videodemuxer: added seek functionality (#311)
* demux add seek functionality
* addressed review comments
[ROCm/rocdecode commit: 77e4c3150a]
这个提交包含在：
Rajy Rawther
GitHub
@@ -40,6 +40,83 @@ extern "C" {
|
||||
* \brief AMD The rocDecode video demuxer API.
|
||||
*/
|
||||
|
||||
/**
|
||||
* @brief Enum for Seek mode
|
||||
*
|
||||
*/
|
||||
typedef enum SeekModeEnum {
|
||||
SEEK_MODE_EXACT_FRAME = 0,
|
||||
SEEK_MODE_PREV_KEY_FRAME = 1,
|
||||
SEEK_MODE_NUM,
|
||||
} SeekMode;
|
||||
|
||||
/**
|
||||
* @brief Enum for Seek Criteria
|
||||
*
|
||||
*/
|
||||
typedef enum SeekCriteriaEnum {
|
||||
SEEK_CRITERIA_FRAME_NUM = 0,
|
||||
SEEK_CRITERIA_TIME_STAMP = 1,
|
||||
SEEK_CRITERIA_NUM,
|
||||
} SeekCriteria;
|
||||
|
||||
struct PacketData {
|
||||
int32_t key;
|
||||
int64_t pts;
|
||||
int64_t dts;
|
||||
uint64_t pos;
|
||||
uintptr_t bsl_data;
|
||||
uint64_t bsl;
|
||||
uint64_t duration;
|
||||
};
|
||||
|
||||
class VideoSeekContext {
|
||||
public:
|
||||
VideoSeekContext()
|
||||
: use_seek_(false), seek_frame_(0), seek_mode_(SEEK_MODE_PREV_KEY_FRAME), seek_crit_(SEEK_CRITERIA_FRAME_NUM),
|
||||
out_frame_pts_(0), out_frame_duration_(0), num_frames_decoded_(0U) {}
|
||||
|
||||
VideoSeekContext(uint64_t frame_id)
|
||||
: use_seek_(true), seek_frame_(frame_id), seek_mode_(SEEK_MODE_PREV_KEY_FRAME),
|
||||
seek_crit_(SEEK_CRITERIA_FRAME_NUM), out_frame_pts_(0), out_frame_duration_(0), num_frames_decoded_(0U) {}
|
||||
|
||||
VideoSeekContext& operator=(const VideoSeekContext& other) {
|
||||
use_seek_ = other.use_seek_;
|
||||
seek_frame_ = other.seek_frame_;
|
||||
seek_mode_ = other.seek_mode_;
|
||||
seek_crit_ = other.seek_crit_;
|
||||
out_frame_pts_ = other.out_frame_pts_;
|
||||
out_frame_duration_ = other.out_frame_duration_;
|
||||
num_frames_decoded_ = other.num_frames_decoded_;
|
||||
return *this;
|
||||
}
|
||||
|
||||
/* Will be set to false when not seeking, true otherwise;
|
||||
*/
|
||||
bool use_seek_;
|
||||
|
||||
/* Frame we want to get. Set by user.
|
||||
* Shall be set to frame timestamp in case seek is done by time.
|
||||
*/
|
||||
uint64_t seek_frame_;
|
||||
|
||||
/* Mode in which we seek. */
|
||||
SeekMode seek_mode_;
|
||||
|
||||
/* Criteria by which we seek. */
|
||||
SeekCriteria seek_crit_;
|
||||
|
||||
/* PTS of frame found after seek. */
|
||||
int64_t out_frame_pts_;
|
||||
|
||||
/* Duration of frame found after seek. */
|
||||
int64_t out_frame_duration_;
|
||||
|
||||
/* Number of frames that were decoded during seek. */
|
||||
uint64_t num_frames_decoded_;
|
||||
|
||||
};
|
||||
|
||||
|
||||
// Video Demuxer Interface class
|
||||
class VideoDemuxer {
|
||||
@@ -54,13 +131,29 @@ class VideoDemuxer {
|
||||
VideoDemuxer(StreamProvider *stream_provider) : VideoDemuxer(CreateFmtContextUtil(stream_provider)) {av_io_ctx_ = av_fmt_input_ctx_->pb;}
|
||||
~VideoDemuxer();
|
||||
bool Demux(uint8_t **video, int *video_size, int64_t *pts = nullptr);
|
||||
bool Seek(VideoSeekContext& seek_ctx, uint8_t** pp_video, int* video_size);
|
||||
const uint32_t GetWidth() const { return width_;}
|
||||
const uint32_t GetHeight() const { return height_;}
|
||||
const uint32_t GetChromaHeight() const { return chroma_height_;}
|
||||
const uint32_t GetBitDepth() const { return bit_depth_;}
|
||||
const uint32_t GetBytePerPixel() const { return byte_per_pixel_;}
|
||||
const uint32_t GetBitRate() const { return bit_rate_;}
|
||||
const double GetFrameRate() const {return frame_rate_.den != 0 ? frame_rate_.num / frame_rate_.den : 0;};
|
||||
const double GetFrameRate() const {return frame_rate_;};
|
||||
bool IsVFR() const { return frame_rate_ != avg_frame_rate_; };
|
||||
int64_t TsFromTime(double ts_sec) {
|
||||
// Convert integer timestamp representation to AV_TIME_BASE and switch to fixed_point
|
||||
auto const ts_tbu = llround(ts_sec * AV_TIME_BASE);
|
||||
|
||||
// Rescale the timestamp to value represented in stream base units;
|
||||
AVRational time_factor = {1, AV_TIME_BASE};
|
||||
return av_rescale_q(ts_tbu, time_factor, av_fmt_input_ctx_->streams[av_stream_]->time_base);
|
||||
}
|
||||
|
||||
int64_t TsFromFrameNumber(int64_t frame_num) {
|
||||
auto const ts_sec = static_cast<double>(frame_num) / frame_rate_;
|
||||
return TsFromTime(ts_sec);
|
||||
}
|
||||
|
||||
private:
|
||||
VideoDemuxer(AVFormatContext *av_fmt_input_ctx);
|
||||
AVFormatContext *CreateFmtContextUtil(StreamProvider *stream_provider);
|
||||
@@ -73,12 +166,14 @@ class VideoDemuxer {
|
||||
AVBSFContext *av_bsf_ctx_ = nullptr;
|
||||
AVCodecID av_video_codec_id_;
|
||||
AVPixelFormat chroma_format_;
|
||||
AVRational frame_rate_ = {};
|
||||
double frame_rate_ = 0.0;
|
||||
double avg_frame_rate_ = 0.0;
|
||||
uint8_t *data_with_header_ = nullptr;
|
||||
int av_stream_ = 0;
|
||||
bool is_h264_ = false;
|
||||
bool is_hevc_ = false;
|
||||
bool is_mpeg4_ = false;
|
||||
bool is_seekable_ = false;
|
||||
int64_t default_time_scale_ = 1000;
|
||||
double time_base_ = 0.0;
|
||||
uint32_t frame_count_ = 0;
|
||||
@@ -197,7 +292,10 @@ VideoDemuxer::VideoDemuxer(AVFormatContext *av_fmt_input_ctx) : av_fmt_input_ctx
|
||||
height_ = av_fmt_input_ctx_->streams[av_stream_]->codecpar->height;
|
||||
chroma_format_ = (AVPixelFormat)av_fmt_input_ctx_->streams[av_stream_]->codecpar->format;
|
||||
bit_rate_ = av_fmt_input_ctx_->streams[av_stream_]->codecpar->bit_rate;
|
||||
frame_rate_ = av_fmt_input_ctx_->streams[av_stream_]->r_frame_rate;
|
||||
if (av_fmt_input_ctx_->streams[av_stream_]->r_frame_rate.den != 0)
|
||||
frame_rate_ = static_cast<double>(av_fmt_input_ctx_->streams[av_stream_]->r_frame_rate.num) / static_cast<double>(av_fmt_input_ctx_->streams[av_stream_]->r_frame_rate.den);
|
||||
if (av_fmt_input_ctx_->streams[av_stream_]->avg_frame_rate.den != 0)
|
||||
avg_frame_rate_ = static_cast<double>(av_fmt_input_ctx_->streams[av_stream_]->avg_frame_rate.num) / static_cast<double>(av_fmt_input_ctx_->streams[av_stream_]->avg_frame_rate.den);
|
||||
|
||||
switch (chroma_format_) {
|
||||
case AV_PIX_FMT_YUV420P10LE:
|
||||
@@ -255,6 +353,9 @@ VideoDemuxer::VideoDemuxer(AVFormatContext *av_fmt_input_ctx) : av_fmt_input_ctx
|
||||
|| !strcmp(av_fmt_input_ctx_->iformat->long_name, "FLV (Flash Video)")
|
||||
|| !strcmp(av_fmt_input_ctx_->iformat->long_name, "Matroska / WebM"));
|
||||
|
||||
// Check if the input file allow seek functionality.
|
||||
is_seekable_ = av_fmt_input_ctx_->iformat->read_seek || av_fmt_input_ctx_->iformat->read_seek2;
|
||||
|
||||
if (is_h264_) {
|
||||
const AVBitStreamFilter *bsf = av_bsf_get_by_name("h264_mp4toannexb");
|
||||
if (!bsf) {
|
||||
@@ -293,6 +394,132 @@ VideoDemuxer::VideoDemuxer(AVFormatContext *av_fmt_input_ctx) : av_fmt_input_ctx
|
||||
}
|
||||
}
|
||||
|
||||
bool VideoDemuxer::Seek(VideoSeekContext& seek_ctx, uint8_t** pp_video, int* video_size) {
|
||||
/* !!! IMPORTANT !!!
|
||||
* Across this function, packet decode timestamp (DTS) values are used to
|
||||
* compare given timestamp against. This is done because DTS values shall
|
||||
* monotonically increase during the course of decoding unlike PTS values
|
||||
* which may be affected by frame reordering due to B frames.
|
||||
*/
|
||||
|
||||
if (!is_seekable_) {
|
||||
std::cerr << "ERROR: Seek isn't supported for this input." << std::endl;
|
||||
return false;
|
||||
}
|
||||
|
||||
if (IsVFR() && (SEEK_CRITERIA_FRAME_NUM == seek_ctx.seek_crit_)) {
|
||||
std::cerr << "ERROR: Can't seek by frame number in VFR sequences. Seek by timestamp instead." << std::endl;
|
||||
return false;
|
||||
}
|
||||
|
||||
// Seek for single frame;
|
||||
auto seek_frame = [&](VideoSeekContext const& seek_ctx, int flags) {
|
||||
bool seek_backward = true;
|
||||
int64_t timestamp = 0;
|
||||
int ret = 0;
|
||||
|
||||
switch (seek_ctx.seek_crit_) {
|
||||
case SEEK_CRITERIA_FRAME_NUM:
|
||||
timestamp = TsFromFrameNumber(seek_ctx.seek_frame_);
|
||||
ret = av_seek_frame(av_fmt_input_ctx_, av_stream_, timestamp, seek_backward ? AVSEEK_FLAG_BACKWARD | flags : flags);
|
||||
break;
|
||||
case SEEK_CRITERIA_TIME_STAMP:
|
||||
timestamp = TsFromTime(seek_ctx.seek_frame_);
|
||||
ret = av_seek_frame(av_fmt_input_ctx_, av_stream_, timestamp, seek_backward ? AVSEEK_FLAG_BACKWARD | flags : flags);
|
||||
break;
|
||||
default:
|
||||
std::cerr << "ERROR: Invalid seek mode" << std::endl;
|
||||
ret = -1;
|
||||
}
|
||||
|
||||
if (ret < 0) {
|
||||
throw std::runtime_error("ERROR: seeking for frame: ");
|
||||
}
|
||||
};
|
||||
|
||||
// Check if frame satisfies seek conditions;
|
||||
auto is_seek_done = [&](PacketData& pkt_data, VideoSeekContext const& seek_ctx) {
|
||||
int64_t target_ts = 0;
|
||||
|
||||
switch (seek_ctx.seek_crit_) {
|
||||
case SEEK_CRITERIA_FRAME_NUM:
|
||||
target_ts = TsFromFrameNumber(seek_ctx.seek_frame_);
|
||||
break;
|
||||
case SEEK_CRITERIA_TIME_STAMP:
|
||||
target_ts = TsFromTime(seek_ctx.seek_frame_);
|
||||
break;
|
||||
default:
|
||||
std::cerr << "ERROR::Invalid seek criteria" << std::endl;
|
||||
return -1;
|
||||
}
|
||||
|
||||
if (pkt_data.dts == target_ts) {
|
||||
return 0;
|
||||
}
|
||||
else if (pkt_data.dts > target_ts) {
|
||||
return 1;
|
||||
}
|
||||
else {
|
||||
return -1;
|
||||
};
|
||||
};
|
||||
|
||||
/* This will seek for exact frame number;
|
||||
* Note that decoder may not be able to decode such frame; */
|
||||
auto seek_for_exact_frame = [&](PacketData& pkt_data, VideoSeekContext& seek_ctx) {
|
||||
// Repetititive seek until seek condition is satisfied;
|
||||
VideoSeekContext tmp_ctx(seek_ctx.seek_frame_);
|
||||
seek_frame(tmp_ctx, AVSEEK_FLAG_ANY);
|
||||
|
||||
int seek_done = 0;
|
||||
do {
|
||||
if (!Demux(pp_video, video_size)) {
|
||||
break;
|
||||
}
|
||||
seek_done = is_seek_done(pkt_data, seek_ctx);
|
||||
|
||||
// We've gone too far and need to seek backwards;
|
||||
if (seek_done > 0) {
|
||||
tmp_ctx.seek_frame_--;
|
||||
seek_frame(tmp_ctx, AVSEEK_FLAG_ANY);
|
||||
}
|
||||
// Need to read more frames until we reach requested number;
|
||||
else if (seek_done < 0) {
|
||||
continue;
|
||||
}
|
||||
} while (seek_done != 0);
|
||||
|
||||
seek_ctx.out_frame_pts_ = pkt_data.pts;
|
||||
seek_ctx.out_frame_duration_ = pkt_data.duration;
|
||||
};
|
||||
|
||||
// Seek for closest key frame in the past;
|
||||
auto seek_for_prev_key_frame = [&](PacketData& pkt_data, VideoSeekContext& seek_ctx) {
|
||||
seek_frame(seek_ctx.seek_crit_, AVSEEK_FLAG_BACKWARD);
|
||||
Demux(pp_video, video_size);
|
||||
seek_ctx.out_frame_pts_ = pkt_data.pts;
|
||||
seek_ctx.out_frame_duration_ = pkt_data.duration;
|
||||
};
|
||||
|
||||
PacketData pktData;
|
||||
pktData.bsl_data = size_t(*pp_video);
|
||||
pktData.bsl = *video_size;
|
||||
|
||||
switch (seek_ctx.seek_mode_) {
|
||||
case SEEK_MODE_EXACT_FRAME:
|
||||
seek_for_exact_frame(pktData, seek_ctx);
|
||||
break;
|
||||
case SEEK_MODE_PREV_KEY_FRAME:
|
||||
seek_for_prev_key_frame(pktData, seek_ctx);
|
||||
break;
|
||||
default:
|
||||
throw std::runtime_error("ERROR::Unsupported seek mode");
|
||||
break;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
AVFormatContext *VideoDemuxer::CreateFmtContextUtil(StreamProvider *stream_provider) {
|
||||
AVFormatContext *ctx = nullptr;
|
||||
if (!(ctx = avformat_alloc_context())) {
|
||||
|
||||
在新工单中引用
屏蔽一个用户