]> git.sesse.net Git - nageru/blobdiff - nageru/ffmpeg_capture.cpp
A hack to make Kaeru overlay CEF scores on a transcoded file. Probably not useful...
[nageru] / nageru / ffmpeg_capture.cpp
index 87d1c007e9142a07e21b2c49820c706709b45c26..984ce5e8b8add69da4b0a9fd8a68d232c7ab6b41 100644 (file)
@@ -143,7 +143,7 @@ AVPixelFormat decide_dst_format(AVPixelFormat src_format, bmusb::PixelFormat dst
        return av_get_pix_fmt(best_format);
 }
 
-YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *frame, bool is_mjpeg)
+YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *frame, bool is_mjpeg, AVColorSpace *last_colorspace, AVChromaLocation *last_chroma_location)
 {
        YCbCrFormat format;
        AVColorSpace colorspace = frame->colorspace;
@@ -163,11 +163,14 @@ YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *f
                format.luma_coefficients = (frame->height >= 720 ? YCBCR_REC_709 : YCBCR_REC_601);
                break;
        default:
-               fprintf(stderr, "Unknown Y'CbCr coefficient enum %d from FFmpeg; choosing Rec. 709.\n",
-                       colorspace);
+               if (colorspace != *last_colorspace) {
+                       fprintf(stderr, "Unknown Y'CbCr coefficient enum %d from FFmpeg; choosing Rec. 709.\n",
+                               colorspace);
+               }
                format.luma_coefficients = YCBCR_REC_709;
                break;
        }
+       *last_colorspace = colorspace;
 
        format.full_range = is_full_range(desc);
        format.num_levels = 1 << desc->comp[0].depth;
@@ -200,12 +203,15 @@ YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *f
                format.cb_y_position = 1.0;
                break;
        default:
-               fprintf(stderr, "Unknown chroma location coefficient enum %d from FFmpeg; choosing center.\n",
-                       frame->chroma_location);
+               if (frame->chroma_location != *last_chroma_location) {
+                       fprintf(stderr, "Unknown chroma location coefficient enum %d from FFmpeg; choosing center.\n",
+                               frame->chroma_location);
+               }
                format.cb_x_position = 0.5;
                format.cb_y_position = 0.5;
                break;
        }
+       *last_chroma_location = frame->chroma_location;
 
        if (is_mjpeg && !format.full_range) {
                // Limited-range MJPEG is only detected by FFmpeg whenever a special
@@ -252,8 +258,9 @@ RGBTriplet get_neutral_color(AVDictionary *metadata)
 }  // namespace
 
 FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned height)
-       : filename(filename), width(width), height(height), video_timebase{1, 1}
+       : width(width), height(height), video_timebase{1, 1}
 {
+       filenames.push_back(filename);
        description = "Video: " + filename;
 
        last_frame = steady_clock::now();
@@ -261,6 +268,17 @@ FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned he
        avformat_network_init();  // In case someone wants this.
 }
 
+FFmpegCapture::FFmpegCapture(const std::vector<std::string> &filenames, unsigned width, unsigned height)
+       : filenames(filenames), width(width), height(height), video_timebase{1, 1}
+{
+       description = "Video: " + filenames[0];
+
+       last_frame = steady_clock::now();
+       play_once = true;
+
+       avformat_network_init();  // In case someone wants this.
+}
+
 #ifdef HAVE_SRT
 FFmpegCapture::FFmpegCapture(int srt_sock, const string &stream_id)
        : srt_sock(srt_sock),
@@ -351,11 +369,14 @@ void FFmpegCapture::producer_thread_func()
        snprintf(thread_name, sizeof(thread_name), "FFmpeg_C_%d", card_index);
        pthread_setname_np(pthread_self(), thread_name);
 
+       printf("CAP\n");
        while (!producer_thread_should_quit.should_quit()) {
                string filename_copy;
+               printf("CAP %zu LEFT\n", filenames.size());
                {
                        lock_guard<mutex> lock(filename_mu);
-                       filename_copy = filename;
+                       filename_copy = filenames.front();
+                       filenames.erase(filenames.begin());
                }
 
                string pathname;
@@ -364,6 +385,7 @@ void FFmpegCapture::producer_thread_func()
                } else {
                        pathname = description;
                }
+               printf("CAP %s\n", pathname.c_str());
                if (pathname.empty()) {
                        send_disconnected_frame();
                        if (play_once) {
@@ -385,7 +407,7 @@ void FFmpegCapture::producer_thread_func()
                        continue;
                }
 
-               if (play_once) {
+               if (play_once && filenames.empty()) {
                        send_disconnected_frame();
                        break;
                }
@@ -432,9 +454,11 @@ void FFmpegCapture::send_disconnected_frame()
                        memset(video_frame.data + frame_width * frame_height, 128, frame_width * frame_height);  // Valid for both NV12 and planar.
                }
 
-               frame_callback(-1, AVRational{1, TIMEBASE}, -1, AVRational{1, TIMEBASE}, timecode++,
-                       video_frame, /*video_offset=*/0, video_format,
-                       FrameAllocator::Frame(), /*audio_offset=*/0, AudioFormat());
+               if (frame_callback != nullptr) {
+                       frame_callback(-1, AVRational{1, TIMEBASE}, -1, AVRational{1, TIMEBASE}, timecode++,
+                               video_frame, /*video_offset=*/0, video_format,
+                               FrameAllocator::Frame(), /*audio_offset=*/0, AudioFormat());
+               }
                last_frame_was_connected = false;
        }
 
@@ -446,6 +470,27 @@ void FFmpegCapture::send_disconnected_frame()
        }
 }
 
+AVPixelFormat get_vaapi_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
+{
+       for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
+               for (int i = 0;; ++i) {  // Termination condition inside loop.
+                       const AVCodecHWConfig *config = avcodec_get_hw_config(ctx->codec, i);
+                       if (config == nullptr) {  // End of list.
+                               fprintf(stderr, "Decoder %s does not support device.\n", ctx->codec->name);
+                               break;
+                       }
+                       if (config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX &&
+                           config->device_type == AV_HWDEVICE_TYPE_VAAPI &&
+                           config->pix_fmt == *fmt_ptr) {
+                               return config->pix_fmt;
+                       }
+               }
+       }
+
+       // We found no VA-API formats, so take the best software format.
+       return fmt[0];
+}
+
 bool FFmpegCapture::play_video(const string &pathname)
 {
        // Note: Call before open, not after; otherwise, there's a race.
@@ -459,6 +504,8 @@ bool FFmpegCapture::play_video(const string &pathname)
        } else {
                last_modified = buf.st_mtim;
        }
+       last_colorspace = static_cast<AVColorSpace>(-1);
+       last_chroma_location = static_cast<AVChromaLocation>(-1);
 
        AVFormatContextWithCloser format_ctx;
        if (srt_sock == -1) {
@@ -469,7 +516,7 @@ bool FFmpegCapture::play_video(const string &pathname)
        } else {
 #ifdef HAVE_SRT
                // SRT socket, already opened.
-               AVInputFormat *mpegts_fmt = av_find_input_format("mpegts");
+               const AVInputFormat *mpegts_fmt = av_find_input_format("mpegts");
                format_ctx = avformat_open_input_unique(&FFmpegCapture::read_srt_thunk, this,
                        mpegts_fmt, /*options=*/nullptr,
                        AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
@@ -499,7 +546,8 @@ bool FFmpegCapture::play_video(const string &pathname)
 
        // Open video decoder.
        const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
-       AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
+       const AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
+
        video_timebase = format_ctx->streams[video_stream_index]->time_base;
        AVCodecContextWithDeleter video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
        if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
@@ -510,6 +558,18 @@ bool FFmpegCapture::play_video(const string &pathname)
                fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
                return false;
        }
+
+       // Seemingly, it's not too easy to make something that just initializes
+       // “whatever goes”, so we don't get VDPAU or CUDA here without enumerating
+       // through several different types. VA-API will do for now.
+       AVBufferRef *hw_device_ctx = nullptr;
+       if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) < 0 || true) {
+               fprintf(stderr, "Failed to initialize VA-API for FFmpeg acceleration. Decoding video in software.\n");
+       } else {
+               video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
+               video_codec_ctx->get_format = get_vaapi_hw_format;
+       }
+
        if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
                fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
                return false;
@@ -530,7 +590,7 @@ bool FFmpegCapture::play_video(const string &pathname)
                        fprintf(stderr, "%s: Cannot fill audio codec parameters\n", pathname.c_str());
                        return false;
                }
-               AVCodec *audio_codec = avcodec_find_decoder(audio_codecpar->codec_id);
+               const AVCodec *audio_codec = avcodec_find_decoder(audio_codecpar->codec_id);
                if (audio_codec == nullptr) {
                        fprintf(stderr, "%s: Cannot find audio decoder\n", pathname.c_str());
                        return false;
@@ -572,7 +632,7 @@ bool FFmpegCapture::play_video(const string &pathname)
                                // so don't try).
                                return true;
                        }
-                       if (av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
+                       if (true || av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
                                fprintf(stderr, "%s: Rewind failed, not looping.\n", pathname.c_str());
                                return true;
                        }
@@ -692,9 +752,11 @@ bool FFmpegCapture::play_video(const string &pathname)
                                        timecode += MAX_FPS * 2 + 1;
                                }
                                last_neutral_color = get_neutral_color(frame->metadata);
-                               frame_callback(frame->pts, video_timebase, audio_pts, audio_timebase, timecode++,
-                                       video_frame.get_and_release(), 0, video_format,
-                                       audio_frame.get_and_release(), 0, audio_format);
+                               if (frame_callback != nullptr) {
+                                       frame_callback(frame->pts, video_timebase, audio_pts, audio_timebase, timecode++,
+                                               video_frame.get_and_release(), 0, video_format,
+                                               audio_frame.get_and_release(), 0, audio_format);
+                               }
                                first_frame = false;
                                last_frame = steady_clock::now();
                                last_frame_was_connected = true;
@@ -799,13 +861,16 @@ AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCo
                        if (pkt.stream_index == audio_stream_index && audio_callback != nullptr) {
                                audio_callback(&pkt, format_ctx->streams[audio_stream_index]->time_base);
                        }
-                       if (pkt.stream_index == video_stream_index) {
+                       if (pkt.stream_index == video_stream_index && video_callback != nullptr) {
+                               video_callback(&pkt, format_ctx->streams[video_stream_index]->time_base);
+                       }
+                       if (pkt.stream_index == video_stream_index && global_flags.transcode_video) {
                                if (avcodec_send_packet(video_codec_ctx, &pkt) < 0) {
                                        fprintf(stderr, "%s: Cannot send packet to video codec.\n", pathname.c_str());
                                        *error = true;
                                        return AVFrameWithDeleter(nullptr);
                                }
-                       } else if (pkt.stream_index == audio_stream_index) {
+                       } else if (pkt.stream_index == audio_stream_index && global_flags.transcode_audio) {
                                has_audio = true;
                                if (avcodec_send_packet(audio_codec_ctx, &pkt) < 0) {
                                        fprintf(stderr, "%s: Cannot send packet to audio codec.\n", pathname.c_str());
@@ -842,6 +907,21 @@ AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCo
                // Decode video, if we have a frame.
                int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
                if (err == 0) {
+                       if (video_avframe->format == AV_PIX_FMT_VAAPI) {
+                               // Get the frame down to the CPU. (TODO: See if we can keep it
+                               // on the GPU all the way, since it will be going up again later.
+                               // However, this only works if the OpenGL GPU is the same one.)
+                               AVFrameWithDeleter sw_frame = av_frame_alloc_unique();
+                               int err = av_hwframe_transfer_data(sw_frame.get(), video_avframe.get(), 0);
+                               if (err != 0) {
+                                       fprintf(stderr, "%s: Cannot transfer hardware video frame to software.\n", pathname.c_str());
+                                       *error = true;
+                                       return AVFrameWithDeleter(nullptr);
+                               }
+                               sw_frame->pts = video_avframe->pts;
+                               sw_frame->pkt_duration = video_avframe->pkt_duration;
+                               video_avframe = move(sw_frame);
+                       }
                        frame_finished = true;
                        break;
                } else if (err != AVERROR(EAGAIN)) {
@@ -1002,7 +1082,7 @@ UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &
                video_frame->len = (frame_width(frame) * 2) * frame_height(frame);
 
                const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
-               current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg);
+               current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
        } else {
                assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
                const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
@@ -1021,7 +1101,7 @@ UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &
 
                video_frame->len = frame_width(frame) * frame_height(frame) + 2 * chroma_width * chroma_height;
 
-               current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg);
+               current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
        }
        sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
 
@@ -1052,7 +1132,7 @@ unsigned FFmpegCapture::frame_height(const AVFrame *frame) const
        if (height == 0) {
                return frame->height;
        } else {
-               return width;
+               return height;
        }
 }