|
Arvind wrote: > I am working on ADPCM G.726 Coder for dual channel. > After completeing coding and testing sucessfully , MIPS(MCPS) > performance is found to be around 15( for full duplex). > The assembly code is fully optimized as it's almost completely serial > and without any software pipeling. > I wanted information from anyone who has worked on this before as to > how many MIPS that had done it in. Also Radisys Comm.,claims to have > done ADPCM in 7.5MCPS (full duplex)!!!!!!!!!!!!!. > Any suggestions on how I can improve on the performance > figures. Am I right to recall that g726 is almost the same as g721 but with a selection of rates? If not, ignore the rest of this message. A couple of years ago, I wrote a g721 for a processor almost the same as a 320c55. It took about 4.5 mips, duplex. I can't tell you details of the tricks I used because the work was done for a client who now owns all the detailed implementation information. However, there is one general point: I spent nearly a month fiddling with C code for the algorithm before going near an assembler. This mostly consisted of experimenting with different ways of performing the arithmetic, checking which alternatives were possible without breaking the bit-exactness. I always write my own C code, even when code already exists as part of the standard. I think it's the only way of getting sufficiently familiar with the algorithm. Tim. ------------------------ Tim Thorpe Tel: +44 1223 240366 Fax: +44 1223 414402 ------------------------ |