Lubomir Bourdev writes:

Hi Ulli,

Thanks for spending time to review GIL.

[...]

...

A traditional STL style solution would probably be easier:

transform(source.begin(), source.end(), destination.begin(), SomeFunctor());

(since implementing a view is more involved than implementing a functor, at least in terms of reading documentation).

They are not quite equivalent. Views can transform both the domain and the range of images. For example, a view may change the way its horizontal and/or vertical iterators advance, and similarly may perform a transformation upon dereferencing (which your SomeFunctor does). We find it appealing to represent this as a single object (view) rather than splitting it into multiple parts, such as ranges and accessors. Yes, I know these are conceptually different transformations, and they are implemented in different classes; they are just combined in the View. It is appealing because conceptually it just represents an abstract image. And you can pipe views nicely together. Another difference is in performance: copy_pixels(src,dst) is usually faster than the STL equivalent std::copy(src.begin(), src.end(), dst.begin()) This is because the one-dimensional iterators are slower, as they have to deal with potential padding at the end of rows, and also because we provide performance specializations.

Seems like no one have mentioned segmented iterators so far. They can make your 1D iterators as fast as 2D counterparts. Oleg Abrosimov.