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))
77 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
80 resource_pool.reset(new ResourcePool);
81 theme.reset(new Theme("theme.lua", resource_pool.get()));
82 output_channel[OUTPUT_LIVE].parent = this;
83 output_channel[OUTPUT_PREVIEW].parent = this;
84 output_channel[OUTPUT_INPUT0].parent = this;
85 output_channel[OUTPUT_INPUT1].parent = this;
87 ImageFormat inout_format;
88 inout_format.color_space = COLORSPACE_sRGB;
89 inout_format.gamma_curve = GAMMA_sRGB;
91 // Display chain; shows the live output produced by the main chain (its RGBA version).
92 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
94 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
95 display_chain->add_input(display_input);
96 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
97 display_chain->set_dither_bits(0); // Don't bother.
98 display_chain->finalize();
100 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, "test.mp4"));
102 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
103 printf("Configuring card %d...\n", card_index);
104 CaptureCard *card = &cards[card_index];
105 card->usb = new BMUSBCapture(0x1edb, card_index == 0 ? 0xbd3b : 0xbd4f);
106 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
107 card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44, WIDTH, HEIGHT));
108 card->usb->set_video_frame_allocator(card->frame_allocator.get());
109 card->surface = create_surface(format);
110 card->usb->set_dequeue_thread_callbacks(
112 eglBindAPI(EGL_OPENGL_API);
113 card->context = create_context();
114 if (!make_current(card->context, card->surface)) {
115 printf("failed to create bmusb context\n");
121 resource_pool->clean_context();
123 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
124 card->usb->configure_card();
127 BMUSBCapture::start_bm_thread();
129 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
130 cards[card_index].usb->start_bm_capture();
133 //chain->enable_phase_timing(true);
135 // Set up stuff for NV12 conversion.
138 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
139 string cbcr_frag_shader =
142 "uniform sampler2D cbcr_tex; \n"
144 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
146 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
151 resource_pool->release_glsl_program(cbcr_program_num);
152 BMUSBCapture::stop_bm_thread();
154 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
156 unique_lock<mutex> lock(bmusb_mutex);
157 cards[card_index].should_quit = true; // Unblock thread.
158 cards[card_index].new_data_ready_changed.notify_all();
160 cards[card_index].usb->stop_dequeue_thread();
166 int unwrap_timecode(uint16_t current_wrapped, int last)
168 uint16_t last_wrapped = last & 0xffff;
169 if (current_wrapped > last_wrapped) {
170 return (last & ~0xffff) | current_wrapped;
172 return 0x10000 + ((last & ~0xffff) | current_wrapped);
178 void Mixer::bm_frame(int card_index, uint16_t timecode,
179 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
180 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
182 CaptureCard *card = &cards[card_index];
184 if (audio_frame.len - audio_offset > 30000) {
185 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",
186 card_index, int(audio_frame.len), int(audio_offset),
187 timecode, int(video_frame.len), int(video_offset), video_format);
188 if (video_frame.owner) {
189 video_frame.owner->release_frame(video_frame);
191 if (audio_frame.owner) {
192 audio_frame.owner->release_frame(audio_frame);
197 // Convert the audio to stereo fp32 and add it.
198 size_t num_samples = (audio_frame.len - audio_offset) / 8 / 3;
200 audio.resize(num_samples * 2);
201 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
203 int unwrapped_timecode = timecode;
204 int dropped_frames = 0;
205 if (card->last_timecode != -1) {
206 unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
207 dropped_frames = unwrapped_timecode - card->last_timecode - 1;
209 card->last_timecode = unwrapped_timecode;
213 unique_lock<mutex> lock(card->audio_mutex);
215 int unwrapped_timecode = timecode;
216 if (dropped_frames > 60 * 2) {
217 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
219 card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
220 } else if (dropped_frames > 0) {
221 // Insert silence as needed.
222 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
223 card_index, dropped_frames, timecode);
224 vector<float> silence;
225 silence.resize((48000 / 60) * 2);
226 for (int i = 0; i < dropped_frames; ++i) {
227 card->resampler->add_input_samples((unwrapped_timecode - dropped_frames + i) / 60.0, silence.data(), (48000 / 60));
230 card->resampler->add_input_samples(unwrapped_timecode / 60.0, audio.data(), num_samples);
233 // Done with the audio, so release it.
234 if (audio_frame.owner) {
235 audio_frame.owner->release_frame(audio_frame);
239 // Wait until the previous frame was consumed.
240 unique_lock<mutex> lock(bmusb_mutex);
241 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
242 if (card->should_quit) return;
245 if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
246 if (video_frame.len != 0) {
247 printf("Card %d: Dropping video frame with wrong length (%ld)\n",
248 card_index, video_frame.len - video_offset);
250 if (video_frame.owner) {
251 video_frame.owner->release_frame(video_frame);
254 // Still send on the information that we _had_ a frame, even though it's corrupted,
255 // so that pts can go up accordingly.
257 unique_lock<mutex> lock(bmusb_mutex);
258 card->new_data_ready = true;
259 card->new_frame = RefCountedFrame(FrameAllocator::Frame());
260 card->new_data_ready_fence = nullptr;
261 card->dropped_frames = dropped_frames;
262 card->new_data_ready_changed.notify_all();
267 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
268 GLuint pbo = userdata->pbo;
270 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
272 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
274 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
277 // Upload the textures.
278 glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
280 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));
282 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
284 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
286 glBindTexture(GL_TEXTURE_2D, 0);
288 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
290 assert(fence != nullptr);
293 unique_lock<mutex> lock(bmusb_mutex);
294 card->new_data_ready = true;
295 card->new_frame = RefCountedFrame(video_frame);
296 card->new_data_ready_fence = fence;
297 card->dropped_frames = dropped_frames;
298 card->new_data_ready_changed.notify_all();
302 void Mixer::thread_func()
304 eglBindAPI(EGL_OPENGL_API);
305 QOpenGLContext *context = create_context();
306 if (!make_current(context, mixer_surface)) {
311 struct timespec start, now;
312 clock_gettime(CLOCK_MONOTONIC, &start);
315 int dropped_frames = 0;
317 while (!should_quit) {
318 CaptureCard card_copy[NUM_CARDS];
321 unique_lock<mutex> lock(bmusb_mutex);
323 // The first card is the master timer, so wait for it to have a new frame.
324 // TODO: Make configurable, and with a timeout.
325 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
327 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
328 CaptureCard *card = &cards[card_index];
329 card_copy[card_index].usb = card->usb;
330 card_copy[card_index].new_data_ready = card->new_data_ready;
331 card_copy[card_index].new_frame = card->new_frame;
332 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
333 card_copy[card_index].new_frame_audio = move(card->new_frame_audio);
334 card_copy[card_index].dropped_frames = card->dropped_frames;
335 card->new_data_ready = false;
336 card->new_data_ready_changed.notify_all();
340 // Resample the audio as needed, including from previously dropped frames.
341 vector<float> samples_out;
342 // TODO: Allow using audio from the other card(s) as well.
343 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
344 for (unsigned card_index = 0; card_index < NUM_CARDS; ++card_index) {
345 samples_out.resize((48000 / 60) * 2);
347 unique_lock<mutex> lock(cards[card_index].audio_mutex);
348 if (!cards[card_index].resampler->get_output_samples(pts(), &samples_out[0], 48000 / 60)) {
349 printf("Card %d reported previous underrun.\n", card_index);
352 if (card_index == 0) {
353 h264_encoder->add_audio(pts_int, move(samples_out));
356 if (frame_num != card_copy[0].dropped_frames) {
357 // For dropped frames, increase the pts.
359 pts_int += TIMEBASE / 60;
363 // If the first card is reporting a corrupted or otherwise dropped frame,
364 // just increase the pts (skipping over this frame) and don't try to compute anything new.
365 if (card_copy[0].new_frame->len == 0) {
367 pts_int += TIMEBASE / 60;
371 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
372 CaptureCard *card = &card_copy[card_index];
373 if (!card->new_data_ready || card->new_frame->len == 0)
376 assert(card->new_frame != nullptr);
377 bmusb_current_rendering_frame[card_index] = card->new_frame;
380 // The new texture might still be uploaded,
381 // tell the GPU to wait until it's there.
382 if (card->new_data_ready_fence) {
383 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
385 glDeleteSync(card->new_data_ready_fence);
388 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
389 theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
392 // Get the main chain from the theme, and set its state immediately.
393 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
394 EffectChain *chain = theme_main_chain.first;
395 theme_main_chain.second();
397 GLuint y_tex, cbcr_tex;
398 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
401 // Render main chain.
402 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
403 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
404 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
405 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
406 resource_pool->release_fbo(fbo);
408 subsample_chroma(cbcr_full_tex, cbcr_tex);
409 resource_pool->release_2d_texture(cbcr_full_tex);
411 // Set the right state for rgba_tex.
412 glBindFramebuffer(GL_FRAMEBUFFER, 0);
413 glBindTexture(GL_TEXTURE_2D, rgba_tex);
414 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
415 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
416 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
418 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
421 // Make sure the H.264 gets a reference to all the
422 // input frames needed, so that they are not released back
423 // until the rendering is done.
424 vector<RefCountedFrame> input_frames;
425 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
426 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
428 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampler.h. TODO: Make less hard-coded.
429 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
431 pts_int += TIMEBASE / 60;
433 // The live frame just shows the RGBA texture we just rendered.
434 // It owns rgba_tex now.
435 DisplayFrame live_frame;
436 live_frame.chain = display_chain.get();
437 live_frame.setup_chain = [this, rgba_tex]{
438 display_input->set_texture_num(rgba_tex);
440 live_frame.ready_fence = fence;
441 live_frame.input_frames = {};
442 live_frame.temp_textures = { rgba_tex };
443 output_channel[OUTPUT_LIVE].output_frame(live_frame);
445 // Set up preview and any additional channels.
446 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
447 DisplayFrame display_frame;
448 pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT); // FIXME: dimensions
449 display_frame.chain = chain.first;
450 display_frame.setup_chain = chain.second;
451 display_frame.ready_fence = fence;
452 display_frame.input_frames = { bmusb_current_rendering_frame[0], bmusb_current_rendering_frame[1] }; // FIXME: possible to do better?
453 display_frame.temp_textures = {};
454 output_channel[i].output_frame(display_frame);
457 clock_gettime(CLOCK_MONOTONIC, &now);
458 double elapsed = now.tv_sec - start.tv_sec +
459 1e-9 * (now.tv_nsec - start.tv_nsec);
460 if (frame % 100 == 0) {
461 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
462 frame, dropped_frames, elapsed, frame / elapsed,
463 1e3 * elapsed / frame);
464 // chain->print_phase_timing();
468 // Reset every 100 frames, so that local variations in frame times
469 // (especially for the first few frames, when the shaders are
470 // compiled etc.) don't make it hard to measure for the entire
471 // remaining duration of the program.
472 if (frame == 10000) {
480 resource_pool->clean_context();
483 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
486 glGenVertexArrays(1, &vao);
495 glBindVertexArray(vao);
499 GLuint fbo = resource_pool->create_fbo(dst_tex);
500 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
501 glViewport(0, 0, WIDTH/2, HEIGHT/2);
504 glUseProgram(cbcr_program_num);
507 glActiveTexture(GL_TEXTURE0);
509 glBindTexture(GL_TEXTURE_2D, src_tex);
511 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
513 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
515 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
518 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
519 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
521 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
522 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
524 glDrawArrays(GL_TRIANGLES, 0, 3);
527 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
528 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
533 resource_pool->release_fbo(fbo);
534 glDeleteVertexArrays(1, &vao);
537 void Mixer::release_display_frame(DisplayFrame *frame)
539 for (GLuint texnum : frame->temp_textures) {
540 resource_pool->release_2d_texture(texnum);
542 frame->temp_textures.clear();
543 frame->ready_fence.reset();
544 frame->input_frames.clear();
549 mixer_thread = thread(&Mixer::thread_func, this);
558 void Mixer::transition_clicked(int transition_num)
560 theme->transition_clicked(transition_num, pts());
563 void Mixer::channel_clicked(int preview_num)
565 theme->channel_clicked(preview_num);
568 Mixer::OutputChannel::~OutputChannel()
570 if (has_current_frame) {
571 parent->release_display_frame(¤t_frame);
573 if (has_ready_frame) {
574 parent->release_display_frame(&ready_frame);
578 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
580 // Store this frame for display. Remove the ready frame if any
581 // (it was seemingly never used).
583 unique_lock<mutex> lock(frame_mutex);
584 if (has_ready_frame) {
585 parent->release_display_frame(&ready_frame);
588 has_ready_frame = true;
591 if (has_new_frame_ready_callback) {
592 new_frame_ready_callback();
596 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
598 unique_lock<mutex> lock(frame_mutex);
599 if (!has_current_frame && !has_ready_frame) {
603 if (has_current_frame && has_ready_frame) {
604 // We have a new ready frame. Toss the current one.
605 parent->release_display_frame(¤t_frame);
606 has_current_frame = false;
608 if (has_ready_frame) {
609 assert(!has_current_frame);
610 current_frame = ready_frame;
611 ready_frame.ready_fence.reset(); // Drop the refcount.
612 ready_frame.input_frames.clear(); // Drop the refcounts.
613 has_current_frame = true;
614 has_ready_frame = false;
617 *frame = current_frame;
621 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
623 new_frame_ready_callback = callback;
624 has_new_frame_ready_callback = true;