On Wed, Jun 29, 2016 at 05:10:31PM +0530, Jerin Jacob wrote:
> On Wed, Jun 29, 2016 at 01:25:35PM +0200, Thomas Monjalon wrote:
> > 2016-06-29 16:48, Jerin Jacob:
> > > On Mon, Jun 27, 2016 at 04:59:42PM +0200, Thomas Monjalon wrote:
> > > > 2016-06-27 20:18, Jerin Jacob:
> > > > > On Mon, Jun 27, 2016 at 04:19:57PM +0200, Thomas Monjalon wrote:
> > > > > > 2016-06-27 17:24, Jerin Jacob:
> > > > > > > --- a/config/common_base
> > > > > > > +++ b/config/common_base
> > > > > > > @@ -267,6 +267,7 @@ CONFIG_RTE_LIBRTE_VIRTIO_DEBUG_RX=n
> > > > > > > CONFIG_RTE_LIBRTE_VIRTIO_DEBUG_TX=n
> > > > > > > CONFIG_RTE_LIBRTE_VIRTIO_DEBUG_DRIVER=n
> > > > > > > CONFIG_RTE_LIBRTE_VIRTIO_DEBUG_DUMP=n
> > > > > > > +CONFIG_RTE_LIBRTE_VIRTIO_INC_VECTOR=y
> > > > > >
> > > > > > I don't remember what means INC_VECTOR?
> > > > > > Why a config option is needed for vector implementations?
> > > > >
> > > > > I thought of adding additional configuration option(INC_VECTOR)
> > > > > _apart_ from
> > > > > cpu flag based scheme in the patch because even though if a given
> > > > > platform
> > > > > has cpu instruction support, in some platforms scalar version may
> > > > > perform well wrt vector version(based on instruction latency,
> > > > > emulation required or not
> > > > > etc). So a top level flag INC_VECTOR, can override the vector
> > > > > selection
> > > > > for a given platform if required.
> > > >
> > > > Isn't it a runtime driver option needed to disable vector virtio?
> > > >
> > > > > Regarding INC_VECTOR(INC in vector configuration name, I have no
> > > > > idea, I
> > > > > followed the existing flags)
> > > > > $ grep "INC_VECTOR" config/common_base
> > > > > CONFIG_RTE_IXGBE_INC_VECTOR=y
> > > > > CONFIG_RTE_LIBRTE_I40E_INC_VECTOR=y
> > > > > CONFIG_RTE_LIBRTE_FM10K_INC_VECTOR=y
> > > >
> > > > If the flag is really needed I would suggest VIRTIO_VECTOR.
> > >
> > > OK I will change to VIRTIO_VECTOR
> >
> > I would prefer a runtime option.
+1
>
> OK
>
> The platform I test their was NO need for additional VIRTIO_VECTOR
> configuration as NEON versions outperforms than scalar version.
>
> I thought of adding this option to override for any platform if it need
> to accommodate such platform differences NEON vs scalar versions.
>
> I will change completely to run-time detection based on cpuflags for IA and
> ARM.
> Any objections?
Nope from me.
--yliu
