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 : mixer_surface(create_surface(format)),
75 h264_encoder_surface(create_surface(format))
79 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
82 resource_pool.reset(new ResourcePool);
83 theme.reset(new Theme("theme.lua", resource_pool.get()));
84 output_channel[OUTPUT_LIVE].parent = this;
85 output_channel[OUTPUT_PREVIEW].parent = this;
86 output_channel[OUTPUT_INPUT0].parent = this;
87 output_channel[OUTPUT_INPUT1].parent = this;
89 ImageFormat inout_format;
90 inout_format.color_space = COLORSPACE_sRGB;
91 inout_format.gamma_curve = GAMMA_sRGB;
93 // Display chain; shows the live output produced by the main chain (its RGBA version).
94 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
96 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
97 display_chain->add_input(display_input);
98 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
99 display_chain->set_dither_bits(0); // Don't bother.
100 display_chain->finalize();
102 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, "test.ts", &httpd));
104 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
105 printf("Configuring card %d...\n", card_index);
106 CaptureCard *card = &cards[card_index];
107 card->usb = new BMUSBCapture(card_index);
108 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
109 card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44, WIDTH, HEIGHT));
110 card->usb->set_video_frame_allocator(card->frame_allocator.get());
111 card->surface = create_surface(format);
112 card->usb->set_dequeue_thread_callbacks(
114 eglBindAPI(EGL_OPENGL_API);
115 card->context = create_context();
116 if (!make_current(card->context, card->surface)) {
117 printf("failed to create bmusb context\n");
122 resource_pool->clean_context();
124 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
125 card->usb->configure_card();
128 BMUSBCapture::start_bm_thread();
130 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
131 cards[card_index].usb->start_bm_capture();
134 //chain->enable_phase_timing(true);
136 // Set up stuff for NV12 conversion.
139 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
140 string cbcr_frag_shader =
143 "uniform sampler2D cbcr_tex; \n"
145 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
147 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
152 resource_pool->release_glsl_program(cbcr_program_num);
153 BMUSBCapture::stop_bm_thread();
155 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
157 unique_lock<mutex> lock(bmusb_mutex);
158 cards[card_index].should_quit = true; // Unblock thread.
159 cards[card_index].new_data_ready_changed.notify_all();
161 cards[card_index].usb->stop_dequeue_thread();
167 int unwrap_timecode(uint16_t current_wrapped, int last)
169 uint16_t last_wrapped = last & 0xffff;
170 if (current_wrapped > last_wrapped) {
171 return (last & ~0xffff) | current_wrapped;
173 return 0x10000 + ((last & ~0xffff) | current_wrapped);
179 void Mixer::bm_frame(int card_index, uint16_t timecode,
180 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
181 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
183 CaptureCard *card = &cards[card_index];
185 if (audio_frame.len - audio_offset > 30000) {
186 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",
187 card_index, int(audio_frame.len), int(audio_offset),
188 timecode, int(video_frame.len), int(video_offset), video_format);
189 if (video_frame.owner) {
190 video_frame.owner->release_frame(video_frame);
192 if (audio_frame.owner) {
193 audio_frame.owner->release_frame(audio_frame);
198 // Convert the audio to stereo fp32 and add it.
199 size_t num_samples = (audio_frame.len - audio_offset) / 8 / 3;
201 audio.resize(num_samples * 2);
202 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
204 int unwrapped_timecode = timecode;
205 int dropped_frames = 0;
206 if (card->last_timecode != -1) {
207 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
208 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
210 card->last_timecode = unwrapped_timecode;
214 unique_lock<mutex> lock(card->audio_mutex);
216 int unwrapped_timecode = timecode;
217 if (dropped_frames > 60 * 2) {
218 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
220 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
221 } else if (dropped_frames > 0) {
222 // Insert silence as needed.
223 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
224 card_index, dropped_frames, timecode);
225 vector<float> silence;
226 silence.resize((48000 / 60) * 2);
227 for (int i = 0; i < dropped_frames; ++i) {
228 card->resampler->add_input_samples((unwrapped_timecode - dropped_frames + i) / 60.0, silence.data(), (48000 / 60));
231 card->resampler->add_input_samples(unwrapped_timecode / 60.0, audio.data(), num_samples);
234 // Done with the audio, so release it.
235 if (audio_frame.owner) {
236 audio_frame.owner->release_frame(audio_frame);
240 // Wait until the previous frame was consumed.
241 unique_lock<mutex> lock(bmusb_mutex);
242 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
243 if (card->should_quit) return;
246 if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
247 if (video_frame.len != 0) {
248 printf("Card %d: Dropping video frame with wrong length (%ld)\n",
249 card_index, video_frame.len - video_offset);
251 if (video_frame.owner) {
252 video_frame.owner->release_frame(video_frame);
255 // Still send on the information that we _had_ a frame, even though it's corrupted,
256 // so that pts can go up accordingly.
258 unique_lock<mutex> lock(bmusb_mutex);
259 card->new_data_ready = true;
260 card->new_frame = RefCountedFrame(FrameAllocator::Frame());
261 card->new_data_ready_fence = nullptr;
262 card->dropped_frames = dropped_frames;
263 card->new_data_ready_changed.notify_all();
268 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
269 GLuint pbo = userdata->pbo;
271 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
273 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
275 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
278 // Upload the textures.
279 glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
281 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));
283 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
285 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
287 glBindTexture(GL_TEXTURE_2D, 0);
289 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
291 assert(fence != nullptr);
294 unique_lock<mutex> lock(bmusb_mutex);
295 card->new_data_ready = true;
296 card->new_frame = RefCountedFrame(video_frame);
297 card->new_data_ready_fence = fence;
298 card->dropped_frames = dropped_frames;
299 card->new_data_ready_changed.notify_all();
303 void Mixer::thread_func()
305 eglBindAPI(EGL_OPENGL_API);
306 QOpenGLContext *context = create_context();
307 if (!make_current(context, mixer_surface)) {
312 struct timespec start, now;
313 clock_gettime(CLOCK_MONOTONIC, &start);
316 int dropped_frames = 0;
318 while (!should_quit) {
319 CaptureCard card_copy[NUM_CARDS];
322 unique_lock<mutex> lock(bmusb_mutex);
324 // The first card is the master timer, so wait for it to have a new frame.
325 // TODO: Make configurable, and with a timeout.
326 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
328 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
329 CaptureCard *card = &cards[card_index];
330 card_copy[card_index].usb = card->usb;
331 card_copy[card_index].new_data_ready = card->new_data_ready;
332 card_copy[card_index].new_frame = card->new_frame;
333 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
334 card_copy[card_index].new_frame_audio = move(card->new_frame_audio);
335 card_copy[card_index].dropped_frames = card->dropped_frames;
336 card->new_data_ready = false;
337 card->new_data_ready_changed.notify_all();
341 // Resample the audio as needed, including from previously dropped frames.
342 vector<float> samples_out;
343 // TODO: Allow using audio from the other card(s) as well.
344 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
345 for (unsigned card_index = 0; card_index < NUM_CARDS; ++card_index) {
346 samples_out.resize((48000 / 60) * 2);
348 unique_lock<mutex> lock(cards[card_index].audio_mutex);
349 if (!cards[card_index].resampler->get_output_samples(pts(), &samples_out[0], 48000 / 60)) {
350 printf("Card %d reported previous underrun.\n", card_index);
353 if (card_index == 0) {
354 h264_encoder->add_audio(pts_int, move(samples_out));
357 if (frame_num != card_copy[0].dropped_frames) {
358 // For dropped frames, increase the pts.
360 pts_int += TIMEBASE / 60;
364 // If the first card is reporting a corrupted or otherwise dropped frame,
365 // just increase the pts (skipping over this frame) and don't try to compute anything new.
366 if (card_copy[0].new_frame->len == 0) {
368 pts_int += TIMEBASE / 60;
372 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
373 CaptureCard *card = &card_copy[card_index];
374 if (!card->new_data_ready || card->new_frame->len == 0)
377 assert(card->new_frame != nullptr);
378 bmusb_current_rendering_frame[card_index] = card->new_frame;
381 // The new texture might still be uploaded,
382 // tell the GPU to wait until it's there.
383 if (card->new_data_ready_fence) {
384 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
386 glDeleteSync(card->new_data_ready_fence);
389 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
390 theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
393 // Get the main chain from the theme, and set its state immediately.
394 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
395 EffectChain *chain = theme_main_chain.first;
396 theme_main_chain.second();
398 GLuint y_tex, cbcr_tex;
399 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
402 // Render main chain.
403 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
404 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
405 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
406 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
407 resource_pool->release_fbo(fbo);
409 subsample_chroma(cbcr_full_tex, cbcr_tex);
410 resource_pool->release_2d_texture(cbcr_full_tex);
412 // Set the right state for rgba_tex.
413 glBindFramebuffer(GL_FRAMEBUFFER, 0);
414 glBindTexture(GL_TEXTURE_2D, rgba_tex);
415 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
416 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
417 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
419 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
422 // Make sure the H.264 gets a reference to all the
423 // input frames needed, so that they are not released back
424 // until the rendering is done.
425 vector<RefCountedFrame> input_frames;
426 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
427 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
429 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampler.h. TODO: Make less hard-coded.
430 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
432 pts_int += TIMEBASE / 60;
434 // The live frame just shows the RGBA texture we just rendered.
435 // It owns rgba_tex now.
436 DisplayFrame live_frame;
437 live_frame.chain = display_chain.get();
438 live_frame.setup_chain = [this, rgba_tex]{
439 display_input->set_texture_num(rgba_tex);
441 live_frame.ready_fence = fence;
442 live_frame.input_frames = {};
443 live_frame.temp_textures = { rgba_tex };
444 output_channel[OUTPUT_LIVE].output_frame(live_frame);
446 // Set up preview and any additional channels.
447 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
448 DisplayFrame display_frame;
449 pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT); // FIXME: dimensions
450 display_frame.chain = chain.first;
451 display_frame.setup_chain = chain.second;
452 display_frame.ready_fence = fence;
453 display_frame.input_frames = { bmusb_current_rendering_frame[0], bmusb_current_rendering_frame[1] }; // FIXME: possible to do better?
454 display_frame.temp_textures = {};
455 output_channel[i].output_frame(display_frame);
458 clock_gettime(CLOCK_MONOTONIC, &now);
459 double elapsed = now.tv_sec - start.tv_sec +
460 1e-9 * (now.tv_nsec - start.tv_nsec);
461 if (frame % 100 == 0) {
462 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
463 frame, dropped_frames, elapsed, frame / elapsed,
464 1e3 * elapsed / frame);
465 // chain->print_phase_timing();
469 // Reset every 100 frames, so that local variations in frame times
470 // (especially for the first few frames, when the shaders are
471 // compiled etc.) don't make it hard to measure for the entire
472 // remaining duration of the program.
473 if (frame == 10000) {
481 resource_pool->clean_context();
484 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
487 glGenVertexArrays(1, &vao);
496 glBindVertexArray(vao);
500 GLuint fbo = resource_pool->create_fbo(dst_tex);
501 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
502 glViewport(0, 0, WIDTH/2, HEIGHT/2);
505 glUseProgram(cbcr_program_num);
508 glActiveTexture(GL_TEXTURE0);
510 glBindTexture(GL_TEXTURE_2D, src_tex);
512 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
514 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
516 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
519 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
520 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
522 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
523 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
525 glDrawArrays(GL_TRIANGLES, 0, 3);
528 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
529 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
534 resource_pool->release_fbo(fbo);
535 glDeleteVertexArrays(1, &vao);
538 void Mixer::release_display_frame(DisplayFrame *frame)
540 for (GLuint texnum : frame->temp_textures) {
541 resource_pool->release_2d_texture(texnum);
543 frame->temp_textures.clear();
544 frame->ready_fence.reset();
545 frame->input_frames.clear();
550 mixer_thread = thread(&Mixer::thread_func, this);
559 void Mixer::transition_clicked(int transition_num)
561 theme->transition_clicked(transition_num, pts());
564 void Mixer::channel_clicked(int preview_num)
566 theme->channel_clicked(preview_num);
569 Mixer::OutputChannel::~OutputChannel()
571 if (has_current_frame) {
572 parent->release_display_frame(¤t_frame);
574 if (has_ready_frame) {
575 parent->release_display_frame(&ready_frame);
579 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
581 // Store this frame for display. Remove the ready frame if any
582 // (it was seemingly never used).
584 unique_lock<mutex> lock(frame_mutex);
585 if (has_ready_frame) {
586 parent->release_display_frame(&ready_frame);
589 has_ready_frame = true;
592 if (has_new_frame_ready_callback) {
593 new_frame_ready_callback();
597 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
599 unique_lock<mutex> lock(frame_mutex);
600 if (!has_current_frame && !has_ready_frame) {
604 if (has_current_frame && has_ready_frame) {
605 // We have a new ready frame. Toss the current one.
606 parent->release_display_frame(¤t_frame);
607 has_current_frame = false;
609 if (has_ready_frame) {
610 assert(!has_current_frame);
611 current_frame = ready_frame;
612 ready_frame.ready_fence.reset(); // Drop the refcount.
613 ready_frame.input_frames.clear(); // Drop the refcounts.
614 has_current_frame = true;
615 has_ready_frame = false;
618 *frame = current_frame;
622 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
624 new_frame_ready_callback = callback;
625 has_new_frame_ready_callback = true;