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

   1 #include "jpeg_frame_view.h"
   2
   3 #include <jpeglib.h>
   4 #include <stdint.h>
   5 #include <unistd.h>
   6
   7 #include <atomic>
   8 #include <condition_variable>
   9 #include <deque>
  10 #include <mutex>
  11 #include <thread>
  12 #include <utility>
  13
  14 #include <QMouseEvent>
  15
  16 #include <movit/resource_pool.h>
  17 #include <movit/init.h>
  18 #include <movit/util.h>
  19
  20 #include "defs.h"
  21 #include "post_to_main_thread.h"
  22 #include "video_stream.h"
  23
  24 using namespace movit;
  25 using namespace std;
  26
  27 // Just an arbitrary order for std::map.
  28 struct JPEGIDLexicalOrder
  29 {
  30         bool operator() (const JPEGID &a, const JPEGID &b) const
  31         {
  32                 if (a.stream_idx != b.stream_idx)
  33                         return a.stream_idx < b.stream_idx;
  34                 if (a.pts != b.pts)
  35                         return a.pts < b.pts;
  36                 return a.interpolated < b.interpolated;
  37         }
  38 };
  39
  40 struct LRUFrame {
  41         shared_ptr<Frame> frame;
  42         size_t last_used;
  43 };
  44
  45 mutex cache_mu;
  46 map<JPEGID, LRUFrame, JPEGIDLexicalOrder> cache;  // Under cache_mu.
  47 condition_variable any_pending_decodes, cache_updated;
  48 deque<pair<JPEGID, JPEGFrameView *>> pending_decodes;  // Under cache_mu.
  49 atomic<size_t> event_counter{0};
  50 extern QGLWidget *global_share_widget;
  51
  52 // TODO: Decode using VA-API if available.
  53 shared_ptr<Frame> decode_jpeg(const string &filename)
  54 {
  55         shared_ptr<Frame> frame(new Frame);
  56
  57         jpeg_decompress_struct dinfo;
  58         jpeg_error_mgr jerr;
  59         dinfo.err = jpeg_std_error(&jerr);
  60         jpeg_create_decompress(&dinfo);
  61
  62         FILE *fp = fopen(filename.c_str(), "rb");
  63         if (fp == nullptr) {
  64                 perror(filename.c_str());
  65                 exit(1);
  66         }
  67         jpeg_stdio_src(&dinfo, fp);
  68
  69         jpeg_read_header(&dinfo, true);
  70
  71         if (dinfo.num_components != 3) {
  72                 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
  73                         dinfo.num_components,
  74                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
  75                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
  76                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
  77                 exit(1);
  78         }
  79         if (dinfo.comp_info[0].h_samp_factor != dinfo.max_h_samp_factor ||
  80             dinfo.comp_info[0].v_samp_factor != dinfo.max_v_samp_factor ||  // Y' must not be subsampled.
  81             dinfo.comp_info[1].h_samp_factor != dinfo.comp_info[2].h_samp_factor ||
  82             dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[2].v_samp_factor ||  // Cb and Cr must be identically subsampled.
  83             (dinfo.max_h_samp_factor % dinfo.comp_info[1].h_samp_factor) != 0 ||
  84             (dinfo.max_v_samp_factor % dinfo.comp_info[1].v_samp_factor) != 0) {  // No 2:3 subsampling or other weirdness.
  85                 fprintf(stderr, "Unsupported subsampling scheme. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
  86                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
  87                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
  88                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
  89                 exit(1);
  90         }
  91         dinfo.raw_data_out = true;
  92
  93         jpeg_start_decompress(&dinfo);
  94
  95         frame->width = dinfo.output_width;
  96         frame->height = dinfo.output_height;
  97         frame->chroma_subsampling_x = dinfo.max_h_samp_factor / dinfo.comp_info[1].h_samp_factor;
  98         frame->chroma_subsampling_y = dinfo.max_v_samp_factor / dinfo.comp_info[1].v_samp_factor;
  99
 100         unsigned h_mcu_size = DCTSIZE * dinfo.max_h_samp_factor;
 101         unsigned v_mcu_size = DCTSIZE * dinfo.max_v_samp_factor;
 102         unsigned mcu_width_blocks = (dinfo.output_width + h_mcu_size - 1) / h_mcu_size;
 103         unsigned mcu_height_blocks = (dinfo.output_height + v_mcu_size - 1) / v_mcu_size;
 104
 105         unsigned luma_width_blocks = mcu_width_blocks * dinfo.comp_info[0].h_samp_factor;
 106         unsigned chroma_width_blocks = mcu_width_blocks * dinfo.comp_info[1].h_samp_factor;
 107         unsigned luma_height_blocks = mcu_height_blocks * dinfo.comp_info[0].v_samp_factor;
 108         unsigned chroma_height_blocks = mcu_height_blocks * dinfo.comp_info[1].v_samp_factor;
 109
 110         // TODO: Decode into a PBO.
 111         frame->y.reset(new uint8_t[luma_width_blocks * luma_height_blocks * DCTSIZE2]);
 112         frame->cb.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 113         frame->cr.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 114         frame->pitch_y = luma_width_blocks * DCTSIZE;
 115         frame->pitch_chroma = chroma_width_blocks * DCTSIZE;
 116
 117         JSAMPROW yptr[v_mcu_size], cbptr[v_mcu_size], crptr[v_mcu_size];
 118         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 119         for (unsigned y = 0; y < mcu_height_blocks; ++y) {
 120                 // NOTE: The last elements of cbptr/crptr will be unused for vertically subsampled chroma.
 121                 for (unsigned yy = 0; yy < v_mcu_size; ++yy) {
 122                         yptr[yy] = frame->y.get() + (y * DCTSIZE * dinfo.max_v_samp_factor + yy) * frame->pitch_y;
 123                         cbptr[yy] = frame->cb.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 124                         crptr[yy] = frame->cr.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 125                 }
 126
 127                 jpeg_read_raw_data(&dinfo, data, v_mcu_size);
 128         }
 129
 130         (void) jpeg_finish_decompress(&dinfo);
 131         jpeg_destroy_decompress(&dinfo);
 132         fclose(fp);
 133
 134         return frame;
 135 }
 136
 137 void prune_cache()
 138 {
 139         // Assumes cache_mu is held.
 140         vector<size_t> lru_timestamps;
 141         for (const auto &key_and_value : cache) {
 142                 lru_timestamps.push_back(key_and_value.second.last_used);
 143         }
 144
 145         size_t cutoff_point = CACHE_SIZE / 10;  // Prune away the 10% oldest ones.
 146         nth_element(lru_timestamps.begin(), lru_timestamps.begin() + cutoff_point, lru_timestamps.end());
 147         size_t must_be_used_after = lru_timestamps[cutoff_point];
 148         for (auto it = cache.begin(); it != cache.end(); ) {
 149                 if (it->second.last_used < must_be_used_after) {
 150                         it = cache.erase(it);
 151                 } else {
 152                         ++it;
 153                 }
 154         }
 155 }
 156
 157 shared_ptr<Frame> decode_jpeg_with_cache(JPEGID id, CacheMissBehavior cache_miss_behavior, bool *did_decode)
 158 {
 159         *did_decode = false;
 160         {
 161                 unique_lock<mutex> lock(cache_mu);
 162                 auto it = cache.find(id);
 163                 if (it != cache.end()) {
 164                         it->second.last_used = event_counter++;
 165                         return it->second.frame;
 166                 }
 167         }
 168
 169         if (cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE) {
 170                 return nullptr;
 171         }
 172
 173         assert(!id.interpolated);
 174         *did_decode = true;
 175         shared_ptr<Frame> frame = decode_jpeg(filename_for_frame(id.stream_idx, id.pts));
 176
 177         unique_lock<mutex> lock(cache_mu);
 178         cache[id] = LRUFrame{ frame, event_counter++ };
 179
 180         if (cache.size() > CACHE_SIZE) {
 181                 prune_cache();
 182         }
 183         return frame;
 184 }
 185
 186 void jpeg_decoder_thread()
 187 {
 188         size_t num_decoded = 0, num_dropped = 0;
 189
 190         pthread_setname_np(pthread_self(), "JPEGDecoder");
 191         for ( ;; ) {
 192                 JPEGID id;
 193                 JPEGFrameView *dest;
 194                 CacheMissBehavior cache_miss_behavior = DECODE_IF_NOT_IN_CACHE;
 195                 {
 196                         unique_lock<mutex> lock(cache_mu);  // TODO: Perhaps under another lock?
 197                         any_pending_decodes.wait(lock, [] {
 198                                 return !pending_decodes.empty();
 199                         });
 200                         id = pending_decodes.front().first;
 201                         dest = pending_decodes.front().second;
 202                         pending_decodes.pop_front();
 203
 204                         size_t num_pending = 0;
 205                         for (const pair<JPEGID, JPEGFrameView *> &decode : pending_decodes) {
 206                                 if (decode.second == dest) {
 207                                         ++num_pending;
 208                                 }
 209                         }
 210                         if (num_pending > 3) {
 211                                 cache_miss_behavior = RETURN_NULLPTR_IF_NOT_IN_CACHE;
 212                         }
 213                 }
 214
 215                 bool found_in_cache;
 216                 shared_ptr<Frame> frame;
 217                 if (id.interpolated) {
 218                         // Interpolated frames are never decoded by us,
 219                         // put directly into the cache from VideoStream.
 220                         unique_lock<mutex> lock(cache_mu);
 221                         cache_updated.wait(lock, [id] {
 222                                 return cache.count(id) != 0;
 223                         });
 224                         found_in_cache = true;  // Don't count it as a decode.
 225
 226                         auto it = cache.find(id);
 227                         assert(it != cache.end());
 228
 229                         it->second.last_used = event_counter++;
 230                         frame = it->second.frame;
 231                         if (frame == nullptr) {
 232                                 // We inserted a nullptr as signal that the frame was never
 233                                 // interpolated and that we should stop waiting.
 234                                 // But don't let it linger in the cache anymore.
 235                                 cache.erase(it);
 236                         }
 237                 } else {
 238                         frame = decode_jpeg_with_cache(id, cache_miss_behavior, &found_in_cache);
 239                 }
 240
 241                 if (frame == nullptr) {
 242                         assert(id.interpolated || cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
 243                         ++num_dropped;
 244                         continue;
 245                 }
 246
 247                 if (!found_in_cache) {
 248                         ++num_decoded;
 249                         if (num_decoded % 1000 == 0) {
 250                                 fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
 251                                         num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
 252                         }
 253                 }
 254
 255                 // TODO: Could we get jitter between non-interpolated and interpolated frames here?
 256                 dest->setDecodedFrame(frame);
 257         }
 258 }
 259
 260 JPEGFrameView::JPEGFrameView(QWidget *parent)
 261         : QGLWidget(parent, global_share_widget) {
 262 }
 263
 264 void JPEGFrameView::setFrame(unsigned stream_idx, int64_t pts, bool interpolated)
 265 {
 266         current_stream_idx = stream_idx;
 267
 268         unique_lock<mutex> lock(cache_mu);
 269         pending_decodes.emplace_back(JPEGID{ stream_idx, pts, interpolated }, this);
 270         any_pending_decodes.notify_all();
 271 }
 272
 273 void JPEGFrameView::insert_interpolated_frame(unsigned stream_idx, int64_t pts, shared_ptr<Frame> frame)
 274 {
 275         JPEGID id{ stream_idx, pts, true };
 276
 277         // We rely on the frame not being evicted from the cache before
 278         // jpeg_decoder_thread() sees it and can display it (otherwise,
 279         // that thread would hang). With a default cache of 1000 elements,
 280         // that would sound like a reasonable assumption.
 281         unique_lock<mutex> lock(cache_mu);
 282         cache[id] = LRUFrame{ std::move(frame), event_counter++ };
 283         cache_updated.notify_all();
 284 }
 285
 286 ResourcePool *resource_pool = nullptr;
 287
 288 void JPEGFrameView::initializeGL()
 289 {
 290         glDisable(GL_BLEND);
 291         glDisable(GL_DEPTH_TEST);
 292         check_error();
 293
 294         static once_flag once;
 295         call_once(once, [] {
 296                 resource_pool = new ResourcePool;
 297                 std::thread(&jpeg_decoder_thread).detach();
 298         });
 299
 300         chain.reset(new EffectChain(1280, 720, resource_pool));
 301         ImageFormat image_format;
 302         image_format.color_space = COLORSPACE_sRGB;
 303         image_format.gamma_curve = GAMMA_sRGB;
 304         ycbcr_format.luma_coefficients = YCBCR_REC_709;
 305         ycbcr_format.full_range = false;
 306         ycbcr_format.num_levels = 256;
 307         ycbcr_format.chroma_subsampling_x = 2;
 308         ycbcr_format.chroma_subsampling_y = 1;
 309         ycbcr_format.cb_x_position = 0.0f;  // H.264 -- _not_ JPEG, even though our input is MJPEG-encoded
 310         ycbcr_format.cb_y_position = 0.5f;  // Irrelevant.
 311         ycbcr_format.cr_x_position = 0.0f;
 312         ycbcr_format.cr_y_position = 0.5f;
 313         ycbcr_input = (movit::YCbCrInput *)chain->add_input(new YCbCrInput(image_format, ycbcr_format, 1280, 720));
 314
 315         ImageFormat inout_format;
 316         inout_format.color_space = COLORSPACE_sRGB;
 317         inout_format.gamma_curve = GAMMA_sRGB;
 318
 319         check_error();
 320         chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 321         check_error();
 322         chain->set_dither_bits(8);
 323         check_error();
 324         chain->finalize();
 325         check_error();
 326
 327         overlay_chain.reset(new EffectChain(overlay_width, overlay_height, resource_pool));
 328         overlay_input = (movit::FlatInput *)overlay_chain->add_input(new FlatInput(image_format, FORMAT_GRAYSCALE, GL_UNSIGNED_BYTE, overlay_width, overlay_height));
 329
 330         overlay_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 331         overlay_chain->finalize();
 332 }
 333
 334 void JPEGFrameView::resizeGL(int width, int height)
 335 {
 336         check_error();
 337         glViewport(0, 0, width, height);
 338         check_error();
 339 }
 340
 341 void JPEGFrameView::paintGL()
 342 {
 343         glViewport(0, 0, width(), height());
 344         if (current_frame == nullptr) {
 345                 glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
 346                 glClear(GL_COLOR_BUFFER_BIT);
 347                 return;
 348         }
 349
 350         check_error();
 351         chain->render_to_screen();
 352
 353         if (overlay_image != nullptr) {
 354                 if (overlay_input_needs_refresh) {
 355                         overlay_input->set_pixel_data(overlay_image->bits());
 356                 }
 357                 glViewport(width() - overlay_width, 0, overlay_width, overlay_height);
 358                 overlay_chain->render_to_screen();
 359         }
 360 }
 361
 362 void JPEGFrameView::setDecodedFrame(std::shared_ptr<Frame> frame)
 363 {
 364         post_to_main_thread([this, frame] {
 365                 current_frame = frame;
 366                 ycbcr_format.chroma_subsampling_x = frame->chroma_subsampling_x;
 367                 ycbcr_format.chroma_subsampling_y = frame->chroma_subsampling_y;
 368                 ycbcr_input->change_ycbcr_format(ycbcr_format);
 369                 ycbcr_input->set_width(frame->width);
 370                 ycbcr_input->set_height(frame->height);
 371                 ycbcr_input->set_pixel_data(0, frame->y.get());
 372                 ycbcr_input->set_pixel_data(1, frame->cb.get());
 373                 ycbcr_input->set_pixel_data(2, frame->cr.get());
 374                 ycbcr_input->set_pitch(0, frame->pitch_y);
 375                 ycbcr_input->set_pitch(1, frame->pitch_chroma);
 376                 ycbcr_input->set_pitch(2, frame->pitch_chroma);
 377                 update();
 378         });
 379 }
 380
 381 void JPEGFrameView::mousePressEvent(QMouseEvent *event)
 382 {
 383         if (event->type() == QEvent::MouseButtonPress && event->button() == Qt::LeftButton) {
 384                 emit clicked();
 385         }
 386 }
 387
 388 void JPEGFrameView::set_overlay(const string &text)
 389 {
 390         if (text.empty()) {
 391                 overlay_image.reset();
 392                 return;
 393         }
 394
 395         overlay_image.reset(new QImage(overlay_width, overlay_height, QImage::Format_Grayscale8));
 396         overlay_image->fill(0);
 397         QPainter painter(overlay_image.get());
 398
 399         painter.setPen(Qt::white);
 400         QFont font = painter.font();
 401         font.setPointSize(12);
 402         painter.setFont(font);
 403
 404         painter.drawText(QRectF(0, 0, overlay_width, overlay_height), Qt::AlignCenter, QString::fromStdString(text));
 405
 406         // Don't refresh immediately; we might not have an OpenGL context here.
 407         overlay_input_needs_refresh = true;
 408 }