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 }  // namespace
  78
  79 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
  80         : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT),
  81           num_cards(num_cards),
  82           mixer_surface(create_surface(format)),
  83           h264_encoder_surface(create_surface(format)),
  84           level_compressor(OUTPUT_FREQUENCY),
  85           limiter(OUTPUT_FREQUENCY),
  86           compressor(OUTPUT_FREQUENCY)
  87 {
  88         httpd.start(9095);
  89
  90         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
  91         check_error();
  92
  93         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
  94         // will be halved when sampling them, and we need to compensate here.
  95         movit_texel_subpixel_precision /= 2.0;
  96
  97         resource_pool.reset(new ResourcePool);
  98         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
  99         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 100                 output_channel[i].parent = this;
 101         }
 102
 103         ImageFormat inout_format;
 104         inout_format.color_space = COLORSPACE_sRGB;
 105         inout_format.gamma_curve = GAMMA_sRGB;
 106
 107         // Display chain; shows the live output produced by the main chain (its RGBA version).
 108         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 109         check_error();
 110         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 111         display_chain->add_input(display_input);
 112         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 113         display_chain->set_dither_bits(0);  // Don't bother.
 114         display_chain->finalize();
 115
 116         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
 117
 118         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 119                 printf("Configuring card %d...\n", card_index);
 120                 CaptureCard *card = &cards[card_index];
 121                 card->usb = new BMUSBCapture(card_index);
 122                 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 123                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 124                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
 125                 card->surface = create_surface(format);
 126                 card->usb->set_dequeue_thread_callbacks(
 127                         [card]{
 128                                 eglBindAPI(EGL_OPENGL_API);
 129                                 card->context = create_context(card->surface);
 130                                 if (!make_current(card->context, card->surface)) {
 131                                         printf("failed to create bmusb context\n");
 132                                         exit(1);
 133                                 }
 134                         },
 135                         [this]{
 136                                 resource_pool->clean_context();
 137                         });
 138                 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 139                 card->usb->configure_card();
 140         }
 141
 142         BMUSBCapture::start_bm_thread();
 143
 144         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 145                 cards[card_index].usb->start_bm_capture();
 146         }
 147
 148         // Set up stuff for NV12 conversion.
 149
 150         // Cb/Cr shader.
 151         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 152         string cbcr_frag_shader =
 153                 "#version 130 \n"
 154                 "in vec2 tc0; \n"
 155                 "uniform sampler2D cbcr_tex; \n"
 156                 "void main() { \n"
 157                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 158                 "} \n";
 159         vector<string> frag_shader_outputs;
 160         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 161
 162         r128.init(2, OUTPUT_FREQUENCY);
 163         r128.integr_start();
 164
 165         locut.init(FILTER_HPF, 2);
 166
 167         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 168         // and there's a limit to how important the peak meter is.
 169         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
 170
 171         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 172 }
 173
 174 Mixer::~Mixer()
 175 {
 176         resource_pool->release_glsl_program(cbcr_program_num);
 177         BMUSBCapture::stop_bm_thread();
 178
 179         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 180                 {
 181                         unique_lock<mutex> lock(bmusb_mutex);
 182                         cards[card_index].should_quit = true;  // Unblock thread.
 183                         cards[card_index].new_data_ready_changed.notify_all();
 184                 }
 185                 cards[card_index].usb->stop_dequeue_thread();
 186         }
 187
 188         h264_encoder.reset(nullptr);
 189 }
 190
 191 namespace {
 192
 193 int unwrap_timecode(uint16_t current_wrapped, int last)
 194 {
 195         uint16_t last_wrapped = last & 0xffff;
 196         if (current_wrapped > last_wrapped) {
 197                 return (last & ~0xffff) | current_wrapped;
 198         } else {
 199                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 200         }
 201 }
 202
 203 float find_peak(const float *samples, size_t num_samples)
 204 {
 205         float m = fabs(samples[0]);
 206         for (size_t i = 1; i < num_samples; ++i) {
 207                 m = std::max(m, fabs(samples[i]));
 208         }
 209         return m;
 210 }
 211
 212 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 213 {
 214         size_t num_samples = in.size() / 2;
 215         out_l->resize(num_samples);
 216         out_r->resize(num_samples);
 217
 218         const float *inptr = in.data();
 219         float *lptr = &(*out_l)[0];
 220         float *rptr = &(*out_r)[0];
 221         for (size_t i = 0; i < num_samples; ++i) {
 222                 *lptr++ = *inptr++;
 223                 *rptr++ = *inptr++;
 224         }
 225 }
 226
 227 }  // namespace
 228
 229 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 230                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 231                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 232 {
 233         CaptureCard *card = &cards[card_index];
 234
 235         unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
 236         bool interlaced;
 237
 238         decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom,
 239                             &frame_rate_nom, &frame_rate_den, &interlaced);  // Ignore return value for now.
 240         int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom;
 241
 242         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 243         if (num_samples > OUTPUT_FREQUENCY / 10) {
 244                 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",
 245                         card_index, int(audio_frame.len), int(audio_offset),
 246                         timecode, int(video_frame.len), int(video_offset), video_format);
 247                 if (video_frame.owner) {
 248                         video_frame.owner->release_frame(video_frame);
 249                 }
 250                 if (audio_frame.owner) {
 251                         audio_frame.owner->release_frame(audio_frame);
 252                 }
 253                 return;
 254         }
 255
 256         int64_t local_pts = card->next_local_pts;
 257         int dropped_frames = 0;
 258         if (card->last_timecode != -1) {
 259                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 260         }
 261
 262         // Convert the audio to stereo fp32 and add it.
 263         vector<float> audio;
 264         audio.resize(num_samples * 2);
 265         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 266
 267         // Add the audio.
 268         {
 269                 unique_lock<mutex> lock(card->audio_mutex);
 270
 271                 // Number of samples per frame if we need to insert silence.
 272                 // (Could be nonintegral, but resampling will save us then.)
 273                 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
 274
 275                 if (dropped_frames > MAX_FPS * 2) {
 276                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 277                                 card_index, card->last_timecode, timecode);
 278                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 279                         dropped_frames = 0;
 280                 } else if (dropped_frames > 0) {
 281                         // Insert silence as needed.
 282                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 283                                 card_index, dropped_frames, timecode);
 284                         vector<float> silence(silence_samples * 2, 0.0f);
 285                         for (int i = 0; i < dropped_frames; ++i) {
 286                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 287                                 // Note that if the format changed in the meantime, we have
 288                                 // no way of detecting that; we just have to assume the frame length
 289                                 // is always the same.
 290                                 local_pts += frame_length;
 291                         }
 292                 }
 293                 if (num_samples == 0) {
 294                         audio.resize(silence_samples * 2);
 295                         num_samples = silence_samples;
 296                 }
 297                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 298                 card->next_local_pts = local_pts + frame_length;
 299         }
 300
 301         card->last_timecode = timecode;
 302
 303         // Done with the audio, so release it.
 304         if (audio_frame.owner) {
 305                 audio_frame.owner->release_frame(audio_frame);
 306         }
 307
 308         {
 309                 // Wait until the previous frame was consumed.
 310                 unique_lock<mutex> lock(bmusb_mutex);
 311                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 312                 if (card->should_quit) return;
 313         }
 314
 315         if (video_frame.len - video_offset == 0 ||
 316             video_frame.len - video_offset != size_t(width * (height + extra_lines_top + extra_lines_bottom) * 2)) {
 317                 if (video_frame.len != 0) {
 318                         printf("Card %d: Dropping video frame with wrong length (%ld)\n",
 319                                 card_index, video_frame.len - video_offset);
 320                 }
 321                 if (video_frame.owner) {
 322                         video_frame.owner->release_frame(video_frame);
 323                 }
 324
 325                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 326                 // so that pts can go up accordingly.
 327                 {
 328                         unique_lock<mutex> lock(bmusb_mutex);
 329                         card->new_data_ready = true;
 330                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 331                         card->new_frame_length = frame_length;
 332                         card->new_frame_interlaced = false;
 333                         card->new_data_ready_fence = nullptr;
 334                         card->dropped_frames = dropped_frames;
 335                         card->new_data_ready_changed.notify_all();
 336                 }
 337                 return;
 338         }
 339
 340         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 341
 342         unsigned num_fields = interlaced ? 2 : 1;
 343         timespec frame_upload_start;
 344         if (interlaced) {
 345                 // NOTE: This isn't deinterlacing. This is just sending the two fields along
 346                 // as separate frames without considering anything like the half-field offset.
 347                 // We'll need to add a proper deinterlacer on the receiving side to get this right.
 348                 assert(height % 2 == 0);
 349                 height /= 2;
 350                 assert(frame_length % 2 == 0);
 351                 frame_length /= 2;
 352                 num_fields = 2;
 353                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 354         }
 355         RefCountedFrame new_frame(video_frame);
 356
 357         // Upload the textures.
 358         size_t cbcr_width = width / 2;
 359         size_t cbcr_offset = video_offset / 2;
 360         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 361
 362         for (unsigned field = 0; field < num_fields; ++field) {
 363                 unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22);
 364
 365                 if (userdata->tex_y[field] == 0 ||
 366                     userdata->tex_cbcr[field] == 0 ||
 367                     width != userdata->last_width[field] ||
 368                     height != userdata->last_height[field]) {
 369                         // We changed resolution since last use of this texture, so we need to create
 370                         // a new object. Note that this each card has its own PBOFrameAllocator,
 371                         // we don't need to worry about these flip-flopping between resolutions.
 372                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 373                         check_error();
 374                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 375                         check_error();
 376                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 377                         check_error();
 378                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 379                         check_error();
 380                         userdata->last_width[field] = width;
 381                         userdata->last_height[field] = height;
 382                 }
 383
 384                 GLuint pbo = userdata->pbo;
 385                 check_error();
 386                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 387                 check_error();
 388                 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 389                 check_error();
 390
 391                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 392                 check_error();
 393                 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)));
 394                 check_error();
 395                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 396                 check_error();
 397                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line));
 398                 check_error();
 399                 glBindTexture(GL_TEXTURE_2D, 0);
 400                 check_error();
 401                 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 402                 check_error();
 403                 assert(fence != nullptr);
 404
 405                 if (field == 1) {
 406                         // Don't upload the second field as fast as we can; wait until
 407                         // the field time has approximately passed. (Otherwise, we could
 408                         // get timing jitter against the other sources, and possibly also
 409                         // against the video display, although the latter is not as critical.)
 410                         // This requires our system clock to be reasonably close to the
 411                         // video clock, but that's not an unreasonable assumption.
 412                         timespec second_field_start;
 413                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 414                                 frame_length * 1000000000 / TIMEBASE;
 415                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 416                                 second_field_start.tv_nsec / 1000000000;
 417                         second_field_start.tv_nsec %= 1000000000;
 418
 419                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 420                                                &second_field_start, nullptr) == -1 &&
 421                                errno == EINTR) ;
 422                 }
 423
 424                 {
 425                         unique_lock<mutex> lock(bmusb_mutex);
 426                         card->new_data_ready = true;
 427                         card->new_frame = new_frame;
 428                         card->new_frame_length = frame_length;
 429                         card->new_frame_field = field;
 430                         card->new_frame_interlaced = interlaced;
 431                         card->new_data_ready_fence = fence;
 432                         card->dropped_frames = dropped_frames;
 433                         card->new_data_ready_changed.notify_all();
 434
 435                         if (field != num_fields - 1) {
 436                                 // Wait until the previous frame was consumed.
 437                                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 438                                 if (card->should_quit) return;
 439                         }
 440                 }
 441         }
 442 }
 443
 444 void Mixer::thread_func()
 445 {
 446         eglBindAPI(EGL_OPENGL_API);
 447         QOpenGLContext *context = create_context(mixer_surface);
 448         if (!make_current(context, mixer_surface)) {
 449                 printf("oops\n");
 450                 exit(1);
 451         }
 452
 453         struct timespec start, now;
 454         clock_gettime(CLOCK_MONOTONIC, &start);
 455
 456         int frame = 0;
 457         int stats_dropped_frames = 0;
 458
 459         while (!should_quit) {
 460                 CaptureCard card_copy[MAX_CARDS];
 461                 int num_samples[MAX_CARDS];
 462
 463                 {
 464                         unique_lock<mutex> lock(bmusb_mutex);
 465
 466                         // The first card is the master timer, so wait for it to have a new frame.
 467                         // TODO: Make configurable, and with a timeout.
 468                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 469
 470                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 471                                 CaptureCard *card = &cards[card_index];
 472                                 card_copy[card_index].usb = card->usb;
 473                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 474                                 card_copy[card_index].new_frame = card->new_frame;
 475                                 card_copy[card_index].new_frame_length = card->new_frame_length;
 476                                 card_copy[card_index].new_frame_field = card->new_frame_field;
 477                                 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
 478                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 479                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 480                                 card->new_data_ready = false;
 481                                 card->new_data_ready_changed.notify_all();
 482
 483                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
 484                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 485                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 486                                 assert(num_samples[card_index] >= 0);
 487                         }
 488                 }
 489
 490                 // Resample the audio as needed, including from previously dropped frames.
 491                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 492                         {
 493                                 // Signal to the audio thread to process this frame.
 494                                 unique_lock<mutex> lock(audio_mutex);
 495                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
 496                                 audio_task_queue_changed.notify_one();
 497                         }
 498                         if (frame_num != card_copy[0].dropped_frames) {
 499                                 // For dropped frames, increase the pts. Note that if the format changed
 500                                 // in the meantime, we have no way of detecting that; we just have to
 501                                 // assume the frame length is always the same.
 502                                 ++stats_dropped_frames;
 503                                 pts_int += card_copy[0].new_frame_length;
 504                         }
 505                 }
 506
 507                 if (audio_level_callback != nullptr) {
 508                         unique_lock<mutex> lock(r128_mutex);
 509                         double loudness_s = r128.loudness_S();
 510                         double loudness_i = r128.integrated();
 511                         double loudness_range_low = r128.range_min();
 512                         double loudness_range_high = r128.range_max();
 513
 514                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 515                                              loudness_i, loudness_range_low, loudness_range_high,
 516                                              last_gain_staging_db);
 517                 }
 518
 519                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 520                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 521                                 ++card_copy[card_index].dropped_frames;
 522                         }
 523                         if (card_copy[card_index].dropped_frames > 0) {
 524                                 printf("Card %u dropped %d frames before this\n",
 525                                         card_index, int(card_copy[card_index].dropped_frames));
 526                         }
 527                 }
 528
 529                 // If the first card is reporting a corrupted or otherwise dropped frame,
 530                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 531                 if (card_copy[0].new_frame->len == 0) {
 532                         ++stats_dropped_frames;
 533                         pts_int += card_copy[0].new_frame_length;
 534                         continue;
 535                 }
 536
 537                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 538                         CaptureCard *card = &card_copy[card_index];
 539                         if (!card->new_data_ready || card->new_frame->len == 0)
 540                                 continue;
 541
 542                         assert(card->new_frame != nullptr);
 543                         insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
 544                         check_error();
 545
 546                         // The new texture might still be uploaded,
 547                         // tell the GPU to wait until it's there.
 548                         if (card->new_data_ready_fence) {
 549                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 550                                 check_error();
 551                                 glDeleteSync(card->new_data_ready_fence);
 552                                 check_error();
 553                         }
 554                 }
 555
 556                 // Get the main chain from the theme, and set its state immediately.
 557                 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 558                 EffectChain *chain = theme_main_chain.chain;
 559                 theme_main_chain.setup_chain();
 560                 //theme_main_chain.chain->enable_phase_timing(true);
 561
 562                 GLuint y_tex, cbcr_tex;
 563                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 564                 assert(got_frame);
 565
 566                 // Render main chain.
 567                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 568                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 569                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 570                 check_error();
 571                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 572                 resource_pool->release_fbo(fbo);
 573
 574                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 575                 resource_pool->release_2d_texture(cbcr_full_tex);
 576
 577                 // Set the right state for rgba_tex.
 578                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 579                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 580                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 581                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 582                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 583
 584                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 585                 check_error();
 586
 587                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 588                 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
 589                 ++frame;
 590                 pts_int += card_copy[0].new_frame_length;
 591
 592                 // The live frame just shows the RGBA texture we just rendered.
 593                 // It owns rgba_tex now.
 594                 DisplayFrame live_frame;
 595                 live_frame.chain = display_chain.get();
 596                 live_frame.setup_chain = [this, rgba_tex]{
 597                         display_input->set_texture_num(rgba_tex);
 598                 };
 599                 live_frame.ready_fence = fence;
 600                 live_frame.input_frames = {};
 601                 live_frame.temp_textures = { rgba_tex };
 602                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 603
 604                 // Set up preview and any additional channels.
 605                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 606                         DisplayFrame display_frame;
 607                         Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 608                         display_frame.chain = chain.chain;
 609                         display_frame.setup_chain = chain.setup_chain;
 610                         display_frame.ready_fence = fence;
 611                         display_frame.input_frames = chain.input_frames;
 612                         display_frame.temp_textures = {};
 613                         output_channel[i].output_frame(display_frame);
 614                 }
 615
 616                 clock_gettime(CLOCK_MONOTONIC, &now);
 617                 double elapsed = now.tv_sec - start.tv_sec +
 618                         1e-9 * (now.tv_nsec - start.tv_nsec);
 619                 if (frame % 100 == 0) {
 620                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 621                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 622                                 1e3 * elapsed / frame);
 623                 //      chain->print_phase_timing();
 624                 }
 625
 626 #if 0
 627                 // Reset every 100 frames, so that local variations in frame times
 628                 // (especially for the first few frames, when the shaders are
 629                 // compiled etc.) don't make it hard to measure for the entire
 630                 // remaining duration of the program.
 631                 if (frame == 10000) {
 632                         frame = 0;
 633                         start = now;
 634                 }
 635 #endif
 636                 check_error();
 637         }
 638
 639         resource_pool->clean_context();
 640 }
 641
 642 void Mixer::audio_thread_func()
 643 {
 644         while (!should_quit) {
 645                 AudioTask task;
 646
 647                 {
 648                         unique_lock<mutex> lock(audio_mutex);
 649                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 650                         task = audio_task_queue.front();
 651                         audio_task_queue.pop();
 652                 }
 653
 654                 process_audio_one_frame(task.pts_int, task.num_samples);
 655         }
 656 }
 657
 658 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 659 {
 660         vector<float> samples_card;
 661         vector<float> samples_out;
 662         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 663                 samples_card.resize(num_samples * 2);
 664                 {
 665                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 666                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 667                                 printf("Card %d reported previous underrun.\n", card_index);
 668                         }
 669                 }
 670                 // TODO: Allow using audio from the other card(s) as well.
 671                 if (card_index == 0) {
 672                         samples_out = move(samples_card);
 673                 }
 674         }
 675
 676         // Cut away everything under 120 Hz (or whatever the cutoff is);
 677         // we don't need it for voice, and it will reduce headroom
 678         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 679         // should be dampened.)
 680         locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 681
 682         // Apply a level compressor to get the general level right.
 683         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 684         // (or more precisely, near it, since we don't use infinite ratio),
 685         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 686         // entirely arbitrary, but from practical tests with speech, it seems to
 687         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 688         float ref_level_dbfs = -14.0f;
 689         {
 690                 float threshold = 0.01f;   // -40 dBFS.
 691                 float ratio = 20.0f;
 692                 float attack_time = 0.5f;
 693                 float release_time = 20.0f;
 694                 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 695                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 696                 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 697         }
 698
 699 #if 0
 700         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 701                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 702                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 703                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 704 #endif
 705
 706 //      float limiter_att, compressor_att;
 707
 708         // The real compressor.
 709         if (compressor_enabled) {
 710                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 711                 float ratio = 20.0f;
 712                 float attack_time = 0.005f;
 713                 float release_time = 0.040f;
 714                 float makeup_gain = 2.0f;  // +6 dB.
 715                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 716 //              compressor_att = compressor.get_attenuation();
 717         }
 718
 719         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 720         // Note that since ratio is not infinite, we could go slightly higher than this.
 721         if (limiter_enabled) {
 722                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 723                 float ratio = 30.0f;
 724                 float attack_time = 0.0f;  // Instant.
 725                 float release_time = 0.020f;
 726                 float makeup_gain = 1.0f;  // 0 dB.
 727                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 728 //              limiter_att = limiter.get_attenuation();
 729         }
 730
 731 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 732
 733         // Upsample 4x to find interpolated peak.
 734         peak_resampler.inp_data = samples_out.data();
 735         peak_resampler.inp_count = samples_out.size() / 2;
 736
 737         vector<float> interpolated_samples_out;
 738         interpolated_samples_out.resize(samples_out.size());
 739         while (peak_resampler.inp_count > 0) {  // About four iterations.
 740                 peak_resampler.out_data = &interpolated_samples_out[0];
 741                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 742                 peak_resampler.process();
 743                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 744                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 745         }
 746
 747         // Find R128 levels.
 748         vector<float> left, right;
 749         deinterleave_samples(samples_out, &left, &right);
 750         float *ptrs[] = { left.data(), right.data() };
 751         {
 752                 unique_lock<mutex> lock(r128_mutex);
 753                 r128.process(left.size(), ptrs);
 754         }
 755
 756         // Send the samples to the sound card.
 757         if (alsa) {
 758                 alsa->write(samples_out);
 759         }
 760
 761         // And finally add them to the output.
 762         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 763 }
 764
 765 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 766 {
 767         GLuint vao;
 768         glGenVertexArrays(1, &vao);
 769         check_error();
 770
 771         float vertices[] = {
 772                 0.0f, 2.0f,
 773                 0.0f, 0.0f,
 774                 2.0f, 0.0f
 775         };
 776
 777         glBindVertexArray(vao);
 778         check_error();
 779
 780         // Extract Cb/Cr.
 781         GLuint fbo = resource_pool->create_fbo(dst_tex);
 782         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 783         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 784         check_error();
 785
 786         glUseProgram(cbcr_program_num);
 787         check_error();
 788
 789         glActiveTexture(GL_TEXTURE0);
 790         check_error();
 791         glBindTexture(GL_TEXTURE_2D, src_tex);
 792         check_error();
 793         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 794         check_error();
 795         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 796         check_error();
 797         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 798         check_error();
 799
 800         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 801         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 802
 803         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 804         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 805
 806         glDrawArrays(GL_TRIANGLES, 0, 3);
 807         check_error();
 808
 809         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 810         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 811
 812         glUseProgram(0);
 813         check_error();
 814
 815         resource_pool->release_fbo(fbo);
 816         glDeleteVertexArrays(1, &vao);
 817 }
 818
 819 void Mixer::release_display_frame(DisplayFrame *frame)
 820 {
 821         for (GLuint texnum : frame->temp_textures) {
 822                 resource_pool->release_2d_texture(texnum);
 823         }
 824         frame->temp_textures.clear();
 825         frame->ready_fence.reset();
 826         frame->input_frames.clear();
 827 }
 828
 829 void Mixer::start()
 830 {
 831         mixer_thread = thread(&Mixer::thread_func, this);
 832         audio_thread = thread(&Mixer::audio_thread_func, this);
 833 }
 834
 835 void Mixer::quit()
 836 {
 837         should_quit = true;
 838         mixer_thread.join();
 839         audio_thread.join();
 840 }
 841
 842 void Mixer::transition_clicked(int transition_num)
 843 {
 844         theme->transition_clicked(transition_num, pts());
 845 }
 846
 847 void Mixer::channel_clicked(int preview_num)
 848 {
 849         theme->channel_clicked(preview_num);
 850 }
 851
 852 void Mixer::reset_meters()
 853 {
 854         peak_resampler.reset();
 855         peak = 0.0f;
 856         r128.reset();
 857         r128.integr_start();
 858 }
 859
 860 Mixer::OutputChannel::~OutputChannel()
 861 {
 862         if (has_current_frame) {
 863                 parent->release_display_frame(&current_frame);
 864         }
 865         if (has_ready_frame) {
 866                 parent->release_display_frame(&ready_frame);
 867         }
 868 }
 869
 870 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 871 {
 872         // Store this frame for display. Remove the ready frame if any
 873         // (it was seemingly never used).
 874         {
 875                 unique_lock<mutex> lock(frame_mutex);
 876                 if (has_ready_frame) {
 877                         parent->release_display_frame(&ready_frame);
 878                 }
 879                 ready_frame = frame;
 880                 has_ready_frame = true;
 881         }
 882
 883         if (has_new_frame_ready_callback) {
 884                 new_frame_ready_callback();
 885         }
 886 }
 887
 888 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 889 {
 890         unique_lock<mutex> lock(frame_mutex);
 891         if (!has_current_frame && !has_ready_frame) {
 892                 return false;
 893         }
 894
 895         if (has_current_frame && has_ready_frame) {
 896                 // We have a new ready frame. Toss the current one.
 897                 parent->release_display_frame(&current_frame);
 898                 has_current_frame = false;
 899         }
 900         if (has_ready_frame) {
 901                 assert(!has_current_frame);
 902                 current_frame = ready_frame;
 903                 ready_frame.ready_fence.reset();  // Drop the refcount.
 904                 ready_frame.input_frames.clear();  // Drop the refcounts.
 905                 has_current_frame = true;
 906                 has_ready_frame = false;
 907         }
 908
 909         *frame = current_frame;
 910         return true;
 911 }
 912
 913 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 914 {
 915         new_frame_ready_callback = callback;
 916         has_new_frame_ready_callback = true;
 917 }