Shouldn't we use a run-time cpu detection instead and switch to the optimized SIMD versions then? just dlload an so. I have a bunch of snippets from the xvid sources. (an optimized assembler routines also :)
i already implemented the runtime detection ... the problem is just that gcc wants to have the march flag to compile the inline assembler properly (to assign input and output registers to be exact) ...
i'm not sure, how to solve that in a portable way. i could only pass pointers to the assembly code, but iirc, that results in a few obsolete instructions ... one could rewrite the whole function in assembler, but that wouldn't be a clean solution either ... or it's possible to compile only the file with the simd code with -march ...
if someone has an idea of a solution, i'm interested to hear it ...
cheers ... tim