Date: Fri, 25 Jul 1997 17:09:00 -0700 (PDT) From: lamaster@george.arc.nasa.gov To: freebsd-multimedia@FreeBSD.ORG Message-ID: <199707260009.RAA15761@george.arc.nasa.gov>
next in thread | raw e-mail | index | archive | help
Tony, Kimball, writes: >> > Quoth Petri Helenius on Fri, 25 July: >> > : Are you speaking of encoder or decoder? The encoders are not cheap nor >> > : likely to be in the immediate future >> > >> > Is the mpeg2 format documented online? I understand that conventional >> > wisdom requires hardware assist, but that just represents an >> > algorithmic challenge.... >> > >> I didn't attend the demonstration myself but there are multiple >> sources claiming that MPEG-2 decode can be done at the ballpark >> power of 233-266 Pentium_II (using MMX extensions) I didn't see a followup which answered the specifics here, so, I hope the following may help muddy the water. *Realtime* MPEG-2 *encoding* remains an expensive proposition, with *quality* MPEG-2 realtime hardware encoders costing > $50K U.S. or more. MPEG-2 is not symmetrical at all in terms of the amount of work it takes to encode versus decode. Encoding is much more expensive, and the higher quality you want [ie the better the image for a given bandwidth] the more expensive it gets. Broadcast-quality encoders are available from only a few vendors. Of course, it may be possible to do a simple lower-quality H/W encoder that doesn't do B-frames, etc., for - a few $K U.S. perhaps, but, I haven't seen it yet. Realtime MPEG-2 decoding, OTOH, should be just about within reach now on the latest processors from several vendors. MPEG-2 decoding has been demonstrated realtime on a modified PPro-200 (some boards like this were floating around about a year ago, with the details secret - but, I saw a non-secret *demo* of the capability), or, presumably, on a Pentium II - 266 using MMX, which should be roughly equivalent to the demo'd PPro 200. Not to mention 300 MHz Ultra 2's, 200 MHz SGI R10k-based systems, etc., using old-fashioned (efficient) programming. MPEG-1 decoding, for example, takes roughly a Pentium-166, a 75 MHz, SuperSPARC, or an R4600-133 MHz, with the minimum requirements then being SPECint92 of > ~110 perhaps, and memory bandwidth > ~60 MB/sec. on these systems. In fact, presumably it can be done with even less CPU power if using certain MPEG-1 assist-capable boards (e.g. supposedly the ATI Mach64 boards can somehow assist MPEG-1 decode with the correct API/library/drivers.) Getting back to MPEG-II decode, it should be is roughly 4X the MPEG-1 requirment, plus more for audio, so, very roughly, on the newer systems, SPECint95 > ~9, and B/W > ~240 MB/sec. This would be a 300 MHz Sun Ultra-3, an SGI Octane, etc., or, a 266 MHz Intel Pentium II, perhaps. That is, the integer performance is there, especially with MMX as a boost, although it may come up a little short on memory bandwidth. In any case, these processors are in the ballpark for doing MPEG-2 decode in software. However, it does take careful optimization of performance-critical loops. "Multimedia" instructions, such as MMX on Intel, but also available on HP, Sun, and eventually on SGI, will help with certain of these performance-critical functions (e.g. conversion from the MPEG-2 color space to RGB). See the IEEE Micro article last year on Multimedia instruction sets. Just to get a vague idea of what a FreeBSD/XFree86/PentiumPro200 system can do, for example, I have tested the H.261 software codec at various combinations of ~1 Mbps, at various H.261 quality levels, at frame rates of 10 fps, 20 fps, 24 fps, etc., and the PPro200 (ASUS Natoma, FreeBSD 3.0-current, XFree86 3.3, vic2.8, 352x288 image, doubled in size on the display side) uses about 20-33% of a PPro-200 CPU, total, for everything, including network overhead. Screen updates at 640x480 are handled OK, but, use much more of the CPU than they could/should (e.g. XF86_S3 using 8% of the CPU at 4 fps, 16% at 8 fps, for full-screen updates - extrapolating to 30 fps, it could use 60% of the CPU.). The H.261 codec is computationally friendly, so the system can handle it, but MPEG-2 decode will want most of the CPU, so it would be nice if the X server were better optimized for painting video images in a window. To summarize, software MPEG-2 decoding should be just within reach now on the fastest commodity microprocessor systems. In order to make it happen, highly optimized MPEG-2 software decoders, and (X Window System) servers are required, but most of the other software components already exist. [Things missing: e.g. I'm not sure if the MPEG-2 RTP encapsulation definition is complete&final.] -Hugh LaMaster Hugh LaMaster, M/S 258-5, ASCII Email: hlamaster@mail.arc.nasa.gov NASA Ames Research Center Or: lamaster@nas.nasa.gov Moffett Field, CA 94035-1000 No Junkmail: USC 18 section 2701 Phone: 415/604-1056 Disclaimer: Unofficial, personal *opinion*.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199707260009.RAA15761>