[FFmpeg-devel] [PATCH 1/3] swscale/input: Fix several invalid shifts related to rgb2yuv constants

[Andreas Rheinhardt]

On Wed, Jan 22, 2020 at 6:44 PM Michael Niedermayer <michael at niedermayer.cc>
wrote:

> On Wed, Jan 22, 2020 at 03:42:25PM +0100, Andreas Rheinhardt wrote:
> > On Tue, Jan 21, 2020 at 11:18 PM Michael Niedermayer
> <michael at niedermayer.cc>
> > wrote:
> >
> > > Fixes: Invalid shifts
> > > Fixes: #8140
> > > Fixes: #8146
> > >
> > > Signed-off-by: Michael Niedermayer <michael at niedermayer.cc>
> > > ---
> > >  libswscale/input.c | 8 ++++----
> > >  1 file changed, 4 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/libswscale/input.c b/libswscale/input.c
> > > index 064f8da314..159f70307d 100644
> > > --- a/libswscale/input.c
> > > +++ b/libswscale/input.c
> > > @@ -286,8 +286,8 @@ static av_always_inline void
> > > rgb16_32ToUV_c_template(int16_t *dstU,
> > >                                                       int gsh, int bsh,
> > > int S,
> > >                                                       int32_t *rgb2yuv)
> > >  {
> > > -    const int ru       = rgb2yuv[RU_IDX] << rsh, gu = rgb2yuv[GU_IDX]
> <<
> > > gsh, bu = rgb2yuv[BU_IDX] << bsh,
> > > -              rv       = rgb2yuv[RV_IDX] << rsh, gv = rgb2yuv[GV_IDX]
> <<
> > > gsh, bv = rgb2yuv[BV_IDX] << bsh;
> > > +    const int ru       = rgb2yuv[RU_IDX] * (1 << rsh), gu =
> > > rgb2yuv[GU_IDX] * (1 << gsh), bu = rgb2yuv[BU_IDX] * (1 << bsh),
> > > +              rv       = rgb2yuv[RV_IDX] * (1 << rsh), gv =
> > > rgb2yuv[GV_IDX] * (1 << gsh), bv = rgb2yuv[BV_IDX] * (1 << bsh);
> > >      const unsigned rnd = (256u<<((S)-1)) + (1<<(S-7));
> > >      int i;
> > >
> > > @@ -314,8 +314,8 @@ static av_always_inline void
> > > rgb16_32ToUV_half_c_template(int16_t *dstU,
> > >                                                            int gsh, int
> > > bsh, int S,
> > >                                                            int32_t
> > > *rgb2yuv)
> > >  {
> > > -    const int ru       = rgb2yuv[RU_IDX] << rsh, gu = rgb2yuv[GU_IDX]
> <<
> > > gsh, bu = rgb2yuv[BU_IDX] << bsh,
> > > -              rv       = rgb2yuv[RV_IDX] << rsh, gv = rgb2yuv[GV_IDX]
> <<
> > > gsh, bv = rgb2yuv[BV_IDX] << bsh,
> > > +    const int ru       = rgb2yuv[RU_IDX] * (1 << rsh), gu =
> > > rgb2yuv[GU_IDX] * (1 << gsh), bu = rgb2yuv[BU_IDX] * (1 << bsh),
> > > +              rv       = rgb2yuv[RV_IDX] * (1 << rsh), gv =
> > > rgb2yuv[GV_IDX] * (1 << gsh), bv = rgb2yuv[BV_IDX] * (1 << bsh),
> > >                maskgx   = ~(maskr | maskb);
> > >      const unsigned rnd = (256U<<(S)) + (1<<(S-6));
> > >      int i;
> > > --
> >
> > When I looked into this last year, I was unable to ascertain that the
> > multiplication wouldn't overflow (in which case one would need to go the
> > unsigned route, most easily by making the rgb2yuv parameter a uint32_t
> *).
> > Can you?
>
> If there is a situation where these overflow then this needs to be looked
> at with a testcase. Simply using unsigned would likely not produce a
> outcome the user wants
>
> The outcome would not change at all (presuming the ordinary implementation
specified conversion unsigned->int): rgb2yuv is only used to initialize
these const ints that you see in the diff and using unsigned would just
mean that the conversion needs to wrap around.

But I agree that there is no reason to use unsigned for the calculation if
there is no testcase showing an overflow. So proceed as you wish.

- Andreas