changeset 5364:4a743a02eb57

draft v0.1

--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/DOCS/tech/video-filters.txt	Wed Mar 27 02:08:14 2002 +0000
@@ -0,0 +1,52 @@
+First draft on video filter layer by A'rpi
+=================================
+
+Main goals:
+- it should be in libmpcodecs
+- it should reduce number of memcpys and memoru allocation to minimum
+- it should do postprocessing, colorspace convert, scale and crop/expand/flip
+  [is there any other type of filters? maybe we could generalize these]
+- maybe: sub/osd
+
+Implementation:
+
+The video filter plugin should implement the get_image() call (called from
+the video decoder) and call get_image() for the destination image.
+
+so, it would have:
+
+optional:
+    int get_image(mp_image_t* mpi)   // called through control()
+
+    if the filter doesn't implement this call, the core (vd.c) will alloc
+    buffer using memalign() as for disabled direct rendering.
+
+mandatory:
+    mp_vfilter_data* init()
+    
+    void uninit(mp_vfilter_data* vfd)
+    
+    int control(mp_vfilter_data* vfd, ...)
+
+    mp_image_t* process(mp_vfilter_data* vfd, mp_image_t* mpi,(*get_image)())
+    
+    this function will receive an mpi (maybe previously returned by get_image())
+    and should return the filtered mpi. it should use mpcodecs_get_image() to
+    allocate a new mpi.
+    the returned mpi may be EXPORT type, so just doing some pointer/stride
+    tricks (usually enough for crop/expand/flip) without duplicating the
+    image buffer or doing the modifications in the incoming buffer (if its
+    type allows it - so no flags PRESERVE/READABLE set)
+
+mp_vfilter_data would be a transparent (to caller) pointer, used by the
+filter instance to store its internal data, tables, parameters etc.
+it would allow a fliter to be used more than one times in the queue.
+for example: resize in 2 step, postproc+resize+postproc, expand+scale+crop etc
+
+the tricky part of these filters is finding and implementing Direct
+Filtering(C) using get_image().
+
+anyway, as first step you can skip get_image() implementation, and just
+leave it to vd.c core. in most cases it will be enough, direct filtering
+only helps with simple filters (crop/expand/flip) in most cases.
+