crc32() optimization

Wolfgang Denk wd at denx.de
Sun Nov 10 20:37:33 EST 2002


In message <20021111013114.GB27214 at buici.com> you wrote:
>
> > > What's "Duff's Device"?
> > 
> > It's a tricky way to implement general loop unrolling directly in  C.
> > Applied  to your problem, code that looks like this (instead of 8 any
> > other loop count may be used, but  you  need  to  adjust  the  "case"
> > statements then):
> > 
> > 	register int n = (len + (8-1)) / 8;
> > 
> > 	switch (len % 8) {
> > 	case 0: do {	val = crc32_table ... ;
> > 	case 7:		val = crc32_table ... ;
> > 	case 6:		val = crc32_table ... ;
> > 	case 5:		val = crc32_table ... ;
> > 	case 4:		val = crc32_table ... ;
> > 	case 3:		val = crc32_table ... ;
> > 	case 2:		val = crc32_table ... ;
> > 	case 1:		val = crc32_table ... ;
> > 		} while (--n > 0);
> > 	}
> 
> This doesn't look right to me.  You are decrementing n but using the
> modulus of len in the switch.  The len modulus is correct when n == 1,
> but not when n > 1.  The idea makes sense, but the implementation
> appears to be missing a detail.

You don't understand. The  switch  is  only  needed  for  the  first,
partial loop where we want less than N statements; then we're nunning
the remaining fully unrolled loos in the do{}while loop.

> As for performance problems, I believe that the trouble is evident
> from the assembler output.  The reason that the unrolled loop is more
> efficient than the simple loop is mainly because you don't jump as
> often.  We all know that jumps tend to perturb the instruction fetch
> queue and cache.

Did you enable optimization?

> Here is an example.  I know it is wrong, but it shows how the compiler
> implements the switch.  I've marked the problem jump.  It is executed

Irrelevant here.

> So, the version I posted may not be significantly better that the
> original one that unrolled 6 times.  It is just that unrolling to a

You save the extra while{} loop.


Best regards,

Wolfgang Denk

-- 
Software Engineering:  Embedded and Realtime Systems,  Embedded Linux
Phone: (+49)-8142-4596-87  Fax: (+49)-8142-4596-88  Email: wd at denx.de
See us @ electronica 2002 in Munich, Nov 12-15, Hall A3, Booth A3.325 




More information about the linux-mtd mailing list