git.sesse.net Git - nageru/blob - mixer.cpp

   1 #undef Success
   2
   3 #include "mixer.h"
   4
   5 #include <assert.h>
   6 #include <epoxy/egl.h>
   7 #include <init.h>
   8 #include <movit/effect_chain.h>
   9 #include <movit/effect_util.h>
  10 #include <movit/flat_input.h>
  11 #include <movit/image_format.h>
  12 #include <movit/resource_pool.h>
  13 #include <stdint.h>
  14 #include <stdio.h>
  15 #include <stdlib.h>
  16 #include <sys/time.h>
  17 #include <time.h>
  18 #include <util.h>
  19 #include <algorithm>
  20 #include <cmath>
  21 #include <condition_variable>
  22 #include <cstddef>
  23 #include <memory>
  24 #include <mutex>
  25 #include <string>
  26 #include <thread>
  27 #include <utility>
  28 #include <vector>
  29
  30 #include "bmusb/bmusb.h"
  31 #include "context.h"
  32 #include "defs.h"
  33 #include "h264encode.h"
  34 #include "pbo_frame_allocator.h"
  35 #include "ref_counted_gl_sync.h"
  36 #include "timebase.h"
  37
  38 class QOpenGLContext;
  39
  40 using namespace movit;
  41 using namespace std;
  42 using namespace std::placeholders;
  43
  44 Mixer *global_mixer = nullptr;
  45
  46 namespace {
  47
  48 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  49 {
  50         for (size_t i = 0; i < num_samples; ++i) {
  51                 for (size_t j = 0; j < out_channels; ++j) {
  52                         uint32_t s1 = *src++;
  53                         uint32_t s2 = *src++;
  54                         uint32_t s3 = *src++;
  55                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  56                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  57                 }
  58                 src += 3 * (in_channels - out_channels);
  59         }
  60 }
  61
  62 }  // namespace
  63
  64 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
  65         : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT),
  66           num_cards(num_cards),
  67           mixer_surface(create_surface(format)),
  68           h264_encoder_surface(create_surface(format)),
  69           level_compressor(OUTPUT_FREQUENCY),
  70           limiter(OUTPUT_FREQUENCY),
  71           compressor(OUTPUT_FREQUENCY)
  72 {
  73         httpd.start(9095);
  74
  75         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
  76         check_error();
  77
  78         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
  79         // will be halved when sampling them, and we need to compensate here.
  80         movit_texel_subpixel_precision /= 2.0;
  81
  82         resource_pool.reset(new ResourcePool);
  83         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
  84         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
  85                 output_channel[i].parent = this;
  86         }
  87
  88         ImageFormat inout_format;
  89         inout_format.color_space = COLORSPACE_sRGB;
  90         inout_format.gamma_curve = GAMMA_sRGB;
  91
  92         // Display chain; shows the live output produced by the main chain (its RGBA version).
  93         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
  94         check_error();
  95         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
  96         display_chain->add_input(display_input);
  97         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
  98         display_chain->set_dither_bits(0);  // Don't bother.
  99         display_chain->finalize();
 100
 101         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
 102
 103         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 104                 printf("Configuring card %d...\n", card_index);
 105                 CaptureCard *card = &cards[card_index];
 106                 card->usb = new BMUSBCapture(card_index);
 107                 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 108                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 109                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
 110                 card->surface = create_surface(format);
 111                 card->usb->set_dequeue_thread_callbacks(
 112                         [card]{
 113                                 eglBindAPI(EGL_OPENGL_API);
 114                                 card->context = create_context(card->surface);
 115                                 if (!make_current(card->context, card->surface)) {
 116                                         printf("failed to create bmusb context\n");
 117                                         exit(1);
 118                                 }
 119                         },
 120                         [this]{
 121                                 resource_pool->clean_context();
 122                         });
 123                 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 124                 card->usb->configure_card();
 125         }
 126
 127         BMUSBCapture::start_bm_thread();
 128
 129         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 130                 cards[card_index].usb->start_bm_capture();
 131         }
 132
 133         //chain->enable_phase_timing(true);
 134
 135         // Set up stuff for NV12 conversion.
 136
 137         // Cb/Cr shader.
 138         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 139         string cbcr_frag_shader =
 140                 "#version 130 \n"
 141                 "in vec2 tc0; \n"
 142                 "uniform sampler2D cbcr_tex; \n"
 143                 "void main() { \n"
 144                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 145                 "} \n";
 146         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
 147
 148         r128.init(2, OUTPUT_FREQUENCY);
 149         r128.integr_start();
 150
 151         locut.init(FILTER_HPF, 2);
 152
 153         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 154         // and there's a limit to how important the peak meter is.
 155         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
 156
 157         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 158 }
 159
 160 Mixer::~Mixer()
 161 {
 162         resource_pool->release_glsl_program(cbcr_program_num);
 163         BMUSBCapture::stop_bm_thread();
 164
 165         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 166                 {
 167                         unique_lock<mutex> lock(bmusb_mutex);
 168                         cards[card_index].should_quit = true;  // Unblock thread.
 169                         cards[card_index].new_data_ready_changed.notify_all();
 170                 }
 171                 cards[card_index].usb->stop_dequeue_thread();
 172         }
 173
 174         h264_encoder.reset(nullptr);
 175 }
 176
 177 namespace {
 178
 179 int unwrap_timecode(uint16_t current_wrapped, int last)
 180 {
 181         uint16_t last_wrapped = last & 0xffff;
 182         if (current_wrapped > last_wrapped) {
 183                 return (last & ~0xffff) | current_wrapped;
 184         } else {
 185                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 186         }
 187 }
 188
 189 float find_peak(const float *samples, size_t num_samples)
 190 {
 191         float m = fabs(samples[0]);
 192         for (size_t i = 1; i < num_samples; ++i) {
 193                 m = std::max(m, fabs(samples[i]));
 194         }
 195         return m;
 196 }
 197
 198 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 199 {
 200         size_t num_samples = in.size() / 2;
 201         out_l->resize(num_samples);
 202         out_r->resize(num_samples);
 203
 204         const float *inptr = in.data();
 205         float *lptr = &(*out_l)[0];
 206         float *rptr = &(*out_r)[0];
 207         for (size_t i = 0; i < num_samples; ++i) {
 208                 *lptr++ = *inptr++;
 209                 *rptr++ = *inptr++;
 210         }
 211 }
 212
 213 }  // namespace
 214
 215 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 216                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 217                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 218 {
 219         CaptureCard *card = &cards[card_index];
 220
 221         int width, height, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
 222         bool interlaced;
 223
 224         decode_video_format(video_format, &width, &height, &extra_lines_top, &extra_lines_bottom,
 225                             &frame_rate_nom, &frame_rate_den, &interlaced);  // Ignore return value for now.
 226         int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom;
 227
 228         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 229         if (num_samples > OUTPUT_FREQUENCY / 10) {
 230                 printf("Card %d: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n",
 231                         card_index, int(audio_frame.len), int(audio_offset),
 232                         timecode, int(video_frame.len), int(video_offset), video_format);
 233                 if (video_frame.owner) {
 234                         video_frame.owner->release_frame(video_frame);
 235                 }
 236                 if (audio_frame.owner) {
 237                         audio_frame.owner->release_frame(audio_frame);
 238                 }
 239                 return;
 240         }
 241
 242         int64_t local_pts = card->next_local_pts;
 243         int dropped_frames = 0;
 244         if (card->last_timecode != -1) {
 245                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 246         }
 247
 248         // Convert the audio to stereo fp32 and add it.
 249         vector<float> audio;
 250         audio.resize(num_samples * 2);
 251         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 252
 253         // Add the audio.
 254         {
 255                 unique_lock<mutex> lock(card->audio_mutex);
 256
 257                 // Number of samples per frame if we need to insert silence.
 258                 // (Could be nonintegral, but resampling will save us then.)
 259                 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
 260
 261                 if (dropped_frames > MAX_FPS * 2) {
 262                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 263                                 card_index, card->last_timecode, timecode);
 264                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 265                         dropped_frames = 0;
 266                 } else if (dropped_frames > 0) {
 267                         // Insert silence as needed.
 268                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 269                                 card_index, dropped_frames, timecode);
 270                         vector<float> silence;
 271                         silence.resize(silence_samples * 2);
 272                         for (int i = 0; i < dropped_frames; ++i) {
 273                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 274                                 // Note that if the format changed in the meantime, we have
 275                                 // no way of detecting that; we just have to assume the frame length
 276                                 // is always the same.
 277                                 local_pts += frame_length;
 278                         }
 279                 }
 280                 if (num_samples == 0) {
 281                         audio.resize(silence_samples * 2);
 282                         num_samples = silence_samples;
 283                 }
 284                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 285                 card->next_local_pts = local_pts + frame_length;
 286         }
 287
 288         card->last_timecode = timecode;
 289
 290         // Done with the audio, so release it.
 291         if (audio_frame.owner) {
 292                 audio_frame.owner->release_frame(audio_frame);
 293         }
 294
 295         {
 296                 // Wait until the previous frame was consumed.
 297                 unique_lock<mutex> lock(bmusb_mutex);
 298                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 299                 if (card->should_quit) return;
 300         }
 301
 302         if (video_frame.len - video_offset == 0 ||
 303             video_frame.len - video_offset != size_t(width * (height + extra_lines_top + extra_lines_bottom) * 2) ||
 304             width != WIDTH || height != HEIGHT) {  // TODO: Remove this once the rest of the infrastructure is in place.
 305                 if (video_frame.len != 0) {
 306                         printf("Card %d: Dropping video frame with wrong length (%ld)\n",
 307                                 card_index, video_frame.len - video_offset);
 308                 }
 309                 if (video_frame.owner) {
 310                         video_frame.owner->release_frame(video_frame);
 311                 }
 312
 313                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 314                 // so that pts can go up accordingly.
 315                 {
 316                         unique_lock<mutex> lock(bmusb_mutex);
 317                         card->new_data_ready = true;
 318                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 319                         card->new_frame_length = frame_length;
 320                         card->new_data_ready_fence = nullptr;
 321                         card->dropped_frames = dropped_frames;
 322                         card->new_data_ready_changed.notify_all();
 323                 }
 324                 return;
 325         }
 326
 327         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
 328         GLuint pbo = userdata->pbo;
 329         check_error();
 330         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 331         check_error();
 332         glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
 333         check_error();
 334         //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 335         //check_error();
 336
 337         // Upload the textures.
 338         size_t cbcr_width = width / 2;
 339         size_t cbcr_offset = video_offset / 2;
 340         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 341
 342         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
 343         check_error();
 344         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * extra_lines_top * sizeof(uint16_t)));
 345         check_error();
 346         glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
 347         check_error();
 348         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * extra_lines_top));
 349         check_error();
 350         glBindTexture(GL_TEXTURE_2D, 0);
 351         check_error();
 352         GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 353         check_error();
 354         assert(fence != nullptr);
 355
 356         {
 357                 unique_lock<mutex> lock(bmusb_mutex);
 358                 card->new_data_ready = true;
 359                 card->new_frame = RefCountedFrame(video_frame);
 360                 card->new_frame_length = frame_length;
 361                 card->new_data_ready_fence = fence;
 362                 card->dropped_frames = dropped_frames;
 363                 card->new_data_ready_changed.notify_all();
 364         }
 365 }
 366
 367 void Mixer::thread_func()
 368 {
 369         eglBindAPI(EGL_OPENGL_API);
 370         QOpenGLContext *context = create_context(mixer_surface);
 371         if (!make_current(context, mixer_surface)) {
 372                 printf("oops\n");
 373                 exit(1);
 374         }
 375
 376         struct timespec start, now;
 377         clock_gettime(CLOCK_MONOTONIC, &start);
 378
 379         int frame = 0;
 380         int stats_dropped_frames = 0;
 381
 382         while (!should_quit) {
 383                 CaptureCard card_copy[MAX_CARDS];
 384                 int num_samples[MAX_CARDS];
 385
 386                 {
 387                         unique_lock<mutex> lock(bmusb_mutex);
 388
 389                         // The first card is the master timer, so wait for it to have a new frame.
 390                         // TODO: Make configurable, and with a timeout.
 391                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 392
 393                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 394                                 CaptureCard *card = &cards[card_index];
 395                                 card_copy[card_index].usb = card->usb;
 396                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 397                                 card_copy[card_index].new_frame = card->new_frame;
 398                                 card_copy[card_index].new_frame_length = card->new_frame_length;
 399                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 400                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 401                                 card->new_data_ready = false;
 402                                 card->new_data_ready_changed.notify_all();
 403
 404                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
 405                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 406                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 407                                 assert(num_samples[card_index] >= 0);
 408                         }
 409                 }
 410
 411                 // Resample the audio as needed, including from previously dropped frames.
 412                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 413                         {
 414                                 // Signal to the audio thread to process this frame.
 415                                 unique_lock<mutex> lock(audio_mutex);
 416                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
 417                                 audio_task_queue_changed.notify_one();
 418                         }
 419                         if (frame_num != card_copy[0].dropped_frames) {
 420                                 // For dropped frames, increase the pts. Note that if the format changed
 421                                 // in the meantime, we have no way of detecting that; we just have to
 422                                 // assume the frame length is always the same.
 423                                 ++stats_dropped_frames;
 424                                 pts_int += card_copy[0].new_frame_length;
 425                         }
 426                 }
 427
 428                 if (audio_level_callback != nullptr) {
 429                         double loudness_s = r128.loudness_S();
 430                         double loudness_i = r128.integrated();
 431                         double loudness_range_low = r128.range_min();
 432                         double loudness_range_high = r128.range_max();
 433
 434                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 435                                              loudness_i, loudness_range_low, loudness_range_high,
 436                                              last_gain_staging_db);
 437                 }
 438
 439                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 440                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 441                                 ++card_copy[card_index].dropped_frames;
 442                         }
 443                         if (card_copy[card_index].dropped_frames > 0) {
 444                                 printf("Card %u dropped %d frames before this\n",
 445                                         card_index, int(card_copy[card_index].dropped_frames));
 446                         }
 447                 }
 448
 449                 // If the first card is reporting a corrupted or otherwise dropped frame,
 450                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 451                 if (card_copy[0].new_frame->len == 0) {
 452                         ++stats_dropped_frames;
 453                         pts_int += card_copy[0].new_frame_length;
 454                         continue;
 455                 }
 456
 457                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 458                         CaptureCard *card = &card_copy[card_index];
 459                         if (!card->new_data_ready || card->new_frame->len == 0)
 460                                 continue;
 461
 462                         assert(card->new_frame != nullptr);
 463                         bmusb_current_rendering_frame[card_index] = card->new_frame;
 464                         check_error();
 465
 466                         // The new texture might still be uploaded,
 467                         // tell the GPU to wait until it's there.
 468                         if (card->new_data_ready_fence) {
 469                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 470                                 check_error();
 471                                 glDeleteSync(card->new_data_ready_fence);
 472                                 check_error();
 473                         }
 474                         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
 475                         theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
 476                 }
 477
 478                 // Get the main chain from the theme, and set its state immediately.
 479                 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
 480                 EffectChain *chain = theme_main_chain.first;
 481                 theme_main_chain.second();
 482
 483                 GLuint y_tex, cbcr_tex;
 484                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 485                 assert(got_frame);
 486
 487                 // Render main chain.
 488                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 489                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 490                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 491                 check_error();
 492                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 493                 resource_pool->release_fbo(fbo);
 494
 495                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 496                 resource_pool->release_2d_texture(cbcr_full_tex);
 497
 498                 // Set the right state for rgba_tex.
 499                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 500                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 501                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 502                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 503                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 504
 505                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 506                 check_error();
 507
 508                 // Make sure the H.264 gets a reference to all the
 509                 // input frames needed, so that they are not released back
 510                 // until the rendering is done.
 511                 vector<RefCountedFrame> input_frames;
 512                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 513                         input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 514                 }
 515                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 516                 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
 517                 ++frame;
 518                 pts_int += card_copy[0].new_frame_length;
 519
 520                 // The live frame just shows the RGBA texture we just rendered.
 521                 // It owns rgba_tex now.
 522                 DisplayFrame live_frame;
 523                 live_frame.chain = display_chain.get();
 524                 live_frame.setup_chain = [this, rgba_tex]{
 525                         display_input->set_texture_num(rgba_tex);
 526                 };
 527                 live_frame.ready_fence = fence;
 528                 live_frame.input_frames = {};
 529                 live_frame.temp_textures = { rgba_tex };
 530                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 531
 532                 // Set up preview and any additional channels.
 533                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 534                         DisplayFrame display_frame;
 535                         pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT);  // FIXME: dimensions
 536                         display_frame.chain = chain.first;
 537                         display_frame.setup_chain = chain.second;
 538                         display_frame.ready_fence = fence;
 539
 540                         // FIXME: possible to do better?
 541                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 542                                 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 543                         }
 544                         display_frame.temp_textures = {};
 545                         output_channel[i].output_frame(display_frame);
 546                 }
 547
 548                 clock_gettime(CLOCK_MONOTONIC, &now);
 549                 double elapsed = now.tv_sec - start.tv_sec +
 550                         1e-9 * (now.tv_nsec - start.tv_nsec);
 551                 if (frame % 100 == 0) {
 552                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 553                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 554                                 1e3 * elapsed / frame);
 555                 //      chain->print_phase_timing();
 556                 }
 557
 558 #if 0
 559                 // Reset every 100 frames, so that local variations in frame times
 560                 // (especially for the first few frames, when the shaders are
 561                 // compiled etc.) don't make it hard to measure for the entire
 562                 // remaining duration of the program.
 563                 if (frame == 10000) {
 564                         frame = 0;
 565                         start = now;
 566                 }
 567 #endif
 568                 check_error();
 569         }
 570
 571         resource_pool->clean_context();
 572 }
 573
 574 void Mixer::audio_thread_func()
 575 {
 576         while (!should_quit) {
 577                 AudioTask task;
 578
 579                 {
 580                         unique_lock<mutex> lock(audio_mutex);
 581                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 582                         task = audio_task_queue.front();
 583                         audio_task_queue.pop();
 584                 }
 585
 586                 process_audio_one_frame(task.pts_int, task.num_samples);
 587         }
 588 }
 589
 590 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 591 {
 592         vector<float> samples_card;
 593         vector<float> samples_out;
 594         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 595                 samples_card.resize(num_samples * 2);
 596                 {
 597                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 598                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 599                                 printf("Card %d reported previous underrun.\n", card_index);
 600                         }
 601                 }
 602                 // TODO: Allow using audio from the other card(s) as well.
 603                 if (card_index == 0) {
 604                         samples_out = move(samples_card);
 605                 }
 606         }
 607
 608         // Cut away everything under 120 Hz (or whatever the cutoff is);
 609         // we don't need it for voice, and it will reduce headroom
 610         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 611         // should be dampened.)
 612         locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 613
 614         // Apply a level compressor to get the general level right.
 615         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 616         // (or more precisely, near it, since we don't use infinite ratio),
 617         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 618         // entirely arbitrary, but from practical tests with speech, it seems to
 619         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 620         float ref_level_dbfs = -14.0f;
 621         {
 622                 float threshold = 0.01f;   // -40 dBFS.
 623                 float ratio = 20.0f;
 624                 float attack_time = 0.5f;
 625                 float release_time = 20.0f;
 626                 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 627                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 628                 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 629         }
 630
 631 #if 0
 632         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 633                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 634                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 635                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 636 #endif
 637
 638 //      float limiter_att, compressor_att;
 639
 640         // The real compressor.
 641         if (compressor_enabled) {
 642                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 643                 float ratio = 20.0f;
 644                 float attack_time = 0.005f;
 645                 float release_time = 0.040f;
 646                 float makeup_gain = 2.0f;  // +6 dB.
 647                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 648 //              compressor_att = compressor.get_attenuation();
 649         }
 650
 651         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 652         // Note that since ratio is not infinite, we could go slightly higher than this.
 653         if (limiter_enabled) {
 654                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 655                 float ratio = 30.0f;
 656                 float attack_time = 0.0f;  // Instant.
 657                 float release_time = 0.020f;
 658                 float makeup_gain = 1.0f;  // 0 dB.
 659                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 660 //              limiter_att = limiter.get_attenuation();
 661         }
 662
 663 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 664
 665         // Upsample 4x to find interpolated peak.
 666         peak_resampler.inp_data = samples_out.data();
 667         peak_resampler.inp_count = samples_out.size() / 2;
 668
 669         vector<float> interpolated_samples_out;
 670         interpolated_samples_out.resize(samples_out.size());
 671         while (peak_resampler.inp_count > 0) {  // About four iterations.
 672                 peak_resampler.out_data = &interpolated_samples_out[0];
 673                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 674                 peak_resampler.process();
 675                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 676                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 677         }
 678
 679         // Find R128 levels.
 680         vector<float> left, right;
 681         deinterleave_samples(samples_out, &left, &right);
 682         float *ptrs[] = { left.data(), right.data() };
 683         r128.process(left.size(), ptrs);
 684
 685         // Send the samples to the sound card.
 686         if (alsa) {
 687                 alsa->write(samples_out);
 688         }
 689
 690         // And finally add them to the output.
 691         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 692 }
 693
 694 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 695 {
 696         GLuint vao;
 697         glGenVertexArrays(1, &vao);
 698         check_error();
 699
 700         float vertices[] = {
 701                 0.0f, 2.0f,
 702                 0.0f, 0.0f,
 703                 2.0f, 0.0f
 704         };
 705
 706         glBindVertexArray(vao);
 707         check_error();
 708
 709         // Extract Cb/Cr.
 710         GLuint fbo = resource_pool->create_fbo(dst_tex);
 711         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 712         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 713         check_error();
 714
 715         glUseProgram(cbcr_program_num);
 716         check_error();
 717
 718         glActiveTexture(GL_TEXTURE0);
 719         check_error();
 720         glBindTexture(GL_TEXTURE_2D, src_tex);
 721         check_error();
 722         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 723         check_error();
 724         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 725         check_error();
 726         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 727         check_error();
 728
 729         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 730         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 731
 732         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 733         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 734
 735         glDrawArrays(GL_TRIANGLES, 0, 3);
 736         check_error();
 737
 738         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 739         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 740
 741         glUseProgram(0);
 742         check_error();
 743
 744         resource_pool->release_fbo(fbo);
 745         glDeleteVertexArrays(1, &vao);
 746 }
 747
 748 void Mixer::release_display_frame(DisplayFrame *frame)
 749 {
 750         for (GLuint texnum : frame->temp_textures) {
 751                 resource_pool->release_2d_texture(texnum);
 752         }
 753         frame->temp_textures.clear();
 754         frame->ready_fence.reset();
 755         frame->input_frames.clear();
 756 }
 757
 758 void Mixer::start()
 759 {
 760         mixer_thread = thread(&Mixer::thread_func, this);
 761         audio_thread = thread(&Mixer::audio_thread_func, this);
 762 }
 763
 764 void Mixer::quit()
 765 {
 766         should_quit = true;
 767         mixer_thread.join();
 768         audio_thread.join();
 769 }
 770
 771 void Mixer::transition_clicked(int transition_num)
 772 {
 773         theme->transition_clicked(transition_num, pts());
 774 }
 775
 776 void Mixer::channel_clicked(int preview_num)
 777 {
 778         theme->channel_clicked(preview_num);
 779 }
 780
 781 void Mixer::reset_meters()
 782 {
 783         peak_resampler.reset();
 784         peak = 0.0f;
 785         r128.reset();
 786         r128.integr_start();
 787 }
 788
 789 Mixer::OutputChannel::~OutputChannel()
 790 {
 791         if (has_current_frame) {
 792                 parent->release_display_frame(&current_frame);
 793         }
 794         if (has_ready_frame) {
 795                 parent->release_display_frame(&ready_frame);
 796         }
 797 }
 798
 799 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 800 {
 801         // Store this frame for display. Remove the ready frame if any
 802         // (it was seemingly never used).
 803         {
 804                 unique_lock<mutex> lock(frame_mutex);
 805                 if (has_ready_frame) {
 806                         parent->release_display_frame(&ready_frame);
 807                 }
 808                 ready_frame = frame;
 809                 has_ready_frame = true;
 810         }
 811
 812         if (has_new_frame_ready_callback) {
 813                 new_frame_ready_callback();
 814         }
 815 }
 816
 817 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 818 {
 819         unique_lock<mutex> lock(frame_mutex);
 820         if (!has_current_frame && !has_ready_frame) {
 821                 return false;
 822         }
 823
 824         if (has_current_frame && has_ready_frame) {
 825                 // We have a new ready frame. Toss the current one.
 826                 parent->release_display_frame(&current_frame);
 827                 has_current_frame = false;
 828         }
 829         if (has_ready_frame) {
 830                 assert(!has_current_frame);
 831                 current_frame = ready_frame;
 832                 ready_frame.ready_fence.reset();  // Drop the refcount.
 833                 ready_frame.input_frames.clear();  // Drop the refcounts.
 834                 has_current_frame = true;
 835                 has_ready_frame = false;
 836         }
 837
 838         *frame = current_frame;
 839         return true;
 840 }
 841
 842 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 843 {
 844         new_frame_ready_callback = callback;
 845         has_new_frame_ready_callback = true;
 846 }