DPDK patches and discussions
 help / color / mirror / Atom feed
* [dpdk-dev] [RFC PATCH 0/1] net/mlx5: add vectorized Rx/Tx burst for ARM
@ 2017-08-25 18:40 Yongseok Koh
  2017-08-25 18:40 ` [dpdk-dev] [RFC PATCH 1/1] " Yongseok Koh
  0 siblings, 1 reply; 4+ messages in thread
From: Yongseok Koh @ 2017-08-25 18:40 UTC (permalink / raw)
  To: adrien.mazarguil, nelio.laranjeiro; +Cc: dev, Yongseok Koh

The SSE(x86) Rx/Tx burst functions added in v17.08 would be ported for ARM NEON
in v17.11. Although this is still ongoing effort (more implementation and
further optimization), this intrim patch can be applied on top of v17.08 and
forward packts.

One of topics to discuss is that I used inilne assembly for performance critical
code blocks because I don't think intrinsics for NEON aren't well optimized yet,
especially vqtbl2q_u8()/vqtbl3q_u8()/vqtbl4q_u8() and gcc's register
optimization. And older gcc doesn't even have vld1q_u8_x4(). I used it to get
rid of hotspots shown in profiling result. I'm not sure whether inline assembly
is allowed in DPDK community. But, I believe there's no reason to prohibit it.

In my patch, some of functions are commented out as I'm not done migrating those
yet. But this is functional (Rx/Tx). For Tx, "--txqflags=0xf01" is needed
because I haven't ported txq_scatter_v() yet.

Yongseok Koh (1):
  net/mlx5: add vectorized Rx/Tx burst for ARM

 drivers/net/mlx5/Makefile             |    2 +
 drivers/net/mlx5/mlx5_ethdev.c        |    4 +-
 drivers/net/mlx5/mlx5_prm.h           |   15 +
 drivers/net/mlx5/mlx5_rxq.c           |   61 ++
 drivers/net/mlx5/mlx5_rxtx.h          |    3 +-
 drivers/net/mlx5/mlx5_rxtx_vec_neon.c | 1464 +++++++++++++++++++++++++++++++++
 6 files changed, 1546 insertions(+), 3 deletions(-)
 create mode 100644 drivers/net/mlx5/mlx5_rxtx_vec_neon.c

-- 
2.11.0

^ permalink raw reply	[flat|nested] 4+ messages in thread

* [dpdk-dev] [RFC PATCH 1/1] net/mlx5: add vectorized Rx/Tx burst for ARM
  2017-08-25 18:40 [dpdk-dev] [RFC PATCH 0/1] net/mlx5: add vectorized Rx/Tx burst for ARM Yongseok Koh
@ 2017-08-25 18:40 ` Yongseok Koh
  2017-09-04 12:37   ` Nélio Laranjeiro
  0 siblings, 1 reply; 4+ messages in thread
From: Yongseok Koh @ 2017-08-25 18:40 UTC (permalink / raw)
  To: adrien.mazarguil, nelio.laranjeiro; +Cc: dev, Yongseok Koh

New Rx/Tx burst functions are added using NEON vector instructions for ARM
CPU.

Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
---
 drivers/net/mlx5/Makefile             |    2 +
 drivers/net/mlx5/mlx5_ethdev.c        |    4 +-
 drivers/net/mlx5/mlx5_prm.h           |   15 +
 drivers/net/mlx5/mlx5_rxq.c           |   61 ++
 drivers/net/mlx5/mlx5_rxtx.h          |    3 +-
 drivers/net/mlx5/mlx5_rxtx_vec_neon.c | 1464 +++++++++++++++++++++++++++++++++
 6 files changed, 1546 insertions(+), 3 deletions(-)
 create mode 100644 drivers/net/mlx5/mlx5_rxtx_vec_neon.c

diff --git a/drivers/net/mlx5/Makefile b/drivers/net/mlx5/Makefile
index 8736de5d3..616d769f8 100644
--- a/drivers/net/mlx5/Makefile
+++ b/drivers/net/mlx5/Makefile
@@ -41,6 +41,8 @@ SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_txq.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxtx.c
 ifeq ($(CONFIG_RTE_ARCH_X86_64),y)
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxtx_vec_sse.c
+else ifeq ($(CONFIG_RTE_ARCH_ARM64),y)
+SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxtx_vec_neon.c
 endif
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_trigger.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_ethdev.c
diff --git a/drivers/net/mlx5/mlx5_ethdev.c b/drivers/net/mlx5/mlx5_ethdev.c
index b0eb3cdfc..b387c6fb3 100644
--- a/drivers/net/mlx5/mlx5_ethdev.c
+++ b/drivers/net/mlx5/mlx5_ethdev.c
@@ -1516,7 +1516,7 @@ priv_select_tx_function(struct priv *priv)
 				priv->dev->tx_pkt_burst = mlx5_tx_burst_raw_vec;
 			else
 				priv->dev->tx_pkt_burst = mlx5_tx_burst_vec;
-			DEBUG("selected Enhanced MPW TX vectorized function");
+			WARN("selected Enhanced MPW TX vectorized function");
 		} else {
 			priv->dev->tx_pkt_burst = mlx5_tx_burst_empw;
 			DEBUG("selected Enhanced MPW TX function");
@@ -1542,7 +1542,7 @@ priv_select_rx_function(struct priv *priv)
 	if (priv_check_vec_rx_support(priv) > 0) {
 		priv_prep_vec_rx_function(priv);
 		priv->dev->rx_pkt_burst = mlx5_rx_burst_vec;
-		DEBUG("selected RX vectorized function");
+		WARN("selected RX vectorized function");
 	} else {
 		priv->dev->rx_pkt_burst = mlx5_rx_burst;
 	}
diff --git a/drivers/net/mlx5/mlx5_prm.h b/drivers/net/mlx5/mlx5_prm.h
index 608072f7e..01e95b466 100644
--- a/drivers/net/mlx5/mlx5_prm.h
+++ b/drivers/net/mlx5/mlx5_prm.h
@@ -224,6 +224,20 @@ struct mlx5_mpw {
 };
 
 /* CQ element structure - should be equal to the cache line size */
+#if 0
+struct mlx5_cqe { // 16B
+       uint16_t hdr_type_etc;
+       uint8_t pkt_info;
+       uint8_t sop_drop_qpn; /* flow_tag */
+       uint16_t byte_cnt;
+       uint16_t vlan_info;
+       uint32_t rx_hash_res;
+       uint8_t timestamp;
+       uint8_t wqe_counter;
+       uint8_t rsvd4;
+       uint8_t op_own;
+};
+#else
 struct mlx5_cqe {
 #if (RTE_CACHE_LINE_SIZE == 128)
 	uint8_t padding[64];
@@ -243,6 +257,7 @@ struct mlx5_cqe {
 	uint8_t rsvd4;
 	uint8_t op_own;
 };
+#endif
 
 /**
  * Convert a user mark to flow mark.
diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
index 74387a797..30654abd3 100644
--- a/drivers/net/mlx5/mlx5_rxq.c
+++ b/drivers/net/mlx5/mlx5_rxq.c
@@ -786,6 +786,61 @@ rxq_cleanup(struct rxq_ctrl *rxq_ctrl)
 	memset(rxq_ctrl, 0, sizeof(*rxq_ctrl));
 }
 
+#ifdef SW_EMULATION
+#define MLX5_EMUL_MCQE_PER_COMP 64
+int rxq_cqe_comp_en = 1;
+/**
+ *
+ * Filling in CQEs to emulate packet arrival
+ *
+ * @param tmpl
+ *   Pointer to RX queue control template.
+ */
+	static inline void
+emulate_rxq_cqe_setup(struct rxq_ctrl *tmpl)
+{
+	struct rxq *rxq = &tmpl->rxq;
+	const unsigned int cqe_n = 1 << rxq->cqe_n;
+	volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[0];
+	unsigned int i;
+
+	for (i = 0; i < cqe_n; i++)
+		cqe[i].op_own = MLX5_CQE_INVALIDATE;
+	if (rxq_cqe_comp_en) {
+		volatile struct mlx5_mini_cqe8 *mcqe;
+		unsigned int j;
+
+		for (i = 0; i < cqe_n;) {
+			if (!(i % MLX5_EMUL_MCQE_PER_COMP)) {
+				/* Fill in title CQE */
+				cqe[i].op_own = 0xc; /* Compressed */
+				/* number of mini CQEs */
+				cqe[i].byte_cnt = htonl(MLX5_EMUL_MCQE_PER_COMP);
+				/* Fill in mini CQEs */
+				mcqe = (volatile void *)&cqe[i + 1].pkt_info;
+				for (j = 0; j < MLX5_EMUL_MCQE_PER_COMP; j++) {
+					mcqe[j % 8].byte_cnt =
+						rxq->crc_present ?
+						htonl(64) : htonl(60);
+					if ((j % 8) == 7) {
+						i += 8;
+						mcqe = (volatile void *)
+						       &cqe[i].pkt_info;
+					}
+				}
+			} else {
+				ERROR("Error on building compressed CQEs");
+			}
+		}
+	} else {
+		for (i = 0; i < cqe_n; i++) {
+			cqe[i].op_own = 0;
+			cqe[i].byte_cnt = rxq->crc_present ? htonl(64) : htonl(60);
+		}
+	}
+}
+#endif /* SW_EMULATION */
+
 /**
  * Initialize RX queue.
  *
@@ -1064,6 +1119,10 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
 		      (void *)dev, strerror(ret));
 		goto error;
 	}
+#ifdef SW_EMULATION
+	rxq_cqe_comp_en = priv->cqe_comp;
+	emulate_rxq_cqe_setup(&tmpl);
+#endif
 	/* Reuse buffers from original queue if possible. */
 	if (rxq_ctrl->rxq.elts_n) {
 		assert(1 << rxq_ctrl->rxq.elts_n == desc);
@@ -1092,7 +1151,9 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
 	/* Update doorbell counter. */
 	rxq_ctrl->rxq.rq_ci = desc >> rxq_ctrl->rxq.sges_n;
 	rte_wmb();
+#ifndef SW_EMULATION
 	*rxq_ctrl->rxq.rq_db = htonl(rxq_ctrl->rxq.rq_ci);
+#endif
 	DEBUG("%p: rxq updated with %p", (void *)rxq_ctrl, (void *)&tmpl);
 	assert(ret == 0);
 	return 0;
diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
index 7de1d1086..6aae00b77 100644
--- a/drivers/net/mlx5/mlx5_rxtx.h
+++ b/drivers/net/mlx5/mlx5_rxtx.h
@@ -602,11 +602,12 @@ mlx5_tx_dbrec(struct txq *txq, volatile struct mlx5_wqe *wqe)
 	uint64_t *dst = (uint64_t *)((uintptr_t)txq->bf_reg);
 	volatile uint64_t *src = ((volatile uint64_t *)wqe);
 
-	rte_wmb();
+	rte_compiler_barrier();
 	*txq->qp_db = htonl(txq->wqe_ci);
 	/* Ensure ordering between DB record and BF copy. */
 	rte_wmb();
 	*dst = *src;
+	rte_wmb();
 }
 
 #endif /* RTE_PMD_MLX5_RXTX_H_ */
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_neon.c b/drivers/net/mlx5/mlx5_rxtx_vec_neon.c
new file mode 100644
index 000000000..e5bce23c2
--- /dev/null
+++ b/drivers/net/mlx5/mlx5_rxtx_vec_neon.c
@@ -0,0 +1,1464 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2017 6WIND S.A.
+ *   Copyright 2017 Mellanox.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <assert.h>
+#include <stdint.h>
+#include <string.h>
+#include <stdlib.h>
+#include <arm_neon.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-Wpedantic"
+#endif
+#include <infiniband/verbs.h>
+#include <infiniband/mlx5_hw.h>
+#include <infiniband/arch.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-Wpedantic"
+#endif
+
+#include <rte_mbuf.h>
+#include <rte_mempool.h>
+#include <rte_prefetch.h>
+
+#include "mlx5.h"
+#include "mlx5_utils.h"
+#include "mlx5_rxtx.h"
+#include "mlx5_autoconf.h"
+#include "mlx5_defs.h"
+#include "mlx5_prm.h"
+
+#pragma GCC diagnostic ignored "-Wcast-qual"
+
+/**
+ * Fill in buffer descriptors in a multi-packet send descriptor.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param dseg
+ *   Pointer to buffer descriptor to be writen.
+ * @param pkts
+ *   Pointer to array of packets to be sent.
+ * @param n
+ *   Number of packets to be filled.
+ */
+static inline void
+txq_wr_dseg_v(struct txq *txq, uint8_t *dseg,
+	      struct rte_mbuf **pkts, unsigned int n)
+{
+	unsigned int pos;
+	uintptr_t addr;
+	const uint8x16_t dseg_shuf_m = {
+		 3,  2,  1,  0, /* length, bswap32 */
+		 4,  5,  6,  7, /* lkey */
+		15, 14, 13, 12, /* addr, bswap64 */
+		11, 10,  9,  8
+	};
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	uint32_t tx_byte = 0;
+#endif
+
+	for (pos = 0; pos < n; ++pos, dseg += MLX5_WQE_DWORD_SIZE) {
+		uint8x16_t desc;
+		struct rte_mbuf *pkt = pkts[pos];
+
+		addr = rte_pktmbuf_mtod(pkt, uintptr_t);
+		desc = vreinterpretq_u8_u32((uint32x4_t) {
+				DATA_LEN(pkt),
+				mlx5_tx_mb2mr(txq, pkt),
+				addr,
+				addr >> 32 });
+		desc = vqtbl1q_u8(desc, dseg_shuf_m);
+		vst1q_u8(dseg, desc);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		tx_byte += DATA_LEN(pkt);
+#endif
+	}
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	txq->stats.obytes += tx_byte;
+#endif
+}
+
+#if 0
+/**
+ * Count the number of continuous single segment packets. The first packet must
+ * be a single segment packet.
+ *
+ * @param pkts
+ *   Pointer to array of packets.
+ * @param pkts_n
+ *   Number of packets.
+ *
+ * @return
+ *   Number of continuous single segment packets.
+ */
+static inline unsigned int
+txq_check_multiseg(struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	unsigned int pos;
+
+	if (!pkts_n)
+		return 0;
+	assert(NB_SEGS(pkts[0]) == 1);
+	/* Count the number of continuous single segment packets. */
+	for (pos = 1; pos < pkts_n; ++pos)
+		if (NB_SEGS(pkts[pos]) > 1)
+			break;
+	return pos;
+}
+
+/**
+ * Count the number of packets having same ol_flags and calculate cs_flags.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param pkts
+ *   Pointer to array of packets.
+ * @param pkts_n
+ *   Number of packets.
+ * @param cs_flags
+ *   Pointer of flags to be returned.
+ *
+ * @return
+ *   Number of packets having same ol_flags.
+ */
+static inline unsigned int
+txq_calc_offload(struct txq *txq, struct rte_mbuf **pkts, uint16_t pkts_n,
+		 uint8_t *cs_flags)
+{
+	unsigned int pos;
+	const uint64_t ol_mask =
+		PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM |
+		PKT_TX_UDP_CKSUM | PKT_TX_TUNNEL_GRE |
+		PKT_TX_TUNNEL_VXLAN | PKT_TX_OUTER_IP_CKSUM;
+
+	if (!pkts_n)
+		return 0;
+	/* Count the number of packets having same ol_flags. */
+	for (pos = 1; pos < pkts_n; ++pos)
+		if ((pkts[pos]->ol_flags ^ pkts[0]->ol_flags) & ol_mask)
+			break;
+	/* Should open another MPW session for the rest. */
+	if (pkts[0]->ol_flags &
+	    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
+		const uint64_t is_tunneled =
+			pkts[0]->ol_flags &
+			(PKT_TX_TUNNEL_GRE |
+			 PKT_TX_TUNNEL_VXLAN);
+
+		if (is_tunneled && txq->tunnel_en) {
+			*cs_flags = MLX5_ETH_WQE_L3_INNER_CSUM |
+				    MLX5_ETH_WQE_L4_INNER_CSUM;
+			if (pkts[0]->ol_flags & PKT_TX_OUTER_IP_CKSUM)
+				*cs_flags |= MLX5_ETH_WQE_L3_CSUM;
+		} else {
+			*cs_flags = MLX5_ETH_WQE_L3_CSUM |
+				    MLX5_ETH_WQE_L4_CSUM;
+		}
+	}
+	return pos;
+}
+
+/**
+ * Send multi-segmented packets until it encounters a single segment packet in
+ * the pkts list.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param pkts
+ *   Pointer to array of packets to be sent.
+ * @param pkts_n
+ *   Number of packets to be sent.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+static uint16_t
+txq_scatter_v(struct txq *txq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	uint16_t elts_head = txq->elts_head;
+	const uint16_t elts_n = 1 << txq->elts_n;
+	const uint16_t elts_m = elts_n - 1;
+	const uint16_t wq_n = 1 << txq->wqe_n;
+	const uint16_t wq_mask = wq_n - 1;
+	const unsigned int nb_dword_per_wqebb =
+		MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE;
+	const unsigned int nb_dword_in_hdr =
+		sizeof(struct mlx5_wqe) / MLX5_WQE_DWORD_SIZE;
+	unsigned int n;
+	volatile struct mlx5_wqe *wqe = NULL;
+
+	assert(elts_n > pkts_n);
+	mlx5_tx_complete(txq);
+	if (unlikely(!pkts_n))
+		return 0;
+	for (n = 0; n < pkts_n; ++n) {
+		struct rte_mbuf *buf = pkts[n];
+		unsigned int segs_n = buf->nb_segs;
+		unsigned int ds = nb_dword_in_hdr;
+		unsigned int len = PKT_LEN(buf);
+		uint16_t wqe_ci = txq->wqe_ci;
+		const __m128i shuf_mask_ctrl =
+			_mm_set_epi8(15, 14, 13, 12,
+				      8,  9, 10, 11, /* bswap32 */
+				      4,  5,  6,  7, /* bswap32 */
+				      0,  1,  2,  3  /* bswap32 */);
+		uint8_t cs_flags = 0;
+		uint16_t max_elts;
+		uint16_t max_wqe;
+		__m128i *t_wqe, *dseg;
+		__m128i ctrl;
+
+		assert(segs_n);
+		max_elts = elts_n - (elts_head - txq->elts_tail);
+		max_wqe = wq_n - (txq->wqe_ci - txq->wqe_pi);
+		/*
+		 * A MPW session consumes 2 WQEs at most to
+		 * include MLX5_MPW_DSEG_MAX pointers.
+		 */
+		if (segs_n == 1 ||
+		    max_elts < segs_n || max_wqe < 2)
+			break;
+		wqe = &((volatile struct mlx5_wqe64 *)
+			 txq->wqes)[wqe_ci & wq_mask].hdr;
+		if (buf->ol_flags &
+		     (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
+			const uint64_t is_tunneled = buf->ol_flags &
+						      (PKT_TX_TUNNEL_GRE |
+						       PKT_TX_TUNNEL_VXLAN);
+
+			if (is_tunneled && txq->tunnel_en) {
+				cs_flags = MLX5_ETH_WQE_L3_INNER_CSUM |
+					   MLX5_ETH_WQE_L4_INNER_CSUM;
+				if (buf->ol_flags & PKT_TX_OUTER_IP_CKSUM)
+					cs_flags |= MLX5_ETH_WQE_L3_CSUM;
+			} else {
+				cs_flags = MLX5_ETH_WQE_L3_CSUM |
+					   MLX5_ETH_WQE_L4_CSUM;
+			}
+		}
+		/* Title WQEBB pointer. */
+		t_wqe = (__m128i *)wqe;
+		dseg = (__m128i *)(wqe + 1);
+		do {
+			if (!(ds++ % nb_dword_per_wqebb)) {
+				dseg = (__m128i *)
+					&((volatile struct mlx5_wqe64 *)
+					   txq->wqes)[++wqe_ci & wq_mask];
+			}
+			txq_wr_dseg_v(txq, dseg++, &buf, 1);
+			(*txq->elts)[elts_head++ & elts_m] = buf;
+			buf = buf->next;
+		} while (--segs_n);
+		++wqe_ci;
+		/* Fill CTRL in the header. */
+		ctrl = _mm_set_epi32(0, 0, txq->qp_num_8s | ds,
+				     MLX5_OPC_MOD_MPW << 24 |
+				     txq->wqe_ci << 8 | MLX5_OPCODE_TSO);
+		ctrl = _mm_shuffle_epi8(ctrl, shuf_mask_ctrl);
+		_mm_store_si128(t_wqe, ctrl);
+		/* Fill ESEG in the header. */
+		_mm_store_si128(t_wqe + 1,
+				_mm_set_epi16(0, 0, 0, 0,
+					      htons(len), cs_flags,
+					      0, 0));
+		txq->wqe_ci = wqe_ci;
+	}
+	if (!n)
+		return 0;
+	txq->elts_comp += (uint16_t)(elts_head - txq->elts_head);
+	txq->elts_head = elts_head;
+	if (txq->elts_comp >= MLX5_TX_COMP_THRESH) {
+		wqe->ctrl[2] = htonl(8);
+		wqe->ctrl[3] = txq->elts_head;
+		txq->elts_comp = 0;
+		++txq->cq_pi;
+	}
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	txq->stats.opackets += n;
+#endif
+	mlx5_tx_dbrec(txq, wqe);
+	return n;
+}
+#endif
+
+/**
+ * Send burst of packets with Enhanced MPW. If it encounters a multi-seg packet,
+ * it returns to make it processed by txq_scatter_v(). All the packets in
+ * the pkts list should be single segment packets having same offload flags.
+ * This must be checked by txq_check_multiseg() and txq_calc_offload().
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param pkts
+ *   Pointer to array of packets to be sent.
+ * @param pkts_n
+ *   Number of packets to be sent (<= MLX5_VPMD_TX_MAX_BURST).
+ * @param cs_flags
+ *   Checksum offload flags to be written in the descriptor.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+static inline uint16_t
+txq_burst_v(struct txq *txq, struct rte_mbuf **pkts, uint16_t pkts_n,
+	    uint8_t cs_flags)
+{
+	struct rte_mbuf **elts;
+	uint16_t elts_head = txq->elts_head;
+	const uint16_t elts_n = 1 << txq->elts_n;
+	const uint16_t elts_m = elts_n - 1;
+	const unsigned int nb_dword_per_wqebb =
+		MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE;
+	const unsigned int nb_dword_in_hdr =
+		sizeof(struct mlx5_wqe) / MLX5_WQE_DWORD_SIZE;
+	unsigned int n = 0;
+	unsigned int pos;
+	uint16_t max_elts;
+	uint16_t max_wqe;
+	uint32_t comp_req = 0;
+	const uint16_t wq_n = 1 << txq->wqe_n;
+	const uint16_t wq_mask = wq_n - 1;
+	uint16_t wq_idx = txq->wqe_ci & wq_mask;
+	volatile struct mlx5_wqe64 *wq =
+		&((volatile struct mlx5_wqe64 *)txq->wqes)[wq_idx];
+	volatile struct mlx5_wqe *wqe = (volatile struct mlx5_wqe *)wq;
+	const uint8x16_t ctrl_shuf_m = {
+		 3,  2,  1,  0, /* bswap32 */
+		 7,  6,  5,  4, /* bswap32 */
+		11, 10,  9,  8, /* bswap32 */
+		12, 13, 14, 15
+	};
+	uint8x16_t *t_wqe;
+	uint8_t *dseg;
+	uint8x16_t ctrl;
+
+	/* Make sure all packets can fit into a single WQE. */
+	assert(elts_n > pkts_n);
+	mlx5_tx_complete(txq);
+	max_elts = (elts_n - (elts_head - txq->elts_tail));
+	max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
+	pkts_n = RTE_MIN((unsigned int)RTE_MIN(pkts_n, max_wqe), max_elts);
+	if (unlikely(!pkts_n))
+		return 0;
+	elts = &(*txq->elts)[elts_head & elts_m];
+	/* Loop for available tailroom first. */
+	n = RTE_MIN(elts_n - (elts_head & elts_m), pkts_n);
+	for (pos = 0; pos < (n & -2); pos += 2)
+		vst1q_u64((void *)&elts[pos], vld1q_u64((void *)&pkts[pos]));
+	if (n & 1)
+		elts[pos] = pkts[pos];
+	/* Check if it crosses the end of the queue. */
+	if (unlikely(n < pkts_n)) {
+		elts = &(*txq->elts)[0];
+		for (pos = 0; pos < pkts_n - n; ++pos)
+			elts[pos] = pkts[n + pos];
+	}
+	txq->elts_head += pkts_n;
+	/* Save title WQEBB pointer. */
+	t_wqe = (uint8x16_t *)wqe;
+	dseg = (uint8_t *)(wqe + 1);
+	/* Calculate the number of entries to the end. */
+	n = RTE_MIN(
+		(wq_n - wq_idx) * nb_dword_per_wqebb - nb_dword_in_hdr,
+		pkts_n);
+	/* Fill DSEGs. */
+	txq_wr_dseg_v(txq, dseg, pkts, n);
+	/* Check if it crosses the end of the queue. */
+	if (n < pkts_n) {
+		dseg = (uint8_t *)txq->wqes;
+		txq_wr_dseg_v(txq, dseg, &pkts[n], pkts_n - n);
+	}
+	if (txq->elts_comp + pkts_n < MLX5_TX_COMP_THRESH) {
+		txq->elts_comp += pkts_n;
+	} else {
+		/* Request a completion. */
+		txq->elts_comp = 0;
+		++txq->cq_pi;
+		comp_req = 8;
+	}
+	/* Fill CTRL in the header. */
+	ctrl = vreinterpretq_u8_u32((uint32x4_t) {
+			MLX5_OPC_MOD_ENHANCED_MPSW << 24 |
+			txq->wqe_ci << 8 | MLX5_OPCODE_ENHANCED_MPSW,
+			txq->qp_num_8s | (pkts_n + 2),
+			comp_req,
+			txq->elts_head });
+	ctrl = vqtbl1q_u8(ctrl, ctrl_shuf_m);
+	vst1q_u8((void *)t_wqe, ctrl);
+	/* Fill ESEG in the header. */
+	vst1q_u8((void *)(t_wqe + 1),
+		 (uint8x16_t) { 0, 0, 0, 0,
+				cs_flags, 0, 0, 0,
+				0, 0, 0, 0,
+				0, 0, 0, 0 });
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	txq->stats.opackets += pkts_n;
+#endif
+	txq->wqe_ci += (nb_dword_in_hdr + pkts_n + (nb_dword_per_wqebb - 1)) /
+		       nb_dword_per_wqebb;
+	/* Ring QP doorbell. */
+	mlx5_tx_dbrec(txq, wqe);
+	return pkts_n;
+}
+
+/**
+ * DPDK callback for vectorized TX.
+ *
+ * @param dpdk_txq
+ *   Generic pointer to TX queue structure.
+ * @param[in] pkts
+ *   Packets to transmit.
+ * @param pkts_n
+ *   Number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+uint16_t
+mlx5_tx_burst_raw_vec(void *dpdk_txq, struct rte_mbuf **pkts,
+		      uint16_t pkts_n)
+{
+	struct txq *txq = (struct txq *)dpdk_txq;
+	uint16_t nb_tx = 0;
+
+	while (pkts_n > nb_tx) {
+		uint16_t n;
+		uint16_t ret;
+
+		n = RTE_MIN((uint16_t)(pkts_n - nb_tx), MLX5_VPMD_TX_MAX_BURST);
+		ret = txq_burst_v(txq, &pkts[nb_tx], n, 0);
+		nb_tx += ret;
+		if (!ret)
+			break;
+	}
+	return nb_tx;
+}
+
+#if 0
+/**
+ * DPDK callback for vectorized TX with multi-seg packets and offload.
+ *
+ * @param dpdk_txq
+ *   Generic pointer to TX queue structure.
+ * @param[in] pkts
+ *   Packets to transmit.
+ * @param pkts_n
+ *   Number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+uint16_t
+mlx5_tx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	struct txq *txq = (struct txq *)dpdk_txq;
+	uint16_t nb_tx = 0;
+
+	while (pkts_n > nb_tx) {
+		uint8_t cs_flags = 0;
+		uint16_t n;
+		uint16_t ret;
+
+		/* Transmit multi-seg packets in the head of pkts list. */
+		if (!(txq->flags & ETH_TXQ_FLAGS_NOMULTSEGS) &&
+		    NB_SEGS(pkts[nb_tx]) > 1)
+			nb_tx += txq_scatter_v(txq,
+					       &pkts[nb_tx],
+					       pkts_n - nb_tx);
+		n = RTE_MIN((uint16_t)(pkts_n - nb_tx), MLX5_VPMD_TX_MAX_BURST);
+		if (!(txq->flags & ETH_TXQ_FLAGS_NOMULTSEGS))
+			n = txq_check_multiseg(&pkts[nb_tx], n);
+		if (!(txq->flags & ETH_TXQ_FLAGS_NOOFFLOADS))
+			n = txq_calc_offload(txq, &pkts[nb_tx], n, &cs_flags);
+		ret = txq_burst_v(txq, &pkts[nb_tx], n, cs_flags);
+		nb_tx += ret;
+		if (!ret)
+			break;
+	}
+	return nb_tx;
+}
+#endif
+
+/**
+ * Store free buffers to RX SW ring.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param pkts
+ *   Pointer to array of packets to be stored.
+ * @param pkts_n
+ *   Number of packets to be stored.
+ */
+static inline void
+rxq_copy_mbuf_v(struct rxq *rxq, struct rte_mbuf **pkts, uint16_t n)
+{
+	const uint16_t q_mask = (1 << rxq->elts_n) - 1;
+	struct rte_mbuf **elts = &(*rxq->elts)[rxq->rq_pi & q_mask];
+	unsigned int pos;
+	uint16_t p = n & -2;
+
+	for (pos = 0; pos < p; pos += 2) {
+		uint64x2_t mbp;
+
+		mbp = vld1q_u64((void *)&elts[pos]);
+		vst1q_u64((void *)&pkts[pos], mbp);
+	}
+	if (n & 1)
+		pkts[pos] = elts[pos];
+}
+
+/**
+ * Replenish buffers for RX in bulk.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param n
+ *   Number of buffers to be replenished.
+ */
+static inline void
+rxq_replenish_bulk_mbuf(struct rxq *rxq, uint16_t n)
+{
+	const uint16_t q_n = 1 << rxq->elts_n;
+	const uint16_t q_mask = q_n - 1;
+	const uint16_t elts_idx = rxq->rq_ci & q_mask;
+	struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
+	volatile struct mlx5_wqe_data_seg *wq = &(*rxq->wqes)[elts_idx];
+	unsigned int i;
+
+	assert(n >= MLX5_VPMD_RXQ_RPLNSH_THRESH);
+	assert(n <= (uint16_t)(q_n - (rxq->rq_ci - rxq->rq_pi)));
+	assert(MLX5_VPMD_RXQ_RPLNSH_THRESH > MLX5_VPMD_DESCS_PER_LOOP);
+	/* Not to cross queue end. */
+	n = RTE_MIN(n - MLX5_VPMD_DESCS_PER_LOOP, q_n - elts_idx);
+	if (rte_mempool_get_bulk(rxq->mp, (void *)elts, n) < 0) {
+		rxq->stats.rx_nombuf += n;
+		return;
+	}
+	for (i = 0; i < n; ++i)
+		wq[i].addr = htonll((uintptr_t)elts[i]->buf_addr +
+				    RTE_PKTMBUF_HEADROOM);
+	rxq->rq_ci += n;
+#ifdef SW_EMULATION
+	*rxq->rq_db = 0;
+#else
+	*rxq->rq_db = htonl(rxq->rq_ci);
+#endif
+}
+
+/**
+ * Decompress a compressed completion and fill in mbufs in RX SW ring with data
+ * extracted from the title completion descriptor.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param cq
+ *   Pointer to completion array having a compressed completion at first.
+ * @param elts
+ *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
+ *   the title completion descriptor to be copied to the rest of mbufs.
+ */
+static inline void
+rxq_cq_decompress_v(struct rxq *rxq,
+		    volatile struct mlx5_cqe *cq,
+		    struct rte_mbuf **elts)
+{
+	volatile struct mlx5_mini_cqe8 *mcq = (void *)&(cq + 1)->pkt_info;
+	struct rte_mbuf *t_pkt = elts[0]; /* Title packet is pre-built. */
+	unsigned int pos;
+#ifdef SW_EMULATION
+	unsigned int inv = 2;
+#else
+	unsigned int i;
+	unsigned int inv = 0;
+#endif
+	/* Mask to shuffle from extracted mini CQE to mbuf. */
+	const uint8x16_t mcqe_shuf_m1 = {
+		-1, -1, -1, -1, /* skip packet_type */
+		 7,  6, -1, -1, /* pkt_len, bswap16 */
+		 7,  6,         /* data_len, bswap16 */
+		-1, -1,         /* skip vlan_tci */
+		 3,  2,  1,  0  /* hash.rss, bswap32 */
+	};
+	const uint8x16_t mcqe_shuf_m2 = {
+		-1, -1, -1, -1, /* skip packet_type */
+		15, 14, -1, -1, /* pkt_len, bswap16 */
+		15, 14,         /* data_len, bswap16 */
+		-1, -1,         /* skip vlan_tci */
+		11, 10,  9,  8  /* hash.rss, bswap32 */
+	};
+	/* Restore the compressed count. Must be 16 bits. */
+	const uint16_t mcqe_n = t_pkt->data_len +
+				(rxq->crc_present * ETHER_CRC_LEN);
+	const uint64x2_t rearm =
+		vld1q_u64((void *)&t_pkt->rearm_data);
+	const uint32x4_t rxdf_mask = {
+		0xffffffff, /* packet_type */
+		0,          /* skip pkt_len */
+		0xffff0000, /* vlan_tci, skip data_len */
+		0,          /* skip hash.rss */
+	};
+	const uint8x16_t rxdf =
+		vandq_u8(vld1q_u8((void *)&t_pkt->rx_descriptor_fields1),
+			 vreinterpretq_u8_u32(rxdf_mask));
+	const uint16x8_t crc_adj = {
+		0, 0,
+		rxq->crc_present * ETHER_CRC_LEN, 0,
+		rxq->crc_present * ETHER_CRC_LEN, 0,
+		0, 0
+	};
+	const uint32_t flow_tag = t_pkt->hash.fdir.hi;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	uint32_t rcvd_byte = 0;
+#endif
+	/* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
+	const uint8x8_t len_shuf_m = {
+		 7,  6,         /* 1st mCQE */
+		15, 14,         /* 2nd mCQE */
+		23, 22,         /* 3rd mCQE */
+		31, 30          /* 4th mCQE */
+	};
+
+	/* Compile time sanity check for this function. */
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
+			 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
+			 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
+			 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
+	/*
+	 * A. load mCQEs into a 128bit register.
+	 * B. store rearm data to mbuf.
+	 * C. combine data from mCQEs with rx_descriptor_fields1.
+	 * D. store rx_descriptor_fields1.
+	 * E. store flow tag (rte_flow mark).
+	 */
+	for (pos = 0; pos < mcqe_n; ) {
+		uint8_t *p = (void *)&mcq[pos % 8];
+		uint8_t *e0 = (void *)&elts[pos]->rearm_data;
+		uint8_t *e1 = (void *)&elts[pos + 1]->rearm_data;
+		uint8_t *e2 = (void *)&elts[pos + 2]->rearm_data;
+		uint8_t *e3 = (void *)&elts[pos + 3]->rearm_data;
+		uint16x4_t byte_cnt;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		uint16x4_t invalid_mask =
+			vcreate_u16(mcqe_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
+				    -1UL << ((mcqe_n - pos) *
+					     sizeof(uint16_t) * 8) : 0);
+#endif
+
+		if (!(pos & 0x7) && pos + 8 < mcqe_n)
+			rte_prefetch0((void *)(cq + pos + 8));
+		__asm__ volatile (
+		/* A.1 load mCQEs into a 128bit register. */
+		"ld1 {v16.16b - v17.16b}, [%[mcq]]\n\t"
+		/* B.1 store rearm data to mbuf. */
+		"st1 {%[rearm].2d}, [%[e0]]\n\t"
+		"add %[e0], %[e0], #16\n\t"
+		"st1 {%[rearm].2d}, [%[e1]]\n\t"
+		"add %[e1], %[e1], #16\n\t"
+		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
+		"tbl v18.16b, {v16.16b}, %[mcqe_shuf_m1].16b\n\t"
+		"tbl v19.16b, {v16.16b}, %[mcqe_shuf_m2].16b\n\t"
+		"sub v18.8h, v18.8h, %[crc_adj].8h\n\t"
+		"sub v19.8h, v19.8h, %[crc_adj].8h\n\t"
+		"orr v18.16b, v18.16b, %[rxdf].16b\n\t"
+		"orr v19.16b, v19.16b, %[rxdf].16b\n\t"
+		/* D.1 store rx_descriptor_fields1. */
+		"st1 {v18.2d}, [%[e0]]\n\t"
+		"st1 {v19.2d}, [%[e1]]\n\t"
+		/* B.1 store rearm data to mbuf. */
+		"st1 {%[rearm].2d}, [%[e2]]\n\t"
+		"add %[e2], %[e2], #16\n\t"
+		"st1 {%[rearm].2d}, [%[e3]]\n\t"
+		"add %[e3], %[e3], #16\n\t"
+		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
+		"tbl v18.16b, {v17.16b}, %[mcqe_shuf_m1].16b\n\t"
+		"tbl v19.16b, {v17.16b}, %[mcqe_shuf_m2].16b\n\t"
+		"sub v18.8h, v18.8h, %[crc_adj].8h\n\t"
+		"sub v19.8h, v19.8h, %[crc_adj].8h\n\t"
+		"orr v18.16b, v18.16b, %[rxdf].16b\n\t"
+		"orr v19.16b, v19.16b, %[rxdf].16b\n\t"
+		/* D.1 store rx_descriptor_fields1. */
+		"st1 {v18.2d}, [%[e2]]\n\t"
+		"st1 {v19.2d}, [%[e3]]\n\t"
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		"tbl %[byte_cnt].8b, {v16.16b - v17.16b}, %[len_shuf_m].8b\n\t"
+#endif
+		:[byte_cnt]"=&w"(byte_cnt)
+		:[mcq]"r"(p), [rxdf]"w"(rxdf), [rearm]"w"(rearm),
+		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
+		 [mcqe_shuf_m1]"w"(mcqe_shuf_m1),
+		 [mcqe_shuf_m2]"w"(mcqe_shuf_m2),
+		 [crc_adj]"w"(crc_adj), [len_shuf_m]"w"(len_shuf_m)
+		:"memory", "v16", "v17", "v18", "v19");
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
+		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
+#endif
+		if (rxq->mark) {
+			/* E.1 store flow tag (rte_flow mark). */
+			elts[pos]->hash.fdir.hi = flow_tag;
+			elts[pos + 1]->hash.fdir.hi = flow_tag;
+			elts[pos + 2]->hash.fdir.hi = flow_tag;
+			elts[pos + 3]->hash.fdir.hi = flow_tag;
+		}
+		pos += MLX5_VPMD_DESCS_PER_LOOP;
+		/* Move to next CQE and invalidate consumed CQEs. */
+		if (!(pos & 0x7) && pos < mcqe_n) {
+			mcq = (void *)&(cq + pos)->pkt_info;
+#ifdef SW_EMULATION
+			for (; (inv & 7) != 0; ++inv)
+				cq[inv].op_own = MLX5_CQE_INVALIDATE;
+			++inv;
+#else
+			for (i = 0; i < 8; ++i)
+				cq[inv++].op_own = MLX5_CQE_INVALIDATE;
+#endif
+		}
+	}
+	/* Invalidate the rest of CQEs. */
+	for (; inv < mcqe_n; ++inv)
+		cq[inv].op_own = MLX5_CQE_INVALIDATE;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	rxq->stats.ipackets += mcqe_n;
+	rxq->stats.ibytes += rcvd_byte;
+#endif
+	rxq->cq_ci += mcqe_n;
+}
+
+#if 0
+/**
+ * Calculate packet type and offload flag for mbuf and store it.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param cqes[4]
+ *   Array of four 16bytes completions extracted from the original completion
+ *   descriptor.
+ * @param op_err
+ *   Opcode vector having responder error status. Each field is 4B.
+ * @param pkts
+ *   Pointer to array of packets to be filled.
+ */
+static inline void
+rxq_cq_to_ptype_oflags_v(struct rxq *rxq, __m128i cqes[4], __m128i op_err,
+			 struct rte_mbuf **pkts)
+{
+	__m128i pinfo0, pinfo1;
+	__m128i pinfo, ptype;
+	__m128i ol_flags = _mm_set1_epi32(rxq->rss_hash * PKT_RX_RSS_HASH);
+	__m128i cv_flags;
+	const __m128i zero = _mm_setzero_si128();
+	const __m128i ptype_mask =
+		_mm_set_epi32(0xfd06, 0xfd06, 0xfd06, 0xfd06);
+	const __m128i ptype_ol_mask =
+		_mm_set_epi32(0x106, 0x106, 0x106, 0x106);
+	const __m128i pinfo_mask =
+		_mm_set_epi32(0x3, 0x3, 0x3, 0x3);
+	const __m128i cv_flag_sel =
+		_mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0, 0,
+			     (uint8_t)((PKT_RX_IP_CKSUM_GOOD |
+					PKT_RX_L4_CKSUM_GOOD) >> 1),
+			     0,
+			     (uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1),
+			     0,
+			     (uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1),
+			     (uint8_t)(PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED),
+			     0);
+	const __m128i cv_mask =
+		_mm_set_epi32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
+			      PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED,
+			      PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
+			      PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED,
+			      PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
+			      PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED,
+			      PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
+			      PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED);
+	const __m128i mbuf_init =
+		_mm_loadl_epi64((__m128i *)&rxq->mbuf_initializer);
+	__m128i rearm0, rearm1, rearm2, rearm3;
+
+	/* Extract pkt_info field. */
+	pinfo0 = _mm_unpacklo_epi32(cqes[0], cqes[1]);
+	pinfo1 = _mm_unpacklo_epi32(cqes[2], cqes[3]);
+	pinfo = _mm_unpacklo_epi64(pinfo0, pinfo1);
+	/* Extract hdr_type_etc field. */
+	pinfo0 = _mm_unpackhi_epi32(cqes[0], cqes[1]);
+	pinfo1 = _mm_unpackhi_epi32(cqes[2], cqes[3]);
+	ptype = _mm_unpacklo_epi64(pinfo0, pinfo1);
+	if (rxq->mark) {
+		const __m128i pinfo_ft_mask =
+			_mm_set_epi32(0xffffff00, 0xffffff00,
+				      0xffffff00, 0xffffff00);
+		const __m128i fdir_flags = _mm_set1_epi32(PKT_RX_FDIR);
+		const __m128i fdir_id_flags = _mm_set1_epi32(PKT_RX_FDIR_ID);
+		__m128i flow_tag, invalid_mask;
+
+		flow_tag = _mm_and_si128(pinfo, pinfo_ft_mask);
+		/* Check if flow tag is non-zero then set PKT_RX_FDIR. */
+		invalid_mask = _mm_cmpeq_epi32(flow_tag, zero);
+		ol_flags = _mm_or_si128(ol_flags,
+					_mm_andnot_si128(invalid_mask,
+							 fdir_flags));
+		/* Mask out invalid entries. */
+		flow_tag = _mm_andnot_si128(invalid_mask, flow_tag);
+		/* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
+		ol_flags = _mm_or_si128(ol_flags,
+					_mm_andnot_si128(
+						_mm_cmpeq_epi32(flow_tag,
+								pinfo_ft_mask),
+						fdir_id_flags));
+	}
+	/*
+	 * Merge the two fields to generate the following:
+	 * bit[1]     = l3_ok
+	 * bit[2]     = l4_ok
+	 * bit[8]     = cv
+	 * bit[11:10] = l3_hdr_type
+	 * bit[14:12] = l4_hdr_type
+	 * bit[15]    = ip_frag
+	 * bit[16]    = tunneled
+	 * bit[17]    = outer_l3_type
+	 */
+	ptype = _mm_and_si128(ptype, ptype_mask);
+	pinfo = _mm_and_si128(pinfo, pinfo_mask);
+	pinfo = _mm_slli_epi32(pinfo, 16);
+	/* Make pinfo has merged fields for ol_flags calculation. */
+	pinfo = _mm_or_si128(ptype, pinfo);
+	ptype = _mm_srli_epi32(pinfo, 10);
+	ptype = _mm_packs_epi32(ptype, zero);
+	/* Errored packets will have RTE_PTYPE_ALL_MASK. */
+	op_err = _mm_srli_epi16(op_err, 8);
+	ptype = _mm_or_si128(ptype, op_err);
+	pkts[0]->packet_type = mlx5_ptype_table[_mm_extract_epi8(ptype, 0)];
+	pkts[1]->packet_type = mlx5_ptype_table[_mm_extract_epi8(ptype, 2)];
+	pkts[2]->packet_type = mlx5_ptype_table[_mm_extract_epi8(ptype, 4)];
+	pkts[3]->packet_type = mlx5_ptype_table[_mm_extract_epi8(ptype, 6)];
+	/* Fill flags for checksum and VLAN. */
+	pinfo = _mm_and_si128(pinfo, ptype_ol_mask);
+	pinfo = _mm_shuffle_epi8(cv_flag_sel, pinfo);
+	/* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
+	cv_flags = _mm_slli_epi32(pinfo, 9);
+	cv_flags = _mm_or_si128(pinfo, cv_flags);
+	/* Move back flags to start from byte[0]. */
+	cv_flags = _mm_srli_epi32(cv_flags, 8);
+	/* Mask out garbage bits. */
+	cv_flags = _mm_and_si128(cv_flags, cv_mask);
+	/* Merge to ol_flags. */
+	ol_flags = _mm_or_si128(ol_flags, cv_flags);
+	/* Merge mbuf_init and ol_flags. */
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
+			 offsetof(struct rte_mbuf, rearm_data) + 8);
+	rearm0 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(ol_flags, 8), 0x30);
+	rearm1 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(ol_flags, 4), 0x30);
+	rearm2 = _mm_blend_epi16(mbuf_init, ol_flags, 0x30);
+	rearm3 = _mm_blend_epi16(mbuf_init, _mm_srli_si128(ol_flags, 4), 0x30);
+	/* Write 8B rearm_data and 8B ol_flags. */
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
+			 RTE_ALIGN(offsetof(struct rte_mbuf, rearm_data), 16));
+	_mm_store_si128((__m128i *)&pkts[0]->rearm_data, rearm0);
+	_mm_store_si128((__m128i *)&pkts[1]->rearm_data, rearm1);
+	_mm_store_si128((__m128i *)&pkts[2]->rearm_data, rearm2);
+	_mm_store_si128((__m128i *)&pkts[3]->rearm_data, rearm3);
+}
+#endif
+
+/**
+ * Skip error packets.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param[out] pkts
+ *   Array to store received packets.
+ * @param pkts_n
+ *   Maximum number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully received (<= pkts_n).
+ */
+static uint16_t
+rxq_handle_pending_error(struct rxq *rxq, struct rte_mbuf **pkts,
+			 uint16_t pkts_n)
+{
+	uint16_t n = 0;
+	unsigned int i;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	uint32_t err_bytes = 0;
+#endif
+
+	for (i = 0; i < pkts_n; ++i) {
+		struct rte_mbuf *pkt = pkts[i];
+
+		if (pkt->packet_type == RTE_PTYPE_ALL_MASK) {
+#ifdef MLX5_PMD_SOFT_COUNTERS
+			err_bytes += PKT_LEN(pkt);
+#endif
+			rte_pktmbuf_free_seg(pkt);
+		} else {
+			pkts[n++] = pkt;
+		}
+	}
+	rxq->stats.idropped += (pkts_n - n);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	/* Correct counters of errored completions. */
+	rxq->stats.ipackets -= (pkts_n - n);
+	rxq->stats.ibytes -= err_bytes;
+#endif
+	rxq->pending_err = 0;
+	return n;
+}
+
+/**
+ * Receive burst of packets. An errored completion also consumes a mbuf, but the
+ * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
+ * before returning to application.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param[out] pkts
+ *   Array to store received packets.
+ * @param pkts_n
+ *   Maximum number of packets in array.
+ *
+ * @return
+ *   Number of packets received including errors (<= pkts_n).
+ */
+static inline uint16_t
+rxq_burst_v(struct rxq *rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	const uint16_t q_n = 1 << rxq->cqe_n;
+	const uint16_t q_mask = q_n - 1;
+	volatile struct mlx5_cqe *cq;
+	struct rte_mbuf **elts;
+	unsigned int pos;
+	uint64_t n;
+	uint16_t repl_n;
+	uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
+	uint16_t nocmp_n = 0;
+	uint16_t rcvd_pkt = 0;
+	unsigned int cq_idx = rxq->cq_ci & q_mask;
+	unsigned int elts_idx;
+#ifdef SW_EMULATION
+	const uint16x4_t ownership = vdup_n_u16(1);
+#else
+	const uint16x4_t ownership = vdup_n_u16(!(rxq->cq_ci & (q_mask + 1)));
+#endif
+	const uint16x4_t owner_check = vcreate_u16(0x0001000100010001);
+	const uint16x4_t opcode_check = vcreate_u16(0x00f000f000f000f0);
+	const uint16x4_t format_check = vcreate_u16(0x000c000c000c000c);
+	const uint16x4_t resp_err_check = vcreate_u16(0x00e000e000e000e0);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	uint32_t rcvd_byte = 0;
+#endif
+	/* Mask to generate 16B length vector. */
+	const uint8x8_t len_shuf_m = {
+		52, 53,         /* 4th CQE */
+		36, 37,         /* 3rd CQE */
+		20, 21,         /* 2nd CQE */
+		 4,  5          /* 1st CQE */
+	};
+	/* Mask to extract 16B data from a 64B CQE. */
+	const uint8x16_t cqe_shuf_m = {
+		29, 28,         /* hdr_type_etc, bswap16 */
+		 0,             /* pkt_info */
+		-1,             /* null */
+		47, 46,         /* byte_cnt, bswap16 */
+		31, 30,         /* vlan_info, bswap16 */
+		15, 14, 13, 12, /* rx_hash_res, bswap32 */
+		57, 58, 59,     /* flow_tag */
+		63              /* op_own */
+	};
+	/* Mask to generate 16B data for mbuf. */
+	const uint8x16_t mb_shuf_m = {
+		 4,  5, -1, -1, /* pkt_len */
+		 4,  5,         /* data_len */
+		 6,  7,         /* vlan_tci */
+		 8,  9, 10, 11, /* hash.rss */
+		12, 13, 14, -1  /* hash.fdir.hi */
+	};
+	/* Mask to generate 16B owner vector. */
+	const uint8x8_t owner_shuf_m = {
+		63, -1,         /* 4th CQE */
+		47, -1,         /* 3rd CQE */
+		31, -1,         /* 2nd CQE */
+		15, -1          /* 1st CQE */
+	};
+	const uint16x8_t crc_adj = {
+		rxq->crc_present * ETHER_CRC_LEN,
+		0,
+		rxq->crc_present * ETHER_CRC_LEN,
+		0, 0, 0, 0, 0
+	};
+	const uint32x4_t flow_mark_adj = { 0, 0, 0, rxq->mark * (-1) };
+
+	/* Compile time sanity check for this function. */
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
+			 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
+			 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
+#if (RTE_CACHE_LINE_SIZE == 128)
+	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, pkt_info) != 64);
+#else
+	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, pkt_info) != 0);
+#endif
+	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, rx_hash_res) !=
+			 offsetof(struct mlx5_cqe, pkt_info) + 12);
+	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, rsvd1) +
+			  sizeof(((struct mlx5_cqe *)0)->rsvd1) !=
+			 offsetof(struct mlx5_cqe, hdr_type_etc));
+	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, vlan_info) !=
+			 offsetof(struct mlx5_cqe, hdr_type_etc) + 2);
+	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, rsvd2) +
+			  sizeof(((struct mlx5_cqe *)0)->rsvd2) !=
+			 offsetof(struct mlx5_cqe, byte_cnt));
+	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, sop_drop_qpn) !=
+			 RTE_ALIGN(offsetof(struct mlx5_cqe, sop_drop_qpn), 8));
+	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, op_own) !=
+			 offsetof(struct mlx5_cqe, sop_drop_qpn) + 7);
+	assert(rxq->sges_n == 0);
+	assert(rxq->cqe_n == rxq->elts_n);
+	cq = &(*rxq->cqes)[cq_idx];
+	rte_prefetch0(cq);
+	rte_prefetch0(cq + 1);
+	rte_prefetch0(cq + 2);
+	rte_prefetch0(cq + 3);
+	pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
+	/*
+	 * Order of indexes:
+	 *   rq_ci >= cq_ci >= rq_pi
+	 * Definition of indexes:
+	 *   rq_ci - cq_ci := # of buffers owned by HW (posted).
+	 *   cq_ci - rq_pi := # of buffers not returned to app (decompressed).
+	 *   N - (rq_ci - rq_pi) := # of buffers consumed (to be replenished).
+	 */
+	repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
+	if (repl_n >= MLX5_VPMD_RXQ_RPLNSH_THRESH)
+		rxq_replenish_bulk_mbuf(rxq, repl_n);
+	/* See if there're unreturned mbufs from compressed CQE. */
+	rcvd_pkt = rxq->cq_ci - rxq->rq_pi;
+	if (rcvd_pkt > 0) {
+		rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
+		rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
+		rxq->rq_pi += rcvd_pkt;
+		pkts += rcvd_pkt;
+	}
+	elts_idx = rxq->rq_pi & q_mask;
+	elts = &(*rxq->elts)[elts_idx];
+	/* Not to overflow pkts array. */
+	pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
+	/* Not to cross queue end. */
+	pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
+	if (!pkts_n)
+		return rcvd_pkt;
+	/* At this point, there shouldn't be any remained packets. */
+	assert(rxq->rq_pi == rxq->cq_ci);
+	/*
+	 * A. copy 4 mbuf pointers from elts ring to returing pkts.
+	 * B. load 64B CQE and extract necessary fields
+	 *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
+	 *    following structure:
+	 *        struct {
+	 *          uint16_t hdr_type_etc;
+	 *          uint8_t  pkt_info;
+	 *          uint8_t  rsvd;
+	 *          uint16_t byte_cnt;
+	 *          uint16_t vlan_info;
+	 *          uint32_t rx_has_res;
+	 *          uint8_t  flow_tag[3];
+	 *          uint8_t  op_own;
+	 *        } c;
+	 * C. fill in mbuf.
+	 * D. get valid CQEs.
+	 * E. find compressed CQE.
+	 */
+	for (pos = 0;
+	     pos < pkts_n;
+	     pos += MLX5_VPMD_DESCS_PER_LOOP) {
+		uint16x4_t op_own;
+		uint16x4_t opcode, owner_mask, invalid_mask;
+		uint16x4_t comp_mask;
+		uint16x4_t mask;
+		uint16x4_t byte_cnt;
+		uint8_t *p0, *p1, *p2, *p3;
+		uint8_t *e0 = (void *)&elts[pos]->pkt_len;
+		uint8_t *e1 = (void *)&elts[pos + 1]->pkt_len;
+		uint8_t *e2 = (void *)&elts[pos + 2]->pkt_len;
+		uint8_t *e3 = (void *)&elts[pos + 3]->pkt_len;
+		void *elts_p = (void *)&elts[pos];
+		void *pkts_p = (void *)&pkts[pos];
+
+		/* A.0 do not cross the end of CQ. */
+		mask = vcreate_u16(pkts_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
+				   -1UL >> ((pkts_n - pos) *
+					    sizeof(uint16_t) * 8) : 0);
+		p0 = (void *)&cq[pos].pkt_info;
+		p1 = p0 + (pkts_n - pos > 1) * sizeof(struct mlx5_cqe);
+		p2 = p1 + (pkts_n - pos > 2) * sizeof(struct mlx5_cqe);
+		p3 = p2 + (pkts_n - pos > 3) * sizeof(struct mlx5_cqe);
+		/* Prefetch next 4 CQEs. */
+		if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
+			rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP]);
+			rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 1]);
+			rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 2]);
+			rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 3]);
+		}
+		__asm__ volatile (
+		/* B.1 (CQE 3) load a block having op_own. */
+		"ld1 {v19.16b}, [%[p3]]\n\t"
+		"sub %[p3], %[p3], #48\n\t"
+		/* B.2 (CQE 3) load the rest blocks. */
+		"ld1 {v16.16b - v18.16b}, [%[p3]]\n\t"
+		/* B.3 (CQE 3) extract 16B fields. */
+		"tbl v23.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b\n\t"
+		/* B.1 (CQE 2) load a block having op_own. */
+		"ld1 {v19.16b}, [%[p2]]\n\t"
+		"sub %[p2], %[p2], #48\n\t"
+		/* C.1 (CQE 3) generate final structure for mbuf. */
+		"tbl v15.16b, {v23.16b}, %[mb_shuf_m].16b\n\t"
+		/* B.2 (CQE 2) load the rest blocks. */
+		"ld1 {v16.16b - v18.16b}, [%[p2]]\n\t"
+		/* B.3 (CQE 2) extract 16B fields. */
+		"tbl v22.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b\n\t"
+		/* B.1 (CQE 1) load a block having op_own. */
+		"ld1 {v19.16b}, [%[p1]]\n\t"
+		"sub %[p1], %[p1], #48\n\t"
+		/* C.1 (CQE 2) generate final structure for mbuf. */
+		"tbl v14.16b, {v22.16b}, %[mb_shuf_m].16b\n\t"
+		/* B.2 (CQE 1) load the rest blocks. */
+		"ld1 {v16.16b - v18.16b}, [%[p1]]\n\t"
+		/* B.3 (CQE 1) extract 16B fields. */
+		"tbl v21.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b\n\t"
+		/* B.1 (CQE 0) load a block having op_own. */
+		"ld1 {v19.16b}, [%[p0]]\n\t"
+		"sub %[p0], %[p0], #48\n\t"
+		/* C.1 (CQE 1) generate final structure for mbuf. */
+		"tbl v13.16b, {v21.16b}, %[mb_shuf_m].16b\n\t"
+		/* B.2 (CQE 0) load the rest blocks. */
+		"ld1 {v16.16b - v18.16b}, [%[p0]]\n\t"
+		/* B.3 (CQE 0) extract 16B fields. */
+		"tbl v20.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b\n\t"
+		/* A.1 load mbuf pointers. */
+		"ld1 {v24.2d - v25.2d}, [%[elts_p]]\n\t"
+		/* D.1 extract op_own byte. */
+		"tbl %[op_own].8b, {v20.16b - v23.16b}, %[owner_shuf_m].8b\n\t"
+		/* C.2 (CQE 3) adjust CRC length. */
+		"sub v15.8h, v15.8h, %[crc_adj].8h\n\t"
+		/* C.3 (CQE 3) adjust flow mark. */
+		"add v15.4s, v15.4s, %[flow_mark_adj].4s\n\t"
+		/* C.4 (CQE 3) fill in mbuf - rx_descriptor_fields1. */
+		"st1 {v15.2d}, [%[e3]]\n\t"
+		/* C.2 (CQE 2) adjust CRC length. */
+		"sub v14.8h, v14.8h, %[crc_adj].8h\n\t"
+		/* C.3 (CQE 2) adjust flow mark. */
+		"add v14.4s, v14.4s, %[flow_mark_adj].4s\n\t"
+		/* C.4 (CQE 2) fill in mbuf - rx_descriptor_fields1. */
+		"st1 {v14.2d}, [%[e2]]\n\t"
+		/* C.1 (CQE 0) generate final structure for mbuf. */
+		"tbl v12.16b, {v20.16b}, %[mb_shuf_m].16b\n\t"
+		/* C.2 (CQE 1) adjust CRC length. */
+		"sub v13.8h, v13.8h, %[crc_adj].8h\n\t"
+		/* C.3 (CQE 1) adjust flow mark. */
+		"add v13.4s, v13.4s, %[flow_mark_adj].4s\n\t"
+		/* C.4 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
+		"st1 {v13.2d}, [%[e1]]\n\t"
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		/* Extract byte_cnt */
+		"tbl %[byte_cnt].8b, {v20.16b - v23.16b}, %[len_shuf_m].8b\n\t"
+#endif
+		/* A.2 copy mbuf pointers. */
+		"st1 {v24.2d - v25.2d}, [%[pkts_p]]\n\t"
+		/* C.2 (CQE 0) adjust CRC length. */
+		"sub v12.8h, v12.8h, %[crc_adj].8h\n\t"
+		/* C.3 (CQE 0) adjust flow mark. */
+		"add v12.4s, v12.4s, %[flow_mark_adj].4s\n\t"
+		/* C.4 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
+		"st1 {v12.2d}, [%[e0]]\n\t"
+		:[op_own]"=&w"(op_own), [byte_cnt]"=&w"(byte_cnt)
+		:[p3]"r"(p3 + 48), [p2]"r"(p2 + 48),
+		 [p1]"r"(p1 + 48), [p0]"r"(p0 + 48),
+		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
+		 [elts_p]"r"(elts_p), [pkts_p]"r"(pkts_p),
+		 [cqe_shuf_m]"w"(cqe_shuf_m), [mb_shuf_m]"w"(mb_shuf_m),
+		 [owner_shuf_m]"w"(owner_shuf_m), [len_shuf_m]"w"(len_shuf_m),
+		 [crc_adj]"w"(crc_adj), [flow_mark_adj]"w"(flow_mark_adj)
+		:"memory",
+		 "v12", "v13", "v14", "v15",
+		 "v16", "v17", "v18", "v19",
+		 "v20", "v21", "v22", "v23",
+		 "v24", "v25");
+		/* D.2 flip owner bit to mark CQEs from last round. */
+		owner_mask = vand_u16(op_own, owner_check);
+		owner_mask = vceq_u16(owner_mask, ownership);
+		/* D.3 get mask for invalidated CQEs. */
+		opcode = vand_u16(op_own, opcode_check);
+		invalid_mask = vceq_u16(opcode_check, opcode);
+		/* E.1 find compressed CQE format. */
+		comp_mask = vand_u16(op_own, format_check);
+		comp_mask = vceq_u16(comp_mask, format_check);
+		/* D.4 mask out beyond boundary. */
+		invalid_mask = vorr_u16(invalid_mask, mask);
+		/* D.5 merge invalid_mask with invalid owner. */
+		invalid_mask = vorr_u16(invalid_mask, owner_mask);
+		/* E.2 mask out invalid entries. */
+		comp_mask = vbic_u16(comp_mask, invalid_mask);
+		/* E.3 get the first compressed CQE. */
+		comp_idx = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
+					  comp_mask), 0)) / (sizeof(uint16_t) * 8);
+		/* D.6 mask out entries after the compressed CQE. */
+		mask = vcreate_u16(comp_idx < MLX5_VPMD_DESCS_PER_LOOP ?
+				   -1UL >> (comp_idx * sizeof(uint16_t) * 8) : 0);
+		invalid_mask = vorr_u16(invalid_mask, mask);
+		/* D.7 count non-compressed valid CQEs. */
+		n = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
+				   invalid_mask), 0)) / (sizeof(uint16_t) * 8);
+		nocmp_n += n;
+		/* D.2 get the final invalid mask. */
+		mask = vcreate_u16(n < MLX5_VPMD_DESCS_PER_LOOP ?
+				   -1UL >> (n * sizeof(uint16_t) * 8) : 0);
+		/* TODO: the following isn't needed.  */
+		invalid_mask = vorr_u16(invalid_mask, mask);
+		/* D.3 check error in opcode. */
+		opcode = vceq_u16(resp_err_check, opcode);
+		opcode = vbic_u16(opcode, invalid_mask);
+		/* D.4 mark if any error is set */
+		rxq->pending_err |=
+			!!vget_lane_u64(vreinterpret_u64_u16(opcode), 0);
+		/* C.5 fill in mbuf - rearm_data and packet_type. */
+		/* TODO:
+		rxq_cq_to_ptype_oflags_v(rxq, cqes, opcode, &pkts[pos]); */
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		/* Add up received bytes count. */
+		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
+		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
+#endif
+		/*
+		 * Break the loop unless more valid CQE is expected, or if
+		 * there's a compressed CQE.
+		 */
+		if (n != MLX5_VPMD_DESCS_PER_LOOP)
+			break;
+	}
+	/* If no new CQE seen, return without updating cq_db. */
+	if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
+		return rcvd_pkt;
+	/* Update the consumer indexes for non-compressed CQEs. */
+	assert(nocmp_n <= pkts_n);
+	rxq->cq_ci += nocmp_n;
+	rxq->rq_pi += nocmp_n;
+	rcvd_pkt += nocmp_n;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	rxq->stats.ipackets += nocmp_n;
+	rxq->stats.ibytes += rcvd_byte;
+#endif
+	/* Decompress the last CQE if compressed. */
+	if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
+		assert(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
+		rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
+		/* Return more packets if needed. */
+		if (nocmp_n < pkts_n) {
+			uint16_t n = rxq->cq_ci - rxq->rq_pi;
+
+			n = RTE_MIN(n, pkts_n - nocmp_n);
+			rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
+			rxq->rq_pi += n;
+			rcvd_pkt += n;
+		}
+	}
+	rte_compiler_barrier();
+#ifdef SW_EMULATION
+	*rxq->cq_db = 0;
+#else
+	*rxq->cq_db = htonl(rxq->cq_ci);
+#endif
+	return rcvd_pkt;
+}
+
+/**
+ * DPDK callback for vectorized RX.
+ *
+ * @param dpdk_rxq
+ *   Generic pointer to RX queue structure.
+ * @param[out] pkts
+ *   Array to store received packets.
+ * @param pkts_n
+ *   Maximum number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully received (<= pkts_n).
+ */
+uint16_t
+mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	struct rxq *rxq = dpdk_rxq;
+	uint16_t nb_rx;
+
+	nb_rx = rxq_burst_v(rxq, pkts, pkts_n);
+	if (unlikely(rxq->pending_err))
+		nb_rx = rxq_handle_pending_error(rxq, pkts, nb_rx);
+	return nb_rx;
+}
+
+/**
+ * Check Tx queue flags are set for raw vectorized Tx.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ *
+ * @return
+ *   1 if supported, negative errno value if not.
+ */
+int __attribute__((cold))
+priv_check_raw_vec_tx_support(struct priv *priv)
+{
+	uint16_t i;
+
+	/* All the configured queues should support. */
+	for (i = 0; i < priv->txqs_n; ++i) {
+		struct txq *txq = (*priv->txqs)[i];
+
+		if (!(txq->flags & ETH_TXQ_FLAGS_NOMULTSEGS) ||
+		    !(txq->flags & ETH_TXQ_FLAGS_NOOFFLOADS))
+			break;
+	}
+	if (i != priv->txqs_n)
+		return -ENOTSUP;
+	return 1;
+}
+
+/**
+ * Check a device can support vectorized TX.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ *
+ * @return
+ *   1 if supported, negative errno value if not.
+ */
+int __attribute__((cold))
+priv_check_vec_tx_support(struct priv *priv)
+{
+	if (!priv->tx_vec_en ||
+	    priv->txqs_n > MLX5_VPMD_MIN_TXQS ||
+	    priv->mps != MLX5_MPW_ENHANCED ||
+	    priv->tso)
+		return -ENOTSUP;
+	return 1;
+}
+
+/**
+ * Check a RX queue can support vectorized RX.
+ *
+ * @param rxq
+ *   Pointer to RX queue.
+ *
+ * @return
+ *   1 if supported, negative errno value if not.
+ */
+int __attribute__((cold))
+rxq_check_vec_support(struct rxq *rxq)
+{
+	struct rxq_ctrl *ctrl = container_of(rxq, struct rxq_ctrl, rxq);
+
+	if (!ctrl->priv->rx_vec_en || rxq->sges_n != 0)
+		return -ENOTSUP;
+	return 1;
+}
+
+/**
+ * Check a device can support vectorized RX.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ *
+ * @return
+ *   1 if supported, negative errno value if not.
+ */
+int __attribute__((cold))
+priv_check_vec_rx_support(struct priv *priv)
+{
+	uint16_t i;
+
+	if (!priv->rx_vec_en)
+		return -ENOTSUP;
+	/* All the configured queues should support. */
+	for (i = 0; i < priv->rxqs_n; ++i) {
+		struct rxq *rxq = (*priv->rxqs)[i];
+
+		if (rxq_check_vec_support(rxq) < 0)
+			break;
+	}
+	if (i != priv->rxqs_n)
+		return -ENOTSUP;
+	return 1;
+}
+
+/**
+ * Prepare for vectorized RX.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ */
+void
+priv_prep_vec_rx_function(struct priv *priv)
+{
+	uint16_t i;
+
+	for (i = 0; i < priv->rxqs_n; ++i) {
+		struct rxq *rxq = (*priv->rxqs)[i];
+		struct rte_mbuf *mbuf_init = &rxq->fake_mbuf;
+		const uint16_t desc = 1 << rxq->elts_n;
+		int j;
+
+		assert(rxq->elts_n == rxq->cqe_n);
+		/* Initialize default rearm_data for vPMD. */
+		mbuf_init->data_off = RTE_PKTMBUF_HEADROOM;
+		rte_mbuf_refcnt_set(mbuf_init, 1);
+		mbuf_init->nb_segs = 1;
+		mbuf_init->port = rxq->port_id;
+		/*
+		 * prevent compiler reordering:
+		 * rearm_data covers previous fields.
+		 */
+		rte_compiler_barrier();
+		rxq->mbuf_initializer =
+			*(uint64_t *)&mbuf_init->rearm_data;
+		/* Padding with a fake mbuf for vectorized Rx. */
+		for (j = 0; j < MLX5_VPMD_DESCS_PER_LOOP; ++j)
+			(*rxq->elts)[desc + j] = &rxq->fake_mbuf;
+		/* Mark that it need to be cleaned up for rxq_alloc_elts(). */
+		rxq->trim_elts = 1;
+	}
+}
-- 
2.11.0

^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [dpdk-dev] [RFC PATCH 1/1] net/mlx5: add vectorized Rx/Tx burst for ARM
  2017-08-25 18:40 ` [dpdk-dev] [RFC PATCH 1/1] " Yongseok Koh
@ 2017-09-04 12:37   ` Nélio Laranjeiro
  2017-09-11 23:28     ` Yongseok Koh
  0 siblings, 1 reply; 4+ messages in thread
From: Nélio Laranjeiro @ 2017-09-04 12:37 UTC (permalink / raw)
  To: Yongseok Koh; +Cc: adrien.mazarguil, dev

Hi Yongseok,

Some questions/comments below,

On Fri, Aug 25, 2017 at 11:40:23AM -0700, Yongseok Koh wrote:
> New Rx/Tx burst functions are added using NEON vector instructions for ARM
> CPU.
> 
> Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
> ---
>  drivers/net/mlx5/Makefile             |    2 +
>  drivers/net/mlx5/mlx5_ethdev.c        |    4 +-
>  drivers/net/mlx5/mlx5_prm.h           |   15 +
>  drivers/net/mlx5/mlx5_rxq.c           |   61 ++
>  drivers/net/mlx5/mlx5_rxtx.h          |    3 +-
>  drivers/net/mlx5/mlx5_rxtx_vec_neon.c | 1464 +++++++++++++++++++++++++++++++++
>  6 files changed, 1546 insertions(+), 3 deletions(-)
>  create mode 100644 drivers/net/mlx5/mlx5_rxtx_vec_neon.c
> 
> diff --git a/drivers/net/mlx5/Makefile b/drivers/net/mlx5/Makefile
> index 8736de5d3..616d769f8 100644
> --- a/drivers/net/mlx5/Makefile
> +++ b/drivers/net/mlx5/Makefile
> @@ -41,6 +41,8 @@ SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_txq.c
>  SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxtx.c
>  ifeq ($(CONFIG_RTE_ARCH_X86_64),y)
>  SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxtx_vec_sse.c
> +else ifeq ($(CONFIG_RTE_ARCH_ARM64),y)
> +SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxtx_vec_neon.c
>  endif
>  SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_trigger.c
>  SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_ethdev.c
> diff --git a/drivers/net/mlx5/mlx5_ethdev.c b/drivers/net/mlx5/mlx5_ethdev.c
> index b0eb3cdfc..b387c6fb3 100644
> --- a/drivers/net/mlx5/mlx5_ethdev.c
> +++ b/drivers/net/mlx5/mlx5_ethdev.c
> @@ -1516,7 +1516,7 @@ priv_select_tx_function(struct priv *priv)
>  				priv->dev->tx_pkt_burst = mlx5_tx_burst_raw_vec;
>  			else
>  				priv->dev->tx_pkt_burst = mlx5_tx_burst_vec;
> -			DEBUG("selected Enhanced MPW TX vectorized function");
> +			WARN("selected Enhanced MPW TX vectorized function");
>  		} else {
>  			priv->dev->tx_pkt_burst = mlx5_tx_burst_empw;
>  			DEBUG("selected Enhanced MPW TX function");
> @@ -1542,7 +1542,7 @@ priv_select_rx_function(struct priv *priv)
>  	if (priv_check_vec_rx_support(priv) > 0) {
>  		priv_prep_vec_rx_function(priv);
>  		priv->dev->rx_pkt_burst = mlx5_rx_burst_vec;
> -		DEBUG("selected RX vectorized function");
> +		WARN("selected RX vectorized function");

This should remain in DEBUG level.

>  	} else {
>  		priv->dev->rx_pkt_burst = mlx5_rx_burst;
>  	}
> diff --git a/drivers/net/mlx5/mlx5_prm.h b/drivers/net/mlx5/mlx5_prm.h
> index 608072f7e..01e95b466 100644
> --- a/drivers/net/mlx5/mlx5_prm.h
> +++ b/drivers/net/mlx5/mlx5_prm.h
> @@ -224,6 +224,20 @@ struct mlx5_mpw {
>  };
>  
>  /* CQ element structure - should be equal to the cache line size */
> +#if 0
> +struct mlx5_cqe { // 16B
> +       uint16_t hdr_type_etc;
> +       uint8_t pkt_info;
> +       uint8_t sop_drop_qpn; /* flow_tag */
> +       uint16_t byte_cnt;
> +       uint16_t vlan_info;
> +       uint32_t rx_hash_res;
> +       uint8_t timestamp;
> +       uint8_t wqe_counter;
> +       uint8_t rsvd4;
> +       uint8_t op_own;
> +};

Seems this structure will never be used due to the #if 0.

> +#else
>  struct mlx5_cqe {
>  #if (RTE_CACHE_LINE_SIZE == 128)
>  	uint8_t padding[64];
> @@ -243,6 +257,7 @@ struct mlx5_cqe {
>  	uint8_t rsvd4;
>  	uint8_t op_own;
>  };
> +#endif
>  
>  /**
>   * Convert a user mark to flow mark.
> diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
> index 74387a797..30654abd3 100644
> --- a/drivers/net/mlx5/mlx5_rxq.c
> +++ b/drivers/net/mlx5/mlx5_rxq.c
> @@ -786,6 +786,61 @@ rxq_cleanup(struct rxq_ctrl *rxq_ctrl)
>  	memset(rxq_ctrl, 0, sizeof(*rxq_ctrl));
>  }
>  
> +#ifdef SW_EMULATION
> +#define MLX5_EMUL_MCQE_PER_COMP 64
> +int rxq_cqe_comp_en = 1;
> +/**
> + *
> + * Filling in CQEs to emulate packet arrival
> + *
> + * @param tmpl
> + *   Pointer to RX queue control template.
> + */
> +	static inline void
> +emulate_rxq_cqe_setup(struct rxq_ctrl *tmpl)
> +{
> +	struct rxq *rxq = &tmpl->rxq;
> +	const unsigned int cqe_n = 1 << rxq->cqe_n;
> +	volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[0];
> +	unsigned int i;
> +
> +	for (i = 0; i < cqe_n; i++)
> +		cqe[i].op_own = MLX5_CQE_INVALIDATE;
> +	if (rxq_cqe_comp_en) {
> +		volatile struct mlx5_mini_cqe8 *mcqe;
> +		unsigned int j;
> +
> +		for (i = 0; i < cqe_n;) {
> +			if (!(i % MLX5_EMUL_MCQE_PER_COMP)) {
> +				/* Fill in title CQE */
> +				cqe[i].op_own = 0xc; /* Compressed */
> +				/* number of mini CQEs */
> +				cqe[i].byte_cnt = htonl(MLX5_EMUL_MCQE_PER_COMP);
> +				/* Fill in mini CQEs */
> +				mcqe = (volatile void *)&cqe[i + 1].pkt_info;
> +				for (j = 0; j < MLX5_EMUL_MCQE_PER_COMP; j++) {
> +					mcqe[j % 8].byte_cnt =
> +						rxq->crc_present ?
> +						htonl(64) : htonl(60);
> +					if ((j % 8) == 7) {
> +						i += 8;
> +						mcqe = (volatile void *)
> +						       &cqe[i].pkt_info;
> +					}
> +				}
> +			} else {
> +				ERROR("Error on building compressed CQEs");
> +			}
> +		}
> +	} else {
> +		for (i = 0; i < cqe_n; i++) {
> +			cqe[i].op_own = 0;
> +			cqe[i].byte_cnt = rxq->crc_present ? htonl(64) : htonl(60);
> +		}
> +	}
> +}
> +#endif /* SW_EMULATION */
> +
>  /**
>   * Initialize RX queue.
>   *
> @@ -1064,6 +1119,10 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
>  		      (void *)dev, strerror(ret));
>  		goto error;
>  	}
> +#ifdef SW_EMULATION
> +	rxq_cqe_comp_en = priv->cqe_comp;
> +	emulate_rxq_cqe_setup(&tmpl);
> +#endif
>  	/* Reuse buffers from original queue if possible. */
>  	if (rxq_ctrl->rxq.elts_n) {
>  		assert(1 << rxq_ctrl->rxq.elts_n == desc);
> @@ -1092,7 +1151,9 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
>  	/* Update doorbell counter. */
>  	rxq_ctrl->rxq.rq_ci = desc >> rxq_ctrl->rxq.sges_n;
>  	rte_wmb();
> +#ifndef SW_EMULATION
>  	*rxq_ctrl->rxq.rq_db = htonl(rxq_ctrl->rxq.rq_ci);
> +#endif
>  	DEBUG("%p: rxq updated with %p", (void *)rxq_ctrl, (void *)&tmpl);
>  	assert(ret == 0);
>  	return 0;

What is the purpose of this SW_EMULATION?

> diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
> index 7de1d1086..6aae00b77 100644
> --- a/drivers/net/mlx5/mlx5_rxtx.h
> +++ b/drivers/net/mlx5/mlx5_rxtx.h
> @@ -602,11 +602,12 @@ mlx5_tx_dbrec(struct txq *txq, volatile struct mlx5_wqe *wqe)
>  	uint64_t *dst = (uint64_t *)((uintptr_t)txq->bf_reg);
>  	volatile uint64_t *src = ((volatile uint64_t *)wqe);
>  
> -	rte_wmb();
> +	rte_compiler_barrier();
>  	*txq->qp_db = htonl(txq->wqe_ci);
>  	/* Ensure ordering between DB record and BF copy. */
>  	rte_wmb();
>  	*dst = *src;
> +	rte_wmb();
>  }

Is this better to have the rte_compiler_barrier() instead of the rte_io_wmb()?

>  #endif /* RTE_PMD_MLX5_RXTX_H_ */
> diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_neon.c b/drivers/net/mlx5/mlx5_rxtx_vec_neon.c
> new file mode 100644
> index 000000000..e5bce23c2
> --- /dev/null
> +++ b/drivers/net/mlx5/mlx5_rxtx_vec_neon.c
> @@ -0,0 +1,1464 @@
> +/*-
> + *   BSD LICENSE
> + *
> + *   Copyright 2017 6WIND S.A.
> + *   Copyright 2017 Mellanox.
> + *
> + *   Redistribution and use in source and binary forms, with or without
> + *   modification, are permitted provided that the following conditions
> + *   are met:
> + *
> + *     * Redistributions of source code must retain the above copyright
> + *       notice, this list of conditions and the following disclaimer.
> + *     * Redistributions in binary form must reproduce the above copyright
> + *       notice, this list of conditions and the following disclaimer in
> + *       the documentation and/or other materials provided with the
> + *       distribution.
> + *     * Neither the name of 6WIND S.A. nor the names of its
> + *       contributors may be used to endorse or promote products derived
> + *       from this software without specific prior written permission.
> + *
> + *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
> + *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
> + *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
> + *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
> + *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
> + *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
> + *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
> + *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
> + *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
> + *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
> + *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
> + */
> +
> +#include <assert.h>
> +#include <stdint.h>
> +#include <string.h>
> +#include <stdlib.h>
> +#include <arm_neon.h>
> +
> +/* Verbs header. */
> +/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
> +#ifdef PEDANTIC
> +#pragma GCC diagnostic ignored "-Wpedantic"
> +#endif
> +#include <infiniband/verbs.h>
> +#include <infiniband/mlx5_hw.h>
> +#include <infiniband/arch.h>
> +#ifdef PEDANTIC
> +#pragma GCC diagnostic error "-Wpedantic"
> +#endif

Should this patch be included before the upstream re-work?

> +
> +#include <rte_mbuf.h>
> +#include <rte_mempool.h>
> +#include <rte_prefetch.h>
> +
> +#include "mlx5.h"
> +#include "mlx5_utils.h"
> +#include "mlx5_rxtx.h"
> +#include "mlx5_autoconf.h"
> +#include "mlx5_defs.h"
> +#include "mlx5_prm.h"
> +
> +#pragma GCC diagnostic ignored "-Wcast-qual"
> +
> +/**
> + * Fill in buffer descriptors in a multi-packet send descriptor.
> + *
> + * @param txq
> + *   Pointer to TX queue structure.
> + * @param dseg
> + *   Pointer to buffer descriptor to be writen.
> + * @param pkts
> + *   Pointer to array of packets to be sent.
> + * @param n
> + *   Number of packets to be filled.
> + */
> +static inline void
> +txq_wr_dseg_v(struct txq *txq, uint8_t *dseg,
> +	      struct rte_mbuf **pkts, unsigned int n)
> +{
> +	unsigned int pos;
> +	uintptr_t addr;
> +	const uint8x16_t dseg_shuf_m = {
> +		 3,  2,  1,  0, /* length, bswap32 */
> +		 4,  5,  6,  7, /* lkey */
> +		15, 14, 13, 12, /* addr, bswap64 */
> +		11, 10,  9,  8
> +	};
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +	uint32_t tx_byte = 0;
> +#endif
> +
> +	for (pos = 0; pos < n; ++pos, dseg += MLX5_WQE_DWORD_SIZE) {
> +		uint8x16_t desc;
> +		struct rte_mbuf *pkt = pkts[pos];
> +
> +		addr = rte_pktmbuf_mtod(pkt, uintptr_t);
> +		desc = vreinterpretq_u8_u32((uint32x4_t) {
> +				DATA_LEN(pkt),
> +				mlx5_tx_mb2mr(txq, pkt),
> +				addr,
> +				addr >> 32 });
> +		desc = vqtbl1q_u8(desc, dseg_shuf_m);
> +		vst1q_u8(dseg, desc);
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +		tx_byte += DATA_LEN(pkt);
> +#endif
> +	}
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +	txq->stats.obytes += tx_byte;
> +#endif
> +}
> +
> +#if 0

#if 0?

It does not help to read an RFC with embed code blocks not even compiled.

>[...]
> +/**
> + * Send burst of packets with Enhanced MPW. If it encounters a multi-seg packet,
> + * it returns to make it processed by txq_scatter_v(). All the packets in
> + * the pkts list should be single segment packets having same offload flags.
> + * This must be checked by txq_check_multiseg() and txq_calc_offload().
> + *
> + * @param txq
> + *   Pointer to TX queue structure.
> + * @param pkts
> + *   Pointer to array of packets to be sent.
> + * @param pkts_n
> + *   Number of packets to be sent (<= MLX5_VPMD_TX_MAX_BURST).
> + * @param cs_flags
> + *   Checksum offload flags to be written in the descriptor.
> + *
> + * @return
> + *   Number of packets successfully transmitted (<= pkts_n).
> + */
> +static inline uint16_t
> +txq_burst_v(struct txq *txq, struct rte_mbuf **pkts, uint16_t pkts_n,
> +	    uint8_t cs_flags)
> +{
> +	struct rte_mbuf **elts;
> +	uint16_t elts_head = txq->elts_head;
> +	const uint16_t elts_n = 1 << txq->elts_n;
> +	const uint16_t elts_m = elts_n - 1;
> +	const unsigned int nb_dword_per_wqebb =
> +		MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE;
> +	const unsigned int nb_dword_in_hdr =
> +		sizeof(struct mlx5_wqe) / MLX5_WQE_DWORD_SIZE;
> +	unsigned int n = 0;
> +	unsigned int pos;
> +	uint16_t max_elts;
> +	uint16_t max_wqe;
> +	uint32_t comp_req = 0;
> +	const uint16_t wq_n = 1 << txq->wqe_n;
> +	const uint16_t wq_mask = wq_n - 1;
> +	uint16_t wq_idx = txq->wqe_ci & wq_mask;
> +	volatile struct mlx5_wqe64 *wq =
> +		&((volatile struct mlx5_wqe64 *)txq->wqes)[wq_idx];
> +	volatile struct mlx5_wqe *wqe = (volatile struct mlx5_wqe *)wq;
> +	const uint8x16_t ctrl_shuf_m = {
> +		 3,  2,  1,  0, /* bswap32 */
> +		 7,  6,  5,  4, /* bswap32 */
> +		11, 10,  9,  8, /* bswap32 */
> +		12, 13, 14, 15
> +	};
> +	uint8x16_t *t_wqe;
> +	uint8_t *dseg;
> +	uint8x16_t ctrl;
> +
> +	/* Make sure all packets can fit into a single WQE. */
> +	assert(elts_n > pkts_n);
> +	mlx5_tx_complete(txq);
> +	max_elts = (elts_n - (elts_head - txq->elts_tail));
> +	max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
> +	pkts_n = RTE_MIN((unsigned int)RTE_MIN(pkts_n, max_wqe), max_elts);
> +	if (unlikely(!pkts_n))
> +		return 0;
> +	elts = &(*txq->elts)[elts_head & elts_m];
> +	/* Loop for available tailroom first. */
> +	n = RTE_MIN(elts_n - (elts_head & elts_m), pkts_n);
> +	for (pos = 0; pos < (n & -2); pos += 2)
> +		vst1q_u64((void *)&elts[pos], vld1q_u64((void *)&pkts[pos]));
> +	if (n & 1)
> +		elts[pos] = pkts[pos];
> +	/* Check if it crosses the end of the queue. */
> +	if (unlikely(n < pkts_n)) {
> +		elts = &(*txq->elts)[0];
> +		for (pos = 0; pos < pkts_n - n; ++pos)
> +			elts[pos] = pkts[n + pos];
> +	}
> +	txq->elts_head += pkts_n;
> +	/* Save title WQEBB pointer. */
> +	t_wqe = (uint8x16_t *)wqe;
> +	dseg = (uint8_t *)(wqe + 1);
> +	/* Calculate the number of entries to the end. */
> +	n = RTE_MIN(
> +		(wq_n - wq_idx) * nb_dword_per_wqebb - nb_dword_in_hdr,
> +		pkts_n);
> +	/* Fill DSEGs. */
> +	txq_wr_dseg_v(txq, dseg, pkts, n);
> +	/* Check if it crosses the end of the queue. */
> +	if (n < pkts_n) {
> +		dseg = (uint8_t *)txq->wqes;
> +		txq_wr_dseg_v(txq, dseg, &pkts[n], pkts_n - n);
> +	}
> +	if (txq->elts_comp + pkts_n < MLX5_TX_COMP_THRESH) {
> +		txq->elts_comp += pkts_n;
> +	} else {
> +		/* Request a completion. */
> +		txq->elts_comp = 0;
> +		++txq->cq_pi;
> +		comp_req = 8;
> +	}
> +	/* Fill CTRL in the header. */
> +	ctrl = vreinterpretq_u8_u32((uint32x4_t) {
> +			MLX5_OPC_MOD_ENHANCED_MPSW << 24 |
> +			txq->wqe_ci << 8 | MLX5_OPCODE_ENHANCED_MPSW,
> +			txq->qp_num_8s | (pkts_n + 2),
> +			comp_req,
> +			txq->elts_head });
> +	ctrl = vqtbl1q_u8(ctrl, ctrl_shuf_m);
> +	vst1q_u8((void *)t_wqe, ctrl);
> +	/* Fill ESEG in the header. */
> +	vst1q_u8((void *)(t_wqe + 1),
> +		 (uint8x16_t) { 0, 0, 0, 0,
> +				cs_flags, 0, 0, 0,
> +				0, 0, 0, 0,
> +				0, 0, 0, 0 });
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +	txq->stats.opackets += pkts_n;
> +#endif
> +	txq->wqe_ci += (nb_dword_in_hdr + pkts_n + (nb_dword_per_wqebb - 1)) /
> +		       nb_dword_per_wqebb;
> +	/* Ring QP doorbell. */
> +	mlx5_tx_dbrec(txq, wqe);
> +	return pkts_n;
> +}
> +
> +/**
> + * DPDK callback for vectorized TX.
> + *
> + * @param dpdk_txq
> + *   Generic pointer to TX queue structure.
> + * @param[in] pkts
> + *   Packets to transmit.
> + * @param pkts_n
> + *   Number of packets in array.
> + *
> + * @return
> + *   Number of packets successfully transmitted (<= pkts_n).
> + */
> +uint16_t
> +mlx5_tx_burst_raw_vec(void *dpdk_txq, struct rte_mbuf **pkts,
> +		      uint16_t pkts_n)
> +{
> +	struct txq *txq = (struct txq *)dpdk_txq;
> +	uint16_t nb_tx = 0;
> +
> +	while (pkts_n > nb_tx) {
> +		uint16_t n;
> +		uint16_t ret;
> +
> +		n = RTE_MIN((uint16_t)(pkts_n - nb_tx), MLX5_VPMD_TX_MAX_BURST);
> +		ret = txq_burst_v(txq, &pkts[nb_tx], n, 0);
> +		nb_tx += ret;
> +		if (!ret)
> +			break;
> +	}
> +	return nb_tx;
> +}
> +
> +#if 0
> +/**
> + * DPDK callback for vectorized TX with multi-seg packets and offload.
> + *
> + * @param dpdk_txq
> + *   Generic pointer to TX queue structure.
> + * @param[in] pkts
> + *   Packets to transmit.
> + * @param pkts_n
> + *   Number of packets in array.
> + *
> + * @return
> + *   Number of packets successfully transmitted (<= pkts_n).
> + */
> +uint16_t
> +mlx5_tx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
> +{
> +	struct txq *txq = (struct txq *)dpdk_txq;
> +	uint16_t nb_tx = 0;
> +
> +	while (pkts_n > nb_tx) {
> +		uint8_t cs_flags = 0;
> +		uint16_t n;
> +		uint16_t ret;
> +
> +		/* Transmit multi-seg packets in the head of pkts list. */
> +		if (!(txq->flags & ETH_TXQ_FLAGS_NOMULTSEGS) &&
> +		    NB_SEGS(pkts[nb_tx]) > 1)
> +			nb_tx += txq_scatter_v(txq,
> +					       &pkts[nb_tx],
> +					       pkts_n - nb_tx);
> +		n = RTE_MIN((uint16_t)(pkts_n - nb_tx), MLX5_VPMD_TX_MAX_BURST);
> +		if (!(txq->flags & ETH_TXQ_FLAGS_NOMULTSEGS))
> +			n = txq_check_multiseg(&pkts[nb_tx], n);
> +		if (!(txq->flags & ETH_TXQ_FLAGS_NOOFFLOADS))
> +			n = txq_calc_offload(txq, &pkts[nb_tx], n, &cs_flags);
> +		ret = txq_burst_v(txq, &pkts[nb_tx], n, cs_flags);
> +		nb_tx += ret;
> +		if (!ret)
> +			break;
> +	}
> +	return nb_tx;
> +}
> +#endif
> +
> +/**
> + * Store free buffers to RX SW ring.
> + *
> + * @param rxq
> + *   Pointer to RX queue structure.
> + * @param pkts
> + *   Pointer to array of packets to be stored.
> + * @param pkts_n
> + *   Number of packets to be stored.
> + */
> +static inline void
> +rxq_copy_mbuf_v(struct rxq *rxq, struct rte_mbuf **pkts, uint16_t n)
> +{
> +	const uint16_t q_mask = (1 << rxq->elts_n) - 1;
> +	struct rte_mbuf **elts = &(*rxq->elts)[rxq->rq_pi & q_mask];
> +	unsigned int pos;
> +	uint16_t p = n & -2;
> +
> +	for (pos = 0; pos < p; pos += 2) {
> +		uint64x2_t mbp;
> +
> +		mbp = vld1q_u64((void *)&elts[pos]);
> +		vst1q_u64((void *)&pkts[pos], mbp);
> +	}
> +	if (n & 1)
> +		pkts[pos] = elts[pos];
> +}
> +
> +/**
> + * Replenish buffers for RX in bulk.
> + *
> + * @param rxq
> + *   Pointer to RX queue structure.
> + * @param n
> + *   Number of buffers to be replenished.
> + */
> +static inline void
> +rxq_replenish_bulk_mbuf(struct rxq *rxq, uint16_t n)
> +{
> +	const uint16_t q_n = 1 << rxq->elts_n;
> +	const uint16_t q_mask = q_n - 1;
> +	const uint16_t elts_idx = rxq->rq_ci & q_mask;
> +	struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
> +	volatile struct mlx5_wqe_data_seg *wq = &(*rxq->wqes)[elts_idx];
> +	unsigned int i;
> +
> +	assert(n >= MLX5_VPMD_RXQ_RPLNSH_THRESH);
> +	assert(n <= (uint16_t)(q_n - (rxq->rq_ci - rxq->rq_pi)));
> +	assert(MLX5_VPMD_RXQ_RPLNSH_THRESH > MLX5_VPMD_DESCS_PER_LOOP);
> +	/* Not to cross queue end. */
> +	n = RTE_MIN(n - MLX5_VPMD_DESCS_PER_LOOP, q_n - elts_idx);
> +	if (rte_mempool_get_bulk(rxq->mp, (void *)elts, n) < 0) {
> +		rxq->stats.rx_nombuf += n;
> +		return;
> +	}
> +	for (i = 0; i < n; ++i)
> +		wq[i].addr = htonll((uintptr_t)elts[i]->buf_addr +
> +				    RTE_PKTMBUF_HEADROOM);
> +	rxq->rq_ci += n;
> +#ifdef SW_EMULATION
> +	*rxq->rq_db = 0;
> +#else
> +	*rxq->rq_db = htonl(rxq->rq_ci);
> +#endif
> +}
> +
> +/**
> + * Decompress a compressed completion and fill in mbufs in RX SW ring with data
> + * extracted from the title completion descriptor.
> + *
> + * @param rxq
> + *   Pointer to RX queue structure.
> + * @param cq
> + *   Pointer to completion array having a compressed completion at first.
> + * @param elts
> + *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
> + *   the title completion descriptor to be copied to the rest of mbufs.
> + */
> +static inline void
> +rxq_cq_decompress_v(struct rxq *rxq,
> +		    volatile struct mlx5_cqe *cq,
> +		    struct rte_mbuf **elts)
> +{
> +	volatile struct mlx5_mini_cqe8 *mcq = (void *)&(cq + 1)->pkt_info;
> +	struct rte_mbuf *t_pkt = elts[0]; /* Title packet is pre-built. */
> +	unsigned int pos;
> +#ifdef SW_EMULATION
> +	unsigned int inv = 2;
> +#else
> +	unsigned int i;
> +	unsigned int inv = 0;
> +#endif
> +	/* Mask to shuffle from extracted mini CQE to mbuf. */
> +	const uint8x16_t mcqe_shuf_m1 = {
> +		-1, -1, -1, -1, /* skip packet_type */
> +		 7,  6, -1, -1, /* pkt_len, bswap16 */
> +		 7,  6,         /* data_len, bswap16 */
> +		-1, -1,         /* skip vlan_tci */
> +		 3,  2,  1,  0  /* hash.rss, bswap32 */
> +	};
> +	const uint8x16_t mcqe_shuf_m2 = {
> +		-1, -1, -1, -1, /* skip packet_type */
> +		15, 14, -1, -1, /* pkt_len, bswap16 */
> +		15, 14,         /* data_len, bswap16 */
> +		-1, -1,         /* skip vlan_tci */
> +		11, 10,  9,  8  /* hash.rss, bswap32 */
> +	};
> +	/* Restore the compressed count. Must be 16 bits. */
> +	const uint16_t mcqe_n = t_pkt->data_len +
> +				(rxq->crc_present * ETHER_CRC_LEN);
> +	const uint64x2_t rearm =
> +		vld1q_u64((void *)&t_pkt->rearm_data);
> +	const uint32x4_t rxdf_mask = {
> +		0xffffffff, /* packet_type */
> +		0,          /* skip pkt_len */
> +		0xffff0000, /* vlan_tci, skip data_len */
> +		0,          /* skip hash.rss */
> +	};
> +	const uint8x16_t rxdf =
> +		vandq_u8(vld1q_u8((void *)&t_pkt->rx_descriptor_fields1),
> +			 vreinterpretq_u8_u32(rxdf_mask));
> +	const uint16x8_t crc_adj = {
> +		0, 0,
> +		rxq->crc_present * ETHER_CRC_LEN, 0,
> +		rxq->crc_present * ETHER_CRC_LEN, 0,
> +		0, 0
> +	};
> +	const uint32_t flow_tag = t_pkt->hash.fdir.hi;
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +	uint32_t rcvd_byte = 0;
> +#endif
> +	/* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
> +	const uint8x8_t len_shuf_m = {
> +		 7,  6,         /* 1st mCQE */
> +		15, 14,         /* 2nd mCQE */
> +		23, 22,         /* 3rd mCQE */
> +		31, 30          /* 4th mCQE */
> +	};
> +
> +	/* Compile time sanity check for this function. */
> +	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
> +			 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
> +	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
> +			 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
> +	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
> +			 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
> +	/*
> +	 * A. load mCQEs into a 128bit register.
> +	 * B. store rearm data to mbuf.
> +	 * C. combine data from mCQEs with rx_descriptor_fields1.
> +	 * D. store rx_descriptor_fields1.
> +	 * E. store flow tag (rte_flow mark).
> +	 */
> +	for (pos = 0; pos < mcqe_n; ) {
> +		uint8_t *p = (void *)&mcq[pos % 8];
> +		uint8_t *e0 = (void *)&elts[pos]->rearm_data;
> +		uint8_t *e1 = (void *)&elts[pos + 1]->rearm_data;
> +		uint8_t *e2 = (void *)&elts[pos + 2]->rearm_data;
> +		uint8_t *e3 = (void *)&elts[pos + 3]->rearm_data;
> +		uint16x4_t byte_cnt;
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +		uint16x4_t invalid_mask =
> +			vcreate_u16(mcqe_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
> +				    -1UL << ((mcqe_n - pos) *
> +					     sizeof(uint16_t) * 8) : 0);
> +#endif
> +
> +		if (!(pos & 0x7) && pos + 8 < mcqe_n)
> +			rte_prefetch0((void *)(cq + pos + 8));
> +		__asm__ volatile (
> +		/* A.1 load mCQEs into a 128bit register. */
> +		"ld1 {v16.16b - v17.16b}, [%[mcq]]\n\t"
> +		/* B.1 store rearm data to mbuf. */
> +		"st1 {%[rearm].2d}, [%[e0]]\n\t"
> +		"add %[e0], %[e0], #16\n\t"
> +		"st1 {%[rearm].2d}, [%[e1]]\n\t"
> +		"add %[e1], %[e1], #16\n\t"
> +		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
> +		"tbl v18.16b, {v16.16b}, %[mcqe_shuf_m1].16b\n\t"
> +		"tbl v19.16b, {v16.16b}, %[mcqe_shuf_m2].16b\n\t"
> +		"sub v18.8h, v18.8h, %[crc_adj].8h\n\t"
> +		"sub v19.8h, v19.8h, %[crc_adj].8h\n\t"
> +		"orr v18.16b, v18.16b, %[rxdf].16b\n\t"
> +		"orr v19.16b, v19.16b, %[rxdf].16b\n\t"
> +		/* D.1 store rx_descriptor_fields1. */
> +		"st1 {v18.2d}, [%[e0]]\n\t"
> +		"st1 {v19.2d}, [%[e1]]\n\t"
> +		/* B.1 store rearm data to mbuf. */
> +		"st1 {%[rearm].2d}, [%[e2]]\n\t"
> +		"add %[e2], %[e2], #16\n\t"
> +		"st1 {%[rearm].2d}, [%[e3]]\n\t"
> +		"add %[e3], %[e3], #16\n\t"
> +		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
> +		"tbl v18.16b, {v17.16b}, %[mcqe_shuf_m1].16b\n\t"
> +		"tbl v19.16b, {v17.16b}, %[mcqe_shuf_m2].16b\n\t"
> +		"sub v18.8h, v18.8h, %[crc_adj].8h\n\t"
> +		"sub v19.8h, v19.8h, %[crc_adj].8h\n\t"
> +		"orr v18.16b, v18.16b, %[rxdf].16b\n\t"
> +		"orr v19.16b, v19.16b, %[rxdf].16b\n\t"
> +		/* D.1 store rx_descriptor_fields1. */
> +		"st1 {v18.2d}, [%[e2]]\n\t"
> +		"st1 {v19.2d}, [%[e3]]\n\t"
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +		"tbl %[byte_cnt].8b, {v16.16b - v17.16b}, %[len_shuf_m].8b\n\t"
> +#endif
> +		:[byte_cnt]"=&w"(byte_cnt)
> +		:[mcq]"r"(p), [rxdf]"w"(rxdf), [rearm]"w"(rearm),
> +		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
> +		 [mcqe_shuf_m1]"w"(mcqe_shuf_m1),
> +		 [mcqe_shuf_m2]"w"(mcqe_shuf_m2),
> +		 [crc_adj]"w"(crc_adj), [len_shuf_m]"w"(len_shuf_m)
> +		:"memory", "v16", "v17", "v18", "v19");

Is not there a better way instead of writing all those assembly instructions,
maybe by using a set of macros?

> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
> +		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
> +#endif
> +		if (rxq->mark) {
> +			/* E.1 store flow tag (rte_flow mark). */
> +			elts[pos]->hash.fdir.hi = flow_tag;
> +			elts[pos + 1]->hash.fdir.hi = flow_tag;
> +			elts[pos + 2]->hash.fdir.hi = flow_tag;
> +			elts[pos + 3]->hash.fdir.hi = flow_tag;
> +		}
> +		pos += MLX5_VPMD_DESCS_PER_LOOP;
> +		/* Move to next CQE and invalidate consumed CQEs. */
> +		if (!(pos & 0x7) && pos < mcqe_n) {
> +			mcq = (void *)&(cq + pos)->pkt_info;
> +#ifdef SW_EMULATION
> +			for (; (inv & 7) != 0; ++inv)
> +				cq[inv].op_own = MLX5_CQE_INVALIDATE;
> +			++inv;
> +#else
> +			for (i = 0; i < 8; ++i)
> +				cq[inv++].op_own = MLX5_CQE_INVALIDATE;
> +#endif
> +		}
> +	}
> +	/* Invalidate the rest of CQEs. */
> +	for (; inv < mcqe_n; ++inv)
> +		cq[inv].op_own = MLX5_CQE_INVALIDATE;
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +	rxq->stats.ipackets += mcqe_n;
> +	rxq->stats.ibytes += rcvd_byte;
> +#endif
> +	rxq->cq_ci += mcqe_n;
> +}
> +
>[...]
> +
> +/**
> + * Skip error packets.
> + *
> + * @param rxq
> + *   Pointer to RX queue structure.
> + * @param[out] pkts
> + *   Array to store received packets.
> + * @param pkts_n
> + *   Maximum number of packets in array.
> + *
> + * @return
> + *   Number of packets successfully received (<= pkts_n).
> + */
> +static uint16_t
> +rxq_handle_pending_error(struct rxq *rxq, struct rte_mbuf **pkts,
> +			 uint16_t pkts_n)
> +{
> +	uint16_t n = 0;
> +	unsigned int i;
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +	uint32_t err_bytes = 0;
> +#endif
> +
> +	for (i = 0; i < pkts_n; ++i) {
> +		struct rte_mbuf *pkt = pkts[i];
> +
> +		if (pkt->packet_type == RTE_PTYPE_ALL_MASK) {
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +			err_bytes += PKT_LEN(pkt);
> +#endif
> +			rte_pktmbuf_free_seg(pkt);
> +		} else {
> +			pkts[n++] = pkt;
> +		}
> +	}
> +	rxq->stats.idropped += (pkts_n - n);
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +	/* Correct counters of errored completions. */
> +	rxq->stats.ipackets -= (pkts_n - n);
> +	rxq->stats.ibytes -= err_bytes;
> +#endif
> +	rxq->pending_err = 0;
> +	return n;
> +}
> +
> +/**
> + * Receive burst of packets. An errored completion also consumes a mbuf, but the
> + * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
> + * before returning to application.
> + *
> + * @param rxq
> + *   Pointer to RX queue structure.
> + * @param[out] pkts
> + *   Array to store received packets.
> + * @param pkts_n
> + *   Maximum number of packets in array.
> + *
> + * @return
> + *   Number of packets received including errors (<= pkts_n).
> + */
> +static inline uint16_t
> +rxq_burst_v(struct rxq *rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
> +{
> +	const uint16_t q_n = 1 << rxq->cqe_n;
> +	const uint16_t q_mask = q_n - 1;
> +	volatile struct mlx5_cqe *cq;
> +	struct rte_mbuf **elts;
> +	unsigned int pos;
> +	uint64_t n;
> +	uint16_t repl_n;
> +	uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
> +	uint16_t nocmp_n = 0;
> +	uint16_t rcvd_pkt = 0;
> +	unsigned int cq_idx = rxq->cq_ci & q_mask;
> +	unsigned int elts_idx;
> +#ifdef SW_EMULATION
> +	const uint16x4_t ownership = vdup_n_u16(1);
> +#else
> +	const uint16x4_t ownership = vdup_n_u16(!(rxq->cq_ci & (q_mask + 1)));
> +#endif
> +	const uint16x4_t owner_check = vcreate_u16(0x0001000100010001);
> +	const uint16x4_t opcode_check = vcreate_u16(0x00f000f000f000f0);
> +	const uint16x4_t format_check = vcreate_u16(0x000c000c000c000c);
> +	const uint16x4_t resp_err_check = vcreate_u16(0x00e000e000e000e0);
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +	uint32_t rcvd_byte = 0;
> +#endif
> +	/* Mask to generate 16B length vector. */
> +	const uint8x8_t len_shuf_m = {
> +		52, 53,         /* 4th CQE */
> +		36, 37,         /* 3rd CQE */
> +		20, 21,         /* 2nd CQE */
> +		 4,  5          /* 1st CQE */
> +	};
> +	/* Mask to extract 16B data from a 64B CQE. */
> +	const uint8x16_t cqe_shuf_m = {
> +		29, 28,         /* hdr_type_etc, bswap16 */
> +		 0,             /* pkt_info */
> +		-1,             /* null */
> +		47, 46,         /* byte_cnt, bswap16 */
> +		31, 30,         /* vlan_info, bswap16 */
> +		15, 14, 13, 12, /* rx_hash_res, bswap32 */
> +		57, 58, 59,     /* flow_tag */
> +		63              /* op_own */
> +	};
> +	/* Mask to generate 16B data for mbuf. */
> +	const uint8x16_t mb_shuf_m = {
> +		 4,  5, -1, -1, /* pkt_len */
> +		 4,  5,         /* data_len */
> +		 6,  7,         /* vlan_tci */
> +		 8,  9, 10, 11, /* hash.rss */
> +		12, 13, 14, -1  /* hash.fdir.hi */
> +	};
> +	/* Mask to generate 16B owner vector. */
> +	const uint8x8_t owner_shuf_m = {
> +		63, -1,         /* 4th CQE */
> +		47, -1,         /* 3rd CQE */
> +		31, -1,         /* 2nd CQE */
> +		15, -1          /* 1st CQE */
> +	};
> +	const uint16x8_t crc_adj = {
> +		rxq->crc_present * ETHER_CRC_LEN,
> +		0,
> +		rxq->crc_present * ETHER_CRC_LEN,
> +		0, 0, 0, 0, 0
> +	};
> +	const uint32x4_t flow_mark_adj = { 0, 0, 0, rxq->mark * (-1) };
> +
> +	/* Compile time sanity check for this function. */
> +	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
> +			 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
> +	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
> +			 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
> +#if (RTE_CACHE_LINE_SIZE == 128)
> +	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, pkt_info) != 64);
> +#else
> +	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, pkt_info) != 0);
> +#endif
> +	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, rx_hash_res) !=
> +			 offsetof(struct mlx5_cqe, pkt_info) + 12);
> +	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, rsvd1) +
> +			  sizeof(((struct mlx5_cqe *)0)->rsvd1) !=
> +			 offsetof(struct mlx5_cqe, hdr_type_etc));
> +	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, vlan_info) !=
> +			 offsetof(struct mlx5_cqe, hdr_type_etc) + 2);
> +	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, rsvd2) +
> +			  sizeof(((struct mlx5_cqe *)0)->rsvd2) !=
> +			 offsetof(struct mlx5_cqe, byte_cnt));
> +	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, sop_drop_qpn) !=
> +			 RTE_ALIGN(offsetof(struct mlx5_cqe, sop_drop_qpn), 8));
> +	RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, op_own) !=
> +			 offsetof(struct mlx5_cqe, sop_drop_qpn) + 7);
> +	assert(rxq->sges_n == 0);
> +	assert(rxq->cqe_n == rxq->elts_n);
> +	cq = &(*rxq->cqes)[cq_idx];
> +	rte_prefetch0(cq);
> +	rte_prefetch0(cq + 1);
> +	rte_prefetch0(cq + 2);
> +	rte_prefetch0(cq + 3);
> +	pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
> +	/*
> +	 * Order of indexes:
> +	 *   rq_ci >= cq_ci >= rq_pi
> +	 * Definition of indexes:
> +	 *   rq_ci - cq_ci := # of buffers owned by HW (posted).
> +	 *   cq_ci - rq_pi := # of buffers not returned to app (decompressed).
> +	 *   N - (rq_ci - rq_pi) := # of buffers consumed (to be replenished).
> +	 */
> +	repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
> +	if (repl_n >= MLX5_VPMD_RXQ_RPLNSH_THRESH)
> +		rxq_replenish_bulk_mbuf(rxq, repl_n);
> +	/* See if there're unreturned mbufs from compressed CQE. */
> +	rcvd_pkt = rxq->cq_ci - rxq->rq_pi;
> +	if (rcvd_pkt > 0) {
> +		rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
> +		rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
> +		rxq->rq_pi += rcvd_pkt;
> +		pkts += rcvd_pkt;
> +	}
> +	elts_idx = rxq->rq_pi & q_mask;
> +	elts = &(*rxq->elts)[elts_idx];
> +	/* Not to overflow pkts array. */
> +	pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
> +	/* Not to cross queue end. */
> +	pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
> +	if (!pkts_n)
> +		return rcvd_pkt;
> +	/* At this point, there shouldn't be any remained packets. */
> +	assert(rxq->rq_pi == rxq->cq_ci);
> +	/*
> +	 * A. copy 4 mbuf pointers from elts ring to returing pkts.
> +	 * B. load 64B CQE and extract necessary fields
> +	 *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
> +	 *    following structure:
> +	 *        struct {
> +	 *          uint16_t hdr_type_etc;
> +	 *          uint8_t  pkt_info;
> +	 *          uint8_t  rsvd;
> +	 *          uint16_t byte_cnt;
> +	 *          uint16_t vlan_info;
> +	 *          uint32_t rx_has_res;
> +	 *          uint8_t  flow_tag[3];
> +	 *          uint8_t  op_own;
> +	 *        } c;
> +	 * C. fill in mbuf.
> +	 * D. get valid CQEs.
> +	 * E. find compressed CQE.
> +	 */
> +	for (pos = 0;
> +	     pos < pkts_n;
> +	     pos += MLX5_VPMD_DESCS_PER_LOOP) {
> +		uint16x4_t op_own;
> +		uint16x4_t opcode, owner_mask, invalid_mask;
> +		uint16x4_t comp_mask;
> +		uint16x4_t mask;
> +		uint16x4_t byte_cnt;
> +		uint8_t *p0, *p1, *p2, *p3;
> +		uint8_t *e0 = (void *)&elts[pos]->pkt_len;
> +		uint8_t *e1 = (void *)&elts[pos + 1]->pkt_len;
> +		uint8_t *e2 = (void *)&elts[pos + 2]->pkt_len;
> +		uint8_t *e3 = (void *)&elts[pos + 3]->pkt_len;
> +		void *elts_p = (void *)&elts[pos];
> +		void *pkts_p = (void *)&pkts[pos];
> +
> +		/* A.0 do not cross the end of CQ. */
> +		mask = vcreate_u16(pkts_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
> +				   -1UL >> ((pkts_n - pos) *
> +					    sizeof(uint16_t) * 8) : 0);
> +		p0 = (void *)&cq[pos].pkt_info;
> +		p1 = p0 + (pkts_n - pos > 1) * sizeof(struct mlx5_cqe);
> +		p2 = p1 + (pkts_n - pos > 2) * sizeof(struct mlx5_cqe);
> +		p3 = p2 + (pkts_n - pos > 3) * sizeof(struct mlx5_cqe);
> +		/* Prefetch next 4 CQEs. */
> +		if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
> +			rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP]);
> +			rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 1]);
> +			rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 2]);
> +			rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 3]);
> +		}
> +		__asm__ volatile (
> +		/* B.1 (CQE 3) load a block having op_own. */
> +		"ld1 {v19.16b}, [%[p3]]\n\t"
> +		"sub %[p3], %[p3], #48\n\t"
> +		/* B.2 (CQE 3) load the rest blocks. */
> +		"ld1 {v16.16b - v18.16b}, [%[p3]]\n\t"
> +		/* B.3 (CQE 3) extract 16B fields. */
> +		"tbl v23.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b\n\t"
> +		/* B.1 (CQE 2) load a block having op_own. */
> +		"ld1 {v19.16b}, [%[p2]]\n\t"
> +		"sub %[p2], %[p2], #48\n\t"
> +		/* C.1 (CQE 3) generate final structure for mbuf. */
> +		"tbl v15.16b, {v23.16b}, %[mb_shuf_m].16b\n\t"
> +		/* B.2 (CQE 2) load the rest blocks. */
> +		"ld1 {v16.16b - v18.16b}, [%[p2]]\n\t"
> +		/* B.3 (CQE 2) extract 16B fields. */
> +		"tbl v22.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b\n\t"
> +		/* B.1 (CQE 1) load a block having op_own. */
> +		"ld1 {v19.16b}, [%[p1]]\n\t"
> +		"sub %[p1], %[p1], #48\n\t"
> +		/* C.1 (CQE 2) generate final structure for mbuf. */
> +		"tbl v14.16b, {v22.16b}, %[mb_shuf_m].16b\n\t"
> +		/* B.2 (CQE 1) load the rest blocks. */
> +		"ld1 {v16.16b - v18.16b}, [%[p1]]\n\t"
> +		/* B.3 (CQE 1) extract 16B fields. */
> +		"tbl v21.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b\n\t"
> +		/* B.1 (CQE 0) load a block having op_own. */
> +		"ld1 {v19.16b}, [%[p0]]\n\t"
> +		"sub %[p0], %[p0], #48\n\t"
> +		/* C.1 (CQE 1) generate final structure for mbuf. */
> +		"tbl v13.16b, {v21.16b}, %[mb_shuf_m].16b\n\t"
> +		/* B.2 (CQE 0) load the rest blocks. */
> +		"ld1 {v16.16b - v18.16b}, [%[p0]]\n\t"
> +		/* B.3 (CQE 0) extract 16B fields. */
> +		"tbl v20.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b\n\t"
> +		/* A.1 load mbuf pointers. */
> +		"ld1 {v24.2d - v25.2d}, [%[elts_p]]\n\t"
> +		/* D.1 extract op_own byte. */
> +		"tbl %[op_own].8b, {v20.16b - v23.16b}, %[owner_shuf_m].8b\n\t"
> +		/* C.2 (CQE 3) adjust CRC length. */
> +		"sub v15.8h, v15.8h, %[crc_adj].8h\n\t"
> +		/* C.3 (CQE 3) adjust flow mark. */
> +		"add v15.4s, v15.4s, %[flow_mark_adj].4s\n\t"
> +		/* C.4 (CQE 3) fill in mbuf - rx_descriptor_fields1. */
> +		"st1 {v15.2d}, [%[e3]]\n\t"
> +		/* C.2 (CQE 2) adjust CRC length. */
> +		"sub v14.8h, v14.8h, %[crc_adj].8h\n\t"
> +		/* C.3 (CQE 2) adjust flow mark. */
> +		"add v14.4s, v14.4s, %[flow_mark_adj].4s\n\t"
> +		/* C.4 (CQE 2) fill in mbuf - rx_descriptor_fields1. */
> +		"st1 {v14.2d}, [%[e2]]\n\t"
> +		/* C.1 (CQE 0) generate final structure for mbuf. */
> +		"tbl v12.16b, {v20.16b}, %[mb_shuf_m].16b\n\t"
> +		/* C.2 (CQE 1) adjust CRC length. */
> +		"sub v13.8h, v13.8h, %[crc_adj].8h\n\t"
> +		/* C.3 (CQE 1) adjust flow mark. */
> +		"add v13.4s, v13.4s, %[flow_mark_adj].4s\n\t"
> +		/* C.4 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
> +		"st1 {v13.2d}, [%[e1]]\n\t"
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +		/* Extract byte_cnt */
> +		"tbl %[byte_cnt].8b, {v20.16b - v23.16b}, %[len_shuf_m].8b\n\t"
> +#endif
> +		/* A.2 copy mbuf pointers. */
> +		"st1 {v24.2d - v25.2d}, [%[pkts_p]]\n\t"
> +		/* C.2 (CQE 0) adjust CRC length. */
> +		"sub v12.8h, v12.8h, %[crc_adj].8h\n\t"
> +		/* C.3 (CQE 0) adjust flow mark. */
> +		"add v12.4s, v12.4s, %[flow_mark_adj].4s\n\t"
> +		/* C.4 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
> +		"st1 {v12.2d}, [%[e0]]\n\t"
> +		:[op_own]"=&w"(op_own), [byte_cnt]"=&w"(byte_cnt)
> +		:[p3]"r"(p3 + 48), [p2]"r"(p2 + 48),
> +		 [p1]"r"(p1 + 48), [p0]"r"(p0 + 48),
> +		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
> +		 [elts_p]"r"(elts_p), [pkts_p]"r"(pkts_p),
> +		 [cqe_shuf_m]"w"(cqe_shuf_m), [mb_shuf_m]"w"(mb_shuf_m),
> +		 [owner_shuf_m]"w"(owner_shuf_m), [len_shuf_m]"w"(len_shuf_m),
> +		 [crc_adj]"w"(crc_adj), [flow_mark_adj]"w"(flow_mark_adj)
> +		:"memory",
> +		 "v12", "v13", "v14", "v15",
> +		 "v16", "v17", "v18", "v19",
> +		 "v20", "v21", "v22", "v23",
> +		 "v24", "v25");
> +		/* D.2 flip owner bit to mark CQEs from last round. */
> +		owner_mask = vand_u16(op_own, owner_check);
> +		owner_mask = vceq_u16(owner_mask, ownership);
> +		/* D.3 get mask for invalidated CQEs. */
> +		opcode = vand_u16(op_own, opcode_check);
> +		invalid_mask = vceq_u16(opcode_check, opcode);
> +		/* E.1 find compressed CQE format. */
> +		comp_mask = vand_u16(op_own, format_check);
> +		comp_mask = vceq_u16(comp_mask, format_check);
> +		/* D.4 mask out beyond boundary. */
> +		invalid_mask = vorr_u16(invalid_mask, mask);
> +		/* D.5 merge invalid_mask with invalid owner. */
> +		invalid_mask = vorr_u16(invalid_mask, owner_mask);
> +		/* E.2 mask out invalid entries. */
> +		comp_mask = vbic_u16(comp_mask, invalid_mask);
> +		/* E.3 get the first compressed CQE. */
> +		comp_idx = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
> +					  comp_mask), 0)) / (sizeof(uint16_t) * 8);
> +		/* D.6 mask out entries after the compressed CQE. */
> +		mask = vcreate_u16(comp_idx < MLX5_VPMD_DESCS_PER_LOOP ?
> +				   -1UL >> (comp_idx * sizeof(uint16_t) * 8) : 0);
> +		invalid_mask = vorr_u16(invalid_mask, mask);
> +		/* D.7 count non-compressed valid CQEs. */
> +		n = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
> +				   invalid_mask), 0)) / (sizeof(uint16_t) * 8);
> +		nocmp_n += n;
> +		/* D.2 get the final invalid mask. */
> +		mask = vcreate_u16(n < MLX5_VPMD_DESCS_PER_LOOP ?
> +				   -1UL >> (n * sizeof(uint16_t) * 8) : 0);
> +		/* TODO: the following isn't needed.  */
> +		invalid_mask = vorr_u16(invalid_mask, mask);
> +		/* D.3 check error in opcode. */
> +		opcode = vceq_u16(resp_err_check, opcode);
> +		opcode = vbic_u16(opcode, invalid_mask);
> +		/* D.4 mark if any error is set */
> +		rxq->pending_err |=
> +			!!vget_lane_u64(vreinterpret_u64_u16(opcode), 0);
> +		/* C.5 fill in mbuf - rearm_data and packet_type. */
> +		/* TODO:
> +		rxq_cq_to_ptype_oflags_v(rxq, cqes, opcode, &pkts[pos]); */
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +		/* Add up received bytes count. */
> +		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
> +		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
> +#endif
> +		/*
> +		 * Break the loop unless more valid CQE is expected, or if
> +		 * there's a compressed CQE.
> +		 */
> +		if (n != MLX5_VPMD_DESCS_PER_LOOP)
> +			break;
> +	}
> +	/* If no new CQE seen, return without updating cq_db. */
> +	if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
> +		return rcvd_pkt;
> +	/* Update the consumer indexes for non-compressed CQEs. */
> +	assert(nocmp_n <= pkts_n);
> +	rxq->cq_ci += nocmp_n;
> +	rxq->rq_pi += nocmp_n;
> +	rcvd_pkt += nocmp_n;
> +#ifdef MLX5_PMD_SOFT_COUNTERS
> +	rxq->stats.ipackets += nocmp_n;
> +	rxq->stats.ibytes += rcvd_byte;
> +#endif
> +	/* Decompress the last CQE if compressed. */
> +	if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
> +		assert(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
> +		rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
> +		/* Return more packets if needed. */
> +		if (nocmp_n < pkts_n) {
> +			uint16_t n = rxq->cq_ci - rxq->rq_pi;
> +
> +			n = RTE_MIN(n, pkts_n - nocmp_n);
> +			rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
> +			rxq->rq_pi += n;
> +			rcvd_pkt += n;
> +		}
> +	}
> +	rte_compiler_barrier();
> +#ifdef SW_EMULATION
> +	*rxq->cq_db = 0;
> +#else
> +	*rxq->cq_db = htonl(rxq->cq_ci);
> +#endif
> +	return rcvd_pkt;
> +}
> +
> +/**
> + * DPDK callback for vectorized RX.
> + *
> + * @param dpdk_rxq
> + *   Generic pointer to RX queue structure.
> + * @param[out] pkts
> + *   Array to store received packets.
> + * @param pkts_n
> + *   Maximum number of packets in array.
> + *
> + * @return
> + *   Number of packets successfully received (<= pkts_n).
> + */
> +uint16_t
> +mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
> +{
> +	struct rxq *rxq = dpdk_rxq;
> +	uint16_t nb_rx;
> +
> +	nb_rx = rxq_burst_v(rxq, pkts, pkts_n);
> +	if (unlikely(rxq->pending_err))
> +		nb_rx = rxq_handle_pending_error(rxq, pkts, nb_rx);
> +	return nb_rx;
> +}
> +
> +/**
> + * Check Tx queue flags are set for raw vectorized Tx.
> + *
> + * @param priv
> + *   Pointer to private structure.
> + *
> + * @return
> + *   1 if supported, negative errno value if not.
> + */
> +int __attribute__((cold))
> +priv_check_raw_vec_tx_support(struct priv *priv)
> +{
> +	uint16_t i;
> +
> +	/* All the configured queues should support. */
> +	for (i = 0; i < priv->txqs_n; ++i) {
> +		struct txq *txq = (*priv->txqs)[i];
> +
> +		if (!(txq->flags & ETH_TXQ_FLAGS_NOMULTSEGS) ||
> +		    !(txq->flags & ETH_TXQ_FLAGS_NOOFFLOADS))
> +			break;
> +	}
> +	if (i != priv->txqs_n)
> +		return -ENOTSUP;
> +	return 1;
> +}
> +
> +/**
> + * Check a device can support vectorized TX.
> + *
> + * @param priv
> + *   Pointer to private structure.
> + *
> + * @return
> + *   1 if supported, negative errno value if not.
> + */
> +int __attribute__((cold))
> +priv_check_vec_tx_support(struct priv *priv)
> +{
> +	if (!priv->tx_vec_en ||
> +	    priv->txqs_n > MLX5_VPMD_MIN_TXQS ||
> +	    priv->mps != MLX5_MPW_ENHANCED ||
> +	    priv->tso)
> +		return -ENOTSUP;
> +	return 1;
> +}
> +
> +/**
> + * Check a RX queue can support vectorized RX.
> + *
> + * @param rxq
> + *   Pointer to RX queue.
> + *
> + * @return
> + *   1 if supported, negative errno value if not.
> + */
> +int __attribute__((cold))
> +rxq_check_vec_support(struct rxq *rxq)
> +{
> +	struct rxq_ctrl *ctrl = container_of(rxq, struct rxq_ctrl, rxq);
> +
> +	if (!ctrl->priv->rx_vec_en || rxq->sges_n != 0)
> +		return -ENOTSUP;
> +	return 1;
> +}
> +
> +/**
> + * Check a device can support vectorized RX.
> + *
> + * @param priv
> + *   Pointer to private structure.
> + *
> + * @return
> + *   1 if supported, negative errno value if not.
> + */
> +int __attribute__((cold))
> +priv_check_vec_rx_support(struct priv *priv)
> +{
> +	uint16_t i;
> +
> +	if (!priv->rx_vec_en)
> +		return -ENOTSUP;
> +	/* All the configured queues should support. */
> +	for (i = 0; i < priv->rxqs_n; ++i) {
> +		struct rxq *rxq = (*priv->rxqs)[i];
> +
> +		if (rxq_check_vec_support(rxq) < 0)
> +			break;
> +	}
> +	if (i != priv->rxqs_n)
> +		return -ENOTSUP;
> +	return 1;
> +}
> +
> +/**
> + * Prepare for vectorized RX.
> + *
> + * @param priv
> + *   Pointer to private structure.
> + */
> +void
> +priv_prep_vec_rx_function(struct priv *priv)
> +{
> +	uint16_t i;
> +
> +	for (i = 0; i < priv->rxqs_n; ++i) {
> +		struct rxq *rxq = (*priv->rxqs)[i];
> +		struct rte_mbuf *mbuf_init = &rxq->fake_mbuf;
> +		const uint16_t desc = 1 << rxq->elts_n;
> +		int j;
> +
> +		assert(rxq->elts_n == rxq->cqe_n);
> +		/* Initialize default rearm_data for vPMD. */
> +		mbuf_init->data_off = RTE_PKTMBUF_HEADROOM;
> +		rte_mbuf_refcnt_set(mbuf_init, 1);
> +		mbuf_init->nb_segs = 1;
> +		mbuf_init->port = rxq->port_id;
> +		/*
> +		 * prevent compiler reordering:
> +		 * rearm_data covers previous fields.
> +		 */
> +		rte_compiler_barrier();
> +		rxq->mbuf_initializer =
> +			*(uint64_t *)&mbuf_init->rearm_data;
> +		/* Padding with a fake mbuf for vectorized Rx. */
> +		for (j = 0; j < MLX5_VPMD_DESCS_PER_LOOP; ++j)
> +			(*rxq->elts)[desc + j] = &rxq->fake_mbuf;
> +		/* Mark that it need to be cleaned up for rxq_alloc_elts(). */
> +		rxq->trim_elts = 1;
> +	}
> +}
> -- 
> 2.11.0

It follows the same algorithm as for x86 side, unless the fact you seem to
need a lot of assembly instruction due to the missing instrasics I suppose.

Do you already have some good feedbacks in performance?

Thanks,

-- 
Nélio Laranjeiro
6WIND

^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [dpdk-dev] [RFC PATCH 1/1] net/mlx5: add vectorized Rx/Tx burst for ARM
  2017-09-04 12:37   ` Nélio Laranjeiro
@ 2017-09-11 23:28     ` Yongseok Koh
  0 siblings, 0 replies; 4+ messages in thread
From: Yongseok Koh @ 2017-09-11 23:28 UTC (permalink / raw)
  To: Nélio Laranjeiro; +Cc: adrien.mazarguil, dev

Hi Nelio,

Sorry for delayed reply. I was on vacation for a week.

On Mon, Sep 04, 2017 at 02:37:05PM +0200, Nélio Laranjeiro wrote:
> Hi Yongseok,
> 
> Some questions/comments below,
> 
> On Fri, Aug 25, 2017 at 11:40:23AM -0700, Yongseok Koh wrote:
> > New Rx/Tx burst functions are added using NEON vector instructions for ARM
> > CPU.
> > 
> > Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
[...]
> > @@ -1542,7 +1542,7 @@ priv_select_rx_function(struct priv *priv)
> >  	if (priv_check_vec_rx_support(priv) > 0) {
> >  		priv_prep_vec_rx_function(priv);
> >  		priv->dev->rx_pkt_burst = mlx5_rx_burst_vec;
> > -		DEBUG("selected RX vectorized function");
> > +		WARN("selected RX vectorized function");
> 
> This should remain in DEBUG level.
My bad, that was just for debugging purpose while I was testing the code. Will
change it back when I send out a patchset for integration.

> >  	} else {
> >  		priv->dev->rx_pkt_burst = mlx5_rx_burst;
> >  	}
> > diff --git a/drivers/net/mlx5/mlx5_prm.h b/drivers/net/mlx5/mlx5_prm.h
> > index 608072f7e..01e95b466 100644
> > --- a/drivers/net/mlx5/mlx5_prm.h
> > +++ b/drivers/net/mlx5/mlx5_prm.h
> > @@ -224,6 +224,20 @@ struct mlx5_mpw {
> >  };
> >  
> >  /* CQ element structure - should be equal to the cache line size */
> > +#if 0
> > +struct mlx5_cqe { // 16B
> > +       uint16_t hdr_type_etc;
> > +       uint8_t pkt_info;
> > +       uint8_t sop_drop_qpn; /* flow_tag */
> > +       uint16_t byte_cnt;
> > +       uint16_t vlan_info;
> > +       uint32_t rx_hash_res;
> > +       uint8_t timestamp;
> > +       uint8_t wqe_counter;
> > +       uint8_t rsvd4;
> > +       uint8_t op_own;
> > +};
> 
> Seems this structure will never be used due to the #if 0.
This code was just for SW emulation to emulate the result of 16B CQE size. I
should've trimmed all the debugging/testing code before I sent out this RFC.

[...]
> > @@ -1064,6 +1119,10 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
> >  		      (void *)dev, strerror(ret));
> >  		goto error;
> >  	}
> > +#ifdef SW_EMULATION
> > +	rxq_cqe_comp_en = priv->cqe_comp;
> > +	emulate_rxq_cqe_setup(&tmpl);
> > +#endif
> >  	/* Reuse buffers from original queue if possible. */
> >  	if (rxq_ctrl->rxq.elts_n) {
> >  		assert(1 << rxq_ctrl->rxq.elts_n == desc);
> > @@ -1092,7 +1151,9 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
> >  	/* Update doorbell counter. */
> >  	rxq_ctrl->rxq.rq_ci = desc >> rxq_ctrl->rxq.sges_n;
> >  	rte_wmb();
> > +#ifndef SW_EMULATION
> >  	*rxq_ctrl->rxq.rq_db = htonl(rxq_ctrl->rxq.rq_ci);
> > +#endif
> >  	DEBUG("%p: rxq updated with %p", (void *)rxq_ctrl, (void *)&tmpl);
> >  	assert(ret == 0);
> >  	return 0;
> 
> What is the purpose of this SW_EMULATION?
That's to emulate packet Rx without getting device involved. And it was to
optimize cycle budget from pure SW perspective. This will get removed in a
formal patch. Please understand that it is RFC code.

> > diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
> > index 7de1d1086..6aae00b77 100644
> > --- a/drivers/net/mlx5/mlx5_rxtx.h
> > +++ b/drivers/net/mlx5/mlx5_rxtx.h
> > @@ -602,11 +602,12 @@ mlx5_tx_dbrec(struct txq *txq, volatile struct mlx5_wqe *wqe)
> >  	uint64_t *dst = (uint64_t *)((uintptr_t)txq->bf_reg);
> >  	volatile uint64_t *src = ((volatile uint64_t *)wqe);
> >  
> > -	rte_wmb();
> > +	rte_compiler_barrier();
> >  	*txq->qp_db = htonl(txq->wqe_ci);
> >  	/* Ensure ordering between DB record and BF copy. */
> >  	rte_wmb();
> >  	*dst = *src;
> > +	rte_wmb();
> >  }
> 
> Is this better to have the rte_compiler_barrier() instead of the rte_io_wmb()?
You are right. I'm also aware of Shahaf's patch - "net/mlx5: replace memory
barrier type". This code block was to include Shahaf's patch for testing and
I've already included Shahaf's final patch. As his patch's been merged, will
remove this hunk in a formal patch.

[...]
> > +/* Verbs header. */
> > +/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
> > +#ifdef PEDANTIC
> > +#pragma GCC diagnostic ignored "-Wpedantic"
> > +#endif
> > +#include <infiniband/verbs.h>
> > +#include <infiniband/mlx5_hw.h>
> > +#include <infiniband/arch.h>
> > +#ifdef PEDANTIC
> > +#pragma GCC diagnostic error "-Wpedantic"
> > +#endif
>
> Should this patch be included before the upstream re-work?
No, after the rework. Will change it properly then.

[...]
> > +		vst1q_u8(dseg, desc);
> > +#ifdef MLX5_PMD_SOFT_COUNTERS
> > +		tx_byte += DATA_LEN(pkt);
> > +#endif
> > +	}
> > +#ifdef MLX5_PMD_SOFT_COUNTERS
> > +	txq->stats.obytes += tx_byte;
> > +#endif
> > +}
> > +
> > +#if 0
> 
> #if 0?
> 
> It does not help to read an RFC with embed code blocks not even compiled.
Right. That was my mistake in rushing to meet the submission deadline. Sorry for
the mess. Now, I've completed porting the commented part and I'll be able to
send out a formal patchset soon.

[...]
> > +		__asm__ volatile (
> > +		/* A.1 load mCQEs into a 128bit register. */
> > +		"ld1 {v16.16b - v17.16b}, [%[mcq]]\n\t"
> > +		/* B.1 store rearm data to mbuf. */
> > +		"st1 {%[rearm].2d}, [%[e0]]\n\t"
> > +		"add %[e0], %[e0], #16\n\t"
> > +		"st1 {%[rearm].2d}, [%[e1]]\n\t"
> > +		"add %[e1], %[e1], #16\n\t"
> > +		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
> > +		"tbl v18.16b, {v16.16b}, %[mcqe_shuf_m1].16b\n\t"
> > +		"tbl v19.16b, {v16.16b}, %[mcqe_shuf_m2].16b\n\t"
> > +		"sub v18.8h, v18.8h, %[crc_adj].8h\n\t"
> > +		"sub v19.8h, v19.8h, %[crc_adj].8h\n\t"
> > +		"orr v18.16b, v18.16b, %[rxdf].16b\n\t"
> > +		"orr v19.16b, v19.16b, %[rxdf].16b\n\t"
> > +		/* D.1 store rx_descriptor_fields1. */
> > +		"st1 {v18.2d}, [%[e0]]\n\t"
> > +		"st1 {v19.2d}, [%[e1]]\n\t"
> > +		/* B.1 store rearm data to mbuf. */
> > +		"st1 {%[rearm].2d}, [%[e2]]\n\t"
> > +		"add %[e2], %[e2], #16\n\t"
> > +		"st1 {%[rearm].2d}, [%[e3]]\n\t"
> > +		"add %[e3], %[e3], #16\n\t"
> > +		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
> > +		"tbl v18.16b, {v17.16b}, %[mcqe_shuf_m1].16b\n\t"
> > +		"tbl v19.16b, {v17.16b}, %[mcqe_shuf_m2].16b\n\t"
> > +		"sub v18.8h, v18.8h, %[crc_adj].8h\n\t"
> > +		"sub v19.8h, v19.8h, %[crc_adj].8h\n\t"
> > +		"orr v18.16b, v18.16b, %[rxdf].16b\n\t"
> > +		"orr v19.16b, v19.16b, %[rxdf].16b\n\t"
> > +		/* D.1 store rx_descriptor_fields1. */
> > +		"st1 {v18.2d}, [%[e2]]\n\t"
> > +		"st1 {v19.2d}, [%[e3]]\n\t"
> > +#ifdef MLX5_PMD_SOFT_COUNTERS
> > +		"tbl %[byte_cnt].8b, {v16.16b - v17.16b}, %[len_shuf_m].8b\n\t"
> > +#endif
> > +		:[byte_cnt]"=&w"(byte_cnt)
> > +		:[mcq]"r"(p), [rxdf]"w"(rxdf), [rearm]"w"(rearm),
> > +		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
> > +		 [mcqe_shuf_m1]"w"(mcqe_shuf_m1),
> > +		 [mcqe_shuf_m2]"w"(mcqe_shuf_m2),
> > +		 [crc_adj]"w"(crc_adj), [len_shuf_m]"w"(len_shuf_m)
> > +		:"memory", "v16", "v17", "v18", "v19");
> 
> Is not there a better way instead of writing all those assembly instructions,
> maybe by using a set of macros?
[...]
> It follows the same algorithm as for x86 side, unless the fact you seem to
> need a lot of assembly instruction due to the missing instrasics I suppose.
Yes, the main reason for this inline assembly was due to lack of load
instrinsics - vld1q_u64_x3() and inefficiency of a shuffle intrinsic
vqtbl3q_u8() in the gcc-5.4.0-6ubuntu1~16.04.4.

I'm quite sure that using vld1q_u64_x3() is more efficient than three
vld1q_u64() although vld1q_u64_x3() internally executes 3 micro-ops in NEON.
This is used to load the first three 16byte blocks from a CQE.

And in vqtbl3q_u8() of arm_neon.h, it uses a redundant load instruction to load
data into 3 consecutive SIMD registers. This caused hotspots in perf-annotate.

> Do you already have some good feedbacks in performance?
Currently, I'm seeing 45% improvement comparing with regular rx/tx_burst in case
of single core. Also, when I firstly wrote the code, I could get ~20% benefit by
removing the hottest instructions with inline assembly.

I also concerned about difficulty for maintenance and code reading, so I tried
to minimize the assembly part as much as possible and it is limited to
fetching/manipulating CQE data on Rx, which is most critical in performance.

Thanks,
Yongseok

^ permalink raw reply	[flat|nested] 4+ messages in thread

end of thread, other threads:[~2017-09-11 23:28 UTC | newest]

Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2017-08-25 18:40 [dpdk-dev] [RFC PATCH 0/1] net/mlx5: add vectorized Rx/Tx burst for ARM Yongseok Koh
2017-08-25 18:40 ` [dpdk-dev] [RFC PATCH 1/1] " Yongseok Koh
2017-09-04 12:37   ` Nélio Laranjeiro
2017-09-11 23:28     ` Yongseok Koh

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).