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

   1 #undef Success
   2
   3 #include "mixer.h"
   4
   5 #include <assert.h>
   6 #include <endian.h>
   7 #include <epoxy/egl.h>
   8 #include <movit/effect_chain.h>
   9 #include <movit/effect_util.h>
  10 #include <movit/flat_input.h>
  11 #include <movit/image_format.h>
  12 #include <movit/init.h>
  13 #include <movit/resource_pool.h>
  14 #include <movit/util.h>
  15 #include <stdint.h>
  16 #include <stdio.h>
  17 #include <stdlib.h>
  18 #include <sys/resource.h>
  19 #include <sys/time.h>
  20 #include <time.h>
  21 #include <algorithm>
  22 #include <chrono>
  23 #include <cmath>
  24 #include <condition_variable>
  25 #include <cstddef>
  26 #include <memory>
  27 #include <mutex>
  28 #include <string>
  29 #include <thread>
  30 #include <utility>
  31 #include <vector>
  32
  33 #include "bmusb/bmusb.h"
  34 #include "bmusb/fake_capture.h"
  35 #include "context.h"
  36 #include "db.h"
  37 #include "decklink_capture.h"
  38 #include "defs.h"
  39 #include "disk_space_estimator.h"
  40 #include "flags.h"
  41 #include "pbo_frame_allocator.h"
  42 #include "ref_counted_gl_sync.h"
  43 #include "timebase.h"
  44 #include "video_encoder.h"
  45
  46 class QOpenGLContext;
  47
  48 using namespace movit;
  49 using namespace std;
  50 using namespace std::chrono;
  51 using namespace std::placeholders;
  52 using namespace bmusb;
  53
  54 Mixer *global_mixer = nullptr;
  55 bool uses_mlock = false;
  56
  57 namespace {
  58
  59 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  60 {
  61         if (interlaced) {
  62                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  63                         input_state->buffered_frames[card_index][frame_num] =
  64                                 input_state->buffered_frames[card_index][frame_num - 1];
  65                 }
  66                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  67         } else {
  68                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  69                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  70                 }
  71         }
  72 }
  73
  74 }  // namespace
  75
  76 void QueueLengthPolicy::update_policy(int queue_length)
  77 {
  78         if (queue_length < 0) {  // Starvation.
  79                 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
  80                         ++safe_queue_length;
  81                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
  82                                 card_index, safe_queue_length);
  83                 }
  84                 frames_with_at_least_one = 0;
  85                 been_at_safe_point_since_last_starvation = false;
  86                 return;
  87         }
  88         if (queue_length > 0) {
  89                 if (queue_length >= int(safe_queue_length)) {
  90                         been_at_safe_point_since_last_starvation = true;
  91                 }
  92                 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
  93                         --safe_queue_length;
  94                         fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
  95                                 card_index, safe_queue_length);
  96                         frames_with_at_least_one = 0;
  97                 }
  98         } else {
  99                 frames_with_at_least_one = 0;
 100         }
 101 }
 102
 103 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 104         : httpd(),
 105           num_cards(num_cards),
 106           mixer_surface(create_surface(format)),
 107           h264_encoder_surface(create_surface(format)),
 108           audio_mixer(num_cards)
 109 {
 110         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 111         check_error();
 112
 113         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 114         // will be halved when sampling them, and we need to compensate here.
 115         movit_texel_subpixel_precision /= 2.0;
 116
 117         resource_pool.reset(new ResourcePool);
 118         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 119         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 120                 output_channel[i].parent = this;
 121                 output_channel[i].channel = i;
 122         }
 123
 124         ImageFormat inout_format;
 125         inout_format.color_space = COLORSPACE_sRGB;
 126         inout_format.gamma_curve = GAMMA_sRGB;
 127
 128         // Display chain; shows the live output produced by the main chain (its RGBA version).
 129         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 130         check_error();
 131         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 132         display_chain->add_input(display_input);
 133         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 134         display_chain->set_dither_bits(0);  // Don't bother.
 135         display_chain->finalize();
 136
 137         video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd, global_disk_space_estimator));
 138
 139         // Start listening for clients only once VideoEncoder has written its header, if any.
 140         httpd.start(9095);
 141
 142         // First try initializing the then PCI devices, then USB, then
 143         // fill up with fake cards until we have the desired number of cards.
 144         unsigned num_pci_devices = 0;
 145         unsigned card_index = 0;
 146
 147         {
 148                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 149                 if (decklink_iterator != nullptr) {
 150                         for ( ; card_index < num_cards; ++card_index) {
 151                                 IDeckLink *decklink;
 152                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 153                                         break;
 154                                 }
 155
 156                                 configure_card(card_index, new DeckLinkCapture(decklink, card_index), /*is_fake_capture=*/false);
 157                                 ++num_pci_devices;
 158                         }
 159                         decklink_iterator->Release();
 160                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 161                 } else {
 162                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 163                 }
 164         }
 165         unsigned num_usb_devices = BMUSBCapture::num_cards();
 166         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 167                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 168                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 169                 configure_card(card_index, capture, /*is_fake_capture=*/false);
 170         }
 171         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 172
 173         unsigned num_fake_cards = 0;
 174         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 175                 FakeCapture *capture = new FakeCapture(WIDTH, HEIGHT, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 176                 configure_card(card_index, capture, /*is_fake_capture=*/true);
 177         }
 178
 179         if (num_fake_cards > 0) {
 180                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 181         }
 182
 183         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 184         BMUSBCapture::start_bm_thread();
 185
 186         for (card_index = 0; card_index < num_cards; ++card_index) {
 187                 cards[card_index].queue_length_policy.reset(card_index);
 188                 cards[card_index].capture->start_bm_capture();
 189         }
 190
 191         // Set up stuff for NV12 conversion.
 192
 193         // Cb/Cr shader.
 194         string cbcr_vert_shader =
 195                 "#version 130 \n"
 196                 " \n"
 197                 "in vec2 position; \n"
 198                 "in vec2 texcoord; \n"
 199                 "out vec2 tc0; \n"
 200                 "uniform vec2 foo_chroma_offset_0; \n"
 201                 " \n"
 202                 "void main() \n"
 203                 "{ \n"
 204                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 205                 "    // \n"
 206                 "    //   2.000  0.000  0.000 -1.000 \n"
 207                 "    //   0.000  2.000  0.000 -1.000 \n"
 208                 "    //   0.000  0.000 -2.000 -1.000 \n"
 209                 "    //   0.000  0.000  0.000  1.000 \n"
 210                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 211                 "    vec2 flipped_tc = texcoord; \n"
 212                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
 213                 "} \n";
 214         string cbcr_frag_shader =
 215                 "#version 130 \n"
 216                 "in vec2 tc0; \n"
 217                 "uniform sampler2D cbcr_tex; \n"
 218                 "out vec4 FragColor; \n"
 219                 "void main() { \n"
 220                 "    FragColor = texture(cbcr_tex, tc0); \n"
 221                 "} \n";
 222         vector<string> frag_shader_outputs;
 223         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 224
 225         float vertices[] = {
 226                 0.0f, 2.0f,
 227                 0.0f, 0.0f,
 228                 2.0f, 0.0f
 229         };
 230         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 231         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 232         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 233
 234         if (global_flags.enable_alsa_output) {
 235                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 236         }
 237 }
 238
 239 Mixer::~Mixer()
 240 {
 241         resource_pool->release_glsl_program(cbcr_program_num);
 242         glDeleteBuffers(1, &cbcr_vbo);
 243         BMUSBCapture::stop_bm_thread();
 244
 245         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 246                 {
 247                         unique_lock<mutex> lock(bmusb_mutex);
 248                         cards[card_index].should_quit = true;  // Unblock thread.
 249                         cards[card_index].new_frames_changed.notify_all();
 250                 }
 251                 cards[card_index].capture->stop_dequeue_thread();
 252         }
 253
 254         video_encoder.reset(nullptr);
 255 }
 256
 257 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture)
 258 {
 259         printf("Configuring card %d...\n", card_index);
 260
 261         CaptureCard *card = &cards[card_index];
 262         if (card->capture != nullptr) {
 263                 card->capture->stop_dequeue_thread();
 264                 delete card->capture;
 265         }
 266         card->capture = capture;
 267         card->is_fake_capture = is_fake_capture;
 268         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 269         if (card->frame_allocator == nullptr) {
 270                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 271         }
 272         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 273         if (card->surface == nullptr) {
 274                 card->surface = create_surface_with_same_format(mixer_surface);
 275         }
 276         while (!card->new_frames.empty()) card->new_frames.pop();
 277         card->fractional_samples = 0;
 278         card->last_timecode = -1;
 279         card->capture->configure_card();
 280
 281         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 282         audio_mixer.reset_resampler(device);
 283         audio_mixer.set_display_name(device, card->capture->get_description());
 284         audio_mixer.trigger_state_changed_callback();
 285 }
 286
 287
 288 namespace {
 289
 290 int unwrap_timecode(uint16_t current_wrapped, int last)
 291 {
 292         uint16_t last_wrapped = last & 0xffff;
 293         if (current_wrapped > last_wrapped) {
 294                 return (last & ~0xffff) | current_wrapped;
 295         } else {
 296                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 297         }
 298 }
 299
 300 }  // namespace
 301
 302 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 303                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 304                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 305 {
 306         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 307         CaptureCard *card = &cards[card_index];
 308
 309         if (is_mode_scanning[card_index]) {
 310                 if (video_format.has_signal) {
 311                         // Found a stable signal, so stop scanning.
 312                         is_mode_scanning[card_index] = false;
 313                 } else {
 314                         static constexpr double switch_time_s = 0.5;  // Should be enough time for the signal to stabilize.
 315                         steady_clock::time_point now = steady_clock::now();
 316                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 317                         if (sec_since_last_switch > switch_time_s) {
 318                                 // It isn't this mode; try the next one.
 319                                 mode_scanlist_index[card_index]++;
 320                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 321                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 322                                 last_mode_scan_change[card_index] = now;
 323                         }
 324                 }
 325         }
 326
 327         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 328         assert(frame_length > 0);
 329
 330         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;
 331         if (num_samples > OUTPUT_FREQUENCY / 10) {
 332                 printf("Card %d: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n",
 333                         card_index, int(audio_frame.len), int(audio_offset),
 334                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 335                 if (video_frame.owner) {
 336                         video_frame.owner->release_frame(video_frame);
 337                 }
 338                 if (audio_frame.owner) {
 339                         audio_frame.owner->release_frame(audio_frame);
 340                 }
 341                 return;
 342         }
 343
 344         int dropped_frames = 0;
 345         if (card->last_timecode != -1) {
 346                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 347         }
 348
 349         // Number of samples per frame if we need to insert silence.
 350         // (Could be nonintegral, but resampling will save us then.)
 351         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 352
 353         if (dropped_frames > MAX_FPS * 2) {
 354                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 355                         card_index, card->last_timecode, timecode);
 356                 audio_mixer.reset_resampler(device);
 357                 dropped_frames = 0;
 358         } else if (dropped_frames > 0) {
 359                 // Insert silence as needed.
 360                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 361                         card_index, dropped_frames, timecode);
 362
 363                 bool success;
 364                 do {
 365                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
 366                 } while (!success);
 367         }
 368
 369         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length);
 370
 371         // Done with the audio, so release it.
 372         if (audio_frame.owner) {
 373                 audio_frame.owner->release_frame(audio_frame);
 374         }
 375
 376         card->last_timecode = timecode;
 377
 378         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 379         if (video_frame.len - video_offset == 0 ||
 380             video_frame.len - video_offset != expected_length) {
 381                 if (video_frame.len != 0) {
 382                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 383                                 card_index, video_frame.len - video_offset, expected_length);
 384                 }
 385                 if (video_frame.owner) {
 386                         video_frame.owner->release_frame(video_frame);
 387                 }
 388
 389                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 390                 // so that pts can go up accordingly.
 391                 {
 392                         unique_lock<mutex> lock(bmusb_mutex);
 393                         CaptureCard::NewFrame new_frame;
 394                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 395                         new_frame.length = frame_length;
 396                         new_frame.interlaced = false;
 397                         new_frame.dropped_frames = dropped_frames;
 398                         card->new_frames.push(move(new_frame));
 399                         card->new_frames_changed.notify_all();
 400                 }
 401                 return;
 402         }
 403
 404         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 405
 406         unsigned num_fields = video_format.interlaced ? 2 : 1;
 407         steady_clock::time_point frame_upload_start;
 408         if (video_format.interlaced) {
 409                 // Send the two fields along as separate frames; the other side will need to add
 410                 // a deinterlacer to actually get this right.
 411                 assert(video_format.height % 2 == 0);
 412                 video_format.height /= 2;
 413                 assert(frame_length % 2 == 0);
 414                 frame_length /= 2;
 415                 num_fields = 2;
 416                 frame_upload_start = steady_clock::now();
 417         }
 418         userdata->last_interlaced = video_format.interlaced;
 419         userdata->last_has_signal = video_format.has_signal;
 420         userdata->last_is_connected = video_format.is_connected;
 421         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 422         userdata->last_frame_rate_den = video_format.frame_rate_den;
 423         RefCountedFrame frame(video_frame);
 424
 425         // Upload the textures.
 426         size_t cbcr_width = video_format.width / 2;
 427         size_t cbcr_offset = video_offset / 2;
 428         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 429
 430         for (unsigned field = 0; field < num_fields; ++field) {
 431                 // Put the actual texture upload in a lambda that is executed in the main thread.
 432                 // It is entirely possible to do this in the same thread (and it might even be
 433                 // faster, depending on the GPU and driver), but it appears to be trickling
 434                 // driver bugs very easily.
 435                 //
 436                 // Note that this means we must hold on to the actual frame data in <userdata>
 437                 // until the upload command is run, but we hold on to <frame> much longer than that
 438                 // (in fact, all the way until we no longer use the texture in rendering).
 439                 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() {
 440                         unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
 441
 442                         if (userdata->tex_y[field] == 0 ||
 443                             userdata->tex_cbcr[field] == 0 ||
 444                             video_format.width != userdata->last_width[field] ||
 445                             video_format.height != userdata->last_height[field]) {
 446                                 // We changed resolution since last use of this texture, so we need to create
 447                                 // a new object. Note that this each card has its own PBOFrameAllocator,
 448                                 // we don't need to worry about these flip-flopping between resolutions.
 449                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 450                                 check_error();
 451                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 452                                 check_error();
 453                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 454                                 check_error();
 455                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 456                                 check_error();
 457                                 userdata->last_width[field] = video_format.width;
 458                                 userdata->last_height[field] = video_format.height;
 459                         }
 460
 461                         GLuint pbo = userdata->pbo;
 462                         check_error();
 463                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
 464                         check_error();
 465
 466                         size_t field_y_start = y_offset + video_format.width * field_start_line;
 467                         size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 468
 469                         if (global_flags.flush_pbos) {
 470                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
 471                                 check_error();
 472                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
 473                                 check_error();
 474                         }
 475
 476                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 477                         check_error();
 478                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
 479                         check_error();
 480                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 481                         check_error();
 482                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
 483                         check_error();
 484                         glBindTexture(GL_TEXTURE_2D, 0);
 485                         check_error();
 486                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 487                         check_error();
 488                 };
 489
 490                 if (field == 1) {
 491                         // Don't upload the second field as fast as we can; wait until
 492                         // the field time has approximately passed. (Otherwise, we could
 493                         // get timing jitter against the other sources, and possibly also
 494                         // against the video display, although the latter is not as critical.)
 495                         // This requires our system clock to be reasonably close to the
 496                         // video clock, but that's not an unreasonable assumption.
 497                         steady_clock::time_point second_field_start = frame_upload_start +
 498                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 499                         this_thread::sleep_until(second_field_start);
 500                 }
 501
 502                 {
 503                         unique_lock<mutex> lock(bmusb_mutex);
 504                         CaptureCard::NewFrame new_frame;
 505                         new_frame.frame = frame;
 506                         new_frame.length = frame_length;
 507                         new_frame.field = field;
 508                         new_frame.interlaced = video_format.interlaced;
 509                         new_frame.upload_func = upload_func;
 510                         new_frame.dropped_frames = dropped_frames;
 511                         card->new_frames.push(move(new_frame));
 512                         card->new_frames_changed.notify_all();
 513                 }
 514         }
 515 }
 516
 517 void Mixer::bm_hotplug_add(libusb_device *dev)
 518 {
 519         lock_guard<mutex> lock(hotplug_mutex);
 520         hotplugged_cards.push_back(dev);
 521 }
 522
 523 void Mixer::bm_hotplug_remove(unsigned card_index)
 524 {
 525         cards[card_index].new_frames_changed.notify_all();
 526 }
 527
 528 void Mixer::thread_func()
 529 {
 530         eglBindAPI(EGL_OPENGL_API);
 531         QOpenGLContext *context = create_context(mixer_surface);
 532         if (!make_current(context, mixer_surface)) {
 533                 printf("oops\n");
 534                 exit(1);
 535         }
 536
 537         steady_clock::time_point start, now;
 538         start = steady_clock::now();
 539
 540         int frame = 0;
 541         int stats_dropped_frames = 0;
 542
 543         while (!should_quit) {
 544                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
 545                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 546                 int num_samples[MAX_VIDEO_CARDS] = { 0 };
 547
 548                 unsigned master_card_index = theme->map_signal(master_clock_channel);
 549                 assert(master_card_index < num_cards);
 550
 551                 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
 552                 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
 553                 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
 554
 555                 handle_hotplugged_cards();
 556
 557                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 558                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 559                                 continue;
 560                         }
 561                         if (new_frames[card_index].frame->len == 0) {
 562                                 ++new_frames[card_index].dropped_frames;
 563                         }
 564                         if (new_frames[card_index].dropped_frames > 0) {
 565                                 printf("Card %u dropped %d frames before this\n",
 566                                         card_index, int(new_frames[card_index].dropped_frames));
 567                         }
 568                 }
 569
 570                 // If the first card is reporting a corrupted or otherwise dropped frame,
 571                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 572                 if (new_frames[master_card_index].frame->len == 0) {
 573                         ++stats_dropped_frames;
 574                         pts_int += new_frames[master_card_index].length;
 575                         continue;
 576                 }
 577
 578                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 579                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 580                                 continue;
 581
 582                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 583                         assert(new_frame->frame != nullptr);
 584                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 585                         check_error();
 586
 587                         // The new texture might need uploading before use.
 588                         if (new_frame->upload_func) {
 589                                 new_frame->upload_func();
 590                                 new_frame->upload_func = nullptr;
 591                         }
 592                 }
 593
 594                 int64_t frame_duration = new_frames[master_card_index].length;
 595                 render_one_frame(frame_duration);
 596                 ++frame;
 597                 pts_int += frame_duration;
 598
 599                 now = steady_clock::now();
 600                 double elapsed = duration<double>(now - start).count();
 601                 if (frame % 100 == 0) {
 602                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 603                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 604                                 1e3 * elapsed / frame);
 605                 //      chain->print_phase_timing();
 606
 607                         // Check our memory usage, to see if we are close to our mlockall()
 608                         // limit (if at all set).
 609                         rusage used;
 610                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 611                                 perror("getrusage(RUSAGE_SELF)");
 612                                 assert(false);
 613                         }
 614
 615                         if (uses_mlock) {
 616                                 rlimit limit;
 617                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 618                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 619                                         assert(false);
 620                                 }
 621
 622                                 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 623                                         long(used.ru_maxrss / 1024),
 624                                         long(limit.rlim_cur / 1048576),
 625                                         float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 626                         } else {
 627                                 printf(", using %ld MB memory (not locked)",
 628                                         long(used.ru_maxrss / 1024));
 629                         }
 630
 631                         printf("\n");
 632                 }
 633
 634
 635                 if (should_cut.exchange(false)) {  // Test and clear.
 636                         video_encoder->do_cut(frame);
 637                 }
 638
 639 #if 0
 640                 // Reset every 100 frames, so that local variations in frame times
 641                 // (especially for the first few frames, when the shaders are
 642                 // compiled etc.) don't make it hard to measure for the entire
 643                 // remaining duration of the program.
 644                 if (frame == 10000) {
 645                         frame = 0;
 646                         start = now;
 647                 }
 648 #endif
 649                 check_error();
 650         }
 651
 652         resource_pool->clean_context();
 653 }
 654
 655 void Mixer::get_one_frame_from_each_card(unsigned master_card_index, CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS], bool has_new_frame[MAX_VIDEO_CARDS], int num_samples[MAX_VIDEO_CARDS])
 656 {
 657 start:
 658         // The first card is the master timer, so wait for it to have a new frame.
 659         // TODO: Add a timeout.
 660         unique_lock<mutex> lock(bmusb_mutex);
 661         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(); });
 662
 663         if (cards[master_card_index].new_frames.empty()) {
 664                 // We were woken up, but not due to a new frame. Deal with it
 665                 // and then restart.
 666                 assert(cards[master_card_index].capture->get_disconnected());
 667                 handle_hotplugged_cards();
 668                 goto start;
 669         }
 670
 671         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 672                 CaptureCard *card = &cards[card_index];
 673                 if (card->new_frames.empty()) {
 674                         assert(card_index != master_card_index);
 675                         card->queue_length_policy.update_policy(-1);
 676                         continue;
 677                 }
 678                 new_frames[card_index] = move(card->new_frames.front());
 679                 has_new_frame[card_index] = true;
 680                 card->new_frames.pop();
 681                 card->new_frames_changed.notify_all();
 682
 683                 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
 684                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 685                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 686                 assert(num_samples[card_index] >= 0);
 687
 688                 if (card_index == master_card_index) {
 689                         // We don't use the queue length policy for the master card,
 690                         // but we will if it stops being the master. Thus, clear out
 691                         // the policy in case we switch in the future.
 692                         card->queue_length_policy.reset(card_index);
 693                 } else {
 694                         // If we have excess frames compared to the policy for this card,
 695                         // drop frames from the head.
 696                         card->queue_length_policy.update_policy(card->new_frames.size());
 697                         while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
 698                                 card->new_frames.pop();
 699                         }
 700                 }
 701         }
 702 }
 703
 704 void Mixer::handle_hotplugged_cards()
 705 {
 706         // Check for cards that have been disconnected since last frame.
 707         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 708                 CaptureCard *card = &cards[card_index];
 709                 if (card->capture->get_disconnected()) {
 710                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 711                         FakeCapture *capture = new FakeCapture(WIDTH, HEIGHT, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 712                         configure_card(card_index, capture, /*is_fake_capture=*/true);
 713                         card->queue_length_policy.reset(card_index);
 714                         card->capture->start_bm_capture();
 715                 }
 716         }
 717
 718         // Check for cards that have been connected since last frame.
 719         vector<libusb_device *> hotplugged_cards_copy;
 720         {
 721                 lock_guard<mutex> lock(hotplug_mutex);
 722                 swap(hotplugged_cards, hotplugged_cards_copy);
 723         }
 724         for (libusb_device *new_dev : hotplugged_cards_copy) {
 725                 // Look for a fake capture card where we can stick this in.
 726                 int free_card_index = -1;
 727                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 728                         if (cards[card_index].is_fake_capture) {
 729                                 free_card_index = int(card_index);
 730                                 break;
 731                         }
 732                 }
 733
 734                 if (free_card_index == -1) {
 735                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
 736                         libusb_unref_device(new_dev);
 737                 } else {
 738                         // BMUSBCapture takes ownership.
 739                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
 740                         CaptureCard *card = &cards[free_card_index];
 741                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
 742                         configure_card(free_card_index, capture, /*is_fake_capture=*/false);
 743                         card->queue_length_policy.reset(free_card_index);
 744                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
 745                         capture->start_bm_capture();
 746                 }
 747         }
 748 }
 749
 750
 751 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
 752 {
 753         // Resample the audio as needed, including from previously dropped frames.
 754         assert(num_cards > 0);
 755         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 756                 const bool dropped_frame = (frame_num != dropped_frames);
 757                 {
 758                         // Signal to the audio thread to process this frame.
 759                         // Note that if the frame is a dropped frame, we signal that
 760                         // we don't want to use this frame as base for adjusting
 761                         // the resampler rate. The reason for this is that the timing
 762                         // of these frames is often way too late; they typically don't
 763                         // “arrive” before we synthesize them. Thus, we could end up
 764                         // in a situation where we have inserted e.g. five audio frames
 765                         // into the queue before we then start pulling five of them
 766                         // back out. This makes ResamplingQueue overestimate the delay,
 767                         // causing undue resampler changes. (We _do_ use the last,
 768                         // non-dropped frame; perhaps we should just discard that as well,
 769                         // since dropped frames are expected to be rare, and it might be
 770                         // better to just wait until we have a slightly more normal situation).
 771                         unique_lock<mutex> lock(audio_mutex);
 772                         bool adjust_rate = !dropped_frame;
 773                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate});
 774                         audio_task_queue_changed.notify_one();
 775                 }
 776                 if (dropped_frame) {
 777                         // For dropped frames, increase the pts. Note that if the format changed
 778                         // in the meantime, we have no way of detecting that; we just have to
 779                         // assume the frame length is always the same.
 780                         pts_int += length_per_frame;
 781                 }
 782         }
 783 }
 784
 785 void Mixer::render_one_frame(int64_t duration)
 786 {
 787         // Get the main chain from the theme, and set its state immediately.
 788         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 789         EffectChain *chain = theme_main_chain.chain;
 790         theme_main_chain.setup_chain();
 791         //theme_main_chain.chain->enable_phase_timing(true);
 792
 793         GLuint y_tex, cbcr_tex;
 794         bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
 795         assert(got_frame);
 796
 797         // Render main chain.
 798         GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 799         GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 800         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 801         check_error();
 802         chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 803         resource_pool->release_fbo(fbo);
 804
 805         subsample_chroma(cbcr_full_tex, cbcr_tex);
 806         resource_pool->release_2d_texture(cbcr_full_tex);
 807
 808         // Set the right state for rgba_tex.
 809         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 810         glBindTexture(GL_TEXTURE_2D, rgba_tex);
 811         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 812         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 813         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 814
 815         const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 816         RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
 817
 818         // The live frame just shows the RGBA texture we just rendered.
 819         // It owns rgba_tex now.
 820         DisplayFrame live_frame;
 821         live_frame.chain = display_chain.get();
 822         live_frame.setup_chain = [this, rgba_tex]{
 823                 display_input->set_texture_num(rgba_tex);
 824         };
 825         live_frame.ready_fence = fence;
 826         live_frame.input_frames = {};
 827         live_frame.temp_textures = { rgba_tex };
 828         output_channel[OUTPUT_LIVE].output_frame(live_frame);
 829
 830         // Set up preview and any additional channels.
 831         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 832                 DisplayFrame display_frame;
 833                 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 834                 display_frame.chain = chain.chain;
 835                 display_frame.setup_chain = chain.setup_chain;
 836                 display_frame.ready_fence = fence;
 837                 display_frame.input_frames = chain.input_frames;
 838                 display_frame.temp_textures = {};
 839                 output_channel[i].output_frame(display_frame);
 840         }
 841 }
 842
 843 void Mixer::audio_thread_func()
 844 {
 845         while (!should_quit) {
 846                 AudioTask task;
 847
 848                 {
 849                         unique_lock<mutex> lock(audio_mutex);
 850                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
 851                         if (should_quit) {
 852                                 return;
 853                         }
 854                         task = audio_task_queue.front();
 855                         audio_task_queue.pop();
 856                 }
 857
 858                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
 859                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
 860                 vector<float> samples_out = audio_mixer.get_output(double(task.pts_int) / TIMEBASE, task.num_samples, rate_adjustment_policy);
 861
 862                 // Send the samples to the sound card, then add them to the output.
 863                 if (alsa) {
 864                         alsa->write(samples_out);
 865                 }
 866                 video_encoder->add_audio(task.pts_int, move(samples_out));
 867         }
 868 }
 869
 870 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 871 {
 872         GLuint vao;
 873         glGenVertexArrays(1, &vao);
 874         check_error();
 875
 876         glBindVertexArray(vao);
 877         check_error();
 878
 879         // Extract Cb/Cr.
 880         GLuint fbo = resource_pool->create_fbo(dst_tex);
 881         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 882         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 883         check_error();
 884
 885         glUseProgram(cbcr_program_num);
 886         check_error();
 887
 888         glActiveTexture(GL_TEXTURE0);
 889         check_error();
 890         glBindTexture(GL_TEXTURE_2D, src_tex);
 891         check_error();
 892         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 893         check_error();
 894         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 895         check_error();
 896         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 897         check_error();
 898
 899         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 900         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 901
 902         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
 903         check_error();
 904
 905         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
 906                 glEnableVertexAttribArray(attr_index);
 907                 check_error();
 908                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 909                 check_error();
 910         }
 911
 912         glDrawArrays(GL_TRIANGLES, 0, 3);
 913         check_error();
 914
 915         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
 916                 glDisableVertexAttribArray(attr_index);
 917                 check_error();
 918         }
 919
 920         glUseProgram(0);
 921         check_error();
 922         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 923         check_error();
 924
 925         resource_pool->release_fbo(fbo);
 926         glDeleteVertexArrays(1, &vao);
 927 }
 928
 929 void Mixer::release_display_frame(DisplayFrame *frame)
 930 {
 931         for (GLuint texnum : frame->temp_textures) {
 932                 resource_pool->release_2d_texture(texnum);
 933         }
 934         frame->temp_textures.clear();
 935         frame->ready_fence.reset();
 936         frame->input_frames.clear();
 937 }
 938
 939 void Mixer::start()
 940 {
 941         mixer_thread = thread(&Mixer::thread_func, this);
 942         audio_thread = thread(&Mixer::audio_thread_func, this);
 943 }
 944
 945 void Mixer::quit()
 946 {
 947         should_quit = true;
 948         audio_task_queue_changed.notify_one();
 949         mixer_thread.join();
 950         audio_thread.join();
 951 }
 952
 953 void Mixer::transition_clicked(int transition_num)
 954 {
 955         theme->transition_clicked(transition_num, pts());
 956 }
 957
 958 void Mixer::channel_clicked(int preview_num)
 959 {
 960         theme->channel_clicked(preview_num);
 961 }
 962
 963 void Mixer::start_mode_scanning(unsigned card_index)
 964 {
 965         assert(card_index < num_cards);
 966         if (is_mode_scanning[card_index]) {
 967                 return;
 968         }
 969         is_mode_scanning[card_index] = true;
 970         mode_scanlist[card_index].clear();
 971         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
 972                 mode_scanlist[card_index].push_back(mode.first);
 973         }
 974         assert(!mode_scanlist[card_index].empty());
 975         mode_scanlist_index[card_index] = 0;
 976         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
 977         last_mode_scan_change[card_index] = steady_clock::now();
 978 }
 979
 980 Mixer::OutputChannel::~OutputChannel()
 981 {
 982         if (has_current_frame) {
 983                 parent->release_display_frame(&current_frame);
 984         }
 985         if (has_ready_frame) {
 986                 parent->release_display_frame(&ready_frame);
 987         }
 988 }
 989
 990 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 991 {
 992         // Store this frame for display. Remove the ready frame if any
 993         // (it was seemingly never used).
 994         {
 995                 unique_lock<mutex> lock(frame_mutex);
 996                 if (has_ready_frame) {
 997                         parent->release_display_frame(&ready_frame);
 998                 }
 999                 ready_frame = frame;
1000                 has_ready_frame = true;
1001         }
1002
1003         if (new_frame_ready_callback) {
1004                 new_frame_ready_callback();
1005         }
1006
1007         // Reduce the number of callbacks by filtering duplicates. The reason
1008         // why we bother doing this is that Qt seemingly can get into a state
1009         // where its builds up an essentially unbounded queue of signals,
1010         // consuming more and more memory, and there's no good way of collapsing
1011         // user-defined signals or limiting the length of the queue.
1012         if (transition_names_updated_callback) {
1013                 vector<string> transition_names = global_mixer->get_transition_names();
1014                 bool changed = false;
1015                 if (transition_names.size() != last_transition_names.size()) {
1016                         changed = true;
1017                 } else {
1018                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1019                                 if (transition_names[i] != last_transition_names[i]) {
1020                                         changed = true;
1021                                         break;
1022                                 }
1023                         }
1024                 }
1025                 if (changed) {
1026                         transition_names_updated_callback(transition_names);
1027                         last_transition_names = transition_names;
1028                 }
1029         }
1030         if (name_updated_callback) {
1031                 string name = global_mixer->get_channel_name(channel);
1032                 if (name != last_name) {
1033                         name_updated_callback(name);
1034                         last_name = name;
1035                 }
1036         }
1037         if (color_updated_callback) {
1038                 string color = global_mixer->get_channel_color(channel);
1039                 if (color != last_color) {
1040                         color_updated_callback(color);
1041                         last_color = color;
1042                 }
1043         }
1044 }
1045
1046 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1047 {
1048         unique_lock<mutex> lock(frame_mutex);
1049         if (!has_current_frame && !has_ready_frame) {
1050                 return false;
1051         }
1052
1053         if (has_current_frame && has_ready_frame) {
1054                 // We have a new ready frame. Toss the current one.
1055                 parent->release_display_frame(&current_frame);
1056                 has_current_frame = false;
1057         }
1058         if (has_ready_frame) {
1059                 assert(!has_current_frame);
1060                 current_frame = ready_frame;
1061                 ready_frame.ready_fence.reset();  // Drop the refcount.
1062                 ready_frame.input_frames.clear();  // Drop the refcounts.
1063                 has_current_frame = true;
1064                 has_ready_frame = false;
1065         }
1066
1067         *frame = current_frame;
1068         return true;
1069 }
1070
1071 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1072 {
1073         new_frame_ready_callback = callback;
1074 }
1075
1076 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1077 {
1078         transition_names_updated_callback = callback;
1079 }
1080
1081 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1082 {
1083         name_updated_callback = callback;
1084 }
1085
1086 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1087 {
1088         color_updated_callback = callback;
1089 }
1090
1091 mutex RefCountedGLsync::fence_lock;