Just started to work with the Intel Performance Primitives. Looks like a fantastic software library !
As a first test, I tried ippiFilterMedianBorder. The u8 and u16 versions are blazingly fast, ippiFilterMedianBorderr_32f_C1R is quite slow and ippiFilterMedianBorderr_32f_C3R and ippiFilterMedianBorderr_32f_AC4R are missing.
This suggests that the u8 and u16 implementation is based on histograms (e.g. as discussed in Fast Median and Bilateral Filtering,
Ben Weiss†, Shell & Slate Software Corp or in Median Filtering in Constant Time, Simon Perreault and Patrick Hébert).
However, for a fast floating-point implementation, may I suggest the "dancing links" solution described in "Median Filtering is Equivalent to Sorting
Jukka Suomela, Helsinki Institute for Information Technology HIIT" ?
The Intel document "[APMF] Fast Algorithms for Median Filtering, Application Note, Intel Corp. Document number 79835, 2001." doesn't seem to be available anywhere on the web.