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

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