[1167] in linux-net channel archive
Re: 8390 eth_copy_and_sum() improvements
daemon@ATHENA.MIT.EDU (Tom May)
Thu Oct 5 02:23:42 1995
Date: Wed, 4 Oct 1995 18:25:57 -0700
From: ftom@netcom.com (Tom May)
To: gpg109@rsphy1.anu.edu.au
CC: iialan@www.linux.org.uk, linux-net@vger.rutgers.edu
In-reply-to: <9510040615.AA16476@rsphy9.anu.edu.au> (message from Paul Gortmaker on Wed, 4 Oct 1995 16:15:26 +1000 (EST))
>My eth_copy_and_sum() changes will be going into the 1.3.31 kernel for
>the 8390 cards that can use it (wd, ultra, 3c503, e2100, ac3200). When
>I originally did these patches about two months ago, you wrote:
>> Before you benchmark, the csum_partial_copy() routine should be
>> changed to use the inner loop from the original csum_partial_copyffs()
>> function. I sent Alan a patch for this a while back. I suppose I
>> should benchmark it to make sure it is worthwhile.
>I was just wondering if this has been done yet? Was/is it worthwhile?
Hmm, I think I was referring to the style of loop that loads up two
registers at once in typical Pentium style. The current code uses a
single register, so I guess it hasn't been changed yet. I guess I
still need to benchmark it; in theory using two registers is better,
but this is the nasty real world.
>(I avoid assembly like the plague whenever possible ;-)
The benefits of understanding assembly and how processors work are
huge, even if you don't write assembly.
Tom.
