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 Mixer::Mixer(const QSurfaceFormat &format)
  54         : mixer_surface(create_surface(format)),
  55           h264_encoder_surface(create_surface(format))
  56 {
  57         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
  58         check_error();
  59
  60         resource_pool.reset(new ResourcePool);
  61         theme.reset(new Theme("theme.lua", resource_pool.get()));
  62         output_channel[OUTPUT_LIVE].parent = this;
  63         output_channel[OUTPUT_PREVIEW].parent = this;
  64         output_channel[OUTPUT_INPUT0].parent = this;
  65         output_channel[OUTPUT_INPUT1].parent = this;
  66
  67         ImageFormat inout_format;
  68         inout_format.color_space = COLORSPACE_sRGB;
  69         inout_format.gamma_curve = GAMMA_sRGB;
  70
  71         // Display chain; shows the live output produced by the main chain (its RGBA version).
  72         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
  73         check_error();
  74         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
  75         display_chain->add_input(display_input);
  76         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
  77         display_chain->set_dither_bits(0);  // Don't bother.
  78         display_chain->finalize();
  79
  80         h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, "test.mp4"));
  81
  82         for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
  83                 printf("Configuring card %d...\n", card_index);
  84                 CaptureCard *card = &cards[card_index];
  85                 card->usb = new BMUSBCapture(0x1edb, card_index == 0 ? 0xbd3b : 0xbd4f);
  86                 card->usb->set_frame_callback(std::bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
  87                 card->frame_allocator.reset(new PBOFrameAllocator(1280 * 750 * 2 + 44, 1280, 720));
  88                 card->usb->set_video_frame_allocator(card->frame_allocator.get());
  89                 card->surface = create_surface(format);
  90                 card->usb->set_dequeue_thread_callbacks(
  91                         [card]{
  92                                 eglBindAPI(EGL_OPENGL_API);
  93                                 card->context = create_context();
  94                                 if (!make_current(card->context, card->surface)) {
  95                                         printf("failed to create bmusb context\n");
  96                                         exit(1);
  97                                 }
  98                                 printf("inited!\n");
  99                         },
 100                         [this]{
 101                                 resource_pool->clean_context();
 102                         });
 103                 card->usb->configure_card();
 104         }
 105
 106         BMUSBCapture::start_bm_thread();
 107
 108         for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
 109                 cards[card_index].usb->start_bm_capture();
 110         }
 111
 112         //chain->enable_phase_timing(true);
 113
 114         // Set up stuff for NV12 conversion.
 115
 116         // Cb/Cr shader.
 117         string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
 118         string cbcr_frag_shader =
 119                 "#version 130 \n"
 120                 "in vec2 tc0; \n"
 121                 "uniform sampler2D cbcr_tex; \n"
 122                 "void main() { \n"
 123                 "    gl_FragColor = texture2D(cbcr_tex, tc0); \n"
 124                 "} \n";
 125         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
 126 }
 127
 128 Mixer::~Mixer()
 129 {
 130         resource_pool->release_glsl_program(cbcr_program_num);
 131         BMUSBCapture::stop_bm_thread();
 132
 133         for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
 134                 cards[card_index].new_data_ready = false;  // Unblock thread.
 135                 cards[card_index].new_data_ready_changed.notify_all();
 136                 cards[card_index].usb->stop_dequeue_thread();
 137         }
 138 }
 139
 140 void Mixer::bm_frame(int card_index, uint16_t timecode,
 141                      FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
 142                      FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
 143 {
 144         CaptureCard *card = &cards[card_index];
 145
 146         if (video_frame.len - video_offset != 1280 * 750 * 2) {
 147                 printf("dropping frame with wrong length (%ld)\n", video_frame.len - video_offset);
 148                 FILE *fp = fopen("frame.raw", "wb");
 149                 fwrite(video_frame.data, video_frame.len, 1, fp);
 150                 fclose(fp);
 151                 //exit(1);
 152                 card->usb->get_video_frame_allocator()->release_frame(video_frame);
 153                 card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
 154                 return;
 155         }
 156         {
 157                 // Wait until the previous frame was consumed.
 158                 std::unique_lock<std::mutex> lock(bmusb_mutex);
 159                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready; });
 160         }
 161         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
 162         GLuint pbo = userdata->pbo;
 163         check_error();
 164         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 165         check_error();
 166         glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
 167         check_error();
 168         //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 169         //check_error();
 170
 171         // Upload the textures.
 172         glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
 173         check_error();
 174         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280, 720, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((1280 * 750 * 2 + 44) / 2 + 1280 * 25 + 22));
 175         check_error();
 176         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
 177         check_error();
 178         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280/2, 720, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(1280 * 25 + 22));
 179         check_error();
 180         glBindTexture(GL_TEXTURE_2D, 0);
 181         check_error();
 182         GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 183         check_error();
 184         assert(fence != nullptr);
 185         {
 186                 std::unique_lock<std::mutex> lock(bmusb_mutex);
 187                 card->new_data_ready = true;
 188                 card->new_frame = RefCountedFrame(video_frame);
 189                 card->new_data_ready_fence = fence;
 190                 card->new_data_ready_changed.notify_all();
 191         }
 192
 193         // Video frame will be released when last user of card->new_frame goes out of scope.
 194         card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
 195 }
 196
 197 void Mixer::thread_func()
 198 {
 199         eglBindAPI(EGL_OPENGL_API);
 200         QOpenGLContext *context = create_context();
 201         if (!make_current(context, mixer_surface)) {
 202                 printf("oops\n");
 203                 exit(1);
 204         }
 205
 206         struct timespec start, now;
 207         clock_gettime(CLOCK_MONOTONIC, &start);
 208
 209         while (!should_quit) {
 210                 ++frame;
 211
 212                 CaptureCard card_copy[NUM_CARDS];
 213
 214                 {
 215                         std::unique_lock<std::mutex> lock(bmusb_mutex);
 216
 217                         // The first card is the master timer, so wait for it to have a new frame.
 218                         // TODO: Make configurable, and with a timeout.
 219                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 220
 221                         for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
 222                                 CaptureCard *card = &cards[card_index];
 223                                 card_copy[card_index].usb = card->usb;
 224                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 225                                 card_copy[card_index].new_frame = card->new_frame;
 226                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 227                                 card->new_data_ready = false;
 228                                 card->new_data_ready_changed.notify_all();
 229                         }
 230                 }
 231
 232                 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
 233                         CaptureCard *card = &card_copy[card_index];
 234                         if (!card->new_data_ready)
 235                                 continue;
 236
 237                         assert(card->new_frame != nullptr);
 238                         bmusb_current_rendering_frame[card_index] = card->new_frame;
 239                         check_error();
 240
 241                         // The new texture might still be uploaded,
 242                         // tell the GPU to wait until it's there.
 243                         if (card->new_data_ready_fence)
 244                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 245                         check_error();
 246                         glDeleteSync(card->new_data_ready_fence);
 247                         check_error();
 248                         const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
 249                         theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
 250                 }
 251
 252                 // Get the main chain from the theme, and set its state immediately.
 253                 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, frame / 60.0f, WIDTH, HEIGHT);
 254                 EffectChain *chain = theme_main_chain.first;
 255                 theme_main_chain.second();
 256
 257                 GLuint y_tex, cbcr_tex;
 258                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 259                 assert(got_frame);
 260
 261                 // Render main chain.
 262                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 263                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 264                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 265                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 266                 resource_pool->release_fbo(fbo);
 267
 268                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 269                 resource_pool->release_2d_texture(cbcr_full_tex);
 270
 271                 // Set the right state for rgba_tex.
 272                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 273                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 274                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 275                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 276                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 277
 278                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 279                 check_error();
 280
 281                 // Make sure the H.264 gets a reference to all the
 282                 // input frames needed, so that they are not released back
 283                 // until the rendering is done.
 284                 vector<RefCountedFrame> input_frames;
 285                 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
 286                         input_frames.push_back(bmusb_current_rendering_frame[card_index]);
 287                 }
 288                 h264_encoder->end_frame(fence, input_frames);
 289
 290                 // The live frame just shows the RGBA texture we just rendered.
 291                 // It owns rgba_tex now.
 292                 DisplayFrame live_frame;
 293                 live_frame.chain = display_chain.get();
 294                 live_frame.setup_chain = [this, rgba_tex]{
 295                         display_input->set_texture_num(rgba_tex);
 296                 };
 297                 live_frame.ready_fence = fence;
 298                 live_frame.input_frames = {};
 299                 live_frame.temp_textures = { rgba_tex };
 300                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 301
 302                 // Set up preview and any additional channels.
 303                 for (unsigned i = 1; i < theme->get_num_channels() + 2; ++i) {
 304                         DisplayFrame display_frame;
 305                         pair<EffectChain *, function<void()>> chain = theme->get_chain(i, frame / 60.0f, WIDTH, HEIGHT);  // FIXME: dimensions
 306                         display_frame.chain = chain.first;
 307                         display_frame.setup_chain = chain.second;
 308                         display_frame.ready_fence = fence;
 309                         display_frame.input_frames = { bmusb_current_rendering_frame[0], bmusb_current_rendering_frame[1] };  // FIXME: possible to do better?
 310                         display_frame.temp_textures = {};
 311                         output_channel[i].output_frame(display_frame);
 312                 }
 313
 314                 clock_gettime(CLOCK_MONOTONIC, &now);
 315                 double elapsed = now.tv_sec - start.tv_sec +
 316                         1e-9 * (now.tv_nsec - start.tv_nsec);
 317                 if (frame % 100 == 0) {
 318                         printf("%d frames in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 319                                 frame, elapsed, frame / elapsed,
 320                                 1e3 * elapsed / frame);
 321                 //      chain->print_phase_timing();
 322                 }
 323
 324                 // Reset every 100 frames, so that local variations in frame times
 325                 // (especially for the first few frames, when the shaders are
 326                 // compiled etc.) don't make it hard to measure for the entire
 327                 // remaining duration of the program.
 328                 if (frame == 10000) {
 329                         frame = 0;
 330                         start = now;
 331                 }
 332                 check_error();
 333         }
 334
 335         resource_pool->clean_context();
 336 }
 337
 338 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 339 {
 340         GLuint vao;
 341         glGenVertexArrays(1, &vao);
 342         check_error();
 343
 344         float vertices[] = {
 345                 0.0f, 2.0f,
 346                 0.0f, 0.0f,
 347                 2.0f, 0.0f
 348         };
 349
 350         glBindVertexArray(vao);
 351         check_error();
 352
 353         // Extract Cb/Cr.
 354         GLuint fbo = resource_pool->create_fbo(dst_tex);
 355         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 356         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 357         check_error();
 358
 359         glUseProgram(cbcr_program_num);
 360         check_error();
 361
 362         glActiveTexture(GL_TEXTURE0);
 363         check_error();
 364         glBindTexture(GL_TEXTURE_2D, src_tex);
 365         check_error();
 366         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 367         check_error();
 368         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 369         check_error();
 370         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 371         check_error();
 372
 373         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 374         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 375
 376         GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
 377         GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
 378
 379         glDrawArrays(GL_TRIANGLES, 0, 3);
 380         check_error();
 381
 382         cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
 383         cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
 384
 385         glUseProgram(0);
 386         check_error();
 387
 388         resource_pool->release_fbo(fbo);
 389         glDeleteVertexArrays(1, &vao);
 390 }
 391
 392 void Mixer::release_display_frame(DisplayFrame *frame)
 393 {
 394         for (GLuint texnum : frame->temp_textures) {
 395                 resource_pool->release_2d_texture(texnum);
 396         }
 397         frame->temp_textures.clear();
 398         frame->ready_fence.reset();
 399         frame->input_frames.clear();
 400 }
 401
 402 void Mixer::start()
 403 {
 404         mixer_thread = std::thread(&Mixer::thread_func, this);
 405 }
 406
 407 void Mixer::quit()
 408 {
 409         should_quit = true;
 410         mixer_thread.join();
 411 }
 412
 413 void Mixer::transition_clicked(int transition_num)
 414 {
 415         theme->transition_clicked(transition_num, frame / 60.0);
 416 }
 417
 418 void Mixer::channel_clicked(int preview_num)
 419 {
 420         theme->channel_clicked(preview_num);
 421 }
 422
 423 Mixer::OutputChannel::~OutputChannel()
 424 {
 425         if (has_current_frame) {
 426                 parent->release_display_frame(&current_frame);
 427         }
 428         if (has_ready_frame) {
 429                 parent->release_display_frame(&ready_frame);
 430         }
 431 }
 432
 433 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 434 {
 435         // Store this frame for display. Remove the ready frame if any
 436         // (it was seemingly never used).
 437         {
 438                 std::unique_lock<std::mutex> lock(frame_mutex);
 439                 if (has_ready_frame) {
 440                         parent->release_display_frame(&ready_frame);
 441                 }
 442                 ready_frame = frame;
 443                 has_ready_frame = true;
 444         }
 445
 446         if (has_new_frame_ready_callback) {
 447                 new_frame_ready_callback();
 448         }
 449 }
 450
 451 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 452 {
 453         std::unique_lock<std::mutex> lock(frame_mutex);
 454         if (!has_current_frame && !has_ready_frame) {
 455                 return false;
 456         }
 457
 458         if (has_current_frame && has_ready_frame) {
 459                 // We have a new ready frame. Toss the current one.
 460                 parent->release_display_frame(&current_frame);
 461                 has_current_frame = false;
 462         }
 463         if (has_ready_frame) {
 464                 assert(!has_current_frame);
 465                 current_frame = ready_frame;
 466                 ready_frame.ready_fence.reset();  // Drop the refcount.
 467                 ready_frame.input_frames.clear();  // Drop the refcounts.
 468                 has_current_frame = true;
 469                 has_ready_frame = false;
 470         }
 471
 472         *frame = current_frame;
 473         return true;
 474 }
 475
 476 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
 477 {
 478         new_frame_ready_callback = callback;
 479         has_new_frame_ready_callback = true;
 480 }