git.sesse.net Git - nageru/blob - mixer.cpp

   1 #define WIDTH 1280
   2 #define HEIGHT 720
   3
   4 #undef Success
   5
   6 #include "mixer.h"
   7
   8 #include <assert.h>
   9 #include <effect.h>
  10 #include <effect_chain.h>
  11 #include <effect_util.h>
  12 #include <epoxy/egl.h>
  13 #include <features.h>
  14 #include <image_format.h>
  15 #include <init.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>
  21 #include <stdint.h>
  22 #include <stdio.h>
  23 #include <stdlib.h>
  24 #include <sys/time.h>
  25 #include <time.h>
  26 #include <util.h>
  27 #include <white_balance_effect.h>
  28 #include <ycbcr.h>
  29 #include <ycbcr_input.h>
  30 #include <cmath>
  31 #include <condition_variable>
  32 #include <cstddef>
  33 #include <memory>
  34 #include <mutex>
  35 #include <string>
  36 #include <thread>
  37 #include <vector>
  38
  39 #include "bmusb.h"
  40 #include "context.h"
  41 #include "h264encode.h"
  42 #include "pbo_frame_allocator.h"
  43 #include "ref_counted_gl_sync.h"
  44 #include "timebase.h"
  45
  46 class QOpenGLContext;
  47
  48 using namespace movit;
  49 using namespace std;
  50 using namespace std::placeholders;
  51
  52 Mixer *global_mixer = nullptr;
  53
  54 namespace {
  55
  56 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  57 {
  58         for (size_t i = 0; i < num_samples; ++i) {
  59                 for (size_t j = 0; j < out_channels; ++j) {
  60                         uint32_t s1 = *src++;
  61                         uint32_t s2 = *src++;
  62                         uint32_t s3 = *src++;
  63                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  64                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  65                 }
  66                 src += 3 * (in_channels - out_channels);
  67         }
  68 }
  69
  70 }  // namespace
  71
  72 Mixer::Mixer(const QSurfaceFormat &format)
  73         : mixer_surface(create_surface(format)),
  74           h264_encoder_surface(create_surface(format))
  75 {
  76         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
  77         check_error();
  78
  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;
  85
  86         ImageFormat inout_format;
  87         inout_format.color_space = COLORSPACE_sRGB;
  88         inout_format.gamma_curve = GAMMA_sRGB;
  89
  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()));
  92         check_error();
  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();
  98
  99         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, "test.mp4"));
 100
 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(
 110                         [card]{
 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");
 115                                         exit(1);
 116                                 }
 117                                 printf("inited!\n");
 118                         },
 119                         [this]{
 120                                 resource_pool->clean_context();
 121                         });
 122                 card->resampler = new Resampler(48000.0, 48000.0, 2);
 123                 card->usb->configure_card();
 124         }
 125
 126         BMUSBCapture::start_bm_thread();
 127
 128         for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
 129                 cards[card_index].usb->start_bm_capture();
 130         }
 131
 132         //chain->enable_phase_timing(true);
 133
 134         // Set up stuff for NV12 conversion.
 135
 136         // Cb/Cr shader.
 137         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 138         string cbcr_frag_shader =
 139                 "#version 130 \n"
 140                 "in vec2 tc0; \n"
 141                 "uniform sampler2D cbcr_tex; \n"
 142                 "void main() { \n"
 143                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 144                 "} \n";
 145         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
 146 }
 147
 148 Mixer::~Mixer()
 149 {
 150         resource_pool->release_glsl_program(cbcr_program_num);
 151         BMUSBCapture::stop_bm_thread();
 152
 153         for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
 154                 {
 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();
 158                 }
 159                 cards[card_index].usb->stop_dequeue_thread();
 160         }
 161 }
 162
 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)
 166 {
 167         CaptureCard *card = &cards[card_index];
 168
 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                 FILE *fp = fopen("frame.raw", "wb");
 172                 fwrite(video_frame.data, video_frame.len, 1, fp);
 173                 fclose(fp);
 174                 //exit(1);
 175                 card->usb->get_video_frame_allocator()->release_frame(video_frame);
 176                 card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
 177                 return;
 178         }
 179         {
 180                 // Wait until the previous frame was consumed.
 181                 std::unique_lock<std::mutex> lock(bmusb_mutex);
 182                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 183                 if (card->should_quit) return;
 184         }
 185         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
 186         GLuint pbo = userdata->pbo;
 187         check_error();
 188         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 189         check_error();
 190         glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
 191         check_error();
 192         //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 193         //check_error();
 194
 195         // Upload the textures.
 196         glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
 197         check_error();
 198         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280, 720, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((1280 * 750 * 2 + 44) / 2 + 1280 * 25 + 22));
 199         check_error();
 200         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
 201         check_error();
 202         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280/2, 720, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(1280 * 25 + 22));
 203         check_error();
 204         glBindTexture(GL_TEXTURE_2D, 0);
 205         check_error();
 206         GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 207         check_error();
 208         assert(fence != nullptr);
 209
 210         // Convert the audio to stereo fp32 and store it next to the video.
 211         size_t num_samples = (audio_frame.len - audio_offset) / 8 / 3;
 212         vector<float> audio;
 213         audio.resize(num_samples * 2);
 214         convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
 215
 216         {
 217                 std::unique_lock<std::mutex> lock(bmusb_mutex);
 218                 card->new_data_ready = true;
 219                 card->new_frame = RefCountedFrame(video_frame);
 220                 card->new_data_ready_fence = fence;
 221                 card->new_frame_audio = move(audio);
 222                 card->new_data_ready_changed.notify_all();
 223         }
 224
 225         // Video frame will be released when last user of card->new_frame goes out of scope.
 226         card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
 227 }
 228
 229 void Mixer::thread_func()
 230 {
 231         eglBindAPI(EGL_OPENGL_API);
 232         QOpenGLContext *context = create_context();
 233         if (!make_current(context, mixer_surface)) {
 234                 printf("oops\n");
 235                 exit(1);
 236         }
 237
 238         struct timespec start, now;
 239         clock_gettime(CLOCK_MONOTONIC, &start);
 240
 241         while (!should_quit) {
 242                 ++frame;
 243
 244                 CaptureCard card_copy[NUM_CARDS];
 245
 246                 {
 247                         std::unique_lock<std::mutex> lock(bmusb_mutex);
 248
 249                         // The first card is the master timer, so wait for it to have a new frame.
 250                         // TODO: Make configurable, and with a timeout.
 251                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 252
 253                         for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
 254                                 CaptureCard *card = &cards[card_index];
 255                                 card_copy[card_index].usb = card->usb;
 256                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 257                                 card_copy[card_index].new_frame = card->new_frame;
 258                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 259                                 card_copy[card_index].new_frame_audio = move(card->new_frame_audio);
 260                                 card->new_data_ready = false;
 261                                 card->new_data_ready_changed.notify_all();
 262                         }
 263                 }
 264
 265                 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
 266                         CaptureCard *card = &card_copy[card_index];
 267                         if (!card->new_data_ready)
 268                                 continue;
 269
 270                         assert(card->new_frame != nullptr);
 271                         bmusb_current_rendering_frame[card_index] = card->new_frame;
 272                         check_error();
 273
 274                         // The new texture might still be uploaded,
 275                         // tell the GPU to wait until it's there.
 276                         if (card->new_data_ready_fence)
 277                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 278                         check_error();
 279                         glDeleteSync(card->new_data_ready_fence);
 280                         check_error();
 281                         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
 282                         theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
 283                 }
 284
 285                 // Get the main chain from the theme, and set its state immediately.
 286                 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, frame / 60.0f, WIDTH, HEIGHT);
 287                 EffectChain *chain = theme_main_chain.first;
 288                 theme_main_chain.second();
 289
 290                 GLuint y_tex, cbcr_tex;
 291                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 292                 assert(got_frame);
 293
 294                 // Render main chain.
 295                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 296                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 297                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 298                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 299                 resource_pool->release_fbo(fbo);
 300
 301                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 302                 resource_pool->release_2d_texture(cbcr_full_tex);
 303
 304                 // Set the right state for rgba_tex.
 305                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 306                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 307                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 308                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 309                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 310
 311                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 312                 check_error();
 313
 314                 // Resample the audio as needed.
 315                 // TODO: Allow using audio from the other card(s) as well.
 316                 double pts = frame / 60.0;
 317                 cards[0].resampler->add_input_samples(pts, card_copy[0].new_frame_audio.data(), card_copy[0].new_frame_audio.size() / 2);
 318                 vector<float> samples_out;
 319                 samples_out.resize((48000 / 60) * 2);
 320                 cards[0].resampler->get_output_samples(pts, &samples_out[0], 48000 / 60);
 321
 322                 // Make sure the H.264 gets a reference to all the
 323                 // input frames needed, so that they are not released back
 324                 // until the rendering is done.
 325                 vector<RefCountedFrame> input_frames;
 326                 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
 327                         input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 328                 }
 329                 h264_encoder->end_frame(fence, frame * (TIMEBASE / 60), move(samples_out), input_frames);
 330
 331                 // The live frame just shows the RGBA texture we just rendered.
 332                 // It owns rgba_tex now.
 333                 DisplayFrame live_frame;
 334                 live_frame.chain = display_chain.get();
 335                 live_frame.setup_chain = [this, rgba_tex]{
 336                         display_input->set_texture_num(rgba_tex);
 337                 };
 338                 live_frame.ready_fence = fence;
 339                 live_frame.input_frames = {};
 340                 live_frame.temp_textures = { rgba_tex };
 341                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 342
 343                 // Set up preview and any additional channels.
 344                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 345                         DisplayFrame display_frame;
 346                         pair<EffectChain *, function<void()>> chain = theme->get_chain(i, frame / 60.0f, WIDTH, HEIGHT);  // FIXME: dimensions
 347                         display_frame.chain = chain.first;
 348                         display_frame.setup_chain = chain.second;
 349                         display_frame.ready_fence = fence;
 350                         display_frame.input_frames = { bmusb_current_rendering_frame[0], bmusb_current_rendering_frame[1] };  // FIXME: possible to do better?
 351                         display_frame.temp_textures = {};
 352                         output_channel[i].output_frame(display_frame);
 353                 }
 354
 355                 clock_gettime(CLOCK_MONOTONIC, &now);
 356                 double elapsed = now.tv_sec - start.tv_sec +
 357                         1e-9 * (now.tv_nsec - start.tv_nsec);
 358                 if (frame % 100 == 0) {
 359                         printf("%d frames in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 360                                 frame, elapsed, frame / elapsed,
 361                                 1e3 * elapsed / frame);
 362                 //      chain->print_phase_timing();
 363                 }
 364
 365 #if 0
 366                 // Reset every 100 frames, so that local variations in frame times
 367                 // (especially for the first few frames, when the shaders are
 368                 // compiled etc.) don't make it hard to measure for the entire
 369                 // remaining duration of the program.
 370                 if (frame == 10000) {
 371                         frame = 0;
 372                         start = now;
 373                 }
 374 #endif
 375                 check_error();
 376         }
 377
 378         resource_pool->clean_context();
 379 }
 380
 381 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 382 {
 383         GLuint vao;
 384         glGenVertexArrays(1, &vao);
 385         check_error();
 386
 387         float vertices[] = {
 388                 0.0f, 2.0f,
 389                 0.0f, 0.0f,
 390                 2.0f, 0.0f
 391         };
 392
 393         glBindVertexArray(vao);
 394         check_error();
 395
 396         // Extract Cb/Cr.
 397         GLuint fbo = resource_pool->create_fbo(dst_tex);
 398         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 399         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 400         check_error();
 401
 402         glUseProgram(cbcr_program_num);
 403         check_error();
 404
 405         glActiveTexture(GL_TEXTURE0);
 406         check_error();
 407         glBindTexture(GL_TEXTURE_2D, src_tex);
 408         check_error();
 409         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 410         check_error();
 411         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 412         check_error();
 413         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 414         check_error();
 415
 416         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 417         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 418
 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.
 421
 422         glDrawArrays(GL_TRIANGLES, 0, 3);
 423         check_error();
 424
 425         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 426         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 427
 428         glUseProgram(0);
 429         check_error();
 430
 431         resource_pool->release_fbo(fbo);
 432         glDeleteVertexArrays(1, &vao);
 433 }
 434
 435 void Mixer::release_display_frame(DisplayFrame *frame)
 436 {
 437         for (GLuint texnum : frame->temp_textures) {
 438                 resource_pool->release_2d_texture(texnum);
 439         }
 440         frame->temp_textures.clear();
 441         frame->ready_fence.reset();
 442         frame->input_frames.clear();
 443 }
 444
 445 void Mixer::start()
 446 {
 447         mixer_thread = std::thread(&Mixer::thread_func, this);
 448 }
 449
 450 void Mixer::quit()
 451 {
 452         should_quit = true;
 453         mixer_thread.join();
 454 }
 455
 456 void Mixer::transition_clicked(int transition_num)
 457 {
 458         theme->transition_clicked(transition_num, frame / 60.0);
 459 }
 460
 461 void Mixer::channel_clicked(int preview_num)
 462 {
 463         theme->channel_clicked(preview_num);
 464 }
 465
 466 Mixer::OutputChannel::~OutputChannel()
 467 {
 468         if (has_current_frame) {
 469                 parent->release_display_frame(&current_frame);
 470         }
 471         if (has_ready_frame) {
 472                 parent->release_display_frame(&ready_frame);
 473         }
 474 }
 475
 476 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 477 {
 478         // Store this frame for display. Remove the ready frame if any
 479         // (it was seemingly never used).
 480         {
 481                 std::unique_lock<std::mutex> lock(frame_mutex);
 482                 if (has_ready_frame) {
 483                         parent->release_display_frame(&ready_frame);
 484                 }
 485                 ready_frame = frame;
 486                 has_ready_frame = true;
 487         }
 488
 489         if (has_new_frame_ready_callback) {
 490                 new_frame_ready_callback();
 491         }
 492 }
 493
 494 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 495 {
 496         std::unique_lock<std::mutex> lock(frame_mutex);
 497         if (!has_current_frame && !has_ready_frame) {
 498                 return false;
 499         }
 500
 501         if (has_current_frame && has_ready_frame) {
 502                 // We have a new ready frame. Toss the current one.
 503                 parent->release_display_frame(&current_frame);
 504                 has_current_frame = false;
 505         }
 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;
 513         }
 514
 515         *frame = current_frame;
 516         return true;
 517 }
 518
 519 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 520 {
 521         new_frame_ready_callback = callback;
 522         has_new_frame_ready_callback = true;
 523 }