Skip to content
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.


MythTV Video Filter Plugins V0.2

Using Filters

To use filters, create an instance of FilterManager.  FilterManager
scans for available filters, and provides the following methods:

FilterInfo *GetFilterInfoByName(QString Name):
Returns a pointer to a FilterInfo describing the named filter, or NULL
if no such filter is found.

QPtrList<FilterInfo> GetAllFilterInfo(void):
Returns a QPtrList of FilterInfo containing all known filters.  This may
be useful for providing the user with a list of filters to select from.
The list is a duplicate of the FilterManager's internal list, and items
can be removed from it or reordered as desired.  This is a list of
pointers, so the FilterInfo structures should be duplicated if they need
to be modified.

VideoFilter *LoadFilter(FilterInfo *Filt, VideoFrameType inpixfmt,
                        VideoFrameType outpixfmt, int &width,
                        int &height, char *opts):
Returns a pointer to a single VideoFilter, corresponding to the given
FilterInfo, or NULL if the filter cannot be loaded.  Inpixfmt and
outpixfmt give the input and output pixel formats requested of the
filter.  The filter load will fail if the filter cannot provide the
requested format conversion.  Width and height are the dimensions of the
video which will be input to the filter.  If the filter alters the frame
dimensions of the video, width and height will be modified accordingly.
Opts is a NULL-terminated string of options for the video filter, and
its interpretation depends on the filter.

FilterChain *LoadFilters(QString filters, VideoFrameType &inpixfmt,
                         VideoFrameType &outpixfmt, int &width,
                         int &height, int &bufsize):
Returns a pointer to a FilterChain object for the requested filters, or
NULL if such a chain cannot be built.  Filters is a comma-separated
list of filter names, which may each optionally be followed by
=<options> to specify filter options.  Inpixfmt and outpixfmt specify
input and output pixelf formats for the filter chain.  If either is set
to -1 (or FMT_NONE), FilterManager will attempt to choose the format
that best fits the requested chain.  These will only be modified if they
are set to FMT_NONE.  Bufsize will be set to the largest buffer size, in
bytes, required by the filter chain.  This is the preferred method for
loading filters, the FilterChain object provides methods for using and
destroying the filters.

FilterChain objects provide the following method:

void ProcessFrame(VideoFrame *Frame):
Applies filters loaded in the FilterChain to the video frame.  Will
return immediately if Frame==NULL, but does no other checking on the
frame's validity.  It is expected that a FilterChain will never be used
to process video which is not in format and size requested when it was

FilterChain objects also provide all of the methods of QPtrList, most of
which need not be used.  The clear(void) method may occasionally be
useful, for emptying the filter chain and destroying the filters in it.
FilterChains will automatically be cleared upon deletion.

Adding new filters

Writing a new filter is fairly simple.  A filter will need to provide a
structure based on this one:

typedef struct VideoFilter_
  // Public part of filter interface

  int (*filter)(Frame *);
  void (*cleanup)(struct VideoFilter_ *);

  // Members filled in by FilterManager

  void *handle;
  VideoFrameType inpixfmt;
  VideoFrameType outpixfmt;
  char *opts;
  FilterInfo *info;

  // Any private data or functions for this filter
  // follows after this point.

} VideoFilter;

This structure is defined in filter.h and may be used as-is by filters
which do not need to store any private data.  To avoid potential errors
from copying or retyping this structure definition, you may want to use
something like this for filters with private data:

struct MyFilter
  VideoFilter vf;
  int foo myprivateint;

This saves hassle and helps to prevent errors, and allows your code to
recompile without change if new fields are added to VideoFilter (unless
the filter function needs to set the new fields).

To create new filter instances, you must provide a function with the
following prototype:
VideoFilter *new_filter(VideoFrameType inpixfmt,
                        VideoFrameType outpixfmt, int *width,
                        int *height, char *options);

The function name may differ, as the name of the function is provided by
the library's filter_table array.  The parameters are the same as those
used by FilterManager::LoadFilter, minus the FilterInfo, with width and
height passed as pointers to allow their modification.

This function must allocate a new VideoFilter structure, and fill in the
filter and cleanup members with pointers to the filter and cleanup
functions for this filter.  If the filter does not allocate any memory
(aside from that used by the VideoFilter structure), cleanup may be set
to NULL.  None of the other public members of the VideoFilter structure
should be modified by this function.

Here is a sample new_filter function, taken from the invert filter:

VideoFilter *new_filter(VideoFrameType inpixfmt,
                        VideoFrameType outpixfmt, int *width,
                        int *height, char *options)
    VideoFilter *filter;


    if( (inpixfmt != outpixfmt) || (inpixfmt != FMT_YV12 &&
        inpixfmt != FMT_RGB24 && inpixfmt != FMT_YUV422P) )
        return NULL;

    filter = malloc(sizeof(VideoFilter));

    if (filter == NULL)
        fprintf(stderr,"Couldn't allocate memory for filter\n");
        return NULL;
    filter->filter = &invert;
    filter->cleanup = NULL;
    return filter;

The invert filter processes one frame at a time, and does not use any
buffers, so it does not need a cleanup function.  The memory used by the
VideoFilter structure will be freed by the FilterChain into which the
filter is loaded.

Here's the filter function from the invert filter.  It's very simple,
and can invert video in most formats, but cannot convert between

int invert(VideoFilter *vf, VideoFrame *frame)
    int size = frame->size;
    unsigned char *buf = frame->buf;

    while (size--)
        *buf = 255 - (*buf);

    return 0;

As a special case, a filter's init function may return a pointer to a
VideoFilter structure in which the filter function pointer is set to
NULL.  This will cause the filter to be removed from the chain, while
leaving the rest of the chain intact.  See the "force" filter directory
for an example of this feature.

A filter library must also provide an array of FilterInfo structures,
named filter_table.  The FilterInfo structure is defined as follow:

typedef struct FilterInfo_
    char *symbol;    // the symol name of the new filter function
    char *name;      // name of the filter, should not contain any '='
                     // or ',' characters
    char *descript   // a description of the filter
    FmtConv *formats // a pointer to a list of FmtConv, describing the
                     // format conversions which this filter can perform
    char *libname    // the path to the library containing the filter,
                     // will be filled in by FilterManager and should be
                     // set to NULL

The FmtConv struct is defined as follows:

typedef struct FmtConv_
    VideoFrameType in;  // input format for this conversion
    VideoFrameType out; // output format for this conversion

The format conversion list must be terminated with the marker value with
FMT_NULL, and should be ordered by preference, with the filter's fastest
and highest-quality conversions listed first.  The filter_table list
must be terminated by the marker value FILT_NULL.

This is much easier to understand than it is to describe.  Here is an
example filter_table definition, also from the invert filter:

FmtConv FmtList[] =
    { FMT_YV12, FMT_YV12 },
    { FMT_YUV422P, FMT_YUV422P },
    { FMT_RGB24, FMT_RGB24 },

FilterInfo filter_table[] =
        symbol:     "new_filter",
        name:       "invert",
        descript:   "inverts the colors of the input video",
        formats:    FmtList,
        libname:    NULL

Because only the name of the filter's new filter function is needed, and
not a pointer to the function itself, multi-filter libraries can easily
put all of the filter definitions together in a separate source file
from filter implementations.

filter.h also provides several macros for use in benchmarking filters.
To support benchmarking of your filter, add TF_STRUCT in your filter
structure definition, call TF_INIT() with a pointer to your filter
structure in your filter init, add TF_VARS to the variable declaration
of your filter function, and TF_START before the actual filter code, and
call TF_END with a pointer to the filter structure and an informational
string to be prefix to the output before your filter function returns.
Look at the quickdnr, kerneldeint, and denoise3d filters for examples.

The benchmarking macros will do nothing unless activated by defining
TIME_FILTER.  You can also adjust the reporting interval, in frames, by
defining TF_INTERVAL.  If you set TF_TYPE_TSC, the P5 TSC registers are
use for timing, otherwise gettimeofday is used.  If you use the TSC
method, you can use contrib/development/tsc-calibrate/tsc-calibrate.c to
determine the correct value for TF_TSC_TICKS for your system.  If this
value is set, TSC benchmarking will report speeds in frames per second.
Otherwise, they will be reported in ticks per frame.  The gettimeofday
benchmark will always report in frames per second, but may incur more
overhead and/or be slightly less accurate.

This API is subject to change, as the needs of MythTV and of filter
writers may change in the future.

Please email any suggestions or questions to andrewmahone AT eml DOT cc
Something went wrong with that request. Please try again.