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, 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         if (audio_frame.len - audio_offset > 30000) {
 224                 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",
 225                         card_index, int(audio_frame.len), int(audio_offset),
 226                         timecode, int(video_frame.len), int(video_offset), video_format);
 227                 if (video_frame.owner) {
 228                         video_frame.owner->release_frame(video_frame);
 229                 }
 230                 if (audio_frame.owner) {
 231                         audio_frame.owner->release_frame(audio_frame);
 232                 }
 233                 return;
 234         }
 235
 236         int unwrapped_timecode = timecode;
 237         int dropped_frames = 0;
 238         if (card->last_timecode != -1) {
 239                 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
 240                 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
 241         }
 242         card->last_timecode = unwrapped_timecode;
 243
 244         // Convert the audio to stereo fp32 and add it.
 245         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 246         vector<float> audio;
 247         audio.resize(num_samples * 2);
 248         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 249
 250         // Add the audio.
 251         {
 252                 unique_lock<mutex> lock(card->audio_mutex);
 253
 254                 int unwrapped_timecode = timecode;
 255                 if (dropped_frames > FPS * 2) {
 256                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
 257                                 card_index);
 258                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 259                 } else if (dropped_frames > 0) {
 260                         // Insert silence as needed.
 261                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 262                                 card_index, dropped_frames, timecode);
 263                         vector<float> silence;
 264                         silence.resize((OUTPUT_FREQUENCY / FPS) * 2);
 265                         for (int i = 0; i < dropped_frames; ++i) {
 266                                 card->resampling_queue->add_input_samples((unwrapped_timecode - dropped_frames + i) / double(FPS), silence.data(), (OUTPUT_FREQUENCY / FPS));
 267                         }
 268                 }
 269                 card->resampling_queue->add_input_samples(unwrapped_timecode / double(FPS), audio.data(), num_samples);
 270         }
 271
 272         // Done with the audio, so release it.
 273         if (audio_frame.owner) {
 274                 audio_frame.owner->release_frame(audio_frame);
 275         }
 276
 277         {
 278                 // Wait until the previous frame was consumed.
 279                 unique_lock<mutex> lock(bmusb_mutex);
 280                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 281                 if (card->should_quit) return;
 282         }
 283
 284         if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
 285                 if (video_frame.len != 0) {
 286                         printf("Card %d: Dropping video frame with wrong length (%ld)\n",
 287                                 card_index, video_frame.len - video_offset);
 288                 }
 289                 if (video_frame.owner) {
 290                         video_frame.owner->release_frame(video_frame);
 291                 }
 292
 293                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 294                 // so that pts can go up accordingly.
 295                 {
 296                         unique_lock<mutex> lock(bmusb_mutex);
 297                         card->new_data_ready = true;
 298                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 299                         card->new_data_ready_fence = nullptr;
 300                         card->dropped_frames = dropped_frames;
 301                         card->new_data_ready_changed.notify_all();
 302                 }
 303                 return;
 304         }
 305
 306         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
 307         GLuint pbo = userdata->pbo;
 308         check_error();
 309         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 310         check_error();
 311         glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
 312         check_error();
 313         //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 314         //check_error();
 315
 316         // Upload the textures.
 317         glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
 318         check_error();
 319         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH, HEIGHT, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44) / 2 + WIDTH * 25 + 22));
 320         check_error();
 321         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
 322         check_error();
 323         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
 324         check_error();
 325         glBindTexture(GL_TEXTURE_2D, 0);
 326         check_error();
 327         GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 328         check_error();
 329         assert(fence != nullptr);
 330
 331         {
 332                 unique_lock<mutex> lock(bmusb_mutex);
 333                 card->new_data_ready = true;
 334                 card->new_frame = RefCountedFrame(video_frame);
 335                 card->new_data_ready_fence = fence;
 336                 card->dropped_frames = dropped_frames;
 337                 card->new_data_ready_changed.notify_all();
 338         }
 339 }
 340
 341 void Mixer::thread_func()
 342 {
 343         eglBindAPI(EGL_OPENGL_API);
 344         QOpenGLContext *context = create_context(mixer_surface);
 345         if (!make_current(context, mixer_surface)) {
 346                 printf("oops\n");
 347                 exit(1);
 348         }
 349
 350         struct timespec start, now;
 351         clock_gettime(CLOCK_MONOTONIC, &start);
 352
 353         int frame = 0;
 354         int dropped_frames = 0;
 355
 356         while (!should_quit) {
 357                 CaptureCard card_copy[MAX_CARDS];
 358
 359                 {
 360                         unique_lock<mutex> lock(bmusb_mutex);
 361
 362                         // The first card is the master timer, so wait for it to have a new frame.
 363                         // TODO: Make configurable, and with a timeout.
 364                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 365
 366                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 367                                 CaptureCard *card = &cards[card_index];
 368                                 card_copy[card_index].usb = card->usb;
 369                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 370                                 card_copy[card_index].new_frame = card->new_frame;
 371                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 372                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 373                                 card->new_data_ready = false;
 374                                 card->new_data_ready_changed.notify_all();
 375                         }
 376                 }
 377
 378                 // Resample the audio as needed, including from previously dropped frames.
 379                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 380                         {
 381                                 // Signal to the audio thread to process this frame.
 382                                 unique_lock<mutex> lock(audio_mutex);
 383                                 audio_pts_queue.push(pts_int);
 384                                 audio_pts_queue_changed.notify_one();
 385                         }
 386                         if (frame_num != card_copy[0].dropped_frames) {
 387                                 // For dropped frames, increase the pts.
 388                                 ++dropped_frames;
 389                                 pts_int += TIMEBASE / FPS;
 390                         }
 391                 }
 392
 393                 if (audio_level_callback != nullptr) {
 394                         double loudness_s = r128.loudness_S();
 395                         double loudness_i = r128.integrated();
 396                         double loudness_range_low = r128.range_min();
 397                         double loudness_range_high = r128.range_max();
 398
 399                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 400                                              loudness_i, loudness_range_low, loudness_range_high,
 401                                              last_gain_staging_db);
 402                 }
 403
 404                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 405                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 406                                 ++card_copy[card_index].dropped_frames;
 407                         }
 408                         if (card_copy[card_index].dropped_frames > 0) {
 409                                 printf("Card %u dropped %d frames before this\n",
 410                                         card_index, int(card_copy[card_index].dropped_frames));
 411                         }
 412                 }
 413
 414                 // If the first card is reporting a corrupted or otherwise dropped frame,
 415                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 416                 if (card_copy[0].new_frame->len == 0) {
 417                         ++dropped_frames;
 418                         pts_int += TIMEBASE / FPS;
 419                         continue;
 420                 }
 421
 422                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 423                         CaptureCard *card = &card_copy[card_index];
 424                         if (!card->new_data_ready || card->new_frame->len == 0)
 425                                 continue;
 426
 427                         assert(card->new_frame != nullptr);
 428                         bmusb_current_rendering_frame[card_index] = card->new_frame;
 429                         check_error();
 430
 431                         // The new texture might still be uploaded,
 432                         // tell the GPU to wait until it's there.
 433                         if (card->new_data_ready_fence) {
 434                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 435                                 check_error();
 436                                 glDeleteSync(card->new_data_ready_fence);
 437                                 check_error();
 438                         }
 439                         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
 440                         theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
 441                 }
 442
 443                 // Get the main chain from the theme, and set its state immediately.
 444                 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
 445                 EffectChain *chain = theme_main_chain.first;
 446                 theme_main_chain.second();
 447
 448                 GLuint y_tex, cbcr_tex;
 449                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 450                 assert(got_frame);
 451
 452                 // Render main chain.
 453                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 454                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 455                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 456                 check_error();
 457                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 458                 resource_pool->release_fbo(fbo);
 459
 460                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 461                 resource_pool->release_2d_texture(cbcr_full_tex);
 462
 463                 // Set the right state for rgba_tex.
 464                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 465                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 466                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 467                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 468                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 469
 470                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 471                 check_error();
 472
 473                 // Make sure the H.264 gets a reference to all the
 474                 // input frames needed, so that they are not released back
 475                 // until the rendering is done.
 476                 vector<RefCountedFrame> input_frames;
 477                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 478                         input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 479                 }
 480                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 481                 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
 482                 ++frame;
 483                 pts_int += TIMEBASE / FPS;
 484
 485                 // The live frame just shows the RGBA texture we just rendered.
 486                 // It owns rgba_tex now.
 487                 DisplayFrame live_frame;
 488                 live_frame.chain = display_chain.get();
 489                 live_frame.setup_chain = [this, rgba_tex]{
 490                         display_input->set_texture_num(rgba_tex);
 491                 };
 492                 live_frame.ready_fence = fence;
 493                 live_frame.input_frames = {};
 494                 live_frame.temp_textures = { rgba_tex };
 495                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 496
 497                 // Set up preview and any additional channels.
 498                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 499                         DisplayFrame display_frame;
 500                         pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT);  // FIXME: dimensions
 501                         display_frame.chain = chain.first;
 502                         display_frame.setup_chain = chain.second;
 503                         display_frame.ready_fence = fence;
 504
 505                         // FIXME: possible to do better?
 506                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 507                                 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 508                         }
 509                         display_frame.temp_textures = {};
 510                         output_channel[i].output_frame(display_frame);
 511                 }
 512
 513                 clock_gettime(CLOCK_MONOTONIC, &now);
 514                 double elapsed = now.tv_sec - start.tv_sec +
 515                         1e-9 * (now.tv_nsec - start.tv_nsec);
 516                 if (frame % 100 == 0) {
 517                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 518                                 frame, dropped_frames, elapsed, frame / elapsed,
 519                                 1e3 * elapsed / frame);
 520                 //      chain->print_phase_timing();
 521                 }
 522
 523 #if 0
 524                 // Reset every 100 frames, so that local variations in frame times
 525                 // (especially for the first few frames, when the shaders are
 526                 // compiled etc.) don't make it hard to measure for the entire
 527                 // remaining duration of the program.
 528                 if (frame == 10000) {
 529                         frame = 0;
 530                         start = now;
 531                 }
 532 #endif
 533                 check_error();
 534         }
 535
 536         resource_pool->clean_context();
 537 }
 538
 539 void Mixer::audio_thread_func()
 540 {
 541         while (!should_quit) {
 542                 int64_t frame_pts_int;
 543
 544                 {
 545                         unique_lock<mutex> lock(audio_mutex);
 546                         audio_pts_queue_changed.wait(lock, [this]{ return !audio_pts_queue.empty(); });
 547                         frame_pts_int = audio_pts_queue.front();
 548                         audio_pts_queue.pop();
 549                 }
 550
 551                 process_audio_one_frame(frame_pts_int);
 552         }
 553 }
 554
 555 void Mixer::process_audio_one_frame(int64_t frame_pts_int)
 556 {
 557         vector<float> samples_card;
 558         vector<float> samples_out;
 559         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 560                 samples_card.resize((OUTPUT_FREQUENCY / FPS) * 2);
 561                 {
 562                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 563                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], OUTPUT_FREQUENCY / FPS)) {
 564                                 printf("Card %d reported previous underrun.\n", card_index);
 565                         }
 566                 }
 567                 // TODO: Allow using audio from the other card(s) as well.
 568                 if (card_index == 0) {
 569                         samples_out = move(samples_card);
 570                 }
 571         }
 572
 573         // Cut away everything under 120 Hz (or whatever the cutoff is);
 574         // we don't need it for voice, and it will reduce headroom
 575         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 576         // should be dampened.)
 577         locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 578
 579         // Apply a level compressor to get the general level right.
 580         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 581         // (or more precisely, near it, since we don't use infinite ratio),
 582         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 583         // entirely arbitrary, but from practical tests with speech, it seems to
 584         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 585         float ref_level_dbfs = -14.0f;
 586         {
 587                 float threshold = 0.01f;   // -40 dBFS.
 588                 float ratio = 20.0f;
 589                 float attack_time = 0.5f;
 590                 float release_time = 20.0f;
 591                 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 592                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 593                 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 594         }
 595
 596 #if 0
 597         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 598                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 599                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 600                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 601 #endif
 602
 603 //      float limiter_att, compressor_att;
 604
 605         // The real compressor.
 606         if (compressor_enabled) {
 607                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 608                 float ratio = 20.0f;
 609                 float attack_time = 0.005f;
 610                 float release_time = 0.040f;
 611                 float makeup_gain = 2.0f;  // +6 dB.
 612                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 613 //              compressor_att = compressor.get_attenuation();
 614         }
 615
 616         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 617         // Note that since ratio is not infinite, we could go slightly higher than this.
 618         if (limiter_enabled) {
 619                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 620                 float ratio = 30.0f;
 621                 float attack_time = 0.0f;  // Instant.
 622                 float release_time = 0.020f;
 623                 float makeup_gain = 1.0f;  // 0 dB.
 624                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 625 //              limiter_att = limiter.get_attenuation();
 626         }
 627
 628 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 629
 630         // Upsample 4x to find interpolated peak.
 631         peak_resampler.inp_data = samples_out.data();
 632         peak_resampler.inp_count = samples_out.size() / 2;
 633
 634         vector<float> interpolated_samples_out;
 635         interpolated_samples_out.resize(samples_out.size());
 636         while (peak_resampler.inp_count > 0) {  // About four iterations.
 637                 peak_resampler.out_data = &interpolated_samples_out[0];
 638                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 639                 peak_resampler.process();
 640                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 641                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 642         }
 643
 644         // Find R128 levels.
 645         vector<float> left, right;
 646         deinterleave_samples(samples_out, &left, &right);
 647         float *ptrs[] = { left.data(), right.data() };
 648         r128.process(left.size(), ptrs);
 649
 650         // Send the samples to the sound card.
 651         if (alsa) {
 652                 alsa->write(samples_out);
 653         }
 654
 655         // And finally add them to the output.
 656         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 657 }
 658
 659 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 660 {
 661         GLuint vao;
 662         glGenVertexArrays(1, &vao);
 663         check_error();
 664
 665         float vertices[] = {
 666                 0.0f, 2.0f,
 667                 0.0f, 0.0f,
 668                 2.0f, 0.0f
 669         };
 670
 671         glBindVertexArray(vao);
 672         check_error();
 673
 674         // Extract Cb/Cr.
 675         GLuint fbo = resource_pool->create_fbo(dst_tex);
 676         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 677         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 678         check_error();
 679
 680         glUseProgram(cbcr_program_num);
 681         check_error();
 682
 683         glActiveTexture(GL_TEXTURE0);
 684         check_error();
 685         glBindTexture(GL_TEXTURE_2D, src_tex);
 686         check_error();
 687         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 688         check_error();
 689         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 690         check_error();
 691         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 692         check_error();
 693
 694         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 695         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 696
 697         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 698         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 699
 700         glDrawArrays(GL_TRIANGLES, 0, 3);
 701         check_error();
 702
 703         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 704         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 705
 706         glUseProgram(0);
 707         check_error();
 708
 709         resource_pool->release_fbo(fbo);
 710         glDeleteVertexArrays(1, &vao);
 711 }
 712
 713 void Mixer::release_display_frame(DisplayFrame *frame)
 714 {
 715         for (GLuint texnum : frame->temp_textures) {
 716                 resource_pool->release_2d_texture(texnum);
 717         }
 718         frame->temp_textures.clear();
 719         frame->ready_fence.reset();
 720         frame->input_frames.clear();
 721 }
 722
 723 void Mixer::start()
 724 {
 725         mixer_thread = thread(&Mixer::thread_func, this);
 726         audio_thread = thread(&Mixer::audio_thread_func, this);
 727 }
 728
 729 void Mixer::quit()
 730 {
 731         should_quit = true;
 732         mixer_thread.join();
 733         audio_thread.join();
 734 }
 735
 736 void Mixer::transition_clicked(int transition_num)
 737 {
 738         theme->transition_clicked(transition_num, pts());
 739 }
 740
 741 void Mixer::channel_clicked(int preview_num)
 742 {
 743         theme->channel_clicked(preview_num);
 744 }
 745
 746 void Mixer::reset_meters()
 747 {
 748         peak_resampler.reset();
 749         peak = 0.0f;
 750         r128.reset();
 751         r128.integr_start();
 752 }
 753
 754 Mixer::OutputChannel::~OutputChannel()
 755 {
 756         if (has_current_frame) {
 757                 parent->release_display_frame(&current_frame);
 758         }
 759         if (has_ready_frame) {
 760                 parent->release_display_frame(&ready_frame);
 761         }
 762 }
 763
 764 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 765 {
 766         // Store this frame for display. Remove the ready frame if any
 767         // (it was seemingly never used).
 768         {
 769                 unique_lock<mutex> lock(frame_mutex);
 770                 if (has_ready_frame) {
 771                         parent->release_display_frame(&ready_frame);
 772                 }
 773                 ready_frame = frame;
 774                 has_ready_frame = true;
 775         }
 776
 777         if (has_new_frame_ready_callback) {
 778                 new_frame_ready_callback();
 779         }
 780 }
 781
 782 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 783 {
 784         unique_lock<mutex> lock(frame_mutex);
 785         if (!has_current_frame && !has_ready_frame) {
 786                 return false;
 787         }
 788
 789         if (has_current_frame && has_ready_frame) {
 790                 // We have a new ready frame. Toss the current one.
 791                 parent->release_display_frame(&current_frame);
 792                 has_current_frame = false;
 793         }
 794         if (has_ready_frame) {
 795                 assert(!has_current_frame);
 796                 current_frame = ready_frame;
 797                 ready_frame.ready_fence.reset();  // Drop the refcount.
 798                 ready_frame.input_frames.clear();  // Drop the refcounts.
 799                 has_current_frame = true;
 800                 has_ready_frame = false;
 801         }
 802
 803         *frame = current_frame;
 804         return true;
 805 }
 806
 807 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 808 {
 809         new_frame_ready_callback = callback;
 810         has_new_frame_ready_callback = true;
 811 }