10 #include <effect_chain.h>
11 #include <effect_util.h>
12 #include <epoxy/egl.h>
14 #include <image_format.h>
16 #include <overlay_effect.h>
17 #include <padding_effect.h>
18 #include <resample_effect.h>
19 #include <resource_pool.h>
20 #include <saturation_effect.h>
27 #include <white_balance_effect.h>
29 #include <ycbcr_input.h>
31 #include <condition_variable>
39 #include "bmusb/bmusb.h"
41 #include "h264encode.h"
42 #include "pbo_frame_allocator.h"
43 #include "ref_counted_gl_sync.h"
48 using namespace movit;
50 using namespace std::placeholders;
52 Mixer *global_mixer = nullptr;
56 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
58 for (size_t i = 0; i < num_samples; ++i) {
59 for (size_t j = 0; j < out_channels; ++j) {
63 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
64 dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
66 src += 3 * (in_channels - out_channels);
72 Mixer::Mixer(const QSurfaceFormat &format)
73 : mixer_surface(create_surface(format)),
74 h264_encoder_surface(create_surface(format))
76 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
79 resource_pool.reset(new ResourcePool);
80 theme.reset(new Theme("theme.lua", resource_pool.get()));
81 output_channel[OUTPUT_LIVE].parent = this;
82 output_channel[OUTPUT_PREVIEW].parent = this;
83 output_channel[OUTPUT_INPUT0].parent = this;
84 output_channel[OUTPUT_INPUT1].parent = this;
86 ImageFormat inout_format;
87 inout_format.color_space = COLORSPACE_sRGB;
88 inout_format.gamma_curve = GAMMA_sRGB;
90 // Display chain; shows the live output produced by the main chain (its RGBA version).
91 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
93 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
94 display_chain->add_input(display_input);
95 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
96 display_chain->set_dither_bits(0); // Don't bother.
97 display_chain->finalize();
99 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, "test.mp4"));
101 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
102 printf("Configuring card %d...\n", card_index);
103 CaptureCard *card = &cards[card_index];
104 card->usb = new BMUSBCapture(0x1edb, card_index == 0 ? 0xbd3b : 0xbd4f);
105 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
106 card->frame_allocator.reset(new PBOFrameAllocator(1280 * 750 * 2 + 44, 1280, 720));
107 card->usb->set_video_frame_allocator(card->frame_allocator.get());
108 card->surface = create_surface(format);
109 card->usb->set_dequeue_thread_callbacks(
111 eglBindAPI(EGL_OPENGL_API);
112 card->context = create_context();
113 if (!make_current(card->context, card->surface)) {
114 printf("failed to create bmusb context\n");
120 resource_pool->clean_context();
122 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
123 card->usb->configure_card();
126 BMUSBCapture::start_bm_thread();
128 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
129 cards[card_index].usb->start_bm_capture();
132 //chain->enable_phase_timing(true);
134 // Set up stuff for NV12 conversion.
137 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
138 string cbcr_frag_shader =
141 "uniform sampler2D cbcr_tex; \n"
143 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
145 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
150 resource_pool->release_glsl_program(cbcr_program_num);
151 BMUSBCapture::stop_bm_thread();
153 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
155 unique_lock<mutex> lock(bmusb_mutex);
156 cards[card_index].should_quit = true; // Unblock thread.
157 cards[card_index].new_data_ready_changed.notify_all();
159 cards[card_index].usb->stop_dequeue_thread();
165 int unwrap_timecode(uint16_t current_wrapped, int last)
167 uint16_t last_wrapped = last & 0xffff;
168 if (current_wrapped > last_wrapped) {
169 return (last & ~0xffff) | current_wrapped;
171 return 0x10000 + ((last & ~0xffff) | current_wrapped);
177 void Mixer::bm_frame(int card_index, uint16_t timecode,
178 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
179 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
181 CaptureCard *card = &cards[card_index];
183 if (audio_frame.len - audio_offset > 30000) {
184 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",
185 card_index, int(audio_frame.len), int(audio_offset),
186 timecode, int(video_frame.len), int(video_offset), video_format);
187 if (video_frame.owner) {
188 video_frame.owner->release_frame(video_frame);
190 if (audio_frame.owner) {
191 audio_frame.owner->release_frame(audio_frame);
196 // Convert the audio to stereo fp32 and add it.
197 size_t num_samples = (audio_frame.len - audio_offset) / 8 / 3;
199 audio.resize(num_samples * 2);
200 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
202 int unwrapped_timecode = timecode;
203 int dropped_frames = 0;
204 if (card->last_timecode != -1) {
205 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
206 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
208 card->last_timecode = unwrapped_timecode;
212 unique_lock<mutex> lock(card->audio_mutex);
214 int unwrapped_timecode = timecode;
215 if (dropped_frames > 60 * 2) {
216 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
218 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
219 } else if (dropped_frames > 0) {
220 // Insert silence as needed.
221 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
222 card_index, dropped_frames, timecode);
223 vector<float> silence;
224 silence.resize((48000 / 60) * 2);
225 for (int i = 0; i < dropped_frames; ++i) {
226 card->resampler->add_input_samples((unwrapped_timecode - dropped_frames + i) / 60.0, silence.data(), (48000 / 60));
229 card->resampler->add_input_samples(unwrapped_timecode / 60.0, audio.data(), num_samples);
232 // Done with the audio, so release it.
233 if (audio_frame.owner) {
234 audio_frame.owner->release_frame(audio_frame);
238 // Wait until the previous frame was consumed.
239 unique_lock<mutex> lock(bmusb_mutex);
240 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
241 if (card->should_quit) return;
244 if (video_frame.len - video_offset != 1280 * 750 * 2) {
245 if (video_frame.len != 0) {
246 printf("Card %d: Dropping video frame with wrong length (%ld)\n",
247 card_index, video_frame.len - video_offset);
249 if (video_frame.owner) {
250 video_frame.owner->release_frame(video_frame);
253 // Still send on the information that we _had_ a frame, even though it's corrupted,
254 // so that pts can go up accordingly.
256 unique_lock<mutex> lock(bmusb_mutex);
257 card->new_data_ready = true;
258 card->new_frame = RefCountedFrame(FrameAllocator::Frame());
259 card->new_data_ready_fence = nullptr;
260 card->dropped_frames = dropped_frames;
261 card->new_data_ready_changed.notify_all();
266 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
267 GLuint pbo = userdata->pbo;
269 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
271 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
273 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
276 // Upload the textures.
277 glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
279 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280, 720, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((1280 * 750 * 2 + 44) / 2 + 1280 * 25 + 22));
281 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
283 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280/2, 720, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(1280 * 25 + 22));
285 glBindTexture(GL_TEXTURE_2D, 0);
287 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
289 assert(fence != nullptr);
292 unique_lock<mutex> lock(bmusb_mutex);
293 card->new_data_ready = true;
294 card->new_frame = RefCountedFrame(video_frame);
295 card->new_data_ready_fence = fence;
296 card->dropped_frames = dropped_frames;
297 card->new_data_ready_changed.notify_all();
301 void Mixer::thread_func()
303 eglBindAPI(EGL_OPENGL_API);
304 QOpenGLContext *context = create_context();
305 if (!make_current(context, mixer_surface)) {
310 struct timespec start, now;
311 clock_gettime(CLOCK_MONOTONIC, &start);
314 int dropped_frames = 0;
316 while (!should_quit) {
317 CaptureCard card_copy[NUM_CARDS];
320 unique_lock<mutex> lock(bmusb_mutex);
322 // The first card is the master timer, so wait for it to have a new frame.
323 // TODO: Make configurable, and with a timeout.
324 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
326 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
327 CaptureCard *card = &cards[card_index];
328 card_copy[card_index].usb = card->usb;
329 card_copy[card_index].new_data_ready = card->new_data_ready;
330 card_copy[card_index].new_frame = card->new_frame;
331 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
332 card_copy[card_index].new_frame_audio = move(card->new_frame_audio);
333 card_copy[card_index].dropped_frames = card->dropped_frames;
334 card->new_data_ready = false;
335 card->new_data_ready_changed.notify_all();
339 // Resample the audio as needed, including from previously dropped frames.
340 vector<float> samples_out;
341 // TODO: Allow using audio from the other card(s) as well.
342 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
343 for (unsigned card_index = 0; card_index < NUM_CARDS; ++card_index) {
344 samples_out.resize((48000 / 60) * 2);
346 unique_lock<mutex> lock(cards[card_index].audio_mutex);
347 if (!cards[card_index].resampler->get_output_samples(pts(), &samples_out[0], 48000 / 60)) {
348 printf("Card %d reported previous underrun.\n", card_index);
351 if (card_index == 0) {
352 h264_encoder->add_audio(pts_int, move(samples_out));
355 if (frame_num != card_copy[0].dropped_frames) {
356 // For dropped frames, increase the pts.
358 pts_int += TIMEBASE / 60;
362 // If the first card is reporting a corrupted or otherwise dropped frame,
363 // just increase the pts (skipping over this frame) and don't try to compute anything new.
364 if (card_copy[0].new_frame->len == 0) {
366 pts_int += TIMEBASE / 60;
370 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
371 CaptureCard *card = &card_copy[card_index];
372 if (!card->new_data_ready || card->new_frame->len == 0)
375 assert(card->new_frame != nullptr);
376 bmusb_current_rendering_frame[card_index] = card->new_frame;
379 // The new texture might still be uploaded,
380 // tell the GPU to wait until it's there.
381 if (card->new_data_ready_fence) {
382 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
384 glDeleteSync(card->new_data_ready_fence);
387 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
388 theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
391 // Get the main chain from the theme, and set its state immediately.
392 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
393 EffectChain *chain = theme_main_chain.first;
394 theme_main_chain.second();
396 GLuint y_tex, cbcr_tex;
397 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
400 // Render main chain.
401 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
402 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
403 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
404 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
405 resource_pool->release_fbo(fbo);
407 subsample_chroma(cbcr_full_tex, cbcr_tex);
408 resource_pool->release_2d_texture(cbcr_full_tex);
410 // Set the right state for rgba_tex.
411 glBindFramebuffer(GL_FRAMEBUFFER, 0);
412 glBindTexture(GL_TEXTURE_2D, rgba_tex);
413 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
414 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
415 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
417 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
420 // Make sure the H.264 gets a reference to all the
421 // input frames needed, so that they are not released back
422 // until the rendering is done.
423 vector<RefCountedFrame> input_frames;
424 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
425 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
427 h264_encoder->end_frame(fence, pts_int, input_frames);
429 pts_int += TIMEBASE / 60;
431 // The live frame just shows the RGBA texture we just rendered.
432 // It owns rgba_tex now.
433 DisplayFrame live_frame;
434 live_frame.chain = display_chain.get();
435 live_frame.setup_chain = [this, rgba_tex]{
436 display_input->set_texture_num(rgba_tex);
438 live_frame.ready_fence = fence;
439 live_frame.input_frames = {};
440 live_frame.temp_textures = { rgba_tex };
441 output_channel[OUTPUT_LIVE].output_frame(live_frame);
443 // Set up preview and any additional channels.
444 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
445 DisplayFrame display_frame;
446 pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT); // FIXME: dimensions
447 display_frame.chain = chain.first;
448 display_frame.setup_chain = chain.second;
449 display_frame.ready_fence = fence;
450 display_frame.input_frames = { bmusb_current_rendering_frame[0], bmusb_current_rendering_frame[1] }; // FIXME: possible to do better?
451 display_frame.temp_textures = {};
452 output_channel[i].output_frame(display_frame);
455 clock_gettime(CLOCK_MONOTONIC, &now);
456 double elapsed = now.tv_sec - start.tv_sec +
457 1e-9 * (now.tv_nsec - start.tv_nsec);
458 if (frame % 100 == 0) {
459 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
460 frame, dropped_frames, elapsed, frame / elapsed,
461 1e3 * elapsed / frame);
462 // chain->print_phase_timing();
466 // Reset every 100 frames, so that local variations in frame times
467 // (especially for the first few frames, when the shaders are
468 // compiled etc.) don't make it hard to measure for the entire
469 // remaining duration of the program.
470 if (frame == 10000) {
478 resource_pool->clean_context();
481 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
484 glGenVertexArrays(1, &vao);
493 glBindVertexArray(vao);
497 GLuint fbo = resource_pool->create_fbo(dst_tex);
498 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
499 glViewport(0, 0, WIDTH/2, HEIGHT/2);
502 glUseProgram(cbcr_program_num);
505 glActiveTexture(GL_TEXTURE0);
507 glBindTexture(GL_TEXTURE_2D, src_tex);
509 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
511 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
513 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
516 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
517 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
519 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
520 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
522 glDrawArrays(GL_TRIANGLES, 0, 3);
525 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
526 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
531 resource_pool->release_fbo(fbo);
532 glDeleteVertexArrays(1, &vao);
535 void Mixer::release_display_frame(DisplayFrame *frame)
537 for (GLuint texnum : frame->temp_textures) {
538 resource_pool->release_2d_texture(texnum);
540 frame->temp_textures.clear();
541 frame->ready_fence.reset();
542 frame->input_frames.clear();
547 mixer_thread = thread(&Mixer::thread_func, this);
556 void Mixer::transition_clicked(int transition_num)
558 theme->transition_clicked(transition_num, pts());
561 void Mixer::channel_clicked(int preview_num)
563 theme->channel_clicked(preview_num);
566 Mixer::OutputChannel::~OutputChannel()
568 if (has_current_frame) {
569 parent->release_display_frame(¤t_frame);
571 if (has_ready_frame) {
572 parent->release_display_frame(&ready_frame);
576 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
578 // Store this frame for display. Remove the ready frame if any
579 // (it was seemingly never used).
581 unique_lock<mutex> lock(frame_mutex);
582 if (has_ready_frame) {
583 parent->release_display_frame(&ready_frame);
586 has_ready_frame = true;
589 if (has_new_frame_ready_callback) {
590 new_frame_ready_callback();
594 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
596 unique_lock<mutex> lock(frame_mutex);
597 if (!has_current_frame && !has_ready_frame) {
601 if (has_current_frame && has_ready_frame) {
602 // We have a new ready frame. Toss the current one.
603 parent->release_display_frame(¤t_frame);
604 has_current_frame = false;
606 if (has_ready_frame) {
607 assert(!has_current_frame);
608 current_frame = ready_frame;
609 ready_frame.ready_fence.reset(); // Drop the refcount.
610 ready_frame.input_frames.clear(); // Drop the refcounts.
611 has_current_frame = true;
612 has_ready_frame = false;
615 *frame = current_frame;
619 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
621 new_frame_ready_callback = callback;
622 has_new_frame_ready_callback = true;