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 // Just an arbitrary order for std::map.
  31 struct JPEGIDLexicalOrder
  32 {
  33         bool operator() (const JPEGID &a, const JPEGID &b) const
  34         {
  35                 if (a.stream_idx != b.stream_idx)
  36                         return a.stream_idx < b.stream_idx;
  37                 if (a.pts != b.pts)
  38                         return a.pts < b.pts;
  39                 return a.interpolated < b.interpolated;
  40         }
  41 };
  42
  43 struct LRUFrame {
  44         shared_ptr<Frame> frame;
  45         size_t last_used;
  46 };
  47
  48 struct PendingDecode {
  49         JPEGID primary, secondary;
  50         float fade_alpha;  // Irrelevant if secondary.stream_idx == -1.
  51         JPEGFrameView *destination;
  52 };
  53
  54 thread JPEGFrameView::jpeg_decoder_thread;
  55 mutex cache_mu;
  56 map<JPEGID, LRUFrame, JPEGIDLexicalOrder> cache;  // Under cache_mu.
  57 condition_variable any_pending_decodes;
  58 deque<PendingDecode> pending_decodes;  // Under cache_mu.
  59 atomic<size_t> event_counter{0};
  60 extern QGLWidget *global_share_widget;
  61 extern atomic<bool> should_quit;
  62
  63 shared_ptr<Frame> decode_jpeg(const string &filename)
  64 {
  65         shared_ptr<Frame> frame;
  66         if (vaapi_jpeg_decoding_usable) {
  67                 frame = decode_jpeg_vaapi(filename);
  68                 if (frame != nullptr) {
  69                         return frame;
  70                 }
  71                 fprintf(stderr, "VA-API hardware decoding failed; falling back to software.\n");
  72         }
  73
  74         frame.reset(new Frame);
  75
  76         jpeg_decompress_struct dinfo;
  77         jpeg_error_mgr jerr;
  78         dinfo.err = jpeg_std_error(&jerr);
  79         jpeg_create_decompress(&dinfo);
  80         JPEGDestroyer destroy_dinfo(&dinfo);
  81
  82         FILE *fp = fopen(filename.c_str(), "rb");
  83         if (fp == nullptr) {
  84                 perror(filename.c_str());
  85                 exit(1);
  86         }
  87         jpeg_stdio_src(&dinfo, fp);
  88
  89         jpeg_read_header(&dinfo, true);
  90
  91         if (dinfo.num_components != 3) {
  92                 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
  93                         dinfo.num_components,
  94                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
  95                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
  96                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
  97                 exit(1);
  98         }
  99         if (dinfo.comp_info[0].h_samp_factor != dinfo.max_h_samp_factor ||
 100             dinfo.comp_info[0].v_samp_factor != dinfo.max_v_samp_factor ||  // Y' must not be subsampled.
 101             dinfo.comp_info[1].h_samp_factor != dinfo.comp_info[2].h_samp_factor ||
 102             dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[2].v_samp_factor ||  // Cb and Cr must be identically subsampled.
 103             (dinfo.max_h_samp_factor % dinfo.comp_info[1].h_samp_factor) != 0 ||
 104             (dinfo.max_v_samp_factor % dinfo.comp_info[1].v_samp_factor) != 0) {  // No 2:3 subsampling or other weirdness.
 105                 fprintf(stderr, "Unsupported subsampling scheme. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 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         dinfo.raw_data_out = true;
 112
 113         jpeg_start_decompress(&dinfo);
 114
 115         frame->width = dinfo.output_width;
 116         frame->height = dinfo.output_height;
 117         frame->chroma_subsampling_x = dinfo.max_h_samp_factor / dinfo.comp_info[1].h_samp_factor;
 118         frame->chroma_subsampling_y = dinfo.max_v_samp_factor / dinfo.comp_info[1].v_samp_factor;
 119
 120         unsigned h_mcu_size = DCTSIZE * dinfo.max_h_samp_factor;
 121         unsigned v_mcu_size = DCTSIZE * dinfo.max_v_samp_factor;
 122         unsigned mcu_width_blocks = (dinfo.output_width + h_mcu_size - 1) / h_mcu_size;
 123         unsigned mcu_height_blocks = (dinfo.output_height + v_mcu_size - 1) / v_mcu_size;
 124
 125         unsigned luma_width_blocks = mcu_width_blocks * dinfo.comp_info[0].h_samp_factor;
 126         unsigned chroma_width_blocks = mcu_width_blocks * dinfo.comp_info[1].h_samp_factor;
 127         unsigned luma_height_blocks = mcu_height_blocks * dinfo.comp_info[0].v_samp_factor;
 128         unsigned chroma_height_blocks = mcu_height_blocks * dinfo.comp_info[1].v_samp_factor;
 129
 130         // TODO: Decode into a PBO.
 131         frame->y.reset(new uint8_t[luma_width_blocks * luma_height_blocks * DCTSIZE2]);
 132         frame->cb.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 133         frame->cr.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 134         frame->pitch_y = luma_width_blocks * DCTSIZE;
 135         frame->pitch_chroma = chroma_width_blocks * DCTSIZE;
 136
 137         JSAMPROW yptr[v_mcu_size], cbptr[v_mcu_size], crptr[v_mcu_size];
 138         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 139         for (unsigned y = 0; y < mcu_height_blocks; ++y) {
 140                 // NOTE: The last elements of cbptr/crptr will be unused for vertically subsampled chroma.
 141                 for (unsigned yy = 0; yy < v_mcu_size; ++yy) {
 142                         yptr[yy] = frame->y.get() + (y * DCTSIZE * dinfo.max_v_samp_factor + yy) * frame->pitch_y;
 143                         cbptr[yy] = frame->cb.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 144                         crptr[yy] = frame->cr.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 145                 }
 146
 147                 jpeg_read_raw_data(&dinfo, data, v_mcu_size);
 148         }
 149
 150         (void)jpeg_finish_decompress(&dinfo);
 151         fclose(fp);
 152
 153         return frame;
 154 }
 155
 156 void prune_cache()
 157 {
 158         // Assumes cache_mu is held.
 159         vector<size_t> lru_timestamps;
 160         for (const auto &key_and_value : cache) {
 161                 lru_timestamps.push_back(key_and_value.second.last_used);
 162         }
 163
 164         size_t cutoff_point = CACHE_SIZE / 10;  // Prune away the 10% oldest ones.
 165         nth_element(lru_timestamps.begin(), lru_timestamps.begin() + cutoff_point, lru_timestamps.end());
 166         size_t must_be_used_after = lru_timestamps[cutoff_point];
 167         for (auto it = cache.begin(); it != cache.end(); ) {
 168                 if (it->second.last_used < must_be_used_after) {
 169                         it = cache.erase(it);
 170                 } else {
 171                         ++it;
 172                 }
 173         }
 174 }
 175
 176 shared_ptr<Frame> decode_jpeg_with_cache(JPEGID id, CacheMissBehavior cache_miss_behavior, bool *did_decode)
 177 {
 178         *did_decode = false;
 179         {
 180                 unique_lock<mutex> lock(cache_mu);
 181                 auto it = cache.find(id);
 182                 if (it != cache.end()) {
 183                         it->second.last_used = event_counter++;
 184                         return it->second.frame;
 185                 }
 186         }
 187
 188         if (cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE) {
 189                 return nullptr;
 190         }
 191
 192         assert(!id.interpolated);
 193         *did_decode = true;
 194         shared_ptr<Frame> frame = decode_jpeg(filename_for_frame(id.stream_idx, id.pts));
 195
 196         unique_lock<mutex> lock(cache_mu);
 197         cache[id] = LRUFrame{ frame, event_counter++ };
 198
 199         if (cache.size() > CACHE_SIZE) {
 200                 prune_cache();
 201         }
 202         return frame;
 203 }
 204
 205 void jpeg_decoder_thread_func()
 206 {
 207         size_t num_decoded = 0, num_dropped = 0;
 208
 209         pthread_setname_np(pthread_self(), "JPEGDecoder");
 210         while (!should_quit.load()) {
 211                 PendingDecode decode;
 212                 CacheMissBehavior cache_miss_behavior = DECODE_IF_NOT_IN_CACHE;
 213                 {
 214                         unique_lock<mutex> lock(cache_mu);  // TODO: Perhaps under another lock?
 215                         any_pending_decodes.wait(lock, [] {
 216                                 return !pending_decodes.empty() || should_quit.load();
 217                         });
 218                         if (should_quit.load())
 219                                 break;
 220                         decode = pending_decodes.front();
 221                         pending_decodes.pop_front();
 222
 223                         size_t num_pending = 0;
 224                         for (const PendingDecode &other_decode : pending_decodes) {
 225                                 if (other_decode.destination == decode.destination) {
 226                                         ++num_pending;
 227                                 }
 228                         }
 229                         if (num_pending > 3) {
 230                                 cache_miss_behavior = RETURN_NULLPTR_IF_NOT_IN_CACHE;
 231                         }
 232                 }
 233
 234                 shared_ptr<Frame> primary_frame, secondary_frame;
 235                 bool drop = false;
 236                 for (int subframe_idx = 0; subframe_idx < 2; ++subframe_idx) {
 237                         const JPEGID &id = (subframe_idx == 0 ? decode.primary : decode.secondary);
 238                         if (id.stream_idx == (unsigned)-1) {
 239                                 // No secondary frame.
 240                                 continue;
 241                         }
 242
 243                         bool found_in_cache;
 244                         shared_ptr<Frame> frame;
 245                         if (id.interpolated) {
 246                                 // Interpolated frames are never decoded by us,
 247                                 // put directly into the cache from VideoStream.
 248                                 unique_lock<mutex> lock(cache_mu);
 249                                 auto it = cache.find(id);
 250                                 if (it != cache.end()) {
 251                                         it->second.last_used = event_counter++;
 252                                         frame = it->second.frame;
 253                                 } else {
 254                                         // This can only really happen if it disappeared out of the
 255                                         // LRU really, really fast. Which shouldn't happen.
 256                                         fprintf(stderr, "WARNING: Interpolated JPEG was supposed to be in the cache, but was not\n");
 257                                 }
 258                                 found_in_cache = true;  // Don't count it as a decode.
 259                         } else {
 260                                 frame = decode_jpeg_with_cache(id, cache_miss_behavior, &found_in_cache);
 261                         }
 262
 263                         if (frame == nullptr) {
 264                                 assert(id.interpolated || cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
 265                                 drop = true;
 266                                 break;
 267                         }
 268
 269                         if (!found_in_cache) {
 270                                 ++num_decoded;
 271                                 if (num_decoded % 1000 == 0) {
 272                                         fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
 273                                                 num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
 274                                 }
 275                         }
 276                         if (subframe_idx == 0) {
 277                                 primary_frame = move(frame);
 278                         } else {
 279                                 secondary_frame = move(frame);
 280                         }
 281                 }
 282                 if (drop) {
 283                         ++num_dropped;
 284                         continue;
 285                 }
 286
 287                 // TODO: Could we get jitter between non-interpolated and interpolated frames here?
 288                 decode.destination->setDecodedFrame(primary_frame, secondary_frame, decode.fade_alpha);
 289         }
 290 }
 291
 292 void JPEGFrameView::shutdown()
 293 {
 294         any_pending_decodes.notify_all();
 295         jpeg_decoder_thread.join();
 296 }
 297
 298 JPEGFrameView::JPEGFrameView(QWidget *parent)
 299         : QGLWidget(parent, global_share_widget)
 300 {
 301 }
 302
 303 void JPEGFrameView::setFrame(unsigned stream_idx, int64_t pts, bool interpolated, int secondary_stream_idx, int64_t secondary_pts, float fade_alpha)
 304 {
 305         current_stream_idx = stream_idx;  // TODO: Does this interact with fades?
 306
 307         unique_lock<mutex> lock(cache_mu);
 308         PendingDecode decode;
 309         if (interpolated && secondary_stream_idx != -1) {
 310                 // The frame will already be faded for us, so ask for only one; we shouldn't fade it against anything.
 311                 decode.primary = create_jpegid_for_interpolated_fade(stream_idx, pts, secondary_stream_idx, secondary_pts);
 312                 decode.secondary = JPEGID{ (unsigned)-1, -1, /*interpolated=*/false };
 313         } else {
 314                 decode.primary = JPEGID{ stream_idx, pts, interpolated };
 315                 decode.secondary = JPEGID{ (unsigned)secondary_stream_idx, secondary_pts, /*interpolated=*/false };
 316         }
 317         decode.fade_alpha = fade_alpha;
 318         decode.destination = this;
 319         pending_decodes.push_back(decode);
 320         any_pending_decodes.notify_all();
 321 }
 322
 323 void JPEGFrameView::insert_interpolated_frame(JPEGID id, shared_ptr<Frame> frame)
 324 {
 325         // We rely on the frame not being evicted from the cache before
 326         // jpeg_decoder_thread() sees it and can display it (otherwise,
 327         // that thread would hang). With a default cache of 1000 elements,
 328         // that would sound like a reasonable assumption.
 329         unique_lock<mutex> lock(cache_mu);
 330         cache[id] = LRUFrame{ std::move(frame), event_counter++ };
 331         prune_cache();
 332 }
 333
 334 ResourcePool *resource_pool = nullptr;
 335
 336 void JPEGFrameView::initializeGL()
 337 {
 338         glDisable(GL_BLEND);
 339         glDisable(GL_DEPTH_TEST);
 340         check_error();
 341
 342         static once_flag once;
 343         call_once(once, [] {
 344                 resource_pool = new ResourcePool;
 345                 jpeg_decoder_thread = std::thread(jpeg_decoder_thread_func);
 346         });
 347
 348         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_RGBA, resource_pool));
 349
 350         ImageFormat inout_format;
 351         inout_format.color_space = COLORSPACE_sRGB;
 352         inout_format.gamma_curve = GAMMA_sRGB;
 353
 354         overlay_chain.reset(new EffectChain(overlay_base_width, overlay_base_height, resource_pool));
 355         overlay_input = (movit::FlatInput *)overlay_chain->add_input(new FlatInput(inout_format, FORMAT_GRAYSCALE, GL_UNSIGNED_BYTE, overlay_base_width, overlay_base_height));
 356
 357         overlay_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 358         overlay_chain->finalize();
 359 }
 360
 361 void JPEGFrameView::resizeGL(int width, int height)
 362 {
 363         check_error();
 364         glViewport(0, 0, width, height);
 365         check_error();
 366
 367         // Save these, as width() and height() will lie with DPI scaling.
 368         gl_width = width;
 369         gl_height = height;
 370 }
 371
 372 void JPEGFrameView::paintGL()
 373 {
 374         glViewport(0, 0, gl_width, gl_height);
 375         if (current_frame == nullptr) {
 376                 glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
 377                 glClear(GL_COLOR_BUFFER_BIT);
 378                 return;
 379         }
 380
 381         check_error();
 382         current_chain->render_to_screen();
 383
 384         if (overlay_image != nullptr) {
 385                 if (overlay_input_needs_refresh) {
 386                         overlay_input->set_width(overlay_width);
 387                         overlay_input->set_height(overlay_height);
 388                         overlay_input->set_pixel_data(overlay_image->bits());
 389                 }
 390                 glViewport(gl_width - overlay_width, 0, overlay_width, overlay_height);
 391                 overlay_chain->render_to_screen();
 392         }
 393 }
 394
 395 namespace {
 396
 397 }  // namespace
 398
 399 void JPEGFrameView::setDecodedFrame(shared_ptr<Frame> frame, shared_ptr<Frame> secondary_frame, float fade_alpha)
 400 {
 401         post_to_main_thread([this, frame, secondary_frame, fade_alpha] {
 402                 current_frame = frame;
 403                 current_secondary_frame = secondary_frame;
 404
 405                 if (secondary_frame != nullptr) {
 406                         current_chain = ycbcr_converter->prepare_chain_for_fade(frame, secondary_frame, fade_alpha);
 407                 } else {
 408                         current_chain = ycbcr_converter->prepare_chain_for_conversion(frame);
 409                 }
 410                 update();
 411         });
 412 }
 413
 414 void JPEGFrameView::mousePressEvent(QMouseEvent *event)
 415 {
 416         if (event->type() == QEvent::MouseButtonPress && event->button() == Qt::LeftButton) {
 417                 emit clicked();
 418         }
 419 }
 420
 421 void JPEGFrameView::set_overlay(const string &text)
 422 {
 423         if (text.empty()) {
 424                 overlay_image.reset();
 425                 return;
 426         }
 427
 428         float dpr = QGuiApplication::primaryScreen()->devicePixelRatio();
 429         overlay_width = lrint(overlay_base_width * dpr);
 430         overlay_height = lrint(overlay_base_height * dpr);
 431
 432         overlay_image.reset(new QImage(overlay_width, overlay_height, QImage::Format_Grayscale8));
 433         overlay_image->setDevicePixelRatio(dpr);
 434         overlay_image->fill(0);
 435         QPainter painter(overlay_image.get());
 436
 437         painter.setPen(Qt::white);
 438         QFont font = painter.font();
 439         font.setPointSize(12);
 440         painter.setFont(font);
 441
 442         painter.drawText(QRectF(0, 0, overlay_base_width, overlay_base_height), Qt::AlignCenter, QString::fromStdString(text));
 443
 444         // Don't refresh immediately; we might not have an OpenGL context here.
 445         overlay_input_needs_refresh = true;
 446 }