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