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 }  // namespace
  63
  64 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
  65         : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT),
  66           num_cards(num_cards),
  67           mixer_surface(create_surface(format)),
  68           h264_encoder_surface(create_surface(format)),
  69           level_compressor(OUTPUT_FREQUENCY),
  70           limiter(OUTPUT_FREQUENCY),
  71           compressor(OUTPUT_FREQUENCY)
  72 {
  73         httpd.start(9095);
  74
  75         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
  76         check_error();
  77
  78         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
  79         // will be halved when sampling them, and we need to compensate here.
  80         movit_texel_subpixel_precision /= 2.0;
  81
  82         resource_pool.reset(new ResourcePool);
  83         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
  84         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
  85                 output_channel[i].parent = this;
  86         }
  87
  88         ImageFormat inout_format;
  89         inout_format.color_space = COLORSPACE_sRGB;
  90         inout_format.gamma_curve = GAMMA_sRGB;
  91
  92         // Display chain; shows the live output produced by the main chain (its RGBA version).
  93         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
  94         check_error();
  95         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
  96         display_chain->add_input(display_input);
  97         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
  98         display_chain->set_dither_bits(0);  // Don't bother.
  99         display_chain->finalize();
 100
 101         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
 102
 103         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 104                 printf("Configuring card %d...\n", card_index);
 105                 CaptureCard *card = &cards[card_index];
 106                 card->usb = new BMUSBCapture(card_index);
 107                 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 108                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 109                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
 110                 card->surface = create_surface(format);
 111                 card->usb->set_dequeue_thread_callbacks(
 112                         [card]{
 113                                 eglBindAPI(EGL_OPENGL_API);
 114                                 card->context = create_context(card->surface);
 115                                 if (!make_current(card->context, card->surface)) {
 116                                         printf("failed to create bmusb context\n");
 117                                         exit(1);
 118                                 }
 119                         },
 120                         [this]{
 121                                 resource_pool->clean_context();
 122                         });
 123                 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 124                 card->usb->configure_card();
 125         }
 126
 127         BMUSBCapture::start_bm_thread();
 128
 129         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 130                 cards[card_index].usb->start_bm_capture();
 131         }
 132
 133         //chain->enable_phase_timing(true);
 134
 135         // Set up stuff for NV12 conversion.
 136
 137         // Cb/Cr shader.
 138         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 139         string cbcr_frag_shader =
 140                 "#version 130 \n"
 141                 "in vec2 tc0; \n"
 142                 "uniform sampler2D cbcr_tex; \n"
 143                 "void main() { \n"
 144                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 145                 "} \n";
 146         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
 147
 148         r128.init(2, OUTPUT_FREQUENCY);
 149         r128.integr_start();
 150
 151         locut.init(FILTER_HPF, 2);
 152
 153         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 154         // and there's a limit to how important the peak meter is.
 155         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
 156
 157         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 158 }
 159
 160 Mixer::~Mixer()
 161 {
 162         resource_pool->release_glsl_program(cbcr_program_num);
 163         BMUSBCapture::stop_bm_thread();
 164
 165         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 166                 {
 167                         unique_lock<mutex> lock(bmusb_mutex);
 168                         cards[card_index].should_quit = true;  // Unblock thread.
 169                         cards[card_index].new_data_ready_changed.notify_all();
 170                 }
 171                 cards[card_index].usb->stop_dequeue_thread();
 172         }
 173
 174         h264_encoder.reset(nullptr);
 175 }
 176
 177 namespace {
 178
 179 int unwrap_timecode(uint16_t current_wrapped, int last)
 180 {
 181         uint16_t last_wrapped = last & 0xffff;
 182         if (current_wrapped > last_wrapped) {
 183                 return (last & ~0xffff) | current_wrapped;
 184         } else {
 185                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 186         }
 187 }
 188
 189 float find_peak(const float *samples, size_t num_samples)
 190 {
 191         float m = fabs(samples[0]);
 192         for (size_t i = 1; i < num_samples; ++i) {
 193                 m = std::max(m, fabs(samples[i]));
 194         }
 195         return m;
 196 }
 197
 198 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 199 {
 200         size_t num_samples = in.size() / 2;
 201         out_l->resize(num_samples);
 202         out_r->resize(num_samples);
 203
 204         const float *inptr = in.data();
 205         float *lptr = &(*out_l)[0];
 206         float *rptr = &(*out_r)[0];
 207         for (size_t i = 0; i < num_samples; ++i) {
 208                 *lptr++ = *inptr++;
 209                 *rptr++ = *inptr++;
 210         }
 211 }
 212
 213 }  // namespace
 214
 215 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 216                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 217                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 218 {
 219         CaptureCard *card = &cards[card_index];
 220
 221         unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
 222         bool interlaced;
 223
 224         decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom,
 225                             &frame_rate_nom, &frame_rate_den, &interlaced);  // Ignore return value for now.
 226         int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom;
 227
 228         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 229         if (num_samples > OUTPUT_FREQUENCY / 10) {
 230                 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",
 231                         card_index, int(audio_frame.len), int(audio_offset),
 232                         timecode, int(video_frame.len), int(video_offset), video_format);
 233                 if (video_frame.owner) {
 234                         video_frame.owner->release_frame(video_frame);
 235                 }
 236                 if (audio_frame.owner) {
 237                         audio_frame.owner->release_frame(audio_frame);
 238                 }
 239                 return;
 240         }
 241
 242         int64_t local_pts = card->next_local_pts;
 243         int dropped_frames = 0;
 244         if (card->last_timecode != -1) {
 245                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 246         }
 247
 248         // Convert the audio to stereo fp32 and add it.
 249         vector<float> audio;
 250         audio.resize(num_samples * 2);
 251         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 252
 253         // Add the audio.
 254         {
 255                 unique_lock<mutex> lock(card->audio_mutex);
 256
 257                 // Number of samples per frame if we need to insert silence.
 258                 // (Could be nonintegral, but resampling will save us then.)
 259                 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
 260
 261                 if (dropped_frames > MAX_FPS * 2) {
 262                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 263                                 card_index, card->last_timecode, timecode);
 264                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 265                         dropped_frames = 0;
 266                 } else if (dropped_frames > 0) {
 267                         // Insert silence as needed.
 268                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 269                                 card_index, dropped_frames, timecode);
 270                         vector<float> silence(silence_samples * 2, 0.0f);
 271                         for (int i = 0; i < dropped_frames; ++i) {
 272                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 273                                 // Note that if the format changed in the meantime, we have
 274                                 // no way of detecting that; we just have to assume the frame length
 275                                 // is always the same.
 276                                 local_pts += frame_length;
 277                         }
 278                 }
 279                 if (num_samples == 0) {
 280                         audio.resize(silence_samples * 2);
 281                         num_samples = silence_samples;
 282                 }
 283                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 284                 card->next_local_pts = local_pts + frame_length;
 285         }
 286
 287         card->last_timecode = timecode;
 288
 289         // Done with the audio, so release it.
 290         if (audio_frame.owner) {
 291                 audio_frame.owner->release_frame(audio_frame);
 292         }
 293
 294         {
 295                 // Wait until the previous frame was consumed.
 296                 unique_lock<mutex> lock(bmusb_mutex);
 297                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 298                 if (card->should_quit) return;
 299         }
 300
 301         if (video_frame.len - video_offset == 0 ||
 302             video_frame.len - video_offset != size_t(width * (height + extra_lines_top + extra_lines_bottom) * 2)) {
 303                 if (video_frame.len != 0) {
 304                         printf("Card %d: Dropping video frame with wrong length (%ld)\n",
 305                                 card_index, video_frame.len - video_offset);
 306                 }
 307                 if (video_frame.owner) {
 308                         video_frame.owner->release_frame(video_frame);
 309                 }
 310
 311                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 312                 // so that pts can go up accordingly.
 313                 {
 314                         unique_lock<mutex> lock(bmusb_mutex);
 315                         card->new_data_ready = true;
 316                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 317                         card->new_frame_length = frame_length;
 318                         card->new_frame_interlaced = false;
 319                         card->new_data_ready_fence = nullptr;
 320                         card->dropped_frames = dropped_frames;
 321                         card->new_data_ready_changed.notify_all();
 322                 }
 323                 return;
 324         }
 325
 326         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 327
 328         unsigned num_fields = interlaced ? 2 : 1;
 329         timespec frame_upload_start;
 330         if (interlaced) {
 331                 // NOTE: This isn't deinterlacing. This is just sending the two fields along
 332                 // as separate frames without considering anything like the half-field offset.
 333                 // We'll need to add a proper deinterlacer on the receiving side to get this right.
 334                 assert(height % 2 == 0);
 335                 height /= 2;
 336                 assert(frame_length % 2 == 0);
 337                 frame_length /= 2;
 338                 num_fields = 2;
 339                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 340         }
 341         RefCountedFrame new_frame(video_frame);
 342
 343         // Upload the textures.
 344         size_t cbcr_width = width / 2;
 345         size_t cbcr_offset = video_offset / 2;
 346         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 347
 348         for (unsigned field = 0; field < num_fields; ++field) {
 349                 unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22);
 350
 351                 if (userdata->tex_y[field] == 0 ||
 352                     userdata->tex_cbcr[field] == 0 ||
 353                     width != userdata->last_width[field] ||
 354                     height != userdata->last_height[field]) {
 355                         // We changed resolution since last use of this texture, so we need to create
 356                         // a new object. Note that this each card has its own PBOFrameAllocator,
 357                         // we don't need to worry about these flip-flopping between resolutions.
 358                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 359                         check_error();
 360                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 361                         check_error();
 362                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 363                         check_error();
 364                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 365                         check_error();
 366                         userdata->last_width[field] = width;
 367                         userdata->last_height[field] = height;
 368                 }
 369
 370                 GLuint pbo = userdata->pbo;
 371                 check_error();
 372                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 373                 check_error();
 374                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
 375                 check_error();
 376                 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 377                 //check_error();
 378
 379                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 380                 check_error();
 381                 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)));
 382                 check_error();
 383                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 384                 check_error();
 385                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line));
 386                 check_error();
 387                 glBindTexture(GL_TEXTURE_2D, 0);
 388                 check_error();
 389                 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 390                 check_error();
 391                 assert(fence != nullptr);
 392
 393                 if (field == 1) {
 394                         // Don't upload the second field as fast as we can; wait until
 395                         // the field time has approximately passed. (Otherwise, we could
 396                         // get timing jitter against the other sources, and possibly also
 397                         // against the video display, although the latter is not as critical.)
 398                         // This requires our system clock to be reasonably close to the
 399                         // video clock, but that's not an unreasonable assumption.
 400                         timespec second_field_start;
 401                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 402                                 frame_length * 1000000000 / TIMEBASE;
 403                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 404                                 second_field_start.tv_nsec / 1000000000;
 405                         second_field_start.tv_nsec %= 1000000000;
 406
 407                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 408                                                &second_field_start, nullptr) == -1 &&
 409                                errno == EINTR) ;
 410                 }
 411
 412                 {
 413                         unique_lock<mutex> lock(bmusb_mutex);
 414                         card->new_data_ready = true;
 415                         card->new_frame = new_frame;
 416                         card->new_frame_length = frame_length;
 417                         card->new_frame_field = field;
 418                         card->new_frame_interlaced = interlaced;
 419                         card->new_data_ready_fence = fence;
 420                         card->dropped_frames = dropped_frames;
 421                         card->new_data_ready_changed.notify_all();
 422
 423                         if (field != num_fields - 1) {
 424                                 // Wait until the previous frame was consumed.
 425                                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 426                                 if (card->should_quit) return;
 427                         }
 428                 }
 429         }
 430 }
 431
 432 void Mixer::thread_func()
 433 {
 434         eglBindAPI(EGL_OPENGL_API);
 435         QOpenGLContext *context = create_context(mixer_surface);
 436         if (!make_current(context, mixer_surface)) {
 437                 printf("oops\n");
 438                 exit(1);
 439         }
 440
 441         struct timespec start, now;
 442         clock_gettime(CLOCK_MONOTONIC, &start);
 443
 444         int frame = 0;
 445         int stats_dropped_frames = 0;
 446
 447         while (!should_quit) {
 448                 CaptureCard card_copy[MAX_CARDS];
 449                 int num_samples[MAX_CARDS];
 450
 451                 {
 452                         unique_lock<mutex> lock(bmusb_mutex);
 453
 454                         // The first card is the master timer, so wait for it to have a new frame.
 455                         // TODO: Make configurable, and with a timeout.
 456                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 457
 458                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 459                                 CaptureCard *card = &cards[card_index];
 460                                 card_copy[card_index].usb = card->usb;
 461                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 462                                 card_copy[card_index].new_frame = card->new_frame;
 463                                 card_copy[card_index].new_frame_length = card->new_frame_length;
 464                                 card_copy[card_index].new_frame_field = card->new_frame_field;
 465                                 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
 466                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 467                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 468                                 card->new_data_ready = false;
 469                                 card->new_data_ready_changed.notify_all();
 470
 471                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
 472                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 473                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 474                                 assert(num_samples[card_index] >= 0);
 475                         }
 476                 }
 477
 478                 // Resample the audio as needed, including from previously dropped frames.
 479                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 480                         {
 481                                 // Signal to the audio thread to process this frame.
 482                                 unique_lock<mutex> lock(audio_mutex);
 483                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
 484                                 audio_task_queue_changed.notify_one();
 485                         }
 486                         if (frame_num != card_copy[0].dropped_frames) {
 487                                 // For dropped frames, increase the pts. Note that if the format changed
 488                                 // in the meantime, we have no way of detecting that; we just have to
 489                                 // assume the frame length is always the same.
 490                                 ++stats_dropped_frames;
 491                                 pts_int += card_copy[0].new_frame_length;
 492                         }
 493                 }
 494
 495                 if (audio_level_callback != nullptr) {
 496                         double loudness_s = r128.loudness_S();
 497                         double loudness_i = r128.integrated();
 498                         double loudness_range_low = r128.range_min();
 499                         double loudness_range_high = r128.range_max();
 500
 501                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 502                                              loudness_i, loudness_range_low, loudness_range_high,
 503                                              last_gain_staging_db);
 504                 }
 505
 506                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 507                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 508                                 ++card_copy[card_index].dropped_frames;
 509                         }
 510                         if (card_copy[card_index].dropped_frames > 0) {
 511                                 printf("Card %u dropped %d frames before this\n",
 512                                         card_index, int(card_copy[card_index].dropped_frames));
 513                         }
 514                 }
 515
 516                 // If the first card is reporting a corrupted or otherwise dropped frame,
 517                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 518                 if (card_copy[0].new_frame->len == 0) {
 519                         ++stats_dropped_frames;
 520                         pts_int += card_copy[0].new_frame_length;
 521                         continue;
 522                 }
 523
 524                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 525                         CaptureCard *card = &card_copy[card_index];
 526                         if (!card->new_data_ready || card->new_frame->len == 0)
 527                                 continue;
 528
 529                         assert(card->new_frame != nullptr);
 530                         if (card->new_frame_interlaced) {
 531                                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
 532                                         buffered_frames[card_index][frame_num] = buffered_frames[card_index][frame_num - 1];
 533                                 }
 534                                 buffered_frames[card_index][0] = { card->new_frame, card->new_frame_field };
 535                         } else {
 536                                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
 537                                         buffered_frames[card_index][frame_num] = { card->new_frame, card->new_frame_field };
 538                                 }
 539                         }
 540                         check_error();
 541
 542                         // The new texture might still be uploaded,
 543                         // tell the GPU to wait until it's there.
 544                         if (card->new_data_ready_fence) {
 545                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 546                                 check_error();
 547                                 glDeleteSync(card->new_data_ready_fence);
 548                                 check_error();
 549                         }
 550                 }
 551
 552                 // Get the main chain from the theme, and set its state immediately.
 553                 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
 554                 EffectChain *chain = theme_main_chain.chain;
 555                 theme_main_chain.setup_chain();
 556
 557                 GLuint y_tex, cbcr_tex;
 558                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 559                 assert(got_frame);
 560
 561                 // Render main chain.
 562                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 563                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 564                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 565                 check_error();
 566                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 567                 resource_pool->release_fbo(fbo);
 568
 569                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 570                 resource_pool->release_2d_texture(cbcr_full_tex);
 571
 572                 // Set the right state for rgba_tex.
 573                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 574                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 575                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 576                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 577                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 578
 579                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 580                 check_error();
 581
 582                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 583                 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
 584                 ++frame;
 585                 pts_int += card_copy[0].new_frame_length;
 586
 587                 // The live frame just shows the RGBA texture we just rendered.
 588                 // It owns rgba_tex now.
 589                 DisplayFrame live_frame;
 590                 live_frame.chain = display_chain.get();
 591                 live_frame.setup_chain = [this, rgba_tex]{
 592                         display_input->set_texture_num(rgba_tex);
 593                 };
 594                 live_frame.ready_fence = fence;
 595                 live_frame.input_frames = {};
 596                 live_frame.temp_textures = { rgba_tex };
 597                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 598
 599                 // Set up preview and any additional channels.
 600                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 601                         DisplayFrame display_frame;
 602                         Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT);  // FIXME: dimensions
 603                         display_frame.chain = chain.chain;
 604                         display_frame.setup_chain = chain.setup_chain;
 605                         display_frame.ready_fence = fence;
 606                         display_frame.input_frames = chain.input_frames;
 607                         display_frame.temp_textures = {};
 608                         output_channel[i].output_frame(display_frame);
 609                 }
 610
 611                 clock_gettime(CLOCK_MONOTONIC, &now);
 612                 double elapsed = now.tv_sec - start.tv_sec +
 613                         1e-9 * (now.tv_nsec - start.tv_nsec);
 614                 if (frame % 100 == 0) {
 615                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 616                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 617                                 1e3 * elapsed / frame);
 618                 //      chain->print_phase_timing();
 619                 }
 620
 621 #if 0
 622                 // Reset every 100 frames, so that local variations in frame times
 623                 // (especially for the first few frames, when the shaders are
 624                 // compiled etc.) don't make it hard to measure for the entire
 625                 // remaining duration of the program.
 626                 if (frame == 10000) {
 627                         frame = 0;
 628                         start = now;
 629                 }
 630 #endif
 631                 check_error();
 632         }
 633
 634         resource_pool->clean_context();
 635 }
 636
 637 void Mixer::audio_thread_func()
 638 {
 639         while (!should_quit) {
 640                 AudioTask task;
 641
 642                 {
 643                         unique_lock<mutex> lock(audio_mutex);
 644                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 645                         task = audio_task_queue.front();
 646                         audio_task_queue.pop();
 647                 }
 648
 649                 process_audio_one_frame(task.pts_int, task.num_samples);
 650         }
 651 }
 652
 653 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 654 {
 655         vector<float> samples_card;
 656         vector<float> samples_out;
 657         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 658                 samples_card.resize(num_samples * 2);
 659                 {
 660                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 661                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 662                                 printf("Card %d reported previous underrun.\n", card_index);
 663                         }
 664                 }
 665                 // TODO: Allow using audio from the other card(s) as well.
 666                 if (card_index == 0) {
 667                         samples_out = move(samples_card);
 668                 }
 669         }
 670
 671         // Cut away everything under 120 Hz (or whatever the cutoff is);
 672         // we don't need it for voice, and it will reduce headroom
 673         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 674         // should be dampened.)
 675         locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 676
 677         // Apply a level compressor to get the general level right.
 678         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 679         // (or more precisely, near it, since we don't use infinite ratio),
 680         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 681         // entirely arbitrary, but from practical tests with speech, it seems to
 682         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 683         float ref_level_dbfs = -14.0f;
 684         {
 685                 float threshold = 0.01f;   // -40 dBFS.
 686                 float ratio = 20.0f;
 687                 float attack_time = 0.5f;
 688                 float release_time = 20.0f;
 689                 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 690                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 691                 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 692         }
 693
 694 #if 0
 695         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 696                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 697                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 698                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 699 #endif
 700
 701 //      float limiter_att, compressor_att;
 702
 703         // The real compressor.
 704         if (compressor_enabled) {
 705                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 706                 float ratio = 20.0f;
 707                 float attack_time = 0.005f;
 708                 float release_time = 0.040f;
 709                 float makeup_gain = 2.0f;  // +6 dB.
 710                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 711 //              compressor_att = compressor.get_attenuation();
 712         }
 713
 714         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 715         // Note that since ratio is not infinite, we could go slightly higher than this.
 716         if (limiter_enabled) {
 717                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 718                 float ratio = 30.0f;
 719                 float attack_time = 0.0f;  // Instant.
 720                 float release_time = 0.020f;
 721                 float makeup_gain = 1.0f;  // 0 dB.
 722                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 723 //              limiter_att = limiter.get_attenuation();
 724         }
 725
 726 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 727
 728         // Upsample 4x to find interpolated peak.
 729         peak_resampler.inp_data = samples_out.data();
 730         peak_resampler.inp_count = samples_out.size() / 2;
 731
 732         vector<float> interpolated_samples_out;
 733         interpolated_samples_out.resize(samples_out.size());
 734         while (peak_resampler.inp_count > 0) {  // About four iterations.
 735                 peak_resampler.out_data = &interpolated_samples_out[0];
 736                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 737                 peak_resampler.process();
 738                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 739                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 740         }
 741
 742         // Find R128 levels.
 743         vector<float> left, right;
 744         deinterleave_samples(samples_out, &left, &right);
 745         float *ptrs[] = { left.data(), right.data() };
 746         r128.process(left.size(), ptrs);
 747
 748         // Send the samples to the sound card.
 749         if (alsa) {
 750                 alsa->write(samples_out);
 751         }
 752
 753         // And finally add them to the output.
 754         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 755 }
 756
 757 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 758 {
 759         GLuint vao;
 760         glGenVertexArrays(1, &vao);
 761         check_error();
 762
 763         float vertices[] = {
 764                 0.0f, 2.0f,
 765                 0.0f, 0.0f,
 766                 2.0f, 0.0f
 767         };
 768
 769         glBindVertexArray(vao);
 770         check_error();
 771
 772         // Extract Cb/Cr.
 773         GLuint fbo = resource_pool->create_fbo(dst_tex);
 774         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 775         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 776         check_error();
 777
 778         glUseProgram(cbcr_program_num);
 779         check_error();
 780
 781         glActiveTexture(GL_TEXTURE0);
 782         check_error();
 783         glBindTexture(GL_TEXTURE_2D, src_tex);
 784         check_error();
 785         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 786         check_error();
 787         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 788         check_error();
 789         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 790         check_error();
 791
 792         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 793         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 794
 795         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 796         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 797
 798         glDrawArrays(GL_TRIANGLES, 0, 3);
 799         check_error();
 800
 801         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 802         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 803
 804         glUseProgram(0);
 805         check_error();
 806
 807         resource_pool->release_fbo(fbo);
 808         glDeleteVertexArrays(1, &vao);
 809 }
 810
 811 void Mixer::release_display_frame(DisplayFrame *frame)
 812 {
 813         for (GLuint texnum : frame->temp_textures) {
 814                 resource_pool->release_2d_texture(texnum);
 815         }
 816         frame->temp_textures.clear();
 817         frame->ready_fence.reset();
 818         frame->input_frames.clear();
 819 }
 820
 821 void Mixer::start()
 822 {
 823         mixer_thread = thread(&Mixer::thread_func, this);
 824         audio_thread = thread(&Mixer::audio_thread_func, this);
 825 }
 826
 827 void Mixer::quit()
 828 {
 829         should_quit = true;
 830         mixer_thread.join();
 831         audio_thread.join();
 832 }
 833
 834 void Mixer::transition_clicked(int transition_num)
 835 {
 836         theme->transition_clicked(transition_num, pts());
 837 }
 838
 839 void Mixer::channel_clicked(int preview_num)
 840 {
 841         theme->channel_clicked(preview_num);
 842 }
 843
 844 void Mixer::reset_meters()
 845 {
 846         peak_resampler.reset();
 847         peak = 0.0f;
 848         r128.reset();
 849         r128.integr_start();
 850 }
 851
 852 Mixer::OutputChannel::~OutputChannel()
 853 {
 854         if (has_current_frame) {
 855                 parent->release_display_frame(&current_frame);
 856         }
 857         if (has_ready_frame) {
 858                 parent->release_display_frame(&ready_frame);
 859         }
 860 }
 861
 862 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 863 {
 864         // Store this frame for display. Remove the ready frame if any
 865         // (it was seemingly never used).
 866         {
 867                 unique_lock<mutex> lock(frame_mutex);
 868                 if (has_ready_frame) {
 869                         parent->release_display_frame(&ready_frame);
 870                 }
 871                 ready_frame = frame;
 872                 has_ready_frame = true;
 873         }
 874
 875         if (has_new_frame_ready_callback) {
 876                 new_frame_ready_callback();
 877         }
 878 }
 879
 880 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 881 {
 882         unique_lock<mutex> lock(frame_mutex);
 883         if (!has_current_frame && !has_ready_frame) {
 884                 return false;
 885         }
 886
 887         if (has_current_frame && has_ready_frame) {
 888                 // We have a new ready frame. Toss the current one.
 889                 parent->release_display_frame(&current_frame);
 890                 has_current_frame = false;
 891         }
 892         if (has_ready_frame) {
 893                 assert(!has_current_frame);
 894                 current_frame = ready_frame;
 895                 ready_frame.ready_fence.reset();  // Drop the refcount.
 896                 ready_frame.input_frames.clear();  // Drop the refcounts.
 897                 has_current_frame = true;
 898                 has_ready_frame = false;
 899         }
 900
 901         *frame = current_frame;
 902         return true;
 903 }
 904
 905 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 906 {
 907         new_frame_ready_callback = callback;
 908         has_new_frame_ready_callback = true;
 909 }