Is this combination supposed to work? It appears that this border filter results in memory corruption if the kernel specified is 16s rather than 32f (in ippiFilterBorderInit_16s rather than ippiFilterBorderInit_32f). Other pixel types (e.g., 16u, 16s, 8u) work with both types of kernel, but the 32f pixel type appears to operate correctly only with a 32f kernel. The workaround is easy enough (i.e., use a float kernel for float pixels), but the documentation and return code do not indicate that the 16s kernel is not supported for 32f pixels. This is with the 2018 update 1 version of IPP.
ippiFilterBorder_32f_C1R with ipp16s kernel is not supported because internal implementation uses float data type. We add BadStatus in next IPP release.