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

   1 #include "jpeg_frame_view.h"
   2
   3 #include "defs.h"
   4 #include "post_to_main_thread.h"
   5 #include "video_stream.h"
   6 #include "ycbcr_converter.h"
   7
   8 #include <QMouseEvent>
   9 #include <QScreen>
  10 #include <atomic>
  11 #include <condition_variable>
  12 #include <deque>
  13 #include <jpeglib.h>
  14 #include <movit/init.h>
  15 #include <movit/resource_pool.h>
  16 #include <movit/util.h>
  17 #include <mutex>
  18 #include <stdint.h>
  19 #include <thread>
  20 #include <unistd.h>
  21 #include <utility>
  22
  23 // Must come after the Qt stuff.
  24 #include "vaapi_jpeg_decoder.h"
  25
  26 using namespace movit;
  27 using namespace std;
  28
  29 // Just an arbitrary order for std::map.
  30 struct JPEGIDLexicalOrder
  31 {
  32         bool operator() (const JPEGID &a, const JPEGID &b) const
  33         {
  34                 if (a.stream_idx != b.stream_idx)
  35                         return a.stream_idx < b.stream_idx;
  36                 if (a.pts != b.pts)
  37                         return a.pts < b.pts;
  38                 return a.interpolated < b.interpolated;
  39         }
  40 };
  41
  42 struct LRUFrame {
  43         shared_ptr<Frame> frame;
  44         size_t last_used;
  45 };
  46
  47 struct PendingDecode {
  48         JPEGID primary, secondary;
  49         float fade_alpha;  // Irrelevant if secondary.stream_idx == -1.
  50         JPEGFrameView *destination;
  51 };
  52
  53 thread JPEGFrameView::jpeg_decoder_thread;
  54 mutex cache_mu;
  55 map<JPEGID, LRUFrame, JPEGIDLexicalOrder> cache;  // Under cache_mu.
  56 condition_variable any_pending_decodes, cache_updated;
  57 deque<PendingDecode> pending_decodes;  // Under cache_mu.
  58 atomic<size_t> event_counter{0};
  59 extern QGLWidget *global_share_widget;
  60 extern atomic<bool> should_quit;
  61
  62 shared_ptr<Frame> decode_jpeg(const string &filename)
  63 {
  64         shared_ptr<Frame> frame;
  65         if (vaapi_jpeg_decoding_usable) {
  66                 frame = decode_jpeg_vaapi(filename);
  67                 if (frame != nullptr) {
  68                         return frame;
  69                 }
  70                 fprintf(stderr, "VA-API hardware decoding failed; falling back to software.\n");
  71         }
  72
  73         frame.reset(new Frame);
  74
  75         jpeg_decompress_struct dinfo;
  76         jpeg_error_mgr jerr;
  77         dinfo.err = jpeg_std_error(&jerr);
  78         jpeg_create_decompress(&dinfo);
  79
  80         FILE *fp = fopen(filename.c_str(), "rb");
  81         if (fp == nullptr) {
  82                 perror(filename.c_str());
  83                 exit(1);
  84         }
  85         jpeg_stdio_src(&dinfo, fp);
  86
  87         jpeg_read_header(&dinfo, true);
  88
  89         if (dinfo.num_components != 3) {
  90                 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
  91                         dinfo.num_components,
  92                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
  93                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
  94                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
  95                 exit(1);
  96         }
  97         if (dinfo.comp_info[0].h_samp_factor != dinfo.max_h_samp_factor ||
  98             dinfo.comp_info[0].v_samp_factor != dinfo.max_v_samp_factor ||  // Y' must not be subsampled.
  99             dinfo.comp_info[1].h_samp_factor != dinfo.comp_info[2].h_samp_factor ||
 100             dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[2].v_samp_factor ||  // Cb and Cr must be identically subsampled.
 101             (dinfo.max_h_samp_factor % dinfo.comp_info[1].h_samp_factor) != 0 ||
 102             (dinfo.max_v_samp_factor % dinfo.comp_info[1].v_samp_factor) != 0) {  // No 2:3 subsampling or other weirdness.
 103                 fprintf(stderr, "Unsupported subsampling scheme. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 104                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 105                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 106                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 107                 exit(1);
 108         }
 109         dinfo.raw_data_out = true;
 110
 111         jpeg_start_decompress(&dinfo);
 112
 113         frame->width = dinfo.output_width;
 114         frame->height = dinfo.output_height;
 115         frame->chroma_subsampling_x = dinfo.max_h_samp_factor / dinfo.comp_info[1].h_samp_factor;
 116         frame->chroma_subsampling_y = dinfo.max_v_samp_factor / dinfo.comp_info[1].v_samp_factor;
 117
 118         unsigned h_mcu_size = DCTSIZE * dinfo.max_h_samp_factor;
 119         unsigned v_mcu_size = DCTSIZE * dinfo.max_v_samp_factor;
 120         unsigned mcu_width_blocks = (dinfo.output_width + h_mcu_size - 1) / h_mcu_size;
 121         unsigned mcu_height_blocks = (dinfo.output_height + v_mcu_size - 1) / v_mcu_size;
 122
 123         unsigned luma_width_blocks = mcu_width_blocks * dinfo.comp_info[0].h_samp_factor;
 124         unsigned chroma_width_blocks = mcu_width_blocks * dinfo.comp_info[1].h_samp_factor;
 125         unsigned luma_height_blocks = mcu_height_blocks * dinfo.comp_info[0].v_samp_factor;
 126         unsigned chroma_height_blocks = mcu_height_blocks * dinfo.comp_info[1].v_samp_factor;
 127
 128         // TODO: Decode into a PBO.
 129         frame->y.reset(new uint8_t[luma_width_blocks * luma_height_blocks * DCTSIZE2]);
 130         frame->cb.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 131         frame->cr.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 132         frame->pitch_y = luma_width_blocks * DCTSIZE;
 133         frame->pitch_chroma = chroma_width_blocks * DCTSIZE;
 134
 135         JSAMPROW yptr[v_mcu_size], cbptr[v_mcu_size], crptr[v_mcu_size];
 136         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 137         for (unsigned y = 0; y < mcu_height_blocks; ++y) {
 138                 // NOTE: The last elements of cbptr/crptr will be unused for vertically subsampled chroma.
 139                 for (unsigned yy = 0; yy < v_mcu_size; ++yy) {
 140                         yptr[yy] = frame->y.get() + (y * DCTSIZE * dinfo.max_v_samp_factor + yy) * frame->pitch_y;
 141                         cbptr[yy] = frame->cb.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 142                         crptr[yy] = frame->cr.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 143                 }
 144
 145                 jpeg_read_raw_data(&dinfo, data, v_mcu_size);
 146         }
 147
 148         (void)jpeg_finish_decompress(&dinfo);
 149         jpeg_destroy_decompress(&dinfo);
 150         fclose(fp);
 151
 152         return frame;
 153 }
 154
 155 void prune_cache()
 156 {
 157         // Assumes cache_mu is held.
 158         vector<size_t> lru_timestamps;
 159         for (const auto &key_and_value : cache) {
 160                 lru_timestamps.push_back(key_and_value.second.last_used);
 161         }
 162
 163         size_t cutoff_point = CACHE_SIZE / 10;  // Prune away the 10% oldest ones.
 164         nth_element(lru_timestamps.begin(), lru_timestamps.begin() + cutoff_point, lru_timestamps.end());
 165         size_t must_be_used_after = lru_timestamps[cutoff_point];
 166         for (auto it = cache.begin(); it != cache.end(); ) {
 167                 if (it->second.last_used < must_be_used_after) {
 168                         it = cache.erase(it);
 169                 } else {
 170                         ++it;
 171                 }
 172         }
 173 }
 174
 175 shared_ptr<Frame> decode_jpeg_with_cache(JPEGID id, CacheMissBehavior cache_miss_behavior, bool *did_decode)
 176 {
 177         *did_decode = false;
 178         {
 179                 unique_lock<mutex> lock(cache_mu);
 180                 auto it = cache.find(id);
 181                 if (it != cache.end()) {
 182                         it->second.last_used = event_counter++;
 183                         return it->second.frame;
 184                 }
 185         }
 186
 187         if (cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE) {
 188                 return nullptr;
 189         }
 190
 191         assert(!id.interpolated);
 192         *did_decode = true;
 193         shared_ptr<Frame> frame = decode_jpeg(filename_for_frame(id.stream_idx, id.pts));
 194
 195         unique_lock<mutex> lock(cache_mu);
 196         cache[id] = LRUFrame{ frame, event_counter++ };
 197
 198         if (cache.size() > CACHE_SIZE) {
 199                 prune_cache();
 200         }
 201         return frame;
 202 }
 203
 204 void jpeg_decoder_thread_func()
 205 {
 206         size_t num_decoded = 0, num_dropped = 0;
 207
 208         pthread_setname_np(pthread_self(), "JPEGDecoder");
 209         while (!should_quit.load()) {
 210                 PendingDecode decode;
 211                 CacheMissBehavior cache_miss_behavior = DECODE_IF_NOT_IN_CACHE;
 212                 {
 213                         unique_lock<mutex> lock(cache_mu);  // TODO: Perhaps under another lock?
 214                         any_pending_decodes.wait(lock, [] {
 215                                 return !pending_decodes.empty() || should_quit.load();
 216                         });
 217                         if (should_quit.load())
 218                                 break;
 219                         decode = pending_decodes.front();
 220                         pending_decodes.pop_front();
 221
 222                         size_t num_pending = 0;
 223                         for (const PendingDecode &other_decode : pending_decodes) {
 224                                 if (other_decode.destination == decode.destination) {
 225                                         ++num_pending;
 226                                 }
 227                         }
 228                         if (num_pending > 3) {
 229                                 cache_miss_behavior = RETURN_NULLPTR_IF_NOT_IN_CACHE;
 230                         }
 231                 }
 232
 233                 shared_ptr<Frame> primary_frame, secondary_frame;
 234                 bool drop = false;
 235                 for (int subframe_idx = 0; subframe_idx < 2; ++subframe_idx) {
 236                         const JPEGID &id = (subframe_idx == 0 ? decode.primary : decode.secondary);
 237                         if (id.stream_idx == (unsigned)-1) {
 238                                 // No secondary frame.
 239                                 continue;
 240                         }
 241
 242                         bool found_in_cache;
 243                         shared_ptr<Frame> frame;
 244                         if (id.interpolated) {
 245                                 // Interpolated frames are never decoded by us,
 246                                 // put directly into the cache from VideoStream.
 247                                 unique_lock<mutex> lock(cache_mu);
 248                                 cache_updated.wait(lock, [id] {
 249                                         return cache.count(id) != 0 || should_quit.load();
 250                                 });
 251                                 if (should_quit.load())
 252                                         break;
 253                                 found_in_cache = true;  // Don't count it as a decode.
 254
 255                                 auto it = cache.find(id);
 256                                 assert(it != cache.end());
 257
 258                                 it->second.last_used = event_counter++;
 259                                 frame = it->second.frame;
 260                                 if (frame == nullptr) {
 261                                         // We inserted a nullptr as signal that the frame was never
 262                                         // interpolated and that we should stop waiting.
 263                                         // But don't let it linger in the cache anymore.
 264                                         cache.erase(it);
 265                                 }
 266                         } else {
 267                                 frame = decode_jpeg_with_cache(id, cache_miss_behavior, &found_in_cache);
 268                         }
 269
 270                         if (frame == nullptr) {
 271                                 assert(id.interpolated || cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
 272                                 drop = true;
 273                                 break;
 274                         }
 275
 276                         if (!found_in_cache) {
 277                                 ++num_decoded;
 278                                 if (num_decoded % 1000 == 0) {
 279                                         fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
 280                                                 num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
 281                                 }
 282                         }
 283                         if (subframe_idx == 0) {
 284                                 primary_frame = move(frame);
 285                         } else {
 286                                 secondary_frame = move(frame);
 287                         }
 288                 }
 289                 if (drop) {
 290                         ++num_dropped;
 291                         continue;
 292                 }
 293
 294                 // TODO: Could we get jitter between non-interpolated and interpolated frames here?
 295                 decode.destination->setDecodedFrame(primary_frame, secondary_frame, decode.fade_alpha);
 296         }
 297 }
 298
 299 void JPEGFrameView::shutdown()
 300 {
 301         any_pending_decodes.notify_all();
 302         jpeg_decoder_thread.join();
 303 }
 304
 305 JPEGFrameView::JPEGFrameView(QWidget *parent)
 306         : QGLWidget(parent, global_share_widget)
 307 {
 308 }
 309
 310 void JPEGFrameView::setFrame(unsigned stream_idx, int64_t pts, bool interpolated, int secondary_stream_idx, int64_t secondary_pts, float fade_alpha)
 311 {
 312         current_stream_idx = stream_idx;  // TODO: Does this interact with fades?
 313
 314         unique_lock<mutex> lock(cache_mu);
 315         PendingDecode decode;
 316         if (interpolated && secondary_stream_idx != -1) {
 317                 // The frame will already be faded for us, so ask for only one; we shouldn't fade it against anything.
 318                 decode.primary = create_jpegid_for_interpolated_fade(stream_idx, pts, secondary_stream_idx, secondary_pts);
 319                 decode.secondary = JPEGID{ (unsigned)-1, -1, /*interpolated=*/false };
 320         } else {
 321                 decode.primary = JPEGID{ stream_idx, pts, interpolated };
 322                 decode.secondary = JPEGID{ (unsigned)secondary_stream_idx, secondary_pts, /*interpolated=*/false };
 323         }
 324         decode.fade_alpha = fade_alpha;
 325         decode.destination = this;
 326         pending_decodes.push_back(decode);
 327         any_pending_decodes.notify_all();
 328 }
 329
 330 void JPEGFrameView::insert_interpolated_frame(JPEGID id, shared_ptr<Frame> frame)
 331 {
 332         // We rely on the frame not being evicted from the cache before
 333         // jpeg_decoder_thread() sees it and can display it (otherwise,
 334         // that thread would hang). With a default cache of 1000 elements,
 335         // that would sound like a reasonable assumption.
 336         unique_lock<mutex> lock(cache_mu);
 337         cache[id] = LRUFrame{ std::move(frame), event_counter++ };
 338         cache_updated.notify_all();
 339 }
 340
 341 ResourcePool *resource_pool = nullptr;
 342
 343 void JPEGFrameView::initializeGL()
 344 {
 345         glDisable(GL_BLEND);
 346         glDisable(GL_DEPTH_TEST);
 347         check_error();
 348
 349         static once_flag once;
 350         call_once(once, [] {
 351                 resource_pool = new ResourcePool;
 352                 jpeg_decoder_thread = std::thread(jpeg_decoder_thread_func);
 353         });
 354
 355         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_RGBA, resource_pool));
 356
 357         ImageFormat inout_format;
 358         inout_format.color_space = COLORSPACE_sRGB;
 359         inout_format.gamma_curve = GAMMA_sRGB;
 360
 361         overlay_chain.reset(new EffectChain(overlay_base_width, overlay_base_height, resource_pool));
 362         overlay_input = (movit::FlatInput *)overlay_chain->add_input(new FlatInput(inout_format, FORMAT_GRAYSCALE, GL_UNSIGNED_BYTE, overlay_base_width, overlay_base_height));
 363
 364         overlay_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 365         overlay_chain->finalize();
 366 }
 367
 368 void JPEGFrameView::resizeGL(int width, int height)
 369 {
 370         check_error();
 371         glViewport(0, 0, width, height);
 372         check_error();
 373
 374         // Save these, as width() and height() will lie with DPI scaling.
 375         gl_width = width;
 376         gl_height = height;
 377 }
 378
 379 void JPEGFrameView::paintGL()
 380 {
 381         glViewport(0, 0, gl_width, gl_height);
 382         if (current_frame == nullptr) {
 383                 glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
 384                 glClear(GL_COLOR_BUFFER_BIT);
 385                 return;
 386         }
 387
 388         check_error();
 389         current_chain->render_to_screen();
 390
 391         if (overlay_image != nullptr) {
 392                 if (overlay_input_needs_refresh) {
 393                         overlay_input->set_width(overlay_width);
 394                         overlay_input->set_height(overlay_height);
 395                         overlay_input->set_pixel_data(overlay_image->bits());
 396                 }
 397                 glViewport(gl_width - overlay_width, 0, overlay_width, overlay_height);
 398                 overlay_chain->render_to_screen();
 399         }
 400 }
 401
 402 namespace {
 403
 404 }  // namespace
 405
 406 void JPEGFrameView::setDecodedFrame(shared_ptr<Frame> frame, shared_ptr<Frame> secondary_frame, float fade_alpha)
 407 {
 408         post_to_main_thread([this, frame, secondary_frame, fade_alpha] {
 409                 current_frame = frame;
 410                 current_secondary_frame = secondary_frame;
 411
 412                 if (secondary_frame != nullptr) {
 413                         current_chain = ycbcr_converter->prepare_chain_for_fade(frame, secondary_frame, fade_alpha);
 414                 } else {
 415                         current_chain = ycbcr_converter->prepare_chain_for_conversion(frame);
 416                 }
 417                 update();
 418         });
 419 }
 420
 421 void JPEGFrameView::mousePressEvent(QMouseEvent *event)
 422 {
 423         if (event->type() == QEvent::MouseButtonPress && event->button() == Qt::LeftButton) {
 424                 emit clicked();
 425         }
 426 }
 427
 428 void JPEGFrameView::set_overlay(const string &text)
 429 {
 430         if (text.empty()) {
 431                 overlay_image.reset();
 432                 return;
 433         }
 434
 435         float dpr = QGuiApplication::primaryScreen()->devicePixelRatio();
 436         overlay_width = lrint(overlay_base_width * dpr);
 437         overlay_height = lrint(overlay_base_height * dpr);
 438
 439         overlay_image.reset(new QImage(overlay_width, overlay_height, QImage::Format_Grayscale8));
 440         overlay_image->setDevicePixelRatio(dpr);
 441         overlay_image->fill(0);
 442         QPainter painter(overlay_image.get());
 443
 444         painter.setPen(Qt::white);
 445         QFont font = painter.font();
 446         font.setPointSize(12);
 447         painter.setFont(font);
 448
 449         painter.drawText(QRectF(0, 0, overlay_base_width, overlay_base_height), Qt::AlignCenter, QString::fromStdString(text));
 450
 451         // Don't refresh immediately; we might not have an OpenGL context here.
 452         overlay_input_needs_refresh = true;
 453 }