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, unsigned num_cards)
74 : httpd("test.ts", WIDTH, HEIGHT),
76 mixer_surface(create_surface(format)),
77 h264_encoder_surface(create_surface(format))
81 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
84 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
85 // will be halved when sampling them, and we need to compensate here.
86 movit_texel_subpixel_precision /= 2.0;
88 resource_pool.reset(new ResourcePool);
89 theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
90 output_channel[OUTPUT_LIVE].parent = this;
91 output_channel[OUTPUT_PREVIEW].parent = this;
92 output_channel[OUTPUT_INPUT0].parent = this;
93 output_channel[OUTPUT_INPUT1].parent = this;
95 ImageFormat inout_format;
96 inout_format.color_space = COLORSPACE_sRGB;
97 inout_format.gamma_curve = GAMMA_sRGB;
99 // Display chain; shows the live output produced by the main chain (its RGBA version).
100 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
102 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
103 display_chain->add_input(display_input);
104 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
105 display_chain->set_dither_bits(0); // Don't bother.
106 display_chain->finalize();
108 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
110 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
111 printf("Configuring card %d...\n", card_index);
112 CaptureCard *card = &cards[card_index];
113 card->usb = new BMUSBCapture(card_index);
114 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
115 card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44, WIDTH, HEIGHT));
116 card->usb->set_video_frame_allocator(card->frame_allocator.get());
117 card->surface = create_surface(format);
118 card->usb->set_dequeue_thread_callbacks(
120 eglBindAPI(EGL_OPENGL_API);
121 card->context = create_context();
122 if (!make_current(card->context, card->surface)) {
123 printf("failed to create bmusb context\n");
128 resource_pool->clean_context();
130 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
131 card->usb->configure_card();
134 BMUSBCapture::start_bm_thread();
136 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
137 cards[card_index].usb->start_bm_capture();
140 //chain->enable_phase_timing(true);
142 // Set up stuff for NV12 conversion.
145 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
146 string cbcr_frag_shader =
149 "uniform sampler2D cbcr_tex; \n"
151 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
153 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
161 resource_pool->release_glsl_program(cbcr_program_num);
162 BMUSBCapture::stop_bm_thread();
164 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
166 unique_lock<mutex> lock(bmusb_mutex);
167 cards[card_index].should_quit = true; // Unblock thread.
168 cards[card_index].new_data_ready_changed.notify_all();
170 cards[card_index].usb->stop_dequeue_thread();
176 int unwrap_timecode(uint16_t current_wrapped, int last)
178 uint16_t last_wrapped = last & 0xffff;
179 if (current_wrapped > last_wrapped) {
180 return (last & ~0xffff) | current_wrapped;
182 return 0x10000 + ((last & ~0xffff) | current_wrapped);
186 float find_peak(const vector<float> &samples)
188 float m = fabs(samples[0]);
189 for (size_t i = 1; i < samples.size(); ++i) {
190 m = std::max(m, fabs(samples[i]));
195 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
197 size_t num_samples = in.size() / 2;
198 out_l->resize(num_samples);
199 out_r->resize(num_samples);
201 const float *inptr = in.data();
202 float *lptr = &(*out_l)[0];
203 float *rptr = &(*out_r)[0];
204 for (size_t i = 0; i < num_samples; ++i) {
212 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
213 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
214 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
216 CaptureCard *card = &cards[card_index];
218 if (audio_frame.len - audio_offset > 30000) {
219 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",
220 card_index, int(audio_frame.len), int(audio_offset),
221 timecode, int(video_frame.len), int(video_offset), video_format);
222 if (video_frame.owner) {
223 video_frame.owner->release_frame(video_frame);
225 if (audio_frame.owner) {
226 audio_frame.owner->release_frame(audio_frame);
231 int unwrapped_timecode = timecode;
232 int dropped_frames = 0;
233 if (card->last_timecode != -1) {
234 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
235 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
237 card->last_timecode = unwrapped_timecode;
239 // Convert the audio to stereo fp32 and add it.
240 size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
242 audio.resize(num_samples * 2);
243 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
247 unique_lock<mutex> lock(card->audio_mutex);
249 int unwrapped_timecode = timecode;
250 if (dropped_frames > 60 * 2) {
251 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
253 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
254 } else if (dropped_frames > 0) {
255 // Insert silence as needed.
256 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
257 card_index, dropped_frames, timecode);
258 vector<float> silence;
259 silence.resize((48000 / 60) * 2);
260 for (int i = 0; i < dropped_frames; ++i) {
261 card->resampler->add_input_samples((unwrapped_timecode - dropped_frames + i) / 60.0, silence.data(), (48000 / 60));
264 card->resampler->add_input_samples(unwrapped_timecode / 60.0, audio.data(), num_samples);
267 // Done with the audio, so release it.
268 if (audio_frame.owner) {
269 audio_frame.owner->release_frame(audio_frame);
273 // Wait until the previous frame was consumed.
274 unique_lock<mutex> lock(bmusb_mutex);
275 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
276 if (card->should_quit) return;
279 if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
280 if (video_frame.len != 0) {
281 printf("Card %d: Dropping video frame with wrong length (%ld)\n",
282 card_index, video_frame.len - video_offset);
284 if (video_frame.owner) {
285 video_frame.owner->release_frame(video_frame);
288 // Still send on the information that we _had_ a frame, even though it's corrupted,
289 // so that pts can go up accordingly.
291 unique_lock<mutex> lock(bmusb_mutex);
292 card->new_data_ready = true;
293 card->new_frame = RefCountedFrame(FrameAllocator::Frame());
294 card->new_data_ready_fence = nullptr;
295 card->dropped_frames = dropped_frames;
296 card->new_data_ready_changed.notify_all();
301 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
302 GLuint pbo = userdata->pbo;
304 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
306 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
308 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
311 // Upload the textures.
312 glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
314 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));
316 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
318 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
320 glBindTexture(GL_TEXTURE_2D, 0);
322 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
324 assert(fence != nullptr);
327 unique_lock<mutex> lock(bmusb_mutex);
328 card->new_data_ready = true;
329 card->new_frame = RefCountedFrame(video_frame);
330 card->new_data_ready_fence = fence;
331 card->dropped_frames = dropped_frames;
332 card->new_data_ready_changed.notify_all();
336 void Mixer::thread_func()
338 eglBindAPI(EGL_OPENGL_API);
339 QOpenGLContext *context = create_context();
340 if (!make_current(context, mixer_surface)) {
345 struct timespec start, now;
346 clock_gettime(CLOCK_MONOTONIC, &start);
349 int dropped_frames = 0;
351 while (!should_quit) {
352 CaptureCard card_copy[MAX_CARDS];
355 unique_lock<mutex> lock(bmusb_mutex);
357 // The first card is the master timer, so wait for it to have a new frame.
358 // TODO: Make configurable, and with a timeout.
359 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
361 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
362 CaptureCard *card = &cards[card_index];
363 card_copy[card_index].usb = card->usb;
364 card_copy[card_index].new_data_ready = card->new_data_ready;
365 card_copy[card_index].new_frame = card->new_frame;
366 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
367 card_copy[card_index].new_frame_audio = move(card->new_frame_audio);
368 card_copy[card_index].dropped_frames = card->dropped_frames;
369 card->new_data_ready = false;
370 card->new_data_ready_changed.notify_all();
374 // Resample the audio as needed, including from previously dropped frames.
375 vector<float> samples_out;
376 // TODO: Allow using audio from the other card(s) as well.
377 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
378 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
379 samples_out.resize((48000 / 60) * 2);
381 unique_lock<mutex> lock(cards[card_index].audio_mutex);
382 if (!cards[card_index].resampler->get_output_samples(pts(), &samples_out[0], 48000 / 60)) {
383 printf("Card %d reported previous underrun.\n", card_index);
386 if (card_index == 0) {
387 vector<float> left, right;
388 peak = std::max(peak, find_peak(samples_out));
389 deinterleave_samples(samples_out, &left, &right);
390 float *ptrs[] = { left.data(), right.data() };
391 r128.process(left.size(), ptrs);
392 h264_encoder->add_audio(pts_int, move(samples_out));
395 if (frame_num != card_copy[0].dropped_frames) {
396 // For dropped frames, increase the pts.
398 pts_int += TIMEBASE / 60;
402 if (audio_level_callback != nullptr) {
403 double loudness_s = r128.loudness_S();
404 double loudness_i = r128.integrated();
405 double loudness_range_low = r128.range_min();
406 double loudness_range_high = r128.range_max();
408 audio_level_callback(loudness_s, 20.0 * log10(peak),
409 loudness_i, loudness_range_low, loudness_range_high);
412 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
413 if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
414 ++card_copy[card_index].dropped_frames;
416 if (card_copy[card_index].dropped_frames > 0) {
417 printf("Card %u dropped %d frames before this\n",
418 card_index, int(card_copy[card_index].dropped_frames));
422 // If the first card is reporting a corrupted or otherwise dropped frame,
423 // just increase the pts (skipping over this frame) and don't try to compute anything new.
424 if (card_copy[0].new_frame->len == 0) {
426 pts_int += TIMEBASE / 60;
430 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
431 CaptureCard *card = &card_copy[card_index];
432 if (!card->new_data_ready || card->new_frame->len == 0)
435 assert(card->new_frame != nullptr);
436 bmusb_current_rendering_frame[card_index] = card->new_frame;
439 // The new texture might still be uploaded,
440 // tell the GPU to wait until it's there.
441 if (card->new_data_ready_fence) {
442 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
444 glDeleteSync(card->new_data_ready_fence);
447 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
448 theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
451 // Get the main chain from the theme, and set its state immediately.
452 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
453 EffectChain *chain = theme_main_chain.first;
454 theme_main_chain.second();
456 GLuint y_tex, cbcr_tex;
457 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
460 // Render main chain.
461 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
462 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
463 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
465 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
466 resource_pool->release_fbo(fbo);
468 subsample_chroma(cbcr_full_tex, cbcr_tex);
469 resource_pool->release_2d_texture(cbcr_full_tex);
471 // Set the right state for rgba_tex.
472 glBindFramebuffer(GL_FRAMEBUFFER, 0);
473 glBindTexture(GL_TEXTURE_2D, rgba_tex);
474 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
475 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
476 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
478 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
481 // Make sure the H.264 gets a reference to all the
482 // input frames needed, so that they are not released back
483 // until the rendering is done.
484 vector<RefCountedFrame> input_frames;
485 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
486 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
488 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampler.h. TODO: Make less hard-coded.
489 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
491 pts_int += TIMEBASE / 60;
493 // The live frame just shows the RGBA texture we just rendered.
494 // It owns rgba_tex now.
495 DisplayFrame live_frame;
496 live_frame.chain = display_chain.get();
497 live_frame.setup_chain = [this, rgba_tex]{
498 display_input->set_texture_num(rgba_tex);
500 live_frame.ready_fence = fence;
501 live_frame.input_frames = {};
502 live_frame.temp_textures = { rgba_tex };
503 output_channel[OUTPUT_LIVE].output_frame(live_frame);
505 // Set up preview and any additional channels.
506 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
507 DisplayFrame display_frame;
508 pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT); // FIXME: dimensions
509 display_frame.chain = chain.first;
510 display_frame.setup_chain = chain.second;
511 display_frame.ready_fence = fence;
513 // FIXME: possible to do better?
514 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
515 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
517 display_frame.temp_textures = {};
518 output_channel[i].output_frame(display_frame);
521 clock_gettime(CLOCK_MONOTONIC, &now);
522 double elapsed = now.tv_sec - start.tv_sec +
523 1e-9 * (now.tv_nsec - start.tv_nsec);
524 if (frame % 100 == 0) {
525 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
526 frame, dropped_frames, elapsed, frame / elapsed,
527 1e3 * elapsed / frame);
528 // chain->print_phase_timing();
532 // Reset every 100 frames, so that local variations in frame times
533 // (especially for the first few frames, when the shaders are
534 // compiled etc.) don't make it hard to measure for the entire
535 // remaining duration of the program.
536 if (frame == 10000) {
544 resource_pool->clean_context();
547 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
550 glGenVertexArrays(1, &vao);
559 glBindVertexArray(vao);
563 GLuint fbo = resource_pool->create_fbo(dst_tex);
564 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
565 glViewport(0, 0, WIDTH/2, HEIGHT/2);
568 glUseProgram(cbcr_program_num);
571 glActiveTexture(GL_TEXTURE0);
573 glBindTexture(GL_TEXTURE_2D, src_tex);
575 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
577 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
579 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
582 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
583 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
585 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
586 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
588 glDrawArrays(GL_TRIANGLES, 0, 3);
591 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
592 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
597 resource_pool->release_fbo(fbo);
598 glDeleteVertexArrays(1, &vao);
601 void Mixer::release_display_frame(DisplayFrame *frame)
603 for (GLuint texnum : frame->temp_textures) {
604 resource_pool->release_2d_texture(texnum);
606 frame->temp_textures.clear();
607 frame->ready_fence.reset();
608 frame->input_frames.clear();
613 mixer_thread = thread(&Mixer::thread_func, this);
622 void Mixer::transition_clicked(int transition_num)
624 theme->transition_clicked(transition_num, pts());
627 void Mixer::channel_clicked(int preview_num)
629 theme->channel_clicked(preview_num);
632 Mixer::OutputChannel::~OutputChannel()
634 if (has_current_frame) {
635 parent->release_display_frame(¤t_frame);
637 if (has_ready_frame) {
638 parent->release_display_frame(&ready_frame);
642 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
644 // Store this frame for display. Remove the ready frame if any
645 // (it was seemingly never used).
647 unique_lock<mutex> lock(frame_mutex);
648 if (has_ready_frame) {
649 parent->release_display_frame(&ready_frame);
652 has_ready_frame = true;
655 if (has_new_frame_ready_callback) {
656 new_frame_ready_callback();
660 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
662 unique_lock<mutex> lock(frame_mutex);
663 if (!has_current_frame && !has_ready_frame) {
667 if (has_current_frame && has_ready_frame) {
668 // We have a new ready frame. Toss the current one.
669 parent->release_display_frame(¤t_frame);
670 has_current_frame = false;
672 if (has_ready_frame) {
673 assert(!has_current_frame);
674 current_frame = ready_frame;
675 ready_frame.ready_fence.reset(); // Drop the refcount.
676 ready_frame.input_frames.clear(); // Drop the refcounts.
677 has_current_frame = true;
678 has_ready_frame = false;
681 *frame = current_frame;
685 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
687 new_frame_ready_callback = callback;
688 has_new_frame_ready_callback = true;