height % 2 == 0);
 886                 video_format.height /= 2;
 887                 cbcr_height /= 2;
 888                 assert(frame_length % 2 == 0);
 889                 frame_length /= 2;
 890                 num_fields = 2;
 891                 if (video_format.second_field_start == 1) {
 892                         interlaced_stride = true;
 893                 }
 894                 frame_upload_start = steady_clock::now();
 895         }
 896         assert(userdata != nullptr);
 897         userdata->last_interlaced = video_format.interlaced;
 898         userdata->last_has_signal = video_format.has_signal;
 899         userdata->last_is_connected = video_format.is_connected;
 900         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 901         userdata->last_frame_rate_den = video_format.frame_rate_den;
 902         RefCountedFrame frame(video_frame);
 903
 904         // Upload the textures.
 905         for (unsigned field = 0; field < num_fields; ++field) {
 906                 // Put the actual texture upload in a lambda that is executed in the main thread.
 907                 // It is entirely possible to do this in the same thread (and it might even be
 908                 // faster, depending on the GPU and driver), but it appears to be trickling
 909                 // driver bugs very easily.
 910                 //
 911                 // Note that this means we must hold on to the actual frame data in <userdata>
 912                 // until the upload command is run, but we hold on to <frame> much longer than that
 913                 // (in fact, all the way until we no longer use the texture in rendering).
 914                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, cbcr_height, interlaced_stride, userdata]() {
 915                         unsigned field_start_line;
 916                         if (field == 1) {
 917                                 field_start_line = video_format.second_field_start;
 918                         } else {
 919                                 field_start_line = video_format.extra_lines_top;
 920                         }
 921
 922                         // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
 923                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 924                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, cbcr_width, cbcr_height, v210_width);
 925
 926                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 927                         check_error();
 928
 929                         switch (userdata->pixel_format) {
 930                         case PixelFormat_10BitYCbCr: {
 931                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 932                                 upload_texture(userdata->tex_v210[field], v210_width, video_format.height, video_format.stride, interlaced_stride, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, field_start);
 933                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 934                                 break;
 935                         }
 936                         case PixelFormat_8BitYCbCr: {
 937                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 938                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 939
 940                                 // Make up our own strides, since we are interleaving.
 941                                 upload_texture(userdata->tex_y[field], video_format.width, video_format.height, video_format.width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_y_start);
 942                                 upload_texture(userdata->tex_cbcr[field], cbcr_width, cbcr_height, cbcr_width * sizeof(uint16_t), interlaced_stride, GL_RG, GL_UNSIGNED_BYTE, field_cbcr_start);
 943                                 break;
 944                         }
 945                         case PixelFormat_8BitYCbCrPlanar: {
 946                                 assert(field_start_line == 0);  // We don't really support interlaced here.
 947                                 size_t field_y_start = y_offset;
 948                                 size_t field_cb_start = cbcr_offset;
 949                                 size_t field_cr_start = cbcr_offset + cbcr_width * cbcr_height;
 950
 951                                 // Make up our own strides, since we are interleaving.
 952                                 upload_texture(userdata->tex_y[field], video_format.width, video_format.height, video_format.width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_y_start);
 953                                 upload_texture(userdata->tex_cb[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cb_start);
 954                                 upload_texture(userdata->tex_cr[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cr_start);
 955                                 break;
 956                         }
 957                         case PixelFormat_8BitBGRA: {
 958                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 959                                 upload_texture(userdata->tex_rgba[field], video_format.width, video_format.height, video_format.stride, interlaced_stride, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, field_start);
 960                                 // These could be asked to deliver mipmaps at any time.
 961                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 962                                 check_error();
 963                                 glGenerateMipmap(GL_TEXTURE_2D);
 964                                 check_error();
 965                                 glBindTexture(GL_TEXTURE_2D, 0);
 966                                 check_error();
 967                                 break;
 968                         }
 969                         default:
 970                                 assert(false);
 971                         }
 972
 973                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 974                         check_error();
 975                 };
 976
 977                 if (field == 1) {
 978                         // Don't upload the second field as fast as we can; wait until
 979                         // the field time has approximately passed. (Otherwise, we could
 980                         // get timing jitter against the other sources, and possibly also
 981                         // against the video display, although the latter is not as critical.)
 982                         // This requires our system clock to be reasonably close to the
 983                         // video clock, but that's not an unreasonable assumption.
 984                         steady_clock::time_point second_field_start = frame_upload_start +
 985                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 986                         this_thread::sleep_until(second_field_start);
 987                 }
 988
 989                 {
 990                         lock_guard<mutex> lock(card_mutex);
 991                         CaptureCard::NewFrame new_frame;
 992                         new_frame.frame = frame;
 993                         new_frame.length = frame_length;
 994                         new_frame.field = field;
 995                         new_frame.interlaced = video_format.interlaced;
 996                         new_frame.upload_func = upload_func;
 997                         new_frame.dropped_frames = dropped_frames;
 998                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
 999                         new_frame.video_format = video_format;
1000                         new_frame.y_offset = y_offset;
1001                         new_frame.cbcr_offset = cbcr_offset;
1002                         card->new_frames.push_back(move(new_frame));
1003                         card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
1004                         card->may_have_dropped_last_frame = false;
1005                 }
1006                 card->new_frames_changed.notify_all();
1007         }
1008 }
1009
1010 void Mixer::bm_hotplug_add(libusb_device *dev)
1011 {
1012         lock_guard<mutex> lock(hotplug_mutex);
1013         hotplugged_cards.push_back(dev);
1014 }
1015
1016 void Mixer::bm_hotplug_remove(unsigned card_index)
1017 {
1018         cards[card_index].new_frames_changed.notify_all();
1019 }
1020
1021 void Mixer::thread_func()
1022 {
1023         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
1024
1025         eglBindAPI(EGL_OPENGL_API);
1026         QOpenGLContext *context = create_context(mixer_surface);
1027         if (!make_current(context, mixer_surface)) {
1028                 printf("oops\n");
1029                 abort();
1030         }
1031
1032         // Start the actual capture. (We don't want to do it before we're actually ready
1033         // to process output frames.)
1034         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1035                 if (int(card_index) != output_card_index) {
1036                         cards[card_index].capture->start_bm_capture();
1037                 }
1038         }
1039
1040         BasicStats basic_stats(/*verbose=*/true, /*use_opengl=*/true);
1041         int stats_dropped_frames = 0;
1042
1043         while (!should_quit) {
1044                 if (desired_output_card_index != output_card_index) {
1045                         set_output_card_internal(desired_output_card_index);
1046                 }
1047                 if (output_card_index != -1 &&
1048                     desired_output_video_mode != output_video_mode) {
1049                         DeckLinkOutput *output = cards[output_card_index].output.get();
1050                         output->end_output();
1051                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
1052                         output->start_output(desired_output_video_mode, pts_int);
1053                 }
1054
1055                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
1056                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
1057
1058                 bool master_card_is_output;
1059                 unsigned master_card_index;
1060                 if (output_card_index != -1) {
1061                         master_card_is_output = true;
1062                         master_card_index = output_card_index;
1063                 } else {
1064                         master_card_is_output = false;
1065                         master_card_index = theme->map_signal(master_clock_channel);
1066                         assert(master_card_index < num_cards + num_video_inputs);
1067                 }
1068
1069                 vector<int32_t> raw_audio[MAX_VIDEO_CARDS];  // For MJPEG encoding.
1070                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame, raw_audio);
1071                 schedule_audio_resampling_tasks(output_frame_info.dropped_frames, output_frame_info.num_samples, output_frame_info.frame_duration, output_frame_info.is_preroll, output_frame_info.frame_timestamp);
1072                 stats_dropped_frames += output_frame_info.dropped_frames;
1073
1074                 handle_hotplugged_cards();
1075
1076                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1077                         DeviceSpec device = card_index_to_device(card_index, num_cards);
1078                         if (card_index == master_card_index || !has_new_frame[card_index]) {
1079                                 continue;
1080                         }
1081                         if (new_frames[card_index].frame->len == 0) {
1082                                 ++new_frames[card_index].dropped_frames;
1083                         }
1084                         if (new_frames[card_index].dropped_frames > 0) {
1085                                 printf("%s dropped %d frames before this\n",
1086                                         spec_to_string(device).c_str(), int(new_frames[card_index].dropped_frames));
1087                         }
1088                 }
1089
1090                 // If the first card is reporting a corrupted or otherwise dropped frame,
1091                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
1092                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
1093                         ++stats_dropped_frames;
1094                         pts_int += new_frames[master_card_index].length;
1095                         continue;
1096                 }
1097
1098                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1099                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
1100                                 continue;
1101
1102                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
1103                         assert(new_frame->frame != nullptr);
1104                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
1105                         check_error();
1106
1107                         // The new texture might need uploading before use.
1108                         if (new_frame->upload_func) {
1109                                 new_frame->upload_func();
1110                                 new_frame->upload_func = nullptr;
1111                         }
1112
1113                         if (new_frame->frame->data_copy != nullptr) {
1114                                 int mjpeg_card_index = mjpeg_encoder->get_mjpeg_stream_for_card(card_index);
1115                                 if (mjpeg_card_index != -1) {
1116                                         mjpeg_encoder->upload_frame(pts_int, mjpeg_card_index, new_frame->frame, new_frame->video_format, new_frame->y_offset, new_frame->cbcr_offset, move(raw_audio[card_index]));
1117                                 }
1118                         }
1119                 }
1120
1121                 int64_t frame_duration = output_frame_info.frame_duration;
1122                 render_one_frame(frame_duration);
1123                 {
1124                         lock_guard<mutex> lock(frame_num_mutex);
1125                         ++frame_num;
1126                 }
1127                 frame_num_updated.notify_all();
1128                 pts_int += frame_duration;
1129
1130                 basic_stats.update(frame_num, stats_dropped_frames);
1131                 // if (frame_num % 100 == 0) chain->print_phase_timing();
1132
1133                 if (should_cut.exchange(false)) {  // Test and clear.
1134                         video_encoder->do_cut(frame_num);
1135                 }
1136
1137 #if 0
1138                 // Reset every 100 frames, so that local variations in frame times
1139                 // (especially for the first few frames, when the shaders are
1140                 // compiled etc.) don't make it hard to measure for the entire
1141                 // remaining duration of the program.
1142                 if (frame == 10000) {
1143                         frame = 0;
1144                         start = now;
1145                 }
1146 #endif
1147                 check_error();
1148         }
1149
1150         resource_pool->clean_context();
1151 }
1152
1153 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
1154 {
1155         if (output_card_index != -1) {
1156                 // The output card (ie., cards[output_card_index].output) is the master clock,
1157                 // so no input card (ie., cards[card_index].capture) is.
1158                 return false;
1159         }
1160         return (card_index == master_card_index);
1161 }
1162
1163 void Mixer::trim_queue(CaptureCard *card, size_t safe_queue_length)
1164 {
1165         // Count the number of frames in the queue, including any frames
1166         // we dropped. It's hard to know exactly how we should deal with
1167         // dropped (corrupted) input frames; they don't help our goal of
1168         // avoiding starvation, but they still add to the problem of latency.
1169         // Since dropped frames is going to mean a bump in the signal anyway,
1170         // we err on the side of having more stable latency instead.
1171         unsigned queue_length = 0;
1172         for (const CaptureCard::NewFrame &frame : card->new_frames) {
1173                 queue_length += frame.dropped_frames + 1;
1174         }
1175
1176         // If needed, drop frames until the queue is below the safe limit.
1177         // We prefer to drop from the head, because all else being equal,
1178         // we'd like more recent frames (less latency).
1179         unsigned dropped_frames = 0;
1180         while (queue_length > safe_queue_length) {
1181                 assert(!card->new_frames.empty());
1182                 assert(queue_length > card->new_frames.front().dropped_frames);
1183                 queue_length -= card->new_frames.front().dropped_frames;
1184
1185                 if (queue_length <= safe_queue_length) {
1186                         // No need to drop anything.
1187                         break;
1188                 }
1189
1190                 card->new_frames.pop_front();
1191                 card->new_frames_changed.notify_all();
1192                 --queue_length;
1193                 ++dropped_frames;
1194
1195                 if (queue_length == 0 && card->is_cef_capture) {
1196                         card->may_have_dropped_last_frame = true;
1197                 }
1198         }
1199
1200         card->metric_input_dropped_frames_jitter += dropped_frames;
1201         card->metric_input_queue_length_frames = queue_length;
1202
1203 #if 0
1204         if (dropped_frames > 0) {
1205                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
1206                         card_index, dropped_frames);
1207         }
1208 #endif
1209 }
1210
1211 pair<string, string> Mixer::get_channels_json()
1212 {
1213         Channels ret;
1214         for (int channel_idx = 2; channel_idx < theme->get_num_channels(); ++channel_idx) {
1215                 Channel *channel = ret.add_channel();
1216                 channel->set_index(channel_idx);
1217                 channel->set_name(theme->get_channel_name(channel_idx));
1218                 channel->set_color(theme->get_channel_color(channel_idx));
1219         }
1220         string contents;
1221         google::protobuf::util::MessageToJsonString(ret, &contents);  // Ignore any errors.
1222         return make_pair(contents, "text/json");
1223 }
1224
1225 pair<string, string> Mixer::get_channel_color_http(unsigned channel_idx)
1226 {
1227         return make_pair(theme->get_channel_color(channel_idx), "text/plain");
1228 }
1229
1230 Mixer::OutputFrameInfo Mixer::get_one_frame_from_each_card(unsigned master_card_index, bool master_card_is_output, CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS], bool has_new_frame[MAX_VIDEO_CARDS], vector<int32_t> raw_audio[MAX_VIDEO_CARDS])
1231 {
1232         OutputFrameInfo output_frame_info;
1233 start:
1234         unique_lock<mutex> lock(card_mutex, defer_lock);
1235         if (master_card_is_output) {
1236                 // Clocked to the output, so wait for it to be ready for the next frame.
1237                 cards[master_card_index].output->wait_for_frame(pts_int, &output_frame_info.dropped_frames, &output_frame_info.frame_duration, &output_frame_info.is_preroll, &output_frame_info.frame_timestamp);
1238                 lock.lock();
1239         } else {
1240                 // Wait for the master card to have a new frame.
1241                 // TODO: Add a timeout.
1242                 output_frame_info.is_preroll = false;
1243                 lock.lock();
1244                 cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty() || cards[master_card_index].capture->get_disconnected(); });
1245         }
1246
1247         if (master_card_is_output) {
1248                 handle_hotplugged_cards();
1249         } else if (cards[master_card_index].new_frames.empty()) {
1250                 // We were woken up, but not due to a new frame. Deal with it
1251                 // and then restart.
1252                 assert(cards[master_card_index].capture->get_disconnected());
1253                 handle_hotplugged_cards();
1254                 lock.unlock();
1255                 goto start;
1256         }
1257
1258         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1259                 CaptureCard *card = &cards[card_index];
1260                 if (card->new_frames.empty()) {  // Starvation.
1261                         ++card->metric_input_duped_frames;
1262 #ifdef HAVE_CEF
1263                         if (card->is_cef_capture && card->may_have_dropped_last_frame) {
1264                                 // Unlike other sources, CEF is not guaranteed to send us a steady
1265                                 // stream of frames, so we'll have to ask it to repaint the frame
1266                                 // we dropped. (may_have_dropped_last_frame is set whenever we
1267                                 // trim the queue completely away, and cleared when we actually
1268                                 // get a new frame.)
1269                                 ((CEFCapture *)card->capture.get())->request_new_frame();
1270                         }
1271 #endif
1272                 } else {
1273                         new_frames[card_index] = move(card->new_frames.front());
1274                         has_new_frame[card_index] = true;
1275                         card->new_frames.pop_front();
1276                         card->new_frames_changed.notify_all();
1277                 }
1278
1279                 raw_audio[card_index] = move(card->new_raw_audio);
1280         }
1281
1282         if (!master_card_is_output) {
1283                 output_frame_info.frame_timestamp = new_frames[master_card_index].received_timestamp;
1284                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
1285                 output_frame_info.frame_duration = new_frames[master_card_index].length;
1286         }
1287
1288         if (!output_frame_info.is_preroll) {
1289                 output_jitter_history.frame_arrived(output_frame_info.frame_timestamp, output_frame_info.frame_duration, output_frame_info.dropped_frames);
1290         }
1291
1292         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1293                 CaptureCard *card = &cards[card_index];
1294                 if (has_new_frame[card_index] &&
1295                     !input_card_is_master_clock(card_index, master_card_index) &&
1296                     !output_frame_info.is_preroll) {
1297                         card->queue_length_policy.update_policy(
1298                                 output_frame_info.frame_timestamp,
1299                                 card->jitter_history.get_expected_next_frame(),
1300                                 new_frames[master_card_index].length,
1301                                 output_frame_info.frame_duration,
1302                                 card->jitter_history.estimate_max_jitter(),
1303                                 output_jitter_history.estimate_max_jitter());
1304                         trim_queue(card, min<int>(global_flags.max_input_queue_frames,
1305                                                   card->queue_length_policy.get_safe_queue_length()));
1306                 }
1307         }
1308
1309         // This might get off by a fractional sample when changing master card
1310         // between ones with different frame rates, but that's fine.
1311         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
1312         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
1313         fractional_samples = num_samples_times_timebase % TIMEBASE;
1314         assert(output_frame_info.num_samples >= 0);
1315
1316         return output_frame_info;
1317 }
1318
1319 void Mixer::handle_hotplugged_cards()
1320 {
1321         // Check for cards that have been disconnected since last frame.
1322         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1323                 CaptureCard *card = &cards[card_index];
1324                 if (card->capture->get_disconnected()) {
1325                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1326                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1327                         configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1328                         card->queue_length_policy.reset(card_index);
1329                         card->capture->start_bm_capture();
1330                 }
1331         }
1332
1333         // Check for cards that have been connected since last frame.
1334         vector<libusb_device *> hotplugged_cards_copy;
1335         {
1336                 lock_guard<mutex> lock(hotplug_mutex);
1337                 swap(hotplugged_cards, hotplugged_cards_copy);
1338         }
1339         for (libusb_device *new_dev : hotplugged_cards_copy) {
1340                 // Look for a fake capture card where we can stick this in.
1341                 int free_card_index = -1;
1342                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1343                         if (cards[card_index].is_fake_capture) {
1344                                 free_card_index = card_index;
1345                                 break;
1346                         }
1347                 }
1348
1349                 if (free_card_index == -1) {
1350                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1351                         libusb_unref_device(new_dev);
1352                 } else {
1353                         // BMUSBCapture takes ownership.
1354                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1355                         CaptureCard *card = &cards[free_card_index];
1356                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1357                         configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1358                         card->queue_length_policy.reset(free_card_index);
1359                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1360                         capture->start_bm_capture();
1361                 }
1362         }
1363 }
1364
1365
1366 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame, bool is_preroll, steady_clock::time_point frame_timestamp)
1367 {
1368         // Resample the audio as needed, including from previously dropped frames.
1369         assert(num_cards > 0);
1370         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1371                 const bool dropped_frame = (frame_num != dropped_frames);
1372                 {
1373                         // Signal to the audio thread to process this frame.
1374                         // Note that if the frame is a dropped frame, we signal that
1375                         // we don't want to use this frame as base for adjusting
1376                         // the resampler rate. The reason for this is that the timing
1377                         // of these frames is often way too late; they typically don't
1378                         // “arrive” before we synthesize them. Thus, we could end up
1379                         // in a situation where we have inserted e.g. five audio frames
1380                         // into the queue before we then start pulling five of them
1381                         // back out. This makes ResamplingQueue overestimate the delay,
1382                         // causing undue resampler changes. (We _do_ use the last,
1383                         // non-dropped frame; perhaps we should just discard that as well,
1384                         // since dropped frames are expected to be rare, and it might be
1385                         // better to just wait until we have a slightly more normal situation).
1386                         lock_guard<mutex> lock(audio_mutex);
1387                         bool adjust_rate = !dropped_frame && !is_preroll;
1388                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1389                         audio_task_queue_changed.notify_one();
1390                 }
1391                 if (dropped_frame) {
1392                         // For dropped frames, increase the pts. Note that if the format changed
1393                         // in the meantime, we have no way of detecting that; we just have to
1394                         // assume the frame length is always the same.
1395                         pts_int += length_per_frame;
1396                 }
1397         }
1398 }
1399
1400 void Mixer::render_one_frame(int64_t duration)
1401 {
1402         // Determine the time code for this frame before we start rendering.
1403         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1404         if (display_timecode_on_stdout) {
1405                 printf("Timecode: '%s'\n", timecode_text.c_str());
1406         }
1407
1408         // Update Y'CbCr settings for all cards.
1409         {
1410                 lock_guard<mutex> lock(card_mutex);
1411                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1412                         YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index];
1413                         input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto;
1414                         input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients;
1415                         input_state.full_range[card_index] = interpretation->full_range;
1416                 }
1417         }
1418
1419         // Get the main chain from the theme, and set its state immediately.
1420         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1421         EffectChain *chain = theme_main_chain.chain;
1422         theme_main_chain.setup_chain();
1423         //theme_main_chain.chain->enable_phase_timing(true);
1424
1425         // If HDMI/SDI output is active and the user has requested auto mode,
1426         // its mode overrides the existing Y'CbCr setting for the chain.
1427         YCbCrLumaCoefficients ycbcr_output_coefficients;
1428         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1429                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1430         } else {
1431                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1432         }
1433
1434         // TODO: Reduce the duplication against theme.cpp.
1435         YCbCrFormat output_ycbcr_format;
1436         output_ycbcr_format.chroma_subsampling_x = 1;
1437         output_ycbcr_format.chroma_subsampling_y = 1;
1438         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1439         output_ycbcr_format.full_range = false;
1440         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1441         chain->change_ycbcr_output_format(output_ycbcr_format);
1442
1443         // Render main chain. If we're using zerocopy Quick Sync encoding
1444         // (the default case), we take an extra copy of the created outputs,
1445         // so that we can display it back to the screen later (it's less memory
1446         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1447         // Ideally, we'd like to avoid taking copies and just use the main textures
1448         // for display as well, but they're just views into VA-API memory and must be
1449         // unmapped during encoding, so we can't use them for display, unfortunately.
1450         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1451         GLuint y_copy_tex, cbcr_copy_tex = 0;
1452         GLuint y_display_tex, cbcr_display_tex;
1453         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1454         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1455         const bool is_zerocopy = video_encoder->is_zerocopy();
1456         if (is_zerocopy) {
1457                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1458                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1459                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1460
1461                 y_display_tex = y_copy_tex;
1462                 cbcr_display_tex = cbcr_copy_tex;
1463
1464                 // y_tex and cbcr_tex will be given by VideoEncoder.
1465         } else {
1466                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1467                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1468                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1469
1470                 y_display_tex = y_tex;
1471                 cbcr_display_tex = cbcr_tex;
1472         }
1473
1474         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1475         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1476         assert(got_frame);
1477
1478         GLuint fbo;
1479         if (is_zerocopy) {
1480                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1481         } else {
1482                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1483         }
1484         check_error();
1485         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1486
1487         if (display_timecode_in_stream) {
1488                 // Render the timecode on top.
1489                 timecode_renderer->render_timecode(fbo, timecode_text);
1490         }
1491
1492         resource_pool->release_fbo(fbo);
1493
1494         if (is_zerocopy) {
1495                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1496         } else {
1497                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1498         }
1499         if (output_card_index != -1) {
1500                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1501         }
1502         resource_pool->release_2d_texture(cbcr_full_tex);
1503
1504         // Set the right state for the Y' and CbCr textures we use for display.
1505         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1506         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1507         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1508         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1509         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1510
1511         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1512         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1513         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1514         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1515
1516         RefCountedGLsync fence = video_encoder->end_frame();
1517
1518         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1519         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1520         DisplayFrame live_frame;
1521         live_frame.chain = display_chain.get();
1522         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1523                 display_input->set_texture_num(0, y_display_tex);
1524                 display_input->set_texture_num(1, cbcr_display_tex);
1525         };
1526         live_frame.ready_fence = fence;
1527         live_frame.input_frames = {};
1528         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1529         output_channel[OUTPUT_LIVE].output_frame(move(live_frame));
1530
1531         // Set up preview and any additional channels.
1532         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1533                 DisplayFrame display_frame;
1534                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1535                 display_frame.chain = move(chain.chain);
1536                 display_frame.setup_chain = move(chain.setup_chain);
1537                 display_frame.ready_fence = fence;
1538                 display_frame.input_frames = move(chain.input_frames);
1539                 display_frame.temp_textures = {};
1540                 output_channel[i].output_frame(move(display_frame));
1541         }
1542 }
1543
1544 void Mixer::audio_thread_func()
1545 {
1546         pthread_setname_np(pthread_self(), "Mixer_Audio");
1547
1548         while (!should_quit) {
1549                 AudioTask task;
1550
1551                 {
1552                         unique_lock<mutex> lock(audio_mutex);
1553                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1554                         if (should_quit) {
1555                                 return;
1556                         }
1557                         task = audio_task_queue.front();
1558                         audio_task_queue.pop();
1559                 }
1560
1561                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1562                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1563                 vector<float> samples_out = audio_mixer->get_output(
1564                         task.frame_timestamp,
1565                         task.num_samples,
1566                         rate_adjustment_policy);
1567
1568                 // Send the samples to the sound card, then add them to the output.
1569                 if (alsa) {
1570                         alsa->write(samples_out);
1571                 }
1572                 if (output_card_index != -1) {
1573                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1574                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1575                 }
1576                 video_encoder->add_audio(task.pts_int, move(samples_out));
1577         }
1578 }
1579
1580 void Mixer::release_display_frame(DisplayFrame *frame)
1581 {
1582         for (GLuint texnum : frame->temp_textures) {
1583                 resource_pool->release_2d_texture(texnum);
1584         }
1585         frame->temp_textures.clear();
1586         frame->ready_fence.reset();
1587         frame->input_frames.clear();
1588 }
1589
1590 void Mixer::start()
1591 {
1592         mixer_thread = thread(&Mixer::thread_func, this);
1593         audio_thread = thread(&Mixer::audio_thread_func, this);
1594 }
1595
1596 void Mixer::quit()
1597 {
1598         should_quit = true;
1599         audio_task_queue_changed.notify_one();
1600         mixer_thread.join();
1601         audio_thread.join();
1602 }
1603
1604 void Mixer::transition_clicked(int transition_num)
1605 {
1606         theme->transition_clicked(transition_num, pts());
1607 }
1608
1609 void Mixer::channel_clicked(int preview_num)
1610 {
1611         theme->channel_clicked(preview_num);
1612 }
1613
1614 YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const
1615 {
1616         lock_guard<mutex> lock(card_mutex);
1617         return ycbcr_interpretation[card_index];
1618 }
1619
1620 void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation)
1621 {
1622         lock_guard<mutex> lock(card_mutex);
1623         ycbcr_interpretation[card_index] = interpretation;
1624 }
1625
1626 void Mixer::start_mode_scanning(unsigned card_index)
1627 {
1628         assert(card_index < num_cards);
1629         if (is_mode_scanning[card_index]) {
1630                 return;
1631         }
1632         is_mode_scanning[card_index] = true;
1633         mode_scanlist[card_index].clear();
1634         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1635                 mode_scanlist[card_index].push_back(mode.first);
1636         }
1637         assert(!mode_scanlist[card_index].empty());
1638         mode_scanlist_index[card_index] = 0;
1639         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1640         last_mode_scan_change[card_index] = steady_clock::now();
1641 }
1642
1643 map<uint32_t, VideoMode> Mixer::get_available_output_video_modes() const
1644 {
1645         assert(desired_output_card_index != -1);
1646         lock_guard<mutex> lock(card_mutex);
1647         return cards[desired_output_card_index].output->get_available_video_modes();
1648 }
1649
1650 string Mixer::get_ffmpeg_filename(unsigned card_index) const
1651 {
1652         assert(card_index >= num_cards && card_index < num_cards + num_video_inputs);
1653         return ((FFmpegCapture *)(cards[card_index].capture.get()))->get_filename();
1654 }
1655
1656 void Mixer::set_ffmpeg_filename(unsigned card_index, const string &filename) {
1657         assert(card_index >= num_cards && card_index < num_cards + num_video_inputs);
1658         ((FFmpegCapture *)(cards[card_index].capture.get()))->change_filename(filename);
1659 }
1660
1661 void Mixer::wait_for_next_frame()
1662 {
1663         unique_lock<mutex> lock(frame_num_mutex);
1664         unsigned old_frame_num = frame_num;
1665         frame_num_updated.wait_for(lock, seconds(1),  // Timeout is just in case.
1666                 [old_frame_num, this]{ return this->frame_num > old_frame_num; });
1667 }
1668
1669 Mixer::OutputChannel::~OutputChannel()
1670 {
1671         if (has_current_frame) {
1672                 parent->release_display_frame(&current_frame);
1673         }
1674         if (has_ready_frame) {
1675                 parent->release_display_frame(&ready_frame);
1676         }
1677 }
1678
1679 void Mixer::OutputChannel::output_frame(DisplayFrame &&frame)
1680 {
1681         // Store this frame for display. Remove the ready frame if any
1682         // (it was seemingly never used).
1683         {
1684                 lock_guard<mutex> lock(frame_mutex);
1685                 if (has_ready_frame) {
1686                         parent->release_display_frame(&ready_frame);
1687                 }
1688                 ready_frame = move(frame);
1689                 has_ready_frame = true;
1690
1691                 // Call the callbacks under the mutex (they should be short),
1692                 // so that we don't race against a callback removal.
1693                 for (const auto &key_and_callback : new_frame_ready_callbacks) {
1694                         key_and_callback.second();
1695                 }
1696         }
1697
1698         // Reduce the number of callbacks by filtering duplicates. The reason
1699         // why we bother doing this is that Qt seemingly can get into a state
1700         // where its builds up an essentially unbounded queue of signals,
1701         // consuming more and more memory, and there's no good way of collapsing
1702         // user-defined signals or limiting the length of the queue.
1703         if (transition_names_updated_callback) {
1704                 vector<string> transition_names = global_mixer->get_transition_names();
1705                 bool changed = false;
1706                 if (transition_names.size() != last_transition_names.size()) {
1707                         changed = true;
1708                 } else {
1709                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1710                                 if (transition_names[i] != last_transition_names[i]) {
1711                                         changed = true;
1712                                         break;
1713                                 }
1714                         }
1715                 }
1716                 if (changed) {
1717                         transition_names_updated_callback(transition_names);
1718                         last_transition_names = transition_names;
1719                 }
1720         }
1721         if (name_updated_callback) {
1722                 string name = global_mixer->get_channel_name(channel);
1723                 if (name != last_name) {
1724                         name_updated_callback(name);
1725                         last_name = name;
1726                 }
1727         }
1728         if (color_updated_callback) {
1729                 string color = global_mixer->get_channel_color(channel);
1730                 if (color != last_color) {
1731                         color_updated_callback(color);
1732                         last_color = color;
1733                 }
1734         }
1735 }
1736
1737 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1738 {
1739         lock_guard<mutex> lock(frame_mutex);
1740         if (!has_current_frame && !has_ready_frame) {
1741                 return false;
1742         }
1743
1744         if (has_current_frame && has_ready_frame) {
1745                 // We have a new ready frame. Toss the current one.
1746                 parent->release_display_frame(&current_frame);
1747                 has_current_frame = false;
1748         }
1749         if (has_ready_frame) {
1750                 assert(!has_current_frame);
1751                 current_frame = move(ready_frame);
1752                 ready_frame.ready_fence.reset();  // Drop the refcount.
1753                 ready_frame.input_frames.clear();  // Drop the refcounts.
1754                 has_current_frame = true;
1755                 has_ready_frame = false;
1756         }
1757
1758         *frame = current_frame;
1759         return true;
1760 }
1761
1762 void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback)
1763 {
1764         lock_guard<mutex> lock(frame_mutex);
1765         new_frame_ready_callbacks[key] = callback;
1766 }
1767
1768 void Mixer::OutputChannel::remove_frame_ready_callback(void *key)
1769 {
1770         lock_guard<mutex> lock(frame_mutex);
1771         new_frame_ready_callbacks.erase(key);
1772 }
1773
1774 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1775 {
1776         transition_names_updated_callback = callback;
1777 }
1778
1779 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1780 {
1781         name_updated_callback = callback;
1782 }
1783
1784 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1785 {
1786         color_updated_callback = callback;
1787 }
1788
1789 mutex RefCountedGLsync::fence_lock;