RE: [PATCH v3 net-next] net: Implement fast csum_partial for x86_64

From	David Laight <David.Laight@ACULAB.COM>
Newsgroups	linux.kernel
Subject	RE: [PATCH v3 net-next] net: Implement fast csum_partial for x86_64
Date	2016-02-10 16:30 +0100
Message-ID	<r0ESm-3Vo-15@gated-at.bofh.it> (permalink)
References	<r0Brs-1FM-1@gated-at.bofh.it> <r0EfD-3qm-7@gated-at.bofh.it>
Organization	linux.* mail to news gateway

Show all headers | View raw

From: George Spelvin
> Sent: 10 February 2016 14:44
...
> > I think the fastest loop is:
> > 10:	adcq	0(%rdi,%rcx,8),%rax
> > 	inc	%rcx
> > 	jnz	10b
> > That loop looks like it will have no overhead on recent cpu.
> 
> Well, it should execute at 1 instruction/cycle.

I presume you do mean 1 adc/cycle.
If it doesn't unrolling once might help.

> (No, a scaled offset doesn't take extra time.)
Maybe I'm remembering the 386 book.

> To break that requires ADCX/ADOX:
> 
> 10:	adcxq	0(%rdi,%rcx),%rax
> 	adoxq	8(%rdi,%rcx),%rdx
>  	leaq	16(%rcx),%rcx
> 	jrcxz	11f
>  	j	10b
> 11:

Getting 2 adc/cycle probably does require a little unrolling.
With luck the adcxq, adoxq and leaq will execute together.
The jrcxz is two clocks - so definitely needs a second adcoxq/adcxq pair.

Experiments would be needed to confirm guesses though.

	David

Back to linux.kernel | Previous | Next — Previous in thread | Find similar | Unroll thread

Thread

RE: [PATCH v3 net-next] net: Implement fast csum_partial for x86_64 David Laight <David.Laight@ACULAB.COM> - 2016-02-09 12:00 +0100
  RE: [PATCH v3 net-next] net: Implement fast csum_partial for x86_64 "George Spelvin" <linux@horizon.com> - 2016-02-10 02:00 +0100
    RE: [PATCH v3 net-next] net: Implement fast csum_partial for x86_64 David Laight <David.Laight@ACULAB.COM> - 2016-02-10 12:50 +0100
      RE: [PATCH v3 net-next] net: Implement fast csum_partial for x86_64 "George Spelvin" <linux@horizon.com> - 2016-02-10 15:50 +0100
        RE: [PATCH v3 net-next] net: Implement fast csum_partial for x86_64 David Laight <David.Laight@ACULAB.COM> - 2016-02-10 16:30 +0100

csiph-web