Re: [libcvd-members] FAST crashes (again!)

[Edward Rosten]

On Tue, 21 Aug 2007, Ethan Eade wrote:

> > for(int x=16; x < xend; x++)
> >   {
> >    const byte* p = &I[y][x];
> >    __m128i lo, hi;
> >     {
> >       const __m128i here = load_si128<Aligned>((const __m128i*)(p));
> >
> > Notice how p is is being incremented in 1-byte steps, and is then loaded as 
> > if it were 16-byte aligned. Hmm! I think one of the 9 or 10 had this 
> > incremented in 16-byte steps earlier.


> Yes, it seems Ed changed the code to work with SubImage (hence the change to 
> a double loop), but it should be 'x+=16'.

Yes. I also attempted to fix the crashes and edge effects that Georg was 
experiencing on oddly sized images. I didn't understand the code well 
enough to fix it, so I rewrote the bits I didn't understand. I have no 
idea if it is slower, and rather oddly it worked when I tested it. I do 
not see how this could have possibly been the case.

I generally prefer if stuff works on subimages, but not if there's a 
significant penalty. If it's too much slower, you can put back in the old 
code (minus the bugs :-)

But yes, it should be x+=16.


-Ed