[FFmpeg-devel] [PATCH v6 3/3] aadec: improve seeking in mp3 content

[Karsten Otto]

MP3 frames may not be aligned to aa chunk boundaries. When seeking,
calculate the expected frame offset in the target chunk. Adjust the
timestamp and truncate the next packet accordingly.

This solution works for the majority of tested audio material. For
some rare encodings with mp3 padding or embedded id3 tags, it will
mispredict the correct offset, and at worst skip an extra frame.
---
This revision includes some comments documenting how/why it works.

 libavformat/aadec.c | 21 ++++++++++++++++++---
 1 file changed, 18 insertions(+), 3 deletions(-)

diff --git a/libavformat/aadec.c b/libavformat/aadec.c
index 17ad20686b..f7e92dbb4b 100644
--- a/libavformat/aadec.c
+++ b/libavformat/aadec.c
@@ -37,6 +37,7 @@
 #define TEA_BLOCK_SIZE 8
 #define CHAPTER_HEADER_SIZE 8
 #define TIMEPREC 1000
+#define MP3_FRAME_SIZE 104
 
 typedef struct AADemuxContext {
     AVClass *class;
@@ -50,6 +51,7 @@ typedef struct AADemuxContext {
     int64_t current_chapter_size;
     int64_t content_start;
     int64_t content_end;
+    int seek_offset;
 } AADemuxContext;
 
 static int get_second_size(char *codec_name)
@@ -177,6 +179,7 @@ static int aa_read_header(AVFormatContext *s)
         st->codecpar->sample_rate = 22050;
         st->need_parsing = AVSTREAM_PARSE_FULL_RAW;
         avpriv_set_pts_info(st, 64, 8, 32000 * TIMEPREC);
+        // encoded audio frame is MP3_FRAME_SIZE bytes (+1 with padding, unlikely)
     } else if (!strcmp(codec_name, "acelp85")) {
         st->codecpar->codec_id = AV_CODEC_ID_SIPR;
         st->codecpar->block_align = 19;
@@ -228,6 +231,7 @@ static int aa_read_header(AVFormatContext *s)
     ff_update_cur_dts(s, st, 0);
     avio_seek(pb, start, SEEK_SET);
     c->current_chapter_size = 0;
+    c->seek_offset = 0;
 
     return 0;
 }
@@ -266,6 +270,10 @@ static int aa_read_packet(AVFormatContext *s, AVPacket *pkt)
     // is this the last block in this chapter?
     if (c->current_chapter_size / c->current_codec_second_size == 0) {
         c->current_codec_second_size = c->current_chapter_size % c->current_codec_second_size;
+        // normally, seek_offset max MP3_FRAME_SIZE-1 <= current_codec_second_size;
+        // in short last block, estimate may be off for file with padding
+        if (c->seek_offset > c->current_codec_second_size)
+            c->seek_offset = 0;
     }
 
     // decrypt c->current_codec_second_size bytes
@@ -293,12 +301,14 @@ static int aa_read_packet(AVFormatContext *s, AVPacket *pkt)
     if (c->current_chapter_size <= 0)
         c->current_chapter_size = 0;
 
-    ret = av_new_packet(pkt, written);
+    // at this point, seek_offset <= written == current_codec_second_size
+    ret = av_new_packet(pkt, written - c->seek_offset);
     if (ret < 0)
         return ret;
-    memcpy(pkt->data, buf, written);
+    memcpy(pkt->data, buf + c->seek_offset, written - c->seek_offset);
     pkt->pos = pos;
 
+    c->seek_offset = 0;
     return 0;
 }
 
@@ -342,7 +352,12 @@ static int aa_read_seek(AVFormatContext *s,
     c->current_chapter_size = chapter_size - chapter_pos;
     c->chapter_idx = 1 + chapter_idx;
 
-    ff_update_cur_dts(s, s->streams[0], ch->start + chapter_pos * TIMEPREC);
+    // estimate extra offset of first unaligned frame in block (assume no padding)
+    if (s->streams[0]->codecpar->codec_id == AV_CODEC_ID_MP3) {
+        c->seek_offset = (MP3_FRAME_SIZE - chapter_pos % MP3_FRAME_SIZE) % MP3_FRAME_SIZE;
+    }
+
+    ff_update_cur_dts(s, s->streams[0], ch->start + (chapter_pos + c->seek_offset) * TIMEPREC);
 
     return 1;
 }
-- 
2.14.3 (Apple Git-98)