]> git.sesse.net Git - nageru/blob - mixer.cpp
Send the audio format explicitly down, not just the ID.
[nageru] / mixer.cpp
1 #undef Success
2
3 #include "mixer.h"
4
5 #include <assert.h>
6 #include <epoxy/egl.h>
7 #include <movit/effect_chain.h>
8 #include <movit/effect_util.h>
9 #include <movit/flat_input.h>
10 #include <movit/image_format.h>
11 #include <movit/init.h>
12 #include <movit/resource_pool.h>
13 #include <movit/util.h>
14 #include <stdint.h>
15 #include <stdio.h>
16 #include <stdlib.h>
17 #include <sys/time.h>
18 #include <time.h>
19 #include <algorithm>
20 #include <cmath>
21 #include <condition_variable>
22 #include <cstddef>
23 #include <memory>
24 #include <mutex>
25 #include <string>
26 #include <thread>
27 #include <utility>
28 #include <vector>
29 #include <arpa/inet.h>
30
31 #include "bmusb/bmusb.h"
32 #include "context.h"
33 #include "decklink_capture.h"
34 #include "defs.h"
35 #include "flags.h"
36 #include "h264encode.h"
37 #include "pbo_frame_allocator.h"
38 #include "ref_counted_gl_sync.h"
39 #include "timebase.h"
40
41 class QOpenGLContext;
42
43 using namespace movit;
44 using namespace std;
45 using namespace std::placeholders;
46
47 Mixer *global_mixer = nullptr;
48
49 namespace {
50
51 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
52 {
53         for (size_t i = 0; i < num_samples; ++i) {
54                 for (size_t j = 0; j < out_channels; ++j) {
55                         uint32_t s1 = *src++;
56                         uint32_t s2 = *src++;
57                         uint32_t s3 = *src++;
58                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
59                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
60                 }
61                 src += 3 * (in_channels - out_channels);
62         }
63 }
64
65 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
66 {
67         if (interlaced) {
68                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
69                         input_state->buffered_frames[card_index][frame_num] =
70                                 input_state->buffered_frames[card_index][frame_num - 1];
71                 }
72                 input_state->buffered_frames[card_index][0] = { frame, field_num };
73         } else {
74                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
75                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
76                 }
77         }
78 }
79
80 string generate_local_dump_filename(int frame)
81 {
82         time_t now = time(NULL);
83         tm now_tm;
84         localtime_r(&now, &now_tm);
85
86         char timestamp[256];
87         strftime(timestamp, sizeof(timestamp), "%F-%T%z", &now_tm);
88
89         // Use the frame number to disambiguate between two cuts starting
90         // on the same second.
91         char filename[256];
92         snprintf(filename, sizeof(filename), "%s%s-f%02d%s",
93                 LOCAL_DUMP_PREFIX, timestamp, frame % 100, LOCAL_DUMP_SUFFIX);
94         return filename;
95 }
96
97 }  // namespace
98
99 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
100         : httpd(WIDTH, HEIGHT),
101           num_cards(num_cards),
102           mixer_surface(create_surface(format)),
103           h264_encoder_surface(create_surface(format)),
104           correlation(OUTPUT_FREQUENCY),
105           level_compressor(OUTPUT_FREQUENCY),
106           limiter(OUTPUT_FREQUENCY),
107           compressor(OUTPUT_FREQUENCY)
108 {
109         httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str());
110         httpd.start(9095);
111
112         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
113         check_error();
114
115         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
116         // will be halved when sampling them, and we need to compensate here.
117         movit_texel_subpixel_precision /= 2.0;
118
119         resource_pool.reset(new ResourcePool);
120         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
121         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
122                 output_channel[i].parent = this;
123         }
124
125         ImageFormat inout_format;
126         inout_format.color_space = COLORSPACE_sRGB;
127         inout_format.gamma_curve = GAMMA_sRGB;
128
129         // Display chain; shows the live output produced by the main chain (its RGBA version).
130         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
131         check_error();
132         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
133         display_chain->add_input(display_input);
134         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
135         display_chain->set_dither_bits(0);  // Don't bother.
136         display_chain->finalize();
137
138         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
139
140         // First try initializing the PCI devices, then USB, until we have the desired number of cards.
141         unsigned num_pci_devices = 0, num_usb_devices = 0;
142         unsigned card_index = 0;
143
144         IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
145         if (decklink_iterator != nullptr) {
146                 for ( ; card_index < num_cards; ++card_index) {
147                         IDeckLink *decklink;
148                         if (decklink_iterator->Next(&decklink) != S_OK) {
149                                 break;
150                         }
151
152                         configure_card(card_index, format, new DeckLinkCapture(decklink, card_index));
153                         ++num_pci_devices;
154                 }
155                 decklink_iterator->Release();
156                 fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
157         } else {
158                 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
159         }
160         for ( ; card_index < num_cards; ++card_index) {
161                 configure_card(card_index, format, new BMUSBCapture(card_index - num_pci_devices));
162                 ++num_usb_devices;
163         }
164
165         if (num_usb_devices > 0) {
166                 BMUSBCapture::start_bm_thread();
167         }
168
169         for (card_index = 0; card_index < num_cards; ++card_index) {
170                 cards[card_index].capture->start_bm_capture();
171         }
172
173         // Set up stuff for NV12 conversion.
174
175         // Cb/Cr shader.
176         string cbcr_vert_shader =
177                 "#version 130 \n"
178                 " \n"
179                 "in vec2 position; \n"
180                 "in vec2 texcoord; \n"
181                 "out vec2 tc0; \n"
182                 "uniform vec2 foo_chroma_offset_0; \n"
183                 " \n"
184                 "void main() \n"
185                 "{ \n"
186                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
187                 "    // \n"
188                 "    //   2.000  0.000  0.000 -1.000 \n"
189                 "    //   0.000  2.000  0.000 -1.000 \n"
190                 "    //   0.000  0.000 -2.000 -1.000 \n"
191                 "    //   0.000  0.000  0.000  1.000 \n"
192                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
193                 "    vec2 flipped_tc = texcoord; \n"
194                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
195                 "} \n";
196         string cbcr_frag_shader =
197                 "#version 130 \n"
198                 "in vec2 tc0; \n"
199                 "uniform sampler2D cbcr_tex; \n"
200                 "out vec4 FragColor; \n"
201                 "void main() { \n"
202                 "    FragColor = texture(cbcr_tex, tc0); \n"
203                 "} \n";
204         vector<string> frag_shader_outputs;
205         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
206
207         float vertices[] = {
208                 0.0f, 2.0f,
209                 0.0f, 0.0f,
210                 2.0f, 0.0f
211         };
212         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
213         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
214         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
215
216         r128.init(2, OUTPUT_FREQUENCY);
217         r128.integr_start();
218
219         locut.init(FILTER_HPF, 2);
220
221         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
222         // and there's a limit to how important the peak meter is.
223         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
224
225         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
226 }
227
228 Mixer::~Mixer()
229 {
230         resource_pool->release_glsl_program(cbcr_program_num);
231         glDeleteBuffers(1, &cbcr_vbo);
232         BMUSBCapture::stop_bm_thread();
233
234         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
235                 {
236                         unique_lock<mutex> lock(bmusb_mutex);
237                         cards[card_index].should_quit = true;  // Unblock thread.
238                         cards[card_index].new_data_ready_changed.notify_all();
239                 }
240                 cards[card_index].capture->stop_dequeue_thread();
241         }
242
243         h264_encoder.reset(nullptr);
244 }
245
246 void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture)
247 {
248         printf("Configuring card %d...\n", card_index);
249
250         CaptureCard *card = &cards[card_index];
251         card->capture = capture;
252         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
253         card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
254         card->capture->set_video_frame_allocator(card->frame_allocator.get());
255         card->surface = create_surface(format);
256         card->capture->set_dequeue_thread_callbacks(
257                 [card]{
258                         eglBindAPI(EGL_OPENGL_API);
259                         card->context = create_context(card->surface);
260                         if (!make_current(card->context, card->surface)) {
261                                 printf("failed to create bmusb context\n");
262                                 exit(1);
263                         }
264                 },
265                 [this]{
266                         resource_pool->clean_context();
267                 });
268         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
269         card->capture->configure_card();
270 }
271
272
273 namespace {
274
275 int unwrap_timecode(uint16_t current_wrapped, int last)
276 {
277         uint16_t last_wrapped = last & 0xffff;
278         if (current_wrapped > last_wrapped) {
279                 return (last & ~0xffff) | current_wrapped;
280         } else {
281                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
282         }
283 }
284
285 float find_peak(const float *samples, size_t num_samples)
286 {
287         float m = fabs(samples[0]);
288         for (size_t i = 1; i < num_samples; ++i) {
289                 m = max(m, fabs(samples[i]));
290         }
291         return m;
292 }
293
294 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
295 {
296         size_t num_samples = in.size() / 2;
297         out_l->resize(num_samples);
298         out_r->resize(num_samples);
299
300         const float *inptr = in.data();
301         float *lptr = &(*out_l)[0];
302         float *rptr = &(*out_r)[0];
303         for (size_t i = 0; i < num_samples; ++i) {
304                 *lptr++ = *inptr++;
305                 *rptr++ = *inptr++;
306         }
307 }
308
309 }  // namespace
310
311 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
312                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
313                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
314 {
315         CaptureCard *card = &cards[card_index];
316
317         int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom;
318
319         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
320         if (num_samples > OUTPUT_FREQUENCY / 10) {
321                 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",
322                         card_index, int(audio_frame.len), int(audio_offset),
323                         timecode, int(video_frame.len), int(video_offset), video_format.id);
324                 if (video_frame.owner) {
325                         video_frame.owner->release_frame(video_frame);
326                 }
327                 if (audio_frame.owner) {
328                         audio_frame.owner->release_frame(audio_frame);
329                 }
330                 return;
331         }
332
333         int64_t local_pts = card->next_local_pts;
334         int dropped_frames = 0;
335         if (card->last_timecode != -1) {
336                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
337         }
338
339         // Convert the audio to stereo fp32 and add it.
340         vector<float> audio;
341         audio.resize(num_samples * 2);
342         switch (audio_format.bits_per_sample) {
343         case 24:
344                 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
345                 break;
346         default:
347                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
348                 assert(false);
349         }
350
351         // Add the audio.
352         {
353                 unique_lock<mutex> lock(card->audio_mutex);
354
355                 // Number of samples per frame if we need to insert silence.
356                 // (Could be nonintegral, but resampling will save us then.)
357                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
358
359                 if (dropped_frames > MAX_FPS * 2) {
360                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
361                                 card_index, card->last_timecode, timecode);
362                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
363                         dropped_frames = 0;
364                 } else if (dropped_frames > 0) {
365                         // Insert silence as needed.
366                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
367                                 card_index, dropped_frames, timecode);
368                         vector<float> silence(silence_samples * 2, 0.0f);
369                         for (int i = 0; i < dropped_frames; ++i) {
370                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
371                                 // Note that if the format changed in the meantime, we have
372                                 // no way of detecting that; we just have to assume the frame length
373                                 // is always the same.
374                                 local_pts += frame_length;
375                         }
376                 }
377                 if (num_samples == 0) {
378                         audio.resize(silence_samples * 2);
379                         num_samples = silence_samples;
380                 }
381                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
382                 card->next_local_pts = local_pts + frame_length;
383         }
384
385         card->last_timecode = timecode;
386
387         // Done with the audio, so release it.
388         if (audio_frame.owner) {
389                 audio_frame.owner->release_frame(audio_frame);
390         }
391
392         {
393                 // Wait until the previous frame was consumed.
394                 unique_lock<mutex> lock(bmusb_mutex);
395                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
396                 if (card->should_quit) return;
397         }
398
399         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
400         if (video_frame.len - video_offset == 0 ||
401             video_frame.len - video_offset != expected_length) {
402                 if (video_frame.len != 0) {
403                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
404                                 card_index, video_frame.len - video_offset, expected_length);
405                 }
406                 if (video_frame.owner) {
407                         video_frame.owner->release_frame(video_frame);
408                 }
409
410                 // Still send on the information that we _had_ a frame, even though it's corrupted,
411                 // so that pts can go up accordingly.
412                 {
413                         unique_lock<mutex> lock(bmusb_mutex);
414                         card->new_data_ready = true;
415                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
416                         card->new_frame_length = frame_length;
417                         card->new_frame_interlaced = false;
418                         card->new_data_ready_fence = nullptr;
419                         card->dropped_frames = dropped_frames;
420                         card->new_data_ready_changed.notify_all();
421                 }
422                 return;
423         }
424
425         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
426
427         unsigned num_fields = video_format.interlaced ? 2 : 1;
428         timespec frame_upload_start;
429         if (video_format.interlaced) {
430                 // Send the two fields along as separate frames; the other side will need to add
431                 // a deinterlacer to actually get this right.
432                 assert(video_format.height % 2 == 0);
433                 video_format.height /= 2;
434                 assert(frame_length % 2 == 0);
435                 frame_length /= 2;
436                 num_fields = 2;
437                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
438         }
439         userdata->last_interlaced = video_format.interlaced;
440         userdata->last_has_signal = video_format.has_signal;
441         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
442         userdata->last_frame_rate_den = video_format.frame_rate_den;
443         RefCountedFrame new_frame(video_frame);
444
445         // Upload the textures.
446         size_t cbcr_width = video_format.width / 2;
447         size_t cbcr_offset = video_offset / 2;
448         size_t y_offset = video_frame.size / 2 + video_offset / 2;
449
450         for (unsigned field = 0; field < num_fields; ++field) {
451                 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
452
453                 if (userdata->tex_y[field] == 0 ||
454                     userdata->tex_cbcr[field] == 0 ||
455                     video_format.width != userdata->last_width[field] ||
456                     video_format.height != userdata->last_height[field]) {
457                         // We changed resolution since last use of this texture, so we need to create
458                         // a new object. Note that this each card has its own PBOFrameAllocator,
459                         // we don't need to worry about these flip-flopping between resolutions.
460                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
461                         check_error();
462                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
463                         check_error();
464                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
465                         check_error();
466                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
467                         check_error();
468                         userdata->last_width[field] = video_format.width;
469                         userdata->last_height[field] = video_format.height;
470                 }
471
472                 GLuint pbo = userdata->pbo;
473                 check_error();
474                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
475                 check_error();
476                 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
477                 check_error();
478
479                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
480                 check_error();
481                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t)));
482                 check_error();
483                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
484                 check_error();
485                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + video_format.width * field_start_line));
486                 check_error();
487                 glBindTexture(GL_TEXTURE_2D, 0);
488                 check_error();
489                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
490                 check_error();
491                 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
492                 check_error();
493                 assert(fence != nullptr);
494
495                 if (field == 1) {
496                         // Don't upload the second field as fast as we can; wait until
497                         // the field time has approximately passed. (Otherwise, we could
498                         // get timing jitter against the other sources, and possibly also
499                         // against the video display, although the latter is not as critical.)
500                         // This requires our system clock to be reasonably close to the
501                         // video clock, but that's not an unreasonable assumption.
502                         timespec second_field_start;
503                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
504                                 frame_length * 1000000000 / TIMEBASE;
505                         second_field_start.tv_sec = frame_upload_start.tv_sec +
506                                 second_field_start.tv_nsec / 1000000000;
507                         second_field_start.tv_nsec %= 1000000000;
508
509                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
510                                                &second_field_start, nullptr) == -1 &&
511                                errno == EINTR) ;
512                 }
513
514                 {
515                         unique_lock<mutex> lock(bmusb_mutex);
516                         card->new_data_ready = true;
517                         card->new_frame = new_frame;
518                         card->new_frame_length = frame_length;
519                         card->new_frame_field = field;
520                         card->new_frame_interlaced = video_format.interlaced;
521                         card->new_data_ready_fence = fence;
522                         card->dropped_frames = dropped_frames;
523                         card->new_data_ready_changed.notify_all();
524
525                         if (field != num_fields - 1) {
526                                 // Wait until the previous frame was consumed.
527                                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
528                                 if (card->should_quit) return;
529                         }
530                 }
531         }
532 }
533
534 void Mixer::thread_func()
535 {
536         eglBindAPI(EGL_OPENGL_API);
537         QOpenGLContext *context = create_context(mixer_surface);
538         if (!make_current(context, mixer_surface)) {
539                 printf("oops\n");
540                 exit(1);
541         }
542
543         struct timespec start, now;
544         clock_gettime(CLOCK_MONOTONIC, &start);
545
546         int frame = 0;
547         int stats_dropped_frames = 0;
548
549         while (!should_quit) {
550                 CaptureCard card_copy[MAX_CARDS];
551                 int num_samples[MAX_CARDS];
552
553                 {
554                         unique_lock<mutex> lock(bmusb_mutex);
555
556                         // The first card is the master timer, so wait for it to have a new frame.
557                         // TODO: Make configurable, and with a timeout.
558                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
559
560                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
561                                 CaptureCard *card = &cards[card_index];
562                                 card_copy[card_index].new_data_ready = card->new_data_ready;
563                                 card_copy[card_index].new_frame = card->new_frame;
564                                 card_copy[card_index].new_frame_length = card->new_frame_length;
565                                 card_copy[card_index].new_frame_field = card->new_frame_field;
566                                 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
567                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
568                                 card_copy[card_index].dropped_frames = card->dropped_frames;
569                                 card->new_data_ready = false;
570                                 card->new_data_ready_changed.notify_all();
571
572                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
573                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
574                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
575                                 assert(num_samples[card_index] >= 0);
576                         }
577                 }
578
579                 // Resample the audio as needed, including from previously dropped frames.
580                 assert(num_cards > 0);
581                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
582                         {
583                                 // Signal to the audio thread to process this frame.
584                                 unique_lock<mutex> lock(audio_mutex);
585                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
586                                 audio_task_queue_changed.notify_one();
587                         }
588                         if (frame_num != card_copy[0].dropped_frames) {
589                                 // For dropped frames, increase the pts. Note that if the format changed
590                                 // in the meantime, we have no way of detecting that; we just have to
591                                 // assume the frame length is always the same.
592                                 ++stats_dropped_frames;
593                                 pts_int += card_copy[0].new_frame_length;
594                         }
595                 }
596
597                 if (audio_level_callback != nullptr) {
598                         unique_lock<mutex> lock(compressor_mutex);
599                         double loudness_s = r128.loudness_S();
600                         double loudness_i = r128.integrated();
601                         double loudness_range_low = r128.range_min();
602                         double loudness_range_high = r128.range_max();
603
604                         audio_level_callback(loudness_s, 20.0 * log10(peak),
605                                              loudness_i, loudness_range_low, loudness_range_high,
606                                              gain_staging_db, 20.0 * log10(final_makeup_gain),
607                                              correlation.get_correlation());
608                 }
609
610                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
611                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
612                                 ++card_copy[card_index].dropped_frames;
613                         }
614                         if (card_copy[card_index].dropped_frames > 0) {
615                                 printf("Card %u dropped %d frames before this\n",
616                                         card_index, int(card_copy[card_index].dropped_frames));
617                         }
618                 }
619
620                 // If the first card is reporting a corrupted or otherwise dropped frame,
621                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
622                 if (card_copy[0].new_frame->len == 0) {
623                         ++stats_dropped_frames;
624                         pts_int += card_copy[0].new_frame_length;
625                         continue;
626                 }
627
628                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
629                         CaptureCard *card = &card_copy[card_index];
630                         if (!card->new_data_ready || card->new_frame->len == 0)
631                                 continue;
632
633                         assert(card->new_frame != nullptr);
634                         insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
635                         check_error();
636
637                         // The new texture might still be uploaded,
638                         // tell the GPU to wait until it's there.
639                         if (card->new_data_ready_fence) {
640                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
641                                 check_error();
642                                 glDeleteSync(card->new_data_ready_fence);
643                                 check_error();
644                         }
645                 }
646
647                 // Get the main chain from the theme, and set its state immediately.
648                 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
649                 EffectChain *chain = theme_main_chain.chain;
650                 theme_main_chain.setup_chain();
651                 //theme_main_chain.chain->enable_phase_timing(true);
652
653                 GLuint y_tex, cbcr_tex;
654                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
655                 assert(got_frame);
656
657                 // Render main chain.
658                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
659                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
660                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
661                 check_error();
662                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
663                 resource_pool->release_fbo(fbo);
664
665                 subsample_chroma(cbcr_full_tex, cbcr_tex);
666                 resource_pool->release_2d_texture(cbcr_full_tex);
667
668                 // Set the right state for rgba_tex.
669                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
670                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
671                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
672                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
673                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
674
675                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
676                 check_error();
677
678                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
679                 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
680                 ++frame;
681                 pts_int += card_copy[0].new_frame_length;
682
683                 // The live frame just shows the RGBA texture we just rendered.
684                 // It owns rgba_tex now.
685                 DisplayFrame live_frame;
686                 live_frame.chain = display_chain.get();
687                 live_frame.setup_chain = [this, rgba_tex]{
688                         display_input->set_texture_num(rgba_tex);
689                 };
690                 live_frame.ready_fence = fence;
691                 live_frame.input_frames = {};
692                 live_frame.temp_textures = { rgba_tex };
693                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
694
695                 // Set up preview and any additional channels.
696                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
697                         DisplayFrame display_frame;
698                         Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
699                         display_frame.chain = chain.chain;
700                         display_frame.setup_chain = chain.setup_chain;
701                         display_frame.ready_fence = fence;
702                         display_frame.input_frames = chain.input_frames;
703                         display_frame.temp_textures = {};
704                         output_channel[i].output_frame(display_frame);
705                 }
706
707                 clock_gettime(CLOCK_MONOTONIC, &now);
708                 double elapsed = now.tv_sec - start.tv_sec +
709                         1e-9 * (now.tv_nsec - start.tv_nsec);
710                 if (frame % 100 == 0) {
711                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
712                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
713                                 1e3 * elapsed / frame);
714                 //      chain->print_phase_timing();
715                 }
716
717                 if (should_cut.exchange(false)) {  // Test and clear.
718                         string filename = generate_local_dump_filename(frame);
719                         printf("Starting new recording: %s\n", filename.c_str());
720                         h264_encoder->shutdown();
721                         httpd.close_output_file();
722                         httpd.open_output_file(filename.c_str());
723                         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
724                 }
725
726 #if 0
727                 // Reset every 100 frames, so that local variations in frame times
728                 // (especially for the first few frames, when the shaders are
729                 // compiled etc.) don't make it hard to measure for the entire
730                 // remaining duration of the program.
731                 if (frame == 10000) {
732                         frame = 0;
733                         start = now;
734                 }
735 #endif
736                 check_error();
737         }
738
739         resource_pool->clean_context();
740 }
741
742 void Mixer::audio_thread_func()
743 {
744         while (!should_quit) {
745                 AudioTask task;
746
747                 {
748                         unique_lock<mutex> lock(audio_mutex);
749                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
750                         task = audio_task_queue.front();
751                         audio_task_queue.pop();
752                 }
753
754                 process_audio_one_frame(task.pts_int, task.num_samples);
755         }
756 }
757
758 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
759 {
760         vector<float> samples_card;
761         vector<float> samples_out;
762
763         // TODO: Allow mixing audio from several sources.
764         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
765         assert(selected_audio_card < num_cards);
766
767         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
768                 samples_card.resize(num_samples * 2);
769                 {
770                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
771                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
772                                 printf("Card %d reported previous underrun.\n", card_index);
773                         }
774                 }
775                 if (card_index == selected_audio_card) {
776                         samples_out = move(samples_card);
777                 }
778         }
779
780         // Cut away everything under 120 Hz (or whatever the cutoff is);
781         // we don't need it for voice, and it will reduce headroom
782         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
783         // should be dampened.)
784         if (locut_enabled) {
785                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
786         }
787
788         // Apply a level compressor to get the general level right.
789         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
790         // (or more precisely, near it, since we don't use infinite ratio),
791         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
792         // entirely arbitrary, but from practical tests with speech, it seems to
793         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
794         {
795                 unique_lock<mutex> lock(compressor_mutex);
796                 if (level_compressor_enabled) {
797                         float threshold = 0.01f;   // -40 dBFS.
798                         float ratio = 20.0f;
799                         float attack_time = 0.5f;
800                         float release_time = 20.0f;
801                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
802                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
803                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
804                 } else {
805                         // Just apply the gain we already had.
806                         float g = pow(10.0f, gain_staging_db / 20.0f);
807                         for (size_t i = 0; i < samples_out.size(); ++i) {
808                                 samples_out[i] *= g;
809                         }
810                 }
811         }
812
813 #if 0
814         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
815                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
816                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
817                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
818 #endif
819
820 //      float limiter_att, compressor_att;
821
822         // The real compressor.
823         if (compressor_enabled) {
824                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
825                 float ratio = 20.0f;
826                 float attack_time = 0.005f;
827                 float release_time = 0.040f;
828                 float makeup_gain = 2.0f;  // +6 dB.
829                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
830 //              compressor_att = compressor.get_attenuation();
831         }
832
833         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
834         // Note that since ratio is not infinite, we could go slightly higher than this.
835         if (limiter_enabled) {
836                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
837                 float ratio = 30.0f;
838                 float attack_time = 0.0f;  // Instant.
839                 float release_time = 0.020f;
840                 float makeup_gain = 1.0f;  // 0 dB.
841                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
842 //              limiter_att = limiter.get_attenuation();
843         }
844
845 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
846
847         // Upsample 4x to find interpolated peak.
848         peak_resampler.inp_data = samples_out.data();
849         peak_resampler.inp_count = samples_out.size() / 2;
850
851         vector<float> interpolated_samples_out;
852         interpolated_samples_out.resize(samples_out.size());
853         while (peak_resampler.inp_count > 0) {  // About four iterations.
854                 peak_resampler.out_data = &interpolated_samples_out[0];
855                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
856                 peak_resampler.process();
857                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
858                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
859                 peak_resampler.out_data = nullptr;
860         }
861
862         // At this point, we are most likely close to +0 LU, but all of our
863         // measurements have been on raw sample values, not R128 values.
864         // So we have a final makeup gain to get us to +0 LU; the gain
865         // adjustments required should be relatively small, and also, the
866         // offset shouldn't change much (only if the type of audio changes
867         // significantly). Thus, we shoot for updating this value basically
868         // “whenever we process buffers”, since the R128 calculation isn't exactly
869         // something we get out per-sample.
870         //
871         // Note that there's a feedback loop here, so we choose a very slow filter
872         // (half-time of 100 seconds).
873         double target_loudness_factor, alpha;
874         {
875                 unique_lock<mutex> lock(compressor_mutex);
876                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
877                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
878                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
879
880                 // If we're outside +/- 5 LU uncorrected, we don't count it as
881                 // a normal signal (probably silence) and don't change the
882                 // correction factor; just apply what we already have.
883                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
884                         alpha = 0.0;
885                 } else {
886                         // Formula adapted from
887                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
888                         const double half_time_s = 100.0;
889                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
890                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
891                 }
892
893                 double m = final_makeup_gain;
894                 for (size_t i = 0; i < samples_out.size(); i += 2) {
895                         samples_out[i + 0] *= m;
896                         samples_out[i + 1] *= m;
897                         m += (target_loudness_factor - m) * alpha;
898                 }
899                 final_makeup_gain = m;
900         }
901
902         // Find R128 levels and L/R correlation.
903         vector<float> left, right;
904         deinterleave_samples(samples_out, &left, &right);
905         float *ptrs[] = { left.data(), right.data() };
906         {
907                 unique_lock<mutex> lock(compressor_mutex);
908                 r128.process(left.size(), ptrs);
909                 correlation.process_samples(samples_out);
910         }
911
912         // Send the samples to the sound card.
913         if (alsa) {
914                 alsa->write(samples_out);
915         }
916
917         // And finally add them to the output.
918         h264_encoder->add_audio(frame_pts_int, move(samples_out));
919 }
920
921 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
922 {
923         GLuint vao;
924         glGenVertexArrays(1, &vao);
925         check_error();
926
927         glBindVertexArray(vao);
928         check_error();
929
930         // Extract Cb/Cr.
931         GLuint fbo = resource_pool->create_fbo(dst_tex);
932         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
933         glViewport(0, 0, WIDTH/2, HEIGHT/2);
934         check_error();
935
936         glUseProgram(cbcr_program_num);
937         check_error();
938
939         glActiveTexture(GL_TEXTURE0);
940         check_error();
941         glBindTexture(GL_TEXTURE_2D, src_tex);
942         check_error();
943         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
944         check_error();
945         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
946         check_error();
947         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
948         check_error();
949
950         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
951         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
952
953         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
954         check_error();
955
956         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
957                 glEnableVertexAttribArray(attr_index);
958                 check_error();
959                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
960                 check_error();
961         }
962
963         glDrawArrays(GL_TRIANGLES, 0, 3);
964         check_error();
965
966         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
967                 glDisableVertexAttribArray(attr_index);
968                 check_error();
969         }
970
971         glUseProgram(0);
972         check_error();
973         glBindFramebuffer(GL_FRAMEBUFFER, 0);
974         check_error();
975
976         resource_pool->release_fbo(fbo);
977         glDeleteVertexArrays(1, &vao);
978 }
979
980 void Mixer::release_display_frame(DisplayFrame *frame)
981 {
982         for (GLuint texnum : frame->temp_textures) {
983                 resource_pool->release_2d_texture(texnum);
984         }
985         frame->temp_textures.clear();
986         frame->ready_fence.reset();
987         frame->input_frames.clear();
988 }
989
990 void Mixer::start()
991 {
992         mixer_thread = thread(&Mixer::thread_func, this);
993         audio_thread = thread(&Mixer::audio_thread_func, this);
994 }
995
996 void Mixer::quit()
997 {
998         should_quit = true;
999         mixer_thread.join();
1000         audio_thread.join();
1001 }
1002
1003 void Mixer::transition_clicked(int transition_num)
1004 {
1005         theme->transition_clicked(transition_num, pts());
1006 }
1007
1008 void Mixer::channel_clicked(int preview_num)
1009 {
1010         theme->channel_clicked(preview_num);
1011 }
1012
1013 void Mixer::reset_meters()
1014 {
1015         peak_resampler.reset();
1016         peak = 0.0f;
1017         r128.reset();
1018         r128.integr_start();
1019         correlation.reset();
1020 }
1021
1022 Mixer::OutputChannel::~OutputChannel()
1023 {
1024         if (has_current_frame) {
1025                 parent->release_display_frame(&current_frame);
1026         }
1027         if (has_ready_frame) {
1028                 parent->release_display_frame(&ready_frame);
1029         }
1030 }
1031
1032 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1033 {
1034         // Store this frame for display. Remove the ready frame if any
1035         // (it was seemingly never used).
1036         {
1037                 unique_lock<mutex> lock(frame_mutex);
1038                 if (has_ready_frame) {
1039                         parent->release_display_frame(&ready_frame);
1040                 }
1041                 ready_frame = frame;
1042                 has_ready_frame = true;
1043         }
1044
1045         if (has_new_frame_ready_callback) {
1046                 new_frame_ready_callback();
1047         }
1048 }
1049
1050 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1051 {
1052         unique_lock<mutex> lock(frame_mutex);
1053         if (!has_current_frame && !has_ready_frame) {
1054                 return false;
1055         }
1056
1057         if (has_current_frame && has_ready_frame) {
1058                 // We have a new ready frame. Toss the current one.
1059                 parent->release_display_frame(&current_frame);
1060                 has_current_frame = false;
1061         }
1062         if (has_ready_frame) {
1063                 assert(!has_current_frame);
1064                 current_frame = ready_frame;
1065                 ready_frame.ready_fence.reset();  // Drop the refcount.
1066                 ready_frame.input_frames.clear();  // Drop the refcounts.
1067                 has_current_frame = true;
1068                 has_ready_frame = false;
1069         }
1070
1071         *frame = current_frame;
1072         return true;
1073 }
1074
1075 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1076 {
1077         new_frame_ready_callback = callback;
1078         has_new_frame_ready_callback = true;
1079 }