[FFmpeg-cvslog] r20552 - trunk/libavcodec/mjpegdec.c

[Reimar Döffinger]

Hello,
I unfortunately missed the reply...

On Mon, Nov 23, 2009 at 07:44:41PM +0100, Michael Niedermayer wrote:
> On Sun, Nov 22, 2009 at 01:10:36PM +0100, Reimar D?ffinger wrote:
> > On Sun, Nov 22, 2009 at 12:26:26PM +0100, Michael Niedermayer wrote:
> > > On Fri, Nov 20, 2009 at 03:57:59PM +0100, Reimar D?ffinger wrote:
> > > > On Fri, Nov 20, 2009 at 02:21:53PM +0100, Michael Niedermayer wrote:
> > > > > I dont think this is correct
> > > > > The test should be s->height % (s->v_max*8)
> > > > > 
> > > > > besides this the fliped code also looks suspect
> > > > 
> > > > Hmm. Does anyone actually have or can create a sample that should be flipped
> > > > and where the height is not divisible by 8 or so?
> > > 
> > > you can just set flipped = 1 with a normal jpeg to test it i guess 
> > 
> > I'm not sure about that, at least I do not know whether the padding is
> > on top or on bottom for the flipped image, and using a normal jpeg can
> > only tell me how that is for normal JPEGs, but not those created as
> > flipped by default I think?
> > 
> > > > Either way, I am quite confident that the check here should be exactly
> > > > for the same thing as whatever is used for the flipping code.
> > > 
> > > i doubt (height / 3) % 8 is correct
> > > example: height=25
> > > height/3 == 8
> > > actually written height= 48
> > 
> > And data[c] is incremented by
> > s->v_scount[i] * (8 * s->mb_height - 1)
> > lines, so all those 48 lines still fit in it even with EMU_EDGE.
> 
> hmm
> 3 * (8 * 2 - 1) = 45
> data->  0,0
>         :
>         0,45
>         0,46
>         0,47
> after this
>         0,0
>         :
> data->  0,45
>         0,46
>         0,47

I admit I completely fail to understand what the does, what v_count means etc. pp.
However, could it be that the code should be
> s->v_scount[i] * 8 * s->mb_height - 1
instead of
> s->v_scount[i] * (8 * s->mb_height - 1)
My point is just, that as long as we do not have to crop, if the picture fits in
the one way it should also fit in flipped.
So the only issue is when we have to decode more than the picture is high, but regardless
of the value of v_scount that is always the case if the height is not a multiple of 8,
or am I misunderstanding how the code works?