[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <063D6719AE5E284EB5DD2968C1650D6D1CCD6FBB@AcuExch.aculab.com>
Date: Fri, 5 Feb 2016 10:07:32 +0000
From: David Laight <David.Laight@...LAB.COM>
To: 'Ingo Molnar' <mingo@...nel.org>, Tom Herbert <tom@...bertland.com>
CC: Linus Torvalds <torvalds@...ux-foundation.org>,
David Miller <davem@...emloft.net>,
Network Development <netdev@...r.kernel.org>,
"Thomas Gleixner" <tglx@...utronix.de>,
Ingo Molnar <mingo@...hat.com>, Peter Anvin <hpa@...or.com>,
the arch/x86 maintainers <x86@...nel.org>,
kernel-team <kernel-team@...com>,
Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
Peter Zijlstra <a.p.zijlstra@...llo.nl>
Subject: RE: [PATCH v3 net-next] net: Implement fast csum_partial for x86_64
From: Ingo Molnar
...
> As Linus noticed, data lookup tables are the intelligent solution: if you manage
> to offload the logic into arithmetics and not affect the control flow then that's
> a big win. The inherent branching will be hidden by executing on massively
> parallel arithmetics units which effectively execute everything fed to them in a
> single cycle.
Not necessarily, in real-life they are likely to be a cache miss.
With the parallel execution units you just need to worry about the
register dependency chains.
In this case the 'adc' have dependencies on both the result register
and the flags register.
The best loop might even be:
10: adc %rax,0(%rdi,%rcx)
lea %rcx,8(%rcx)
jcxz 20f
jmp 10b
20:
You then need to insert just enough copies of the adc so they take as long
as the other instructions.
David
Powered by blists - more mailing lists