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 <movit/effect_chain.h>
   8 #include <movit/effect_util.h>
   9 #include <movit/flat_input.h>
  10 #include <movit/image_format.h>
  11 #include <movit/init.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 "flags.h"
  34 #include "h264encode.h"
  35 #include "pbo_frame_allocator.h"
  36 #include "ref_counted_gl_sync.h"
  37 #include "timebase.h"
  38
  39 class QOpenGLContext;
  40
  41 using namespace movit;
  42 using namespace std;
  43 using namespace std::placeholders;
  44
  45 Mixer *global_mixer = nullptr;
  46
  47 namespace {
  48
  49 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  50 {
  51         for (size_t i = 0; i < num_samples; ++i) {
  52                 for (size_t j = 0; j < out_channels; ++j) {
  53                         uint32_t s1 = *src++;
  54                         uint32_t s2 = *src++;
  55                         uint32_t s3 = *src++;
  56                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  57                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  58                 }
  59                 src += 3 * (in_channels - out_channels);
  60         }
  61 }
  62
  63 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  64 {
  65         if (interlaced) {
  66                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  67                         input_state->buffered_frames[card_index][frame_num] =
  68                                 input_state->buffered_frames[card_index][frame_num - 1];
  69                 }
  70                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  71         } else {
  72                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  73                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  74                 }
  75         }
  76 }
  77
  78 string generate_local_dump_filename(int frame)
  79 {
  80         time_t now = time(NULL);
  81         tm now_tm;
  82         localtime_r(&now, &now_tm);
  83
  84         char timestamp[256];
  85         strftime(timestamp, sizeof(timestamp), "%F-%T%z", &now_tm);
  86
  87         // Use the frame number to disambiguate between two cuts starting
  88         // on the same second.
  89         char filename[256];
  90         snprintf(filename, sizeof(filename), "%s%s-f%02d%s",
  91                 LOCAL_DUMP_PREFIX, timestamp, frame % 100, LOCAL_DUMP_SUFFIX);
  92         return filename;
  93 }
  94
  95 }  // namespace
  96
  97 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
  98         : httpd(WIDTH, HEIGHT),
  99           num_cards(num_cards),
 100           mixer_surface(create_surface(format)),
 101           h264_encoder_surface(create_surface(format)),
 102           correlation(OUTPUT_FREQUENCY),
 103           level_compressor(OUTPUT_FREQUENCY),
 104           limiter(OUTPUT_FREQUENCY),
 105           compressor(OUTPUT_FREQUENCY)
 106 {
 107         httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str());
 108         httpd.start(9095);
 109
 110         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 111         check_error();
 112
 113         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 114         // will be halved when sampling them, and we need to compensate here.
 115         movit_texel_subpixel_precision /= 2.0;
 116
 117         resource_pool.reset(new ResourcePool);
 118         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
 119         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 120                 output_channel[i].parent = this;
 121         }
 122
 123         ImageFormat inout_format;
 124         inout_format.color_space = COLORSPACE_sRGB;
 125         inout_format.gamma_curve = GAMMA_sRGB;
 126
 127         // Display chain; shows the live output produced by the main chain (its RGBA version).
 128         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 129         check_error();
 130         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 131         display_chain->add_input(display_input);
 132         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 133         display_chain->set_dither_bits(0);  // Don't bother.
 134         display_chain->finalize();
 135
 136         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 137
 138         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 139                 printf("Configuring card %d...\n", card_index);
 140                 CaptureCard *card = &cards[card_index];
 141                 card->usb = new BMUSBCapture(card_index);
 142                 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 143                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 144                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
 145                 card->surface = create_surface(format);
 146                 card->usb->set_dequeue_thread_callbacks(
 147                         [card]{
 148                                 eglBindAPI(EGL_OPENGL_API);
 149                                 card->context = create_context(card->surface);
 150                                 if (!make_current(card->context, card->surface)) {
 151                                         printf("failed to create bmusb context\n");
 152                                         exit(1);
 153                                 }
 154                         },
 155                         [this]{
 156                                 resource_pool->clean_context();
 157                         });
 158                 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 159                 card->usb->configure_card();
 160         }
 161
 162         BMUSBCapture::start_bm_thread();
 163
 164         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 165                 cards[card_index].usb->start_bm_capture();
 166         }
 167
 168         // Set up stuff for NV12 conversion.
 169
 170         // Cb/Cr shader.
 171         string cbcr_vert_shader =
 172                 "#version 130 \n"
 173                 " \n"
 174                 "in vec2 position; \n"
 175                 "in vec2 texcoord; \n"
 176                 "out vec2 tc0; \n"
 177                 "uniform vec2 foo_chroma_offset_0; \n"
 178                 " \n"
 179                 "void main() \n"
 180                 "{ \n"
 181                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 182                 "    // \n"
 183                 "    //   2.000  0.000  0.000 -1.000 \n"
 184                 "    //   0.000  2.000  0.000 -1.000 \n"
 185                 "    //   0.000  0.000 -2.000 -1.000 \n"
 186                 "    //   0.000  0.000  0.000  1.000 \n"
 187                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 188                 "    vec2 flipped_tc = texcoord; \n"
 189                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
 190                 "} \n";
 191         string cbcr_frag_shader =
 192                 "#version 130 \n"
 193                 "in vec2 tc0; \n"
 194                 "uniform sampler2D cbcr_tex; \n"
 195                 "out vec4 FragColor; \n"
 196                 "void main() { \n"
 197                 "    FragColor = texture(cbcr_tex, tc0); \n"
 198                 "} \n";
 199         vector<string> frag_shader_outputs;
 200         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 201
 202         r128.init(2, OUTPUT_FREQUENCY);
 203         r128.integr_start();
 204
 205         locut.init(FILTER_HPF, 2);
 206
 207         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 208         // and there's a limit to how important the peak meter is.
 209         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 210
 211         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 212 }
 213
 214 Mixer::~Mixer()
 215 {
 216         resource_pool->release_glsl_program(cbcr_program_num);
 217         BMUSBCapture::stop_bm_thread();
 218
 219         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 220                 {
 221                         unique_lock<mutex> lock(bmusb_mutex);
 222                         cards[card_index].should_quit = true;  // Unblock thread.
 223                         cards[card_index].new_data_ready_changed.notify_all();
 224                 }
 225                 cards[card_index].usb->stop_dequeue_thread();
 226         }
 227
 228         h264_encoder.reset(nullptr);
 229 }
 230
 231 namespace {
 232
 233 int unwrap_timecode(uint16_t current_wrapped, int last)
 234 {
 235         uint16_t last_wrapped = last & 0xffff;
 236         if (current_wrapped > last_wrapped) {
 237                 return (last & ~0xffff) | current_wrapped;
 238         } else {
 239                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 240         }
 241 }
 242
 243 float find_peak(const float *samples, size_t num_samples)
 244 {
 245         float m = fabs(samples[0]);
 246         for (size_t i = 1; i < num_samples; ++i) {
 247                 m = max(m, fabs(samples[i]));
 248         }
 249         return m;
 250 }
 251
 252 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 253 {
 254         size_t num_samples = in.size() / 2;
 255         out_l->resize(num_samples);
 256         out_r->resize(num_samples);
 257
 258         const float *inptr = in.data();
 259         float *lptr = &(*out_l)[0];
 260         float *rptr = &(*out_r)[0];
 261         for (size_t i = 0; i < num_samples; ++i) {
 262                 *lptr++ = *inptr++;
 263                 *rptr++ = *inptr++;
 264         }
 265 }
 266
 267 }  // namespace
 268
 269 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 270                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 271                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 272 {
 273         CaptureCard *card = &cards[card_index];
 274
 275         unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
 276         bool interlaced;
 277
 278         decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom,
 279                             &frame_rate_nom, &frame_rate_den, &interlaced);  // Ignore return value for now.
 280         int64_t frame_length = int64_t(TIMEBASE * frame_rate_den) / frame_rate_nom;
 281
 282         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 283         if (num_samples > OUTPUT_FREQUENCY / 10) {
 284                 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",
 285                         card_index, int(audio_frame.len), int(audio_offset),
 286                         timecode, int(video_frame.len), int(video_offset), video_format);
 287                 if (video_frame.owner) {
 288                         video_frame.owner->release_frame(video_frame);
 289                 }
 290                 if (audio_frame.owner) {
 291                         audio_frame.owner->release_frame(audio_frame);
 292                 }
 293                 return;
 294         }
 295
 296         int64_t local_pts = card->next_local_pts;
 297         int dropped_frames = 0;
 298         if (card->last_timecode != -1) {
 299                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 300         }
 301
 302         // Convert the audio to stereo fp32 and add it.
 303         vector<float> audio;
 304         audio.resize(num_samples * 2);
 305         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 306
 307         // Add the audio.
 308         {
 309                 unique_lock<mutex> lock(card->audio_mutex);
 310
 311                 // Number of samples per frame if we need to insert silence.
 312                 // (Could be nonintegral, but resampling will save us then.)
 313                 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
 314
 315                 if (dropped_frames > MAX_FPS * 2) {
 316                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 317                                 card_index, card->last_timecode, timecode);
 318                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 319                         dropped_frames = 0;
 320                 } else if (dropped_frames > 0) {
 321                         // Insert silence as needed.
 322                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 323                                 card_index, dropped_frames, timecode);
 324                         vector<float> silence(silence_samples * 2, 0.0f);
 325                         for (int i = 0; i < dropped_frames; ++i) {
 326                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 327                                 // Note that if the format changed in the meantime, we have
 328                                 // no way of detecting that; we just have to assume the frame length
 329                                 // is always the same.
 330                                 local_pts += frame_length;
 331                         }
 332                 }
 333                 if (num_samples == 0) {
 334                         audio.resize(silence_samples * 2);
 335                         num_samples = silence_samples;
 336                 }
 337                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 338                 card->next_local_pts = local_pts + frame_length;
 339         }
 340
 341         card->last_timecode = timecode;
 342
 343         // Done with the audio, so release it.
 344         if (audio_frame.owner) {
 345                 audio_frame.owner->release_frame(audio_frame);
 346         }
 347
 348         {
 349                 // Wait until the previous frame was consumed.
 350                 unique_lock<mutex> lock(bmusb_mutex);
 351                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 352                 if (card->should_quit) return;
 353         }
 354
 355         size_t expected_length = width * (height + extra_lines_top + extra_lines_bottom) * 2;
 356         if (video_frame.len - video_offset == 0 ||
 357             video_frame.len - video_offset != expected_length) {
 358                 if (video_frame.len != 0) {
 359                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 360                                 card_index, video_frame.len - video_offset, expected_length);
 361                 }
 362                 if (video_frame.owner) {
 363                         video_frame.owner->release_frame(video_frame);
 364                 }
 365
 366                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 367                 // so that pts can go up accordingly.
 368                 {
 369                         unique_lock<mutex> lock(bmusb_mutex);
 370                         card->new_data_ready = true;
 371                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 372                         card->new_frame_length = frame_length;
 373                         card->new_frame_interlaced = false;
 374                         card->new_data_ready_fence = nullptr;
 375                         card->dropped_frames = dropped_frames;
 376                         card->new_data_ready_changed.notify_all();
 377                 }
 378                 return;
 379         }
 380
 381         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 382
 383         unsigned num_fields = interlaced ? 2 : 1;
 384         timespec frame_upload_start;
 385         if (interlaced) {
 386                 // Send the two fields along as separate frames; the other side will need to add
 387                 // a deinterlacer to actually get this right.
 388                 assert(height % 2 == 0);
 389                 height /= 2;
 390                 assert(frame_length % 2 == 0);
 391                 frame_length /= 2;
 392                 num_fields = 2;
 393                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 394         }
 395         userdata->last_interlaced = interlaced;
 396         userdata->last_frame_rate_nom = frame_rate_nom;
 397         userdata->last_frame_rate_den = frame_rate_den;
 398         RefCountedFrame new_frame(video_frame);
 399
 400         // Upload the textures.
 401         size_t cbcr_width = width / 2;
 402         size_t cbcr_offset = video_offset / 2;
 403         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 404
 405         for (unsigned field = 0; field < num_fields; ++field) {
 406                 unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22);
 407
 408                 if (userdata->tex_y[field] == 0 ||
 409                     userdata->tex_cbcr[field] == 0 ||
 410                     width != userdata->last_width[field] ||
 411                     height != userdata->last_height[field]) {
 412                         // We changed resolution since last use of this texture, so we need to create
 413                         // a new object. Note that this each card has its own PBOFrameAllocator,
 414                         // we don't need to worry about these flip-flopping between resolutions.
 415                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 416                         check_error();
 417                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 418                         check_error();
 419                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 420                         check_error();
 421                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 422                         check_error();
 423                         userdata->last_width[field] = width;
 424                         userdata->last_height[field] = height;
 425                 }
 426
 427                 GLuint pbo = userdata->pbo;
 428                 check_error();
 429                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 430                 check_error();
 431                 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 432                 check_error();
 433
 434                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 435                 check_error();
 436                 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)));
 437                 check_error();
 438                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 439                 check_error();
 440                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line));
 441                 check_error();
 442                 glBindTexture(GL_TEXTURE_2D, 0);
 443                 check_error();
 444                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 445                 check_error();
 446                 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 447                 check_error();
 448                 assert(fence != nullptr);
 449
 450                 if (field == 1) {
 451                         // Don't upload the second field as fast as we can; wait until
 452                         // the field time has approximately passed. (Otherwise, we could
 453                         // get timing jitter against the other sources, and possibly also
 454                         // against the video display, although the latter is not as critical.)
 455                         // This requires our system clock to be reasonably close to the
 456                         // video clock, but that's not an unreasonable assumption.
 457                         timespec second_field_start;
 458                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 459                                 frame_length * 1000000000 / TIMEBASE;
 460                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 461                                 second_field_start.tv_nsec / 1000000000;
 462                         second_field_start.tv_nsec %= 1000000000;
 463
 464                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 465                                                &second_field_start, nullptr) == -1 &&
 466                                errno == EINTR) ;
 467                 }
 468
 469                 {
 470                         unique_lock<mutex> lock(bmusb_mutex);
 471                         card->new_data_ready = true;
 472                         card->new_frame = new_frame;
 473                         card->new_frame_length = frame_length;
 474                         card->new_frame_field = field;
 475                         card->new_frame_interlaced = interlaced;
 476                         card->new_data_ready_fence = fence;
 477                         card->dropped_frames = dropped_frames;
 478                         card->new_data_ready_changed.notify_all();
 479
 480                         if (field != num_fields - 1) {
 481                                 // Wait until the previous frame was consumed.
 482                                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 483                                 if (card->should_quit) return;
 484                         }
 485                 }
 486         }
 487 }
 488
 489 void Mixer::thread_func()
 490 {
 491         eglBindAPI(EGL_OPENGL_API);
 492         QOpenGLContext *context = create_context(mixer_surface);
 493         if (!make_current(context, mixer_surface)) {
 494                 printf("oops\n");
 495                 exit(1);
 496         }
 497
 498         struct timespec start, now;
 499         clock_gettime(CLOCK_MONOTONIC, &start);
 500
 501         int frame = 0;
 502         int stats_dropped_frames = 0;
 503
 504         while (!should_quit) {
 505                 CaptureCard card_copy[MAX_CARDS];
 506                 int num_samples[MAX_CARDS];
 507
 508                 {
 509                         unique_lock<mutex> lock(bmusb_mutex);
 510
 511                         // The first card is the master timer, so wait for it to have a new frame.
 512                         // TODO: Make configurable, and with a timeout.
 513                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 514
 515                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 516                                 CaptureCard *card = &cards[card_index];
 517                                 card_copy[card_index].usb = card->usb;
 518                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 519                                 card_copy[card_index].new_frame = card->new_frame;
 520                                 card_copy[card_index].new_frame_length = card->new_frame_length;
 521                                 card_copy[card_index].new_frame_field = card->new_frame_field;
 522                                 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
 523                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 524                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 525                                 card->new_data_ready = false;
 526                                 card->new_data_ready_changed.notify_all();
 527
 528                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
 529                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 530                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 531                                 assert(num_samples[card_index] >= 0);
 532                         }
 533                 }
 534
 535                 // Resample the audio as needed, including from previously dropped frames.
 536                 assert(num_cards > 0);
 537                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 538                         {
 539                                 // Signal to the audio thread to process this frame.
 540                                 unique_lock<mutex> lock(audio_mutex);
 541                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
 542                                 audio_task_queue_changed.notify_one();
 543                         }
 544                         if (frame_num != card_copy[0].dropped_frames) {
 545                                 // For dropped frames, increase the pts. Note that if the format changed
 546                                 // in the meantime, we have no way of detecting that; we just have to
 547                                 // assume the frame length is always the same.
 548                                 ++stats_dropped_frames;
 549                                 pts_int += card_copy[0].new_frame_length;
 550                         }
 551                 }
 552
 553                 if (audio_level_callback != nullptr) {
 554                         unique_lock<mutex> lock(compressor_mutex);
 555                         double loudness_s = r128.loudness_S();
 556                         double loudness_i = r128.integrated();
 557                         double loudness_range_low = r128.range_min();
 558                         double loudness_range_high = r128.range_max();
 559
 560                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 561                                              loudness_i, loudness_range_low, loudness_range_high,
 562                                              gain_staging_db, 20.0 * log10(final_makeup_gain),
 563                                              correlation.get_correlation());
 564                 }
 565
 566                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 567                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 568                                 ++card_copy[card_index].dropped_frames;
 569                         }
 570                         if (card_copy[card_index].dropped_frames > 0) {
 571                                 printf("Card %u dropped %d frames before this\n",
 572                                         card_index, int(card_copy[card_index].dropped_frames));
 573                         }
 574                 }
 575
 576                 // If the first card is reporting a corrupted or otherwise dropped frame,
 577                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 578                 if (card_copy[0].new_frame->len == 0) {
 579                         ++stats_dropped_frames;
 580                         pts_int += card_copy[0].new_frame_length;
 581                         continue;
 582                 }
 583
 584                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 585                         CaptureCard *card = &card_copy[card_index];
 586                         if (!card->new_data_ready || card->new_frame->len == 0)
 587                                 continue;
 588
 589                         assert(card->new_frame != nullptr);
 590                         insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
 591                         check_error();
 592
 593                         // The new texture might still be uploaded,
 594                         // tell the GPU to wait until it's there.
 595                         if (card->new_data_ready_fence) {
 596                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 597                                 check_error();
 598                                 glDeleteSync(card->new_data_ready_fence);
 599                                 check_error();
 600                         }
 601                 }
 602
 603                 // Get the main chain from the theme, and set its state immediately.
 604                 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 605                 EffectChain *chain = theme_main_chain.chain;
 606                 theme_main_chain.setup_chain();
 607                 //theme_main_chain.chain->enable_phase_timing(true);
 608
 609                 GLuint y_tex, cbcr_tex;
 610                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 611                 assert(got_frame);
 612
 613                 // Render main chain.
 614                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 615                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 616                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 617                 check_error();
 618                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 619                 resource_pool->release_fbo(fbo);
 620
 621                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 622                 resource_pool->release_2d_texture(cbcr_full_tex);
 623
 624                 // Set the right state for rgba_tex.
 625                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 626                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 627                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 628                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 629                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 630
 631                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 632                 check_error();
 633
 634                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 635                 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
 636                 ++frame;
 637                 pts_int += card_copy[0].new_frame_length;
 638
 639                 // The live frame just shows the RGBA texture we just rendered.
 640                 // It owns rgba_tex now.
 641                 DisplayFrame live_frame;
 642                 live_frame.chain = display_chain.get();
 643                 live_frame.setup_chain = [this, rgba_tex]{
 644                         display_input->set_texture_num(rgba_tex);
 645                 };
 646                 live_frame.ready_fence = fence;
 647                 live_frame.input_frames = {};
 648                 live_frame.temp_textures = { rgba_tex };
 649                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 650
 651                 // Set up preview and any additional channels.
 652                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 653                         DisplayFrame display_frame;
 654                         Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 655                         display_frame.chain = chain.chain;
 656                         display_frame.setup_chain = chain.setup_chain;
 657                         display_frame.ready_fence = fence;
 658                         display_frame.input_frames = chain.input_frames;
 659                         display_frame.temp_textures = {};
 660                         output_channel[i].output_frame(display_frame);
 661                 }
 662
 663                 clock_gettime(CLOCK_MONOTONIC, &now);
 664                 double elapsed = now.tv_sec - start.tv_sec +
 665                         1e-9 * (now.tv_nsec - start.tv_nsec);
 666                 if (frame % 100 == 0) {
 667                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 668                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 669                                 1e3 * elapsed / frame);
 670                 //      chain->print_phase_timing();
 671                 }
 672
 673                 if (should_cut.exchange(false)) {  // Test and clear.
 674                         string filename = generate_local_dump_filename(frame);
 675                         printf("Starting new recording: %s\n", filename.c_str());
 676                         h264_encoder->shutdown();
 677                         httpd.close_output_file();
 678                         httpd.open_output_file(filename.c_str());
 679                         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 680                 }
 681
 682 #if 0
 683                 // Reset every 100 frames, so that local variations in frame times
 684                 // (especially for the first few frames, when the shaders are
 685                 // compiled etc.) don't make it hard to measure for the entire
 686                 // remaining duration of the program.
 687                 if (frame == 10000) {
 688                         frame = 0;
 689                         start = now;
 690                 }
 691 #endif
 692                 check_error();
 693         }
 694
 695         resource_pool->clean_context();
 696 }
 697
 698 void Mixer::audio_thread_func()
 699 {
 700         while (!should_quit) {
 701                 AudioTask task;
 702
 703                 {
 704                         unique_lock<mutex> lock(audio_mutex);
 705                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 706                         task = audio_task_queue.front();
 707                         audio_task_queue.pop();
 708                 }
 709
 710                 process_audio_one_frame(task.pts_int, task.num_samples);
 711         }
 712 }
 713
 714 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 715 {
 716         vector<float> samples_card;
 717         vector<float> samples_out;
 718
 719         // TODO: Allow mixing audio from several sources.
 720         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 721         assert(selected_audio_card < num_cards);
 722
 723         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 724                 samples_card.resize(num_samples * 2);
 725                 {
 726                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 727                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 728                                 printf("Card %d reported previous underrun.\n", card_index);
 729                         }
 730                 }
 731                 if (card_index == selected_audio_card) {
 732                         samples_out = move(samples_card);
 733                 }
 734         }
 735
 736         // Cut away everything under 120 Hz (or whatever the cutoff is);
 737         // we don't need it for voice, and it will reduce headroom
 738         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 739         // should be dampened.)
 740         if (locut_enabled) {
 741                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 742         }
 743
 744         // Apply a level compressor to get the general level right.
 745         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 746         // (or more precisely, near it, since we don't use infinite ratio),
 747         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 748         // entirely arbitrary, but from practical tests with speech, it seems to
 749         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 750         {
 751                 unique_lock<mutex> lock(compressor_mutex);
 752                 if (level_compressor_enabled) {
 753                         float threshold = 0.01f;   // -40 dBFS.
 754                         float ratio = 20.0f;
 755                         float attack_time = 0.5f;
 756                         float release_time = 20.0f;
 757                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 758                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 759                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 760                 } else {
 761                         // Just apply the gain we already had.
 762                         float g = pow(10.0f, gain_staging_db / 20.0f);
 763                         for (size_t i = 0; i < samples_out.size(); ++i) {
 764                                 samples_out[i] *= g;
 765                         }
 766                 }
 767         }
 768
 769 #if 0
 770         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 771                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 772                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 773                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 774 #endif
 775
 776 //      float limiter_att, compressor_att;
 777
 778         // The real compressor.
 779         if (compressor_enabled) {
 780                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 781                 float ratio = 20.0f;
 782                 float attack_time = 0.005f;
 783                 float release_time = 0.040f;
 784                 float makeup_gain = 2.0f;  // +6 dB.
 785                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 786 //              compressor_att = compressor.get_attenuation();
 787         }
 788
 789         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 790         // Note that since ratio is not infinite, we could go slightly higher than this.
 791         if (limiter_enabled) {
 792                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 793                 float ratio = 30.0f;
 794                 float attack_time = 0.0f;  // Instant.
 795                 float release_time = 0.020f;
 796                 float makeup_gain = 1.0f;  // 0 dB.
 797                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 798 //              limiter_att = limiter.get_attenuation();
 799         }
 800
 801 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 802
 803         // Upsample 4x to find interpolated peak.
 804         peak_resampler.inp_data = samples_out.data();
 805         peak_resampler.inp_count = samples_out.size() / 2;
 806
 807         vector<float> interpolated_samples_out;
 808         interpolated_samples_out.resize(samples_out.size());
 809         while (peak_resampler.inp_count > 0) {  // About four iterations.
 810                 peak_resampler.out_data = &interpolated_samples_out[0];
 811                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 812                 peak_resampler.process();
 813                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 814                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 815                 peak_resampler.out_data = nullptr;
 816         }
 817
 818         // At this point, we are most likely close to +0 LU, but all of our
 819         // measurements have been on raw sample values, not R128 values.
 820         // So we have a final makeup gain to get us to +0 LU; the gain
 821         // adjustments required should be relatively small, and also, the
 822         // offset shouldn't change much (only if the type of audio changes
 823         // significantly). Thus, we shoot for updating this value basically
 824         // “whenever we process buffers”, since the R128 calculation isn't exactly
 825         // something we get out per-sample.
 826         //
 827         // Note that there's a feedback loop here, so we choose a very slow filter
 828         // (half-time of 100 seconds).
 829         double target_loudness_factor, alpha;
 830         {
 831                 unique_lock<mutex> lock(compressor_mutex);
 832                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
 833                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
 834                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
 835
 836                 // If we're outside +/- 5 LU uncorrected, we don't count it as
 837                 // a normal signal (probably silence) and don't change the
 838                 // correction factor; just apply what we already have.
 839                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 840                         alpha = 0.0;
 841                 } else {
 842                         // Formula adapted from
 843                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 844                         const double half_time_s = 100.0;
 845                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 846                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 847                 }
 848
 849                 double m = final_makeup_gain;
 850                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 851                         samples_out[i + 0] *= m;
 852                         samples_out[i + 1] *= m;
 853                         m += (target_loudness_factor - m) * alpha;
 854                 }
 855                 final_makeup_gain = m;
 856         }
 857
 858         // Find R128 levels and L/R correlation.
 859         vector<float> left, right;
 860         deinterleave_samples(samples_out, &left, &right);
 861         float *ptrs[] = { left.data(), right.data() };
 862         {
 863                 unique_lock<mutex> lock(compressor_mutex);
 864                 r128.process(left.size(), ptrs);
 865                 correlation.process_samples(samples_out);
 866         }
 867
 868         // Send the samples to the sound card.
 869         if (alsa) {
 870                 alsa->write(samples_out);
 871         }
 872
 873         // And finally add them to the output.
 874         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 875 }
 876
 877 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 878 {
 879         GLuint vao;
 880         glGenVertexArrays(1, &vao);
 881         check_error();
 882
 883         float vertices[] = {
 884                 0.0f, 2.0f,
 885                 0.0f, 0.0f,
 886                 2.0f, 0.0f
 887         };
 888
 889         glBindVertexArray(vao);
 890         check_error();
 891
 892         // Extract Cb/Cr.
 893         GLuint fbo = resource_pool->create_fbo(dst_tex);
 894         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 895         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 896         check_error();
 897
 898         glUseProgram(cbcr_program_num);
 899         check_error();
 900
 901         glActiveTexture(GL_TEXTURE0);
 902         check_error();
 903         glBindTexture(GL_TEXTURE_2D, src_tex);
 904         check_error();
 905         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 906         check_error();
 907         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 908         check_error();
 909         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 910         check_error();
 911
 912         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 913         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 914
 915         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 916         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 917
 918         glDrawArrays(GL_TRIANGLES, 0, 3);
 919         check_error();
 920
 921         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 922         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 923
 924         glUseProgram(0);
 925         check_error();
 926
 927         resource_pool->release_fbo(fbo);
 928         glDeleteVertexArrays(1, &vao);
 929 }
 930
 931 void Mixer::release_display_frame(DisplayFrame *frame)
 932 {
 933         for (GLuint texnum : frame->temp_textures) {
 934                 resource_pool->release_2d_texture(texnum);
 935         }
 936         frame->temp_textures.clear();
 937         frame->ready_fence.reset();
 938         frame->input_frames.clear();
 939 }
 940
 941 void Mixer::start()
 942 {
 943         mixer_thread = thread(&Mixer::thread_func, this);
 944         audio_thread = thread(&Mixer::audio_thread_func, this);
 945 }
 946
 947 void Mixer::quit()
 948 {
 949         should_quit = true;
 950         mixer_thread.join();
 951         audio_thread.join();
 952 }
 953
 954 void Mixer::transition_clicked(int transition_num)
 955 {
 956         theme->transition_clicked(transition_num, pts());
 957 }
 958
 959 void Mixer::channel_clicked(int preview_num)
 960 {
 961         theme->channel_clicked(preview_num);
 962 }
 963
 964 void Mixer::reset_meters()
 965 {
 966         peak_resampler.reset();
 967         peak = 0.0f;
 968         r128.reset();
 969         r128.integr_start();
 970         correlation.reset();
 971 }
 972
 973 Mixer::OutputChannel::~OutputChannel()
 974 {
 975         if (has_current_frame) {
 976                 parent->release_display_frame(&current_frame);
 977         }
 978         if (has_ready_frame) {
 979                 parent->release_display_frame(&ready_frame);
 980         }
 981 }
 982
 983 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 984 {
 985         // Store this frame for display. Remove the ready frame if any
 986         // (it was seemingly never used).
 987         {
 988                 unique_lock<mutex> lock(frame_mutex);
 989                 if (has_ready_frame) {
 990                         parent->release_display_frame(&ready_frame);
 991                 }
 992                 ready_frame = frame;
 993                 has_ready_frame = true;
 994         }
 995
 996         if (has_new_frame_ready_callback) {
 997                 new_frame_ready_callback();
 998         }
 999 }
1000
1001 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1002 {
1003         unique_lock<mutex> lock(frame_mutex);
1004         if (!has_current_frame && !has_ready_frame) {
1005                 return false;
1006         }
1007
1008         if (has_current_frame && has_ready_frame) {
1009                 // We have a new ready frame. Toss the current one.
1010                 parent->release_display_frame(&current_frame);
1011                 has_current_frame = false;
1012         }
1013         if (has_ready_frame) {
1014                 assert(!has_current_frame);
1015                 current_frame = ready_frame;
1016                 ready_frame.ready_fence.reset();  // Drop the refcount.
1017                 ready_frame.input_frames.clear();  // Drop the refcounts.
1018                 has_current_frame = true;
1019                 has_ready_frame = false;
1020         }
1021
1022         *frame = current_frame;
1023         return true;
1024 }
1025
1026 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1027 {
1028         new_frame_ready_callback = callback;
1029         has_new_frame_ready_callback = true;
1030 }