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 <init.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/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
  30 #include "bmusb/bmusb.h"
  31 #include "context.h"
  32 #include "defs.h"
  33 #include "h264encode.h"
  34 #include "pbo_frame_allocator.h"
  35 #include "ref_counted_gl_sync.h"
  36 #include "timebase.h"
  37
  38 class QOpenGLContext;
  39
  40 using namespace movit;
  41 using namespace std;
  42 using namespace std::placeholders;
  43
  44 Mixer *global_mixer = nullptr;
  45
  46 namespace {
  47
  48 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  49 {
  50         for (size_t i = 0; i < num_samples; ++i) {
  51                 for (size_t j = 0; j < out_channels; ++j) {
  52                         uint32_t s1 = *src++;
  53                         uint32_t s2 = *src++;
  54                         uint32_t s3 = *src++;
  55                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  56                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  57                 }
  58                 src += 3 * (in_channels - out_channels);
  59         }
  60 }
  61
  62 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  63 {
  64         if (interlaced) {
  65                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  66                         input_state->buffered_frames[card_index][frame_num] =
  67                                 input_state->buffered_frames[card_index][frame_num - 1];
  68                 }
  69                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  70         } else {
  71                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  72                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  73                 }
  74         }
  75 }
  76
  77 string generate_local_dump_filename(int frame)
  78 {
  79         time_t now = time(NULL);
  80         tm now_tm;
  81         localtime_r(&now, &now_tm);
  82
  83         char timestamp[256];
  84         strftime(timestamp, sizeof(timestamp), "%F-%T%z", &now_tm);
  85
  86         // Use the frame number to disambiguate between two cuts starting
  87         // on the same second.
  88         char filename[256];
  89         snprintf(filename, sizeof(filename), "%s%s-f%02d%s",
  90                 LOCAL_DUMP_PREFIX, timestamp, frame % 100, LOCAL_DUMP_SUFFIX);
  91         return filename;
  92 }
  93
  94 }  // namespace
  95
  96 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
  97         : httpd(WIDTH, HEIGHT),
  98           num_cards(num_cards),
  99           mixer_surface(create_surface(format)),
 100           h264_encoder_surface(create_surface(format)),
 101           correlation(OUTPUT_FREQUENCY),
 102           level_compressor(OUTPUT_FREQUENCY),
 103           limiter(OUTPUT_FREQUENCY),
 104           compressor(OUTPUT_FREQUENCY)
 105 {
 106         httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str());
 107         httpd.start(9095);
 108
 109         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 110         check_error();
 111
 112         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 113         // will be halved when sampling them, and we need to compensate here.
 114         movit_texel_subpixel_precision /= 2.0;
 115
 116         resource_pool.reset(new ResourcePool);
 117         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
 118         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 119                 output_channel[i].parent = this;
 120         }
 121
 122         ImageFormat inout_format;
 123         inout_format.color_space = COLORSPACE_sRGB;
 124         inout_format.gamma_curve = GAMMA_sRGB;
 125
 126         // Display chain; shows the live output produced by the main chain (its RGBA version).
 127         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 128         check_error();
 129         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 130         display_chain->add_input(display_input);
 131         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 132         display_chain->set_dither_bits(0);  // Don't bother.
 133         display_chain->finalize();
 134
 135         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
 136
 137         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 138                 printf("Configuring card %d...\n", card_index);
 139                 CaptureCard *card = &cards[card_index];
 140                 card->usb = new BMUSBCapture(card_index);
 141                 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 142                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 143                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
 144                 card->surface = create_surface(format);
 145                 card->usb->set_dequeue_thread_callbacks(
 146                         [card]{
 147                                 eglBindAPI(EGL_OPENGL_API);
 148                                 card->context = create_context(card->surface);
 149                                 if (!make_current(card->context, card->surface)) {
 150                                         printf("failed to create bmusb context\n");
 151                                         exit(1);
 152                                 }
 153                         },
 154                         [this]{
 155                                 resource_pool->clean_context();
 156                         });
 157                 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 158                 card->usb->configure_card();
 159         }
 160
 161         BMUSBCapture::start_bm_thread();
 162
 163         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 164                 cards[card_index].usb->start_bm_capture();
 165         }
 166
 167         // Set up stuff for NV12 conversion.
 168
 169         // Cb/Cr shader.
 170         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 171         string cbcr_frag_shader =
 172                 "#version 130 \n"
 173                 "in vec2 tc0; \n"
 174                 "uniform sampler2D cbcr_tex; \n"
 175                 "void main() { \n"
 176                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 177                 "} \n";
 178         vector<string> frag_shader_outputs;
 179         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 180
 181         r128.init(2, OUTPUT_FREQUENCY);
 182         r128.integr_start();
 183
 184         locut.init(FILTER_HPF, 2);
 185
 186         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 187         // and there's a limit to how important the peak meter is.
 188         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 189
 190         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 191 }
 192
 193 Mixer::~Mixer()
 194 {
 195         resource_pool->release_glsl_program(cbcr_program_num);
 196         BMUSBCapture::stop_bm_thread();
 197
 198         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 199                 {
 200                         unique_lock<mutex> lock(bmusb_mutex);
 201                         cards[card_index].should_quit = true;  // Unblock thread.
 202                         cards[card_index].new_data_ready_changed.notify_all();
 203                 }
 204                 cards[card_index].usb->stop_dequeue_thread();
 205         }
 206
 207         h264_encoder.reset(nullptr);
 208 }
 209
 210 namespace {
 211
 212 int unwrap_timecode(uint16_t current_wrapped, int last)
 213 {
 214         uint16_t last_wrapped = last & 0xffff;
 215         if (current_wrapped > last_wrapped) {
 216                 return (last & ~0xffff) | current_wrapped;
 217         } else {
 218                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 219         }
 220 }
 221
 222 float find_peak(const float *samples, size_t num_samples)
 223 {
 224         float m = fabs(samples[0]);
 225         for (size_t i = 1; i < num_samples; ++i) {
 226                 m = max(m, fabs(samples[i]));
 227         }
 228         return m;
 229 }
 230
 231 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 232 {
 233         size_t num_samples = in.size() / 2;
 234         out_l->resize(num_samples);
 235         out_r->resize(num_samples);
 236
 237         const float *inptr = in.data();
 238         float *lptr = &(*out_l)[0];
 239         float *rptr = &(*out_r)[0];
 240         for (size_t i = 0; i < num_samples; ++i) {
 241                 *lptr++ = *inptr++;
 242                 *rptr++ = *inptr++;
 243         }
 244 }
 245
 246 }  // namespace
 247
 248 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 249                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 250                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 251 {
 252         CaptureCard *card = &cards[card_index];
 253
 254         unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
 255         bool interlaced;
 256
 257         decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom,
 258                             &frame_rate_nom, &frame_rate_den, &interlaced);  // Ignore return value for now.
 259         int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom;
 260
 261         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 262         if (num_samples > OUTPUT_FREQUENCY / 10) {
 263                 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",
 264                         card_index, int(audio_frame.len), int(audio_offset),
 265                         timecode, int(video_frame.len), int(video_offset), video_format);
 266                 if (video_frame.owner) {
 267                         video_frame.owner->release_frame(video_frame);
 268                 }
 269                 if (audio_frame.owner) {
 270                         audio_frame.owner->release_frame(audio_frame);
 271                 }
 272                 return;
 273         }
 274
 275         int64_t local_pts = card->next_local_pts;
 276         int dropped_frames = 0;
 277         if (card->last_timecode != -1) {
 278                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 279         }
 280
 281         // Convert the audio to stereo fp32 and add it.
 282         vector<float> audio;
 283         audio.resize(num_samples * 2);
 284         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 285
 286         // Add the audio.
 287         {
 288                 unique_lock<mutex> lock(card->audio_mutex);
 289
 290                 // Number of samples per frame if we need to insert silence.
 291                 // (Could be nonintegral, but resampling will save us then.)
 292                 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
 293
 294                 if (dropped_frames > MAX_FPS * 2) {
 295                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 296                                 card_index, card->last_timecode, timecode);
 297                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 298                         dropped_frames = 0;
 299                 } else if (dropped_frames > 0) {
 300                         // Insert silence as needed.
 301                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 302                                 card_index, dropped_frames, timecode);
 303                         vector<float> silence(silence_samples * 2, 0.0f);
 304                         for (int i = 0; i < dropped_frames; ++i) {
 305                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 306                                 // Note that if the format changed in the meantime, we have
 307                                 // no way of detecting that; we just have to assume the frame length
 308                                 // is always the same.
 309                                 local_pts += frame_length;
 310                         }
 311                 }
 312                 if (num_samples == 0) {
 313                         audio.resize(silence_samples * 2);
 314                         num_samples = silence_samples;
 315                 }
 316                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 317                 card->next_local_pts = local_pts + frame_length;
 318         }
 319
 320         card->last_timecode = timecode;
 321
 322         // Done with the audio, so release it.
 323         if (audio_frame.owner) {
 324                 audio_frame.owner->release_frame(audio_frame);
 325         }
 326
 327         {
 328                 // Wait until the previous frame was consumed.
 329                 unique_lock<mutex> lock(bmusb_mutex);
 330                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 331                 if (card->should_quit) return;
 332         }
 333
 334         size_t expected_length = width * (height + extra_lines_top + extra_lines_bottom) * 2;
 335         if (video_frame.len - video_offset == 0 ||
 336             video_frame.len - video_offset != expected_length) {
 337                 if (video_frame.len != 0) {
 338                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 339                                 card_index, video_frame.len - video_offset, expected_length);
 340                 }
 341                 if (video_frame.owner) {
 342                         video_frame.owner->release_frame(video_frame);
 343                 }
 344
 345                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 346                 // so that pts can go up accordingly.
 347                 {
 348                         unique_lock<mutex> lock(bmusb_mutex);
 349                         card->new_data_ready = true;
 350                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 351                         card->new_frame_length = frame_length;
 352                         card->new_frame_interlaced = false;
 353                         card->new_data_ready_fence = nullptr;
 354                         card->dropped_frames = dropped_frames;
 355                         card->new_data_ready_changed.notify_all();
 356                 }
 357                 return;
 358         }
 359
 360         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 361
 362         unsigned num_fields = interlaced ? 2 : 1;
 363         timespec frame_upload_start;
 364         if (interlaced) {
 365                 // Send the two fields along as separate frames; the other side will need to add
 366                 // a deinterlacer to actually get this right.
 367                 assert(height % 2 == 0);
 368                 height /= 2;
 369                 assert(frame_length % 2 == 0);
 370                 frame_length /= 2;
 371                 num_fields = 2;
 372                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 373         }
 374         userdata->last_interlaced = interlaced;
 375         userdata->last_frame_rate_nom = frame_rate_nom;
 376         userdata->last_frame_rate_den = frame_rate_den;
 377         RefCountedFrame new_frame(video_frame);
 378
 379         // Upload the textures.
 380         size_t cbcr_width = width / 2;
 381         size_t cbcr_offset = video_offset / 2;
 382         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 383
 384         for (unsigned field = 0; field < num_fields; ++field) {
 385                 unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22);
 386
 387                 if (userdata->tex_y[field] == 0 ||
 388                     userdata->tex_cbcr[field] == 0 ||
 389                     width != userdata->last_width[field] ||
 390                     height != userdata->last_height[field]) {
 391                         // We changed resolution since last use of this texture, so we need to create
 392                         // a new object. Note that this each card has its own PBOFrameAllocator,
 393                         // we don't need to worry about these flip-flopping between resolutions.
 394                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 395                         check_error();
 396                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 397                         check_error();
 398                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 399                         check_error();
 400                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 401                         check_error();
 402                         userdata->last_width[field] = width;
 403                         userdata->last_height[field] = height;
 404                 }
 405
 406                 GLuint pbo = userdata->pbo;
 407                 check_error();
 408                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 409                 check_error();
 410                 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 411                 check_error();
 412
 413                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 414                 check_error();
 415                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t)));
 416                 check_error();
 417                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 418                 check_error();
 419                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line));
 420                 check_error();
 421                 glBindTexture(GL_TEXTURE_2D, 0);
 422                 check_error();
 423                 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 424                 check_error();
 425                 assert(fence != nullptr);
 426
 427                 if (field == 1) {
 428                         // Don't upload the second field as fast as we can; wait until
 429                         // the field time has approximately passed. (Otherwise, we could
 430                         // get timing jitter against the other sources, and possibly also
 431                         // against the video display, although the latter is not as critical.)
 432                         // This requires our system clock to be reasonably close to the
 433                         // video clock, but that's not an unreasonable assumption.
 434                         timespec second_field_start;
 435                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 436                                 frame_length * 1000000000 / TIMEBASE;
 437                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 438                                 second_field_start.tv_nsec / 1000000000;
 439                         second_field_start.tv_nsec %= 1000000000;
 440
 441                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 442                                                &second_field_start, nullptr) == -1 &&
 443                                errno == EINTR) ;
 444                 }
 445
 446                 {
 447                         unique_lock<mutex> lock(bmusb_mutex);
 448                         card->new_data_ready = true;
 449                         card->new_frame = new_frame;
 450                         card->new_frame_length = frame_length;
 451                         card->new_frame_field = field;
 452                         card->new_frame_interlaced = interlaced;
 453                         card->new_data_ready_fence = fence;
 454                         card->dropped_frames = dropped_frames;
 455                         card->new_data_ready_changed.notify_all();
 456
 457                         if (field != num_fields - 1) {
 458                                 // Wait until the previous frame was consumed.
 459                                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 460                                 if (card->should_quit) return;
 461                         }
 462                 }
 463         }
 464 }
 465
 466 void Mixer::thread_func()
 467 {
 468         eglBindAPI(EGL_OPENGL_API);
 469         QOpenGLContext *context = create_context(mixer_surface);
 470         if (!make_current(context, mixer_surface)) {
 471                 printf("oops\n");
 472                 exit(1);
 473         }
 474
 475         struct timespec start, now;
 476         clock_gettime(CLOCK_MONOTONIC, &start);
 477
 478         int frame = 0;
 479         int stats_dropped_frames = 0;
 480
 481         while (!should_quit) {
 482                 CaptureCard card_copy[MAX_CARDS];
 483                 int num_samples[MAX_CARDS];
 484
 485                 {
 486                         unique_lock<mutex> lock(bmusb_mutex);
 487
 488                         // The first card is the master timer, so wait for it to have a new frame.
 489                         // TODO: Make configurable, and with a timeout.
 490                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 491
 492                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 493                                 CaptureCard *card = &cards[card_index];
 494                                 card_copy[card_index].usb = card->usb;
 495                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 496                                 card_copy[card_index].new_frame = card->new_frame;
 497                                 card_copy[card_index].new_frame_length = card->new_frame_length;
 498                                 card_copy[card_index].new_frame_field = card->new_frame_field;
 499                                 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
 500                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 501                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 502                                 card->new_data_ready = false;
 503                                 card->new_data_ready_changed.notify_all();
 504
 505                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
 506                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 507                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 508                                 assert(num_samples[card_index] >= 0);
 509                         }
 510                 }
 511
 512                 // Resample the audio as needed, including from previously dropped frames.
 513                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 514                         {
 515                                 // Signal to the audio thread to process this frame.
 516                                 unique_lock<mutex> lock(audio_mutex);
 517                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
 518                                 audio_task_queue_changed.notify_one();
 519                         }
 520                         if (frame_num != card_copy[0].dropped_frames) {
 521                                 // For dropped frames, increase the pts. Note that if the format changed
 522                                 // in the meantime, we have no way of detecting that; we just have to
 523                                 // assume the frame length is always the same.
 524                                 ++stats_dropped_frames;
 525                                 pts_int += card_copy[0].new_frame_length;
 526                         }
 527                 }
 528
 529                 if (audio_level_callback != nullptr) {
 530                         unique_lock<mutex> lock(compressor_mutex);
 531                         double loudness_s = r128.loudness_S();
 532                         double loudness_i = r128.integrated();
 533                         double loudness_range_low = r128.range_min();
 534                         double loudness_range_high = r128.range_max();
 535
 536                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 537                                              loudness_i, loudness_range_low, loudness_range_high,
 538                                              gain_staging_db, 20.0 * log10(final_makeup_gain),
 539                                              correlation.get_correlation());
 540                 }
 541
 542                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 543                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 544                                 ++card_copy[card_index].dropped_frames;
 545                         }
 546                         if (card_copy[card_index].dropped_frames > 0) {
 547                                 printf("Card %u dropped %d frames before this\n",
 548                                         card_index, int(card_copy[card_index].dropped_frames));
 549                         }
 550                 }
 551
 552                 // If the first card is reporting a corrupted or otherwise dropped frame,
 553                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 554                 if (card_copy[0].new_frame->len == 0) {
 555                         ++stats_dropped_frames;
 556                         pts_int += card_copy[0].new_frame_length;
 557                         continue;
 558                 }
 559
 560                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 561                         CaptureCard *card = &card_copy[card_index];
 562                         if (!card->new_data_ready || card->new_frame->len == 0)
 563                                 continue;
 564
 565                         assert(card->new_frame != nullptr);
 566                         insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
 567                         check_error();
 568
 569                         // The new texture might still be uploaded,
 570                         // tell the GPU to wait until it's there.
 571                         if (card->new_data_ready_fence) {
 572                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 573                                 check_error();
 574                                 glDeleteSync(card->new_data_ready_fence);
 575                                 check_error();
 576                         }
 577                 }
 578
 579                 // Get the main chain from the theme, and set its state immediately.
 580                 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 581                 EffectChain *chain = theme_main_chain.chain;
 582                 theme_main_chain.setup_chain();
 583                 //theme_main_chain.chain->enable_phase_timing(true);
 584
 585                 GLuint y_tex, cbcr_tex;
 586                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 587                 assert(got_frame);
 588
 589                 // Render main chain.
 590                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 591                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 592                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 593                 check_error();
 594                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 595                 resource_pool->release_fbo(fbo);
 596
 597                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 598                 resource_pool->release_2d_texture(cbcr_full_tex);
 599
 600                 // Set the right state for rgba_tex.
 601                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 602                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 603                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 604                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 605                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 606
 607                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 608                 check_error();
 609
 610                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 611                 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
 612                 ++frame;
 613                 pts_int += card_copy[0].new_frame_length;
 614
 615                 // The live frame just shows the RGBA texture we just rendered.
 616                 // It owns rgba_tex now.
 617                 DisplayFrame live_frame;
 618                 live_frame.chain = display_chain.get();
 619                 live_frame.setup_chain = [this, rgba_tex]{
 620                         display_input->set_texture_num(rgba_tex);
 621                 };
 622                 live_frame.ready_fence = fence;
 623                 live_frame.input_frames = {};
 624                 live_frame.temp_textures = { rgba_tex };
 625                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 626
 627                 // Set up preview and any additional channels.
 628                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 629                         DisplayFrame display_frame;
 630                         Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 631                         display_frame.chain = chain.chain;
 632                         display_frame.setup_chain = chain.setup_chain;
 633                         display_frame.ready_fence = fence;
 634                         display_frame.input_frames = chain.input_frames;
 635                         display_frame.temp_textures = {};
 636                         output_channel[i].output_frame(display_frame);
 637                 }
 638
 639                 clock_gettime(CLOCK_MONOTONIC, &now);
 640                 double elapsed = now.tv_sec - start.tv_sec +
 641                         1e-9 * (now.tv_nsec - start.tv_nsec);
 642                 if (frame % 100 == 0) {
 643                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 644                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 645                                 1e3 * elapsed / frame);
 646                 //      chain->print_phase_timing();
 647                 }
 648
 649                 if (should_cut.exchange(false)) {  // Test and clear.
 650                         string filename = generate_local_dump_filename(frame);
 651                         printf("Starting new recording: %s\n", filename.c_str());
 652                         h264_encoder->shutdown();
 653                         httpd.close_output_file();
 654                         httpd.open_output_file(filename.c_str());
 655                         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
 656                 }
 657
 658 #if 0
 659                 // Reset every 100 frames, so that local variations in frame times
 660                 // (especially for the first few frames, when the shaders are
 661                 // compiled etc.) don't make it hard to measure for the entire
 662                 // remaining duration of the program.
 663                 if (frame == 10000) {
 664                         frame = 0;
 665                         start = now;
 666                 }
 667 #endif
 668                 check_error();
 669         }
 670
 671         resource_pool->clean_context();
 672 }
 673
 674 void Mixer::audio_thread_func()
 675 {
 676         while (!should_quit) {
 677                 AudioTask task;
 678
 679                 {
 680                         unique_lock<mutex> lock(audio_mutex);
 681                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 682                         task = audio_task_queue.front();
 683                         audio_task_queue.pop();
 684                 }
 685
 686                 process_audio_one_frame(task.pts_int, task.num_samples);
 687         }
 688 }
 689
 690 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 691 {
 692         vector<float> samples_card;
 693         vector<float> samples_out;
 694         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 695                 samples_card.resize(num_samples * 2);
 696                 {
 697                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 698                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 699                                 printf("Card %d reported previous underrun.\n", card_index);
 700                         }
 701                 }
 702                 // TODO: Allow using audio from the other card(s) as well.
 703                 if (card_index == 0) {
 704                         samples_out = move(samples_card);
 705                 }
 706         }
 707
 708         // Cut away everything under 120 Hz (or whatever the cutoff is);
 709         // we don't need it for voice, and it will reduce headroom
 710         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 711         // should be dampened.)
 712         locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 713
 714         // Apply a level compressor to get the general level right.
 715         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 716         // (or more precisely, near it, since we don't use infinite ratio),
 717         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 718         // entirely arbitrary, but from practical tests with speech, it seems to
 719         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 720         {
 721                 unique_lock<mutex> lock(compressor_mutex);
 722                 if (level_compressor_enabled) {
 723                         float threshold = 0.01f;   // -40 dBFS.
 724                         float ratio = 20.0f;
 725                         float attack_time = 0.5f;
 726                         float release_time = 20.0f;
 727                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 728                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 729                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 730                 } else {
 731                         // Just apply the gain we already had.
 732                         float g = pow(10.0f, gain_staging_db / 20.0f);
 733                         for (size_t i = 0; i < samples_out.size(); ++i) {
 734                                 samples_out[i] *= g;
 735                         }
 736                 }
 737         }
 738
 739 #if 0
 740         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 741                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 742                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 743                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 744 #endif
 745
 746 //      float limiter_att, compressor_att;
 747
 748         // The real compressor.
 749         if (compressor_enabled) {
 750                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 751                 float ratio = 20.0f;
 752                 float attack_time = 0.005f;
 753                 float release_time = 0.040f;
 754                 float makeup_gain = 2.0f;  // +6 dB.
 755                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 756 //              compressor_att = compressor.get_attenuation();
 757         }
 758
 759         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 760         // Note that since ratio is not infinite, we could go slightly higher than this.
 761         if (limiter_enabled) {
 762                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 763                 float ratio = 30.0f;
 764                 float attack_time = 0.0f;  // Instant.
 765                 float release_time = 0.020f;
 766                 float makeup_gain = 1.0f;  // 0 dB.
 767                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 768 //              limiter_att = limiter.get_attenuation();
 769         }
 770
 771 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 772
 773         // Upsample 4x to find interpolated peak.
 774         peak_resampler.inp_data = samples_out.data();
 775         peak_resampler.inp_count = samples_out.size() / 2;
 776
 777         vector<float> interpolated_samples_out;
 778         interpolated_samples_out.resize(samples_out.size());
 779         while (peak_resampler.inp_count > 0) {  // About four iterations.
 780                 peak_resampler.out_data = &interpolated_samples_out[0];
 781                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 782                 peak_resampler.process();
 783                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 784                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 785         }
 786
 787         // At this point, we are most likely close to +0 LU, but all of our
 788         // measurements have been on raw sample values, not R128 values.
 789         // So we have a final makeup gain to get us to +0 LU; the gain
 790         // adjustments required should be relatively small, and also, the
 791         // offset shouldn't change much (only if the type of audio changes
 792         // significantly). Thus, we shoot for updating this value basically
 793         // “whenever we process buffers”, since the R128 calculation isn't exactly
 794         // something we get out per-sample.
 795         //
 796         // Note that there's a feedback loop here, so we choose a very slow filter
 797         // (half-time of 100 seconds).
 798         double target_loudness_factor, alpha;
 799         {
 800                 unique_lock<mutex> lock(compressor_mutex);
 801                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
 802                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
 803                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
 804
 805                 // If we're outside +/- 5 LU uncorrected, we don't count it as
 806                 // a normal signal (probably silence) and don't change the
 807                 // correction factor; just apply what we already have.
 808                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 809                         alpha = 0.0;
 810                 } else {
 811                         // Formula adapted from
 812                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 813                         const double half_time_s = 100.0;
 814                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 815                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 816                 }
 817
 818                 double m = final_makeup_gain;
 819                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 820                         samples_out[i + 0] *= m;
 821                         samples_out[i + 1] *= m;
 822                         m += (target_loudness_factor - m) * alpha;
 823                 }
 824                 final_makeup_gain = m;
 825         }
 826
 827         // Find R128 levels and L/R correlation.
 828         vector<float> left, right;
 829         deinterleave_samples(samples_out, &left, &right);
 830         float *ptrs[] = { left.data(), right.data() };
 831         {
 832                 unique_lock<mutex> lock(compressor_mutex);
 833                 r128.process(left.size(), ptrs);
 834                 correlation.process_samples(samples_out);
 835         }
 836
 837         // Send the samples to the sound card.
 838         if (alsa) {
 839                 alsa->write(samples_out);
 840         }
 841
 842         // And finally add them to the output.
 843         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 844 }
 845
 846 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 847 {
 848         GLuint vao;
 849         glGenVertexArrays(1, &vao);
 850         check_error();
 851
 852         float vertices[] = {
 853                 0.0f, 2.0f,
 854                 0.0f, 0.0f,
 855                 2.0f, 0.0f
 856         };
 857
 858         glBindVertexArray(vao);
 859         check_error();
 860
 861         // Extract Cb/Cr.
 862         GLuint fbo = resource_pool->create_fbo(dst_tex);
 863         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 864         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 865         check_error();
 866
 867         glUseProgram(cbcr_program_num);
 868         check_error();
 869
 870         glActiveTexture(GL_TEXTURE0);
 871         check_error();
 872         glBindTexture(GL_TEXTURE_2D, src_tex);
 873         check_error();
 874         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 875         check_error();
 876         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 877         check_error();
 878         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 879         check_error();
 880
 881         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 882         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 883
 884         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 885         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 886
 887         glDrawArrays(GL_TRIANGLES, 0, 3);
 888         check_error();
 889
 890         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 891         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 892
 893         glUseProgram(0);
 894         check_error();
 895
 896         resource_pool->release_fbo(fbo);
 897         glDeleteVertexArrays(1, &vao);
 898 }
 899
 900 void Mixer::release_display_frame(DisplayFrame *frame)
 901 {
 902         for (GLuint texnum : frame->temp_textures) {
 903                 resource_pool->release_2d_texture(texnum);
 904         }
 905         frame->temp_textures.clear();
 906         frame->ready_fence.reset();
 907         frame->input_frames.clear();
 908 }
 909
 910 void Mixer::start()
 911 {
 912         mixer_thread = thread(&Mixer::thread_func, this);
 913         audio_thread = thread(&Mixer::audio_thread_func, this);
 914 }
 915
 916 void Mixer::quit()
 917 {
 918         should_quit = true;
 919         mixer_thread.join();
 920         audio_thread.join();
 921 }
 922
 923 void Mixer::transition_clicked(int transition_num)
 924 {
 925         theme->transition_clicked(transition_num, pts());
 926 }
 927
 928 void Mixer::channel_clicked(int preview_num)
 929 {
 930         theme->channel_clicked(preview_num);
 931 }
 932
 933 void Mixer::reset_meters()
 934 {
 935         peak_resampler.reset();
 936         peak = 0.0f;
 937         r128.reset();
 938         r128.integr_start();
 939         correlation.reset();
 940 }
 941
 942 Mixer::OutputChannel::~OutputChannel()
 943 {
 944         if (has_current_frame) {
 945                 parent->release_display_frame(&current_frame);
 946         }
 947         if (has_ready_frame) {
 948                 parent->release_display_frame(&ready_frame);
 949         }
 950 }
 951
 952 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 953 {
 954         // Store this frame for display. Remove the ready frame if any
 955         // (it was seemingly never used).
 956         {
 957                 unique_lock<mutex> lock(frame_mutex);
 958                 if (has_ready_frame) {
 959                         parent->release_display_frame(&ready_frame);
 960                 }
 961                 ready_frame = frame;
 962                 has_ready_frame = true;
 963         }
 964
 965         if (has_new_frame_ready_callback) {
 966                 new_frame_ready_callback();
 967         }
 968 }
 969
 970 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 971 {
 972         unique_lock<mutex> lock(frame_mutex);
 973         if (!has_current_frame && !has_ready_frame) {
 974                 return false;
 975         }
 976
 977         if (has_current_frame && has_ready_frame) {
 978                 // We have a new ready frame. Toss the current one.
 979                 parent->release_display_frame(&current_frame);
 980                 has_current_frame = false;
 981         }
 982         if (has_ready_frame) {
 983                 assert(!has_current_frame);
 984                 current_frame = ready_frame;
 985                 ready_frame.ready_fence.reset();  // Drop the refcount.
 986                 ready_frame.input_frames.clear();  // Drop the refcounts.
 987                 has_current_frame = true;
 988                 has_ready_frame = false;
 989         }
 990
 991         *frame = current_frame;
 992         return true;
 993 }
 994
 995 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 996 {
 997         new_frame_ready_callback = callback;
 998         has_new_frame_ready_callback = true;
 999 }