X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;ds=sidebyside;f=futatabi%2Fvideo_stream.cpp;h=987bef542ac45c3b9ba623320b4a298ac96e507c;hb=2cb648106d32b9968f2026536fbead096308c7d1;hp=27e2c4a81e499ddc031b21434c05fd0490f1f37f;hpb=94a228f7473325130704e96c82937b690703615b;p=nageru diff --git a/futatabi/video_stream.cpp b/futatabi/video_stream.cpp index 27e2c4a..987bef5 100644 --- a/futatabi/video_stream.cpp +++ b/futatabi/video_stream.cpp @@ -6,14 +6,16 @@ extern "C" { } #include "chroma_subsampler.h" -#include "shared/context.h" +#include "exif_parser.h" #include "flags.h" #include "flow.h" -#include "shared/httpd.h" #include "jpeg_frame_view.h" #include "movit/util.h" -#include "shared/mux.h" #include "player.h" +#include "shared/context.h" +#include "shared/httpd.h" +#include "shared/shared_defs.h" +#include "shared/mux.h" #include "util.h" #include "ycbcr_converter.h" @@ -21,6 +23,7 @@ extern "C" { #include #include +using namespace movit; using namespace std; using namespace std::chrono; @@ -28,7 +31,7 @@ extern HTTPD *global_httpd; struct VectorDestinationManager { jpeg_destination_mgr pub; - std::vector dest; + string dest; VectorDestinationManager() { @@ -62,7 +65,7 @@ struct VectorDestinationManager { { dest.resize(bytes_used + 4096); dest.resize(dest.capacity()); - pub.next_output_byte = dest.data() + bytes_used; + pub.next_output_byte = (uint8_t *)dest.data() + bytes_used; pub.free_in_buffer = dest.size() - bytes_used; } @@ -78,7 +81,7 @@ struct VectorDestinationManager { }; static_assert(std::is_standard_layout::value, ""); -vector encode_jpeg(const uint8_t *y_data, const uint8_t *cb_data, const uint8_t *cr_data, unsigned width, unsigned height) +string encode_jpeg(const uint8_t *y_data, const uint8_t *cb_data, const uint8_t *cr_data, unsigned width, unsigned height, const string exif_data) { VectorDestinationManager dest; @@ -111,6 +114,10 @@ vector encode_jpeg(const uint8_t *y_data, const uint8_t *cb_data, const // (and nothing else). jpeg_write_marker(&cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601")); + if (!exif_data.empty()) { + jpeg_write_marker(&cinfo, JPEG_APP0 + 1, (const JOCTET *)exif_data.data(), exif_data.size()); + } + JSAMPROW yptr[8], cbptr[8], crptr[8]; JSAMPARRAY data[3] = { yptr, cbptr, crptr }; for (unsigned y = 0; y < height; y += 8) { @@ -233,7 +240,13 @@ VideoStream::VideoStream(AVFormatContext *file_avctx) unique_ptr cb_or_cr(new uint8_t[(global_flags.width / 2) * global_flags.height]); memset(y.get(), 16, global_flags.width * global_flags.height); memset(cb_or_cr.get(), 128, (global_flags.width / 2) * global_flags.height); - last_frame = encode_jpeg(y.get(), cb_or_cr.get(), cb_or_cr.get(), global_flags.width, global_flags.height); + last_frame = encode_jpeg(y.get(), cb_or_cr.get(), cb_or_cr.get(), global_flags.width, global_flags.height, /*exif_data=*/""); + + if (file_avctx != nullptr) { + with_subtitles = Mux::WITHOUT_SUBTITLES; + } else { + with_subtitles = Mux::WITH_SUBTITLES; + } } VideoStream::~VideoStream() @@ -241,6 +254,30 @@ VideoStream::~VideoStream() if (last_flow_tex != 0) { compute_flow->release_texture(last_flow_tex); } + + for (const unique_ptr &resource : interpolate_resources) { + glUnmapNamedBuffer(resource->pbo); + check_error(); + glDeleteBuffers(1, &resource->pbo); + check_error(); + glDeleteFramebuffers(2, resource->input_fbos); + check_error(); + glDeleteFramebuffers(1, &resource->fade_fbo); + check_error(); + glDeleteTextures(1, &resource->input_tex); + check_error(); + glDeleteTextures(1, &resource->gray_tex); + check_error(); + glDeleteTextures(1, &resource->fade_y_output_tex); + check_error(); + glDeleteTextures(1, &resource->fade_cbcr_output_tex); + check_error(); + glDeleteTextures(1, &resource->cb_tex); + check_error(); + glDeleteTextures(1, &resource->cr_tex); + check_error(); + } + assert(interpolate_resources.size() == num_interpolate_slots); } void VideoStream::start() @@ -262,11 +299,19 @@ void VideoStream::start() avctx->flags = AVFMT_FLAG_CUSTOM_IO; } + AVCodecParameters *audio_codecpar = avcodec_parameters_alloc(); + + audio_codecpar->codec_type = AVMEDIA_TYPE_AUDIO; + audio_codecpar->codec_id = AV_CODEC_ID_PCM_S32LE; + audio_codecpar->channel_layout = AV_CH_LAYOUT_STEREO; + audio_codecpar->channels = 2; + audio_codecpar->sample_rate = OUTPUT_FREQUENCY; + size_t width = global_flags.width, height = global_flags.height; // Doesn't matter for MJPEG. - mux.reset(new Mux(avctx, width, height, Mux::CODEC_MJPEG, /*video_extradata=*/"", /*audio_codec_parameters=*/nullptr, - AVCOL_SPC_BT709, Mux::WITHOUT_AUDIO, - COARSE_TIMEBASE, /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {})); + mux.reset(new Mux(avctx, width, height, Mux::CODEC_MJPEG, /*video_extradata=*/"", audio_codecpar, + AVCOL_SPC_BT709, COARSE_TIMEBASE, /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}, with_subtitles)); + avcodec_parameters_free(&audio_codecpar); encode_thread = thread(&VideoStream::encode_thread_func, this); } @@ -283,7 +328,7 @@ void VideoStream::clear_queue() deque q; { - unique_lock lock(queue_lock); + lock_guard lock(queue_lock); q = move(frame_queue); } @@ -308,23 +353,22 @@ void VideoStream::clear_queue() void VideoStream::schedule_original_frame(steady_clock::time_point local_pts, int64_t output_pts, function &&display_func, QueueSpotHolder &&queue_spot_holder, - FrameOnDisk frame) + FrameOnDisk frame, const string &subtitle, bool include_audio) { - fprintf(stderr, "output_pts=%ld original input_pts=%ld\n", output_pts, frame.pts); - - // Preload the file from disk, so that the encoder thread does not get stalled. - // TODO: Consider sending it through the queue instead. - (void)frame_reader.read_frame(frame); + fprintf(stderr, "output_pts=%" PRId64 " original input_pts=%" PRId64 "\n", output_pts, frame.pts); QueuedFrame qf; qf.local_pts = local_pts; qf.type = QueuedFrame::ORIGINAL; qf.output_pts = output_pts; - qf.frame1 = frame; qf.display_func = move(display_func); qf.queue_spot_holder = move(queue_spot_holder); + qf.subtitle = subtitle; + FrameReader::Frame read_frame = frame_reader.read_frame(frame, /*read_video=*/true, include_audio); + qf.encoded_jpeg.reset(new string(move(read_frame.video))); + qf.audio = move(read_frame.audio); - unique_lock lock(queue_lock); + lock_guard lock(queue_lock); frame_queue.push_back(move(qf)); queue_changed.notify_all(); } @@ -333,9 +377,9 @@ void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64 function &&display_func, QueueSpotHolder &&queue_spot_holder, FrameOnDisk frame1_spec, FrameOnDisk frame2_spec, - float fade_alpha) + float fade_alpha, const string &subtitle) { - fprintf(stderr, "output_pts=%ld faded input_pts=%ld,%ld fade_alpha=%.2f\n", output_pts, frame1_spec.pts, frame2_spec.pts, fade_alpha); + fprintf(stderr, "output_pts=%" PRId64 " faded input_pts=%" PRId64 ",%" PRId64 " fade_alpha=%.2f\n", output_pts, frame1_spec.pts, frame2_spec.pts, fade_alpha); // Get the temporary OpenGL resources we need for doing the fade. // (We share these with interpolated frames, which is slightly @@ -343,7 +387,7 @@ void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64 // separate pools around.) BorrowedInterpolatedFrameResources resources; { - unique_lock lock(queue_lock); + lock_guard lock(queue_lock); if (interpolate_resources.empty()) { fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n"); return; @@ -366,6 +410,7 @@ void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64 qf.frame1 = frame1_spec; qf.display_func = move(display_func); qf.queue_spot_holder = move(queue_spot_holder); + qf.subtitle = subtitle; qf.secondary_frame = frame2_spec; @@ -392,7 +437,7 @@ void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64 qf.resources = move(resources); qf.local_pts = local_pts; - unique_lock lock(queue_lock); + lock_guard lock(queue_lock); frame_queue.push_back(move(qf)); queue_changed.notify_all(); } @@ -401,18 +446,19 @@ void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts int64_t output_pts, function)> &&display_func, QueueSpotHolder &&queue_spot_holder, FrameOnDisk frame1, FrameOnDisk frame2, - float alpha, FrameOnDisk secondary_frame, float fade_alpha) + float alpha, FrameOnDisk secondary_frame, float fade_alpha, const string &subtitle, + bool play_audio) { if (secondary_frame.pts != -1) { - fprintf(stderr, "output_pts=%ld interpolated input_pts1=%ld input_pts2=%ld alpha=%.3f secondary_pts=%ld fade_alpha=%.2f\n", output_pts, frame1.pts, frame2.pts, alpha, secondary_frame.pts, fade_alpha); + fprintf(stderr, "output_pts=%" PRId64 " interpolated input_pts1=%" PRId64 " input_pts2=%" PRId64 " alpha=%.3f secondary_pts=%" PRId64 " fade_alpha=%.2f\n", output_pts, frame1.pts, frame2.pts, alpha, secondary_frame.pts, fade_alpha); } else { - fprintf(stderr, "output_pts=%ld interpolated input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, frame1.pts, frame2.pts, alpha); + fprintf(stderr, "output_pts=%" PRId64 " interpolated input_pts1=%" PRId64 " input_pts2=%" PRId64 " alpha=%.3f\n", output_pts, frame1.pts, frame2.pts, alpha); } // Get the temporary OpenGL resources we need for doing the interpolation. BorrowedInterpolatedFrameResources resources; { - unique_lock lock(queue_lock); + lock_guard lock(queue_lock); if (interpolate_resources.empty()) { fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n"); return; @@ -427,6 +473,11 @@ void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts qf.display_decoded_func = move(display_func); qf.queue_spot_holder = move(queue_spot_holder); qf.local_pts = local_pts; + qf.subtitle = subtitle; + + if (play_audio) { + qf.audio = frame_reader.read_frame(frame1, /*read_video=*/false, /*read_audio=*/true).audio; + } check_error(); @@ -436,6 +487,9 @@ void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts bool did_decode; shared_ptr frame = decode_jpeg_with_cache(frame_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode); ycbcr_converter->prepare_chain_for_conversion(frame)->render_to_fbo(resources->input_fbos[frame_no], global_flags.width, global_flags.height); + if (frame_no == 1) { + qf.exif_data = frame->exif_data; // Use the white point from the last frame. + } } glGenerateTextureMipmap(resources->input_tex); @@ -474,12 +528,16 @@ void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts shared_ptr frame2 = decode_jpeg_with_cache(secondary_frame, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode); // Then fade against it, putting it into the fade Y' and CbCr textures. - ycbcr_semiplanar_converter->prepare_chain_for_fade_from_texture(qf.output_tex, global_flags.width, global_flags.height, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, global_flags.width, global_flags.height); + RGBTriplet neutral_color = get_neutral_color(qf.exif_data); + ycbcr_semiplanar_converter->prepare_chain_for_fade_from_texture(qf.output_tex, neutral_color, global_flags.width, global_flags.height, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, global_flags.width, global_flags.height); // Subsample and split Cb/Cr. chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex); interpolate_no_split->release_texture(qf.output_tex); + + // We already applied the white balance, so don't have the client redo it. + qf.exif_data.clear(); } else { tie(qf.output_tex, qf.cbcr_tex) = interpolate->exec(resources->input_tex, resources->gray_tex, flow_tex, global_flags.width, global_flags.height, alpha); check_error(); @@ -518,22 +576,37 @@ void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts check_error(); qf.resources = move(resources); - unique_lock lock(queue_lock); + lock_guard lock(queue_lock); frame_queue.push_back(move(qf)); queue_changed.notify_all(); } void VideoStream::schedule_refresh_frame(steady_clock::time_point local_pts, int64_t output_pts, function &&display_func, - QueueSpotHolder &&queue_spot_holder) + QueueSpotHolder &&queue_spot_holder, const string &subtitle) { QueuedFrame qf; qf.type = QueuedFrame::REFRESH; qf.output_pts = output_pts; qf.display_func = move(display_func); qf.queue_spot_holder = move(queue_spot_holder); + qf.subtitle = subtitle; + + lock_guard lock(queue_lock); + frame_queue.push_back(move(qf)); + queue_changed.notify_all(); +} + +void VideoStream::schedule_silence(steady_clock::time_point local_pts, int64_t output_pts, + int64_t length_pts, QueueSpotHolder &&queue_spot_holder) +{ + QueuedFrame qf; + qf.type = QueuedFrame::SILENCE; + qf.output_pts = output_pts; + qf.queue_spot_holder = move(queue_spot_holder); + qf.silence_length_pts = length_pts; - unique_lock lock(queue_lock); + lock_guard lock(queue_lock); frame_queue.push_back(move(qf)); queue_changed.notify_all(); } @@ -577,7 +650,7 @@ void VideoStream::encode_thread_func() bool ok = make_current(context, surface); if (!ok) { fprintf(stderr, "Video stream couldn't get an OpenGL context\n"); - exit(1); + abort(); } while (!should_quit) { @@ -586,7 +659,7 @@ void VideoStream::encode_thread_func() unique_lock lock(queue_lock); // Wait until we have a frame to play. - queue_changed.wait(lock, [this]{ + queue_changed.wait(lock, [this] { return !frame_queue.empty() || should_quit; }); if (should_quit) { @@ -600,7 +673,7 @@ void VideoStream::encode_thread_func() if (output_fast_forward) { aborted = frame_queue.empty() || frame_queue.front().local_pts != frame_start; } else { - aborted = queue_changed.wait_until(lock, frame_start, [this, frame_start]{ + aborted = queue_changed.wait_until(lock, frame_start, [this, frame_start] { return frame_queue.empty() || frame_queue.front().local_pts != frame_start; }); } @@ -612,9 +685,23 @@ void VideoStream::encode_thread_func() frame_queue.pop_front(); } + // Hack: We mux the subtitle packet one time unit before the actual frame, + // so that Nageru is sure to get it first. + if (!qf.subtitle.empty() && with_subtitles == Mux::WITH_SUBTITLES) { + AVPacket pkt; + av_init_packet(&pkt); + pkt.stream_index = mux->get_subtitle_stream_idx(); + assert(pkt.stream_index != -1); + pkt.data = (uint8_t *)qf.subtitle.data(); + pkt.size = qf.subtitle.size(); + pkt.flags = 0; + pkt.duration = lrint(TIMEBASE / global_flags.output_framerate); // Doesn't really matter for Nageru. + mux->add_packet(pkt, qf.output_pts - 1, qf.output_pts - 1); + } + if (qf.type == QueuedFrame::ORIGINAL) { // Send the JPEG frame on, unchanged. - string jpeg = frame_reader.read_frame(qf.frame1); + string jpeg = move(*qf.encoded_jpeg); AVPacket pkt; av_init_packet(&pkt); pkt.stream_index = 0; @@ -622,15 +709,17 @@ void VideoStream::encode_thread_func() pkt.size = jpeg.size(); pkt.flags = AV_PKT_FLAG_KEY; mux->add_packet(pkt, qf.output_pts, qf.output_pts); + last_frame = move(jpeg); - last_frame.assign(&jpeg[0], &jpeg[0] + jpeg.size()); + add_audio_or_silence(qf); } else if (qf.type == QueuedFrame::FADED) { glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED); shared_ptr frame = frame_from_pbo(qf.resources->pbo_contents, global_flags.width, global_flags.height); + assert(frame->exif_data.empty()); // Now JPEG encode it, and send it on to the stream. - vector jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), global_flags.width, global_flags.height); + string jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), global_flags.width, global_flags.height, /*exif_data=*/""); AVPacket pkt; av_init_packet(&pkt); @@ -640,6 +729,8 @@ void VideoStream::encode_thread_func() pkt.flags = AV_PKT_FLAG_KEY; mux->add_packet(pkt, qf.output_pts, qf.output_pts); last_frame = move(jpeg); + + add_audio_or_silence(qf); } else if (qf.type == QueuedFrame::INTERPOLATED || qf.type == QueuedFrame::FADED_INTERPOLATED) { glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED); @@ -650,7 +741,7 @@ void VideoStream::encode_thread_func() } // Now JPEG encode it, and send it on to the stream. - vector jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), global_flags.width, global_flags.height); + string jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), global_flags.width, global_flags.height, move(qf.exif_data)); if (qf.flow_tex != 0) { compute_flow->release_texture(qf.flow_tex); } @@ -667,6 +758,8 @@ void VideoStream::encode_thread_func() pkt.flags = AV_PKT_FLAG_KEY; mux->add_packet(pkt, qf.output_pts, qf.output_pts); last_frame = move(jpeg); + + add_audio_or_silence(qf); } else if (qf.type == QueuedFrame::REFRESH) { AVPacket pkt; av_init_packet(&pkt); @@ -675,6 +768,10 @@ void VideoStream::encode_thread_func() pkt.size = last_frame.size(); pkt.flags = AV_PKT_FLAG_KEY; mux->add_packet(pkt, qf.output_pts, qf.output_pts); + + add_audio_or_silence(qf); // Definitely silence. + } else if (qf.type == QueuedFrame::SILENCE) { + add_silence(qf.output_pts, qf.silence_length_pts); } else { assert(false); } @@ -700,11 +797,47 @@ int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType ty type = AVIO_DATA_MARKER_SYNC_POINT; } + HTTPD::StreamID stream_id{ HTTPD::MAIN_STREAM, 0 }; if (type == AVIO_DATA_MARKER_HEADER) { stream_mux_header.append((char *)buf, buf_size); - global_httpd->set_header(HTTPD::MAIN_STREAM, stream_mux_header); + global_httpd->set_header(stream_id, stream_mux_header); } else { - global_httpd->add_data(HTTPD::MAIN_STREAM, (char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 }); + global_httpd->add_data(stream_id, (char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 }); } return buf_size; } + +void VideoStream::add_silence(int64_t pts, int64_t length_pts) +{ + // At 59.94, this will never quite add up (even discounting refresh frames, + // which have unpredictable length), but hopefully, the player in the other + // end should be able to stretch silence easily enough. + long num_samples = lrint(length_pts * double(OUTPUT_FREQUENCY) / double(TIMEBASE)) * 2; + uint8_t *zero = (uint8_t *)calloc(num_samples, sizeof(int32_t)); + + AVPacket pkt; + av_init_packet(&pkt); + pkt.stream_index = 1; + pkt.data = zero; + pkt.size = num_samples * sizeof(int32_t); + pkt.flags = AV_PKT_FLAG_KEY; + mux->add_packet(pkt, pts, pts); + + free(zero); +} + +void VideoStream::add_audio_or_silence(const QueuedFrame &qf) +{ + if (qf.audio.empty()) { + int64_t frame_length = lrint(double(TIMEBASE) / global_flags.output_framerate); + add_silence(qf.output_pts, frame_length); + } else { + AVPacket pkt; + av_init_packet(&pkt); + pkt.stream_index = 1; + pkt.data = (uint8_t *)qf.audio.data(); + pkt.size = qf.audio.size(); + pkt.flags = AV_PKT_FLAG_KEY; + mux->add_packet(pkt, qf.output_pts, qf.output_pts); + } +}