[FFmpeg-devel] [PATCH] WebM mux/demux
James Zern
jzern
Sat May 29 01:17:26 CEST 2010
On Thu, May 20, 2010 at 05:24, David Conrad <lessen42 at gmail.com> wrote:
> On May 19, 2010, at 3:26 PM, James Zern wrote:
>
>> On Wed, May 19, 2010 at 15:14, David Conrad <lessen42 at gmail.com> wrote:
>>> On May 19, 2010, at 2:51 PM, James Zern wrote:
>>>
>>>> Related thread: '[PATCH] VP8 de/encode via libvpx'
>>>> VP8 and associated WebM container made public today [1].
>>>>
>>>> The attached adds WebM de/muxing to matroska.
>>>> For both some extra ifdef's/checks were added to only allow VP8/vorbis
>>>> as that is the current supported configuration in web browsers. If
>>>> these are too busy in this respect they be cleaned up.
>>>
>>> Yes, the only #ifdef should be around the AVOutputFormat.
>>>
>>>> + ? ?if (!strcmp("webm", s->oformat->name)) {
>>>> + ? ? ? ?mkv->max_cluster_size = 500*1024*1024; ? // 500 MB
>>>> + ? ? ? ?mkv->max_cluster_pts = 30000; ? ? ? ? ? ?// 30 seconds
>>>> + ? ?} else {
>>>
>>> Why? Seeking can only seek to a cluster, then do a linear scan to find the needed packet, and 500 MB makes that incredibly inefficient.
>> This was added as a worst case, the recommendation is for keyframes on
>> the 2-5s range.
>
>>>> +static int mkv_copy_packet(MatroskaMuxContext *mkv, const AVPacket *pkt)
>>>> +{
>>>> + ? ?uint8_t *data = mkv->cur_audio_pkt.data;
>>>> + ? ?mkv->cur_audio_pkt = *pkt;
>>>> + ? ?mkv->cur_audio_pkt.data = av_fast_realloc(data, &mkv->audio_buffer_size, pkt->size);
>>>> + ? ?if (mkv->cur_audio_pkt.data == NULL)
>>>> + ? ? ? ?return AVERROR(ENOMEM);
>>>> +
>>>> + ? ?memcpy(mkv->cur_audio_pkt.data, pkt->data, pkt->size);
>>>> + ? ?mkv->cur_audio_pkt.size = pkt->size;
>>>> + ? ?return 0;
>>>> +}
>>>
>>>> +
>>>> +static int mkv_write_packet2(AVFormatContext *s, AVPacket *pkt)
>>>> +{
>>>
>>> webm and mkv shouldn't use separate write_packets, the interleaving changes should be done for mkv too. Maybe call the current mkv_write_packet mkv_write_packet_internal and move all cluster handling to the outer one.
>>>
>> OK. The thought initially was not to force the keyframe change on the
>> matroska muxer, but I think this can be reworked.
>
> If the conceptual goal is 1 cluster == 1 gop, the main reason I didn't do it for mkv is because I was worried about overhead. But after making the the streaming changes I think the overhead of this shouldn't be terribly important. At any rate, the only difference I want in muxing webm is the doctype and checks to not write unsupported elements (namely non-vp8+vorbis.)
>
The attached adds the webm doctype, audio packet buffering and
keyframe clustering decision.
-------------- next part --------------
Index: Changelog
===================================================================
--- Changelog (revision 23368)
+++ Changelog (working copy)
@@ -4,7 +4,7 @@ releases are sorted from youngest to old
version <next>:
-- WebM support in Matroska demuxer
+- WebM support in Matroska de/muxer
- low overhead Ogg muxing
- MMS-TCP support
- VP8 decoding via libvpx
Index: libavcodec/Makefile
===================================================================
--- libavcodec/Makefile (revision 23368)
+++ libavcodec/Makefile (working copy)
@@ -512,6 +512,8 @@ OBJS-$(CONFIG_OGG_DEMUXER) +
dirac.o mpeg12data.o
OBJS-$(CONFIG_OGG_MUXER) += xiph.o flacdec.o flacdata.o flac.o
OBJS-$(CONFIG_RTP_MUXER) += mpegvideo.o
+OBJS-$(CONFIG_WEBM_MUXER) += xiph.o mpeg4audio.o \
+ flacdec.o flacdata.o flac.o
# external codec libraries
OBJS-$(CONFIG_LIBDIRAC_DECODER) += libdiracdec.o
Index: doc/general.texi
===================================================================
--- doc/general.texi (revision 23368)
+++ doc/general.texi (working copy)
@@ -236,7 +236,7 @@ library:
@item VC-1 test bitstream @tab X @tab X
@item WAV @tab X @tab X
@item WavPack @tab @tab X
- at item WebM @tab @tab X
+ at item WebM @tab X @tab X
@item Wing Commander III movie @tab @tab X
@tab Multimedia format used in Origin's Wing Commander III computer game.
@item Westwood Studios audio @tab @tab X
Index: libavformat/matroskaenc.c
===================================================================
--- libavformat/matroskaenc.c (revision 23368)
+++ libavformat/matroskaenc.c (working copy)
@@ -81,6 +81,8 @@ typedef struct MatroskaMuxContext {
mkv_track *tracks;
struct AVMD5 *md5_ctx;
+ unsigned int audio_buffer_size;
+ AVPacket cur_audio_pkt;
} MatroskaMuxContext;
@@ -565,6 +567,11 @@ static int mkv_write_tracks(AVFormatCont
switch (codec->codec_type) {
case AVMEDIA_TYPE_VIDEO:
+ if (codec->codec_id != CODEC_ID_VP8 && !strcmp("webm", s->oformat->name)) {
+ av_log(s, AV_LOG_ERROR, "Only VP8 video and Vorbis audio are supported for WebM.\n");
+ return AVERROR(EINVAL);
+ }
+
put_ebml_uint(pb, MATROSKA_ID_TRACKTYPE, MATROSKA_TRACK_TYPE_VIDEO);
put_ebml_uint(pb, MATROSKA_ID_TRACKDEFAULTDURATION, av_q2d(codec->time_base)*1E9);
@@ -597,6 +604,11 @@ static int mkv_write_tracks(AVFormatCont
break;
case AVMEDIA_TYPE_AUDIO:
+ if (codec->codec_id != CODEC_ID_VORBIS && !strcmp("webm", s->oformat->name)) {
+ av_log(s, AV_LOG_ERROR, "Only VP8 video and Vorbis audio are supported for WebM.\n");
+ return AVERROR(EINVAL);
+ }
+
put_ebml_uint(pb, MATROSKA_ID_TRACKTYPE, MATROSKA_TRACK_TYPE_AUDIO);
if (!native_id)
@@ -693,7 +705,7 @@ static int mkv_write_header(AVFormatCont
put_ebml_uint (pb, EBML_ID_EBMLREADVERSION , 1);
put_ebml_uint (pb, EBML_ID_EBMLMAXIDLENGTH , 4);
put_ebml_uint (pb, EBML_ID_EBMLMAXSIZELENGTH , 8);
- put_ebml_string (pb, EBML_ID_DOCTYPE , "matroska");
+ put_ebml_string (pb, EBML_ID_DOCTYPE , s->oformat->name);
put_ebml_uint (pb, EBML_ID_DOCTYPEVERSION , 2);
put_ebml_uint (pb, EBML_ID_DOCTYPEREADVERSION , 2);
end_ebml_master(pb, ebml_header);
@@ -746,6 +758,10 @@ static int mkv_write_header(AVFormatCont
if (mkv->cues == NULL)
return AVERROR(ENOMEM);
+ av_init_packet(&mkv->cur_audio_pkt);
+ mkv->cur_audio_pkt.size = 0;
+ mkv->audio_buffer_size = 0;
+
put_flush_packet(pb);
return 0;
}
@@ -861,7 +877,7 @@ static void mkv_flush_dynbuf(AVFormatCon
mkv->dyn_bc = NULL;
}
-static int mkv_write_packet(AVFormatContext *s, AVPacket *pkt)
+static int mkv_write_packet_internal(AVFormatContext *s, AVPacket *pkt)
{
MatroskaMuxContext *mkv = s->priv_data;
ByteIOContext *pb = s->pb;
@@ -910,9 +926,39 @@ static int mkv_write_packet(AVFormatCont
if (ret < 0) return ret;
}
- // start a new cluster every 5 MB or 5 sec, or 32k / 1 sec for streaming
- if ((url_is_streamed(s->pb) && (url_ftell(pb) > 32*1024 || ts > mkv->cluster_pts + 1000))
- || url_ftell(pb) > mkv->cluster_pos + 5*1024*1024 || ts > mkv->cluster_pts + 5000) {
+ mkv->duration = FFMAX(mkv->duration, ts + duration);
+ return 0;
+}
+
+static int mkv_copy_packet(MatroskaMuxContext *mkv, const AVPacket *pkt)
+{
+ uint8_t *data = mkv->cur_audio_pkt.data;
+ mkv->cur_audio_pkt = *pkt;
+ mkv->cur_audio_pkt.data = av_fast_realloc(data, &mkv->audio_buffer_size, pkt->size);
+ if (mkv->cur_audio_pkt.data == NULL)
+ return AVERROR(ENOMEM);
+
+ memcpy(mkv->cur_audio_pkt.data, pkt->data, pkt->size);
+ mkv->cur_audio_pkt.size = pkt->size;
+ return 0;
+}
+
+static int mkv_write_packet(AVFormatContext *s, AVPacket *pkt)
+{
+ MatroskaMuxContext *mkv = s->priv_data;
+ ByteIOContext *pb = url_is_streamed(s->pb) ? mkv->dyn_bc : s->pb;
+ AVCodecContext *codec = s->streams[pkt->stream_index]->codec;
+ int keyframe = !!(pkt->flags & AV_PKT_FLAG_KEY);
+ int ret = 0;
+ int64_t ts = mkv->tracks[pkt->stream_index].write_dts ? pkt->dts : pkt->pts;
+
+ // start a new cluster every 5 MB or 5 sec, or 32k / 1 sec for streaming or
+ // after 4k and on a keyframe for WebM
+ if (mkv->cluster_pos &&
+ ((url_is_streamed(s->pb) && (url_ftell(pb) > 32*1024 || ts > mkv->cluster_pts + 1000))
+ || url_ftell(pb) > mkv->cluster_pos + 5*1024*1024 || ts > mkv->cluster_pts + 5000
+ || (codec->codec_type == AVMEDIA_TYPE_VIDEO && keyframe &&
+ url_ftell(s->pb) > mkv->cluster_pos + 4*1024))) {
av_log(s, AV_LOG_DEBUG, "Starting new cluster at offset %" PRIu64
" bytes, pts %" PRIu64 "\n", url_ftell(pb), ts);
end_ebml_master(pb, mkv->cluster);
@@ -921,8 +967,24 @@ static int mkv_write_packet(AVFormatCont
mkv_flush_dynbuf(s);
}
- mkv->duration = FFMAX(mkv->duration, ts + duration);
- return 0;
+ // check if we have an audio packet cached
+ if ((codec->codec_type == AVMEDIA_TYPE_AUDIO || codec->codec_type == AVMEDIA_TYPE_VIDEO)
+ && mkv->cur_audio_pkt.size > 0) {
+ ret = mkv_write_packet_internal(s, &mkv->cur_audio_pkt);
+ mkv->cur_audio_pkt.size = 0;
+ if (ret < 0) {
+ av_log(s, AV_LOG_ERROR, "Could not write cached audio packet ret:%d\n", ret);
+ return ret;
+ }
+ }
+
+ // buffer an audio packet to ensure the packet containing the video
+ // keyframe's timecode is contained in the same cluster for WebM
+ if (codec->codec_type == AVMEDIA_TYPE_AUDIO)
+ ret = mkv_copy_packet(mkv, pkt);
+ else
+ ret = mkv_write_packet_internal(s, pkt);
+ return ret;
}
static int mkv_write_trailer(AVFormatContext *s)
@@ -932,6 +994,16 @@ static int mkv_write_trailer(AVFormatCon
int64_t currentpos, second_seekhead, cuespos;
int ret;
+ // check if we have an audio packet cached
+ if (mkv->cur_audio_pkt.size > 0) {
+ ret = mkv_write_packet_internal(s, &mkv->cur_audio_pkt);
+ mkv->cur_audio_pkt.size = 0;
+ if (ret < 0) {
+ av_log(s, AV_LOG_ERROR, "Could not write cached audio packet ret:%d\n", ret);
+ return ret;
+ }
+ }
+
if (mkv->dyn_bc) {
end_ebml_master(mkv->dyn_bc, mkv->cluster);
mkv_flush_dynbuf(s);
@@ -970,10 +1042,12 @@ static int mkv_write_trailer(AVFormatCon
end_ebml_master(pb, mkv->segment);
av_free(mkv->md5_ctx);
av_free(mkv->tracks);
+ av_destruct_packet(&mkv->cur_audio_pkt);
put_flush_packet(pb);
return 0;
}
+#if CONFIG_MATROSKA_MUXER
AVOutputFormat matroska_muxer = {
"matroska",
NULL_IF_CONFIG_SMALL("Matroska file format"),
@@ -989,7 +1063,23 @@ AVOutputFormat matroska_muxer = {
.codec_tag = (const AVCodecTag* const []){ff_codec_bmp_tags, ff_codec_wav_tags, 0},
.subtitle_codec = CODEC_ID_TEXT,
};
-
+#endif
+#if CONFIG_WEBM_MUXER
+AVOutputFormat webm_muxer = {
+ "webm",
+ NULL_IF_CONFIG_SMALL("WebM file format"),
+ "video/webm",
+ "webm",
+ sizeof(MatroskaMuxContext),
+ CODEC_ID_VORBIS,
+ CODEC_ID_VP8,
+ mkv_write_header,
+ mkv_write_packet,
+ mkv_write_trailer,
+ .flags = AVFMT_GLOBALHEADER | AVFMT_VARIABLE_FPS,
+};
+#endif
+#if CONFIG_MATROSKA_AUDIO_MUXER
AVOutputFormat matroska_audio_muxer = {
"matroska",
NULL_IF_CONFIG_SMALL("Matroska file format"),
@@ -1004,3 +1094,4 @@ AVOutputFormat matroska_audio_muxer = {
.flags = AVFMT_GLOBALHEADER,
.codec_tag = (const AVCodecTag* const []){ff_codec_wav_tags, 0},
};
+#endif
Index: libavformat/avformat.h
===================================================================
--- libavformat/avformat.h (revision 23368)
+++ libavformat/avformat.h (working copy)
@@ -22,7 +22,7 @@
#define AVFORMAT_AVFORMAT_H
#define LIBAVFORMAT_VERSION_MAJOR 52
-#define LIBAVFORMAT_VERSION_MINOR 67
+#define LIBAVFORMAT_VERSION_MINOR 68
#define LIBAVFORMAT_VERSION_MICRO 0
#define LIBAVFORMAT_VERSION_INT AV_VERSION_INT(LIBAVFORMAT_VERSION_MAJOR, \
Index: libavformat/Makefile
===================================================================
--- libavformat/Makefile (revision 23368)
+++ libavformat/Makefile (working copy)
@@ -258,6 +258,9 @@ OBJS-$(CONFIG_W64_DEMUXER)
OBJS-$(CONFIG_WAV_DEMUXER) += wav.o riff.o raw.o
OBJS-$(CONFIG_WAV_MUXER) += wav.o riff.o
OBJS-$(CONFIG_WC3_DEMUXER) += wc3movie.o
+OBJS-$(CONFIG_WEBM_MUXER) += matroskaenc.o matroska.o \
+ riff.o isom.o avc.o \
+ flacenc_header.o
OBJS-$(CONFIG_WSAUD_DEMUXER) += westwood.o
OBJS-$(CONFIG_WSVQA_DEMUXER) += westwood.o
OBJS-$(CONFIG_WV_DEMUXER) += wv.o apetag.o id3v1.o
Index: libavformat/allformats.c
===================================================================
--- libavformat/allformats.c (revision 23368)
+++ libavformat/allformats.c (working copy)
@@ -206,6 +206,7 @@ void av_register_all(void)
REGISTER_DEMUXER (W64, w64);
REGISTER_MUXDEMUX (WAV, wav);
REGISTER_DEMUXER (WC3, wc3);
+ REGISTER_MUXER (WEBM, webm);
REGISTER_DEMUXER (WSAUD, wsaud);
REGISTER_DEMUXER (WSVQA, wsvqa);
REGISTER_DEMUXER (WV, wv);
More information about the ffmpeg-devel
mailing list