11 #include <effect_chain.h>
12 #include <effect_util.h>
13 #include <epoxy/egl.h>
15 #include <image_format.h>
17 #include <overlay_effect.h>
18 #include <padding_effect.h>
19 #include <resample_effect.h>
20 #include <resource_pool.h>
21 #include <saturation_effect.h>
28 #include <white_balance_effect.h>
30 #include <ycbcr_input.h>
32 #include <condition_variable>
40 #include "bmusb/bmusb.h"
42 #include "h264encode.h"
43 #include "pbo_frame_allocator.h"
44 #include "ref_counted_gl_sync.h"
49 using namespace movit;
51 using namespace std::placeholders;
53 Mixer *global_mixer = nullptr;
57 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
59 for (size_t i = 0; i < num_samples; ++i) {
60 for (size_t j = 0; j < out_channels; ++j) {
64 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
65 dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
67 src += 3 * (in_channels - out_channels);
73 Mixer::Mixer(const QSurfaceFormat &format)
74 : httpd("test.ts", WIDTH, HEIGHT),
75 mixer_surface(create_surface(format)),
76 h264_encoder_surface(create_surface(format))
80 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
83 resource_pool.reset(new ResourcePool);
84 theme.reset(new Theme("theme.lua", resource_pool.get()));
85 output_channel[OUTPUT_LIVE].parent = this;
86 output_channel[OUTPUT_PREVIEW].parent = this;
87 output_channel[OUTPUT_INPUT0].parent = this;
88 output_channel[OUTPUT_INPUT1].parent = this;
90 ImageFormat inout_format;
91 inout_format.color_space = COLORSPACE_sRGB;
92 inout_format.gamma_curve = GAMMA_sRGB;
94 // Display chain; shows the live output produced by the main chain (its RGBA version).
95 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
97 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
98 display_chain->add_input(display_input);
99 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
100 display_chain->set_dither_bits(0); // Don't bother.
101 display_chain->finalize();
103 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
105 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
106 printf("Configuring card %d...\n", card_index);
107 CaptureCard *card = &cards[card_index];
108 card->usb = new BMUSBCapture(card_index);
109 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
110 card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44, WIDTH, HEIGHT));
111 card->usb->set_video_frame_allocator(card->frame_allocator.get());
112 card->surface = create_surface(format);
113 card->usb->set_dequeue_thread_callbacks(
115 eglBindAPI(EGL_OPENGL_API);
116 card->context = create_context();
117 if (!make_current(card->context, card->surface)) {
118 printf("failed to create bmusb context\n");
123 resource_pool->clean_context();
125 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
126 card->usb->configure_card();
129 BMUSBCapture::start_bm_thread();
131 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
132 cards[card_index].usb->start_bm_capture();
135 //chain->enable_phase_timing(true);
137 // Set up stuff for NV12 conversion.
140 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
141 string cbcr_frag_shader =
144 "uniform sampler2D cbcr_tex; \n"
146 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
148 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
150 r128_state = ebur128_init(2, 48000, EBUR128_MODE_SAMPLE_PEAK | EBUR128_MODE_M | EBUR128_MODE_S | EBUR128_MODE_I | EBUR128_MODE_LRA);
155 resource_pool->release_glsl_program(cbcr_program_num);
156 BMUSBCapture::stop_bm_thread();
158 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
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();
164 cards[card_index].usb->stop_dequeue_thread();
167 ebur128_destroy(&r128_state);
172 int unwrap_timecode(uint16_t current_wrapped, int last)
174 uint16_t last_wrapped = last & 0xffff;
175 if (current_wrapped > last_wrapped) {
176 return (last & ~0xffff) | current_wrapped;
178 return 0x10000 + ((last & ~0xffff) | current_wrapped);
184 void Mixer::bm_frame(int card_index, uint16_t timecode,
185 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
186 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
188 CaptureCard *card = &cards[card_index];
190 if (audio_frame.len - audio_offset > 30000) {
191 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",
192 card_index, int(audio_frame.len), int(audio_offset),
193 timecode, int(video_frame.len), int(video_offset), video_format);
194 if (video_frame.owner) {
195 video_frame.owner->release_frame(video_frame);
197 if (audio_frame.owner) {
198 audio_frame.owner->release_frame(audio_frame);
203 // Convert the audio to stereo fp32 and add it.
204 size_t num_samples = (audio_frame.len - audio_offset) / 8 / 3;
206 audio.resize(num_samples * 2);
207 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
209 int unwrapped_timecode = timecode;
210 int dropped_frames = 0;
211 if (card->last_timecode != -1) {
212 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
213 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
215 card->last_timecode = unwrapped_timecode;
219 unique_lock<mutex> lock(card->audio_mutex);
221 int unwrapped_timecode = timecode;
222 if (dropped_frames > 60 * 2) {
223 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
225 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
226 } else if (dropped_frames > 0) {
227 // Insert silence as needed.
228 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
229 card_index, dropped_frames, timecode);
230 vector<float> silence;
231 silence.resize((48000 / 60) * 2);
232 for (int i = 0; i < dropped_frames; ++i) {
233 card->resampler->add_input_samples((unwrapped_timecode - dropped_frames + i) / 60.0, silence.data(), (48000 / 60));
236 card->resampler->add_input_samples(unwrapped_timecode / 60.0, audio.data(), num_samples);
239 // Done with the audio, so release it.
240 if (audio_frame.owner) {
241 audio_frame.owner->release_frame(audio_frame);
245 // Wait until the previous frame was consumed.
246 unique_lock<mutex> lock(bmusb_mutex);
247 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
248 if (card->should_quit) return;
251 if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
252 if (video_frame.len != 0) {
253 printf("Card %d: Dropping video frame with wrong length (%ld)\n",
254 card_index, video_frame.len - video_offset);
256 if (video_frame.owner) {
257 video_frame.owner->release_frame(video_frame);
260 // Still send on the information that we _had_ a frame, even though it's corrupted,
261 // so that pts can go up accordingly.
263 unique_lock<mutex> lock(bmusb_mutex);
264 card->new_data_ready = true;
265 card->new_frame = RefCountedFrame(FrameAllocator::Frame());
266 card->new_data_ready_fence = nullptr;
267 card->dropped_frames = dropped_frames;
268 card->new_data_ready_changed.notify_all();
273 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
274 GLuint pbo = userdata->pbo;
276 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
278 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
280 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
283 // Upload the textures.
284 glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
286 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));
288 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
290 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
292 glBindTexture(GL_TEXTURE_2D, 0);
294 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
296 assert(fence != nullptr);
299 unique_lock<mutex> lock(bmusb_mutex);
300 card->new_data_ready = true;
301 card->new_frame = RefCountedFrame(video_frame);
302 card->new_data_ready_fence = fence;
303 card->dropped_frames = dropped_frames;
304 card->new_data_ready_changed.notify_all();
308 void Mixer::thread_func()
310 eglBindAPI(EGL_OPENGL_API);
311 QOpenGLContext *context = create_context();
312 if (!make_current(context, mixer_surface)) {
317 struct timespec start, now;
318 clock_gettime(CLOCK_MONOTONIC, &start);
321 int dropped_frames = 0;
323 while (!should_quit) {
324 CaptureCard card_copy[NUM_CARDS];
327 unique_lock<mutex> lock(bmusb_mutex);
329 // The first card is the master timer, so wait for it to have a new frame.
330 // TODO: Make configurable, and with a timeout.
331 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
333 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
334 CaptureCard *card = &cards[card_index];
335 card_copy[card_index].usb = card->usb;
336 card_copy[card_index].new_data_ready = card->new_data_ready;
337 card_copy[card_index].new_frame = card->new_frame;
338 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
339 card_copy[card_index].new_frame_audio = move(card->new_frame_audio);
340 card_copy[card_index].dropped_frames = card->dropped_frames;
341 card->new_data_ready = false;
342 card->new_data_ready_changed.notify_all();
346 // Resample the audio as needed, including from previously dropped frames.
347 vector<float> samples_out;
348 // TODO: Allow using audio from the other card(s) as well.
349 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
350 for (unsigned card_index = 0; card_index < NUM_CARDS; ++card_index) {
351 samples_out.resize((48000 / 60) * 2);
353 unique_lock<mutex> lock(cards[card_index].audio_mutex);
354 if (!cards[card_index].resampler->get_output_samples(pts(), &samples_out[0], 48000 / 60)) {
355 printf("Card %d reported previous underrun.\n", card_index);
358 if (card_index == 0) {
359 ebur128_add_frames_float(r128_state, samples_out.data(), samples_out.size() / 2);
360 h264_encoder->add_audio(pts_int, move(samples_out));
363 if (frame_num != card_copy[0].dropped_frames) {
364 // For dropped frames, increase the pts.
366 pts_int += TIMEBASE / 60;
370 if (audio_level_callback != nullptr) {
371 double loudness_s, peak_level_l, peak_level_r;
372 ebur128_loudness_shortterm(r128_state, &loudness_s);
373 ebur128_sample_peak(r128_state, 0, &peak_level_l);
374 ebur128_sample_peak(r128_state, 1, &peak_level_r);
375 audio_level_callback(loudness_s, 20.0 * log10(max(peak_level_l, peak_level_r)));
378 // If the first card is reporting a corrupted or otherwise dropped frame,
379 // just increase the pts (skipping over this frame) and don't try to compute anything new.
380 if (card_copy[0].new_frame->len == 0) {
382 pts_int += TIMEBASE / 60;
386 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
387 CaptureCard *card = &card_copy[card_index];
388 if (!card->new_data_ready || card->new_frame->len == 0)
391 assert(card->new_frame != nullptr);
392 bmusb_current_rendering_frame[card_index] = card->new_frame;
395 // The new texture might still be uploaded,
396 // tell the GPU to wait until it's there.
397 if (card->new_data_ready_fence) {
398 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
400 glDeleteSync(card->new_data_ready_fence);
403 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
404 theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
407 // Get the main chain from the theme, and set its state immediately.
408 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
409 EffectChain *chain = theme_main_chain.first;
410 theme_main_chain.second();
412 GLuint y_tex, cbcr_tex;
413 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
416 // Render main chain.
417 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
418 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
419 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
420 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
421 resource_pool->release_fbo(fbo);
423 subsample_chroma(cbcr_full_tex, cbcr_tex);
424 resource_pool->release_2d_texture(cbcr_full_tex);
426 // Set the right state for rgba_tex.
427 glBindFramebuffer(GL_FRAMEBUFFER, 0);
428 glBindTexture(GL_TEXTURE_2D, rgba_tex);
429 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
430 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
431 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
433 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
436 // Make sure the H.264 gets a reference to all the
437 // input frames needed, so that they are not released back
438 // until the rendering is done.
439 vector<RefCountedFrame> input_frames;
440 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
441 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
443 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampler.h. TODO: Make less hard-coded.
444 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
446 pts_int += TIMEBASE / 60;
448 // The live frame just shows the RGBA texture we just rendered.
449 // It owns rgba_tex now.
450 DisplayFrame live_frame;
451 live_frame.chain = display_chain.get();
452 live_frame.setup_chain = [this, rgba_tex]{
453 display_input->set_texture_num(rgba_tex);
455 live_frame.ready_fence = fence;
456 live_frame.input_frames = {};
457 live_frame.temp_textures = { rgba_tex };
458 output_channel[OUTPUT_LIVE].output_frame(live_frame);
460 // Set up preview and any additional channels.
461 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
462 DisplayFrame display_frame;
463 pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT); // FIXME: dimensions
464 display_frame.chain = chain.first;
465 display_frame.setup_chain = chain.second;
466 display_frame.ready_fence = fence;
467 display_frame.input_frames = { bmusb_current_rendering_frame[0], bmusb_current_rendering_frame[1] }; // FIXME: possible to do better?
468 display_frame.temp_textures = {};
469 output_channel[i].output_frame(display_frame);
472 clock_gettime(CLOCK_MONOTONIC, &now);
473 double elapsed = now.tv_sec - start.tv_sec +
474 1e-9 * (now.tv_nsec - start.tv_nsec);
475 if (frame % 100 == 0) {
476 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
477 frame, dropped_frames, elapsed, frame / elapsed,
478 1e3 * elapsed / frame);
479 // chain->print_phase_timing();
483 // Reset every 100 frames, so that local variations in frame times
484 // (especially for the first few frames, when the shaders are
485 // compiled etc.) don't make it hard to measure for the entire
486 // remaining duration of the program.
487 if (frame == 10000) {
495 resource_pool->clean_context();
498 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
501 glGenVertexArrays(1, &vao);
510 glBindVertexArray(vao);
514 GLuint fbo = resource_pool->create_fbo(dst_tex);
515 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
516 glViewport(0, 0, WIDTH/2, HEIGHT/2);
519 glUseProgram(cbcr_program_num);
522 glActiveTexture(GL_TEXTURE0);
524 glBindTexture(GL_TEXTURE_2D, src_tex);
526 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
528 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
530 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
533 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
534 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
536 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
537 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
539 glDrawArrays(GL_TRIANGLES, 0, 3);
542 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
543 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
548 resource_pool->release_fbo(fbo);
549 glDeleteVertexArrays(1, &vao);
552 void Mixer::release_display_frame(DisplayFrame *frame)
554 for (GLuint texnum : frame->temp_textures) {
555 resource_pool->release_2d_texture(texnum);
557 frame->temp_textures.clear();
558 frame->ready_fence.reset();
559 frame->input_frames.clear();
564 mixer_thread = thread(&Mixer::thread_func, this);
573 void Mixer::transition_clicked(int transition_num)
575 theme->transition_clicked(transition_num, pts());
578 void Mixer::channel_clicked(int preview_num)
580 theme->channel_clicked(preview_num);
583 Mixer::OutputChannel::~OutputChannel()
585 if (has_current_frame) {
586 parent->release_display_frame(¤t_frame);
588 if (has_ready_frame) {
589 parent->release_display_frame(&ready_frame);
593 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
595 // Store this frame for display. Remove the ready frame if any
596 // (it was seemingly never used).
598 unique_lock<mutex> lock(frame_mutex);
599 if (has_ready_frame) {
600 parent->release_display_frame(&ready_frame);
603 has_ready_frame = true;
606 if (has_new_frame_ready_callback) {
607 new_frame_ready_callback();
611 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
613 unique_lock<mutex> lock(frame_mutex);
614 if (!has_current_frame && !has_ready_frame) {
618 if (has_current_frame && has_ready_frame) {
619 // We have a new ready frame. Toss the current one.
620 parent->release_display_frame(¤t_frame);
621 has_current_frame = false;
623 if (has_ready_frame) {
624 assert(!has_current_frame);
625 current_frame = ready_frame;
626 ready_frame.ready_fence.reset(); // Drop the refcount.
627 ready_frame.input_frames.clear(); // Drop the refcounts.
628 has_current_frame = true;
629 has_ready_frame = false;
632 *frame = current_frame;
636 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
638 new_frame_ready_callback = callback;
639 has_new_frame_ready_callback = true;