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