]> git.sesse.net Git - nageru/blob - mixer.cpp
3b602bfb07597d41fefc0e213b66d496799a5a23
[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, uint16_t 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         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
343
344         // Add the audio.
345         {
346                 unique_lock<mutex> lock(card->audio_mutex);
347
348                 // Number of samples per frame if we need to insert silence.
349                 // (Could be nonintegral, but resampling will save us then.)
350                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
351
352                 if (dropped_frames > MAX_FPS * 2) {
353                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
354                                 card_index, card->last_timecode, timecode);
355                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
356                         dropped_frames = 0;
357                 } else if (dropped_frames > 0) {
358                         // Insert silence as needed.
359                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
360                                 card_index, dropped_frames, timecode);
361                         vector<float> silence(silence_samples * 2, 0.0f);
362                         for (int i = 0; i < dropped_frames; ++i) {
363                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
364                                 // Note that if the format changed in the meantime, we have
365                                 // no way of detecting that; we just have to assume the frame length
366                                 // is always the same.
367                                 local_pts += frame_length;
368                         }
369                 }
370                 if (num_samples == 0) {
371                         audio.resize(silence_samples * 2);
372                         num_samples = silence_samples;
373                 }
374                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
375                 card->next_local_pts = local_pts + frame_length;
376         }
377
378         card->last_timecode = timecode;
379
380         // Done with the audio, so release it.
381         if (audio_frame.owner) {
382                 audio_frame.owner->release_frame(audio_frame);
383         }
384
385         {
386                 // Wait until the previous frame was consumed.
387                 unique_lock<mutex> lock(bmusb_mutex);
388                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
389                 if (card->should_quit) return;
390         }
391
392         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
393         if (video_frame.len - video_offset == 0 ||
394             video_frame.len - video_offset != expected_length) {
395                 if (video_frame.len != 0) {
396                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
397                                 card_index, video_frame.len - video_offset, expected_length);
398                 }
399                 if (video_frame.owner) {
400                         video_frame.owner->release_frame(video_frame);
401                 }
402
403                 // Still send on the information that we _had_ a frame, even though it's corrupted,
404                 // so that pts can go up accordingly.
405                 {
406                         unique_lock<mutex> lock(bmusb_mutex);
407                         card->new_data_ready = true;
408                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
409                         card->new_frame_length = frame_length;
410                         card->new_frame_interlaced = false;
411                         card->new_data_ready_fence = nullptr;
412                         card->dropped_frames = dropped_frames;
413                         card->new_data_ready_changed.notify_all();
414                 }
415                 return;
416         }
417
418         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
419
420         unsigned num_fields = video_format.interlaced ? 2 : 1;
421         timespec frame_upload_start;
422         if (video_format.interlaced) {
423                 // Send the two fields along as separate frames; the other side will need to add
424                 // a deinterlacer to actually get this right.
425                 assert(video_format.height % 2 == 0);
426                 video_format.height /= 2;
427                 assert(frame_length % 2 == 0);
428                 frame_length /= 2;
429                 num_fields = 2;
430                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
431         }
432         userdata->last_interlaced = video_format.interlaced;
433         userdata->last_has_signal = video_format.has_signal;
434         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
435         userdata->last_frame_rate_den = video_format.frame_rate_den;
436         RefCountedFrame new_frame(video_frame);
437
438         // Upload the textures.
439         size_t cbcr_width = video_format.width / 2;
440         size_t cbcr_offset = video_offset / 2;
441         size_t y_offset = video_frame.size / 2 + video_offset / 2;
442
443         for (unsigned field = 0; field < num_fields; ++field) {
444                 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
445
446                 if (userdata->tex_y[field] == 0 ||
447                     userdata->tex_cbcr[field] == 0 ||
448                     video_format.width != userdata->last_width[field] ||
449                     video_format.height != userdata->last_height[field]) {
450                         // We changed resolution since last use of this texture, so we need to create
451                         // a new object. Note that this each card has its own PBOFrameAllocator,
452                         // we don't need to worry about these flip-flopping between resolutions.
453                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
454                         check_error();
455                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
456                         check_error();
457                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
458                         check_error();
459                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
460                         check_error();
461                         userdata->last_width[field] = video_format.width;
462                         userdata->last_height[field] = video_format.height;
463                 }
464
465                 GLuint pbo = userdata->pbo;
466                 check_error();
467                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
468                 check_error();
469                 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
470                 check_error();
471
472                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
473                 check_error();
474                 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)));
475                 check_error();
476                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
477                 check_error();
478                 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));
479                 check_error();
480                 glBindTexture(GL_TEXTURE_2D, 0);
481                 check_error();
482                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
483                 check_error();
484                 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
485                 check_error();
486                 assert(fence != nullptr);
487
488                 if (field == 1) {
489                         // Don't upload the second field as fast as we can; wait until
490                         // the field time has approximately passed. (Otherwise, we could
491                         // get timing jitter against the other sources, and possibly also
492                         // against the video display, although the latter is not as critical.)
493                         // This requires our system clock to be reasonably close to the
494                         // video clock, but that's not an unreasonable assumption.
495                         timespec second_field_start;
496                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
497                                 frame_length * 1000000000 / TIMEBASE;
498                         second_field_start.tv_sec = frame_upload_start.tv_sec +
499                                 second_field_start.tv_nsec / 1000000000;
500                         second_field_start.tv_nsec %= 1000000000;
501
502                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
503                                                &second_field_start, nullptr) == -1 &&
504                                errno == EINTR) ;
505                 }
506
507                 {
508                         unique_lock<mutex> lock(bmusb_mutex);
509                         card->new_data_ready = true;
510                         card->new_frame = new_frame;
511                         card->new_frame_length = frame_length;
512                         card->new_frame_field = field;
513                         card->new_frame_interlaced = video_format.interlaced;
514                         card->new_data_ready_fence = fence;
515                         card->dropped_frames = dropped_frames;
516                         card->new_data_ready_changed.notify_all();
517
518                         if (field != num_fields - 1) {
519                                 // Wait until the previous frame was consumed.
520                                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
521                                 if (card->should_quit) return;
522                         }
523                 }
524         }
525 }
526
527 void Mixer::thread_func()
528 {
529         eglBindAPI(EGL_OPENGL_API);
530         QOpenGLContext *context = create_context(mixer_surface);
531         if (!make_current(context, mixer_surface)) {
532                 printf("oops\n");
533                 exit(1);
534         }
535
536         struct timespec start, now;
537         clock_gettime(CLOCK_MONOTONIC, &start);
538
539         int frame = 0;
540         int stats_dropped_frames = 0;
541
542         while (!should_quit) {
543                 CaptureCard card_copy[MAX_CARDS];
544                 int num_samples[MAX_CARDS];
545
546                 {
547                         unique_lock<mutex> lock(bmusb_mutex);
548
549                         // The first card is the master timer, so wait for it to have a new frame.
550                         // TODO: Make configurable, and with a timeout.
551                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
552
553                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
554                                 CaptureCard *card = &cards[card_index];
555                                 card_copy[card_index].new_data_ready = card->new_data_ready;
556                                 card_copy[card_index].new_frame = card->new_frame;
557                                 card_copy[card_index].new_frame_length = card->new_frame_length;
558                                 card_copy[card_index].new_frame_field = card->new_frame_field;
559                                 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
560                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
561                                 card_copy[card_index].dropped_frames = card->dropped_frames;
562                                 card->new_data_ready = false;
563                                 card->new_data_ready_changed.notify_all();
564
565                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
566                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
567                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
568                                 assert(num_samples[card_index] >= 0);
569                         }
570                 }
571
572                 // Resample the audio as needed, including from previously dropped frames.
573                 assert(num_cards > 0);
574                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
575                         {
576                                 // Signal to the audio thread to process this frame.
577                                 unique_lock<mutex> lock(audio_mutex);
578                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
579                                 audio_task_queue_changed.notify_one();
580                         }
581                         if (frame_num != card_copy[0].dropped_frames) {
582                                 // For dropped frames, increase the pts. Note that if the format changed
583                                 // in the meantime, we have no way of detecting that; we just have to
584                                 // assume the frame length is always the same.
585                                 ++stats_dropped_frames;
586                                 pts_int += card_copy[0].new_frame_length;
587                         }
588                 }
589
590                 if (audio_level_callback != nullptr) {
591                         unique_lock<mutex> lock(compressor_mutex);
592                         double loudness_s = r128.loudness_S();
593                         double loudness_i = r128.integrated();
594                         double loudness_range_low = r128.range_min();
595                         double loudness_range_high = r128.range_max();
596
597                         audio_level_callback(loudness_s, 20.0 * log10(peak),
598                                              loudness_i, loudness_range_low, loudness_range_high,
599                                              gain_staging_db, 20.0 * log10(final_makeup_gain),
600                                              correlation.get_correlation());
601                 }
602
603                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
604                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
605                                 ++card_copy[card_index].dropped_frames;
606                         }
607                         if (card_copy[card_index].dropped_frames > 0) {
608                                 printf("Card %u dropped %d frames before this\n",
609                                         card_index, int(card_copy[card_index].dropped_frames));
610                         }
611                 }
612
613                 // If the first card is reporting a corrupted or otherwise dropped frame,
614                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
615                 if (card_copy[0].new_frame->len == 0) {
616                         ++stats_dropped_frames;
617                         pts_int += card_copy[0].new_frame_length;
618                         continue;
619                 }
620
621                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
622                         CaptureCard *card = &card_copy[card_index];
623                         if (!card->new_data_ready || card->new_frame->len == 0)
624                                 continue;
625
626                         assert(card->new_frame != nullptr);
627                         insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
628                         check_error();
629
630                         // The new texture might still be uploaded,
631                         // tell the GPU to wait until it's there.
632                         if (card->new_data_ready_fence) {
633                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
634                                 check_error();
635                                 glDeleteSync(card->new_data_ready_fence);
636                                 check_error();
637                         }
638                 }
639
640                 // Get the main chain from the theme, and set its state immediately.
641                 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
642                 EffectChain *chain = theme_main_chain.chain;
643                 theme_main_chain.setup_chain();
644                 //theme_main_chain.chain->enable_phase_timing(true);
645
646                 GLuint y_tex, cbcr_tex;
647                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
648                 assert(got_frame);
649
650                 // Render main chain.
651                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
652                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
653                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
654                 check_error();
655                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
656                 resource_pool->release_fbo(fbo);
657
658                 subsample_chroma(cbcr_full_tex, cbcr_tex);
659                 resource_pool->release_2d_texture(cbcr_full_tex);
660
661                 // Set the right state for rgba_tex.
662                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
663                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
664                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
665                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
666                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
667
668                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
669                 check_error();
670
671                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
672                 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
673                 ++frame;
674                 pts_int += card_copy[0].new_frame_length;
675
676                 // The live frame just shows the RGBA texture we just rendered.
677                 // It owns rgba_tex now.
678                 DisplayFrame live_frame;
679                 live_frame.chain = display_chain.get();
680                 live_frame.setup_chain = [this, rgba_tex]{
681                         display_input->set_texture_num(rgba_tex);
682                 };
683                 live_frame.ready_fence = fence;
684                 live_frame.input_frames = {};
685                 live_frame.temp_textures = { rgba_tex };
686                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
687
688                 // Set up preview and any additional channels.
689                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
690                         DisplayFrame display_frame;
691                         Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
692                         display_frame.chain = chain.chain;
693                         display_frame.setup_chain = chain.setup_chain;
694                         display_frame.ready_fence = fence;
695                         display_frame.input_frames = chain.input_frames;
696                         display_frame.temp_textures = {};
697                         output_channel[i].output_frame(display_frame);
698                 }
699
700                 clock_gettime(CLOCK_MONOTONIC, &now);
701                 double elapsed = now.tv_sec - start.tv_sec +
702                         1e-9 * (now.tv_nsec - start.tv_nsec);
703                 if (frame % 100 == 0) {
704                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
705                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
706                                 1e3 * elapsed / frame);
707                 //      chain->print_phase_timing();
708                 }
709
710                 if (should_cut.exchange(false)) {  // Test and clear.
711                         string filename = generate_local_dump_filename(frame);
712                         printf("Starting new recording: %s\n", filename.c_str());
713                         h264_encoder->shutdown();
714                         httpd.close_output_file();
715                         httpd.open_output_file(filename.c_str());
716                         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
717                 }
718
719 #if 0
720                 // Reset every 100 frames, so that local variations in frame times
721                 // (especially for the first few frames, when the shaders are
722                 // compiled etc.) don't make it hard to measure for the entire
723                 // remaining duration of the program.
724                 if (frame == 10000) {
725                         frame = 0;
726                         start = now;
727                 }
728 #endif
729                 check_error();
730         }
731
732         resource_pool->clean_context();
733 }
734
735 void Mixer::audio_thread_func()
736 {
737         while (!should_quit) {
738                 AudioTask task;
739
740                 {
741                         unique_lock<mutex> lock(audio_mutex);
742                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
743                         task = audio_task_queue.front();
744                         audio_task_queue.pop();
745                 }
746
747                 process_audio_one_frame(task.pts_int, task.num_samples);
748         }
749 }
750
751 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
752 {
753         vector<float> samples_card;
754         vector<float> samples_out;
755
756         // TODO: Allow mixing audio from several sources.
757         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
758         assert(selected_audio_card < num_cards);
759
760         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
761                 samples_card.resize(num_samples * 2);
762                 {
763                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
764                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
765                                 printf("Card %d reported previous underrun.\n", card_index);
766                         }
767                 }
768                 if (card_index == selected_audio_card) {
769                         samples_out = move(samples_card);
770                 }
771         }
772
773         // Cut away everything under 120 Hz (or whatever the cutoff is);
774         // we don't need it for voice, and it will reduce headroom
775         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
776         // should be dampened.)
777         if (locut_enabled) {
778                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
779         }
780
781         // Apply a level compressor to get the general level right.
782         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
783         // (or more precisely, near it, since we don't use infinite ratio),
784         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
785         // entirely arbitrary, but from practical tests with speech, it seems to
786         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
787         {
788                 unique_lock<mutex> lock(compressor_mutex);
789                 if (level_compressor_enabled) {
790                         float threshold = 0.01f;   // -40 dBFS.
791                         float ratio = 20.0f;
792                         float attack_time = 0.5f;
793                         float release_time = 20.0f;
794                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
795                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
796                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
797                 } else {
798                         // Just apply the gain we already had.
799                         float g = pow(10.0f, gain_staging_db / 20.0f);
800                         for (size_t i = 0; i < samples_out.size(); ++i) {
801                                 samples_out[i] *= g;
802                         }
803                 }
804         }
805
806 #if 0
807         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
808                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
809                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
810                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
811 #endif
812
813 //      float limiter_att, compressor_att;
814
815         // The real compressor.
816         if (compressor_enabled) {
817                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
818                 float ratio = 20.0f;
819                 float attack_time = 0.005f;
820                 float release_time = 0.040f;
821                 float makeup_gain = 2.0f;  // +6 dB.
822                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
823 //              compressor_att = compressor.get_attenuation();
824         }
825
826         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
827         // Note that since ratio is not infinite, we could go slightly higher than this.
828         if (limiter_enabled) {
829                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
830                 float ratio = 30.0f;
831                 float attack_time = 0.0f;  // Instant.
832                 float release_time = 0.020f;
833                 float makeup_gain = 1.0f;  // 0 dB.
834                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
835 //              limiter_att = limiter.get_attenuation();
836         }
837
838 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
839
840         // Upsample 4x to find interpolated peak.
841         peak_resampler.inp_data = samples_out.data();
842         peak_resampler.inp_count = samples_out.size() / 2;
843
844         vector<float> interpolated_samples_out;
845         interpolated_samples_out.resize(samples_out.size());
846         while (peak_resampler.inp_count > 0) {  // About four iterations.
847                 peak_resampler.out_data = &interpolated_samples_out[0];
848                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
849                 peak_resampler.process();
850                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
851                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
852                 peak_resampler.out_data = nullptr;
853         }
854
855         // At this point, we are most likely close to +0 LU, but all of our
856         // measurements have been on raw sample values, not R128 values.
857         // So we have a final makeup gain to get us to +0 LU; the gain
858         // adjustments required should be relatively small, and also, the
859         // offset shouldn't change much (only if the type of audio changes
860         // significantly). Thus, we shoot for updating this value basically
861         // “whenever we process buffers”, since the R128 calculation isn't exactly
862         // something we get out per-sample.
863         //
864         // Note that there's a feedback loop here, so we choose a very slow filter
865         // (half-time of 100 seconds).
866         double target_loudness_factor, alpha;
867         {
868                 unique_lock<mutex> lock(compressor_mutex);
869                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
870                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
871                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
872
873                 // If we're outside +/- 5 LU uncorrected, we don't count it as
874                 // a normal signal (probably silence) and don't change the
875                 // correction factor; just apply what we already have.
876                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
877                         alpha = 0.0;
878                 } else {
879                         // Formula adapted from
880                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
881                         const double half_time_s = 100.0;
882                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
883                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
884                 }
885
886                 double m = final_makeup_gain;
887                 for (size_t i = 0; i < samples_out.size(); i += 2) {
888                         samples_out[i + 0] *= m;
889                         samples_out[i + 1] *= m;
890                         m += (target_loudness_factor - m) * alpha;
891                 }
892                 final_makeup_gain = m;
893         }
894
895         // Find R128 levels and L/R correlation.
896         vector<float> left, right;
897         deinterleave_samples(samples_out, &left, &right);
898         float *ptrs[] = { left.data(), right.data() };
899         {
900                 unique_lock<mutex> lock(compressor_mutex);
901                 r128.process(left.size(), ptrs);
902                 correlation.process_samples(samples_out);
903         }
904
905         // Send the samples to the sound card.
906         if (alsa) {
907                 alsa->write(samples_out);
908         }
909
910         // And finally add them to the output.
911         h264_encoder->add_audio(frame_pts_int, move(samples_out));
912 }
913
914 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
915 {
916         GLuint vao;
917         glGenVertexArrays(1, &vao);
918         check_error();
919
920         glBindVertexArray(vao);
921         check_error();
922
923         // Extract Cb/Cr.
924         GLuint fbo = resource_pool->create_fbo(dst_tex);
925         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
926         glViewport(0, 0, WIDTH/2, HEIGHT/2);
927         check_error();
928
929         glUseProgram(cbcr_program_num);
930         check_error();
931
932         glActiveTexture(GL_TEXTURE0);
933         check_error();
934         glBindTexture(GL_TEXTURE_2D, src_tex);
935         check_error();
936         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
937         check_error();
938         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
939         check_error();
940         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
941         check_error();
942
943         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
944         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
945
946         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
947         check_error();
948
949         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
950                 glEnableVertexAttribArray(attr_index);
951                 check_error();
952                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
953                 check_error();
954         }
955
956         glDrawArrays(GL_TRIANGLES, 0, 3);
957         check_error();
958
959         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
960                 glDisableVertexAttribArray(attr_index);
961                 check_error();
962         }
963
964         glUseProgram(0);
965         check_error();
966         glBindFramebuffer(GL_FRAMEBUFFER, 0);
967         check_error();
968
969         resource_pool->release_fbo(fbo);
970         glDeleteVertexArrays(1, &vao);
971 }
972
973 void Mixer::release_display_frame(DisplayFrame *frame)
974 {
975         for (GLuint texnum : frame->temp_textures) {
976                 resource_pool->release_2d_texture(texnum);
977         }
978         frame->temp_textures.clear();
979         frame->ready_fence.reset();
980         frame->input_frames.clear();
981 }
982
983 void Mixer::start()
984 {
985         mixer_thread = thread(&Mixer::thread_func, this);
986         audio_thread = thread(&Mixer::audio_thread_func, this);
987 }
988
989 void Mixer::quit()
990 {
991         should_quit = true;
992         mixer_thread.join();
993         audio_thread.join();
994 }
995
996 void Mixer::transition_clicked(int transition_num)
997 {
998         theme->transition_clicked(transition_num, pts());
999 }
1000
1001 void Mixer::channel_clicked(int preview_num)
1002 {
1003         theme->channel_clicked(preview_num);
1004 }
1005
1006 void Mixer::reset_meters()
1007 {
1008         peak_resampler.reset();
1009         peak = 0.0f;
1010         r128.reset();
1011         r128.integr_start();
1012         correlation.reset();
1013 }
1014
1015 Mixer::OutputChannel::~OutputChannel()
1016 {
1017         if (has_current_frame) {
1018                 parent->release_display_frame(&current_frame);
1019         }
1020         if (has_ready_frame) {
1021                 parent->release_display_frame(&ready_frame);
1022         }
1023 }
1024
1025 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1026 {
1027         // Store this frame for display. Remove the ready frame if any
1028         // (it was seemingly never used).
1029         {
1030                 unique_lock<mutex> lock(frame_mutex);
1031                 if (has_ready_frame) {
1032                         parent->release_display_frame(&ready_frame);
1033                 }
1034                 ready_frame = frame;
1035                 has_ready_frame = true;
1036         }
1037
1038         if (has_new_frame_ready_callback) {
1039                 new_frame_ready_callback();
1040         }
1041 }
1042
1043 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1044 {
1045         unique_lock<mutex> lock(frame_mutex);
1046         if (!has_current_frame && !has_ready_frame) {
1047                 return false;
1048         }
1049
1050         if (has_current_frame && has_ready_frame) {
1051                 // We have a new ready frame. Toss the current one.
1052                 parent->release_display_frame(&current_frame);
1053                 has_current_frame = false;
1054         }
1055         if (has_ready_frame) {
1056                 assert(!has_current_frame);
1057                 current_frame = ready_frame;
1058                 ready_frame.ready_fence.reset();  // Drop the refcount.
1059                 ready_frame.input_frames.clear();  // Drop the refcounts.
1060                 has_current_frame = true;
1061                 has_ready_frame = false;
1062         }
1063
1064         *frame = current_frame;
1065         return true;
1066 }
1067
1068 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1069 {
1070         new_frame_ready_callback = callback;
1071         has_new_frame_ready_callback = true;
1072 }