git.sesse.net Git - nageru/blob - futatabi/jpeg_frame_view.cpp

   1 #include "jpeg_frame_view.h"
   2
   3 #include "defs.h"
   4 #include "jpeg_destroyer.h"
   5 #include "jpeglib_error_wrapper.h"
   6 #include "shared/post_to_main_thread.h"
   7 #include "video_stream.h"
   8 #include "ycbcr_converter.h"
   9
  10 #include <QMouseEvent>
  11 #include <QScreen>
  12 #include <atomic>
  13 #include <condition_variable>
  14 #include <deque>
  15 #include <jpeglib.h>
  16 #include <movit/init.h>
  17 #include <movit/resource_pool.h>
  18 #include <movit/util.h>
  19 #include <mutex>
  20 #include <stdint.h>
  21 #include <thread>
  22 #include <unistd.h>
  23 #include <utility>
  24
  25 // Must come after the Qt stuff.
  26 #include "vaapi_jpeg_decoder.h"
  27
  28 using namespace movit;
  29 using namespace std;
  30
  31 namespace {
  32
  33 // Just an arbitrary order for std::map.
  34 struct FrameOnDiskLexicalOrder
  35 {
  36         bool operator() (const FrameOnDisk &a, const FrameOnDisk &b) const
  37         {
  38                 if (a.pts != b.pts)
  39                         return a.pts < b.pts;
  40                 if (a.offset != b.offset)
  41                         return a.offset < b.offset;
  42                 if (a.filename_idx != b.filename_idx)
  43                         return a.filename_idx < b.filename_idx;
  44                 assert(a.size == b.size);
  45                 return false;
  46         }
  47 };
  48
  49 inline size_t frame_size(const Frame &frame)
  50 {
  51         size_t y_size = frame.width * frame.height;
  52         size_t cbcr_size = y_size / frame.chroma_subsampling_x / frame.chroma_subsampling_y;
  53         return y_size + cbcr_size * 2;
  54 }
  55
  56 struct LRUFrame {
  57         shared_ptr<Frame> frame;
  58         size_t last_used;
  59 };
  60
  61 struct PendingDecode {
  62         JPEGFrameView *destination;
  63
  64         // For actual decodes (only if frame below is nullptr).
  65         FrameOnDisk primary, secondary;
  66         float fade_alpha;  // Irrelevant if secondary.stream_idx == -1.
  67
  68         // Already-decoded frames are also sent through PendingDecode,
  69         // so that they get drawn in the right order. If frame is nullptr,
  70         // it's a real decode.
  71         shared_ptr<Frame> frame;
  72 };
  73
  74 }  // namespace
  75
  76 thread JPEGFrameView::jpeg_decoder_thread;
  77 mutex cache_mu;
  78 map<FrameOnDisk, LRUFrame, FrameOnDiskLexicalOrder> cache;  // Under cache_mu.
  79 size_t cache_bytes_used = 0;  // Under cache_mu.
  80 condition_variable any_pending_decodes;
  81 deque<PendingDecode> pending_decodes;  // Under cache_mu.
  82 atomic<size_t> event_counter{0};
  83 extern QGLWidget *global_share_widget;
  84 extern atomic<bool> should_quit;
  85
  86 shared_ptr<Frame> decode_jpeg(const string &jpeg)
  87 {
  88         shared_ptr<Frame> frame;
  89         if (vaapi_jpeg_decoding_usable) {
  90                 frame = decode_jpeg_vaapi(jpeg);
  91                 if (frame != nullptr) {
  92                         return frame;
  93                 }
  94                 fprintf(stderr, "VA-API hardware decoding failed; falling back to software.\n");
  95         }
  96
  97         frame.reset(new Frame);
  98
  99         jpeg_decompress_struct dinfo;
 100         JPEGWrapErrorManager error_mgr(&dinfo);
 101         if (!error_mgr.run([&dinfo]{ jpeg_create_decompress(&dinfo); })) {
 102                 return get_black_frame();
 103         }
 104         JPEGDestroyer destroy_dinfo(&dinfo);
 105
 106         if (!error_mgr.run([&dinfo, &jpeg]{
 107                 jpeg_mem_src(&dinfo, reinterpret_cast<const unsigned char *>(jpeg.data()), jpeg.size());
 108                 jpeg_read_header(&dinfo, true);
 109         })) {
 110                 return get_black_frame();
 111         }
 112
 113         if (dinfo.num_components != 3) {
 114                 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 115                         dinfo.num_components,
 116                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 117                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 118                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 119                 return get_black_frame();
 120         }
 121         if (dinfo.comp_info[0].h_samp_factor != dinfo.max_h_samp_factor ||
 122             dinfo.comp_info[0].v_samp_factor != dinfo.max_v_samp_factor ||  // Y' must not be subsampled.
 123             dinfo.comp_info[1].h_samp_factor != dinfo.comp_info[2].h_samp_factor ||
 124             dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[2].v_samp_factor ||  // Cb and Cr must be identically subsampled.
 125             (dinfo.max_h_samp_factor % dinfo.comp_info[1].h_samp_factor) != 0 ||
 126             (dinfo.max_v_samp_factor % dinfo.comp_info[1].v_samp_factor) != 0) {  // No 2:3 subsampling or other weirdness.
 127                 fprintf(stderr, "Unsupported subsampling scheme. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 128                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 129                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 130                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 131                 exit(1);
 132         }
 133         dinfo.raw_data_out = true;
 134
 135         if (!error_mgr.run([&dinfo]{
 136                 jpeg_start_decompress(&dinfo);
 137         })) {
 138                 return get_black_frame();
 139         }
 140
 141         frame->width = dinfo.output_width;
 142         frame->height = dinfo.output_height;
 143         frame->chroma_subsampling_x = dinfo.max_h_samp_factor / dinfo.comp_info[1].h_samp_factor;
 144         frame->chroma_subsampling_y = dinfo.max_v_samp_factor / dinfo.comp_info[1].v_samp_factor;
 145
 146         unsigned h_mcu_size = DCTSIZE * dinfo.max_h_samp_factor;
 147         unsigned v_mcu_size = DCTSIZE * dinfo.max_v_samp_factor;
 148         unsigned mcu_width_blocks = (dinfo.output_width + h_mcu_size - 1) / h_mcu_size;
 149         unsigned mcu_height_blocks = (dinfo.output_height + v_mcu_size - 1) / v_mcu_size;
 150
 151         unsigned luma_width_blocks = mcu_width_blocks * dinfo.comp_info[0].h_samp_factor;
 152         unsigned chroma_width_blocks = mcu_width_blocks * dinfo.comp_info[1].h_samp_factor;
 153         unsigned luma_height_blocks = mcu_height_blocks * dinfo.comp_info[0].v_samp_factor;
 154         unsigned chroma_height_blocks = mcu_height_blocks * dinfo.comp_info[1].v_samp_factor;
 155
 156         // TODO: Decode into a PBO.
 157         frame->y.reset(new uint8_t[luma_width_blocks * luma_height_blocks * DCTSIZE2]);
 158         frame->cb.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 159         frame->cr.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 160         frame->pitch_y = luma_width_blocks * DCTSIZE;
 161         frame->pitch_chroma = chroma_width_blocks * DCTSIZE;
 162
 163         if (!error_mgr.run([&dinfo, &frame, v_mcu_size, mcu_height_blocks] {
 164                 JSAMPROW yptr[v_mcu_size], cbptr[v_mcu_size], crptr[v_mcu_size];
 165                 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 166                 for (unsigned y = 0; y < mcu_height_blocks; ++y) {
 167                         // NOTE: The last elements of cbptr/crptr will be unused for vertically subsampled chroma.
 168                         for (unsigned yy = 0; yy < v_mcu_size; ++yy) {
 169                                 yptr[yy] = frame->y.get() + (y * DCTSIZE * dinfo.max_v_samp_factor + yy) * frame->pitch_y;
 170                                 cbptr[yy] = frame->cb.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 171                                 crptr[yy] = frame->cr.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 172                         }
 173
 174                         jpeg_read_raw_data(&dinfo, data, v_mcu_size);
 175                 }
 176
 177                 (void)jpeg_finish_decompress(&dinfo);
 178         })) {
 179                 return get_black_frame();
 180         }
 181
 182         return frame;
 183 }
 184
 185 void prune_cache()
 186 {
 187         // Assumes cache_mu is held.
 188         int64_t bytes_still_to_remove = cache_bytes_used - (size_t(CACHE_SIZE_MB) * 1024 * 1024) * 9 / 10;
 189         if (bytes_still_to_remove <= 0) return;
 190
 191         vector<pair<size_t, size_t>> lru_timestamps_and_size;
 192         for (const auto &key_and_value : cache) {
 193                 lru_timestamps_and_size.emplace_back(
 194                         key_and_value.second.last_used,
 195                         frame_size(*key_and_value.second.frame));
 196         }
 197         sort(lru_timestamps_and_size.begin(), lru_timestamps_and_size.end());
 198
 199         // Remove the oldest ones until we are below 90% of the cache used.
 200         size_t lru_cutoff_point = 0;
 201         for (const pair<size_t, size_t> &it : lru_timestamps_and_size) {
 202                 lru_cutoff_point = it.first;
 203                 bytes_still_to_remove -= it.second;
 204                 if (bytes_still_to_remove <= 0) break;
 205         }
 206
 207         for (auto it = cache.begin(); it != cache.end(); ) {
 208                 if (it->second.last_used <= lru_cutoff_point) {
 209                         cache_bytes_used -= frame_size(*it->second.frame);
 210                         it = cache.erase(it);
 211                 } else {
 212                         ++it;
 213                 }
 214         }
 215 }
 216
 217 shared_ptr<Frame> decode_jpeg_with_cache(FrameOnDisk frame_spec, CacheMissBehavior cache_miss_behavior, FrameReader *frame_reader, bool *did_decode)
 218 {
 219         *did_decode = false;
 220         {
 221                 unique_lock<mutex> lock(cache_mu);
 222                 auto it = cache.find(frame_spec);
 223                 if (it != cache.end()) {
 224                         it->second.last_used = event_counter++;
 225                         return it->second.frame;
 226                 }
 227         }
 228
 229         if (cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE) {
 230                 return nullptr;
 231         }
 232
 233         *did_decode = true;
 234         shared_ptr<Frame> frame = decode_jpeg(frame_reader->read_frame(frame_spec));
 235
 236         unique_lock<mutex> lock(cache_mu);
 237         cache_bytes_used += frame_size(*frame);
 238         cache[frame_spec] = LRUFrame{ frame, event_counter++ };
 239
 240         if (cache_bytes_used > size_t(CACHE_SIZE_MB) * 1024 * 1024) {
 241                 prune_cache();
 242         }
 243         return frame;
 244 }
 245
 246 void JPEGFrameView::jpeg_decoder_thread_func()
 247 {
 248         size_t num_decoded = 0, num_dropped = 0;
 249
 250         pthread_setname_np(pthread_self(), "JPEGDecoder");
 251         while (!should_quit.load()) {
 252                 PendingDecode decode;
 253                 CacheMissBehavior cache_miss_behavior = DECODE_IF_NOT_IN_CACHE;
 254                 {
 255                         unique_lock<mutex> lock(cache_mu);  // TODO: Perhaps under another lock?
 256                         any_pending_decodes.wait(lock, [] {
 257                                 return !pending_decodes.empty() || should_quit.load();
 258                         });
 259                         if (should_quit.load())
 260                                 break;
 261                         decode = pending_decodes.front();
 262                         pending_decodes.pop_front();
 263
 264                         size_t num_pending = 0;
 265                         for (const PendingDecode &other_decode : pending_decodes) {
 266                                 if (other_decode.destination == decode.destination) {
 267                                         ++num_pending;
 268                                 }
 269                         }
 270                         if (num_pending > 3) {
 271                                 cache_miss_behavior = RETURN_NULLPTR_IF_NOT_IN_CACHE;
 272                         }
 273                 }
 274
 275                 if (decode.frame != nullptr) {
 276                         // Already decoded, so just show it.
 277                         decode.destination->setDecodedFrame(decode.frame, nullptr, 1.0f);
 278                         continue;
 279                 }
 280
 281                 shared_ptr<Frame> primary_frame, secondary_frame;
 282                 bool drop = false;
 283                 for (int subframe_idx = 0; subframe_idx < 2; ++subframe_idx) {
 284                         const FrameOnDisk &frame_spec = (subframe_idx == 0 ? decode.primary : decode.secondary);
 285                         if (frame_spec.pts == -1) {
 286                                 // No secondary frame.
 287                                 continue;
 288                         }
 289
 290                         bool found_in_cache;
 291                         shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, cache_miss_behavior, &decode.destination->frame_reader, &found_in_cache);
 292
 293                         if (frame == nullptr) {
 294                                 assert(cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
 295                                 drop = true;
 296                                 break;
 297                         }
 298
 299                         if (!found_in_cache) {
 300                                 ++num_decoded;
 301                                 if (num_decoded % 1000 == 0) {
 302                                         fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
 303                                                 num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
 304                                 }
 305                         }
 306                         if (subframe_idx == 0) {
 307                                 primary_frame = std::move(frame);
 308                         } else {
 309                                 secondary_frame = std::move(frame);
 310                         }
 311                 }
 312                 if (drop) {
 313                         ++num_dropped;
 314                         continue;
 315                 }
 316
 317                 // TODO: Could we get jitter between non-interpolated and interpolated frames here?
 318                 decode.destination->setDecodedFrame(primary_frame, secondary_frame, decode.fade_alpha);
 319         }
 320 }
 321
 322 void JPEGFrameView::shutdown()
 323 {
 324         any_pending_decodes.notify_all();
 325         jpeg_decoder_thread.join();
 326 }
 327
 328 JPEGFrameView::JPEGFrameView(QWidget *parent)
 329         : QGLWidget(parent, global_share_widget)
 330 {
 331 }
 332
 333 void JPEGFrameView::setFrame(unsigned stream_idx, FrameOnDisk frame, FrameOnDisk secondary_frame, float fade_alpha)
 334 {
 335         current_stream_idx = stream_idx;  // TODO: Does this interact with fades?
 336
 337         unique_lock<mutex> lock(cache_mu);
 338         PendingDecode decode;
 339         decode.primary = frame;
 340         decode.secondary = secondary_frame;
 341         decode.fade_alpha = fade_alpha;
 342         decode.destination = this;
 343         pending_decodes.push_back(decode);
 344         any_pending_decodes.notify_all();
 345 }
 346
 347 void JPEGFrameView::setFrame(shared_ptr<Frame> frame)
 348 {
 349         unique_lock<mutex> lock(cache_mu);
 350         PendingDecode decode;
 351         decode.frame = std::move(frame);
 352         decode.destination = this;
 353         pending_decodes.push_back(decode);
 354         any_pending_decodes.notify_all();
 355 }
 356
 357 ResourcePool *resource_pool = nullptr;
 358
 359 void JPEGFrameView::initializeGL()
 360 {
 361         glDisable(GL_BLEND);
 362         glDisable(GL_DEPTH_TEST);
 363         check_error();
 364
 365         static once_flag once;
 366         call_once(once, [] {
 367                 resource_pool = new ResourcePool;
 368                 jpeg_decoder_thread = std::thread(jpeg_decoder_thread_func);
 369         });
 370
 371         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_RGBA, resource_pool));
 372
 373         ImageFormat inout_format;
 374         inout_format.color_space = COLORSPACE_sRGB;
 375         inout_format.gamma_curve = GAMMA_sRGB;
 376
 377         overlay_chain.reset(new EffectChain(overlay_base_width, overlay_base_height, resource_pool));
 378         overlay_input = (movit::FlatInput *)overlay_chain->add_input(new FlatInput(inout_format, FORMAT_GRAYSCALE, GL_UNSIGNED_BYTE, overlay_base_width, overlay_base_height));
 379
 380         overlay_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 381         overlay_chain->finalize();
 382 }
 383
 384 void JPEGFrameView::resizeGL(int width, int height)
 385 {
 386         check_error();
 387         glViewport(0, 0, width, height);
 388         check_error();
 389
 390         // Save these, as width() and height() will lie with DPI scaling.
 391         gl_width = width;
 392         gl_height = height;
 393 }
 394
 395 void JPEGFrameView::paintGL()
 396 {
 397         glViewport(0, 0, gl_width, gl_height);
 398         if (current_frame == nullptr) {
 399                 glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
 400                 glClear(GL_COLOR_BUFFER_BIT);
 401                 return;
 402         }
 403
 404         check_error();
 405         current_chain->render_to_screen();
 406
 407         if (overlay_image != nullptr) {
 408                 if (overlay_input_needs_refresh) {
 409                         overlay_input->set_width(overlay_width);
 410                         overlay_input->set_height(overlay_height);
 411                         overlay_input->set_pixel_data(overlay_image->bits());
 412                 }
 413                 glViewport(gl_width - overlay_width, 0, overlay_width, overlay_height);
 414                 overlay_chain->render_to_screen();
 415         }
 416 }
 417
 418 namespace {
 419
 420 }  // namespace
 421
 422 void JPEGFrameView::setDecodedFrame(shared_ptr<Frame> frame, shared_ptr<Frame> secondary_frame, float fade_alpha)
 423 {
 424         post_to_main_thread([this, frame, secondary_frame, fade_alpha] {
 425                 current_frame = frame;
 426                 current_secondary_frame = secondary_frame;
 427
 428                 if (secondary_frame != nullptr) {
 429                         current_chain = ycbcr_converter->prepare_chain_for_fade(frame, secondary_frame, fade_alpha);
 430                 } else {
 431                         current_chain = ycbcr_converter->prepare_chain_for_conversion(frame);
 432                 }
 433                 update();
 434         });
 435 }
 436
 437 void JPEGFrameView::mousePressEvent(QMouseEvent *event)
 438 {
 439         if (event->type() == QEvent::MouseButtonPress && event->button() == Qt::LeftButton) {
 440                 emit clicked();
 441         }
 442 }
 443
 444 void JPEGFrameView::set_overlay(const string &text)
 445 {
 446         if (text.empty()) {
 447                 overlay_image.reset();
 448                 return;
 449         }
 450
 451         float dpr = QGuiApplication::primaryScreen()->devicePixelRatio();
 452         overlay_width = lrint(overlay_base_width * dpr);
 453         overlay_height = lrint(overlay_base_height * dpr);
 454
 455         overlay_image.reset(new QImage(overlay_width, overlay_height, QImage::Format_Grayscale8));
 456         overlay_image->setDevicePixelRatio(dpr);
 457         overlay_image->fill(0);
 458         QPainter painter(overlay_image.get());
 459
 460         painter.setPen(Qt::white);
 461         QFont font = painter.font();
 462         font.setPointSize(12);
 463         painter.setFont(font);
 464
 465         painter.drawText(QRectF(0, 0, overlay_base_width, overlay_base_height), Qt::AlignCenter, QString::fromStdString(text));
 466
 467         // Don't refresh immediately; we might not have an OpenGL context here.
 468         overlay_input_needs_refresh = true;
 469 }
 470
 471 shared_ptr<Frame> get_black_frame()
 472 {
 473         static shared_ptr<Frame> black_frame;
 474         static once_flag flag;
 475         call_once(flag, [] {
 476                 black_frame.reset(new Frame);
 477                 black_frame->y.reset(new uint8_t[1280 * 720]);
 478                 black_frame->cb.reset(new uint8_t[(1280 / 2) * (720 / 2)]);
 479                 black_frame->cr.reset(new uint8_t[(1280 / 2) * (720 / 2)]);
 480                 black_frame->width = 1280;
 481                 black_frame->height = 720;
 482                 black_frame->chroma_subsampling_x = 2;
 483                 black_frame->chroma_subsampling_y = 2;
 484                 black_frame->pitch_y = 1280;
 485                 black_frame->pitch_chroma = 1280 / 2;
 486         });
 487         return black_frame;
 488 }