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                 configure_card(card_index, format, new BMUSBCapture(card_index));
 140         }
 141
 142         BMUSBCapture::start_bm_thread();
 143
 144         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 145                 cards[card_index].capture->start_bm_capture();
 146         }
 147
 148         // Set up stuff for NV12 conversion.
 149
 150         // Cb/Cr shader.
 151         string cbcr_vert_shader =
 152                 "#version 130 \n"
 153                 " \n"
 154                 "in vec2 position; \n"
 155                 "in vec2 texcoord; \n"
 156                 "out vec2 tc0; \n"
 157                 "uniform vec2 foo_chroma_offset_0; \n"
 158                 " \n"
 159                 "void main() \n"
 160                 "{ \n"
 161                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 162                 "    // \n"
 163                 "    //   2.000  0.000  0.000 -1.000 \n"
 164                 "    //   0.000  2.000  0.000 -1.000 \n"
 165                 "    //   0.000  0.000 -2.000 -1.000 \n"
 166                 "    //   0.000  0.000  0.000  1.000 \n"
 167                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 168                 "    vec2 flipped_tc = texcoord; \n"
 169                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
 170                 "} \n";
 171         string cbcr_frag_shader =
 172                 "#version 130 \n"
 173                 "in vec2 tc0; \n"
 174                 "uniform sampler2D cbcr_tex; \n"
 175                 "out vec4 FragColor; \n"
 176                 "void main() { \n"
 177                 "    FragColor = texture(cbcr_tex, tc0); \n"
 178                 "} \n";
 179         vector<string> frag_shader_outputs;
 180         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 181
 182         float vertices[] = {
 183                 0.0f, 2.0f,
 184                 0.0f, 0.0f,
 185                 2.0f, 0.0f
 186         };
 187         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 188         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 189         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 190
 191         r128.init(2, OUTPUT_FREQUENCY);
 192         r128.integr_start();
 193
 194         locut.init(FILTER_HPF, 2);
 195
 196         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 197         // and there's a limit to how important the peak meter is.
 198         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 199
 200         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 201 }
 202
 203 Mixer::~Mixer()
 204 {
 205         resource_pool->release_glsl_program(cbcr_program_num);
 206         glDeleteBuffers(1, &cbcr_vbo);
 207         BMUSBCapture::stop_bm_thread();
 208
 209         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 210                 {
 211                         unique_lock<mutex> lock(bmusb_mutex);
 212                         cards[card_index].should_quit = true;  // Unblock thread.
 213                         cards[card_index].new_data_ready_changed.notify_all();
 214                 }
 215                 cards[card_index].capture->stop_dequeue_thread();
 216         }
 217
 218         h264_encoder.reset(nullptr);
 219 }
 220
 221 void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture)
 222 {
 223         printf("Configuring card %d...\n", card_index);
 224
 225         CaptureCard *card = &cards[card_index];
 226         card->capture = capture;
 227         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 228         card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 229         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 230         card->surface = create_surface(format);
 231         card->capture->set_dequeue_thread_callbacks(
 232                 [card]{
 233                         eglBindAPI(EGL_OPENGL_API);
 234                         card->context = create_context(card->surface);
 235                         if (!make_current(card->context, card->surface)) {
 236                                 printf("failed to create bmusb context\n");
 237                                 exit(1);
 238                         }
 239                 },
 240                 [this]{
 241                         resource_pool->clean_context();
 242                 });
 243         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 244         card->capture->configure_card();
 245 }
 246
 247
 248 namespace {
 249
 250 int unwrap_timecode(uint16_t current_wrapped, int last)
 251 {
 252         uint16_t last_wrapped = last & 0xffff;
 253         if (current_wrapped > last_wrapped) {
 254                 return (last & ~0xffff) | current_wrapped;
 255         } else {
 256                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 257         }
 258 }
 259
 260 float find_peak(const float *samples, size_t num_samples)
 261 {
 262         float m = fabs(samples[0]);
 263         for (size_t i = 1; i < num_samples; ++i) {
 264                 m = max(m, fabs(samples[i]));
 265         }
 266         return m;
 267 }
 268
 269 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 270 {
 271         size_t num_samples = in.size() / 2;
 272         out_l->resize(num_samples);
 273         out_r->resize(num_samples);
 274
 275         const float *inptr = in.data();
 276         float *lptr = &(*out_l)[0];
 277         float *rptr = &(*out_r)[0];
 278         for (size_t i = 0; i < num_samples; ++i) {
 279                 *lptr++ = *inptr++;
 280                 *rptr++ = *inptr++;
 281         }
 282 }
 283
 284 }  // namespace
 285
 286 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 287                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 288                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 289 {
 290         CaptureCard *card = &cards[card_index];
 291
 292         int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom;
 293
 294         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 295         if (num_samples > OUTPUT_FREQUENCY / 10) {
 296                 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",
 297                         card_index, int(audio_frame.len), int(audio_offset),
 298                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 299                 if (video_frame.owner) {
 300                         video_frame.owner->release_frame(video_frame);
 301                 }
 302                 if (audio_frame.owner) {
 303                         audio_frame.owner->release_frame(audio_frame);
 304                 }
 305                 return;
 306         }
 307
 308         int64_t local_pts = card->next_local_pts;
 309         int dropped_frames = 0;
 310         if (card->last_timecode != -1) {
 311                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 312         }
 313
 314         // Convert the audio to stereo fp32 and add it.
 315         vector<float> audio;
 316         audio.resize(num_samples * 2);
 317         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 318
 319         // Add the audio.
 320         {
 321                 unique_lock<mutex> lock(card->audio_mutex);
 322
 323                 // Number of samples per frame if we need to insert silence.
 324                 // (Could be nonintegral, but resampling will save us then.)
 325                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 326
 327                 if (dropped_frames > MAX_FPS * 2) {
 328                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 329                                 card_index, card->last_timecode, timecode);
 330                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 331                         dropped_frames = 0;
 332                 } else if (dropped_frames > 0) {
 333                         // Insert silence as needed.
 334                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 335                                 card_index, dropped_frames, timecode);
 336                         vector<float> silence(silence_samples * 2, 0.0f);
 337                         for (int i = 0; i < dropped_frames; ++i) {
 338                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 339                                 // Note that if the format changed in the meantime, we have
 340                                 // no way of detecting that; we just have to assume the frame length
 341                                 // is always the same.
 342                                 local_pts += frame_length;
 343                         }
 344                 }
 345                 if (num_samples == 0) {
 346                         audio.resize(silence_samples * 2);
 347                         num_samples = silence_samples;
 348                 }
 349                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 350                 card->next_local_pts = local_pts + frame_length;
 351         }
 352
 353         card->last_timecode = timecode;
 354
 355         // Done with the audio, so release it.
 356         if (audio_frame.owner) {
 357                 audio_frame.owner->release_frame(audio_frame);
 358         }
 359
 360         {
 361                 // Wait until the previous frame was consumed.
 362                 unique_lock<mutex> lock(bmusb_mutex);
 363                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 364                 if (card->should_quit) return;
 365         }
 366
 367         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 368         if (video_frame.len - video_offset == 0 ||
 369             video_frame.len - video_offset != expected_length) {
 370                 if (video_frame.len != 0) {
 371                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 372                                 card_index, video_frame.len - video_offset, expected_length);
 373                 }
 374                 if (video_frame.owner) {
 375                         video_frame.owner->release_frame(video_frame);
 376                 }
 377
 378                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 379                 // so that pts can go up accordingly.
 380                 {
 381                         unique_lock<mutex> lock(bmusb_mutex);
 382                         card->new_data_ready = true;
 383                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 384                         card->new_frame_length = frame_length;
 385                         card->new_frame_interlaced = false;
 386                         card->new_data_ready_fence = nullptr;
 387                         card->dropped_frames = dropped_frames;
 388                         card->new_data_ready_changed.notify_all();
 389                 }
 390                 return;
 391         }
 392
 393         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 394
 395         unsigned num_fields = video_format.interlaced ? 2 : 1;
 396         timespec frame_upload_start;
 397         if (video_format.interlaced) {
 398                 // Send the two fields along as separate frames; the other side will need to add
 399                 // a deinterlacer to actually get this right.
 400                 assert(video_format.height % 2 == 0);
 401                 video_format.height /= 2;
 402                 assert(frame_length % 2 == 0);
 403                 frame_length /= 2;
 404                 num_fields = 2;
 405                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 406         }
 407         userdata->last_interlaced = video_format.interlaced;
 408         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 409         userdata->last_frame_rate_den = video_format.frame_rate_den;
 410         RefCountedFrame new_frame(video_frame);
 411
 412         // Upload the textures.
 413         size_t cbcr_width = video_format.width / 2;
 414         size_t cbcr_offset = video_offset / 2;
 415         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 416
 417         for (unsigned field = 0; field < num_fields; ++field) {
 418                 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
 419
 420                 if (userdata->tex_y[field] == 0 ||
 421                     userdata->tex_cbcr[field] == 0 ||
 422                     video_format.width != userdata->last_width[field] ||
 423                     video_format.height != userdata->last_height[field]) {
 424                         // We changed resolution since last use of this texture, so we need to create
 425                         // a new object. Note that this each card has its own PBOFrameAllocator,
 426                         // we don't need to worry about these flip-flopping between resolutions.
 427                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 428                         check_error();
 429                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 430                         check_error();
 431                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 432                         check_error();
 433                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 434                         check_error();
 435                         userdata->last_width[field] = video_format.width;
 436                         userdata->last_height[field] = video_format.height;
 437                 }
 438
 439                 GLuint pbo = userdata->pbo;
 440                 check_error();
 441                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 442                 check_error();
 443                 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 444                 check_error();
 445
 446                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 447                 check_error();
 448                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t)));
 449                 check_error();
 450                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 451                 check_error();
 452                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + video_format.width * field_start_line));
 453                 check_error();
 454                 glBindTexture(GL_TEXTURE_2D, 0);
 455                 check_error();
 456                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 457                 check_error();
 458                 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 459                 check_error();
 460                 assert(fence != nullptr);
 461
 462                 if (field == 1) {
 463                         // Don't upload the second field as fast as we can; wait until
 464                         // the field time has approximately passed. (Otherwise, we could
 465                         // get timing jitter against the other sources, and possibly also
 466                         // against the video display, although the latter is not as critical.)
 467                         // This requires our system clock to be reasonably close to the
 468                         // video clock, but that's not an unreasonable assumption.
 469                         timespec second_field_start;
 470                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 471                                 frame_length * 1000000000 / TIMEBASE;
 472                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 473                                 second_field_start.tv_nsec / 1000000000;
 474                         second_field_start.tv_nsec %= 1000000000;
 475
 476                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 477                                                &second_field_start, nullptr) == -1 &&
 478                                errno == EINTR) ;
 479                 }
 480
 481                 {
 482                         unique_lock<mutex> lock(bmusb_mutex);
 483                         card->new_data_ready = true;
 484                         card->new_frame = new_frame;
 485                         card->new_frame_length = frame_length;
 486                         card->new_frame_field = field;
 487                         card->new_frame_interlaced = video_format.interlaced;
 488                         card->new_data_ready_fence = fence;
 489                         card->dropped_frames = dropped_frames;
 490                         card->new_data_ready_changed.notify_all();
 491
 492                         if (field != num_fields - 1) {
 493                                 // Wait until the previous frame was consumed.
 494                                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 495                                 if (card->should_quit) return;
 496                         }
 497                 }
 498         }
 499 }
 500
 501 void Mixer::thread_func()
 502 {
 503         eglBindAPI(EGL_OPENGL_API);
 504         QOpenGLContext *context = create_context(mixer_surface);
 505         if (!make_current(context, mixer_surface)) {
 506                 printf("oops\n");
 507                 exit(1);
 508         }
 509
 510         struct timespec start, now;
 511         clock_gettime(CLOCK_MONOTONIC, &start);
 512
 513         int frame = 0;
 514         int stats_dropped_frames = 0;
 515
 516         while (!should_quit) {
 517                 CaptureCard card_copy[MAX_CARDS];
 518                 int num_samples[MAX_CARDS];
 519
 520                 {
 521                         unique_lock<mutex> lock(bmusb_mutex);
 522
 523                         // The first card is the master timer, so wait for it to have a new frame.
 524                         // TODO: Make configurable, and with a timeout.
 525                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 526
 527                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 528                                 CaptureCard *card = &cards[card_index];
 529                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 530                                 card_copy[card_index].new_frame = card->new_frame;
 531                                 card_copy[card_index].new_frame_length = card->new_frame_length;
 532                                 card_copy[card_index].new_frame_field = card->new_frame_field;
 533                                 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
 534                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 535                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 536                                 card->new_data_ready = false;
 537                                 card->new_data_ready_changed.notify_all();
 538
 539                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
 540                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 541                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 542                                 assert(num_samples[card_index] >= 0);
 543                         }
 544                 }
 545
 546                 // Resample the audio as needed, including from previously dropped frames.
 547                 assert(num_cards > 0);
 548                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 549                         {
 550                                 // Signal to the audio thread to process this frame.
 551                                 unique_lock<mutex> lock(audio_mutex);
 552                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
 553                                 audio_task_queue_changed.notify_one();
 554                         }
 555                         if (frame_num != card_copy[0].dropped_frames) {
 556                                 // For dropped frames, increase the pts. Note that if the format changed
 557                                 // in the meantime, we have no way of detecting that; we just have to
 558                                 // assume the frame length is always the same.
 559                                 ++stats_dropped_frames;
 560                                 pts_int += card_copy[0].new_frame_length;
 561                         }
 562                 }
 563
 564                 if (audio_level_callback != nullptr) {
 565                         unique_lock<mutex> lock(compressor_mutex);
 566                         double loudness_s = r128.loudness_S();
 567                         double loudness_i = r128.integrated();
 568                         double loudness_range_low = r128.range_min();
 569                         double loudness_range_high = r128.range_max();
 570
 571                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 572                                              loudness_i, loudness_range_low, loudness_range_high,
 573                                              gain_staging_db, 20.0 * log10(final_makeup_gain),
 574                                              correlation.get_correlation());
 575                 }
 576
 577                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 578                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 579                                 ++card_copy[card_index].dropped_frames;
 580                         }
 581                         if (card_copy[card_index].dropped_frames > 0) {
 582                                 printf("Card %u dropped %d frames before this\n",
 583                                         card_index, int(card_copy[card_index].dropped_frames));
 584                         }
 585                 }
 586
 587                 // If the first card is reporting a corrupted or otherwise dropped frame,
 588                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 589                 if (card_copy[0].new_frame->len == 0) {
 590                         ++stats_dropped_frames;
 591                         pts_int += card_copy[0].new_frame_length;
 592                         continue;
 593                 }
 594
 595                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 596                         CaptureCard *card = &card_copy[card_index];
 597                         if (!card->new_data_ready || card->new_frame->len == 0)
 598                                 continue;
 599
 600                         assert(card->new_frame != nullptr);
 601                         insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
 602                         check_error();
 603
 604                         // The new texture might still be uploaded,
 605                         // tell the GPU to wait until it's there.
 606                         if (card->new_data_ready_fence) {
 607                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 608                                 check_error();
 609                                 glDeleteSync(card->new_data_ready_fence);
 610                                 check_error();
 611                         }
 612                 }
 613
 614                 // Get the main chain from the theme, and set its state immediately.
 615                 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 616                 EffectChain *chain = theme_main_chain.chain;
 617                 theme_main_chain.setup_chain();
 618                 //theme_main_chain.chain->enable_phase_timing(true);
 619
 620                 GLuint y_tex, cbcr_tex;
 621                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 622                 assert(got_frame);
 623
 624                 // Render main chain.
 625                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 626                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 627                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 628                 check_error();
 629                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 630                 resource_pool->release_fbo(fbo);
 631
 632                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 633                 resource_pool->release_2d_texture(cbcr_full_tex);
 634
 635                 // Set the right state for rgba_tex.
 636                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 637                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 638                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 639                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 640                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 641
 642                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 643                 check_error();
 644
 645                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 646                 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
 647                 ++frame;
 648                 pts_int += card_copy[0].new_frame_length;
 649
 650                 // The live frame just shows the RGBA texture we just rendered.
 651                 // It owns rgba_tex now.
 652                 DisplayFrame live_frame;
 653                 live_frame.chain = display_chain.get();
 654                 live_frame.setup_chain = [this, rgba_tex]{
 655                         display_input->set_texture_num(rgba_tex);
 656                 };
 657                 live_frame.ready_fence = fence;
 658                 live_frame.input_frames = {};
 659                 live_frame.temp_textures = { rgba_tex };
 660                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 661
 662                 // Set up preview and any additional channels.
 663                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 664                         DisplayFrame display_frame;
 665                         Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 666                         display_frame.chain = chain.chain;
 667                         display_frame.setup_chain = chain.setup_chain;
 668                         display_frame.ready_fence = fence;
 669                         display_frame.input_frames = chain.input_frames;
 670                         display_frame.temp_textures = {};
 671                         output_channel[i].output_frame(display_frame);
 672                 }
 673
 674                 clock_gettime(CLOCK_MONOTONIC, &now);
 675                 double elapsed = now.tv_sec - start.tv_sec +
 676                         1e-9 * (now.tv_nsec - start.tv_nsec);
 677                 if (frame % 100 == 0) {
 678                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 679                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 680                                 1e3 * elapsed / frame);
 681                 //      chain->print_phase_timing();
 682                 }
 683
 684                 if (should_cut.exchange(false)) {  // Test and clear.
 685                         string filename = generate_local_dump_filename(frame);
 686                         printf("Starting new recording: %s\n", filename.c_str());
 687                         h264_encoder->shutdown();
 688                         httpd.close_output_file();
 689                         httpd.open_output_file(filename.c_str());
 690                         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 691                 }
 692
 693 #if 0
 694                 // Reset every 100 frames, so that local variations in frame times
 695                 // (especially for the first few frames, when the shaders are
 696                 // compiled etc.) don't make it hard to measure for the entire
 697                 // remaining duration of the program.
 698                 if (frame == 10000) {
 699                         frame = 0;
 700                         start = now;
 701                 }
 702 #endif
 703                 check_error();
 704         }
 705
 706         resource_pool->clean_context();
 707 }
 708
 709 void Mixer::audio_thread_func()
 710 {
 711         while (!should_quit) {
 712                 AudioTask task;
 713
 714                 {
 715                         unique_lock<mutex> lock(audio_mutex);
 716                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 717                         task = audio_task_queue.front();
 718                         audio_task_queue.pop();
 719                 }
 720
 721                 process_audio_one_frame(task.pts_int, task.num_samples);
 722         }
 723 }
 724
 725 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 726 {
 727         vector<float> samples_card;
 728         vector<float> samples_out;
 729
 730         // TODO: Allow mixing audio from several sources.
 731         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 732         assert(selected_audio_card < num_cards);
 733
 734         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 735                 samples_card.resize(num_samples * 2);
 736                 {
 737                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 738                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 739                                 printf("Card %d reported previous underrun.\n", card_index);
 740                         }
 741                 }
 742                 if (card_index == selected_audio_card) {
 743                         samples_out = move(samples_card);
 744                 }
 745         }
 746
 747         // Cut away everything under 120 Hz (or whatever the cutoff is);
 748         // we don't need it for voice, and it will reduce headroom
 749         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 750         // should be dampened.)
 751         if (locut_enabled) {
 752                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 753         }
 754
 755         // Apply a level compressor to get the general level right.
 756         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 757         // (or more precisely, near it, since we don't use infinite ratio),
 758         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 759         // entirely arbitrary, but from practical tests with speech, it seems to
 760         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 761         {
 762                 unique_lock<mutex> lock(compressor_mutex);
 763                 if (level_compressor_enabled) {
 764                         float threshold = 0.01f;   // -40 dBFS.
 765                         float ratio = 20.0f;
 766                         float attack_time = 0.5f;
 767                         float release_time = 20.0f;
 768                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 769                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 770                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 771                 } else {
 772                         // Just apply the gain we already had.
 773                         float g = pow(10.0f, gain_staging_db / 20.0f);
 774                         for (size_t i = 0; i < samples_out.size(); ++i) {
 775                                 samples_out[i] *= g;
 776                         }
 777                 }
 778         }
 779
 780 #if 0
 781         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 782                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 783                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 784                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 785 #endif
 786
 787 //      float limiter_att, compressor_att;
 788
 789         // The real compressor.
 790         if (compressor_enabled) {
 791                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 792                 float ratio = 20.0f;
 793                 float attack_time = 0.005f;
 794                 float release_time = 0.040f;
 795                 float makeup_gain = 2.0f;  // +6 dB.
 796                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 797 //              compressor_att = compressor.get_attenuation();
 798         }
 799
 800         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 801         // Note that since ratio is not infinite, we could go slightly higher than this.
 802         if (limiter_enabled) {
 803                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 804                 float ratio = 30.0f;
 805                 float attack_time = 0.0f;  // Instant.
 806                 float release_time = 0.020f;
 807                 float makeup_gain = 1.0f;  // 0 dB.
 808                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 809 //              limiter_att = limiter.get_attenuation();
 810         }
 811
 812 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 813
 814         // Upsample 4x to find interpolated peak.
 815         peak_resampler.inp_data = samples_out.data();
 816         peak_resampler.inp_count = samples_out.size() / 2;
 817
 818         vector<float> interpolated_samples_out;
 819         interpolated_samples_out.resize(samples_out.size());
 820         while (peak_resampler.inp_count > 0) {  // About four iterations.
 821                 peak_resampler.out_data = &interpolated_samples_out[0];
 822                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 823                 peak_resampler.process();
 824                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 825                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 826                 peak_resampler.out_data = nullptr;
 827         }
 828
 829         // At this point, we are most likely close to +0 LU, but all of our
 830         // measurements have been on raw sample values, not R128 values.
 831         // So we have a final makeup gain to get us to +0 LU; the gain
 832         // adjustments required should be relatively small, and also, the
 833         // offset shouldn't change much (only if the type of audio changes
 834         // significantly). Thus, we shoot for updating this value basically
 835         // “whenever we process buffers”, since the R128 calculation isn't exactly
 836         // something we get out per-sample.
 837         //
 838         // Note that there's a feedback loop here, so we choose a very slow filter
 839         // (half-time of 100 seconds).
 840         double target_loudness_factor, alpha;
 841         {
 842                 unique_lock<mutex> lock(compressor_mutex);
 843                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
 844                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
 845                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
 846
 847                 // If we're outside +/- 5 LU uncorrected, we don't count it as
 848                 // a normal signal (probably silence) and don't change the
 849                 // correction factor; just apply what we already have.
 850                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 851                         alpha = 0.0;
 852                 } else {
 853                         // Formula adapted from
 854                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 855                         const double half_time_s = 100.0;
 856                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 857                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 858                 }
 859
 860                 double m = final_makeup_gain;
 861                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 862                         samples_out[i + 0] *= m;
 863                         samples_out[i + 1] *= m;
 864                         m += (target_loudness_factor - m) * alpha;
 865                 }
 866                 final_makeup_gain = m;
 867         }
 868
 869         // Find R128 levels and L/R correlation.
 870         vector<float> left, right;
 871         deinterleave_samples(samples_out, &left, &right);
 872         float *ptrs[] = { left.data(), right.data() };
 873         {
 874                 unique_lock<mutex> lock(compressor_mutex);
 875                 r128.process(left.size(), ptrs);
 876                 correlation.process_samples(samples_out);
 877         }
 878
 879         // Send the samples to the sound card.
 880         if (alsa) {
 881                 alsa->write(samples_out);
 882         }
 883
 884         // And finally add them to the output.
 885         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 886 }
 887
 888 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 889 {
 890         GLuint vao;
 891         glGenVertexArrays(1, &vao);
 892         check_error();
 893
 894         glBindVertexArray(vao);
 895         check_error();
 896
 897         // Extract Cb/Cr.
 898         GLuint fbo = resource_pool->create_fbo(dst_tex);
 899         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 900         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 901         check_error();
 902
 903         glUseProgram(cbcr_program_num);
 904         check_error();
 905
 906         glActiveTexture(GL_TEXTURE0);
 907         check_error();
 908         glBindTexture(GL_TEXTURE_2D, src_tex);
 909         check_error();
 910         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 911         check_error();
 912         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 913         check_error();
 914         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 915         check_error();
 916
 917         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 918         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 919
 920         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
 921         check_error();
 922
 923         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
 924                 glEnableVertexAttribArray(attr_index);
 925                 check_error();
 926                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 927                 check_error();
 928         }
 929
 930         glDrawArrays(GL_TRIANGLES, 0, 3);
 931         check_error();
 932
 933         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
 934                 glDisableVertexAttribArray(attr_index);
 935                 check_error();
 936         }
 937
 938         glUseProgram(0);
 939         check_error();
 940         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 941         check_error();
 942
 943         resource_pool->release_fbo(fbo);
 944         glDeleteVertexArrays(1, &vao);
 945 }
 946
 947 void Mixer::release_display_frame(DisplayFrame *frame)
 948 {
 949         for (GLuint texnum : frame->temp_textures) {
 950                 resource_pool->release_2d_texture(texnum);
 951         }
 952         frame->temp_textures.clear();
 953         frame->ready_fence.reset();
 954         frame->input_frames.clear();
 955 }
 956
 957 void Mixer::start()
 958 {
 959         mixer_thread = thread(&Mixer::thread_func, this);
 960         audio_thread = thread(&Mixer::audio_thread_func, this);
 961 }
 962
 963 void Mixer::quit()
 964 {
 965         should_quit = true;
 966         mixer_thread.join();
 967         audio_thread.join();
 968 }
 969
 970 void Mixer::transition_clicked(int transition_num)
 971 {
 972         theme->transition_clicked(transition_num, pts());
 973 }
 974
 975 void Mixer::channel_clicked(int preview_num)
 976 {
 977         theme->channel_clicked(preview_num);
 978 }
 979
 980 void Mixer::reset_meters()
 981 {
 982         peak_resampler.reset();
 983         peak = 0.0f;
 984         r128.reset();
 985         r128.integr_start();
 986         correlation.reset();
 987 }
 988
 989 Mixer::OutputChannel::~OutputChannel()
 990 {
 991         if (has_current_frame) {
 992                 parent->release_display_frame(&current_frame);
 993         }
 994         if (has_ready_frame) {
 995                 parent->release_display_frame(&ready_frame);
 996         }
 997 }
 998
 999 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1000 {
1001         // Store this frame for display. Remove the ready frame if any
1002         // (it was seemingly never used).
1003         {
1004                 unique_lock<mutex> lock(frame_mutex);
1005                 if (has_ready_frame) {
1006                         parent->release_display_frame(&ready_frame);
1007                 }
1008                 ready_frame = frame;
1009                 has_ready_frame = true;
1010         }
1011
1012         if (has_new_frame_ready_callback) {
1013                 new_frame_ready_callback();
1014         }
1015 }
1016
1017 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1018 {
1019         unique_lock<mutex> lock(frame_mutex);
1020         if (!has_current_frame && !has_ready_frame) {
1021                 return false;
1022         }
1023
1024         if (has_current_frame && has_ready_frame) {
1025                 // We have a new ready frame. Toss the current one.
1026                 parent->release_display_frame(&current_frame);
1027                 has_current_frame = false;
1028         }
1029         if (has_ready_frame) {
1030                 assert(!has_current_frame);
1031                 current_frame = ready_frame;
1032                 ready_frame.ready_fence.reset();  // Drop the refcount.
1033                 ready_frame.input_frames.clear();  // Drop the refcounts.
1034                 has_current_frame = true;
1035                 has_ready_frame = false;
1036         }
1037
1038         *frame = current_frame;
1039         return true;
1040 }
1041
1042 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1043 {
1044         new_frame_ready_callback = callback;
1045         has_new_frame_ready_callback = true;
1046 }