using namespace std;
using namespace std::chrono;
using namespace bmusb;
+using namespace movit;
namespace {
return (buf.st_mtim.tv_sec != ts.tv_sec || buf.st_mtim.tv_nsec != ts.tv_nsec);
}
+bool is_full_range(const AVPixFmtDescriptor *desc)
+{
+ // This is horrible, but there's no better way that I know of.
+ return (strchr(desc->name, 'j') != nullptr);
+}
+
+AVPixelFormat decide_dst_format(AVPixelFormat src_format, bmusb::PixelFormat dst_format_type)
+{
+ if (dst_format_type == bmusb::PixelFormat_8BitBGRA) {
+ return AV_PIX_FMT_BGRA;
+ }
+
+ assert(dst_format_type == bmusb::PixelFormat_8BitYCbCrPlanar);
+
+ // If this is a non-Y'CbCr format, just convert to 4:4:4 Y'CbCr
+ // and be done with it. It's too strange to spend a lot of time on.
+ // (Let's hope there's no alpha.)
+ const AVPixFmtDescriptor *src_desc = av_pix_fmt_desc_get(src_format);
+ if (src_desc == nullptr ||
+ src_desc->nb_components != 3 ||
+ (src_desc->flags & AV_PIX_FMT_FLAG_RGB)) {
+ return AV_PIX_FMT_YUV444P;
+ }
+
+ // The best for us would be Cb and Cr together if possible,
+ // but FFmpeg doesn't support that except in the special case of
+ // NV12, so we need to go to planar even for the case of NV12.
+ // Thus, look for the closest (but no worse) 8-bit planar Y'CbCr format
+ // that matches in color range. (This will also include the case of
+ // the source format already being acceptable.)
+ bool src_full_range = is_full_range(src_desc);
+ const char *best_format = "yuv444p";
+ unsigned best_score = numeric_limits<unsigned>::max();
+ for (const AVPixFmtDescriptor *desc = av_pix_fmt_desc_next(nullptr);
+ desc;
+ desc = av_pix_fmt_desc_next(desc)) {
+ // Find planar Y'CbCr formats only.
+ if (desc->nb_components != 3) continue;
+ if (desc->flags & AV_PIX_FMT_FLAG_RGB) continue;
+ if (!(desc->flags & AV_PIX_FMT_FLAG_PLANAR)) continue;
+ if (desc->comp[0].plane != 0 ||
+ desc->comp[1].plane != 1 ||
+ desc->comp[2].plane != 2) continue;
+
+ // 8-bit formats only.
+ if (desc->flags & AV_PIX_FMT_FLAG_BE) continue;
+ 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;
+ if (chroma_w_diff < 0 || chroma_h_diff < 0)
+ continue;
+
+ // Matching full/limited range only.
+ if (is_full_range(desc) != src_full_range)
+ continue;
+
+ // Pick something with as little excess chroma resolution as possible.
+ unsigned score = (1 << (chroma_w_diff)) << chroma_h_diff;
+ if (score < best_score) {
+ best_score = score;
+ best_format = desc->name;
+ }
+ }
+ return av_get_pix_fmt(best_format);
+}
+
+YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *frame)
+{
+ YCbCrFormat format;
+ AVColorSpace colorspace = av_frame_get_colorspace(frame);
+ switch (colorspace) {
+ case AVCOL_SPC_BT709:
+ format.luma_coefficients = YCBCR_REC_709;
+ break;
+ case AVCOL_SPC_BT470BG:
+ case AVCOL_SPC_SMPTE170M:
+ case AVCOL_SPC_SMPTE240M:
+ format.luma_coefficients = YCBCR_REC_601;
+ break;
+ case AVCOL_SPC_BT2020_NCL:
+ format.luma_coefficients = YCBCR_REC_2020;
+ break;
+ case AVCOL_SPC_UNSPECIFIED:
+ 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);
+ format.luma_coefficients = YCBCR_REC_709;
+ break;
+ }
+
+ format.full_range = is_full_range(desc);
+ format.num_levels = 1 << desc->comp[0].depth;
+ format.chroma_subsampling_x = 1 << desc->log2_chroma_w;
+ format.chroma_subsampling_y = 1 << desc->log2_chroma_h;
+
+ switch (frame->chroma_location) {
+ case AVCHROMA_LOC_LEFT:
+ format.cb_x_position = 0.0;
+ format.cb_y_position = 0.5;
+ break;
+ case AVCHROMA_LOC_CENTER:
+ format.cb_x_position = 0.5;
+ format.cb_y_position = 0.5;
+ break;
+ case AVCHROMA_LOC_TOPLEFT:
+ format.cb_x_position = 0.0;
+ format.cb_y_position = 0.0;
+ break;
+ case AVCHROMA_LOC_TOP:
+ format.cb_x_position = 0.5;
+ format.cb_y_position = 0.0;
+ break;
+ case AVCHROMA_LOC_BOTTOMLEFT:
+ format.cb_x_position = 0.0;
+ format.cb_y_position = 1.0;
+ break;
+ case AVCHROMA_LOC_BOTTOM:
+ format.cb_x_position = 0.5;
+ format.cb_y_position = 1.0;
+ break;
+ default:
+ fprintf(stderr, "Unknown chroma location coefficient enum %d from FFmpeg; choosing Rec. 709.\n",
+ frame->chroma_location);
+ format.cb_x_position = 0.5;
+ format.cb_y_position = 0.5;
+ break;
+ }
+
+ format.cr_x_position = format.cb_x_position;
+ format.cr_y_position = format.cb_y_position;
+ return format;
+}
+
} // namespace
FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned height)
- : filename(filename), width(width), height(height)
+ : filename(filename), width(width), height(height), video_timebase{1, 1}
{
// Not really used for anything.
description = "Video: " + filename;
}
const AVCodecParameters *codecpar = format_ctx->streams[video_stream_index]->codecpar;
- AVRational video_timebase = format_ctx->streams[video_stream_index]->time_base;
+ video_timebase = format_ctx->streams[video_stream_index]->time_base;
AVCodecContextWithDeleter codec_ctx = avcodec_alloc_context3_unique(nullptr);
if (avcodec_parameters_to_context(codec_ctx.get(), codecpar) < 0) {
fprintf(stderr, "%s: Cannot fill codec parameters\n", pathname.c_str());
codec_ctx.get(), avcodec_close);
internal_rewind();
- double rate = 1.0;
-
- unique_ptr<SwsContext, decltype(sws_freeContext)*> sws_ctx(nullptr, sws_freeContext);
- int sws_last_width = -1, sws_last_height = -1;
// Main loop.
while (!producer_thread_should_quit.should_quit()) {
- // Process any queued commands from other threads.
- vector<QueuedCommand> commands;
- {
- lock_guard<mutex> lock(queue_mu);
- swap(commands, command_queue);
+ if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) {
+ return true;
}
- for (const QueuedCommand &cmd : commands) {
- switch (cmd.command) {
- case QueuedCommand::REWIND:
- if (av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
- fprintf(stderr, "%s: Rewind failed, stopping play.\n", pathname.c_str());
- }
- // If the file has changed since last time, return to get it reloaded.
- // Note that depending on how you move the file into place, you might
- // end up corrupting the one you're already playing, so this path
- // might not trigger.
- if (changed_since(pathname, last_modified)) {
- return true;
- }
- internal_rewind();
- break;
- case QueuedCommand::CHANGE_RATE:
- start = next_frame_start;
- pts_origin = last_pts;
- rate = cmd.new_rate;
- break;
- }
+ bool error;
+ AVFrameWithDeleter frame = decode_frame(format_ctx.get(), codec_ctx.get(), pathname, video_stream_index, &error);
+ if (error) {
+ return false;
}
-
- // Read packets until we have a frame or there are none left.
- int frame_finished = 0;
- AVFrameWithDeleter frame = av_frame_alloc_unique();
- bool eof = false;
- do {
- AVPacket pkt;
- unique_ptr<AVPacket, decltype(av_packet_unref)*> pkt_cleanup(
- &pkt, av_packet_unref);
- av_init_packet(&pkt);
- pkt.data = nullptr;
- pkt.size = 0;
- if (av_read_frame(format_ctx.get(), &pkt) == 0) {
- if (pkt.stream_index != video_stream_index) {
- // Ignore audio for now.
- continue;
- }
- if (avcodec_send_packet(codec_ctx.get(), &pkt) < 0) {
- fprintf(stderr, "%s: Cannot send packet to codec.\n", pathname.c_str());
- return false;
- }
- } else {
- eof = true; // Or error, but ignore that for the time being.
- }
-
- int err = avcodec_receive_frame(codec_ctx.get(), frame.get());
- if (err == 0) {
- frame_finished = true;
- break;
- } else if (err != AVERROR(EAGAIN)) {
- fprintf(stderr, "%s: Cannot receive frame from codec.\n", pathname.c_str());
- return false;
- }
- } while (!eof);
-
- if (!frame_finished) {
+ if (frame == nullptr) {
// EOF. Loop back to the start if we can.
if (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());
continue;
}
- if (sws_ctx == nullptr || sws_last_width != frame->width || sws_last_height != frame->height) {
- sws_ctx.reset(
- sws_getContext(frame->width, frame->height, (AVPixelFormat)frame->format,
- width, height, AV_PIX_FMT_BGRA,
- SWS_BICUBIC, nullptr, nullptr, nullptr));
- sws_last_width = frame->width;
- sws_last_height = frame->height;
- }
- if (sws_ctx == nullptr) {
- fprintf(stderr, "%s: Could not create scaler context\n", pathname.c_str());
+ VideoFormat video_format = construct_video_format(frame.get(), video_timebase);
+ FrameAllocator::Frame video_frame = make_video_frame(frame.get(), pathname, &error);
+ if (error) {
return false;
}
- VideoFormat video_format;
- video_format.width = width;
- video_format.height = height;
- video_format.stride = width * 4;
- video_format.frame_rate_nom = video_timebase.den;
- video_format.frame_rate_den = av_frame_get_pkt_duration(frame.get()) * video_timebase.num;
- if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
- // Invalid frame rate.
- video_format.frame_rate_nom = 60;
- video_format.frame_rate_den = 1;
- }
- video_format.has_signal = true;
- video_format.is_connected = true;
-
- next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
- last_pts = frame->pts;
-
- FrameAllocator::Frame video_frame = video_frame_allocator->alloc_frame();
- if (video_frame.data != nullptr) {
- uint8_t *pic_data[4] = { video_frame.data, nullptr, nullptr, nullptr };
- int linesizes[4] = { int(video_format.stride), 0, 0, 0 };
- sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
- video_frame.len = video_format.stride * height;
- video_frame.received_timestamp = next_frame_start;
- }
-
FrameAllocator::Frame audio_frame;
AudioFormat audio_format;
audio_format.bits_per_sample = 32;
audio_format.num_channels = 8;
- producer_thread_should_quit.sleep_until(next_frame_start);
- frame_callback(timecode++,
- video_frame, 0, video_format,
- audio_frame, 0, audio_format);
+ for ( ;; ) {
+ next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
+ video_frame.received_timestamp = next_frame_start;
+ bool finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
+ if (finished_wakeup) {
+ frame_callback(timecode++,
+ video_frame, 0, video_format,
+ audio_frame, 0, audio_format);
+ break;
+ } else {
+ if (producer_thread_should_quit.should_quit()) break;
+
+ bool rewound = false;
+ if (process_queued_commands(format_ctx.get(), pathname, last_modified, &rewound)) {
+ return true;
+ }
+ // If we just rewound, drop this frame on the floor and be done.
+ if (rewound) {
+ video_frame_allocator->release_frame(video_frame);
+ break;
+ }
+ // OK, we didn't, so probably a rate change. Recalculate next_frame_start,
+ // but if it's now in the past, we'll reset the origin, so that we don't
+ // generate a huge backlog of frames that we need to run through quickly.
+ next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
+ steady_clock::time_point now = steady_clock::now();
+ if (next_frame_start < now) {
+ pts_origin = frame->pts;
+ start = next_frame_start = now;
+ }
+ }
+ }
+ last_pts = frame->pts;
}
return true;
}
pts_origin = last_pts = 0;
start = next_frame_start = steady_clock::now();
}
+
+bool FFmpegCapture::process_queued_commands(AVFormatContext *format_ctx, const std::string &pathname, timespec last_modified, bool *rewound)
+{
+ // Process any queued commands from other threads.
+ vector<QueuedCommand> commands;
+ {
+ lock_guard<mutex> lock(queue_mu);
+ swap(commands, command_queue);
+ }
+ for (const QueuedCommand &cmd : commands) {
+ switch (cmd.command) {
+ case QueuedCommand::REWIND:
+ if (av_seek_frame(format_ctx, /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
+ fprintf(stderr, "%s: Rewind failed, stopping play.\n", pathname.c_str());
+ }
+ // If the file has changed since last time, return to get it reloaded.
+ // Note that depending on how you move the file into place, you might
+ // end up corrupting the one you're already playing, so this path
+ // might not trigger.
+ if (changed_since(pathname, last_modified)) {
+ return true;
+ }
+ internal_rewind();
+ if (rewound != nullptr) {
+ *rewound = true;
+ }
+ break;
+
+ case QueuedCommand::CHANGE_RATE:
+ // Change the origin to the last played frame.
+ start = compute_frame_start(last_pts, pts_origin, video_timebase, start, rate);
+ pts_origin = last_pts;
+ rate = cmd.new_rate;
+ break;
+ }
+ }
+ return false;
+}
+
+AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCodecContext *codec_ctx, const std::string &pathname, int video_stream_index, bool *error)
+{
+ *error = false;
+
+ // Read packets until we have a frame or there are none left.
+ bool frame_finished = false;
+ AVFrameWithDeleter frame = av_frame_alloc_unique();
+ bool eof = false;
+ do {
+ AVPacket pkt;
+ unique_ptr<AVPacket, decltype(av_packet_unref)*> pkt_cleanup(
+ &pkt, av_packet_unref);
+ av_init_packet(&pkt);
+ pkt.data = nullptr;
+ pkt.size = 0;
+ if (av_read_frame(format_ctx, &pkt) == 0) {
+ if (pkt.stream_index != video_stream_index) {
+ // Ignore audio for now.
+ continue;
+ }
+ if (avcodec_send_packet(codec_ctx, &pkt) < 0) {
+ fprintf(stderr, "%s: Cannot send packet to codec.\n", pathname.c_str());
+ *error = true;
+ return AVFrameWithDeleter(nullptr);
+ }
+ } else {
+ eof = true; // Or error, but ignore that for the time being.
+ }
+
+ int err = avcodec_receive_frame(codec_ctx, frame.get());
+ if (err == 0) {
+ frame_finished = true;
+ break;
+ } else if (err != AVERROR(EAGAIN)) {
+ fprintf(stderr, "%s: Cannot receive frame from codec.\n", pathname.c_str());
+ *error = true;
+ return AVFrameWithDeleter(nullptr);
+ }
+ } while (!eof);
+
+ if (frame_finished)
+ return frame;
+ else
+ return AVFrameWithDeleter(nullptr);
+}
+
+VideoFormat FFmpegCapture::construct_video_format(const AVFrame *frame, AVRational video_timebase)
+{
+ VideoFormat video_format;
+ video_format.width = width;
+ video_format.height = height;
+ if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
+ video_format.stride = width * 4;
+ } else {
+ assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
+ video_format.stride = width;
+ }
+ video_format.frame_rate_nom = video_timebase.den;
+ video_format.frame_rate_den = av_frame_get_pkt_duration(frame) * video_timebase.num;
+ if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
+ // Invalid frame rate.
+ video_format.frame_rate_nom = 60;
+ video_format.frame_rate_den = 1;
+ }
+ video_format.has_signal = true;
+ video_format.is_connected = true;
+ return video_format;
+}
+
+FrameAllocator::Frame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &pathname, bool *error)
+{
+ *error = false;
+
+ FrameAllocator::Frame video_frame = video_frame_allocator->alloc_frame();
+ if (video_frame.data == nullptr) {
+ return video_frame;
+ }
+
+ if (sws_ctx == nullptr ||
+ sws_last_width != frame->width ||
+ sws_last_height != frame->height ||
+ sws_last_src_format != frame->format) {
+ sws_dst_format = decide_dst_format(AVPixelFormat(frame->format), pixel_format);
+ sws_ctx.reset(
+ sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
+ width, height, sws_dst_format,
+ SWS_BICUBIC, nullptr, nullptr, nullptr));
+ sws_last_width = frame->width;
+ sws_last_height = frame->height;
+ sws_last_src_format = frame->format;
+ }
+ if (sws_ctx == nullptr) {
+ fprintf(stderr, "%s: Could not create scaler context\n", pathname.c_str());
+ *error = true;
+ return video_frame;
+ }
+
+ uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
+ int linesizes[4] = { 0, 0, 0, 0 };
+ if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
+ pic_data[0] = video_frame.data;
+ linesizes[0] = width * 4;
+ video_frame.len = (width * 4) * height;
+ } else {
+ assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
+ const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
+
+ int chroma_width = AV_CEIL_RSHIFT(int(width), desc->log2_chroma_w);
+ int chroma_height = AV_CEIL_RSHIFT(int(height), desc->log2_chroma_h);
+
+ pic_data[0] = video_frame.data;
+ linesizes[0] = width;
+
+ pic_data[1] = pic_data[0] + width * height;
+ linesizes[1] = chroma_width;
+
+ pic_data[2] = pic_data[1] + chroma_width * chroma_height;
+ linesizes[2] = chroma_width;
+
+ video_frame.len = width * height + 2 * chroma_width * chroma_height;
+
+ current_frame_ycbcr_format = decode_ycbcr_format(desc, frame);
+ }
+ sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
+
+ return video_frame;
+}
projects / nageru / blobdiff