Brian Ristuccia wrote: > > On Mon, Aug 07, 2000 at 08:31:59PM +0100, Andrew Stevens wrote: > > > > [stuff deleted] > > > > I gather the latest mp1e apparently may do B and P frame (which should > > make it rather better than MJPEG). However, as is typical, I can't look at > > the source code to see if it does much (or any) motion compensation. > > However, even without it should get another 50% or so data reduction over > > mjpeg. Its certainly blazingly fast according to Brian Ristuccia. This > > means they're (a) being incredibly clever and I shall cease developing > > mpeg2enc and hacking MMX/SSE grab a copy of mp1e and watch and compress > > movies ;-) (b) they're not doing much motion compensation and I'll see if > > I can merge my motion compensation goodies into mp1e for the best of both > > worlds. The code base *can't* be more obscure than mpeg2enc ;-). > > > > I'd love to be more specific, (mp1e sounds like the answer to this maidens > > prayer), but the web-site (appears to be down) to look at the 1.8 source > > code. Aaaaaargh.... anyone care to put me out of my misery. I gotta look > > at that code. > > > > I don't know if mp1e is doing motion compensation, but it does make good use > of MMX. My copy is at http://osiris.978.org/~brianr/tars/ Thanks Brian my agony is over. It doesn't do motion compensation. The MMXery is indeed clever. However what's even cleverer is the cunning way the author telescopes a lot of arithmetic into a cunning table driven combined quantization/fdct routine. My hat is taken off... I'd only got halfways to where he'd got in a *lot* of solid thinking. Wish he'd put in the odd comment though... those tables weren't filled in at random and its hard to do an error analysis without knowing the math thinking behind the cunningly chosen constants. Andrew
