net/i40e: add AVX512 vector path
authorLeyi Rong <leyi.rong@intel.com>
Thu, 14 Jan 2021 06:39:50 +0000 (14:39 +0800)
committerFerruh Yigit <ferruh.yigit@intel.com>
Tue, 19 Jan 2021 02:30:11 +0000 (03:30 +0100)
Add AVX512 support for i40e PMD. This patch adds i40e_rxtx_vec_avx512.c
to support i40e AVX512 vPMD.

This patch aims to enable AVX512 on i40e vPMD. Main changes are focus
on Rx path compared with AVX2 vPMD.

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Acked-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
drivers/net/i40e/i40e_rxtx.c
drivers/net/i40e/i40e_rxtx.h
drivers/net/i40e/i40e_rxtx_vec_avx512.c [new file with mode: 0644]
drivers/net/i40e/meson.build

index 46e9276..25963b3 100644 (file)
@@ -1767,6 +1767,10 @@ i40e_dev_supported_ptypes_get(struct rte_eth_dev *dev)
            dev->rx_pkt_burst == i40e_recv_scattered_pkts ||
            dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
            dev->rx_pkt_burst == i40e_recv_pkts_vec ||
+#ifdef CC_AVX512_SUPPORT
+           dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx512 ||
+           dev->rx_pkt_burst == i40e_recv_pkts_vec_avx512 ||
+#endif
            dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 ||
            dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2)
                return ptypes;
@@ -3162,6 +3166,7 @@ i40e_set_rx_function(struct rte_eth_dev *dev)
                I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
        uint16_t rx_using_sse, i;
        bool use_avx2 = false;
+       bool use_avx512 = false;
        /* In order to allow Vector Rx there are a few configuration
         * conditions to be met and Rx Bulk Allocation should be allowed.
         */
@@ -3185,28 +3190,51 @@ i40e_set_rx_function(struct rte_eth_dev *dev)
                                }
                        }
 
-                       use_avx2 = get_avx_supported(0);
+                       use_avx512 = get_avx_supported(1);
+
+                       if (!use_avx512)
+                               use_avx2 = get_avx_supported(0);
                }
        }
 
        if (ad->rx_vec_allowed  &&
                        rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
                if (dev->data->scattered_rx) {
-                       PMD_INIT_LOG(DEBUG,
-                               "Using %sVector Scattered Rx (port %d).",
-                               use_avx2 ? "avx2 " : "",
-                               dev->data->port_id);
-                       dev->rx_pkt_burst = use_avx2 ?
-                               i40e_recv_scattered_pkts_vec_avx2 :
-                               i40e_recv_scattered_pkts_vec;
+                       if (use_avx512) {
+#ifdef CC_AVX512_SUPPORT
+                               PMD_DRV_LOG(NOTICE,
+                                       "Using AVX512 Vector Scattered Rx (port %d).",
+                                       dev->data->port_id);
+                               dev->rx_pkt_burst =
+                                       i40e_recv_scattered_pkts_vec_avx512;
+#endif
+                       } else {
+                               PMD_INIT_LOG(DEBUG,
+                                       "Using %sVector Scattered Rx (port %d).",
+                                       use_avx2 ? "avx2 " : "",
+                                       dev->data->port_id);
+                               dev->rx_pkt_burst = use_avx2 ?
+                                       i40e_recv_scattered_pkts_vec_avx2 :
+                                       i40e_recv_scattered_pkts_vec;
+                       }
                } else {
-                       PMD_INIT_LOG(DEBUG,
-                               "Using %sVector Rx (port %d).",
-                               use_avx2 ? "avx2 " : "",
-                               dev->data->port_id);
-                       dev->rx_pkt_burst = use_avx2 ?
-                               i40e_recv_pkts_vec_avx2 :
-                               i40e_recv_pkts_vec;
+                       if (use_avx512) {
+#ifdef CC_AVX512_SUPPORT
+                               PMD_DRV_LOG(NOTICE,
+                                       "Using AVX512 Vector Rx (port %d).",
+                                       dev->data->port_id);
+                               dev->rx_pkt_burst =
+                                       i40e_recv_pkts_vec_avx512;
+#endif
+                       } else {
+                               PMD_INIT_LOG(DEBUG,
+                                       "Using %sVector Rx (port %d).",
+                                       use_avx2 ? "avx2 " : "",
+                                       dev->data->port_id);
+                               dev->rx_pkt_burst = use_avx2 ?
+                                       i40e_recv_pkts_vec_avx2 :
+                                       i40e_recv_pkts_vec;
+                       }
                }
        } else if (!dev->data->scattered_rx && ad->rx_bulk_alloc_allowed) {
                PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
@@ -3229,6 +3257,10 @@ i40e_set_rx_function(struct rte_eth_dev *dev)
                rx_using_sse =
                        (dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
                         dev->rx_pkt_burst == i40e_recv_pkts_vec ||
+#ifdef CC_AVX512_SUPPORT
+                        dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx512 ||
+                        dev->rx_pkt_burst == i40e_recv_pkts_vec_avx512 ||
+#endif
                         dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 ||
                         dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2);
 
@@ -3249,6 +3281,10 @@ static const struct {
        { i40e_recv_pkts_bulk_alloc,         "Scalar Bulk Alloc" },
        { i40e_recv_pkts,                    "Scalar" },
 #ifdef RTE_ARCH_X86
+#ifdef CC_AVX512_SUPPORT
+       { i40e_recv_scattered_pkts_vec_avx512, "Vector AVX512 Scattered" },
+       { i40e_recv_pkts_vec_avx512,           "Vector AVX512" },
+#endif
        { i40e_recv_scattered_pkts_vec_avx2, "Vector AVX2 Scattered" },
        { i40e_recv_pkts_vec_avx2,           "Vector AVX2" },
        { i40e_recv_scattered_pkts_vec,      "Vector SSE Scattered" },
@@ -3315,6 +3351,7 @@ i40e_set_tx_function(struct rte_eth_dev *dev)
                I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
        int i;
        bool use_avx2 = false;
+       bool use_avx512 = false;
 
        if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
                if (ad->tx_vec_allowed) {
@@ -3328,19 +3365,30 @@ i40e_set_tx_function(struct rte_eth_dev *dev)
                                }
                        }
 
-                       use_avx2 = get_avx_supported(0);
+                       use_avx512 = get_avx_supported(1);
+
+                       if (!use_avx512)
+                               use_avx2 = get_avx_supported(0);
                }
        }
 
        if (ad->tx_simple_allowed) {
                if (ad->tx_vec_allowed &&
                                rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
-                       PMD_INIT_LOG(DEBUG, "Using %sVector Tx (port %d).",
-                                    use_avx2 ? "avx2 " : "",
-                                    dev->data->port_id);
-                       dev->tx_pkt_burst = use_avx2 ?
-                                           i40e_xmit_pkts_vec_avx2 :
-                                           i40e_xmit_pkts_vec;
+                       if (use_avx512) {
+#ifdef CC_AVX512_SUPPORT
+                               PMD_DRV_LOG(NOTICE, "Using AVX512 Vector Tx (port %d).",
+                                           dev->data->port_id);
+                               dev->tx_pkt_burst = i40e_xmit_pkts_vec_avx512;
+#endif
+                       } else {
+                               PMD_INIT_LOG(DEBUG, "Using %sVector Tx (port %d).",
+                                            use_avx2 ? "avx2 " : "",
+                                            dev->data->port_id);
+                               dev->tx_pkt_burst = use_avx2 ?
+                                                   i40e_xmit_pkts_vec_avx2 :
+                                                   i40e_xmit_pkts_vec;
+                       }
                } else {
                        PMD_INIT_LOG(DEBUG, "Simple tx finally be used.");
                        dev->tx_pkt_burst = i40e_xmit_pkts_simple;
@@ -3360,6 +3408,9 @@ static const struct {
        { i40e_xmit_pkts_simple,   "Scalar Simple" },
        { i40e_xmit_pkts,          "Scalar" },
 #ifdef RTE_ARCH_X86
+#ifdef CC_AVX512_SUPPORT
+       { i40e_xmit_pkts_vec_avx512, "Vector AVX512" },
+#endif
        { i40e_xmit_pkts_vec_avx2, "Vector AVX2" },
        { i40e_xmit_pkts_vec,      "Vector SSE" },
 #elif defined(RTE_ARCH_ARM64)
index e149452..f4ad6b6 100644 (file)
@@ -249,6 +249,15 @@ uint16_t i40e_recv_scattered_pkts_vec_avx2(void *rx_queue,
 uint16_t i40e_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
        uint16_t nb_pkts);
 int i40e_get_monitor_addr(void *rx_queue, struct rte_power_monitor_cond *pmc);
+uint16_t i40e_recv_pkts_vec_avx512(void *rx_queue,
+                                  struct rte_mbuf **rx_pkts,
+                                  uint16_t nb_pkts);
+uint16_t i40e_recv_scattered_pkts_vec_avx512(void *rx_queue,
+                                            struct rte_mbuf **rx_pkts,
+                                            uint16_t nb_pkts);
+uint16_t i40e_xmit_pkts_vec_avx512(void *tx_queue,
+                                  struct rte_mbuf **tx_pkts,
+                                  uint16_t nb_pkts);
 
 /* For each value it means, datasheet of hardware can tell more details
  *
diff --git a/drivers/net/i40e/i40e_rxtx_vec_avx512.c b/drivers/net/i40e/i40e_rxtx_vec_avx512.c
new file mode 100644 (file)
index 0000000..94075aa
--- /dev/null
@@ -0,0 +1,1024 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include <stdint.h>
+#include <rte_ethdev_driver.h>
+#include <rte_malloc.h>
+
+#include "base/i40e_prototype.h"
+#include "base/i40e_type.h"
+#include "i40e_ethdev.h"
+#include "i40e_rxtx.h"
+#include "i40e_rxtx_vec_common.h"
+
+#include <rte_vect.h>
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+#define RTE_I40E_DESCS_PER_LOOP_AVX 8
+
+static inline void
+i40e_rxq_rearm(struct i40e_rx_queue *rxq)
+{
+       int i;
+       uint16_t rx_id;
+       volatile union i40e_rx_desc *rxdp;
+       struct i40e_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
+       struct rte_mempool_cache *cache = rte_mempool_default_cache(rxq->mp,
+                       rte_lcore_id());
+
+       rxdp = rxq->rx_ring + rxq->rxrearm_start;
+
+       /* We need to pull 'n' more MBUFs into the software ring from mempool
+        * We inline the mempool function here, so we can vectorize the copy
+        * from the cache into the shadow ring.
+        */
+
+       if (cache->len < RTE_I40E_RXQ_REARM_THRESH) {
+               /* No. Backfill the cache first, and then fill from it */
+               uint32_t req = RTE_I40E_RXQ_REARM_THRESH + (cache->size -
+                               cache->len);
+
+               /* How many do we require
+                * i.e. number to fill the cache + the request
+                */
+               int ret = rte_mempool_ops_dequeue_bulk(rxq->mp,
+                               &cache->objs[cache->len], req);
+               if (ret == 0) {
+                       cache->len += req;
+               } else {
+                       if (rxq->rxrearm_nb + RTE_I40E_RXQ_REARM_THRESH >=
+                                       rxq->nb_rx_desc) {
+                               __m128i dma_addr0;
+
+                               dma_addr0 = _mm_setzero_si128();
+                               for (i = 0; i < RTE_I40E_DESCS_PER_LOOP; i++) {
+                                       rxep[i].mbuf = &rxq->fake_mbuf;
+                                       _mm_store_si128
+                                               ((__m128i *)&rxdp[i].read,
+                                                       dma_addr0);
+                               }
+                       }
+                       rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+                                       RTE_I40E_RXQ_REARM_THRESH;
+                       return;
+               }
+       }
+
+       const __m512i iova_offsets =  _mm512_set1_epi64
+               (offsetof(struct rte_mbuf, buf_iova));
+       const __m512i headroom = _mm512_set1_epi64(RTE_PKTMBUF_HEADROOM);
+
+#ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
+       /* to shuffle the addresses to correct slots. Values 4-7 will contain
+        * zeros, so use 7 for a zero-value.
+        */
+       const __m512i permute_idx = _mm512_set_epi64(7, 7, 3, 1, 7, 7, 2, 0);
+#else
+       const __m512i permute_idx = _mm512_set_epi64(7, 3, 6, 2, 5, 1, 4, 0);
+#endif
+
+       /* Initialize the mbufs in vector, process 8 mbufs in one loop, taking
+        * from mempool cache and populating both shadow and HW rings
+        */
+       for (i = 0; i < RTE_I40E_RXQ_REARM_THRESH / 8; i++) {
+               const __m512i mbuf_ptrs = _mm512_loadu_si512
+                       (&cache->objs[cache->len - 8]);
+               _mm512_store_si512(rxep, mbuf_ptrs);
+
+               /* gather iova of mbuf0-7 into one zmm reg */
+               const __m512i iova_base_addrs = _mm512_i64gather_epi64
+                       (_mm512_add_epi64(mbuf_ptrs, iova_offsets),
+                               0, /* base */
+                               1 /* scale */);
+               const __m512i iova_addrs = _mm512_add_epi64(iova_base_addrs,
+                               headroom);
+#ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
+               const __m512i iovas0 = _mm512_castsi256_si512
+                       (_mm512_extracti64x4_epi64(iova_addrs, 0));
+               const __m512i iovas1 = _mm512_castsi256_si512
+                       (_mm512_extracti64x4_epi64(iova_addrs, 1));
+
+               /* permute leaves desc 2-3 addresses in header address slots 0-1
+                * but these are ignored by driver since header split not
+                * enabled. Similarly for desc 4 & 5.
+                */
+               const __m512i desc_rd_0_1 = _mm512_permutexvar_epi64
+                       (permute_idx, iovas0);
+               const __m512i desc_rd_2_3 = _mm512_bsrli_epi128(desc_rd_0_1, 8);
+
+               const __m512i desc_rd_4_5 = _mm512_permutexvar_epi64
+                       (permute_idx, iovas1);
+               const __m512i desc_rd_6_7 = _mm512_bsrli_epi128(desc_rd_4_5, 8);
+
+               _mm512_store_si512((void *)rxdp, desc_rd_0_1);
+               _mm512_store_si512((void *)(rxdp + 2), desc_rd_2_3);
+               _mm512_store_si512((void *)(rxdp + 4), desc_rd_4_5);
+               _mm512_store_si512((void *)(rxdp + 6), desc_rd_6_7);
+#else
+               /* permute leaves desc 4-7 addresses in header address slots 0-3
+                * but these are ignored by driver since header split not
+                * enabled.
+                */
+               const __m512i desc_rd_0_3 = _mm512_permutexvar_epi64
+                       (permute_idx, iova_addrs);
+               const __m512i desc_rd_4_7 = _mm512_bsrli_epi128(desc_rd_0_3, 8);
+
+               _mm512_store_si512((void *)rxdp, desc_rd_0_3);
+               _mm512_store_si512((void *)(rxdp + 4), desc_rd_4_7);
+#endif
+               rxep += 8, rxdp += 8, cache->len -= 8;
+       }
+
+       rxq->rxrearm_start += RTE_I40E_RXQ_REARM_THRESH;
+       if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+               rxq->rxrearm_start = 0;
+
+       rxq->rxrearm_nb -= RTE_I40E_RXQ_REARM_THRESH;
+
+       rx_id = (uint16_t)((rxq->rxrearm_start == 0) ?
+                            (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1));
+
+       /* Update the tail pointer on the NIC */
+       I40E_PCI_REG_WC_WRITE(rxq->qrx_tail, rx_id);
+}
+
+#ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
+/* Handles 32B descriptor FDIR ID processing:
+ * rxdp: receive descriptor ring, required to load 2nd 16B half of each desc
+ * rx_pkts: required to store metadata back to mbufs
+ * pkt_idx: offset into the burst, increments in vector widths
+ * desc_idx: required to select the correct shift at compile time
+ */
+static inline __m256i
+desc_fdir_processing_32b(volatile union i40e_rx_desc *rxdp,
+                        struct rte_mbuf **rx_pkts,
+                        const uint32_t pkt_idx,
+                        const uint32_t desc_idx)
+{
+       /* 32B desc path: load rxdp.wb.qword2 for EXT_STATUS and FLEXBH_STAT */
+       __m128i *rxdp_desc_0 = (void *)(&rxdp[desc_idx + 0].wb.qword2);
+       __m128i *rxdp_desc_1 = (void *)(&rxdp[desc_idx + 1].wb.qword2);
+       const __m128i desc_qw2_0 = _mm_load_si128(rxdp_desc_0);
+       const __m128i desc_qw2_1 = _mm_load_si128(rxdp_desc_1);
+
+       /* Mask for FLEXBH_STAT, and the FDIR_ID value to compare against. The
+        * remaining data is set to all 1's to pass through data.
+        */
+       const __m256i flexbh_mask = _mm256_set_epi32(-1, -1, -1, 3 << 4,
+                                                    -1, -1, -1, 3 << 4);
+       const __m256i flexbh_id   = _mm256_set_epi32(-1, -1, -1, 1 << 4,
+                                                    -1, -1, -1, 1 << 4);
+
+       /* Load descriptor, check for FLEXBH bits, generate a mask for both
+        * packets in the register.
+        */
+       __m256i desc_qw2_0_1 =
+               _mm256_inserti128_si256(_mm256_castsi128_si256(desc_qw2_0),
+                                       desc_qw2_1, 1);
+       __m256i desc_tmp_msk = _mm256_and_si256(flexbh_mask, desc_qw2_0_1);
+       __m256i fdir_mask = _mm256_cmpeq_epi32(flexbh_id, desc_tmp_msk);
+       __m256i fdir_data = _mm256_alignr_epi8(desc_qw2_0_1, desc_qw2_0_1, 12);
+       __m256i desc_fdir_data = _mm256_and_si256(fdir_mask, fdir_data);
+
+       /* Write data out to the mbuf. There is no store to this area of the
+        * mbuf today, so we cannot combine it with another store.
+        */
+       const uint32_t idx_0 = pkt_idx + desc_idx;
+       const uint32_t idx_1 = pkt_idx + desc_idx + 1;
+
+       rx_pkts[idx_0]->hash.fdir.hi = _mm256_extract_epi32(desc_fdir_data, 0);
+       rx_pkts[idx_1]->hash.fdir.hi = _mm256_extract_epi32(desc_fdir_data, 4);
+
+       /* Create mbuf flags as required for mbuf_flags layout
+        *  (That's high lane [1,3,5,7, 0,2,4,6] as u32 lanes).
+        * Approach:
+        * - Mask away bits not required from the fdir_mask
+        * - Leave the PKT_FDIR_ID bit (1 << 13)
+        * - Position that bit correctly based on packet number
+        * - OR in the resulting bit to mbuf_flags
+        */
+       RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
+       __m256i mbuf_flag_mask = _mm256_set_epi32(0, 0, 0, 1 << 13,
+                                                 0, 0, 0, 1 << 13);
+       __m256i desc_flag_bit =  _mm256_and_si256(mbuf_flag_mask, fdir_mask);
+
+       /* For static-inline function, this will be stripped out
+        * as the desc_idx is a hard-coded constant.
+        */
+       switch (desc_idx) {
+       case 0:
+               return _mm256_alignr_epi8(desc_flag_bit, desc_flag_bit,  4);
+       case 2:
+               return _mm256_alignr_epi8(desc_flag_bit, desc_flag_bit,  8);
+       case 4:
+               return _mm256_alignr_epi8(desc_flag_bit, desc_flag_bit, 12);
+       case 6:
+               return desc_flag_bit;
+       default:
+               break;
+       }
+
+       /* NOT REACHED, see above switch returns */
+       return _mm256_setzero_si256();
+}
+#endif /* RTE_LIBRTE_I40E_16BYTE_RX_DESC */
+
+#define PKTLEN_SHIFT     10
+
+/* Force inline as some compilers will not inline by default. */
+static __rte_always_inline uint16_t
+_recv_raw_pkts_vec_avx512(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
+                         uint16_t nb_pkts, uint8_t *split_packet)
+{
+       const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
+       const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
+                       0, rxq->mbuf_initializer);
+       struct i40e_rx_entry *sw_ring = &rxq->sw_ring[rxq->rx_tail];
+       volatile union i40e_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
+
+       rte_prefetch0(rxdp);
+
+       /* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP_AVX */
+       nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP_AVX);
+
+       /* See if we need to rearm the RX queue - gives the prefetch a bit
+        * of time to act
+        */
+       if (rxq->rxrearm_nb > RTE_I40E_RXQ_REARM_THRESH)
+               i40e_rxq_rearm(rxq);
+
+       /* Before we start moving massive data around, check to see if
+        * there is actually a packet available
+        */
+       if (!(rxdp->wb.qword1.status_error_len &
+                       rte_cpu_to_le_32(1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
+               return 0;
+
+       /* constants used in processing loop */
+       const __m512i crc_adjust =
+               _mm512_set4_epi32
+                       (0,             /* ignore non-length fields */
+                        -rxq->crc_len, /* sub crc on data_len */
+                        -rxq->crc_len, /* sub crc on pkt_len */
+                        0              /* ignore non-length fields */
+                       );
+
+       /* 8 packets DD mask, LSB in each 32-bit value */
+       const __m256i dd_check = _mm256_set1_epi32(1);
+
+       /* 8 packets EOP mask, second-LSB in each 32-bit value */
+       const __m256i eop_check = _mm256_slli_epi32(dd_check,
+                       I40E_RX_DESC_STATUS_EOF_SHIFT);
+
+       /* mask to shuffle from desc. to mbuf (2 descriptors)*/
+       const __m512i shuf_msk =
+               _mm512_set4_epi32
+                       (/* rss hash parsed separately */
+                        /* octet 4~7, 32bits rss */
+                        7 << 24 | 6 << 16 | 5 << 8 | 4,
+                        /* octet 2~3, low 16 bits vlan_macip */
+                        /* octet 14~15, 16 bits data_len */
+                        3 << 24 | 2 << 16 | 15 << 8 | 14,
+                        /* skip hi 16 bits pkt_len, zero out */
+                        /* octet 14~15, 16 bits pkt_len */
+                        0xFFFF << 16 | 15 << 8 | 14,
+                        /* pkt_type set as unknown */
+                        0xFFFFFFFF
+                       );
+       /* compile-time check the above crc and shuffle layout is correct.
+        * NOTE: the first field (lowest address) is given last in set_epi
+        * calls above.
+        */
+       RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
+                       offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
+       RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
+                       offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
+       RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
+                       offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
+       RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
+                       offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
+
+       /* Status/Error flag masks */
+       /* mask everything except RSS, flow director and VLAN flags
+        * bit2 is for VLAN tag, bit11 for flow director indication
+        * bit13:12 for RSS indication. Bits 3-5 of error
+        * field (bits 22-24) are for IP/L4 checksum errors
+        */
+       const __m256i flags_mask = _mm256_set1_epi32
+               ((1 << 2) | (1 << 11) | (3 << 12) | (7 << 22));
+
+       /* data to be shuffled by result of flag mask. If VLAN bit is set,
+        * (bit 2), then position 4 in this array will be used in the
+        * destination
+        */
+       const __m256i vlan_flags_shuf = _mm256_set_epi32
+               (0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0,
+               0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0);
+
+       /* data to be shuffled by result of flag mask, shifted down 11.
+        * If RSS/FDIR bits are set, shuffle moves appropriate flags in
+        * place.
+        */
+       const __m256i rss_flags_shuf = _mm256_set_epi8
+               (0, 0, 0, 0, 0, 0, 0, 0,
+               PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH, 0, 0,
+               0, 0, PKT_RX_FDIR, 0, /* end up 128-bits */
+               0, 0, 0, 0, 0, 0, 0, 0,
+               PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH, 0, 0,
+               0, 0, PKT_RX_FDIR, 0);
+
+       /* data to be shuffled by the result of the flags mask shifted by 22
+        * bits.  This gives use the l3_l4 flags.
+        */
+       const __m256i l3_l4_flags_shuf = _mm256_set_epi8
+               (0, 0, 0, 0, 0, 0, 0, 0,
+               /* shift right 1 bit to make sure it not exceed 255 */
+               (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+                PKT_RX_IP_CKSUM_BAD) >> 1,
+               (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD |
+                PKT_RX_L4_CKSUM_BAD) >> 1,
+               (PKT_RX_EIP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+               (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD) >> 1,
+               (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+               (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
+               PKT_RX_IP_CKSUM_BAD >> 1,
+               (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1,
+               /* second 128-bits */
+               0, 0, 0, 0, 0, 0, 0, 0,
+               (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+                PKT_RX_IP_CKSUM_BAD) >> 1,
+               (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD |
+                PKT_RX_L4_CKSUM_BAD) >> 1,
+               (PKT_RX_EIP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+               (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD) >> 1,
+               (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+               (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
+               PKT_RX_IP_CKSUM_BAD >> 1,
+               (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1);
+
+       const __m256i cksum_mask = _mm256_set1_epi32
+               (PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
+               PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
+               PKT_RX_EIP_CKSUM_BAD);
+
+       uint16_t i, received;
+
+       for (i = 0, received = 0; i < nb_pkts;
+                       i += RTE_I40E_DESCS_PER_LOOP_AVX,
+                       rxdp += RTE_I40E_DESCS_PER_LOOP_AVX) {
+               /* step 1, copy over 8 mbuf pointers to rx_pkts array */
+               _mm256_storeu_si256((void *)&rx_pkts[i],
+                               _mm256_loadu_si256((void *)&sw_ring[i]));
+#ifdef RTE_ARCH_X86_64
+               _mm256_storeu_si256((void *)&rx_pkts[i + 4],
+                               _mm256_loadu_si256((void *)&sw_ring[i + 4]));
+#endif
+
+               __m512i raw_desc0_3, raw_desc4_7;
+               __m256i raw_desc0_1, raw_desc2_3, raw_desc4_5, raw_desc6_7;
+
+               /* load in descriptors, in reverse order */
+               const __m128i raw_desc7 =
+                       _mm_load_si128((void *)(rxdp + 7));
+               rte_compiler_barrier();
+               const __m128i raw_desc6 =
+                       _mm_load_si128((void *)(rxdp + 6));
+               rte_compiler_barrier();
+               const __m128i raw_desc5 =
+                       _mm_load_si128((void *)(rxdp + 5));
+               rte_compiler_barrier();
+               const __m128i raw_desc4 =
+                       _mm_load_si128((void *)(rxdp + 4));
+               rte_compiler_barrier();
+               const __m128i raw_desc3 =
+                       _mm_load_si128((void *)(rxdp + 3));
+               rte_compiler_barrier();
+               const __m128i raw_desc2 =
+                       _mm_load_si128((void *)(rxdp + 2));
+               rte_compiler_barrier();
+               const __m128i raw_desc1 =
+                       _mm_load_si128((void *)(rxdp + 1));
+               rte_compiler_barrier();
+               const __m128i raw_desc0 =
+                       _mm_load_si128((void *)(rxdp + 0));
+
+               raw_desc6_7 =
+                       _mm256_inserti128_si256
+                               (_mm256_castsi128_si256(raw_desc6),
+                                raw_desc7, 1);
+               raw_desc4_5 =
+                       _mm256_inserti128_si256
+                               (_mm256_castsi128_si256(raw_desc4),
+                                raw_desc5, 1);
+               raw_desc2_3 =
+                       _mm256_inserti128_si256
+                               (_mm256_castsi128_si256(raw_desc2),
+                                raw_desc3, 1);
+               raw_desc0_1 =
+                       _mm256_inserti128_si256
+                               (_mm256_castsi128_si256(raw_desc0),
+                                raw_desc1, 1);
+
+               raw_desc4_7 =
+                       _mm512_inserti64x4
+                               (_mm512_castsi256_si512(raw_desc4_5),
+                                raw_desc6_7, 1);
+               raw_desc0_3 =
+                       _mm512_inserti64x4
+                               (_mm512_castsi256_si512(raw_desc0_1),
+                                raw_desc2_3, 1);
+
+               if (split_packet) {
+                       int j;
+
+                       for (j = 0; j < RTE_I40E_DESCS_PER_LOOP_AVX; j++)
+                               rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+               }
+
+               /* convert descriptors 0-7 into mbufs, adjusting length and
+                * re-arranging fields. Then write into the mbuf
+                */
+               const __m512i len4_7 = _mm512_slli_epi32
+                                       (raw_desc4_7, PKTLEN_SHIFT);
+               const __m512i len0_3 = _mm512_slli_epi32
+                                       (raw_desc0_3, PKTLEN_SHIFT);
+               const __m512i desc4_7 = _mm512_mask_blend_epi16
+                                       (0x80808080, raw_desc4_7, len4_7);
+               const __m512i desc0_3 = _mm512_mask_blend_epi16
+                                       (0x80808080, raw_desc0_3, len0_3);
+               __m512i mb4_7 = _mm512_shuffle_epi8(desc4_7, shuf_msk);
+               __m512i mb0_3 = _mm512_shuffle_epi8(desc0_3, shuf_msk);
+
+               mb4_7 = _mm512_add_epi32(mb4_7, crc_adjust);
+               mb0_3 = _mm512_add_epi32(mb0_3, crc_adjust);
+
+               /* to get packet types, shift 64-bit values down 30 bits
+                * and so ptype is in lower 8-bits in each
+                */
+               const __m512i ptypes4_7 = _mm512_srli_epi64(desc4_7, 30);
+               const __m512i ptypes0_3 = _mm512_srli_epi64(desc0_3, 30);
+               const __m256i ptypes6_7 =
+                       _mm512_extracti64x4_epi64(ptypes4_7, 1);
+               const __m256i ptypes4_5 =
+                       _mm512_extracti64x4_epi64(ptypes4_7, 0);
+               const __m256i ptypes2_3 =
+                       _mm512_extracti64x4_epi64(ptypes0_3, 1);
+               const __m256i ptypes0_1 =
+                       _mm512_extracti64x4_epi64(ptypes0_3, 0);
+               const uint8_t ptype7 = _mm256_extract_epi8(ptypes6_7, 24);
+               const uint8_t ptype6 = _mm256_extract_epi8(ptypes6_7, 8);
+               const uint8_t ptype5 = _mm256_extract_epi8(ptypes4_5, 24);
+               const uint8_t ptype4 = _mm256_extract_epi8(ptypes4_5, 8);
+               const uint8_t ptype3 = _mm256_extract_epi8(ptypes2_3, 24);
+               const uint8_t ptype2 = _mm256_extract_epi8(ptypes2_3, 8);
+               const uint8_t ptype1 = _mm256_extract_epi8(ptypes0_1, 24);
+               const uint8_t ptype0 = _mm256_extract_epi8(ptypes0_1, 8);
+
+               const __m512i ptype4_7 = _mm512_set_epi32
+                       (0, 0, 0, ptype_tbl[ptype7],
+                        0, 0, 0, ptype_tbl[ptype6],
+                        0, 0, 0, ptype_tbl[ptype5],
+                        0, 0, 0, ptype_tbl[ptype4]);
+               const __m512i ptype0_3 = _mm512_set_epi32
+                       (0, 0, 0, ptype_tbl[ptype3],
+                        0, 0, 0, ptype_tbl[ptype2],
+                        0, 0, 0, ptype_tbl[ptype1],
+                        0, 0, 0, ptype_tbl[ptype0]);
+
+               mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+               mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+               __m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7, 0);
+               __m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7, 1);
+               __m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3, 0);
+               __m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3, 1);
+
+               /**
+                * use permute/extract to get status content
+                * After the operations, the packets status flags are in the
+                * order (hi->lo): [1, 3, 5, 7, 0, 2, 4, 6]
+                */
+               /* merge the status bits into one register */
+               const __m512i status_permute_msk = _mm512_set_epi32
+                       (0, 0, 0, 0,
+                        0, 0, 0, 0,
+                        22, 30, 6, 14,
+                        18, 26, 2, 10);
+               const __m512i raw_status0_7 = _mm512_permutex2var_epi32
+                       (desc4_7, status_permute_msk, desc0_3);
+               __m256i status0_7 = _mm512_extracti64x4_epi64
+                       (raw_status0_7, 0);
+
+               /* now do flag manipulation */
+
+               /* get only flag/error bits we want */
+               const __m256i flag_bits =
+                       _mm256_and_si256(status0_7, flags_mask);
+               /* set vlan and rss flags */
+               const __m256i vlan_flags =
+                       _mm256_shuffle_epi8(vlan_flags_shuf, flag_bits);
+               const __m256i rss_fdir_bits = _mm256_srli_epi32(flag_bits, 11);
+               const __m256i rss_flags = _mm256_shuffle_epi8(rss_flags_shuf,
+                                                             rss_fdir_bits);
+
+               /* l3_l4_error flags, shuffle, then shift to correct adjustment
+                * of flags in flags_shuf, and finally mask out extra bits
+                */
+               __m256i l3_l4_flags = _mm256_shuffle_epi8(l3_l4_flags_shuf,
+                               _mm256_srli_epi32(flag_bits, 22));
+               l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
+               l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
+
+               /* merge flags */
+               __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+                               _mm256_or_si256(rss_flags, vlan_flags));
+
+               /* If the rxq has FDIR enabled, read and process the FDIR info
+                * from the descriptor. This can cause more loads/stores, so is
+                * not always performed. Branch over the code when not enabled.
+                */
+               if (rxq->fdir_enabled) {
+#ifdef RTE_LIBRTE_I40E_16BYTE_RX_DESC
+                       /* 16B descriptor code path:
+                        * RSS and FDIR ID use the same offset in the desc, so
+                        * only one can be present at a time. The code below
+                        * identifies an FDIR ID match, and zeros the RSS value
+                        * in the mbuf on FDIR match to keep mbuf data clean.
+                        */
+#define FDIR_BLEND_MASK ((1 << 3) | (1 << 7))
+
+                       /* Flags:
+                        * - Take flags, shift bits to null out
+                        * - CMPEQ with known FDIR ID, to get 0xFFFF or 0 mask
+                        * - Strip bits from mask, leaving 0 or 1 for FDIR ID
+                        * - Merge with mbuf_flags
+                        */
+                       /* FLM = 1, FLTSTAT = 0b01, (FLM | FLTSTAT) == 3.
+                        * Shift left by 28 to avoid having to mask.
+                        */
+                       const __m256i fdir =
+                               _mm256_slli_epi32(rss_fdir_bits, 28);
+                       const __m256i fdir_id = _mm256_set1_epi32(3 << 28);
+
+                       /* As above, the fdir_mask to packet mapping is this:
+                        * order (hi->lo): [1, 3, 5, 7, 0, 2, 4, 6]
+                        * Then OR FDIR flags to mbuf_flags on FDIR ID hit.
+                        */
+                       RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
+                       const __m256i pkt_fdir_bit = _mm256_set1_epi32(1 << 13);
+                       const __m256i fdir_mask =
+                               _mm256_cmpeq_epi32(fdir, fdir_id);
+                       __m256i fdir_bits =
+                               _mm256_and_si256(fdir_mask, pkt_fdir_bit);
+
+                       mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_bits);
+
+                       /* Based on FDIR_MASK, clear the RSS or FDIR value.
+                        * The FDIR ID value is masked to zero if not a hit,
+                        * otherwise the mb0_1 register RSS field is zeroed.
+                        */
+                       const __m256i fdir_zero_mask = _mm256_setzero_si256();
+                       __m256i tmp0_1 = _mm256_blend_epi32(fdir_zero_mask,
+                                               fdir_mask, FDIR_BLEND_MASK);
+                       __m256i fdir_mb0_1 = _mm256_and_si256(mb0_1, fdir_mask);
+
+                       mb0_1 = _mm256_andnot_si256(tmp0_1, mb0_1);
+
+                       /* Write to mbuf: no stores to combine with, so just a
+                        * scalar store to push data here.
+                        */
+                       rx_pkts[i + 0]->hash.fdir.hi =
+                               _mm256_extract_epi32(fdir_mb0_1, 3);
+                       rx_pkts[i + 1]->hash.fdir.hi =
+                               _mm256_extract_epi32(fdir_mb0_1, 7);
+
+                       /* Same as above, only shift the fdir_mask to align
+                        * the packet FDIR mask with the FDIR_ID desc lane.
+                        */
+                       __m256i tmp2_3 =
+                               _mm256_alignr_epi8(fdir_mask, fdir_mask, 12);
+                       __m256i fdir_mb2_3 = _mm256_and_si256(mb2_3, tmp2_3);
+
+                       tmp2_3 = _mm256_blend_epi32(fdir_zero_mask, tmp2_3,
+                                                   FDIR_BLEND_MASK);
+                       mb2_3 = _mm256_andnot_si256(tmp2_3, mb2_3);
+                       rx_pkts[i + 2]->hash.fdir.hi =
+                               _mm256_extract_epi32(fdir_mb2_3, 3);
+                       rx_pkts[i + 3]->hash.fdir.hi =
+                               _mm256_extract_epi32(fdir_mb2_3, 7);
+
+                       __m256i tmp4_5 =
+                               _mm256_alignr_epi8(fdir_mask, fdir_mask, 8);
+                       __m256i fdir_mb4_5 = _mm256_and_si256(mb4_5, tmp4_5);
+
+                       tmp4_5 = _mm256_blend_epi32(fdir_zero_mask, tmp4_5,
+                                                   FDIR_BLEND_MASK);
+                       mb4_5 = _mm256_andnot_si256(tmp4_5, mb4_5);
+                       rx_pkts[i + 4]->hash.fdir.hi =
+                               _mm256_extract_epi32(fdir_mb4_5, 3);
+                       rx_pkts[i + 5]->hash.fdir.hi =
+                               _mm256_extract_epi32(fdir_mb4_5, 7);
+
+                       __m256i tmp6_7 =
+                               _mm256_alignr_epi8(fdir_mask, fdir_mask, 4);
+                       __m256i fdir_mb6_7 = _mm256_and_si256(mb6_7, tmp6_7);
+
+                       tmp6_7 = _mm256_blend_epi32(fdir_zero_mask, tmp6_7,
+                                                   FDIR_BLEND_MASK);
+                       mb6_7 = _mm256_andnot_si256(tmp6_7, mb6_7);
+                       rx_pkts[i + 6]->hash.fdir.hi =
+                               _mm256_extract_epi32(fdir_mb6_7, 3);
+                       rx_pkts[i + 7]->hash.fdir.hi =
+                               _mm256_extract_epi32(fdir_mb6_7, 7);
+
+                       /* End of 16B descriptor handling */
+#else
+                       /* 32B descriptor FDIR ID mark handling. Returns bits
+                        * to be OR-ed into the mbuf olflags.
+                        */
+                       __m256i fdir_add_flags;
+
+                       fdir_add_flags =
+                               desc_fdir_processing_32b(rxdp, rx_pkts, i, 0);
+                       mbuf_flags =
+                               _mm256_or_si256(mbuf_flags, fdir_add_flags);
+
+                       fdir_add_flags =
+                               desc_fdir_processing_32b(rxdp, rx_pkts, i, 2);
+                       mbuf_flags =
+                               _mm256_or_si256(mbuf_flags, fdir_add_flags);
+
+                       fdir_add_flags =
+                               desc_fdir_processing_32b(rxdp, rx_pkts, i, 4);
+                       mbuf_flags =
+                               _mm256_or_si256(mbuf_flags, fdir_add_flags);
+
+                       fdir_add_flags =
+                               desc_fdir_processing_32b(rxdp, rx_pkts, i, 6);
+                       mbuf_flags =
+                               _mm256_or_si256(mbuf_flags, fdir_add_flags);
+                       /* End 32B desc handling */
+#endif /* RTE_LIBRTE_I40E_16BYTE_RX_DESC */
+
+               } /* if() on FDIR enabled */
+
+               /* At this point, we have the 8 sets of flags in the low 16-bits
+                * of each 32-bit value in vlan0.
+                * We want to extract these, and merge them with the mbuf init data
+                * so we can do a single write to the mbuf to set the flags
+                * and all the other initialization fields. Extracting the
+                * appropriate flags means that we have to do a shift and blend for
+                * each mbuf before we do the write. However, we can also
+                * add in the previously computed rx_descriptor fields to
+                * make a single 256-bit write per mbuf
+                */
+               /* check the structure matches expectations */
+               RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
+                               offsetof(struct rte_mbuf, rearm_data) + 8);
+               RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
+                               RTE_ALIGN(offsetof(struct rte_mbuf, rearm_data), 16));
+               /* build up data and do writes */
+               __m256i rearm0, rearm1, rearm2, rearm3, rearm4, rearm5,
+                               rearm6, rearm7;
+               rearm6 = _mm256_blend_epi32
+                       (mbuf_init, _mm256_slli_si256(mbuf_flags, 8), 0x04);
+               rearm4 = _mm256_blend_epi32
+                       (mbuf_init, _mm256_slli_si256(mbuf_flags, 4), 0x04);
+               rearm2 = _mm256_blend_epi32
+                       (mbuf_init, mbuf_flags, 0x04);
+               rearm0 = _mm256_blend_epi32
+                       (mbuf_init, _mm256_srli_si256(mbuf_flags, 4), 0x04);
+               /* permute to add in the rx_descriptor e.g. rss fields */
+               rearm6 = _mm256_permute2f128_si256(rearm6, mb6_7, 0x20);
+               rearm4 = _mm256_permute2f128_si256(rearm4, mb4_5, 0x20);
+               rearm2 = _mm256_permute2f128_si256(rearm2, mb2_3, 0x20);
+               rearm0 = _mm256_permute2f128_si256(rearm0, mb0_1, 0x20);
+               /* write to mbuf */
+               _mm256_storeu_si256
+                       ((__m256i *)&rx_pkts[i + 6]->rearm_data, rearm6);
+               _mm256_storeu_si256
+                       ((__m256i *)&rx_pkts[i + 4]->rearm_data, rearm4);
+               _mm256_storeu_si256
+                       ((__m256i *)&rx_pkts[i + 2]->rearm_data, rearm2);
+               _mm256_storeu_si256
+                       ((__m256i *)&rx_pkts[i + 0]->rearm_data, rearm0);
+
+               /* repeat for the odd mbufs */
+               const __m256i odd_flags = _mm256_castsi128_si256
+                       (_mm256_extracti128_si256(mbuf_flags, 1));
+               rearm7 = _mm256_blend_epi32
+                       (mbuf_init, _mm256_slli_si256(odd_flags, 8), 0x04);
+               rearm5 = _mm256_blend_epi32
+                       (mbuf_init, _mm256_slli_si256(odd_flags, 4), 0x04);
+               rearm3 = _mm256_blend_epi32
+                       (mbuf_init, odd_flags, 0x04);
+               rearm1 = _mm256_blend_epi32
+                       (mbuf_init, _mm256_srli_si256(odd_flags, 4), 0x04);
+               /* since odd mbufs are already in hi 128-bits use blend */
+               rearm7 = _mm256_blend_epi32(rearm7, mb6_7, 0xF0);
+               rearm5 = _mm256_blend_epi32(rearm5, mb4_5, 0xF0);
+               rearm3 = _mm256_blend_epi32(rearm3, mb2_3, 0xF0);
+               rearm1 = _mm256_blend_epi32(rearm1, mb0_1, 0xF0);
+               /* again write to mbufs */
+               _mm256_storeu_si256
+                       ((__m256i *)&rx_pkts[i + 7]->rearm_data, rearm7);
+               _mm256_storeu_si256
+                       ((__m256i *)&rx_pkts[i + 5]->rearm_data, rearm5);
+               _mm256_storeu_si256
+                       ((__m256i *)&rx_pkts[i + 3]->rearm_data, rearm3);
+               _mm256_storeu_si256
+                       ((__m256i *)&rx_pkts[i + 1]->rearm_data, rearm1);
+
+               /* extract and record EOP bit */
+               if (split_packet) {
+                       const __m128i eop_mask =
+                               _mm_set1_epi16
+                               (1 << I40E_RX_DESC_STATUS_EOF_SHIFT);
+                       const __m256i eop_bits256 =
+                               _mm256_and_si256(status0_7, eop_check);
+                       /* pack status bits into a single 128-bit register */
+                       const __m128i eop_bits =
+                               _mm_packus_epi32
+                               (_mm256_castsi256_si128(eop_bits256),
+                               _mm256_extractf128_si256(eop_bits256, 1));
+                       /* flip bits, and mask out the EOP bit, which is now
+                        * a split-packet bit i.e. !EOP, rather than EOP one.
+                        */
+                       __m128i split_bits = _mm_andnot_si128(eop_bits,
+                                       eop_mask);
+                       /* eop bits are out of order, so we need to shuffle them
+                        * back into order again. In doing so, only use low 8
+                        * bits, which acts like another pack instruction
+                        * The original order is (hi->lo): 1,3,5,7,0,2,4,6
+                        * [Since we use epi8, the 16-bit positions are
+                        * multiplied by 2 in the eop_shuffle value.]
+                        */
+                       __m128i eop_shuffle = _mm_set_epi8
+                               (0xFF, 0xFF, 0xFF, 0xFF, /* zero hi 64b */
+                               0xFF, 0xFF, 0xFF, 0xFF,
+                               8, 0, 10, 2, /* move values to lo 64b */
+                               12, 4, 14, 6);
+                       split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+                       *(uint64_t *)split_packet =
+                               _mm_cvtsi128_si64(split_bits);
+                       split_packet += RTE_I40E_DESCS_PER_LOOP_AVX;
+               }
+
+               /* perform dd_check */
+               status0_7 = _mm256_and_si256(status0_7, dd_check);
+               status0_7 = _mm256_packs_epi32
+                       (status0_7, _mm256_setzero_si256());
+
+               uint64_t burst = __builtin_popcountll
+                               (_mm_cvtsi128_si64
+                                       (_mm256_extracti128_si256
+                                               (status0_7, 1)));
+               burst += __builtin_popcountll(_mm_cvtsi128_si64
+                               (_mm256_castsi256_si128(status0_7)));
+               received += burst;
+               if (burst != RTE_I40E_DESCS_PER_LOOP_AVX)
+                       break;
+       }
+
+       /* update tail pointers */
+       rxq->rx_tail += received;
+       rxq->rx_tail &= (rxq->nb_rx_desc - 1);
+       if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep avx2 aligned */
+               rxq->rx_tail--;
+               received--;
+       }
+       rxq->rxrearm_nb += received;
+       return received;
+}
+
+/**
+ * Notice:
+ * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+i40e_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+                         uint16_t nb_pkts)
+{
+       return _recv_raw_pkts_vec_avx512(rx_queue, rx_pkts, nb_pkts, NULL);
+}
+
+/**
+ * vPMD receive routine that reassembles single burst of 32 scattered packets
+ * Notice:
+ * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+i40e_recv_scattered_burst_vec_avx512(void *rx_queue,
+                                    struct rte_mbuf **rx_pkts,
+                                    uint16_t nb_pkts)
+{
+       struct i40e_rx_queue *rxq = rx_queue;
+       uint8_t split_flags[RTE_I40E_VPMD_RX_BURST] = {0};
+
+       /* get some new buffers */
+       uint16_t nb_bufs = _recv_raw_pkts_vec_avx512(rxq, rx_pkts, nb_pkts,
+                       split_flags);
+       if (nb_bufs == 0)
+               return 0;
+
+       /* happy day case, full burst + no packets to be joined */
+       const uint64_t *split_fl64 = (uint64_t *)split_flags;
+
+       if (!rxq->pkt_first_seg &&
+           split_fl64[0] == 0 && split_fl64[1] == 0 &&
+           split_fl64[2] == 0 && split_fl64[3] == 0)
+               return nb_bufs;
+
+       /* reassemble any packets that need reassembly*/
+       unsigned int i = 0;
+
+       if (!rxq->pkt_first_seg) {
+               /* find the first split flag, and only reassemble then*/
+               while (i < nb_bufs && !split_flags[i])
+                       i++;
+               if (i == nb_bufs)
+                       return nb_bufs;
+               rxq->pkt_first_seg = rx_pkts[i];
+       }
+       return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
+               &split_flags[i]);
+}
+
+/**
+ * vPMD receive routine that reassembles scattered packets.
+ * Main receive routine that can handle arbitrary burst sizes
+ * Notice:
+ * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+i40e_recv_scattered_pkts_vec_avx512(void *rx_queue,
+                                   struct rte_mbuf **rx_pkts,
+                                   uint16_t nb_pkts)
+{
+       uint16_t retval = 0;
+
+       while (nb_pkts > RTE_I40E_VPMD_RX_BURST) {
+               uint16_t burst = i40e_recv_scattered_burst_vec_avx512(rx_queue,
+                               rx_pkts + retval, RTE_I40E_VPMD_RX_BURST);
+               retval += burst;
+               nb_pkts -= burst;
+               if (burst < RTE_I40E_VPMD_RX_BURST)
+                       return retval;
+       }
+       return retval + i40e_recv_scattered_burst_vec_avx512(rx_queue,
+                               rx_pkts + retval, nb_pkts);
+}
+
+static inline void
+vtx1(volatile struct i40e_tx_desc *txdp, struct rte_mbuf *pkt, uint64_t flags)
+{
+       uint64_t high_qw = (I40E_TX_DESC_DTYPE_DATA |
+               ((uint64_t)flags  << I40E_TXD_QW1_CMD_SHIFT) |
+               ((uint64_t)pkt->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT));
+
+       __m128i descriptor = _mm_set_epi64x(high_qw,
+                               pkt->buf_iova + pkt->data_off);
+       _mm_store_si128((__m128i *)txdp, descriptor);
+}
+
+static inline void
+vtx(volatile struct i40e_tx_desc *txdp,
+       struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+       const uint64_t hi_qw_tmpl = (I40E_TX_DESC_DTYPE_DATA |
+                       ((uint64_t)flags  << I40E_TXD_QW1_CMD_SHIFT));
+
+       /* if unaligned on 32-bit boundary, do one to align */
+       if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+               vtx1(txdp, *pkt, flags);
+               nb_pkts--, txdp++, pkt++;
+       }
+
+       /* do two at a time while possible, in bursts */
+       for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
+               uint64_t hi_qw3 =
+                       hi_qw_tmpl |
+                       ((uint64_t)pkt[3]->data_len <<
+                        I40E_TXD_QW1_TX_BUF_SZ_SHIFT);
+               uint64_t hi_qw2 =
+                       hi_qw_tmpl |
+                       ((uint64_t)pkt[2]->data_len <<
+                        I40E_TXD_QW1_TX_BUF_SZ_SHIFT);
+               uint64_t hi_qw1 =
+                       hi_qw_tmpl |
+                       ((uint64_t)pkt[1]->data_len <<
+                        I40E_TXD_QW1_TX_BUF_SZ_SHIFT);
+               uint64_t hi_qw0 =
+                       hi_qw_tmpl |
+                       ((uint64_t)pkt[0]->data_len <<
+                        I40E_TXD_QW1_TX_BUF_SZ_SHIFT);
+
+               __m256i desc2_3 = _mm256_set_epi64x
+                       (hi_qw3, pkt[3]->buf_iova + pkt[3]->data_off,
+                       hi_qw2, pkt[2]->buf_iova + pkt[2]->data_off);
+               __m256i desc0_1 = _mm256_set_epi64x
+                       (hi_qw1, pkt[1]->buf_iova + pkt[1]->data_off,
+                       hi_qw0, pkt[0]->buf_iova + pkt[0]->data_off);
+               _mm256_store_si256((void *)(txdp + 2), desc2_3);
+               _mm256_store_si256((void *)txdp, desc0_1);
+       }
+
+       /* do any last ones */
+       while (nb_pkts) {
+               vtx1(txdp, *pkt, flags);
+               txdp++, pkt++, nb_pkts--;
+       }
+}
+
+static inline uint16_t
+i40e_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+                                uint16_t nb_pkts)
+{
+       struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
+       volatile struct i40e_tx_desc *txdp;
+       struct i40e_tx_entry *txep;
+       uint16_t n, nb_commit, tx_id;
+       uint64_t flags = I40E_TD_CMD;
+       uint64_t rs = I40E_TX_DESC_CMD_RS | I40E_TD_CMD;
+
+       /* cross rx_thresh boundary is not allowed */
+       nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+
+       if (txq->nb_tx_free < txq->tx_free_thresh)
+               i40e_tx_free_bufs(txq);
+
+       nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
+       if (unlikely(nb_pkts == 0))
+               return 0;
+
+       tx_id = txq->tx_tail;
+       txdp = &txq->tx_ring[tx_id];
+       txep = &txq->sw_ring[tx_id];
+
+       txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
+
+       n = (uint16_t)(txq->nb_tx_desc - tx_id);
+       if (nb_commit >= n) {
+               tx_backlog_entry(txep, tx_pkts, n);
+
+               vtx(txdp, tx_pkts, n - 1, flags);
+               tx_pkts += (n - 1);
+               txdp += (n - 1);
+
+               vtx1(txdp, *tx_pkts++, rs);
+
+               nb_commit = (uint16_t)(nb_commit - n);
+
+               tx_id = 0;
+               txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
+
+               /* avoid reach the end of ring */
+               txdp = &txq->tx_ring[tx_id];
+               txep = &txq->sw_ring[tx_id];
+       }
+
+       tx_backlog_entry(txep, tx_pkts, nb_commit);
+
+       vtx(txdp, tx_pkts, nb_commit, flags);
+
+       tx_id = (uint16_t)(tx_id + nb_commit);
+       if (tx_id > txq->tx_next_rs) {
+               txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
+                       rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
+                                               I40E_TXD_QW1_CMD_SHIFT);
+               txq->tx_next_rs =
+                       (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
+       }
+
+       txq->tx_tail = tx_id;
+
+       I40E_PCI_REG_WC_WRITE(txq->qtx_tail, txq->tx_tail);
+
+       return nb_pkts;
+}
+
+uint16_t
+i40e_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+                         uint16_t nb_pkts)
+{
+       uint16_t nb_tx = 0;
+       struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
+
+       while (nb_pkts) {
+               uint16_t ret, num;
+
+               num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+               ret = i40e_xmit_fixed_burst_vec_avx512
+                               (tx_queue, &tx_pkts[nb_tx], num);
+               nb_tx += ret;
+               nb_pkts -= ret;
+               if (ret < num)
+                       break;
+       }
+
+       return nb_tx;
+}
index 882168c..e4eb925 100644 (file)
@@ -45,6 +45,30 @@ if arch_subdir == 'x86'
                                c_args: [cflags, '-mavx2'])
                objs += i40e_avx2_lib.extract_objects('i40e_rxtx_vec_avx2.c')
        endif
+
+       i40e_avx512_cpu_support = (
+               cc.get_define('__AVX512F__', args: machine_args) != '' and
+               cc.get_define('__AVX512BW__', args: machine_args) != '')
+
+       i40e_avx512_cc_support = (
+               not machine_args.contains('-mno-avx512f') and
+               cc.has_argument('-mavx512f') and
+               cc.has_argument('-mavx512bw'))
+
+       if i40e_avx512_cpu_support == true or i40e_avx512_cc_support == true
+               cflags += ['-DCC_AVX512_SUPPORT']
+               avx512_args = [cflags, '-mavx512f', '-mavx512bw']
+               if cc.has_argument('-march=skylake-avx512')
+                       avx512_args += '-march=skylake-avx512'
+               endif
+               i40e_avx512_lib = static_library('i40e_avx512_lib',
+                               'i40e_rxtx_vec_avx512.c',
+                               dependencies: [static_rte_ethdev,
+                                       static_rte_kvargs, static_rte_hash],
+                               include_directories: includes,
+                               c_args: avx512_args)
+               objs += i40e_avx512_lib.extract_objects('i40e_rxtx_vec_avx512.c')
+       endif
 elif arch_subdir == 'ppc'
        dpdk_conf.set('RTE_LIBRTE_I40E_INC_VECTOR', 1)
        sources += files('i40e_rxtx_vec_altivec.c')