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;
53 Mixer::Mixer(const QSurfaceFormat &format)
54 : mixer_surface(create_surface(format)),
55 h264_encoder_surface(create_surface(format))
57 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
60 resource_pool.reset(new ResourcePool);
61 output_channel[OUTPUT_LIVE].parent = this;
62 output_channel[OUTPUT_PREVIEW].parent = this;
64 ImageFormat inout_format;
65 inout_format.color_space = COLORSPACE_sRGB;
66 inout_format.gamma_curve = GAMMA_sRGB;
68 YCbCrFormat input_ycbcr_format;
69 input_ycbcr_format.chroma_subsampling_x = 2;
70 input_ycbcr_format.chroma_subsampling_y = 1;
71 input_ycbcr_format.cb_x_position = 0.0;
72 input_ycbcr_format.cr_x_position = 0.0;
73 input_ycbcr_format.cb_y_position = 0.5;
74 input_ycbcr_format.cr_y_position = 0.5;
75 input_ycbcr_format.luma_coefficients = YCBCR_REC_601;
76 input_ycbcr_format.full_range = false;
78 YCbCrFormat output_ycbcr_format;
79 output_ycbcr_format.chroma_subsampling_x = 1;
80 output_ycbcr_format.chroma_subsampling_y = 1;
81 output_ycbcr_format.luma_coefficients = YCBCR_REC_601;
82 output_ycbcr_format.full_range = false;
85 chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
87 input[0] = new YCbCrInput(inout_format, input_ycbcr_format, WIDTH, HEIGHT, YCBCR_INPUT_SPLIT_Y_AND_CBCR);
88 chain->add_input(input[0]);
89 input[1] = new YCbCrInput(inout_format, input_ycbcr_format, WIDTH, HEIGHT, YCBCR_INPUT_SPLIT_Y_AND_CBCR);
90 chain->add_input(input[1]);
91 resample_effect = chain->add_effect(new ResampleEffect(), input[0]);
92 padding_effect = chain->add_effect(new IntegralPaddingEffect());
93 float border_color[] = { 0.0f, 0.0f, 0.0f, 1.0f };
94 CHECK(padding_effect->set_vec4("border_color", border_color));
96 resample2_effect = chain->add_effect(new ResampleEffect(), input[1]);
97 Effect *saturation_effect = chain->add_effect(new SaturationEffect());
98 CHECK(saturation_effect->set_float("saturation", 0.3f));
99 Effect *wb_effect = chain->add_effect(new WhiteBalanceEffect());
100 CHECK(wb_effect->set_float("output_color_temperature", 3500.0));
101 padding2_effect = chain->add_effect(new IntegralPaddingEffect());
103 chain->add_effect(new OverlayEffect(), padding_effect, padding2_effect);
105 chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
106 chain->add_ycbcr_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, output_ycbcr_format, YCBCR_OUTPUT_SPLIT_Y_AND_CBCR);
107 chain->set_dither_bits(8);
108 chain->set_output_origin(OUTPUT_ORIGIN_TOP_LEFT);
111 // Display chain; shows the live output produced by the main chain (its RGBA version).
112 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
114 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
115 display_chain->add_input(display_input);
116 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
117 display_chain->set_dither_bits(0); // Don't bother.
118 display_chain->finalize();
120 // Preview chain (always shows just first input for now).
121 preview_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
123 preview_input = new YCbCrInput(inout_format, input_ycbcr_format, WIDTH, HEIGHT, YCBCR_INPUT_SPLIT_Y_AND_CBCR);
124 preview_chain->add_input(preview_input);
125 preview_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
126 preview_chain->set_dither_bits(0); // Don't bother.
127 preview_chain->finalize();
129 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, "test.mp4"));
131 printf("Configuring first card...\n");
132 cards[0].usb = new BMUSBCapture(0x1edb, 0xbd3b); // 0xbd4f
133 cards[0].usb->set_frame_callback(std::bind(&Mixer::bm_frame, this, 0, _1, _2, _3, _4, _5, _6, _7));
134 cards[0].frame_allocator.reset(new PBOFrameAllocator(1280 * 750 * 2 + 44, 1280, 720));
135 cards[0].usb->set_video_frame_allocator(cards[0].frame_allocator.get());
136 cards[0].usb->configure_card();
137 cards[0].surface = create_surface(format);
139 cards[1].surface = create_surface(format);
142 if (NUM_CARDS == 2) {
143 printf("Configuring second card...\n");
144 cards[1].usb = new BMUSBCapture(0x1edb, 0xbd4f);
145 cards[1].usb->set_frame_callback(std::bind(&Mixer::bm_frame, this, 1, _1, _2, _3, _4, _5, _6, _7));
146 cards[1].frame_allocator.reset(new PBOFrameAllocator(1280 * 750 * 2 + 44, 1280, 720));
147 cards[1].usb->set_video_frame_allocator(cards[1].frame_allocator.get());
148 cards[1].usb->configure_card();
151 BMUSBCapture::start_bm_thread();
153 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
154 cards[card_index].usb->start_bm_capture();
155 input[card_index]->set_pixel_data(0, nullptr, 0);
156 input[card_index]->set_pixel_data(1, nullptr, 0);
159 //chain->enable_phase_timing(true);
161 // Set up stuff for NV12 conversion.
164 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
165 string cbcr_frag_shader =
168 "uniform sampler2D cbcr_tex; \n"
170 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
172 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
177 resource_pool->release_glsl_program(cbcr_program_num);
178 BMUSBCapture::stop_bm_thread();
181 void Mixer::bm_frame(int card_index, uint16_t timecode,
182 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
183 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
185 CaptureCard *card = &cards[card_index];
186 if (!card->thread_initialized) {
187 printf("initializing context for bmusb thread %d\n", card_index);
188 eglBindAPI(EGL_OPENGL_API);
189 card->context = create_context();
190 if (!make_current(card->context, card->surface)) {
191 printf("failed to create bmusb context\n");
194 card->thread_initialized = true;
197 if (video_frame.len - video_offset != 1280 * 750 * 2) {
198 printf("dropping frame with wrong length (%ld)\n", video_frame.len - video_offset);
199 FILE *fp = fopen("frame.raw", "wb");
200 fwrite(video_frame.data, video_frame.len, 1, fp);
203 card->usb->get_video_frame_allocator()->release_frame(video_frame);
204 card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
208 // Wait until the previous frame was consumed.
209 std::unique_lock<std::mutex> lock(bmusb_mutex);
210 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready; });
212 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
213 GLuint pbo = userdata->pbo;
215 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
217 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
219 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
222 // Upload the textures.
223 glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
225 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280, 720, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((1280 * 750 * 2 + 44) / 2 + 1280 * 25 + 22));
227 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
229 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280/2, 720, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(1280 * 25 + 22));
231 glBindTexture(GL_TEXTURE_2D, 0);
233 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
235 assert(fence != nullptr);
237 std::unique_lock<std::mutex> lock(bmusb_mutex);
238 card->new_data_ready = true;
239 card->new_frame = RefCountedFrame(video_frame);
240 card->new_data_ready_fence = fence;
241 card->new_data_ready_changed.notify_all();
244 // Video frame will be released when last user of card->new_frame goes out of scope.
245 card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
248 void Mixer::place_rectangle(Effect *resample_effect, Effect *padding_effect, float x0, float y0, float x1, float y1)
256 if (x0 > 1280.0 || x1 < 0.0 || y0 > 720.0 || y1 < 0.0) {
257 CHECK(resample_effect->set_int("width", 1));
258 CHECK(resample_effect->set_int("height", 1));
259 CHECK(resample_effect->set_float("zoom_x", 1280.0));
260 CHECK(resample_effect->set_float("zoom_y", 720.0));
261 CHECK(padding_effect->set_int("left", 2000));
262 CHECK(padding_effect->set_int("top", 2000));
266 // Clip. (TODO: Clip on upper/left sides, too.)
268 srcx1 = (1280.0 - x0) / (x1 - x0);
272 srcy1 = (720.0 - y0) / (y1 - y0);
276 float x_subpixel_offset = x0 - floor(x0);
277 float y_subpixel_offset = y0 - floor(y0);
279 // Resampling must be to an integral number of pixels. Round up,
280 // and then add an extra pixel so we have some leeway for the border.
281 int width = int(ceil(x1 - x0)) + 1;
282 int height = int(ceil(y1 - y0)) + 1;
283 CHECK(resample_effect->set_int("width", width));
284 CHECK(resample_effect->set_int("height", height));
286 // Correct the discrepancy with zoom. (This will leave a small
287 // excess edge of pixels and subpixels, which we'll correct for soon.)
288 float zoom_x = (x1 - x0) / (width * (srcx1 - srcx0));
289 float zoom_y = (y1 - y0) / (height * (srcy1 - srcy0));
290 CHECK(resample_effect->set_float("zoom_x", zoom_x));
291 CHECK(resample_effect->set_float("zoom_y", zoom_y));
292 CHECK(resample_effect->set_float("zoom_center_x", 0.0f));
293 CHECK(resample_effect->set_float("zoom_center_y", 0.0f));
295 // Padding must also be to a whole-pixel offset.
296 CHECK(padding_effect->set_int("left", floor(x0)));
297 CHECK(padding_effect->set_int("top", floor(y0)));
299 // Correct _that_ discrepancy by subpixel offset in the resampling.
300 CHECK(resample_effect->set_float("left", -x_subpixel_offset / zoom_x));
301 CHECK(resample_effect->set_float("top", -y_subpixel_offset / zoom_y));
303 // Finally, adjust the border so it is exactly where we want it.
304 CHECK(padding_effect->set_float("border_offset_left", x_subpixel_offset));
305 CHECK(padding_effect->set_float("border_offset_right", x1 - (floor(x0) + width)));
306 CHECK(padding_effect->set_float("border_offset_top", y_subpixel_offset));
307 CHECK(padding_effect->set_float("border_offset_bottom", y1 - (floor(y0) + height)));
310 void Mixer::thread_func()
312 eglBindAPI(EGL_OPENGL_API);
313 QOpenGLContext *context = create_context();
314 if (!make_current(context, mixer_surface)) {
319 struct timespec start, now;
320 clock_gettime(CLOCK_MONOTONIC, &start);
322 while (!should_quit) {
325 //int width0 = lrintf(848 * (1.0 + 0.2 * sin(frame * 0.02)));
327 int height0 = lrintf(width0 * 9.0 / 16.0);
329 //float top0 = 96 + 48 * sin(frame * 0.005);
330 //float left0 = 96 + 48 * cos(frame * 0.006);
333 float bottom0 = top0 + height0;
334 float right0 = left0 + width0;
339 float bottom1 = 720 - 48;
340 float right1 = 1280 - 16;
341 float top1 = bottom1 - height1;
342 float left1 = right1 - width1;
344 if (current_source == SOURCE_INPUT1) {
351 bottom1 = HEIGHT + 20;
354 } else if (current_source == SOURCE_INPUT2) {
361 bottom0 = HEIGHT + 20;
365 float t = 0.5 + 0.5 * cos(frame * 0.006);
366 float scale0 = 1.0 + t * (1280.0 / 848.0 - 1.0);
367 float tx0 = 0.0 + t * (-16.0 * scale0);
368 float ty0 = 0.0 + t * (-48.0 * scale0);
370 top0 = top0 * scale0 + ty0;
371 bottom0 = bottom0 * scale0 + ty0;
372 left0 = left0 * scale0 + tx0;
373 right0 = right0 * scale0 + tx0;
375 top1 = top1 * scale0 + ty0;
376 bottom1 = bottom1 * scale0 + ty0;
377 left1 = left1 * scale0 + tx0;
378 right1 = right1 * scale0 + tx0;
381 place_rectangle(resample_effect, padding_effect, left0, top0, right0, bottom0);
382 place_rectangle(resample2_effect, padding2_effect, left1, top1, right1, bottom1);
384 CaptureCard card_copy[NUM_CARDS];
387 std::unique_lock<std::mutex> lock(bmusb_mutex);
389 // The first card is the master timer, so wait for it to have a new frame.
390 // TODO: Make configurable, and with a timeout.
391 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
393 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
394 CaptureCard *card = &cards[card_index];
395 card_copy[card_index].usb = card->usb;
396 card_copy[card_index].new_data_ready = card->new_data_ready;
397 card_copy[card_index].new_frame = card->new_frame;
398 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
399 card->new_data_ready = false;
400 card->new_data_ready_changed.notify_all();
404 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
405 CaptureCard *card = &card_copy[card_index];
406 if (!card->new_data_ready)
409 bmusb_current_rendering_frame[card_index] = card->new_frame;
412 // The new texture might still be uploaded,
413 // tell the GPU to wait until it's there.
414 if (card->new_data_ready_fence)
415 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
417 glDeleteSync(card->new_data_ready_fence);
419 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
420 input[card_index]->set_texture_num(0, userdata->tex_y);
421 input[card_index]->set_texture_num(1, userdata->tex_cbcr);
423 if (NUM_CARDS == 1) {
424 // Set to the other one, too.
425 input[1]->set_texture_num(0, userdata->tex_y);
426 input[1]->set_texture_num(1, userdata->tex_cbcr);
430 GLuint y_tex, cbcr_tex;
431 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
434 // Render main chain.
435 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
436 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
437 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
438 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
439 resource_pool->release_fbo(fbo);
441 subsample_chroma(cbcr_full_tex, cbcr_tex);
442 resource_pool->release_2d_texture(cbcr_full_tex);
444 // Set the right state for rgba_tex.
445 glBindFramebuffer(GL_FRAMEBUFFER, 0);
446 glBindTexture(GL_TEXTURE_2D, rgba_tex);
447 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
448 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
449 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
451 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
454 // Make sure the H.264 gets a reference to all the
455 // input frames needed, so that they are not released back
456 // until the rendering is done.
457 vector<RefCountedFrame> input_frames;
458 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
459 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
461 h264_encoder->end_frame(fence, input_frames);
463 // The live frame just shows the RGBA texture we just rendered.
464 // It owns rgba_tex now.
465 DisplayFrame live_frame;
466 live_frame.chain = display_chain.get();
467 live_frame.setup_chain = [this, rgba_tex]{
468 display_input->set_texture_num(rgba_tex);
470 live_frame.ready_fence = fence;
471 live_frame.input_frames = {};
472 live_frame.temp_textures = { rgba_tex };
473 output_channel[OUTPUT_LIVE].output_frame(live_frame);
475 // The preview frame shows the first input. Note that the textures
476 // are owned by the input frame, not the display frame.
477 const PBOFrameAllocator::Userdata *input0_userdata = (const PBOFrameAllocator::Userdata *)bmusb_current_rendering_frame[0]->userdata;
478 GLuint input0_y_tex = input0_userdata->tex_y;
479 GLuint input0_cbcr_tex = input0_userdata->tex_cbcr;
480 DisplayFrame preview_frame;
481 preview_frame.chain = preview_chain.get();
482 preview_frame.setup_chain = [this, input0_y_tex, input0_cbcr_tex]{
483 preview_input->set_texture_num(0, input0_y_tex);
484 preview_input->set_texture_num(1, input0_cbcr_tex);
486 preview_frame.ready_fence = fence;
487 preview_frame.input_frames = { bmusb_current_rendering_frame[0] };
488 preview_frame.temp_textures = {};
489 output_channel[OUTPUT_PREVIEW].output_frame(preview_frame);
491 clock_gettime(CLOCK_MONOTONIC, &now);
492 double elapsed = now.tv_sec - start.tv_sec +
493 1e-9 * (now.tv_nsec - start.tv_nsec);
494 if (frame % 100 == 0) {
495 printf("%d frames in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
496 frame, elapsed, frame / elapsed,
497 1e3 * elapsed / frame);
498 // chain->print_phase_timing();
501 // Reset every 100 frames, so that local variations in frame times
502 // (especially for the first few frames, when the shaders are
503 // compiled etc.) don't make it hard to measure for the entire
504 // remaining duration of the program.
505 if (frame == 10000) {
513 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
516 glGenVertexArrays(1, &vao);
525 glBindVertexArray(vao);
529 GLuint fbo = resource_pool->create_fbo(dst_tex);
530 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
531 glViewport(0, 0, WIDTH/2, HEIGHT/2);
534 glUseProgram(cbcr_program_num);
537 glActiveTexture(GL_TEXTURE0);
539 glBindTexture(GL_TEXTURE_2D, src_tex);
541 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
543 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
545 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
548 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
549 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
551 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
552 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
554 glDrawArrays(GL_TRIANGLES, 0, 3);
557 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
558 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
563 resource_pool->release_fbo(fbo);
564 glDeleteVertexArrays(1, &vao);
567 void Mixer::release_display_frame(DisplayFrame *frame)
569 for (GLuint texnum : frame->temp_textures) {
570 resource_pool->release_2d_texture(texnum);
572 frame->temp_textures.clear();
573 frame->ready_fence.reset();
574 frame->input_frames.clear();
579 mixer_thread = std::thread(&Mixer::thread_func, this);
588 void Mixer::cut(Source source)
590 current_source = source;
593 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
595 // Store this frame for display. Remove the ready frame if any
596 // (it was seemingly never used).
598 std::unique_lock<std::mutex> lock(frame_mutex);
599 if (has_ready_frame) {
600 parent->release_display_frame(&ready_frame);
603 has_ready_frame = true;
606 if (has_new_frame_ready_callback) {
607 new_frame_ready_callback();
611 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
613 std::unique_lock<std::mutex> lock(frame_mutex);
614 if (!has_current_frame && !has_ready_frame) {
618 if (has_current_frame && has_ready_frame) {
619 // We have a new ready frame. Toss the current one.
620 parent->release_display_frame(¤t_frame);
621 has_current_frame = false;
623 if (has_ready_frame) {
624 assert(!has_current_frame);
625 current_frame = ready_frame;
626 ready_frame.ready_fence.reset(); // Drop the refcount.
627 ready_frame.input_frames.clear(); // Drop the refcounts.
628 has_current_frame = true;
629 has_ready_frame = false;
632 *frame = current_frame;
636 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
638 new_frame_ready_callback = callback;
639 has_new_frame_ready_callback = true;