[FFmpeg-devel,3/6] avcodec/pictordec: Optimize picmemset() for single plane full lines

Submitted by Michael Niedermayer on Aug. 3, 2019, 11:09 a.m.

Details

Message ID 20190803110949.GQ3219@michaelspb
State Accepted
Commit 5231e89eb9eedc119d4f762469355f83e3628f20
Headers show

Commit Message

Michael Niedermayer Aug. 3, 2019, 11:09 a.m.
On Sat, Aug 03, 2019 at 12:43:32PM +1000, Peter Ross wrote:
> On Sat, Aug 03, 2019 at 01:49:54AM +0200, Michael Niedermayer wrote:
> > Fixes: Timeout (72sec -> 1sec)
> > Fixes: 15512/clusterfuzz-testcase-minimized-ffmpeg_AV_CODEC_ID_PICTOR_fuzzer-5663942342344704
> > 
> > Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
> > Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
> > ---
> >  libavcodec/pictordec.c | 16 +++++++++++++++-
> >  1 file changed, 15 insertions(+), 1 deletion(-)
> > 
> > diff --git a/libavcodec/pictordec.c b/libavcodec/pictordec.c
> > index 2e6fcdca52..5beb03cd5d 100644
> > --- a/libavcodec/pictordec.c
> > +++ b/libavcodec/pictordec.c
> > @@ -66,6 +66,7 @@ static void picmemset(PicContext *s, AVFrame *frame, unsigned value, int run,
> >      int xl = *x;
> >      int yl = *y;
> >      int planel = *plane;
> > +    int pixels_per_value = 8/bits_per_plane;
> >      value   <<= shift;
> >  
> >      d = frame->data[0] + yl * frame->linesize[0];
> > @@ -74,7 +75,7 @@ static void picmemset(PicContext *s, AVFrame *frame, unsigned value, int run,
> >          for (j = 8-bits_per_plane; j >= 0; j -= bits_per_plane) {
> >              d[xl] |= (value >> j) & mask;
> >              xl += 1;
> > -            if (xl == s->width) {
> > +            while (xl == s->width) {
> >                  yl -= 1;
> >                  xl = 0;
> >                  if (yl < 0) {
> > @@ -86,6 +87,19 @@ static void picmemset(PicContext *s, AVFrame *frame, unsigned value, int run,
> >                     mask  <<= bits_per_plane;
> >                  }
> >                  d = frame->data[0] + yl * frame->linesize[0];
> > +                if (s->nb_planes == 1 &&
> > +                    run*pixels_per_value >= s->width &&
> > +                    pixels_per_value < s->width) {
> > +                    int j;
> > +                    for (j = 8-bits_per_plane; j >= 0; j -= bits_per_plane) {
> 
> suggest naming it 'k' to avoid confusion with earlier for loop.

actually, looking at this again, i think we should use the same j,
This also now excludes s->width % pixels_per_value != 0 for which i suspect there
is no testcase. Ill add support for this in case the fuzzer finds a case
that way we then also have a testcase for implementing that corner case.

heres the new code:



[...]

Comments

Michael Niedermayer Aug. 23, 2019, 1:47 p.m.
On Sat, Aug 03, 2019 at 01:09:49PM +0200, Michael Niedermayer wrote:
> On Sat, Aug 03, 2019 at 12:43:32PM +1000, Peter Ross wrote:
> > On Sat, Aug 03, 2019 at 01:49:54AM +0200, Michael Niedermayer wrote:
> > > Fixes: Timeout (72sec -> 1sec)
> > > Fixes: 15512/clusterfuzz-testcase-minimized-ffmpeg_AV_CODEC_ID_PICTOR_fuzzer-5663942342344704
> > > 
> > > Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
> > > Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
> > > ---
> > >  libavcodec/pictordec.c | 16 +++++++++++++++-
> > >  1 file changed, 15 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/libavcodec/pictordec.c b/libavcodec/pictordec.c
> > > index 2e6fcdca52..5beb03cd5d 100644
> > > --- a/libavcodec/pictordec.c
> > > +++ b/libavcodec/pictordec.c
> > > @@ -66,6 +66,7 @@ static void picmemset(PicContext *s, AVFrame *frame, unsigned value, int run,
> > >      int xl = *x;
> > >      int yl = *y;
> > >      int planel = *plane;
> > > +    int pixels_per_value = 8/bits_per_plane;
> > >      value   <<= shift;
> > >  
> > >      d = frame->data[0] + yl * frame->linesize[0];
> > > @@ -74,7 +75,7 @@ static void picmemset(PicContext *s, AVFrame *frame, unsigned value, int run,
> > >          for (j = 8-bits_per_plane; j >= 0; j -= bits_per_plane) {
> > >              d[xl] |= (value >> j) & mask;
> > >              xl += 1;
> > > -            if (xl == s->width) {
> > > +            while (xl == s->width) {
> > >                  yl -= 1;
> > >                  xl = 0;
> > >                  if (yl < 0) {
> > > @@ -86,6 +87,19 @@ static void picmemset(PicContext *s, AVFrame *frame, unsigned value, int run,
> > >                     mask  <<= bits_per_plane;
> > >                  }
> > >                  d = frame->data[0] + yl * frame->linesize[0];
> > > +                if (s->nb_planes == 1 &&
> > > +                    run*pixels_per_value >= s->width &&
> > > +                    pixels_per_value < s->width) {
> > > +                    int j;
> > > +                    for (j = 8-bits_per_plane; j >= 0; j -= bits_per_plane) {
> > 
> > suggest naming it 'k' to avoid confusion with earlier for loop.
> 
> actually, looking at this again, i think we should use the same j,
> This also now excludes s->width % pixels_per_value != 0 for which i suspect there
> is no testcase. Ill add support for this in case the fuzzer finds a case
> that way we then also have a testcase for implementing that corner case.
> 
> heres the new code:
> 
> --- a/libavcodec/pictordec.c
> +++ b/libavcodec/pictordec.c
> @@ -66,6 +66,7 @@ static void picmemset(PicContext *s, AVFrame *frame, unsigned value, int run,
>      int xl = *x;
>      int yl = *y;
>      int planel = *plane;
> +    int pixels_per_value = 8/bits_per_plane;
>      value   <<= shift;
>  
>      d = frame->data[0] + yl * frame->linesize[0];
> @@ -74,7 +75,7 @@ static void picmemset(PicContext *s, AVFrame *frame, unsigned value, int run,
>          for (j = 8-bits_per_plane; j >= 0; j -= bits_per_plane) {
>              d[xl] |= (value >> j) & mask;
>              xl += 1;
> -            if (xl == s->width) {
> +            while (xl == s->width) {
>                  yl -= 1;
>                  xl = 0;
>                  if (yl < 0) {
> @@ -86,6 +87,19 @@ static void picmemset(PicContext *s, AVFrame *frame, unsigned value, int run,
>                     mask  <<= bits_per_plane;
>                  }
>                  d = frame->data[0] + yl * frame->linesize[0];
> +                if (s->nb_planes == 1 &&
> +                    run*pixels_per_value >= s->width &&
> +                    pixels_per_value < s->width &&
> +                    s->width % pixels_per_value == 0
> +                    ) {
> +                    for (; xl < pixels_per_value; xl ++) {
> +                        j = (j < bits_per_plane ? 8 : j) - bits_per_plane;
> +                        d[xl] |= (value >> j) & mask;
> +                    }
> +                    av_memcpy_backptr(d+xl, pixels_per_value, s->width - xl);
> +                    run -= s->width / pixels_per_value;
> +                    xl = s->width;
> +                }
>              }
>          }
>          run--;


will apply

[...]

Patch hide | download patch | download mbox

--- a/libavcodec/pictordec.c
+++ b/libavcodec/pictordec.c
@@ -66,6 +66,7 @@  static void picmemset(PicContext *s, AVFrame *frame, unsigned value, int run,
     int xl = *x;
     int yl = *y;
     int planel = *plane;
+    int pixels_per_value = 8/bits_per_plane;
     value   <<= shift;
 
     d = frame->data[0] + yl * frame->linesize[0];
@@ -74,7 +75,7 @@  static void picmemset(PicContext *s, AVFrame *frame, unsigned value, int run,
         for (j = 8-bits_per_plane; j >= 0; j -= bits_per_plane) {
             d[xl] |= (value >> j) & mask;
             xl += 1;
-            if (xl == s->width) {
+            while (xl == s->width) {
                 yl -= 1;
                 xl = 0;
                 if (yl < 0) {
@@ -86,6 +87,19 @@  static void picmemset(PicContext *s, AVFrame *frame, unsigned value, int run,
                    mask  <<= bits_per_plane;
                 }
                 d = frame->data[0] + yl * frame->linesize[0];
+                if (s->nb_planes == 1 &&
+                    run*pixels_per_value >= s->width &&
+                    pixels_per_value < s->width &&
+                    s->width % pixels_per_value == 0
+                    ) {
+                    for (; xl < pixels_per_value; xl ++) {
+                        j = (j < bits_per_plane ? 8 : j) - bits_per_plane;
+                        d[xl] |= (value >> j) & mask;
+                    }
+                    av_memcpy_backptr(d+xl, pixels_per_value, s->width - xl);
+                    run -= s->width / pixels_per_value;
+                    xl = s->width;
+                }
             }
         }
         run--;