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

   1 #include "jpeg_frame_view.h"
   2
   3 #include "defs.h"
   4 #include "flags.h"
   5 #include "jpeg_destroyer.h"
   6 #include "jpeglib_error_wrapper.h"
   7 #include "shared/post_to_main_thread.h"
   8 #include "video_stream.h"
   9 #include "ycbcr_converter.h"
  10
  11 #include <QMouseEvent>
  12 #include <QScreen>
  13 #include <atomic>
  14 #include <condition_variable>
  15 #include <deque>
  16 #include <jpeglib.h>
  17 #include <movit/init.h>
  18 #include <movit/resource_pool.h>
  19 #include <movit/util.h>
  20 #include <mutex>
  21 #include <stdint.h>
  22 #include <thread>
  23 #include <unistd.h>
  24 #include <utility>
  25
  26 // Must come after the Qt stuff.
  27 #include "vaapi_jpeg_decoder.h"
  28
  29 using namespace movit;
  30 using namespace std;
  31
  32 namespace {
  33
  34 // Just an arbitrary order for std::map.
  35 struct FrameOnDiskLexicalOrder
  36 {
  37         bool operator() (const FrameOnDisk &a, const FrameOnDisk &b) const
  38         {
  39                 if (a.pts != b.pts)
  40                         return a.pts < b.pts;
  41                 if (a.offset != b.offset)
  42                         return a.offset < b.offset;
  43                 if (a.filename_idx != b.filename_idx)
  44                         return a.filename_idx < b.filename_idx;
  45                 assert(a.size == b.size);
  46                 return false;
  47         }
  48 };
  49
  50 inline size_t frame_size(const Frame &frame)
  51 {
  52         size_t y_size = frame.width * frame.height;
  53         size_t cbcr_size = y_size / frame.chroma_subsampling_x / frame.chroma_subsampling_y;
  54         return y_size + cbcr_size * 2;
  55 }
  56
  57 struct LRUFrame {
  58         shared_ptr<Frame> frame;
  59         size_t last_used;
  60 };
  61
  62 struct PendingDecode {
  63         JPEGFrameView *destination;
  64
  65         // For actual decodes (only if frame below is nullptr).
  66         FrameOnDisk primary, secondary;
  67         float fade_alpha;  // Irrelevant if secondary.stream_idx == -1.
  68
  69         // Already-decoded frames are also sent through PendingDecode,
  70         // so that they get drawn in the right order. If frame is nullptr,
  71         // it's a real decode.
  72         shared_ptr<Frame> frame;
  73 };
  74
  75 }  // namespace
  76
  77 thread JPEGFrameView::jpeg_decoder_thread;
  78 mutex cache_mu;
  79 map<FrameOnDisk, LRUFrame, FrameOnDiskLexicalOrder> cache;  // Under cache_mu.
  80 size_t cache_bytes_used = 0;  // Under cache_mu.
  81 condition_variable any_pending_decodes;
  82 deque<PendingDecode> pending_decodes;  // Under cache_mu.
  83 atomic<size_t> event_counter{0};
  84 extern QGLWidget *global_share_widget;
  85 extern atomic<bool> should_quit;
  86
  87 shared_ptr<Frame> decode_jpeg(const string &jpeg)
  88 {
  89         shared_ptr<Frame> frame;
  90         if (vaapi_jpeg_decoding_usable) {
  91                 frame = decode_jpeg_vaapi(jpeg);
  92                 if (frame != nullptr) {
  93                         return frame;
  94                 }
  95                 fprintf(stderr, "VA-API hardware decoding failed; falling back to software.\n");
  96         }
  97
  98         frame.reset(new Frame);
  99
 100         jpeg_decompress_struct dinfo;
 101         JPEGWrapErrorManager error_mgr(&dinfo);
 102         if (!error_mgr.run([&dinfo]{ jpeg_create_decompress(&dinfo); })) {
 103                 return get_black_frame();
 104         }
 105         JPEGDestroyer destroy_dinfo(&dinfo);
 106
 107         if (!error_mgr.run([&dinfo, &jpeg]{
 108                 jpeg_mem_src(&dinfo, reinterpret_cast<const unsigned char *>(jpeg.data()), jpeg.size());
 109                 jpeg_read_header(&dinfo, true);
 110         })) {
 111                 return get_black_frame();
 112         }
 113
 114         if (dinfo.num_components != 3) {
 115                 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 116                         dinfo.num_components,
 117                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 118                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 119                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 120                 return get_black_frame();
 121         }
 122         if (dinfo.comp_info[0].h_samp_factor != dinfo.max_h_samp_factor ||
 123             dinfo.comp_info[0].v_samp_factor != dinfo.max_v_samp_factor ||  // Y' must not be subsampled.
 124             dinfo.comp_info[1].h_samp_factor != dinfo.comp_info[2].h_samp_factor ||
 125             dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[2].v_samp_factor ||  // Cb and Cr must be identically subsampled.
 126             (dinfo.max_h_samp_factor % dinfo.comp_info[1].h_samp_factor) != 0 ||
 127             (dinfo.max_v_samp_factor % dinfo.comp_info[1].v_samp_factor) != 0) {  // No 2:3 subsampling or other weirdness.
 128                 fprintf(stderr, "Unsupported subsampling scheme. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 129                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 130                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 131                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 132                 exit(1);
 133         }
 134         dinfo.raw_data_out = true;
 135
 136         if (!error_mgr.run([&dinfo]{
 137                 jpeg_start_decompress(&dinfo);
 138         })) {
 139                 return get_black_frame();
 140         }
 141
 142         frame->width = dinfo.output_width;
 143         frame->height = dinfo.output_height;
 144         frame->chroma_subsampling_x = dinfo.max_h_samp_factor / dinfo.comp_info[1].h_samp_factor;
 145         frame->chroma_subsampling_y = dinfo.max_v_samp_factor / dinfo.comp_info[1].v_samp_factor;
 146
 147         unsigned h_mcu_size = DCTSIZE * dinfo.max_h_samp_factor;
 148         unsigned v_mcu_size = DCTSIZE * dinfo.max_v_samp_factor;
 149         unsigned mcu_width_blocks = (dinfo.output_width + h_mcu_size - 1) / h_mcu_size;
 150         unsigned mcu_height_blocks = (dinfo.output_height + v_mcu_size - 1) / v_mcu_size;
 151
 152         unsigned luma_width_blocks = mcu_width_blocks * dinfo.comp_info[0].h_samp_factor;
 153         unsigned chroma_width_blocks = mcu_width_blocks * dinfo.comp_info[1].h_samp_factor;
 154         unsigned luma_height_blocks = mcu_height_blocks * dinfo.comp_info[0].v_samp_factor;
 155         unsigned chroma_height_blocks = mcu_height_blocks * dinfo.comp_info[1].v_samp_factor;
 156
 157         // TODO: Decode into a PBO.
 158         frame->y.reset(new uint8_t[luma_width_blocks * luma_height_blocks * DCTSIZE2]);
 159         frame->cb.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 160         frame->cr.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 161         frame->pitch_y = luma_width_blocks * DCTSIZE;
 162         frame->pitch_chroma = chroma_width_blocks * DCTSIZE;
 163
 164         if (!error_mgr.run([&dinfo, &frame, v_mcu_size, mcu_height_blocks] {
 165                 JSAMPROW yptr[v_mcu_size], cbptr[v_mcu_size], crptr[v_mcu_size];
 166                 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 167                 for (unsigned y = 0; y < mcu_height_blocks; ++y) {
 168                         // NOTE: The last elements of cbptr/crptr will be unused for vertically subsampled chroma.
 169                         for (unsigned yy = 0; yy < v_mcu_size; ++yy) {
 170                                 yptr[yy] = frame->y.get() + (y * DCTSIZE * dinfo.max_v_samp_factor + yy) * frame->pitch_y;
 171                                 cbptr[yy] = frame->cb.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 172                                 crptr[yy] = frame->cr.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 173                         }
 174
 175                         jpeg_read_raw_data(&dinfo, data, v_mcu_size);
 176                 }
 177
 178                 (void)jpeg_finish_decompress(&dinfo);
 179         })) {
 180                 return get_black_frame();
 181         }
 182
 183         return frame;
 184 }
 185
 186 void prune_cache()
 187 {
 188         // Assumes cache_mu is held.
 189         int64_t bytes_still_to_remove = cache_bytes_used - (size_t(CACHE_SIZE_MB) * 1024 * 1024) * 9 / 10;
 190         if (bytes_still_to_remove <= 0) return;
 191
 192         vector<pair<size_t, size_t>> lru_timestamps_and_size;
 193         for (const auto &key_and_value : cache) {
 194                 lru_timestamps_and_size.emplace_back(
 195                         key_and_value.second.last_used,
 196                         frame_size(*key_and_value.second.frame));
 197         }
 198         sort(lru_timestamps_and_size.begin(), lru_timestamps_and_size.end());
 199
 200         // Remove the oldest ones until we are below 90% of the cache used.
 201         size_t lru_cutoff_point = 0;
 202         for (const pair<size_t, size_t> &it : lru_timestamps_and_size) {
 203                 lru_cutoff_point = it.first;
 204                 bytes_still_to_remove -= it.second;
 205                 if (bytes_still_to_remove <= 0) break;
 206         }
 207
 208         for (auto it = cache.begin(); it != cache.end(); ) {
 209                 if (it->second.last_used <= lru_cutoff_point) {
 210                         cache_bytes_used -= frame_size(*it->second.frame);
 211                         it = cache.erase(it);
 212                 } else {
 213                         ++it;
 214                 }
 215         }
 216 }
 217
 218 shared_ptr<Frame> decode_jpeg_with_cache(FrameOnDisk frame_spec, CacheMissBehavior cache_miss_behavior, FrameReader *frame_reader, bool *did_decode)
 219 {
 220         *did_decode = false;
 221         {
 222                 unique_lock<mutex> lock(cache_mu);
 223                 auto it = cache.find(frame_spec);
 224                 if (it != cache.end()) {
 225                         it->second.last_used = event_counter++;
 226                         return it->second.frame;
 227                 }
 228         }
 229
 230         if (cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE) {
 231                 return nullptr;
 232         }
 233
 234         *did_decode = true;
 235         shared_ptr<Frame> frame = decode_jpeg(frame_reader->read_frame(frame_spec));
 236
 237         unique_lock<mutex> lock(cache_mu);
 238         cache_bytes_used += frame_size(*frame);
 239         cache[frame_spec] = LRUFrame{ frame, event_counter++ };
 240
 241         if (cache_bytes_used > size_t(CACHE_SIZE_MB) * 1024 * 1024) {
 242                 prune_cache();
 243         }
 244         return frame;
 245 }
 246
 247 void JPEGFrameView::jpeg_decoder_thread_func()
 248 {
 249         size_t num_decoded = 0, num_dropped = 0;
 250
 251         pthread_setname_np(pthread_self(), "JPEGDecoder");
 252         while (!should_quit.load()) {
 253                 PendingDecode decode;
 254                 CacheMissBehavior cache_miss_behavior = DECODE_IF_NOT_IN_CACHE;
 255                 {
 256                         unique_lock<mutex> lock(cache_mu);  // TODO: Perhaps under another lock?
 257                         any_pending_decodes.wait(lock, [] {
 258                                 return !pending_decodes.empty() || should_quit.load();
 259                         });
 260                         if (should_quit.load())
 261                                 break;
 262                         decode = pending_decodes.front();
 263                         pending_decodes.pop_front();
 264
 265                         size_t num_pending = 0;
 266                         for (const PendingDecode &other_decode : pending_decodes) {
 267                                 if (other_decode.destination == decode.destination) {
 268                                         ++num_pending;
 269                                 }
 270                         }
 271                         if (num_pending > 3) {
 272                                 cache_miss_behavior = RETURN_NULLPTR_IF_NOT_IN_CACHE;
 273                         }
 274                 }
 275
 276                 if (decode.frame != nullptr) {
 277                         // Already decoded, so just show it.
 278                         decode.destination->setDecodedFrame(decode.frame, nullptr, 1.0f);
 279                         continue;
 280                 }
 281
 282                 shared_ptr<Frame> primary_frame, secondary_frame;
 283                 bool drop = false;
 284                 for (int subframe_idx = 0; subframe_idx < 2; ++subframe_idx) {
 285                         const FrameOnDisk &frame_spec = (subframe_idx == 0 ? decode.primary : decode.secondary);
 286                         if (frame_spec.pts == -1) {
 287                                 // No secondary frame.
 288                                 continue;
 289                         }
 290
 291                         bool found_in_cache;
 292                         shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, cache_miss_behavior, &decode.destination->frame_reader, &found_in_cache);
 293
 294                         if (frame == nullptr) {
 295                                 assert(cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
 296                                 drop = true;
 297                                 break;
 298                         }
 299
 300                         if (!found_in_cache) {
 301                                 ++num_decoded;
 302                                 if (num_decoded % 1000 == 0) {
 303                                         fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
 304                                                 num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
 305                                 }
 306                         }
 307                         if (subframe_idx == 0) {
 308                                 primary_frame = std::move(frame);
 309                         } else {
 310                                 secondary_frame = std::move(frame);
 311                         }
 312                 }
 313                 if (drop) {
 314                         ++num_dropped;
 315                         continue;
 316                 }
 317
 318                 // TODO: Could we get jitter between non-interpolated and interpolated frames here?
 319                 decode.destination->setDecodedFrame(primary_frame, secondary_frame, decode.fade_alpha);
 320         }
 321 }
 322
 323 void JPEGFrameView::shutdown()
 324 {
 325         any_pending_decodes.notify_all();
 326         jpeg_decoder_thread.join();
 327 }
 328
 329 JPEGFrameView::JPEGFrameView(QWidget *parent)
 330         : QGLWidget(parent, global_share_widget)
 331 {
 332 }
 333
 334 void JPEGFrameView::setFrame(unsigned stream_idx, FrameOnDisk frame, FrameOnDisk secondary_frame, float fade_alpha)
 335 {
 336         current_stream_idx = stream_idx;  // TODO: Does this interact with fades?
 337
 338         unique_lock<mutex> lock(cache_mu);
 339         PendingDecode decode;
 340         decode.primary = frame;
 341         decode.secondary = secondary_frame;
 342         decode.fade_alpha = fade_alpha;
 343         decode.destination = this;
 344         pending_decodes.push_back(decode);
 345         any_pending_decodes.notify_all();
 346 }
 347
 348 void JPEGFrameView::setFrame(shared_ptr<Frame> frame)
 349 {
 350         unique_lock<mutex> lock(cache_mu);
 351         PendingDecode decode;
 352         decode.frame = std::move(frame);
 353         decode.destination = this;
 354         pending_decodes.push_back(decode);
 355         any_pending_decodes.notify_all();
 356 }
 357
 358 ResourcePool *resource_pool = nullptr;
 359
 360 void JPEGFrameView::initializeGL()
 361 {
 362         glDisable(GL_BLEND);
 363         glDisable(GL_DEPTH_TEST);
 364         check_error();
 365
 366         static once_flag once;
 367         call_once(once, [] {
 368                 resource_pool = new ResourcePool;
 369                 jpeg_decoder_thread = std::thread(jpeg_decoder_thread_func);
 370         });
 371
 372         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_RGBA, resource_pool));
 373
 374         ImageFormat inout_format;
 375         inout_format.color_space = COLORSPACE_sRGB;
 376         inout_format.gamma_curve = GAMMA_sRGB;
 377
 378         overlay_chain.reset(new EffectChain(overlay_base_width, overlay_base_height, resource_pool));
 379         overlay_input = (movit::FlatInput *)overlay_chain->add_input(new FlatInput(inout_format, FORMAT_GRAYSCALE, GL_UNSIGNED_BYTE, overlay_base_width, overlay_base_height));
 380
 381         overlay_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 382         overlay_chain->finalize();
 383 }
 384
 385 void JPEGFrameView::resizeGL(int width, int height)
 386 {
 387         check_error();
 388         glViewport(0, 0, width, height);
 389         check_error();
 390
 391         // Save these, as width() and height() will lie with DPI scaling.
 392         gl_width = width;
 393         gl_height = height;
 394 }
 395
 396 void JPEGFrameView::paintGL()
 397 {
 398         glViewport(0, 0, gl_width, gl_height);
 399         if (current_frame == nullptr) {
 400                 glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
 401                 glClear(GL_COLOR_BUFFER_BIT);
 402                 return;
 403         }
 404
 405         check_error();
 406         current_chain->render_to_screen();
 407
 408         if (overlay_image != nullptr) {
 409                 if (overlay_input_needs_refresh) {
 410                         overlay_input->set_width(overlay_width);
 411                         overlay_input->set_height(overlay_height);
 412                         overlay_input->set_pixel_data(overlay_image->bits());
 413                 }
 414                 glViewport(gl_width - overlay_width, 0, overlay_width, overlay_height);
 415                 overlay_chain->render_to_screen();
 416         }
 417 }
 418
 419 namespace {
 420
 421 }  // namespace
 422
 423 void JPEGFrameView::setDecodedFrame(shared_ptr<Frame> frame, shared_ptr<Frame> secondary_frame, float fade_alpha)
 424 {
 425         post_to_main_thread([this, frame, secondary_frame, fade_alpha] {
 426                 current_frame = frame;
 427                 current_secondary_frame = secondary_frame;
 428
 429                 if (secondary_frame != nullptr) {
 430                         current_chain = ycbcr_converter->prepare_chain_for_fade(frame, secondary_frame, fade_alpha);
 431                 } else {
 432                         current_chain = ycbcr_converter->prepare_chain_for_conversion(frame);
 433                 }
 434                 update();
 435         });
 436 }
 437
 438 void JPEGFrameView::mousePressEvent(QMouseEvent *event)
 439 {
 440         if (event->type() == QEvent::MouseButtonPress && event->button() == Qt::LeftButton) {
 441                 emit clicked();
 442         }
 443 }
 444
 445 void JPEGFrameView::set_overlay(const string &text)
 446 {
 447         if (text.empty()) {
 448                 overlay_image.reset();
 449                 return;
 450         }
 451
 452         float dpr = QGuiApplication::primaryScreen()->devicePixelRatio();
 453         overlay_width = lrint(overlay_base_width * dpr);
 454         overlay_height = lrint(overlay_base_height * dpr);
 455
 456         overlay_image.reset(new QImage(overlay_width, overlay_height, QImage::Format_Grayscale8));
 457         overlay_image->setDevicePixelRatio(dpr);
 458         overlay_image->fill(0);
 459         QPainter painter(overlay_image.get());
 460
 461         painter.setPen(Qt::white);
 462         QFont font = painter.font();
 463         font.setPointSize(12);
 464         painter.setFont(font);
 465
 466         painter.drawText(QRectF(0, 0, overlay_base_width, overlay_base_height), Qt::AlignCenter, QString::fromStdString(text));
 467
 468         // Don't refresh immediately; we might not have an OpenGL context here.
 469         overlay_input_needs_refresh = true;
 470 }
 471
 472 shared_ptr<Frame> get_black_frame()
 473 {
 474         static shared_ptr<Frame> black_frame;
 475         static once_flag flag;
 476         call_once(flag, [] {
 477                 black_frame.reset(new Frame);
 478                 black_frame->y.reset(new uint8_t[global_flags.width * global_flags.height]);
 479                 black_frame->cb.reset(new uint8_t[(global_flags.width / 2) * (global_flags.height / 2)]);
 480                 black_frame->cr.reset(new uint8_t[(global_flags.width / 2) * (global_flags.height / 2)]);
 481                 black_frame->width = global_flags.width;
 482                 black_frame->height = global_flags.height;
 483                 black_frame->chroma_subsampling_x = 2;
 484                 black_frame->chroma_subsampling_y = 2;
 485                 black_frame->pitch_y = global_flags.width;
 486                 black_frame->pitch_chroma = global_flags.width / 2;
 487         });
 488         return black_frame;
 489 }