[dpdk-dev] [PATCH v3 2/3] lpm: add support for NEON

Thomas Monjalon thomas.monjalon at 6wind.com
Fri Feb 12 09:42:43 CET 2016


2016-02-12 12:17, Jerin Jacob:
> On Thu, Feb 11, 2016 at 12:46:33PM +0100, Thomas Monjalon wrote:
> > 2016-01-29 09:40, Jerin Jacob:
> > > --- a/lib/librte_lpm/Makefile
> > > +++ b/lib/librte_lpm/Makefile
> > > +ifneq ($(filter y,$(CONFIG_RTE_ARCH_ARM) $(CONFIG_RTE_ARCH_ARM64)),)
> > > +SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_neon.h
> > > +else
> > >  SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_sse.h
> > > +endif
> > [...]
> > > --- a/lib/librte_lpm/rte_lpm.h
> > > +++ b/lib/librte_lpm/rte_lpm.h
> > > +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64)
> > > +#include "rte_lpm_neon.h"
> > > +#else
> > >  #include "rte_lpm_sse.h"
> > > +#endif
> >
> > Instead of defaulting to x86 SSE, it would be better to replace
> > "else" by "elif X86/SSE".
> > I suggest using RTE_ARCH_X86 or RTE_CPUFLAG_SSEx.
> 
> Some architectures(tile)[1] are planning to emulate SSE instruction used
> in LPM for LPM library support.So that way it makes sense to use SSE as default.

Not sure it is a great idea to emulate instructions of another arch.

> But if anyone has any objections then I can add the check else let
> keep in existing way.

If Tile wants to use x86 code, it's better to do it explicitly (X86 || TILE).

> [1]
> http://dpdk.org/ml/archives/dev/2016-January/031147.html



More information about the dev mailing list