git.sesse.net Git - nageru/blob - 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 <movit/util.h>
  14 #include <stdint.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <sys/time.h>
  18 #include <time.h>
  19 #include <algorithm>
  20 #include <cmath>
  21 #include <condition_variable>
  22 #include <cstddef>
  23 #include <memory>
  24 #include <mutex>
  25 #include <string>
  26 #include <thread>
  27 #include <utility>
  28 #include <vector>
  29 #include <arpa/inet.h>
  30
  31 #include "bmusb/bmusb.h"
  32 #include "context.h"
  33 #include "decklink_capture.h"
  34 #include "defs.h"
  35 #include "flags.h"
  36 #include "h264encode.h"
  37 #include "pbo_frame_allocator.h"
  38 #include "ref_counted_gl_sync.h"
  39 #include "timebase.h"
  40
  41 class QOpenGLContext;
  42
  43 using namespace movit;
  44 using namespace std;
  45 using namespace std::placeholders;
  46
  47 Mixer *global_mixer = nullptr;
  48
  49 namespace {
  50
  51 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  52 {
  53         for (size_t i = 0; i < num_samples; ++i) {
  54                 for (size_t j = 0; j < out_channels; ++j) {
  55                         uint32_t s1 = *src++;
  56                         uint32_t s2 = *src++;
  57                         uint32_t s3 = *src++;
  58                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  59                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  60                 }
  61                 src += 3 * (in_channels - out_channels);
  62         }
  63 }
  64
  65 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  66 {
  67         if (interlaced) {
  68                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  69                         input_state->buffered_frames[card_index][frame_num] =
  70                                 input_state->buffered_frames[card_index][frame_num - 1];
  71                 }
  72                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  73         } else {
  74                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  75                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  76                 }
  77         }
  78 }
  79
  80 string generate_local_dump_filename(int frame)
  81 {
  82         time_t now = time(NULL);
  83         tm now_tm;
  84         localtime_r(&now, &now_tm);
  85
  86         char timestamp[256];
  87         strftime(timestamp, sizeof(timestamp), "%F-%T%z", &now_tm);
  88
  89         // Use the frame number to disambiguate between two cuts starting
  90         // on the same second.
  91         char filename[256];
  92         snprintf(filename, sizeof(filename), "%s%s-f%02d%s",
  93                 LOCAL_DUMP_PREFIX, timestamp, frame % 100, LOCAL_DUMP_SUFFIX);
  94         return filename;
  95 }
  96
  97 }  // namespace
  98
  99 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 100         : httpd(WIDTH, HEIGHT),
 101           num_cards(num_cards),
 102           mixer_surface(create_surface(format)),
 103           h264_encoder_surface(create_surface(format)),
 104           correlation(OUTPUT_FREQUENCY),
 105           level_compressor(OUTPUT_FREQUENCY),
 106           limiter(OUTPUT_FREQUENCY),
 107           compressor(OUTPUT_FREQUENCY)
 108 {
 109         httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str());
 110         httpd.start(9095);
 111
 112         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 113         check_error();
 114
 115         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 116         // will be halved when sampling them, and we need to compensate here.
 117         movit_texel_subpixel_precision /= 2.0;
 118
 119         resource_pool.reset(new ResourcePool);
 120         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
 121         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 122                 output_channel[i].parent = this;
 123         }
 124
 125         ImageFormat inout_format;
 126         inout_format.color_space = COLORSPACE_sRGB;
 127         inout_format.gamma_curve = GAMMA_sRGB;
 128
 129         // Display chain; shows the live output produced by the main chain (its RGBA version).
 130         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 131         check_error();
 132         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 133         display_chain->add_input(display_input);
 134         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 135         display_chain->set_dither_bits(0);  // Don't bother.
 136         display_chain->finalize();
 137
 138         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 139
 140         // First try initializing the PCI devices, then USB, until we have the desired number of cards.
 141         unsigned num_pci_devices = 0, num_usb_devices = 0;
 142         unsigned card_index = 0;
 143
 144         IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 145         if (decklink_iterator != nullptr) {
 146                 for ( ; card_index < num_cards; ++card_index) {
 147                         IDeckLink *decklink;
 148                         if (decklink_iterator->Next(&decklink) != S_OK) {
 149                                 break;
 150                         }
 151
 152                         configure_card(card_index, format, new DeckLinkCapture(decklink, card_index));
 153                         ++num_pci_devices;
 154                 }
 155                 decklink_iterator->Release();
 156                 fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
 157         } else {
 158                 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 159         }
 160         for ( ; card_index < num_cards; ++card_index) {
 161                 configure_card(card_index, format, new BMUSBCapture(card_index - num_pci_devices));
 162                 ++num_usb_devices;
 163         }
 164
 165         if (num_usb_devices > 0) {
 166                 BMUSBCapture::start_bm_thread();
 167         }
 168
 169         for (card_index = 0; card_index < num_cards; ++card_index) {
 170                 cards[card_index].capture->start_bm_capture();
 171         }
 172
 173         // Set up stuff for NV12 conversion.
 174
 175         // Cb/Cr shader.
 176         string cbcr_vert_shader =
 177                 "#version 130 \n"
 178                 " \n"
 179                 "in vec2 position; \n"
 180                 "in vec2 texcoord; \n"
 181                 "out vec2 tc0; \n"
 182                 "uniform vec2 foo_chroma_offset_0; \n"
 183                 " \n"
 184                 "void main() \n"
 185                 "{ \n"
 186                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 187                 "    // \n"
 188                 "    //   2.000  0.000  0.000 -1.000 \n"
 189                 "    //   0.000  2.000  0.000 -1.000 \n"
 190                 "    //   0.000  0.000 -2.000 -1.000 \n"
 191                 "    //   0.000  0.000  0.000  1.000 \n"
 192                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 193                 "    vec2 flipped_tc = texcoord; \n"
 194                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
 195                 "} \n";
 196         string cbcr_frag_shader =
 197                 "#version 130 \n"
 198                 "in vec2 tc0; \n"
 199                 "uniform sampler2D cbcr_tex; \n"
 200                 "out vec4 FragColor; \n"
 201                 "void main() { \n"
 202                 "    FragColor = texture(cbcr_tex, tc0); \n"
 203                 "} \n";
 204         vector<string> frag_shader_outputs;
 205         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 206
 207         float vertices[] = {
 208                 0.0f, 2.0f,
 209                 0.0f, 0.0f,
 210                 2.0f, 0.0f
 211         };
 212         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 213         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 214         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 215
 216         r128.init(2, OUTPUT_FREQUENCY);
 217         r128.integr_start();
 218
 219         locut.init(FILTER_HPF, 2);
 220
 221         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 222         // and there's a limit to how important the peak meter is.
 223         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 224
 225         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 226 }
 227
 228 Mixer::~Mixer()
 229 {
 230         resource_pool->release_glsl_program(cbcr_program_num);
 231         glDeleteBuffers(1, &cbcr_vbo);
 232         BMUSBCapture::stop_bm_thread();
 233
 234         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 235                 {
 236                         unique_lock<mutex> lock(bmusb_mutex);
 237                         cards[card_index].should_quit = true;  // Unblock thread.
 238                         cards[card_index].new_data_ready_changed.notify_all();
 239                 }
 240                 cards[card_index].capture->stop_dequeue_thread();
 241         }
 242
 243         h264_encoder.reset(nullptr);
 244 }
 245
 246 void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture)
 247 {
 248         printf("Configuring card %d...\n", card_index);
 249
 250         CaptureCard *card = &cards[card_index];
 251         card->capture = capture;
 252         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 253         card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 254         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 255         card->surface = create_surface(format);
 256         card->capture->set_dequeue_thread_callbacks(
 257                 [card]{
 258                         eglBindAPI(EGL_OPENGL_API);
 259                         card->context = create_context(card->surface);
 260                         if (!make_current(card->context, card->surface)) {
 261                                 printf("failed to create bmusb context\n");
 262                                 exit(1);
 263                         }
 264                 },
 265                 [this]{
 266                         resource_pool->clean_context();
 267                 });
 268         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 269         card->capture->configure_card();
 270 }
 271
 272
 273 namespace {
 274
 275 int unwrap_timecode(uint16_t current_wrapped, int last)
 276 {
 277         uint16_t last_wrapped = last & 0xffff;
 278         if (current_wrapped > last_wrapped) {
 279                 return (last & ~0xffff) | current_wrapped;
 280         } else {
 281                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 282         }
 283 }
 284
 285 float find_peak(const float *samples, size_t num_samples)
 286 {
 287         float m = fabs(samples[0]);
 288         for (size_t i = 1; i < num_samples; ++i) {
 289                 m = max(m, fabs(samples[i]));
 290         }
 291         return m;
 292 }
 293
 294 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 295 {
 296         size_t num_samples = in.size() / 2;
 297         out_l->resize(num_samples);
 298         out_r->resize(num_samples);
 299
 300         const float *inptr = in.data();
 301         float *lptr = &(*out_l)[0];
 302         float *rptr = &(*out_r)[0];
 303         for (size_t i = 0; i < num_samples; ++i) {
 304                 *lptr++ = *inptr++;
 305                 *rptr++ = *inptr++;
 306         }
 307 }
 308
 309 }  // namespace
 310
 311 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 312                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 313                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 314 {
 315         CaptureCard *card = &cards[card_index];
 316
 317         int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom;
 318
 319         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 320         if (num_samples > OUTPUT_FREQUENCY / 10) {
 321                 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",
 322                         card_index, int(audio_frame.len), int(audio_offset),
 323                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 324                 if (video_frame.owner) {
 325                         video_frame.owner->release_frame(video_frame);
 326                 }
 327                 if (audio_frame.owner) {
 328                         audio_frame.owner->release_frame(audio_frame);
 329                 }
 330                 return;
 331         }
 332
 333         int64_t local_pts = card->next_local_pts;
 334         int dropped_frames = 0;
 335         if (card->last_timecode != -1) {
 336                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 337         }
 338
 339         // Convert the audio to stereo fp32 and add it.
 340         vector<float> audio;
 341         audio.resize(num_samples * 2);
 342         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 343
 344         // Add the audio.
 345         {
 346                 unique_lock<mutex> lock(card->audio_mutex);
 347
 348                 // Number of samples per frame if we need to insert silence.
 349                 // (Could be nonintegral, but resampling will save us then.)
 350                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 351
 352                 if (dropped_frames > MAX_FPS * 2) {
 353                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 354                                 card_index, card->last_timecode, timecode);
 355                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 356                         dropped_frames = 0;
 357                 } else if (dropped_frames > 0) {
 358                         // Insert silence as needed.
 359                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 360                                 card_index, dropped_frames, timecode);
 361                         vector<float> silence(silence_samples * 2, 0.0f);
 362                         for (int i = 0; i < dropped_frames; ++i) {
 363                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 364                                 // Note that if the format changed in the meantime, we have
 365                                 // no way of detecting that; we just have to assume the frame length
 366                                 // is always the same.
 367                                 local_pts += frame_length;
 368                         }
 369                 }
 370                 if (num_samples == 0) {
 371                         audio.resize(silence_samples * 2);
 372                         num_samples = silence_samples;
 373                 }
 374                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 375                 card->next_local_pts = local_pts + frame_length;
 376         }
 377
 378         card->last_timecode = timecode;
 379
 380         // Done with the audio, so release it.
 381         if (audio_frame.owner) {
 382                 audio_frame.owner->release_frame(audio_frame);
 383         }
 384
 385         {
 386                 // Wait until the previous frame was consumed.
 387                 unique_lock<mutex> lock(bmusb_mutex);
 388                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 389                 if (card->should_quit) return;
 390         }
 391
 392         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 393         if (video_frame.len - video_offset == 0 ||
 394             video_frame.len - video_offset != expected_length) {
 395                 if (video_frame.len != 0) {
 396                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 397                                 card_index, video_frame.len - video_offset, expected_length);
 398                 }
 399                 if (video_frame.owner) {
 400                         video_frame.owner->release_frame(video_frame);
 401                 }
 402
 403                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 404                 // so that pts can go up accordingly.
 405                 {
 406                         unique_lock<mutex> lock(bmusb_mutex);
 407                         card->new_data_ready = true;
 408                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 409                         card->new_frame_length = frame_length;
 410                         card->new_frame_interlaced = false;
 411                         card->new_data_ready_fence = nullptr;
 412                         card->dropped_frames = dropped_frames;
 413                         card->new_data_ready_changed.notify_all();
 414                 }
 415                 return;
 416         }
 417
 418         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 419
 420         unsigned num_fields = video_format.interlaced ? 2 : 1;
 421         timespec frame_upload_start;
 422         if (video_format.interlaced) {
 423                 // Send the two fields along as separate frames; the other side will need to add
 424                 // a deinterlacer to actually get this right.
 425                 assert(video_format.height % 2 == 0);
 426                 video_format.height /= 2;
 427                 assert(frame_length % 2 == 0);
 428                 frame_length /= 2;
 429                 num_fields = 2;
 430                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 431         }
 432         userdata->last_interlaced = video_format.interlaced;
 433         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 434         userdata->last_frame_rate_den = video_format.frame_rate_den;
 435         RefCountedFrame new_frame(video_frame);
 436
 437         // Upload the textures.
 438         size_t cbcr_width = video_format.width / 2;
 439         size_t cbcr_offset = video_offset / 2;
 440         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 441
 442         for (unsigned field = 0; field < num_fields; ++field) {
 443                 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
 444
 445                 if (userdata->tex_y[field] == 0 ||
 446                     userdata->tex_cbcr[field] == 0 ||
 447                     video_format.width != userdata->last_width[field] ||
 448                     video_format.height != userdata->last_height[field]) {
 449                         // We changed resolution since last use of this texture, so we need to create
 450                         // a new object. Note that this each card has its own PBOFrameAllocator,
 451                         // we don't need to worry about these flip-flopping between resolutions.
 452                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 453                         check_error();
 454                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 455                         check_error();
 456                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 457                         check_error();
 458                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 459                         check_error();
 460                         userdata->last_width[field] = video_format.width;
 461                         userdata->last_height[field] = video_format.height;
 462                 }
 463
 464                 GLuint pbo = userdata->pbo;
 465                 check_error();
 466                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 467                 check_error();
 468                 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 469                 check_error();
 470
 471                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 472                 check_error();
 473                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t)));
 474                 check_error();
 475                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 476                 check_error();
 477                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + video_format.width * field_start_line));
 478                 check_error();
 479                 glBindTexture(GL_TEXTURE_2D, 0);
 480                 check_error();
 481                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 482                 check_error();
 483                 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 484                 check_error();
 485                 assert(fence != nullptr);
 486
 487                 if (field == 1) {
 488                         // Don't upload the second field as fast as we can; wait until
 489                         // the field time has approximately passed. (Otherwise, we could
 490                         // get timing jitter against the other sources, and possibly also
 491                         // against the video display, although the latter is not as critical.)
 492                         // This requires our system clock to be reasonably close to the
 493                         // video clock, but that's not an unreasonable assumption.
 494                         timespec second_field_start;
 495                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 496                                 frame_length * 1000000000 / TIMEBASE;
 497                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 498                                 second_field_start.tv_nsec / 1000000000;
 499                         second_field_start.tv_nsec %= 1000000000;
 500
 501                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 502                                                &second_field_start, nullptr) == -1 &&
 503                                errno == EINTR) ;
 504                 }
 505
 506                 {
 507                         unique_lock<mutex> lock(bmusb_mutex);
 508                         card->new_data_ready = true;
 509                         card->new_frame = new_frame;
 510                         card->new_frame_length = frame_length;
 511                         card->new_frame_field = field;
 512                         card->new_frame_interlaced = video_format.interlaced;
 513                         card->new_data_ready_fence = fence;
 514                         card->dropped_frames = dropped_frames;
 515                         card->new_data_ready_changed.notify_all();
 516
 517                         if (field != num_fields - 1) {
 518                                 // Wait until the previous frame was consumed.
 519                                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 520                                 if (card->should_quit) return;
 521                         }
 522                 }
 523         }
 524 }
 525
 526 void Mixer::thread_func()
 527 {
 528         eglBindAPI(EGL_OPENGL_API);
 529         QOpenGLContext *context = create_context(mixer_surface);
 530         if (!make_current(context, mixer_surface)) {
 531                 printf("oops\n");
 532                 exit(1);
 533         }
 534
 535         struct timespec start, now;
 536         clock_gettime(CLOCK_MONOTONIC, &start);
 537
 538         int frame = 0;
 539         int stats_dropped_frames = 0;
 540
 541         while (!should_quit) {
 542                 CaptureCard card_copy[MAX_CARDS];
 543                 int num_samples[MAX_CARDS];
 544
 545                 {
 546                         unique_lock<mutex> lock(bmusb_mutex);
 547
 548                         // The first card is the master timer, so wait for it to have a new frame.
 549                         // TODO: Make configurable, and with a timeout.
 550                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 551
 552                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 553                                 CaptureCard *card = &cards[card_index];
 554                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 555                                 card_copy[card_index].new_frame = card->new_frame;
 556                                 card_copy[card_index].new_frame_length = card->new_frame_length;
 557                                 card_copy[card_index].new_frame_field = card->new_frame_field;
 558                                 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
 559                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 560                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 561                                 card->new_data_ready = false;
 562                                 card->new_data_ready_changed.notify_all();
 563
 564                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
 565                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 566                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 567                                 assert(num_samples[card_index] >= 0);
 568                         }
 569                 }
 570
 571                 // Resample the audio as needed, including from previously dropped frames.
 572                 assert(num_cards > 0);
 573                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 574                         {
 575                                 // Signal to the audio thread to process this frame.
 576                                 unique_lock<mutex> lock(audio_mutex);
 577                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
 578                                 audio_task_queue_changed.notify_one();
 579                         }
 580                         if (frame_num != card_copy[0].dropped_frames) {
 581                                 // For dropped frames, increase the pts. Note that if the format changed
 582                                 // in the meantime, we have no way of detecting that; we just have to
 583                                 // assume the frame length is always the same.
 584                                 ++stats_dropped_frames;
 585                                 pts_int += card_copy[0].new_frame_length;
 586                         }
 587                 }
 588
 589                 if (audio_level_callback != nullptr) {
 590                         unique_lock<mutex> lock(compressor_mutex);
 591                         double loudness_s = r128.loudness_S();
 592                         double loudness_i = r128.integrated();
 593                         double loudness_range_low = r128.range_min();
 594                         double loudness_range_high = r128.range_max();
 595
 596                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 597                                              loudness_i, loudness_range_low, loudness_range_high,
 598                                              gain_staging_db, 20.0 * log10(final_makeup_gain),
 599                                              correlation.get_correlation());
 600                 }
 601
 602                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 603                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 604                                 ++card_copy[card_index].dropped_frames;
 605                         }
 606                         if (card_copy[card_index].dropped_frames > 0) {
 607                                 printf("Card %u dropped %d frames before this\n",
 608                                         card_index, int(card_copy[card_index].dropped_frames));
 609                         }
 610                 }
 611
 612                 // If the first card is reporting a corrupted or otherwise dropped frame,
 613                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 614                 if (card_copy[0].new_frame->len == 0) {
 615                         ++stats_dropped_frames;
 616                         pts_int += card_copy[0].new_frame_length;
 617                         continue;
 618                 }
 619
 620                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 621                         CaptureCard *card = &card_copy[card_index];
 622                         if (!card->new_data_ready || card->new_frame->len == 0)
 623                                 continue;
 624
 625                         assert(card->new_frame != nullptr);
 626                         insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
 627                         check_error();
 628
 629                         // The new texture might still be uploaded,
 630                         // tell the GPU to wait until it's there.
 631                         if (card->new_data_ready_fence) {
 632                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 633                                 check_error();
 634                                 glDeleteSync(card->new_data_ready_fence);
 635                                 check_error();
 636                         }
 637                 }
 638
 639                 // Get the main chain from the theme, and set its state immediately.
 640                 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 641                 EffectChain *chain = theme_main_chain.chain;
 642                 theme_main_chain.setup_chain();
 643                 //theme_main_chain.chain->enable_phase_timing(true);
 644
 645                 GLuint y_tex, cbcr_tex;
 646                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 647                 assert(got_frame);
 648
 649                 // Render main chain.
 650                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 651                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 652                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 653                 check_error();
 654                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 655                 resource_pool->release_fbo(fbo);
 656
 657                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 658                 resource_pool->release_2d_texture(cbcr_full_tex);
 659
 660                 // Set the right state for rgba_tex.
 661                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 662                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 663                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 664                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 665                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 666
 667                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 668                 check_error();
 669
 670                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 671                 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
 672                 ++frame;
 673                 pts_int += card_copy[0].new_frame_length;
 674
 675                 // The live frame just shows the RGBA texture we just rendered.
 676                 // It owns rgba_tex now.
 677                 DisplayFrame live_frame;
 678                 live_frame.chain = display_chain.get();
 679                 live_frame.setup_chain = [this, rgba_tex]{
 680                         display_input->set_texture_num(rgba_tex);
 681                 };
 682                 live_frame.ready_fence = fence;
 683                 live_frame.input_frames = {};
 684                 live_frame.temp_textures = { rgba_tex };
 685                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 686
 687                 // Set up preview and any additional channels.
 688                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 689                         DisplayFrame display_frame;
 690                         Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 691                         display_frame.chain = chain.chain;
 692                         display_frame.setup_chain = chain.setup_chain;
 693                         display_frame.ready_fence = fence;
 694                         display_frame.input_frames = chain.input_frames;
 695                         display_frame.temp_textures = {};
 696                         output_channel[i].output_frame(display_frame);
 697                 }
 698
 699                 clock_gettime(CLOCK_MONOTONIC, &now);
 700                 double elapsed = now.tv_sec - start.tv_sec +
 701                         1e-9 * (now.tv_nsec - start.tv_nsec);
 702                 if (frame % 100 == 0) {
 703                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 704                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 705                                 1e3 * elapsed / frame);
 706                 //      chain->print_phase_timing();
 707                 }
 708
 709                 if (should_cut.exchange(false)) {  // Test and clear.
 710                         string filename = generate_local_dump_filename(frame);
 711                         printf("Starting new recording: %s\n", filename.c_str());
 712                         h264_encoder->shutdown();
 713                         httpd.close_output_file();
 714                         httpd.open_output_file(filename.c_str());
 715                         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 716                 }
 717
 718 #if 0
 719                 // Reset every 100 frames, so that local variations in frame times
 720                 // (especially for the first few frames, when the shaders are
 721                 // compiled etc.) don't make it hard to measure for the entire
 722                 // remaining duration of the program.
 723                 if (frame == 10000) {
 724                         frame = 0;
 725                         start = now;
 726                 }
 727 #endif
 728                 check_error();
 729         }
 730
 731         resource_pool->clean_context();
 732 }
 733
 734 void Mixer::audio_thread_func()
 735 {
 736         while (!should_quit) {
 737                 AudioTask task;
 738
 739                 {
 740                         unique_lock<mutex> lock(audio_mutex);
 741                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 742                         task = audio_task_queue.front();
 743                         audio_task_queue.pop();
 744                 }
 745
 746                 process_audio_one_frame(task.pts_int, task.num_samples);
 747         }
 748 }
 749
 750 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 751 {
 752         vector<float> samples_card;
 753         vector<float> samples_out;
 754
 755         // TODO: Allow mixing audio from several sources.
 756         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 757         assert(selected_audio_card < num_cards);
 758
 759         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 760                 samples_card.resize(num_samples * 2);
 761                 {
 762                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 763                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 764                                 printf("Card %d reported previous underrun.\n", card_index);
 765                         }
 766                 }
 767                 if (card_index == selected_audio_card) {
 768                         samples_out = move(samples_card);
 769                 }
 770         }
 771
 772         // Cut away everything under 120 Hz (or whatever the cutoff is);
 773         // we don't need it for voice, and it will reduce headroom
 774         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 775         // should be dampened.)
 776         if (locut_enabled) {
 777                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 778         }
 779
 780         // Apply a level compressor to get the general level right.
 781         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 782         // (or more precisely, near it, since we don't use infinite ratio),
 783         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 784         // entirely arbitrary, but from practical tests with speech, it seems to
 785         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 786         {
 787                 unique_lock<mutex> lock(compressor_mutex);
 788                 if (level_compressor_enabled) {
 789                         float threshold = 0.01f;   // -40 dBFS.
 790                         float ratio = 20.0f;
 791                         float attack_time = 0.5f;
 792                         float release_time = 20.0f;
 793                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 794                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 795                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 796                 } else {
 797                         // Just apply the gain we already had.
 798                         float g = pow(10.0f, gain_staging_db / 20.0f);
 799                         for (size_t i = 0; i < samples_out.size(); ++i) {
 800                                 samples_out[i] *= g;
 801                         }
 802                 }
 803         }
 804
 805 #if 0
 806         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 807                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 808                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 809                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 810 #endif
 811
 812 //      float limiter_att, compressor_att;
 813
 814         // The real compressor.
 815         if (compressor_enabled) {
 816                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 817                 float ratio = 20.0f;
 818                 float attack_time = 0.005f;
 819                 float release_time = 0.040f;
 820                 float makeup_gain = 2.0f;  // +6 dB.
 821                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 822 //              compressor_att = compressor.get_attenuation();
 823         }
 824
 825         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 826         // Note that since ratio is not infinite, we could go slightly higher than this.
 827         if (limiter_enabled) {
 828                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 829                 float ratio = 30.0f;
 830                 float attack_time = 0.0f;  // Instant.
 831                 float release_time = 0.020f;
 832                 float makeup_gain = 1.0f;  // 0 dB.
 833                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 834 //              limiter_att = limiter.get_attenuation();
 835         }
 836
 837 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 838
 839         // Upsample 4x to find interpolated peak.
 840         peak_resampler.inp_data = samples_out.data();
 841         peak_resampler.inp_count = samples_out.size() / 2;
 842
 843         vector<float> interpolated_samples_out;
 844         interpolated_samples_out.resize(samples_out.size());
 845         while (peak_resampler.inp_count > 0) {  // About four iterations.
 846                 peak_resampler.out_data = &interpolated_samples_out[0];
 847                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 848                 peak_resampler.process();
 849                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 850                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 851                 peak_resampler.out_data = nullptr;
 852         }
 853
 854         // At this point, we are most likely close to +0 LU, but all of our
 855         // measurements have been on raw sample values, not R128 values.
 856         // So we have a final makeup gain to get us to +0 LU; the gain
 857         // adjustments required should be relatively small, and also, the
 858         // offset shouldn't change much (only if the type of audio changes
 859         // significantly). Thus, we shoot for updating this value basically
 860         // “whenever we process buffers”, since the R128 calculation isn't exactly
 861         // something we get out per-sample.
 862         //
 863         // Note that there's a feedback loop here, so we choose a very slow filter
 864         // (half-time of 100 seconds).
 865         double target_loudness_factor, alpha;
 866         {
 867                 unique_lock<mutex> lock(compressor_mutex);
 868                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
 869                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
 870                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
 871
 872                 // If we're outside +/- 5 LU uncorrected, we don't count it as
 873                 // a normal signal (probably silence) and don't change the
 874                 // correction factor; just apply what we already have.
 875                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 876                         alpha = 0.0;
 877                 } else {
 878                         // Formula adapted from
 879                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 880                         const double half_time_s = 100.0;
 881                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 882                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 883                 }
 884
 885                 double m = final_makeup_gain;
 886                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 887                         samples_out[i + 0] *= m;
 888                         samples_out[i + 1] *= m;
 889                         m += (target_loudness_factor - m) * alpha;
 890                 }
 891                 final_makeup_gain = m;
 892         }
 893
 894         // Find R128 levels and L/R correlation.
 895         vector<float> left, right;
 896         deinterleave_samples(samples_out, &left, &right);
 897         float *ptrs[] = { left.data(), right.data() };
 898         {
 899                 unique_lock<mutex> lock(compressor_mutex);
 900                 r128.process(left.size(), ptrs);
 901                 correlation.process_samples(samples_out);
 902         }
 903
 904         // Send the samples to the sound card.
 905         if (alsa) {
 906                 alsa->write(samples_out);
 907         }
 908
 909         // And finally add them to the output.
 910         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 911 }
 912
 913 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 914 {
 915         GLuint vao;
 916         glGenVertexArrays(1, &vao);
 917         check_error();
 918
 919         glBindVertexArray(vao);
 920         check_error();
 921
 922         // Extract Cb/Cr.
 923         GLuint fbo = resource_pool->create_fbo(dst_tex);
 924         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 925         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 926         check_error();
 927
 928         glUseProgram(cbcr_program_num);
 929         check_error();
 930
 931         glActiveTexture(GL_TEXTURE0);
 932         check_error();
 933         glBindTexture(GL_TEXTURE_2D, src_tex);
 934         check_error();
 935         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 936         check_error();
 937         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 938         check_error();
 939         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 940         check_error();
 941
 942         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 943         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 944
 945         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
 946         check_error();
 947
 948         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
 949                 glEnableVertexAttribArray(attr_index);
 950                 check_error();
 951                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 952                 check_error();
 953         }
 954
 955         glDrawArrays(GL_TRIANGLES, 0, 3);
 956         check_error();
 957
 958         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
 959                 glDisableVertexAttribArray(attr_index);
 960                 check_error();
 961         }
 962
 963         glUseProgram(0);
 964         check_error();
 965         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 966         check_error();
 967
 968         resource_pool->release_fbo(fbo);
 969         glDeleteVertexArrays(1, &vao);
 970 }
 971
 972 void Mixer::release_display_frame(DisplayFrame *frame)
 973 {
 974         for (GLuint texnum : frame->temp_textures) {
 975                 resource_pool->release_2d_texture(texnum);
 976         }
 977         frame->temp_textures.clear();
 978         frame->ready_fence.reset();
 979         frame->input_frames.clear();
 980 }
 981
 982 void Mixer::start()
 983 {
 984         mixer_thread = thread(&Mixer::thread_func, this);
 985         audio_thread = thread(&Mixer::audio_thread_func, this);
 986 }
 987
 988 void Mixer::quit()
 989 {
 990         should_quit = true;
 991         mixer_thread.join();
 992         audio_thread.join();
 993 }
 994
 995 void Mixer::transition_clicked(int transition_num)
 996 {
 997         theme->transition_clicked(transition_num, pts());
 998 }
 999
1000 void Mixer::channel_clicked(int preview_num)
1001 {
1002         theme->channel_clicked(preview_num);
1003 }
1004
1005 void Mixer::reset_meters()
1006 {
1007         peak_resampler.reset();
1008         peak = 0.0f;
1009         r128.reset();
1010         r128.integr_start();
1011         correlation.reset();
1012 }
1013
1014 Mixer::OutputChannel::~OutputChannel()
1015 {
1016         if (has_current_frame) {
1017                 parent->release_display_frame(&current_frame);
1018         }
1019         if (has_ready_frame) {
1020                 parent->release_display_frame(&ready_frame);
1021         }
1022 }
1023
1024 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1025 {
1026         // Store this frame for display. Remove the ready frame if any
1027         // (it was seemingly never used).
1028         {
1029                 unique_lock<mutex> lock(frame_mutex);
1030                 if (has_ready_frame) {
1031                         parent->release_display_frame(&ready_frame);
1032                 }
1033                 ready_frame = frame;
1034                 has_ready_frame = true;
1035         }
1036
1037         if (has_new_frame_ready_callback) {
1038                 new_frame_ready_callback();
1039         }
1040 }
1041
1042 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1043 {
1044         unique_lock<mutex> lock(frame_mutex);
1045         if (!has_current_frame && !has_ready_frame) {
1046                 return false;
1047         }
1048
1049         if (has_current_frame && has_ready_frame) {
1050                 // We have a new ready frame. Toss the current one.
1051                 parent->release_display_frame(&current_frame);
1052                 has_current_frame = false;
1053         }
1054         if (has_ready_frame) {
1055                 assert(!has_current_frame);
1056                 current_frame = ready_frame;
1057                 ready_frame.ready_fence.reset();  // Drop the refcount.
1058                 ready_frame.input_frames.clear();  // Drop the refcounts.
1059                 has_current_frame = true;
1060                 has_ready_frame = false;
1061         }
1062
1063         *frame = current_frame;
1064         return true;
1065 }
1066
1067 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1068 {
1069         new_frame_ready_callback = callback;
1070         has_new_frame_ready_callback = true;
1071 }