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(std::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 = 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 std::unique_lock<std::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();
163 void Mixer::bm_frame(int card_index, uint16_t timecode,
164 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
165 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
167 CaptureCard *card = &cards[card_index];
169 if (video_frame.len - video_offset != 1280 * 750 * 2) {
170 printf("dropping frame with wrong length (%ld)\n", video_frame.len - video_offset);
171 card->usb->get_video_frame_allocator()->release_frame(video_frame);
172 card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
175 if (audio_frame.len - audio_offset > 30000) {
176 printf("dropping frame with implausible audio length (%ld)\n", audio_frame.len - audio_offset);
177 card->usb->get_video_frame_allocator()->release_frame(video_frame);
178 card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
183 // Wait until the previous frame was consumed.
184 std::unique_lock<std::mutex> lock(bmusb_mutex);
185 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
186 if (card->should_quit) return;
188 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
189 GLuint pbo = userdata->pbo;
191 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
193 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
195 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
198 // Upload the textures.
199 glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
201 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280, 720, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((1280 * 750 * 2 + 44) / 2 + 1280 * 25 + 22));
203 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
205 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280/2, 720, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(1280 * 25 + 22));
207 glBindTexture(GL_TEXTURE_2D, 0);
209 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
211 assert(fence != nullptr);
213 // Convert the audio to stereo fp32 and store it next to the video.
214 size_t num_samples = (audio_frame.len - audio_offset) / 8 / 3;
216 audio.resize(num_samples * 2);
217 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
220 std::unique_lock<std::mutex> lock(bmusb_mutex);
221 card->new_data_ready = true;
222 card->new_frame = RefCountedFrame(video_frame);
223 card->new_data_ready_fence = fence;
224 card->new_frame_audio = move(audio);
225 card->new_data_ready_changed.notify_all();
228 // Video frame will be released when last user of card->new_frame goes out of scope.
229 card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
232 void Mixer::thread_func()
234 eglBindAPI(EGL_OPENGL_API);
235 QOpenGLContext *context = create_context();
236 if (!make_current(context, mixer_surface)) {
241 struct timespec start, now;
242 clock_gettime(CLOCK_MONOTONIC, &start);
244 while (!should_quit) {
247 CaptureCard card_copy[NUM_CARDS];
250 std::unique_lock<std::mutex> lock(bmusb_mutex);
252 // The first card is the master timer, so wait for it to have a new frame.
253 // TODO: Make configurable, and with a timeout.
254 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
256 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
257 CaptureCard *card = &cards[card_index];
258 card_copy[card_index].usb = card->usb;
259 card_copy[card_index].new_data_ready = card->new_data_ready;
260 card_copy[card_index].new_frame = card->new_frame;
261 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
262 card_copy[card_index].new_frame_audio = move(card->new_frame_audio);
263 card->new_data_ready = false;
264 card->new_data_ready_changed.notify_all();
268 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
269 CaptureCard *card = &card_copy[card_index];
270 if (!card->new_data_ready)
273 assert(card->new_frame != nullptr);
274 bmusb_current_rendering_frame[card_index] = card->new_frame;
277 // The new texture might still be uploaded,
278 // tell the GPU to wait until it's there.
279 if (card->new_data_ready_fence)
280 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
282 glDeleteSync(card->new_data_ready_fence);
284 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
285 theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
288 // Get the main chain from the theme, and set its state immediately.
289 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, frame / 60.0f, WIDTH, HEIGHT);
290 EffectChain *chain = theme_main_chain.first;
291 theme_main_chain.second();
293 GLuint y_tex, cbcr_tex;
294 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
297 // Render main chain.
298 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
299 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
300 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
301 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
302 resource_pool->release_fbo(fbo);
304 subsample_chroma(cbcr_full_tex, cbcr_tex);
305 resource_pool->release_2d_texture(cbcr_full_tex);
307 // Set the right state for rgba_tex.
308 glBindFramebuffer(GL_FRAMEBUFFER, 0);
309 glBindTexture(GL_TEXTURE_2D, rgba_tex);
310 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
311 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
312 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
314 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
317 // Resample the audio as needed.
318 // TODO: Allow using audio from the other card(s) as well.
319 double pts = frame / 60.0;
320 cards[0].resampler->add_input_samples(pts, card_copy[0].new_frame_audio.data(), card_copy[0].new_frame_audio.size() / 2);
321 vector<float> samples_out;
322 samples_out.resize((48000 / 60) * 2);
323 cards[0].resampler->get_output_samples(pts, &samples_out[0], 48000 / 60);
325 // Make sure the H.264 gets a reference to all the
326 // input frames needed, so that they are not released back
327 // until the rendering is done.
328 vector<RefCountedFrame> input_frames;
329 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
330 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
332 h264_encoder->end_frame(fence, frame * (TIMEBASE / 60), move(samples_out), input_frames);
334 // The live frame just shows the RGBA texture we just rendered.
335 // It owns rgba_tex now.
336 DisplayFrame live_frame;
337 live_frame.chain = display_chain.get();
338 live_frame.setup_chain = [this, rgba_tex]{
339 display_input->set_texture_num(rgba_tex);
341 live_frame.ready_fence = fence;
342 live_frame.input_frames = {};
343 live_frame.temp_textures = { rgba_tex };
344 output_channel[OUTPUT_LIVE].output_frame(live_frame);
346 // Set up preview and any additional channels.
347 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
348 DisplayFrame display_frame;
349 pair<EffectChain *, function<void()>> chain = theme->get_chain(i, frame / 60.0f, WIDTH, HEIGHT); // FIXME: dimensions
350 display_frame.chain = chain.first;
351 display_frame.setup_chain = chain.second;
352 display_frame.ready_fence = fence;
353 display_frame.input_frames = { bmusb_current_rendering_frame[0], bmusb_current_rendering_frame[1] }; // FIXME: possible to do better?
354 display_frame.temp_textures = {};
355 output_channel[i].output_frame(display_frame);
358 clock_gettime(CLOCK_MONOTONIC, &now);
359 double elapsed = now.tv_sec - start.tv_sec +
360 1e-9 * (now.tv_nsec - start.tv_nsec);
361 if (frame % 100 == 0) {
362 printf("%d frames in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
363 frame, elapsed, frame / elapsed,
364 1e3 * elapsed / frame);
365 // chain->print_phase_timing();
369 // Reset every 100 frames, so that local variations in frame times
370 // (especially for the first few frames, when the shaders are
371 // compiled etc.) don't make it hard to measure for the entire
372 // remaining duration of the program.
373 if (frame == 10000) {
381 resource_pool->clean_context();
384 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
387 glGenVertexArrays(1, &vao);
396 glBindVertexArray(vao);
400 GLuint fbo = resource_pool->create_fbo(dst_tex);
401 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
402 glViewport(0, 0, WIDTH/2, HEIGHT/2);
405 glUseProgram(cbcr_program_num);
408 glActiveTexture(GL_TEXTURE0);
410 glBindTexture(GL_TEXTURE_2D, src_tex);
412 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
414 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);
419 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
420 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
422 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
423 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
425 glDrawArrays(GL_TRIANGLES, 0, 3);
428 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
429 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
434 resource_pool->release_fbo(fbo);
435 glDeleteVertexArrays(1, &vao);
438 void Mixer::release_display_frame(DisplayFrame *frame)
440 for (GLuint texnum : frame->temp_textures) {
441 resource_pool->release_2d_texture(texnum);
443 frame->temp_textures.clear();
444 frame->ready_fence.reset();
445 frame->input_frames.clear();
450 mixer_thread = std::thread(&Mixer::thread_func, this);
459 void Mixer::transition_clicked(int transition_num)
461 theme->transition_clicked(transition_num, frame / 60.0);
464 void Mixer::channel_clicked(int preview_num)
466 theme->channel_clicked(preview_num);
469 Mixer::OutputChannel::~OutputChannel()
471 if (has_current_frame) {
472 parent->release_display_frame(¤t_frame);
474 if (has_ready_frame) {
475 parent->release_display_frame(&ready_frame);
479 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
481 // Store this frame for display. Remove the ready frame if any
482 // (it was seemingly never used).
484 std::unique_lock<std::mutex> lock(frame_mutex);
485 if (has_ready_frame) {
486 parent->release_display_frame(&ready_frame);
489 has_ready_frame = true;
492 if (has_new_frame_ready_callback) {
493 new_frame_ready_callback();
497 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
499 std::unique_lock<std::mutex> lock(frame_mutex);
500 if (!has_current_frame && !has_ready_frame) {
504 if (has_current_frame && has_ready_frame) {
505 // We have a new ready frame. Toss the current one.
506 parent->release_display_frame(¤t_frame);
507 has_current_frame = false;
509 if (has_ready_frame) {
510 assert(!has_current_frame);
511 current_frame = ready_frame;
512 ready_frame.ready_fence.reset(); // Drop the refcount.
513 ready_frame.input_frames.clear(); // Drop the refcounts.
514 has_current_frame = true;
515 has_ready_frame = false;
518 *frame = current_frame;
522 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
524 new_frame_ready_callback = callback;
525 has_new_frame_ready_callback = true;