[FFmpeg-devel] [PATCH 1/2] lavf/matroska: Add functions for WebM DASH Manifest

Vignesh Venkatasubramanian vigneshv at google.com
Mon Jul 7 21:52:37 CEST 2014


Add functions and logic to matroskadec for use by the WebM DASH Manifest
XML Muxer. The actual muxer is added in a future patch.

Signed-off-by: Vignesh Venkatasubramanian <vigneshv at google.com>
---
 libavformat/matroska.h    |  12 ++
 libavformat/matroskadec.c | 368 ++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 380 insertions(+)

diff --git a/libavformat/matroska.h b/libavformat/matroska.h
index 3bb5aee..e01b9de 100644
--- a/libavformat/matroska.h
+++ b/libavformat/matroska.h
@@ -284,4 +284,16 @@ extern const AVMetadataConv ff_mkv_metadata_conv[];
 extern const char * const ff_matroska_video_stereo_mode[MATROSKA_VIDEO_STEREO_MODE_COUNT];
 extern const char * const ff_matroska_video_stereo_plane[MATROSKA_VIDEO_STEREO_PLANE_COUNT];
 
+/* AVStream Metadata tag keys for WebM Dash Manifest */
+#define INITIALIZATION_RANGE "webm_dash_manifest_initialization_range"
+#define CUES_START "webm_dash_manifest_cues_start"
+#define CUES_END "webm_dash_manifest_cues_end"
+#define FILENAME "webm_dash_manifest_file_name"
+#define BANDWIDTH "webm_dash_manifest_bandwidth"
+#define DURATION "webm_dash_manifest_duration"
+#define CLUSTER_KEYFRAME "webm_dash_manifest_cluster_keyframe"
+#define CUE_TIMESTAMPS "webm_dash_manifest_cue_timestamps"
+#define TRACK_NUMBER "webm_dash_manifest_track_number"
+#define CODEC_PRIVATE_SIZE "webm_dash_manifest_codec_priv_size"
+
 #endif /* AVFORMAT_MATROSKA_H */
diff --git a/libavformat/matroskadec.c b/libavformat/matroskadec.c
index 213767c..6fa8229 100644
--- a/libavformat/matroskadec.c
+++ b/libavformat/matroskadec.c
@@ -3068,6 +3068,365 @@ static int matroska_read_close(AVFormatContext *s)
     return 0;
 }
 
+typedef struct {
+    int64_t start_time_ns;
+    int64_t end_time_ns;
+    int64_t start_offset;
+    int64_t end_offset;
+} CueDesc;
+
+/* This function searches all the Cues and returns the CueDesc corresponding the
+ * the timestamp ts. Returned CueDesc will be such that start_time_ns <= ts <
+ * end_time_ns. All 4 fields will be set to -1 if ts >= file's duration.
+ */
+static CueDesc get_cue_desc(AVFormatContext *s, int64_t ts, int64_t cues_start) {
+    MatroskaDemuxContext *matroska = s->priv_data;
+    CueDesc cue_desc;
+    int i;
+    int nb_index_entries = s->streams[0]->nb_index_entries;
+    AVIndexEntry *index_entries = s->streams[0]->index_entries;
+    if (ts >= matroska->duration * matroska->time_scale) return (CueDesc) {-1, -1, -1, -1};
+    for (i = 1; i < nb_index_entries; i++) {
+        if (index_entries[i - 1].timestamp * matroska->time_scale <= ts &&
+            index_entries[i].timestamp * matroska->time_scale > ts) {
+            break;
+        }
+    }
+    --i;
+    cue_desc.start_time_ns = index_entries[i].timestamp * matroska->time_scale;
+    cue_desc.start_offset = index_entries[i].pos - matroska->segment_start;
+    if (i != nb_index_entries - 1) {
+        cue_desc.end_time_ns = index_entries[i + 1].timestamp * matroska->time_scale;
+        cue_desc.end_offset = index_entries[i + 1].pos - matroska->segment_start;
+    } else {
+        cue_desc.end_time_ns = matroska->duration * matroska->time_scale;
+        // FIXME: this needs special handling for files where Cues appear
+        // before Clusters. the current logic assumes Cues appear after
+        // Clusters.
+        cue_desc.end_offset = cues_start - matroska->segment_start;
+    }
+    return cue_desc;
+}
+
+static int webm_clusters_start_with_keyframe(AVFormatContext *s)
+{
+    MatroskaDemuxContext *matroska = s->priv_data;
+    int64_t cluster_pos, before_pos;
+    int index, rv = 1;
+    if (s->streams[0]->nb_index_entries <= 0) return 0;
+    // seek to the first cluster using cues.
+    index = av_index_search_timestamp(s->streams[0], 0, 0);
+    if (index < 0)  return 0;
+    cluster_pos = s->streams[0]->index_entries[index].pos;
+    before_pos = avio_tell(s->pb);
+    while (1) {
+        int64_t cluster_id = 0, cluster_length = 0;
+        AVPacket *pkt;
+        avio_seek(s->pb, cluster_pos, SEEK_SET);
+        // read cluster id and length
+        ebml_read_num(matroska, matroska->ctx->pb, 4, &cluster_id);
+        ebml_read_length(matroska, matroska->ctx->pb, &cluster_length);
+        if (cluster_id != 0xF43B675) { // done with all clusters
+            break;
+        }
+        avio_seek(s->pb, cluster_pos, SEEK_SET);
+        matroska->current_id = 0;
+        matroska_clear_queue(matroska);
+        if (matroska_parse_cluster(matroska) < 0 ||
+            matroska->num_packets <= 0) {
+            break;
+        }
+        pkt = matroska->packets[0];
+        cluster_pos += cluster_length + 12; // 12 is the offset of the cluster id and length.
+        if (!(pkt->flags & AV_PKT_FLAG_KEY)) {
+            rv = 0;
+            break;
+        }
+    }
+    avio_seek(s->pb, before_pos, SEEK_SET);
+    return rv;
+}
+
+static int buffer_size_after_time_downloaded(int64_t time_ns, double search_sec, int64_t bps,
+                                             double min_buffer, double* buffer,
+                                             double* sec_to_download, AVFormatContext *s,
+                                             int64_t cues_start)
+{
+    double nano_seconds_per_second = 1000000000.0;
+    double time_sec = time_ns / nano_seconds_per_second;
+    int rv = 0;
+    int64_t time_to_search_ns = (int64_t)(search_sec * nano_seconds_per_second);
+    int64_t end_time_ns = time_ns + time_to_search_ns;
+    double sec_downloaded = 0.0;
+    CueDesc desc_curr = get_cue_desc(s, time_ns, cues_start);
+    if (desc_curr.start_time_ns == -1)
+      return -1;
+    *sec_to_download = 0.0;
+
+    // Check for non cue start time.
+    if (time_ns > desc_curr.start_time_ns) {
+      int64_t cue_nano = desc_curr.end_time_ns - time_ns;
+      double percent = (double)(cue_nano) / (desc_curr.end_time_ns - desc_curr.start_time_ns);
+      double cueBytes = (desc_curr.end_offset - desc_curr.start_offset) * percent;
+      double timeToDownload = (cueBytes * 8.0) / bps;
+
+      sec_downloaded += (cue_nano / nano_seconds_per_second) - timeToDownload;
+      *sec_to_download += timeToDownload;
+
+      // Check if the search ends within the first cue.
+      if (desc_curr.end_time_ns >= end_time_ns) {
+          double desc_end_time_sec = desc_curr.end_time_ns / nano_seconds_per_second;
+          double percent_to_sub = search_sec / (desc_end_time_sec - time_sec);
+          sec_downloaded = percent_to_sub * sec_downloaded;
+          *sec_to_download = percent_to_sub * *sec_to_download;
+      }
+
+      if ((sec_downloaded + *buffer) <= min_buffer) {
+          return 1;
+      }
+
+      // Get the next Cue.
+      desc_curr = get_cue_desc(s, desc_curr.end_time_ns, cues_start);
+    }
+
+    while (desc_curr.start_time_ns != -1) {
+        int64_t desc_bytes = desc_curr.end_offset - desc_curr.start_offset;
+        int64_t desc_ns = desc_curr.end_time_ns - desc_curr.start_time_ns;
+        double desc_sec = desc_ns / nano_seconds_per_second;
+        double bits = (desc_bytes * 8.0);
+        double time_to_download = bits / bps;
+
+        sec_downloaded += desc_sec - time_to_download;
+        *sec_to_download += time_to_download;
+
+        if (desc_curr.end_time_ns >= end_time_ns) {
+            double desc_end_time_sec = desc_curr.end_time_ns / nano_seconds_per_second;
+            double percent_to_sub = search_sec / (desc_end_time_sec - time_sec);
+            sec_downloaded = percent_to_sub * sec_downloaded;
+            *sec_to_download = percent_to_sub * *sec_to_download;
+
+            if ((sec_downloaded + *buffer) <= min_buffer)
+                rv = 1;
+            break;
+        }
+
+        if ((sec_downloaded + *buffer) <= min_buffer) {
+            rv = 1;
+            break;
+        }
+
+        desc_curr = get_cue_desc(s, desc_curr.end_time_ns, cues_start);
+    }
+    *buffer = *buffer + sec_downloaded;
+    return rv;
+}
+
+/* This function computes the bandwidth of the WebM file with the help of
+ * buffer_size_after_time_downloaded() function. Both of these functions are
+ * adapted from WebM Tools project and are adapted to work with FFmpeg's
+ * Matroska parsing mechanism.
+ *
+ * Returns the bandwidth of the file on success; -1 on error.
+ * */
+static int64_t webm_dash_manifest_compute_bandwidth(AVFormatContext *s, int64_t cues_start)
+{
+    MatroskaDemuxContext *matroska = s->priv_data;
+    AVStream *st = s->streams[0];
+    double bandwidth = 0.0;
+    for (int i = 0; i < st->nb_index_entries; i++) {
+        int64_t prebuffer_ns = 1000000000;
+        int64_t time_ns = st->index_entries[i].timestamp * matroska->time_scale;
+        double nano_seconds_per_second = 1000000000.0;
+        int64_t prebuffered_ns = time_ns + prebuffer_ns;
+        double prebuffer_bytes = 0.0;
+        int64_t temp_prebuffer_ns = prebuffer_ns;
+        int64_t pre_bytes, pre_ns;
+        double pre_sec, prebuffer, bits_per_second;
+        CueDesc desc_beg = get_cue_desc(s, time_ns, cues_start);
+
+        // Start with the first Cue.
+        CueDesc desc_end = desc_beg;
+
+        // Figure out how much data we have downloaded for the prebuffer. This will
+        // be used later to adjust the bits per sample to try.
+        while (desc_end.start_time_ns != -1 && desc_end.end_time_ns < prebuffered_ns) {
+            // Prebuffered the entire Cue.
+            prebuffer_bytes += desc_end.end_offset - desc_end.start_offset;
+            temp_prebuffer_ns -= desc_end.end_time_ns - desc_end.start_time_ns;
+            desc_end = get_cue_desc(s, desc_end.end_time_ns, cues_start);
+        }
+        if (desc_end.start_time_ns == -1) {
+            // The prebuffer is larger than the duration.
+            return (matroska->duration * matroska->time_scale >= prebuffered_ns) ? -1 : 0;
+        }
+
+        // The prebuffer ends in the last Cue. Estimate how much data was
+        // prebuffered.
+        pre_bytes = desc_end.end_offset - desc_end.start_offset;
+        pre_ns = desc_end.end_time_ns - desc_end.start_time_ns;
+        pre_sec = pre_ns / nano_seconds_per_second;
+        prebuffer_bytes +=
+            pre_bytes * ((temp_prebuffer_ns / nano_seconds_per_second) / pre_sec);
+
+        prebuffer = prebuffer_ns / nano_seconds_per_second;
+
+        // Set this to 0.0 in case our prebuffer buffers the entire video.
+        bits_per_second = 0.0;
+        do {
+            int64_t desc_bytes = desc_end.end_offset - desc_beg.start_offset;
+            int64_t desc_ns = desc_end.end_time_ns - desc_beg.start_time_ns;
+            double desc_sec = desc_ns / nano_seconds_per_second;
+            double calc_bits_per_second = (desc_bytes * 8) / desc_sec;
+
+            // Drop the bps by the percentage of bytes buffered.
+            double percent = (desc_bytes - prebuffer_bytes) / desc_bytes;
+            double mod_bits_per_second = calc_bits_per_second * percent;
+
+            if (prebuffer < desc_sec) {
+                double search_sec =
+                    (double)(matroska->duration * matroska->time_scale) / nano_seconds_per_second;
+
+                // Add 1 so the bits per second should be a little bit greater than file
+                // datarate.
+                int64_t bps = (int64_t)(mod_bits_per_second) + 1;
+                const double min_buffer = 0.0;
+                double buffer = prebuffer;
+                double sec_to_download = 0.0;
+
+                int rv = buffer_size_after_time_downloaded(prebuffered_ns, search_sec, bps,
+                                                           min_buffer, &buffer, &sec_to_download,
+                                                           s, cues_start);
+                if (rv < 0) {
+                    return -1;
+                } else if (rv == 0) {
+                    bits_per_second = (double)(bps);
+                    break;
+                }
+            }
+
+            desc_end = get_cue_desc(s, desc_end.end_time_ns, cues_start);
+        } while (desc_end.start_time_ns != -1);
+        if (bandwidth < bits_per_second) bandwidth = bits_per_second;
+    }
+    return (int64_t)bandwidth;
+}
+
+static int webm_dash_manifest_cues(AVFormatContext *s)
+{
+    MatroskaDemuxContext *matroska = s->priv_data;
+    EbmlList *seekhead_list = &matroska->seekhead;
+    MatroskaSeekhead *seekhead = seekhead_list->elem;
+    char *buf;
+    int64_t cues_start, cues_end, before_pos, bandwidth;
+    int i;
+
+    // determine cues start and end positions
+    for (i = 0; i < seekhead_list->nb_elem; i++)
+        if (seekhead[i].id == MATROSKA_ID_CUES)
+            break;
+
+    if (i >= seekhead_list->nb_elem) return -1;
+
+    before_pos = avio_tell(matroska->ctx->pb);
+    cues_start = seekhead[i].pos + matroska->segment_start;
+    if (avio_seek(matroska->ctx->pb, cues_start, SEEK_SET) == cues_start) {
+        uint64_t cues_length = 0, cues_id = 0;
+        ebml_read_num(matroska, matroska->ctx->pb, 4, &cues_id);
+        ebml_read_length(matroska, matroska->ctx->pb, &cues_length);
+        cues_end = cues_start + cues_length + 11; // 11 is the offset of Cues ID.
+    }
+    avio_seek(matroska->ctx->pb, before_pos, SEEK_SET);
+
+    // parse the cues
+    matroska_parse_cues(matroska);
+
+    // cues start
+    buf = av_asprintf("%" PRId64, cues_start);
+    if (!buf) return AVERROR(ENOMEM);
+    av_dict_set(&s->streams[0]->metadata, CUES_START, buf, 0);
+    av_free(buf);
+
+    // cues end
+    buf = av_asprintf("%" PRId64, cues_end);
+    if (!buf) return AVERROR(ENOMEM);
+    av_dict_set(&s->streams[0]->metadata, CUES_END, buf, 0);
+    av_free(buf);
+
+    // bandwidth
+    bandwidth = webm_dash_manifest_compute_bandwidth(s, cues_start);
+    if (bandwidth < 0) return -1;
+    buf = av_asprintf("%" PRId64, bandwidth);
+    if (!buf) return AVERROR(ENOMEM);
+    av_dict_set(&s->streams[0]->metadata, BANDWIDTH, buf, 0);
+    av_free(buf);
+
+    // check if all clusters start with key frames
+    buf = av_asprintf("%d", webm_clusters_start_with_keyframe(s));
+    if (!buf) return AVERROR(ENOMEM);
+    av_dict_set(&s->streams[0]->metadata, CLUSTER_KEYFRAME, buf, 0);
+    av_free(buf);
+
+    // store cue point timestamps as a comma separated list for checking subsegment alignment in
+    // the muxer. assumes that each timestamp cannot be more than 20 characters long.
+    buf = av_malloc(s->streams[0]->nb_index_entries * 20 * sizeof(char));
+    if (!buf) return -1;
+    strcpy(buf, "");
+    for (i = 0; i < s->streams[0]->nb_index_entries; i++) {
+        snprintf(buf, (i + 1) * 20 * sizeof(char),
+                 "%s%" PRId64, buf, s->streams[0]->index_entries[i].timestamp);
+        if (i != s->streams[0]->nb_index_entries - 1)
+            strncat(buf, ",", sizeof(char));
+    }
+    av_dict_set(&s->streams[0]->metadata, CUE_TIMESTAMPS, buf, 0);
+    av_free(buf);
+
+    return 0;
+}
+
+static int webm_dash_manifest_read_header(AVFormatContext *s)
+{
+    char *buf;
+    int ret = matroska_read_header(s);
+    MatroskaTrack *tracks;
+    MatroskaDemuxContext *matroska = s->priv_data;
+    if (ret) {
+        av_log(s, AV_LOG_ERROR, "Failed to read file headers\n");
+        return -1;
+    }
+
+    // initialization range
+    buf = av_asprintf("%" PRId64, avio_tell(s->pb) - 5); // 5 is the offset of Cluster ID.
+    if (!buf) return AVERROR(ENOMEM);
+    av_dict_set(&s->streams[0]->metadata, INITIALIZATION_RANGE, buf, 0);
+    av_free(buf);
+
+    // basename of the file
+    buf = strrchr(s->filename, '/');
+    if (buf == NULL) return -1;
+    av_dict_set(&s->streams[0]->metadata, FILENAME, ++buf, 0);
+
+    // duration
+    buf = av_asprintf("%g", matroska->duration);
+    if (!buf) return AVERROR(ENOMEM);
+    av_dict_set(&s->streams[0]->metadata, DURATION, buf, 0);
+    av_free(buf);
+
+    // track number
+    tracks = matroska->tracks.elem;
+    buf = av_asprintf("%" PRId64, tracks[0].num);
+    if (!buf) return AVERROR(ENOMEM);
+    av_dict_set(&s->streams[0]->metadata, TRACK_NUMBER, buf, 0);
+    av_free(buf);
+
+    // parse the cues and populate Cue related fields
+    return webm_dash_manifest_cues(s);
+}
+
+static int webm_dash_manifest_read_packet(AVFormatContext *s, AVPacket *pkt)
+{
+    return AVERROR_EOF;
+}
+
 AVInputFormat ff_matroska_demuxer = {
     .name           = "matroska,webm",
     .long_name      = NULL_IF_CONFIG_SMALL("Matroska / WebM"),
@@ -3078,3 +3437,12 @@ AVInputFormat ff_matroska_demuxer = {
     .read_close     = matroska_read_close,
     .read_seek      = matroska_read_seek,
 };
+
+AVInputFormat ff_webm_dash_manifest_demuxer = {
+    .name           = "webm_dash_manifest",
+    .long_name      = NULL_IF_CONFIG_SMALL("WebM DASH Manifest"),
+    .priv_data_size = sizeof(MatroskaDemuxContext),
+    .read_header    = webm_dash_manifest_read_header,
+    .read_packet    = webm_dash_manifest_read_packet,
+    .read_close     = matroska_read_close,
+};
-- 
2.0.0.526.g5318336



More information about the ffmpeg-devel mailing list