[PATCH] - trivial - Improve appletalk checksum calculation

[PATCH] - trivial - Improve appletalk checksum calculation

[Joe Perches]

It's a bit after 2.6.1 now...

Removes unnecessary if, uses 16 bit rotate left.
Performance improves ~30%

Signed-off-by: Joe Perches <joe@perches.com>

diff --git a/net/appletalk/ddp.c b/net/appletalk/ddp.c
index 7c0b515..1c50f4c 100644
--- a/net/appletalk/ddp.c
+++ b/net/appletalk/ddp.c
@@ -925,15 +925,9 @@ static int atrtr_ioctl(unsigned int cmd, void __user *arg)
 static unsigned long atalk_sum_partial(const unsigned char *data,
 				       int len, unsigned long sum)
 {
-	/* This ought to be unwrapped neatly. I'll trust gcc for now */
 	while (len--) {
-		sum += *data;
-		sum <<= 1;
-		if (sum & 0x10000) {
-			sum++;
-			sum &= 0xffff;
-		}
-		data++;
+		sum += *data++;
+		sum = ((sum & 0x8000)>>15) | ((sum & 0x7fff)<<1);
 	}
 	return sum;
 }

Re: [PATCH] - trivial - Improve appletalk checksum calculation

[David Miller]

From: Joe Perches <joe@perches.com>
Date: Mon, 22 Oct 2007 12:36:19 -0700

> It's a bit after 2.6.1 now...
> 
> Removes unnecessary if, uses 16 bit rotate left.
> Performance improves ~30%
> 
> Signed-off-by: Joe Perches <joe@perches.com>

I'm not sure your transformation is equivalent.  Did you
actually test this with appletalk traffic to make sure
the sum is computed correctly?  Or more simply, did you
write a test program to send some test data through the
old and new functions?

Here is what I think the problem is:

>  static unsigned long atalk_sum_partial(const unsigned char *data,
>  				       int len, unsigned long sum)
>  {
> -	/* This ought to be unwrapped neatly. I'll trust gcc for now */
>  	while (len--) {
> -		sum += *data;
> -		sum <<= 1;
> -		if (sum & 0x10000) {
> -			sum++;
> -			sum &= 0xffff;
> -		}
> -		data++;
> +		sum += *data++;
> +		sum = ((sum & 0x8000)>>15) | ((sum & 0x7fff)<<1);
>  	}
>  	return sum;
>  }

The old code is handling overflow, so that when bit 16 gets set by and
addition, that overflow bit gets cleared and then added back into the
sum.

Your code is rotating bit 15 down by one bit and bits 0-14 up by one
bit, my remedial math knowledge tell me that's likely not the same
thing. :-)

Re: [PATCH] - trivial - Improve appletalk checksum calculation

[Joe Perches]

On Mon, 2007-10-22 at 17:35 -0700, David Miller wrote:
> Your code is rotating bit 15 down by one bit and bits 0-14 up by one
> bit.

Yes, a 16 bit rotate left.

There was a discussion a few years ago:
http://oss.sgi.com/archives/netdev/2003-10/msg00734.html

>From the spec:

Implementers of DDP should treat generating the checksum as an optional
feature. The 16-bit DDP checksum is computed as follows:

CkSum := 0 ;
FOR each datagram byte starting with the byte immediately following this
Checksum field
REPEAT the following algorithm:
          CkSum := CkSum + byte; (unsigned addition)
          Rotate CkSum left one bit, rotating the
              most significant bit in least significant bit;
IF, at the end, CkSum = 0 THEN
          CkSum := $FFFF (all ones).

Reception of a datagram with CkSum equal to 0 implies that a checksum is
not performed.

Re: [PATCH] - trivial - Improve appletalk checksum calculation

[David Miller]

From: Joe Perches <joe@perches.com>
Date: Mon, 22 Oct 2007 18:36:28 -0700

> On Mon, 2007-10-22 at 17:35 -0700, David Miller wrote:
> > Your code is rotating bit 15 down by one bit and bits 0-14 up by one
> > bit.
> 
> Yes, a 16 bit rotate left.
> 
> There was a discussion a few years ago:
> http://oss.sgi.com/archives/netdev/2003-10/msg00734.html
> 
> >From the spec:
 ...
> Reception of a datagram with CkSum equal to 0 implies that a checksum is
> not performed.

Ok, but again did you test it?

Re: [PATCH] - trivial - Improve appletalk checksum calculation

[Joe Perches]

On Mon, 2007-10-22 at 18:43 -0700, David Miller wrote:
> Ok, but again did you test it?

Nope.  Stephen Hemminger did in 2003.

Re: [PATCH] - trivial - Improve appletalk checksum calculation

[Herbert Xu]

Joe Perches <joe@perches.com> wrote:
> On Mon, 2007-10-22 at 18:43 -0700, David Miller wrote:
>> Ok, but again did you test it?
> 
> Nope.  Stephen Hemminger did in 2003.

But your code differs significantly from Stephen's version.
However, if it is correct it does look like a good improvement.

So please write a simple test program.  It can't that bad since
there are only 65536 values to test :)

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH] - trivial - Improve appletalk checksum calculation

[Urs Thuermann]

Herbert Xu <herbert@gondor.apana.org.au> writes:

> But your code differs significantly from Stephen's version.
> However, if it is correct it does look like a good improvement.
> 
> So please write a simple test program.  It can't that bad since
> there are only 65536 values to test :)

I think the code is simple enough to see immediately that it should
calculate the same checksum.  But I have written a short test program
to show this and tested it on i686 (gcc-4.2.2) and powerpc (gcc-3.4.5)
without optimization and with -O6.

BTW, shouldn't unsigned long be replaced by unsigned int to avoid
64-bit operations one some platforms?

urs


#include <stdio.h>

unsigned long f1(unsigned long sum)
{
    sum <<= 1;
    if (sum & 0x10000) {
	sum++;
	sum &= 0xffff;
    }
    return sum;
}

unsigned long f2(unsigned long sum)
{
    sum = ((sum & 0x8000)>>15) | ((sum & 0x7fff)<<1);
    return sum;
}

int main()
{
    unsigned long s;

    for (s = 0; s < 65536; s++) {
	if (f1(s) != f2(s) || f1(s) > 65535)
	    printf("%ld\n", s);
    }
    return 0;
}

Re: [PATCH] - trivial - Improve appletalk checksum calculation

[Joe Perches]

On Sun, 2007-10-28 at 21:18 +0100, Urs Thuermann wrote:

> I have written a short test program to show this and tested it on i686
> (gcc-4.2.2) and powerpc (gcc-3.4.5) without optimization and with -O6.

Thanks for that.

> BTW, shouldn't unsigned long be replaced by unsigned int to avoid
> 64-bit operations one some platforms?

probably unsigned long should be u16.

Compiled, untested.

diff --git a/net/appletalk/ddp.c b/net/appletalk/ddp.c
index 7c0b515..c3890cc 100644
--- a/net/appletalk/ddp.c
+++ b/net/appletalk/ddp.c
@@ -922,25 +922,19 @@ static int atrtr_ioctl(unsigned int cmd, void __user *arg)
  * Checksum: This is 'optional'. It's quite likely also a good
  * candidate for assembler hackery 8)
  */
-static unsigned long atalk_sum_partial(const unsigned char *data,
-				       int len, unsigned long sum)
+static u16 atalk_sum_partial(const unsigned char *data,
+			     int len, u16 sum)
 {
-	/* This ought to be unwrapped neatly. I'll trust gcc for now */
 	while (len--) {
-		sum += *data;
-		sum <<= 1;
-		if (sum & 0x10000) {
-			sum++;
-			sum &= 0xffff;
-		}
-		data++;
+		sum += *data++;
+		sum = ((sum & 0x8000) >> 15) | ((sum & 0x7fff) << 1);
 	}
 	return sum;
 }
 
 /*  Checksum skb data --  similar to skb_checksum  */
-static unsigned long atalk_sum_skb(const struct sk_buff *skb, int offset,
-				   int len, unsigned long sum)
+static u16 atalk_sum_skb(const struct sk_buff *skb, int offset,
+			 int len, u16 sum)
 {
 	int start = skb_headlen(skb);
 	int i, copy;
@@ -1010,13 +1004,13 @@ static unsigned long atalk_sum_skb(const struct sk_buff *skb, int offset,
 
 static __be16 atalk_checksum(const struct sk_buff *skb, int len)
 {
-	unsigned long sum;
+	u16 sum;
 
 	/* skip header 4 bytes */
 	sum = atalk_sum_skb(skb, 4, len-4, 0);
 
 	/* Use 0xFFFF for 0. 0 itself means none */
-	return sum ? htons((unsigned short)sum) : htons(0xFFFF);
+	return cpu_to_be16(sum ? sum : 0xFFFF);
 }
 
 static struct proto ddp_proto = {
      * 

Re: [PATCH] - trivial - Improve appletalk checksum calculation

[Joe Perches]

On Sun, 2007-10-28 at 14:01 -0700, Joe Perches wrote:
> probably unsigned long should be u16.

Or not.
x86 u16 performs much worse than u32 (~ 50% worse)

Re: [PATCH] - trivial - Improve appletalk checksum calculation

[Stephen Hemminger]

On Mon, 22 Oct 2007 12:36:19 -0700
Joe Perches <joe@perches.com> wrote:

> It's a bit after 2.6.1 now...
> 
> Removes unnecessary if, uses 16 bit rotate left.
> Performance improves ~30%
> 
> Signed-off-by: Joe Perches <joe@perches.com>
> 
> diff --git a/net/appletalk/ddp.c b/net/appletalk/ddp.c
> index 7c0b515..1c50f4c 100644
> --- a/net/appletalk/ddp.c
> +++ b/net/appletalk/ddp.c
> @@ -925,15 +925,9 @@ static int atrtr_ioctl(unsigned int cmd, void __user *arg)
>  static unsigned long atalk_sum_partial(const unsigned char *data,
>  				       int len, unsigned long sum)
>  {
> -	/* This ought to be unwrapped neatly. I'll trust gcc for now */
>  	while (len--) {
> -		sum += *data;
> -		sum <<= 1;
> -		if (sum & 0x10000) {
> -			sum++;
> -			sum &= 0xffff;
> -		}
> -		data++;
> +		sum += *data++;
> +		sum = ((sum & 0x8000)>>15) | ((sum & 0x7fff)<<1);
>  	}
>  	return sum;
>  }
> 

The end of the message you quoted was:

> Corrected fast code is:
> 
>         while (len--) {
>                 sum += *data++;
>                 sum <<= 1;
>                 sum = (((sum & 0x10000) >> 16) + sum) & 0xffff;
>         }
> 
> At least it is correct on the standalone random data test, and the
> new code is 30% faster for the cached memory case (13.7 clks/byte vs 18 
> clks/byte).

Your code looks different...

-- 
Stephen Hemminger <shemminger@linux-foundation.org>

Re: [PATCH] - trivial - Improve appletalk checksum calculation

[Joe Perches]

On Mon, 2007-10-22 at 20:30 -0700, Stephen Hemminger wrote:
> > Corrected fast code is:
> > 
> >         while (len--) {
> >                 sum += *data++;
> >                 sum <<= 1;
> >                 sum = (((sum & 0x10000) >> 16) + sum) & 0xffff;
> >         }
> > 
> > At least it is correct on the standalone random data test, and the
> > new code is 30% faster for the cached memory case (13.7 clks/byte vs 18 
> > clks/byte).
> Your code looks different...

Both are 16 bit rotate lefts.
Which looks clearer?