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         float vertices[] = {
 203                 0.0f, 2.0f,
 204                 0.0f, 0.0f,
 205                 2.0f, 0.0f
 206         };
 207         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 208         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 209         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 210
 211         r128.init(2, OUTPUT_FREQUENCY);
 212         r128.integr_start();
 213
 214         locut.init(FILTER_HPF, 2);
 215
 216         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 217         // and there's a limit to how important the peak meter is.
 218         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 219
 220         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 221 }
 222
 223 Mixer::~Mixer()
 224 {
 225         resource_pool->release_glsl_program(cbcr_program_num);
 226         glDeleteBuffers(1, &cbcr_vbo);
 227         BMUSBCapture::stop_bm_thread();
 228
 229         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 230                 {
 231                         unique_lock<mutex> lock(bmusb_mutex);
 232                         cards[card_index].should_quit = true;  // Unblock thread.
 233                         cards[card_index].new_data_ready_changed.notify_all();
 234                 }
 235                 cards[card_index].usb->stop_dequeue_thread();
 236         }
 237
 238         h264_encoder.reset(nullptr);
 239 }
 240
 241 namespace {
 242
 243 int unwrap_timecode(uint16_t current_wrapped, int last)
 244 {
 245         uint16_t last_wrapped = last & 0xffff;
 246         if (current_wrapped > last_wrapped) {
 247                 return (last & ~0xffff) | current_wrapped;
 248         } else {
 249                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 250         }
 251 }
 252
 253 float find_peak(const float *samples, size_t num_samples)
 254 {
 255         float m = fabs(samples[0]);
 256         for (size_t i = 1; i < num_samples; ++i) {
 257                 m = max(m, fabs(samples[i]));
 258         }
 259         return m;
 260 }
 261
 262 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 263 {
 264         size_t num_samples = in.size() / 2;
 265         out_l->resize(num_samples);
 266         out_r->resize(num_samples);
 267
 268         const float *inptr = in.data();
 269         float *lptr = &(*out_l)[0];
 270         float *rptr = &(*out_r)[0];
 271         for (size_t i = 0; i < num_samples; ++i) {
 272                 *lptr++ = *inptr++;
 273                 *rptr++ = *inptr++;
 274         }
 275 }
 276
 277 }  // namespace
 278
 279 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 280                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 281                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 282 {
 283         CaptureCard *card = &cards[card_index];
 284
 285         unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
 286         bool interlaced;
 287
 288         decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom,
 289                             &frame_rate_nom, &frame_rate_den, &interlaced);  // Ignore return value for now.
 290         int64_t frame_length = int64_t(TIMEBASE * frame_rate_den) / frame_rate_nom;
 291
 292         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 293         if (num_samples > OUTPUT_FREQUENCY / 10) {
 294                 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",
 295                         card_index, int(audio_frame.len), int(audio_offset),
 296                         timecode, int(video_frame.len), int(video_offset), video_format);
 297                 if (video_frame.owner) {
 298                         video_frame.owner->release_frame(video_frame);
 299                 }
 300                 if (audio_frame.owner) {
 301                         audio_frame.owner->release_frame(audio_frame);
 302                 }
 303                 return;
 304         }
 305
 306         int64_t local_pts = card->next_local_pts;
 307         int dropped_frames = 0;
 308         if (card->last_timecode != -1) {
 309                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 310         }
 311
 312         // Convert the audio to stereo fp32 and add it.
 313         vector<float> audio;
 314         audio.resize(num_samples * 2);
 315         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 316
 317         // Add the audio.
 318         {
 319                 unique_lock<mutex> lock(card->audio_mutex);
 320
 321                 // Number of samples per frame if we need to insert silence.
 322                 // (Could be nonintegral, but resampling will save us then.)
 323                 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
 324
 325                 if (dropped_frames > MAX_FPS * 2) {
 326                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 327                                 card_index, card->last_timecode, timecode);
 328                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 329                         dropped_frames = 0;
 330                 } else if (dropped_frames > 0) {
 331                         // Insert silence as needed.
 332                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 333                                 card_index, dropped_frames, timecode);
 334                         vector<float> silence(silence_samples * 2, 0.0f);
 335                         for (int i = 0; i < dropped_frames; ++i) {
 336                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 337                                 // Note that if the format changed in the meantime, we have
 338                                 // no way of detecting that; we just have to assume the frame length
 339                                 // is always the same.
 340                                 local_pts += frame_length;
 341                         }
 342                 }
 343                 if (num_samples == 0) {
 344                         audio.resize(silence_samples * 2);
 345                         num_samples = silence_samples;
 346                 }
 347                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 348                 card->next_local_pts = local_pts + frame_length;
 349         }
 350
 351         card->last_timecode = timecode;
 352
 353         // Done with the audio, so release it.
 354         if (audio_frame.owner) {
 355                 audio_frame.owner->release_frame(audio_frame);
 356         }
 357
 358         {
 359                 // Wait until the previous frame was consumed.
 360                 unique_lock<mutex> lock(bmusb_mutex);
 361                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 362                 if (card->should_quit) return;
 363         }
 364
 365         size_t expected_length = width * (height + extra_lines_top + extra_lines_bottom) * 2;
 366         if (video_frame.len - video_offset == 0 ||
 367             video_frame.len - video_offset != expected_length) {
 368                 if (video_frame.len != 0) {
 369                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 370                                 card_index, video_frame.len - video_offset, expected_length);
 371                 }
 372                 if (video_frame.owner) {
 373                         video_frame.owner->release_frame(video_frame);
 374                 }
 375
 376                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 377                 // so that pts can go up accordingly.
 378                 {
 379                         unique_lock<mutex> lock(bmusb_mutex);
 380                         card->new_data_ready = true;
 381                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 382                         card->new_frame_length = frame_length;
 383                         card->new_frame_interlaced = false;
 384                         card->new_data_ready_fence = nullptr;
 385                         card->dropped_frames = dropped_frames;
 386                         card->new_data_ready_changed.notify_all();
 387                 }
 388                 return;
 389         }
 390
 391         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 392
 393         unsigned num_fields = interlaced ? 2 : 1;
 394         timespec frame_upload_start;
 395         if (interlaced) {
 396                 // Send the two fields along as separate frames; the other side will need to add
 397                 // a deinterlacer to actually get this right.
 398                 assert(height % 2 == 0);
 399                 height /= 2;
 400                 assert(frame_length % 2 == 0);
 401                 frame_length /= 2;
 402                 num_fields = 2;
 403                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 404         }
 405         userdata->last_interlaced = interlaced;
 406         userdata->last_frame_rate_nom = frame_rate_nom;
 407         userdata->last_frame_rate_den = frame_rate_den;
 408         RefCountedFrame new_frame(video_frame);
 409
 410         // Upload the textures.
 411         size_t cbcr_width = width / 2;
 412         size_t cbcr_offset = video_offset / 2;
 413         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 414
 415         for (unsigned field = 0; field < num_fields; ++field) {
 416                 unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22);
 417
 418                 if (userdata->tex_y[field] == 0 ||
 419                     userdata->tex_cbcr[field] == 0 ||
 420                     width != userdata->last_width[field] ||
 421                     height != userdata->last_height[field]) {
 422                         // We changed resolution since last use of this texture, so we need to create
 423                         // a new object. Note that this each card has its own PBOFrameAllocator,
 424                         // we don't need to worry about these flip-flopping between resolutions.
 425                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 426                         check_error();
 427                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 428                         check_error();
 429                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 430                         check_error();
 431                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 432                         check_error();
 433                         userdata->last_width[field] = width;
 434                         userdata->last_height[field] = height;
 435                 }
 436
 437                 GLuint pbo = userdata->pbo;
 438                 check_error();
 439                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 440                 check_error();
 441                 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 442                 check_error();
 443
 444                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 445                 check_error();
 446                 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)));
 447                 check_error();
 448                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 449                 check_error();
 450                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line));
 451                 check_error();
 452                 glBindTexture(GL_TEXTURE_2D, 0);
 453                 check_error();
 454                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 455                 check_error();
 456                 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 457                 check_error();
 458                 assert(fence != nullptr);
 459
 460                 if (field == 1) {
 461                         // Don't upload the second field as fast as we can; wait until
 462                         // the field time has approximately passed. (Otherwise, we could
 463                         // get timing jitter against the other sources, and possibly also
 464                         // against the video display, although the latter is not as critical.)
 465                         // This requires our system clock to be reasonably close to the
 466                         // video clock, but that's not an unreasonable assumption.
 467                         timespec second_field_start;
 468                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 469                                 frame_length * 1000000000 / TIMEBASE;
 470                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 471                                 second_field_start.tv_nsec / 1000000000;
 472                         second_field_start.tv_nsec %= 1000000000;
 473
 474                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 475                                                &second_field_start, nullptr) == -1 &&
 476                                errno == EINTR) ;
 477                 }
 478
 479                 {
 480                         unique_lock<mutex> lock(bmusb_mutex);
 481                         card->new_data_ready = true;
 482                         card->new_frame = new_frame;
 483                         card->new_frame_length = frame_length;
 484                         card->new_frame_field = field;
 485                         card->new_frame_interlaced = interlaced;
 486                         card->new_data_ready_fence = fence;
 487                         card->dropped_frames = dropped_frames;
 488                         card->new_data_ready_changed.notify_all();
 489
 490                         if (field != num_fields - 1) {
 491                                 // Wait until the previous frame was consumed.
 492                                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 493                                 if (card->should_quit) return;
 494                         }
 495                 }
 496         }
 497 }
 498
 499 void Mixer::thread_func()
 500 {
 501         eglBindAPI(EGL_OPENGL_API);
 502         QOpenGLContext *context = create_context(mixer_surface);
 503         if (!make_current(context, mixer_surface)) {
 504                 printf("oops\n");
 505                 exit(1);
 506         }
 507
 508         struct timespec start, now;
 509         clock_gettime(CLOCK_MONOTONIC, &start);
 510
 511         int frame = 0;
 512         int stats_dropped_frames = 0;
 513
 514         while (!should_quit) {
 515                 CaptureCard card_copy[MAX_CARDS];
 516                 int num_samples[MAX_CARDS];
 517
 518                 {
 519                         unique_lock<mutex> lock(bmusb_mutex);
 520
 521                         // The first card is the master timer, so wait for it to have a new frame.
 522                         // TODO: Make configurable, and with a timeout.
 523                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 524
 525                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 526                                 CaptureCard *card = &cards[card_index];
 527                                 card_copy[card_index].usb = card->usb;
 528                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 529                                 card_copy[card_index].new_frame = card->new_frame;
 530                                 card_copy[card_index].new_frame_length = card->new_frame_length;
 531                                 card_copy[card_index].new_frame_field = card->new_frame_field;
 532                                 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
 533                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 534                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 535                                 card->new_data_ready = false;
 536                                 card->new_data_ready_changed.notify_all();
 537
 538                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
 539                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 540                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 541                                 assert(num_samples[card_index] >= 0);
 542                         }
 543                 }
 544
 545                 // Resample the audio as needed, including from previously dropped frames.
 546                 assert(num_cards > 0);
 547                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 548                         {
 549                                 // Signal to the audio thread to process this frame.
 550                                 unique_lock<mutex> lock(audio_mutex);
 551                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
 552                                 audio_task_queue_changed.notify_one();
 553                         }
 554                         if (frame_num != card_copy[0].dropped_frames) {
 555                                 // For dropped frames, increase the pts. Note that if the format changed
 556                                 // in the meantime, we have no way of detecting that; we just have to
 557                                 // assume the frame length is always the same.
 558                                 ++stats_dropped_frames;
 559                                 pts_int += card_copy[0].new_frame_length;
 560                         }
 561                 }
 562
 563                 if (audio_level_callback != nullptr) {
 564                         unique_lock<mutex> lock(compressor_mutex);
 565                         double loudness_s = r128.loudness_S();
 566                         double loudness_i = r128.integrated();
 567                         double loudness_range_low = r128.range_min();
 568                         double loudness_range_high = r128.range_max();
 569
 570                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 571                                              loudness_i, loudness_range_low, loudness_range_high,
 572                                              gain_staging_db, 20.0 * log10(final_makeup_gain),
 573                                              correlation.get_correlation());
 574                 }
 575
 576                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 577                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 578                                 ++card_copy[card_index].dropped_frames;
 579                         }
 580                         if (card_copy[card_index].dropped_frames > 0) {
 581                                 printf("Card %u dropped %d frames before this\n",
 582                                         card_index, int(card_copy[card_index].dropped_frames));
 583                         }
 584                 }
 585
 586                 // If the first card is reporting a corrupted or otherwise dropped frame,
 587                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 588                 if (card_copy[0].new_frame->len == 0) {
 589                         ++stats_dropped_frames;
 590                         pts_int += card_copy[0].new_frame_length;
 591                         continue;
 592                 }
 593
 594                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 595                         CaptureCard *card = &card_copy[card_index];
 596                         if (!card->new_data_ready || card->new_frame->len == 0)
 597                                 continue;
 598
 599                         assert(card->new_frame != nullptr);
 600                         insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
 601                         check_error();
 602
 603                         // The new texture might still be uploaded,
 604                         // tell the GPU to wait until it's there.
 605                         if (card->new_data_ready_fence) {
 606                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 607                                 check_error();
 608                                 glDeleteSync(card->new_data_ready_fence);
 609                                 check_error();
 610                         }
 611                 }
 612
 613                 // Get the main chain from the theme, and set its state immediately.
 614                 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 615                 EffectChain *chain = theme_main_chain.chain;
 616                 theme_main_chain.setup_chain();
 617                 //theme_main_chain.chain->enable_phase_timing(true);
 618
 619                 GLuint y_tex, cbcr_tex;
 620                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 621                 assert(got_frame);
 622
 623                 // Render main chain.
 624                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 625                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 626                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 627                 check_error();
 628                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 629                 resource_pool->release_fbo(fbo);
 630
 631                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 632                 resource_pool->release_2d_texture(cbcr_full_tex);
 633
 634                 // Set the right state for rgba_tex.
 635                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 636                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 637                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 638                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 639                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 640
 641                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 642                 check_error();
 643
 644                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 645                 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
 646                 ++frame;
 647                 pts_int += card_copy[0].new_frame_length;
 648
 649                 // The live frame just shows the RGBA texture we just rendered.
 650                 // It owns rgba_tex now.
 651                 DisplayFrame live_frame;
 652                 live_frame.chain = display_chain.get();
 653                 live_frame.setup_chain = [this, rgba_tex]{
 654                         display_input->set_texture_num(rgba_tex);
 655                 };
 656                 live_frame.ready_fence = fence;
 657                 live_frame.input_frames = {};
 658                 live_frame.temp_textures = { rgba_tex };
 659                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 660
 661                 // Set up preview and any additional channels.
 662                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 663                         DisplayFrame display_frame;
 664                         Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 665                         display_frame.chain = chain.chain;
 666                         display_frame.setup_chain = chain.setup_chain;
 667                         display_frame.ready_fence = fence;
 668                         display_frame.input_frames = chain.input_frames;
 669                         display_frame.temp_textures = {};
 670                         output_channel[i].output_frame(display_frame);
 671                 }
 672
 673                 clock_gettime(CLOCK_MONOTONIC, &now);
 674                 double elapsed = now.tv_sec - start.tv_sec +
 675                         1e-9 * (now.tv_nsec - start.tv_nsec);
 676                 if (frame % 100 == 0) {
 677                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 678                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 679                                 1e3 * elapsed / frame);
 680                 //      chain->print_phase_timing();
 681                 }
 682
 683                 if (should_cut.exchange(false)) {  // Test and clear.
 684                         string filename = generate_local_dump_filename(frame);
 685                         printf("Starting new recording: %s\n", filename.c_str());
 686                         h264_encoder->shutdown();
 687                         httpd.close_output_file();
 688                         httpd.open_output_file(filename.c_str());
 689                         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 690                 }
 691
 692 #if 0
 693                 // Reset every 100 frames, so that local variations in frame times
 694                 // (especially for the first few frames, when the shaders are
 695                 // compiled etc.) don't make it hard to measure for the entire
 696                 // remaining duration of the program.
 697                 if (frame == 10000) {
 698                         frame = 0;
 699                         start = now;
 700                 }
 701 #endif
 702                 check_error();
 703         }
 704
 705         resource_pool->clean_context();
 706 }
 707
 708 void Mixer::audio_thread_func()
 709 {
 710         while (!should_quit) {
 711                 AudioTask task;
 712
 713                 {
 714                         unique_lock<mutex> lock(audio_mutex);
 715                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 716                         task = audio_task_queue.front();
 717                         audio_task_queue.pop();
 718                 }
 719
 720                 process_audio_one_frame(task.pts_int, task.num_samples);
 721         }
 722 }
 723
 724 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 725 {
 726         vector<float> samples_card;
 727         vector<float> samples_out;
 728
 729         // TODO: Allow mixing audio from several sources.
 730         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 731         assert(selected_audio_card < num_cards);
 732
 733         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 734                 samples_card.resize(num_samples * 2);
 735                 {
 736                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 737                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 738                                 printf("Card %d reported previous underrun.\n", card_index);
 739                         }
 740                 }
 741                 if (card_index == selected_audio_card) {
 742                         samples_out = move(samples_card);
 743                 }
 744         }
 745
 746         // Cut away everything under 120 Hz (or whatever the cutoff is);
 747         // we don't need it for voice, and it will reduce headroom
 748         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 749         // should be dampened.)
 750         if (locut_enabled) {
 751                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 752         }
 753
 754         // Apply a level compressor to get the general level right.
 755         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 756         // (or more precisely, near it, since we don't use infinite ratio),
 757         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 758         // entirely arbitrary, but from practical tests with speech, it seems to
 759         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 760         {
 761                 unique_lock<mutex> lock(compressor_mutex);
 762                 if (level_compressor_enabled) {
 763                         float threshold = 0.01f;   // -40 dBFS.
 764                         float ratio = 20.0f;
 765                         float attack_time = 0.5f;
 766                         float release_time = 20.0f;
 767                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 768                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 769                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 770                 } else {
 771                         // Just apply the gain we already had.
 772                         float g = pow(10.0f, gain_staging_db / 20.0f);
 773                         for (size_t i = 0; i < samples_out.size(); ++i) {
 774                                 samples_out[i] *= g;
 775                         }
 776                 }
 777         }
 778
 779 #if 0
 780         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 781                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 782                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 783                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 784 #endif
 785
 786 //      float limiter_att, compressor_att;
 787
 788         // The real compressor.
 789         if (compressor_enabled) {
 790                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 791                 float ratio = 20.0f;
 792                 float attack_time = 0.005f;
 793                 float release_time = 0.040f;
 794                 float makeup_gain = 2.0f;  // +6 dB.
 795                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 796 //              compressor_att = compressor.get_attenuation();
 797         }
 798
 799         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 800         // Note that since ratio is not infinite, we could go slightly higher than this.
 801         if (limiter_enabled) {
 802                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 803                 float ratio = 30.0f;
 804                 float attack_time = 0.0f;  // Instant.
 805                 float release_time = 0.020f;
 806                 float makeup_gain = 1.0f;  // 0 dB.
 807                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 808 //              limiter_att = limiter.get_attenuation();
 809         }
 810
 811 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 812
 813         // Upsample 4x to find interpolated peak.
 814         peak_resampler.inp_data = samples_out.data();
 815         peak_resampler.inp_count = samples_out.size() / 2;
 816
 817         vector<float> interpolated_samples_out;
 818         interpolated_samples_out.resize(samples_out.size());
 819         while (peak_resampler.inp_count > 0) {  // About four iterations.
 820                 peak_resampler.out_data = &interpolated_samples_out[0];
 821                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 822                 peak_resampler.process();
 823                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 824                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 825                 peak_resampler.out_data = nullptr;
 826         }
 827
 828         // At this point, we are most likely close to +0 LU, but all of our
 829         // measurements have been on raw sample values, not R128 values.
 830         // So we have a final makeup gain to get us to +0 LU; the gain
 831         // adjustments required should be relatively small, and also, the
 832         // offset shouldn't change much (only if the type of audio changes
 833         // significantly). Thus, we shoot for updating this value basically
 834         // “whenever we process buffers”, since the R128 calculation isn't exactly
 835         // something we get out per-sample.
 836         //
 837         // Note that there's a feedback loop here, so we choose a very slow filter
 838         // (half-time of 100 seconds).
 839         double target_loudness_factor, alpha;
 840         {
 841                 unique_lock<mutex> lock(compressor_mutex);
 842                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
 843                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
 844                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
 845
 846                 // If we're outside +/- 5 LU uncorrected, we don't count it as
 847                 // a normal signal (probably silence) and don't change the
 848                 // correction factor; just apply what we already have.
 849                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 850                         alpha = 0.0;
 851                 } else {
 852                         // Formula adapted from
 853                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 854                         const double half_time_s = 100.0;
 855                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 856                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 857                 }
 858
 859                 double m = final_makeup_gain;
 860                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 861                         samples_out[i + 0] *= m;
 862                         samples_out[i + 1] *= m;
 863                         m += (target_loudness_factor - m) * alpha;
 864                 }
 865                 final_makeup_gain = m;
 866         }
 867
 868         // Find R128 levels and L/R correlation.
 869         vector<float> left, right;
 870         deinterleave_samples(samples_out, &left, &right);
 871         float *ptrs[] = { left.data(), right.data() };
 872         {
 873                 unique_lock<mutex> lock(compressor_mutex);
 874                 r128.process(left.size(), ptrs);
 875                 correlation.process_samples(samples_out);
 876         }
 877
 878         // Send the samples to the sound card.
 879         if (alsa) {
 880                 alsa->write(samples_out);
 881         }
 882
 883         // And finally add them to the output.
 884         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 885 }
 886
 887 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 888 {
 889         GLuint vao;
 890         glGenVertexArrays(1, &vao);
 891         check_error();
 892
 893         glBindVertexArray(vao);
 894         check_error();
 895
 896         // Extract Cb/Cr.
 897         GLuint fbo = resource_pool->create_fbo(dst_tex);
 898         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 899         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 900         check_error();
 901
 902         glUseProgram(cbcr_program_num);
 903         check_error();
 904
 905         glActiveTexture(GL_TEXTURE0);
 906         check_error();
 907         glBindTexture(GL_TEXTURE_2D, src_tex);
 908         check_error();
 909         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 910         check_error();
 911         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 912         check_error();
 913         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 914         check_error();
 915
 916         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 917         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 918
 919         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
 920         check_error();
 921
 922         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
 923                 glEnableVertexAttribArray(attr_index);
 924                 check_error();
 925                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 926                 check_error();
 927         }
 928
 929         glDrawArrays(GL_TRIANGLES, 0, 3);
 930         check_error();
 931
 932         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
 933                 glDisableVertexAttribArray(attr_index);
 934                 check_error();
 935         }
 936
 937         glUseProgram(0);
 938         check_error();
 939         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 940         check_error();
 941
 942         resource_pool->release_fbo(fbo);
 943         glDeleteVertexArrays(1, &vao);
 944 }
 945
 946 void Mixer::release_display_frame(DisplayFrame *frame)
 947 {
 948         for (GLuint texnum : frame->temp_textures) {
 949                 resource_pool->release_2d_texture(texnum);
 950         }
 951         frame->temp_textures.clear();
 952         frame->ready_fence.reset();
 953         frame->input_frames.clear();
 954 }
 955
 956 void Mixer::start()
 957 {
 958         mixer_thread = thread(&Mixer::thread_func, this);
 959         audio_thread = thread(&Mixer::audio_thread_func, this);
 960 }
 961
 962 void Mixer::quit()
 963 {
 964         should_quit = true;
 965         mixer_thread.join();
 966         audio_thread.join();
 967 }
 968
 969 void Mixer::transition_clicked(int transition_num)
 970 {
 971         theme->transition_clicked(transition_num, pts());
 972 }
 973
 974 void Mixer::channel_clicked(int preview_num)
 975 {
 976         theme->channel_clicked(preview_num);
 977 }
 978
 979 void Mixer::reset_meters()
 980 {
 981         peak_resampler.reset();
 982         peak = 0.0f;
 983         r128.reset();
 984         r128.integr_start();
 985         correlation.reset();
 986 }
 987
 988 Mixer::OutputChannel::~OutputChannel()
 989 {
 990         if (has_current_frame) {
 991                 parent->release_display_frame(&current_frame);
 992         }
 993         if (has_ready_frame) {
 994                 parent->release_display_frame(&ready_frame);
 995         }
 996 }
 997
 998 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 999 {
1000         // Store this frame for display. Remove the ready frame if any
1001         // (it was seemingly never used).
1002         {
1003                 unique_lock<mutex> lock(frame_mutex);
1004                 if (has_ready_frame) {
1005                         parent->release_display_frame(&ready_frame);
1006                 }
1007                 ready_frame = frame;
1008                 has_ready_frame = true;
1009         }
1010
1011         if (has_new_frame_ready_callback) {
1012                 new_frame_ready_callback();
1013         }
1014 }
1015
1016 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1017 {
1018         unique_lock<mutex> lock(frame_mutex);
1019         if (!has_current_frame && !has_ready_frame) {
1020                 return false;
1021         }
1022
1023         if (has_current_frame && has_ready_frame) {
1024                 // We have a new ready frame. Toss the current one.
1025                 parent->release_display_frame(&current_frame);
1026                 has_current_frame = false;
1027         }
1028         if (has_ready_frame) {
1029                 assert(!has_current_frame);
1030                 current_frame = ready_frame;
1031                 ready_frame.ready_fence.reset();  // Drop the refcount.
1032                 ready_frame.input_frames.clear();  // Drop the refcounts.
1033                 has_current_frame = true;
1034                 has_ready_frame = false;
1035         }
1036
1037         *frame = current_frame;
1038         return true;
1039 }
1040
1041 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1042 {
1043         new_frame_ready_callback = callback;
1044         has_new_frame_ready_callback = true;
1045 }