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>
41 #include "h264encode.h"
42 #include "pbo_frame_allocator.h"
43 #include "ref_counted_gl_sync.h"
47 using namespace movit;
49 using namespace std::placeholders;
51 Mixer *global_mixer = nullptr;
55 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
57 for (size_t i = 0; i < num_samples; ++i) {
58 for (size_t j = 0; j < out_channels; ++j) {
62 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
63 dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
65 src += 3 * (in_channels - out_channels);
71 Mixer::Mixer(const QSurfaceFormat &format)
72 : mixer_surface(create_surface(format)),
73 h264_encoder_surface(create_surface(format))
75 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
78 resource_pool.reset(new ResourcePool);
79 theme.reset(new Theme("theme.lua", resource_pool.get()));
80 output_channel[OUTPUT_LIVE].parent = this;
81 output_channel[OUTPUT_PREVIEW].parent = this;
82 output_channel[OUTPUT_INPUT0].parent = this;
83 output_channel[OUTPUT_INPUT1].parent = this;
85 ImageFormat inout_format;
86 inout_format.color_space = COLORSPACE_sRGB;
87 inout_format.gamma_curve = GAMMA_sRGB;
89 // Display chain; shows the live output produced by the main chain (its RGBA version).
90 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
92 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
93 display_chain->add_input(display_input);
94 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
95 display_chain->set_dither_bits(0); // Don't bother.
96 display_chain->finalize();
98 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, "test.mp4"));
100 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
101 printf("Configuring card %d...\n", card_index);
102 CaptureCard *card = &cards[card_index];
103 card->usb = new BMUSBCapture(0x1edb, card_index == 0 ? 0xbd3b : 0xbd4f);
104 card->usb->set_frame_callback(std::bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
105 card->frame_allocator.reset(new PBOFrameAllocator(1280 * 750 * 2 + 44, 1280, 720));
106 card->usb->set_video_frame_allocator(card->frame_allocator.get());
107 card->surface = create_surface(format);
108 card->usb->set_dequeue_thread_callbacks(
110 eglBindAPI(EGL_OPENGL_API);
111 card->context = create_context();
112 if (!make_current(card->context, card->surface)) {
113 printf("failed to create bmusb context\n");
119 resource_pool->clean_context();
121 card->resampler = new Resampler(48000.0, 48000.0, 2);
122 card->usb->configure_card();
125 BMUSBCapture::start_bm_thread();
127 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
128 cards[card_index].usb->start_bm_capture();
131 //chain->enable_phase_timing(true);
133 // Set up stuff for NV12 conversion.
136 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
137 string cbcr_frag_shader =
140 "uniform sampler2D cbcr_tex; \n"
142 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
144 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
149 resource_pool->release_glsl_program(cbcr_program_num);
150 BMUSBCapture::stop_bm_thread();
152 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
153 cards[card_index].new_data_ready = false; // Unblock thread.
154 cards[card_index].new_data_ready_changed.notify_all();
155 cards[card_index].usb->stop_dequeue_thread();
159 void Mixer::bm_frame(int card_index, uint16_t timecode,
160 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
161 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
163 CaptureCard *card = &cards[card_index];
165 if (video_frame.len - video_offset != 1280 * 750 * 2) {
166 printf("dropping frame with wrong length (%ld)\n", video_frame.len - video_offset);
167 FILE *fp = fopen("frame.raw", "wb");
168 fwrite(video_frame.data, video_frame.len, 1, fp);
171 card->usb->get_video_frame_allocator()->release_frame(video_frame);
172 card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
176 // Wait until the previous frame was consumed.
177 std::unique_lock<std::mutex> lock(bmusb_mutex);
178 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready; });
180 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
181 GLuint pbo = userdata->pbo;
183 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
185 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
187 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
190 // Upload the textures.
191 glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
193 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280, 720, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((1280 * 750 * 2 + 44) / 2 + 1280 * 25 + 22));
195 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
197 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280/2, 720, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(1280 * 25 + 22));
199 glBindTexture(GL_TEXTURE_2D, 0);
201 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
203 assert(fence != nullptr);
205 std::unique_lock<std::mutex> lock(bmusb_mutex);
206 card->new_data_ready = true;
207 card->new_frame = RefCountedFrame(video_frame);
208 card->new_data_ready_fence = fence;
209 card->new_data_ready_changed.notify_all();
212 // As a test of the resampler, send the data from card 0 through it and onto disk.
213 // TODO: Send the audio on, and encode it through ffmpeg.
214 if (card_index == 0) {
215 size_t num_samples = (audio_frame.len - audio_offset) / 8 / 3;
216 double pts = timecode / 60.0; // FIXME: Unwrap. And rebase.
217 unique_ptr<float[]> samplesf(new float[num_samples * 2]);
218 convert_fixed24_to_fp32(samplesf.get(), 2, audio_frame.data + audio_offset, 8, num_samples);
219 card->resampler->add_input_samples(pts, samplesf.get(), num_samples);
221 float samples_out[(48000 / 60) * 2];
222 card->resampler->get_output_samples(pts, samples_out, 48000 / 60);
224 static FILE *audiofp = nullptr;
225 if (audiofp == nullptr) {
226 audiofp = fopen("audio.raw", "wb");
228 fwrite(samples_out, sizeof(samples_out), 1, audiofp);
229 //fwrite(samplesf.get(), num_samples * sizeof(float) * 2, 1, audiofp);
231 if (audio_frame.len - audio_offset != 19200) {
232 printf("%d: %d samples (%d bytes)\n", card_index, int(num_samples), int(audio_frame.len - audio_offset));
236 // Video frame will be released when last user of card->new_frame goes out of scope.
237 card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
240 void Mixer::thread_func()
242 eglBindAPI(EGL_OPENGL_API);
243 QOpenGLContext *context = create_context();
244 if (!make_current(context, mixer_surface)) {
249 struct timespec start, now;
250 clock_gettime(CLOCK_MONOTONIC, &start);
252 while (!should_quit) {
255 CaptureCard card_copy[NUM_CARDS];
258 std::unique_lock<std::mutex> lock(bmusb_mutex);
260 // The first card is the master timer, so wait for it to have a new frame.
261 // TODO: Make configurable, and with a timeout.
262 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
264 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
265 CaptureCard *card = &cards[card_index];
266 card_copy[card_index].usb = card->usb;
267 card_copy[card_index].new_data_ready = card->new_data_ready;
268 card_copy[card_index].new_frame = card->new_frame;
269 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
270 card->new_data_ready = false;
271 card->new_data_ready_changed.notify_all();
275 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
276 CaptureCard *card = &card_copy[card_index];
277 if (!card->new_data_ready)
280 assert(card->new_frame != nullptr);
281 bmusb_current_rendering_frame[card_index] = card->new_frame;
284 // The new texture might still be uploaded,
285 // tell the GPU to wait until it's there.
286 if (card->new_data_ready_fence)
287 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
289 glDeleteSync(card->new_data_ready_fence);
291 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
292 theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
295 // Get the main chain from the theme, and set its state immediately.
296 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, frame / 60.0f, WIDTH, HEIGHT);
297 EffectChain *chain = theme_main_chain.first;
298 theme_main_chain.second();
300 GLuint y_tex, cbcr_tex;
301 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
304 // Render main chain.
305 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
306 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
307 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
308 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
309 resource_pool->release_fbo(fbo);
311 subsample_chroma(cbcr_full_tex, cbcr_tex);
312 resource_pool->release_2d_texture(cbcr_full_tex);
314 // Set the right state for rgba_tex.
315 glBindFramebuffer(GL_FRAMEBUFFER, 0);
316 glBindTexture(GL_TEXTURE_2D, rgba_tex);
317 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
318 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
319 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
321 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
324 // Make sure the H.264 gets a reference to all the
325 // input frames needed, so that they are not released back
326 // until the rendering is done.
327 vector<RefCountedFrame> input_frames;
328 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
329 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
331 h264_encoder->end_frame(fence, input_frames);
333 // The live frame just shows the RGBA texture we just rendered.
334 // It owns rgba_tex now.
335 DisplayFrame live_frame;
336 live_frame.chain = display_chain.get();
337 live_frame.setup_chain = [this, rgba_tex]{
338 display_input->set_texture_num(rgba_tex);
340 live_frame.ready_fence = fence;
341 live_frame.input_frames = {};
342 live_frame.temp_textures = { rgba_tex };
343 output_channel[OUTPUT_LIVE].output_frame(live_frame);
345 // Set up preview and any additional channels.
346 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
347 DisplayFrame display_frame;
348 pair<EffectChain *, function<void()>> chain = theme->get_chain(i, frame / 60.0f, WIDTH, HEIGHT); // FIXME: dimensions
349 display_frame.chain = chain.first;
350 display_frame.setup_chain = chain.second;
351 display_frame.ready_fence = fence;
352 display_frame.input_frames = { bmusb_current_rendering_frame[0], bmusb_current_rendering_frame[1] }; // FIXME: possible to do better?
353 display_frame.temp_textures = {};
354 output_channel[i].output_frame(display_frame);
357 clock_gettime(CLOCK_MONOTONIC, &now);
358 double elapsed = now.tv_sec - start.tv_sec +
359 1e-9 * (now.tv_nsec - start.tv_nsec);
360 if (frame % 100 == 0) {
361 printf("%d frames in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
362 frame, elapsed, frame / elapsed,
363 1e3 * elapsed / frame);
364 // chain->print_phase_timing();
367 // Reset every 100 frames, so that local variations in frame times
368 // (especially for the first few frames, when the shaders are
369 // compiled etc.) don't make it hard to measure for the entire
370 // remaining duration of the program.
371 if (frame == 10000) {
378 resource_pool->clean_context();
381 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
384 glGenVertexArrays(1, &vao);
393 glBindVertexArray(vao);
397 GLuint fbo = resource_pool->create_fbo(dst_tex);
398 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
399 glViewport(0, 0, WIDTH/2, HEIGHT/2);
402 glUseProgram(cbcr_program_num);
405 glActiveTexture(GL_TEXTURE0);
407 glBindTexture(GL_TEXTURE_2D, src_tex);
409 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
411 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
413 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
416 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
417 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
419 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
420 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
422 glDrawArrays(GL_TRIANGLES, 0, 3);
425 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
426 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
431 resource_pool->release_fbo(fbo);
432 glDeleteVertexArrays(1, &vao);
435 void Mixer::release_display_frame(DisplayFrame *frame)
437 for (GLuint texnum : frame->temp_textures) {
438 resource_pool->release_2d_texture(texnum);
440 frame->temp_textures.clear();
441 frame->ready_fence.reset();
442 frame->input_frames.clear();
447 mixer_thread = std::thread(&Mixer::thread_func, this);
456 void Mixer::transition_clicked(int transition_num)
458 theme->transition_clicked(transition_num, frame / 60.0);
461 void Mixer::channel_clicked(int preview_num)
463 theme->channel_clicked(preview_num);
466 Mixer::OutputChannel::~OutputChannel()
468 if (has_current_frame) {
469 parent->release_display_frame(¤t_frame);
471 if (has_ready_frame) {
472 parent->release_display_frame(&ready_frame);
476 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
478 // Store this frame for display. Remove the ready frame if any
479 // (it was seemingly never used).
481 std::unique_lock<std::mutex> lock(frame_mutex);
482 if (has_ready_frame) {
483 parent->release_display_frame(&ready_frame);
486 has_ready_frame = true;
489 if (has_new_frame_ready_callback) {
490 new_frame_ready_callback();
494 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
496 std::unique_lock<std::mutex> lock(frame_mutex);
497 if (!has_current_frame && !has_ready_frame) {
501 if (has_current_frame && has_ready_frame) {
502 // We have a new ready frame. Toss the current one.
503 parent->release_display_frame(¤t_frame);
504 has_current_frame = false;
506 if (has_ready_frame) {
507 assert(!has_current_frame);
508 current_frame = ready_frame;
509 ready_frame.ready_fence.reset(); // Drop the refcount.
510 ready_frame.input_frames.clear(); // Drop the refcounts.
511 has_current_frame = true;
512 has_ready_frame = false;
515 *frame = current_frame;
519 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
521 new_frame_ready_callback = callback;
522 has_new_frame_ready_callback = true;