[FFmpeg-devel] [PATCH v5 2/3] vf_find_rect.c: use the optimized sad function to improve the find performance

[Michael Niedermayer]

On Sat, Jun 15, 2019 at 06:39:38AM +0800, Lance Wang wrote:
> On Sat, Jun 15, 2019 at 3:02 AM Michael Niedermayer <michael at niedermayer.cc>
> wrote:
> 
> > Hi
> >
> > On Wed, Jun 12, 2019 at 06:57:30PM +0800, lance.lmwang at gmail.com wrote:
> > > From: Limin Wang <lance.lmwang at gmail.com>
> > >
> > > benchmark on x86_64: 6.4 -> 16 with below command:
> > > ./ffmpeg  -i 1920x1080.mp4 -vf
> > find_rect=./find.tif,cover_rect=./cover.jpg:mode=cover -f null -
> > > 6.4 fps -> 16fps
> > >
> > > Signed-off-by: Limin Wang <lance.lmwang at gmail.com>
> > > ---
> > >  libavfilter/vf_find_rect.c | 53 +++++++++++++++-----------------------
> > >  1 file changed, 21 insertions(+), 32 deletions(-)
> > >
> > > diff --git a/libavfilter/vf_find_rect.c b/libavfilter/vf_find_rect.c
> > > index ee6c3f4b45..ed15885bc2 100644
> > > --- a/libavfilter/vf_find_rect.c
> > > +++ b/libavfilter/vf_find_rect.c
> > > @@ -26,6 +26,7 @@
> > >  #include "libavutil/imgutils.h"
> > >  #include "libavutil/opt.h"
> > >  #include "internal.h"
> > > +#include "scene_sad.h"
> > >
> > >  #include "lavfutils.h"
> > >  #include "lswsutils.h"
> > > @@ -36,6 +37,8 @@ typedef struct FOCContext {
> > >      AVClass *class;
> > >      float threshold;
> > >      int mipmaps;
> > > +    ff_scene_sad_fn sad;
> > > +    int bitdepth;
> > >      int xmin, ymin, xmax, ymax;
> > >      char *obj_filename;
> > >      int last_x, last_y;
> > > @@ -103,54 +106,40 @@ static AVFrame *downscale(AVFrame *in)
> > >      return frame;
> > >  }
> > >
> > > -static float compare(const AVFrame *haystack, const AVFrame *obj, int
> > offx, int offy)
> > > +static float compare_sad(FOCContext *foc, AVFrame *haystack, AVFrame
> > *obj, int offx, int offy)
> > >  {
> > > -    int x,y;
> > > -    int o_sum_v = 0;
> > > -    int h_sum_v = 0;
> > > -    int64_t oo_sum_v = 0;
> > > -    int64_t hh_sum_v = 0;
> > > -    int64_t oh_sum_v = 0;
> > > -    float c;
> > > +    uint64_t sad = 0;
> > >      int n = obj->height * obj->width;
> > > -    const uint8_t *odat = obj     ->data[0];
> > > +    double mafd;
> > > +    const uint8_t *odat = obj->data[0];
> > >      const uint8_t *hdat = haystack->data[0] + offx + offy *
> > haystack->linesize[0];
> > > -    int64_t o_sigma, h_sigma;
> > > -
> > > -    for(y = 0; y < obj->height; y++) {
> > > -        for(x = 0; x < obj->width; x++) {
> > > -            int o_v = odat[x];
> > > -            int h_v = hdat[x];
> > > -            o_sum_v += o_v;
> > > -            h_sum_v += h_v;
> > > -            oo_sum_v += o_v * o_v;
> > > -            hh_sum_v += h_v * h_v;
> > > -            oh_sum_v += o_v * h_v;
> > > -        }
> > > -        odat += obj->linesize[0];
> > > -        hdat += haystack->linesize[0];
> > > -    }
> > > -    o_sigma = n*oo_sum_v - o_sum_v*(int64_t)o_sum_v;
> > > -    h_sigma = n*hh_sum_v - h_sum_v*(int64_t)h_sum_v;
> > >
> > > -    if (o_sigma == 0 || h_sigma == 0)
> > > -        return 1.0;
> > > +    foc->sad(hdat, haystack->linesize[0], odat, obj->linesize[0],
> > > +            obj->width, obj->height, &sad);
> > > +    emms_c();
> > > +    mafd = (double)sad / n / (1ULL << foc->bitdepth);
> >
> > mixing floating point and MMX in the same function is likely not
> > safe
> >
> 
> The code is changed from vf_freezedetect.c,  it's OK on my testing system.
> That's
> why we had to use emms_c to avoid it.

Nothing stops the compiler from moving any floating
point operations, registers, ... around
putting mmx + emms in a function with floating point is asking for
problems, its better not to do this 
worse such problem would only show up on 32bit mmx systems, so if an issue
occured it would be specific to older systems, thats never good for finding
and fixing issues quickly

thx

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

No great genius has ever existed without some touch of madness. -- Aristotle
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 181 bytes
Desc: not available
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20190616/f7182915/attachment.sig>