ffmpeg-4.2.2/doc/writing_filters.txt

This document is a tutorial/initiation for writing simple filters in
libavfilter.
Foreword: just like everything else in FFmpeg, libavfilter is monolithic, which
means that it is highly recommended that you submit your filters to the FFmpeg
development mailing-list and make sure that they are applied. Otherwise, your filters
are likely to have a very short lifetime due to more or less regular internal API
changes, and a limited distribution, review, and testing.
Bootstrap
=========
Let's say you want to write a new simple video filter called "foobar" which
takes one frame in input, changes the pixels in whatever fashion you fancy, and
outputs the modified frame. The most simple way of doing this is to take a
similar filter.  We'll pick edgedetect, but any other should do. You can look
for others using the `./ffmpeg -v 0 -filters|grep ' V->V '` command.
 - sed 's/edgedetect/foobar/g;s/EdgeDetect/Foobar/g' libavfilter/vf_edgedetect.c > libavfilter/vf_foobar.c
 - edit libavfilter/Makefile, and add an entry for "foobar" following the
   pattern of the other filters.
 - edit libavfilter/allfilters.c, and add an entry for "foobar" following the
   pattern of the other filters.
 - ./configure ...
 - make -j<whatever> ffmpeg
 - ./ffmpeg -i http://samples.ffmpeg.org/image-samples/lena.pnm -vf foobar foobar.png
   Note here: you can obviously use a random local image instead of a remote URL.
If everything went right, you should get a foobar.png with Lena edge-detected.
That's it, your new playground is ready.
Some little details about what's going on:
libavfilter/allfilters.c:this file is parsed by the configure script, which in turn
will define variables for the build system and the C:
    --- after running configure ---
    $ grep FOOBAR ffbuild/config.mak
    CONFIG_FOOBAR_FILTER=yes
    $ grep FOOBAR config.h
    #define CONFIG_FOOBAR_FILTER 1
CONFIG_FOOBAR_FILTER=yes from the ffbuild/config.mak is later used to enable
the filter in libavfilter/Makefile and CONFIG_FOOBAR_FILTER=1 from the config.h
will be used for registering the filter in libavfilter/allfilters.c.
Filter code layout
==================
You now need some theory about the general code layout of a filter. Open your
libavfilter/vf_foobar.c. This section will detail the important parts of the
code you need to understand before messing with it.
Copyright
---------
First chunk is the copyright. Most filters are LGPL, and we are assuming
vf_foobar is as well. We are also assuming vf_foobar is not an edge detector
filter, so you can update the boilerplate with your credits.
Doxy
----
Next chunk is the Doxygen about the file. See https://ffmpeg.org/doxygen/trunk/.
Detail here what the filter is, does, and add some references if you feel like
it.
Context
-------
Skip the headers and scroll down to the definition of FoobarContext. This is
your local state context. It is already filled with 0 when you get it so do not
worry about uninitialized reads into this context. This is where you put all
"global" information that you need; typically the variables storing the user options.
You'll notice the first field "const AVClass *class"; it's the only field you
need to keep assuming you have a context. There is some magic you don't need to
care about around this field, just let it be (in the first position) for now.
Options
-------
Then comes the options array. This is what will define the user accessible
options. For example, -vf foobar=mode=colormix:high=0.4:low=0.1. Most options
have the following pattern:
  name, description, offset, type, default value, minimum value, maximum value, flags
 - name is the option name, keep it simple and lowercase
 - description are short, in lowercase, without period, and describe what they
   do, for example "set the foo of the bar"
 - offset is the offset of the field in your local context, see the OFFSET()
   macro; the option parser will use that information to fill the fields
   according to the user input
 - type is any of AV_OPT_TYPE_* defined in libavutil/opt.h
 - default value is an union where you pick the appropriate type; "{.dbl=0.3}",
   "{.i64=0x234}", "{.str=NULL}", ...
 - min and max values define the range of available values, inclusive
 - flags are AVOption generic flags. See AV_OPT_FLAG_* definitions
When in doubt, just look at the other AVOption definitions all around the codebase,
there are tons of examples.
Class
-----
AVFILTER_DEFINE_CLASS(foobar) will define a unique foobar_class with some kind
of signature referencing the options, etc. which will be referenced in the
definition of the AVFilter.
Filter definition
-----------------
At the end of the file, you will find foobar_inputs, foobar_outputs and
the AVFilter ff_vf_foobar. Don't forget to update the AVFilter.description with
a description of what the filter does, starting with a capitalized letter and
ending with a period. You'd better drop the AVFilter.flags entry for now, and
re-add them later depending on the capabilities of your filter.
Callbacks
---------
Let's now study the common callbacks. Before going into details, note that all
these callbacks are explained in details in libavfilter/avfilter.h, so in
doubt, refer to the doxy in that file.
init()
~~~~~~
First one to be called is init(). It's flagged as cold because not called
often. Look for "cold" on
http://gcc.gnu.org/onlinedocs/gcc/Function-Attributes.html for more
information.
As the name suggests, init() is where you eventually initialize and allocate
your buffers, pre-compute your data, etc. Note that at this point, your local
context already has the user options initialized, but you still haven't any
clue about the kind of data input you will get, so this function is often
mainly used to sanitize the user options.
Some init()s will also define the number of inputs or outputs dynamically
according to the user options. A good example of this is the split filter, but
we won't cover this here since vf_foobar is just a simple 1:1 filter.
uninit()
~~~~~~~~
Similarly, there is the uninit() callback, doing what the name suggests. Free
everything you allocated here.
query_formats()
~~~~~~~~~~~~~~~
This follows the init() and is used for the format negotiation. Basically
you specify here what pixel format(s) (gray, rgb 32, yuv 4:2:0, ...) you accept
for your inputs, and what you can output. All pixel formats are defined in
libavutil/pixfmt.h. If you don't change the pixel format between the input and
the output, you just have to define a pixel formats array and call
ff_set_common_formats(). For more complex negotiation, you can refer to other
filters such as vf_scale.
config_props()
~~~~~~~~~~~~~~
This callback is not necessary, but you will probably have one or more
config_props() anyway. It's not a callback for the filter itself but for its
inputs or outputs (they're called "pads" - AVFilterPad - in libavfilter's
lexicon).
Inside the input config_props(), you are at a point where you know which pixel
format has been picked after query_formats(), and more information such as the
video width and height (inlink->{w,h}). So if you need to update your internal
context state depending on your input you can do it here. In edgedetect you can
see that this callback is used to allocate buffers depending on these
information. They will be destroyed in uninit().
Inside the output config_props(), you can define what you want to change in the
output. Typically, if your filter is going to double the size of the video, you
will update outlink->w and outlink->h.
filter_frame()
~~~~~~~~~~~~~~
This is the callback you are waiting for from the beginning: it is where you
process the received frames. Along with the frame, you get the input link from
where the frame comes from.
    static int filter_frame(AVFilterLink *inlink, AVFrame *in) { ... }
You can get the filter context through that input link:
    AVFilterContext *ctx = inlink->dst;
Then access your internal state context:
    FoobarContext *foobar = ctx->priv;
And also the output link where you will send your frame when you are done:
    AVFilterLink *outlink = ctx->outputs[0];
Here, we are picking the first output. You can have several, but in our case we
only have one since we are in a 1:1 input-output situation.
If you want to define a simple pass-through filter, you can just do:
    return ff_filter_frame(outlink, in);
But of course, you probably want to change the data of that frame.
This can be done by accessing frame->data[] and frame->linesize[].  Important
note here: the width does NOT match the linesize. The linesize is always
greater or equal to the width. The padding created should not be changed or
even read. Typically, keep in mind that a previous filter in your chain might
have altered the frame dimension but not the linesize. Imagine a crop filter
that halves the video size: the linesizes won't be changed, just the width.
    <-------------- linesize ------------------------>
    +-------------------------------+----------------+ ^
    |                               |                | |
    |                               |                | |
    |           picture             |    padding     | | height
    |                               |                | |
    |                               |                | |
    +-------------------------------+----------------+ v
    <----------- width ------------->
Before modifying the "in" frame, you have to make sure it is writable, or get a
new one. Multiple scenarios are possible here depending on the kind of
processing you are doing.
Let's say you want to change one pixel depending on multiple pixels (typically
the surrounding ones) of the input. In that case, you can't do an in-place
processing of the input so you will need to allocate a new frame, with the same
properties as the input one, and send that new frame to the next filter:
    AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
    if (!out) {
        av_frame_free(&in);
        return AVERROR(ENOMEM);
    }
    av_frame_copy_props(out, in);
    // out->data[...] = foobar(in->data[...])
    av_frame_free(&in);
    return ff_filter_frame(outlink, out);
In-place processing
~~~~~~~~~~~~~~~~~~~
If you can just alter the input frame, you probably just want to do that
instead:
    av_frame_make_writable(in);
    // in->data[...] = foobar(in->data[...])
    return ff_filter_frame(outlink, in);
You may wonder why a frame might not be writable. The answer is that for
example a previous filter might still own the frame data: imagine a filter
prior to yours in the filtergraph that needs to cache the frame. You must not
alter that frame, otherwise it will make that previous filter buggy. This is
where av_frame_make_writable() helps (it won't have any effect if the frame
already is writable).
The problem with using av_frame_make_writable() is that in the worst case it
will copy the whole input frame before you change it all over again with your
filter: if the frame is not writable, av_frame_make_writable() will allocate
new buffers, and copy the input frame data. You don't want that, and you can
avoid it by just allocating a new buffer if necessary, and process from in to
out in your filter, saving the memcpy. Generally, this is done following this
scheme:
    int direct = 0;
    AVFrame *out;
    if (av_frame_is_writable(in)) {
        direct = 1;
        out = in;
    } else {
        out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
        if (!out) {
            av_frame_free(&in);
            return AVERROR(ENOMEM);
        }
        av_frame_copy_props(out, in);
    }
    // out->data[...] = foobar(in->data[...])
    if (!direct)
        av_frame_free(&in);
    return ff_filter_frame(outlink, out);
Of course, this will only work if you can do in-place processing. To test if
your filter handles well the permissions, you can use the perms filter. For
example with:
    -vf perms=random,foobar
Make sure no automatic pixel conversion is inserted between perms and foobar,
otherwise the frames permissions might change again and the test will be
meaningless: add av_log(0,0,"direct=%d\n",direct) in your code to check that.
You can avoid the issue with something like:
    -vf format=rgb24,perms=random,foobar
...assuming your filter accepts rgb24 of course. This will make sure the
necessary conversion is inserted before the perms filter.
Timeline
~~~~~~~~
Adding timeline support
(http://ffmpeg.org/ffmpeg-filters.html#Timeline-editing) is often an easy
feature to add. In the most simple case, you just have to add
AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC to the AVFilter.flags. You can typically
do this when your filter does not need to save the previous context frames, or
basically if your filter just alters whatever goes in and doesn't need
previous/future information. See for instance commit 86cb986ce that adds
timeline support to the fieldorder filter.
In some cases, you might need to reset your context somehow. This is handled by
the AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL flag which is used if the filter
must not process the frames but still wants to keep track of the frames going
through (to keep them in cache for when it's enabled again). See for example
commit 69d72140a that adds timeline support to the phase filter.
Threading
~~~~~~~~~
libavfilter does not yet support frame threading, but you can add slice
threading to your filters.
Let's say the foobar filter has the following frame processing function:
    dst = out->data[0];
    src = in ->data[0];
    for (y = 0; y < inlink->h; y++) {
        for (x = 0; x < inlink->w; x++)
            dst[x] = foobar(src[x]);
        dst += out->linesize[0];
        src += in ->linesize[0];
    }
The first thing is to make this function work into slices. The new code will
look like this:
    for (y = slice_start; y < slice_end; y++) {
        for (x = 0; x < inlink->w; x++)
            dst[x] = foobar(src[x]);
        dst += out->linesize[0];
        src += in ->linesize[0];
    }
The source and destination pointers, and slice_start/slice_end will be defined
according to the number of jobs. Generally, it looks like this:
    const int slice_start = (in->height *  jobnr   ) / nb_jobs;
    const int slice_end   = (in->height * (jobnr+1)) / nb_jobs;
    uint8_t       *dst = out->data[0] + slice_start * out->linesize[0];
    const uint8_t *src =  in->data[0] + slice_start *  in->linesize[0];
This new code will be isolated in a new filter_slice():
    static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs) { ... }
Note that we need our input and output frame to define slice_{start,end} and
dst/src, which are not available in that callback. They will be transmitted
through the opaque void *arg. You have to define a structure which contains
everything you need:
    typedef struct ThreadData {
        AVFrame *in, *out;
    } ThreadData;
If you need some more information from your local context, put them here.
In you filter_slice function, you access it like that:
    const ThreadData *td = arg;
Then in your filter_frame() callback, you need to call the threading
distributor with something like this:
    ThreadData td;
    // ...
    td.in  = in;
    td.out = out;
    ctx->internal->execute(ctx, filter_slice, &td, NULL, FFMIN(outlink->h, ff_filter_get_nb_threads(ctx)));
    // ...
    return ff_filter_frame(outlink, out);
Last step is to add AVFILTER_FLAG_SLICE_THREADS flag to AVFilter.flags.
For more example of slice threading additions, you can try to run git log -p
--grep 'slice threading' libavfilter/
Finalization
~~~~~~~~~~~~
When your awesome filter is finished, you have a few more steps before you're
done:
 - write its documentation in doc/filters.texi, and test the output with make
   doc/ffmpeg-filters.html.
 - add a FATE test, generally by adding an entry in
   tests/fate/filter-video.mak, add running make fate-filter-foobar GEN=1 to
   generate the data.
 - add an entry in the Changelog
 - edit libavfilter/version.h and increase LIBAVFILTER_VERSION_MINOR by one
   (and reset LIBAVFILTER_VERSION_MICRO to 100)
 - git add ... && git commit -m "avfilter: add foobar filter." && git format-patch -1
When all of this is done, you can submit your patch to the ffmpeg-devel
mailing-list for review.  If you need any help, feel free to come on our IRC
channel, #ffmpeg-devel on irc.freenode.net.