by theover » Wed Apr 24, 2013 12:04 pm
Well, the question of FPGA versus Adaptiva chip should be a matter of which parts of the encoding/decoding algorithms/signal processing blocks fit better on either. That is not that easy to determine,but in general random logic and small, high bandwidth connected not huge signal processing blocks with moderate amounts of very fast local memory fit on the FPGA, and can be tightly connected to the ARM, and pretty ok connected with the Epiphany chips interface. I know that if the Xilinx 7020 is used that is a pretty powerful FPGA, although there sure are bigger Kintex-7 and such (which however probably take almost forever to program). The pipelined floating point (I don't recall by heart if there is double precision like the latest Nvidia Cuda Kepler graphics processors) and possibly fixed point computation rate of the Epiphany is considerable, and if the Risc architecture functions well in practice, small grained task switching and multiple instructions acting on data streams should be fairly efficient.
My remark about the graphics card is about advanced video filters, which make High Definition motion (also deinterlace) and sharpness pretty good in practice. Those filters aren't Open Source (only the interface in VDPAU) and when done on full HD work fine on recent cards, but do take serious energy from for instance my GT640, and on my Asus 2d gen. I7 notebook the 540 card under windows can do motion filtering, but on HD only just (and it gets hot doing so).
My work with the Schroedinger (early Linux command line tool and recently self-compiled ffmpeg) was done on TV recordings, High Defintion BBC and ITV recordings, and transcoding of my own (Sony) .m2t high definition video files. On TV and HD fiulms the effect is a much nicer image, more visually balanced, emphasis of th right details and dither, and the transcoded results can be yet again transcoded to h264 with good results. On my own HD recordings, going to Schroediner makes the result (and the frames) more film-like (requires good HD video filming skills), especially when preceded by VDPAU mpeg-2 decoding and deinterlacing. When the result of that is put into h264, youtube is clueless how to co *completely* mess up the result, is a nice side-effect.
Of course it would be nice to have a networked box (I use a gigaether switch at the moment) which can do real-time HD h264 and realtime Schroedinger en and decoding, but that challenges the network data-rate: maybe 4 b/s or 10Gb/s interfaces wuld be needed, especially forraw frames at more than 420 and low frame rates. I can decode up to 422 1080HD Schroedinger in software, but that does require power (I mean Watts).
Recently i had some discussion with people into HD recording concerts in the Amsterdam Concert Hall, because of a demonstration of their cameras and remote control and recording equipment. Suffice it to say I believe high quality encoders shoul dbe in demand, to capture and nicely compress all those wonderful pixels and their color-bits accurately, and with an eye for filmic correctness. of course full feature film multi level very high bandwidth Dirac/Schroeding variations wouldbe welcome too, up to Ultra HD or at least in 4k as you're I suppose talking about when you mention the Red system, but that may be a bit ambitious. I recall, I had an early Blackfin (Analog Devices) DSP board where the software from Analog's site or CD included wavelet transform code. At the time it seemed hopeless to me but nowadays Schroedinger is going strong! Working with a nice 4k (Red) system is a dream for me at this moment, but for film and High Q (HD) TV is a good challenge, which deserves a filmic approach, which partially can come from a good Schoedinger codec box!
To answer some of your suggestions, I once worked on a EE graphics project (before NVidia took some of the ideas along somehow) for accelerating ray-tracing/radiosity computer graphics computations, so I kno my share of architectural design of such machines, which is a serious job and far from easy. Of course wavelets are a generality coming from Physics (in nuclear physics they are in un-windowed version a possible functional basis for solving multi-dimensional differential equations like the well known Schroedinger equation, one of the most important physics equations since Einstein c.s.) whic can also be used to analize sounds, and all kinds of coherent multi-dimensional signals.
T.V
