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           level_compressor(OUTPUT_FREQUENCY),
 102           limiter(OUTPUT_FREQUENCY),
 103           compressor(OUTPUT_FREQUENCY)
 104 {
 105         httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str());
 106         httpd.start(9095);
 107
 108         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 109         check_error();
 110
 111         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 112         // will be halved when sampling them, and we need to compensate here.
 113         movit_texel_subpixel_precision /= 2.0;
 114
 115         resource_pool.reset(new ResourcePool);
 116         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
 117         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 118                 output_channel[i].parent = this;
 119         }
 120
 121         ImageFormat inout_format;
 122         inout_format.color_space = COLORSPACE_sRGB;
 123         inout_format.gamma_curve = GAMMA_sRGB;
 124
 125         // Display chain; shows the live output produced by the main chain (its RGBA version).
 126         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 127         check_error();
 128         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 129         display_chain->add_input(display_input);
 130         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 131         display_chain->set_dither_bits(0);  // Don't bother.
 132         display_chain->finalize();
 133
 134         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
 135
 136         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 137                 printf("Configuring card %d...\n", card_index);
 138                 CaptureCard *card = &cards[card_index];
 139                 card->usb = new BMUSBCapture(card_index);
 140                 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 141                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 142                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
 143                 card->surface = create_surface(format);
 144                 card->usb->set_dequeue_thread_callbacks(
 145                         [card]{
 146                                 eglBindAPI(EGL_OPENGL_API);
 147                                 card->context = create_context(card->surface);
 148                                 if (!make_current(card->context, card->surface)) {
 149                                         printf("failed to create bmusb context\n");
 150                                         exit(1);
 151                                 }
 152                         },
 153                         [this]{
 154                                 resource_pool->clean_context();
 155                         });
 156                 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 157                 card->usb->configure_card();
 158         }
 159
 160         BMUSBCapture::start_bm_thread();
 161
 162         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 163                 cards[card_index].usb->start_bm_capture();
 164         }
 165
 166         // Set up stuff for NV12 conversion.
 167
 168         // Cb/Cr shader.
 169         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 170         string cbcr_frag_shader =
 171                 "#version 130 \n"
 172                 "in vec2 tc0; \n"
 173                 "uniform sampler2D cbcr_tex; \n"
 174                 "void main() { \n"
 175                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 176                 "} \n";
 177         vector<string> frag_shader_outputs;
 178         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 179
 180         r128.init(2, OUTPUT_FREQUENCY);
 181         r128.integr_start();
 182
 183         locut.init(FILTER_HPF, 2);
 184
 185         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 186         // and there's a limit to how important the peak meter is.
 187         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
 188
 189         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 190 }
 191
 192 Mixer::~Mixer()
 193 {
 194         resource_pool->release_glsl_program(cbcr_program_num);
 195         BMUSBCapture::stop_bm_thread();
 196
 197         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 198                 {
 199                         unique_lock<mutex> lock(bmusb_mutex);
 200                         cards[card_index].should_quit = true;  // Unblock thread.
 201                         cards[card_index].new_data_ready_changed.notify_all();
 202                 }
 203                 cards[card_index].usb->stop_dequeue_thread();
 204         }
 205
 206         h264_encoder.reset(nullptr);
 207 }
 208
 209 namespace {
 210
 211 int unwrap_timecode(uint16_t current_wrapped, int last)
 212 {
 213         uint16_t last_wrapped = last & 0xffff;
 214         if (current_wrapped > last_wrapped) {
 215                 return (last & ~0xffff) | current_wrapped;
 216         } else {
 217                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 218         }
 219 }
 220
 221 float find_peak(const float *samples, size_t num_samples)
 222 {
 223         float m = fabs(samples[0]);
 224         for (size_t i = 1; i < num_samples; ++i) {
 225                 m = max(m, fabs(samples[i]));
 226         }
 227         return m;
 228 }
 229
 230 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 231 {
 232         size_t num_samples = in.size() / 2;
 233         out_l->resize(num_samples);
 234         out_r->resize(num_samples);
 235
 236         const float *inptr = in.data();
 237         float *lptr = &(*out_l)[0];
 238         float *rptr = &(*out_r)[0];
 239         for (size_t i = 0; i < num_samples; ++i) {
 240                 *lptr++ = *inptr++;
 241                 *rptr++ = *inptr++;
 242         }
 243 }
 244
 245 }  // namespace
 246
 247 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 248                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 249                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 250 {
 251         CaptureCard *card = &cards[card_index];
 252
 253         unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
 254         bool interlaced;
 255
 256         decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom,
 257                             &frame_rate_nom, &frame_rate_den, &interlaced);  // Ignore return value for now.
 258         int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom;
 259
 260         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 261         if (num_samples > OUTPUT_FREQUENCY / 10) {
 262                 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",
 263                         card_index, int(audio_frame.len), int(audio_offset),
 264                         timecode, int(video_frame.len), int(video_offset), video_format);
 265                 if (video_frame.owner) {
 266                         video_frame.owner->release_frame(video_frame);
 267                 }
 268                 if (audio_frame.owner) {
 269                         audio_frame.owner->release_frame(audio_frame);
 270                 }
 271                 return;
 272         }
 273
 274         int64_t local_pts = card->next_local_pts;
 275         int dropped_frames = 0;
 276         if (card->last_timecode != -1) {
 277                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 278         }
 279
 280         // Convert the audio to stereo fp32 and add it.
 281         vector<float> audio;
 282         audio.resize(num_samples * 2);
 283         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 284
 285         // Add the audio.
 286         {
 287                 unique_lock<mutex> lock(card->audio_mutex);
 288
 289                 // Number of samples per frame if we need to insert silence.
 290                 // (Could be nonintegral, but resampling will save us then.)
 291                 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
 292
 293                 if (dropped_frames > MAX_FPS * 2) {
 294                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 295                                 card_index, card->last_timecode, timecode);
 296                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 297                         dropped_frames = 0;
 298                 } else if (dropped_frames > 0) {
 299                         // Insert silence as needed.
 300                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 301                                 card_index, dropped_frames, timecode);
 302                         vector<float> silence(silence_samples * 2, 0.0f);
 303                         for (int i = 0; i < dropped_frames; ++i) {
 304                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 305                                 // Note that if the format changed in the meantime, we have
 306                                 // no way of detecting that; we just have to assume the frame length
 307                                 // is always the same.
 308                                 local_pts += frame_length;
 309                         }
 310                 }
 311                 if (num_samples == 0) {
 312                         audio.resize(silence_samples * 2);
 313                         num_samples = silence_samples;
 314                 }
 315                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 316                 card->next_local_pts = local_pts + frame_length;
 317         }
 318
 319         card->last_timecode = timecode;
 320
 321         // Done with the audio, so release it.
 322         if (audio_frame.owner) {
 323                 audio_frame.owner->release_frame(audio_frame);
 324         }
 325
 326         {
 327                 // Wait until the previous frame was consumed.
 328                 unique_lock<mutex> lock(bmusb_mutex);
 329                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 330                 if (card->should_quit) return;
 331         }
 332
 333         size_t expected_length = width * (height + extra_lines_top + extra_lines_bottom) * 2;
 334         if (video_frame.len - video_offset == 0 ||
 335             video_frame.len - video_offset != expected_length) {
 336                 if (video_frame.len != 0) {
 337                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 338                                 card_index, video_frame.len - video_offset, expected_length);
 339                 }
 340                 if (video_frame.owner) {
 341                         video_frame.owner->release_frame(video_frame);
 342                 }
 343
 344                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 345                 // so that pts can go up accordingly.
 346                 {
 347                         unique_lock<mutex> lock(bmusb_mutex);
 348                         card->new_data_ready = true;
 349                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 350                         card->new_frame_length = frame_length;
 351                         card->new_frame_interlaced = false;
 352                         card->new_data_ready_fence = nullptr;
 353                         card->dropped_frames = dropped_frames;
 354                         card->new_data_ready_changed.notify_all();
 355                 }
 356                 return;
 357         }
 358
 359         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 360
 361         unsigned num_fields = interlaced ? 2 : 1;
 362         timespec frame_upload_start;
 363         if (interlaced) {
 364                 // Send the two fields along as separate frames; the other side will need to add
 365                 // a deinterlacer to actually get this right.
 366                 assert(height % 2 == 0);
 367                 height /= 2;
 368                 assert(frame_length % 2 == 0);
 369                 frame_length /= 2;
 370                 num_fields = 2;
 371                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 372         }
 373         userdata->last_interlaced = interlaced;
 374         userdata->last_frame_rate_nom = frame_rate_nom;
 375         userdata->last_frame_rate_den = frame_rate_den;
 376         RefCountedFrame new_frame(video_frame);
 377
 378         // Upload the textures.
 379         size_t cbcr_width = width / 2;
 380         size_t cbcr_offset = video_offset / 2;
 381         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 382
 383         for (unsigned field = 0; field < num_fields; ++field) {
 384                 unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22);
 385
 386                 if (userdata->tex_y[field] == 0 ||
 387                     userdata->tex_cbcr[field] == 0 ||
 388                     width != userdata->last_width[field] ||
 389                     height != userdata->last_height[field]) {
 390                         // We changed resolution since last use of this texture, so we need to create
 391                         // a new object. Note that this each card has its own PBOFrameAllocator,
 392                         // we don't need to worry about these flip-flopping between resolutions.
 393                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 394                         check_error();
 395                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 396                         check_error();
 397                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 398                         check_error();
 399                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 400                         check_error();
 401                         userdata->last_width[field] = width;
 402                         userdata->last_height[field] = height;
 403                 }
 404
 405                 GLuint pbo = userdata->pbo;
 406                 check_error();
 407                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 408                 check_error();
 409                 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 410                 check_error();
 411
 412                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 413                 check_error();
 414                 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)));
 415                 check_error();
 416                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 417                 check_error();
 418                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line));
 419                 check_error();
 420                 glBindTexture(GL_TEXTURE_2D, 0);
 421                 check_error();
 422                 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 423                 check_error();
 424                 assert(fence != nullptr);
 425
 426                 if (field == 1) {
 427                         // Don't upload the second field as fast as we can; wait until
 428                         // the field time has approximately passed. (Otherwise, we could
 429                         // get timing jitter against the other sources, and possibly also
 430                         // against the video display, although the latter is not as critical.)
 431                         // This requires our system clock to be reasonably close to the
 432                         // video clock, but that's not an unreasonable assumption.
 433                         timespec second_field_start;
 434                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 435                                 frame_length * 1000000000 / TIMEBASE;
 436                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 437                                 second_field_start.tv_nsec / 1000000000;
 438                         second_field_start.tv_nsec %= 1000000000;
 439
 440                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 441                                                &second_field_start, nullptr) == -1 &&
 442                                errno == EINTR) ;
 443                 }
 444
 445                 {
 446                         unique_lock<mutex> lock(bmusb_mutex);
 447                         card->new_data_ready = true;
 448                         card->new_frame = new_frame;
 449                         card->new_frame_length = frame_length;
 450                         card->new_frame_field = field;
 451                         card->new_frame_interlaced = interlaced;
 452                         card->new_data_ready_fence = fence;
 453                         card->dropped_frames = dropped_frames;
 454                         card->new_data_ready_changed.notify_all();
 455
 456                         if (field != num_fields - 1) {
 457                                 // Wait until the previous frame was consumed.
 458                                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 459                                 if (card->should_quit) return;
 460                         }
 461                 }
 462         }
 463 }
 464
 465 void Mixer::thread_func()
 466 {
 467         eglBindAPI(EGL_OPENGL_API);
 468         QOpenGLContext *context = create_context(mixer_surface);
 469         if (!make_current(context, mixer_surface)) {
 470                 printf("oops\n");
 471                 exit(1);
 472         }
 473
 474         struct timespec start, now;
 475         clock_gettime(CLOCK_MONOTONIC, &start);
 476
 477         int frame = 0;
 478         int stats_dropped_frames = 0;
 479
 480         while (!should_quit) {
 481                 CaptureCard card_copy[MAX_CARDS];
 482                 int num_samples[MAX_CARDS];
 483
 484                 {
 485                         unique_lock<mutex> lock(bmusb_mutex);
 486
 487                         // The first card is the master timer, so wait for it to have a new frame.
 488                         // TODO: Make configurable, and with a timeout.
 489                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 490
 491                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 492                                 CaptureCard *card = &cards[card_index];
 493                                 card_copy[card_index].usb = card->usb;
 494                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 495                                 card_copy[card_index].new_frame = card->new_frame;
 496                                 card_copy[card_index].new_frame_length = card->new_frame_length;
 497                                 card_copy[card_index].new_frame_field = card->new_frame_field;
 498                                 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
 499                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 500                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 501                                 card->new_data_ready = false;
 502                                 card->new_data_ready_changed.notify_all();
 503
 504                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
 505                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 506                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 507                                 assert(num_samples[card_index] >= 0);
 508                         }
 509                 }
 510
 511                 // Resample the audio as needed, including from previously dropped frames.
 512                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 513                         {
 514                                 // Signal to the audio thread to process this frame.
 515                                 unique_lock<mutex> lock(audio_mutex);
 516                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
 517                                 audio_task_queue_changed.notify_one();
 518                         }
 519                         if (frame_num != card_copy[0].dropped_frames) {
 520                                 // For dropped frames, increase the pts. Note that if the format changed
 521                                 // in the meantime, we have no way of detecting that; we just have to
 522                                 // assume the frame length is always the same.
 523                                 ++stats_dropped_frames;
 524                                 pts_int += card_copy[0].new_frame_length;
 525                         }
 526                 }
 527
 528                 if (audio_level_callback != nullptr) {
 529                         unique_lock<mutex> lock(r128_mutex);
 530                         double loudness_s = r128.loudness_S();
 531                         double loudness_i = r128.integrated();
 532                         double loudness_range_low = r128.range_min();
 533                         double loudness_range_high = r128.range_max();
 534
 535                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 536                                              loudness_i, loudness_range_low, loudness_range_high,
 537                                              last_gain_staging_db);
 538                 }
 539
 540                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 541                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 542                                 ++card_copy[card_index].dropped_frames;
 543                         }
 544                         if (card_copy[card_index].dropped_frames > 0) {
 545                                 printf("Card %u dropped %d frames before this\n",
 546                                         card_index, int(card_copy[card_index].dropped_frames));
 547                         }
 548                 }
 549
 550                 // If the first card is reporting a corrupted or otherwise dropped frame,
 551                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 552                 if (card_copy[0].new_frame->len == 0) {
 553                         ++stats_dropped_frames;
 554                         pts_int += card_copy[0].new_frame_length;
 555                         continue;
 556                 }
 557
 558                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 559                         CaptureCard *card = &card_copy[card_index];
 560                         if (!card->new_data_ready || card->new_frame->len == 0)
 561                                 continue;
 562
 563                         assert(card->new_frame != nullptr);
 564                         insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
 565                         check_error();
 566
 567                         // The new texture might still be uploaded,
 568                         // tell the GPU to wait until it's there.
 569                         if (card->new_data_ready_fence) {
 570                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 571                                 check_error();
 572                                 glDeleteSync(card->new_data_ready_fence);
 573                                 check_error();
 574                         }
 575                 }
 576
 577                 // Get the main chain from the theme, and set its state immediately.
 578                 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 579                 EffectChain *chain = theme_main_chain.chain;
 580                 theme_main_chain.setup_chain();
 581                 //theme_main_chain.chain->enable_phase_timing(true);
 582
 583                 GLuint y_tex, cbcr_tex;
 584                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 585                 assert(got_frame);
 586
 587                 // Render main chain.
 588                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 589                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 590                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 591                 check_error();
 592                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 593                 resource_pool->release_fbo(fbo);
 594
 595                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 596                 resource_pool->release_2d_texture(cbcr_full_tex);
 597
 598                 // Set the right state for rgba_tex.
 599                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 600                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 601                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 602                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 603                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 604
 605                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 606                 check_error();
 607
 608                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 609                 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
 610                 ++frame;
 611                 pts_int += card_copy[0].new_frame_length;
 612
 613                 // The live frame just shows the RGBA texture we just rendered.
 614                 // It owns rgba_tex now.
 615                 DisplayFrame live_frame;
 616                 live_frame.chain = display_chain.get();
 617                 live_frame.setup_chain = [this, rgba_tex]{
 618                         display_input->set_texture_num(rgba_tex);
 619                 };
 620                 live_frame.ready_fence = fence;
 621                 live_frame.input_frames = {};
 622                 live_frame.temp_textures = { rgba_tex };
 623                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 624
 625                 // Set up preview and any additional channels.
 626                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 627                         DisplayFrame display_frame;
 628                         Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 629                         display_frame.chain = chain.chain;
 630                         display_frame.setup_chain = chain.setup_chain;
 631                         display_frame.ready_fence = fence;
 632                         display_frame.input_frames = chain.input_frames;
 633                         display_frame.temp_textures = {};
 634                         output_channel[i].output_frame(display_frame);
 635                 }
 636
 637                 clock_gettime(CLOCK_MONOTONIC, &now);
 638                 double elapsed = now.tv_sec - start.tv_sec +
 639                         1e-9 * (now.tv_nsec - start.tv_nsec);
 640                 if (frame % 100 == 0) {
 641                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 642                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 643                                 1e3 * elapsed / frame);
 644                 //      chain->print_phase_timing();
 645                 }
 646
 647                 if (should_cut.exchange(false)) {  // Test and clear.
 648                         string filename = generate_local_dump_filename(frame);
 649                         printf("Starting new recording: %s\n", filename.c_str());
 650                         h264_encoder->shutdown();
 651                         httpd.close_output_file();
 652                         httpd.open_output_file(filename.c_str());
 653                         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
 654                 }
 655
 656 #if 0
 657                 // Reset every 100 frames, so that local variations in frame times
 658                 // (especially for the first few frames, when the shaders are
 659                 // compiled etc.) don't make it hard to measure for the entire
 660                 // remaining duration of the program.
 661                 if (frame == 10000) {
 662                         frame = 0;
 663                         start = now;
 664                 }
 665 #endif
 666                 check_error();
 667         }
 668
 669         resource_pool->clean_context();
 670 }
 671
 672 void Mixer::audio_thread_func()
 673 {
 674         while (!should_quit) {
 675                 AudioTask task;
 676
 677                 {
 678                         unique_lock<mutex> lock(audio_mutex);
 679                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 680                         task = audio_task_queue.front();
 681                         audio_task_queue.pop();
 682                 }
 683
 684                 process_audio_one_frame(task.pts_int, task.num_samples);
 685         }
 686 }
 687
 688 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 689 {
 690         vector<float> samples_card;
 691         vector<float> samples_out;
 692         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 693                 samples_card.resize(num_samples * 2);
 694                 {
 695                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 696                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 697                                 printf("Card %d reported previous underrun.\n", card_index);
 698                         }
 699                 }
 700                 // TODO: Allow using audio from the other card(s) as well.
 701                 if (card_index == 0) {
 702                         samples_out = move(samples_card);
 703                 }
 704         }
 705
 706         // Cut away everything under 120 Hz (or whatever the cutoff is);
 707         // we don't need it for voice, and it will reduce headroom
 708         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 709         // should be dampened.)
 710         locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 711
 712         // Apply a level compressor to get the general level right.
 713         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 714         // (or more precisely, near it, since we don't use infinite ratio),
 715         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 716         // entirely arbitrary, but from practical tests with speech, it seems to
 717         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 718         if (level_compressor_enabled) {
 719                 float threshold = 0.01f;   // -40 dBFS.
 720                 float ratio = 20.0f;
 721                 float attack_time = 0.5f;
 722                 float release_time = 20.0f;
 723                 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 724                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 725                 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 726         }
 727
 728 #if 0
 729         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 730                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 731                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 732                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 733 #endif
 734
 735 //      float limiter_att, compressor_att;
 736
 737         // The real compressor.
 738         if (compressor_enabled) {
 739                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 740                 float ratio = 20.0f;
 741                 float attack_time = 0.005f;
 742                 float release_time = 0.040f;
 743                 float makeup_gain = 2.0f;  // +6 dB.
 744                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 745 //              compressor_att = compressor.get_attenuation();
 746         }
 747
 748         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 749         // Note that since ratio is not infinite, we could go slightly higher than this.
 750         if (limiter_enabled) {
 751                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 752                 float ratio = 30.0f;
 753                 float attack_time = 0.0f;  // Instant.
 754                 float release_time = 0.020f;
 755                 float makeup_gain = 1.0f;  // 0 dB.
 756                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 757 //              limiter_att = limiter.get_attenuation();
 758         }
 759
 760 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 761
 762         // Upsample 4x to find interpolated peak.
 763         peak_resampler.inp_data = samples_out.data();
 764         peak_resampler.inp_count = samples_out.size() / 2;
 765
 766         vector<float> interpolated_samples_out;
 767         interpolated_samples_out.resize(samples_out.size());
 768         while (peak_resampler.inp_count > 0) {  // About four iterations.
 769                 peak_resampler.out_data = &interpolated_samples_out[0];
 770                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 771                 peak_resampler.process();
 772                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 773                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 774         }
 775
 776         // Find R128 levels.
 777         vector<float> left, right;
 778         deinterleave_samples(samples_out, &left, &right);
 779         float *ptrs[] = { left.data(), right.data() };
 780         {
 781                 unique_lock<mutex> lock(r128_mutex);
 782                 r128.process(left.size(), ptrs);
 783         }
 784
 785         // Send the samples to the sound card.
 786         if (alsa) {
 787                 alsa->write(samples_out);
 788         }
 789
 790         // And finally add them to the output.
 791         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 792 }
 793
 794 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 795 {
 796         GLuint vao;
 797         glGenVertexArrays(1, &vao);
 798         check_error();
 799
 800         float vertices[] = {
 801                 0.0f, 2.0f,
 802                 0.0f, 0.0f,
 803                 2.0f, 0.0f
 804         };
 805
 806         glBindVertexArray(vao);
 807         check_error();
 808
 809         // Extract Cb/Cr.
 810         GLuint fbo = resource_pool->create_fbo(dst_tex);
 811         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 812         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 813         check_error();
 814
 815         glUseProgram(cbcr_program_num);
 816         check_error();
 817
 818         glActiveTexture(GL_TEXTURE0);
 819         check_error();
 820         glBindTexture(GL_TEXTURE_2D, src_tex);
 821         check_error();
 822         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 823         check_error();
 824         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 825         check_error();
 826         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 827         check_error();
 828
 829         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 830         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 831
 832         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 833         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 834
 835         glDrawArrays(GL_TRIANGLES, 0, 3);
 836         check_error();
 837
 838         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 839         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 840
 841         glUseProgram(0);
 842         check_error();
 843
 844         resource_pool->release_fbo(fbo);
 845         glDeleteVertexArrays(1, &vao);
 846 }
 847
 848 void Mixer::release_display_frame(DisplayFrame *frame)
 849 {
 850         for (GLuint texnum : frame->temp_textures) {
 851                 resource_pool->release_2d_texture(texnum);
 852         }
 853         frame->temp_textures.clear();
 854         frame->ready_fence.reset();
 855         frame->input_frames.clear();
 856 }
 857
 858 void Mixer::start()
 859 {
 860         mixer_thread = thread(&Mixer::thread_func, this);
 861         audio_thread = thread(&Mixer::audio_thread_func, this);
 862 }
 863
 864 void Mixer::quit()
 865 {
 866         should_quit = true;
 867         mixer_thread.join();
 868         audio_thread.join();
 869 }
 870
 871 void Mixer::transition_clicked(int transition_num)
 872 {
 873         theme->transition_clicked(transition_num, pts());
 874 }
 875
 876 void Mixer::channel_clicked(int preview_num)
 877 {
 878         theme->channel_clicked(preview_num);
 879 }
 880
 881 void Mixer::reset_meters()
 882 {
 883         peak_resampler.reset();
 884         peak = 0.0f;
 885         r128.reset();
 886         r128.integr_start();
 887 }
 888
 889 Mixer::OutputChannel::~OutputChannel()
 890 {
 891         if (has_current_frame) {
 892                 parent->release_display_frame(&current_frame);
 893         }
 894         if (has_ready_frame) {
 895                 parent->release_display_frame(&ready_frame);
 896         }
 897 }
 898
 899 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 900 {
 901         // Store this frame for display. Remove the ready frame if any
 902         // (it was seemingly never used).
 903         {
 904                 unique_lock<mutex> lock(frame_mutex);
 905                 if (has_ready_frame) {
 906                         parent->release_display_frame(&ready_frame);
 907                 }
 908                 ready_frame = frame;
 909                 has_ready_frame = true;
 910         }
 911
 912         if (has_new_frame_ready_callback) {
 913                 new_frame_ready_callback();
 914         }
 915 }
 916
 917 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 918 {
 919         unique_lock<mutex> lock(frame_mutex);
 920         if (!has_current_frame && !has_ready_frame) {
 921                 return false;
 922         }
 923
 924         if (has_current_frame && has_ready_frame) {
 925                 // We have a new ready frame. Toss the current one.
 926                 parent->release_display_frame(&current_frame);
 927                 has_current_frame = false;
 928         }
 929         if (has_ready_frame) {
 930                 assert(!has_current_frame);
 931                 current_frame = ready_frame;
 932                 ready_frame.ready_fence.reset();  // Drop the refcount.
 933                 ready_frame.input_frames.clear();  // Drop the refcounts.
 934                 has_current_frame = true;
 935                 has_ready_frame = false;
 936         }
 937
 938         *frame = current_frame;
 939         return true;
 940 }
 941
 942 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 943 {
 944         new_frame_ready_callback = callback;
 945         has_new_frame_ready_callback = true;
 946 }