[v1] app/testpmd: optimized MAC swap by using neon intrinsics

Message ID 1552292065-15239-1-git-send-email-ruifeng.wang@arm.com
State Superseded, archived
Delegated to: Ferruh Yigit
Headers show
Series
  • [v1] app/testpmd: optimized MAC swap by using neon intrinsics
Related show

Checks

Context Check Description
ci/Intel-compilation success Compilation OK
ci/intel-Performance-Testing success Performance Testing PASS
ci/mellanox-Performance-Testing success Performance Testing PASS
ci/checkpatch success coding style OK

Commit Message

Ruifeng Wang (Arm Technology China) March 11, 2019, 8:14 a.m.
Improved MAC swap performance for ARM platform.
The improvement was achieved by using neon intrinsics
to save CPU cycles and doing swap for four packets
at a time.
The optimization had 15% - 20% throughput boost
in testpmd MAC swap mode.

Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
Reviewed-by: Gavin Hu <gavin.hu@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
---
 app/test-pmd/macswap.c      |  4 +-
 app/test-pmd/macswap_neon.h | 93 +++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 96 insertions(+), 1 deletion(-)
 create mode 100644 app/test-pmd/macswap_neon.h

Comments

Jerin Jacob Kollanukkaran March 11, 2019, 2:16 p.m. | #1
On Mon, 2019-03-11 at 16:14 +0800, Ruifeng Wang wrote:
> -------------------------------------------------------------------
> ---
> Improved MAC swap performance for ARM platform.
> The improvement was achieved by using neon intrinsics
> to save CPU cycles and doing swap for four packets
> at a time.
> The optimization had 15% - 20% throughput boost
> in testpmd MAC swap mode.
> 
> Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
> Reviewed-by: Gavin Hu <gavin.hu@arm.com>
> Reviewed-by: Phil Yang <phil.yang@arm.com>
> ---
>  app/test-pmd/macswap.c      |  4 +-
>  app/test-pmd/macswap_neon.h | 93
> +++++++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 96 insertions(+), 1 deletion(-)
>  create mode 100644 app/test-pmd/macswap_neon.h
> 
> diff --git a/app/test-pmd/macswap.c b/app/test-pmd/macswap.c
> 
> +static inline void
> +do_macswap(struct rte_mbuf *pkts[], uint16_t nb,
> +		struct rte_port *txp)
> +{
> +	struct ether_hdr *eth_hdr[4];
> +	struct rte_mbuf *mb[4];
> +	uint64_t ol_flags;
> +	int i;
> +	int r;
> +	uint8x16_t v0, v1, v2, v3;
> +	/**
> +	 * Index map be used to shuffle the 16 bytes.
> +	 * byte 0-5 will be swapped with byte 6-11.
> +	 * byte 12-15 will keep unchanged.
> +	 */
> +	uint8x16_t idx_map = {6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5,
> +				12, 13, 14, 15};

Nit: I think, we can make it as "const uint8x16_t idx_map".

Other than that it looks good to me.
Regarding the performance, I have tested with two SoCs.

octeontx: +13% improvement
octeontx2: +46% improvement


Acked-by: Jerin Jacob <jerinj@marvell.com>
Honnappa Nagarahalli March 11, 2019, 3:33 p.m. | #2
> Improved MAC swap performance for ARM platform.
> The improvement was achieved by using neon intrinsics to save CPU cycles and
> doing swap for four packets at a time.
> The optimization had 15% - 20% throughput boost in testpmd MAC swap mode.
> 
> Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
> Reviewed-by: Gavin Hu <gavin.hu@arm.com>
> Reviewed-by: Phil Yang <phil.yang@arm.com>
> ---
>  app/test-pmd/macswap.c      |  4 +-
>  app/test-pmd/macswap_neon.h | 93
> +++++++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 96 insertions(+), 1 deletion(-)  create mode 100644 app/test-
> pmd/macswap_neon.h
> 
> diff --git a/app/test-pmd/macswap.c b/app/test-pmd/macswap.c index
> cbb41b7..71af916 100644
> --- a/app/test-pmd/macswap.c
> +++ b/app/test-pmd/macswap.c
> @@ -66,8 +66,10 @@
>  #include <rte_flow.h>
> 
>  #include "testpmd.h"
> -#ifdef RTE_ARCH_X86
> +#if defined(RTE_ARCH_X86)
>  #include "macswap_sse.h"
> +#elif defined(RTE_MACHINE_CPUFLAG_NEON) #include "macswap_neon.h"
>  #else
>  #include "macswap.h"
>  #endif
> diff --git a/app/test-pmd/macswap_neon.h b/app/test-pmd/macswap_neon.h
> new file mode 100644 index 0000000..bad1b9b
> --- /dev/null
> +++ b/app/test-pmd/macswap_neon.h
> @@ -0,0 +1,93 @@
> +/* SPDX-License-Identifier: BSD-3-Clause
> + * Copyright(c) 2019 Arm Limited
This change is derived from Intel's work in macswap_sse.h. Can you please add a header similar to lib/librte_lpm/rte_lpm_neon.h?

> + */
> +
> +#ifndef _MACSWAP_NEON_H_
> +#define _MACSWAP_NEON_H_
> +

<snip>
Ruifeng Wang (Arm Technology China) March 12, 2019, 1:34 a.m. | #3
Regards,
/Ruifeng

> -----Original Message-----
> From: Jerin Jacob Kollanukkaran <jerinj@marvell.com>
> Sent: 2019年3月11日 22:17
> To: Ruifeng Wang (Arm Technology China) <Ruifeng.Wang@arm.com>;
> jingjing.wu@intel.com; bernard.iremonger@intel.com;
> wenzhuo.lu@intel.com
> Cc: Honnappa Nagarahalli <Honnappa.Nagarahalli@arm.com>; nd
> <nd@arm.com>; hemant.agrawal@nxp.com; dev@dpdk.org
> Subject: Re: [PATCH v1] app/testpmd: optimized MAC swap by using neon
> intrinsics
> 
> On Mon, 2019-03-11 at 16:14 +0800, Ruifeng Wang wrote:
> > -------------------------------------------------------------------
> > ---
> > Improved MAC swap performance for ARM platform.
> > The improvement was achieved by using neon intrinsics to save CPU
> > cycles and doing swap for four packets at a time.
> > The optimization had 15% - 20% throughput boost in testpmd MAC swap
> > mode.
> >
> > Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
> > Reviewed-by: Gavin Hu <gavin.hu@arm.com>
> > Reviewed-by: Phil Yang <phil.yang@arm.com>
> > ---
> >  app/test-pmd/macswap.c      |  4 +-
> >  app/test-pmd/macswap_neon.h | 93
> > +++++++++++++++++++++++++++++++++++++++++++++
> >  2 files changed, 96 insertions(+), 1 deletion(-)  create mode 100644
> > app/test-pmd/macswap_neon.h
> >
> > diff --git a/app/test-pmd/macswap.c b/app/test-pmd/macswap.c
> >
> > +static inline void
> > +do_macswap(struct rte_mbuf *pkts[], uint16_t nb,
> > +		struct rte_port *txp)
> > +{
> > +	struct ether_hdr *eth_hdr[4];
> > +	struct rte_mbuf *mb[4];
> > +	uint64_t ol_flags;
> > +	int i;
> > +	int r;
> > +	uint8x16_t v0, v1, v2, v3;
> > +	/**
> > +	 * Index map be used to shuffle the 16 bytes.
> > +	 * byte 0-5 will be swapped with byte 6-11.
> > +	 * byte 12-15 will keep unchanged.
> > +	 */
> > +	uint8x16_t idx_map = {6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5,
> > +				12, 13, 14, 15};
> 
> Nit: I think, we can make it as "const uint8x16_t idx_map".
> 
> Other than that it looks good to me.
> Regarding the performance, I have tested with two SoCs.
> 
> octeontx: +13% improvement
> octeontx2: +46% improvement
> 
> 
> Acked-by: Jerin Jacob <jerinj@marvell.com>
> 
Thanks Jerin for your test and data.
The code change will be included in v2.
Ruifeng Wang (Arm Technology China) March 12, 2019, 1:37 a.m. | #4
Regards,
/Ruifeng

> -----Original Message-----
> From: Honnappa Nagarahalli <Honnappa.Nagarahalli@arm.com>
> Sent: 2019年3月11日 23:33
> To: Ruifeng Wang (Arm Technology China) <Ruifeng.Wang@arm.com>;
> wenzhuo.lu@intel.com; jingjing.wu@intel.com;
> bernard.iremonger@intel.com
> Cc: dev@dpdk.org; jerinj@marvell.com; hemant.agrawal@nxp.com; nd
> <nd@arm.com>; Ruifeng Wang (Arm Technology China)
> <Ruifeng.Wang@arm.com>; Honnappa Nagarahalli
> <Honnappa.Nagarahalli@arm.com>; nd <nd@arm.com>
> Subject: RE: [PATCH v1] app/testpmd: optimized MAC swap by using neon
> intrinsics
> 
> > Improved MAC swap performance for ARM platform.
> > The improvement was achieved by using neon intrinsics to save CPU
> > cycles and doing swap for four packets at a time.
> > The optimization had 15% - 20% throughput boost in testpmd MAC swap
> mode.
> >
> > Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
> > Reviewed-by: Gavin Hu <gavin.hu@arm.com>
> > Reviewed-by: Phil Yang <phil.yang@arm.com>
> > ---
> >  app/test-pmd/macswap.c      |  4 +-
> >  app/test-pmd/macswap_neon.h | 93
> > +++++++++++++++++++++++++++++++++++++++++++++
> >  2 files changed, 96 insertions(+), 1 deletion(-)  create mode 100644
> > app/test- pmd/macswap_neon.h
> >
> > diff --git a/app/test-pmd/macswap.c b/app/test-pmd/macswap.c index
> > cbb41b7..71af916 100644
> > --- a/app/test-pmd/macswap.c
> > +++ b/app/test-pmd/macswap.c
> > @@ -66,8 +66,10 @@
> >  #include <rte_flow.h>
> >
> >  #include "testpmd.h"
> > -#ifdef RTE_ARCH_X86
> > +#if defined(RTE_ARCH_X86)
> >  #include "macswap_sse.h"
> > +#elif defined(RTE_MACHINE_CPUFLAG_NEON) #include
> "macswap_neon.h"
> >  #else
> >  #include "macswap.h"
> >  #endif
> > diff --git a/app/test-pmd/macswap_neon.h b/app/test-
> pmd/macswap_neon.h
> > new file mode 100644 index 0000000..bad1b9b
> > --- /dev/null
> > +++ b/app/test-pmd/macswap_neon.h
> > @@ -0,0 +1,93 @@
> > +/* SPDX-License-Identifier: BSD-3-Clause
> > + * Copyright(c) 2019 Arm Limited
> This change is derived from Intel's work in macswap_sse.h. Can you please
> add a header similar to lib/librte_lpm/rte_lpm_neon.h?
> 
Sure. Will update file header in v2.
Thanks for your suggestion.

> > + */
> > +
> > +#ifndef _MACSWAP_NEON_H_
> > +#define _MACSWAP_NEON_H_
> > +
> 
> <snip>

Patch

diff --git a/app/test-pmd/macswap.c b/app/test-pmd/macswap.c
index cbb41b7..71af916 100644
--- a/app/test-pmd/macswap.c
+++ b/app/test-pmd/macswap.c
@@ -66,8 +66,10 @@ 
 #include <rte_flow.h>
 
 #include "testpmd.h"
-#ifdef RTE_ARCH_X86
+#if defined(RTE_ARCH_X86)
 #include "macswap_sse.h"
+#elif defined(RTE_MACHINE_CPUFLAG_NEON)
+#include "macswap_neon.h"
 #else
 #include "macswap.h"
 #endif
diff --git a/app/test-pmd/macswap_neon.h b/app/test-pmd/macswap_neon.h
new file mode 100644
index 0000000..bad1b9b
--- /dev/null
+++ b/app/test-pmd/macswap_neon.h
@@ -0,0 +1,93 @@ 
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2019 Arm Limited
+ */
+
+#ifndef _MACSWAP_NEON_H_
+#define _MACSWAP_NEON_H_
+
+#include "macswap_common.h"
+#include "rte_vect.h"
+
+static inline void
+do_macswap(struct rte_mbuf *pkts[], uint16_t nb,
+		struct rte_port *txp)
+{
+	struct ether_hdr *eth_hdr[4];
+	struct rte_mbuf *mb[4];
+	uint64_t ol_flags;
+	int i;
+	int r;
+	uint8x16_t v0, v1, v2, v3;
+	/**
+	 * Index map be used to shuffle the 16 bytes.
+	 * byte 0-5 will be swapped with byte 6-11.
+	 * byte 12-15 will keep unchanged.
+	 */
+	uint8x16_t idx_map = {6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5,
+				12, 13, 14, 15};
+
+	ol_flags = ol_flags_init(txp->dev_conf.txmode.offloads);
+	vlan_qinq_set(pkts, nb, ol_flags,
+			txp->tx_vlan_id, txp->tx_vlan_id_outer);
+
+	i = 0;
+	r = nb;
+
+	while (r >= 4) {
+		if (r >= 8) {
+			rte_prefetch0(rte_pktmbuf_mtod(pkts[i + 4], void *));
+			rte_prefetch0(rte_pktmbuf_mtod(pkts[i + 5], void *));
+			rte_prefetch0(rte_pktmbuf_mtod(pkts[i + 6], void *));
+			rte_prefetch0(rte_pktmbuf_mtod(pkts[i + 7], void *));
+		}
+
+		mb[0] = pkts[i++];
+		eth_hdr[0] = rte_pktmbuf_mtod(mb[0], struct ether_hdr *);
+
+		mb[1] = pkts[i++];
+		eth_hdr[1] = rte_pktmbuf_mtod(mb[1], struct ether_hdr *);
+
+		mb[2] = pkts[i++];
+		eth_hdr[2] = rte_pktmbuf_mtod(mb[2], struct ether_hdr *);
+
+		mb[3] = pkts[i++];
+		eth_hdr[3] = rte_pktmbuf_mtod(mb[3], struct ether_hdr *);
+
+		v0 = vld1q_u8((uint8_t const *)eth_hdr[0]);
+		v1 = vld1q_u8((uint8_t const *)eth_hdr[1]);
+		v2 = vld1q_u8((uint8_t const *)eth_hdr[2]);
+		v3 = vld1q_u8((uint8_t const *)eth_hdr[3]);
+
+		v0 = vqtbl1q_u8(v0, idx_map);
+		v1 = vqtbl1q_u8(v1, idx_map);
+		v2 = vqtbl1q_u8(v2, idx_map);
+		v3 = vqtbl1q_u8(v3, idx_map);
+
+		vst1q_u8((uint8_t *)eth_hdr[0], v0);
+		vst1q_u8((uint8_t *)eth_hdr[1], v1);
+		vst1q_u8((uint8_t *)eth_hdr[2], v2);
+		vst1q_u8((uint8_t *)eth_hdr[3], v3);
+
+		mbuf_field_set(mb[0], ol_flags);
+		mbuf_field_set(mb[1], ol_flags);
+		mbuf_field_set(mb[2], ol_flags);
+		mbuf_field_set(mb[3], ol_flags);
+		r -= 4;
+	}
+
+	for ( ; i < nb; i++) {
+		if (i < nb - 1)
+			rte_prefetch0(rte_pktmbuf_mtod(pkts[i+1], void *));
+		mb[0] = pkts[i];
+		eth_hdr[0] = rte_pktmbuf_mtod(mb[0], struct ether_hdr *);
+
+		/* Swap dest and src mac addresses. */
+		v0 = vld1q_u8((uint8_t const *)eth_hdr[0]);
+		v0 = vqtbl1q_u8(v0, idx_map);
+		vst1q_u8((uint8_t *)eth_hdr[0], v0);
+
+		mbuf_field_set(mb[0], ol_flags);
+	}
+}
+
+#endif /* _MACSWAP_NEON_H_ */