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/metrics.h"
   8 #include "shared/post_to_main_thread.h"
   9 #include "video_stream.h"
  10 #include "ycbcr_converter.h"
  11
  12 #include <QMouseEvent>
  13 #include <QScreen>
  14 #include <atomic>
  15 #include <condition_variable>
  16 #include <deque>
  17 #include <jpeglib.h>
  18 #include <movit/init.h>
  19 #include <movit/resource_pool.h>
  20 #include <movit/util.h>
  21 #include <mutex>
  22 #include <stdint.h>
  23 #include <thread>
  24 #include <unistd.h>
  25 #include <utility>
  26
  27 // Must come after the Qt stuff.
  28 #include "vaapi_jpeg_decoder.h"
  29
  30 using namespace movit;
  31 using namespace std;
  32
  33 namespace {
  34
  35 // Just an arbitrary order for std::map.
  36 struct FrameOnDiskLexicalOrder {
  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 // There can be multiple JPEGFrameView instances, so make all the metrics static.
  63 once_flag jpeg_metrics_inited;
  64 atomic<int64_t> metric_jpeg_cache_used_bytes{ 0 };  // Same value as cache_bytes_used.
  65 atomic<int64_t> metric_jpeg_cache_limit_bytes{ size_t(CACHE_SIZE_MB) * 1024 * 1024 };
  66 atomic<int64_t> metric_jpeg_cache_given_up_frames{ 0 };
  67 atomic<int64_t> metric_jpeg_cache_hit_frames{ 0 };
  68 atomic<int64_t> metric_jpeg_cache_miss_frames{ 0 };
  69 atomic<int64_t> metric_jpeg_software_decode_frames{ 0 };
  70 atomic<int64_t> metric_jpeg_software_fail_frames{ 0 };
  71 atomic<int64_t> metric_jpeg_vaapi_decode_frames{ 0 };
  72 atomic<int64_t> metric_jpeg_vaapi_fail_frames{ 0 };
  73
  74 }  // namespace
  75
  76 mutex cache_mu;
  77 map<FrameOnDisk, LRUFrame, FrameOnDiskLexicalOrder> cache;  // Under cache_mu.
  78 size_t cache_bytes_used = 0;  // Under cache_mu.
  79 atomic<size_t> event_counter{ 0 };
  80 extern QGLWidget *global_share_widget;
  81 extern atomic<bool> should_quit;
  82
  83 shared_ptr<Frame> decode_jpeg(const string &jpeg)
  84 {
  85         shared_ptr<Frame> frame;
  86         if (vaapi_jpeg_decoding_usable) {
  87                 frame = decode_jpeg_vaapi(jpeg);
  88                 if (frame != nullptr) {
  89                         ++metric_jpeg_vaapi_decode_frames;
  90                         return frame;
  91                 }
  92                 fprintf(stderr, "VA-API hardware decoding failed; falling back to software.\n");
  93                 ++metric_jpeg_vaapi_fail_frames;
  94         }
  95
  96         frame.reset(new Frame);
  97
  98         jpeg_decompress_struct dinfo;
  99         JPEGWrapErrorManager error_mgr(&dinfo);
 100         if (!error_mgr.run([&dinfo] { jpeg_create_decompress(&dinfo); })) {
 101                 return get_black_frame();
 102         }
 103         JPEGDestroyer destroy_dinfo(&dinfo);
 104
 105         if (!error_mgr.run([&dinfo, &jpeg] {
 106                     jpeg_mem_src(&dinfo, reinterpret_cast<const unsigned char *>(jpeg.data()), jpeg.size());
 107                     jpeg_read_header(&dinfo, true);
 108             })) {
 109                 return get_black_frame();
 110         }
 111
 112         if (dinfo.num_components != 3) {
 113                 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 114                         dinfo.num_components,
 115                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 116                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 117                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 118                 return get_black_frame();
 119         }
 120         if (dinfo.comp_info[0].h_samp_factor != dinfo.max_h_samp_factor ||
 121             dinfo.comp_info[0].v_samp_factor != dinfo.max_v_samp_factor ||  // Y' must not be subsampled.
 122             dinfo.comp_info[1].h_samp_factor != dinfo.comp_info[2].h_samp_factor ||
 123             dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[2].v_samp_factor ||  // Cb and Cr must be identically subsampled.
 124             (dinfo.max_h_samp_factor % dinfo.comp_info[1].h_samp_factor) != 0 ||
 125             (dinfo.max_v_samp_factor % dinfo.comp_info[1].v_samp_factor) != 0) {  // No 2:3 subsampling or other weirdness.
 126                 fprintf(stderr, "Unsupported subsampling scheme. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 127                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 128                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 129                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 130                 abort();
 131         }
 132         dinfo.raw_data_out = true;
 133
 134         if (!error_mgr.run([&dinfo] {
 135                     jpeg_start_decompress(&dinfo);
 136             })) {
 137                 return get_black_frame();
 138         }
 139
 140         frame->width = dinfo.output_width;
 141         frame->height = dinfo.output_height;
 142         frame->chroma_subsampling_x = dinfo.max_h_samp_factor / dinfo.comp_info[1].h_samp_factor;
 143         frame->chroma_subsampling_y = dinfo.max_v_samp_factor / dinfo.comp_info[1].v_samp_factor;
 144
 145         unsigned h_mcu_size = DCTSIZE * dinfo.max_h_samp_factor;
 146         unsigned v_mcu_size = DCTSIZE * dinfo.max_v_samp_factor;
 147         unsigned mcu_width_blocks = (dinfo.output_width + h_mcu_size - 1) / h_mcu_size;
 148         unsigned mcu_height_blocks = (dinfo.output_height + v_mcu_size - 1) / v_mcu_size;
 149
 150         unsigned luma_width_blocks = mcu_width_blocks * dinfo.comp_info[0].h_samp_factor;
 151         unsigned chroma_width_blocks = mcu_width_blocks * dinfo.comp_info[1].h_samp_factor;
 152         unsigned luma_height_blocks = mcu_height_blocks * dinfo.comp_info[0].v_samp_factor;
 153         unsigned chroma_height_blocks = mcu_height_blocks * dinfo.comp_info[1].v_samp_factor;
 154
 155         // TODO: Decode into a PBO.
 156         frame->y.reset(new uint8_t[luma_width_blocks * luma_height_blocks * DCTSIZE2]);
 157         frame->cb.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 158         frame->cr.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 159         frame->pitch_y = luma_width_blocks * DCTSIZE;
 160         frame->pitch_chroma = chroma_width_blocks * DCTSIZE;
 161
 162         if (!error_mgr.run([&dinfo, &frame, v_mcu_size, mcu_height_blocks] {
 163                     JSAMPROW yptr[v_mcu_size], cbptr[v_mcu_size], crptr[v_mcu_size];
 164                     JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 165                     for (unsigned y = 0; y < mcu_height_blocks; ++y) {
 166                             // NOTE: The last elements of cbptr/crptr will be unused for vertically subsampled chroma.
 167                             for (unsigned yy = 0; yy < v_mcu_size; ++yy) {
 168                                     yptr[yy] = frame->y.get() + (y * DCTSIZE * dinfo.max_v_samp_factor + yy) * frame->pitch_y;
 169                                     cbptr[yy] = frame->cb.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 170                                     crptr[yy] = frame->cr.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 171                             }
 172
 173                             jpeg_read_raw_data(&dinfo, data, v_mcu_size);
 174                     }
 175
 176                     (void)jpeg_finish_decompress(&dinfo);
 177             })) {
 178                 return get_black_frame();
 179         }
 180
 181         ++metric_jpeg_software_decode_frames;
 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)
 190                 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)
 206                         break;
 207         }
 208
 209         for (auto it = cache.begin(); it != cache.end();) {
 210                 if (it->second.last_used <= lru_cutoff_point) {
 211                         cache_bytes_used -= frame_size(*it->second.frame);
 212                         metric_jpeg_cache_used_bytes = cache_bytes_used;
 213                         it = cache.erase(it);
 214                 } else {
 215                         ++it;
 216                 }
 217         }
 218 }
 219
 220 shared_ptr<Frame> decode_jpeg_with_cache(FrameOnDisk frame_spec, CacheMissBehavior cache_miss_behavior, FrameReader *frame_reader, bool *did_decode)
 221 {
 222         *did_decode = false;
 223         {
 224                 lock_guard<mutex> lock(cache_mu);
 225                 auto it = cache.find(frame_spec);
 226                 if (it != cache.end()) {
 227                         ++metric_jpeg_cache_hit_frames;
 228                         it->second.last_used = event_counter++;
 229                         return it->second.frame;
 230                 }
 231         }
 232
 233         if (cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE) {
 234                 ++metric_jpeg_cache_given_up_frames;
 235                 return nullptr;
 236         }
 237
 238         ++metric_jpeg_cache_miss_frames;
 239
 240         *did_decode = true;
 241         shared_ptr<Frame> frame = decode_jpeg(frame_reader->read_frame(frame_spec));
 242
 243         lock_guard<mutex> lock(cache_mu);
 244         cache_bytes_used += frame_size(*frame);
 245         metric_jpeg_cache_used_bytes = cache_bytes_used;
 246         cache[frame_spec] = LRUFrame{ frame, event_counter++ };
 247
 248         if (cache_bytes_used > size_t(CACHE_SIZE_MB) * 1024 * 1024) {
 249                 prune_cache();
 250         }
 251         return frame;
 252 }
 253
 254 void JPEGFrameView::jpeg_decoder_thread_func()
 255 {
 256         size_t num_decoded = 0, num_dropped = 0;
 257
 258         pthread_setname_np(pthread_self(), "JPEGDecoder");
 259         while (!should_quit.load()) {
 260                 PendingDecode decode;
 261                 CacheMissBehavior cache_miss_behavior = DECODE_IF_NOT_IN_CACHE;
 262                 {
 263                         unique_lock<mutex> lock(cache_mu);  // TODO: Perhaps under another lock?
 264                         any_pending_decodes.wait(lock, [this] {
 265                                 return !pending_decodes.empty() || should_quit.load();
 266                         });
 267                         if (should_quit.load())
 268                                 break;
 269                         decode = pending_decodes.front();
 270                         pending_decodes.pop_front();
 271
 272                         if (pending_decodes.size() > 3) {
 273                                 cache_miss_behavior = RETURN_NULLPTR_IF_NOT_IN_CACHE;
 274                         }
 275                 }
 276
 277                 if (decode.frame != nullptr) {
 278                         // Already decoded, so just show it.
 279                         setDecodedFrame(decode.frame, nullptr, 1.0f);
 280                         continue;
 281                 }
 282
 283                 shared_ptr<Frame> primary_frame, secondary_frame;
 284                 bool drop = false;
 285                 for (int subframe_idx = 0; subframe_idx < 2; ++subframe_idx) {
 286                         const FrameOnDisk &frame_spec = (subframe_idx == 0 ? decode.primary : decode.secondary);
 287                         if (frame_spec.pts == -1) {
 288                                 // No secondary frame.
 289                                 continue;
 290                         }
 291
 292                         bool found_in_cache;
 293                         shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, cache_miss_behavior, &frame_reader, &found_in_cache);
 294
 295                         if (frame == nullptr) {
 296                                 assert(cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
 297                                 drop = true;
 298                                 break;
 299                         }
 300
 301                         if (!found_in_cache) {
 302                                 ++num_decoded;
 303                                 if (num_decoded % 1000 == 0) {
 304                                         fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
 305                                                 num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
 306                                 }
 307                         }
 308                         if (subframe_idx == 0) {
 309                                 primary_frame = std::move(frame);
 310                         } else {
 311                                 secondary_frame = std::move(frame);
 312                         }
 313                 }
 314                 if (drop) {
 315                         ++num_dropped;
 316                         continue;
 317                 }
 318
 319                 // TODO: Could we get jitter between non-interpolated and interpolated frames here?
 320                 setDecodedFrame(primary_frame, secondary_frame, decode.fade_alpha);
 321         }
 322 }
 323
 324 JPEGFrameView::~JPEGFrameView()
 325 {
 326         any_pending_decodes.notify_all();
 327         jpeg_decoder_thread.join();
 328 }
 329
 330 JPEGFrameView::JPEGFrameView(QWidget *parent)
 331         : QGLWidget(parent, global_share_widget)
 332 {
 333         call_once(jpeg_metrics_inited, [] {
 334                 global_metrics.add("jpeg_cache_used_bytes", &metric_jpeg_cache_used_bytes, Metrics::TYPE_GAUGE);
 335                 global_metrics.add("jpeg_cache_limit_bytes", &metric_jpeg_cache_limit_bytes, Metrics::TYPE_GAUGE);
 336                 global_metrics.add("jpeg_cache_frames", { { "action", "given_up" } }, &metric_jpeg_cache_given_up_frames);
 337                 global_metrics.add("jpeg_cache_frames", { { "action", "hit" } }, &metric_jpeg_cache_hit_frames);
 338                 global_metrics.add("jpeg_cache_frames", { { "action", "miss" } }, &metric_jpeg_cache_miss_frames);
 339                 global_metrics.add("jpeg_decode_frames", { { "decoder", "software" }, { "result", "decode" } }, &metric_jpeg_software_decode_frames);
 340                 global_metrics.add("jpeg_decode_frames", { { "decoder", "software" }, { "result", "fail" } }, &metric_jpeg_software_fail_frames);
 341                 global_metrics.add("jpeg_decode_frames", { { "decoder", "vaapi" }, { "result", "decode" } }, &metric_jpeg_vaapi_decode_frames);
 342                 global_metrics.add("jpeg_decode_frames", { { "decoder", "vaapi" }, { "result", "fail" } }, &metric_jpeg_vaapi_fail_frames);
 343         });
 344 }
 345
 346 void JPEGFrameView::setFrame(unsigned stream_idx, FrameOnDisk frame, FrameOnDisk secondary_frame, float fade_alpha)
 347 {
 348         current_stream_idx = stream_idx;  // TODO: Does this interact with fades?
 349
 350         lock_guard<mutex> lock(cache_mu);
 351         PendingDecode decode;
 352         decode.primary = frame;
 353         decode.secondary = secondary_frame;
 354         decode.fade_alpha = fade_alpha;
 355         pending_decodes.push_back(decode);
 356         any_pending_decodes.notify_all();
 357 }
 358
 359 void JPEGFrameView::setFrame(shared_ptr<Frame> frame)
 360 {
 361         lock_guard<mutex> lock(cache_mu);
 362         PendingDecode decode;
 363         decode.frame = std::move(frame);
 364         pending_decodes.push_back(decode);
 365         any_pending_decodes.notify_all();
 366 }
 367
 368 void JPEGFrameView::initializeGL()
 369 {
 370         glDisable(GL_BLEND);
 371         glDisable(GL_DEPTH_TEST);
 372         check_error();
 373
 374         resource_pool = new ResourcePool;
 375         jpeg_decoder_thread = std::thread(&JPEGFrameView::jpeg_decoder_thread_func, this);
 376
 377         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_RGBA, resource_pool));
 378
 379         ImageFormat inout_format;
 380         inout_format.color_space = COLORSPACE_sRGB;
 381         inout_format.gamma_curve = GAMMA_sRGB;
 382
 383         overlay_chain.reset(new EffectChain(overlay_base_width, overlay_base_height, resource_pool));
 384         overlay_input = (movit::FlatInput *)overlay_chain->add_input(new FlatInput(inout_format, FORMAT_GRAYSCALE, GL_UNSIGNED_BYTE, overlay_base_width, overlay_base_height));
 385
 386         overlay_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 387         overlay_chain->finalize();
 388 }
 389
 390 void JPEGFrameView::resizeGL(int width, int height)
 391 {
 392         check_error();
 393         glViewport(0, 0, width, height);
 394         check_error();
 395
 396         // Save these, as width() and height() will lie with DPI scaling.
 397         gl_width = width;
 398         gl_height = height;
 399 }
 400
 401 void JPEGFrameView::paintGL()
 402 {
 403         glViewport(0, 0, gl_width, gl_height);
 404         if (current_frame == nullptr) {
 405                 glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
 406                 glClear(GL_COLOR_BUFFER_BIT);
 407                 return;
 408         }
 409
 410         check_error();
 411         current_chain->render_to_screen();
 412
 413         if (overlay_image != nullptr) {
 414                 if (overlay_input_needs_refresh) {
 415                         overlay_input->set_width(overlay_width);
 416                         overlay_input->set_height(overlay_height);
 417                         overlay_input->set_pixel_data(overlay_image->bits());
 418                 }
 419                 glViewport(gl_width - overlay_width, 0, overlay_width, overlay_height);
 420                 overlay_chain->render_to_screen();
 421         }
 422 }
 423
 424 namespace {
 425
 426 }  // namespace
 427
 428 void JPEGFrameView::setDecodedFrame(shared_ptr<Frame> frame, shared_ptr<Frame> secondary_frame, float fade_alpha)
 429 {
 430         post_to_main_thread([this, frame, secondary_frame, fade_alpha] {
 431                 current_frame = frame;
 432                 current_secondary_frame = secondary_frame;
 433
 434                 if (secondary_frame != nullptr) {
 435                         current_chain = ycbcr_converter->prepare_chain_for_fade(frame, secondary_frame, fade_alpha);
 436                 } else {
 437                         current_chain = ycbcr_converter->prepare_chain_for_conversion(frame);
 438                 }
 439                 update();
 440         });
 441 }
 442
 443 void JPEGFrameView::mousePressEvent(QMouseEvent *event)
 444 {
 445         if (event->type() == QEvent::MouseButtonPress && event->button() == Qt::LeftButton) {
 446                 emit clicked();
 447         }
 448 }
 449
 450 void JPEGFrameView::set_overlay(const string &text)
 451 {
 452         if (text.empty()) {
 453                 overlay_image.reset();
 454                 return;
 455         }
 456
 457         float dpr = QGuiApplication::primaryScreen()->devicePixelRatio();
 458         overlay_width = lrint(overlay_base_width * dpr);
 459         overlay_height = lrint(overlay_base_height * dpr);
 460
 461         overlay_image.reset(new QImage(overlay_width, overlay_height, QImage::Format_Grayscale8));
 462         overlay_image->setDevicePixelRatio(dpr);
 463         overlay_image->fill(0);
 464         QPainter painter(overlay_image.get());
 465
 466         painter.setPen(Qt::white);
 467         QFont font = painter.font();
 468         font.setPointSize(12);
 469         painter.setFont(font);
 470
 471         painter.drawText(QRectF(0, 0, overlay_base_width, overlay_base_height), Qt::AlignCenter, QString::fromStdString(text));
 472
 473         // Don't refresh immediately; we might not have an OpenGL context here.
 474         overlay_input_needs_refresh = true;
 475 }
 476
 477 shared_ptr<Frame> get_black_frame()
 478 {
 479         static shared_ptr<Frame> black_frame;
 480         static once_flag flag;
 481         call_once(flag, [] {
 482                 black_frame.reset(new Frame);
 483                 black_frame->y.reset(new uint8_t[global_flags.width * global_flags.height]);
 484                 black_frame->cb.reset(new uint8_t[(global_flags.width / 2) * (global_flags.height / 2)]);
 485                 black_frame->cr.reset(new uint8_t[(global_flags.width / 2) * (global_flags.height / 2)]);
 486                 black_frame->width = global_flags.width;
 487                 black_frame->height = global_flags.height;
 488                 black_frame->chroma_subsampling_x = 2;
 489                 black_frame->chroma_subsampling_y = 2;
 490                 black_frame->pitch_y = global_flags.width;
 491                 black_frame->pitch_chroma = global_flags.width / 2;
 492         });
 493         ++metric_jpeg_software_fail_frames;
 494         return black_frame;
 495 }