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

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