[FFmpeg-devel] [PATCH 1/4] ssim: refactor a weird double loop.

[Ronald S. Bultje]

Hi,

On Sun, Jul 12, 2015 at 6:48 AM, Paul B Mahol <onemda at gmail.com> wrote:

> Dana 12. 7. 2015. 01:56 osoba "Ronald S. Bultje" <rsbultje at gmail.com>
> napisala je:
> >
> > ---
> >  libavfilter/vf_ssim.c | 5 ++---
> >  1 file changed, 2 insertions(+), 3 deletions(-)
> >
> > diff --git a/libavfilter/vf_ssim.c b/libavfilter/vf_ssim.c
> > index 0721ddd..3ef122f 100644
> > --- a/libavfilter/vf_ssim.c
> > +++ b/libavfilter/vf_ssim.c
> > @@ -134,7 +134,7 @@ static float ssim_end1(int s1, int s2, int ss, int
> s12)
> >           / ((float)(fs1 * fs1 + fs2 * fs2 + ssim_c1) * (float)(vars +
> ssim_c2));
> >  }
> >
> > -static float ssim_end4(int sum0[5][4], int sum1[5][4], int width)
> > +static float ssim_endn(int (*sum0)[4], int (*sum1)[4], int width)
> >  {
> >      float ssim = 0.0;
> >      int i;
> > @@ -169,8 +169,7 @@ static float ssim_plane(uint8_t *main, int
> main_stride,
> >                                  &sum0[x]);
> >          }
> >
> > -        for (x = 0; x < width - 1; x += 4)
> > -            ssim += ssim_end4(sum0 + x, sum1 + x, FFMIN(4, width - x -
> 1));
> > +        ssim += ssim_endn(sum0, sum1, width - 1);
> >      }
> >
> >      return ssim / ((height - 1) * (width - 1));
> > --
> > 2.1.2
> >
> >
>
> Why? There was reason behind this code I guess.
>

I think it's for simd code simplification. See, I'm guessing the code you
took from libvpx had an extra condition to do only 4-sized chunks through a
function pointer, and then the odd tail in c code. If you do this, the simd
code has a fixed size (always 4), which makes the implementation much more
trivial: 4 16-byte loads, add, transpose4x4d, and then ssim_end1 to get 4
results, which you horizontal-add and return.

The disadvantage is overhead. First, call overhead since each 4-element
chunk requires a function call, second overhead for function initialization
(anything outside the main loop, either before or after). This includes the
horizontal-add, which is relatively expensive. Third, it limits us to
16-byte: no avx(2). Doing a variable-size function makes the simd slightly
more complex, but is more future-proof (avx/2) and theoretically faster.

Does this change results?


No.

Ronald