[FFmpeg-devel] [PATCH] delogo filter: new "uglarm" interpolation mode added

[Nicolas George]

Nicolas George (2019-01-02):
> Uwe Freese (2019-01-01):
> > > This can be optimized, and since it is the inner loop of the filter, I
> > > think it is worth it. You can declare a pointer that will stay the same
> > > for the whole y loop:
> > > 
> > > 	double *xweight = uglarmtable + (y - logo_y1) * (logo_w - 1);
> > > 
> > > and then use it in that loop:
> > > 
> > > 	interpd += ... * xweight[abs(bx - (x - logo_x1))];
> > > 
> > > It avoids a lot of multiplications.
> 
> > Secondly, I don't exactly understand how *xweight in your example should be
> > used (and would mean smaller or easier code).
> 
> XXX

Sorry, forgot to fill that part before sending.

Completely untested code:

    double *table_t, *table_b, *table_l, *table_r;

    table_t = uglarmtable + x + table_stride * y;
    table_b = uglarmtable + x + table_stride * (logo_h - y - 1);
    table_l = uglarmtable + table_stride * (y - 1) + x;
    table_r = uglarmtable + table_stride * (y - 1) + logo_w - x - 1;
    /* top+bottom on the left of the current point */
    for (bx = 0; bx < x; bx++) {
	weightsum += table_t;
	weightsum += table_b;
	table_t--;
	table_b--;
    }
    /* top+bottom on the right of the current point */
    for (; bx < logo_w; bx++) {
	weightsum += table_t;
	weightsum += table_b;
	table_t++;
	table_b++;
    }
    /* left+right above the current point */
    for (by = 1; by < y; by++) {
	weightsum += *table_l;
	weightsum += *table_r;
	table_l -= table_stride;
	table_r -= table_stride;
    }
    /* left+right below the current point */
    for (; by < logo_w - 1; by++) {
	weightsum += *table_l;
	weightsum += *table_r;
	table_l += table_stride;
	table_r += table_stride;
    }
    av_assert2(table_t == uglarmtable + (logo_w - x) + table_stride * y);
    av_assert2(table_b == uglarmtable + (logo_w - x) + table_stride * (logo_h - y - 1));
    av_assert2(table_l == uglarmtable + table_stride * (logo_h - y - 1) + x);
    av_assert2(table_r == uglarmtable + table_stride * (logo_h - y - 1) + logo_w - x - 1;

That makes more lines, but the lines are way simpler: no tricky
arithmetic, all blocks almost identical, with the changes easy to see
and understand.

Regards,

-- 
  Nicolas George