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
   6 #include <atomic>
   7 #include <condition_variable>
   8 #include <deque>
   9 #include <mutex>
  10 #include <thread>
  11 #include <utility>
  12
  13 #include <movit/resource_pool.h>
  14 #include <movit/init.h>
  15 #include <movit/util.h>
  16
  17 #include "defs.h"
  18 #include "post_to_main_thread.h"
  19
  20 using namespace movit;
  21 using namespace std;
  22
  23 struct JPEGID {
  24         unsigned stream_idx;
  25         int64_t pts;
  26 };
  27 bool operator< (const JPEGID &a, const JPEGID &b) {
  28         return make_pair(a.stream_idx, a.pts) < make_pair(b.stream_idx, b.pts);
  29 }
  30
  31 struct LRUFrame {
  32         shared_ptr<Frame> frame;
  33         size_t last_used;
  34 };
  35
  36 mutex cache_mu;
  37 map<JPEGID, LRUFrame> cache;  // Under cache_mu.
  38 condition_variable any_pending_decodes;
  39 deque<pair<JPEGID, JPEGFrameView *>> pending_decodes;  // Under cache_mu.
  40 atomic<size_t> event_counter{0};
  41 extern QGLWidget *global_share_widget;
  42
  43 // TODO: Decode using VA-API if available.
  44 shared_ptr<Frame> decode_jpeg(const string &filename)
  45 {
  46         shared_ptr<Frame> frame(new Frame);
  47
  48         jpeg_decompress_struct dinfo;
  49         jpeg_error_mgr jerr;
  50         dinfo.err = jpeg_std_error(&jerr);
  51         jpeg_create_decompress(&dinfo);
  52
  53         FILE *fp = fopen(filename.c_str(), "rb");
  54         if (fp == nullptr) {
  55                 perror(filename.c_str());
  56                 exit(1);
  57         }
  58         jpeg_stdio_src(&dinfo, fp);
  59
  60         jpeg_read_header(&dinfo, true);
  61
  62         if (dinfo.num_components != 3) {
  63                 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
  64                         dinfo.num_components,
  65                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
  66                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
  67                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
  68                 exit(1);
  69         }
  70         if (dinfo.comp_info[0].h_samp_factor != dinfo.max_h_samp_factor ||
  71             dinfo.comp_info[0].v_samp_factor != dinfo.max_v_samp_factor ||  // Y' must not be subsampled.
  72             dinfo.comp_info[1].h_samp_factor != dinfo.comp_info[2].h_samp_factor ||
  73             dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[2].v_samp_factor ||  // Cb and Cr must be identically subsampled.
  74             (dinfo.max_h_samp_factor % dinfo.comp_info[1].h_samp_factor) != 0 ||
  75             (dinfo.max_v_samp_factor % dinfo.comp_info[1].v_samp_factor) != 0) {  // No 2:3 subsampling or other weirdness.
  76                 fprintf(stderr, "Unsupported subsampling scheme. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
  77                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
  78                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
  79                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
  80                 exit(1);
  81         }
  82         dinfo.raw_data_out = true;
  83
  84         jpeg_start_decompress(&dinfo);
  85
  86         frame->width = dinfo.output_width;
  87         frame->height = dinfo.output_height;
  88         frame->chroma_subsampling_x = dinfo.max_h_samp_factor / dinfo.comp_info[1].h_samp_factor;
  89         frame->chroma_subsampling_y = dinfo.max_v_samp_factor / dinfo.comp_info[1].v_samp_factor;
  90
  91         unsigned h_mcu_size = DCTSIZE * dinfo.max_h_samp_factor;
  92         unsigned v_mcu_size = DCTSIZE * dinfo.max_v_samp_factor;
  93         unsigned mcu_width_blocks = (dinfo.output_width + h_mcu_size - 1) / h_mcu_size;
  94         unsigned mcu_height_blocks = (dinfo.output_height + v_mcu_size - 1) / v_mcu_size;
  95
  96         unsigned luma_width_blocks = mcu_width_blocks * dinfo.comp_info[0].h_samp_factor;
  97         unsigned chroma_width_blocks = mcu_width_blocks * dinfo.comp_info[1].h_samp_factor;
  98         unsigned luma_height_blocks = mcu_height_blocks * dinfo.comp_info[0].v_samp_factor;
  99         unsigned chroma_height_blocks = mcu_height_blocks * dinfo.comp_info[1].v_samp_factor;
 100
 101         // TODO: Decode into a PBO.
 102         frame->y.reset(new uint8_t[luma_width_blocks * luma_height_blocks * DCTSIZE2]);
 103         frame->cb.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 104         frame->cr.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
 105         frame->pitch_y = luma_width_blocks * DCTSIZE;
 106         frame->pitch_chroma = chroma_width_blocks * DCTSIZE;
 107
 108         JSAMPROW yptr[v_mcu_size], cbptr[v_mcu_size], crptr[v_mcu_size];
 109         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 110         for (unsigned y = 0; y < mcu_height_blocks; ++y) {
 111                 // NOTE: The last elements of cbptr/crptr will be unused for vertically subsampled chroma.
 112                 for (unsigned yy = 0; yy < v_mcu_size; ++yy) {
 113                         yptr[yy] = frame->y.get() + (y * DCTSIZE * dinfo.max_v_samp_factor + yy) * frame->pitch_y;
 114                         cbptr[yy] = frame->cb.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 115                         crptr[yy] = frame->cr.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
 116                 }
 117
 118                 jpeg_read_raw_data(&dinfo, data, v_mcu_size);
 119         }
 120
 121         (void) jpeg_finish_decompress(&dinfo);
 122         jpeg_destroy_decompress(&dinfo);
 123         fclose(fp);
 124
 125         return frame;
 126 }
 127
 128 void prune_cache()
 129 {
 130         // Assumes cache_mu is held.
 131         vector<size_t> lru_timestamps;
 132         for (const auto &key_and_value : cache) {
 133                 lru_timestamps.push_back(key_and_value.second.last_used);
 134         }
 135
 136         size_t cutoff_point = CACHE_SIZE / 10;  // Prune away the 10% oldest ones.
 137         nth_element(lru_timestamps.begin(), lru_timestamps.begin() + cutoff_point, lru_timestamps.end());
 138         size_t must_be_used_after = lru_timestamps[cutoff_point];
 139         for (auto it = cache.begin(); it != cache.end(); ) {
 140                 if (it->second.last_used < must_be_used_after) {
 141                         it = cache.erase(it);
 142                 } else {
 143                         ++it;
 144                 }
 145         }
 146 }
 147
 148 void jpeg_decoder_thread()
 149 {
 150         size_t num_decoded = 0, num_dropped = 0;
 151
 152         pthread_setname_np(pthread_self(), "JPEGDecoder");
 153         for ( ;; ) {
 154                 JPEGID id;
 155                 JPEGFrameView *dest;
 156                 shared_ptr<Frame> frame;
 157                 {
 158                         unique_lock<mutex> lock(cache_mu);
 159                         any_pending_decodes.wait(lock, [] {
 160                                 return !pending_decodes.empty();
 161                         });
 162                         id = pending_decodes.front().first;
 163                         dest = pending_decodes.front().second;
 164                         pending_decodes.pop_front();
 165
 166                         auto it = cache.find(id);
 167                         if (it != cache.end()) {
 168                                 frame = it->second.frame;
 169                                 it->second.last_used = event_counter++;
 170                         }
 171                 }
 172
 173                 if (frame == nullptr) {
 174                         // Not found in the cache, so we need to do a decode or drop the request.
 175                         // Prune the queue if there are too many pending for this destination.
 176                         // TODO: Could we get starvation here?
 177                         size_t num_pending = 0;
 178                         for (const pair<JPEGID, JPEGFrameView *> &decode : pending_decodes) {
 179                                 if (decode.second == dest) {
 180                                         ++num_pending;
 181                                 }
 182                         }
 183                         if (num_pending > 3) {
 184                                 ++num_dropped;
 185                                 continue;
 186                         }
 187
 188                         frame = decode_jpeg(filename_for_frame(id.stream_idx, id.pts));
 189
 190                         unique_lock<mutex> lock(cache_mu);
 191                         cache[id] = LRUFrame{ frame, event_counter++ };
 192
 193                         if (cache.size() > CACHE_SIZE) {
 194                                 prune_cache();
 195                         }
 196                         ++num_decoded;
 197                         if (num_decoded % 1000 == 0) {
 198                                 fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
 199                                         num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
 200                         }
 201                 }
 202
 203                 dest->setDecodedFrame(frame);
 204         }
 205 }
 206
 207 JPEGFrameView::JPEGFrameView(QWidget *parent)
 208         : QGLWidget(parent, global_share_widget) {
 209 }
 210
 211 void JPEGFrameView::update_frame()
 212 {
 213         unique_lock<mutex> lock(cache_mu);
 214         pending_decodes.emplace_back(JPEGID{ stream_idx, pts }, this);
 215         any_pending_decodes.notify_all();
 216 }
 217
 218 ResourcePool *resource_pool = nullptr;
 219
 220 void JPEGFrameView::initializeGL()
 221 {
 222         glDisable(GL_BLEND);
 223         glDisable(GL_DEPTH_TEST);
 224         glDepthMask(GL_FALSE);
 225         check_error();
 226
 227         static once_flag once;
 228         call_once(once, [] {
 229                 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 230                 resource_pool = new ResourcePool;
 231
 232                 std::thread(&jpeg_decoder_thread).detach();
 233         });
 234
 235         chain.reset(new EffectChain(1280, 720, resource_pool));
 236         ImageFormat image_format;
 237         image_format.color_space = COLORSPACE_sRGB;
 238         image_format.gamma_curve = GAMMA_sRGB;
 239         ycbcr_format.luma_coefficients = YCBCR_REC_709;
 240         ycbcr_format.full_range = false;
 241         ycbcr_format.num_levels = 256;
 242         ycbcr_format.chroma_subsampling_x = 2;
 243         ycbcr_format.chroma_subsampling_y = 1;
 244         ycbcr_format.cb_x_position = 0.0f;  // H.264 -- _not_ JPEG, even though our input is MJPEG-encoded
 245         ycbcr_format.cb_y_position = 0.5f;  // Irrelevant.
 246         ycbcr_format.cr_x_position = 0.0f;
 247         ycbcr_format.cr_y_position = 0.5f;
 248         ycbcr_input = (movit::YCbCrInput *)chain->add_input(new YCbCrInput(image_format, ycbcr_format, 1280, 720));
 249
 250         ImageFormat inout_format;
 251         inout_format.color_space = COLORSPACE_sRGB;
 252         inout_format.gamma_curve = GAMMA_sRGB;
 253
 254         check_error();
 255         chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 256         check_error();
 257         chain->set_dither_bits(8);
 258         check_error();
 259         chain->finalize();
 260         check_error();
 261 }
 262
 263 void JPEGFrameView::resizeGL(int width, int height)
 264 {
 265         check_error();
 266         glViewport(0, 0, width, height);
 267         check_error();
 268 }
 269
 270 void JPEGFrameView::paintGL()
 271 {
 272         if (current_frame == nullptr) {
 273                 glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
 274                 glClear(GL_COLOR_BUFFER_BIT);
 275                 return;
 276         }
 277
 278         check_error();
 279         chain->render_to_screen();
 280 }
 281
 282 void JPEGFrameView::setDecodedFrame(std::shared_ptr<Frame> frame)
 283 {
 284         post_to_main_thread([this, frame] {
 285                 current_frame = frame;
 286                 ycbcr_format.chroma_subsampling_x = frame->chroma_subsampling_x;
 287                 ycbcr_format.chroma_subsampling_y = frame->chroma_subsampling_y;
 288                 ycbcr_input->change_ycbcr_format(ycbcr_format);
 289                 ycbcr_input->set_width(frame->width);
 290                 ycbcr_input->set_height(frame->height);
 291                 ycbcr_input->set_pixel_data(0, frame->y.get());
 292                 ycbcr_input->set_pixel_data(1, frame->cb.get());
 293                 ycbcr_input->set_pixel_data(2, frame->cr.get());
 294                 ycbcr_input->set_pitch(0, frame->pitch_y);
 295                 ycbcr_input->set_pitch(1, frame->pitch_chroma);
 296                 ycbcr_input->set_pitch(2, frame->pitch_chroma);
 297                 update();
 298         });
 299 }