On Tue, Aug 13, 2002 at 12:14:12PM +0100, Steve Harris wrote: > The big advantage (IMHO) of DSP processors is the guaranteed execution > time, even if a current GPU can do more flops on average, you can;t rely > on getting the same ammount of work done each block. For things like > reverb algorithms that you want to use as much processor power as possible > it nive to know where the limits are. > > Not to mention the CPU has to run the host software, and it would be nice > to let it gen on with that.
And the host CPU also runs the waiting cycles needed in certain situations (e.g. IDE, PLIP, spinlocks, etc) not to mention swapping, NFS-data and so on. The host software also means several deamons and even newly exec()ed programs at certain events. A DSP is usally running only one program[1] which is disturbed at most by some interrupts, which it is designed to handle better than the IA-32 in most cases. BTW: These are the main reasons, why it is non-sense to put a OS the size of Linux or worse on a DSP. Some kind of minimal OS might be considered, but a General Purpose OS would defeat most of the DSP advantages. So there are just some standards missing for attaching a DSP to an application under Linux, and hence this is part of our work ;-) Regards Ingo Oeser [1] Ok, some may use threads or if the programmer is really good, than he is even able to implement it in proper co-routines instead. -- Science is what we can tell a computer. Art is everything else. --- D.E.Knuth
