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
  45 class QOpenGLContext;
  46
  47 using namespace movit;
  48 using namespace std;
  49 using namespace std::placeholders;
  50
  51 Mixer *global_mixer = nullptr;
  52
  53 namespace {
  54
  55 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  56 {
  57         for (size_t i = 0; i < num_samples; ++i) {
  58                 for (size_t j = 0; j < out_channels; ++j) {
  59                         uint32_t s1 = *src++;
  60                         uint32_t s2 = *src++;
  61                         uint32_t s3 = *src++;
  62                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  63                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  64                 }
  65                 src += 3 * (in_channels - out_channels);
  66         }
  67 }
  68
  69 }  // namespace
  70
  71 Mixer::Mixer(const QSurfaceFormat &format)
  72         : mixer_surface(create_surface(format)),
  73           h264_encoder_surface(create_surface(format))
  74 {
  75         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
  76         check_error();
  77
  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;
  84
  85         ImageFormat inout_format;
  86         inout_format.color_space = COLORSPACE_sRGB;
  87         inout_format.gamma_curve = GAMMA_sRGB;
  88
  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()));
  91         check_error();
  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();
  97
  98         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, "test.mp4"));
  99
 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(
 109                         [card]{
 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");
 114                                         exit(1);
 115                                 }
 116                                 printf("inited!\n");
 117                         },
 118                         [this]{
 119                                 resource_pool->clean_context();
 120                         });
 121                 card->resampler = new Resampler(48000.0, 48000.0, 2);
 122                 card->usb->configure_card();
 123         }
 124
 125         BMUSBCapture::start_bm_thread();
 126
 127         for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
 128                 cards[card_index].usb->start_bm_capture();
 129         }
 130
 131         //chain->enable_phase_timing(true);
 132
 133         // Set up stuff for NV12 conversion.
 134
 135         // Cb/Cr shader.
 136         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 137         string cbcr_frag_shader =
 138                 "#version 130 \n"
 139                 "in vec2 tc0; \n"
 140                 "uniform sampler2D cbcr_tex; \n"
 141                 "void main() { \n"
 142                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 143                 "} \n";
 144         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
 145 }
 146
 147 Mixer::~Mixer()
 148 {
 149         resource_pool->release_glsl_program(cbcr_program_num);
 150         BMUSBCapture::stop_bm_thread();
 151
 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();
 156         }
 157 }
 158
 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)
 162 {
 163         CaptureCard *card = &cards[card_index];
 164
 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);
 169                 fclose(fp);
 170                 //exit(1);
 171                 card->usb->get_video_frame_allocator()->release_frame(video_frame);
 172                 card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
 173                 return;
 174         }
 175         {
 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; });
 179         }
 180         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
 181         GLuint pbo = userdata->pbo;
 182         check_error();
 183         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 184         check_error();
 185         glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
 186         check_error();
 187         //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 188         //check_error();
 189
 190         // Upload the textures.
 191         glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
 192         check_error();
 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));
 194         check_error();
 195         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
 196         check_error();
 197         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280/2, 720, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(1280 * 25 + 22));
 198         check_error();
 199         glBindTexture(GL_TEXTURE_2D, 0);
 200         check_error();
 201         GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 202         check_error();
 203         assert(fence != nullptr);
 204         {
 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();
 210         }
 211
 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);
 220
 221                 float samples_out[(48000 / 60) * 2];
 222                 card->resampler->get_output_samples(pts, samples_out, 48000 / 60);
 223
 224                 static FILE *audiofp = nullptr;
 225                 if (audiofp == nullptr) {
 226                         audiofp = fopen("audio.raw", "wb");
 227                 }
 228                 fwrite(samples_out, sizeof(samples_out), 1, audiofp);
 229                 //fwrite(samplesf.get(), num_samples * sizeof(float) * 2, 1, audiofp);
 230
 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));
 233                 }
 234         }
 235
 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);
 238 }
 239
 240 void Mixer::thread_func()
 241 {
 242         eglBindAPI(EGL_OPENGL_API);
 243         QOpenGLContext *context = create_context();
 244         if (!make_current(context, mixer_surface)) {
 245                 printf("oops\n");
 246                 exit(1);
 247         }
 248
 249         struct timespec start, now;
 250         clock_gettime(CLOCK_MONOTONIC, &start);
 251
 252         while (!should_quit) {
 253                 ++frame;
 254
 255                 CaptureCard card_copy[NUM_CARDS];
 256
 257                 {
 258                         std::unique_lock<std::mutex> lock(bmusb_mutex);
 259
 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; });
 263
 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();
 272                         }
 273                 }
 274
 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)
 278                                 continue;
 279
 280                         assert(card->new_frame != nullptr);
 281                         bmusb_current_rendering_frame[card_index] = card->new_frame;
 282                         check_error();
 283
 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);
 288                         check_error();
 289                         glDeleteSync(card->new_data_ready_fence);
 290                         check_error();
 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);
 293                 }
 294
 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();
 299
 300                 GLuint y_tex, cbcr_tex;
 301                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 302                 assert(got_frame);
 303
 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);
 310
 311                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 312                 resource_pool->release_2d_texture(cbcr_full_tex);
 313
 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);
 320
 321                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 322                 check_error();
 323
 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]);
 330                 }
 331                 h264_encoder->end_frame(fence, input_frames);
 332
 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);
 339                 };
 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);
 344
 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);
 355                 }
 356
 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();
 365                 }
 366
 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) {
 372                         frame = 0;
 373                         start = now;
 374                 }
 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 }