🚀 go-pugleaf

In article <3985@phri.UUCP>, roy@phri.UUCP (Roy Smith) writes:
> >       SUM = 0.0
> >       DO 10 I = 1, 1000000
> >       SUM = SUM + VEC(I)
> > 10    CONTINUE
> >
> > A data cache is *no use at all* for this problem.  You will get a
> > cache miss on every data access.  
> 
> 	Well, not quite.  True, you're going to miss on all the VEC(I)
> references, but you should hit on all the I and SUM references.  Rewrite
> the loop as follows (and no, I don't care if I got the test-at-the-top or
> test-at-the-bottom backwards)
> 
> 10	IF (I .GT. 1000000) GOTO 20	# fetch I
> 	I = I + 1			# fetch I, store I
> 	SUM = SUM + VEC(I)		# fetch I, fetch SUM, store SUM
> 	GOTO 10
> 20	CONTINUE
> 
> and you can count 4 data fetches (as pseudo-commented above), all of which
> should be cache hits.  This far outnumbers the 1 cache miss per loop due to
> the main data array accesses.  Some FORTRAN compilers have even been known
> to put constants in memory and refer to them indirectly; if that is true
> here, you get another two cache hits for the 1 and the 1000000.  You could
> argue, of course, that I and SUM might be in registers, but that's a whole
> other argument.
All perfectly right, but the cache miss rate also depends on the cache
line size - how many words are fetched per miss - which might be (say)
128 bytes or 16 words.  In this case, you'd get a miss rate of 1/16 on
the vec(i) references.  Also, some mainframe caches are smart enough to
look ahead and prefetch the next sequential line (usually only on I-fetch,
but sometimes on data as well).  This could result in an actual miss rate
of zero for the example.  Even without this, you have an operand miss
rate of 1/80 or .0125.  Quite respectable.
-- 
Mike Taylor                               ...!{hplabs,amdcad,sun}!amdahl!mat

[ This may not reflect my opinion, let alone anyone else's.  ]