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           limiter(OUTPUT_FREQUENCY),
  75           compressor(OUTPUT_FREQUENCY)
  76 {
  77         httpd.start(9095);
  78
  79         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
  80         check_error();
  81
  82         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
  83         // will be halved when sampling them, and we need to compensate here.
  84         movit_texel_subpixel_precision /= 2.0;
  85
  86         resource_pool.reset(new ResourcePool);
  87         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
  88         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
  89                 output_channel[i].parent = this;
  90         }
  91
  92         ImageFormat inout_format;
  93         inout_format.color_space = COLORSPACE_sRGB;
  94         inout_format.gamma_curve = GAMMA_sRGB;
  95
  96         // Display chain; shows the live output produced by the main chain (its RGBA version).
  97         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
  98         check_error();
  99         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 100         display_chain->add_input(display_input);
 101         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 102         display_chain->set_dither_bits(0);  // Don't bother.
 103         display_chain->finalize();
 104
 105         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
 106
 107         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 108                 printf("Configuring card %d...\n", card_index);
 109                 CaptureCard *card = &cards[card_index];
 110                 card->usb = new BMUSBCapture(card_index);
 111                 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 112                 card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44, WIDTH, HEIGHT));
 113                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
 114                 card->surface = create_surface(format);
 115                 card->usb->set_dequeue_thread_callbacks(
 116                         [card]{
 117                                 eglBindAPI(EGL_OPENGL_API);
 118                                 card->context = create_context(card->surface);
 119                                 if (!make_current(card->context, card->surface)) {
 120                                         printf("failed to create bmusb context\n");
 121                                         exit(1);
 122                                 }
 123                         },
 124                         [this]{
 125                                 resource_pool->clean_context();
 126                         });
 127                 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 128                 card->usb->configure_card();
 129         }
 130
 131         BMUSBCapture::start_bm_thread();
 132
 133         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 134                 cards[card_index].usb->start_bm_capture();
 135         }
 136
 137         //chain->enable_phase_timing(true);
 138
 139         // Set up stuff for NV12 conversion.
 140
 141         // Cb/Cr shader.
 142         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 143         string cbcr_frag_shader =
 144                 "#version 130 \n"
 145                 "in vec2 tc0; \n"
 146                 "uniform sampler2D cbcr_tex; \n"
 147                 "void main() { \n"
 148                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 149                 "} \n";
 150         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
 151
 152         r128.init(2, OUTPUT_FREQUENCY);
 153         r128.integr_start();
 154
 155         locut.init(FILTER_HPF, 2);
 156 }
 157
 158 Mixer::~Mixer()
 159 {
 160         resource_pool->release_glsl_program(cbcr_program_num);
 161         BMUSBCapture::stop_bm_thread();
 162
 163         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 164                 {
 165                         unique_lock<mutex> lock(bmusb_mutex);
 166                         cards[card_index].should_quit = true;  // Unblock thread.
 167                         cards[card_index].new_data_ready_changed.notify_all();
 168                 }
 169                 cards[card_index].usb->stop_dequeue_thread();
 170         }
 171 }
 172
 173 namespace {
 174
 175 int unwrap_timecode(uint16_t current_wrapped, int last)
 176 {
 177         uint16_t last_wrapped = last & 0xffff;
 178         if (current_wrapped > last_wrapped) {
 179                 return (last & ~0xffff) | current_wrapped;
 180         } else {
 181                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 182         }
 183 }
 184
 185 float find_peak(const vector<float> &samples)
 186 {
 187         float m = fabs(samples[0]);
 188         for (size_t i = 1; i < samples.size(); ++i) {
 189                 m = std::max(m, fabs(samples[i]));
 190         }
 191         return m;
 192 }
 193
 194 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 195 {
 196         size_t num_samples = in.size() / 2;
 197         out_l->resize(num_samples);
 198         out_r->resize(num_samples);
 199
 200         const float *inptr = in.data();
 201         float *lptr = &(*out_l)[0];
 202         float *rptr = &(*out_r)[0];
 203         for (size_t i = 0; i < num_samples; ++i) {
 204                 *lptr++ = *inptr++;
 205                 *rptr++ = *inptr++;
 206         }
 207 }
 208
 209 }  // namespace
 210
 211 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 212                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 213                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 214 {
 215         CaptureCard *card = &cards[card_index];
 216
 217         if (audio_frame.len - audio_offset > 30000) {
 218                 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",
 219                         card_index, int(audio_frame.len), int(audio_offset),
 220                         timecode, int(video_frame.len), int(video_offset), video_format);
 221                 if (video_frame.owner) {
 222                         video_frame.owner->release_frame(video_frame);
 223                 }
 224                 if (audio_frame.owner) {
 225                         audio_frame.owner->release_frame(audio_frame);
 226                 }
 227                 return;
 228         }
 229
 230         int unwrapped_timecode = timecode;
 231         int dropped_frames = 0;
 232         if (card->last_timecode != -1) {
 233                 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
 234                 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
 235         }
 236         card->last_timecode = unwrapped_timecode;
 237
 238         // Convert the audio to stereo fp32 and add it.
 239         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 240         vector<float> audio;
 241         audio.resize(num_samples * 2);
 242         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 243
 244         // Add the audio.
 245         {
 246                 unique_lock<mutex> lock(card->audio_mutex);
 247
 248                 int unwrapped_timecode = timecode;
 249                 if (dropped_frames > FPS * 2) {
 250                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
 251                                 card_index);
 252                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 253                 } else if (dropped_frames > 0) {
 254                         // Insert silence as needed.
 255                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 256                                 card_index, dropped_frames, timecode);
 257                         vector<float> silence;
 258                         silence.resize((OUTPUT_FREQUENCY / FPS) * 2);
 259                         for (int i = 0; i < dropped_frames; ++i) {
 260                                 card->resampling_queue->add_input_samples((unwrapped_timecode - dropped_frames + i) / double(FPS), silence.data(), (OUTPUT_FREQUENCY / FPS));
 261                         }
 262                 }
 263                 card->resampling_queue->add_input_samples(unwrapped_timecode / double(FPS), audio.data(), num_samples);
 264         }
 265
 266         // Done with the audio, so release it.
 267         if (audio_frame.owner) {
 268                 audio_frame.owner->release_frame(audio_frame);
 269         }
 270
 271         {
 272                 // Wait until the previous frame was consumed.
 273                 unique_lock<mutex> lock(bmusb_mutex);
 274                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 275                 if (card->should_quit) return;
 276         }
 277
 278         if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
 279                 if (video_frame.len != 0) {
 280                         printf("Card %d: Dropping video frame with wrong length (%ld)\n",
 281                                 card_index, video_frame.len - video_offset);
 282                 }
 283                 if (video_frame.owner) {
 284                         video_frame.owner->release_frame(video_frame);
 285                 }
 286
 287                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 288                 // so that pts can go up accordingly.
 289                 {
 290                         unique_lock<mutex> lock(bmusb_mutex);
 291                         card->new_data_ready = true;
 292                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 293                         card->new_data_ready_fence = nullptr;
 294                         card->dropped_frames = dropped_frames;
 295                         card->new_data_ready_changed.notify_all();
 296                 }
 297                 return;
 298         }
 299
 300         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
 301         GLuint pbo = userdata->pbo;
 302         check_error();
 303         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 304         check_error();
 305         glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
 306         check_error();
 307         //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 308         //check_error();
 309
 310         // Upload the textures.
 311         glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
 312         check_error();
 313         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));
 314         check_error();
 315         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
 316         check_error();
 317         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
 318         check_error();
 319         glBindTexture(GL_TEXTURE_2D, 0);
 320         check_error();
 321         GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 322         check_error();
 323         assert(fence != nullptr);
 324
 325         {
 326                 unique_lock<mutex> lock(bmusb_mutex);
 327                 card->new_data_ready = true;
 328                 card->new_frame = RefCountedFrame(video_frame);
 329                 card->new_data_ready_fence = fence;
 330                 card->dropped_frames = dropped_frames;
 331                 card->new_data_ready_changed.notify_all();
 332         }
 333 }
 334
 335 void Mixer::thread_func()
 336 {
 337         eglBindAPI(EGL_OPENGL_API);
 338         QOpenGLContext *context = create_context(mixer_surface);
 339         if (!make_current(context, mixer_surface)) {
 340                 printf("oops\n");
 341                 exit(1);
 342         }
 343
 344         struct timespec start, now;
 345         clock_gettime(CLOCK_MONOTONIC, &start);
 346
 347         int frame = 0;
 348         int dropped_frames = 0;
 349
 350         while (!should_quit) {
 351                 CaptureCard card_copy[MAX_CARDS];
 352
 353                 {
 354                         unique_lock<mutex> lock(bmusb_mutex);
 355
 356                         // The first card is the master timer, so wait for it to have a new frame.
 357                         // TODO: Make configurable, and with a timeout.
 358                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 359
 360                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 361                                 CaptureCard *card = &cards[card_index];
 362                                 card_copy[card_index].usb = card->usb;
 363                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 364                                 card_copy[card_index].new_frame = card->new_frame;
 365                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 366                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 367                                 card->new_data_ready = false;
 368                                 card->new_data_ready_changed.notify_all();
 369                         }
 370                 }
 371
 372                 // Resample the audio as needed, including from previously dropped frames.
 373                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 374                         process_audio_one_frame();
 375                         if (frame_num != card_copy[0].dropped_frames) {
 376                                 // For dropped frames, increase the pts.
 377                                 ++dropped_frames;
 378                                 pts_int += TIMEBASE / FPS;
 379                         }
 380                 }
 381
 382                 if (audio_level_callback != nullptr) {
 383                         double loudness_s = r128.loudness_S();
 384                         double loudness_i = r128.integrated();
 385                         double loudness_range_low = r128.range_min();
 386                         double loudness_range_high = r128.range_max();
 387
 388                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 389                                              loudness_i, loudness_range_low, loudness_range_high,
 390                                              last_gain_staging_db);
 391                 }
 392
 393                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 394                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 395                                 ++card_copy[card_index].dropped_frames;
 396                         }
 397                         if (card_copy[card_index].dropped_frames > 0) {
 398                                 printf("Card %u dropped %d frames before this\n",
 399                                         card_index, int(card_copy[card_index].dropped_frames));
 400                         }
 401                 }
 402
 403                 // If the first card is reporting a corrupted or otherwise dropped frame,
 404                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 405                 if (card_copy[0].new_frame->len == 0) {
 406                         ++dropped_frames;
 407                         pts_int += TIMEBASE / FPS;
 408                         continue;
 409                 }
 410
 411                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 412                         CaptureCard *card = &card_copy[card_index];
 413                         if (!card->new_data_ready || card->new_frame->len == 0)
 414                                 continue;
 415
 416                         assert(card->new_frame != nullptr);
 417                         bmusb_current_rendering_frame[card_index] = card->new_frame;
 418                         check_error();
 419
 420                         // The new texture might still be uploaded,
 421                         // tell the GPU to wait until it's there.
 422                         if (card->new_data_ready_fence) {
 423                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 424                                 check_error();
 425                                 glDeleteSync(card->new_data_ready_fence);
 426                                 check_error();
 427                         }
 428                         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
 429                         theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
 430                 }
 431
 432                 // Get the main chain from the theme, and set its state immediately.
 433                 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
 434                 EffectChain *chain = theme_main_chain.first;
 435                 theme_main_chain.second();
 436
 437                 GLuint y_tex, cbcr_tex;
 438                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 439                 assert(got_frame);
 440
 441                 // Render main chain.
 442                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 443                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 444                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 445                 check_error();
 446                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 447                 resource_pool->release_fbo(fbo);
 448
 449                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 450                 resource_pool->release_2d_texture(cbcr_full_tex);
 451
 452                 // Set the right state for rgba_tex.
 453                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 454                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 455                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 456                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 457                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 458
 459                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 460                 check_error();
 461
 462                 // Make sure the H.264 gets a reference to all the
 463                 // input frames needed, so that they are not released back
 464                 // until the rendering is done.
 465                 vector<RefCountedFrame> input_frames;
 466                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 467                         input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 468                 }
 469                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 470                 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
 471                 ++frame;
 472                 pts_int += TIMEBASE / FPS;
 473
 474                 // The live frame just shows the RGBA texture we just rendered.
 475                 // It owns rgba_tex now.
 476                 DisplayFrame live_frame;
 477                 live_frame.chain = display_chain.get();
 478                 live_frame.setup_chain = [this, rgba_tex]{
 479                         display_input->set_texture_num(rgba_tex);
 480                 };
 481                 live_frame.ready_fence = fence;
 482                 live_frame.input_frames = {};
 483                 live_frame.temp_textures = { rgba_tex };
 484                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 485
 486                 // Set up preview and any additional channels.
 487                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 488                         DisplayFrame display_frame;
 489                         pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT);  // FIXME: dimensions
 490                         display_frame.chain = chain.first;
 491                         display_frame.setup_chain = chain.second;
 492                         display_frame.ready_fence = fence;
 493
 494                         // FIXME: possible to do better?
 495                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 496                                 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 497                         }
 498                         display_frame.temp_textures = {};
 499                         output_channel[i].output_frame(display_frame);
 500                 }
 501
 502                 clock_gettime(CLOCK_MONOTONIC, &now);
 503                 double elapsed = now.tv_sec - start.tv_sec +
 504                         1e-9 * (now.tv_nsec - start.tv_nsec);
 505                 if (frame % 100 == 0) {
 506                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 507                                 frame, dropped_frames, elapsed, frame / elapsed,
 508                                 1e3 * elapsed / frame);
 509                 //      chain->print_phase_timing();
 510                 }
 511
 512 #if 0
 513                 // Reset every 100 frames, so that local variations in frame times
 514                 // (especially for the first few frames, when the shaders are
 515                 // compiled etc.) don't make it hard to measure for the entire
 516                 // remaining duration of the program.
 517                 if (frame == 10000) {
 518                         frame = 0;
 519                         start = now;
 520                 }
 521 #endif
 522                 check_error();
 523         }
 524
 525         resource_pool->clean_context();
 526 }
 527
 528 void Mixer::process_audio_one_frame()
 529 {
 530         vector<float> samples_card;
 531         vector<float> samples_out;
 532         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 533                 samples_card.resize((OUTPUT_FREQUENCY / FPS) * 2);
 534                 {
 535                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 536                         if (!cards[card_index].resampling_queue->get_output_samples(pts(), &samples_card[0], OUTPUT_FREQUENCY / FPS)) {
 537                                 printf("Card %d reported previous underrun.\n", card_index);
 538                         }
 539                 }
 540                 // TODO: Allow using audio from the other card(s) as well.
 541                 if (card_index == 0) {
 542                         samples_out = move(samples_card);
 543                 }
 544         }
 545
 546         // Cut away everything under 150 Hz; we don't need it for voice,
 547         // and it will reduce headroom and confuse the compressor.
 548         // (In particular, any hums at 50 or 60 Hz should be dampened.)
 549         locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 550
 551         // Apply a level compressor to get the general level right.
 552         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 553         // (or more precisely, near it, since we don't use infinite ratio),
 554         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 555         // entirely arbitrary, but from practical tests with speech, it seems to
 556         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 557         float ref_level_dbfs = -14.0f;
 558         {
 559                 float threshold = 0.01f;   // -40 dBFS.
 560                 float ratio = 20.0f;
 561                 float attack_time = 0.5f;
 562                 float release_time = 20.0f;
 563                 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 564                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 565                 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 566         }
 567
 568 #if 0
 569         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 570                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 571                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 572                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 573 #endif
 574
 575 //      float limiter_att, compressor_att;
 576
 577         // Then a limiter at +0 dB (so, -14 dBFS) to take out the worst peaks only.
 578         // Note that since ratio is not infinite, we could go slightly higher than this.
 579         // Probably more tuning is warranted here.
 580         {
 581                 float threshold = pow(10.0f, (ref_level_dbfs + 0.0f) / 20.0f);  // +0 dB.
 582                 float ratio = 30.0f;
 583                 float attack_time = 0.0f;  // Instant.
 584                 float release_time = 0.005f;
 585                 float makeup_gain = 1.0f;  // 0 dB.
 586                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 587 //              limiter_att = limiter.get_attenuation();
 588         }
 589
 590         // Finally, the real compressor.
 591         {
 592                 float threshold = pow(10.0f, (ref_level_dbfs - 12.0f) / 20.0f);  // -12 dB.
 593                 float ratio = 20.0f;
 594                 float attack_time = 0.005f;
 595                 float release_time = 0.040f;
 596                 float makeup_gain = 2.0f;  // +6 dB.
 597                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 598 //              compressor_att = compressor.get_attenuation();
 599         }
 600
 601 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 602
 603         // Find peak and R128 levels.
 604         peak = max<float>(peak, find_peak(samples_out));
 605         vector<float> left, right;
 606         deinterleave_samples(samples_out, &left, &right);
 607         float *ptrs[] = { left.data(), right.data() };
 608         r128.process(left.size(), ptrs);
 609
 610         // Actually add the samples to the output.
 611         h264_encoder->add_audio(pts_int, move(samples_out));
 612 }
 613
 614 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 615 {
 616         GLuint vao;
 617         glGenVertexArrays(1, &vao);
 618         check_error();
 619
 620         float vertices[] = {
 621                 0.0f, 2.0f,
 622                 0.0f, 0.0f,
 623                 2.0f, 0.0f
 624         };
 625
 626         glBindVertexArray(vao);
 627         check_error();
 628
 629         // Extract Cb/Cr.
 630         GLuint fbo = resource_pool->create_fbo(dst_tex);
 631         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 632         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 633         check_error();
 634
 635         glUseProgram(cbcr_program_num);
 636         check_error();
 637
 638         glActiveTexture(GL_TEXTURE0);
 639         check_error();
 640         glBindTexture(GL_TEXTURE_2D, src_tex);
 641         check_error();
 642         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 643         check_error();
 644         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 645         check_error();
 646         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 647         check_error();
 648
 649         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 650         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 651
 652         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 653         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 654
 655         glDrawArrays(GL_TRIANGLES, 0, 3);
 656         check_error();
 657
 658         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 659         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 660
 661         glUseProgram(0);
 662         check_error();
 663
 664         resource_pool->release_fbo(fbo);
 665         glDeleteVertexArrays(1, &vao);
 666 }
 667
 668 void Mixer::release_display_frame(DisplayFrame *frame)
 669 {
 670         for (GLuint texnum : frame->temp_textures) {
 671                 resource_pool->release_2d_texture(texnum);
 672         }
 673         frame->temp_textures.clear();
 674         frame->ready_fence.reset();
 675         frame->input_frames.clear();
 676 }
 677
 678 void Mixer::start()
 679 {
 680         mixer_thread = thread(&Mixer::thread_func, this);
 681 }
 682
 683 void Mixer::quit()
 684 {
 685         should_quit = true;
 686         mixer_thread.join();
 687 }
 688
 689 void Mixer::transition_clicked(int transition_num)
 690 {
 691         theme->transition_clicked(transition_num, pts());
 692 }
 693
 694 void Mixer::channel_clicked(int preview_num)
 695 {
 696         theme->channel_clicked(preview_num);
 697 }
 698
 699 void Mixer::reset_meters()
 700 {
 701         peak = 0.0f;
 702         r128.reset();
 703         r128.integr_start();
 704 }
 705
 706 Mixer::OutputChannel::~OutputChannel()
 707 {
 708         if (has_current_frame) {
 709                 parent->release_display_frame(&current_frame);
 710         }
 711         if (has_ready_frame) {
 712                 parent->release_display_frame(&ready_frame);
 713         }
 714 }
 715
 716 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 717 {
 718         // Store this frame for display. Remove the ready frame if any
 719         // (it was seemingly never used).
 720         {
 721                 unique_lock<mutex> lock(frame_mutex);
 722                 if (has_ready_frame) {
 723                         parent->release_display_frame(&ready_frame);
 724                 }
 725                 ready_frame = frame;
 726                 has_ready_frame = true;
 727         }
 728
 729         if (has_new_frame_ready_callback) {
 730                 new_frame_ready_callback();
 731         }
 732 }
 733
 734 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 735 {
 736         unique_lock<mutex> lock(frame_mutex);
 737         if (!has_current_frame && !has_ready_frame) {
 738                 return false;
 739         }
 740
 741         if (has_current_frame && has_ready_frame) {
 742                 // We have a new ready frame. Toss the current one.
 743                 parent->release_display_frame(&current_frame);
 744                 has_current_frame = false;
 745         }
 746         if (has_ready_frame) {
 747                 assert(!has_current_frame);
 748                 current_frame = ready_frame;
 749                 ready_frame.ready_fence.reset();  // Drop the refcount.
 750                 ready_frame.input_frames.clear();  // Drop the refcounts.
 751                 has_current_frame = true;
 752                 has_ready_frame = false;
 753         }
 754
 755         *frame = current_frame;
 756         return true;
 757 }
 758
 759 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 760 {
 761         new_frame_ready_callback = callback;
 762         has_new_frame_ready_callback = true;
 763 }