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 "chroma_subsampler.h"
  12 #include "context.h"
  13 #include "flow.h"
  14 #include "httpd.h"
  15 #include "jpeg_frame_view.h"
  16 #include "movit/util.h"
  17 #include "mux.h"
  18 #include "player.h"
  19 #include "util.h"
  20
  21 #include <epoxy/glx.h>
  22
  23 using namespace std;
  24
  25 extern HTTPD *global_httpd;
  26
  27 namespace {
  28
  29 string read_file(const string &filename)
  30 {
  31         FILE *fp = fopen(filename.c_str(), "rb");
  32         if (fp == nullptr) {
  33                 perror(filename.c_str());
  34                 return "";
  35         }
  36
  37         fseek(fp, 0, SEEK_END);
  38         long len = ftell(fp);
  39         rewind(fp);
  40
  41         string ret;
  42         ret.resize(len);
  43         fread(&ret[0], len, 1, fp);
  44         fclose(fp);
  45         return ret;
  46 }
  47
  48 }  // namespace
  49
  50 struct VectorDestinationManager {
  51         jpeg_destination_mgr pub;
  52         std::vector<uint8_t> dest;
  53
  54         VectorDestinationManager()
  55         {
  56                 pub.init_destination = init_destination_thunk;
  57                 pub.empty_output_buffer = empty_output_buffer_thunk;
  58                 pub.term_destination = term_destination_thunk;
  59         }
  60
  61         static void init_destination_thunk(j_compress_ptr ptr)
  62         {
  63                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
  64         }
  65
  66         inline void init_destination()
  67         {
  68                 make_room(0);
  69         }
  70
  71         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
  72         {
  73                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
  74         }
  75
  76         inline bool empty_output_buffer()
  77         {
  78                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
  79                 return true;
  80         }
  81
  82         inline void make_room(size_t bytes_used)
  83         {
  84                 dest.resize(bytes_used + 4096);
  85                 dest.resize(dest.capacity());
  86                 pub.next_output_byte = dest.data() + bytes_used;
  87                 pub.free_in_buffer = dest.size() - bytes_used;
  88         }
  89
  90         static void term_destination_thunk(j_compress_ptr ptr)
  91         {
  92                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
  93         }
  94
  95         inline void term_destination()
  96         {
  97                 dest.resize(dest.size() - pub.free_in_buffer);
  98         }
  99 };
 100 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
 101
 102 vector<uint8_t> encode_jpeg(const uint8_t *y_data, const uint8_t *cb_data, const uint8_t *cr_data, unsigned width, unsigned height)
 103 {
 104         VectorDestinationManager dest;
 105
 106         jpeg_compress_struct cinfo;
 107         jpeg_error_mgr jerr;
 108         cinfo.err = jpeg_std_error(&jerr);
 109         jpeg_create_compress(&cinfo);
 110
 111         cinfo.dest = (jpeg_destination_mgr *)&dest;
 112         cinfo.input_components = 3;
 113         cinfo.in_color_space = JCS_RGB;
 114         jpeg_set_defaults(&cinfo);
 115         constexpr int quality = 90;
 116         jpeg_set_quality(&cinfo, quality, /*force_baseline=*/false);
 117
 118         cinfo.image_width = width;
 119         cinfo.image_height = height;
 120         cinfo.raw_data_in = true;
 121         jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 122         cinfo.comp_info[0].h_samp_factor = 2;
 123         cinfo.comp_info[0].v_samp_factor = 1;
 124         cinfo.comp_info[1].h_samp_factor = 1;
 125         cinfo.comp_info[1].v_samp_factor = 1;
 126         cinfo.comp_info[2].h_samp_factor = 1;
 127         cinfo.comp_info[2].v_samp_factor = 1;
 128         cinfo.CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 129         jpeg_start_compress(&cinfo, true);
 130
 131         JSAMPROW yptr[8], cbptr[8], crptr[8];
 132         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 133         for (unsigned y = 0; y < height; y += 8) {
 134                 for (unsigned yy = 0; yy < 8; ++yy) {
 135                         yptr[yy] = const_cast<JSAMPROW>(&y_data[(height - y - yy - 1) * width]);
 136                         cbptr[yy] = const_cast<JSAMPROW>(&cb_data[(height - y - yy - 1) * width/2]);
 137                         crptr[yy] = const_cast<JSAMPROW>(&cr_data[(height - y - yy - 1) * width/2]);
 138                 }
 139
 140                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 141         }
 142
 143         jpeg_finish_compress(&cinfo);
 144         jpeg_destroy_compress(&cinfo);
 145
 146         return move(dest.dest);
 147 }
 148
 149 VideoStream::VideoStream()
 150 {
 151         using namespace movit;
 152         // TODO: deduplicate code against JPEGFrameView?
 153         ycbcr_convert_chain.reset(new EffectChain(1280, 720));
 154         ImageFormat image_format;
 155         image_format.color_space = COLORSPACE_sRGB;
 156         image_format.gamma_curve = GAMMA_sRGB;
 157         ycbcr_format.luma_coefficients = YCBCR_REC_709;
 158         ycbcr_format.full_range = true;  // JPEG.
 159         ycbcr_format.num_levels = 256;
 160         ycbcr_format.chroma_subsampling_x = 2;
 161         ycbcr_format.chroma_subsampling_y = 1;
 162         ycbcr_format.cb_x_position = 0.0f;  // H.264 -- _not_ JPEG, even though our input is MJPEG-encoded
 163         ycbcr_format.cb_y_position = 0.5f;  // Irrelevant.
 164         ycbcr_format.cr_x_position = 0.0f;
 165         ycbcr_format.cr_y_position = 0.5f;
 166         ycbcr_input = (movit::YCbCrInput *)ycbcr_convert_chain->add_input(new YCbCrInput(image_format, ycbcr_format, 1280, 720));
 167
 168         YCbCrFormat ycbcr_output_format = ycbcr_format;
 169         ycbcr_output_format.chroma_subsampling_x = 1;
 170
 171         ImageFormat inout_format;
 172         inout_format.color_space = COLORSPACE_sRGB;
 173         inout_format.gamma_curve = GAMMA_sRGB;
 174
 175         check_error();
 176
 177         // One full Y'CbCr texture (for interpolation), one that's just Y (throwing away the
 178         // Cb and Cr channels). The second copy is sort of redundant, but it's the easiest way
 179         // of getting the gray data into a layered texture.
 180         ycbcr_convert_chain->add_ycbcr_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, ycbcr_output_format);
 181         check_error();
 182         ycbcr_convert_chain->add_ycbcr_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, ycbcr_output_format);
 183         check_error();
 184         ycbcr_convert_chain->set_dither_bits(8);
 185         check_error();
 186         ycbcr_convert_chain->finalize();
 187         check_error();
 188
 189         GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots], cb_tex[num_interpolate_slots], cr_tex[num_interpolate_slots];
 190         glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, input_tex);
 191         glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, gray_tex);
 192         glCreateTextures(GL_TEXTURE_2D, 10, cb_tex);
 193         glCreateTextures(GL_TEXTURE_2D, 10, cr_tex);
 194         check_error();
 195         constexpr size_t width = 1280, height = 720;  // FIXME: adjustable width, height
 196         int levels = find_num_levels(width, height);
 197         for (size_t i = 0; i < num_interpolate_slots; ++i) {
 198                 glTextureStorage3D(input_tex[i], levels, GL_RGBA8, width, height, 2);
 199                 check_error();
 200                 glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
 201                 check_error();
 202                 glTextureStorage2D(cb_tex[i], 1, GL_R8, width / 2, height);
 203                 check_error();
 204                 glTextureStorage2D(cr_tex[i], 1, GL_R8, width / 2, height);
 205                 check_error();
 206
 207                 InterpolatedFrameResources resource;
 208                 resource.input_tex = input_tex[i];
 209                 resource.gray_tex = gray_tex[i];
 210                 resource.cb_tex = cb_tex[i];
 211                 resource.cr_tex = cr_tex[i];
 212                 glCreateFramebuffers(2, resource.input_fbos);
 213                 check_error();
 214
 215                 glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
 216                 check_error();
 217                 glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0);
 218                 check_error();
 219                 glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
 220                 check_error();
 221                 glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1);
 222                 check_error();
 223
 224                 GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 225                 glNamedFramebufferDrawBuffers(resource.input_fbos[0], 2, bufs);
 226                 check_error();
 227                 glNamedFramebufferDrawBuffers(resource.input_fbos[1], 2, bufs);
 228                 check_error();
 229
 230                 glCreateBuffers(1, &resource.pbo);
 231                 check_error();
 232                 glNamedBufferStorage(resource.pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 233                 check_error();
 234                 resource.pbo_contents = glMapNamedBufferRange(resource.pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 235                 interpolate_resources.push_back(resource);
 236         }
 237
 238         check_error();
 239
 240         compute_flow.reset(new DISComputeFlow(width, height, operating_point2));
 241         interpolate.reset(new Interpolate(width, height, operating_point2, /*split_ycbcr_output=*/true));
 242         chroma_subsampler.reset(new ChromaSubsampler);
 243         check_error();
 244 }
 245
 246 VideoStream::~VideoStream() {}
 247
 248 void VideoStream::start()
 249 {
 250         AVFormatContext *avctx = avformat_alloc_context();
 251         avctx->oformat = av_guess_format("nut", nullptr, nullptr);
 252
 253         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 254         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 255         avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
 256         avctx->pb->ignore_boundary_point = 1;
 257
 258         Mux::Codec video_codec = Mux::CODEC_MJPEG;
 259
 260         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 261
 262         string video_extradata;
 263
 264         constexpr int width = 1280, height = 720;  // Doesn't matter for MJPEG.
 265         stream_mux.reset(new Mux(avctx, width, height, video_codec, video_extradata, /*audio_codec_parameters=*/nullptr, COARSE_TIMEBASE,
 266                 /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}));
 267
 268
 269         encode_thread = thread(&VideoStream::encode_thread_func, this);
 270 }
 271
 272 void VideoStream::stop()
 273 {
 274         encode_thread.join();
 275 }
 276
 277 void VideoStream::schedule_original_frame(int64_t output_pts, unsigned stream_idx, int64_t input_pts)
 278 {
 279         fprintf(stderr, "output_pts=%ld  original      input_pts=%ld\n", output_pts, input_pts);
 280
 281         QueuedFrame qf;
 282         qf.type = QueuedFrame::ORIGINAL;
 283         qf.output_pts = output_pts;
 284         qf.stream_idx = stream_idx;
 285         qf.input_first_pts = input_pts;
 286
 287         unique_lock<mutex> lock(queue_lock);
 288         frame_queue.push_back(qf);
 289         queue_nonempty.notify_all();
 290 }
 291
 292 void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned stream_idx, int64_t input_first_pts, int64_t input_second_pts, float alpha)
 293 {
 294         fprintf(stderr, "output_pts=%ld  interpolated  input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, input_first_pts, input_second_pts, alpha);
 295
 296         // Get the temporary OpenGL resources we need for doing the interpolation.
 297         InterpolatedFrameResources resources;
 298         {
 299                 unique_lock<mutex> lock(queue_lock);
 300                 if (interpolate_resources.empty()) {
 301                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 302                         JPEGFrameView::insert_interpolated_frame(stream_idx, output_pts, nullptr);
 303                         return;
 304                 }
 305                 resources = interpolate_resources.front();
 306                 interpolate_resources.pop_front();
 307         }
 308
 309         QueuedFrame qf;
 310         qf.type = QueuedFrame::INTERPOLATED;
 311         qf.output_pts = output_pts;
 312         qf.stream_idx = stream_idx;
 313         qf.resources = resources;
 314
 315         check_error();
 316
 317         // Convert frame0 and frame1 to OpenGL textures.
 318         // TODO: Deduplicate against JPEGFrameView::setDecodedFrame?
 319         for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
 320                 JPEGID jpeg_id;
 321                 jpeg_id.stream_idx = stream_idx;
 322                 jpeg_id.pts = frame_no == 1 ? input_second_pts : input_first_pts;
 323                 jpeg_id.interpolated = false;
 324                 bool did_decode;
 325                 shared_ptr<Frame> frame = decode_jpeg_with_cache(jpeg_id, DECODE_IF_NOT_IN_CACHE, &did_decode);
 326                 ycbcr_format.chroma_subsampling_x = frame->chroma_subsampling_x;
 327                 ycbcr_format.chroma_subsampling_y = frame->chroma_subsampling_y;
 328                 ycbcr_input->change_ycbcr_format(ycbcr_format);
 329                 ycbcr_input->set_width(frame->width);
 330                 ycbcr_input->set_height(frame->height);
 331                 ycbcr_input->set_pixel_data(0, frame->y.get());
 332                 ycbcr_input->set_pixel_data(1, frame->cb.get());
 333                 ycbcr_input->set_pixel_data(2, frame->cr.get());
 334                 ycbcr_input->set_pitch(0, frame->pitch_y);
 335                 ycbcr_input->set_pitch(1, frame->pitch_chroma);
 336                 ycbcr_input->set_pitch(2, frame->pitch_chroma);
 337                 ycbcr_convert_chain->render_to_fbo(resources.input_fbos[frame_no], 1280, 720);
 338         }
 339
 340         glGenerateTextureMipmap(resources.input_tex);
 341         check_error();
 342         glGenerateTextureMipmap(resources.gray_tex);
 343         check_error();
 344
 345         // Compute the interpolated frame.
 346         qf.flow_tex = compute_flow->exec(resources.gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 347         check_error();
 348         tie(qf.output_tex, qf.cbcr_tex) = interpolate->exec(resources.input_tex, resources.gray_tex, qf.flow_tex, 1280, 720, alpha);
 349         check_error();
 350
 351         // Subsample and split Cb/Cr.
 352         chroma_subsampler->subsample_chroma(qf.cbcr_tex, 1280, 720, resources.cb_tex, resources.cr_tex);
 353
 354         // We could have released qf.flow_tex here, but to make sure we don't cause a stall
 355         // when trying to reuse it for the next frame, we can just as well hold on to it
 356         // and release it only when the readback is done.
 357
 358         // Read it down (asynchronously) to the CPU.
 359         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 360         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources.pbo);
 361         check_error();
 362         glGetTextureImage(qf.output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
 363         check_error();
 364         glGetTextureImage(resources.cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720));
 365         check_error();
 366         glGetTextureImage(resources.cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3 - 640 * 720, BUFFER_OFFSET(1280 * 720 + 640 * 720));
 367         check_error();
 368         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 369
 370         // Set a fence we can wait for to make sure the CPU sees the read.
 371         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 372         check_error();
 373         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 374         check_error();
 375
 376         unique_lock<mutex> lock(queue_lock);
 377         frame_queue.push_back(qf);
 378         queue_nonempty.notify_all();
 379 }
 380
 381 void VideoStream::encode_thread_func()
 382 {
 383         pthread_setname_np(pthread_self(), "VideoStream");
 384         QSurface *surface = create_surface();
 385         QOpenGLContext *context = create_context(surface);
 386         bool ok = make_current(context, surface);
 387         if (!ok) {
 388                 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
 389                 exit(1);
 390         }
 391
 392         for ( ;; ) {
 393                 QueuedFrame qf;
 394                 {
 395                         unique_lock<mutex> lock(queue_lock);
 396                         queue_nonempty.wait(lock, [this]{
 397                                 return !frame_queue.empty();
 398                         });
 399                         qf = frame_queue.front();
 400                         frame_queue.pop_front();
 401                 }
 402
 403                 if (qf.type == QueuedFrame::ORIGINAL) {
 404                         // Send the JPEG frame on, unchanged.
 405                         string jpeg = read_file(filename_for_frame(qf.stream_idx, qf.input_first_pts));
 406                         AVPacket pkt;
 407                         av_init_packet(&pkt);
 408                         pkt.stream_index = 0;
 409                         pkt.data = (uint8_t *)jpeg.data();
 410                         pkt.size = jpeg.size();
 411                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 412                 } else if (qf.type == QueuedFrame::INTERPOLATED) {
 413                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 414
 415                         const uint8_t *y = (const uint8_t *)qf.resources.pbo_contents;
 416                         const uint8_t *cb = (const uint8_t *)qf.resources.pbo_contents + 1280 * 720;
 417                         const uint8_t *cr = (const uint8_t *)qf.resources.pbo_contents + 1280 * 720 + 640 * 720;
 418
 419                         // Send a copy of the frame on to display.
 420                         shared_ptr<Frame> frame(new Frame);
 421                         frame->y.reset(new uint8_t[1280 * 720]);
 422                         frame->cb.reset(new uint8_t[640 * 720]);
 423                         frame->cr.reset(new uint8_t[640 * 720]);
 424                         for (unsigned yy = 0; yy < 720; ++yy) {
 425                                 memcpy(frame->y.get() + 1280 * yy, y + 1280 * (719 - yy), 1280);
 426                                 memcpy(frame->cb.get() + 640 * yy, cb + 640 * (719 - yy), 640);
 427                                 memcpy(frame->cr.get() + 640 * yy, cr + 640 * (719 - yy), 640);
 428                         }
 429                         frame->width = 1280;
 430                         frame->height = 720;
 431                         frame->chroma_subsampling_x = 2;
 432                         frame->chroma_subsampling_y = 1;
 433                         frame->pitch_y = 1280;
 434                         frame->pitch_chroma = 640;
 435                         JPEGFrameView::insert_interpolated_frame(qf.stream_idx, qf.output_pts, std::move(frame));
 436
 437                         // Now JPEG encode it, and send it on to the stream.
 438                         vector<uint8_t> jpeg = encode_jpeg(y, cb, cr, 1280, 720);
 439                         compute_flow->release_texture(qf.flow_tex);
 440                         interpolate->release_texture(qf.output_tex);
 441                         interpolate->release_texture(qf.cbcr_tex);
 442
 443                         AVPacket pkt;
 444                         av_init_packet(&pkt);
 445                         pkt.stream_index = 0;
 446                         pkt.data = (uint8_t *)jpeg.data();
 447                         pkt.size = jpeg.size();
 448                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 449
 450                         // Put the frame resources back.
 451                         unique_lock<mutex> lock(queue_lock);
 452                         interpolate_resources.push_back(qf.resources);
 453                 }
 454         }
 455 }
 456
 457 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 458 {
 459         VideoStream *video_stream = (VideoStream *)opaque;
 460         return video_stream->write_packet2(buf, buf_size, type, time);
 461 }
 462
 463 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 464 {
 465         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
 466                 seen_sync_markers = true;
 467         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
 468                 // We don't know if this is a keyframe or not (the muxer could
 469                 // avoid marking it), so we just have to make the best of it.
 470                 type = AVIO_DATA_MARKER_SYNC_POINT;
 471         }
 472
 473         if (type == AVIO_DATA_MARKER_HEADER) {
 474                 stream_mux_header.append((char *)buf, buf_size);
 475                 global_httpd->set_header(stream_mux_header);
 476         } else {
 477                 global_httpd->add_data((char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
 478         }
 479         return buf_size;
 480 }
 481