Ullrich Koethe wrote:
Joel de Guzman wrote:
I had a look at Fusion, but I'm not sure whether it would be helpful in this context. TinyVector is based on three design goals: it should support the std::vector interface (except for resize etc.),
Like boost::array?
Similar, but more sophisticated. E.g. it has separate view templates that don't own the data, and provides a number of mathematical functions with loop unrolling.
There are some code in libs/fusion/example/performance that shows huge performance improvements with applying fusion algorithms on boost::array over std algorithms. This is due to the fact that fusion does the loop unrolling for you, automatically since it does compile time recursion instead of loops.
Fusion may be more helpful in heterogeneous pixel types, but they often have very specific requirements that are perhaps better handled explicitly.
Like what, for example?
How about RGBA? How would a Fusion version of alpha blending two pixels look like? The formula is
RGB' = alpha2*RGB2 + (1-alpha2)*alpha1*RGB1 alpha' = alpha2 + (1-alpha2)*alpha1
where pixel 2 is in front of pixel 1. Alternatively, with premultiplied alpha we have
RGBA' = RGBA2 + (1-alpha2)*RGBA1
When the individual channels are represented by uint8, we must add functionality to stay in the range 0...255.
Ok, that will be the challenge. Fusion can definitely work on homogeneous as well as heterogeneous structures. And this case will be easier to optimize because of the use of ints (uint8) and simpler _flat_ structures unlike Andy's more complex use of matrices and doubles (c++ can't actually have compile time floats). I bet Fusion's compile time metaprogramming will work in our favor this time. I'll get back to you on this but it might take some time, ok?. Regards, -- Joel de Guzman http://www.boost-consulting.com http://spirit.sf.net