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

   1 #define WIDTH 1280
   2 #define HEIGHT 720
   3 #define EXTRAHEIGHT 30
   4
   5 #undef Success
   6
   7 #include "mixer.h"
   8
   9 #include <assert.h>
  10 #include <epoxy/egl.h>
  11 #include <init.h>
  12 #include <movit/effect_chain.h>
  13 #include <movit/effect_util.h>
  14 #include <movit/flat_input.h>
  15 #include <movit/image_format.h>
  16 #include <movit/resource_pool.h>
  17 #include <stdint.h>
  18 #include <stdio.h>
  19 #include <stdlib.h>
  20 #include <sys/time.h>
  21 #include <time.h>
  22 #include <util.h>
  23 #include <algorithm>
  24 #include <cmath>
  25 #include <condition_variable>
  26 #include <cstddef>
  27 #include <memory>
  28 #include <mutex>
  29 #include <string>
  30 #include <thread>
  31 #include <utility>
  32 #include <vector>
  33
  34 #include "bmusb/bmusb.h"
  35 #include "context.h"
  36 #include "defs.h"
  37 #include "h264encode.h"
  38 #include "pbo_frame_allocator.h"
  39 #include "ref_counted_gl_sync.h"
  40 #include "timebase.h"
  41
  42 class QOpenGLContext;
  43
  44 using namespace movit;
  45 using namespace std;
  46 using namespace std::placeholders;
  47
  48 Mixer *global_mixer = nullptr;
  49
  50 namespace {
  51
  52 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  53 {
  54         for (size_t i = 0; i < num_samples; ++i) {
  55                 for (size_t j = 0; j < out_channels; ++j) {
  56                         uint32_t s1 = *src++;
  57                         uint32_t s2 = *src++;
  58                         uint32_t s3 = *src++;
  59                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  60                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  61                 }
  62                 src += 3 * (in_channels - out_channels);
  63         }
  64 }
  65
  66 }  // namespace
  67
  68 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
  69         : httpd("test.ts", WIDTH, HEIGHT),
  70           num_cards(num_cards),
  71           mixer_surface(create_surface(format)),
  72           h264_encoder_surface(create_surface(format)),
  73           level_compressor(OUTPUT_FREQUENCY)
  74 {
  75         httpd.start(9095);
  76
  77         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
  78         check_error();
  79
  80         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
  81         // will be halved when sampling them, and we need to compensate here.
  82         movit_texel_subpixel_precision /= 2.0;
  83
  84         resource_pool.reset(new ResourcePool);
  85         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
  86         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
  87                 output_channel[i].parent = this;
  88         }
  89
  90         ImageFormat inout_format;
  91         inout_format.color_space = COLORSPACE_sRGB;
  92         inout_format.gamma_curve = GAMMA_sRGB;
  93
  94         // Display chain; shows the live output produced by the main chain (its RGBA version).
  95         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
  96         check_error();
  97         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
  98         display_chain->add_input(display_input);
  99         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 100         display_chain->set_dither_bits(0);  // Don't bother.
 101         display_chain->finalize();
 102
 103         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
 104
 105         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 106                 printf("Configuring card %d...\n", card_index);
 107                 CaptureCard *card = &cards[card_index];
 108                 card->usb = new BMUSBCapture(card_index);
 109                 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 110                 card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44, WIDTH, HEIGHT));
 111                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
 112                 card->surface = create_surface(format);
 113                 card->usb->set_dequeue_thread_callbacks(
 114                         [card]{
 115                                 eglBindAPI(EGL_OPENGL_API);
 116                                 card->context = create_context(card->surface);
 117                                 if (!make_current(card->context, card->surface)) {
 118                                         printf("failed to create bmusb context\n");
 119                                         exit(1);
 120                                 }
 121                         },
 122                         [this]{
 123                                 resource_pool->clean_context();
 124                         });
 125                 card->resampler.reset(new Resampler(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 126                 card->usb->configure_card();
 127         }
 128
 129         BMUSBCapture::start_bm_thread();
 130
 131         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 132                 cards[card_index].usb->start_bm_capture();
 133         }
 134
 135         //chain->enable_phase_timing(true);
 136
 137         // Set up stuff for NV12 conversion.
 138
 139         // Cb/Cr shader.
 140         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 141         string cbcr_frag_shader =
 142                 "#version 130 \n"
 143                 "in vec2 tc0; \n"
 144                 "uniform sampler2D cbcr_tex; \n"
 145                 "void main() { \n"
 146                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 147                 "} \n";
 148         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
 149
 150         r128.init(2, OUTPUT_FREQUENCY);
 151         r128.integr_start();
 152
 153         locut.init(FILTER_HPF, 2);
 154 }
 155
 156 Mixer::~Mixer()
 157 {
 158         resource_pool->release_glsl_program(cbcr_program_num);
 159         BMUSBCapture::stop_bm_thread();
 160
 161         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 162                 {
 163                         unique_lock<mutex> lock(bmusb_mutex);
 164                         cards[card_index].should_quit = true;  // Unblock thread.
 165                         cards[card_index].new_data_ready_changed.notify_all();
 166                 }
 167                 cards[card_index].usb->stop_dequeue_thread();
 168         }
 169 }
 170
 171 namespace {
 172
 173 int unwrap_timecode(uint16_t current_wrapped, int last)
 174 {
 175         uint16_t last_wrapped = last & 0xffff;
 176         if (current_wrapped > last_wrapped) {
 177                 return (last & ~0xffff) | current_wrapped;
 178         } else {
 179                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 180         }
 181 }
 182
 183 float find_peak(const vector<float> &samples)
 184 {
 185         float m = fabs(samples[0]);
 186         for (size_t i = 1; i < samples.size(); ++i) {
 187                 m = std::max(m, fabs(samples[i]));
 188         }
 189         return m;
 190 }
 191
 192 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 193 {
 194         size_t num_samples = in.size() / 2;
 195         out_l->resize(num_samples);
 196         out_r->resize(num_samples);
 197
 198         const float *inptr = in.data();
 199         float *lptr = &(*out_l)[0];
 200         float *rptr = &(*out_r)[0];
 201         for (size_t i = 0; i < num_samples; ++i) {
 202                 *lptr++ = *inptr++;
 203                 *rptr++ = *inptr++;
 204         }
 205 }
 206
 207 }  // namespace
 208
 209 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 210                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 211                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 212 {
 213         CaptureCard *card = &cards[card_index];
 214
 215         if (audio_frame.len - audio_offset > 30000) {
 216                 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",
 217                         card_index, int(audio_frame.len), int(audio_offset),
 218                         timecode, int(video_frame.len), int(video_offset), video_format);
 219                 if (video_frame.owner) {
 220                         video_frame.owner->release_frame(video_frame);
 221                 }
 222                 if (audio_frame.owner) {
 223                         audio_frame.owner->release_frame(audio_frame);
 224                 }
 225                 return;
 226         }
 227
 228         int unwrapped_timecode = timecode;
 229         int dropped_frames = 0;
 230         if (card->last_timecode != -1) {
 231                 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
 232                 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
 233         }
 234         card->last_timecode = unwrapped_timecode;
 235
 236         // Convert the audio to stereo fp32 and add it.
 237         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 238         vector<float> audio;
 239         audio.resize(num_samples * 2);
 240         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 241
 242         // Add the audio.
 243         {
 244                 unique_lock<mutex> lock(card->audio_mutex);
 245
 246                 int unwrapped_timecode = timecode;
 247                 if (dropped_frames > FPS * 2) {
 248                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
 249                                 card_index);
 250                         card->resampler.reset(new Resampler(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 251                 } else if (dropped_frames > 0) {
 252                         // Insert silence as needed.
 253                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 254                                 card_index, dropped_frames, timecode);
 255                         vector<float> silence;
 256                         silence.resize((OUTPUT_FREQUENCY / FPS) * 2);
 257                         for (int i = 0; i < dropped_frames; ++i) {
 258                                 card->resampler->add_input_samples((unwrapped_timecode - dropped_frames + i) / double(FPS), silence.data(), (OUTPUT_FREQUENCY / FPS));
 259                         }
 260                 }
 261                 card->resampler->add_input_samples(unwrapped_timecode / double(FPS), audio.data(), num_samples);
 262         }
 263
 264         // Done with the audio, so release it.
 265         if (audio_frame.owner) {
 266                 audio_frame.owner->release_frame(audio_frame);
 267         }
 268
 269         {
 270                 // Wait until the previous frame was consumed.
 271                 unique_lock<mutex> lock(bmusb_mutex);
 272                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 273                 if (card->should_quit) return;
 274         }
 275
 276         if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
 277                 if (video_frame.len != 0) {
 278                         printf("Card %d: Dropping video frame with wrong length (%ld)\n",
 279                                 card_index, video_frame.len - video_offset);
 280                 }
 281                 if (video_frame.owner) {
 282                         video_frame.owner->release_frame(video_frame);
 283                 }
 284
 285                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 286                 // so that pts can go up accordingly.
 287                 {
 288                         unique_lock<mutex> lock(bmusb_mutex);
 289                         card->new_data_ready = true;
 290                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 291                         card->new_data_ready_fence = nullptr;
 292                         card->dropped_frames = dropped_frames;
 293                         card->new_data_ready_changed.notify_all();
 294                 }
 295                 return;
 296         }
 297
 298         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
 299         GLuint pbo = userdata->pbo;
 300         check_error();
 301         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 302         check_error();
 303         glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
 304         check_error();
 305         //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 306         //check_error();
 307
 308         // Upload the textures.
 309         glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
 310         check_error();
 311         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH, HEIGHT, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44) / 2 + WIDTH * 25 + 22));
 312         check_error();
 313         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
 314         check_error();
 315         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
 316         check_error();
 317         glBindTexture(GL_TEXTURE_2D, 0);
 318         check_error();
 319         GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 320         check_error();
 321         assert(fence != nullptr);
 322
 323         {
 324                 unique_lock<mutex> lock(bmusb_mutex);
 325                 card->new_data_ready = true;
 326                 card->new_frame = RefCountedFrame(video_frame);
 327                 card->new_data_ready_fence = fence;
 328                 card->dropped_frames = dropped_frames;
 329                 card->new_data_ready_changed.notify_all();
 330         }
 331 }
 332
 333 void Mixer::thread_func()
 334 {
 335         eglBindAPI(EGL_OPENGL_API);
 336         QOpenGLContext *context = create_context(mixer_surface);
 337         if (!make_current(context, mixer_surface)) {
 338                 printf("oops\n");
 339                 exit(1);
 340         }
 341
 342         struct timespec start, now;
 343         clock_gettime(CLOCK_MONOTONIC, &start);
 344
 345         int frame = 0;
 346         int dropped_frames = 0;
 347
 348         while (!should_quit) {
 349                 CaptureCard card_copy[MAX_CARDS];
 350
 351                 {
 352                         unique_lock<mutex> lock(bmusb_mutex);
 353
 354                         // The first card is the master timer, so wait for it to have a new frame.
 355                         // TODO: Make configurable, and with a timeout.
 356                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 357
 358                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 359                                 CaptureCard *card = &cards[card_index];
 360                                 card_copy[card_index].usb = card->usb;
 361                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 362                                 card_copy[card_index].new_frame = card->new_frame;
 363                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 364                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 365                                 card->new_data_ready = false;
 366                                 card->new_data_ready_changed.notify_all();
 367                         }
 368                 }
 369
 370                 // Resample the audio as needed, including from previously dropped frames.
 371                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 372                         process_audio_one_frame();
 373                         if (frame_num != card_copy[0].dropped_frames) {
 374                                 // For dropped frames, increase the pts.
 375                                 ++dropped_frames;
 376                                 pts_int += TIMEBASE / FPS;
 377                         }
 378                 }
 379
 380                 if (audio_level_callback != nullptr) {
 381                         double loudness_s = r128.loudness_S();
 382                         double loudness_i = r128.integrated();
 383                         double loudness_range_low = r128.range_min();
 384                         double loudness_range_high = r128.range_max();
 385
 386                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 387                                              loudness_i, loudness_range_low, loudness_range_high,
 388                                              last_gain_staging_db);
 389                 }
 390
 391                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 392                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 393                                 ++card_copy[card_index].dropped_frames;
 394                         }
 395                         if (card_copy[card_index].dropped_frames > 0) {
 396                                 printf("Card %u dropped %d frames before this\n",
 397                                         card_index, int(card_copy[card_index].dropped_frames));
 398                         }
 399                 }
 400
 401                 // If the first card is reporting a corrupted or otherwise dropped frame,
 402                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 403                 if (card_copy[0].new_frame->len == 0) {
 404                         ++dropped_frames;
 405                         pts_int += TIMEBASE / FPS;
 406                         continue;
 407                 }
 408
 409                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 410                         CaptureCard *card = &card_copy[card_index];
 411                         if (!card->new_data_ready || card->new_frame->len == 0)
 412                                 continue;
 413
 414                         assert(card->new_frame != nullptr);
 415                         bmusb_current_rendering_frame[card_index] = card->new_frame;
 416                         check_error();
 417
 418                         // The new texture might still be uploaded,
 419                         // tell the GPU to wait until it's there.
 420                         if (card->new_data_ready_fence) {
 421                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 422                                 check_error();
 423                                 glDeleteSync(card->new_data_ready_fence);
 424                                 check_error();
 425                         }
 426                         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
 427                         theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
 428                 }
 429
 430                 // Get the main chain from the theme, and set its state immediately.
 431                 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
 432                 EffectChain *chain = theme_main_chain.first;
 433                 theme_main_chain.second();
 434
 435                 GLuint y_tex, cbcr_tex;
 436                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 437                 assert(got_frame);
 438
 439                 // Render main chain.
 440                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 441                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 442                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 443                 check_error();
 444                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 445                 resource_pool->release_fbo(fbo);
 446
 447                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 448                 resource_pool->release_2d_texture(cbcr_full_tex);
 449
 450                 // Set the right state for rgba_tex.
 451                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 452                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 453                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 454                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 455                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 456
 457                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 458                 check_error();
 459
 460                 // Make sure the H.264 gets a reference to all the
 461                 // input frames needed, so that they are not released back
 462                 // until the rendering is done.
 463                 vector<RefCountedFrame> input_frames;
 464                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 465                         input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 466                 }
 467                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampler.h. TODO: Make less hard-coded.
 468                 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
 469                 ++frame;
 470                 pts_int += TIMEBASE / FPS;
 471
 472                 // The live frame just shows the RGBA texture we just rendered.
 473                 // It owns rgba_tex now.
 474                 DisplayFrame live_frame;
 475                 live_frame.chain = display_chain.get();
 476                 live_frame.setup_chain = [this, rgba_tex]{
 477                         display_input->set_texture_num(rgba_tex);
 478                 };
 479                 live_frame.ready_fence = fence;
 480                 live_frame.input_frames = {};
 481                 live_frame.temp_textures = { rgba_tex };
 482                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 483
 484                 // Set up preview and any additional channels.
 485                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 486                         DisplayFrame display_frame;
 487                         pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT);  // FIXME: dimensions
 488                         display_frame.chain = chain.first;
 489                         display_frame.setup_chain = chain.second;
 490                         display_frame.ready_fence = fence;
 491
 492                         // FIXME: possible to do better?
 493                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 494                                 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 495                         }
 496                         display_frame.temp_textures = {};
 497                         output_channel[i].output_frame(display_frame);
 498                 }
 499
 500                 clock_gettime(CLOCK_MONOTONIC, &now);
 501                 double elapsed = now.tv_sec - start.tv_sec +
 502                         1e-9 * (now.tv_nsec - start.tv_nsec);
 503                 if (frame % 100 == 0) {
 504                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 505                                 frame, dropped_frames, elapsed, frame / elapsed,
 506                                 1e3 * elapsed / frame);
 507                 //      chain->print_phase_timing();
 508                 }
 509
 510 #if 0
 511                 // Reset every 100 frames, so that local variations in frame times
 512                 // (especially for the first few frames, when the shaders are
 513                 // compiled etc.) don't make it hard to measure for the entire
 514                 // remaining duration of the program.
 515                 if (frame == 10000) {
 516                         frame = 0;
 517                         start = now;
 518                 }
 519 #endif
 520                 check_error();
 521         }
 522
 523         resource_pool->clean_context();
 524 }
 525
 526 void Mixer::process_audio_one_frame()
 527 {
 528         vector<float> samples_card;
 529         vector<float> samples_out;
 530         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 531                 samples_card.resize((OUTPUT_FREQUENCY / FPS) * 2);
 532                 {
 533                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 534                         if (!cards[card_index].resampler->get_output_samples(pts(), &samples_card[0], OUTPUT_FREQUENCY / FPS)) {
 535                                 printf("Card %d reported previous underrun.\n", card_index);
 536                         }
 537                 }
 538                 // TODO: Allow using audio from the other card(s) as well.
 539                 if (card_index == 0) {
 540                         samples_out = move(samples_card);
 541                 }
 542         }
 543
 544         // Cut away everything under 150 Hz; we don't need it for voice,
 545         // and it will reduce headroom and confuse the compressor.
 546         // (In particular, any hums at 50 or 60 Hz should be dampened.)
 547         locut.render(samples_out.data(), samples_out.size() / 2, 150.0 * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 548
 549         // Apply a level compressor to get the general level right.
 550         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 551         // (or more precisely, near it, since we don't use infinite ratio),
 552         // then apply a makeup gain to get it to -12 dBFS. -12 dBFS is, of course,
 553         // entirely arbitrary, but from practical tests with speech, it seems to
 554         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 555         //
 556         // TODO: Hook this up to a UI, so we can see the effects, and/or turn it off
 557         // to control the gain manually instead. For now, there's only the #if-ed out
 558         // code below.
 559         //
 560         // TODO: Add the actual compressors/limiters (for taking care of transients)
 561         // later in the chain.
 562         float threshold = 0.01f;   // -40 dBFS.
 563         float ratio = 20.0f;
 564         float attack_time = 0.1f;
 565         float release_time = 10.0f;
 566         float makeup_gain = pow(10.0f, 28.0f / 20.0f);  // +28 dB takes us to -12 dBFS.
 567         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 568         last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 569
 570 #if 0
 571         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 572                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 573                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 574                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 575 #endif
 576
 577         // Find peak and R128 levels.
 578         peak = std::max(peak, find_peak(samples_out));
 579         vector<float> left, right;
 580         deinterleave_samples(samples_out, &left, &right);
 581         float *ptrs[] = { left.data(), right.data() };
 582         r128.process(left.size(), ptrs);
 583
 584         // Actually add the samples to the output.
 585         h264_encoder->add_audio(pts_int, move(samples_out));
 586 }
 587
 588 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 589 {
 590         GLuint vao;
 591         glGenVertexArrays(1, &vao);
 592         check_error();
 593
 594         float vertices[] = {
 595                 0.0f, 2.0f,
 596                 0.0f, 0.0f,
 597                 2.0f, 0.0f
 598         };
 599
 600         glBindVertexArray(vao);
 601         check_error();
 602
 603         // Extract Cb/Cr.
 604         GLuint fbo = resource_pool->create_fbo(dst_tex);
 605         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 606         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 607         check_error();
 608
 609         glUseProgram(cbcr_program_num);
 610         check_error();
 611
 612         glActiveTexture(GL_TEXTURE0);
 613         check_error();
 614         glBindTexture(GL_TEXTURE_2D, src_tex);
 615         check_error();
 616         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 617         check_error();
 618         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 619         check_error();
 620         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 621         check_error();
 622
 623         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 624         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 625
 626         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 627         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 628
 629         glDrawArrays(GL_TRIANGLES, 0, 3);
 630         check_error();
 631
 632         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 633         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 634
 635         glUseProgram(0);
 636         check_error();
 637
 638         resource_pool->release_fbo(fbo);
 639         glDeleteVertexArrays(1, &vao);
 640 }
 641
 642 void Mixer::release_display_frame(DisplayFrame *frame)
 643 {
 644         for (GLuint texnum : frame->temp_textures) {
 645                 resource_pool->release_2d_texture(texnum);
 646         }
 647         frame->temp_textures.clear();
 648         frame->ready_fence.reset();
 649         frame->input_frames.clear();
 650 }
 651
 652 void Mixer::start()
 653 {
 654         mixer_thread = thread(&Mixer::thread_func, this);
 655 }
 656
 657 void Mixer::quit()
 658 {
 659         should_quit = true;
 660         mixer_thread.join();
 661 }
 662
 663 void Mixer::transition_clicked(int transition_num)
 664 {
 665         theme->transition_clicked(transition_num, pts());
 666 }
 667
 668 void Mixer::channel_clicked(int preview_num)
 669 {
 670         theme->channel_clicked(preview_num);
 671 }
 672
 673 Mixer::OutputChannel::~OutputChannel()
 674 {
 675         if (has_current_frame) {
 676                 parent->release_display_frame(&current_frame);
 677         }
 678         if (has_ready_frame) {
 679                 parent->release_display_frame(&ready_frame);
 680         }
 681 }
 682
 683 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 684 {
 685         // Store this frame for display. Remove the ready frame if any
 686         // (it was seemingly never used).
 687         {
 688                 unique_lock<mutex> lock(frame_mutex);
 689                 if (has_ready_frame) {
 690                         parent->release_display_frame(&ready_frame);
 691                 }
 692                 ready_frame = frame;
 693                 has_ready_frame = true;
 694         }
 695
 696         if (has_new_frame_ready_callback) {
 697                 new_frame_ready_callback();
 698         }
 699 }
 700
 701 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 702 {
 703         unique_lock<mutex> lock(frame_mutex);
 704         if (!has_current_frame && !has_ready_frame) {
 705                 return false;
 706         }
 707
 708         if (has_current_frame && has_ready_frame) {
 709                 // We have a new ready frame. Toss the current one.
 710                 parent->release_display_frame(&current_frame);
 711                 has_current_frame = false;
 712         }
 713         if (has_ready_frame) {
 714                 assert(!has_current_frame);
 715                 current_frame = ready_frame;
 716                 ready_frame.ready_fence.reset();  // Drop the refcount.
 717                 ready_frame.input_frames.clear();  // Drop the refcounts.
 718                 has_current_frame = true;
 719                 has_ready_frame = false;
 720         }
 721
 722         *frame = current_frame;
 723         return true;
 724 }
 725
 726 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 727 {
 728         new_frame_ready_callback = callback;
 729         has_new_frame_ready_callback = true;
 730 }