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 *pixel_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 and deinterleave on the GPU.
 131
 132         unique_ptr<uint8_t[]> ydata(new uint8_t[width * 8]);
 133         unique_ptr<uint8_t[]> cbdata(new uint8_t[(width/2) * 8]);
 134         unique_ptr<uint8_t[]> crdata(new uint8_t[(width/2) * 8]);
 135         JSAMPROW yptr[8], cbptr[8], crptr[8];
 136         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 137         for (unsigned yy = 0; yy < 8; ++yy) {
 138                 yptr[yy] = ydata.get() + yy * width;
 139                 cbptr[yy] = cbdata.get() + yy * (width / 2);
 140                 crptr[yy] = crdata.get() + yy * (width / 2);
 141         }
 142         for (unsigned y = 0; y < height; y += 8) {
 143                 uint8_t *yptr = ydata.get();
 144                 uint8_t *cbptr = cbdata.get();
 145                 uint8_t *crptr = crdata.get();
 146                 for (unsigned yy = 0; yy < 8; ++yy) {
 147                         const uint8_t *sptr = &pixel_data[(height - y - yy - 1) * width * 4];
 148                         for (unsigned x = 0; x < width; x += 2) {
 149                                 *yptr++ = sptr[0];
 150                                 *yptr++ = sptr[4];
 151                                 *cbptr++ = (sptr[1] + sptr[5]) / 2;
 152                                 *crptr++ = (sptr[2] + sptr[6]) / 2;
 153                                 sptr += 8;
 154                         }
 155                 }
 156
 157                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 158         }
 159
 160         jpeg_finish_compress(&cinfo);
 161         jpeg_destroy_compress(&cinfo);
 162
 163         return move(dest.dest);
 164 }
 165
 166 VideoStream::VideoStream()
 167 {
 168         using namespace movit;
 169         // TODO: deduplicate code against JPEGFrameView?
 170         ycbcr_convert_chain.reset(new EffectChain(1280, 720));
 171         ImageFormat image_format;
 172         image_format.color_space = COLORSPACE_sRGB;
 173         image_format.gamma_curve = GAMMA_sRGB;
 174         ycbcr_format.luma_coefficients = YCBCR_REC_709;
 175         ycbcr_format.full_range = true;  // JPEG.
 176         ycbcr_format.num_levels = 256;
 177         ycbcr_format.chroma_subsampling_x = 2;
 178         ycbcr_format.chroma_subsampling_y = 1;
 179         ycbcr_format.cb_x_position = 0.0f;  // H.264 -- _not_ JPEG, even though our input is MJPEG-encoded
 180         ycbcr_format.cb_y_position = 0.5f;  // Irrelevant.
 181         ycbcr_format.cr_x_position = 0.0f;
 182         ycbcr_format.cr_y_position = 0.5f;
 183         ycbcr_input = (movit::YCbCrInput *)ycbcr_convert_chain->add_input(new YCbCrInput(image_format, ycbcr_format, 1280, 720));
 184
 185         YCbCrFormat ycbcr_output_format = ycbcr_format;
 186         ycbcr_output_format.chroma_subsampling_x = 1;
 187
 188         ImageFormat inout_format;
 189         inout_format.color_space = COLORSPACE_sRGB;
 190         inout_format.gamma_curve = GAMMA_sRGB;
 191
 192         check_error();
 193
 194         // One full Y'CbCr texture (for interpolation), one that's just Y (throwing away the
 195         // Cb and Cr channels). The second copy is sort of redundant, but it's the easiest way
 196         // of getting the gray data into a layered texture.
 197         ycbcr_convert_chain->add_ycbcr_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, ycbcr_output_format);
 198         check_error();
 199         ycbcr_convert_chain->add_ycbcr_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, ycbcr_output_format);
 200         check_error();
 201         ycbcr_convert_chain->set_dither_bits(8);
 202         check_error();
 203         ycbcr_convert_chain->finalize();
 204         check_error();
 205
 206         GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots];
 207         glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, input_tex);
 208         glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, gray_tex);
 209         check_error();
 210         constexpr size_t width = 1280, height = 720;  // FIXME: adjustable width, height
 211         int levels = find_num_levels(width, height);
 212         for (size_t i = 0; i < num_interpolate_slots; ++i) {
 213                 glTextureStorage3D(input_tex[i], levels, GL_RGBA8, width, height, 2);
 214                 check_error();
 215                 glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
 216                 check_error();
 217
 218                 InterpolatedFrameResources resource;
 219                 resource.input_tex = input_tex[i];
 220                 resource.gray_tex = gray_tex[i];
 221                 glCreateFramebuffers(2, resource.input_fbos);
 222                 check_error();
 223
 224                 glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
 225                 check_error();
 226                 glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0);
 227                 check_error();
 228                 glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
 229                 check_error();
 230                 glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1);
 231                 check_error();
 232
 233                 GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 234                 glNamedFramebufferDrawBuffers(resource.input_fbos[0], 2, bufs);
 235                 check_error();
 236                 glNamedFramebufferDrawBuffers(resource.input_fbos[1], 2, bufs);
 237                 check_error();
 238
 239                 glCreateBuffers(1, &resource.pbo);
 240                 check_error();
 241                 glNamedBufferStorage(resource.pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 242                 check_error();
 243                 resource.pbo_contents = glMapNamedBufferRange(resource.pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 244                 interpolate_resources.push_back(resource);
 245         }
 246
 247         check_error();
 248
 249         compute_flow.reset(new DISComputeFlow(width, height, operating_point3));
 250         interpolate.reset(new Interpolate(width, height, operating_point3));
 251         check_error();
 252 }
 253
 254 VideoStream::~VideoStream() {}
 255
 256 void VideoStream::start()
 257 {
 258         AVFormatContext *avctx = avformat_alloc_context();
 259         avctx->oformat = av_guess_format("nut", nullptr, nullptr);
 260
 261         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 262         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 263         avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
 264         avctx->pb->ignore_boundary_point = 1;
 265
 266         Mux::Codec video_codec = Mux::CODEC_MJPEG;
 267
 268         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 269
 270         string video_extradata;
 271
 272         constexpr int width = 1280, height = 720;  // Doesn't matter for MJPEG.
 273         stream_mux.reset(new Mux(avctx, width, height, video_codec, video_extradata, /*audio_codec_parameters=*/nullptr, COARSE_TIMEBASE,
 274                 /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}));
 275
 276
 277         encode_thread = thread(&VideoStream::encode_thread_func, this);
 278 }
 279
 280 void VideoStream::stop()
 281 {
 282         encode_thread.join();
 283 }
 284
 285 void VideoStream::schedule_original_frame(int64_t output_pts, unsigned stream_idx, int64_t input_pts)
 286 {
 287         fprintf(stderr, "output_pts=%ld  original      input_pts=%ld\n", output_pts, input_pts);
 288
 289         QueuedFrame qf;
 290         qf.type = QueuedFrame::ORIGINAL;
 291         qf.output_pts = output_pts;
 292         qf.stream_idx = stream_idx;
 293         qf.input_first_pts = input_pts;
 294
 295         unique_lock<mutex> lock(queue_lock);
 296         frame_queue.push_back(qf);
 297         queue_nonempty.notify_all();
 298 }
 299
 300 void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned stream_idx, int64_t input_first_pts, int64_t input_second_pts, float alpha)
 301 {
 302         fprintf(stderr, "output_pts=%ld  interpolated  input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, input_first_pts, input_second_pts, alpha);
 303
 304         // Get the temporary OpenGL resources we need for doing the interpolation.
 305         InterpolatedFrameResources resources;
 306         {
 307                 unique_lock<mutex> lock(queue_lock);
 308                 if (interpolate_resources.empty()) {
 309                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 310                         return;
 311                 }
 312                 resources = interpolate_resources.front();
 313                 interpolate_resources.pop_front();
 314         }
 315
 316         QueuedFrame qf;
 317         qf.type = QueuedFrame::INTERPOLATED;
 318         qf.output_pts = output_pts;
 319         qf.stream_idx = stream_idx;
 320         qf.resources = resources;
 321
 322         check_error();
 323
 324         // Convert frame0 and frame1 to OpenGL textures.
 325         // TODO: Deduplicate against JPEGFrameView::setDecodedFrame?
 326         for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
 327                 JPEGID jpeg_id;
 328                 jpeg_id.stream_idx = stream_idx;
 329                 jpeg_id.pts = frame_no == 1 ? input_second_pts : input_first_pts;
 330                 bool did_decode;
 331                 shared_ptr<Frame> frame = decode_jpeg_with_cache(jpeg_id, DECODE_IF_NOT_IN_CACHE, &did_decode);
 332                 ycbcr_format.chroma_subsampling_x = frame->chroma_subsampling_x;
 333                 ycbcr_format.chroma_subsampling_y = frame->chroma_subsampling_y;
 334                 ycbcr_input->change_ycbcr_format(ycbcr_format);
 335                 ycbcr_input->set_width(frame->width);
 336                 ycbcr_input->set_height(frame->height);
 337                 ycbcr_input->set_pixel_data(0, frame->y.get());
 338                 ycbcr_input->set_pixel_data(1, frame->cb.get());
 339                 ycbcr_input->set_pixel_data(2, frame->cr.get());
 340                 ycbcr_input->set_pitch(0, frame->pitch_y);
 341                 ycbcr_input->set_pitch(1, frame->pitch_chroma);
 342                 ycbcr_input->set_pitch(2, frame->pitch_chroma);
 343                 ycbcr_convert_chain->render_to_fbo(resources.input_fbos[frame_no], 1280, 720);
 344         }
 345
 346         glGenerateTextureMipmap(resources.input_tex);
 347         check_error();
 348         glGenerateTextureMipmap(resources.gray_tex);
 349         check_error();
 350
 351         // Compute the interpolated frame.
 352         qf.flow_tex = compute_flow->exec(resources.gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 353         check_error();
 354         qf.output_tex = interpolate->exec(resources.input_tex, resources.gray_tex, qf.flow_tex, 1280, 720, alpha);
 355         check_error();
 356
 357         // We could have released qf.flow_tex here, but to make sure we don't cause a stall
 358         // when trying to reuse it for the next frame, we can just as well hold on to it
 359         // and release it only when the readback is done.
 360
 361         // Read it down (asynchronously) to the CPU.
 362         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 363         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources.pbo);
 364         check_error();
 365         glGetTextureImage(qf.output_tex, 0, GL_RGBA, GL_UNSIGNED_BYTE, 1280 * 720 * 4, nullptr);
 366         check_error();
 367         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 368
 369         // Set a fence we can wait for to make sure the CPU sees the read.
 370         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 371         check_error();
 372         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 373         check_error();
 374
 375         unique_lock<mutex> lock(queue_lock);
 376         frame_queue.push_back(qf);
 377         queue_nonempty.notify_all();
 378 }
 379
 380 void VideoStream::encode_thread_func()
 381 {
 382         pthread_setname_np(pthread_self(), "VideoStream");
 383         QSurface *surface = create_surface();
 384         QOpenGLContext *context = create_context(surface);
 385         bool ok = make_current(context, surface);
 386         if (!ok) {
 387                 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
 388                 exit(1);
 389         }
 390
 391         for ( ;; ) {
 392                 QueuedFrame qf;
 393                 {
 394                         unique_lock<mutex> lock(queue_lock);
 395                         queue_nonempty.wait(lock, [this]{
 396                                 return !frame_queue.empty();
 397                         });
 398                         qf = frame_queue.front();
 399                         frame_queue.pop_front();
 400                 }
 401
 402                 if (qf.type == QueuedFrame::ORIGINAL) {
 403                         // Send the JPEG frame on, unchanged.
 404                         string jpeg = read_file(filename_for_frame(qf.stream_idx, qf.input_first_pts));
 405                         AVPacket pkt;
 406                         av_init_packet(&pkt);
 407                         pkt.stream_index = 0;
 408                         pkt.data = (uint8_t *)jpeg.data();
 409                         pkt.size = jpeg.size();
 410                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 411                 } else if (qf.type == QueuedFrame::INTERPOLATED) {
 412                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 413
 414                         vector<uint8_t> jpeg = encode_jpeg((const uint8_t *)qf.resources.pbo_contents, 1280, 720);
 415                         compute_flow->release_texture(qf.flow_tex);
 416                         interpolate->release_texture(qf.output_tex);
 417
 418                         AVPacket pkt;
 419                         av_init_packet(&pkt);
 420                         pkt.stream_index = 0;
 421                         pkt.data = (uint8_t *)jpeg.data();
 422                         pkt.size = jpeg.size();
 423                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 424
 425                         // Put the frame resources back.
 426                         unique_lock<mutex> lock(queue_lock);
 427                         interpolate_resources.push_back(qf.resources);
 428                 }
 429         }
 430 }
 431
 432 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 433 {
 434         VideoStream *video_stream = (VideoStream *)opaque;
 435         return video_stream->write_packet2(buf, buf_size, type, time);
 436 }
 437
 438 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 439 {
 440         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
 441                 seen_sync_markers = true;
 442         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
 443                 // We don't know if this is a keyframe or not (the muxer could
 444                 // avoid marking it), so we just have to make the best of it.
 445                 type = AVIO_DATA_MARKER_SYNC_POINT;
 446         }
 447
 448         if (type == AVIO_DATA_MARKER_HEADER) {
 449                 stream_mux_header.append((char *)buf, buf_size);
 450                 global_httpd->set_header(stream_mux_header);
 451         } else {
 452                 global_httpd->add_data((char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
 453         }
 454         return buf_size;
 455 }
 456