Quote:
Originally Posted by systematika  Also AVX is totally going to rip the crap out of DSPs and may even make them totally obsolete in any application other than dedicated systems. Advanced Vector Extensions - Wikipedia, the free encyclopedia
It's nice to have that load off for more important stuff though, and you get that functionality for the price of the smallest waves bundle WITH the bundle included as well! |
I'm not so sure that any SIMD type of scheme is going really whack DSPs. Look at the CUDA system, which is SIMD on a stupid scale. But it's still not really that great for audio processing. Most or much of audio processing requires feeding the previous samples or the results of previous samples forward into subsequent calculations, so they aren't nearly as conducive to SIMD type processing as other types of applications like graphics. They will generally have much more divergence which falls to the application programmer to deal with manually.
I still say that if you want to create the killer audio processing system, then create a board with 256 very simple but very fast DSPs that just handle core DSP functionality, with a queing system in both directions. You queue up your problem packets with input and output buffers, the manager on the board farms the queued packets out as fast as possible to available DSPs and queues back up the output to you. It would have a number of built in higher level algorithms in firmware that the packet can request be used. It would also support chaining, where the resulting data can be turned around on chip and feed back in for another operation with the previous output buffer becoming the input buffer. Some of the higher level algorithms might just be pre-fab chains. And of course you could provide some flags and comparison values to be used in the process, via the info you queue up.
I think that ultimately that would be vastly faster than any vector based system for audio processing. I could be wrong but I think it would be the case due to the nature of audio.