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

   1 #include "video_stream.h"
   2
   3 extern "C" {
   4 #include <libavformat/avformat.h>
   5 #include <libavformat/avio.h>
   6 }
   7
   8 #include <jpeglib.h>
   9 #include <unistd.h>
  10
  11 #include "context.h"
  12 #include "flow.h"
  13 #include "httpd.h"
  14 #include "jpeg_frame_view.h"
  15 #include "movit/util.h"
  16 #include "mux.h"
  17 #include "player.h"
  18 #include "util.h"
  19
  20 #include <epoxy/glx.h>
  21
  22 using namespace std;
  23
  24 extern HTTPD *global_httpd;
  25
  26 namespace {
  27
  28 string read_file(const string &filename)
  29 {
  30         FILE *fp = fopen(filename.c_str(), "rb");
  31         if (fp == nullptr) {
  32                 perror(filename.c_str());
  33                 return "";
  34         }
  35
  36         fseek(fp, 0, SEEK_END);
  37         long len = ftell(fp);
  38         rewind(fp);
  39
  40         string ret;
  41         ret.resize(len);
  42         fread(&ret[0], len, 1, fp);
  43         fclose(fp);
  44         return ret;
  45 }
  46
  47 }  // namespace
  48
  49 struct VectorDestinationManager {
  50         jpeg_destination_mgr pub;
  51         std::vector<uint8_t> dest;
  52
  53         VectorDestinationManager()
  54         {
  55                 pub.init_destination = init_destination_thunk;
  56                 pub.empty_output_buffer = empty_output_buffer_thunk;
  57                 pub.term_destination = term_destination_thunk;
  58         }
  59
  60         static void init_destination_thunk(j_compress_ptr ptr)
  61         {
  62                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
  63         }
  64
  65         inline void init_destination()
  66         {
  67                 make_room(0);
  68         }
  69
  70         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
  71         {
  72                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
  73         }
  74
  75         inline bool empty_output_buffer()
  76         {
  77                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
  78                 return true;
  79         }
  80
  81         inline void make_room(size_t bytes_used)
  82         {
  83                 dest.resize(bytes_used + 4096);
  84                 dest.resize(dest.capacity());
  85                 pub.next_output_byte = dest.data() + bytes_used;
  86                 pub.free_in_buffer = dest.size() - bytes_used;
  87         }
  88
  89         static void term_destination_thunk(j_compress_ptr ptr)
  90         {
  91                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
  92         }
  93
  94         inline void term_destination()
  95         {
  96                 dest.resize(dest.size() - pub.free_in_buffer);
  97         }
  98 };
  99 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
 100
 101 vector<uint8_t> encode_jpeg(const uint8_t *y_data, const uint8_t *cbcr_data, unsigned width, unsigned height)
 102 {
 103         VectorDestinationManager dest;
 104
 105         jpeg_compress_struct cinfo;
 106         jpeg_error_mgr jerr;
 107         cinfo.err = jpeg_std_error(&jerr);
 108         jpeg_create_compress(&cinfo);
 109
 110         cinfo.dest = (jpeg_destination_mgr *)&dest;
 111         cinfo.input_components = 3;
 112         cinfo.in_color_space = JCS_RGB;
 113         jpeg_set_defaults(&cinfo);
 114         constexpr int quality = 90;
 115         jpeg_set_quality(&cinfo, quality, /*force_baseline=*/false);
 116
 117         cinfo.image_width = width;
 118         cinfo.image_height = height;
 119         cinfo.raw_data_in = true;
 120         jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 121         cinfo.comp_info[0].h_samp_factor = 2;
 122         cinfo.comp_info[0].v_samp_factor = 1;
 123         cinfo.comp_info[1].h_samp_factor = 1;
 124         cinfo.comp_info[1].v_samp_factor = 1;
 125         cinfo.comp_info[2].h_samp_factor = 1;
 126         cinfo.comp_info[2].v_samp_factor = 1;
 127         cinfo.CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 128         jpeg_start_compress(&cinfo, true);
 129
 130         // TODO: Subsample on the GPU.
 131         unique_ptr<uint8_t[]> cbdata(new uint8_t[(width/2) * 8]);
 132         unique_ptr<uint8_t[]> crdata(new uint8_t[(width/2) * 8]);
 133         JSAMPROW yptr[8], cbptr[8], crptr[8];
 134         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 135         for (unsigned yy = 0; yy < 8; ++yy) {
 136                 cbptr[yy] = cbdata.get() + yy * (width / 2);
 137                 crptr[yy] = crdata.get() + yy * (width / 2);
 138         }
 139         for (unsigned y = 0; y < height; y += 8) {
 140                 uint8_t *cbptr = cbdata.get();
 141                 uint8_t *crptr = crdata.get();
 142                 for (unsigned yy = 0; yy < 8; ++yy) {
 143                         yptr[yy] = const_cast<JSAMPROW>(&y_data[(height - y - yy - 1) * width]);
 144                         const uint8_t *sptr = &cbcr_data[(height - y - yy - 1) * width * 2];
 145                         for (unsigned x = 0; x < width; x += 2) {
 146                                 *cbptr++ = (sptr[0] + sptr[2]) / 2;
 147                                 *crptr++ = (sptr[1] + sptr[3]) / 2;
 148                                 sptr += 4;
 149                         }
 150                 }
 151
 152                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 153         }
 154
 155         jpeg_finish_compress(&cinfo);
 156         jpeg_destroy_compress(&cinfo);
 157
 158         return move(dest.dest);
 159 }
 160
 161 VideoStream::VideoStream()
 162 {
 163         using namespace movit;
 164         // TODO: deduplicate code against JPEGFrameView?
 165         ycbcr_convert_chain.reset(new EffectChain(1280, 720));
 166         ImageFormat image_format;
 167         image_format.color_space = COLORSPACE_sRGB;
 168         image_format.gamma_curve = GAMMA_sRGB;
 169         ycbcr_format.luma_coefficients = YCBCR_REC_709;
 170         ycbcr_format.full_range = true;  // JPEG.
 171         ycbcr_format.num_levels = 256;
 172         ycbcr_format.chroma_subsampling_x = 2;
 173         ycbcr_format.chroma_subsampling_y = 1;
 174         ycbcr_format.cb_x_position = 0.0f;  // H.264 -- _not_ JPEG, even though our input is MJPEG-encoded
 175         ycbcr_format.cb_y_position = 0.5f;  // Irrelevant.
 176         ycbcr_format.cr_x_position = 0.0f;
 177         ycbcr_format.cr_y_position = 0.5f;
 178         ycbcr_input = (movit::YCbCrInput *)ycbcr_convert_chain->add_input(new YCbCrInput(image_format, ycbcr_format, 1280, 720));
 179
 180         YCbCrFormat ycbcr_output_format = ycbcr_format;
 181         ycbcr_output_format.chroma_subsampling_x = 1;
 182
 183         ImageFormat inout_format;
 184         inout_format.color_space = COLORSPACE_sRGB;
 185         inout_format.gamma_curve = GAMMA_sRGB;
 186
 187         check_error();
 188
 189         // One full Y'CbCr texture (for interpolation), one that's just Y (throwing away the
 190         // Cb and Cr channels). The second copy is sort of redundant, but it's the easiest way
 191         // of getting the gray data into a layered texture.
 192         ycbcr_convert_chain->add_ycbcr_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, ycbcr_output_format);
 193         check_error();
 194         ycbcr_convert_chain->add_ycbcr_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, ycbcr_output_format);
 195         check_error();
 196         ycbcr_convert_chain->set_dither_bits(8);
 197         check_error();
 198         ycbcr_convert_chain->finalize();
 199         check_error();
 200
 201         GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots];
 202         glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, input_tex);
 203         glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, gray_tex);
 204         check_error();
 205         constexpr size_t width = 1280, height = 720;  // FIXME: adjustable width, height
 206         int levels = find_num_levels(width, height);
 207         for (size_t i = 0; i < num_interpolate_slots; ++i) {
 208                 glTextureStorage3D(input_tex[i], levels, GL_RGBA8, width, height, 2);
 209                 check_error();
 210                 glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
 211                 check_error();
 212
 213                 InterpolatedFrameResources resource;
 214                 resource.input_tex = input_tex[i];
 215                 resource.gray_tex = gray_tex[i];
 216                 glCreateFramebuffers(2, resource.input_fbos);
 217                 check_error();
 218
 219                 glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
 220                 check_error();
 221                 glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0);
 222                 check_error();
 223                 glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
 224                 check_error();
 225                 glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1);
 226                 check_error();
 227
 228                 GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 229                 glNamedFramebufferDrawBuffers(resource.input_fbos[0], 2, bufs);
 230                 check_error();
 231                 glNamedFramebufferDrawBuffers(resource.input_fbos[1], 2, bufs);
 232                 check_error();
 233
 234                 glCreateBuffers(1, &resource.pbo);
 235                 check_error();
 236                 glNamedBufferStorage(resource.pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 237                 check_error();
 238                 resource.pbo_contents = glMapNamedBufferRange(resource.pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 239                 interpolate_resources.push_back(resource);
 240         }
 241
 242         check_error();
 243
 244         compute_flow.reset(new DISComputeFlow(width, height, operating_point2));
 245         interpolate.reset(new Interpolate(width, height, operating_point2, /*split_ycbcr_output=*/true));
 246         check_error();
 247 }
 248
 249 VideoStream::~VideoStream() {}
 250
 251 void VideoStream::start()
 252 {
 253         AVFormatContext *avctx = avformat_alloc_context();
 254         avctx->oformat = av_guess_format("nut", nullptr, nullptr);
 255
 256         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 257         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 258         avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
 259         avctx->pb->ignore_boundary_point = 1;
 260
 261         Mux::Codec video_codec = Mux::CODEC_MJPEG;
 262
 263         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 264
 265         string video_extradata;
 266
 267         constexpr int width = 1280, height = 720;  // Doesn't matter for MJPEG.
 268         stream_mux.reset(new Mux(avctx, width, height, video_codec, video_extradata, /*audio_codec_parameters=*/nullptr, COARSE_TIMEBASE,
 269                 /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}));
 270
 271
 272         encode_thread = thread(&VideoStream::encode_thread_func, this);
 273 }
 274
 275 void VideoStream::stop()
 276 {
 277         encode_thread.join();
 278 }
 279
 280 void VideoStream::schedule_original_frame(int64_t output_pts, unsigned stream_idx, int64_t input_pts)
 281 {
 282         fprintf(stderr, "output_pts=%ld  original      input_pts=%ld\n", output_pts, input_pts);
 283
 284         QueuedFrame qf;
 285         qf.type = QueuedFrame::ORIGINAL;
 286         qf.output_pts = output_pts;
 287         qf.stream_idx = stream_idx;
 288         qf.input_first_pts = input_pts;
 289
 290         unique_lock<mutex> lock(queue_lock);
 291         frame_queue.push_back(qf);
 292         queue_nonempty.notify_all();
 293 }
 294
 295 void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned stream_idx, int64_t input_first_pts, int64_t input_second_pts, float alpha)
 296 {
 297         fprintf(stderr, "output_pts=%ld  interpolated  input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, input_first_pts, input_second_pts, alpha);
 298
 299         // Get the temporary OpenGL resources we need for doing the interpolation.
 300         InterpolatedFrameResources resources;
 301         {
 302                 unique_lock<mutex> lock(queue_lock);
 303                 if (interpolate_resources.empty()) {
 304                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 305                         return;
 306                 }
 307                 resources = interpolate_resources.front();
 308                 interpolate_resources.pop_front();
 309         }
 310
 311         QueuedFrame qf;
 312         qf.type = QueuedFrame::INTERPOLATED;
 313         qf.output_pts = output_pts;
 314         qf.stream_idx = stream_idx;
 315         qf.resources = resources;
 316
 317         check_error();
 318
 319         // Convert frame0 and frame1 to OpenGL textures.
 320         // TODO: Deduplicate against JPEGFrameView::setDecodedFrame?
 321         for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
 322                 JPEGID jpeg_id;
 323                 jpeg_id.stream_idx = stream_idx;
 324                 jpeg_id.pts = frame_no == 1 ? input_second_pts : input_first_pts;
 325                 bool did_decode;
 326                 shared_ptr<Frame> frame = decode_jpeg_with_cache(jpeg_id, DECODE_IF_NOT_IN_CACHE, &did_decode);
 327                 ycbcr_format.chroma_subsampling_x = frame->chroma_subsampling_x;
 328                 ycbcr_format.chroma_subsampling_y = frame->chroma_subsampling_y;
 329                 ycbcr_input->change_ycbcr_format(ycbcr_format);
 330                 ycbcr_input->set_width(frame->width);
 331                 ycbcr_input->set_height(frame->height);
 332                 ycbcr_input->set_pixel_data(0, frame->y.get());
 333                 ycbcr_input->set_pixel_data(1, frame->cb.get());
 334                 ycbcr_input->set_pixel_data(2, frame->cr.get());
 335                 ycbcr_input->set_pitch(0, frame->pitch_y);
 336                 ycbcr_input->set_pitch(1, frame->pitch_chroma);
 337                 ycbcr_input->set_pitch(2, frame->pitch_chroma);
 338                 ycbcr_convert_chain->render_to_fbo(resources.input_fbos[frame_no], 1280, 720);
 339         }
 340
 341         glGenerateTextureMipmap(resources.input_tex);
 342         check_error();
 343         glGenerateTextureMipmap(resources.gray_tex);
 344         check_error();
 345
 346         // Compute the interpolated frame.
 347         qf.flow_tex = compute_flow->exec(resources.gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 348         check_error();
 349         tie(qf.output_tex, qf.output2_tex) = interpolate->exec(resources.input_tex, resources.gray_tex, qf.flow_tex, 1280, 720, alpha);
 350         check_error();
 351
 352         // We could have released qf.flow_tex here, but to make sure we don't cause a stall
 353         // when trying to reuse it for the next frame, we can just as well hold on to it
 354         // and release it only when the readback is done.
 355
 356         // Read it down (asynchronously) to the CPU.
 357         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 358         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources.pbo);
 359         check_error();
 360         glGetTextureImage(qf.output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
 361         check_error();
 362         glGetTextureImage(qf.output2_tex, 0, GL_RG, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720));
 363         check_error();
 364         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 365
 366         // Set a fence we can wait for to make sure the CPU sees the read.
 367         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 368         check_error();
 369         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 370         check_error();
 371
 372         unique_lock<mutex> lock(queue_lock);
 373         frame_queue.push_back(qf);
 374         queue_nonempty.notify_all();
 375 }
 376
 377 void VideoStream::encode_thread_func()
 378 {
 379         pthread_setname_np(pthread_self(), "VideoStream");
 380         QSurface *surface = create_surface();
 381         QOpenGLContext *context = create_context(surface);
 382         bool ok = make_current(context, surface);
 383         if (!ok) {
 384                 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
 385                 exit(1);
 386         }
 387
 388         for ( ;; ) {
 389                 QueuedFrame qf;
 390                 {
 391                         unique_lock<mutex> lock(queue_lock);
 392                         queue_nonempty.wait(lock, [this]{
 393                                 return !frame_queue.empty();
 394                         });
 395                         qf = frame_queue.front();
 396                         frame_queue.pop_front();
 397                 }
 398
 399                 if (qf.type == QueuedFrame::ORIGINAL) {
 400                         // Send the JPEG frame on, unchanged.
 401                         string jpeg = read_file(filename_for_frame(qf.stream_idx, qf.input_first_pts));
 402                         AVPacket pkt;
 403                         av_init_packet(&pkt);
 404                         pkt.stream_index = 0;
 405                         pkt.data = (uint8_t *)jpeg.data();
 406                         pkt.size = jpeg.size();
 407                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 408                 } else if (qf.type == QueuedFrame::INTERPOLATED) {
 409                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 410
 411                         vector<uint8_t> jpeg = encode_jpeg(
 412                                 (const uint8_t *)qf.resources.pbo_contents,
 413                                 (const uint8_t *)qf.resources.pbo_contents + 1280 * 720,
 414                                 1280, 720);
 415                         compute_flow->release_texture(qf.flow_tex);
 416                         interpolate->release_texture(qf.output_tex);
 417                         interpolate->release_texture(qf.output2_tex);
 418
 419                         AVPacket pkt;
 420                         av_init_packet(&pkt);
 421                         pkt.stream_index = 0;
 422                         pkt.data = (uint8_t *)jpeg.data();
 423                         pkt.size = jpeg.size();
 424                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 425
 426                         // Put the frame resources back.
 427                         unique_lock<mutex> lock(queue_lock);
 428                         interpolate_resources.push_back(qf.resources);
 429                 }
 430         }
 431 }
 432
 433 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 434 {
 435         VideoStream *video_stream = (VideoStream *)opaque;
 436         return video_stream->write_packet2(buf, buf_size, type, time);
 437 }
 438
 439 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 440 {
 441         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
 442                 seen_sync_markers = true;
 443         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
 444                 // We don't know if this is a keyframe or not (the muxer could
 445                 // avoid marking it), so we just have to make the best of it.
 446                 type = AVIO_DATA_MARKER_SYNC_POINT;
 447         }
 448
 449         if (type == AVIO_DATA_MARKER_HEADER) {
 450                 stream_mux_header.append((char *)buf, buf_size);
 451                 global_httpd->set_header(stream_mux_header);
 452         } else {
 453                 global_httpd->add_data((char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
 454         }
 455         return buf_size;
 456 }
 457