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

   1 #define EXTRAHEIGHT 30
   2
   3 #undef Success
   4
   5 #include "mixer.h"
   6
   7 #include <assert.h>
   8 #include <epoxy/egl.h>
   9 #include <init.h>
  10 #include <movit/effect_chain.h>
  11 #include <movit/effect_util.h>
  12 #include <movit/flat_input.h>
  13 #include <movit/image_format.h>
  14 #include <movit/resource_pool.h>
  15 #include <stdint.h>
  16 #include <stdio.h>
  17 #include <stdlib.h>
  18 #include <sys/time.h>
  19 #include <time.h>
  20 #include <util.h>
  21 #include <algorithm>
  22 #include <cmath>
  23 #include <condition_variable>
  24 #include <cstddef>
  25 #include <memory>
  26 #include <mutex>
  27 #include <string>
  28 #include <thread>
  29 #include <utility>
  30 #include <vector>
  31
  32 #include "bmusb/bmusb.h"
  33 #include "context.h"
  34 #include "defs.h"
  35 #include "h264encode.h"
  36 #include "pbo_frame_allocator.h"
  37 #include "ref_counted_gl_sync.h"
  38 #include "timebase.h"
  39
  40 class QOpenGLContext;
  41
  42 using namespace movit;
  43 using namespace std;
  44 using namespace std::placeholders;
  45
  46 Mixer *global_mixer = nullptr;
  47
  48 namespace {
  49
  50 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  51 {
  52         for (size_t i = 0; i < num_samples; ++i) {
  53                 for (size_t j = 0; j < out_channels; ++j) {
  54                         uint32_t s1 = *src++;
  55                         uint32_t s2 = *src++;
  56                         uint32_t s3 = *src++;
  57                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  58                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  59                 }
  60                 src += 3 * (in_channels - out_channels);
  61         }
  62 }
  63
  64 }  // namespace
  65
  66 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
  67         : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT),
  68           num_cards(num_cards),
  69           mixer_surface(create_surface(format)),
  70           h264_encoder_surface(create_surface(format)),
  71           level_compressor(OUTPUT_FREQUENCY),
  72           limiter(OUTPUT_FREQUENCY),
  73           compressor(OUTPUT_FREQUENCY)
  74 {
  75         httpd.start(9095);
  76
  77         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
  78         check_error();
  79
  80         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
  81         // will be halved when sampling them, and we need to compensate here.
  82         movit_texel_subpixel_precision /= 2.0;
  83
  84         resource_pool.reset(new ResourcePool);
  85         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
  86         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
  87                 output_channel[i].parent = this;
  88         }
  89
  90         ImageFormat inout_format;
  91         inout_format.color_space = COLORSPACE_sRGB;
  92         inout_format.gamma_curve = GAMMA_sRGB;
  93
  94         // Display chain; shows the live output produced by the main chain (its RGBA version).
  95         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
  96         check_error();
  97         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
  98         display_chain->add_input(display_input);
  99         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 100         display_chain->set_dither_bits(0);  // Don't bother.
 101         display_chain->finalize();
 102
 103         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
 104
 105         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 106                 printf("Configuring card %d...\n", card_index);
 107                 CaptureCard *card = &cards[card_index];
 108                 card->usb = new BMUSBCapture(card_index);
 109                 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 110                 card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44 + 1, WIDTH, HEIGHT));
 111                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
 112                 card->surface = create_surface(format);
 113                 card->usb->set_dequeue_thread_callbacks(
 114                         [card]{
 115                                 eglBindAPI(EGL_OPENGL_API);
 116                                 card->context = create_context(card->surface);
 117                                 if (!make_current(card->context, card->surface)) {
 118                                         printf("failed to create bmusb context\n");
 119                                         exit(1);
 120                                 }
 121                         },
 122                         [this]{
 123                                 resource_pool->clean_context();
 124                         });
 125                 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 126                 card->usb->configure_card();
 127         }
 128
 129         BMUSBCapture::start_bm_thread();
 130
 131         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 132                 cards[card_index].usb->start_bm_capture();
 133         }
 134
 135         //chain->enable_phase_timing(true);
 136
 137         // Set up stuff for NV12 conversion.
 138
 139         // Cb/Cr shader.
 140         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 141         string cbcr_frag_shader =
 142                 "#version 130 \n"
 143                 "in vec2 tc0; \n"
 144                 "uniform sampler2D cbcr_tex; \n"
 145                 "void main() { \n"
 146                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 147                 "} \n";
 148         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
 149
 150         r128.init(2, OUTPUT_FREQUENCY);
 151         r128.integr_start();
 152
 153         locut.init(FILTER_HPF, 2);
 154
 155         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 156         // and there's a limit to how important the peak meter is.
 157         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
 158
 159         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 160 }
 161
 162 Mixer::~Mixer()
 163 {
 164         resource_pool->release_glsl_program(cbcr_program_num);
 165         BMUSBCapture::stop_bm_thread();
 166
 167         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 168                 {
 169                         unique_lock<mutex> lock(bmusb_mutex);
 170                         cards[card_index].should_quit = true;  // Unblock thread.
 171                         cards[card_index].new_data_ready_changed.notify_all();
 172                 }
 173                 cards[card_index].usb->stop_dequeue_thread();
 174         }
 175
 176         h264_encoder.reset(nullptr);
 177 }
 178
 179 namespace {
 180
 181 int unwrap_timecode(uint16_t current_wrapped, int last)
 182 {
 183         uint16_t last_wrapped = last & 0xffff;
 184         if (current_wrapped > last_wrapped) {
 185                 return (last & ~0xffff) | current_wrapped;
 186         } else {
 187                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 188         }
 189 }
 190
 191 float find_peak(const float *samples, size_t num_samples)
 192 {
 193         float m = fabs(samples[0]);
 194         for (size_t i = 1; i < num_samples; ++i) {
 195                 m = std::max(m, fabs(samples[i]));
 196         }
 197         return m;
 198 }
 199
 200 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 201 {
 202         size_t num_samples = in.size() / 2;
 203         out_l->resize(num_samples);
 204         out_r->resize(num_samples);
 205
 206         const float *inptr = in.data();
 207         float *lptr = &(*out_l)[0];
 208         float *rptr = &(*out_r)[0];
 209         for (size_t i = 0; i < num_samples; ++i) {
 210                 *lptr++ = *inptr++;
 211                 *rptr++ = *inptr++;
 212         }
 213 }
 214
 215 }  // namespace
 216
 217 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 218                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 219                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 220 {
 221         CaptureCard *card = &cards[card_index];
 222
 223         int width, height, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
 224         bool interlaced;
 225
 226         decode_video_format(video_format, &width, &height, &extra_lines_top, &extra_lines_bottom,
 227                             &frame_rate_nom, &frame_rate_den, &interlaced);  // Ignore return value for now.
 228         int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom;
 229
 230         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 231         if (num_samples > OUTPUT_FREQUENCY / 10) {
 232                 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",
 233                         card_index, int(audio_frame.len), int(audio_offset),
 234                         timecode, int(video_frame.len), int(video_offset), video_format);
 235                 if (video_frame.owner) {
 236                         video_frame.owner->release_frame(video_frame);
 237                 }
 238                 if (audio_frame.owner) {
 239                         audio_frame.owner->release_frame(audio_frame);
 240                 }
 241                 return;
 242         }
 243
 244         int64_t local_pts = card->next_local_pts;
 245         int dropped_frames = 0;
 246         if (card->last_timecode != -1) {
 247                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 248         }
 249
 250         // Convert the audio to stereo fp32 and add it.
 251         vector<float> audio;
 252         audio.resize(num_samples * 2);
 253         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 254
 255         // Add the audio.
 256         {
 257                 unique_lock<mutex> lock(card->audio_mutex);
 258
 259                 // Number of samples per frame if we need to insert silence.
 260                 // (Could be nonintegral, but resampling will save us then.)
 261                 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
 262
 263                 if (dropped_frames > MAX_FPS * 2) {
 264                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 265                                 card_index, card->last_timecode, timecode);
 266                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 267                         dropped_frames = 0;
 268                 } else if (dropped_frames > 0) {
 269                         // Insert silence as needed.
 270                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 271                                 card_index, dropped_frames, timecode);
 272                         vector<float> silence;
 273                         silence.resize(silence_samples * 2);
 274                         for (int i = 0; i < dropped_frames; ++i) {
 275                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 276                                 // Note that if the format changed in the meantime, we have
 277                                 // no way of detecting that; we just have to assume the frame length
 278                                 // is always the same.
 279                                 local_pts += frame_length;
 280                         }
 281                 }
 282                 if (num_samples == 0) {
 283                         audio.resize(silence_samples * 2);
 284                         num_samples = silence_samples;
 285                 }
 286                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 287                 card->next_local_pts = local_pts + frame_length;
 288         }
 289
 290         card->last_timecode = timecode;
 291
 292         // Done with the audio, so release it.
 293         if (audio_frame.owner) {
 294                 audio_frame.owner->release_frame(audio_frame);
 295         }
 296
 297         {
 298                 // Wait until the previous frame was consumed.
 299                 unique_lock<mutex> lock(bmusb_mutex);
 300                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 301                 if (card->should_quit) return;
 302         }
 303
 304         if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
 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 skipped_lines = 25;
 339         size_t cbcr_width = WIDTH / 2;
 340         size_t cbcr_offset = video_offset / 2;
 341         size_t y_offset = cbcr_offset + cbcr_width * (HEIGHT + EXTRAHEIGHT) * sizeof(uint16_t) + video_offset / 2;
 342
 343         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
 344         check_error();
 345         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * skipped_lines * sizeof(uint16_t)));
 346         check_error();
 347         glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
 348         check_error();
 349         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH, HEIGHT, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + WIDTH * skipped_lines));
 350         check_error();
 351         glBindTexture(GL_TEXTURE_2D, 0);
 352         check_error();
 353         GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 354         check_error();
 355         assert(fence != nullptr);
 356
 357         {
 358                 unique_lock<mutex> lock(bmusb_mutex);
 359                 card->new_data_ready = true;
 360                 card->new_frame = RefCountedFrame(video_frame);
 361                 card->new_frame_length = frame_length;
 362                 card->new_data_ready_fence = fence;
 363                 card->dropped_frames = dropped_frames;
 364                 card->new_data_ready_changed.notify_all();
 365         }
 366 }
 367
 368 void Mixer::thread_func()
 369 {
 370         eglBindAPI(EGL_OPENGL_API);
 371         QOpenGLContext *context = create_context(mixer_surface);
 372         if (!make_current(context, mixer_surface)) {
 373                 printf("oops\n");
 374                 exit(1);
 375         }
 376
 377         struct timespec start, now;
 378         clock_gettime(CLOCK_MONOTONIC, &start);
 379
 380         int frame = 0;
 381         int stats_dropped_frames = 0;
 382
 383         while (!should_quit) {
 384                 CaptureCard card_copy[MAX_CARDS];
 385                 int num_samples[MAX_CARDS];
 386
 387                 {
 388                         unique_lock<mutex> lock(bmusb_mutex);
 389
 390                         // The first card is the master timer, so wait for it to have a new frame.
 391                         // TODO: Make configurable, and with a timeout.
 392                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 393
 394                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 395                                 CaptureCard *card = &cards[card_index];
 396                                 card_copy[card_index].usb = card->usb;
 397                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 398                                 card_copy[card_index].new_frame = card->new_frame;
 399                                 card_copy[card_index].new_frame_length = card->new_frame_length;
 400                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 401                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 402                                 card->new_data_ready = false;
 403                                 card->new_data_ready_changed.notify_all();
 404
 405                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
 406                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 407                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 408                                 assert(num_samples[card_index] >= 0);
 409                         }
 410                 }
 411
 412                 // Resample the audio as needed, including from previously dropped frames.
 413                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 414                         {
 415                                 // Signal to the audio thread to process this frame.
 416                                 unique_lock<mutex> lock(audio_mutex);
 417                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
 418                                 audio_task_queue_changed.notify_one();
 419                         }
 420                         if (frame_num != card_copy[0].dropped_frames) {
 421                                 // For dropped frames, increase the pts. Note that if the format changed
 422                                 // in the meantime, we have no way of detecting that; we just have to
 423                                 // assume the frame length is always the same.
 424                                 ++stats_dropped_frames;
 425                                 pts_int += card_copy[0].new_frame_length;
 426                         }
 427                 }
 428
 429                 if (audio_level_callback != nullptr) {
 430                         double loudness_s = r128.loudness_S();
 431                         double loudness_i = r128.integrated();
 432                         double loudness_range_low = r128.range_min();
 433                         double loudness_range_high = r128.range_max();
 434
 435                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 436                                              loudness_i, loudness_range_low, loudness_range_high,
 437                                              last_gain_staging_db);
 438                 }
 439
 440                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 441                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 442                                 ++card_copy[card_index].dropped_frames;
 443                         }
 444                         if (card_copy[card_index].dropped_frames > 0) {
 445                                 printf("Card %u dropped %d frames before this\n",
 446                                         card_index, int(card_copy[card_index].dropped_frames));
 447                         }
 448                 }
 449
 450                 // If the first card is reporting a corrupted or otherwise dropped frame,
 451                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 452                 if (card_copy[0].new_frame->len == 0) {
 453                         ++stats_dropped_frames;
 454                         pts_int += card_copy[0].new_frame_length;
 455                         continue;
 456                 }
 457
 458                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 459                         CaptureCard *card = &card_copy[card_index];
 460                         if (!card->new_data_ready || card->new_frame->len == 0)
 461                                 continue;
 462
 463                         assert(card->new_frame != nullptr);
 464                         bmusb_current_rendering_frame[card_index] = card->new_frame;
 465                         check_error();
 466
 467                         // The new texture might still be uploaded,
 468                         // tell the GPU to wait until it's there.
 469                         if (card->new_data_ready_fence) {
 470                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 471                                 check_error();
 472                                 glDeleteSync(card->new_data_ready_fence);
 473                                 check_error();
 474                         }
 475                         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
 476                         theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
 477                 }
 478
 479                 // Get the main chain from the theme, and set its state immediately.
 480                 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
 481                 EffectChain *chain = theme_main_chain.first;
 482                 theme_main_chain.second();
 483
 484                 GLuint y_tex, cbcr_tex;
 485                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 486                 assert(got_frame);
 487
 488                 // Render main chain.
 489                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 490                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 491                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 492                 check_error();
 493                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 494                 resource_pool->release_fbo(fbo);
 495
 496                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 497                 resource_pool->release_2d_texture(cbcr_full_tex);
 498
 499                 // Set the right state for rgba_tex.
 500                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 501                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 502                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 503                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 504                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 505
 506                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 507                 check_error();
 508
 509                 // Make sure the H.264 gets a reference to all the
 510                 // input frames needed, so that they are not released back
 511                 // until the rendering is done.
 512                 vector<RefCountedFrame> input_frames;
 513                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 514                         input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 515                 }
 516                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 517                 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
 518                 ++frame;
 519                 pts_int += card_copy[0].new_frame_length;
 520
 521                 // The live frame just shows the RGBA texture we just rendered.
 522                 // It owns rgba_tex now.
 523                 DisplayFrame live_frame;
 524                 live_frame.chain = display_chain.get();
 525                 live_frame.setup_chain = [this, rgba_tex]{
 526                         display_input->set_texture_num(rgba_tex);
 527                 };
 528                 live_frame.ready_fence = fence;
 529                 live_frame.input_frames = {};
 530                 live_frame.temp_textures = { rgba_tex };
 531                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 532
 533                 // Set up preview and any additional channels.
 534                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 535                         DisplayFrame display_frame;
 536                         pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT);  // FIXME: dimensions
 537                         display_frame.chain = chain.first;
 538                         display_frame.setup_chain = chain.second;
 539                         display_frame.ready_fence = fence;
 540
 541                         // FIXME: possible to do better?
 542                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 543                                 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 544                         }
 545                         display_frame.temp_textures = {};
 546                         output_channel[i].output_frame(display_frame);
 547                 }
 548
 549                 clock_gettime(CLOCK_MONOTONIC, &now);
 550                 double elapsed = now.tv_sec - start.tv_sec +
 551                         1e-9 * (now.tv_nsec - start.tv_nsec);
 552                 if (frame % 100 == 0) {
 553                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 554                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 555                                 1e3 * elapsed / frame);
 556                 //      chain->print_phase_timing();
 557                 }
 558
 559 #if 0
 560                 // Reset every 100 frames, so that local variations in frame times
 561                 // (especially for the first few frames, when the shaders are
 562                 // compiled etc.) don't make it hard to measure for the entire
 563                 // remaining duration of the program.
 564                 if (frame == 10000) {
 565                         frame = 0;
 566                         start = now;
 567                 }
 568 #endif
 569                 check_error();
 570         }
 571
 572         resource_pool->clean_context();
 573 }
 574
 575 void Mixer::audio_thread_func()
 576 {
 577         while (!should_quit) {
 578                 AudioTask task;
 579
 580                 {
 581                         unique_lock<mutex> lock(audio_mutex);
 582                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 583                         task = audio_task_queue.front();
 584                         audio_task_queue.pop();
 585                 }
 586
 587                 process_audio_one_frame(task.pts_int, task.num_samples);
 588         }
 589 }
 590
 591 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 592 {
 593         vector<float> samples_card;
 594         vector<float> samples_out;
 595         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 596                 samples_card.resize(num_samples * 2);
 597                 {
 598                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 599                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 600                                 printf("Card %d reported previous underrun.\n", card_index);
 601                         }
 602                 }
 603                 // TODO: Allow using audio from the other card(s) as well.
 604                 if (card_index == 0) {
 605                         samples_out = move(samples_card);
 606                 }
 607         }
 608
 609         // Cut away everything under 120 Hz (or whatever the cutoff is);
 610         // we don't need it for voice, and it will reduce headroom
 611         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 612         // should be dampened.)
 613         locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 614
 615         // Apply a level compressor to get the general level right.
 616         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 617         // (or more precisely, near it, since we don't use infinite ratio),
 618         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 619         // entirely arbitrary, but from practical tests with speech, it seems to
 620         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 621         float ref_level_dbfs = -14.0f;
 622         {
 623                 float threshold = 0.01f;   // -40 dBFS.
 624                 float ratio = 20.0f;
 625                 float attack_time = 0.5f;
 626                 float release_time = 20.0f;
 627                 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 628                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 629                 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 630         }
 631
 632 #if 0
 633         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 634                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 635                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 636                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 637 #endif
 638
 639 //      float limiter_att, compressor_att;
 640
 641         // The real compressor.
 642         if (compressor_enabled) {
 643                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 644                 float ratio = 20.0f;
 645                 float attack_time = 0.005f;
 646                 float release_time = 0.040f;
 647                 float makeup_gain = 2.0f;  // +6 dB.
 648                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 649 //              compressor_att = compressor.get_attenuation();
 650         }
 651
 652         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 653         // Note that since ratio is not infinite, we could go slightly higher than this.
 654         if (limiter_enabled) {
 655                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 656                 float ratio = 30.0f;
 657                 float attack_time = 0.0f;  // Instant.
 658                 float release_time = 0.020f;
 659                 float makeup_gain = 1.0f;  // 0 dB.
 660                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 661 //              limiter_att = limiter.get_attenuation();
 662         }
 663
 664 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 665
 666         // Upsample 4x to find interpolated peak.
 667         peak_resampler.inp_data = samples_out.data();
 668         peak_resampler.inp_count = samples_out.size() / 2;
 669
 670         vector<float> interpolated_samples_out;
 671         interpolated_samples_out.resize(samples_out.size());
 672         while (peak_resampler.inp_count > 0) {  // About four iterations.
 673                 peak_resampler.out_data = &interpolated_samples_out[0];
 674                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 675                 peak_resampler.process();
 676                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 677                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 678         }
 679
 680         // Find R128 levels.
 681         vector<float> left, right;
 682         deinterleave_samples(samples_out, &left, &right);
 683         float *ptrs[] = { left.data(), right.data() };
 684         r128.process(left.size(), ptrs);
 685
 686         // Send the samples to the sound card.
 687         if (alsa) {
 688                 alsa->write(samples_out);
 689         }
 690
 691         // And finally add them to the output.
 692         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 693 }
 694
 695 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 696 {
 697         GLuint vao;
 698         glGenVertexArrays(1, &vao);
 699         check_error();
 700
 701         float vertices[] = {
 702                 0.0f, 2.0f,
 703                 0.0f, 0.0f,
 704                 2.0f, 0.0f
 705         };
 706
 707         glBindVertexArray(vao);
 708         check_error();
 709
 710         // Extract Cb/Cr.
 711         GLuint fbo = resource_pool->create_fbo(dst_tex);
 712         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 713         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 714         check_error();
 715
 716         glUseProgram(cbcr_program_num);
 717         check_error();
 718
 719         glActiveTexture(GL_TEXTURE0);
 720         check_error();
 721         glBindTexture(GL_TEXTURE_2D, src_tex);
 722         check_error();
 723         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 724         check_error();
 725         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 726         check_error();
 727         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 728         check_error();
 729
 730         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 731         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 732
 733         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 734         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 735
 736         glDrawArrays(GL_TRIANGLES, 0, 3);
 737         check_error();
 738
 739         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 740         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 741
 742         glUseProgram(0);
 743         check_error();
 744
 745         resource_pool->release_fbo(fbo);
 746         glDeleteVertexArrays(1, &vao);
 747 }
 748
 749 void Mixer::release_display_frame(DisplayFrame *frame)
 750 {
 751         for (GLuint texnum : frame->temp_textures) {
 752                 resource_pool->release_2d_texture(texnum);
 753         }
 754         frame->temp_textures.clear();
 755         frame->ready_fence.reset();
 756         frame->input_frames.clear();
 757 }
 758
 759 void Mixer::start()
 760 {
 761         mixer_thread = thread(&Mixer::thread_func, this);
 762         audio_thread = thread(&Mixer::audio_thread_func, this);
 763 }
 764
 765 void Mixer::quit()
 766 {
 767         should_quit = true;
 768         mixer_thread.join();
 769         audio_thread.join();
 770 }
 771
 772 void Mixer::transition_clicked(int transition_num)
 773 {
 774         theme->transition_clicked(transition_num, pts());
 775 }
 776
 777 void Mixer::channel_clicked(int preview_num)
 778 {
 779         theme->channel_clicked(preview_num);
 780 }
 781
 782 void Mixer::reset_meters()
 783 {
 784         peak_resampler.reset();
 785         peak = 0.0f;
 786         r128.reset();
 787         r128.integr_start();
 788 }
 789
 790 Mixer::OutputChannel::~OutputChannel()
 791 {
 792         if (has_current_frame) {
 793                 parent->release_display_frame(&current_frame);
 794         }
 795         if (has_ready_frame) {
 796                 parent->release_display_frame(&ready_frame);
 797         }
 798 }
 799
 800 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 801 {
 802         // Store this frame for display. Remove the ready frame if any
 803         // (it was seemingly never used).
 804         {
 805                 unique_lock<mutex> lock(frame_mutex);
 806                 if (has_ready_frame) {
 807                         parent->release_display_frame(&ready_frame);
 808                 }
 809                 ready_frame = frame;
 810                 has_ready_frame = true;
 811         }
 812
 813         if (has_new_frame_ready_callback) {
 814                 new_frame_ready_callback();
 815         }
 816 }
 817
 818 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 819 {
 820         unique_lock<mutex> lock(frame_mutex);
 821         if (!has_current_frame && !has_ready_frame) {
 822                 return false;
 823         }
 824
 825         if (has_current_frame && has_ready_frame) {
 826                 // We have a new ready frame. Toss the current one.
 827                 parent->release_display_frame(&current_frame);
 828                 has_current_frame = false;
 829         }
 830         if (has_ready_frame) {
 831                 assert(!has_current_frame);
 832                 current_frame = ready_frame;
 833                 ready_frame.ready_fence.reset();  // Drop the refcount.
 834                 ready_frame.input_frames.clear();  // Drop the refcounts.
 835                 has_current_frame = true;
 836                 has_ready_frame = false;
 837         }
 838
 839         *frame = current_frame;
 840         return true;
 841 }
 842
 843 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 844 {
 845         new_frame_ready_callback = callback;
 846         has_new_frame_ready_callback = true;
 847 }