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

   1 #define WIDTH 1280
   2 #define HEIGHT 720
   3 #define EXTRAHEIGHT 30
   4
   5 #undef Success
   6
   7 #include "mixer.h"
   8
   9 #include <assert.h>
  10 #include <epoxy/egl.h>
  11 #include <init.h>
  12 #include <movit/effect_chain.h>
  13 #include <movit/effect_util.h>
  14 #include <movit/flat_input.h>
  15 #include <movit/image_format.h>
  16 #include <movit/resource_pool.h>
  17 #include <stdint.h>
  18 #include <stdio.h>
  19 #include <stdlib.h>
  20 #include <sys/time.h>
  21 #include <time.h>
  22 #include <util.h>
  23 #include <algorithm>
  24 #include <cmath>
  25 #include <condition_variable>
  26 #include <cstddef>
  27 #include <memory>
  28 #include <mutex>
  29 #include <string>
  30 #include <thread>
  31 #include <utility>
  32 #include <vector>
  33
  34 #include "bmusb/bmusb.h"
  35 #include "context.h"
  36 #include "defs.h"
  37 #include "h264encode.h"
  38 #include "pbo_frame_allocator.h"
  39 #include "ref_counted_gl_sync.h"
  40 #include "timebase.h"
  41
  42 class QOpenGLContext;
  43
  44 using namespace movit;
  45 using namespace std;
  46 using namespace std::placeholders;
  47
  48 Mixer *global_mixer = nullptr;
  49
  50 namespace {
  51
  52 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  53 {
  54         for (size_t i = 0; i < num_samples; ++i) {
  55                 for (size_t j = 0; j < out_channels; ++j) {
  56                         uint32_t s1 = *src++;
  57                         uint32_t s2 = *src++;
  58                         uint32_t s3 = *src++;
  59                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  60                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  61                 }
  62                 src += 3 * (in_channels - out_channels);
  63         }
  64 }
  65
  66 }  // namespace
  67
  68 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
  69         : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT),
  70           num_cards(num_cards),
  71           mixer_surface(create_surface(format)),
  72           h264_encoder_surface(create_surface(format)),
  73           level_compressor(OUTPUT_FREQUENCY),
  74           limiter(OUTPUT_FREQUENCY),
  75           compressor(OUTPUT_FREQUENCY)
  76 {
  77         httpd.start(9095);
  78
  79         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
  80         check_error();
  81
  82         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
  83         // will be halved when sampling them, and we need to compensate here.
  84         movit_texel_subpixel_precision /= 2.0;
  85
  86         resource_pool.reset(new ResourcePool);
  87         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
  88         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
  89                 output_channel[i].parent = this;
  90         }
  91
  92         ImageFormat inout_format;
  93         inout_format.color_space = COLORSPACE_sRGB;
  94         inout_format.gamma_curve = GAMMA_sRGB;
  95
  96         // Display chain; shows the live output produced by the main chain (its RGBA version).
  97         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
  98         check_error();
  99         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 100         display_chain->add_input(display_input);
 101         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 102         display_chain->set_dither_bits(0);  // Don't bother.
 103         display_chain->finalize();
 104
 105         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
 106
 107         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 108                 printf("Configuring card %d...\n", card_index);
 109                 CaptureCard *card = &cards[card_index];
 110                 card->usb = new BMUSBCapture(card_index);
 111                 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 112                 card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44, WIDTH, HEIGHT));
 113                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
 114                 card->surface = create_surface(format);
 115                 card->usb->set_dequeue_thread_callbacks(
 116                         [card]{
 117                                 eglBindAPI(EGL_OPENGL_API);
 118                                 card->context = create_context(card->surface);
 119                                 if (!make_current(card->context, card->surface)) {
 120                                         printf("failed to create bmusb context\n");
 121                                         exit(1);
 122                                 }
 123                         },
 124                         [this]{
 125                                 resource_pool->clean_context();
 126                         });
 127                 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 128                 card->usb->configure_card();
 129         }
 130
 131         BMUSBCapture::start_bm_thread();
 132
 133         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 134                 cards[card_index].usb->start_bm_capture();
 135         }
 136
 137         //chain->enable_phase_timing(true);
 138
 139         // Set up stuff for NV12 conversion.
 140
 141         // Cb/Cr shader.
 142         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 143         string cbcr_frag_shader =
 144                 "#version 130 \n"
 145                 "in vec2 tc0; \n"
 146                 "uniform sampler2D cbcr_tex; \n"
 147                 "void main() { \n"
 148                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 149                 "} \n";
 150         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
 151
 152         r128.init(2, OUTPUT_FREQUENCY);
 153         r128.integr_start();
 154
 155         locut.init(FILTER_HPF, 2);
 156
 157         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 158         // and there's a limit to how important the peak meter is.
 159         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
 160
 161         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 162 }
 163
 164 Mixer::~Mixer()
 165 {
 166         resource_pool->release_glsl_program(cbcr_program_num);
 167         BMUSBCapture::stop_bm_thread();
 168
 169         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 170                 {
 171                         unique_lock<mutex> lock(bmusb_mutex);
 172                         cards[card_index].should_quit = true;  // Unblock thread.
 173                         cards[card_index].new_data_ready_changed.notify_all();
 174                 }
 175                 cards[card_index].usb->stop_dequeue_thread();
 176         }
 177
 178         h264_encoder.reset(nullptr);
 179 }
 180
 181 namespace {
 182
 183 int unwrap_timecode(uint16_t current_wrapped, int last)
 184 {
 185         uint16_t last_wrapped = last & 0xffff;
 186         if (current_wrapped > last_wrapped) {
 187                 return (last & ~0xffff) | current_wrapped;
 188         } else {
 189                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 190         }
 191 }
 192
 193 float find_peak(const float *samples, size_t num_samples)
 194 {
 195         float m = fabs(samples[0]);
 196         for (size_t i = 1; i < num_samples; ++i) {
 197                 m = std::max(m, fabs(samples[i]));
 198         }
 199         return m;
 200 }
 201
 202 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 203 {
 204         size_t num_samples = in.size() / 2;
 205         out_l->resize(num_samples);
 206         out_r->resize(num_samples);
 207
 208         const float *inptr = in.data();
 209         float *lptr = &(*out_l)[0];
 210         float *rptr = &(*out_r)[0];
 211         for (size_t i = 0; i < num_samples; ++i) {
 212                 *lptr++ = *inptr++;
 213                 *rptr++ = *inptr++;
 214         }
 215 }
 216
 217 }  // namespace
 218
 219 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 220                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 221                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 222 {
 223         CaptureCard *card = &cards[card_index];
 224
 225         if (audio_frame.len - audio_offset > 30000) {
 226                 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",
 227                         card_index, int(audio_frame.len), int(audio_offset),
 228                         timecode, int(video_frame.len), int(video_offset), video_format);
 229                 if (video_frame.owner) {
 230                         video_frame.owner->release_frame(video_frame);
 231                 }
 232                 if (audio_frame.owner) {
 233                         audio_frame.owner->release_frame(audio_frame);
 234                 }
 235                 return;
 236         }
 237
 238         int unwrapped_timecode = timecode;
 239         int dropped_frames = 0;
 240         if (card->last_timecode != -1) {
 241                 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
 242                 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
 243         }
 244         card->last_timecode = unwrapped_timecode;
 245
 246         // Convert the audio to stereo fp32 and add it.
 247         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 248         vector<float> audio;
 249         audio.resize(num_samples * 2);
 250         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 251
 252         // Add the audio.
 253         {
 254                 unique_lock<mutex> lock(card->audio_mutex);
 255
 256                 int unwrapped_timecode = timecode;
 257                 if (dropped_frames > FPS * 2) {
 258                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
 259                                 card_index);
 260                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 261                 } else if (dropped_frames > 0) {
 262                         // Insert silence as needed.
 263                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 264                                 card_index, dropped_frames, timecode);
 265                         vector<float> silence;
 266                         silence.resize((OUTPUT_FREQUENCY / FPS) * 2);
 267                         for (int i = 0; i < dropped_frames; ++i) {
 268                                 card->resampling_queue->add_input_samples((unwrapped_timecode - dropped_frames + i) / double(FPS), silence.data(), (OUTPUT_FREQUENCY / FPS));
 269                         }
 270                 }
 271                 card->resampling_queue->add_input_samples(unwrapped_timecode / double(FPS), audio.data(), num_samples);
 272         }
 273
 274         // Done with the audio, so release it.
 275         if (audio_frame.owner) {
 276                 audio_frame.owner->release_frame(audio_frame);
 277         }
 278
 279         {
 280                 // Wait until the previous frame was consumed.
 281                 unique_lock<mutex> lock(bmusb_mutex);
 282                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 283                 if (card->should_quit) return;
 284         }
 285
 286         if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
 287                 if (video_frame.len != 0) {
 288                         printf("Card %d: Dropping video frame with wrong length (%ld)\n",
 289                                 card_index, video_frame.len - video_offset);
 290                 }
 291                 if (video_frame.owner) {
 292                         video_frame.owner->release_frame(video_frame);
 293                 }
 294
 295                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 296                 // so that pts can go up accordingly.
 297                 {
 298                         unique_lock<mutex> lock(bmusb_mutex);
 299                         card->new_data_ready = true;
 300                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 301                         card->new_data_ready_fence = nullptr;
 302                         card->dropped_frames = dropped_frames;
 303                         card->new_data_ready_changed.notify_all();
 304                 }
 305                 return;
 306         }
 307
 308         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
 309         GLuint pbo = userdata->pbo;
 310         check_error();
 311         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 312         check_error();
 313         glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
 314         check_error();
 315         //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 316         //check_error();
 317
 318         // Upload the textures.
 319         glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
 320         check_error();
 321         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));
 322         check_error();
 323         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
 324         check_error();
 325         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
 326         check_error();
 327         glBindTexture(GL_TEXTURE_2D, 0);
 328         check_error();
 329         GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 330         check_error();
 331         assert(fence != nullptr);
 332
 333         {
 334                 unique_lock<mutex> lock(bmusb_mutex);
 335                 card->new_data_ready = true;
 336                 card->new_frame = RefCountedFrame(video_frame);
 337                 card->new_data_ready_fence = fence;
 338                 card->dropped_frames = dropped_frames;
 339                 card->new_data_ready_changed.notify_all();
 340         }
 341 }
 342
 343 void Mixer::thread_func()
 344 {
 345         eglBindAPI(EGL_OPENGL_API);
 346         QOpenGLContext *context = create_context(mixer_surface);
 347         if (!make_current(context, mixer_surface)) {
 348                 printf("oops\n");
 349                 exit(1);
 350         }
 351
 352         struct timespec start, now;
 353         clock_gettime(CLOCK_MONOTONIC, &start);
 354
 355         int frame = 0;
 356         int dropped_frames = 0;
 357
 358         while (!should_quit) {
 359                 CaptureCard card_copy[MAX_CARDS];
 360
 361                 {
 362                         unique_lock<mutex> lock(bmusb_mutex);
 363
 364                         // The first card is the master timer, so wait for it to have a new frame.
 365                         // TODO: Make configurable, and with a timeout.
 366                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 367
 368                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 369                                 CaptureCard *card = &cards[card_index];
 370                                 card_copy[card_index].usb = card->usb;
 371                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 372                                 card_copy[card_index].new_frame = card->new_frame;
 373                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 374                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 375                                 card->new_data_ready = false;
 376                                 card->new_data_ready_changed.notify_all();
 377                         }
 378                 }
 379
 380                 // Resample the audio as needed, including from previously dropped frames.
 381                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 382                         {
 383                                 // Signal to the audio thread to process this frame.
 384                                 unique_lock<mutex> lock(audio_mutex);
 385                                 audio_pts_queue.push(pts_int);
 386                                 audio_pts_queue_changed.notify_one();
 387                         }
 388                         if (frame_num != card_copy[0].dropped_frames) {
 389                                 // For dropped frames, increase the pts.
 390                                 ++dropped_frames;
 391                                 pts_int += TIMEBASE / FPS;
 392                         }
 393                 }
 394
 395                 if (audio_level_callback != nullptr) {
 396                         double loudness_s = r128.loudness_S();
 397                         double loudness_i = r128.integrated();
 398                         double loudness_range_low = r128.range_min();
 399                         double loudness_range_high = r128.range_max();
 400
 401                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 402                                              loudness_i, loudness_range_low, loudness_range_high,
 403                                              last_gain_staging_db);
 404                 }
 405
 406                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 407                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 408                                 ++card_copy[card_index].dropped_frames;
 409                         }
 410                         if (card_copy[card_index].dropped_frames > 0) {
 411                                 printf("Card %u dropped %d frames before this\n",
 412                                         card_index, int(card_copy[card_index].dropped_frames));
 413                         }
 414                 }
 415
 416                 // If the first card is reporting a corrupted or otherwise dropped frame,
 417                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 418                 if (card_copy[0].new_frame->len == 0) {
 419                         ++dropped_frames;
 420                         pts_int += TIMEBASE / FPS;
 421                         continue;
 422                 }
 423
 424                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 425                         CaptureCard *card = &card_copy[card_index];
 426                         if (!card->new_data_ready || card->new_frame->len == 0)
 427                                 continue;
 428
 429                         assert(card->new_frame != nullptr);
 430                         bmusb_current_rendering_frame[card_index] = card->new_frame;
 431                         check_error();
 432
 433                         // The new texture might still be uploaded,
 434                         // tell the GPU to wait until it's there.
 435                         if (card->new_data_ready_fence) {
 436                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 437                                 check_error();
 438                                 glDeleteSync(card->new_data_ready_fence);
 439                                 check_error();
 440                         }
 441                         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
 442                         theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
 443                 }
 444
 445                 // Get the main chain from the theme, and set its state immediately.
 446                 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
 447                 EffectChain *chain = theme_main_chain.first;
 448                 theme_main_chain.second();
 449
 450                 GLuint y_tex, cbcr_tex;
 451                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 452                 assert(got_frame);
 453
 454                 // Render main chain.
 455                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 456                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 457                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 458                 check_error();
 459                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 460                 resource_pool->release_fbo(fbo);
 461
 462                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 463                 resource_pool->release_2d_texture(cbcr_full_tex);
 464
 465                 // Set the right state for rgba_tex.
 466                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 467                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 468                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 469                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 470                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 471
 472                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 473                 check_error();
 474
 475                 // Make sure the H.264 gets a reference to all the
 476                 // input frames needed, so that they are not released back
 477                 // until the rendering is done.
 478                 vector<RefCountedFrame> input_frames;
 479                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 480                         input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 481                 }
 482                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 483                 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
 484                 ++frame;
 485                 pts_int += TIMEBASE / FPS;
 486
 487                 // The live frame just shows the RGBA texture we just rendered.
 488                 // It owns rgba_tex now.
 489                 DisplayFrame live_frame;
 490                 live_frame.chain = display_chain.get();
 491                 live_frame.setup_chain = [this, rgba_tex]{
 492                         display_input->set_texture_num(rgba_tex);
 493                 };
 494                 live_frame.ready_fence = fence;
 495                 live_frame.input_frames = {};
 496                 live_frame.temp_textures = { rgba_tex };
 497                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 498
 499                 // Set up preview and any additional channels.
 500                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 501                         DisplayFrame display_frame;
 502                         pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT);  // FIXME: dimensions
 503                         display_frame.chain = chain.first;
 504                         display_frame.setup_chain = chain.second;
 505                         display_frame.ready_fence = fence;
 506
 507                         // FIXME: possible to do better?
 508                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 509                                 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 510                         }
 511                         display_frame.temp_textures = {};
 512                         output_channel[i].output_frame(display_frame);
 513                 }
 514
 515                 clock_gettime(CLOCK_MONOTONIC, &now);
 516                 double elapsed = now.tv_sec - start.tv_sec +
 517                         1e-9 * (now.tv_nsec - start.tv_nsec);
 518                 if (frame % 100 == 0) {
 519                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 520                                 frame, dropped_frames, elapsed, frame / elapsed,
 521                                 1e3 * elapsed / frame);
 522                 //      chain->print_phase_timing();
 523                 }
 524
 525 #if 0
 526                 // Reset every 100 frames, so that local variations in frame times
 527                 // (especially for the first few frames, when the shaders are
 528                 // compiled etc.) don't make it hard to measure for the entire
 529                 // remaining duration of the program.
 530                 if (frame == 10000) {
 531                         frame = 0;
 532                         start = now;
 533                 }
 534 #endif
 535                 check_error();
 536         }
 537
 538         resource_pool->clean_context();
 539 }
 540
 541 void Mixer::audio_thread_func()
 542 {
 543         while (!should_quit) {
 544                 int64_t frame_pts_int;
 545
 546                 {
 547                         unique_lock<mutex> lock(audio_mutex);
 548                         audio_pts_queue_changed.wait(lock, [this]{ return !audio_pts_queue.empty(); });
 549                         frame_pts_int = audio_pts_queue.front();
 550                         audio_pts_queue.pop();
 551                 }
 552
 553                 process_audio_one_frame(frame_pts_int);
 554         }
 555 }
 556
 557 void Mixer::process_audio_one_frame(int64_t frame_pts_int)
 558 {
 559         vector<float> samples_card;
 560         vector<float> samples_out;
 561         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 562                 samples_card.resize((OUTPUT_FREQUENCY / FPS) * 2);
 563                 {
 564                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 565                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], OUTPUT_FREQUENCY / FPS)) {
 566                                 printf("Card %d reported previous underrun.\n", card_index);
 567                         }
 568                 }
 569                 // TODO: Allow using audio from the other card(s) as well.
 570                 if (card_index == 0) {
 571                         samples_out = move(samples_card);
 572                 }
 573         }
 574
 575         // Cut away everything under 120 Hz (or whatever the cutoff is);
 576         // we don't need it for voice, and it will reduce headroom
 577         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 578         // should be dampened.)
 579         locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 580
 581         // Apply a level compressor to get the general level right.
 582         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 583         // (or more precisely, near it, since we don't use infinite ratio),
 584         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 585         // entirely arbitrary, but from practical tests with speech, it seems to
 586         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 587         float ref_level_dbfs = -14.0f;
 588         {
 589                 float threshold = 0.01f;   // -40 dBFS.
 590                 float ratio = 20.0f;
 591                 float attack_time = 0.5f;
 592                 float release_time = 20.0f;
 593                 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 594                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 595                 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 596         }
 597
 598 #if 0
 599         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 600                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 601                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 602                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 603 #endif
 604
 605 //      float limiter_att, compressor_att;
 606
 607         // The real compressor.
 608         if (compressor_enabled) {
 609                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 610                 float ratio = 20.0f;
 611                 float attack_time = 0.005f;
 612                 float release_time = 0.040f;
 613                 float makeup_gain = 2.0f;  // +6 dB.
 614                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 615 //              compressor_att = compressor.get_attenuation();
 616         }
 617
 618         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 619         // Note that since ratio is not infinite, we could go slightly higher than this.
 620         if (limiter_enabled) {
 621                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 622                 float ratio = 30.0f;
 623                 float attack_time = 0.0f;  // Instant.
 624                 float release_time = 0.020f;
 625                 float makeup_gain = 1.0f;  // 0 dB.
 626                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 627 //              limiter_att = limiter.get_attenuation();
 628         }
 629
 630 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 631
 632         // Upsample 4x to find interpolated peak.
 633         peak_resampler.inp_data = samples_out.data();
 634         peak_resampler.inp_count = samples_out.size() / 2;
 635
 636         vector<float> interpolated_samples_out;
 637         interpolated_samples_out.resize(samples_out.size());
 638         while (peak_resampler.inp_count > 0) {  // About four iterations.
 639                 peak_resampler.out_data = &interpolated_samples_out[0];
 640                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 641                 peak_resampler.process();
 642                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 643                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 644         }
 645
 646         // Find R128 levels.
 647         vector<float> left, right;
 648         deinterleave_samples(samples_out, &left, &right);
 649         float *ptrs[] = { left.data(), right.data() };
 650         r128.process(left.size(), ptrs);
 651
 652         // Send the samples to the sound card.
 653         if (alsa) {
 654                 alsa->write(samples_out);
 655         }
 656
 657         // And finally add them to the output.
 658         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 659 }
 660
 661 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 662 {
 663         GLuint vao;
 664         glGenVertexArrays(1, &vao);
 665         check_error();
 666
 667         float vertices[] = {
 668                 0.0f, 2.0f,
 669                 0.0f, 0.0f,
 670                 2.0f, 0.0f
 671         };
 672
 673         glBindVertexArray(vao);
 674         check_error();
 675
 676         // Extract Cb/Cr.
 677         GLuint fbo = resource_pool->create_fbo(dst_tex);
 678         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 679         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 680         check_error();
 681
 682         glUseProgram(cbcr_program_num);
 683         check_error();
 684
 685         glActiveTexture(GL_TEXTURE0);
 686         check_error();
 687         glBindTexture(GL_TEXTURE_2D, src_tex);
 688         check_error();
 689         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 690         check_error();
 691         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 692         check_error();
 693         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 694         check_error();
 695
 696         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 697         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 698
 699         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 700         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 701
 702         glDrawArrays(GL_TRIANGLES, 0, 3);
 703         check_error();
 704
 705         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 706         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 707
 708         glUseProgram(0);
 709         check_error();
 710
 711         resource_pool->release_fbo(fbo);
 712         glDeleteVertexArrays(1, &vao);
 713 }
 714
 715 void Mixer::release_display_frame(DisplayFrame *frame)
 716 {
 717         for (GLuint texnum : frame->temp_textures) {
 718                 resource_pool->release_2d_texture(texnum);
 719         }
 720         frame->temp_textures.clear();
 721         frame->ready_fence.reset();
 722         frame->input_frames.clear();
 723 }
 724
 725 void Mixer::start()
 726 {
 727         mixer_thread = thread(&Mixer::thread_func, this);
 728         audio_thread = thread(&Mixer::audio_thread_func, this);
 729 }
 730
 731 void Mixer::quit()
 732 {
 733         should_quit = true;
 734         mixer_thread.join();
 735         audio_thread.join();
 736 }
 737
 738 void Mixer::transition_clicked(int transition_num)
 739 {
 740         theme->transition_clicked(transition_num, pts());
 741 }
 742
 743 void Mixer::channel_clicked(int preview_num)
 744 {
 745         theme->channel_clicked(preview_num);
 746 }
 747
 748 void Mixer::reset_meters()
 749 {
 750         peak_resampler.reset();
 751         peak = 0.0f;
 752         r128.reset();
 753         r128.integr_start();
 754 }
 755
 756 Mixer::OutputChannel::~OutputChannel()
 757 {
 758         if (has_current_frame) {
 759                 parent->release_display_frame(&current_frame);
 760         }
 761         if (has_ready_frame) {
 762                 parent->release_display_frame(&ready_frame);
 763         }
 764 }
 765
 766 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 767 {
 768         // Store this frame for display. Remove the ready frame if any
 769         // (it was seemingly never used).
 770         {
 771                 unique_lock<mutex> lock(frame_mutex);
 772                 if (has_ready_frame) {
 773                         parent->release_display_frame(&ready_frame);
 774                 }
 775                 ready_frame = frame;
 776                 has_ready_frame = true;
 777         }
 778
 779         if (has_new_frame_ready_callback) {
 780                 new_frame_ready_callback();
 781         }
 782 }
 783
 784 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 785 {
 786         unique_lock<mutex> lock(frame_mutex);
 787         if (!has_current_frame && !has_ready_frame) {
 788                 return false;
 789         }
 790
 791         if (has_current_frame && has_ready_frame) {
 792                 // We have a new ready frame. Toss the current one.
 793                 parent->release_display_frame(&current_frame);
 794                 has_current_frame = false;
 795         }
 796         if (has_ready_frame) {
 797                 assert(!has_current_frame);
 798                 current_frame = ready_frame;
 799                 ready_frame.ready_fence.reset();  // Drop the refcount.
 800                 ready_frame.input_frames.clear();  // Drop the refcounts.
 801                 has_current_frame = true;
 802                 has_ready_frame = false;
 803         }
 804
 805         *frame = current_frame;
 806         return true;
 807 }
 808
 809 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 810 {
 811         new_frame_ready_callback = callback;
 812         has_new_frame_ready_callback = true;
 813 }