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

   1 #define EXTRAHEIGHT 30
   2
   3 #undef Success
   4
   5 #include "mixer.h"
   6
   7 #include <assert.h>
   8 #include <epoxy/egl.h>
   9 #include <init.h>
  10 #include <movit/effect_chain.h>
  11 #include <movit/effect_util.h>
  12 #include <movit/flat_input.h>
  13 #include <movit/image_format.h>
  14 #include <movit/resource_pool.h>
  15 #include <stdint.h>
  16 #include <stdio.h>
  17 #include <stdlib.h>
  18 #include <sys/time.h>
  19 #include <time.h>
  20 #include <util.h>
  21 #include <algorithm>
  22 #include <cmath>
  23 #include <condition_variable>
  24 #include <cstddef>
  25 #include <memory>
  26 #include <mutex>
  27 #include <string>
  28 #include <thread>
  29 #include <utility>
  30 #include <vector>
  31
  32 #include "bmusb/bmusb.h"
  33 #include "context.h"
  34 #include "defs.h"
  35 #include "h264encode.h"
  36 #include "pbo_frame_allocator.h"
  37 #include "ref_counted_gl_sync.h"
  38 #include "timebase.h"
  39
  40 class QOpenGLContext;
  41
  42 using namespace movit;
  43 using namespace std;
  44 using namespace std::placeholders;
  45
  46 Mixer *global_mixer = nullptr;
  47
  48 namespace {
  49
  50 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  51 {
  52         for (size_t i = 0; i < num_samples; ++i) {
  53                 for (size_t j = 0; j < out_channels; ++j) {
  54                         uint32_t s1 = *src++;
  55                         uint32_t s2 = *src++;
  56                         uint32_t s3 = *src++;
  57                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  58                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  59                 }
  60                 src += 3 * (in_channels - out_channels);
  61         }
  62 }
  63
  64 }  // namespace
  65
  66 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
  67         : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT),
  68           num_cards(num_cards),
  69           mixer_surface(create_surface(format)),
  70           h264_encoder_surface(create_surface(format)),
  71           level_compressor(OUTPUT_FREQUENCY),
  72           limiter(OUTPUT_FREQUENCY),
  73           compressor(OUTPUT_FREQUENCY)
  74 {
  75         httpd.start(9095);
  76
  77         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
  78         check_error();
  79
  80         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
  81         // will be halved when sampling them, and we need to compensate here.
  82         movit_texel_subpixel_precision /= 2.0;
  83
  84         resource_pool.reset(new ResourcePool);
  85         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
  86         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
  87                 output_channel[i].parent = this;
  88         }
  89
  90         ImageFormat inout_format;
  91         inout_format.color_space = COLORSPACE_sRGB;
  92         inout_format.gamma_curve = GAMMA_sRGB;
  93
  94         // Display chain; shows the live output produced by the main chain (its RGBA version).
  95         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
  96         check_error();
  97         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
  98         display_chain->add_input(display_input);
  99         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 100         display_chain->set_dither_bits(0);  // Don't bother.
 101         display_chain->finalize();
 102
 103         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
 104
 105         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 106                 printf("Configuring card %d...\n", card_index);
 107                 CaptureCard *card = &cards[card_index];
 108                 card->usb = new BMUSBCapture(card_index);
 109                 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 110                 card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44, WIDTH, HEIGHT));
 111                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
 112                 card->surface = create_surface(format);
 113                 card->usb->set_dequeue_thread_callbacks(
 114                         [card]{
 115                                 eglBindAPI(EGL_OPENGL_API);
 116                                 card->context = create_context(card->surface);
 117                                 if (!make_current(card->context, card->surface)) {
 118                                         printf("failed to create bmusb context\n");
 119                                         exit(1);
 120                                 }
 121                         },
 122                         [this]{
 123                                 resource_pool->clean_context();
 124                         });
 125                 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 126                 card->usb->configure_card();
 127         }
 128
 129         BMUSBCapture::start_bm_thread();
 130
 131         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 132                 cards[card_index].usb->start_bm_capture();
 133         }
 134
 135         //chain->enable_phase_timing(true);
 136
 137         // Set up stuff for NV12 conversion.
 138
 139         // Cb/Cr shader.
 140         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 141         string cbcr_frag_shader =
 142                 "#version 130 \n"
 143                 "in vec2 tc0; \n"
 144                 "uniform sampler2D cbcr_tex; \n"
 145                 "void main() { \n"
 146                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 147                 "} \n";
 148         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
 149
 150         r128.init(2, OUTPUT_FREQUENCY);
 151         r128.integr_start();
 152
 153         locut.init(FILTER_HPF, 2);
 154
 155         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 156         // and there's a limit to how important the peak meter is.
 157         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
 158
 159         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 160 }
 161
 162 Mixer::~Mixer()
 163 {
 164         resource_pool->release_glsl_program(cbcr_program_num);
 165         BMUSBCapture::stop_bm_thread();
 166
 167         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 168                 {
 169                         unique_lock<mutex> lock(bmusb_mutex);
 170                         cards[card_index].should_quit = true;  // Unblock thread.
 171                         cards[card_index].new_data_ready_changed.notify_all();
 172                 }
 173                 cards[card_index].usb->stop_dequeue_thread();
 174         }
 175
 176         h264_encoder.reset(nullptr);
 177 }
 178
 179 namespace {
 180
 181 int unwrap_timecode(uint16_t current_wrapped, int last)
 182 {
 183         uint16_t last_wrapped = last & 0xffff;
 184         if (current_wrapped > last_wrapped) {
 185                 return (last & ~0xffff) | current_wrapped;
 186         } else {
 187                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 188         }
 189 }
 190
 191 float find_peak(const float *samples, size_t num_samples)
 192 {
 193         float m = fabs(samples[0]);
 194         for (size_t i = 1; i < num_samples; ++i) {
 195                 m = std::max(m, fabs(samples[i]));
 196         }
 197         return m;
 198 }
 199
 200 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 201 {
 202         size_t num_samples = in.size() / 2;
 203         out_l->resize(num_samples);
 204         out_r->resize(num_samples);
 205
 206         const float *inptr = in.data();
 207         float *lptr = &(*out_l)[0];
 208         float *rptr = &(*out_r)[0];
 209         for (size_t i = 0; i < num_samples; ++i) {
 210                 *lptr++ = *inptr++;
 211                 *rptr++ = *inptr++;
 212         }
 213 }
 214
 215 // Returns length of a frame with the given format, in TIMEBASE units.
 216 int64_t find_frame_length(uint16_t video_format)
 217 {
 218         if (video_format == 0x0800) {
 219                 // No video signal. These green pseudo-frames seem to come at about 30.13 Hz.
 220                 // It's a strange thing, but what can you do.
 221                 return TIMEBASE * 100 / 3013;
 222         }
 223         if ((video_format & 0xe800) != 0xe800) {
 224                 printf("Video format 0x%04x does not appear to be a video format. Assuming 60 Hz.\n",
 225                         video_format);
 226                 return TIMEBASE / 60;
 227         }
 228
 229         // 0x8 seems to be a flag about availability of deep color on the input,
 230         // except when it's not (e.g. it's the only difference between NTSC 23.98
 231         // and PAL). Rather confusing. But we clear it here nevertheless, because
 232         // usually it doesn't mean anything.
 233         //
 234         // We don't really handle interlaced formats at all yet.
 235         uint16_t normalized_video_format = video_format & ~0xe808;
 236         if (normalized_video_format == 0x0143) {         // 720p50.
 237                 return TIMEBASE / 50;
 238         } else if (normalized_video_format == 0x0103) {  // 720p60.
 239                 return TIMEBASE / 60;
 240         } else if (normalized_video_format == 0x0121) {  // 720p59.94.
 241                 return TIMEBASE * 1001 / 60000;
 242         } else if (normalized_video_format == 0x01c3 ||  // 1080p30.
 243                    normalized_video_format == 0x0003) {  // 1080i60.
 244                 return TIMEBASE / 30;
 245         } else if (normalized_video_format == 0x01e1 ||  // 1080p29.97.
 246                    normalized_video_format == 0x0021 ||  // 1080i59.94.
 247                    video_format == 0xe901 ||             // NTSC (480i59.94, I suppose).
 248                    video_format == 0xe9c1 ||             // Ditto.
 249                    video_format == 0xe801) {             // Ditto.
 250                 return TIMEBASE * 1001 / 30000;
 251         } else if (normalized_video_format == 0x0063 ||  // 1080p25.
 252                    normalized_video_format == 0x0043 ||  // 1080i50.
 253                    video_format == 0xe909) {             // PAL (576i50, I suppose).
 254                 return TIMEBASE / 25;
 255         } else if (normalized_video_format == 0x008e) {  // 1080p24.
 256                 return TIMEBASE / 24;
 257         } else if (normalized_video_format == 0x00a1) {  // 1080p23.98.
 258                 return TIMEBASE * 1001 / 24000;
 259                 return TIMEBASE / 25;
 260         } else {
 261                 printf("Unknown video format 0x%04x. Assuming 60 Hz.\n", video_format);
 262                 return TIMEBASE / 60;
 263         }
 264 }
 265
 266 }  // namespace
 267
 268 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 269                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 270                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 271 {
 272         CaptureCard *card = &cards[card_index];
 273
 274         int64_t frame_length = find_frame_length(video_format);
 275
 276         if (audio_frame.len - audio_offset > 30000) {
 277                 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",
 278                         card_index, int(audio_frame.len), int(audio_offset),
 279                         timecode, int(video_frame.len), int(video_offset), video_format);
 280                 if (video_frame.owner) {
 281                         video_frame.owner->release_frame(video_frame);
 282                 }
 283                 if (audio_frame.owner) {
 284                         audio_frame.owner->release_frame(audio_frame);
 285                 }
 286                 return;
 287         }
 288
 289         int64_t local_pts = card->next_local_pts;
 290         int dropped_frames = 0;
 291         if (card->last_timecode != -1) {
 292                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 293         }
 294         card->last_timecode = timecode;
 295
 296         // Convert the audio to stereo fp32 and add it.
 297         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
 298         vector<float> audio;
 299         audio.resize(num_samples * 2);
 300         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 301
 302         // Add the audio.
 303         {
 304                 unique_lock<mutex> lock(card->audio_mutex);
 305
 306                 if (dropped_frames > MAX_FPS * 2) {
 307                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
 308                                 card_index);
 309                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 310                 } else if (dropped_frames > 0) {
 311                         // Insert silence as needed. (The number of samples could be nonintegral,
 312                         // but resampling will save us then.)
 313                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 314                                 card_index, dropped_frames, timecode);
 315                         vector<float> silence;
 316                         silence.resize((OUTPUT_FREQUENCY * frame_length / TIMEBASE) * 2);
 317                         for (int i = 0; i < dropped_frames; ++i) {
 318                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence.size() / 2);
 319                                 // Note that if the format changed in the meantime, we have
 320                                 // no way of detecting that; we just have to assume the frame length
 321                                 // is always the same.
 322                                 local_pts += frame_length;
 323                         }
 324                 }
 325                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 326                 card->next_local_pts = local_pts + frame_length;
 327         }
 328
 329         // Done with the audio, so release it.
 330         if (audio_frame.owner) {
 331                 audio_frame.owner->release_frame(audio_frame);
 332         }
 333
 334         {
 335                 // Wait until the previous frame was consumed.
 336                 unique_lock<mutex> lock(bmusb_mutex);
 337                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 338                 if (card->should_quit) return;
 339         }
 340
 341         if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
 342                 if (video_frame.len != 0) {
 343                         printf("Card %d: Dropping video frame with wrong length (%ld)\n",
 344                                 card_index, video_frame.len - video_offset);
 345                 }
 346                 if (video_frame.owner) {
 347                         video_frame.owner->release_frame(video_frame);
 348                 }
 349
 350                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 351                 // so that pts can go up accordingly.
 352                 {
 353                         unique_lock<mutex> lock(bmusb_mutex);
 354                         card->new_data_ready = true;
 355                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 356                         card->new_frame_length = frame_length;
 357                         card->new_data_ready_fence = nullptr;
 358                         card->dropped_frames = dropped_frames;
 359                         card->new_data_ready_changed.notify_all();
 360                 }
 361                 return;
 362         }
 363
 364         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
 365         GLuint pbo = userdata->pbo;
 366         check_error();
 367         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 368         check_error();
 369         glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
 370         check_error();
 371         //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 372         //check_error();
 373
 374         // Upload the textures.
 375         glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
 376         check_error();
 377         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));
 378         check_error();
 379         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
 380         check_error();
 381         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
 382         check_error();
 383         glBindTexture(GL_TEXTURE_2D, 0);
 384         check_error();
 385         GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 386         check_error();
 387         assert(fence != nullptr);
 388
 389         {
 390                 unique_lock<mutex> lock(bmusb_mutex);
 391                 card->new_data_ready = true;
 392                 card->new_frame = RefCountedFrame(video_frame);
 393                 card->new_frame_length = frame_length;
 394                 card->new_data_ready_fence = fence;
 395                 card->dropped_frames = dropped_frames;
 396                 card->new_data_ready_changed.notify_all();
 397         }
 398 }
 399
 400 void Mixer::thread_func()
 401 {
 402         eglBindAPI(EGL_OPENGL_API);
 403         QOpenGLContext *context = create_context(mixer_surface);
 404         if (!make_current(context, mixer_surface)) {
 405                 printf("oops\n");
 406                 exit(1);
 407         }
 408
 409         struct timespec start, now;
 410         clock_gettime(CLOCK_MONOTONIC, &start);
 411
 412         int frame = 0;
 413         int dropped_frames = 0;
 414
 415         while (!should_quit) {
 416                 CaptureCard card_copy[MAX_CARDS];
 417                 int num_samples[MAX_CARDS];
 418
 419                 {
 420                         unique_lock<mutex> lock(bmusb_mutex);
 421
 422                         // The first card is the master timer, so wait for it to have a new frame.
 423                         // TODO: Make configurable, and with a timeout.
 424                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 425
 426                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 427                                 CaptureCard *card = &cards[card_index];
 428                                 card_copy[card_index].usb = card->usb;
 429                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 430                                 card_copy[card_index].new_frame = card->new_frame;
 431                                 card_copy[card_index].new_frame_length = card->new_frame_length;
 432                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 433                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 434                                 card->new_data_ready = false;
 435                                 card->new_data_ready_changed.notify_all();
 436
 437                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
 438                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 439                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 440                         }
 441                 }
 442
 443                 // Resample the audio as needed, including from previously dropped frames.
 444                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 445                         {
 446                                 // Signal to the audio thread to process this frame.
 447                                 unique_lock<mutex> lock(audio_mutex);
 448                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
 449                                 audio_task_queue_changed.notify_one();
 450                         }
 451                         if (frame_num != card_copy[0].dropped_frames) {
 452                                 // For dropped frames, increase the pts. Note that if the format changed
 453                                 // in the meantime, we have no way of detecting that; we just have to
 454                                 // assume the frame length is always the same.
 455                                 ++dropped_frames;
 456                                 pts_int += card_copy[0].new_frame_length;
 457                         }
 458                 }
 459
 460                 if (audio_level_callback != nullptr) {
 461                         double loudness_s = r128.loudness_S();
 462                         double loudness_i = r128.integrated();
 463                         double loudness_range_low = r128.range_min();
 464                         double loudness_range_high = r128.range_max();
 465
 466                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 467                                              loudness_i, loudness_range_low, loudness_range_high,
 468                                              last_gain_staging_db);
 469                 }
 470
 471                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 472                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 473                                 ++card_copy[card_index].dropped_frames;
 474                         }
 475                         if (card_copy[card_index].dropped_frames > 0) {
 476                                 printf("Card %u dropped %d frames before this\n",
 477                                         card_index, int(card_copy[card_index].dropped_frames));
 478                         }
 479                 }
 480
 481                 // If the first card is reporting a corrupted or otherwise dropped frame,
 482                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 483                 if (card_copy[0].new_frame->len == 0) {
 484                         ++dropped_frames;
 485                         pts_int += card_copy[0].new_frame_length;
 486                         continue;
 487                 }
 488
 489                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 490                         CaptureCard *card = &card_copy[card_index];
 491                         if (!card->new_data_ready || card->new_frame->len == 0)
 492                                 continue;
 493
 494                         assert(card->new_frame != nullptr);
 495                         bmusb_current_rendering_frame[card_index] = card->new_frame;
 496                         check_error();
 497
 498                         // The new texture might still be uploaded,
 499                         // tell the GPU to wait until it's there.
 500                         if (card->new_data_ready_fence) {
 501                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 502                                 check_error();
 503                                 glDeleteSync(card->new_data_ready_fence);
 504                                 check_error();
 505                         }
 506                         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
 507                         theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
 508                 }
 509
 510                 // Get the main chain from the theme, and set its state immediately.
 511                 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
 512                 EffectChain *chain = theme_main_chain.first;
 513                 theme_main_chain.second();
 514
 515                 GLuint y_tex, cbcr_tex;
 516                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 517                 assert(got_frame);
 518
 519                 // Render main chain.
 520                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 521                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 522                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 523                 check_error();
 524                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 525                 resource_pool->release_fbo(fbo);
 526
 527                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 528                 resource_pool->release_2d_texture(cbcr_full_tex);
 529
 530                 // Set the right state for rgba_tex.
 531                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 532                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 533                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 534                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 535                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 536
 537                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 538                 check_error();
 539
 540                 // Make sure the H.264 gets a reference to all the
 541                 // input frames needed, so that they are not released back
 542                 // until the rendering is done.
 543                 vector<RefCountedFrame> input_frames;
 544                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 545                         input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 546                 }
 547                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 548                 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
 549                 ++frame;
 550                 pts_int += card_copy[0].new_frame_length;
 551
 552                 // The live frame just shows the RGBA texture we just rendered.
 553                 // It owns rgba_tex now.
 554                 DisplayFrame live_frame;
 555                 live_frame.chain = display_chain.get();
 556                 live_frame.setup_chain = [this, rgba_tex]{
 557                         display_input->set_texture_num(rgba_tex);
 558                 };
 559                 live_frame.ready_fence = fence;
 560                 live_frame.input_frames = {};
 561                 live_frame.temp_textures = { rgba_tex };
 562                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 563
 564                 // Set up preview and any additional channels.
 565                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 566                         DisplayFrame display_frame;
 567                         pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT);  // FIXME: dimensions
 568                         display_frame.chain = chain.first;
 569                         display_frame.setup_chain = chain.second;
 570                         display_frame.ready_fence = fence;
 571
 572                         // FIXME: possible to do better?
 573                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 574                                 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 575                         }
 576                         display_frame.temp_textures = {};
 577                         output_channel[i].output_frame(display_frame);
 578                 }
 579
 580                 clock_gettime(CLOCK_MONOTONIC, &now);
 581                 double elapsed = now.tv_sec - start.tv_sec +
 582                         1e-9 * (now.tv_nsec - start.tv_nsec);
 583                 if (frame % 100 == 0) {
 584                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 585                                 frame, dropped_frames, elapsed, frame / elapsed,
 586                                 1e3 * elapsed / frame);
 587                 //      chain->print_phase_timing();
 588                 }
 589
 590 #if 0
 591                 // Reset every 100 frames, so that local variations in frame times
 592                 // (especially for the first few frames, when the shaders are
 593                 // compiled etc.) don't make it hard to measure for the entire
 594                 // remaining duration of the program.
 595                 if (frame == 10000) {
 596                         frame = 0;
 597                         start = now;
 598                 }
 599 #endif
 600                 check_error();
 601         }
 602
 603         resource_pool->clean_context();
 604 }
 605
 606 void Mixer::audio_thread_func()
 607 {
 608         while (!should_quit) {
 609                 AudioTask task;
 610
 611                 {
 612                         unique_lock<mutex> lock(audio_mutex);
 613                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 614                         task = audio_task_queue.front();
 615                         audio_task_queue.pop();
 616                 }
 617
 618                 process_audio_one_frame(task.pts_int, task.num_samples);
 619         }
 620 }
 621
 622 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 623 {
 624         vector<float> samples_card;
 625         vector<float> samples_out;
 626         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 627                 samples_card.resize(num_samples * 2);
 628                 {
 629                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 630                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 631                                 printf("Card %d reported previous underrun.\n", card_index);
 632                         }
 633                 }
 634                 // TODO: Allow using audio from the other card(s) as well.
 635                 if (card_index == 0) {
 636                         samples_out = move(samples_card);
 637                 }
 638         }
 639
 640         // Cut away everything under 120 Hz (or whatever the cutoff is);
 641         // we don't need it for voice, and it will reduce headroom
 642         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 643         // should be dampened.)
 644         locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 645
 646         // Apply a level compressor to get the general level right.
 647         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 648         // (or more precisely, near it, since we don't use infinite ratio),
 649         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 650         // entirely arbitrary, but from practical tests with speech, it seems to
 651         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 652         float ref_level_dbfs = -14.0f;
 653         {
 654                 float threshold = 0.01f;   // -40 dBFS.
 655                 float ratio = 20.0f;
 656                 float attack_time = 0.5f;
 657                 float release_time = 20.0f;
 658                 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 659                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 660                 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 661         }
 662
 663 #if 0
 664         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 665                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 666                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 667                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 668 #endif
 669
 670 //      float limiter_att, compressor_att;
 671
 672         // The real compressor.
 673         if (compressor_enabled) {
 674                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 675                 float ratio = 20.0f;
 676                 float attack_time = 0.005f;
 677                 float release_time = 0.040f;
 678                 float makeup_gain = 2.0f;  // +6 dB.
 679                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 680 //              compressor_att = compressor.get_attenuation();
 681         }
 682
 683         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 684         // Note that since ratio is not infinite, we could go slightly higher than this.
 685         if (limiter_enabled) {
 686                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 687                 float ratio = 30.0f;
 688                 float attack_time = 0.0f;  // Instant.
 689                 float release_time = 0.020f;
 690                 float makeup_gain = 1.0f;  // 0 dB.
 691                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 692 //              limiter_att = limiter.get_attenuation();
 693         }
 694
 695 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 696
 697         // Upsample 4x to find interpolated peak.
 698         peak_resampler.inp_data = samples_out.data();
 699         peak_resampler.inp_count = samples_out.size() / 2;
 700
 701         vector<float> interpolated_samples_out;
 702         interpolated_samples_out.resize(samples_out.size());
 703         while (peak_resampler.inp_count > 0) {  // About four iterations.
 704                 peak_resampler.out_data = &interpolated_samples_out[0];
 705                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 706                 peak_resampler.process();
 707                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 708                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 709         }
 710
 711         // Find R128 levels.
 712         vector<float> left, right;
 713         deinterleave_samples(samples_out, &left, &right);
 714         float *ptrs[] = { left.data(), right.data() };
 715         r128.process(left.size(), ptrs);
 716
 717         // Send the samples to the sound card.
 718         if (alsa) {
 719                 alsa->write(samples_out);
 720         }
 721
 722         // And finally add them to the output.
 723         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 724 }
 725
 726 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 727 {
 728         GLuint vao;
 729         glGenVertexArrays(1, &vao);
 730         check_error();
 731
 732         float vertices[] = {
 733                 0.0f, 2.0f,
 734                 0.0f, 0.0f,
 735                 2.0f, 0.0f
 736         };
 737
 738         glBindVertexArray(vao);
 739         check_error();
 740
 741         // Extract Cb/Cr.
 742         GLuint fbo = resource_pool->create_fbo(dst_tex);
 743         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 744         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 745         check_error();
 746
 747         glUseProgram(cbcr_program_num);
 748         check_error();
 749
 750         glActiveTexture(GL_TEXTURE0);
 751         check_error();
 752         glBindTexture(GL_TEXTURE_2D, src_tex);
 753         check_error();
 754         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 755         check_error();
 756         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 757         check_error();
 758         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 759         check_error();
 760
 761         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 762         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 763
 764         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 765         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 766
 767         glDrawArrays(GL_TRIANGLES, 0, 3);
 768         check_error();
 769
 770         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 771         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 772
 773         glUseProgram(0);
 774         check_error();
 775
 776         resource_pool->release_fbo(fbo);
 777         glDeleteVertexArrays(1, &vao);
 778 }
 779
 780 void Mixer::release_display_frame(DisplayFrame *frame)
 781 {
 782         for (GLuint texnum : frame->temp_textures) {
 783                 resource_pool->release_2d_texture(texnum);
 784         }
 785         frame->temp_textures.clear();
 786         frame->ready_fence.reset();
 787         frame->input_frames.clear();
 788 }
 789
 790 void Mixer::start()
 791 {
 792         mixer_thread = thread(&Mixer::thread_func, this);
 793         audio_thread = thread(&Mixer::audio_thread_func, this);
 794 }
 795
 796 void Mixer::quit()
 797 {
 798         should_quit = true;
 799         mixer_thread.join();
 800         audio_thread.join();
 801 }
 802
 803 void Mixer::transition_clicked(int transition_num)
 804 {
 805         theme->transition_clicked(transition_num, pts());
 806 }
 807
 808 void Mixer::channel_clicked(int preview_num)
 809 {
 810         theme->channel_clicked(preview_num);
 811 }
 812
 813 void Mixer::reset_meters()
 814 {
 815         peak_resampler.reset();
 816         peak = 0.0f;
 817         r128.reset();
 818         r128.integr_start();
 819 }
 820
 821 Mixer::OutputChannel::~OutputChannel()
 822 {
 823         if (has_current_frame) {
 824                 parent->release_display_frame(&current_frame);
 825         }
 826         if (has_ready_frame) {
 827                 parent->release_display_frame(&ready_frame);
 828         }
 829 }
 830
 831 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 832 {
 833         // Store this frame for display. Remove the ready frame if any
 834         // (it was seemingly never used).
 835         {
 836                 unique_lock<mutex> lock(frame_mutex);
 837                 if (has_ready_frame) {
 838                         parent->release_display_frame(&ready_frame);
 839                 }
 840                 ready_frame = frame;
 841                 has_ready_frame = true;
 842         }
 843
 844         if (has_new_frame_ready_callback) {
 845                 new_frame_ready_callback();
 846         }
 847 }
 848
 849 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 850 {
 851         unique_lock<mutex> lock(frame_mutex);
 852         if (!has_current_frame && !has_ready_frame) {
 853                 return false;
 854         }
 855
 856         if (has_current_frame && has_ready_frame) {
 857                 // We have a new ready frame. Toss the current one.
 858                 parent->release_display_frame(&current_frame);
 859                 has_current_frame = false;
 860         }
 861         if (has_ready_frame) {
 862                 assert(!has_current_frame);
 863                 current_frame = ready_frame;
 864                 ready_frame.ready_fence.reset();  // Drop the refcount.
 865                 ready_frame.input_frames.clear();  // Drop the refcounts.
 866                 has_current_frame = true;
 867                 has_ready_frame = false;
 868         }
 869
 870         *frame = current_frame;
 871         return true;
 872 }
 873
 874 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 875 {
 876         new_frame_ready_callback = callback;
 877         has_new_frame_ready_callback = true;
 878 }