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