[FoRK] Reconfigurable Code Re: Software hacks using timestamp counters

Dave Long dave.long at bluewin.ch
Fri Oct 5 06:04:48 PDT 2012


> Do you know anyone else who's attempted this? Any thoughts on  
> whether this will prove useful?


very similar:

http://ecee.colorado.edu/~siek/pohll08.pdf
Siek Karlin Jessup, "Build to Order Linear Algebra Kernels", Workshop  
on Performance Optimization of High-level Languages and Libraries  
(POHLL'08)

The usefulness of this kind of autotuning is probably part of the  
Wheel of Reincarnation.  I've had reasons to do it at least once per  
decade (for instance, loop renesting was worth it for at least one  
SPEC benchmark), and we have folklore that messing with data  
placement and latency is a practice that's gone back all the way to  
the Royal McBee:
> [Mel] wrote the innermost parts of his program loops first so they  
> would get first choice of the optimum address locations on the  
> drum. The optimising assembler wasn't smart enough to do it that way.
>
> Mel never wrote time delay loops either even when the balky  
> Flexowriter required a delay between output characters to work  
> right. He just located instructions on the drum so each successive  
> one was just past the read head when it was needed ...


-Dave

(I'm sure I've mentioned this before on FoRK, but can't dig up the  
reference: at one point I ran across an office management book from  
early last century which recommended a crude hashing system in order  
to avoid "bank contention" at key drawers by the clerical staff.  One  
wonders if ancient persian bureaucrats were given staggered reporting  
intervals so their pony express relay stations could run much closer  
to average than to peak demand...)



More information about the FoRK mailing list