[dpdk-dev] [PATCH v5] examples/l3fwd: em path performance fix

[Jan Viktorin]

On Fri, 18 Mar 2016 12:45:03 +0000
"Kulasek, TomaszX" <tomaszx.kulasek at intel.com> wrote:

> > -----Original Message-----
> > From: Jan Viktorin [mailto:viktorin at rehivetech.com]
> > Sent: Friday, March 18, 2016 12:57
> > To: Thomas Monjalon <thomas.monjalon at 6wind.com>
> > Cc: Jerin Jacob <jerin.jacob at caviumnetworks.com>; Kulasek, TomaszX
> > <tomaszx.kulasek at intel.com>; dev at dpdk.org; jianbo.liu at linaro.org
> > Subject: Re: [dpdk-dev] [PATCH v5] examples/l3fwd: em path performance fix
> > 
> > Hello Thomas, Jerin, Tomasz, all...
> > 
> > On Fri, 18 Mar 2016 12:00:24 +0100
> > Thomas Monjalon <thomas.monjalon at 6wind.com> wrote:
> >   
> > > 2016-03-18 16:22, Jerin Jacob:  
> > > > On Fri, Mar 18, 2016 at 11:04:49AM +0100, Thomas Monjalon wrote:  
> > > > > 2016-03-18 10:52, Tomasz Kulasek:  
> > > > > > +#if !defined(NO_HASH_MULTI_LOOKUP) && defined(__ARM_NEON)  
> > > > >
> > > > > I think we should use CONFIG_RTE_ARCH_ARM_NEON here.
> > > > > Any ARM maintainer to confirm?  
> > > >
> > > > __ARM_NEON should work existing GCC, but it is better to use
> > > > RTE_MACHINE_CPUFLAG_NEON as -it has been generated by probing the
> > > > compiler capabilities.
> > > > -it's future-proof solution to support clang or other gcc versions
> > > > in future  
> > >
> > > I agree to use RTE_MACHINE_CPUFLAG_NEON.
> > >
> > > I just don't understand why CONFIG_RTE_ARCH_ARM_NEON has been  
> > introduced.  
> > > It seems to be used to disable NEON on ARMv7:  
> > 
> > This is true. You should be able to disable the NEON-specific code if it
> > is unwanted. Eg., the memcpy operations are not always faster with NEON.
> > But...
> > 
> > $ git grep ARM_NEON
> > ...
> > lib/librte_eal/common/include/arch/arm/rte_memcpy_32.h:45:#ifdef
> > __ARM_NEON_FP
> > lib/librte_eal/common/include/arch/arm/rte_memcpy_32.h:328:#endif /*
> > __ARM_NEON_FP */ ...
> > 
> > From this point of view, this is wrong and should be fixed to check a
> > different constant.
> >   
> > > 	ifeq ($(CONFIG_RTE_ARCH_ARM_NEON),y)
> > > 	MACHINE_CFLAGS += -mfpu=neon
> > > 	endif  
> > 
> > However, there is another possible way of understanding these options.
> > We can (well, unlikely and I am about to say 'never') have an ARM
> > processor without NEON. This cannot be detected by gcc as it does not know
> > the target processor... So from my point of view:
> > 
> > * CONFIG_RTE_ARCH_ARM_NEON says "my CPU does (not) support NEON" or "I
> >   want to enable/disable NEON" while
> > * RTE_MACHINE_CPUFLAG_NEON says, the _compiler_ supports NEON
> > 
> > I'll send a patch trying to solve this.
> > 
> > Regards
> > Jan  
> 
> Hi
> 
> As I understand with your last patch it's safe and preferred to use RTE_MACHINE_CPUFLAG_NEON for ARM Neon detection? If so, I can include this modification for whole l3fwd in v6 of this patch.

Yes, I'd prefer this approach as well.

Jan

> 
> Tomasz.



-- 
   Jan Viktorin                  E-mail: Viktorin at RehiveTech.com
   System Architect              Web:    www.RehiveTech.com
   RehiveTech
   Brno, Czech Republic