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