[libcvd-members] Subimages

[Edward Rosten]

OK, so I've added preliminary subimage support to libcvd, int the 
subimage2 branch.

The design is pretty striaghtforward, but herre's an overview:

BasicImage now derives from SubImage.

SubImage provides basic operations, like [], .size(), etc, and a new 
.sub_image() function. SubImages work like BasicImage, they're just 
unmanaged smart-ish pointers.

BasicImage provides stronger guarantees (ie contiguous data), and Image 
provides even more (ie SIMD aligned data).

The changes should be 100% compatible, (modulo a recompile), so most stuff 
(eg all the algorithms will only currently work on BasicImages).


SubImage has iterator support, so you can now copy an area of an image 
with std::copy, for instance.

BasicImage<T> still uses T* for iterators.


look at subimage_test.cc, for some examples. Not the use of .fastend() as 
well as .end(). If you compile with -O3 -funroll-loops, then fastend() has 
a 2x speedup over .end(), is nearly as fast as raw pointers and is faster 
than using a double for-loop over an area of an image.

Without -funroll-loops, fastend() is slower than end(), and they're all 
about 1/2 the speed as with -funroll-loops.

Unfortunately, standard algorithms don't get the benefit of fastend(), 
since they require the begin and end iterators to be of the same type.


The implementation is still preliminary (no const iterator support, for 
instance).


I've tried to make it minimally invasive, but 4 bytes have been added to 
the size of all BasicImage derived classes.


Comments/suggestions?

Is this reasonable?

-Ed