net/octeontx2: add Tx vector version
authorNithin Dabilpuram <ndabilpuram@marvell.com>
Sun, 2 Jun 2019 11:37:08 +0000 (17:07 +0530)
committerFerruh Yigit <ferruh.yigit@intel.com>
Thu, 4 Jul 2019 23:52:02 +0000 (01:52 +0200)
Add vector version of packet transmit function.

Signed-off-by: Nithin Dabilpuram <ndabilpuram@marvell.com>
Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
drivers/net/octeontx2/otx2_tx.c

index 0ac5ea6..6bce551 100644 (file)
@@ -80,6 +80,859 @@ nix_xmit_pkts_mseg(void *tx_queue, struct rte_mbuf **tx_pkts,
        return pkts;
 }
 
+#if defined(RTE_ARCH_ARM64)
+
+#define NIX_DESCS_PER_LOOP     4
+static __rte_always_inline uint16_t
+nix_xmit_pkts_vector(void *tx_queue, struct rte_mbuf **tx_pkts,
+                    uint16_t pkts, const uint16_t flags)
+{
+       uint64x2_t dataoff_iova0, dataoff_iova1, dataoff_iova2, dataoff_iova3;
+       uint64x2_t len_olflags0, len_olflags1, len_olflags2, len_olflags3;
+       uint64_t *mbuf0, *mbuf1, *mbuf2, *mbuf3;
+       uint64x2_t senddesc01_w0, senddesc23_w0;
+       uint64x2_t senddesc01_w1, senddesc23_w1;
+       uint64x2_t sgdesc01_w0, sgdesc23_w0;
+       uint64x2_t sgdesc01_w1, sgdesc23_w1;
+       struct otx2_eth_txq *txq = tx_queue;
+       uint64_t *lmt_addr = txq->lmt_addr;
+       rte_iova_t io_addr = txq->io_addr;
+       uint64x2_t ltypes01, ltypes23;
+       uint64x2_t xtmp128, ytmp128;
+       uint64x2_t xmask01, xmask23;
+       uint64x2_t mbuf01, mbuf23;
+       uint64x2_t cmd00, cmd01;
+       uint64x2_t cmd10, cmd11;
+       uint64x2_t cmd20, cmd21;
+       uint64x2_t cmd30, cmd31;
+       uint64_t lmt_status, i;
+
+       pkts = RTE_ALIGN_FLOOR(pkts, NIX_DESCS_PER_LOOP);
+
+       NIX_XMIT_FC_OR_RETURN(txq, pkts);
+
+       /* Reduce the cached count */
+       txq->fc_cache_pkts -= pkts;
+
+       /* Lets commit any changes in the packet */
+       rte_cio_wmb();
+
+       senddesc01_w0 = vld1q_dup_u64(&txq->cmd[0]);
+       senddesc23_w0 = senddesc01_w0;
+       senddesc01_w1 = vdupq_n_u64(0);
+       senddesc23_w1 = senddesc01_w1;
+       sgdesc01_w0 = vld1q_dup_u64(&txq->cmd[2]);
+       sgdesc23_w0 = sgdesc01_w0;
+
+       for (i = 0; i < pkts; i += NIX_DESCS_PER_LOOP) {
+               mbuf01 = vld1q_u64((uint64_t *)tx_pkts);
+               mbuf23 = vld1q_u64((uint64_t *)(tx_pkts + 2));
+
+               /* Clear lower 32bit of SEND_HDR_W0 and SEND_SG_W0 */
+               senddesc01_w0 = vbicq_u64(senddesc01_w0,
+                                         vdupq_n_u64(0xFFFFFFFF));
+               sgdesc01_w0 = vbicq_u64(sgdesc01_w0,
+                                       vdupq_n_u64(0xFFFFFFFF));
+
+               senddesc23_w0 = senddesc01_w0;
+               sgdesc23_w0 = sgdesc01_w0;
+
+               tx_pkts = tx_pkts + NIX_DESCS_PER_LOOP;
+
+               /* Move mbufs to iova */
+               mbuf0 = (uint64_t *)vgetq_lane_u64(mbuf01, 0);
+               mbuf1 = (uint64_t *)vgetq_lane_u64(mbuf01, 1);
+               mbuf2 = (uint64_t *)vgetq_lane_u64(mbuf23, 0);
+               mbuf3 = (uint64_t *)vgetq_lane_u64(mbuf23, 1);
+
+               mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
+                                    offsetof(struct rte_mbuf, buf_iova));
+               mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
+                                    offsetof(struct rte_mbuf, buf_iova));
+               mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
+                                    offsetof(struct rte_mbuf, buf_iova));
+               mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
+                                    offsetof(struct rte_mbuf, buf_iova));
+               /*
+                * Get mbuf's, olflags, iova, pktlen, dataoff
+                * dataoff_iovaX.D[0] = iova,
+                * dataoff_iovaX.D[1](15:0) = mbuf->dataoff
+                * len_olflagsX.D[0] = ol_flags,
+                * len_olflagsX.D[1](63:32) = mbuf->pkt_len
+                */
+               dataoff_iova0  = vld1q_u64(mbuf0);
+               len_olflags0 = vld1q_u64(mbuf0 + 2);
+               dataoff_iova1  = vld1q_u64(mbuf1);
+               len_olflags1 = vld1q_u64(mbuf1 + 2);
+               dataoff_iova2  = vld1q_u64(mbuf2);
+               len_olflags2 = vld1q_u64(mbuf2 + 2);
+               dataoff_iova3  = vld1q_u64(mbuf3);
+               len_olflags3 = vld1q_u64(mbuf3 + 2);
+
+               if (flags & NIX_TX_OFFLOAD_MBUF_NOFF_F) {
+                       struct rte_mbuf *mbuf;
+                       /* Set don't free bit if reference count > 1 */
+                       xmask01 = vdupq_n_u64(0);
+                       xmask23 = xmask01;
+
+                       mbuf = (struct rte_mbuf *)((uintptr_t)mbuf0 -
+                               offsetof(struct rte_mbuf, buf_iova));
+
+                       if (rte_pktmbuf_prefree_seg(mbuf) == NULL)
+                               vsetq_lane_u64(0x80000, xmask01, 0);
+                       else
+                               __mempool_check_cookies(mbuf->pool,
+                                                       (void **)&mbuf,
+                                                       1, 0);
+
+                       mbuf = (struct rte_mbuf *)((uintptr_t)mbuf1 -
+                               offsetof(struct rte_mbuf, buf_iova));
+                       if (rte_pktmbuf_prefree_seg(mbuf) == NULL)
+                               vsetq_lane_u64(0x80000, xmask01, 1);
+                       else
+                               __mempool_check_cookies(mbuf->pool,
+                                                       (void **)&mbuf,
+                                                       1, 0);
+
+                       mbuf = (struct rte_mbuf *)((uintptr_t)mbuf2 -
+                               offsetof(struct rte_mbuf, buf_iova));
+                       if (rte_pktmbuf_prefree_seg(mbuf) == NULL)
+                               vsetq_lane_u64(0x80000, xmask23, 0);
+                       else
+                               __mempool_check_cookies(mbuf->pool,
+                                                       (void **)&mbuf,
+                                                       1, 0);
+
+                       mbuf = (struct rte_mbuf *)((uintptr_t)mbuf3 -
+                               offsetof(struct rte_mbuf, buf_iova));
+                       if (rte_pktmbuf_prefree_seg(mbuf) == NULL)
+                               vsetq_lane_u64(0x80000, xmask23, 1);
+                       else
+                               __mempool_check_cookies(mbuf->pool,
+                                                       (void **)&mbuf,
+                                                       1, 0);
+                       senddesc01_w0 = vorrq_u64(senddesc01_w0, xmask01);
+                       senddesc23_w0 = vorrq_u64(senddesc23_w0, xmask23);
+               } else {
+                       struct rte_mbuf *mbuf;
+                       /* Mark mempool object as "put" since
+                        * it is freed by NIX
+                        */
+                       mbuf = (struct rte_mbuf *)((uintptr_t)mbuf0 -
+                               offsetof(struct rte_mbuf, buf_iova));
+                       __mempool_check_cookies(mbuf->pool, (void **)&mbuf,
+                                               1, 0);
+
+                       mbuf = (struct rte_mbuf *)((uintptr_t)mbuf1 -
+                               offsetof(struct rte_mbuf, buf_iova));
+                       __mempool_check_cookies(mbuf->pool, (void **)&mbuf,
+                                               1, 0);
+
+                       mbuf = (struct rte_mbuf *)((uintptr_t)mbuf2 -
+                               offsetof(struct rte_mbuf, buf_iova));
+                       __mempool_check_cookies(mbuf->pool, (void **)&mbuf,
+                                               1, 0);
+
+                       mbuf = (struct rte_mbuf *)((uintptr_t)mbuf3 -
+                               offsetof(struct rte_mbuf, buf_iova));
+                       __mempool_check_cookies(mbuf->pool, (void **)&mbuf,
+                                               1, 0);
+                       RTE_SET_USED(mbuf);
+               }
+
+               /* Move mbufs to point pool */
+               mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
+                        offsetof(struct rte_mbuf, pool) -
+                        offsetof(struct rte_mbuf, buf_iova));
+               mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
+                        offsetof(struct rte_mbuf, pool) -
+                        offsetof(struct rte_mbuf, buf_iova));
+               mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
+                        offsetof(struct rte_mbuf, pool) -
+                        offsetof(struct rte_mbuf, buf_iova));
+               mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
+                        offsetof(struct rte_mbuf, pool) -
+                        offsetof(struct rte_mbuf, buf_iova));
+
+               if (flags &
+                   (NIX_TX_OFFLOAD_OL3_OL4_CSUM_F |
+                    NIX_TX_OFFLOAD_L3_L4_CSUM_F)) {
+                       /* Get tx_offload for ol2, ol3, l2, l3 lengths */
+                       /*
+                        * E(8):OL2_LEN(7):OL3_LEN(9):E(24):L3_LEN(9):L2_LEN(7)
+                        * E(8):OL2_LEN(7):OL3_LEN(9):E(24):L3_LEN(9):L2_LEN(7)
+                        */
+
+                       asm volatile ("LD1 {%[a].D}[0],[%[in]]\n\t" :
+                                     [a]"+w"(senddesc01_w1) :
+                                     [in]"r"(mbuf0 + 2) : "memory");
+
+                       asm volatile ("LD1 {%[a].D}[1],[%[in]]\n\t" :
+                                     [a]"+w"(senddesc01_w1) :
+                                     [in]"r"(mbuf1 + 2) : "memory");
+
+                       asm volatile ("LD1 {%[b].D}[0],[%[in]]\n\t" :
+                                     [b]"+w"(senddesc23_w1) :
+                                     [in]"r"(mbuf2 + 2) : "memory");
+
+                       asm volatile ("LD1 {%[b].D}[1],[%[in]]\n\t" :
+                                     [b]"+w"(senddesc23_w1) :
+                                     [in]"r"(mbuf3 + 2) : "memory");
+
+                       /* Get pool pointer alone */
+                       mbuf0 = (uint64_t *)*mbuf0;
+                       mbuf1 = (uint64_t *)*mbuf1;
+                       mbuf2 = (uint64_t *)*mbuf2;
+                       mbuf3 = (uint64_t *)*mbuf3;
+               } else {
+                       /* Get pool pointer alone */
+                       mbuf0 = (uint64_t *)*mbuf0;
+                       mbuf1 = (uint64_t *)*mbuf1;
+                       mbuf2 = (uint64_t *)*mbuf2;
+                       mbuf3 = (uint64_t *)*mbuf3;
+               }
+
+               const uint8x16_t shuf_mask2 = {
+                       0x4, 0x5, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+                       0xc, 0xd, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+               };
+               xtmp128 = vzip2q_u64(len_olflags0, len_olflags1);
+               ytmp128 = vzip2q_u64(len_olflags2, len_olflags3);
+
+               /* Clear dataoff_iovaX.D[1] bits other than dataoff(15:0) */
+               const uint64x2_t and_mask0 = {
+                       0xFFFFFFFFFFFFFFFF,
+                       0x000000000000FFFF,
+               };
+
+               dataoff_iova0 = vandq_u64(dataoff_iova0, and_mask0);
+               dataoff_iova1 = vandq_u64(dataoff_iova1, and_mask0);
+               dataoff_iova2 = vandq_u64(dataoff_iova2, and_mask0);
+               dataoff_iova3 = vandq_u64(dataoff_iova3, and_mask0);
+
+               /*
+                * Pick only 16 bits of pktlen preset at bits 63:32
+                * and place them at bits 15:0.
+                */
+               xtmp128 = vqtbl1q_u8(xtmp128, shuf_mask2);
+               ytmp128 = vqtbl1q_u8(ytmp128, shuf_mask2);
+
+               /* Add pairwise to get dataoff + iova in sgdesc_w1 */
+               sgdesc01_w1 = vpaddq_u64(dataoff_iova0, dataoff_iova1);
+               sgdesc23_w1 = vpaddq_u64(dataoff_iova2, dataoff_iova3);
+
+               /* Orr both sgdesc_w0 and senddesc_w0 with 16 bits of
+                * pktlen at 15:0 position.
+                */
+               sgdesc01_w0 = vorrq_u64(sgdesc01_w0, xtmp128);
+               sgdesc23_w0 = vorrq_u64(sgdesc23_w0, ytmp128);
+               senddesc01_w0 = vorrq_u64(senddesc01_w0, xtmp128);
+               senddesc23_w0 = vorrq_u64(senddesc23_w0, ytmp128);
+
+               if ((flags & NIX_TX_OFFLOAD_L3_L4_CSUM_F) &&
+                   !(flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F)) {
+                       /*
+                        * Lookup table to translate ol_flags to
+                        * il3/il4 types. But we still use ol3/ol4 types in
+                        * senddesc_w1 as only one header processing is enabled.
+                        */
+                       const uint8x16_t tbl = {
+                               /* [0-15] = il4type:il3type */
+                               0x04, /* none (IPv6 assumed) */
+                               0x14, /* PKT_TX_TCP_CKSUM (IPv6 assumed) */
+                               0x24, /* PKT_TX_SCTP_CKSUM (IPv6 assumed) */
+                               0x34, /* PKT_TX_UDP_CKSUM (IPv6 assumed) */
+                               0x03, /* PKT_TX_IP_CKSUM */
+                               0x13, /* PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM */
+                               0x23, /* PKT_TX_IP_CKSUM | PKT_TX_SCTP_CKSUM */
+                               0x33, /* PKT_TX_IP_CKSUM | PKT_TX_UDP_CKSUM */
+                               0x02, /* PKT_TX_IPV4  */
+                               0x12, /* PKT_TX_IPV4 | PKT_TX_TCP_CKSUM */
+                               0x22, /* PKT_TX_IPV4 | PKT_TX_SCTP_CKSUM */
+                               0x32, /* PKT_TX_IPV4 | PKT_TX_UDP_CKSUM */
+                               0x03, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM */
+                               0x13, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
+                                      * PKT_TX_TCP_CKSUM
+                                      */
+                               0x23, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
+                                      * PKT_TX_SCTP_CKSUM
+                                      */
+                               0x33, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
+                                      * PKT_TX_UDP_CKSUM
+                                      */
+                       };
+
+                       /* Extract olflags to translate to iltypes */
+                       xtmp128 = vzip1q_u64(len_olflags0, len_olflags1);
+                       ytmp128 = vzip1q_u64(len_olflags2, len_olflags3);
+
+                       /*
+                        * E(47):L3_LEN(9):L2_LEN(7+z)
+                        * E(47):L3_LEN(9):L2_LEN(7+z)
+                        */
+                       senddesc01_w1 = vshlq_n_u64(senddesc01_w1, 1);
+                       senddesc23_w1 = vshlq_n_u64(senddesc23_w1, 1);
+
+                       /* Move OLFLAGS bits 55:52 to 51:48
+                        * with zeros preprended on the byte and rest
+                        * don't care
+                        */
+                       xtmp128 = vshrq_n_u8(xtmp128, 4);
+                       ytmp128 = vshrq_n_u8(ytmp128, 4);
+                       /*
+                        * E(48):L3_LEN(8):L2_LEN(z+7)
+                        * E(48):L3_LEN(8):L2_LEN(z+7)
+                        */
+                       const int8x16_t tshft3 = {
+                               -1, 0, 8, 8, 8, 8, 8, 8,
+                               -1, 0, 8, 8, 8, 8, 8, 8,
+                       };
+
+                       senddesc01_w1 = vshlq_u8(senddesc01_w1, tshft3);
+                       senddesc23_w1 = vshlq_u8(senddesc23_w1, tshft3);
+
+                       /* Do the lookup */
+                       ltypes01 = vqtbl1q_u8(tbl, xtmp128);
+                       ltypes23 = vqtbl1q_u8(tbl, ytmp128);
+
+                       /* Just use ld1q to retrieve aura
+                        * when we don't need tx_offload
+                        */
+                       mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
+                                       offsetof(struct rte_mempool, pool_id));
+
+                       /* Pick only relevant fields i.e Bit 48:55 of iltype
+                        * and place it in ol3/ol4type of senddesc_w1
+                        */
+                       const uint8x16_t shuf_mask0 = {
+                               0xFF, 0xFF, 0xFF, 0xFF, 0x6, 0xFF, 0xFF, 0xFF,
+                               0xFF, 0xFF, 0xFF, 0xFF, 0xE, 0xFF, 0xFF, 0xFF,
+                       };
+
+                       ltypes01 = vqtbl1q_u8(ltypes01, shuf_mask0);
+                       ltypes23 = vqtbl1q_u8(ltypes23, shuf_mask0);
+
+                       /* Prepare ol4ptr, ol3ptr from ol3len, ol2len.
+                        * a [E(32):E(16):OL3(8):OL2(8)]
+                        * a = a + (a << 8)
+                        * a [E(32):E(16):(OL3+OL2):OL2]
+                        * => E(32):E(16)::OL4PTR(8):OL3PTR(8)
+                        */
+                       senddesc01_w1 = vaddq_u8(senddesc01_w1,
+                                                vshlq_n_u16(senddesc01_w1, 8));
+                       senddesc23_w1 = vaddq_u8(senddesc23_w1,
+                                                vshlq_n_u16(senddesc23_w1, 8));
+
+                       /* Create first half of 4W cmd for 4 mbufs (sgdesc) */
+                       cmd01 = vzip1q_u64(sgdesc01_w0, sgdesc01_w1);
+                       cmd11 = vzip2q_u64(sgdesc01_w0, sgdesc01_w1);
+                       cmd21 = vzip1q_u64(sgdesc23_w0, sgdesc23_w1);
+                       cmd31 = vzip2q_u64(sgdesc23_w0, sgdesc23_w1);
+
+                       xmask01 = vdupq_n_u64(0);
+                       xmask23 = xmask01;
+                       asm volatile ("LD1 {%[a].H}[0],[%[in]]\n\t" :
+                               [a]"+w"(xmask01) : [in]"r"(mbuf0) : "memory");
+
+                       asm volatile ("LD1 {%[a].H}[4],[%[in]]\n\t" :
+                                [a]"+w"(xmask01) : [in]"r"(mbuf1) : "memory");
+
+                       asm volatile ("LD1 {%[b].H}[0],[%[in]]\n\t" :
+                                [b]"+w"(xmask23) : [in]"r"(mbuf2) : "memory");
+
+                       asm volatile ("LD1 {%[b].H}[4],[%[in]]\n\t" :
+                                [b]"+w"(xmask23) : [in]"r"(mbuf3) : "memory");
+                       xmask01 = vshlq_n_u64(xmask01, 20);
+                       xmask23 = vshlq_n_u64(xmask23, 20);
+
+                       senddesc01_w0 = vorrq_u64(senddesc01_w0, xmask01);
+                       senddesc23_w0 = vorrq_u64(senddesc23_w0, xmask23);
+                       /* Move ltypes to senddesc*_w1 */
+                       senddesc01_w1 = vorrq_u64(senddesc01_w1, ltypes01);
+                       senddesc23_w1 = vorrq_u64(senddesc23_w1, ltypes23);
+
+                       /* Create first half of 4W cmd for 4 mbufs (sendhdr) */
+                       cmd00 = vzip1q_u64(senddesc01_w0, senddesc01_w1);
+                       cmd10 = vzip2q_u64(senddesc01_w0, senddesc01_w1);
+                       cmd20 = vzip1q_u64(senddesc23_w0, senddesc23_w1);
+                       cmd30 = vzip2q_u64(senddesc23_w0, senddesc23_w1);
+
+               } else if (!(flags & NIX_TX_OFFLOAD_L3_L4_CSUM_F) &&
+                          (flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F)) {
+                       /*
+                        * Lookup table to translate ol_flags to
+                        * ol3/ol4 types.
+                        */
+
+                       const uint8x16_t tbl = {
+                               /* [0-15] = ol4type:ol3type */
+                               0x00, /* none */
+                               0x03, /* OUTER_IP_CKSUM */
+                               0x02, /* OUTER_IPV4 */
+                               0x03, /* OUTER_IPV4 | OUTER_IP_CKSUM */
+                               0x04, /* OUTER_IPV6 */
+                               0x00, /* OUTER_IPV6 | OUTER_IP_CKSUM */
+                               0x00, /* OUTER_IPV6 | OUTER_IPV4 */
+                               0x00, /* OUTER_IPV6 | OUTER_IPV4 |
+                                      * OUTER_IP_CKSUM
+                                      */
+                               0x00, /* OUTER_UDP_CKSUM */
+                               0x33, /* OUTER_UDP_CKSUM | OUTER_IP_CKSUM */
+                               0x32, /* OUTER_UDP_CKSUM | OUTER_IPV4 */
+                               0x33, /* OUTER_UDP_CKSUM | OUTER_IPV4 |
+                                      * OUTER_IP_CKSUM
+                                      */
+                               0x34, /* OUTER_UDP_CKSUM | OUTER_IPV6 */
+                               0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
+                                      * OUTER_IP_CKSUM
+                                      */
+                               0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
+                                      * OUTER_IPV4
+                                      */
+                               0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
+                                      * OUTER_IPV4 | OUTER_IP_CKSUM
+                                      */
+                       };
+
+                       /* Extract olflags to translate to iltypes */
+                       xtmp128 = vzip1q_u64(len_olflags0, len_olflags1);
+                       ytmp128 = vzip1q_u64(len_olflags2, len_olflags3);
+
+                       /*
+                        * E(47):OL3_LEN(9):OL2_LEN(7+z)
+                        * E(47):OL3_LEN(9):OL2_LEN(7+z)
+                        */
+                       const uint8x16_t shuf_mask5 = {
+                               0x6, 0x5, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+                               0xE, 0xD, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+                       };
+                       senddesc01_w1 = vqtbl1q_u8(senddesc01_w1, shuf_mask5);
+                       senddesc23_w1 = vqtbl1q_u8(senddesc23_w1, shuf_mask5);
+
+                       /* Extract outer ol flags only */
+                       const uint64x2_t o_cksum_mask = {
+                               0x1C00020000000000,
+                               0x1C00020000000000,
+                       };
+
+                       xtmp128 = vandq_u64(xtmp128, o_cksum_mask);
+                       ytmp128 = vandq_u64(ytmp128, o_cksum_mask);
+
+                       /* Extract OUTER_UDP_CKSUM bit 41 and
+                        * move it to bit 61
+                        */
+
+                       xtmp128 = xtmp128 | vshlq_n_u64(xtmp128, 20);
+                       ytmp128 = ytmp128 | vshlq_n_u64(ytmp128, 20);
+
+                       /* Shift oltype by 2 to start nibble from BIT(56)
+                        * instead of BIT(58)
+                        */
+                       xtmp128 = vshrq_n_u8(xtmp128, 2);
+                       ytmp128 = vshrq_n_u8(ytmp128, 2);
+                       /*
+                        * E(48):L3_LEN(8):L2_LEN(z+7)
+                        * E(48):L3_LEN(8):L2_LEN(z+7)
+                        */
+                       const int8x16_t tshft3 = {
+                               -1, 0, 8, 8, 8, 8, 8, 8,
+                               -1, 0, 8, 8, 8, 8, 8, 8,
+                       };
+
+                       senddesc01_w1 = vshlq_u8(senddesc01_w1, tshft3);
+                       senddesc23_w1 = vshlq_u8(senddesc23_w1, tshft3);
+
+                       /* Do the lookup */
+                       ltypes01 = vqtbl1q_u8(tbl, xtmp128);
+                       ltypes23 = vqtbl1q_u8(tbl, ytmp128);
+
+                       /* Just use ld1q to retrieve aura
+                        * when we don't need tx_offload
+                        */
+                       mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
+                                       offsetof(struct rte_mempool, pool_id));
+
+                       /* Pick only relevant fields i.e Bit 56:63 of oltype
+                        * and place it in ol3/ol4type of senddesc_w1
+                        */
+                       const uint8x16_t shuf_mask0 = {
+                               0xFF, 0xFF, 0xFF, 0xFF, 0x7, 0xFF, 0xFF, 0xFF,
+                               0xFF, 0xFF, 0xFF, 0xFF, 0xF, 0xFF, 0xFF, 0xFF,
+                       };
+
+                       ltypes01 = vqtbl1q_u8(ltypes01, shuf_mask0);
+                       ltypes23 = vqtbl1q_u8(ltypes23, shuf_mask0);
+
+                       /* Prepare ol4ptr, ol3ptr from ol3len, ol2len.
+                        * a [E(32):E(16):OL3(8):OL2(8)]
+                        * a = a + (a << 8)
+                        * a [E(32):E(16):(OL3+OL2):OL2]
+                        * => E(32):E(16)::OL4PTR(8):OL3PTR(8)
+                        */
+                       senddesc01_w1 = vaddq_u8(senddesc01_w1,
+                                                vshlq_n_u16(senddesc01_w1, 8));
+                       senddesc23_w1 = vaddq_u8(senddesc23_w1,
+                                                vshlq_n_u16(senddesc23_w1, 8));
+
+                       /* Create second half of 4W cmd for 4 mbufs (sgdesc) */
+                       cmd01 = vzip1q_u64(sgdesc01_w0, sgdesc01_w1);
+                       cmd11 = vzip2q_u64(sgdesc01_w0, sgdesc01_w1);
+                       cmd21 = vzip1q_u64(sgdesc23_w0, sgdesc23_w1);
+                       cmd31 = vzip2q_u64(sgdesc23_w0, sgdesc23_w1);
+
+                       xmask01 = vdupq_n_u64(0);
+                       xmask23 = xmask01;
+                       asm volatile ("LD1 {%[a].H}[0],[%[in]]\n\t" :
+                                [a]"+w"(xmask01) : [in]"r"(mbuf0) : "memory");
+
+                       asm volatile ("LD1 {%[a].H}[4],[%[in]]\n\t" :
+                                [a]"+w"(xmask01) : [in]"r"(mbuf1) : "memory");
+
+                       asm volatile ("LD1 {%[b].H}[0],[%[in]]\n\t" :
+                                [b]"+w"(xmask23) : [in]"r"(mbuf2) : "memory");
+
+                       asm volatile ("LD1 {%[b].H}[4],[%[in]]\n\t" :
+                                [b]"+w"(xmask23) : [in]"r"(mbuf3) : "memory");
+                       xmask01 = vshlq_n_u64(xmask01, 20);
+                       xmask23 = vshlq_n_u64(xmask23, 20);
+
+                       senddesc01_w0 = vorrq_u64(senddesc01_w0, xmask01);
+                       senddesc23_w0 = vorrq_u64(senddesc23_w0, xmask23);
+                       /* Move ltypes to senddesc*_w1 */
+                       senddesc01_w1 = vorrq_u64(senddesc01_w1, ltypes01);
+                       senddesc23_w1 = vorrq_u64(senddesc23_w1, ltypes23);
+
+                       /* Create first half of 4W cmd for 4 mbufs (sendhdr) */
+                       cmd00 = vzip1q_u64(senddesc01_w0, senddesc01_w1);
+                       cmd10 = vzip2q_u64(senddesc01_w0, senddesc01_w1);
+                       cmd20 = vzip1q_u64(senddesc23_w0, senddesc23_w1);
+                       cmd30 = vzip2q_u64(senddesc23_w0, senddesc23_w1);
+
+               } else if ((flags & NIX_TX_OFFLOAD_L3_L4_CSUM_F) &&
+                          (flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F)) {
+                       /* Lookup table to translate ol_flags to
+                        * ol4type, ol3type, il4type, il3type of senddesc_w1
+                        */
+                       const uint8x16x2_t tbl = {
+                       {
+                               {
+                                       /* [0-15] = il4type:il3type */
+                                       0x04, /* none (IPv6) */
+                                       0x14, /* PKT_TX_TCP_CKSUM (IPv6) */
+                                       0x24, /* PKT_TX_SCTP_CKSUM (IPv6) */
+                                       0x34, /* PKT_TX_UDP_CKSUM (IPv6) */
+                                       0x03, /* PKT_TX_IP_CKSUM */
+                                       0x13, /* PKT_TX_IP_CKSUM |
+                                              * PKT_TX_TCP_CKSUM
+                                              */
+                                       0x23, /* PKT_TX_IP_CKSUM |
+                                              * PKT_TX_SCTP_CKSUM
+                                              */
+                                       0x33, /* PKT_TX_IP_CKSUM |
+                                              * PKT_TX_UDP_CKSUM
+                                              */
+                                       0x02, /* PKT_TX_IPV4 */
+                                       0x12, /* PKT_TX_IPV4 |
+                                              * PKT_TX_TCP_CKSUM
+                                              */
+                                       0x22, /* PKT_TX_IPV4 |
+                                              * PKT_TX_SCTP_CKSUM
+                                              */
+                                       0x32, /* PKT_TX_IPV4 |
+                                              * PKT_TX_UDP_CKSUM
+                                              */
+                                       0x03, /* PKT_TX_IPV4 |
+                                              * PKT_TX_IP_CKSUM
+                                              */
+                                       0x13, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
+                                              * PKT_TX_TCP_CKSUM
+                                              */
+                                       0x23, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
+                                              * PKT_TX_SCTP_CKSUM
+                                              */
+                                       0x33, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
+                                              * PKT_TX_UDP_CKSUM
+                                              */
+                               },
+
+                               {
+                                       /* [16-31] = ol4type:ol3type */
+                                       0x00, /* none */
+                                       0x03, /* OUTER_IP_CKSUM */
+                                       0x02, /* OUTER_IPV4 */
+                                       0x03, /* OUTER_IPV4 | OUTER_IP_CKSUM */
+                                       0x04, /* OUTER_IPV6 */
+                                       0x00, /* OUTER_IPV6 | OUTER_IP_CKSUM */
+                                       0x00, /* OUTER_IPV6 | OUTER_IPV4 */
+                                       0x00, /* OUTER_IPV6 | OUTER_IPV4 |
+                                              * OUTER_IP_CKSUM
+                                              */
+                                       0x00, /* OUTER_UDP_CKSUM */
+                                       0x33, /* OUTER_UDP_CKSUM |
+                                              * OUTER_IP_CKSUM
+                                              */
+                                       0x32, /* OUTER_UDP_CKSUM |
+                                              * OUTER_IPV4
+                                              */
+                                       0x33, /* OUTER_UDP_CKSUM |
+                                              * OUTER_IPV4 | OUTER_IP_CKSUM
+                                              */
+                                       0x34, /* OUTER_UDP_CKSUM |
+                                              * OUTER_IPV6
+                                              */
+                                       0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
+                                              * OUTER_IP_CKSUM
+                                              */
+                                       0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
+                                              * OUTER_IPV4
+                                              */
+                                       0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
+                                              * OUTER_IPV4 | OUTER_IP_CKSUM
+                                              */
+                               },
+                       }
+                       };
+
+                       /* Extract olflags to translate to oltype & iltype */
+                       xtmp128 = vzip1q_u64(len_olflags0, len_olflags1);
+                       ytmp128 = vzip1q_u64(len_olflags2, len_olflags3);
+
+                       /*
+                        * E(8):OL2_LN(7):OL3_LN(9):E(23):L3_LN(9):L2_LN(7+z)
+                        * E(8):OL2_LN(7):OL3_LN(9):E(23):L3_LN(9):L2_LN(7+z)
+                        */
+                       const uint32x4_t tshft_4 = {
+                               1, 0,
+                               1, 0,
+                       };
+                       senddesc01_w1 = vshlq_u32(senddesc01_w1, tshft_4);
+                       senddesc23_w1 = vshlq_u32(senddesc23_w1, tshft_4);
+
+                       /*
+                        * E(32):L3_LEN(8):L2_LEN(7+Z):OL3_LEN(8):OL2_LEN(7+Z)
+                        * E(32):L3_LEN(8):L2_LEN(7+Z):OL3_LEN(8):OL2_LEN(7+Z)
+                        */
+                       const uint8x16_t shuf_mask5 = {
+                               0x6, 0x5, 0x0, 0x1, 0xFF, 0xFF, 0xFF, 0xFF,
+                               0xE, 0xD, 0x8, 0x9, 0xFF, 0xFF, 0xFF, 0xFF,
+                       };
+                       senddesc01_w1 = vqtbl1q_u8(senddesc01_w1, shuf_mask5);
+                       senddesc23_w1 = vqtbl1q_u8(senddesc23_w1, shuf_mask5);
+
+                       /* Extract outer and inner header ol_flags */
+                       const uint64x2_t oi_cksum_mask = {
+                               0x1CF0020000000000,
+                               0x1CF0020000000000,
+                       };
+
+                       xtmp128 = vandq_u64(xtmp128, oi_cksum_mask);
+                       ytmp128 = vandq_u64(ytmp128, oi_cksum_mask);
+
+                       /* Extract OUTER_UDP_CKSUM bit 41 and
+                        * move it to bit 61
+                        */
+
+                       xtmp128 = xtmp128 | vshlq_n_u64(xtmp128, 20);
+                       ytmp128 = ytmp128 | vshlq_n_u64(ytmp128, 20);
+
+                       /* Shift right oltype by 2 and iltype by 4
+                        * to start oltype nibble from BIT(58)
+                        * instead of BIT(56) and iltype nibble from BIT(48)
+                        * instead of BIT(52).
+                        */
+                       const int8x16_t tshft5 = {
+                               8, 8, 8, 8, 8, 8, -4, -2,
+                               8, 8, 8, 8, 8, 8, -4, -2,
+                       };
+
+                       xtmp128 = vshlq_u8(xtmp128, tshft5);
+                       ytmp128 = vshlq_u8(ytmp128, tshft5);
+                       /*
+                        * E(32):L3_LEN(8):L2_LEN(8):OL3_LEN(8):OL2_LEN(8)
+                        * E(32):L3_LEN(8):L2_LEN(8):OL3_LEN(8):OL2_LEN(8)
+                        */
+                       const int8x16_t tshft3 = {
+                               -1, 0, -1, 0, 0, 0, 0, 0,
+                               -1, 0, -1, 0, 0, 0, 0, 0,
+                       };
+
+                       senddesc01_w1 = vshlq_u8(senddesc01_w1, tshft3);
+                       senddesc23_w1 = vshlq_u8(senddesc23_w1, tshft3);
+
+                       /* Mark Bit(4) of oltype */
+                       const uint64x2_t oi_cksum_mask2 = {
+                               0x1000000000000000,
+                               0x1000000000000000,
+                       };
+
+                       xtmp128 = vorrq_u64(xtmp128, oi_cksum_mask2);
+                       ytmp128 = vorrq_u64(ytmp128, oi_cksum_mask2);
+
+                       /* Do the lookup */
+                       ltypes01 = vqtbl2q_u8(tbl, xtmp128);
+                       ltypes23 = vqtbl2q_u8(tbl, ytmp128);
+
+                       /* Just use ld1q to retrieve aura
+                        * when we don't need tx_offload
+                        */
+                       mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
+                                       offsetof(struct rte_mempool, pool_id));
+
+                       /* Pick only relevant fields i.e Bit 48:55 of iltype and
+                        * Bit 56:63 of oltype and place it in corresponding
+                        * place in senddesc_w1.
+                        */
+                       const uint8x16_t shuf_mask0 = {
+                               0xFF, 0xFF, 0xFF, 0xFF, 0x7, 0x6, 0xFF, 0xFF,
+                               0xFF, 0xFF, 0xFF, 0xFF, 0xF, 0xE, 0xFF, 0xFF,
+                       };
+
+                       ltypes01 = vqtbl1q_u8(ltypes01, shuf_mask0);
+                       ltypes23 = vqtbl1q_u8(ltypes23, shuf_mask0);
+
+                       /* Prepare l4ptr, l3ptr, ol4ptr, ol3ptr from
+                        * l3len, l2len, ol3len, ol2len.
+                        * a [E(32):L3(8):L2(8):OL3(8):OL2(8)]
+                        * a = a + (a << 8)
+                        * a [E:(L3+L2):(L2+OL3):(OL3+OL2):OL2]
+                        * a = a + (a << 16)
+                        * a [E:(L3+L2+OL3+OL2):(L2+OL3+OL2):(OL3+OL2):OL2]
+                        * => E(32):IL4PTR(8):IL3PTR(8):OL4PTR(8):OL3PTR(8)
+                        */
+                       senddesc01_w1 = vaddq_u8(senddesc01_w1,
+                                                vshlq_n_u32(senddesc01_w1, 8));
+                       senddesc23_w1 = vaddq_u8(senddesc23_w1,
+                                                vshlq_n_u32(senddesc23_w1, 8));
+
+                       /* Create second half of 4W cmd for 4 mbufs (sgdesc) */
+                       cmd01 = vzip1q_u64(sgdesc01_w0, sgdesc01_w1);
+                       cmd11 = vzip2q_u64(sgdesc01_w0, sgdesc01_w1);
+                       cmd21 = vzip1q_u64(sgdesc23_w0, sgdesc23_w1);
+                       cmd31 = vzip2q_u64(sgdesc23_w0, sgdesc23_w1);
+
+                       /* Continue preparing l4ptr, l3ptr, ol4ptr, ol3ptr */
+                       senddesc01_w1 = vaddq_u8(senddesc01_w1,
+                                               vshlq_n_u32(senddesc01_w1, 16));
+                       senddesc23_w1 = vaddq_u8(senddesc23_w1,
+                                               vshlq_n_u32(senddesc23_w1, 16));
+
+                       xmask01 = vdupq_n_u64(0);
+                       xmask23 = xmask01;
+                       asm volatile ("LD1 {%[a].H}[0],[%[in]]\n\t" :
+                                [a]"+w"(xmask01) : [in]"r"(mbuf0) : "memory");
+
+                       asm volatile ("LD1 {%[a].H}[4],[%[in]]\n\t" :
+                                [a]"+w"(xmask01) : [in]"r"(mbuf1) : "memory");
+
+                       asm volatile ("LD1 {%[b].H}[0],[%[in]]\n\t" :
+                                [b]"+w"(xmask23) : [in]"r"(mbuf2) : "memory");
+
+                       asm volatile ("LD1 {%[b].H}[4],[%[in]]\n\t" :
+                                [b]"+w"(xmask23) : [in]"r"(mbuf3) : "memory");
+                       xmask01 = vshlq_n_u64(xmask01, 20);
+                       xmask23 = vshlq_n_u64(xmask23, 20);
+
+                       senddesc01_w0 = vorrq_u64(senddesc01_w0, xmask01);
+                       senddesc23_w0 = vorrq_u64(senddesc23_w0, xmask23);
+                       /* Move ltypes to senddesc*_w1 */
+                       senddesc01_w1 = vorrq_u64(senddesc01_w1, ltypes01);
+                       senddesc23_w1 = vorrq_u64(senddesc23_w1, ltypes23);
+
+                       /* Create first half of 4W cmd for 4 mbufs (sendhdr) */
+                       cmd00 = vzip1q_u64(senddesc01_w0, senddesc01_w1);
+                       cmd10 = vzip2q_u64(senddesc01_w0, senddesc01_w1);
+                       cmd20 = vzip1q_u64(senddesc23_w0, senddesc23_w1);
+                       cmd30 = vzip2q_u64(senddesc23_w0, senddesc23_w1);
+               } else {
+                       /* Just use ld1q to retrieve aura
+                        * when we don't need tx_offload
+                        */
+                       mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
+                                       offsetof(struct rte_mempool, pool_id));
+                       xmask01 = vdupq_n_u64(0);
+                       xmask23 = xmask01;
+                       asm volatile ("LD1 {%[a].H}[0],[%[in]]\n\t" :
+                                [a]"+w"(xmask01) : [in]"r"(mbuf0) : "memory");
+
+                       asm volatile ("LD1 {%[a].H}[4],[%[in]]\n\t" :
+                                [a]"+w"(xmask01) : [in]"r"(mbuf1) : "memory");
+
+                       asm volatile ("LD1 {%[b].H}[0],[%[in]]\n\t" :
+                                [b]"+w"(xmask23) : [in]"r"(mbuf2) : "memory");
+
+                       asm volatile ("LD1 {%[b].H}[4],[%[in]]\n\t" :
+                                [b]"+w"(xmask23) : [in]"r"(mbuf3) : "memory");
+                       xmask01 = vshlq_n_u64(xmask01, 20);
+                       xmask23 = vshlq_n_u64(xmask23, 20);
+
+                       senddesc01_w0 = vorrq_u64(senddesc01_w0, xmask01);
+                       senddesc23_w0 = vorrq_u64(senddesc23_w0, xmask23);
+
+                       /* Create 4W cmd for 4 mbufs (sendhdr, sgdesc) */
+                       cmd00 = vzip1q_u64(senddesc01_w0, senddesc01_w1);
+                       cmd01 = vzip1q_u64(sgdesc01_w0, sgdesc01_w1);
+                       cmd10 = vzip2q_u64(senddesc01_w0, senddesc01_w1);
+                       cmd11 = vzip2q_u64(sgdesc01_w0, sgdesc01_w1);
+                       cmd20 = vzip1q_u64(senddesc23_w0, senddesc23_w1);
+                       cmd21 = vzip1q_u64(sgdesc23_w0, sgdesc23_w1);
+                       cmd30 = vzip2q_u64(senddesc23_w0, senddesc23_w1);
+                       cmd31 = vzip2q_u64(sgdesc23_w0, sgdesc23_w1);
+               }
+
+               do {
+                       vst1q_u64(lmt_addr, cmd00);
+                       vst1q_u64(lmt_addr + 2, cmd01);
+                       vst1q_u64(lmt_addr + 4, cmd10);
+                       vst1q_u64(lmt_addr + 6, cmd11);
+                       vst1q_u64(lmt_addr + 8, cmd20);
+                       vst1q_u64(lmt_addr + 10, cmd21);
+                       vst1q_u64(lmt_addr + 12, cmd30);
+                       vst1q_u64(lmt_addr + 14, cmd31);
+                       lmt_status = otx2_lmt_submit(io_addr);
+
+               } while (lmt_status == 0);
+       }
+
+       return pkts;
+}
+
+#else
+static __rte_always_inline uint16_t
+nix_xmit_pkts_vector(void *tx_queue, struct rte_mbuf **tx_pkts,
+                    uint16_t pkts, const uint16_t flags)
+{
+       RTE_SET_USED(tx_queue);
+       RTE_SET_USED(tx_pkts);
+       RTE_SET_USED(pkts);
+       RTE_SET_USED(flags);
+       return 0;
+}
+#endif
+
 #define T(name, f4, f3, f2, f1, f0, sz, flags)                         \
 static uint16_t __rte_noinline __hot                                   \
 otx2_nix_xmit_pkts_ ## name(void *tx_queue,                            \
@@ -107,6 +960,21 @@ otx2_nix_xmit_pkts_mseg_ ## name(void *tx_queue,                   \
 NIX_TX_FASTPATH_MODES
 #undef T
 
+#define T(name, f4, f3, f2, f1, f0, sz, flags)                         \
+static uint16_t __rte_noinline __hot                                   \
+otx2_nix_xmit_pkts_vec_ ## name(void *tx_queue,                                \
+                       struct rte_mbuf **tx_pkts, uint16_t pkts)       \
+{                                                                      \
+       /* VLAN and TSTMP is not supported by vec */                    \
+       if ((flags) & NIX_TX_OFFLOAD_VLAN_QINQ_F ||                     \
+           (flags) & NIX_TX_OFFLOAD_TSTAMP_F)                          \
+               return 0;                                               \
+       return nix_xmit_pkts_vector(tx_queue, tx_pkts, pkts, (flags));  \
+}
+
+NIX_TX_FASTPATH_MODES
+#undef T
+
 static inline void
 pick_tx_func(struct rte_eth_dev *eth_dev,
             const eth_tx_burst_t tx_burst[2][2][2][2][2])
@@ -143,7 +1011,20 @@ NIX_TX_FASTPATH_MODES
 #undef T
        };
 
-       pick_tx_func(eth_dev, nix_eth_tx_burst);
+       const eth_tx_burst_t nix_eth_tx_vec_burst[2][2][2][2][2] = {
+#define T(name, f4, f3, f2, f1, f0, sz, flags)                         \
+       [f4][f3][f2][f1][f0] =  otx2_nix_xmit_pkts_vec_ ## name,
+
+NIX_TX_FASTPATH_MODES
+#undef T
+       };
+
+       if (dev->scalar_ena ||
+           (dev->tx_offload_flags &
+            (NIX_TX_OFFLOAD_VLAN_QINQ_F | NIX_TX_OFFLOAD_TSTAMP_F)))
+               pick_tx_func(eth_dev, nix_eth_tx_burst);
+       else
+               pick_tx_func(eth_dev, nix_eth_tx_vec_burst);
 
        if (dev->tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS)
                pick_tx_func(eth_dev, nix_eth_tx_burst_mseg);