THE GOAL -------- What we are trying to achieve: satisfy: patching of CVS checkout using our patch files placed in our CVS passing of make make distcheck non-srcdir build (ie, mkdir build; cd build; ../configure; make) THE SETUP --------- There is a "mirror" root CVS module that contains "ffmpeg". This directory contains a vendor-branch checkout of upstream FFmpeg CVS of a given day. On head, the following things have been commited on top of this: * patches/, which is a directory with a set of patches, and a series file listing the order, as generated by quilt * .pc/, which is a tree of files that quilt uses to keep control of its state. It contains a list of applied patches, and one directory per patch, containing a tree of hardlinked files that were added to the patchset, and a .pc file listing all files part of the patchset. * the result of having all these patches commited (ie, quilt push -a) to the ffmpeg tree. Both the actually patched CVS ffmpeg code as well as the .pc dir need to be commited to CVS so the state of quilt wrt. the source is in sync. THE WARNING ----------- *** NEVER EVER commit stuff in gst-libs/ext/ffmpeg UNLESS your quilt stack is completely applied ! This means, ALWAYS make sure quilt push -a has been run without problems. What's more, if you want to be on the safe side, make sure that you can unapply and reapply without problems, by running quilt pop -a then quilt push -a. The ONLY exception to this is when you're working on a branch to update the upstream source you're working with. *** THE WAY ------- - If you want to hack on our copy of the FFmpeg code, there are some basic rules you need to respect: - you need to use quilt. If you don't use quilt, you can't hack on it. - we separate patches based on the functionality they patch, and whether or not we want to send stuff upstream. Make sure you work in the right patch. use "quilt applied" to check which patches are applied. - before starting to hack, run cvs diff. There should be NO diffs, and NO files listed with question mark. If there are, somebody before you probably made a mistake. To manage the state correctly, it is vital that none of the files are unknown to CVS. - if you want to add a file to a patchset, you need to: - be in the right patchset - quilt add (file) - cvs add .pc/(patchsetname)/(file) - cvs commit .pc/(patchsetname) (to update the state of quilt in cvs) - edit the file - cvs add the file if it doesn't exist yet - quilt refresh - quilt push -a (This one is IMPORTANT, otherwise you'll have a huge diff) - cvs commit - if you want to add a patchset, you need to: - go over the procedure with thomas to check it's correct - decide where in the stack to put it. ask for help if you don't know. - go there in the patch stack (use quilt pop/push) - quilt new (patchsetname).patch (don't forget .patch !) - quilt add (files) - cvs add .pc/(patchsetname) the whole tree - cvs commit .pc/(patchsetname) - quilt refresh - quilt push -a - cvs commit - cvs diff (to check if any of the files are unknown to CVS; if they are, you need to add them to CVS) THE UPSTREAM ------------ At some points you want to update the upstream snapshot code to a newer date. This is easy if you follow the steps outlined here, but make sure to follow them correctly ! - find a good CVS snapshot date for upstream, one that is known to work. You're going to save yourself quite a bit of trouble if you verify this first ! - check it out to a local directory: cvs -z9 -d:pserver:anonymous@mplayerhq.hu:/cvsroot/ffmpeg export -D '2004-04-11 23:00 GMT' ffmpeg - compile it and test it, make sure it works - in gst-ffmpeg/gst-libs/ext/ffmpeg: - Pre-flight checks: - first make sure you don't have local changes, all files are either in CVS or in .cvsignore patch, the whole quilt stack is applied, and stuff works. - do a quilt pop -a and quilt push -a to verify everything is ok. - Branch and rollback: - tag HEAD with the branch root point: cvs tag BRANCH-UPDATE-CVS-2004-04-11-23-00-ROOT - branch: cvs tag -b BRANCH-UPDATE-CVS-2004-04-11-23-00 - FIXME: lock cvs HEAD - update local copy to branch: cvs update -r BRANCH-UPDATE-CVS-2004-04-11-23-00 - peel off all patches: quilt pop -a - commit this cvs commit - check cvs diff you should only have ? for files that are generated somehow (binaries, build files, ...) you get warnings about cvs not finding files to diff that are in .pc or generated by your patches - if you want, you can now compare this state of CVS (which should be last upstream CVS combined with your local unapplied quilt state) - remember to NOT do cvs update from here on, since you popped your quilt state all your added files that are also in CVS are not locally present. - sync with upstream: - in a temp dir, redo the export: cd .. mkdir tmp cd tmp cvs -z9 -d:pserver:anonymous@mplayerhq.hu:/cvsroot/ffmpeg export -D '2004-04-11 23:00 GMT' ffmpeg - rsync it over the old ffmpeg tree rsync -arv ffmpeg .. - go back and commit this new snapshot cd ../ffmpeg cvs commit - check if any new files got added that you should add to cvs cvs diff This will list a lot of local files missing, from your quilt state, which you shouldn't worry about. Just inspect all the ?'s and add files to cvs that belong to upstream and should be in cvs. - if everything's ok and commited, tag the state: cvs tag UPSTREAM-CVS-2004-04-11-23-00 - reapply and fix quilt patches one by one - try applying one quilt push - if that didn't work, inspect the patch and figure out how to fix it: - if the patch got applied upstream completely, quilt push will tell you the patch looks like a reverse patch. In that case you can remove the patch from your patches file (patches/series), and remove the .pc/$(patchname) and patches/$(patchname).patch files from cvs. - if the patch conflicts somehow, you can force application with quilt push -f and then resolve all the rejects, and fix the patch completely. Then refresh quilt state with quilt refresh - when the patch is succesfully removed or reworked, commit current state to CVS, then check again if nothing is missing using cvs diff, and resolve problems/apply missing files from your quilt state/... - after reapplying your complete quilt state, test locally if the complete gst-ffmpeg module now works. Compile and test. Resolve all problems (upstream errors, missing symbols, missing files, ...) until you have a working module. commit again. - merge to head: - update locally back to head cvs update -A - FIXME: unlock cvs HEAD - merge from your branch cvs update -j BRANCH-UPDATE-CVS-2004-04-11-23-00 - commit cvs commit - check for diffs cvs diff - tag merge point cvs tag BRANCH-UPDATE-CVS-2004-04-11-23-00-MERGE - add upstream date to "THE RECORDS" below - get a drink THE PLUGIN ---------- Some notes on how ffmpeg wrapping inside GStreamer currently works: * gstffmpeg{dec,enc,demux,mux}.c are wrappers for specific element types from their ffmpeg counterpart. If you want to wrap a new type of element in wrapper file. * gstffmpegcolorspace.c is a wrapper for one specific function in ffmpeg: colorspace conversion. This works different from the previously mentioned ones, and we'll come to that in the next item. If you want to wrap one specific function, then that, too, belongs in a new wrapper file. * the important difference between all those is that the colorspace element contains one element, so there is a 1<->1 mapping. This makes for a fairly basic element implementation. gstffmpegcolorspace.c, therefore, doesn't differ much from other colorspace elements. The ffmpeg element types, however, define a whole *list* of elements (in GStreamer, each decoder etc. needs to be its own element). We use a set of tricks for that to keep coding simple: codec mapping and dynamic type creation. * ffmpeg uses CODEC_ID_* enumerations for their codecs. GStreamer uses caps, which consists of a mimetype and a defined set of properties. In ffmpeg, these properties live in a AVCodecContext struct, which contains anything that could configure any codec (which makes it rather messy, but ohwell). To convert from one to the other, we use codec mapping, which is done in gstffmpegcodecmap.[ch]. This is the most important file in the whole ffmpeg wrapping process! It contains functions to go from a codec type (video or audio - used as the output format for decoding or the input format for encoding), a codec id (to identify each format) or a format id (a string identifying a file format - usually the file format extension) to a GstCaps, and the other way around. * to define multiple elements in one source file (which all behave similarly), we dynamically create types for each plugin and let all of them operate on the same struct (GstFFMpegDec, GstFFMpegEnc, ...). The functions in gstffmpeg{dec,enc,demux,mux}.c called gst_ffmpeg*_register() do this. The magic is as follows: for each codec or format, ffmpeg has a single AVCodec or AV{Input,Output}Format, which are packed together in a list of supported codecs/formats. We simply walk through the list, for each of those, we check whether gstffmpegcodecmap.c knows about this single one. If it does, we get the GstCaps for each pad template that belongs to it, and register a type for all of those together. We also leave this inside a caching struct, that will later be used by the base_init() function to fill in information about this specific codec in the class struct of this element (pad templates and codec/format information). Since the actual codec information is the only thing that really makes each codec/format different (they all behave the same through the ffmpeg API), we don't really need to do anything else that is codec-specific, so all other functions are rather simple. * one particular thing that needs mention is how gstffmpeg{mux,demux}.c and gstffmpegprotocol.c interoperate. ffmpeg uses URLProtocols for data input and output. Now, of course, we want to use the *GStreamer* way of doing input and output (filesrc, ...) rather than the ffmpeg way. Therefore, we wrap up a GstPad as a URLProtocol and register this with ffmpeg. This is what gstffmpegprotocol.c does. The URL is called gstreamer://%p, where %p is the address of a GstPad. gstffmpeg{mux,demux}.c then open a file called gstreamer://%p, with %p being their source/sink pad, respectively. This way, we use GStreamer for data input/output through the ffmpeg API. It's rather ugly, but it has worked quite well so far. * there's lots of things that still need doing. See the TODO file for more information. THE RECORDS ----------- - list of snapshots used: CVS-2004-04-11-23-00 THE REMINDERS ------------- * the initial ffmpeg checkout was imported using: - get CVS ffmpeg cvs -z3 -d:pserver:anonymous@mplayerhq.hu:/cvsroot/ffmpeg co -D '2004-03-09 06:00 GMT' ffmpeg