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);
156 resource_pool->release_glsl_program(cbcr_program_num);
157 BMUSBCapture::stop_bm_thread();
159 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
161 unique_lock<mutex> lock(bmusb_mutex);
162 cards[card_index].should_quit = true; // Unblock thread.
163 cards[card_index].new_data_ready_changed.notify_all();
165 cards[card_index].usb->stop_dequeue_thread();
171 int unwrap_timecode(uint16_t current_wrapped, int last)
173 uint16_t last_wrapped = last & 0xffff;
174 if (current_wrapped > last_wrapped) {
175 return (last & ~0xffff) | current_wrapped;
177 return 0x10000 + ((last & ~0xffff) | current_wrapped);
181 float find_peak(const vector<float> &samples)
183 float m = fabs(samples[0]);
184 for (size_t i = 1; i < samples.size(); ++i) {
185 m = std::max(m, fabs(samples[i]));
190 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
192 size_t num_samples = in.size() / 2;
193 out_l->resize(num_samples);
194 out_r->resize(num_samples);
196 const float *inptr = in.data();
197 float *lptr = &(*out_l)[0];
198 float *rptr = &(*out_r)[0];
199 for (size_t i = 0; i < num_samples; ++i) {
207 void Mixer::bm_frame(int card_index, uint16_t timecode,
208 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
209 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
211 CaptureCard *card = &cards[card_index];
213 if (audio_frame.len - audio_offset > 30000) {
214 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",
215 card_index, int(audio_frame.len), int(audio_offset),
216 timecode, int(video_frame.len), int(video_offset), video_format);
217 if (video_frame.owner) {
218 video_frame.owner->release_frame(video_frame);
220 if (audio_frame.owner) {
221 audio_frame.owner->release_frame(audio_frame);
226 // Convert the audio to stereo fp32 and add it.
227 size_t num_samples = (audio_frame.len - audio_offset) / 8 / 3;
229 audio.resize(num_samples * 2);
230 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
232 int unwrapped_timecode = timecode;
233 int dropped_frames = 0;
234 if (card->last_timecode != -1) {
235 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
236 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
238 card->last_timecode = unwrapped_timecode;
242 unique_lock<mutex> lock(card->audio_mutex);
244 int unwrapped_timecode = timecode;
245 if (dropped_frames > 60 * 2) {
246 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
248 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
249 } else if (dropped_frames > 0) {
250 // Insert silence as needed.
251 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
252 card_index, dropped_frames, timecode);
253 vector<float> silence;
254 silence.resize((48000 / 60) * 2);
255 for (int i = 0; i < dropped_frames; ++i) {
256 card->resampler->add_input_samples((unwrapped_timecode - dropped_frames + i) / 60.0, silence.data(), (48000 / 60));
259 card->resampler->add_input_samples(unwrapped_timecode / 60.0, audio.data(), num_samples);
262 // Done with the audio, so release it.
263 if (audio_frame.owner) {
264 audio_frame.owner->release_frame(audio_frame);
268 // Wait until the previous frame was consumed.
269 unique_lock<mutex> lock(bmusb_mutex);
270 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
271 if (card->should_quit) return;
274 if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
275 if (video_frame.len != 0) {
276 printf("Card %d: Dropping video frame with wrong length (%ld)\n",
277 card_index, video_frame.len - video_offset);
279 if (video_frame.owner) {
280 video_frame.owner->release_frame(video_frame);
283 // Still send on the information that we _had_ a frame, even though it's corrupted,
284 // so that pts can go up accordingly.
286 unique_lock<mutex> lock(bmusb_mutex);
287 card->new_data_ready = true;
288 card->new_frame = RefCountedFrame(FrameAllocator::Frame());
289 card->new_data_ready_fence = nullptr;
290 card->dropped_frames = dropped_frames;
291 card->new_data_ready_changed.notify_all();
296 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
297 GLuint pbo = userdata->pbo;
299 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
301 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
303 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
306 // Upload the textures.
307 glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
309 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));
311 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
313 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
315 glBindTexture(GL_TEXTURE_2D, 0);
317 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
319 assert(fence != nullptr);
322 unique_lock<mutex> lock(bmusb_mutex);
323 card->new_data_ready = true;
324 card->new_frame = RefCountedFrame(video_frame);
325 card->new_data_ready_fence = fence;
326 card->dropped_frames = dropped_frames;
327 card->new_data_ready_changed.notify_all();
331 void Mixer::thread_func()
333 eglBindAPI(EGL_OPENGL_API);
334 QOpenGLContext *context = create_context();
335 if (!make_current(context, mixer_surface)) {
340 struct timespec start, now;
341 clock_gettime(CLOCK_MONOTONIC, &start);
344 int dropped_frames = 0;
346 while (!should_quit) {
347 CaptureCard card_copy[NUM_CARDS];
350 unique_lock<mutex> lock(bmusb_mutex);
352 // The first card is the master timer, so wait for it to have a new frame.
353 // TODO: Make configurable, and with a timeout.
354 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
356 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
357 CaptureCard *card = &cards[card_index];
358 card_copy[card_index].usb = card->usb;
359 card_copy[card_index].new_data_ready = card->new_data_ready;
360 card_copy[card_index].new_frame = card->new_frame;
361 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
362 card_copy[card_index].new_frame_audio = move(card->new_frame_audio);
363 card_copy[card_index].dropped_frames = card->dropped_frames;
364 card->new_data_ready = false;
365 card->new_data_ready_changed.notify_all();
369 // Resample the audio as needed, including from previously dropped frames.
370 vector<float> samples_out;
371 // TODO: Allow using audio from the other card(s) as well.
372 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
373 for (unsigned card_index = 0; card_index < NUM_CARDS; ++card_index) {
374 samples_out.resize((48000 / 60) * 2);
376 unique_lock<mutex> lock(cards[card_index].audio_mutex);
377 if (!cards[card_index].resampler->get_output_samples(pts(), &samples_out[0], 48000 / 60)) {
378 printf("Card %d reported previous underrun.\n", card_index);
381 if (card_index == 0) {
382 vector<float> left, right;
383 peak = std::max(peak, find_peak(samples_out));
384 deinterleave_samples(samples_out, &left, &right);
385 float *ptrs[] = { left.data(), right.data() };
386 r128.process(left.size(), ptrs);
387 h264_encoder->add_audio(pts_int, move(samples_out));
390 if (frame_num != card_copy[0].dropped_frames) {
391 // For dropped frames, increase the pts.
393 pts_int += TIMEBASE / 60;
397 if (audio_level_callback != nullptr) {
398 double loudness_s = r128.loudness_S();
399 double loudness_i = r128.integrated();
400 double loudness_range_low = r128.range_min();
401 double loudness_range_high = r128.range_max();
403 audio_level_callback(loudness_s, 20.0 * log10(peak),
404 loudness_i, loudness_range_low, loudness_range_high);
407 // If the first card is reporting a corrupted or otherwise dropped frame,
408 // just increase the pts (skipping over this frame) and don't try to compute anything new.
409 if (card_copy[0].new_frame->len == 0) {
411 pts_int += TIMEBASE / 60;
415 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
416 CaptureCard *card = &card_copy[card_index];
417 if (!card->new_data_ready || card->new_frame->len == 0)
420 assert(card->new_frame != nullptr);
421 bmusb_current_rendering_frame[card_index] = card->new_frame;
424 // The new texture might still be uploaded,
425 // tell the GPU to wait until it's there.
426 if (card->new_data_ready_fence) {
427 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
429 glDeleteSync(card->new_data_ready_fence);
432 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
433 theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
436 // Get the main chain from the theme, and set its state immediately.
437 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
438 EffectChain *chain = theme_main_chain.first;
439 theme_main_chain.second();
441 GLuint y_tex, cbcr_tex;
442 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
445 // Render main chain.
446 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
447 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
448 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
450 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
451 resource_pool->release_fbo(fbo);
453 subsample_chroma(cbcr_full_tex, cbcr_tex);
454 resource_pool->release_2d_texture(cbcr_full_tex);
456 // Set the right state for rgba_tex.
457 glBindFramebuffer(GL_FRAMEBUFFER, 0);
458 glBindTexture(GL_TEXTURE_2D, rgba_tex);
459 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
460 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
461 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
463 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
466 // Make sure the H.264 gets a reference to all the
467 // input frames needed, so that they are not released back
468 // until the rendering is done.
469 vector<RefCountedFrame> input_frames;
470 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
471 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
473 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampler.h. TODO: Make less hard-coded.
474 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
476 pts_int += TIMEBASE / 60;
478 // The live frame just shows the RGBA texture we just rendered.
479 // It owns rgba_tex now.
480 DisplayFrame live_frame;
481 live_frame.chain = display_chain.get();
482 live_frame.setup_chain = [this, rgba_tex]{
483 display_input->set_texture_num(rgba_tex);
485 live_frame.ready_fence = fence;
486 live_frame.input_frames = {};
487 live_frame.temp_textures = { rgba_tex };
488 output_channel[OUTPUT_LIVE].output_frame(live_frame);
490 // Set up preview and any additional channels.
491 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
492 DisplayFrame display_frame;
493 pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT); // FIXME: dimensions
494 display_frame.chain = chain.first;
495 display_frame.setup_chain = chain.second;
496 display_frame.ready_fence = fence;
497 display_frame.input_frames = { bmusb_current_rendering_frame[0], bmusb_current_rendering_frame[1] }; // FIXME: possible to do better?
498 display_frame.temp_textures = {};
499 output_channel[i].output_frame(display_frame);
502 clock_gettime(CLOCK_MONOTONIC, &now);
503 double elapsed = now.tv_sec - start.tv_sec +
504 1e-9 * (now.tv_nsec - start.tv_nsec);
505 if (frame % 100 == 0) {
506 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
507 frame, dropped_frames, elapsed, frame / elapsed,
508 1e3 * elapsed / frame);
509 // chain->print_phase_timing();
513 // Reset every 100 frames, so that local variations in frame times
514 // (especially for the first few frames, when the shaders are
515 // compiled etc.) don't make it hard to measure for the entire
516 // remaining duration of the program.
517 if (frame == 10000) {
525 resource_pool->clean_context();
528 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
531 glGenVertexArrays(1, &vao);
540 glBindVertexArray(vao);
544 GLuint fbo = resource_pool->create_fbo(dst_tex);
545 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
546 glViewport(0, 0, WIDTH/2, HEIGHT/2);
549 glUseProgram(cbcr_program_num);
552 glActiveTexture(GL_TEXTURE0);
554 glBindTexture(GL_TEXTURE_2D, src_tex);
556 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
558 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
560 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
563 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
564 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
566 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
567 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
569 glDrawArrays(GL_TRIANGLES, 0, 3);
572 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
573 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
578 resource_pool->release_fbo(fbo);
579 glDeleteVertexArrays(1, &vao);
582 void Mixer::release_display_frame(DisplayFrame *frame)
584 for (GLuint texnum : frame->temp_textures) {
585 resource_pool->release_2d_texture(texnum);
587 frame->temp_textures.clear();
588 frame->ready_fence.reset();
589 frame->input_frames.clear();
594 mixer_thread = thread(&Mixer::thread_func, this);
603 void Mixer::transition_clicked(int transition_num)
605 theme->transition_clicked(transition_num, pts());
608 void Mixer::channel_clicked(int preview_num)
610 theme->channel_clicked(preview_num);
613 Mixer::OutputChannel::~OutputChannel()
615 if (has_current_frame) {
616 parent->release_display_frame(¤t_frame);
618 if (has_ready_frame) {
619 parent->release_display_frame(&ready_frame);
623 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
625 // Store this frame for display. Remove the ready frame if any
626 // (it was seemingly never used).
628 unique_lock<mutex> lock(frame_mutex);
629 if (has_ready_frame) {
630 parent->release_display_frame(&ready_frame);
633 has_ready_frame = true;
636 if (has_new_frame_ready_callback) {
637 new_frame_ready_callback();
641 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
643 unique_lock<mutex> lock(frame_mutex);
644 if (!has_current_frame && !has_ready_frame) {
648 if (has_current_frame && has_ready_frame) {
649 // We have a new ready frame. Toss the current one.
650 parent->release_display_frame(¤t_frame);
651 has_current_frame = false;
653 if (has_ready_frame) {
654 assert(!has_current_frame);
655 current_frame = ready_frame;
656 ready_frame.ready_fence.reset(); // Drop the refcount.
657 ready_frame.input_frames.clear(); // Drop the refcounts.
658 has_current_frame = true;
659 has_ready_frame = false;
662 *frame = current_frame;
666 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
668 new_frame_ready_callback = callback;
669 has_new_frame_ready_callback = true;