net/mlx5: add vectorized Rx/Tx burst for ARM
authorYongseok Koh <yskoh@mellanox.com>
Mon, 9 Oct 2017 18:47:00 +0000 (11:47 -0700)
committerFerruh Yigit <ferruh.yigit@intel.com>
Thu, 12 Oct 2017 00:36:58 +0000 (01:36 +0100)
Brings vectorization through NEON instructions.

Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
drivers/net/mlx5/Makefile
drivers/net/mlx5/mlx5_rxtx_vec.c
drivers/net/mlx5/mlx5_rxtx_vec.h
drivers/net/mlx5/mlx5_rxtx_vec_neon.h [new file with mode: 0644]

index bd223e8..e7aca04 100644 (file)
@@ -39,7 +39,8 @@ SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxq.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_txq.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxtx.c
-ifeq ($(CONFIG_RTE_ARCH_X86_64),y)
+ifneq ($(filter y,$(CONFIG_RTE_ARCH_X86_64) \
+                 $(CONFIG_RTE_ARCH_ARM64)),)
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxtx_vec.c
 endif
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_trigger.c
index edc6638..ba6c8ce 100644 (file)
 #include "mlx5_defs.h"
 #include "mlx5_prm.h"
 
-#ifdef RTE_ARCH_X86_64
+#if defined RTE_ARCH_X86_64
 #include "mlx5_rxtx_vec_sse.h"
+#elif defined RTE_ARCH_ARM64
+#include "mlx5_rxtx_vec_neon.h"
 #else
 #error "This should not be compiled if SIMD instructions are not supported."
 #endif
index 9656fb7..4261690 100644 (file)
@@ -68,7 +68,11 @@ S_ASSERT_RTE_MBUF(offsetof(struct rte_mbuf, pkt_len) ==
                  offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
 S_ASSERT_RTE_MBUF(offsetof(struct rte_mbuf, data_len) ==
                  offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
+#if (RTE_CACHE_LINE_SIZE == 128)
+S_ASSERT_MLX5_CQE(offsetof(struct mlx5_cqe, pkt_info) == 64);
+#else
 S_ASSERT_MLX5_CQE(offsetof(struct mlx5_cqe, pkt_info) == 0);
+#endif
 S_ASSERT_MLX5_CQE(offsetof(struct mlx5_cqe, rx_hash_res) ==
                  offsetof(struct mlx5_cqe, pkt_info) + 12);
 S_ASSERT_MLX5_CQE(offsetof(struct mlx5_cqe, rsvd1) +
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_neon.h b/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
new file mode 100644 (file)
index 0000000..6dd18b6
--- /dev/null
@@ -0,0 +1,1028 @@
+/*-
+ *   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.
+ */
+
+#ifndef RTE_PMD_MLX5_RXTX_VEC_NEON_H_
+#define RTE_PMD_MLX5_RXTX_VEC_NEON_H_
+
+#include <assert.h>
+#include <stdint.h>
+#include <string.h>
+#include <stdlib.h>
+#include <arm_neon.h>
+
+#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_rxtx_vec.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 mlx5_txq_data *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
+}
+
+/**
+ * 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 mlx5_txq_data *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 uint8x16_t ctrl_shuf_m = {
+                       3,  2,  1,  0, /* bswap32 */
+                       7,  6,  5,  4, /* bswap32 */
+                       11, 10,  9,  8, /* bswap32 */
+                       12, 13, 14, 15
+               };
+               uint8_t cs_flags = 0;
+               uint16_t max_elts;
+               uint16_t max_wqe;
+               uint8x16_t *t_wqe;
+               uint8_t *dseg;
+               uint8x16_t 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 = (uint8x16_t *)wqe;
+               dseg = (uint8_t *)(wqe + 1);
+               do {
+                       if (!(ds++ % nb_dword_per_wqebb)) {
+                               dseg = (uint8_t *)
+                                       &((volatile struct mlx5_wqe64 *)
+                                          txq->wqes)[++wqe_ci & wq_mask];
+                       }
+                       txq_wr_dseg_v(txq, dseg, &buf, 1);
+                       dseg += MLX5_WQE_DWORD_SIZE;
+                       (*txq->elts)[elts_head++ & elts_m] = buf;
+                       buf = buf->next;
+               } while (--segs_n);
+               ++wqe_ci;
+               /* Fill CTRL in the header. */
+               ctrl = vreinterpretq_u8_u32((uint32x4_t) {
+                               MLX5_OPC_MOD_MPW << 24 |
+                               txq->wqe_ci << 8 | MLX5_OPCODE_TSO,
+                               txq->qp_num_8s | ds, 0, 0});
+               ctrl = vqtbl1q_u8(ctrl, ctrl_shuf_m);
+               vst1q_u8((void *)t_wqe, ctrl);
+               /* Fill ESEG in the header. */
+               vst1q_u16((void *)(t_wqe + 1),
+                         (uint16x8_t) { 0, 0, cs_flags, rte_cpu_to_be_16(len),
+                                        0, 0, 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] = rte_cpu_to_be_32(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;
+}
+
+/**
+ * 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 mlx5_txq_data *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;
+}
+
+/**
+ * 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 mlx5_rxq_data *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];
+}
+
+/**
+ * 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 mlx5_rxq_data *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;
+       unsigned int i;
+       unsigned int inv = 0;
+       /* 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 */
+       };
+
+       /*
+        * Not to overflow elts array. Decompress next time after mbuf
+        * replenishment.
+        */
+       if (unlikely(mcqe_n + MLX5_VPMD_DESCS_PER_LOOP >
+                    (uint16_t)(rxq->rq_ci - rxq->cq_ci)))
+               return;
+       /*
+        * 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;
+                       for (i = 0; i < 8; ++i)
+                               cq[inv++].op_own = MLX5_CQE_INVALIDATE;
+               }
+       }
+       /* 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;
+}
+
+/**
+ * Calculate packet type and offload flag for mbuf and store it.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param ptype_info
+ *   Array of four 4bytes packet type info extracted from the original
+ *   completion descriptor.
+ * @param flow_tag
+ *   Array of four 4bytes flow ID 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 mlx5_rxq_data *rxq,
+                        uint32x4_t ptype_info, uint32x4_t flow_tag,
+                        uint16x4_t op_err, struct rte_mbuf **pkts)
+{
+       uint16x4_t ptype;
+       uint32x4_t pinfo, cv_flags;
+       uint32x4_t ol_flags = vdupq_n_u32(rxq->rss_hash * PKT_RX_RSS_HASH);
+       const uint32x4_t ptype_ol_mask = { 0x106, 0x106, 0x106, 0x106 };
+       const uint8x16_t cv_flag_sel = {
+               0,
+               (uint8_t)(PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED),
+               (uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1),
+               0,
+               (uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1),
+               0,
+               (uint8_t)((PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1),
+               0, 0, 0, 0, 0, 0, 0, 0, 0
+       };
+       const uint32x4_t cv_mask =
+               vdupq_n_u32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
+                           PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED);
+       const uint64x1_t mbuf_init = vld1_u64(&rxq->mbuf_initializer);
+       const uint64x1_t r32_mask = vcreate_u64(0xffffffff);
+       uint64x2_t rearm0, rearm1, rearm2, rearm3;
+
+       if (rxq->mark) {
+               const uint32x4_t ft_def = vdupq_n_u32(MLX5_FLOW_MARK_DEFAULT);
+               const uint32x4_t fdir_flags = vdupq_n_u32(PKT_RX_FDIR);
+               const uint32x4_t fdir_id_flags = vdupq_n_u32(PKT_RX_FDIR_ID);
+
+               /* Check if flow tag is non-zero then set PKT_RX_FDIR. */
+               ol_flags = vorrq_u32(ol_flags, vbicq_u32(fdir_flags,
+                                                        vceqzq_u32(flow_tag)));
+               /* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
+               ol_flags = vorrq_u32(ol_flags,
+                                    vbicq_u32(fdir_id_flags,
+                                              vceqq_u32(flow_tag, ft_def)));
+       }
+       /*
+        * ptype_info has 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 = vshrn_n_u32(ptype_info, 10);
+       /* Errored packets will have RTE_PTYPE_ALL_MASK. */
+       ptype = vorr_u16(ptype, op_err);
+       pkts[0]->packet_type =
+               mlx5_ptype_table[vget_lane_u8(vreinterpret_u8_u16(ptype), 6)];
+       pkts[1]->packet_type =
+               mlx5_ptype_table[vget_lane_u8(vreinterpret_u8_u16(ptype), 4)];
+       pkts[2]->packet_type =
+               mlx5_ptype_table[vget_lane_u8(vreinterpret_u8_u16(ptype), 2)];
+       pkts[3]->packet_type =
+               mlx5_ptype_table[vget_lane_u8(vreinterpret_u8_u16(ptype), 0)];
+       /* Fill flags for checksum and VLAN. */
+       pinfo = vandq_u32(ptype_info, ptype_ol_mask);
+       pinfo = vreinterpretq_u32_u8(
+               vqtbl1q_u8(cv_flag_sel, vreinterpretq_u8_u32(pinfo)));
+       /* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
+       cv_flags = vshlq_n_u32(pinfo, 9);
+       cv_flags = vorrq_u32(pinfo, cv_flags);
+       /* Move back flags to start from byte[0]. */
+       cv_flags = vshrq_n_u32(cv_flags, 8);
+       /* Mask out garbage bits. */
+       cv_flags = vandq_u32(cv_flags, cv_mask);
+       /* Merge to ol_flags. */
+       ol_flags = vorrq_u32(ol_flags, cv_flags);
+       /* Merge mbuf_init and ol_flags, and store. */
+       rearm0 = vcombine_u64(mbuf_init,
+                             vshr_n_u64(vget_high_u64(vreinterpretq_u64_u32(
+                                                      ol_flags)), 32));
+       rearm1 = vcombine_u64(mbuf_init,
+                             vand_u64(vget_high_u64(vreinterpretq_u64_u32(
+                                                    ol_flags)), r32_mask));
+       rearm2 = vcombine_u64(mbuf_init,
+                             vshr_n_u64(vget_low_u64(vreinterpretq_u64_u32(
+                                                     ol_flags)), 32));
+       rearm3 = vcombine_u64(mbuf_init,
+                             vand_u64(vget_low_u64(vreinterpretq_u64_u32(
+                                                   ol_flags)), r32_mask));
+       vst1q_u64((void *)&pkts[0]->rearm_data, rearm0);
+       vst1q_u64((void *)&pkts[1]->rearm_data, rearm1);
+       vst1q_u64((void *)&pkts[2]->rearm_data, rearm2);
+       vst1q_u64((void *)&pkts[3]->rearm_data, rearm3);
+}
+
+/**
+ * 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 mlx5_rxq_data *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;
+       const uint16x4_t ownership = vdup_n_u16(!(rxq->cq_ci & (q_mask + 1)));
+       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 = {
+               28, 29,         /* hdr_type_etc */
+                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 */
+       };
+       /* Mask to generate a vector having packet_type/ol_flags. */
+       const uint8x16_t ptype_shuf_m = {
+               48, 49, 50, -1, /* 4th CQE */
+               32, 33, 34, -1, /* 3rd CQE */
+               16, 17, 18, -1, /* 2nd CQE */
+                0,  1,  2, -1  /* 1st CQE */
+       };
+       /* Mask to generate a vector having flow tags. */
+       const uint8x16_t ftag_shuf_m = {
+               60, 61, 62, -1, /* 4th CQE */
+               44, 45, 46, -1, /* 3rd CQE */
+               28, 29, 30, -1, /* 2nd CQE */
+               12, 13, 14, -1  /* 1st CQE */
+       };
+       const uint16x8_t crc_adj = {
+               0, 0, rxq->crc_present * ETHER_CRC_LEN, 0, 0, 0, 0, 0
+       };
+       const uint32x4_t flow_mark_adj = { 0, 0, 0, rxq->mark * (-1) };
+
+       assert(rxq->sges_n == 0);
+       assert(rxq->cqe_n == rxq->elts_n);
+       cq = &(*rxq->cqes)[cq_idx];
+       rte_prefetch_non_temporal(cq);
+       rte_prefetch_non_temporal(cq + 1);
+       rte_prefetch_non_temporal(cq + 2);
+       rte_prefetch_non_temporal(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)
+               mlx5_rx_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];
+       pkts_n = RTE_MIN(pkts_n - rcvd_pkt,
+                        (uint16_t)(rxq->rq_ci - rxq->cq_ci));
+       /* Not to overflow pkts/elts array. */
+       pkts_n = RTE_ALIGN_FLOOR(pkts_n, 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);
+       /*
+        * Note that vectors have reverse order - {v3, v2, v1, v0}, because
+        * there's no instruction to count trailing zeros. __builtin_clzl() is
+        * used instead.
+        *
+        * 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;
+               uint32x4_t ptype_info, flow_tag;
+               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) {
+                       unsigned int next = pos + MLX5_VPMD_DESCS_PER_LOOP;
+                       rte_prefetch_non_temporal(&cq[next]);
+                       rte_prefetch_non_temporal(&cq[next + 1]);
+                       rte_prefetch_non_temporal(&cq[next + 2]);
+                       rte_prefetch_non_temporal(&cq[next + 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.4 (CQE 3) adjust CRC length. */
+               "sub v23.8h, v23.8h, %[crc_adj].8h \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.4 (CQE 2) adjust CRC length. */
+               "sub v22.8h, v22.8h, %[crc_adj].8h \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.4 (CQE 1) adjust CRC length. */
+               "sub v21.8h, v21.8h, %[crc_adj].8h \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"
+               /* B.4 (CQE 0) adjust CRC length. */
+               "sub v20.8h, v20.8h, %[crc_adj].8h \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 flow mark. */
+               "add v15.4s, v15.4s, %[flow_mark_adj].4s \n\t"
+               /* C.3 (CQE 3) fill in mbuf - rx_descriptor_fields1. */
+               "st1 {v15.2d}, [%[e3]] \n\t"
+               /* C.2 (CQE 2) adjust flow mark. */
+               "add v14.4s, v14.4s, %[flow_mark_adj].4s \n\t"
+               /* C.3 (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 flow mark. */
+               "add v13.4s, v13.4s, %[flow_mark_adj].4s \n\t"
+               /* C.3 (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
+               /* Extract ptype_info. */
+               "tbl %[ptype_info].16b, {v20.16b - v23.16b}, %[ptype_shuf_m].16b \n\t"
+               /* Extract flow_tag. */
+               "tbl %[flow_tag].16b, {v20.16b - v23.16b}, %[ftag_shuf_m].16b \n\t"
+               /* A.2 copy mbuf pointers. */
+               "st1 {v24.2d - v25.2d}, [%[pkts_p]] \n\t"
+               /* C.2 (CQE 0) adjust flow mark. */
+               "add v12.4s, v12.4s, %[flow_mark_adj].4s \n\t"
+               /* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
+               "st1 {v12.2d}, [%[e0]] \n\t"
+               :[op_own]"=&w"(op_own),
+                [byte_cnt]"=&w"(byte_cnt),
+                [ptype_info]"=&w"(ptype_info),
+                [flow_tag]"=&w"(flow_tag)
+               :[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),
+                [ptype_shuf_m]"w"(ptype_shuf_m),
+                [ftag_shuf_m]"w"(ftag_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);
+               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.4 fill in mbuf - rearm_data and packet_type. */
+               rxq_cq_to_ptype_oflags_v(rxq, ptype_info, flow_tag,
+                                        opcode, &elts[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();
+       *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
+       return rcvd_pkt;
+}
+
+#endif /* RTE_PMD_MLX5_RXTX_VEC_NEON_H_ */