Where possible, the notation dest = op (dest, src) applies to both floats residing in an MMX register; the operation is executed in parallel on them. The bold instructions were introduced on Athlon.

SIMD-FP solution
pi2fd	dest = float(dwordsrc)
pf2id	dest = dword(floatsrc)
pi2fw	dest[63..32] = float(wordsrc[47..32]) dest[32..0] = float(wordsrc[15..0])
pf2iw	dest[63..48] = 0 dest[47..32] = word(floatsrc[63..32]) dest[31..16] = 0 dest[15..0] = word(floatsrc[31..0])
pfacc	dest.hi = src.hi + src.lo, dest.lo = dest.hi + dest.lo
pfnacc	dest.hi = src.hi - src.lo, dest.lo = dest.hi - dest.lo
pfpnacc	dest.hi = src.hi + src.lo, dest.lo = dest.hi - dest.lo
pfadd	dest = dest + src
pfsub	dest = dest - src
pfsubr	dest = src - dest
pfcmpeq	dest = (dest == src) ? 0xFFFFFFFF : 0
pfcmpge	dest = (dest >= src) ? 0xFFFFFFFF : 0
pfcmpgt	dest = (dest > src) ? 0xFFFFFFFF : 0
pfmin	dest = min (dest, src)
pfmax	dest = max (dest, src)
pfmul	dest = dest * src
pfrcp	dest.hi = dest.lo = approx15(1/src.lo)
pfsqrt	dest.hi = dest.lo = approx15(1/sqrt(src.lo))
pfrcpit1	first iteration of reciprocal approximation
pfrcpit2	second it. of reciprocal and recip. sqrt approx.
pfrsqit1	first it. of recip. sqrt approx.
Extensions to MMX
pavgusb	dest = average (dest, src) (on unsigned bytes)
pmulhrw	used instead of pmulhw, for fixed point math
pswapd	dest.hi = src.lo, dest.lo = src.hi
femms	fast empty MMX state
prefetch	prefetch data to L1 cache
prefetchw	on current processors does not differ from prefetch