Done. I've called it icolconv.s and I've also put in colconv.c some shift
table initialization code needed by the non MMX asm code.
Tested the whole thing under Windows (but MMX untested).
About the asm code: as I already said it, it doesn't read the destination
rectangle dimensions. Moreover, the width must be 4-aligned. The code is
optimized for memory aligned rectangles (src and dest), the performance
penalty for single misalignment can reach 50% I think.
I haven't added the two 15-bit routines yet.
Patch attached. You will have to manually delete src/win/update.s though.