#include <cstdint>
#include <utility>
#include <vector>
+#include <unordered_set>
#include <Eigen/Core>
#include <Eigen/LU>
#include <srt/srt.h>
#endif
-#define FRAME_SIZE (8 << 20) // 8 MB.
-
using namespace std;
using namespace std::chrono;
using namespace bmusb;
if (desc->comp[0].depth != 8) continue;
// Same or better chroma resolution only.
- int chroma_w_diff = desc->log2_chroma_w - src_desc->log2_chroma_w;
- int chroma_h_diff = desc->log2_chroma_h - src_desc->log2_chroma_h;
+ int chroma_w_diff = src_desc->log2_chroma_w - desc->log2_chroma_w;
+ int chroma_h_diff = src_desc->log2_chroma_h - desc->log2_chroma_h;
if (chroma_w_diff < 0 || chroma_h_diff < 0)
continue;
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;
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;
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
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;
}
}
}
+template<AVHWDeviceType type>
+AVPixelFormat get_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
+{
+ 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 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];
+}
+
bool FFmpegCapture::play_video(const string &pathname)
{
// Note: Call before open, not after; otherwise, there's a race.
} 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) {
} 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 });
// 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) {
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 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_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_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) {
fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
return false;
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;
// 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;
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.
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;
// 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;
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());
// 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 ||
+ 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.)
+ 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)) {
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);
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);
}
+
+ // FIXME: Currently, if the video is too high-res for one of the allocated
+ // frames, we simply refuse to scale it here to avoid crashes. It would be better
+ // if we could somehow signal getting larger frames, especially as 4K is a thing now.
+ if (video_frame->len > FRAME_SIZE) {
+ fprintf(stderr, "%s: Decoded frame would be larger than supported FRAME_SIZE (%zu > %u), not decoding.\n", pathname.c_str(), video_frame->len, FRAME_SIZE);
+ *error = true;
+ return video_frame;
+ }
+
sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
return video_frame;
if (height == 0) {
return frame->height;
} else {
- return width;
+ return height;
}
}