At 01:12 PM 3/19/01 +0200, Matti Aarnio wrote: > Repeat after me: NO FLOATING POINT IN THE KERNEL. > (I will point you to a way to do it below, unless you > can be persuaded to do it in smarter way.) Microsoft's NT architects said the same thing -- no floating-point in the kernel. Most of the time, that's a do-able thing. Consider, though, the DSP coding for a host-signal-processing modem. This is a non-trivial problem that has indeed been coded in scaled integer without too many problems...yet when you benchmark integer code against floating-point code, you may well find that the floating-point code is faster because the FPU performance on modern chips have outstripped the integer equivalents. Further, because of the vast amount of processing done on the data that you get far better results from floating-point code than from integer code. Furthermore, the speed benefit holds when you save the full FPU state, because of the sheer number of FPU operations performed during each context switch, particularly when you bottom-half FPU operations so that you don't do any during interrupts (unless you get a yellow alarm from your data buffering). I understand that real-time video people have the same order of complexity to deal with. Sometimes you can't get enough fixed-point bits to do the job right. Satch - : send the line "unsubscribe linux-net" in the body of a message to majordomo@vger.kernel.org
