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