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.input_components = 3;
 120         cinfo.comp_info[0].h_samp_factor = 2;
 121         cinfo.comp_info[0].v_samp_factor = 1;
 122         cinfo.comp_info[1].h_samp_factor = 1;
 123         cinfo.comp_info[1].v_samp_factor = 1;
 124         cinfo.comp_info[2].h_samp_factor = 1;
 125         cinfo.comp_info[2].v_samp_factor = 1;
 126         // cinfo.CCIR601_sampling = true;  // TODO: Subsample ourselves.
 127         jpeg_start_compress(&cinfo, true);
 128
 129         unique_ptr<uint8_t[]> row(new uint8_t[width * 3]);
 130         JSAMPROW row_pointer[1] = { row.get() };
 131         for (unsigned y = 0; y < height; ++y) {
 132                 const uint8_t *sptr = &pixel_data[(height - cinfo.next_scanline - 1) * width * 4];
 133                 uint8_t *dptr = row.get();
 134                 for (unsigned x = 0; x < width; ++x) {
 135                         *dptr++ = *sptr++;
 136                         *dptr++ = *sptr++;
 137                         *dptr++ = *sptr++;
 138                         ++sptr;
 139                 }
 140                 (void) jpeg_write_scanlines(&cinfo, row_pointer, 1);
 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         ImageFormat inout_format;
 169         inout_format.color_space = COLORSPACE_sRGB;
 170         inout_format.gamma_curve = GAMMA_sRGB;
 171
 172         check_error();
 173         ycbcr_convert_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 174         check_error();
 175         ycbcr_convert_chain->set_dither_bits(8);
 176         check_error();
 177         ycbcr_convert_chain->finalize();
 178         check_error();
 179
 180         GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots];
 181         glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, input_tex);
 182         glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, gray_tex);
 183         check_error();
 184         constexpr size_t width = 1280, height = 720;  // FIXME: adjustable width, height
 185         int levels = find_num_levels(width, height);
 186         for (size_t i = 0; i < num_interpolate_slots; ++i) {
 187                 glTextureStorage3D(input_tex[i], levels, GL_RGBA8, width, height, 2);
 188                 check_error();
 189                 glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
 190                 check_error();
 191
 192                 InterpolatedFrameResources resource;
 193                 resource.input_tex = input_tex[i];
 194                 resource.gray_tex = gray_tex[i];
 195                 glCreateFramebuffers(2, resource.input_fbos);
 196                 check_error();
 197
 198                 glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
 199                 check_error();
 200                 glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
 201                 check_error();
 202
 203                 GLuint buf = GL_COLOR_ATTACHMENT0;
 204                 glNamedFramebufferDrawBuffers(resource.input_fbos[0], 1, &buf);
 205                 check_error();
 206                 glNamedFramebufferDrawBuffers(resource.input_fbos[1], 1, &buf);
 207                 check_error();
 208
 209                 glCreateBuffers(1, &resource.pbo);
 210                 check_error();
 211                 glNamedBufferStorage(resource.pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 212                 check_error();
 213                 resource.pbo_contents = glMapNamedBufferRange(resource.pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 214                 interpolate_resources.push_back(resource);
 215         }
 216
 217         check_error();
 218
 219         compute_flow.reset(new DISComputeFlow(width, height, operating_point3));
 220         gray.reset(new GrayscaleConversion);  // NOTE: Must come after DISComputeFlow, since it sets up the VBO!
 221         interpolate.reset(new Interpolate(width, height, operating_point3));
 222         check_error();
 223 }
 224
 225 VideoStream::~VideoStream() {}
 226
 227 void VideoStream::start()
 228 {
 229         AVFormatContext *avctx = avformat_alloc_context();
 230         avctx->oformat = av_guess_format("nut", nullptr, nullptr);
 231
 232         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 233         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 234         avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
 235         avctx->pb->ignore_boundary_point = 1;
 236
 237         Mux::Codec video_codec = Mux::CODEC_MJPEG;
 238
 239         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 240
 241         string video_extradata;
 242
 243         constexpr int width = 1280, height = 720;  // Doesn't matter for MJPEG.
 244         stream_mux.reset(new Mux(avctx, width, height, video_codec, video_extradata, /*audio_codec_parameters=*/nullptr, COARSE_TIMEBASE,
 245                 /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}));
 246
 247
 248         encode_thread = thread(&VideoStream::encode_thread_func, this);
 249 }
 250
 251 void VideoStream::stop()
 252 {
 253         encode_thread.join();
 254 }
 255
 256 void VideoStream::schedule_original_frame(int64_t output_pts, unsigned stream_idx, int64_t input_pts)
 257 {
 258         fprintf(stderr, "output_pts=%ld  original      input_pts=%ld\n", output_pts, input_pts);
 259
 260         QueuedFrame qf;
 261         qf.type = QueuedFrame::ORIGINAL;
 262         qf.output_pts = output_pts;
 263         qf.stream_idx = stream_idx;
 264         qf.input_first_pts = input_pts;
 265
 266         unique_lock<mutex> lock(queue_lock);
 267         frame_queue.push_back(qf);
 268         queue_nonempty.notify_all();
 269 }
 270
 271 void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned stream_idx, int64_t input_first_pts, int64_t input_second_pts, float alpha)
 272 {
 273         fprintf(stderr, "output_pts=%ld  interpolated  input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, input_first_pts, input_second_pts, alpha);
 274
 275         // Get the temporary OpenGL resources we need for doing the interpolation.
 276         InterpolatedFrameResources resources;
 277         {
 278                 unique_lock<mutex> lock(queue_lock);
 279                 if (interpolate_resources.empty()) {
 280                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 281                         return;
 282                 }
 283                 resources = interpolate_resources.front();
 284                 interpolate_resources.pop_front();
 285         }
 286
 287         QueuedFrame qf;
 288         qf.type = QueuedFrame::INTERPOLATED;
 289         qf.output_pts = output_pts;
 290         qf.stream_idx = stream_idx;
 291         qf.resources = resources;
 292
 293         check_error();
 294
 295         // Convert frame0 and frame1 to OpenGL textures.
 296         // TODO: Deduplicate against JPEGFrameView::setDecodedFrame?
 297         for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
 298                 JPEGID jpeg_id;
 299                 jpeg_id.stream_idx = stream_idx;
 300                 jpeg_id.pts = frame_no == 1 ? input_second_pts : input_first_pts;
 301                 bool did_decode;
 302                 shared_ptr<Frame> frame = decode_jpeg_with_cache(jpeg_id, DECODE_IF_NOT_IN_CACHE, &did_decode);
 303                 ycbcr_format.chroma_subsampling_x = frame->chroma_subsampling_x;
 304                 ycbcr_format.chroma_subsampling_y = frame->chroma_subsampling_y;
 305                 ycbcr_input->change_ycbcr_format(ycbcr_format);
 306                 ycbcr_input->set_width(frame->width);
 307                 ycbcr_input->set_height(frame->height);
 308                 ycbcr_input->set_pixel_data(0, frame->y.get());
 309                 ycbcr_input->set_pixel_data(1, frame->cb.get());
 310                 ycbcr_input->set_pixel_data(2, frame->cr.get());
 311                 ycbcr_input->set_pitch(0, frame->pitch_y);
 312                 ycbcr_input->set_pitch(1, frame->pitch_chroma);
 313                 ycbcr_input->set_pitch(2, frame->pitch_chroma);
 314                 ycbcr_convert_chain->render_to_fbo(resources.input_fbos[frame_no], 1280, 720);
 315         }
 316
 317         glGenerateTextureMipmap(resources.input_tex);
 318
 319         // Compute the interpolated frame.
 320         check_error();
 321         gray->exec(resources.input_tex, resources.gray_tex, 1280, 720, /*num_layers=*/2);
 322         check_error();
 323         glGenerateTextureMipmap(resources.gray_tex);
 324         check_error();
 325         qf.flow_tex = compute_flow->exec(resources.gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 326         check_error();
 327
 328         qf.output_tex = interpolate->exec(resources.input_tex, qf.flow_tex, 1280, 720, alpha);
 329         check_error();
 330
 331         // We could have released qf.flow_tex here, but to make sure we don't cause a stall
 332         // when trying to reuse it for the next frame, we can just as well hold on to it
 333         // and release it only when the readback is done.
 334
 335         // Read it down (asynchronously) to the CPU.
 336         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 337         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources.pbo);
 338         check_error();
 339         glGetTextureImage(qf.output_tex, 0, GL_RGBA, GL_UNSIGNED_BYTE, 1280 * 720 * 4, nullptr);
 340         check_error();
 341         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 342
 343         // Set a fence we can wait for to make sure the CPU sees the read.
 344         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 345         check_error();
 346         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 347         check_error();
 348
 349         unique_lock<mutex> lock(queue_lock);
 350         frame_queue.push_back(qf);
 351         queue_nonempty.notify_all();
 352 }
 353
 354 void VideoStream::encode_thread_func()
 355 {
 356         QSurface *surface = create_surface();
 357         QOpenGLContext *context = create_context(surface);
 358         bool ok = make_current(context, surface);
 359         if (!ok) {
 360                 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
 361                 exit(1);
 362         }
 363
 364         for ( ;; ) {
 365                 QueuedFrame qf;
 366                 {
 367                         unique_lock<mutex> lock(queue_lock);
 368                         queue_nonempty.wait(lock, [this]{
 369                                 return !frame_queue.empty();
 370                         });
 371                         qf = frame_queue.front();
 372                         frame_queue.pop_front();
 373                 }
 374
 375                 if (qf.type == QueuedFrame::ORIGINAL) {
 376                         // Send the JPEG frame on, unchanged.
 377                         string jpeg = read_file(filename_for_frame(qf.stream_idx, qf.input_first_pts));
 378                         AVPacket pkt;
 379                         av_init_packet(&pkt);
 380                         pkt.stream_index = 0;
 381                         pkt.data = (uint8_t *)jpeg.data();
 382                         pkt.size = jpeg.size();
 383                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 384                 } else if (qf.type == QueuedFrame::INTERPOLATED) {
 385                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 386
 387                         vector<uint8_t> jpeg = encode_jpeg((const uint8_t *)qf.resources.pbo_contents, 1280, 720);
 388                         compute_flow->release_texture(qf.flow_tex);
 389                         interpolate->release_texture(qf.output_tex);
 390
 391                         AVPacket pkt;
 392                         av_init_packet(&pkt);
 393                         pkt.stream_index = 0;
 394                         pkt.data = (uint8_t *)jpeg.data();
 395                         pkt.size = jpeg.size();
 396                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 397
 398                         // Put the frame resources back.
 399                         unique_lock<mutex> lock(queue_lock);
 400                         interpolate_resources.push_back(qf.resources);
 401                 }
 402         }
 403 }
 404
 405 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 406 {
 407         VideoStream *video_stream = (VideoStream *)opaque;
 408         return video_stream->write_packet2(buf, buf_size, type, time);
 409 }
 410
 411 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 412 {
 413         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
 414                 seen_sync_markers = true;
 415         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
 416                 // We don't know if this is a keyframe or not (the muxer could
 417                 // avoid marking it), so we just have to make the best of it.
 418                 type = AVIO_DATA_MARKER_SYNC_POINT;
 419         }
 420
 421         if (type == AVIO_DATA_MARKER_HEADER) {
 422                 stream_mux_header.append((char *)buf, buf_size);
 423                 global_httpd->set_header(stream_mux_header);
 424         } else {
 425                 global_httpd->add_data((char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
 426         }
 427         return buf_size;
 428 }
 429