]> git.sesse.net Git - nageru/blob - mixer.cpp
Add a compressor (kindly relicensed by Rune Holm). Fixed for now, and only to get...
[nageru] / mixer.cpp
1 #define WIDTH 1280
2 #define HEIGHT 720
3 #define EXTRAHEIGHT 30
4
5 #undef Success
6
7 #include "mixer.h"
8
9 #include <assert.h>
10 #include <epoxy/egl.h>
11 #include <init.h>
12 #include <movit/effect_chain.h>
13 #include <movit/effect_util.h>
14 #include <movit/flat_input.h>
15 #include <movit/image_format.h>
16 #include <movit/resource_pool.h>
17 #include <stdint.h>
18 #include <stdio.h>
19 #include <stdlib.h>
20 #include <sys/time.h>
21 #include <time.h>
22 #include <util.h>
23 #include <algorithm>
24 #include <cmath>
25 #include <condition_variable>
26 #include <cstddef>
27 #include <memory>
28 #include <mutex>
29 #include <string>
30 #include <thread>
31 #include <utility>
32 #include <vector>
33
34 #include "bmusb/bmusb.h"
35 #include "context.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 }  // namespace
66
67 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
68         : httpd("test.ts", WIDTH, HEIGHT),
69           num_cards(num_cards),
70           mixer_surface(create_surface(format)),
71           h264_encoder_surface(create_surface(format)),
72           compressor(48000.0f)
73 {
74         httpd.start(9095);
75
76         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
77         check_error();
78
79         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
80         // will be halved when sampling them, and we need to compensate here.
81         movit_texel_subpixel_precision /= 2.0;
82
83         resource_pool.reset(new ResourcePool);
84         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
85         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
86                 output_channel[i].parent = this;
87         }
88
89         ImageFormat inout_format;
90         inout_format.color_space = COLORSPACE_sRGB;
91         inout_format.gamma_curve = GAMMA_sRGB;
92
93         // Display chain; shows the live output produced by the main chain (its RGBA version).
94         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
95         check_error();
96         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
97         display_chain->add_input(display_input);
98         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
99         display_chain->set_dither_bits(0);  // Don't bother.
100         display_chain->finalize();
101
102         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
103
104         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
105                 printf("Configuring card %d...\n", card_index);
106                 CaptureCard *card = &cards[card_index];
107                 card->usb = new BMUSBCapture(card_index);
108                 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
109                 card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44, WIDTH, HEIGHT));
110                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
111                 card->surface = create_surface(format);
112                 card->usb->set_dequeue_thread_callbacks(
113                         [card]{
114                                 eglBindAPI(EGL_OPENGL_API);
115                                 card->context = create_context(card->surface);
116                                 if (!make_current(card->context, card->surface)) {
117                                         printf("failed to create bmusb context\n");
118                                         exit(1);
119                                 }
120                         },
121                         [this]{
122                                 resource_pool->clean_context();
123                         });
124                 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
125                 card->usb->configure_card();
126         }
127
128         BMUSBCapture::start_bm_thread();
129
130         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
131                 cards[card_index].usb->start_bm_capture();
132         }
133
134         //chain->enable_phase_timing(true);
135
136         // Set up stuff for NV12 conversion.
137
138         // Cb/Cr shader.
139         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
140         string cbcr_frag_shader =
141                 "#version 130 \n"
142                 "in vec2 tc0; \n"
143                 "uniform sampler2D cbcr_tex; \n"
144                 "void main() { \n"
145                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
146                 "} \n";
147         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
148
149         r128.init(2, 48000);
150         r128.integr_start();
151 }
152
153 Mixer::~Mixer()
154 {
155         resource_pool->release_glsl_program(cbcr_program_num);
156         BMUSBCapture::stop_bm_thread();
157
158         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
159                 {
160                         unique_lock<mutex> lock(bmusb_mutex);
161                         cards[card_index].should_quit = true;  // Unblock thread.
162                         cards[card_index].new_data_ready_changed.notify_all();
163                 }
164                 cards[card_index].usb->stop_dequeue_thread();
165         }
166 }
167
168 namespace {
169
170 int unwrap_timecode(uint16_t current_wrapped, int last)
171 {
172         uint16_t last_wrapped = last & 0xffff;
173         if (current_wrapped > last_wrapped) {
174                 return (last & ~0xffff) | current_wrapped;
175         } else {
176                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
177         }
178 }
179
180 float find_peak(const vector<float> &samples)
181 {
182         float m = fabs(samples[0]);
183         for (size_t i = 1; i < samples.size(); ++i) {
184                 m = std::max(m, fabs(samples[i]));
185         }
186         return m;
187 }
188
189 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
190 {
191         size_t num_samples = in.size() / 2;
192         out_l->resize(num_samples);
193         out_r->resize(num_samples);
194
195         const float *inptr = in.data();
196         float *lptr = &(*out_l)[0];
197         float *rptr = &(*out_r)[0];
198         for (size_t i = 0; i < num_samples; ++i) {
199                 *lptr++ = *inptr++;
200                 *rptr++ = *inptr++;
201         }
202 }
203
204 }  // namespace
205
206 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
207                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
208                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
209 {
210         CaptureCard *card = &cards[card_index];
211
212         if (audio_frame.len - audio_offset > 30000) {
213                 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",
214                         card_index, int(audio_frame.len), int(audio_offset),
215                         timecode, int(video_frame.len), int(video_offset), video_format);
216                 if (video_frame.owner) {
217                         video_frame.owner->release_frame(video_frame);
218                 }
219                 if (audio_frame.owner) {
220                         audio_frame.owner->release_frame(audio_frame);
221                 }
222                 return;
223         }
224
225         int unwrapped_timecode = timecode;
226         int dropped_frames = 0;
227         if (card->last_timecode != -1) {
228                 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
229                 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
230         }
231         card->last_timecode = unwrapped_timecode;
232
233         // Convert the audio to stereo fp32 and add it.
234         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
235         vector<float> audio;
236         audio.resize(num_samples * 2);
237         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
238
239         // Add the audio.
240         {
241                 unique_lock<mutex> lock(card->audio_mutex);
242
243                 int unwrapped_timecode = timecode;
244                 if (dropped_frames > 60 * 2) {
245                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
246                                 card_index);
247                         card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
248                 } else if (dropped_frames > 0) {
249                         // Insert silence as needed.
250                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
251                                 card_index, dropped_frames, timecode);
252                         vector<float> silence;
253                         silence.resize((48000 / 60) * 2);
254                         for (int i = 0; i < dropped_frames; ++i) {
255                                 card->resampler->add_input_samples((unwrapped_timecode - dropped_frames + i) / 60.0, silence.data(), (48000 / 60));
256                         }
257                 }
258                 card->resampler->add_input_samples(unwrapped_timecode / 60.0, audio.data(), num_samples);
259         }
260
261         // Done with the audio, so release it.
262         if (audio_frame.owner) {
263                 audio_frame.owner->release_frame(audio_frame);
264         }
265
266         {
267                 // Wait until the previous frame was consumed.
268                 unique_lock<mutex> lock(bmusb_mutex);
269                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
270                 if (card->should_quit) return;
271         }
272
273         if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
274                 if (video_frame.len != 0) {
275                         printf("Card %d: Dropping video frame with wrong length (%ld)\n",
276                                 card_index, video_frame.len - video_offset);
277                 }
278                 if (video_frame.owner) {
279                         video_frame.owner->release_frame(video_frame);
280                 }
281
282                 // Still send on the information that we _had_ a frame, even though it's corrupted,
283                 // so that pts can go up accordingly.
284                 {
285                         unique_lock<mutex> lock(bmusb_mutex);
286                         card->new_data_ready = true;
287                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
288                         card->new_data_ready_fence = nullptr;
289                         card->dropped_frames = dropped_frames;
290                         card->new_data_ready_changed.notify_all();
291                 }
292                 return;
293         }
294
295         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
296         GLuint pbo = userdata->pbo;
297         check_error();
298         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
299         check_error();
300         glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
301         check_error();
302         //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
303         //check_error();
304
305         // Upload the textures.
306         glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
307         check_error();
308         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH, HEIGHT, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44) / 2 + WIDTH * 25 + 22));
309         check_error();
310         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
311         check_error();
312         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
313         check_error();
314         glBindTexture(GL_TEXTURE_2D, 0);
315         check_error();
316         GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);              
317         check_error();
318         assert(fence != nullptr);
319
320         {
321                 unique_lock<mutex> lock(bmusb_mutex);
322                 card->new_data_ready = true;
323                 card->new_frame = RefCountedFrame(video_frame);
324                 card->new_data_ready_fence = fence;
325                 card->dropped_frames = dropped_frames;
326                 card->new_data_ready_changed.notify_all();
327         }
328 }
329
330 void Mixer::thread_func()
331 {
332         eglBindAPI(EGL_OPENGL_API);
333         QOpenGLContext *context = create_context(mixer_surface);
334         if (!make_current(context, mixer_surface)) {
335                 printf("oops\n");
336                 exit(1);
337         }
338
339         struct timespec start, now;
340         clock_gettime(CLOCK_MONOTONIC, &start);
341
342         int frame = 0;
343         int dropped_frames = 0;
344
345         while (!should_quit) {
346                 CaptureCard card_copy[MAX_CARDS];
347
348                 {
349                         unique_lock<mutex> lock(bmusb_mutex);
350
351                         // The first card is the master timer, so wait for it to have a new frame.
352                         // TODO: Make configurable, and with a timeout.
353                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
354
355                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
356                                 CaptureCard *card = &cards[card_index];
357                                 card_copy[card_index].usb = card->usb;
358                                 card_copy[card_index].new_data_ready = card->new_data_ready;
359                                 card_copy[card_index].new_frame = card->new_frame;
360                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
361                                 card_copy[card_index].dropped_frames = card->dropped_frames;
362                                 card->new_data_ready = false;
363                                 card->new_data_ready_changed.notify_all();
364                         }
365                 }
366
367                 // Resample the audio as needed, including from previously dropped frames.
368                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
369                         process_audio_one_frame();
370                         if (frame_num != card_copy[0].dropped_frames) {
371                                 // For dropped frames, increase the pts.
372                                 ++dropped_frames;
373                                 pts_int += TIMEBASE / 60;
374                         }
375                 }
376
377                 if (audio_level_callback != nullptr) {
378                         double loudness_s = r128.loudness_S();
379                         double loudness_i = r128.integrated();
380                         double loudness_range_low = r128.range_min();
381                         double loudness_range_high = r128.range_max();
382
383                         audio_level_callback(loudness_s, 20.0 * log10(peak),
384                                              loudness_i, loudness_range_low, loudness_range_high);
385                 }
386
387                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
388                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
389                                 ++card_copy[card_index].dropped_frames;
390                         }
391                         if (card_copy[card_index].dropped_frames > 0) {
392                                 printf("Card %u dropped %d frames before this\n",
393                                         card_index, int(card_copy[card_index].dropped_frames));
394                         }
395                 }
396
397                 // If the first card is reporting a corrupted or otherwise dropped frame,
398                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
399                 if (card_copy[0].new_frame->len == 0) {
400                         ++dropped_frames;
401                         pts_int += TIMEBASE / 60;
402                         continue;
403                 }
404
405                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
406                         CaptureCard *card = &card_copy[card_index];
407                         if (!card->new_data_ready || card->new_frame->len == 0)
408                                 continue;
409
410                         assert(card->new_frame != nullptr);
411                         bmusb_current_rendering_frame[card_index] = card->new_frame;
412                         check_error();
413
414                         // The new texture might still be uploaded,
415                         // tell the GPU to wait until it's there.
416                         if (card->new_data_ready_fence) {
417                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
418                                 check_error();
419                                 glDeleteSync(card->new_data_ready_fence);
420                                 check_error();
421                         }
422                         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
423                         theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
424                 }
425
426                 // Get the main chain from the theme, and set its state immediately.
427                 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
428                 EffectChain *chain = theme_main_chain.first;
429                 theme_main_chain.second();
430
431                 GLuint y_tex, cbcr_tex;
432                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
433                 assert(got_frame);
434
435                 // Render main chain.
436                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
437                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
438                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
439                 check_error();
440                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
441                 resource_pool->release_fbo(fbo);
442
443                 subsample_chroma(cbcr_full_tex, cbcr_tex);
444                 resource_pool->release_2d_texture(cbcr_full_tex);
445
446                 // Set the right state for rgba_tex.
447                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
448                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
449                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
450                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
451                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
452
453                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
454                 check_error();
455
456                 // Make sure the H.264 gets a reference to all the
457                 // input frames needed, so that they are not released back
458                 // until the rendering is done.
459                 vector<RefCountedFrame> input_frames;
460                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
461                         input_frames.push_back(bmusb_current_rendering_frame[card_index]);
462                 }
463                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampler.h. TODO: Make less hard-coded.
464                 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
465                 ++frame;
466                 pts_int += TIMEBASE / 60;
467
468                 // The live frame just shows the RGBA texture we just rendered.
469                 // It owns rgba_tex now.
470                 DisplayFrame live_frame;
471                 live_frame.chain = display_chain.get();
472                 live_frame.setup_chain = [this, rgba_tex]{
473                         display_input->set_texture_num(rgba_tex);
474                 };
475                 live_frame.ready_fence = fence;
476                 live_frame.input_frames = {};
477                 live_frame.temp_textures = { rgba_tex };
478                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
479
480                 // Set up preview and any additional channels.
481                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
482                         DisplayFrame display_frame;
483                         pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT);  // FIXME: dimensions
484                         display_frame.chain = chain.first;
485                         display_frame.setup_chain = chain.second;
486                         display_frame.ready_fence = fence;
487
488                         // FIXME: possible to do better?
489                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
490                                 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
491                         }
492                         display_frame.temp_textures = {};
493                         output_channel[i].output_frame(display_frame);
494                 }
495
496                 clock_gettime(CLOCK_MONOTONIC, &now);
497                 double elapsed = now.tv_sec - start.tv_sec +
498                         1e-9 * (now.tv_nsec - start.tv_nsec);
499                 if (frame % 100 == 0) {
500                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
501                                 frame, dropped_frames, elapsed, frame / elapsed,
502                                 1e3 * elapsed / frame);
503                 //      chain->print_phase_timing();
504                 }
505
506 #if 0
507                 // Reset every 100 frames, so that local variations in frame times
508                 // (especially for the first few frames, when the shaders are
509                 // compiled etc.) don't make it hard to measure for the entire
510                 // remaining duration of the program.
511                 if (frame == 10000) {
512                         frame = 0;
513                         start = now;
514                 }
515 #endif
516                 check_error();
517         }
518
519         resource_pool->clean_context();
520 }
521
522 void Mixer::process_audio_one_frame()
523 {
524         vector<float> samples_card;
525         vector<float> samples_out;
526         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
527                 samples_card.resize((48000 / 60) * 2);
528                 {
529                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
530                         if (!cards[card_index].resampler->get_output_samples(pts(), &samples_card[0], 48000 / 60)) {
531                                 printf("Card %d reported previous underrun.\n", card_index);
532                         }
533                 }
534                 // TODO: Allow using audio from the other card(s) as well.
535                 if (card_index == 0) {
536                         samples_out = move(samples_card);
537                 }
538         }
539
540         // Apply a level compressor to get the general level right.
541         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
542         // (or more precisely, near it, since we don't use infinite ratio),
543         // then apply a makeup gain to get it to -12 dBFS. -12 dBFS is, of course,
544         // entirely arbitrary, but from practical tests with speech, it seems to
545         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
546         //
547         // TODO: Hook this up to a UI, so we can see the effects, and/or turn it off
548         // to control the gain manually instead. For now, there's only the #if-ed out
549         // code below.
550         //
551         // TODO: Add the actual compressors/limiters (for taking care of transients)
552         // later in the chain.
553         float threshold = 0.01f;   // -40 dBFS.
554         float ratio = 20.0f;
555         float attack_time = 0.1f;
556         float release_time = 10.0f;
557         float makeup_gain = pow(10.0f, 28.0f / 20.0f);  // +28 dB takes us to -12 dBFS.
558         compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
559
560 #if 0
561         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
562                 compressor.get_level(), 20.0 * log10(compressor.get_level()),
563                 compressor.get_attenuation(), 20.0 * log10(compressor.get_attenuation()),
564                 20.0 * log10(compressor.get_level() * compressor.get_attenuation() * makeup_gain));
565 #endif
566
567         // Find peak and R128 levels.
568         peak = std::max(peak, find_peak(samples_out));
569         vector<float> left, right;
570         deinterleave_samples(samples_out, &left, &right);
571         float *ptrs[] = { left.data(), right.data() };
572         r128.process(left.size(), ptrs);
573
574         // Actually add the samples to the output.
575         h264_encoder->add_audio(pts_int, move(samples_out));
576 }
577
578 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
579 {
580         GLuint vao;
581         glGenVertexArrays(1, &vao);
582         check_error();
583
584         float vertices[] = {
585                 0.0f, 2.0f,
586                 0.0f, 0.0f,
587                 2.0f, 0.0f
588         };
589
590         glBindVertexArray(vao);
591         check_error();
592
593         // Extract Cb/Cr.
594         GLuint fbo = resource_pool->create_fbo(dst_tex);
595         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
596         glViewport(0, 0, WIDTH/2, HEIGHT/2);
597         check_error();
598
599         glUseProgram(cbcr_program_num);
600         check_error();
601
602         glActiveTexture(GL_TEXTURE0);
603         check_error();
604         glBindTexture(GL_TEXTURE_2D, src_tex);
605         check_error();
606         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
607         check_error();
608         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
609         check_error();
610         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
611         check_error();
612
613         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
614         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
615
616         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
617         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
618
619         glDrawArrays(GL_TRIANGLES, 0, 3);
620         check_error();
621
622         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
623         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
624
625         glUseProgram(0);
626         check_error();
627
628         resource_pool->release_fbo(fbo);
629         glDeleteVertexArrays(1, &vao);
630 }
631
632 void Mixer::release_display_frame(DisplayFrame *frame)
633 {
634         for (GLuint texnum : frame->temp_textures) {
635                 resource_pool->release_2d_texture(texnum);
636         }
637         frame->temp_textures.clear();
638         frame->ready_fence.reset();
639         frame->input_frames.clear();
640 }
641
642 void Mixer::start()
643 {
644         mixer_thread = thread(&Mixer::thread_func, this);
645 }
646
647 void Mixer::quit()
648 {
649         should_quit = true;
650         mixer_thread.join();
651 }
652
653 void Mixer::transition_clicked(int transition_num)
654 {
655         theme->transition_clicked(transition_num, pts());
656 }
657
658 void Mixer::channel_clicked(int preview_num)
659 {
660         theme->channel_clicked(preview_num);
661 }
662
663 Mixer::OutputChannel::~OutputChannel()
664 {
665         if (has_current_frame) {
666                 parent->release_display_frame(&current_frame);
667         }
668         if (has_ready_frame) {
669                 parent->release_display_frame(&ready_frame);
670         }
671 }
672
673 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
674 {
675         // Store this frame for display. Remove the ready frame if any
676         // (it was seemingly never used).
677         {
678                 unique_lock<mutex> lock(frame_mutex);
679                 if (has_ready_frame) {
680                         parent->release_display_frame(&ready_frame);
681                 }
682                 ready_frame = frame;
683                 has_ready_frame = true;
684         }
685
686         if (has_new_frame_ready_callback) {
687                 new_frame_ready_callback();
688         }
689 }
690
691 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
692 {
693         unique_lock<mutex> lock(frame_mutex);
694         if (!has_current_frame && !has_ready_frame) {
695                 return false;
696         }
697
698         if (has_current_frame && has_ready_frame) {
699                 // We have a new ready frame. Toss the current one.
700                 parent->release_display_frame(&current_frame);
701                 has_current_frame = false;
702         }
703         if (has_ready_frame) {
704                 assert(!has_current_frame);
705                 current_frame = ready_frame;
706                 ready_frame.ready_fence.reset();  // Drop the refcount.
707                 ready_frame.input_frames.clear();  // Drop the refcounts.
708                 has_current_frame = true;
709                 has_ready_frame = false;
710         }
711
712         *frame = current_frame;
713         return true;
714 }
715
716 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
717 {
718         new_frame_ready_callback = callback;
719         has_new_frame_ready_callback = true;
720 }