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