]> git.sesse.net Git - nageru/blob - mixer.cpp
Add a switch to print video latency.
[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/resource.h>
18 #include <algorithm>
19 #include <chrono>
20 #include <condition_variable>
21 #include <cstddef>
22 #include <cstdint>
23 #include <memory>
24 #include <mutex>
25 #include <ratio>
26 #include <string>
27 #include <thread>
28 #include <utility>
29 #include <vector>
30
31 #include "DeckLinkAPI.h"
32 #include "LinuxCOM.h"
33 #include "alsa_output.h"
34 #include "bmusb/bmusb.h"
35 #include "bmusb/fake_capture.h"
36 #include "context.h"
37 #include "decklink_capture.h"
38 #include "defs.h"
39 #include "disk_space_estimator.h"
40 #include "flags.h"
41 #include "input_mapping.h"
42 #include "pbo_frame_allocator.h"
43 #include "ref_counted_gl_sync.h"
44 #include "resampling_queue.h"
45 #include "timebase.h"
46 #include "video_encoder.h"
47
48 class IDeckLink;
49 class QOpenGLContext;
50
51 using namespace movit;
52 using namespace std;
53 using namespace std::chrono;
54 using namespace std::placeholders;
55 using namespace bmusb;
56
57 Mixer *global_mixer = nullptr;
58 bool uses_mlock = false;
59
60 namespace {
61
62 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
63 {
64         if (interlaced) {
65                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
66                         input_state->buffered_frames[card_index][frame_num] =
67                                 input_state->buffered_frames[card_index][frame_num - 1];
68                 }
69                 input_state->buffered_frames[card_index][0] = { frame, field_num };
70         } else {
71                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
72                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
73                 }
74         }
75 }
76
77 }  // namespace
78
79 void QueueLengthPolicy::update_policy(int queue_length)
80 {
81         if (queue_length < 0) {  // Starvation.
82                 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
83                         ++safe_queue_length;
84                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
85                                 card_index, safe_queue_length);
86                 }
87                 frames_with_at_least_one = 0;
88                 been_at_safe_point_since_last_starvation = false;
89                 return;
90         }
91         if (queue_length > 0) {
92                 if (queue_length >= int(safe_queue_length)) {
93                         been_at_safe_point_since_last_starvation = true;
94                 }
95                 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
96                         --safe_queue_length;
97                         fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
98                                 card_index, safe_queue_length);
99                         frames_with_at_least_one = 0;
100                 }
101         } else {
102                 frames_with_at_least_one = 0;
103         }
104 }
105
106 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
107         : httpd(),
108           num_cards(num_cards),
109           mixer_surface(create_surface(format)),
110           h264_encoder_surface(create_surface(format)),
111           audio_mixer(num_cards)
112 {
113         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
114         check_error();
115
116         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
117         // will be halved when sampling them, and we need to compensate here.
118         movit_texel_subpixel_precision /= 2.0;
119
120         resource_pool.reset(new ResourcePool);
121         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
122         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
123                 output_channel[i].parent = this;
124                 output_channel[i].channel = i;
125         }
126
127         ImageFormat inout_format;
128         inout_format.color_space = COLORSPACE_sRGB;
129         inout_format.gamma_curve = GAMMA_sRGB;
130
131         // Display chain; shows the live output produced by the main chain (its RGBA version).
132         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
133         check_error();
134         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, global_flags.width, global_flags.height);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
135         display_chain->add_input(display_input);
136         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
137         display_chain->set_dither_bits(0);  // Don't bother.
138         display_chain->finalize();
139
140         video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, global_flags.width, global_flags.height, &httpd, global_disk_space_estimator));
141
142         // Start listening for clients only once VideoEncoder has written its header, if any.
143         httpd.start(9095);
144
145         // First try initializing the then PCI devices, then USB, then
146         // fill up with fake cards until we have the desired number of cards.
147         unsigned num_pci_devices = 0;
148         unsigned card_index = 0;
149
150         {
151                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
152                 if (decklink_iterator != nullptr) {
153                         for ( ; card_index < num_cards; ++card_index) {
154                                 IDeckLink *decklink;
155                                 if (decklink_iterator->Next(&decklink) != S_OK) {
156                                         break;
157                                 }
158
159                                 configure_card(card_index, new DeckLinkCapture(decklink, card_index), /*is_fake_capture=*/false);
160                                 ++num_pci_devices;
161                         }
162                         decklink_iterator->Release();
163                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
164                 } else {
165                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
166                 }
167         }
168         unsigned num_usb_devices = BMUSBCapture::num_cards();
169         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
170                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
171                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
172                 configure_card(card_index, capture, /*is_fake_capture=*/false);
173         }
174         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
175
176         unsigned num_fake_cards = 0;
177         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
178                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
179                 configure_card(card_index, capture, /*is_fake_capture=*/true);
180         }
181
182         if (num_fake_cards > 0) {
183                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
184         }
185
186         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
187         BMUSBCapture::start_bm_thread();
188
189         for (card_index = 0; card_index < num_cards; ++card_index) {
190                 cards[card_index].queue_length_policy.reset(card_index);
191                 cards[card_index].capture->start_bm_capture();
192         }
193
194         // Set up stuff for NV12 conversion.
195
196         // Cb/Cr shader.
197         string cbcr_vert_shader =
198                 "#version 130 \n"
199                 " \n"
200                 "in vec2 position; \n"
201                 "in vec2 texcoord; \n"
202                 "out vec2 tc0; \n"
203                 "uniform vec2 foo_chroma_offset_0; \n"
204                 " \n"
205                 "void main() \n"
206                 "{ \n"
207                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
208                 "    // \n"
209                 "    //   2.000  0.000  0.000 -1.000 \n"
210                 "    //   0.000  2.000  0.000 -1.000 \n"
211                 "    //   0.000  0.000 -2.000 -1.000 \n"
212                 "    //   0.000  0.000  0.000  1.000 \n"
213                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
214                 "    vec2 flipped_tc = texcoord; \n"
215                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
216                 "} \n";
217         string cbcr_frag_shader =
218                 "#version 130 \n"
219                 "in vec2 tc0; \n"
220                 "uniform sampler2D cbcr_tex; \n"
221                 "out vec4 FragColor; \n"
222                 "void main() { \n"
223                 "    FragColor = texture(cbcr_tex, tc0); \n"
224                 "} \n";
225         vector<string> frag_shader_outputs;
226         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
227
228         float vertices[] = {
229                 0.0f, 2.0f,
230                 0.0f, 0.0f,
231                 2.0f, 0.0f
232         };
233         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
234         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
235         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
236
237         if (global_flags.enable_alsa_output) {
238                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
239         }
240 }
241
242 Mixer::~Mixer()
243 {
244         resource_pool->release_glsl_program(cbcr_program_num);
245         glDeleteBuffers(1, &cbcr_vbo);
246         BMUSBCapture::stop_bm_thread();
247
248         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
249                 {
250                         unique_lock<mutex> lock(bmusb_mutex);
251                         cards[card_index].should_quit = true;  // Unblock thread.
252                         cards[card_index].new_frames_changed.notify_all();
253                 }
254                 cards[card_index].capture->stop_dequeue_thread();
255         }
256
257         video_encoder.reset(nullptr);
258 }
259
260 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture)
261 {
262         printf("Configuring card %d...\n", card_index);
263
264         CaptureCard *card = &cards[card_index];
265         if (card->capture != nullptr) {
266                 card->capture->stop_dequeue_thread();
267                 delete card->capture;
268         }
269         card->capture = capture;
270         card->is_fake_capture = is_fake_capture;
271         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
272         if (card->frame_allocator == nullptr) {
273                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, global_flags.width, global_flags.height));  // 8 MB.
274         }
275         card->capture->set_video_frame_allocator(card->frame_allocator.get());
276         if (card->surface == nullptr) {
277                 card->surface = create_surface_with_same_format(mixer_surface);
278         }
279         while (!card->new_frames.empty()) card->new_frames.pop();
280         card->fractional_samples = 0;
281         card->last_timecode = -1;
282         card->capture->configure_card();
283
284         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
285         audio_mixer.reset_resampler(device);
286         audio_mixer.set_display_name(device, card->capture->get_description());
287         audio_mixer.trigger_state_changed_callback();
288 }
289
290
291 namespace {
292
293 int unwrap_timecode(uint16_t current_wrapped, int last)
294 {
295         uint16_t last_wrapped = last & 0xffff;
296         if (current_wrapped > last_wrapped) {
297                 return (last & ~0xffff) | current_wrapped;
298         } else {
299                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
300         }
301 }
302
303 }  // namespace
304
305 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
306                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
307                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
308 {
309         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
310         CaptureCard *card = &cards[card_index];
311
312         if (is_mode_scanning[card_index]) {
313                 if (video_format.has_signal) {
314                         // Found a stable signal, so stop scanning.
315                         is_mode_scanning[card_index] = false;
316                 } else {
317                         static constexpr double switch_time_s = 0.5;  // Should be enough time for the signal to stabilize.
318                         steady_clock::time_point now = steady_clock::now();
319                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
320                         if (sec_since_last_switch > switch_time_s) {
321                                 // It isn't this mode; try the next one.
322                                 mode_scanlist_index[card_index]++;
323                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
324                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
325                                 last_mode_scan_change[card_index] = now;
326                         }
327                 }
328         }
329
330         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
331         assert(frame_length > 0);
332
333         size_t num_samples = (audio_frame.len > audio_offset) ? (audio_frame.len - audio_offset) / audio_format.num_channels / (audio_format.bits_per_sample / 8) : 0;
334         if (num_samples > OUTPUT_FREQUENCY / 10) {
335                 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",
336                         card_index, int(audio_frame.len), int(audio_offset),
337                         timecode, int(video_frame.len), int(video_offset), video_format.id);
338                 if (video_frame.owner) {
339                         video_frame.owner->release_frame(video_frame);
340                 }
341                 if (audio_frame.owner) {
342                         audio_frame.owner->release_frame(audio_frame);
343                 }
344                 return;
345         }
346
347         int dropped_frames = 0;
348         if (card->last_timecode != -1) {
349                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
350         }
351
352         // Number of samples per frame if we need to insert silence.
353         // (Could be nonintegral, but resampling will save us then.)
354         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
355
356         if (dropped_frames > MAX_FPS * 2) {
357                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
358                         card_index, card->last_timecode, timecode);
359                 audio_mixer.reset_resampler(device);
360                 dropped_frames = 0;
361         } else if (dropped_frames > 0) {
362                 // Insert silence as needed.
363                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
364                         card_index, dropped_frames, timecode);
365
366                 bool success;
367                 do {
368                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
369                 } while (!success);
370         }
371
372         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length);
373
374         // Done with the audio, so release it.
375         if (audio_frame.owner) {
376                 audio_frame.owner->release_frame(audio_frame);
377         }
378
379         card->last_timecode = timecode;
380
381         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
382         if (video_frame.len - video_offset == 0 ||
383             video_frame.len - video_offset != expected_length) {
384                 if (video_frame.len != 0) {
385                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
386                                 card_index, video_frame.len - video_offset, expected_length);
387                 }
388                 if (video_frame.owner) {
389                         video_frame.owner->release_frame(video_frame);
390                 }
391
392                 // Still send on the information that we _had_ a frame, even though it's corrupted,
393                 // so that pts can go up accordingly.
394                 {
395                         unique_lock<mutex> lock(bmusb_mutex);
396                         CaptureCard::NewFrame new_frame;
397                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
398                         new_frame.length = frame_length;
399                         new_frame.interlaced = false;
400                         new_frame.dropped_frames = dropped_frames;
401                         card->new_frames.push(move(new_frame));
402                         card->new_frames_changed.notify_all();
403                 }
404                 return;
405         }
406
407         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
408
409         unsigned num_fields = video_format.interlaced ? 2 : 1;
410         steady_clock::time_point frame_upload_start;
411         if (video_format.interlaced) {
412                 // Send the two fields along as separate frames; the other side will need to add
413                 // a deinterlacer to actually get this right.
414                 assert(video_format.height % 2 == 0);
415                 video_format.height /= 2;
416                 assert(frame_length % 2 == 0);
417                 frame_length /= 2;
418                 num_fields = 2;
419                 frame_upload_start = steady_clock::now();
420         }
421         userdata->last_interlaced = video_format.interlaced;
422         userdata->last_has_signal = video_format.has_signal;
423         userdata->last_is_connected = video_format.is_connected;
424         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
425         userdata->last_frame_rate_den = video_format.frame_rate_den;
426         RefCountedFrame frame(video_frame);
427
428         // Upload the textures.
429         size_t cbcr_width = video_format.width / 2;
430         size_t cbcr_offset = video_offset / 2;
431         size_t y_offset = video_frame.size / 2 + video_offset / 2;
432
433         for (unsigned field = 0; field < num_fields; ++field) {
434                 // Put the actual texture upload in a lambda that is executed in the main thread.
435                 // It is entirely possible to do this in the same thread (and it might even be
436                 // faster, depending on the GPU and driver), but it appears to be trickling
437                 // driver bugs very easily.
438                 //
439                 // Note that this means we must hold on to the actual frame data in <userdata>
440                 // until the upload command is run, but we hold on to <frame> much longer than that
441                 // (in fact, all the way until we no longer use the texture in rendering).
442                 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() {
443                         unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
444
445                         if (userdata->tex_y[field] == 0 ||
446                             userdata->tex_cbcr[field] == 0 ||
447                             video_format.width != userdata->last_width[field] ||
448                             video_format.height != userdata->last_height[field]) {
449                                 // We changed resolution since last use of this texture, so we need to create
450                                 // a new object. Note that this each card has its own PBOFrameAllocator,
451                                 // we don't need to worry about these flip-flopping between resolutions.
452                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
453                                 check_error();
454                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
455                                 check_error();
456                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
457                                 check_error();
458                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
459                                 check_error();
460                                 userdata->last_width[field] = video_format.width;
461                                 userdata->last_height[field] = video_format.height;
462                         }
463
464                         GLuint pbo = userdata->pbo;
465                         check_error();
466                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
467                         check_error();
468
469                         size_t field_y_start = y_offset + video_format.width * field_start_line;
470                         size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
471
472                         if (global_flags.flush_pbos) {
473                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
474                                 check_error();
475                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
476                                 check_error();
477                         }
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(field_cbcr_start));
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(field_y_start));
486                         check_error();
487                         glBindTexture(GL_TEXTURE_2D, 0);
488                         check_error();
489                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
490                         check_error();
491                 };
492
493                 if (field == 1) {
494                         // Don't upload the second field as fast as we can; wait until
495                         // the field time has approximately passed. (Otherwise, we could
496                         // get timing jitter against the other sources, and possibly also
497                         // against the video display, although the latter is not as critical.)
498                         // This requires our system clock to be reasonably close to the
499                         // video clock, but that's not an unreasonable assumption.
500                         steady_clock::time_point second_field_start = frame_upload_start +
501                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
502                         this_thread::sleep_until(second_field_start);
503                 }
504
505                 {
506                         unique_lock<mutex> lock(bmusb_mutex);
507                         CaptureCard::NewFrame new_frame;
508                         new_frame.frame = frame;
509                         new_frame.length = frame_length;
510                         new_frame.field = field;
511                         new_frame.interlaced = video_format.interlaced;
512                         new_frame.upload_func = upload_func;
513                         new_frame.dropped_frames = dropped_frames;
514                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
515                         card->new_frames.push(move(new_frame));
516                         card->new_frames_changed.notify_all();
517                 }
518         }
519 }
520
521 void Mixer::bm_hotplug_add(libusb_device *dev)
522 {
523         lock_guard<mutex> lock(hotplug_mutex);
524         hotplugged_cards.push_back(dev);
525 }
526
527 void Mixer::bm_hotplug_remove(unsigned card_index)
528 {
529         cards[card_index].new_frames_changed.notify_all();
530 }
531
532 void Mixer::thread_func()
533 {
534         eglBindAPI(EGL_OPENGL_API);
535         QOpenGLContext *context = create_context(mixer_surface);
536         if (!make_current(context, mixer_surface)) {
537                 printf("oops\n");
538                 exit(1);
539         }
540
541         steady_clock::time_point start, now;
542         start = steady_clock::now();
543
544         int frame = 0;
545         int stats_dropped_frames = 0;
546
547         while (!should_quit) {
548                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
549                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
550                 int num_samples[MAX_VIDEO_CARDS] = { 0 };
551
552                 unsigned master_card_index = theme->map_signal(master_clock_channel);
553                 assert(master_card_index < num_cards);
554
555                 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
556                 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
557                 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
558
559                 handle_hotplugged_cards();
560
561                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
562                         if (card_index == master_card_index || !has_new_frame[card_index]) {
563                                 continue;
564                         }
565                         if (new_frames[card_index].frame->len == 0) {
566                                 ++new_frames[card_index].dropped_frames;
567                         }
568                         if (new_frames[card_index].dropped_frames > 0) {
569                                 printf("Card %u dropped %d frames before this\n",
570                                         card_index, int(new_frames[card_index].dropped_frames));
571                         }
572                 }
573
574                 // If the first card is reporting a corrupted or otherwise dropped frame,
575                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
576                 if (new_frames[master_card_index].frame->len == 0) {
577                         ++stats_dropped_frames;
578                         pts_int += new_frames[master_card_index].length;
579                         continue;
580                 }
581
582                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
583                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
584                                 continue;
585
586                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
587                         assert(new_frame->frame != nullptr);
588                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
589                         check_error();
590
591                         // The new texture might need uploading before use.
592                         if (new_frame->upload_func) {
593                                 new_frame->upload_func();
594                                 new_frame->upload_func = nullptr;
595                         }
596                 }
597
598                 int64_t frame_duration = new_frames[master_card_index].length;
599                 render_one_frame(frame_duration);
600                 ++frame;
601                 pts_int += frame_duration;
602
603                 now = steady_clock::now();
604                 double elapsed = duration<double>(now - start).count();
605                 if (frame % 100 == 0) {
606                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
607                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
608                                 1e3 * elapsed / frame);
609                 //      chain->print_phase_timing();
610
611                         // Check our memory usage, to see if we are close to our mlockall()
612                         // limit (if at all set).
613                         rusage used;
614                         if (getrusage(RUSAGE_SELF, &used) == -1) {
615                                 perror("getrusage(RUSAGE_SELF)");
616                                 assert(false);
617                         }
618
619                         if (uses_mlock) {
620                                 rlimit limit;
621                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
622                                         perror("getrlimit(RLIMIT_MEMLOCK)");
623                                         assert(false);
624                                 }
625
626                                 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
627                                         long(used.ru_maxrss / 1024),
628                                         long(limit.rlim_cur / 1048576),
629                                         float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
630                         } else {
631                                 printf(", using %ld MB memory (not locked)",
632                                         long(used.ru_maxrss / 1024));
633                         }
634
635                         printf("\n");
636                 }
637
638
639                 if (should_cut.exchange(false)) {  // Test and clear.
640                         video_encoder->do_cut(frame);
641                 }
642
643 #if 0
644                 // Reset every 100 frames, so that local variations in frame times
645                 // (especially for the first few frames, when the shaders are
646                 // compiled etc.) don't make it hard to measure for the entire
647                 // remaining duration of the program.
648                 if (frame == 10000) {
649                         frame = 0;
650                         start = now;
651                 }
652 #endif
653                 check_error();
654         }
655
656         resource_pool->clean_context();
657 }
658
659 void Mixer::get_one_frame_from_each_card(unsigned master_card_index, CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS], bool has_new_frame[MAX_VIDEO_CARDS], int num_samples[MAX_VIDEO_CARDS])
660 {
661 start:
662         // The first card is the master timer, so wait for it to have a new frame.
663         // TODO: Add a timeout.
664         unique_lock<mutex> lock(bmusb_mutex);
665         cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty() || cards[master_card_index].capture->get_disconnected(); });
666
667         if (cards[master_card_index].new_frames.empty()) {
668                 // We were woken up, but not due to a new frame. Deal with it
669                 // and then restart.
670                 assert(cards[master_card_index].capture->get_disconnected());
671                 handle_hotplugged_cards();
672                 goto start;
673         }
674
675         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
676                 CaptureCard *card = &cards[card_index];
677                 if (card->new_frames.empty()) {
678                         assert(card_index != master_card_index);
679                         card->queue_length_policy.update_policy(-1);
680                         continue;
681                 }
682                 new_frames[card_index] = move(card->new_frames.front());
683                 has_new_frame[card_index] = true;
684                 card->new_frames.pop();
685                 card->new_frames_changed.notify_all();
686
687                 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
688                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
689                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
690                 assert(num_samples[card_index] >= 0);
691
692                 if (card_index == master_card_index) {
693                         // We don't use the queue length policy for the master card,
694                         // but we will if it stops being the master. Thus, clear out
695                         // the policy in case we switch in the future.
696                         card->queue_length_policy.reset(card_index);
697                 } else {
698                         // If we have excess frames compared to the policy for this card,
699                         // drop frames from the head.
700                         card->queue_length_policy.update_policy(card->new_frames.size());
701                         while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
702                                 card->new_frames.pop();
703                         }
704                 }
705         }
706 }
707
708 void Mixer::handle_hotplugged_cards()
709 {
710         // Check for cards that have been disconnected since last frame.
711         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
712                 CaptureCard *card = &cards[card_index];
713                 if (card->capture->get_disconnected()) {
714                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
715                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
716                         configure_card(card_index, capture, /*is_fake_capture=*/true);
717                         card->queue_length_policy.reset(card_index);
718                         card->capture->start_bm_capture();
719                 }
720         }
721
722         // Check for cards that have been connected since last frame.
723         vector<libusb_device *> hotplugged_cards_copy;
724         {
725                 lock_guard<mutex> lock(hotplug_mutex);
726                 swap(hotplugged_cards, hotplugged_cards_copy);
727         }
728         for (libusb_device *new_dev : hotplugged_cards_copy) {
729                 // Look for a fake capture card where we can stick this in.
730                 int free_card_index = -1;
731                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
732                         if (cards[card_index].is_fake_capture) {
733                                 free_card_index = int(card_index);
734                                 break;
735                         }
736                 }
737
738                 if (free_card_index == -1) {
739                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
740                         libusb_unref_device(new_dev);
741                 } else {
742                         // BMUSBCapture takes ownership.
743                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
744                         CaptureCard *card = &cards[free_card_index];
745                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
746                         configure_card(free_card_index, capture, /*is_fake_capture=*/false);
747                         card->queue_length_policy.reset(free_card_index);
748                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
749                         capture->start_bm_capture();
750                 }
751         }
752 }
753
754
755 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
756 {
757         // Resample the audio as needed, including from previously dropped frames.
758         assert(num_cards > 0);
759         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
760                 const bool dropped_frame = (frame_num != dropped_frames);
761                 {
762                         // Signal to the audio thread to process this frame.
763                         // Note that if the frame is a dropped frame, we signal that
764                         // we don't want to use this frame as base for adjusting
765                         // the resampler rate. The reason for this is that the timing
766                         // of these frames is often way too late; they typically don't
767                         // “arrive” before we synthesize them. Thus, we could end up
768                         // in a situation where we have inserted e.g. five audio frames
769                         // into the queue before we then start pulling five of them
770                         // back out. This makes ResamplingQueue overestimate the delay,
771                         // causing undue resampler changes. (We _do_ use the last,
772                         // non-dropped frame; perhaps we should just discard that as well,
773                         // since dropped frames are expected to be rare, and it might be
774                         // better to just wait until we have a slightly more normal situation).
775                         unique_lock<mutex> lock(audio_mutex);
776                         bool adjust_rate = !dropped_frame;
777                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate});
778                         audio_task_queue_changed.notify_one();
779                 }
780                 if (dropped_frame) {
781                         // For dropped frames, increase the pts. Note that if the format changed
782                         // in the meantime, we have no way of detecting that; we just have to
783                         // assume the frame length is always the same.
784                         pts_int += length_per_frame;
785                 }
786         }
787 }
788
789 void Mixer::render_one_frame(int64_t duration)
790 {
791         // Get the main chain from the theme, and set its state immediately.
792         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
793         EffectChain *chain = theme_main_chain.chain;
794         theme_main_chain.setup_chain();
795         //theme_main_chain.chain->enable_phase_timing(true);
796
797         GLuint y_tex, cbcr_tex;
798         bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
799         assert(got_frame);
800
801         // Render main chain.
802         GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, global_flags.width, global_flags.height);
803         GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, global_flags.width, global_flags.height);  // Saves texture bandwidth, although dithering gets messed up.
804         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
805         check_error();
806         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
807         resource_pool->release_fbo(fbo);
808
809         subsample_chroma(cbcr_full_tex, cbcr_tex);
810         resource_pool->release_2d_texture(cbcr_full_tex);
811
812         // Set the right state for rgba_tex.
813         glBindFramebuffer(GL_FRAMEBUFFER, 0);
814         glBindTexture(GL_TEXTURE_2D, rgba_tex);
815         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
816         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
817         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
818
819         const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
820         RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
821
822         // The live frame just shows the RGBA texture we just rendered.
823         // It owns rgba_tex now.
824         DisplayFrame live_frame;
825         live_frame.chain = display_chain.get();
826         live_frame.setup_chain = [this, rgba_tex]{
827                 display_input->set_texture_num(rgba_tex);
828         };
829         live_frame.ready_fence = fence;
830         live_frame.input_frames = {};
831         live_frame.temp_textures = { rgba_tex };
832         output_channel[OUTPUT_LIVE].output_frame(live_frame);
833
834         // Set up preview and any additional channels.
835         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
836                 DisplayFrame display_frame;
837                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
838                 display_frame.chain = chain.chain;
839                 display_frame.setup_chain = chain.setup_chain;
840                 display_frame.ready_fence = fence;
841                 display_frame.input_frames = chain.input_frames;
842                 display_frame.temp_textures = {};
843                 output_channel[i].output_frame(display_frame);
844         }
845 }
846
847 void Mixer::audio_thread_func()
848 {
849         while (!should_quit) {
850                 AudioTask task;
851
852                 {
853                         unique_lock<mutex> lock(audio_mutex);
854                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
855                         if (should_quit) {
856                                 return;
857                         }
858                         task = audio_task_queue.front();
859                         audio_task_queue.pop();
860                 }
861
862                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
863                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
864                 vector<float> samples_out = audio_mixer.get_output(
865                         double(task.pts_int) / TIMEBASE,
866                         task.num_samples,
867                         rate_adjustment_policy);
868
869                 // Send the samples to the sound card, then add them to the output.
870                 if (alsa) {
871                         alsa->write(samples_out);
872                 }
873                 video_encoder->add_audio(task.pts_int, move(samples_out));
874         }
875 }
876
877 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
878 {
879         GLuint vao;
880         glGenVertexArrays(1, &vao);
881         check_error();
882
883         glBindVertexArray(vao);
884         check_error();
885
886         // Extract Cb/Cr.
887         GLuint fbo = resource_pool->create_fbo(dst_tex);
888         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
889         glViewport(0, 0, global_flags.width/2, global_flags.height/2);
890         check_error();
891
892         glUseProgram(cbcr_program_num);
893         check_error();
894
895         glActiveTexture(GL_TEXTURE0);
896         check_error();
897         glBindTexture(GL_TEXTURE_2D, src_tex);
898         check_error();
899         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
900         check_error();
901         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
902         check_error();
903         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
904         check_error();
905
906         float chroma_offset_0[] = { -0.5f / global_flags.width, 0.0f };
907         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
908
909         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
910         check_error();
911
912         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
913                 glEnableVertexAttribArray(attr_index);
914                 check_error();
915                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
916                 check_error();
917         }
918
919         glDrawArrays(GL_TRIANGLES, 0, 3);
920         check_error();
921
922         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
923                 glDisableVertexAttribArray(attr_index);
924                 check_error();
925         }
926
927         glUseProgram(0);
928         check_error();
929         glBindFramebuffer(GL_FRAMEBUFFER, 0);
930         check_error();
931
932         resource_pool->release_fbo(fbo);
933         glDeleteVertexArrays(1, &vao);
934 }
935
936 void Mixer::release_display_frame(DisplayFrame *frame)
937 {
938         for (GLuint texnum : frame->temp_textures) {
939                 resource_pool->release_2d_texture(texnum);
940         }
941         frame->temp_textures.clear();
942         frame->ready_fence.reset();
943         frame->input_frames.clear();
944 }
945
946 void Mixer::start()
947 {
948         mixer_thread = thread(&Mixer::thread_func, this);
949         audio_thread = thread(&Mixer::audio_thread_func, this);
950 }
951
952 void Mixer::quit()
953 {
954         should_quit = true;
955         audio_task_queue_changed.notify_one();
956         mixer_thread.join();
957         audio_thread.join();
958 }
959
960 void Mixer::transition_clicked(int transition_num)
961 {
962         theme->transition_clicked(transition_num, pts());
963 }
964
965 void Mixer::channel_clicked(int preview_num)
966 {
967         theme->channel_clicked(preview_num);
968 }
969
970 void Mixer::start_mode_scanning(unsigned card_index)
971 {
972         assert(card_index < num_cards);
973         if (is_mode_scanning[card_index]) {
974                 return;
975         }
976         is_mode_scanning[card_index] = true;
977         mode_scanlist[card_index].clear();
978         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
979                 mode_scanlist[card_index].push_back(mode.first);
980         }
981         assert(!mode_scanlist[card_index].empty());
982         mode_scanlist_index[card_index] = 0;
983         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
984         last_mode_scan_change[card_index] = steady_clock::now();
985 }
986
987 Mixer::OutputChannel::~OutputChannel()
988 {
989         if (has_current_frame) {
990                 parent->release_display_frame(&current_frame);
991         }
992         if (has_ready_frame) {
993                 parent->release_display_frame(&ready_frame);
994         }
995 }
996
997 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
998 {
999         // Store this frame for display. Remove the ready frame if any
1000         // (it was seemingly never used).
1001         {
1002                 unique_lock<mutex> lock(frame_mutex);
1003                 if (has_ready_frame) {
1004                         parent->release_display_frame(&ready_frame);
1005                 }
1006                 ready_frame = frame;
1007                 has_ready_frame = true;
1008         }
1009
1010         if (new_frame_ready_callback) {
1011                 new_frame_ready_callback();
1012         }
1013
1014         // Reduce the number of callbacks by filtering duplicates. The reason
1015         // why we bother doing this is that Qt seemingly can get into a state
1016         // where its builds up an essentially unbounded queue of signals,
1017         // consuming more and more memory, and there's no good way of collapsing
1018         // user-defined signals or limiting the length of the queue.
1019         if (transition_names_updated_callback) {
1020                 vector<string> transition_names = global_mixer->get_transition_names();
1021                 bool changed = false;
1022                 if (transition_names.size() != last_transition_names.size()) {
1023                         changed = true;
1024                 } else {
1025                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1026                                 if (transition_names[i] != last_transition_names[i]) {
1027                                         changed = true;
1028                                         break;
1029                                 }
1030                         }
1031                 }
1032                 if (changed) {
1033                         transition_names_updated_callback(transition_names);
1034                         last_transition_names = transition_names;
1035                 }
1036         }
1037         if (name_updated_callback) {
1038                 string name = global_mixer->get_channel_name(channel);
1039                 if (name != last_name) {
1040                         name_updated_callback(name);
1041                         last_name = name;
1042                 }
1043         }
1044         if (color_updated_callback) {
1045                 string color = global_mixer->get_channel_color(channel);
1046                 if (color != last_color) {
1047                         color_updated_callback(color);
1048                         last_color = color;
1049                 }
1050         }
1051 }
1052
1053 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1054 {
1055         unique_lock<mutex> lock(frame_mutex);
1056         if (!has_current_frame && !has_ready_frame) {
1057                 return false;
1058         }
1059
1060         if (has_current_frame && has_ready_frame) {
1061                 // We have a new ready frame. Toss the current one.
1062                 parent->release_display_frame(&current_frame);
1063                 has_current_frame = false;
1064         }
1065         if (has_ready_frame) {
1066                 assert(!has_current_frame);
1067                 current_frame = ready_frame;
1068                 ready_frame.ready_fence.reset();  // Drop the refcount.
1069                 ready_frame.input_frames.clear();  // Drop the refcounts.
1070                 has_current_frame = true;
1071                 has_ready_frame = false;
1072         }
1073
1074         *frame = current_frame;
1075         return true;
1076 }
1077
1078 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1079 {
1080         new_frame_ready_callback = callback;
1081 }
1082
1083 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1084 {
1085         transition_names_updated_callback = callback;
1086 }
1087
1088 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1089 {
1090         name_updated_callback = callback;
1091 }
1092
1093 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1094 {
1095         color_updated_callback = callback;
1096 }
1097
1098 mutex RefCountedGLsync::fence_lock;