Hi David, On Tue, Feb 9, 2010 at 1:41 AM, David Gay <[email protected]> wrote: > On Mon, Feb 8, 2010 at 11:35 AM, Miklos Maroti <[email protected]> > wrote: >> Janos, Andreas and David, >> >> From scale.h the scale32(x,a,b) call calculates x*a/b with fixed point >> arithmetic without overflow. The offset was added to help rounding, >> but it was not implemented correctly. If you want to calculate x/b, >> then it makes sense to calculate it as (x + b>>1)/b to get the >> rounding effect. However, this is NOT needed here since a/b should be >> close to 1 and in fact the added offset should be a/(2*b). So I think >> this is clearly a bug and should be removed! >> >> Andras, can you please change the code to >> >> async command uint32_t Atm128Calibrate.calibrateMicro(uint32_t n) { >> return scale32(n, cycles, MAGIC); >> } >> >> async command uint32_t Atm128Calibrate.actualMicro(uint32_t n) { >> return scale32(n, MAGIC, cycles); >> } >> >> and rerun your experiments? David, can you confirm that this is indeed a bug? > > Yes, your summary is correct (except that the ideal offset would be > b/2a, which would be ~1/2). > It might be worth forcing the result to be at least 1 as well - do we > want calibrateMicro(1)/actualMicro(1) to sometimes give 0?
Yes, you are right. Attached is a code that calculates ROUND(x*a/b), and I think we should use this instead of FLOOR(x*a/b). We should change this after the next release. I do not think we should care if something becomes 0, since the Alarms take care of that, or people can check it manually, especially then the input must be close to 0 too. Andreas, did you manage to run some tests? Miklos
scale.h
Description: Binary data
_______________________________________________ Tinyos-help mailing list [email protected] https://www.millennium.berkeley.edu/cgi-bin/mailman/listinfo/tinyos-help
