[M4IF Technotes] MS reference sw bug

[dsearles]

The following applies to versions 1.0, and 2.3. Version 2.3
does have some modifications in the functions affected, but
did not completely remove the bug.
Test streams mit013 and mit014 demonstrate the bug. I noticed
green specks in the decoded sequences, that were not present
in the Momusys decoded sequence. I've been running under VC++ 5.0.
The functions decodeIVOP_DataPartitioning(), decodeMBTextureModeOfPVOP_DataPartitioning()
and possibly decodeIVOP_WithShape_DataPartitioning() and
decodePVOP_WithShape_DataPartitioning().
Part of the fix that version 2.3 has is that pointer
m_piMCBPC is initialized to point to an array with one more
than the number of macroblocks in the frame. This is needed
because some things are stored off through the pointer after
it has been incremented past the last macroblock of the frame
(due to mb stuffing after the frame). The real problem, however,
is stuff stored off through pmbmd. This pointer has also been
incremented out of the frame, and out of the memory array
allocated. Even the new decodeMBTextureModeOfPVOP_DataPartitioning()
in version 2.3 stores through pmbmd before the check for the
stuffing code.
The areas that are still questionable to me are in
decodeIVOP_WithShape_DataPartitioning() and
decodePVOP_WithShape_DataPartitioning(). Is mb stuffing
after the "frame" valid? The uses of pmbmd are more complex
in these routines, so the check for mb stuffing needs to
be done earlier, if possible.
Dan Searles
Atmel Multimedia & Communications
3800 Gateway Centre, Suite 311
Morrisville, NC 27560
phone: (919)462-6553 fax: (919)462-0300