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("test.ts", 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
 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
 177 namespace {
 178
 179 int unwrap_timecode(uint16_t current_wrapped, int last)
 180 {
 181         uint16_t last_wrapped = last & 0xffff;
 182         if (current_wrapped > last_wrapped) {
 183                 return (last & ~0xffff) | current_wrapped;
 184         } else {
 185                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 186         }
 187 }
 188
 189 float find_peak(const float *samples, size_t num_samples)
 190 {
 191         float m = fabs(samples[0]);
 192         for (size_t i = 1; i < num_samples; ++i) {
 193                 m = std::max(m, fabs(samples[i]));
 194         }
 195         return m;
 196 }
 197
 198 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 199 {
 200         size_t num_samples = in.size() / 2;
 201         out_l->resize(num_samples);
 202         out_r->resize(num_samples);
 203
 204         const float *inptr = in.data();
 205         float *lptr = &(*out_l)[0];
 206         float *rptr = &(*out_r)[0];
 207         for (size_t i = 0; i < num_samples; ++i) {
 208                 *lptr++ = *inptr++;
 209                 *rptr++ = *inptr++;
 210         }
 211 }
 212
 213 }  // namespace
 214
 215 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 216                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 217                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 218 {
 219         CaptureCard *card = &cards[card_index];
 220
 221         if (audio_frame.len - audio_offset > 30000) {
 222                 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",
 223                         card_index, int(audio_frame.len), int(audio_offset),
 224                         timecode, int(video_frame.len), int(video_offset), video_format);
 225                 if (video_frame.owner) {
 226                         video_frame.owner->release_frame(video_frame);
 227                 }
 228                 if (audio_frame.owner) {
 229                         audio_frame.owner->release_frame(audio_frame);
 230                 }
 231                 return;
 232         }
 233
 234         int unwrapped_timecode = timecode;
 235         int dropped_frames = 0;
 236         if (card->last_timecode != -1) {
 237                 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
 238                 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
 239         }
 240         card->last_timecode = unwrapped_timecode;
 241
 242         // Convert the audio to stereo fp32 and add it.
 243         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 244         vector<float> audio;
 245         audio.resize(num_samples * 2);
 246         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 247
 248         // Add the audio.
 249         {
 250                 unique_lock<mutex> lock(card->audio_mutex);
 251
 252                 int unwrapped_timecode = timecode;
 253                 if (dropped_frames > FPS * 2) {
 254                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
 255                                 card_index);
 256                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 257                 } else if (dropped_frames > 0) {
 258                         // Insert silence as needed.
 259                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 260                                 card_index, dropped_frames, timecode);
 261                         vector<float> silence;
 262                         silence.resize((OUTPUT_FREQUENCY / FPS) * 2);
 263                         for (int i = 0; i < dropped_frames; ++i) {
 264                                 card->resampling_queue->add_input_samples((unwrapped_timecode - dropped_frames + i) / double(FPS), silence.data(), (OUTPUT_FREQUENCY / FPS));
 265                         }
 266                 }
 267                 card->resampling_queue->add_input_samples(unwrapped_timecode / double(FPS), audio.data(), num_samples);
 268         }
 269
 270         // Done with the audio, so release it.
 271         if (audio_frame.owner) {
 272                 audio_frame.owner->release_frame(audio_frame);
 273         }
 274
 275         {
 276                 // Wait until the previous frame was consumed.
 277                 unique_lock<mutex> lock(bmusb_mutex);
 278                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 279                 if (card->should_quit) return;
 280         }
 281
 282         if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
 283                 if (video_frame.len != 0) {
 284                         printf("Card %d: Dropping video frame with wrong length (%ld)\n",
 285                                 card_index, video_frame.len - video_offset);
 286                 }
 287                 if (video_frame.owner) {
 288                         video_frame.owner->release_frame(video_frame);
 289                 }
 290
 291                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 292                 // so that pts can go up accordingly.
 293                 {
 294                         unique_lock<mutex> lock(bmusb_mutex);
 295                         card->new_data_ready = true;
 296                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 297                         card->new_data_ready_fence = nullptr;
 298                         card->dropped_frames = dropped_frames;
 299                         card->new_data_ready_changed.notify_all();
 300                 }
 301                 return;
 302         }
 303
 304         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
 305         GLuint pbo = userdata->pbo;
 306         check_error();
 307         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 308         check_error();
 309         glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
 310         check_error();
 311         //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 312         //check_error();
 313
 314         // Upload the textures.
 315         glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
 316         check_error();
 317         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));
 318         check_error();
 319         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
 320         check_error();
 321         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
 322         check_error();
 323         glBindTexture(GL_TEXTURE_2D, 0);
 324         check_error();
 325         GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 326         check_error();
 327         assert(fence != nullptr);
 328
 329         {
 330                 unique_lock<mutex> lock(bmusb_mutex);
 331                 card->new_data_ready = true;
 332                 card->new_frame = RefCountedFrame(video_frame);
 333                 card->new_data_ready_fence = fence;
 334                 card->dropped_frames = dropped_frames;
 335                 card->new_data_ready_changed.notify_all();
 336         }
 337 }
 338
 339 void Mixer::thread_func()
 340 {
 341         eglBindAPI(EGL_OPENGL_API);
 342         QOpenGLContext *context = create_context(mixer_surface);
 343         if (!make_current(context, mixer_surface)) {
 344                 printf("oops\n");
 345                 exit(1);
 346         }
 347
 348         struct timespec start, now;
 349         clock_gettime(CLOCK_MONOTONIC, &start);
 350
 351         int frame = 0;
 352         int dropped_frames = 0;
 353
 354         while (!should_quit) {
 355                 CaptureCard card_copy[MAX_CARDS];
 356
 357                 {
 358                         unique_lock<mutex> lock(bmusb_mutex);
 359
 360                         // The first card is the master timer, so wait for it to have a new frame.
 361                         // TODO: Make configurable, and with a timeout.
 362                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 363
 364                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 365                                 CaptureCard *card = &cards[card_index];
 366                                 card_copy[card_index].usb = card->usb;
 367                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 368                                 card_copy[card_index].new_frame = card->new_frame;
 369                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 370                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 371                                 card->new_data_ready = false;
 372                                 card->new_data_ready_changed.notify_all();
 373                         }
 374                 }
 375
 376                 // Resample the audio as needed, including from previously dropped frames.
 377                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 378                         process_audio_one_frame();
 379                         if (frame_num != card_copy[0].dropped_frames) {
 380                                 // For dropped frames, increase the pts.
 381                                 ++dropped_frames;
 382                                 pts_int += TIMEBASE / FPS;
 383                         }
 384                 }
 385
 386                 if (audio_level_callback != nullptr) {
 387                         double loudness_s = r128.loudness_S();
 388                         double loudness_i = r128.integrated();
 389                         double loudness_range_low = r128.range_min();
 390                         double loudness_range_high = r128.range_max();
 391
 392                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 393                                              loudness_i, loudness_range_low, loudness_range_high,
 394                                              last_gain_staging_db);
 395                 }
 396
 397                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 398                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 399                                 ++card_copy[card_index].dropped_frames;
 400                         }
 401                         if (card_copy[card_index].dropped_frames > 0) {
 402                                 printf("Card %u dropped %d frames before this\n",
 403                                         card_index, int(card_copy[card_index].dropped_frames));
 404                         }
 405                 }
 406
 407                 // If the first card is reporting a corrupted or otherwise dropped frame,
 408                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 409                 if (card_copy[0].new_frame->len == 0) {
 410                         ++dropped_frames;
 411                         pts_int += TIMEBASE / FPS;
 412                         continue;
 413                 }
 414
 415                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 416                         CaptureCard *card = &card_copy[card_index];
 417                         if (!card->new_data_ready || card->new_frame->len == 0)
 418                                 continue;
 419
 420                         assert(card->new_frame != nullptr);
 421                         bmusb_current_rendering_frame[card_index] = card->new_frame;
 422                         check_error();
 423
 424                         // The new texture might still be uploaded,
 425                         // tell the GPU to wait until it's there.
 426                         if (card->new_data_ready_fence) {
 427                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 428                                 check_error();
 429                                 glDeleteSync(card->new_data_ready_fence);
 430                                 check_error();
 431                         }
 432                         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
 433                         theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
 434                 }
 435
 436                 // Get the main chain from the theme, and set its state immediately.
 437                 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
 438                 EffectChain *chain = theme_main_chain.first;
 439                 theme_main_chain.second();
 440
 441                 GLuint y_tex, cbcr_tex;
 442                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 443                 assert(got_frame);
 444
 445                 // Render main chain.
 446                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 447                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 448                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 449                 check_error();
 450                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 451                 resource_pool->release_fbo(fbo);
 452
 453                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 454                 resource_pool->release_2d_texture(cbcr_full_tex);
 455
 456                 // Set the right state for rgba_tex.
 457                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 458                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 459                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 460                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 461                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 462
 463                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 464                 check_error();
 465
 466                 // Make sure the H.264 gets a reference to all the
 467                 // input frames needed, so that they are not released back
 468                 // until the rendering is done.
 469                 vector<RefCountedFrame> input_frames;
 470                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 471                         input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 472                 }
 473                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 474                 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
 475                 ++frame;
 476                 pts_int += TIMEBASE / FPS;
 477
 478                 // The live frame just shows the RGBA texture we just rendered.
 479                 // It owns rgba_tex now.
 480                 DisplayFrame live_frame;
 481                 live_frame.chain = display_chain.get();
 482                 live_frame.setup_chain = [this, rgba_tex]{
 483                         display_input->set_texture_num(rgba_tex);
 484                 };
 485                 live_frame.ready_fence = fence;
 486                 live_frame.input_frames = {};
 487                 live_frame.temp_textures = { rgba_tex };
 488                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 489
 490                 // Set up preview and any additional channels.
 491                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 492                         DisplayFrame display_frame;
 493                         pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT);  // FIXME: dimensions
 494                         display_frame.chain = chain.first;
 495                         display_frame.setup_chain = chain.second;
 496                         display_frame.ready_fence = fence;
 497
 498                         // FIXME: possible to do better?
 499                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 500                                 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 501                         }
 502                         display_frame.temp_textures = {};
 503                         output_channel[i].output_frame(display_frame);
 504                 }
 505
 506                 clock_gettime(CLOCK_MONOTONIC, &now);
 507                 double elapsed = now.tv_sec - start.tv_sec +
 508                         1e-9 * (now.tv_nsec - start.tv_nsec);
 509                 if (frame % 100 == 0) {
 510                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 511                                 frame, dropped_frames, elapsed, frame / elapsed,
 512                                 1e3 * elapsed / frame);
 513                 //      chain->print_phase_timing();
 514                 }
 515
 516 #if 0
 517                 // Reset every 100 frames, so that local variations in frame times
 518                 // (especially for the first few frames, when the shaders are
 519                 // compiled etc.) don't make it hard to measure for the entire
 520                 // remaining duration of the program.
 521                 if (frame == 10000) {
 522                         frame = 0;
 523                         start = now;
 524                 }
 525 #endif
 526                 check_error();
 527         }
 528
 529         resource_pool->clean_context();
 530 }
 531
 532 void Mixer::process_audio_one_frame()
 533 {
 534         vector<float> samples_card;
 535         vector<float> samples_out;
 536         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 537                 samples_card.resize((OUTPUT_FREQUENCY / FPS) * 2);
 538                 {
 539                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 540                         if (!cards[card_index].resampling_queue->get_output_samples(pts(), &samples_card[0], OUTPUT_FREQUENCY / FPS)) {
 541                                 printf("Card %d reported previous underrun.\n", card_index);
 542                         }
 543                 }
 544                 // TODO: Allow using audio from the other card(s) as well.
 545                 if (card_index == 0) {
 546                         samples_out = move(samples_card);
 547                 }
 548         }
 549
 550         // Cut away everything under 150 Hz; we don't need it for voice,
 551         // and it will reduce headroom and confuse the compressor.
 552         // (In particular, any hums at 50 or 60 Hz should be dampened.)
 553         locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 554
 555         // Apply a level compressor to get the general level right.
 556         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 557         // (or more precisely, near it, since we don't use infinite ratio),
 558         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 559         // entirely arbitrary, but from practical tests with speech, it seems to
 560         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 561         float ref_level_dbfs = -14.0f;
 562         {
 563                 float threshold = 0.01f;   // -40 dBFS.
 564                 float ratio = 20.0f;
 565                 float attack_time = 0.5f;
 566                 float release_time = 20.0f;
 567                 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 568                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 569                 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 570         }
 571
 572 #if 0
 573         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 574                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 575                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 576                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 577 #endif
 578
 579 //      float limiter_att, compressor_att;
 580
 581         // The real compressor.
 582         if (compressor_enabled) {
 583                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 584                 float ratio = 20.0f;
 585                 float attack_time = 0.005f;
 586                 float release_time = 0.040f;
 587                 float makeup_gain = 2.0f;  // +6 dB.
 588                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 589 //              compressor_att = compressor.get_attenuation();
 590         }
 591
 592         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 593         // Note that since ratio is not infinite, we could go slightly higher than this.
 594         if (limiter_enabled) {
 595                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 596                 float ratio = 30.0f;
 597                 float attack_time = 0.0f;  // Instant.
 598                 float release_time = 0.020f;
 599                 float makeup_gain = 1.0f;  // 0 dB.
 600                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 601 //              limiter_att = limiter.get_attenuation();
 602         }
 603
 604 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 605
 606         // Upsample 4x to find interpolated peak.
 607         peak_resampler.inp_data = samples_out.data();
 608         peak_resampler.inp_count = samples_out.size() / 2;
 609
 610         vector<float> interpolated_samples_out;
 611         interpolated_samples_out.resize(samples_out.size());
 612         while (peak_resampler.inp_count > 0) {  // About four iterations.
 613                 peak_resampler.out_data = &interpolated_samples_out[0];
 614                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 615                 peak_resampler.process();
 616                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 617                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 618         }
 619
 620         // Find R128 levels.
 621         vector<float> left, right;
 622         deinterleave_samples(samples_out, &left, &right);
 623         float *ptrs[] = { left.data(), right.data() };
 624         r128.process(left.size(), ptrs);
 625
 626         // Actually add the samples to the output.
 627         h264_encoder->add_audio(pts_int, move(samples_out));
 628 }
 629
 630 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 631 {
 632         GLuint vao;
 633         glGenVertexArrays(1, &vao);
 634         check_error();
 635
 636         float vertices[] = {
 637                 0.0f, 2.0f,
 638                 0.0f, 0.0f,
 639                 2.0f, 0.0f
 640         };
 641
 642         glBindVertexArray(vao);
 643         check_error();
 644
 645         // Extract Cb/Cr.
 646         GLuint fbo = resource_pool->create_fbo(dst_tex);
 647         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 648         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 649         check_error();
 650
 651         glUseProgram(cbcr_program_num);
 652         check_error();
 653
 654         glActiveTexture(GL_TEXTURE0);
 655         check_error();
 656         glBindTexture(GL_TEXTURE_2D, src_tex);
 657         check_error();
 658         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 659         check_error();
 660         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 661         check_error();
 662         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 663         check_error();
 664
 665         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 666         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 667
 668         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 669         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 670
 671         glDrawArrays(GL_TRIANGLES, 0, 3);
 672         check_error();
 673
 674         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 675         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 676
 677         glUseProgram(0);
 678         check_error();
 679
 680         resource_pool->release_fbo(fbo);
 681         glDeleteVertexArrays(1, &vao);
 682 }
 683
 684 void Mixer::release_display_frame(DisplayFrame *frame)
 685 {
 686         for (GLuint texnum : frame->temp_textures) {
 687                 resource_pool->release_2d_texture(texnum);
 688         }
 689         frame->temp_textures.clear();
 690         frame->ready_fence.reset();
 691         frame->input_frames.clear();
 692 }
 693
 694 void Mixer::start()
 695 {
 696         mixer_thread = thread(&Mixer::thread_func, this);
 697 }
 698
 699 void Mixer::quit()
 700 {
 701         should_quit = true;
 702         mixer_thread.join();
 703 }
 704
 705 void Mixer::transition_clicked(int transition_num)
 706 {
 707         theme->transition_clicked(transition_num, pts());
 708 }
 709
 710 void Mixer::channel_clicked(int preview_num)
 711 {
 712         theme->channel_clicked(preview_num);
 713 }
 714
 715 void Mixer::reset_meters()
 716 {
 717         peak_resampler.reset();
 718         peak = 0.0f;
 719         r128.reset();
 720         r128.integr_start();
 721 }
 722
 723 Mixer::OutputChannel::~OutputChannel()
 724 {
 725         if (has_current_frame) {
 726                 parent->release_display_frame(&current_frame);
 727         }
 728         if (has_ready_frame) {
 729                 parent->release_display_frame(&ready_frame);
 730         }
 731 }
 732
 733 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 734 {
 735         // Store this frame for display. Remove the ready frame if any
 736         // (it was seemingly never used).
 737         {
 738                 unique_lock<mutex> lock(frame_mutex);
 739                 if (has_ready_frame) {
 740                         parent->release_display_frame(&ready_frame);
 741                 }
 742                 ready_frame = frame;
 743                 has_ready_frame = true;
 744         }
 745
 746         if (has_new_frame_ready_callback) {
 747                 new_frame_ready_callback();
 748         }
 749 }
 750
 751 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 752 {
 753         unique_lock<mutex> lock(frame_mutex);
 754         if (!has_current_frame && !has_ready_frame) {
 755                 return false;
 756         }
 757
 758         if (has_current_frame && has_ready_frame) {
 759                 // We have a new ready frame. Toss the current one.
 760                 parent->release_display_frame(&current_frame);
 761                 has_current_frame = false;
 762         }
 763         if (has_ready_frame) {
 764                 assert(!has_current_frame);
 765                 current_frame = ready_frame;
 766                 ready_frame.ready_fence.reset();  // Drop the refcount.
 767                 ready_frame.input_frames.clear();  // Drop the refcounts.
 768                 has_current_frame = true;
 769                 has_ready_frame = false;
 770         }
 771
 772         *frame = current_frame;
 773         return true;
 774 }
 775
 776 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 777 {
 778         new_frame_ready_callback = callback;
 779         has_new_frame_ready_callback = true;
 780 }