[dpdk-dev] [PATCH] i40e: improve performance of vector PMD

[dpdk-dev] [PATCH] i40e: improve performance of vector PMD

[Bruce Richardson]

An analysis of the i40e code using Intel® VTune™ Amplifier 2016 showed
that the code was unexpectedly causing stalls due to "Loads blocked by
Store Forwards". This can occur when a load from memory has to wait
due to the prior store being to the same address, but being of a smaller
size i.e. the stored value cannot be directly returned to the loader.
[See ref: https://software.intel.com/en-us/node/544454]

These stalls are due to the way in which the data_len values are handled
in the driver. The lengths are extracted using vector operations, but those
16-bit lengths are then assigned using scalar operations i.e. 16-bit
stores.

These regular 16-bit stores actually have two effects in the code:
* they cause the "Loads blocked by Store Forwards" issues reported
* they also cause the previous loads in the RX function to actually be a
load followed by a store to an address on the stack, because the 16-bit
assignment can't be done to an xmm register.

By converting the 16-bit stores operations into a sequence of SSE blend
operations, we can ensure that the descriptor loads only occur once, and
avoid both the additional store and loads from the stack, as well as the
stalls due to the second loads being blocked.

Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>

---
 drivers/net/i40e/i40e_rxtx_vec.c | 24 ++++++++++--------------
 1 file changed, 10 insertions(+), 14 deletions(-)

diff --git a/drivers/net/i40e/i40e_rxtx_vec.c b/drivers/net/i40e/i40e_rxtx_vec.c
index 047aff5..d0a0cc9 100644
--- a/drivers/net/i40e/i40e_rxtx_vec.c
+++ b/drivers/net/i40e/i40e_rxtx_vec.c
@@ -192,11 +192,7 @@ desc_to_olflags_v(__m128i descs[4], struct rte_mbuf **rx_pkts)
 static inline void
 desc_pktlen_align(__m128i descs[4])
 {
-	__m128i pktlen0, pktlen1, zero;
-	union {
-		uint16_t e[4];
-		uint64_t dword;
-	} vol;
+	__m128i pktlen0, pktlen1;
 
 	/* mask everything except pktlen field*/
 	const __m128i pktlen_msk = _mm_set_epi32(PKTLEN_MASK, PKTLEN_MASK,
@@ -206,18 +202,18 @@ desc_pktlen_align(__m128i descs[4])
 	pktlen1 = _mm_unpackhi_epi32(descs[1], descs[3]);
 	pktlen0 = _mm_unpackhi_epi32(pktlen0, pktlen1);
 
-	zero = _mm_xor_si128(pktlen0, pktlen0);
-
 	pktlen0 = _mm_srli_epi32(pktlen0, PKTLEN_SHIFT);
 	pktlen0 = _mm_and_si128(pktlen0, pktlen_msk);
 
-	pktlen0 = _mm_packs_epi32(pktlen0, zero);
-	vol.dword = _mm_cvtsi128_si64(pktlen0);
-	/* let the descriptor byte 15-14 store the pkt len */
-	*((uint16_t *)&descs[0]+7) = vol.e[0];
-	*((uint16_t *)&descs[1]+7) = vol.e[1];
-	*((uint16_t *)&descs[2]+7) = vol.e[2];
-	*((uint16_t *)&descs[3]+7) = vol.e[3];
+	pktlen0 = _mm_packs_epi32(pktlen0, pktlen0);
+
+	descs[3] = _mm_blend_epi16(descs[3], pktlen0, 0x80);
+	pktlen0 = _mm_slli_epi64(pktlen0, 16);
+	descs[2] = _mm_blend_epi16(descs[2], pktlen0, 0x80);
+	pktlen0 = _mm_slli_epi64(pktlen0, 16);
+	descs[1] = _mm_blend_epi16(descs[1], pktlen0, 0x80);
+	pktlen0 = _mm_slli_epi64(pktlen0, 16);
+	descs[0] = _mm_blend_epi16(descs[0], pktlen0, 0x80);
 }
 
  /*
-- 
2.5.5

Re: [dpdk-dev] [PATCH] i40e: improve performance of vector PMD

[Bruce Richardson]

On Thu, Apr 14, 2016 at 11:15:21AM +0100, Bruce Richardson wrote:
> An analysis of the i40e code using Intel® VTune™ Amplifier 2016 showed
> that the code was unexpectedly causing stalls due to "Loads blocked by
> Store Forwards". This can occur when a load from memory has to wait
> due to the prior store being to the same address, but being of a smaller
> size i.e. the stored value cannot be directly returned to the loader.
> [See ref: https://software.intel.com/en-us/node/544454]
> 
> These stalls are due to the way in which the data_len values are handled
> in the driver. The lengths are extracted using vector operations, but those
> 16-bit lengths are then assigned using scalar operations i.e. 16-bit
> stores.
> 
> These regular 16-bit stores actually have two effects in the code:
> * they cause the "Loads blocked by Store Forwards" issues reported
> * they also cause the previous loads in the RX function to actually be a
> load followed by a store to an address on the stack, because the 16-bit
> assignment can't be done to an xmm register.
> 
> By converting the 16-bit stores operations into a sequence of SSE blend
> operations, we can ensure that the descriptor loads only occur once, and
> avoid both the additional store and loads from the stack, as well as the
> stalls due to the second loads being blocked.
> 
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> 
Self-NAK on this version. The blend instruction used is SSE4.1 so breaks the
"default" build.

Two obvious options to fix this:
1. Keep the old code with SSE4.1 #ifdefs separating old and new
2. Update the vpmd requirement to SSE4.1, and factor that in during runtime
select of the RX code path.

Personally, I prefer the second option. Any objections?

/Bruce

Re: [dpdk-dev] [PATCH] i40e: improve performance of vector PMD

[Ananyev, Konstantin]

> -----Original Message-----
> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Bruce Richardson
> Sent: Thursday, April 14, 2016 2:50 PM
> To: dev@dpdk.org
> Cc: Zhang, Helin; Wu, Jingjing
> Subject: Re: [dpdk-dev] [PATCH] i40e: improve performance of vector PMD
> 
> On Thu, Apr 14, 2016 at 11:15:21AM +0100, Bruce Richardson wrote:
> > An analysis of the i40e code using Intel® VTune™ Amplifier 2016 showed
> > that the code was unexpectedly causing stalls due to "Loads blocked by
> > Store Forwards". This can occur when a load from memory has to wait
> > due to the prior store being to the same address, but being of a smaller
> > size i.e. the stored value cannot be directly returned to the loader.
> > [See ref: https://software.intel.com/en-us/node/544454]
> >
> > These stalls are due to the way in which the data_len values are handled
> > in the driver. The lengths are extracted using vector operations, but those
> > 16-bit lengths are then assigned using scalar operations i.e. 16-bit
> > stores.
> >
> > These regular 16-bit stores actually have two effects in the code:
> > * they cause the "Loads blocked by Store Forwards" issues reported
> > * they also cause the previous loads in the RX function to actually be a
> > load followed by a store to an address on the stack, because the 16-bit
> > assignment can't be done to an xmm register.
> >
> > By converting the 16-bit stores operations into a sequence of SSE blend
> > operations, we can ensure that the descriptor loads only occur once, and
> > avoid both the additional store and loads from the stack, as well as the
> > stalls due to the second loads being blocked.
> >
> > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> >
> Self-NAK on this version. The blend instruction used is SSE4.1 so breaks the
> "default" build.
> 
> Two obvious options to fix this:
> 1. Keep the old code with SSE4.1 #ifdefs separating old and new
> 2. Update the vpmd requirement to SSE4.1, and factor that in during runtime
> select of the RX code path.
> 
> Personally, I prefer the second option. Any objections?

+1 for second one.

> 
> /Bruce

Re: [dpdk-dev] [PATCH] i40e: improve performance of vector PMD

[Iremonger, Bernard]

Hi Bruce,

> -----Original Message-----
> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Ananyev,
> Konstantin
> Sent: Thursday, April 14, 2016 3:00 PM
> To: Richardson, Bruce <bruce.richardson@intel.com>; dev@dpdk.org
> Cc: Zhang, Helin <helin.zhang@intel.com>; Wu, Jingjing
> <jingjing.wu@intel.com>
> Subject: Re: [dpdk-dev] [PATCH] i40e: improve performance of vector PMD
> 
> 
> 
> > -----Original Message-----
> > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Bruce Richardson
> > Sent: Thursday, April 14, 2016 2:50 PM
> > To: dev@dpdk.org
> > Cc: Zhang, Helin; Wu, Jingjing
> > Subject: Re: [dpdk-dev] [PATCH] i40e: improve performance of vector
> > PMD
> >
> > On Thu, Apr 14, 2016 at 11:15:21AM +0100, Bruce Richardson wrote:
> > > An analysis of the i40e code using Intel® VTune™ Amplifier 2016
> > > showed that the code was unexpectedly causing stalls due to "Loads
> > > blocked by Store Forwards". This can occur when a load from memory
> > > has to wait due to the prior store being to the same address, but
> > > being of a smaller size i.e. the stored value cannot be directly returned to
> the loader.
> > > [See ref: https://software.intel.com/en-us/node/544454]
> > >
> > > These stalls are due to the way in which the data_len values are
> > > handled in the driver. The lengths are extracted using vector
> > > operations, but those 16-bit lengths are then assigned using scalar
> > > operations i.e. 16-bit stores.
> > >
> > > These regular 16-bit stores actually have two effects in the code:
> > > * they cause the "Loads blocked by Store Forwards" issues reported
> > > * they also cause the previous loads in the RX function to actually
> > > be a load followed by a store to an address on the stack, because
> > > the 16-bit assignment can't be done to an xmm register.
> > >
> > > By converting the 16-bit stores operations into a sequence of SSE
> > > blend operations, we can ensure that the descriptor loads only occur
> > > once, and avoid both the additional store and loads from the stack,
> > > as well as the stalls due to the second loads being blocked.
> > >
> > > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > >
> > Self-NAK on this version. The blend instruction used is SSE4.1 so
> > breaks the "default" build.
> >
> > Two obvious options to fix this:
> > 1. Keep the old code with SSE4.1 #ifdefs separating old and new 2.
> > Update the vpmd requirement to SSE4.1, and factor that in during
> > runtime select of the RX code path.
> >
> > Personally, I prefer the second option. Any objections?
> 
> +1 for second one.
> 
> >
> > /Bruce

I am using the "default" build when building in VM's, will both options work for me?

Regards,

Bernard.  

[dpdk-dev] [PATCH v2 0/3] improve i40e vpmd

[Bruce Richardson]

This patchset improves the performance of the i40e SSE pmd by removing
operations that triggered CPU stalls. It also shortens the code and
cleans it up a little.

The base requirement for using the SSE code path has been pushed up to
SSE4.1 from SSE3, due to the use of the blend instruction. The instruction
set level is now checked at runtime as part of the driver selection process

Bruce Richardson (3):
  i40e: require SSE4.1 support for vector driver
  i40e: improve performance of vector PMD
  i40e: simplify SSE packet length extraction code

 drivers/net/i40e/Makefile        |  6 ++++
 drivers/net/i40e/i40e_rxtx_vec.c | 59 ++++++++++++++--------------------------
 2 files changed, 27 insertions(+), 38 deletions(-)

-- 
2.5.5

[dpdk-dev] [PATCH v2 1/3] i40e: require SSE4.1 support for vector driver

[Bruce Richardson]

Later commits to improve the driver will make use of the SSE4.1
_mm_blend_epi16 intrinsic, so:
* set the compilation level to always have SSE4.1 support,
* and add in a runtime check for SSE4.1 as part of the condition checks
  for vector driver selection.

Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 drivers/net/i40e/Makefile        | 6 ++++++
 drivers/net/i40e/i40e_rxtx_vec.c | 4 ++++
 2 files changed, 10 insertions(+)

diff --git a/drivers/net/i40e/Makefile b/drivers/net/i40e/Makefile
index 6dd6eaa..56b20d5 100644
--- a/drivers/net/i40e/Makefile
+++ b/drivers/net/i40e/Makefile
@@ -102,6 +102,12 @@ SRCS-$(CONFIG_RTE_LIBRTE_I40E_PMD) += i40e_ethdev_vf.c
 SRCS-$(CONFIG_RTE_LIBRTE_I40E_PMD) += i40e_pf.c
 SRCS-$(CONFIG_RTE_LIBRTE_I40E_PMD) += i40e_fdir.c
 
+# vector PMD driver needs SSE4.1 support
+ifeq ($(findstring RTE_MACHINE_CPUFLAG_SSE4_1,$(CFLAGS)),)
+CFLAGS_i40e_rxtx_vec.o += -msse4.1
+endif
+
+
 # this lib depends upon:
 DEPDIRS-$(CONFIG_RTE_LIBRTE_I40E_PMD) += lib/librte_eal lib/librte_ether
 DEPDIRS-$(CONFIG_RTE_LIBRTE_I40E_PMD) += lib/librte_mempool lib/librte_mbuf
diff --git a/drivers/net/i40e/i40e_rxtx_vec.c b/drivers/net/i40e/i40e_rxtx_vec.c
index 047aff5..1e2fadd 100644
--- a/drivers/net/i40e/i40e_rxtx_vec.c
+++ b/drivers/net/i40e/i40e_rxtx_vec.c
@@ -751,6 +751,10 @@ i40e_rx_vec_dev_conf_condition_check(struct rte_eth_dev *dev)
 	struct rte_eth_rxmode *rxmode = &dev->data->dev_conf.rxmode;
 	struct rte_fdir_conf *fconf = &dev->data->dev_conf.fdir_conf;
 
+	/* need SSE4.1 support */
+	if (!rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE4_1))
+		return -1;
+
 #ifndef RTE_LIBRTE_I40E_RX_OLFLAGS_ENABLE
 	/* whithout rx ol_flags, no VP flag report */
 	if (rxmode->hw_vlan_strip != 0 ||
-- 
2.5.5

[dpdk-dev] [PATCH v2 2/3] i40e: improve performance of vector PMD

[Bruce Richardson]

An analysis of the i40e code using Intel® VTune™ Amplifier 2016 showed
that the code was unexpectedly causing stalls due to "Loads blocked by
Store Forwards". This can occur when a load from memory has to wait
due to the prior store being to the same address, but being of a smaller
size i.e. the stored value cannot be directly returned to the loader.
[See ref: https://software.intel.com/en-us/node/544454]

These stalls are due to the way in which the data_len values are handled
in the driver. The lengths are extracted using vector operations, but those
16-bit lengths are then assigned using scalar operations i.e. 16-bit
stores.

These regular 16-bit stores actually have two effects in the code:
* they cause the "Loads blocked by Store Forwards" issues reported
* they also cause the previous loads in the RX function to actually be a
load followed by a store to an address on the stack, because the 16-bit
assignment can't be done to an xmm register.

By converting the 16-bit store operations into a sequence of SSE blend
operations, we can ensure that the descriptor loads only occur once, and
avoid both the additional stores and loads from the stack, as well as the
stalls due to the blocked loads.

Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 drivers/net/i40e/i40e_rxtx_vec.c | 24 ++++++++++--------------
 1 file changed, 10 insertions(+), 14 deletions(-)

diff --git a/drivers/net/i40e/i40e_rxtx_vec.c b/drivers/net/i40e/i40e_rxtx_vec.c
index 1e2fadd..9f67f9d 100644
--- a/drivers/net/i40e/i40e_rxtx_vec.c
+++ b/drivers/net/i40e/i40e_rxtx_vec.c
@@ -192,11 +192,7 @@ desc_to_olflags_v(__m128i descs[4], struct rte_mbuf **rx_pkts)
 static inline void
 desc_pktlen_align(__m128i descs[4])
 {
-	__m128i pktlen0, pktlen1, zero;
-	union {
-		uint16_t e[4];
-		uint64_t dword;
-	} vol;
+	__m128i pktlen0, pktlen1;
 
 	/* mask everything except pktlen field*/
 	const __m128i pktlen_msk = _mm_set_epi32(PKTLEN_MASK, PKTLEN_MASK,
@@ -206,18 +202,18 @@ desc_pktlen_align(__m128i descs[4])
 	pktlen1 = _mm_unpackhi_epi32(descs[1], descs[3]);
 	pktlen0 = _mm_unpackhi_epi32(pktlen0, pktlen1);
 
-	zero = _mm_xor_si128(pktlen0, pktlen0);
-
 	pktlen0 = _mm_srli_epi32(pktlen0, PKTLEN_SHIFT);
 	pktlen0 = _mm_and_si128(pktlen0, pktlen_msk);
 
-	pktlen0 = _mm_packs_epi32(pktlen0, zero);
-	vol.dword = _mm_cvtsi128_si64(pktlen0);
-	/* let the descriptor byte 15-14 store the pkt len */
-	*((uint16_t *)&descs[0]+7) = vol.e[0];
-	*((uint16_t *)&descs[1]+7) = vol.e[1];
-	*((uint16_t *)&descs[2]+7) = vol.e[2];
-	*((uint16_t *)&descs[3]+7) = vol.e[3];
+	pktlen0 = _mm_packs_epi32(pktlen0, pktlen0);
+
+	descs[3] = _mm_blend_epi16(descs[3], pktlen0, 0x80);
+	pktlen0 = _mm_slli_epi64(pktlen0, 16);
+	descs[2] = _mm_blend_epi16(descs[2], pktlen0, 0x80);
+	pktlen0 = _mm_slli_epi64(pktlen0, 16);
+	descs[1] = _mm_blend_epi16(descs[1], pktlen0, 0x80);
+	pktlen0 = _mm_slli_epi64(pktlen0, 16);
+	descs[0] = _mm_blend_epi16(descs[0], pktlen0, 0x80);
 }
 
  /*
-- 
2.5.5

[dpdk-dev] [PATCH v2 3/3] i40e: simplify SSE packet length extraction code

[Bruce Richardson]

In Table 8-16 of the "Intel® Ethernet Controller XL710 Datasheet" it is
stated that when the whole packet is written to a single buffer, the
header length field in the descriptor will be 0. This means that when
extracting the packet/data_len field from the descriptor in the driver
we do not need to mask out the extra header-length bits.

Inside the vector driver, this reduces the need to pull all four pktlen
fields into a single register to work on. Instead of a shift and mask,
we now need to only do a shift. Therefore, we can work on each descriptor
independently, processing each using one shift intrinsic and a blend.

This change makes the code shorter and easier to read, so we can pull it
into the main descriptor processing loop instead of needing its own
function. This in turn makes the descriptor processing in the loop as a
whole slightly easier to read as it's more linear.

In terms of performance, in testing this change shows little effect, with
single-core perf tests showing a very slight improvement.

Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 drivers/net/i40e/i40e_rxtx_vec.c | 51 ++++++++++++++--------------------------
 1 file changed, 17 insertions(+), 34 deletions(-)

diff --git a/drivers/net/i40e/i40e_rxtx_vec.c b/drivers/net/i40e/i40e_rxtx_vec.c
index 9f67f9d..f7a62a8 100644
--- a/drivers/net/i40e/i40e_rxtx_vec.c
+++ b/drivers/net/i40e/i40e_rxtx_vec.c
@@ -184,37 +184,7 @@ desc_to_olflags_v(__m128i descs[4], struct rte_mbuf **rx_pkts)
 #define desc_to_olflags_v(desc, rx_pkts) do {} while (0)
 #endif
 
-#define PKTLEN_SHIFT     (6)
-#define PKTLEN_MASK      (0x3FFF)
-/* Handling the pkt len field is not aligned with 1byte, so shift is
- * needed to let it align
- */
-static inline void
-desc_pktlen_align(__m128i descs[4])
-{
-	__m128i pktlen0, pktlen1;
-
-	/* mask everything except pktlen field*/
-	const __m128i pktlen_msk = _mm_set_epi32(PKTLEN_MASK, PKTLEN_MASK,
-						PKTLEN_MASK, PKTLEN_MASK);
-
-	pktlen0 = _mm_unpackhi_epi32(descs[0], descs[2]);
-	pktlen1 = _mm_unpackhi_epi32(descs[1], descs[3]);
-	pktlen0 = _mm_unpackhi_epi32(pktlen0, pktlen1);
-
-	pktlen0 = _mm_srli_epi32(pktlen0, PKTLEN_SHIFT);
-	pktlen0 = _mm_and_si128(pktlen0, pktlen_msk);
-
-	pktlen0 = _mm_packs_epi32(pktlen0, pktlen0);
-
-	descs[3] = _mm_blend_epi16(descs[3], pktlen0, 0x80);
-	pktlen0 = _mm_slli_epi64(pktlen0, 16);
-	descs[2] = _mm_blend_epi16(descs[2], pktlen0, 0x80);
-	pktlen0 = _mm_slli_epi64(pktlen0, 16);
-	descs[1] = _mm_blend_epi16(descs[1], pktlen0, 0x80);
-	pktlen0 = _mm_slli_epi64(pktlen0, 16);
-	descs[0] = _mm_blend_epi16(descs[0], pktlen0, 0x80);
-}
+#define PKTLEN_SHIFT     10
 
  /*
  * Notice:
@@ -333,12 +303,17 @@ _recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 			rte_prefetch0(&rx_pkts[pos + 3]->cacheline1);
 		}
 
-		/*shift the pktlen field*/
-		desc_pktlen_align(descs);
-
 		/* avoid compiler reorder optimization */
 		rte_compiler_barrier();
 
+		/* pkt 3,4 shift the pktlen field to be 16-bit aligned*/
+		const __m128i len3 = _mm_slli_epi32(descs[3], PKTLEN_SHIFT);
+		const __m128i len2 = _mm_slli_epi32(descs[2], PKTLEN_SHIFT);
+
+		/* merge the now-aligned packet length fields back in */
+		descs[3] = _mm_blend_epi16(descs[3], len3, 0x80);
+		descs[2] = _mm_blend_epi16(descs[2], len2, 0x80);
+
 		/* D.1 pkt 3,4 convert format from desc to pktmbuf */
 		pkt_mb4 = _mm_shuffle_epi8(descs[3], shuf_msk);
 		pkt_mb3 = _mm_shuffle_epi8(descs[2], shuf_msk);
@@ -354,6 +329,14 @@ _recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 		pkt_mb4 = _mm_add_epi16(pkt_mb4, crc_adjust);
 		pkt_mb3 = _mm_add_epi16(pkt_mb3, crc_adjust);
 
+		/* pkt 1,2 shift the pktlen field to be 16-bit aligned*/
+		const __m128i len1 = _mm_slli_epi32(descs[1], PKTLEN_SHIFT);
+		const __m128i len0 = _mm_slli_epi32(descs[0], PKTLEN_SHIFT);
+
+		/* merge the now-aligned packet length fields back in */
+		descs[1] = _mm_blend_epi16(descs[1], len1, 0x80);
+		descs[0] = _mm_blend_epi16(descs[0], len0, 0x80);
+
 		/* D.1 pkt 1,2 convert format from desc to pktmbuf */
 		pkt_mb2 = _mm_shuffle_epi8(descs[1], shuf_msk);
 		pkt_mb1 = _mm_shuffle_epi8(descs[0], shuf_msk);
-- 
2.5.5

Re: [dpdk-dev] [PATCH v2 0/3] improve i40e vpmd

[Zhe Tao]

On Thu, Apr 14, 2016 at 05:02:34PM +0100, Bruce Richardson wrote:
> This patchset improves the performance of the i40e SSE pmd by removing
> operations that triggered CPU stalls. It also shortens the code and
> cleans it up a little.
> 
> The base requirement for using the SSE code path has been pushed up to
> SSE4.1 from SSE3, due to the use of the blend instruction. The instruction
> set level is now checked at runtime as part of the driver selection process
> 
> Bruce Richardson (3):
>   i40e: require SSE4.1 support for vector driver
>   i40e: improve performance of vector PMD
>   i40e: simplify SSE packet length extraction code
> 
>  drivers/net/i40e/Makefile        |  6 ++++
>  drivers/net/i40e/i40e_rxtx_vec.c | 59 ++++++++++++++--------------------------
>  2 files changed, 27 insertions(+), 38 deletions(-)
> 
> -- 
> 2.5.5
Acked-by: Zhe Tao <zhe.tao@intel.com> 

Re: [dpdk-dev] [PATCH v2 0/3] improve i40e vpmd

[Bruce Richardson]

On Sun, Apr 17, 2016 at 04:32:10PM +0800, Zhe Tao wrote:
> On Thu, Apr 14, 2016 at 05:02:34PM +0100, Bruce Richardson wrote:
> > This patchset improves the performance of the i40e SSE pmd by removing
> > operations that triggered CPU stalls. It also shortens the code and
> > cleans it up a little.
> > 
> > The base requirement for using the SSE code path has been pushed up to
> > SSE4.1 from SSE3, due to the use of the blend instruction. The instruction
> > set level is now checked at runtime as part of the driver selection process
> > 
> > Bruce Richardson (3):
> >   i40e: require SSE4.1 support for vector driver
> >   i40e: improve performance of vector PMD
> >   i40e: simplify SSE packet length extraction code
> > 
> >  drivers/net/i40e/Makefile        |  6 ++++
> >  drivers/net/i40e/i40e_rxtx_vec.c | 59 ++++++++++++++--------------------------
> >  2 files changed, 27 insertions(+), 38 deletions(-)
> > 
> > -- 
> > 2.5.5
> Acked-by: Zhe Tao <zhe.tao@intel.com> 

Applied to dpdk-next-net/rel_16_07

/Bruce