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

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