]> git.sesse.net Git - nageru/blobdiff - nageru/ffmpeg_capture.cpp
Improve selection of software formats on hwaccel fallback.
[nageru] / nageru / ffmpeg_capture.cpp
index 8f174ad018f7e236c896b7458af7e3248101a62f..5cf1d8c9ec75f35a6932a6bd7a282914c3017d92 100644 (file)
@@ -26,6 +26,7 @@ extern "C" {
 #include <cstdint>
 #include <utility>
 #include <vector>
+#include <unordered_set>
 
 #include <Eigen/Core>
 #include <Eigen/LU>
@@ -438,9 +439,11 @@ 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;
        }
 
@@ -452,24 +455,63 @@ void FFmpegCapture::send_disconnected_frame()
        }
 }
 
-AVPixelFormat get_vaapi_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
+template<AVHWDeviceType type>
+AVPixelFormat get_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) {
+       bool found_config_of_right_type = false;
+       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.
+                       break;
+               }
+               if (!(config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX) ||
+                   config->device_type != type) {
+                       // Not interesting for us.
+                       continue;
+               }
+
+               // We have a config of the right type, but does it actually support
+               // the pixel format we want? (Seemingly, FFmpeg's way of signaling errors
+               // is to just replace the pixel format with a software-decoded one,
+               // such as yuv420p.)
+               found_config_of_right_type = true;
+               for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
+                       if (config->pix_fmt == *fmt_ptr) {
+                               fprintf(stderr, "Initialized '%s' hardware decoding for codec '%s'.\n",
+                                       av_hwdevice_get_type_name(type), ctx->codec->name);
+                               if (ctx->profile == FF_PROFILE_H264_BASELINE) {
+                                       fprintf(stderr, "WARNING: Stream claims to be H.264 Baseline, which is generally poorly supported in hardware decoders.\n");
+                                       fprintf(stderr, "         Consider encoding it as Constrained Baseline, Main or High instead.\n");
+                                       fprintf(stderr, "         Decoding might fail and fall back to software.\n");
+                               }
                                return config->pix_fmt;
                        }
                }
+               fprintf(stderr, "Decoder '%s' supports only these pixel formats:", ctx->codec->name);
+               unordered_set<AVPixelFormat> seen;
+               for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
+                       if (!seen.count(*fmt_ptr)) {
+                               fprintf(stderr, " %s", av_get_pix_fmt_name(*fmt_ptr));
+                               seen.insert(*fmt_ptr);
+                       }
+               }
+               fprintf(stderr, " (wanted %s for hardware acceleration)\n", av_get_pix_fmt_name(config->pix_fmt));
+
+       }
+
+       if (!found_config_of_right_type) {
+               fprintf(stderr, "Decoder '%s' does not support device type '%s'.\n", ctx->codec->name, av_hwdevice_get_type_name(type));
        }
 
-       // We found no VA-API formats, so take the best software format.
+       // We found no VA-API formats, so take the first software format.
+       for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
+               if ((av_pix_fmt_desc_get(*fmt_ptr)->flags & AV_PIX_FMT_FLAG_HWACCEL) == 0) {
+                       fprintf(stderr, "Falling back to software format %s.\n", av_get_pix_fmt_name(*fmt_ptr));
+                       return *fmt_ptr;
+               }
+       }
+
+       // Fallback: Just return anything. (Should never really happen.)
        return fmt[0];
 }
 
@@ -498,7 +540,7 @@ 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 });
@@ -528,7 +570,7 @@ 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);
@@ -542,14 +584,24 @@ bool FFmpegCapture::play_video(const string &pathname)
        }
 
        // 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.
+       // “whatever goes”, so we don't get CUDA or VULKAN or whatever here
+       // without enumerating through several different types.
+       // VA-API and VDPAU will do for now. We prioritize VDPAU for the
+       // simple reason that there's a VA-API-via-VDPAU emulation for NVidia
+       // cards that seems to work, but just hangs when trying to transfer the frame.
+       //
+       // Note that we don't actually check codec support beforehand,
+       // so if you have a low-end VDPAU device but a high-end VA-API device,
+       // you lose out on the extra codec support from the latter.
        AVBufferRef *hw_device_ctx = nullptr;
-       if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) < 0) {
-               fprintf(stderr, "Failed to initialize VA-API for FFmpeg acceleration. Decoding video in software.\n");
-       } else {
+       if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VDPAU, nullptr, nullptr, 0) >= 0) {
+               video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
+               video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VDPAU>;
+       } else if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) >= 0) {
                video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
-               video_codec_ctx->get_format = get_vaapi_hw_format;
+               video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VAAPI>;
+       } else {
+               fprintf(stderr, "Failed to initialize VA-API or VDPAU for FFmpeg acceleration. Decoding video in software.\n");
        }
 
        if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
@@ -572,7 +624,7 @@ 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;
@@ -589,6 +641,7 @@ bool FFmpegCapture::play_video(const string &pathname)
 
        // Main loop.
        bool first_frame = true;
+       int consecutive_errors = 0;
        while (!producer_thread_should_quit.should_quit()) {
                if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) {
                        return true;
@@ -605,7 +658,14 @@ bool FFmpegCapture::play_video(const string &pathname)
                AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(), audio_codec_ctx.get(),
                        pathname, video_stream_index, audio_stream_index, subtitle_stream_index, audio_frame.get(), &audio_format, &audio_pts, &error);
                if (error) {
-                       return false;
+                       if (++consecutive_errors >= 100) {
+                               fprintf(stderr, "More than 100 consecutive video frames, aborting playback.\n");
+                               return false;
+                       } else {
+                               continue;
+                       }
+               } else {
+                       consecutive_errors = 0;
                }
                if (frame == nullptr) {
                        // EOF. Loop back to the start if we can.
@@ -713,7 +773,7 @@ bool FFmpegCapture::play_video(const string &pathname)
                                                1e3 * duration<double>(now - next_frame_start).count());
                                        pts_origin = frame->pts;
                                        start = next_frame_start = now;
-                                       timecode += MAX_FPS * 2 + 1;
+                                       timecode += TYPICAL_FPS * 2 + 1;
                                }
                        }
                        bool finished_wakeup;
@@ -731,12 +791,14 @@ bool FFmpegCapture::play_video(const string &pathname)
                                        // Make sure to get the audio resampler reset. (This is a hack;
                                        // ideally, the frame callback should just accept a way to signal
                                        // audio discontinuity.)
-                                       timecode += MAX_FPS * 2 + 1;
+                                       timecode += TYPICAL_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;
@@ -841,13 +903,16 @@ 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());
@@ -884,7 +949,8 @@ 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) {
+                       if (video_avframe->format == AV_PIX_FMT_VAAPI ||
+                           video_avframe->format == AV_PIX_FMT_VDPAU) {
                                // 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.)
@@ -1109,7 +1175,7 @@ unsigned FFmpegCapture::frame_height(const AVFrame *frame) const
        if (height == 0) {
                return frame->height;
        } else {
-               return width;
+               return height;
        }
 }