10 #include <epoxy/egl.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>
25 #include <condition_variable>
34 #include "bmusb/bmusb.h"
36 #include "h264encode.h"
37 #include "pbo_frame_allocator.h"
38 #include "ref_counted_gl_sync.h"
43 using namespace movit;
45 using namespace std::placeholders;
47 Mixer *global_mixer = nullptr;
51 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
53 for (size_t i = 0; i < num_samples; ++i) {
54 for (size_t j = 0; j < out_channels; ++j) {
58 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
59 dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
61 src += 3 * (in_channels - out_channels);
67 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
68 : httpd("test.ts", WIDTH, HEIGHT),
70 mixer_surface(create_surface(format)),
71 h264_encoder_surface(create_surface(format))
75 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
78 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
79 // will be halved when sampling them, and we need to compensate here.
80 movit_texel_subpixel_precision /= 2.0;
82 resource_pool.reset(new ResourcePool);
83 theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
84 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
85 output_channel[i].parent = this;
88 ImageFormat inout_format;
89 inout_format.color_space = COLORSPACE_sRGB;
90 inout_format.gamma_curve = GAMMA_sRGB;
92 // Display chain; shows the live output produced by the main chain (its RGBA version).
93 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
95 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
96 display_chain->add_input(display_input);
97 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
98 display_chain->set_dither_bits(0); // Don't bother.
99 display_chain->finalize();
101 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
103 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
104 printf("Configuring card %d...\n", card_index);
105 CaptureCard *card = &cards[card_index];
106 card->usb = new BMUSBCapture(card_index);
107 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
108 card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44, WIDTH, HEIGHT));
109 card->usb->set_video_frame_allocator(card->frame_allocator.get());
110 card->surface = create_surface(format);
111 card->usb->set_dequeue_thread_callbacks(
113 eglBindAPI(EGL_OPENGL_API);
114 card->context = create_context(card->surface);
115 if (!make_current(card->context, card->surface)) {
116 printf("failed to create bmusb context\n");
121 resource_pool->clean_context();
123 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
124 card->usb->configure_card();
127 BMUSBCapture::start_bm_thread();
129 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
130 cards[card_index].usb->start_bm_capture();
133 //chain->enable_phase_timing(true);
135 // Set up stuff for NV12 conversion.
138 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
139 string cbcr_frag_shader =
142 "uniform sampler2D cbcr_tex; \n"
144 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
146 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
154 resource_pool->release_glsl_program(cbcr_program_num);
155 BMUSBCapture::stop_bm_thread();
157 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
159 unique_lock<mutex> lock(bmusb_mutex);
160 cards[card_index].should_quit = true; // Unblock thread.
161 cards[card_index].new_data_ready_changed.notify_all();
163 cards[card_index].usb->stop_dequeue_thread();
169 int unwrap_timecode(uint16_t current_wrapped, int last)
171 uint16_t last_wrapped = last & 0xffff;
172 if (current_wrapped > last_wrapped) {
173 return (last & ~0xffff) | current_wrapped;
175 return 0x10000 + ((last & ~0xffff) | current_wrapped);
179 float find_peak(const vector<float> &samples)
181 float m = fabs(samples[0]);
182 for (size_t i = 1; i < samples.size(); ++i) {
183 m = std::max(m, fabs(samples[i]));
188 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
190 size_t num_samples = in.size() / 2;
191 out_l->resize(num_samples);
192 out_r->resize(num_samples);
194 const float *inptr = in.data();
195 float *lptr = &(*out_l)[0];
196 float *rptr = &(*out_r)[0];
197 for (size_t i = 0; i < num_samples; ++i) {
205 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
206 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
207 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
209 CaptureCard *card = &cards[card_index];
211 if (audio_frame.len - audio_offset > 30000) {
212 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",
213 card_index, int(audio_frame.len), int(audio_offset),
214 timecode, int(video_frame.len), int(video_offset), video_format);
215 if (video_frame.owner) {
216 video_frame.owner->release_frame(video_frame);
218 if (audio_frame.owner) {
219 audio_frame.owner->release_frame(audio_frame);
224 int unwrapped_timecode = timecode;
225 int dropped_frames = 0;
226 if (card->last_timecode != -1) {
227 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
228 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
230 card->last_timecode = unwrapped_timecode;
232 // Convert the audio to stereo fp32 and add it.
233 size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
235 audio.resize(num_samples * 2);
236 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
240 unique_lock<mutex> lock(card->audio_mutex);
242 int unwrapped_timecode = timecode;
243 if (dropped_frames > 60 * 2) {
244 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
246 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
247 } else if (dropped_frames > 0) {
248 // Insert silence as needed.
249 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
250 card_index, dropped_frames, timecode);
251 vector<float> silence;
252 silence.resize((48000 / 60) * 2);
253 for (int i = 0; i < dropped_frames; ++i) {
254 card->resampler->add_input_samples((unwrapped_timecode - dropped_frames + i) / 60.0, silence.data(), (48000 / 60));
257 card->resampler->add_input_samples(unwrapped_timecode / 60.0, audio.data(), num_samples);
260 // Done with the audio, so release it.
261 if (audio_frame.owner) {
262 audio_frame.owner->release_frame(audio_frame);
266 // Wait until the previous frame was consumed.
267 unique_lock<mutex> lock(bmusb_mutex);
268 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
269 if (card->should_quit) return;
272 if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
273 if (video_frame.len != 0) {
274 printf("Card %d: Dropping video frame with wrong length (%ld)\n",
275 card_index, video_frame.len - video_offset);
277 if (video_frame.owner) {
278 video_frame.owner->release_frame(video_frame);
281 // Still send on the information that we _had_ a frame, even though it's corrupted,
282 // so that pts can go up accordingly.
284 unique_lock<mutex> lock(bmusb_mutex);
285 card->new_data_ready = true;
286 card->new_frame = RefCountedFrame(FrameAllocator::Frame());
287 card->new_data_ready_fence = nullptr;
288 card->dropped_frames = dropped_frames;
289 card->new_data_ready_changed.notify_all();
294 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
295 GLuint pbo = userdata->pbo;
297 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
299 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
301 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
304 // Upload the textures.
305 glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
307 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 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
311 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
313 glBindTexture(GL_TEXTURE_2D, 0);
315 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
317 assert(fence != nullptr);
320 unique_lock<mutex> lock(bmusb_mutex);
321 card->new_data_ready = true;
322 card->new_frame = RefCountedFrame(video_frame);
323 card->new_data_ready_fence = fence;
324 card->dropped_frames = dropped_frames;
325 card->new_data_ready_changed.notify_all();
329 void Mixer::thread_func()
331 eglBindAPI(EGL_OPENGL_API);
332 QOpenGLContext *context = create_context(mixer_surface);
333 if (!make_current(context, mixer_surface)) {
338 struct timespec start, now;
339 clock_gettime(CLOCK_MONOTONIC, &start);
342 int dropped_frames = 0;
344 while (!should_quit) {
345 CaptureCard card_copy[MAX_CARDS];
348 unique_lock<mutex> lock(bmusb_mutex);
350 // The first card is the master timer, so wait for it to have a new frame.
351 // TODO: Make configurable, and with a timeout.
352 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
354 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
355 CaptureCard *card = &cards[card_index];
356 card_copy[card_index].usb = card->usb;
357 card_copy[card_index].new_data_ready = card->new_data_ready;
358 card_copy[card_index].new_frame = card->new_frame;
359 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
360 card_copy[card_index].new_frame_audio = move(card->new_frame_audio);
361 card_copy[card_index].dropped_frames = card->dropped_frames;
362 card->new_data_ready = false;
363 card->new_data_ready_changed.notify_all();
367 // Resample the audio as needed, including from previously dropped frames.
368 vector<float> samples_out;
369 // TODO: Allow using audio from the other card(s) as well.
370 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
371 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
372 samples_out.resize((48000 / 60) * 2);
374 unique_lock<mutex> lock(cards[card_index].audio_mutex);
375 if (!cards[card_index].resampler->get_output_samples(pts(), &samples_out[0], 48000 / 60)) {
376 printf("Card %d reported previous underrun.\n", card_index);
379 if (card_index == 0) {
380 vector<float> left, right;
381 peak = std::max(peak, find_peak(samples_out));
382 deinterleave_samples(samples_out, &left, &right);
383 float *ptrs[] = { left.data(), right.data() };
384 r128.process(left.size(), ptrs);
385 h264_encoder->add_audio(pts_int, move(samples_out));
388 if (frame_num != card_copy[0].dropped_frames) {
389 // For dropped frames, increase the pts.
391 pts_int += TIMEBASE / 60;
395 if (audio_level_callback != nullptr) {
396 double loudness_s = r128.loudness_S();
397 double loudness_i = r128.integrated();
398 double loudness_range_low = r128.range_min();
399 double loudness_range_high = r128.range_max();
401 audio_level_callback(loudness_s, 20.0 * log10(peak),
402 loudness_i, loudness_range_low, loudness_range_high);
405 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
406 if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
407 ++card_copy[card_index].dropped_frames;
409 if (card_copy[card_index].dropped_frames > 0) {
410 printf("Card %u dropped %d frames before this\n",
411 card_index, int(card_copy[card_index].dropped_frames));
415 // If the first card is reporting a corrupted or otherwise dropped frame,
416 // just increase the pts (skipping over this frame) and don't try to compute anything new.
417 if (card_copy[0].new_frame->len == 0) {
419 pts_int += TIMEBASE / 60;
423 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
424 CaptureCard *card = &card_copy[card_index];
425 if (!card->new_data_ready || card->new_frame->len == 0)
428 assert(card->new_frame != nullptr);
429 bmusb_current_rendering_frame[card_index] = card->new_frame;
432 // The new texture might still be uploaded,
433 // tell the GPU to wait until it's there.
434 if (card->new_data_ready_fence) {
435 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
437 glDeleteSync(card->new_data_ready_fence);
440 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
441 theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
444 // Get the main chain from the theme, and set its state immediately.
445 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
446 EffectChain *chain = theme_main_chain.first;
447 theme_main_chain.second();
449 GLuint y_tex, cbcr_tex;
450 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
453 // Render main chain.
454 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
455 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
456 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
458 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
459 resource_pool->release_fbo(fbo);
461 subsample_chroma(cbcr_full_tex, cbcr_tex);
462 resource_pool->release_2d_texture(cbcr_full_tex);
464 // Set the right state for rgba_tex.
465 glBindFramebuffer(GL_FRAMEBUFFER, 0);
466 glBindTexture(GL_TEXTURE_2D, rgba_tex);
467 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
468 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
469 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
471 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
474 // Make sure the H.264 gets a reference to all the
475 // input frames needed, so that they are not released back
476 // until the rendering is done.
477 vector<RefCountedFrame> input_frames;
478 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
479 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
481 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampler.h. TODO: Make less hard-coded.
482 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
484 pts_int += TIMEBASE / 60;
486 // The live frame just shows the RGBA texture we just rendered.
487 // It owns rgba_tex now.
488 DisplayFrame live_frame;
489 live_frame.chain = display_chain.get();
490 live_frame.setup_chain = [this, rgba_tex]{
491 display_input->set_texture_num(rgba_tex);
493 live_frame.ready_fence = fence;
494 live_frame.input_frames = {};
495 live_frame.temp_textures = { rgba_tex };
496 output_channel[OUTPUT_LIVE].output_frame(live_frame);
498 // Set up preview and any additional channels.
499 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
500 DisplayFrame display_frame;
501 pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT); // FIXME: dimensions
502 display_frame.chain = chain.first;
503 display_frame.setup_chain = chain.second;
504 display_frame.ready_fence = fence;
506 // FIXME: possible to do better?
507 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
508 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
510 display_frame.temp_textures = {};
511 output_channel[i].output_frame(display_frame);
514 clock_gettime(CLOCK_MONOTONIC, &now);
515 double elapsed = now.tv_sec - start.tv_sec +
516 1e-9 * (now.tv_nsec - start.tv_nsec);
517 if (frame % 100 == 0) {
518 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
519 frame, dropped_frames, elapsed, frame / elapsed,
520 1e3 * elapsed / frame);
521 // chain->print_phase_timing();
525 // Reset every 100 frames, so that local variations in frame times
526 // (especially for the first few frames, when the shaders are
527 // compiled etc.) don't make it hard to measure for the entire
528 // remaining duration of the program.
529 if (frame == 10000) {
537 resource_pool->clean_context();
540 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
543 glGenVertexArrays(1, &vao);
552 glBindVertexArray(vao);
556 GLuint fbo = resource_pool->create_fbo(dst_tex);
557 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
558 glViewport(0, 0, WIDTH/2, HEIGHT/2);
561 glUseProgram(cbcr_program_num);
564 glActiveTexture(GL_TEXTURE0);
566 glBindTexture(GL_TEXTURE_2D, src_tex);
568 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
570 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
572 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
575 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
576 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
578 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
579 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
581 glDrawArrays(GL_TRIANGLES, 0, 3);
584 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
585 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
590 resource_pool->release_fbo(fbo);
591 glDeleteVertexArrays(1, &vao);
594 void Mixer::release_display_frame(DisplayFrame *frame)
596 for (GLuint texnum : frame->temp_textures) {
597 resource_pool->release_2d_texture(texnum);
599 frame->temp_textures.clear();
600 frame->ready_fence.reset();
601 frame->input_frames.clear();
606 mixer_thread = thread(&Mixer::thread_func, this);
615 void Mixer::transition_clicked(int transition_num)
617 theme->transition_clicked(transition_num, pts());
620 void Mixer::channel_clicked(int preview_num)
622 theme->channel_clicked(preview_num);
625 Mixer::OutputChannel::~OutputChannel()
627 if (has_current_frame) {
628 parent->release_display_frame(¤t_frame);
630 if (has_ready_frame) {
631 parent->release_display_frame(&ready_frame);
635 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
637 // Store this frame for display. Remove the ready frame if any
638 // (it was seemingly never used).
640 unique_lock<mutex> lock(frame_mutex);
641 if (has_ready_frame) {
642 parent->release_display_frame(&ready_frame);
645 has_ready_frame = true;
648 if (has_new_frame_ready_callback) {
649 new_frame_ready_callback();
653 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
655 unique_lock<mutex> lock(frame_mutex);
656 if (!has_current_frame && !has_ready_frame) {
660 if (has_current_frame && has_ready_frame) {
661 // We have a new ready frame. Toss the current one.
662 parent->release_display_frame(¤t_frame);
663 has_current_frame = false;
665 if (has_ready_frame) {
666 assert(!has_current_frame);
667 current_frame = ready_frame;
668 ready_frame.ready_fence.reset(); // Drop the refcount.
669 ready_frame.input_frames.clear(); // Drop the refcounts.
670 has_current_frame = true;
671 has_ready_frame = false;
674 *frame = current_frame;
678 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
680 new_frame_ready_callback = callback;
681 has_new_frame_ready_callback = true;