Code cleanup - emms is not required when MMX block is commented out. Special notes for Michael Niedermayer: Are you still here? If you don't like for(cond;cond;cond) C-constructions and prefer asm ones: "jb 1b" then use .align 16 pseudo assembler instructions else loops rather will be not aligned on correct boundary. (16 it's for K7 for pent should be 8). Your parts have a lot such lacks.