DPDK patches and discussions
 help / color / mirror / Atom feed
* [dpdk-dev] [PATCH 0/3] enable AVX512 for iavf
@ 2020-09-10  5:59 Wenzhuo Lu
  2020-09-10  5:59 ` [dpdk-dev] [PATCH 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
                   ` (8 more replies)
  0 siblings, 9 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-10  5:59 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu

AVX512 instructions is supported by more and more platforms. These
instructions can be used in the data path to enhance the per-core
performance of packet processing.
Comparing with the existing implementation, this path set introduces
some AVX512 instructions into the iavf data path, and we get a better
per-code throughput.

Wenzhuo Lu (3):
  net/iavf: enable AVX512 for legacy RX
  net/iavf: enable AVX512 for flexible RX
  net/iavf: enable AVX512 for TX

 doc/guides/rel_notes/release_20_11.rst  |    3 +
 drivers/net/iavf/iavf_ethdev.c          |    3 +-
 drivers/net/iavf/iavf_rxtx.c            |   69 +-
 drivers/net/iavf/iavf_rxtx.h            |   18 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 1720 +++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |    7 +
 6 files changed, 1808 insertions(+), 12 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH 1/3] net/iavf: enable AVX512 for legacy RX
  2020-09-10  5:59 [dpdk-dev] [PATCH 0/3] enable AVX512 for iavf Wenzhuo Lu
@ 2020-09-10  5:59 ` Wenzhuo Lu
  2020-09-10  9:29   ` Bruce Richardson
  2020-09-10  5:59 ` [dpdk-dev] [PATCH 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
                   ` (7 subsequent siblings)
  8 siblings, 1 reply; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-10  5:59 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the legacy RX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  27 +-
 drivers/net/iavf/iavf_rxtx.h            |   5 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 720 ++++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |   7 +
 4 files changed, 755 insertions(+), 4 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index 05a7dd8..c36e809 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2104,6 +2104,9 @@
 	struct iavf_rx_queue *rxq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_rx_vec_dev_check(dev)) {
 		for (i = 0; i < dev->data->nb_rx_queues; i++) {
@@ -2114,6 +2117,10 @@
 		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+			use_avx512 = true;
+#endif
 
 		if (dev->data->scattered_rx) {
 			PMD_DRV_LOG(DEBUG,
@@ -2121,27 +2128,39 @@
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
 					iavf_recv_scattered_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512;
+#endif
+			}
 		} else {
 			PMD_DRV_LOG(DEBUG, "Using %sVector Rx (port %d).",
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
 					iavf_recv_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512;
+#endif
+			}
 		}
 
 		return;
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 59625a9..cb12888 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -437,6 +437,11 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_tx_vec_dev_check(struct rte_eth_dev *dev);
 int iavf_rxq_vec_setup(struct iavf_rx_queue *rxq);
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
+uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
new file mode 100644
index 0000000..b528ed3
--- /dev/null
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -0,0 +1,720 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include "iavf_rxtx_vec_common.h"
+
+#include <x86intrin.h>
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+#define IAVF_DESCS_PER_LOOP_AVX 8
+#define PKTLEN_SHIFT 10
+
+__attribute__((optimize("unroll-loops"))) __rte_always_inline
+static inline void
+iavf_rxq_rearm(struct iavf_rx_queue *rxq)
+{
+	int i;
+	uint16_t rx_id;
+	volatile union iavf_rx_desc *rxdp;
+	struct rte_mempool_cache *cache = rte_mempool_default_cache(rxq->mp,
+			rte_lcore_id());
+	struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start];
+
+	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.
+	 */
+
+	/* Can this be satisfied from the cache? */
+	if (cache->len < IAVF_RXQ_REARM_THRESH) {
+		/* No. Backfill the cache first, and then fill from it */
+		uint32_t req = IAVF_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 + IAVF_RXQ_REARM_THRESH >=
+					rxq->nb_rx_desc) {
+				__m128i dma_addr0;
+
+				dma_addr0 = _mm_setzero_si128();
+				for (i = 0; i < IAVF_VPMD_DESCS_PER_LOOP; i++) {
+					rxp[i] = &rxq->fake_mbuf;
+					_mm_storeu_si128((__m128i *)&rxdp[i].read,
+							 dma_addr0);
+				}
+			}
+			rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+					IAVF_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_IAVF_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 < IAVF_RXQ_REARM_THRESH / IAVF_DESCS_PER_LOOP_AVX; i++) {
+		const __m512i mbuf_ptrs = _mm512_loadu_si512
+			(&cache->objs[cache->len - IAVF_DESCS_PER_LOOP_AVX]);
+		_mm512_storeu_si512(rxp, mbuf_ptrs);
+
+		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_IAVF_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 6 & 7.
+		 */
+		const __m512i desc0_1 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas0);
+		const __m512i desc2_3 = _mm512_bsrli_epi128(desc0_1, 8);
+
+		const __m512i desc4_5 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas1);
+		const __m512i desc6_7 = _mm512_bsrli_epi128(desc4_5, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_1);
+		_mm512_storeu_si512((void *)(rxdp + 2), desc2_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_5);
+		_mm512_storeu_si512((void *)(rxdp + 6), desc6_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 desc0_3 = _mm512_permutexvar_epi64(permute_idx, iova_addrs);
+		const __m512i desc4_7 = _mm512_bsrli_epi128(desc0_3, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_7);
+#endif
+		rxp += IAVF_DESCS_PER_LOOP_AVX;
+		rxdp += IAVF_DESCS_PER_LOOP_AVX;
+		cache->len -= IAVF_DESCS_PER_LOOP_AVX;
+	}
+
+	rxq->rxrearm_start += IAVF_RXQ_REARM_THRESH;
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= IAVF_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 */
+	IAVF_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
+}
+
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512(struct iavf_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 uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
+			0, rxq->mbuf_initializer);
+	/* struct iavf_rx_entry *sw_ring = &rxq->sw_ring[rxq->rx_tail]; */
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
+	const __mmask32 len_mask = _cvtu32_mask32(0x80808080);
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_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 << IAVF_RX_DESC_STATUS_DD_SHIFT)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi16
+			(/* 1st descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0,          /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0,          /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0,          /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0           /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_DESC_STATUS_EOF_SHIFT);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi8
+			(/* 1st descriptor */
+			 7, 6, 5, 4,  /* octet 4~7, 32bits rss */
+			 3, 2,        /* octet 2~3, low 16 bits vlan_macip */
+			 15, 14,      /* octet 15~14, 16 bits data_len */
+			 0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
+			 15, 14,      /* octet 15~14, low 16 bits pkt_len */
+			 0xFF, 0xFF,  /* pkt_type set as unknown */
+			 0xFF, 0xFF,  /*pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 7, 6, 5, 4,  /* octet 4~7, 32bits rss */
+			 3, 2,        /* octet 2~3, low 16 bits vlan_macip */
+			 15, 14,      /* octet 15~14, 16 bits data_len */
+			 0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
+			 15, 14,      /* octet 15~14, low 16 bits pkt_len */
+			 0xFF, 0xFF,  /* pkt_type set as unknown */
+			 0xFF, 0xFF,  /*pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 7, 6, 5, 4,  /* octet 4~7, 32bits rss */
+			 3, 2,        /* octet 2~3, low 16 bits vlan_macip */
+			 15, 14,      /* octet 15~14, 16 bits data_len */
+			 0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
+			 15, 14,      /* octet 15~14, low 16 bits pkt_len */
+			 0xFF, 0xFF,  /* pkt_type set as unknown */
+			 0xFF, 0xFF,  /*pkt_type set as unknown */
+			 /* 4th descriptor */
+			 7, 6, 5, 4,  /* octet 4~7, 32bits rss */
+			 3, 2,        /* octet 2~3, low 16 bits vlan_macip */
+			 15, 14,      /* octet 15~14, 16 bits data_len */
+			 0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
+			 15, 14,      /* octet 15~14, low 16 bits pkt_len */
+			 0xFF, 0xFF,  /* pkt_type set as unknown */
+			 0xFF, 0xFF   /*pkt_type set as unknown */
+			);
+	/**
+	 * 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 += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 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 desc4_7 = _mm512_mask_blend_epi16(len_mask,
+								raw_desc4_7,
+								len4_7);
+		__m512i mb4_7 = _mm512_shuffle_epi8(desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, 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 __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 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 __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, adjusting length and
+		 * re-arranging fields. Then write into the mbuf
+		 */
+		const __m512i len0_3 = _mm512_slli_epi32(raw_desc0_3,
+							 PKTLEN_SHIFT);
+		const __m512i desc0_3 = _mm512_mask_blend_epi16(len_mask,
+								raw_desc0_3,
+								len0_3);
+		__m512i mb0_3 = _mm512_shuffle_epi8(desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/* get the packet types */
+		const __m512i ptypes0_3 = _mm512_srli_epi64(desc0_3, 30);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		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 ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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_flags =
+			_mm256_shuffle_epi8(rss_flags_shuf,
+					    _mm256_srli_epi32(flag_bits, 11));
+		/**
+		 * 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 */
+		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				_mm256_or_si256(rss_flags, vlan_flags));
+		/**
+		 * 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;
+		const __m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7, 0);
+		const __m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7, 1);
+		const __m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3, 0);
+		const __m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3, 1);
+
+		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 << IAVF_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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
+static inline __m256i
+flex_rxd_to_fdir_flags_vec_avx512(const __m256i fdir_id0_7)
+{
+#define FDID_MIS_MAGIC 0xFFFFFFFF
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR != (1 << 2));
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
+	const __m256i pkt_fdir_bit = _mm256_set1_epi32(PKT_RX_FDIR |
+			PKT_RX_FDIR_ID);
+	/* desc->flow_id field == 0xFFFFFFFF means fdir mismatch */
+	const __m256i fdir_mis_mask = _mm256_set1_epi32(FDID_MIS_MAGIC);
+	__m256i fdir_mask = _mm256_cmpeq_epi32(fdir_id0_7,
+			fdir_mis_mask);
+	/* this XOR op results to bit-reverse the fdir_mask */
+	fdir_mask = _mm256_xor_si256(fdir_mask, fdir_mis_mask);
+	const __m256i fdir_flags = _mm256_and_si256(fdir_mask, pkt_fdir_bit);
+
+	return fdir_flags;
+}
+
+/**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+			  uint16_t nb_pkts)
+{
+	return _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				     uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				    uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst = iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
diff --git a/drivers/net/iavf/meson.build b/drivers/net/iavf/meson.build
index a3fad36..6427885 100644
--- a/drivers/net/iavf/meson.build
+++ b/drivers/net/iavf/meson.build
@@ -34,4 +34,11 @@ if arch_subdir == 'x86'
 				c_args: [cflags, '-mavx2'])
 		objs += iavf_avx2_lib.extract_objects('iavf_rxtx_vec_avx2.c')
 	endif
+
+	if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F')
+		cflags += ['-DCC_AVX512_SUPPORT']
+		cflags += ['-mavx512f']
+		cflags += ['-march=skylake-avx512']
+		sources += files('iavf_rxtx_vec_avx512.c')
+	endif
 endif
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH 2/3] net/iavf: enable AVX512 for flexible RX
  2020-09-10  5:59 [dpdk-dev] [PATCH 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-09-10  5:59 ` [dpdk-dev] [PATCH 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
@ 2020-09-10  5:59 ` Wenzhuo Lu
  2020-09-10  5:59 ` [dpdk-dev] [PATCH 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
                   ` (6 subsequent siblings)
  8 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-10  5:59 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to hande the flexible RX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  10 +
 drivers/net/iavf/iavf_rxtx.h            |   6 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 699 ++++++++++++++++++++++++++++++++
 3 files changed, 715 insertions(+)

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index c36e809..0818107 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2132,6 +2132,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
@@ -2151,6 +2156,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index cb12888..9653e0c 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -439,9 +439,15 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
 uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
 				   uint16_t nb_pkts);
+uint16_t iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					    struct rte_mbuf **rx_pkts,
+					    uint16_t nb_pkts);
 uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 					     struct rte_mbuf **rx_pkts,
 					     uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+						      struct rte_mbuf **rx_pkts,
+						      uint16_t nb_pkts);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index b528ed3..8c33661 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -643,6 +643,624 @@
 	return fdir_flags;
 }
 
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512_flex_rxd(struct iavf_rx_queue *rxq,
+					struct rte_mbuf **rx_pkts,
+					uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
+			0, rxq->mbuf_initializer);
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_flex_desc *rxdp =
+		(union iavf_rx_flex_desc *)rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_rxq_rearm(rxq);
+
+	/* Before we start moving massive data around, check to see if
+	 * there is actually a packet available
+	 */
+	if (!(rxdp->wb.status_error0 &
+			rte_cpu_to_le_32(1 << IAVF_RX_FLEX_DESC_STATUS0_DD_S)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi16
+			(/* 1st descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0,          /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0,          /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0,          /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0           /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi8
+			(/* 1st descriptor */
+			 0xFF, 0xFF,
+			 0xFF, 0xFF,    /* rss hash parsed separately */
+			 11, 10,	/* octet 10~11, 16 bits vlan_macip */
+			 5, 4,		/* octet 4~5, 16 bits data_len */
+			 0xFF, 0xFF,	/* skip hi 16 bits pkt_len, zero out */
+			 5, 4,		/* octet 4~5, 16 bits pkt_len */
+			 0xFF, 0xFF,	/* pkt_type set as unknown */
+			 0xFF, 0xFF,	/*pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 0xFF, 0xFF,
+			 0xFF, 0xFF,    /* rss hash parsed separately */
+			 11, 10,	/* octet 10~11, 16 bits vlan_macip */
+			 5, 4,		/* octet 4~5, 16 bits data_len */
+			 0xFF, 0xFF,	/* skip hi 16 bits pkt_len, zero out */
+			 5, 4,		/* octet 4~5, 16 bits pkt_len */
+			 0xFF, 0xFF,	/* pkt_type set as unknown */
+			 0xFF, 0xFF,	/*pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 0xFF, 0xFF,
+			 0xFF, 0xFF,    /* rss hash parsed separately */
+			 11, 10,	/* octet 10~11, 16 bits vlan_macip */
+			 5, 4,		/* octet 4~5, 16 bits data_len */
+			 0xFF, 0xFF,	/* skip hi 16 bits pkt_len, zero out */
+			 5, 4,		/* octet 4~5, 16 bits pkt_len */
+			 0xFF, 0xFF,	/* pkt_type set as unknown */
+			 0xFF, 0xFF,	/*pkt_type set as unknown */
+			 /* 4th descriptor */
+			 0xFF, 0xFF,
+			 0xFF, 0xFF,    /* rss hash parsed separately */
+			 11, 10,	/* octet 10~11, 16 bits vlan_macip */
+			 5, 4,		/* octet 4~5, 16 bits data_len */
+			 0xFF, 0xFF,	/* skip hi 16 bits pkt_len, zero out */
+			 5, 4,		/* octet 4~5, 16 bits pkt_len */
+			 0xFF, 0xFF,	/* pkt_type set as unknown */
+			 0xFF, 0xFF	/*pkt_type set as unknown */
+			);
+	/**
+	 * 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 Checksum Reports, RSS indication
+	 * and VLAN indication.
+	 * bit6:4 for IP/L4 checksum errors.
+	 * bit12 is for RSS indication.
+	 * bit13 is for VLAN indication.
+	 */
+	const __m256i flags_mask =
+		 _mm256_set1_epi32((7 << 4) | (1 << 12) | (1 << 13));
+	/**
+	 * data to be shuffled by the result of the flags mask shifted by 4
+	 * 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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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);
+	/**
+	 * data to be shuffled by result of flag mask, shifted down 12.
+	 * If RSS(bit12)/VLAN(bit13) are set,
+	 * shuffle moves appropriate flags in place.
+	 */
+	const __m256i rss_vlan_flags_shuf = _mm256_set_epi8(0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0,
+			/* end up 128-bits */
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0);
+
+	uint16_t i, received;
+
+	for (i = 0, received = 0; i < nb_pkts;
+	     i += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb4_7 = _mm512_shuffle_epi8(raw_desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptype_mask =
+			_mm512_set1_epi16(IAVF_RX_FLEX_DESC_PTYPE_M);
+		const __m512i ptypes4_7 =
+			_mm512_and_si512(raw_desc4_7, ptype_mask);
+		const __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 0);
+		const uint16_t ptype7 = _mm256_extract_epi16(ptypes6_7, 9);
+		const uint16_t ptype6 = _mm256_extract_epi16(ptypes6_7, 1);
+		const uint16_t ptype5 = _mm256_extract_epi16(ptypes4_5, 9);
+		const uint16_t ptype4 = _mm256_extract_epi16(ptypes4_5, 1);
+
+		const __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb0_3 = _mm512_shuffle_epi8(raw_desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptypes0_3 =
+			_mm512_and_si512(raw_desc0_3, ptype_mask);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		const uint16_t ptype3 = _mm256_extract_epi16(ptypes2_3, 9);
+		const uint16_t ptype2 = _mm256_extract_epi16(ptypes2_3, 1);
+		const uint16_t ptype1 = _mm256_extract_epi16(ptypes0_1, 9);
+		const uint16_t ptype0 = _mm256_extract_epi16(ptypes0_1, 1);
+
+		const __m512i ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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);
+		/**
+		 * 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, 4));
+		l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
+		l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
+		/* set rss and vlan flags */
+		const __m256i rss_vlan_flag_bits =
+			_mm256_srli_epi32(flag_bits, 12);
+		const __m256i rss_vlan_flags =
+			_mm256_shuffle_epi8(rss_vlan_flags_shuf,
+					    rss_vlan_flag_bits);
+
+		/* merge flags */
+		__m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				rss_vlan_flags);
+
+		if (rxq->fdir_enabled) {
+			const __m512i fdir_permute_mask = _mm512_set_epi32
+				(0, 0, 0, 0,
+				 0, 0, 0, 0,
+				 15, 11, 7, 3,
+				 31, 27, 23, 19);
+			__m512i fdir_tmp = _mm512_permutex2var_epi32
+				(raw_desc4_7, fdir_permute_mask, raw_desc0_3);
+			const __m256i fdir_id0_7 = _mm512_extracti64x4_epi64
+				(fdir_tmp, 0);
+			const __m256i fdir_flags =
+				flex_rxd_to_fdir_flags_vec_avx512(fdir_id0_7);
+
+			/* merge with fdir_flags */
+			mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_flags);
+
+			/* write to mbuf: have to use scalar store here */
+			rx_pkts[i + 0]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 3);
+
+			rx_pkts[i + 1]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 7);
+
+			rx_pkts[i + 2]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 2);
+
+			rx_pkts[i + 3]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 6);
+
+			rx_pkts[i + 4]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 1);
+
+			rx_pkts[i + 5]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 5);
+
+			rx_pkts[i + 6]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 0);
+
+			rx_pkts[i + 7]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 4);
+		} /* if() on fdir_enabled */
+
+		__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);
+
+#ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
+		/**
+		 * needs to load 2nd 16B of each desc for RSS hash parsing,
+		 * will cause performance drop to get into this context.
+		 */
+		if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
+				DEV_RX_OFFLOAD_RSS_HASH) {
+			/* load bottom half of every 32B desc */
+			const __m128i raw_desc_bh7 =
+				_mm_load_si128
+					((void *)(&rxdp[7].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh6 =
+				_mm_load_si128
+					((void *)(&rxdp[6].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh5 =
+				_mm_load_si128
+					((void *)(&rxdp[5].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh4 =
+				_mm_load_si128
+					((void *)(&rxdp[4].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh3 =
+				_mm_load_si128
+					((void *)(&rxdp[3].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh2 =
+				_mm_load_si128
+					((void *)(&rxdp[2].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh1 =
+				_mm_load_si128
+					((void *)(&rxdp[1].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh0 =
+				_mm_load_si128
+					((void *)(&rxdp[0].wb.status_error1));
+
+			__m256i raw_desc_bh6_7 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh6),
+					raw_desc_bh7, 1);
+			__m256i raw_desc_bh4_5 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh4),
+					raw_desc_bh5, 1);
+			__m256i raw_desc_bh2_3 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh2),
+					raw_desc_bh3, 1);
+			__m256i raw_desc_bh0_1 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh0),
+					raw_desc_bh1, 1);
+
+			/**
+			 * to shift the 32b RSS hash value to the
+			 * highest 32b of each 128b before mask
+			 */
+			__m256i rss_hash6_7 =
+				_mm256_slli_epi64(raw_desc_bh6_7, 32);
+			__m256i rss_hash4_5 =
+				_mm256_slli_epi64(raw_desc_bh4_5, 32);
+			__m256i rss_hash2_3 =
+				_mm256_slli_epi64(raw_desc_bh2_3, 32);
+			__m256i rss_hash0_1 =
+				_mm256_slli_epi64(raw_desc_bh0_1, 32);
+
+			__m256i rss_hash_msk =
+				_mm256_set_epi32(0xFFFFFFFF, 0, 0, 0,
+						 0xFFFFFFFF, 0, 0, 0);
+
+			rss_hash6_7 = _mm256_and_si256
+					(rss_hash6_7, rss_hash_msk);
+			rss_hash4_5 = _mm256_and_si256
+					(rss_hash4_5, rss_hash_msk);
+			rss_hash2_3 = _mm256_and_si256
+					(rss_hash2_3, rss_hash_msk);
+			rss_hash0_1 = _mm256_and_si256
+					(rss_hash0_1, rss_hash_msk);
+
+			mb6_7 = _mm256_or_si256(mb6_7, rss_hash6_7);
+			mb4_5 = _mm256_or_si256(mb4_5, rss_hash4_5);
+			mb2_3 = _mm256_or_si256(mb2_3, rss_hash2_3);
+			mb0_1 = _mm256_or_si256(mb0_1, rss_hash0_1);
+		} /* if() on RSS hash parsing */
+#endif
+
+		/**
+		 * 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 <<
+					       IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+			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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
 /**
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -655,6 +1273,18 @@
 }
 
 /**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts)
+{
+	return _iavf_recv_raw_pkts_vec_avx512_flex_rxd(rx_queue, rx_pkts,
+						       nb_pkts, NULL);
+}
+
+/**
  * vPMD receive routine that reassembles single burst of 32 scattered packets
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -718,3 +1348,72 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+/**
+ * vPMD receive routine that reassembles single burst of
+ * 32 scattered packets for flex RxD
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512_flex_rxd(void *rx_queue,
+					      struct rte_mbuf **rx_pkts,
+					      uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_recv_raw_pkts_vec_avx512_flex_rxd(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 for flex RxD.
+ * Main receive routine that can handle arbitrary burst sizes
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst =
+			iavf_recv_scattered_burst_vec_avx512_flex_rxd
+			(rx_queue, rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH 3/3] net/iavf: enable AVX512 for TX
  2020-09-10  5:59 [dpdk-dev] [PATCH 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-09-10  5:59 ` [dpdk-dev] [PATCH 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
  2020-09-10  5:59 ` [dpdk-dev] [PATCH 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
@ 2020-09-10  5:59 ` Wenzhuo Lu
  2020-09-15  1:17   ` Wang, Haiyue
  2020-09-17  1:39 ` [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf Wenzhuo Lu
                   ` (5 subsequent siblings)
  8 siblings, 1 reply; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-10  5:59 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the TX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 doc/guides/rel_notes/release_20_11.rst  |   3 +
 drivers/net/iavf/iavf_ethdev.c          |   3 +-
 drivers/net/iavf/iavf_rxtx.c            |  32 +++-
 drivers/net/iavf/iavf_rxtx.h            |   7 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 301 ++++++++++++++++++++++++++++++++
 5 files changed, 338 insertions(+), 8 deletions(-)

diff --git a/doc/guides/rel_notes/release_20_11.rst b/doc/guides/rel_notes/release_20_11.rst
index df227a1..d40b8d6 100644
--- a/doc/guides/rel_notes/release_20_11.rst
+++ b/doc/guides/rel_notes/release_20_11.rst
@@ -55,6 +55,9 @@ New Features
      Also, make sure to start the actual text at the margin.
      =======================================================
 
+   * **Added support of vector instructions on IAVF.**
+
+     Added support of AVX512 instructions in IAVF RX and TX path.
 
 Removed Items
 -------------
diff --git a/drivers/net/iavf/iavf_ethdev.c b/drivers/net/iavf/iavf_ethdev.c
index c3aa4cd..5bc2851 100644
--- a/drivers/net/iavf/iavf_ethdev.c
+++ b/drivers/net/iavf/iavf_ethdev.c
@@ -528,7 +528,8 @@ static int iavf_config_rx_queues_irqs(struct rte_eth_dev *dev,
 		DEV_TX_OFFLOAD_GRE_TNL_TSO |
 		DEV_TX_OFFLOAD_IPIP_TNL_TSO |
 		DEV_TX_OFFLOAD_GENEVE_TNL_TSO |
-		DEV_TX_OFFLOAD_MULTI_SEGS;
+		DEV_TX_OFFLOAD_MULTI_SEGS |
+		DEV_TX_OFFLOAD_MBUF_FAST_FREE;
 
 	dev_info->default_rxconf = (struct rte_eth_rxconf) {
 		.rx_free_thresh = IAVF_DEFAULT_RX_FREE_THRESH,
diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index 0818107..04dcd48 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2206,18 +2206,18 @@
 	struct iavf_tx_queue *txq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_tx_vec_dev_check(dev)) {
-		for (i = 0; i < dev->data->nb_tx_queues; i++) {
-			txq = dev->data->tx_queues[i];
-			if (!txq)
-				continue;
-			iavf_txq_vec_setup(txq);
-		}
-
 		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+			use_avx512 = true;
+#endif
 
 		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
 			    use_avx2 ? "avx2 " : "",
@@ -2225,8 +2225,26 @@
 		dev->tx_pkt_burst = use_avx2 ?
 				    iavf_xmit_pkts_vec_avx2 :
 				    iavf_xmit_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+		if (use_avx512)
+			dev->tx_pkt_burst = iavf_xmit_pkts_vec_avx512;
+#endif
 		dev->tx_pkt_prepare = NULL;
 
+		for (i = 0; i < dev->data->nb_tx_queues; i++) {
+			txq = dev->data->tx_queues[i];
+			if (!txq)
+				continue;
+#ifdef CC_AVX512_SUPPORT
+			if (use_avx512)
+				iavf_txq_vec_setup_avx512(txq);
+			else
+				iavf_txq_vec_setup(txq);
+#else
+			iavf_txq_vec_setup(txq);
+#endif
+		}
+
 		return;
 	}
 #endif
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 9653e0c..08eebb0 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -122,6 +122,10 @@ struct iavf_tx_entry {
 	uint16_t last_id;
 };
 
+struct iavf_tx_vec_entry {
+	struct rte_mbuf *mbuf;
+};
+
 /* Structure associated with each TX queue. */
 struct iavf_tx_queue {
 	const struct rte_memzone *mz;  /* memzone for Tx ring */
@@ -448,6 +452,9 @@ uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
 						      struct rte_mbuf **rx_pkts,
 						      uint16_t nb_pkts);
+uint16_t iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				   uint16_t nb_pkts);
+int iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index 8c33661..6c75d04 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -1417,3 +1417,304 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+static __rte_always_inline int
+iavf_tx_free_bufs_avx512(struct iavf_tx_queue *txq)
+{
+	struct iavf_tx_vec_entry *txep;
+	uint32_t n;
+	uint32_t i;
+	int nb_free = 0;
+	struct rte_mbuf *m, *free[IAVF_VPMD_TX_MAX_FREE_BUF];
+
+	/* check DD bits on threshold descriptor */
+	if ((txq->tx_ring[txq->next_dd].cmd_type_offset_bsz &
+			rte_cpu_to_le_64(IAVF_TXD_QW1_DTYPE_MASK)) !=
+			rte_cpu_to_le_64(IAVF_TX_DESC_DTYPE_DESC_DONE))
+		return 0;
+
+	n = txq->rs_thresh;
+
+	 /* first buffer to free from S/W ring is at index
+	  * tx_next_dd - (tx_rs_thresh-1)
+	  */
+	txep = (void *)txq->sw_ring;
+	txep += txq->next_dd - (n - 1);
+
+	if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE && (n & 31) == 0) {
+		struct rte_mempool *mp = txep[0].mbuf->pool;
+		struct rte_mempool_cache *cache = rte_mempool_default_cache(mp,
+				rte_lcore_id());
+		void **cache_objs = &cache->objs[cache->len];
+
+		if (n > RTE_MEMPOOL_CACHE_MAX_SIZE) {
+			rte_mempool_ops_enqueue_bulk(mp, (void *)txep, n);
+			goto done;
+		}
+
+		/* The cache follows the following algorithm
+		 *   1. Add the objects to the cache
+		 *   2. Anything greater than the cache min value (if it crosses the
+		 *   cache flush threshold) is flushed to the ring.
+		 */
+		/* Add elements back into the cache */
+		uint32_t copied = 0;
+		/* n is multiple of 32 */
+		while (copied < n) {
+			const __m512i a = _mm512_loadu_si512(&txep[copied]);
+			const __m512i b = _mm512_loadu_si512(&txep[copied + 8]);
+			const __m512i c = _mm512_loadu_si512(&txep[copied + 16]);
+			const __m512i d = _mm512_loadu_si512(&txep[copied + 24]);
+
+			_mm512_storeu_si512(&cache_objs[copied], a);
+			_mm512_storeu_si512(&cache_objs[copied + 8], b);
+			_mm512_storeu_si512(&cache_objs[copied + 16], c);
+			_mm512_storeu_si512(&cache_objs[copied + 24], d);
+			copied += 32;
+		}
+		cache->len += n;
+
+		if (cache->len >= cache->flushthresh) {
+			rte_mempool_ops_enqueue_bulk(mp,
+						     &cache->objs[cache->size],
+						     cache->len - cache->size);
+			cache->len = cache->size;
+		}
+		goto done;
+	}
+
+	m = rte_pktmbuf_prefree_seg(txep[0].mbuf);
+	if (likely(m)) {
+		free[0] = m;
+		nb_free = 1;
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (likely(m)) {
+				if (likely(m->pool == free[0]->pool)) {
+					free[nb_free++] = m;
+				} else {
+					rte_mempool_put_bulk(free[0]->pool,
+							     (void *)free,
+							     nb_free);
+					free[0] = m;
+					nb_free = 1;
+				}
+			}
+		}
+		rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
+	} else {
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (m)
+				rte_mempool_put(m->pool, m);
+		}
+	}
+
+done:
+	/* buffers were freed, update counters */
+	txq->nb_free = (uint16_t)(txq->nb_free + txq->rs_thresh);
+	txq->next_dd = (uint16_t)(txq->next_dd + txq->rs_thresh);
+	if (txq->next_dd >= txq->nb_tx_desc)
+		txq->next_dd = (uint16_t)(txq->rs_thresh - 1);
+
+	return txq->rs_thresh;
+}
+
+static __rte_always_inline void
+tx_backlog_entry_avx512(struct iavf_tx_vec_entry *txep,
+			struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	int i;
+
+	for (i = 0; i < (int)nb_pkts; ++i)
+		txep[i].mbuf = tx_pkts[i];
+}
+
+static inline void
+iavf_vtx1(volatile struct iavf_tx_desc *txdp,
+	  struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(IAVF_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT) |
+		 ((uint64_t)pkt->data_len << IAVF_TXD_QW1_TX_BUF_SZ_SHIFT));
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+				pkt->buf_physaddr + pkt->data_off);
+	_mm_storeu_si128((__m128i *)txdp, descriptor);
+}
+
+static inline void
+iavf_vtx(volatile struct iavf_tx_desc *txdp,
+	 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (IAVF_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT));
+	const __mmask8 len_msk = _cvtu32_mask8(0xAA);
+	const __mmask8 off_msk = _cvtu32_mask8(0x55);
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		iavf_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) {
+		__m512i desc4 =
+			_mm512_set_epi64
+				((uint64_t)pkt[3]->data_len,
+				 pkt[3]->buf_physaddr,
+				 (uint64_t)pkt[2]->data_len,
+				 pkt[2]->buf_physaddr,
+				 (uint64_t)pkt[1]->data_len,
+				 pkt[1]->buf_physaddr,
+				 (uint64_t)pkt[0]->data_len,
+				 pkt[0]->buf_physaddr);
+		__m512i hi_qw_tmpl_4 = _mm512_set1_epi64(hi_qw_tmpl);
+		__m512i data_off_4 =
+			_mm512_set_epi64
+				(0,
+				 pkt[3]->data_off,
+				 0,
+				 pkt[2]->data_off,
+				 0,
+				 pkt[1]->data_off,
+				 0,
+				 pkt[0]->data_off);
+
+		desc4 = _mm512_mask_slli_epi64(desc4, len_msk, desc4, IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
+		desc4 = _mm512_mask_or_epi64(desc4, len_msk, desc4, hi_qw_tmpl_4);
+		desc4 = _mm512_mask_add_epi64(desc4, off_msk, desc4, data_off_4);
+		_mm512_storeu_si512((void *)txdp, desc4);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		iavf_vtx1(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+iavf_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				 uint16_t nb_pkts)
+{
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+	volatile struct iavf_tx_desc *txdp;
+	struct iavf_tx_vec_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	/* bit2 is reserved and must be set to 1 according to Spec */
+	uint64_t flags = IAVF_TX_DESC_CMD_EOP | IAVF_TX_DESC_CMD_ICRC;
+	uint64_t rs = IAVF_TX_DESC_CMD_RS | flags;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
+
+	if (txq->nb_free < txq->free_thresh)
+		iavf_tx_free_bufs_avx512(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = (void *)txq->sw_ring;
+	txep += tx_id;
+
+	txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		tx_backlog_entry_avx512(txep, tx_pkts, n);
+
+		iavf_vtx(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		iavf_vtx1(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = (void *)txq->sw_ring;
+		txep += tx_id;
+	}
+
+	tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
+
+	iavf_vtx(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->next_rs) {
+		txq->tx_ring[txq->next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)IAVF_TX_DESC_CMD_RS) <<
+					 IAVF_TXD_QW1_CMD_SHIFT);
+		txq->next_rs =
+			(uint16_t)(txq->next_rs + txq->rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	IAVF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+			  uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->rs_thresh);
+		ret = iavf_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;
+}
+
+static inline void
+iavf_tx_queue_release_mbufs_avx512(struct iavf_tx_queue *txq)
+{
+	unsigned int i;
+	const uint16_t max_desc = (uint16_t)(txq->nb_tx_desc - 1);
+	struct iavf_tx_vec_entry *swr = (void *)txq->sw_ring;
+
+	if (!txq->sw_ring || txq->nb_free == max_desc)
+		return;
+
+	i = txq->next_dd - txq->rs_thresh + 1;
+	if (txq->tx_tail < i) {
+		for (; i < txq->nb_tx_desc; i++) {
+			rte_pktmbuf_free_seg(swr[i].mbuf);
+			swr[i].mbuf = NULL;
+		}
+		i = 0;
+	}
+}
+
+static const struct iavf_txq_ops avx512_vec_txq_ops = {
+	.release_mbufs = iavf_tx_queue_release_mbufs_avx512,
+};
+
+int __rte_cold
+iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq)
+{
+	txq->ops = &avx512_vec_txq_ops;
+	return 0;
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* Re: [dpdk-dev] [PATCH 1/3] net/iavf: enable AVX512 for legacy RX
  2020-09-10  5:59 ` [dpdk-dev] [PATCH 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
@ 2020-09-10  9:29   ` Bruce Richardson
  2020-09-11  3:06     ` Lu, Wenzhuo
  0 siblings, 1 reply; 39+ messages in thread
From: Bruce Richardson @ 2020-09-10  9:29 UTC (permalink / raw)
  To: Wenzhuo Lu; +Cc: dev, Leyi Rong

On Thu, Sep 10, 2020 at 01:59:03PM +0800, Wenzhuo Lu wrote:
> To enhance the per-core performance, this patch adds some AVX512
> instructions to the data path to handle the legacy RX descriptors.
> 
> Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> Signed-off-by: Leyi Rong <leyi.rong@intel.com>
> ---
>  drivers/net/iavf/iavf_rxtx.c            |  27 +-
>  drivers/net/iavf/iavf_rxtx.h            |   5 +
>  drivers/net/iavf/iavf_rxtx_vec_avx512.c | 720 ++++++++++++++++++++++++++++++++
>  drivers/net/iavf/meson.build            |   7 +
>  4 files changed, 755 insertions(+), 4 deletions(-)
>  create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c
> 
<snip>
> diff --git a/drivers/net/iavf/meson.build b/drivers/net/iavf/meson.build
> index a3fad36..6427885 100644
> --- a/drivers/net/iavf/meson.build
> +++ b/drivers/net/iavf/meson.build
> @@ -34,4 +34,11 @@ if arch_subdir == 'x86'
>  				c_args: [cflags, '-mavx2'])
>  		objs += iavf_avx2_lib.extract_objects('iavf_rxtx_vec_avx2.c')
>  	endif
> +
> +	if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F')
> +		cflags += ['-DCC_AVX512_SUPPORT']
> +		cflags += ['-mavx512f']
> +		cflags += ['-march=skylake-avx512']
> +		sources += files('iavf_rxtx_vec_avx512.c')
> +	endif

This logic is probably not what you want, since if the machine cpuflag
value is set, then AVX512 instructions are already available, meaning you
don't need to add in the mavx512 and march flags. Instead the logic should
be handling the case where it's not enabled, but the compiler supports it.
Use the AVX2 build logic as a reference.

>  endif
> -- 
> 1.9.3
> 

^ permalink raw reply	[flat|nested] 39+ messages in thread

* Re: [dpdk-dev] [PATCH 1/3] net/iavf: enable AVX512 for legacy RX
  2020-09-10  9:29   ` Bruce Richardson
@ 2020-09-11  3:06     ` Lu, Wenzhuo
  0 siblings, 0 replies; 39+ messages in thread
From: Lu, Wenzhuo @ 2020-09-11  3:06 UTC (permalink / raw)
  To: Richardson, Bruce; +Cc: dev, Rong, Leyi

Hi Bruce,


> -----Original Message-----
> From: Bruce Richardson <bruce.richardson@intel.com>
> Sent: Thursday, September 10, 2020 5:30 PM
> To: Lu, Wenzhuo <wenzhuo.lu@intel.com>
> Cc: dev@dpdk.org; Rong, Leyi <leyi.rong@intel.com>
> Subject: Re: [PATCH 1/3] net/iavf: enable AVX512 for legacy RX
> 
> On Thu, Sep 10, 2020 at 01:59:03PM +0800, Wenzhuo Lu wrote:
> > To enhance the per-core performance, this patch adds some AVX512
> > instructions to the data path to handle the legacy RX descriptors.
> >
> > Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
> > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > Signed-off-by: Leyi Rong <leyi.rong@intel.com>
> > ---
> >  drivers/net/iavf/iavf_rxtx.c            |  27 +-
> >  drivers/net/iavf/iavf_rxtx.h            |   5 +
> >  drivers/net/iavf/iavf_rxtx_vec_avx512.c | 720
> ++++++++++++++++++++++++++++++++
> >  drivers/net/iavf/meson.build            |   7 +
> >  4 files changed, 755 insertions(+), 4 deletions(-)  create mode
> > 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c
> >
> <snip>
> > diff --git a/drivers/net/iavf/meson.build
> > b/drivers/net/iavf/meson.build index a3fad36..6427885 100644
> > --- a/drivers/net/iavf/meson.build
> > +++ b/drivers/net/iavf/meson.build
> > @@ -34,4 +34,11 @@ if arch_subdir == 'x86'
> >  				c_args: [cflags, '-mavx2'])
> >  		objs += iavf_avx2_lib.extract_objects('iavf_rxtx_vec_avx2.c')
> >  	endif
> > +
> > +	if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F')
> > +		cflags += ['-DCC_AVX512_SUPPORT']
> > +		cflags += ['-mavx512f']
> > +		cflags += ['-march=skylake-avx512']
> > +		sources += files('iavf_rxtx_vec_avx512.c')
> > +	endif
> 
> This logic is probably not what you want, since if the machine cpuflag value is
> set, then AVX512 instructions are already available, meaning you don't need to
> add in the mavx512 and march flags. Instead the logic should be handling the
> case where it's not enabled, but the compiler supports it.
> Use the AVX2 build logic as a reference.
Thanks for the review. I'll send a  new version.
> 
> >  endif
> > --
> > 1.9.3
> >

^ permalink raw reply	[flat|nested] 39+ messages in thread

* Re: [dpdk-dev] [PATCH 3/3] net/iavf: enable AVX512 for TX
  2020-09-10  5:59 ` [dpdk-dev] [PATCH 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
@ 2020-09-15  1:17   ` Wang, Haiyue
  2020-09-17  1:29     ` Lu, Wenzhuo
  0 siblings, 1 reply; 39+ messages in thread
From: Wang, Haiyue @ 2020-09-15  1:17 UTC (permalink / raw)
  To: Lu, Wenzhuo, dev; +Cc: Lu, Wenzhuo, Richardson, Bruce, Rong, Leyi

> -----Original Message-----
> From: dev <dev-bounces@dpdk.org> On Behalf Of Wenzhuo Lu
> Sent: Thursday, September 10, 2020 13:59
> To: dev@dpdk.org
> Cc: Lu, Wenzhuo <wenzhuo.lu@intel.com>; Richardson, Bruce <bruce.richardson@intel.com>; Rong, Leyi
> <leyi.rong@intel.com>
> Subject: [dpdk-dev] [PATCH 3/3] net/iavf: enable AVX512 for TX
> 
> To enhance the per-core performance, this patch adds some AVX512
> instructions to the data path to handle the TX descriptors.
> 
> Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> Signed-off-by: Leyi Rong <leyi.rong@intel.com>
> ---
>  doc/guides/rel_notes/release_20_11.rst  |   3 +
>  drivers/net/iavf/iavf_ethdev.c          |   3 +-
>  drivers/net/iavf/iavf_rxtx.c            |  32 +++-
>  drivers/net/iavf/iavf_rxtx.h            |   7 +
>  drivers/net/iavf/iavf_rxtx_vec_avx512.c | 301 ++++++++++++++++++++++++++++++++
>  5 files changed, 338 insertions(+), 8 deletions(-)
> 


> +		__m512i desc4 =
> +			_mm512_set_epi64
> +				((uint64_t)pkt[3]->data_len,
> +				 pkt[3]->buf_physaddr,

'buf_physaddr' will be remove in 20.11, we need to use 'buf_iova' instead.


> --
> 1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* Re: [dpdk-dev] [PATCH 3/3] net/iavf: enable AVX512 for TX
  2020-09-15  1:17   ` Wang, Haiyue
@ 2020-09-17  1:29     ` Lu, Wenzhuo
  0 siblings, 0 replies; 39+ messages in thread
From: Lu, Wenzhuo @ 2020-09-17  1:29 UTC (permalink / raw)
  To: Wang, Haiyue, dev; +Cc: Richardson, Bruce, Rong, Leyi

Hi Haiyue,

> -----Original Message-----
> From: Wang, Haiyue <haiyue.wang@intel.com>
> Sent: Tuesday, September 15, 2020 9:17 AM
> To: Lu, Wenzhuo <wenzhuo.lu@intel.com>; dev@dpdk.org
> Cc: Lu, Wenzhuo <wenzhuo.lu@intel.com>; Richardson, Bruce
> <bruce.richardson@intel.com>; Rong, Leyi <leyi.rong@intel.com>
> Subject: RE: [dpdk-dev] [PATCH 3/3] net/iavf: enable AVX512 for TX
> 
> > -----Original Message-----
> > From: dev <dev-bounces@dpdk.org> On Behalf Of Wenzhuo Lu
> > Sent: Thursday, September 10, 2020 13:59
> > To: dev@dpdk.org
> > Cc: Lu, Wenzhuo <wenzhuo.lu@intel.com>; Richardson, Bruce
> > <bruce.richardson@intel.com>; Rong, Leyi <leyi.rong@intel.com>
> > Subject: [dpdk-dev] [PATCH 3/3] net/iavf: enable AVX512 for TX
> >
> > To enhance the per-core performance, this patch adds some AVX512
> > instructions to the data path to handle the TX descriptors.
> >
> > Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
> > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > Signed-off-by: Leyi Rong <leyi.rong@intel.com>
> > ---
> >  doc/guides/rel_notes/release_20_11.rst  |   3 +
> >  drivers/net/iavf/iavf_ethdev.c          |   3 +-
> >  drivers/net/iavf/iavf_rxtx.c            |  32 +++-
> >  drivers/net/iavf/iavf_rxtx.h            |   7 +
> >  drivers/net/iavf/iavf_rxtx_vec_avx512.c | 301
> > ++++++++++++++++++++++++++++++++
> >  5 files changed, 338 insertions(+), 8 deletions(-)
> >
> 
> 
> > +		__m512i desc4 =
> > +			_mm512_set_epi64
> > +				((uint64_t)pkt[3]->data_len,
> > +				 pkt[3]->buf_physaddr,
> 
> 'buf_physaddr' will be remove in 20.11, we need to use 'buf_iova' instead.
Thanks for the reminder. Will change it in V2.

^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf
  2020-09-10  5:59 [dpdk-dev] [PATCH 0/3] enable AVX512 for iavf Wenzhuo Lu
                   ` (2 preceding siblings ...)
  2020-09-10  5:59 ` [dpdk-dev] [PATCH 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
@ 2020-09-17  1:39 ` Wenzhuo Lu
  2020-09-17  1:39   ` [dpdk-dev] [PATCH v2 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
                     ` (3 more replies)
  2020-09-21  8:13 ` [dpdk-dev] [PATCH v3 " Wenzhuo Lu
                   ` (4 subsequent siblings)
  8 siblings, 4 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-17  1:39 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu

AVX512 instructions is supported by more and more platforms. These instructions
can be used in the data path to enhance the per-core performance of packet
processing.
Comparing with the existing implementation, this path set introduces some AVX512
instructions into the iavf data path, and we get a better per-code throughput.

v2:
Update meson.build.
Repalce the deprecated 'buf_physaddr' by 'buf_iova'.

Wenzhuo Lu (3):
  net/iavf: enable AVX512 for legacy RX
  net/iavf: enable AVX512 for flexible RX
  net/iavf: enable AVX512 for TX

 doc/guides/rel_notes/release_20_11.rst  |    3 +
 drivers/net/iavf/iavf_ethdev.c          |    3 +-
 drivers/net/iavf/iavf_rxtx.c            |   69 +-
 drivers/net/iavf/iavf_rxtx.h            |   18 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 1720 +++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |   17 +
 6 files changed, 1818 insertions(+), 12 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v2 1/3] net/iavf: enable AVX512 for legacy RX
  2020-09-17  1:39 ` [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf Wenzhuo Lu
@ 2020-09-17  1:39   ` Wenzhuo Lu
  2020-09-17  1:39   ` [dpdk-dev] [PATCH v2 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
                     ` (2 subsequent siblings)
  3 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-17  1:39 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the legacy RX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  27 +-
 drivers/net/iavf/iavf_rxtx.h            |   5 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 720 ++++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |  17 +
 4 files changed, 765 insertions(+), 4 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index 05a7dd8..c36e809 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2104,6 +2104,9 @@
 	struct iavf_rx_queue *rxq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_rx_vec_dev_check(dev)) {
 		for (i = 0; i < dev->data->nb_rx_queues; i++) {
@@ -2114,6 +2117,10 @@
 		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+			use_avx512 = true;
+#endif
 
 		if (dev->data->scattered_rx) {
 			PMD_DRV_LOG(DEBUG,
@@ -2121,27 +2128,39 @@
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
 					iavf_recv_scattered_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512;
+#endif
+			}
 		} else {
 			PMD_DRV_LOG(DEBUG, "Using %sVector Rx (port %d).",
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
 					iavf_recv_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512;
+#endif
+			}
 		}
 
 		return;
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 59625a9..cb12888 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -437,6 +437,11 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_tx_vec_dev_check(struct rte_eth_dev *dev);
 int iavf_rxq_vec_setup(struct iavf_rx_queue *rxq);
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
+uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
new file mode 100644
index 0000000..b528ed3
--- /dev/null
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -0,0 +1,720 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include "iavf_rxtx_vec_common.h"
+
+#include <x86intrin.h>
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+#define IAVF_DESCS_PER_LOOP_AVX 8
+#define PKTLEN_SHIFT 10
+
+__attribute__((optimize("unroll-loops"))) __rte_always_inline
+static inline void
+iavf_rxq_rearm(struct iavf_rx_queue *rxq)
+{
+	int i;
+	uint16_t rx_id;
+	volatile union iavf_rx_desc *rxdp;
+	struct rte_mempool_cache *cache = rte_mempool_default_cache(rxq->mp,
+			rte_lcore_id());
+	struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start];
+
+	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.
+	 */
+
+	/* Can this be satisfied from the cache? */
+	if (cache->len < IAVF_RXQ_REARM_THRESH) {
+		/* No. Backfill the cache first, and then fill from it */
+		uint32_t req = IAVF_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 + IAVF_RXQ_REARM_THRESH >=
+					rxq->nb_rx_desc) {
+				__m128i dma_addr0;
+
+				dma_addr0 = _mm_setzero_si128();
+				for (i = 0; i < IAVF_VPMD_DESCS_PER_LOOP; i++) {
+					rxp[i] = &rxq->fake_mbuf;
+					_mm_storeu_si128((__m128i *)&rxdp[i].read,
+							 dma_addr0);
+				}
+			}
+			rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+					IAVF_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_IAVF_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 < IAVF_RXQ_REARM_THRESH / IAVF_DESCS_PER_LOOP_AVX; i++) {
+		const __m512i mbuf_ptrs = _mm512_loadu_si512
+			(&cache->objs[cache->len - IAVF_DESCS_PER_LOOP_AVX]);
+		_mm512_storeu_si512(rxp, mbuf_ptrs);
+
+		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_IAVF_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 6 & 7.
+		 */
+		const __m512i desc0_1 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas0);
+		const __m512i desc2_3 = _mm512_bsrli_epi128(desc0_1, 8);
+
+		const __m512i desc4_5 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas1);
+		const __m512i desc6_7 = _mm512_bsrli_epi128(desc4_5, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_1);
+		_mm512_storeu_si512((void *)(rxdp + 2), desc2_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_5);
+		_mm512_storeu_si512((void *)(rxdp + 6), desc6_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 desc0_3 = _mm512_permutexvar_epi64(permute_idx, iova_addrs);
+		const __m512i desc4_7 = _mm512_bsrli_epi128(desc0_3, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_7);
+#endif
+		rxp += IAVF_DESCS_PER_LOOP_AVX;
+		rxdp += IAVF_DESCS_PER_LOOP_AVX;
+		cache->len -= IAVF_DESCS_PER_LOOP_AVX;
+	}
+
+	rxq->rxrearm_start += IAVF_RXQ_REARM_THRESH;
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= IAVF_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 */
+	IAVF_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
+}
+
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512(struct iavf_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 uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
+			0, rxq->mbuf_initializer);
+	/* struct iavf_rx_entry *sw_ring = &rxq->sw_ring[rxq->rx_tail]; */
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
+	const __mmask32 len_mask = _cvtu32_mask32(0x80808080);
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_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 << IAVF_RX_DESC_STATUS_DD_SHIFT)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi16
+			(/* 1st descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0,          /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0,          /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0,          /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0           /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_DESC_STATUS_EOF_SHIFT);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi8
+			(/* 1st descriptor */
+			 7, 6, 5, 4,  /* octet 4~7, 32bits rss */
+			 3, 2,        /* octet 2~3, low 16 bits vlan_macip */
+			 15, 14,      /* octet 15~14, 16 bits data_len */
+			 0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
+			 15, 14,      /* octet 15~14, low 16 bits pkt_len */
+			 0xFF, 0xFF,  /* pkt_type set as unknown */
+			 0xFF, 0xFF,  /*pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 7, 6, 5, 4,  /* octet 4~7, 32bits rss */
+			 3, 2,        /* octet 2~3, low 16 bits vlan_macip */
+			 15, 14,      /* octet 15~14, 16 bits data_len */
+			 0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
+			 15, 14,      /* octet 15~14, low 16 bits pkt_len */
+			 0xFF, 0xFF,  /* pkt_type set as unknown */
+			 0xFF, 0xFF,  /*pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 7, 6, 5, 4,  /* octet 4~7, 32bits rss */
+			 3, 2,        /* octet 2~3, low 16 bits vlan_macip */
+			 15, 14,      /* octet 15~14, 16 bits data_len */
+			 0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
+			 15, 14,      /* octet 15~14, low 16 bits pkt_len */
+			 0xFF, 0xFF,  /* pkt_type set as unknown */
+			 0xFF, 0xFF,  /*pkt_type set as unknown */
+			 /* 4th descriptor */
+			 7, 6, 5, 4,  /* octet 4~7, 32bits rss */
+			 3, 2,        /* octet 2~3, low 16 bits vlan_macip */
+			 15, 14,      /* octet 15~14, 16 bits data_len */
+			 0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
+			 15, 14,      /* octet 15~14, low 16 bits pkt_len */
+			 0xFF, 0xFF,  /* pkt_type set as unknown */
+			 0xFF, 0xFF   /*pkt_type set as unknown */
+			);
+	/**
+	 * 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 += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 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 desc4_7 = _mm512_mask_blend_epi16(len_mask,
+								raw_desc4_7,
+								len4_7);
+		__m512i mb4_7 = _mm512_shuffle_epi8(desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, 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 __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 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 __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, adjusting length and
+		 * re-arranging fields. Then write into the mbuf
+		 */
+		const __m512i len0_3 = _mm512_slli_epi32(raw_desc0_3,
+							 PKTLEN_SHIFT);
+		const __m512i desc0_3 = _mm512_mask_blend_epi16(len_mask,
+								raw_desc0_3,
+								len0_3);
+		__m512i mb0_3 = _mm512_shuffle_epi8(desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/* get the packet types */
+		const __m512i ptypes0_3 = _mm512_srli_epi64(desc0_3, 30);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		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 ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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_flags =
+			_mm256_shuffle_epi8(rss_flags_shuf,
+					    _mm256_srli_epi32(flag_bits, 11));
+		/**
+		 * 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 */
+		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				_mm256_or_si256(rss_flags, vlan_flags));
+		/**
+		 * 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;
+		const __m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7, 0);
+		const __m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7, 1);
+		const __m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3, 0);
+		const __m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3, 1);
+
+		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 << IAVF_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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
+static inline __m256i
+flex_rxd_to_fdir_flags_vec_avx512(const __m256i fdir_id0_7)
+{
+#define FDID_MIS_MAGIC 0xFFFFFFFF
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR != (1 << 2));
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
+	const __m256i pkt_fdir_bit = _mm256_set1_epi32(PKT_RX_FDIR |
+			PKT_RX_FDIR_ID);
+	/* desc->flow_id field == 0xFFFFFFFF means fdir mismatch */
+	const __m256i fdir_mis_mask = _mm256_set1_epi32(FDID_MIS_MAGIC);
+	__m256i fdir_mask = _mm256_cmpeq_epi32(fdir_id0_7,
+			fdir_mis_mask);
+	/* this XOR op results to bit-reverse the fdir_mask */
+	fdir_mask = _mm256_xor_si256(fdir_mask, fdir_mis_mask);
+	const __m256i fdir_flags = _mm256_and_si256(fdir_mask, pkt_fdir_bit);
+
+	return fdir_flags;
+}
+
+/**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+			  uint16_t nb_pkts)
+{
+	return _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				     uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				    uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst = iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
diff --git a/drivers/net/iavf/meson.build b/drivers/net/iavf/meson.build
index a3fad36..00c7a12 100644
--- a/drivers/net/iavf/meson.build
+++ b/drivers/net/iavf/meson.build
@@ -34,4 +34,21 @@ if arch_subdir == 'x86'
 				c_args: [cflags, '-mavx2'])
 		objs += iavf_avx2_lib.extract_objects('iavf_rxtx_vec_avx2.c')
 	endif
+
+	if not machine_args.contains('-mno-avx512f')
+		if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F')
+			cflags += ['-DCC_AVX512_SUPPORT']
+			cflags += ['-march=skylake-avx512']
+			sources += files('iavf_rxtx_vec_avx512.c')
+		elif cc.has_argument('-mavx512f')
+			cflags += ['-DCC_AVX512_SUPPORT']
+			iavf_avx512_lib = static_library('iavf_avx512_lib',
+					'iavf_rxtx_vec_avx512.c',
+					dependencies: [static_rte_ethdev,
+						static_rte_kvargs, static_rte_hash],
+					include_directories: includes,
+					c_args: [cflags, '-mavx512f', '-march=skylake-avx512'])
+			objs += iavf_avx512_lib.extract_objects('iavf_rxtx_vec_avx512.c')
+		endif
+	endif
 endif
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v2 2/3] net/iavf: enable AVX512 for flexible RX
  2020-09-17  1:39 ` [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-09-17  1:39   ` [dpdk-dev] [PATCH v2 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
@ 2020-09-17  1:39   ` Wenzhuo Lu
  2020-09-17  1:39   ` [dpdk-dev] [PATCH v2 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
  2020-09-17  7:37   ` [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf Morten Brørup
  3 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-17  1:39 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the flexible RX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  10 +
 drivers/net/iavf/iavf_rxtx.h            |   6 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 699 ++++++++++++++++++++++++++++++++
 3 files changed, 715 insertions(+)

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index c36e809..0818107 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2132,6 +2132,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
@@ -2151,6 +2156,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index cb12888..9653e0c 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -439,9 +439,15 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
 uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
 				   uint16_t nb_pkts);
+uint16_t iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					    struct rte_mbuf **rx_pkts,
+					    uint16_t nb_pkts);
 uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 					     struct rte_mbuf **rx_pkts,
 					     uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+						      struct rte_mbuf **rx_pkts,
+						      uint16_t nb_pkts);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index b528ed3..8c33661 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -643,6 +643,624 @@
 	return fdir_flags;
 }
 
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512_flex_rxd(struct iavf_rx_queue *rxq,
+					struct rte_mbuf **rx_pkts,
+					uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
+			0, rxq->mbuf_initializer);
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_flex_desc *rxdp =
+		(union iavf_rx_flex_desc *)rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_rxq_rearm(rxq);
+
+	/* Before we start moving massive data around, check to see if
+	 * there is actually a packet available
+	 */
+	if (!(rxdp->wb.status_error0 &
+			rte_cpu_to_le_32(1 << IAVF_RX_FLEX_DESC_STATUS0_DD_S)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi16
+			(/* 1st descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0,          /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0,          /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0,          /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0, 0, 0,       /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 0,             /* ignore high-16bits of pkt_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0, 0           /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi8
+			(/* 1st descriptor */
+			 0xFF, 0xFF,
+			 0xFF, 0xFF,    /* rss hash parsed separately */
+			 11, 10,	/* octet 10~11, 16 bits vlan_macip */
+			 5, 4,		/* octet 4~5, 16 bits data_len */
+			 0xFF, 0xFF,	/* skip hi 16 bits pkt_len, zero out */
+			 5, 4,		/* octet 4~5, 16 bits pkt_len */
+			 0xFF, 0xFF,	/* pkt_type set as unknown */
+			 0xFF, 0xFF,	/*pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 0xFF, 0xFF,
+			 0xFF, 0xFF,    /* rss hash parsed separately */
+			 11, 10,	/* octet 10~11, 16 bits vlan_macip */
+			 5, 4,		/* octet 4~5, 16 bits data_len */
+			 0xFF, 0xFF,	/* skip hi 16 bits pkt_len, zero out */
+			 5, 4,		/* octet 4~5, 16 bits pkt_len */
+			 0xFF, 0xFF,	/* pkt_type set as unknown */
+			 0xFF, 0xFF,	/*pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 0xFF, 0xFF,
+			 0xFF, 0xFF,    /* rss hash parsed separately */
+			 11, 10,	/* octet 10~11, 16 bits vlan_macip */
+			 5, 4,		/* octet 4~5, 16 bits data_len */
+			 0xFF, 0xFF,	/* skip hi 16 bits pkt_len, zero out */
+			 5, 4,		/* octet 4~5, 16 bits pkt_len */
+			 0xFF, 0xFF,	/* pkt_type set as unknown */
+			 0xFF, 0xFF,	/*pkt_type set as unknown */
+			 /* 4th descriptor */
+			 0xFF, 0xFF,
+			 0xFF, 0xFF,    /* rss hash parsed separately */
+			 11, 10,	/* octet 10~11, 16 bits vlan_macip */
+			 5, 4,		/* octet 4~5, 16 bits data_len */
+			 0xFF, 0xFF,	/* skip hi 16 bits pkt_len, zero out */
+			 5, 4,		/* octet 4~5, 16 bits pkt_len */
+			 0xFF, 0xFF,	/* pkt_type set as unknown */
+			 0xFF, 0xFF	/*pkt_type set as unknown */
+			);
+	/**
+	 * 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 Checksum Reports, RSS indication
+	 * and VLAN indication.
+	 * bit6:4 for IP/L4 checksum errors.
+	 * bit12 is for RSS indication.
+	 * bit13 is for VLAN indication.
+	 */
+	const __m256i flags_mask =
+		 _mm256_set1_epi32((7 << 4) | (1 << 12) | (1 << 13));
+	/**
+	 * data to be shuffled by the result of the flags mask shifted by 4
+	 * 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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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);
+	/**
+	 * data to be shuffled by result of flag mask, shifted down 12.
+	 * If RSS(bit12)/VLAN(bit13) are set,
+	 * shuffle moves appropriate flags in place.
+	 */
+	const __m256i rss_vlan_flags_shuf = _mm256_set_epi8(0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0,
+			/* end up 128-bits */
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0);
+
+	uint16_t i, received;
+
+	for (i = 0, received = 0; i < nb_pkts;
+	     i += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb4_7 = _mm512_shuffle_epi8(raw_desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptype_mask =
+			_mm512_set1_epi16(IAVF_RX_FLEX_DESC_PTYPE_M);
+		const __m512i ptypes4_7 =
+			_mm512_and_si512(raw_desc4_7, ptype_mask);
+		const __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 0);
+		const uint16_t ptype7 = _mm256_extract_epi16(ptypes6_7, 9);
+		const uint16_t ptype6 = _mm256_extract_epi16(ptypes6_7, 1);
+		const uint16_t ptype5 = _mm256_extract_epi16(ptypes4_5, 9);
+		const uint16_t ptype4 = _mm256_extract_epi16(ptypes4_5, 1);
+
+		const __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb0_3 = _mm512_shuffle_epi8(raw_desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptypes0_3 =
+			_mm512_and_si512(raw_desc0_3, ptype_mask);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		const uint16_t ptype3 = _mm256_extract_epi16(ptypes2_3, 9);
+		const uint16_t ptype2 = _mm256_extract_epi16(ptypes2_3, 1);
+		const uint16_t ptype1 = _mm256_extract_epi16(ptypes0_1, 9);
+		const uint16_t ptype0 = _mm256_extract_epi16(ptypes0_1, 1);
+
+		const __m512i ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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);
+		/**
+		 * 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, 4));
+		l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
+		l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
+		/* set rss and vlan flags */
+		const __m256i rss_vlan_flag_bits =
+			_mm256_srli_epi32(flag_bits, 12);
+		const __m256i rss_vlan_flags =
+			_mm256_shuffle_epi8(rss_vlan_flags_shuf,
+					    rss_vlan_flag_bits);
+
+		/* merge flags */
+		__m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				rss_vlan_flags);
+
+		if (rxq->fdir_enabled) {
+			const __m512i fdir_permute_mask = _mm512_set_epi32
+				(0, 0, 0, 0,
+				 0, 0, 0, 0,
+				 15, 11, 7, 3,
+				 31, 27, 23, 19);
+			__m512i fdir_tmp = _mm512_permutex2var_epi32
+				(raw_desc4_7, fdir_permute_mask, raw_desc0_3);
+			const __m256i fdir_id0_7 = _mm512_extracti64x4_epi64
+				(fdir_tmp, 0);
+			const __m256i fdir_flags =
+				flex_rxd_to_fdir_flags_vec_avx512(fdir_id0_7);
+
+			/* merge with fdir_flags */
+			mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_flags);
+
+			/* write to mbuf: have to use scalar store here */
+			rx_pkts[i + 0]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 3);
+
+			rx_pkts[i + 1]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 7);
+
+			rx_pkts[i + 2]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 2);
+
+			rx_pkts[i + 3]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 6);
+
+			rx_pkts[i + 4]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 1);
+
+			rx_pkts[i + 5]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 5);
+
+			rx_pkts[i + 6]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 0);
+
+			rx_pkts[i + 7]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 4);
+		} /* if() on fdir_enabled */
+
+		__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);
+
+#ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
+		/**
+		 * needs to load 2nd 16B of each desc for RSS hash parsing,
+		 * will cause performance drop to get into this context.
+		 */
+		if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
+				DEV_RX_OFFLOAD_RSS_HASH) {
+			/* load bottom half of every 32B desc */
+			const __m128i raw_desc_bh7 =
+				_mm_load_si128
+					((void *)(&rxdp[7].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh6 =
+				_mm_load_si128
+					((void *)(&rxdp[6].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh5 =
+				_mm_load_si128
+					((void *)(&rxdp[5].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh4 =
+				_mm_load_si128
+					((void *)(&rxdp[4].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh3 =
+				_mm_load_si128
+					((void *)(&rxdp[3].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh2 =
+				_mm_load_si128
+					((void *)(&rxdp[2].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh1 =
+				_mm_load_si128
+					((void *)(&rxdp[1].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh0 =
+				_mm_load_si128
+					((void *)(&rxdp[0].wb.status_error1));
+
+			__m256i raw_desc_bh6_7 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh6),
+					raw_desc_bh7, 1);
+			__m256i raw_desc_bh4_5 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh4),
+					raw_desc_bh5, 1);
+			__m256i raw_desc_bh2_3 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh2),
+					raw_desc_bh3, 1);
+			__m256i raw_desc_bh0_1 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh0),
+					raw_desc_bh1, 1);
+
+			/**
+			 * to shift the 32b RSS hash value to the
+			 * highest 32b of each 128b before mask
+			 */
+			__m256i rss_hash6_7 =
+				_mm256_slli_epi64(raw_desc_bh6_7, 32);
+			__m256i rss_hash4_5 =
+				_mm256_slli_epi64(raw_desc_bh4_5, 32);
+			__m256i rss_hash2_3 =
+				_mm256_slli_epi64(raw_desc_bh2_3, 32);
+			__m256i rss_hash0_1 =
+				_mm256_slli_epi64(raw_desc_bh0_1, 32);
+
+			__m256i rss_hash_msk =
+				_mm256_set_epi32(0xFFFFFFFF, 0, 0, 0,
+						 0xFFFFFFFF, 0, 0, 0);
+
+			rss_hash6_7 = _mm256_and_si256
+					(rss_hash6_7, rss_hash_msk);
+			rss_hash4_5 = _mm256_and_si256
+					(rss_hash4_5, rss_hash_msk);
+			rss_hash2_3 = _mm256_and_si256
+					(rss_hash2_3, rss_hash_msk);
+			rss_hash0_1 = _mm256_and_si256
+					(rss_hash0_1, rss_hash_msk);
+
+			mb6_7 = _mm256_or_si256(mb6_7, rss_hash6_7);
+			mb4_5 = _mm256_or_si256(mb4_5, rss_hash4_5);
+			mb2_3 = _mm256_or_si256(mb2_3, rss_hash2_3);
+			mb0_1 = _mm256_or_si256(mb0_1, rss_hash0_1);
+		} /* if() on RSS hash parsing */
+#endif
+
+		/**
+		 * 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 <<
+					       IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+			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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
 /**
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -655,6 +1273,18 @@
 }
 
 /**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts)
+{
+	return _iavf_recv_raw_pkts_vec_avx512_flex_rxd(rx_queue, rx_pkts,
+						       nb_pkts, NULL);
+}
+
+/**
  * vPMD receive routine that reassembles single burst of 32 scattered packets
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -718,3 +1348,72 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+/**
+ * vPMD receive routine that reassembles single burst of
+ * 32 scattered packets for flex RxD
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512_flex_rxd(void *rx_queue,
+					      struct rte_mbuf **rx_pkts,
+					      uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_recv_raw_pkts_vec_avx512_flex_rxd(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 for flex RxD.
+ * Main receive routine that can handle arbitrary burst sizes
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst =
+			iavf_recv_scattered_burst_vec_avx512_flex_rxd
+			(rx_queue, rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v2 3/3] net/iavf: enable AVX512 for TX
  2020-09-17  1:39 ` [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-09-17  1:39   ` [dpdk-dev] [PATCH v2 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
  2020-09-17  1:39   ` [dpdk-dev] [PATCH v2 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
@ 2020-09-17  1:39   ` Wenzhuo Lu
  2020-09-17  7:37   ` [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf Morten Brørup
  3 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-17  1:39 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the TX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 doc/guides/rel_notes/release_20_11.rst  |   3 +
 drivers/net/iavf/iavf_ethdev.c          |   3 +-
 drivers/net/iavf/iavf_rxtx.c            |  32 +++-
 drivers/net/iavf/iavf_rxtx.h            |   7 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 301 ++++++++++++++++++++++++++++++++
 5 files changed, 338 insertions(+), 8 deletions(-)

diff --git a/doc/guides/rel_notes/release_20_11.rst b/doc/guides/rel_notes/release_20_11.rst
index df227a1..d40b8d6 100644
--- a/doc/guides/rel_notes/release_20_11.rst
+++ b/doc/guides/rel_notes/release_20_11.rst
@@ -55,6 +55,9 @@ New Features
      Also, make sure to start the actual text at the margin.
      =======================================================
 
+   * **Added support of vector instructions on IAVF.**
+
+     Added support of AVX512 instructions in IAVF RX and TX path.
 
 Removed Items
 -------------
diff --git a/drivers/net/iavf/iavf_ethdev.c b/drivers/net/iavf/iavf_ethdev.c
index c3aa4cd..5bc2851 100644
--- a/drivers/net/iavf/iavf_ethdev.c
+++ b/drivers/net/iavf/iavf_ethdev.c
@@ -528,7 +528,8 @@ static int iavf_config_rx_queues_irqs(struct rte_eth_dev *dev,
 		DEV_TX_OFFLOAD_GRE_TNL_TSO |
 		DEV_TX_OFFLOAD_IPIP_TNL_TSO |
 		DEV_TX_OFFLOAD_GENEVE_TNL_TSO |
-		DEV_TX_OFFLOAD_MULTI_SEGS;
+		DEV_TX_OFFLOAD_MULTI_SEGS |
+		DEV_TX_OFFLOAD_MBUF_FAST_FREE;
 
 	dev_info->default_rxconf = (struct rte_eth_rxconf) {
 		.rx_free_thresh = IAVF_DEFAULT_RX_FREE_THRESH,
diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index 0818107..04dcd48 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2206,18 +2206,18 @@
 	struct iavf_tx_queue *txq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_tx_vec_dev_check(dev)) {
-		for (i = 0; i < dev->data->nb_tx_queues; i++) {
-			txq = dev->data->tx_queues[i];
-			if (!txq)
-				continue;
-			iavf_txq_vec_setup(txq);
-		}
-
 		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+			use_avx512 = true;
+#endif
 
 		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
 			    use_avx2 ? "avx2 " : "",
@@ -2225,8 +2225,26 @@
 		dev->tx_pkt_burst = use_avx2 ?
 				    iavf_xmit_pkts_vec_avx2 :
 				    iavf_xmit_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+		if (use_avx512)
+			dev->tx_pkt_burst = iavf_xmit_pkts_vec_avx512;
+#endif
 		dev->tx_pkt_prepare = NULL;
 
+		for (i = 0; i < dev->data->nb_tx_queues; i++) {
+			txq = dev->data->tx_queues[i];
+			if (!txq)
+				continue;
+#ifdef CC_AVX512_SUPPORT
+			if (use_avx512)
+				iavf_txq_vec_setup_avx512(txq);
+			else
+				iavf_txq_vec_setup(txq);
+#else
+			iavf_txq_vec_setup(txq);
+#endif
+		}
+
 		return;
 	}
 #endif
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 9653e0c..08eebb0 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -122,6 +122,10 @@ struct iavf_tx_entry {
 	uint16_t last_id;
 };
 
+struct iavf_tx_vec_entry {
+	struct rte_mbuf *mbuf;
+};
+
 /* Structure associated with each TX queue. */
 struct iavf_tx_queue {
 	const struct rte_memzone *mz;  /* memzone for Tx ring */
@@ -448,6 +452,9 @@ uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
 						      struct rte_mbuf **rx_pkts,
 						      uint16_t nb_pkts);
+uint16_t iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				   uint16_t nb_pkts);
+int iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index 8c33661..399c73b 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -1417,3 +1417,304 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+static __rte_always_inline int
+iavf_tx_free_bufs_avx512(struct iavf_tx_queue *txq)
+{
+	struct iavf_tx_vec_entry *txep;
+	uint32_t n;
+	uint32_t i;
+	int nb_free = 0;
+	struct rte_mbuf *m, *free[IAVF_VPMD_TX_MAX_FREE_BUF];
+
+	/* check DD bits on threshold descriptor */
+	if ((txq->tx_ring[txq->next_dd].cmd_type_offset_bsz &
+			rte_cpu_to_le_64(IAVF_TXD_QW1_DTYPE_MASK)) !=
+			rte_cpu_to_le_64(IAVF_TX_DESC_DTYPE_DESC_DONE))
+		return 0;
+
+	n = txq->rs_thresh;
+
+	 /* first buffer to free from S/W ring is at index
+	  * tx_next_dd - (tx_rs_thresh-1)
+	  */
+	txep = (void *)txq->sw_ring;
+	txep += txq->next_dd - (n - 1);
+
+	if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE && (n & 31) == 0) {
+		struct rte_mempool *mp = txep[0].mbuf->pool;
+		struct rte_mempool_cache *cache = rte_mempool_default_cache(mp,
+				rte_lcore_id());
+		void **cache_objs = &cache->objs[cache->len];
+
+		if (n > RTE_MEMPOOL_CACHE_MAX_SIZE) {
+			rte_mempool_ops_enqueue_bulk(mp, (void *)txep, n);
+			goto done;
+		}
+
+		/* The cache follows the following algorithm
+		 *   1. Add the objects to the cache
+		 *   2. Anything greater than the cache min value (if it crosses the
+		 *   cache flush threshold) is flushed to the ring.
+		 */
+		/* Add elements back into the cache */
+		uint32_t copied = 0;
+		/* n is multiple of 32 */
+		while (copied < n) {
+			const __m512i a = _mm512_loadu_si512(&txep[copied]);
+			const __m512i b = _mm512_loadu_si512(&txep[copied + 8]);
+			const __m512i c = _mm512_loadu_si512(&txep[copied + 16]);
+			const __m512i d = _mm512_loadu_si512(&txep[copied + 24]);
+
+			_mm512_storeu_si512(&cache_objs[copied], a);
+			_mm512_storeu_si512(&cache_objs[copied + 8], b);
+			_mm512_storeu_si512(&cache_objs[copied + 16], c);
+			_mm512_storeu_si512(&cache_objs[copied + 24], d);
+			copied += 32;
+		}
+		cache->len += n;
+
+		if (cache->len >= cache->flushthresh) {
+			rte_mempool_ops_enqueue_bulk(mp,
+						     &cache->objs[cache->size],
+						     cache->len - cache->size);
+			cache->len = cache->size;
+		}
+		goto done;
+	}
+
+	m = rte_pktmbuf_prefree_seg(txep[0].mbuf);
+	if (likely(m)) {
+		free[0] = m;
+		nb_free = 1;
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (likely(m)) {
+				if (likely(m->pool == free[0]->pool)) {
+					free[nb_free++] = m;
+				} else {
+					rte_mempool_put_bulk(free[0]->pool,
+							     (void *)free,
+							     nb_free);
+					free[0] = m;
+					nb_free = 1;
+				}
+			}
+		}
+		rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
+	} else {
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (m)
+				rte_mempool_put(m->pool, m);
+		}
+	}
+
+done:
+	/* buffers were freed, update counters */
+	txq->nb_free = (uint16_t)(txq->nb_free + txq->rs_thresh);
+	txq->next_dd = (uint16_t)(txq->next_dd + txq->rs_thresh);
+	if (txq->next_dd >= txq->nb_tx_desc)
+		txq->next_dd = (uint16_t)(txq->rs_thresh - 1);
+
+	return txq->rs_thresh;
+}
+
+static __rte_always_inline void
+tx_backlog_entry_avx512(struct iavf_tx_vec_entry *txep,
+			struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	int i;
+
+	for (i = 0; i < (int)nb_pkts; ++i)
+		txep[i].mbuf = tx_pkts[i];
+}
+
+static inline void
+iavf_vtx1(volatile struct iavf_tx_desc *txdp,
+	  struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(IAVF_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT) |
+		 ((uint64_t)pkt->data_len << IAVF_TXD_QW1_TX_BUF_SZ_SHIFT));
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+				pkt->buf_iova + pkt->data_off);
+	_mm_storeu_si128((__m128i *)txdp, descriptor);
+}
+
+static inline void
+iavf_vtx(volatile struct iavf_tx_desc *txdp,
+	 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (IAVF_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT));
+	const __mmask8 len_msk = _cvtu32_mask8(0xAA);
+	const __mmask8 off_msk = _cvtu32_mask8(0x55);
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		iavf_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) {
+		__m512i desc4 =
+			_mm512_set_epi64
+				((uint64_t)pkt[3]->data_len,
+				 pkt[3]->buf_iova,
+				 (uint64_t)pkt[2]->data_len,
+				 pkt[2]->buf_iova,
+				 (uint64_t)pkt[1]->data_len,
+				 pkt[1]->buf_iova,
+				 (uint64_t)pkt[0]->data_len,
+				 pkt[0]->buf_iova);
+		__m512i hi_qw_tmpl_4 = _mm512_set1_epi64(hi_qw_tmpl);
+		__m512i data_off_4 =
+			_mm512_set_epi64
+				(0,
+				 pkt[3]->data_off,
+				 0,
+				 pkt[2]->data_off,
+				 0,
+				 pkt[1]->data_off,
+				 0,
+				 pkt[0]->data_off);
+
+		desc4 = _mm512_mask_slli_epi64(desc4, len_msk, desc4, IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
+		desc4 = _mm512_mask_or_epi64(desc4, len_msk, desc4, hi_qw_tmpl_4);
+		desc4 = _mm512_mask_add_epi64(desc4, off_msk, desc4, data_off_4);
+		_mm512_storeu_si512((void *)txdp, desc4);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		iavf_vtx1(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+iavf_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				 uint16_t nb_pkts)
+{
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+	volatile struct iavf_tx_desc *txdp;
+	struct iavf_tx_vec_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	/* bit2 is reserved and must be set to 1 according to Spec */
+	uint64_t flags = IAVF_TX_DESC_CMD_EOP | IAVF_TX_DESC_CMD_ICRC;
+	uint64_t rs = IAVF_TX_DESC_CMD_RS | flags;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
+
+	if (txq->nb_free < txq->free_thresh)
+		iavf_tx_free_bufs_avx512(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = (void *)txq->sw_ring;
+	txep += tx_id;
+
+	txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		tx_backlog_entry_avx512(txep, tx_pkts, n);
+
+		iavf_vtx(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		iavf_vtx1(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = (void *)txq->sw_ring;
+		txep += tx_id;
+	}
+
+	tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
+
+	iavf_vtx(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->next_rs) {
+		txq->tx_ring[txq->next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)IAVF_TX_DESC_CMD_RS) <<
+					 IAVF_TXD_QW1_CMD_SHIFT);
+		txq->next_rs =
+			(uint16_t)(txq->next_rs + txq->rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	IAVF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+			  uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->rs_thresh);
+		ret = iavf_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;
+}
+
+static inline void
+iavf_tx_queue_release_mbufs_avx512(struct iavf_tx_queue *txq)
+{
+	unsigned int i;
+	const uint16_t max_desc = (uint16_t)(txq->nb_tx_desc - 1);
+	struct iavf_tx_vec_entry *swr = (void *)txq->sw_ring;
+
+	if (!txq->sw_ring || txq->nb_free == max_desc)
+		return;
+
+	i = txq->next_dd - txq->rs_thresh + 1;
+	if (txq->tx_tail < i) {
+		for (; i < txq->nb_tx_desc; i++) {
+			rte_pktmbuf_free_seg(swr[i].mbuf);
+			swr[i].mbuf = NULL;
+		}
+		i = 0;
+	}
+}
+
+static const struct iavf_txq_ops avx512_vec_txq_ops = {
+	.release_mbufs = iavf_tx_queue_release_mbufs_avx512,
+};
+
+int __rte_cold
+iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq)
+{
+	txq->ops = &avx512_vec_txq_ops;
+	return 0;
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* Re: [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf
  2020-09-17  1:39 ` [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf Wenzhuo Lu
                     ` (2 preceding siblings ...)
  2020-09-17  1:39   ` [dpdk-dev] [PATCH v2 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
@ 2020-09-17  7:37   ` Morten Brørup
  2020-09-17  9:13     ` Bruce Richardson
  3 siblings, 1 reply; 39+ messages in thread
From: Morten Brørup @ 2020-09-17  7:37 UTC (permalink / raw)
  To: Wenzhuo Lu, Bruce Richardson, Leyi Rong; +Cc: dev

> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Wenzhuo Lu
> Sent: Thursday, September 17, 2020 3:40 AM
> 
> AVX512 instructions is supported by more and more platforms. These
> instructions
> can be used in the data path to enhance the per-core performance of
> packet
> processing.
> Comparing with the existing implementation, this path set introduces
> some AVX512
> instructions into the iavf data path, and we get a better per-code
> throughput.
> 
> v2:
> Update meson.build.
> Repalce the deprecated 'buf_physaddr' by 'buf_iova'.
> 
> Wenzhuo Lu (3):
>   net/iavf: enable AVX512 for legacy RX
>   net/iavf: enable AVX512 for flexible RX
>   net/iavf: enable AVX512 for TX
> 
>  doc/guides/rel_notes/release_20_11.rst  |    3 +
>  drivers/net/iavf/iavf_ethdev.c          |    3 +-
>  drivers/net/iavf/iavf_rxtx.c            |   69 +-
>  drivers/net/iavf/iavf_rxtx.h            |   18 +
>  drivers/net/iavf/iavf_rxtx_vec_avx512.c | 1720
> +++++++++++++++++++++++++++++++
>  drivers/net/iavf/meson.build            |   17 +
>  6 files changed, 1818 insertions(+), 12 deletions(-)
>  create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c
> 
> --
> 1.9.3
> 

I am not sure I understand the full context here, so please bear with me if I'm completely off...

With this patch set, it looks like the driver manipulates the mempool cache directly, bypassing the libararies encapsulating it.

Isn't that going deeper into a library than expected... What if the implementation of the mempool library changes radically?

And if there are performance gains to be achieved by using vector instructions for manipulating the mempool, perhaps your vector optimizations should go into the mempool library instead?


Med venlig hilsen / kind regards
- Morten Brørup




^ permalink raw reply	[flat|nested] 39+ messages in thread

* Re: [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf
  2020-09-17  7:37   ` [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf Morten Brørup
@ 2020-09-17  9:13     ` Bruce Richardson
  2020-09-17  9:35       ` Morten Brørup
  0 siblings, 1 reply; 39+ messages in thread
From: Bruce Richardson @ 2020-09-17  9:13 UTC (permalink / raw)
  To: Morten Brørup; +Cc: Wenzhuo Lu, Leyi Rong, dev

On Thu, Sep 17, 2020 at 09:37:29AM +0200, Morten Brørup wrote:
> > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Wenzhuo Lu
> > Sent: Thursday, September 17, 2020 3:40 AM
> > 
> > AVX512 instructions is supported by more and more platforms. These
> > instructions
> > can be used in the data path to enhance the per-core performance of
> > packet
> > processing.
> > Comparing with the existing implementation, this path set introduces
> > some AVX512
> > instructions into the iavf data path, and we get a better per-code
> > throughput.
> > 
> > v2:
> > Update meson.build.
> > Repalce the deprecated 'buf_physaddr' by 'buf_iova'.
> > 
> > Wenzhuo Lu (3):
> >   net/iavf: enable AVX512 for legacy RX
> >   net/iavf: enable AVX512 for flexible RX
> >   net/iavf: enable AVX512 for TX
> > 
> >  doc/guides/rel_notes/release_20_11.rst  |    3 +
> >  drivers/net/iavf/iavf_ethdev.c          |    3 +-
> >  drivers/net/iavf/iavf_rxtx.c            |   69 +-
> >  drivers/net/iavf/iavf_rxtx.h            |   18 +
> >  drivers/net/iavf/iavf_rxtx_vec_avx512.c | 1720
> > +++++++++++++++++++++++++++++++
> >  drivers/net/iavf/meson.build            |   17 +
> >  6 files changed, 1818 insertions(+), 12 deletions(-)
> >  create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c
> > 
> > --
> > 1.9.3
> > 
> 
> I am not sure I understand the full context here, so please bear with me if I'm completely off...
> 
> With this patch set, it looks like the driver manipulates the mempool cache directly, bypassing the libararies encapsulating it.
> 
> Isn't that going deeper into a library than expected... What if the implementation of the mempool library changes radically?
> 
> And if there are performance gains to be achieved by using vector instructions for manipulating the mempool, perhaps your vector optimizations should go into the mempool library instead?
> 

Looking specifically at the descriptor re-arm code, the benefit from
working off the mempool cache directly comes from saving loads by merging
the code blocks, rather than directly from the vectorization itself -
though the vectorization doesn't hurt. The original code having a separate
mempool function worked roughly like below:

1. mempool code loads mbuf pointers from cache
2. mempool code writes mbuf pointers to the SW ring for the NIC
3. driver code loads the mempool pointers from the SW ring
4. driver code then does the rest of the descriptor re-arm.

The benefit comes from eliminating step 3, the loads in the driver, which
are dependent upon the previous stores. By having the driver itself read
from the mempool cache (the code still uses mempool functions for every
other part, since everything beyond the cache depends on the
ring/stack/bucket implementation), we can have the stores go out, and while
they are completing reuse the already-loaded data to do the descriptor
rearm.

Hope this clarifies things.

/Bruce


^ permalink raw reply	[flat|nested] 39+ messages in thread

* Re: [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf
  2020-09-17  9:13     ` Bruce Richardson
@ 2020-09-17  9:35       ` Morten Brørup
  0 siblings, 0 replies; 39+ messages in thread
From: Morten Brørup @ 2020-09-17  9:35 UTC (permalink / raw)
  To: Bruce Richardson; +Cc: Wenzhuo Lu, Leyi Rong, dev

> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Bruce Richardson
> Sent: Thursday, September 17, 2020 11:13 AM
> 
> On Thu, Sep 17, 2020 at 09:37:29AM +0200, Morten Brørup wrote:
> > > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Wenzhuo Lu
> > > Sent: Thursday, September 17, 2020 3:40 AM
> > >
> > > AVX512 instructions is supported by more and more platforms. These
> > > instructions
> > > can be used in the data path to enhance the per-core performance of
> > > packet
> > > processing.
> > > Comparing with the existing implementation, this path set
> introduces
> > > some AVX512
> > > instructions into the iavf data path, and we get a better per-code
> > > throughput.
> > >
> > > v2:
> > > Update meson.build.
> > > Repalce the deprecated 'buf_physaddr' by 'buf_iova'.
> > >
> > > Wenzhuo Lu (3):
> > >   net/iavf: enable AVX512 for legacy RX
> > >   net/iavf: enable AVX512 for flexible RX
> > >   net/iavf: enable AVX512 for TX
> > >
> > >  doc/guides/rel_notes/release_20_11.rst  |    3 +
> > >  drivers/net/iavf/iavf_ethdev.c          |    3 +-
> > >  drivers/net/iavf/iavf_rxtx.c            |   69 +-
> > >  drivers/net/iavf/iavf_rxtx.h            |   18 +
> > >  drivers/net/iavf/iavf_rxtx_vec_avx512.c | 1720
> > > +++++++++++++++++++++++++++++++
> > >  drivers/net/iavf/meson.build            |   17 +
> > >  6 files changed, 1818 insertions(+), 12 deletions(-)
> > >  create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c
> > >
> > > --
> > > 1.9.3
> > >
> >
> > I am not sure I understand the full context here, so please bear with
> me if I'm completely off...
> >
> > With this patch set, it looks like the driver manipulates the mempool
> cache directly, bypassing the libararies encapsulating it.
> >
> > Isn't that going deeper into a library than expected... What if the
> implementation of the mempool library changes radically?
> >
> > And if there are performance gains to be achieved by using vector
> instructions for manipulating the mempool, perhaps your vector
> optimizations should go into the mempool library instead?
> >
> 
> Looking specifically at the descriptor re-arm code, the benefit from
> working off the mempool cache directly comes from saving loads by
> merging
> the code blocks, rather than directly from the vectorization itself -
> though the vectorization doesn't hurt. The original code having a
> separate
> mempool function worked roughly like below:
> 
> 1. mempool code loads mbuf pointers from cache
> 2. mempool code writes mbuf pointers to the SW ring for the NIC
> 3. driver code loads the mempool pointers from the SW ring
> 4. driver code then does the rest of the descriptor re-arm.
> 
> The benefit comes from eliminating step 3, the loads in the driver,
> which
> are dependent upon the previous stores. By having the driver itself
> read
> from the mempool cache (the code still uses mempool functions for every
> other part, since everything beyond the cache depends on the
> ring/stack/bucket implementation), we can have the stores go out, and
> while
> they are completing reuse the already-loaded data to do the descriptor
> rearm.
> 
> Hope this clarifies things.
> 
> /Bruce
> 

Thank you for the detailed explanation, Bruce.

It makes sense to me now. So,

Acked-By: Morten Brørup <mb@smartsharesystems.com>


Med venlig hilsen / kind regards
- Morten Brørup




^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v3 0/3] enable AVX512 for iavf
  2020-09-10  5:59 [dpdk-dev] [PATCH 0/3] enable AVX512 for iavf Wenzhuo Lu
                   ` (3 preceding siblings ...)
  2020-09-17  1:39 ` [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf Wenzhuo Lu
@ 2020-09-21  8:13 ` Wenzhuo Lu
  2020-09-21  8:13   ` [dpdk-dev] [PATCH v3 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
                     ` (2 more replies)
  2020-09-27  1:30 ` [dpdk-dev] [PATCH v4 0/3] enable AVX512 for iavf Wenzhuo Lu
                   ` (3 subsequent siblings)
  8 siblings, 3 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-21  8:13 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu

AVX512 instructions is supported by more and more platforms. These instructions
can be used in the data path to enhance the per-core performance of packet
processing.
Comparing with the existing implementation, this path set introduces some AVX512
instructions into the iavf data path, and we get a better per-code throughput.

v2:
Update meson.build.
Repalce the deprecated 'buf_physaddr' by 'buf_iova'.

v3:
Fix compile errors.

Wenzhuo Lu (3):
  net/iavf: enable AVX512 for legacy RX
  net/iavf: enable AVX512 for flexible RX
  net/iavf: enable AVX512 for TX

 doc/guides/rel_notes/release_20_11.rst  |    3 +
 drivers/net/iavf/iavf_ethdev.c          |    3 +-
 drivers/net/iavf/iavf_rxtx.c            |   69 +-
 drivers/net/iavf/iavf_rxtx.h            |   18 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 1702 +++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |   17 +
 6 files changed, 1800 insertions(+), 12 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v3 1/3] net/iavf: enable AVX512 for legacy RX
  2020-09-21  8:13 ` [dpdk-dev] [PATCH v3 " Wenzhuo Lu
@ 2020-09-21  8:13   ` Wenzhuo Lu
  2020-09-21  8:13   ` [dpdk-dev] [PATCH v3 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
  2020-09-21  8:13   ` [dpdk-dev] [PATCH v3 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
  2 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-21  8:13 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the legacy RX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  27 +-
 drivers/net/iavf/iavf_rxtx.h            |   5 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 711 ++++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |  17 +
 4 files changed, 756 insertions(+), 4 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index 05a7dd8..c36e809 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2104,6 +2104,9 @@
 	struct iavf_rx_queue *rxq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_rx_vec_dev_check(dev)) {
 		for (i = 0; i < dev->data->nb_rx_queues; i++) {
@@ -2114,6 +2117,10 @@
 		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+			use_avx512 = true;
+#endif
 
 		if (dev->data->scattered_rx) {
 			PMD_DRV_LOG(DEBUG,
@@ -2121,27 +2128,39 @@
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
 					iavf_recv_scattered_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512;
+#endif
+			}
 		} else {
 			PMD_DRV_LOG(DEBUG, "Using %sVector Rx (port %d).",
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
 					iavf_recv_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512;
+#endif
+			}
 		}
 
 		return;
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 59625a9..cb12888 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -437,6 +437,11 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_tx_vec_dev_check(struct rte_eth_dev *dev);
 int iavf_rxq_vec_setup(struct iavf_rx_queue *rxq);
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
+uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
new file mode 100644
index 0000000..04c2df8
--- /dev/null
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -0,0 +1,711 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include "iavf_rxtx_vec_common.h"
+
+#include <x86intrin.h>
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+#define IAVF_DESCS_PER_LOOP_AVX 8
+#define PKTLEN_SHIFT 10
+
+static inline void
+iavf_rxq_rearm(struct iavf_rx_queue *rxq)
+{
+	int i;
+	uint16_t rx_id;
+	volatile union iavf_rx_desc *rxdp;
+	struct rte_mempool_cache *cache = rte_mempool_default_cache(rxq->mp,
+			rte_lcore_id());
+	struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start];
+
+	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.
+	 */
+
+	/* Can this be satisfied from the cache? */
+	if (cache->len < IAVF_RXQ_REARM_THRESH) {
+		/* No. Backfill the cache first, and then fill from it */
+		uint32_t req = IAVF_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 + IAVF_RXQ_REARM_THRESH >=
+					rxq->nb_rx_desc) {
+				__m128i dma_addr0;
+
+				dma_addr0 = _mm_setzero_si128();
+				for (i = 0; i < IAVF_VPMD_DESCS_PER_LOOP; i++) {
+					rxp[i] = &rxq->fake_mbuf;
+					_mm_storeu_si128((__m128i *)&rxdp[i].read,
+							 dma_addr0);
+				}
+			}
+			rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+					IAVF_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_IAVF_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 < IAVF_RXQ_REARM_THRESH / IAVF_DESCS_PER_LOOP_AVX; i++) {
+		const __m512i mbuf_ptrs = _mm512_loadu_si512
+			(&cache->objs[cache->len - IAVF_DESCS_PER_LOOP_AVX]);
+		_mm512_storeu_si512(rxp, mbuf_ptrs);
+
+		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_IAVF_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 6 & 7.
+		 */
+		const __m512i desc0_1 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas0);
+		const __m512i desc2_3 = _mm512_bsrli_epi128(desc0_1, 8);
+
+		const __m512i desc4_5 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas1);
+		const __m512i desc6_7 = _mm512_bsrli_epi128(desc4_5, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_1);
+		_mm512_storeu_si512((void *)(rxdp + 2), desc2_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_5);
+		_mm512_storeu_si512((void *)(rxdp + 6), desc6_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 desc0_3 = _mm512_permutexvar_epi64(permute_idx, iova_addrs);
+		const __m512i desc4_7 = _mm512_bsrli_epi128(desc0_3, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_7);
+#endif
+		rxp += IAVF_DESCS_PER_LOOP_AVX;
+		rxdp += IAVF_DESCS_PER_LOOP_AVX;
+		cache->len -= IAVF_DESCS_PER_LOOP_AVX;
+	}
+
+	rxq->rxrearm_start += IAVF_RXQ_REARM_THRESH;
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= IAVF_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 */
+	IAVF_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
+}
+
+#define IAVF_RX_LEN_MASK 0x80808080
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512(struct iavf_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 uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
+			0, rxq->mbuf_initializer);
+	/* struct iavf_rx_entry *sw_ring = &rxq->sw_ring[rxq->rx_tail]; */
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_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 << IAVF_RX_DESC_STATUS_DD_SHIFT)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_DESC_STATUS_EOF_SHIFT);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 4th descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF     /* pkt_type set as unknown */
+			);
+	/**
+	 * 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 += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 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 desc4_7 = _mm512_mask_blend_epi16(IAVF_RX_LEN_MASK,
+								raw_desc4_7,
+								len4_7);
+		__m512i mb4_7 = _mm512_shuffle_epi8(desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, 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 __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 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 __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, adjusting length and
+		 * re-arranging fields. Then write into the mbuf
+		 */
+		const __m512i len0_3 = _mm512_slli_epi32(raw_desc0_3,
+							 PKTLEN_SHIFT);
+		const __m512i desc0_3 = _mm512_mask_blend_epi16(IAVF_RX_LEN_MASK,
+								raw_desc0_3,
+								len0_3);
+		__m512i mb0_3 = _mm512_shuffle_epi8(desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/* get the packet types */
+		const __m512i ptypes0_3 = _mm512_srli_epi64(desc0_3, 30);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		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 ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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_flags =
+			_mm256_shuffle_epi8(rss_flags_shuf,
+					    _mm256_srli_epi32(flag_bits, 11));
+		/**
+		 * 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 */
+		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				_mm256_or_si256(rss_flags, vlan_flags));
+		/**
+		 * 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;
+		const __m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7, 0);
+		const __m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7, 1);
+		const __m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3, 0);
+		const __m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3, 1);
+
+		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 << IAVF_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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
+static inline __m256i
+flex_rxd_to_fdir_flags_vec_avx512(const __m256i fdir_id0_7)
+{
+#define FDID_MIS_MAGIC 0xFFFFFFFF
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR != (1 << 2));
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
+	const __m256i pkt_fdir_bit = _mm256_set1_epi32(PKT_RX_FDIR |
+			PKT_RX_FDIR_ID);
+	/* desc->flow_id field == 0xFFFFFFFF means fdir mismatch */
+	const __m256i fdir_mis_mask = _mm256_set1_epi32(FDID_MIS_MAGIC);
+	__m256i fdir_mask = _mm256_cmpeq_epi32(fdir_id0_7,
+			fdir_mis_mask);
+	/* this XOR op results to bit-reverse the fdir_mask */
+	fdir_mask = _mm256_xor_si256(fdir_mask, fdir_mis_mask);
+	const __m256i fdir_flags = _mm256_and_si256(fdir_mask, pkt_fdir_bit);
+
+	return fdir_flags;
+}
+
+/**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+			  uint16_t nb_pkts)
+{
+	return _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				     uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				    uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst = iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
diff --git a/drivers/net/iavf/meson.build b/drivers/net/iavf/meson.build
index a3fad36..00c7a12 100644
--- a/drivers/net/iavf/meson.build
+++ b/drivers/net/iavf/meson.build
@@ -34,4 +34,21 @@ if arch_subdir == 'x86'
 				c_args: [cflags, '-mavx2'])
 		objs += iavf_avx2_lib.extract_objects('iavf_rxtx_vec_avx2.c')
 	endif
+
+	if not machine_args.contains('-mno-avx512f')
+		if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F')
+			cflags += ['-DCC_AVX512_SUPPORT']
+			cflags += ['-march=skylake-avx512']
+			sources += files('iavf_rxtx_vec_avx512.c')
+		elif cc.has_argument('-mavx512f')
+			cflags += ['-DCC_AVX512_SUPPORT']
+			iavf_avx512_lib = static_library('iavf_avx512_lib',
+					'iavf_rxtx_vec_avx512.c',
+					dependencies: [static_rte_ethdev,
+						static_rte_kvargs, static_rte_hash],
+					include_directories: includes,
+					c_args: [cflags, '-mavx512f', '-march=skylake-avx512'])
+			objs += iavf_avx512_lib.extract_objects('iavf_rxtx_vec_avx512.c')
+		endif
+	endif
 endif
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v3 2/3] net/iavf: enable AVX512 for flexible RX
  2020-09-21  8:13 ` [dpdk-dev] [PATCH v3 " Wenzhuo Lu
  2020-09-21  8:13   ` [dpdk-dev] [PATCH v3 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
@ 2020-09-21  8:13   ` Wenzhuo Lu
  2020-09-21  8:13   ` [dpdk-dev] [PATCH v3 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
  2 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-21  8:13 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the flexible RX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  10 +
 drivers/net/iavf/iavf_rxtx.h            |   6 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 687 ++++++++++++++++++++++++++++++++
 3 files changed, 703 insertions(+)

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index c36e809..0818107 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2132,6 +2132,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
@@ -2151,6 +2156,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index cb12888..9653e0c 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -439,9 +439,15 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
 uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
 				   uint16_t nb_pkts);
+uint16_t iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					    struct rte_mbuf **rx_pkts,
+					    uint16_t nb_pkts);
 uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 					     struct rte_mbuf **rx_pkts,
 					     uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+						      struct rte_mbuf **rx_pkts,
+						      uint16_t nb_pkts);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index 04c2df8..2b6c99f 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -634,6 +634,612 @@
 	return fdir_flags;
 }
 
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512_flex_rxd(struct iavf_rx_queue *rxq,
+					struct rte_mbuf **rx_pkts,
+					uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
+			0, rxq->mbuf_initializer);
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_flex_desc *rxdp =
+		(union iavf_rx_flex_desc *)rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_rxq_rearm(rxq);
+
+	/* Before we start moving massive data around, check to see if
+	 * there is actually a packet available
+	 */
+	if (!(rxdp->wb.status_error0 &
+			rte_cpu_to_le_32(1 << IAVF_RX_FLEX_DESC_STATUS0_DD_S)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 4th descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF     /* pkt_type set as unknown */
+			);
+	/**
+	 * 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 Checksum Reports, RSS indication
+	 * and VLAN indication.
+	 * bit6:4 for IP/L4 checksum errors.
+	 * bit12 is for RSS indication.
+	 * bit13 is for VLAN indication.
+	 */
+	const __m256i flags_mask =
+		 _mm256_set1_epi32((7 << 4) | (1 << 12) | (1 << 13));
+	/**
+	 * data to be shuffled by the result of the flags mask shifted by 4
+	 * 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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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);
+	/**
+	 * data to be shuffled by result of flag mask, shifted down 12.
+	 * If RSS(bit12)/VLAN(bit13) are set,
+	 * shuffle moves appropriate flags in place.
+	 */
+	const __m256i rss_vlan_flags_shuf = _mm256_set_epi8(0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0,
+			/* end up 128-bits */
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0);
+
+	uint16_t i, received;
+
+	for (i = 0, received = 0; i < nb_pkts;
+	     i += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb4_7 = _mm512_shuffle_epi8(raw_desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptype_mask =
+			_mm512_set1_epi16(IAVF_RX_FLEX_DESC_PTYPE_M);
+		const __m512i ptypes4_7 =
+			_mm512_and_si512(raw_desc4_7, ptype_mask);
+		const __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 0);
+		const uint16_t ptype7 = _mm256_extract_epi16(ptypes6_7, 9);
+		const uint16_t ptype6 = _mm256_extract_epi16(ptypes6_7, 1);
+		const uint16_t ptype5 = _mm256_extract_epi16(ptypes4_5, 9);
+		const uint16_t ptype4 = _mm256_extract_epi16(ptypes4_5, 1);
+
+		const __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb0_3 = _mm512_shuffle_epi8(raw_desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptypes0_3 =
+			_mm512_and_si512(raw_desc0_3, ptype_mask);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		const uint16_t ptype3 = _mm256_extract_epi16(ptypes2_3, 9);
+		const uint16_t ptype2 = _mm256_extract_epi16(ptypes2_3, 1);
+		const uint16_t ptype1 = _mm256_extract_epi16(ptypes0_1, 9);
+		const uint16_t ptype0 = _mm256_extract_epi16(ptypes0_1, 1);
+
+		const __m512i ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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);
+		/**
+		 * 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, 4));
+		l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
+		l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
+		/* set rss and vlan flags */
+		const __m256i rss_vlan_flag_bits =
+			_mm256_srli_epi32(flag_bits, 12);
+		const __m256i rss_vlan_flags =
+			_mm256_shuffle_epi8(rss_vlan_flags_shuf,
+					    rss_vlan_flag_bits);
+
+		/* merge flags */
+		__m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				rss_vlan_flags);
+
+		if (rxq->fdir_enabled) {
+			const __m512i fdir_permute_mask = _mm512_set_epi32
+				(0, 0, 0, 0,
+				 0, 0, 0, 0,
+				 15, 11, 7, 3,
+				 31, 27, 23, 19);
+			__m512i fdir_tmp = _mm512_permutex2var_epi32
+				(raw_desc4_7, fdir_permute_mask, raw_desc0_3);
+			const __m256i fdir_id0_7 = _mm512_extracti64x4_epi64
+				(fdir_tmp, 0);
+			const __m256i fdir_flags =
+				flex_rxd_to_fdir_flags_vec_avx512(fdir_id0_7);
+
+			/* merge with fdir_flags */
+			mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_flags);
+
+			/* write to mbuf: have to use scalar store here */
+			rx_pkts[i + 0]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 3);
+
+			rx_pkts[i + 1]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 7);
+
+			rx_pkts[i + 2]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 2);
+
+			rx_pkts[i + 3]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 6);
+
+			rx_pkts[i + 4]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 1);
+
+			rx_pkts[i + 5]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 5);
+
+			rx_pkts[i + 6]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 0);
+
+			rx_pkts[i + 7]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 4);
+		} /* if() on fdir_enabled */
+
+		__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);
+
+#ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
+		/**
+		 * needs to load 2nd 16B of each desc for RSS hash parsing,
+		 * will cause performance drop to get into this context.
+		 */
+		if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
+				DEV_RX_OFFLOAD_RSS_HASH) {
+			/* load bottom half of every 32B desc */
+			const __m128i raw_desc_bh7 =
+				_mm_load_si128
+					((void *)(&rxdp[7].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh6 =
+				_mm_load_si128
+					((void *)(&rxdp[6].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh5 =
+				_mm_load_si128
+					((void *)(&rxdp[5].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh4 =
+				_mm_load_si128
+					((void *)(&rxdp[4].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh3 =
+				_mm_load_si128
+					((void *)(&rxdp[3].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh2 =
+				_mm_load_si128
+					((void *)(&rxdp[2].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh1 =
+				_mm_load_si128
+					((void *)(&rxdp[1].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh0 =
+				_mm_load_si128
+					((void *)(&rxdp[0].wb.status_error1));
+
+			__m256i raw_desc_bh6_7 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh6),
+					raw_desc_bh7, 1);
+			__m256i raw_desc_bh4_5 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh4),
+					raw_desc_bh5, 1);
+			__m256i raw_desc_bh2_3 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh2),
+					raw_desc_bh3, 1);
+			__m256i raw_desc_bh0_1 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh0),
+					raw_desc_bh1, 1);
+
+			/**
+			 * to shift the 32b RSS hash value to the
+			 * highest 32b of each 128b before mask
+			 */
+			__m256i rss_hash6_7 =
+				_mm256_slli_epi64(raw_desc_bh6_7, 32);
+			__m256i rss_hash4_5 =
+				_mm256_slli_epi64(raw_desc_bh4_5, 32);
+			__m256i rss_hash2_3 =
+				_mm256_slli_epi64(raw_desc_bh2_3, 32);
+			__m256i rss_hash0_1 =
+				_mm256_slli_epi64(raw_desc_bh0_1, 32);
+
+			__m256i rss_hash_msk =
+				_mm256_set_epi32(0xFFFFFFFF, 0, 0, 0,
+						 0xFFFFFFFF, 0, 0, 0);
+
+			rss_hash6_7 = _mm256_and_si256
+					(rss_hash6_7, rss_hash_msk);
+			rss_hash4_5 = _mm256_and_si256
+					(rss_hash4_5, rss_hash_msk);
+			rss_hash2_3 = _mm256_and_si256
+					(rss_hash2_3, rss_hash_msk);
+			rss_hash0_1 = _mm256_and_si256
+					(rss_hash0_1, rss_hash_msk);
+
+			mb6_7 = _mm256_or_si256(mb6_7, rss_hash6_7);
+			mb4_5 = _mm256_or_si256(mb4_5, rss_hash4_5);
+			mb2_3 = _mm256_or_si256(mb2_3, rss_hash2_3);
+			mb0_1 = _mm256_or_si256(mb0_1, rss_hash0_1);
+		} /* if() on RSS hash parsing */
+#endif
+
+		/**
+		 * 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 <<
+					       IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+			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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
 /**
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -646,6 +1252,18 @@
 }
 
 /**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts)
+{
+	return _iavf_recv_raw_pkts_vec_avx512_flex_rxd(rx_queue, rx_pkts,
+						       nb_pkts, NULL);
+}
+
+/**
  * vPMD receive routine that reassembles single burst of 32 scattered packets
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -709,3 +1327,72 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+/**
+ * vPMD receive routine that reassembles single burst of
+ * 32 scattered packets for flex RxD
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512_flex_rxd(void *rx_queue,
+					      struct rte_mbuf **rx_pkts,
+					      uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_recv_raw_pkts_vec_avx512_flex_rxd(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 for flex RxD.
+ * Main receive routine that can handle arbitrary burst sizes
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst =
+			iavf_recv_scattered_burst_vec_avx512_flex_rxd
+			(rx_queue, rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v3 3/3] net/iavf: enable AVX512 for TX
  2020-09-21  8:13 ` [dpdk-dev] [PATCH v3 " Wenzhuo Lu
  2020-09-21  8:13   ` [dpdk-dev] [PATCH v3 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
  2020-09-21  8:13   ` [dpdk-dev] [PATCH v3 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
@ 2020-09-21  8:13   ` Wenzhuo Lu
  2020-09-21 19:10     ` Morten Brørup
  2 siblings, 1 reply; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-21  8:13 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the TX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 doc/guides/rel_notes/release_20_11.rst  |   3 +
 drivers/net/iavf/iavf_ethdev.c          |   3 +-
 drivers/net/iavf/iavf_rxtx.c            |  32 +++-
 drivers/net/iavf/iavf_rxtx.h            |   7 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 304 ++++++++++++++++++++++++++++++++
 5 files changed, 341 insertions(+), 8 deletions(-)

diff --git a/doc/guides/rel_notes/release_20_11.rst b/doc/guides/rel_notes/release_20_11.rst
index df227a1..d40b8d6 100644
--- a/doc/guides/rel_notes/release_20_11.rst
+++ b/doc/guides/rel_notes/release_20_11.rst
@@ -55,6 +55,9 @@ New Features
      Also, make sure to start the actual text at the margin.
      =======================================================
 
+   * **Added support of vector instructions on IAVF.**
+
+     Added support of AVX512 instructions in IAVF RX and TX path.
 
 Removed Items
 -------------
diff --git a/drivers/net/iavf/iavf_ethdev.c b/drivers/net/iavf/iavf_ethdev.c
index c3aa4cd..5bc2851 100644
--- a/drivers/net/iavf/iavf_ethdev.c
+++ b/drivers/net/iavf/iavf_ethdev.c
@@ -528,7 +528,8 @@ static int iavf_config_rx_queues_irqs(struct rte_eth_dev *dev,
 		DEV_TX_OFFLOAD_GRE_TNL_TSO |
 		DEV_TX_OFFLOAD_IPIP_TNL_TSO |
 		DEV_TX_OFFLOAD_GENEVE_TNL_TSO |
-		DEV_TX_OFFLOAD_MULTI_SEGS;
+		DEV_TX_OFFLOAD_MULTI_SEGS |
+		DEV_TX_OFFLOAD_MBUF_FAST_FREE;
 
 	dev_info->default_rxconf = (struct rte_eth_rxconf) {
 		.rx_free_thresh = IAVF_DEFAULT_RX_FREE_THRESH,
diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index 0818107..04dcd48 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2206,18 +2206,18 @@
 	struct iavf_tx_queue *txq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_tx_vec_dev_check(dev)) {
-		for (i = 0; i < dev->data->nb_tx_queues; i++) {
-			txq = dev->data->tx_queues[i];
-			if (!txq)
-				continue;
-			iavf_txq_vec_setup(txq);
-		}
-
 		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+			use_avx512 = true;
+#endif
 
 		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
 			    use_avx2 ? "avx2 " : "",
@@ -2225,8 +2225,26 @@
 		dev->tx_pkt_burst = use_avx2 ?
 				    iavf_xmit_pkts_vec_avx2 :
 				    iavf_xmit_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+		if (use_avx512)
+			dev->tx_pkt_burst = iavf_xmit_pkts_vec_avx512;
+#endif
 		dev->tx_pkt_prepare = NULL;
 
+		for (i = 0; i < dev->data->nb_tx_queues; i++) {
+			txq = dev->data->tx_queues[i];
+			if (!txq)
+				continue;
+#ifdef CC_AVX512_SUPPORT
+			if (use_avx512)
+				iavf_txq_vec_setup_avx512(txq);
+			else
+				iavf_txq_vec_setup(txq);
+#else
+			iavf_txq_vec_setup(txq);
+#endif
+		}
+
 		return;
 	}
 #endif
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 9653e0c..08eebb0 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -122,6 +122,10 @@ struct iavf_tx_entry {
 	uint16_t last_id;
 };
 
+struct iavf_tx_vec_entry {
+	struct rte_mbuf *mbuf;
+};
+
 /* Structure associated with each TX queue. */
 struct iavf_tx_queue {
 	const struct rte_memzone *mz;  /* memzone for Tx ring */
@@ -448,6 +452,9 @@ uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
 						      struct rte_mbuf **rx_pkts,
 						      uint16_t nb_pkts);
+uint16_t iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				   uint16_t nb_pkts);
+int iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index 2b6c99f..4a33930 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -1396,3 +1396,307 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+static __rte_always_inline int
+iavf_tx_free_bufs_avx512(struct iavf_tx_queue *txq)
+{
+	struct iavf_tx_vec_entry *txep;
+	uint32_t n;
+	uint32_t i;
+	int nb_free = 0;
+	struct rte_mbuf *m, *free[IAVF_VPMD_TX_MAX_FREE_BUF];
+
+	/* check DD bits on threshold descriptor */
+	if ((txq->tx_ring[txq->next_dd].cmd_type_offset_bsz &
+			rte_cpu_to_le_64(IAVF_TXD_QW1_DTYPE_MASK)) !=
+			rte_cpu_to_le_64(IAVF_TX_DESC_DTYPE_DESC_DONE))
+		return 0;
+
+	n = txq->rs_thresh;
+
+	 /* first buffer to free from S/W ring is at index
+	  * tx_next_dd - (tx_rs_thresh-1)
+	  */
+	txep = (void *)txq->sw_ring;
+	txep += txq->next_dd - (n - 1);
+
+	if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE && (n & 31) == 0) {
+		struct rte_mempool *mp = txep[0].mbuf->pool;
+		struct rte_mempool_cache *cache = rte_mempool_default_cache(mp,
+				rte_lcore_id());
+		void **cache_objs = &cache->objs[cache->len];
+
+		if (n > RTE_MEMPOOL_CACHE_MAX_SIZE) {
+			rte_mempool_ops_enqueue_bulk(mp, (void *)txep, n);
+			goto done;
+		}
+
+		/* The cache follows the following algorithm
+		 *   1. Add the objects to the cache
+		 *   2. Anything greater than the cache min value (if it crosses the
+		 *   cache flush threshold) is flushed to the ring.
+		 */
+		/* Add elements back into the cache */
+		uint32_t copied = 0;
+		/* n is multiple of 32 */
+		while (copied < n) {
+			const __m512i a = _mm512_loadu_si512(&txep[copied]);
+			const __m512i b = _mm512_loadu_si512(&txep[copied + 8]);
+			const __m512i c = _mm512_loadu_si512(&txep[copied + 16]);
+			const __m512i d = _mm512_loadu_si512(&txep[copied + 24]);
+
+			_mm512_storeu_si512(&cache_objs[copied], a);
+			_mm512_storeu_si512(&cache_objs[copied + 8], b);
+			_mm512_storeu_si512(&cache_objs[copied + 16], c);
+			_mm512_storeu_si512(&cache_objs[copied + 24], d);
+			copied += 32;
+		}
+		cache->len += n;
+
+		if (cache->len >= cache->flushthresh) {
+			rte_mempool_ops_enqueue_bulk(mp,
+						     &cache->objs[cache->size],
+						     cache->len - cache->size);
+			cache->len = cache->size;
+		}
+		goto done;
+	}
+
+	m = rte_pktmbuf_prefree_seg(txep[0].mbuf);
+	if (likely(m)) {
+		free[0] = m;
+		nb_free = 1;
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (likely(m)) {
+				if (likely(m->pool == free[0]->pool)) {
+					free[nb_free++] = m;
+				} else {
+					rte_mempool_put_bulk(free[0]->pool,
+							     (void *)free,
+							     nb_free);
+					free[0] = m;
+					nb_free = 1;
+				}
+			}
+		}
+		rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
+	} else {
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (m)
+				rte_mempool_put(m->pool, m);
+		}
+	}
+
+done:
+	/* buffers were freed, update counters */
+	txq->nb_free = (uint16_t)(txq->nb_free + txq->rs_thresh);
+	txq->next_dd = (uint16_t)(txq->next_dd + txq->rs_thresh);
+	if (txq->next_dd >= txq->nb_tx_desc)
+		txq->next_dd = (uint16_t)(txq->rs_thresh - 1);
+
+	return txq->rs_thresh;
+}
+
+static __rte_always_inline void
+tx_backlog_entry_avx512(struct iavf_tx_vec_entry *txep,
+			struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	int i;
+
+	for (i = 0; i < (int)nb_pkts; ++i)
+		txep[i].mbuf = tx_pkts[i];
+}
+
+static inline void
+iavf_vtx1(volatile struct iavf_tx_desc *txdp,
+	  struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(IAVF_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT) |
+		 ((uint64_t)pkt->data_len << IAVF_TXD_QW1_TX_BUF_SZ_SHIFT));
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+				pkt->buf_iova + pkt->data_off);
+	_mm_storeu_si128((__m128i *)txdp, descriptor);
+}
+
+#define IAVF_TX_LEN_MASK 0xAA
+#define IAVF_TX_OFF_MASK 0x55
+static inline void
+iavf_vtx(volatile struct iavf_tx_desc *txdp,
+	 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (IAVF_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT));
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		iavf_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) {
+		__m512i desc4 =
+			_mm512_set_epi64
+				((uint64_t)pkt[3]->data_len,
+				 pkt[3]->buf_iova,
+				 (uint64_t)pkt[2]->data_len,
+				 pkt[2]->buf_iova,
+				 (uint64_t)pkt[1]->data_len,
+				 pkt[1]->buf_iova,
+				 (uint64_t)pkt[0]->data_len,
+				 pkt[0]->buf_iova);
+		__m512i hi_qw_tmpl_4 = _mm512_set1_epi64(hi_qw_tmpl);
+		__m512i data_off_4 =
+			_mm512_set_epi64
+				(0,
+				 pkt[3]->data_off,
+				 0,
+				 pkt[2]->data_off,
+				 0,
+				 pkt[1]->data_off,
+				 0,
+				 pkt[0]->data_off);
+
+		desc4 = _mm512_mask_slli_epi64(desc4, IAVF_TX_LEN_MASK, desc4,
+					       IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
+		desc4 = _mm512_mask_or_epi64(desc4, IAVF_TX_LEN_MASK, desc4,
+					     hi_qw_tmpl_4);
+		desc4 = _mm512_mask_add_epi64(desc4, IAVF_TX_OFF_MASK, desc4,
+					      data_off_4);
+		_mm512_storeu_si512((void *)txdp, desc4);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		iavf_vtx1(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+iavf_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				 uint16_t nb_pkts)
+{
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+	volatile struct iavf_tx_desc *txdp;
+	struct iavf_tx_vec_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	/* bit2 is reserved and must be set to 1 according to Spec */
+	uint64_t flags = IAVF_TX_DESC_CMD_EOP | IAVF_TX_DESC_CMD_ICRC;
+	uint64_t rs = IAVF_TX_DESC_CMD_RS | flags;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
+
+	if (txq->nb_free < txq->free_thresh)
+		iavf_tx_free_bufs_avx512(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = (void *)txq->sw_ring;
+	txep += tx_id;
+
+	txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		tx_backlog_entry_avx512(txep, tx_pkts, n);
+
+		iavf_vtx(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		iavf_vtx1(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = (void *)txq->sw_ring;
+		txep += tx_id;
+	}
+
+	tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
+
+	iavf_vtx(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->next_rs) {
+		txq->tx_ring[txq->next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)IAVF_TX_DESC_CMD_RS) <<
+					 IAVF_TXD_QW1_CMD_SHIFT);
+		txq->next_rs =
+			(uint16_t)(txq->next_rs + txq->rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	IAVF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+			  uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->rs_thresh);
+		ret = iavf_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;
+}
+
+static inline void
+iavf_tx_queue_release_mbufs_avx512(struct iavf_tx_queue *txq)
+{
+	unsigned int i;
+	const uint16_t max_desc = (uint16_t)(txq->nb_tx_desc - 1);
+	struct iavf_tx_vec_entry *swr = (void *)txq->sw_ring;
+
+	if (!txq->sw_ring || txq->nb_free == max_desc)
+		return;
+
+	i = txq->next_dd - txq->rs_thresh + 1;
+	if (txq->tx_tail < i) {
+		for (; i < txq->nb_tx_desc; i++) {
+			rte_pktmbuf_free_seg(swr[i].mbuf);
+			swr[i].mbuf = NULL;
+		}
+		i = 0;
+	}
+}
+
+static const struct iavf_txq_ops avx512_vec_txq_ops = {
+	.release_mbufs = iavf_tx_queue_release_mbufs_avx512,
+};
+
+int __rte_cold
+iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq)
+{
+	txq->ops = &avx512_vec_txq_ops;
+	return 0;
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* Re: [dpdk-dev] [PATCH v3 3/3] net/iavf: enable AVX512 for TX
  2020-09-21  8:13   ` [dpdk-dev] [PATCH v3 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
@ 2020-09-21 19:10     ` Morten Brørup
  2020-09-22  1:34       ` Lu, Wenzhuo
  0 siblings, 1 reply; 39+ messages in thread
From: Morten Brørup @ 2020-09-21 19:10 UTC (permalink / raw)
  To: Wenzhuo Lu, dev; +Cc: Bruce Richardson, Leyi Rong

> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Wenzhuo Lu
> Sent: Monday, September 21, 2020 10:14 AM
> 
> To enhance the per-core performance, this patch adds some AVX512
> instructions to the data path to handle the TX descriptors.
> 
> Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> Signed-off-by: Leyi Rong <leyi.rong@intel.com>

[...]

> +static inline void
> +iavf_vtx(volatile struct iavf_tx_desc *txdp,
> +	 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
> +{
> +	const uint64_t hi_qw_tmpl = (IAVF_TX_DESC_DTYPE_DATA |
> +			((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT));
> +
> +	/* if unaligned on 32-bit boundary, do one to align */
> +	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
> +		iavf_vtx1(txdp, *pkt, flags);
> +		nb_pkts--, txdp++, pkt++;
> +	}
> +
> +	/* do two at a time while possible, in bursts */

It looks like four at a time, not two.

> +	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
> +		__m512i desc4 =
> +			_mm512_set_epi64
> +				((uint64_t)pkt[3]->data_len,
> +				 pkt[3]->buf_iova,
> +				 (uint64_t)pkt[2]->data_len,
> +				 pkt[2]->buf_iova,
> +				 (uint64_t)pkt[1]->data_len,
> +				 pkt[1]->buf_iova,
> +				 (uint64_t)pkt[0]->data_len,
> +				 pkt[0]->buf_iova);
> +		__m512i hi_qw_tmpl_4 = _mm512_set1_epi64(hi_qw_tmpl);
> +		__m512i data_off_4 =
> +			_mm512_set_epi64
> +				(0,
> +				 pkt[3]->data_off,
> +				 0,
> +				 pkt[2]->data_off,
> +				 0,
> +				 pkt[1]->data_off,
> +				 0,
> +				 pkt[0]->data_off);
> +
> +		desc4 = _mm512_mask_slli_epi64(desc4, IAVF_TX_LEN_MASK,
> desc4,
> +					       IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
> +		desc4 = _mm512_mask_or_epi64(desc4, IAVF_TX_LEN_MASK,
> desc4,
> +					     hi_qw_tmpl_4);
> +		desc4 = _mm512_mask_add_epi64(desc4, IAVF_TX_OFF_MASK,
> desc4,
> +					      data_off_4);
> +		_mm512_storeu_si512((void *)txdp, desc4);
> +	}
> +
> +	/* do any last ones */
> +	while (nb_pkts) {
> +		iavf_vtx1(txdp, *pkt, flags);
> +		txdp++, pkt++, nb_pkts--;
> +	}
> +}

^ permalink raw reply	[flat|nested] 39+ messages in thread

* Re: [dpdk-dev] [PATCH v3 3/3] net/iavf: enable AVX512 for TX
  2020-09-21 19:10     ` Morten Brørup
@ 2020-09-22  1:34       ` Lu, Wenzhuo
  0 siblings, 0 replies; 39+ messages in thread
From: Lu, Wenzhuo @ 2020-09-22  1:34 UTC (permalink / raw)
  To: Morten Brørup, dev; +Cc: Richardson, Bruce, Rong, Leyi

Hi Morten,

> -----Original Message-----
> From: Morten Brørup <mb@smartsharesystems.com>
> Sent: Tuesday, September 22, 2020 3:11 AM
> To: Lu, Wenzhuo <wenzhuo.lu@intel.com>; dev@dpdk.org
> Cc: Richardson, Bruce <bruce.richardson@intel.com>; Rong, Leyi
> <leyi.rong@intel.com>
> Subject: RE: [dpdk-dev] [PATCH v3 3/3] net/iavf: enable AVX512 for TX
> 
> > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Wenzhuo Lu
> > Sent: Monday, September 21, 2020 10:14 AM
> >
> > To enhance the per-core performance, this patch adds some AVX512
> > instructions to the data path to handle the TX descriptors.
> >
> > Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
> > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > Signed-off-by: Leyi Rong <leyi.rong@intel.com>
> 
> [...]
> 
> > +static inline void
> > +iavf_vtx(volatile struct iavf_tx_desc *txdp,
> > +	 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags) {
> > +	const uint64_t hi_qw_tmpl = (IAVF_TX_DESC_DTYPE_DATA |
> > +			((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT));
> > +
> > +	/* if unaligned on 32-bit boundary, do one to align */
> > +	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
> > +		iavf_vtx1(txdp, *pkt, flags);
> > +		nb_pkts--, txdp++, pkt++;
> > +	}
> > +
> > +	/* do two at a time while possible, in bursts */
> 
> It looks like four at a time, not two.
Sorry for the mistake, will correct it.

> 
> > +	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {

^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v4 0/3] enable AVX512 for iavf
  2020-09-10  5:59 [dpdk-dev] [PATCH 0/3] enable AVX512 for iavf Wenzhuo Lu
                   ` (4 preceding siblings ...)
  2020-09-21  8:13 ` [dpdk-dev] [PATCH v3 " Wenzhuo Lu
@ 2020-09-27  1:30 ` Wenzhuo Lu
  2020-09-27  1:30   ` [dpdk-dev] [PATCH v4 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
                     ` (2 more replies)
  2020-10-21  7:47 ` [dpdk-dev] [PATCH v5 0/3] enable AVX512 for iavf Wenzhuo Lu
                   ` (2 subsequent siblings)
  8 siblings, 3 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-27  1:30 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu

AVX512 instructions is supported by more and more platforms. These instructions can be used in the data path to enhance the per-core performance of packet processing.
Comparing with the existing implementation, this path set introduces some AVX512 instructions into the iavf data path, and we get a better per-code throughput.

v2:
Update meson.build.
Repalce the deprecated 'buf_physaddr' by 'buf_iova'.

v3:
Fix compile errors.

v4:
Fix wrong info in commnets.
Trivial adjustment of the arrangement.

Wenzhuo Lu (3):
  net/iavf: enable AVX512 for legacy RX
  net/iavf: enable AVX512 for flexible RX
  net/iavf: enable AVX512 for TX

 doc/guides/rel_notes/release_20_11.rst  |    3 +
 drivers/net/iavf/iavf_ethdev.c          |    3 +-
 drivers/net/iavf/iavf_rxtx.c            |   69 +-
 drivers/net/iavf/iavf_rxtx.h            |   18 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 1702 +++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |   17 +
 6 files changed, 1800 insertions(+), 12 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v4 1/3] net/iavf: enable AVX512 for legacy RX
  2020-09-27  1:30 ` [dpdk-dev] [PATCH v4 0/3] enable AVX512 for iavf Wenzhuo Lu
@ 2020-09-27  1:30   ` Wenzhuo Lu
  2020-09-27  1:30   ` [dpdk-dev] [PATCH v4 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
  2020-09-27  1:30   ` [dpdk-dev] [PATCH v4 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
  2 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-27  1:30 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the legacy RX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  27 +-
 drivers/net/iavf/iavf_rxtx.h            |   5 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 710 ++++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |  17 +
 4 files changed, 755 insertions(+), 4 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index 05a7dd8..c36e809 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2104,6 +2104,9 @@
 	struct iavf_rx_queue *rxq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_rx_vec_dev_check(dev)) {
 		for (i = 0; i < dev->data->nb_rx_queues; i++) {
@@ -2114,6 +2117,10 @@
 		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+			use_avx512 = true;
+#endif
 
 		if (dev->data->scattered_rx) {
 			PMD_DRV_LOG(DEBUG,
@@ -2121,27 +2128,39 @@
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
 					iavf_recv_scattered_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512;
+#endif
+			}
 		} else {
 			PMD_DRV_LOG(DEBUG, "Using %sVector Rx (port %d).",
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
 					iavf_recv_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512;
+#endif
+			}
 		}
 
 		return;
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 59625a9..cb12888 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -437,6 +437,11 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_tx_vec_dev_check(struct rte_eth_dev *dev);
 int iavf_rxq_vec_setup(struct iavf_rx_queue *rxq);
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
+uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
new file mode 100644
index 0000000..a28c39b
--- /dev/null
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -0,0 +1,710 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include "iavf_rxtx_vec_common.h"
+
+#include <x86intrin.h>
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+#define IAVF_DESCS_PER_LOOP_AVX 8
+#define PKTLEN_SHIFT 10
+
+static inline void
+iavf_rxq_rearm(struct iavf_rx_queue *rxq)
+{
+	int i;
+	uint16_t rx_id;
+	volatile union iavf_rx_desc *rxdp;
+	struct rte_mempool_cache *cache =
+		rte_mempool_default_cache(rxq->mp, rte_lcore_id());
+	struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start];
+
+	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.
+	 */
+
+	/* Can this be satisfied from the cache? */
+	if (cache->len < IAVF_RXQ_REARM_THRESH) {
+		/* No. Backfill the cache first, and then fill from it */
+		uint32_t req = IAVF_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 + IAVF_RXQ_REARM_THRESH >=
+			    rxq->nb_rx_desc) {
+				__m128i dma_addr0;
+
+				dma_addr0 = _mm_setzero_si128();
+				for (i = 0; i < IAVF_VPMD_DESCS_PER_LOOP; i++) {
+					rxp[i] = &rxq->fake_mbuf;
+					_mm_storeu_si128((__m128i *)&rxdp[i].read,
+							 dma_addr0);
+				}
+			}
+			rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+					IAVF_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_IAVF_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 < IAVF_RXQ_REARM_THRESH / IAVF_DESCS_PER_LOOP_AVX; i++) {
+		const __m512i mbuf_ptrs = _mm512_loadu_si512
+			(&cache->objs[cache->len - IAVF_DESCS_PER_LOOP_AVX]);
+		_mm512_storeu_si512(rxp, mbuf_ptrs);
+
+		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_IAVF_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 6 & 7.
+		 */
+		const __m512i desc0_1 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas0);
+		const __m512i desc2_3 = _mm512_bsrli_epi128(desc0_1, 8);
+
+		const __m512i desc4_5 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas1);
+		const __m512i desc6_7 = _mm512_bsrli_epi128(desc4_5, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_1);
+		_mm512_storeu_si512((void *)(rxdp + 2), desc2_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_5);
+		_mm512_storeu_si512((void *)(rxdp + 6), desc6_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 desc0_3 = _mm512_permutexvar_epi64(permute_idx,
+								 iova_addrs);
+		const __m512i desc4_7 = _mm512_bsrli_epi128(desc0_3, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_7);
+#endif
+		rxp += IAVF_DESCS_PER_LOOP_AVX;
+		rxdp += IAVF_DESCS_PER_LOOP_AVX;
+		cache->len -= IAVF_DESCS_PER_LOOP_AVX;
+	}
+
+	rxq->rxrearm_start += IAVF_RXQ_REARM_THRESH;
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= IAVF_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 */
+	IAVF_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
+}
+
+#define IAVF_RX_LEN_MASK 0x80808080
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512(struct iavf_rx_queue *rxq,
+			       struct rte_mbuf **rx_pkts,
+			       uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0, 0,
+						    rxq->mbuf_initializer);
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_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 << IAVF_RX_DESC_STATUS_DD_SHIFT)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_DESC_STATUS_EOF_SHIFT);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 4th descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF     /* pkt_type set as unknown */
+			);
+	/**
+	 * 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 += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 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 desc4_7 = _mm512_mask_blend_epi16(IAVF_RX_LEN_MASK,
+								raw_desc4_7,
+								len4_7);
+		__m512i mb4_7 = _mm512_shuffle_epi8(desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, 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 __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 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 __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, adjusting length and
+		 * re-arranging fields. Then write into the mbuf
+		 */
+		const __m512i len0_3 = _mm512_slli_epi32(raw_desc0_3,
+							 PKTLEN_SHIFT);
+		const __m512i desc0_3 = _mm512_mask_blend_epi16(IAVF_RX_LEN_MASK,
+								raw_desc0_3,
+								len0_3);
+		__m512i mb0_3 = _mm512_shuffle_epi8(desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/* get the packet types */
+		const __m512i ptypes0_3 = _mm512_srli_epi64(desc0_3, 30);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		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 ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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_flags =
+			_mm256_shuffle_epi8(rss_flags_shuf,
+					    _mm256_srli_epi32(flag_bits, 11));
+		/**
+		 * 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 */
+		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				_mm256_or_si256(rss_flags, vlan_flags));
+		/**
+		 * 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;
+		const __m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7, 0);
+		const __m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7, 1);
+		const __m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3, 0);
+		const __m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3, 1);
+
+		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 << IAVF_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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
+static inline __m256i
+flex_rxd_to_fdir_flags_vec_avx512(const __m256i fdir_id0_7)
+{
+#define FDID_MIS_MAGIC 0xFFFFFFFF
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR != (1 << 2));
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
+	const __m256i pkt_fdir_bit = _mm256_set1_epi32(PKT_RX_FDIR |
+						       PKT_RX_FDIR_ID);
+	/* desc->flow_id field == 0xFFFFFFFF means fdir mismatch */
+	const __m256i fdir_mis_mask = _mm256_set1_epi32(FDID_MIS_MAGIC);
+	__m256i fdir_mask = _mm256_cmpeq_epi32(fdir_id0_7,
+					       fdir_mis_mask);
+	/* this XOR op results to bit-reverse the fdir_mask */
+	fdir_mask = _mm256_xor_si256(fdir_mask, fdir_mis_mask);
+	const __m256i fdir_flags = _mm256_and_si256(fdir_mask, pkt_fdir_bit);
+
+	return fdir_flags;
+}
+
+/**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+			  uint16_t nb_pkts)
+{
+	return _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				     uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				    uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst = iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
diff --git a/drivers/net/iavf/meson.build b/drivers/net/iavf/meson.build
index a3fad36..00c7a12 100644
--- a/drivers/net/iavf/meson.build
+++ b/drivers/net/iavf/meson.build
@@ -34,4 +34,21 @@ if arch_subdir == 'x86'
 				c_args: [cflags, '-mavx2'])
 		objs += iavf_avx2_lib.extract_objects('iavf_rxtx_vec_avx2.c')
 	endif
+
+	if not machine_args.contains('-mno-avx512f')
+		if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F')
+			cflags += ['-DCC_AVX512_SUPPORT']
+			cflags += ['-march=skylake-avx512']
+			sources += files('iavf_rxtx_vec_avx512.c')
+		elif cc.has_argument('-mavx512f')
+			cflags += ['-DCC_AVX512_SUPPORT']
+			iavf_avx512_lib = static_library('iavf_avx512_lib',
+					'iavf_rxtx_vec_avx512.c',
+					dependencies: [static_rte_ethdev,
+						static_rte_kvargs, static_rte_hash],
+					include_directories: includes,
+					c_args: [cflags, '-mavx512f', '-march=skylake-avx512'])
+			objs += iavf_avx512_lib.extract_objects('iavf_rxtx_vec_avx512.c')
+		endif
+	endif
 endif
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v4 2/3] net/iavf: enable AVX512 for flexible RX
  2020-09-27  1:30 ` [dpdk-dev] [PATCH v4 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-09-27  1:30   ` [dpdk-dev] [PATCH v4 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
@ 2020-09-27  1:30   ` Wenzhuo Lu
  2020-09-27  1:30   ` [dpdk-dev] [PATCH v4 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
  2 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-27  1:30 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the flexible RX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  10 +
 drivers/net/iavf/iavf_rxtx.h            |   6 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 688 ++++++++++++++++++++++++++++++++
 3 files changed, 704 insertions(+)

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index c36e809..0818107 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2132,6 +2132,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
@@ -2151,6 +2156,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index cb12888..9653e0c 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -439,9 +439,15 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
 uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
 				   uint16_t nb_pkts);
+uint16_t iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					    struct rte_mbuf **rx_pkts,
+					    uint16_t nb_pkts);
 uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 					     struct rte_mbuf **rx_pkts,
 					     uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+						      struct rte_mbuf **rx_pkts,
+						      uint16_t nb_pkts);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index a28c39b..63320e6 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -633,6 +633,612 @@
 	return fdir_flags;
 }
 
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512_flex_rxd(struct iavf_rx_queue *rxq,
+					struct rte_mbuf **rx_pkts,
+					uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0, 0,
+						    rxq->mbuf_initializer);
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_flex_desc *rxdp =
+		(union iavf_rx_flex_desc *)rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_rxq_rearm(rxq);
+
+	/* Before we start moving massive data around, check to see if
+	 * there is actually a packet available
+	 */
+	if (!(rxdp->wb.status_error0 &
+	      rte_cpu_to_le_32(1 << IAVF_RX_FLEX_DESC_STATUS0_DD_S)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 4th descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF     /* pkt_type set as unknown */
+			);
+	/**
+	 * 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 Checksum Reports, RSS indication
+	 * and VLAN indication.
+	 * bit6:4 for IP/L4 checksum errors.
+	 * bit12 is for RSS indication.
+	 * bit13 is for VLAN indication.
+	 */
+	const __m256i flags_mask =
+		_mm256_set1_epi32((7 << 4) | (1 << 12) | (1 << 13));
+	/**
+	 * data to be shuffled by the result of the flags mask shifted by 4
+	 * 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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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);
+	/**
+	 * data to be shuffled by result of flag mask, shifted down 12.
+	 * If RSS(bit12)/VLAN(bit13) are set,
+	 * shuffle moves appropriate flags in place.
+	 */
+	const __m256i rss_vlan_flags_shuf = _mm256_set_epi8(0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0,
+			/* end up 128-bits */
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0);
+
+	uint16_t i, received;
+
+	for (i = 0, received = 0; i < nb_pkts;
+	     i += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb4_7 = _mm512_shuffle_epi8(raw_desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptype_mask =
+			_mm512_set1_epi16(IAVF_RX_FLEX_DESC_PTYPE_M);
+		const __m512i ptypes4_7 =
+			_mm512_and_si512(raw_desc4_7, ptype_mask);
+		const __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 0);
+		const uint16_t ptype7 = _mm256_extract_epi16(ptypes6_7, 9);
+		const uint16_t ptype6 = _mm256_extract_epi16(ptypes6_7, 1);
+		const uint16_t ptype5 = _mm256_extract_epi16(ptypes4_5, 9);
+		const uint16_t ptype4 = _mm256_extract_epi16(ptypes4_5, 1);
+
+		const __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb0_3 = _mm512_shuffle_epi8(raw_desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptypes0_3 =
+			_mm512_and_si512(raw_desc0_3, ptype_mask);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		const uint16_t ptype3 = _mm256_extract_epi16(ptypes2_3, 9);
+		const uint16_t ptype2 = _mm256_extract_epi16(ptypes2_3, 1);
+		const uint16_t ptype1 = _mm256_extract_epi16(ptypes0_1, 9);
+		const uint16_t ptype0 = _mm256_extract_epi16(ptypes0_1, 1);
+
+		const __m512i ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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);
+		/**
+		 * 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, 4));
+		l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
+		l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
+		/* set rss and vlan flags */
+		const __m256i rss_vlan_flag_bits =
+			_mm256_srli_epi32(flag_bits, 12);
+		const __m256i rss_vlan_flags =
+			_mm256_shuffle_epi8(rss_vlan_flags_shuf,
+					    rss_vlan_flag_bits);
+
+		/* merge flags */
+		__m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+						     rss_vlan_flags);
+
+		if (rxq->fdir_enabled) {
+			const __m512i fdir_permute_mask = _mm512_set_epi32
+				(0, 0, 0, 0,
+				 0, 0, 0, 0,
+				 15, 11, 7, 3,
+				 31, 27, 23, 19);
+			__m512i fdir_tmp = _mm512_permutex2var_epi32
+				(raw_desc4_7, fdir_permute_mask, raw_desc0_3);
+			const __m256i fdir_id0_7 = _mm512_extracti64x4_epi64
+				(fdir_tmp, 0);
+			const __m256i fdir_flags =
+				flex_rxd_to_fdir_flags_vec_avx512(fdir_id0_7);
+
+			/* merge with fdir_flags */
+			mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_flags);
+
+			/* write to mbuf: have to use scalar store here */
+			rx_pkts[i + 0]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 3);
+
+			rx_pkts[i + 1]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 7);
+
+			rx_pkts[i + 2]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 2);
+
+			rx_pkts[i + 3]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 6);
+
+			rx_pkts[i + 4]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 1);
+
+			rx_pkts[i + 5]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 5);
+
+			rx_pkts[i + 6]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 0);
+
+			rx_pkts[i + 7]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 4);
+		} /* if() on fdir_enabled */
+
+		__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);
+
+#ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
+		/**
+		 * needs to load 2nd 16B of each desc for RSS hash parsing,
+		 * will cause performance drop to get into this context.
+		 */
+		if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
+		    DEV_RX_OFFLOAD_RSS_HASH) {
+			/* load bottom half of every 32B desc */
+			const __m128i raw_desc_bh7 =
+				_mm_load_si128
+					((void *)(&rxdp[7].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh6 =
+				_mm_load_si128
+					((void *)(&rxdp[6].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh5 =
+				_mm_load_si128
+					((void *)(&rxdp[5].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh4 =
+				_mm_load_si128
+					((void *)(&rxdp[4].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh3 =
+				_mm_load_si128
+					((void *)(&rxdp[3].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh2 =
+				_mm_load_si128
+					((void *)(&rxdp[2].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh1 =
+				_mm_load_si128
+					((void *)(&rxdp[1].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh0 =
+				_mm_load_si128
+					((void *)(&rxdp[0].wb.status_error1));
+
+			__m256i raw_desc_bh6_7 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh6),
+					 raw_desc_bh7, 1);
+			__m256i raw_desc_bh4_5 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh4),
+					 raw_desc_bh5, 1);
+			__m256i raw_desc_bh2_3 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh2),
+					 raw_desc_bh3, 1);
+			__m256i raw_desc_bh0_1 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh0),
+					 raw_desc_bh1, 1);
+
+			/**
+			 * to shift the 32b RSS hash value to the
+			 * highest 32b of each 128b before mask
+			 */
+			__m256i rss_hash6_7 =
+				_mm256_slli_epi64(raw_desc_bh6_7, 32);
+			__m256i rss_hash4_5 =
+				_mm256_slli_epi64(raw_desc_bh4_5, 32);
+			__m256i rss_hash2_3 =
+				_mm256_slli_epi64(raw_desc_bh2_3, 32);
+			__m256i rss_hash0_1 =
+				_mm256_slli_epi64(raw_desc_bh0_1, 32);
+
+			__m256i rss_hash_msk =
+				_mm256_set_epi32(0xFFFFFFFF, 0, 0, 0,
+						 0xFFFFFFFF, 0, 0, 0);
+
+			rss_hash6_7 = _mm256_and_si256
+					(rss_hash6_7, rss_hash_msk);
+			rss_hash4_5 = _mm256_and_si256
+					(rss_hash4_5, rss_hash_msk);
+			rss_hash2_3 = _mm256_and_si256
+					(rss_hash2_3, rss_hash_msk);
+			rss_hash0_1 = _mm256_and_si256
+					(rss_hash0_1, rss_hash_msk);
+
+			mb6_7 = _mm256_or_si256(mb6_7, rss_hash6_7);
+			mb4_5 = _mm256_or_si256(mb4_5, rss_hash4_5);
+			mb2_3 = _mm256_or_si256(mb2_3, rss_hash2_3);
+			mb0_1 = _mm256_or_si256(mb0_1, rss_hash0_1);
+		} /* if() on RSS hash parsing */
+#endif
+
+		/**
+		 * 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 <<
+					       IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+			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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
 /**
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -645,6 +1251,18 @@
 }
 
 /**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts)
+{
+	return _iavf_recv_raw_pkts_vec_avx512_flex_rxd(rx_queue, rx_pkts,
+						       nb_pkts, NULL);
+}
+
+/**
  * vPMD receive routine that reassembles single burst of 32 scattered packets
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -708,3 +1326,73 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+/**
+ * vPMD receive routine that reassembles single burst of
+ * 32 scattered packets for flex RxD
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512_flex_rxd(void *rx_queue,
+					      struct rte_mbuf **rx_pkts,
+					      uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_recv_raw_pkts_vec_avx512_flex_rxd(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 for flex RxD.
+ * Main receive routine that can handle arbitrary burst sizes
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst =
+			iavf_recv_scattered_burst_vec_avx512_flex_rxd
+				(rx_queue, rx_pkts + retval,
+				 IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v4 3/3] net/iavf: enable AVX512 for TX
  2020-09-27  1:30 ` [dpdk-dev] [PATCH v4 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-09-27  1:30   ` [dpdk-dev] [PATCH v4 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
  2020-09-27  1:30   ` [dpdk-dev] [PATCH v4 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
@ 2020-09-27  1:30   ` Wenzhuo Lu
  2 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-09-27  1:30 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the TX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 doc/guides/rel_notes/release_20_11.rst  |   3 +
 drivers/net/iavf/iavf_ethdev.c          |   3 +-
 drivers/net/iavf/iavf_rxtx.c            |  32 +++-
 drivers/net/iavf/iavf_rxtx.h            |   7 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 304 ++++++++++++++++++++++++++++++++
 5 files changed, 341 insertions(+), 8 deletions(-)

diff --git a/doc/guides/rel_notes/release_20_11.rst b/doc/guides/rel_notes/release_20_11.rst
index df227a1..d40b8d6 100644
--- a/doc/guides/rel_notes/release_20_11.rst
+++ b/doc/guides/rel_notes/release_20_11.rst
@@ -55,6 +55,9 @@ New Features
      Also, make sure to start the actual text at the margin.
      =======================================================
 
+   * **Added support of vector instructions on IAVF.**
+
+     Added support of AVX512 instructions in IAVF RX and TX path.
 
 Removed Items
 -------------
diff --git a/drivers/net/iavf/iavf_ethdev.c b/drivers/net/iavf/iavf_ethdev.c
index c3aa4cd..5bc2851 100644
--- a/drivers/net/iavf/iavf_ethdev.c
+++ b/drivers/net/iavf/iavf_ethdev.c
@@ -528,7 +528,8 @@ static int iavf_config_rx_queues_irqs(struct rte_eth_dev *dev,
 		DEV_TX_OFFLOAD_GRE_TNL_TSO |
 		DEV_TX_OFFLOAD_IPIP_TNL_TSO |
 		DEV_TX_OFFLOAD_GENEVE_TNL_TSO |
-		DEV_TX_OFFLOAD_MULTI_SEGS;
+		DEV_TX_OFFLOAD_MULTI_SEGS |
+		DEV_TX_OFFLOAD_MBUF_FAST_FREE;
 
 	dev_info->default_rxconf = (struct rte_eth_rxconf) {
 		.rx_free_thresh = IAVF_DEFAULT_RX_FREE_THRESH,
diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index 0818107..04dcd48 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2206,18 +2206,18 @@
 	struct iavf_tx_queue *txq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_tx_vec_dev_check(dev)) {
-		for (i = 0; i < dev->data->nb_tx_queues; i++) {
-			txq = dev->data->tx_queues[i];
-			if (!txq)
-				continue;
-			iavf_txq_vec_setup(txq);
-		}
-
 		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+			use_avx512 = true;
+#endif
 
 		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
 			    use_avx2 ? "avx2 " : "",
@@ -2225,8 +2225,26 @@
 		dev->tx_pkt_burst = use_avx2 ?
 				    iavf_xmit_pkts_vec_avx2 :
 				    iavf_xmit_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+		if (use_avx512)
+			dev->tx_pkt_burst = iavf_xmit_pkts_vec_avx512;
+#endif
 		dev->tx_pkt_prepare = NULL;
 
+		for (i = 0; i < dev->data->nb_tx_queues; i++) {
+			txq = dev->data->tx_queues[i];
+			if (!txq)
+				continue;
+#ifdef CC_AVX512_SUPPORT
+			if (use_avx512)
+				iavf_txq_vec_setup_avx512(txq);
+			else
+				iavf_txq_vec_setup(txq);
+#else
+			iavf_txq_vec_setup(txq);
+#endif
+		}
+
 		return;
 	}
 #endif
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 9653e0c..08eebb0 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -122,6 +122,10 @@ struct iavf_tx_entry {
 	uint16_t last_id;
 };
 
+struct iavf_tx_vec_entry {
+	struct rte_mbuf *mbuf;
+};
+
 /* Structure associated with each TX queue. */
 struct iavf_tx_queue {
 	const struct rte_memzone *mz;  /* memzone for Tx ring */
@@ -448,6 +452,9 @@ uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
 						      struct rte_mbuf **rx_pkts,
 						      uint16_t nb_pkts);
+uint16_t iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				   uint16_t nb_pkts);
+int iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index 63320e6..0de34f0 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -1396,3 +1396,307 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+static __rte_always_inline int
+iavf_tx_free_bufs_avx512(struct iavf_tx_queue *txq)
+{
+	struct iavf_tx_vec_entry *txep;
+	uint32_t n;
+	uint32_t i;
+	int nb_free = 0;
+	struct rte_mbuf *m, *free[IAVF_VPMD_TX_MAX_FREE_BUF];
+
+	/* check DD bits on threshold descriptor */
+	if ((txq->tx_ring[txq->next_dd].cmd_type_offset_bsz &
+	     rte_cpu_to_le_64(IAVF_TXD_QW1_DTYPE_MASK)) !=
+	    rte_cpu_to_le_64(IAVF_TX_DESC_DTYPE_DESC_DONE))
+		return 0;
+
+	n = txq->rs_thresh;
+
+	 /* first buffer to free from S/W ring is at index
+	  * tx_next_dd - (tx_rs_thresh-1)
+	  */
+	txep = (void *)txq->sw_ring;
+	txep += txq->next_dd - (n - 1);
+
+	if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE && (n & 31) == 0) {
+		struct rte_mempool *mp = txep[0].mbuf->pool;
+		struct rte_mempool_cache *cache = rte_mempool_default_cache(mp,
+								rte_lcore_id());
+		void **cache_objs = &cache->objs[cache->len];
+
+		if (n > RTE_MEMPOOL_CACHE_MAX_SIZE) {
+			rte_mempool_ops_enqueue_bulk(mp, (void *)txep, n);
+			goto done;
+		}
+
+		/* The cache follows the following algorithm
+		 *   1. Add the objects to the cache
+		 *   2. Anything greater than the cache min value (if it crosses the
+		 *   cache flush threshold) is flushed to the ring.
+		 */
+		/* Add elements back into the cache */
+		uint32_t copied = 0;
+		/* n is multiple of 32 */
+		while (copied < n) {
+			const __m512i a = _mm512_loadu_si512(&txep[copied]);
+			const __m512i b = _mm512_loadu_si512(&txep[copied + 8]);
+			const __m512i c = _mm512_loadu_si512(&txep[copied + 16]);
+			const __m512i d = _mm512_loadu_si512(&txep[copied + 24]);
+
+			_mm512_storeu_si512(&cache_objs[copied], a);
+			_mm512_storeu_si512(&cache_objs[copied + 8], b);
+			_mm512_storeu_si512(&cache_objs[copied + 16], c);
+			_mm512_storeu_si512(&cache_objs[copied + 24], d);
+			copied += 32;
+		}
+		cache->len += n;
+
+		if (cache->len >= cache->flushthresh) {
+			rte_mempool_ops_enqueue_bulk(mp,
+						     &cache->objs[cache->size],
+						     cache->len - cache->size);
+			cache->len = cache->size;
+		}
+		goto done;
+	}
+
+	m = rte_pktmbuf_prefree_seg(txep[0].mbuf);
+	if (likely(m)) {
+		free[0] = m;
+		nb_free = 1;
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (likely(m)) {
+				if (likely(m->pool == free[0]->pool)) {
+					free[nb_free++] = m;
+				} else {
+					rte_mempool_put_bulk(free[0]->pool,
+							     (void *)free,
+							     nb_free);
+					free[0] = m;
+					nb_free = 1;
+				}
+			}
+		}
+		rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
+	} else {
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (m)
+				rte_mempool_put(m->pool, m);
+		}
+	}
+
+done:
+	/* buffers were freed, update counters */
+	txq->nb_free = (uint16_t)(txq->nb_free + txq->rs_thresh);
+	txq->next_dd = (uint16_t)(txq->next_dd + txq->rs_thresh);
+	if (txq->next_dd >= txq->nb_tx_desc)
+		txq->next_dd = (uint16_t)(txq->rs_thresh - 1);
+
+	return txq->rs_thresh;
+}
+
+static __rte_always_inline void
+tx_backlog_entry_avx512(struct iavf_tx_vec_entry *txep,
+			struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	int i;
+
+	for (i = 0; i < (int)nb_pkts; ++i)
+		txep[i].mbuf = tx_pkts[i];
+}
+
+static inline void
+iavf_vtx1(volatile struct iavf_tx_desc *txdp,
+	  struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(IAVF_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT) |
+		 ((uint64_t)pkt->data_len << IAVF_TXD_QW1_TX_BUF_SZ_SHIFT));
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+					    pkt->buf_iova + pkt->data_off);
+	_mm_storeu_si128((__m128i *)txdp, descriptor);
+}
+
+#define IAVF_TX_LEN_MASK 0xAA
+#define IAVF_TX_OFF_MASK 0x55
+static inline void
+iavf_vtx(volatile struct iavf_tx_desc *txdp,
+	 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (IAVF_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT));
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		iavf_vtx1(txdp, *pkt, flags);
+		nb_pkts--, txdp++, pkt++;
+	}
+
+	/* do 4 at a time while possible, in bursts */
+	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
+		__m512i desc4 =
+			_mm512_set_epi64
+				((uint64_t)pkt[3]->data_len,
+				 pkt[3]->buf_iova,
+				 (uint64_t)pkt[2]->data_len,
+				 pkt[2]->buf_iova,
+				 (uint64_t)pkt[1]->data_len,
+				 pkt[1]->buf_iova,
+				 (uint64_t)pkt[0]->data_len,
+				 pkt[0]->buf_iova);
+		__m512i hi_qw_tmpl_4 = _mm512_set1_epi64(hi_qw_tmpl);
+		__m512i data_off_4 =
+			_mm512_set_epi64
+				(0,
+				 pkt[3]->data_off,
+				 0,
+				 pkt[2]->data_off,
+				 0,
+				 pkt[1]->data_off,
+				 0,
+				 pkt[0]->data_off);
+
+		desc4 = _mm512_mask_slli_epi64(desc4, IAVF_TX_LEN_MASK, desc4,
+					       IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
+		desc4 = _mm512_mask_or_epi64(desc4, IAVF_TX_LEN_MASK, desc4,
+					     hi_qw_tmpl_4);
+		desc4 = _mm512_mask_add_epi64(desc4, IAVF_TX_OFF_MASK, desc4,
+					      data_off_4);
+		_mm512_storeu_si512((void *)txdp, desc4);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		iavf_vtx1(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+iavf_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				 uint16_t nb_pkts)
+{
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+	volatile struct iavf_tx_desc *txdp;
+	struct iavf_tx_vec_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	/* bit2 is reserved and must be set to 1 according to Spec */
+	uint64_t flags = IAVF_TX_DESC_CMD_EOP | IAVF_TX_DESC_CMD_ICRC;
+	uint64_t rs = IAVF_TX_DESC_CMD_RS | flags;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
+
+	if (txq->nb_free < txq->free_thresh)
+		iavf_tx_free_bufs_avx512(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = (void *)txq->sw_ring;
+	txep += tx_id;
+
+	txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		tx_backlog_entry_avx512(txep, tx_pkts, n);
+
+		iavf_vtx(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		iavf_vtx1(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = (void *)txq->sw_ring;
+		txep += tx_id;
+	}
+
+	tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
+
+	iavf_vtx(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->next_rs) {
+		txq->tx_ring[txq->next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)IAVF_TX_DESC_CMD_RS) <<
+					 IAVF_TXD_QW1_CMD_SHIFT);
+		txq->next_rs =
+			(uint16_t)(txq->next_rs + txq->rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	IAVF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+			  uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->rs_thresh);
+		ret = iavf_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;
+}
+
+static inline void
+iavf_tx_queue_release_mbufs_avx512(struct iavf_tx_queue *txq)
+{
+	unsigned int i;
+	const uint16_t max_desc = (uint16_t)(txq->nb_tx_desc - 1);
+	struct iavf_tx_vec_entry *swr = (void *)txq->sw_ring;
+
+	if (!txq->sw_ring || txq->nb_free == max_desc)
+		return;
+
+	i = txq->next_dd - txq->rs_thresh + 1;
+	if (txq->tx_tail < i) {
+		for (; i < txq->nb_tx_desc; i++) {
+			rte_pktmbuf_free_seg(swr[i].mbuf);
+			swr[i].mbuf = NULL;
+		}
+		i = 0;
+	}
+}
+
+static const struct iavf_txq_ops avx512_vec_txq_ops = {
+	.release_mbufs = iavf_tx_queue_release_mbufs_avx512,
+};
+
+int __rte_cold
+iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq)
+{
+	txq->ops = &avx512_vec_txq_ops;
+	return 0;
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v5 0/3] enable AVX512 for iavf
  2020-09-10  5:59 [dpdk-dev] [PATCH 0/3] enable AVX512 for iavf Wenzhuo Lu
                   ` (5 preceding siblings ...)
  2020-09-27  1:30 ` [dpdk-dev] [PATCH v4 0/3] enable AVX512 for iavf Wenzhuo Lu
@ 2020-10-21  7:47 ` Wenzhuo Lu
  2020-10-21  7:47   ` [dpdk-dev] [PATCH v5 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
                     ` (2 more replies)
  2020-10-28  5:14 ` [dpdk-dev] [PATCH v6 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-10-29  1:24 ` [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf Wenzhuo Lu
  8 siblings, 3 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-10-21  7:47 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu

AVX512 instructions is supported by more and more platforms. These instructions can be used in the data path to enhance the per-core performance of packet processing.
Comparing with the existing implementation, this path set introduces some AVX512 instructions into the iavf data path, and we get a better per-code throughput.

v2:
Update meson.build.
Repalce the deprecated 'buf_physaddr' by 'buf_iova'.

v3:
Fix compile errors.

v4:
Fix wrong info in commnets.
Trivial adjustment of the arrangement.

v5:
Support "max SIMD bitwidth".

Wenzhuo Lu (3):
  net/iavf: enable AVX512 for legacy RX
  net/iavf: enable AVX512 for flexible RX
  net/iavf: enable AVX512 for TX

 doc/guides/rel_notes/release_20_11.rst  |    3 +
 drivers/net/iavf/iavf_ethdev.c          |    3 +-
 drivers/net/iavf/iavf_rxtx.c            |   71 +-
 drivers/net/iavf/iavf_rxtx.h            |   18 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 1702 +++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |   17 +
 6 files changed, 1802 insertions(+), 12 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v5 1/3] net/iavf: enable AVX512 for legacy RX
  2020-10-21  7:47 ` [dpdk-dev] [PATCH v5 0/3] enable AVX512 for iavf Wenzhuo Lu
@ 2020-10-21  7:47   ` Wenzhuo Lu
  2020-10-21  7:47   ` [dpdk-dev] [PATCH v5 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
  2020-10-21  7:47   ` [dpdk-dev] [PATCH v5 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
  2 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-10-21  7:47 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the legacy RX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  28 +-
 drivers/net/iavf/iavf_rxtx.h            |   5 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 691 ++++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |  17 +
 4 files changed, 737 insertions(+), 4 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index edb2dc3..0067b64 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2104,6 +2104,9 @@
 	struct iavf_rx_queue *rxq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_rx_vec_dev_check(dev) &&
 			rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
@@ -2116,6 +2119,11 @@
 		     rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&
 				rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1 &&
+		    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512)
+			use_avx512 = true;
+#endif
 
 		if (dev->data->scattered_rx) {
 			PMD_DRV_LOG(DEBUG,
@@ -2123,27 +2131,39 @@
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
 					iavf_recv_scattered_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512;
+#endif
+			}
 		} else {
 			PMD_DRV_LOG(DEBUG, "Using %sVector Rx (port %d).",
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
 					iavf_recv_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512;
+#endif
+			}
 		}
 
 		return;
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 3d02c65..7c1f05f 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -438,6 +438,11 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_tx_vec_dev_check(struct rte_eth_dev *dev);
 int iavf_rxq_vec_setup(struct iavf_rx_queue *rxq);
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
+uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
new file mode 100644
index 0000000..959067c
--- /dev/null
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -0,0 +1,691 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include "iavf_rxtx_vec_common.h"
+
+#include <x86intrin.h>
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+#define IAVF_DESCS_PER_LOOP_AVX 8
+#define PKTLEN_SHIFT 10
+
+static inline void
+iavf_rxq_rearm(struct iavf_rx_queue *rxq)
+{
+	int i;
+	uint16_t rx_id;
+	volatile union iavf_rx_desc *rxdp;
+	struct rte_mempool_cache *cache =
+		rte_mempool_default_cache(rxq->mp, rte_lcore_id());
+	struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start];
+
+	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.
+	 */
+
+	/* Can this be satisfied from the cache? */
+	if (cache->len < IAVF_RXQ_REARM_THRESH) {
+		/* No. Backfill the cache first, and then fill from it */
+		uint32_t req = IAVF_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 + IAVF_RXQ_REARM_THRESH >=
+			    rxq->nb_rx_desc) {
+				__m128i dma_addr0;
+
+				dma_addr0 = _mm_setzero_si128();
+				for (i = 0; i < IAVF_VPMD_DESCS_PER_LOOP; i++) {
+					rxp[i] = &rxq->fake_mbuf;
+					_mm_storeu_si128((__m128i *)&rxdp[i].read,
+							 dma_addr0);
+				}
+			}
+			rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+					IAVF_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_IAVF_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 < IAVF_RXQ_REARM_THRESH / IAVF_DESCS_PER_LOOP_AVX; i++) {
+		const __m512i mbuf_ptrs = _mm512_loadu_si512
+			(&cache->objs[cache->len - IAVF_DESCS_PER_LOOP_AVX]);
+		_mm512_storeu_si512(rxp, mbuf_ptrs);
+
+		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_IAVF_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 6 & 7.
+		 */
+		const __m512i desc0_1 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas0);
+		const __m512i desc2_3 = _mm512_bsrli_epi128(desc0_1, 8);
+
+		const __m512i desc4_5 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas1);
+		const __m512i desc6_7 = _mm512_bsrli_epi128(desc4_5, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_1);
+		_mm512_storeu_si512((void *)(rxdp + 2), desc2_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_5);
+		_mm512_storeu_si512((void *)(rxdp + 6), desc6_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 desc0_3 = _mm512_permutexvar_epi64(permute_idx,
+								 iova_addrs);
+		const __m512i desc4_7 = _mm512_bsrli_epi128(desc0_3, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_7);
+#endif
+		rxp += IAVF_DESCS_PER_LOOP_AVX;
+		rxdp += IAVF_DESCS_PER_LOOP_AVX;
+		cache->len -= IAVF_DESCS_PER_LOOP_AVX;
+	}
+
+	rxq->rxrearm_start += IAVF_RXQ_REARM_THRESH;
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= IAVF_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 */
+	IAVF_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
+}
+
+#define IAVF_RX_LEN_MASK 0x80808080
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512(struct iavf_rx_queue *rxq,
+			       struct rte_mbuf **rx_pkts,
+			       uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0, 0,
+						    rxq->mbuf_initializer);
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_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 << IAVF_RX_DESC_STATUS_DD_SHIFT)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_DESC_STATUS_EOF_SHIFT);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 4th descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF     /* pkt_type set as unknown */
+			);
+	/**
+	 * 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 += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 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 desc4_7 = _mm512_mask_blend_epi16(IAVF_RX_LEN_MASK,
+								raw_desc4_7,
+								len4_7);
+		__m512i mb4_7 = _mm512_shuffle_epi8(desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, 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 __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 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 __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, adjusting length and
+		 * re-arranging fields. Then write into the mbuf
+		 */
+		const __m512i len0_3 = _mm512_slli_epi32(raw_desc0_3,
+							 PKTLEN_SHIFT);
+		const __m512i desc0_3 = _mm512_mask_blend_epi16(IAVF_RX_LEN_MASK,
+								raw_desc0_3,
+								len0_3);
+		__m512i mb0_3 = _mm512_shuffle_epi8(desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/* get the packet types */
+		const __m512i ptypes0_3 = _mm512_srli_epi64(desc0_3, 30);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		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 ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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_flags =
+			_mm256_shuffle_epi8(rss_flags_shuf,
+					    _mm256_srli_epi32(flag_bits, 11));
+		/**
+		 * 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 */
+		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				_mm256_or_si256(rss_flags, vlan_flags));
+		/**
+		 * 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;
+		const __m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7, 0);
+		const __m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7, 1);
+		const __m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3, 0);
+		const __m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3, 1);
+
+		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 << IAVF_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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
+/**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+			  uint16_t nb_pkts)
+{
+	return _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				     uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				    uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst = iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
diff --git a/drivers/net/iavf/meson.build b/drivers/net/iavf/meson.build
index 33407c5..fd33b72 100644
--- a/drivers/net/iavf/meson.build
+++ b/drivers/net/iavf/meson.build
@@ -34,4 +34,21 @@ if arch_subdir == 'x86'
 				c_args: [cflags, '-mavx2'])
 		objs += iavf_avx2_lib.extract_objects('iavf_rxtx_vec_avx2.c')
 	endif
+
+	if not machine_args.contains('-mno-avx512f')
+		if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F')
+			cflags += ['-DCC_AVX512_SUPPORT']
+			cflags += ['-march=skylake-avx512']
+			sources += files('iavf_rxtx_vec_avx512.c')
+		elif cc.has_argument('-mavx512f')
+			cflags += ['-DCC_AVX512_SUPPORT']
+			iavf_avx512_lib = static_library('iavf_avx512_lib',
+					'iavf_rxtx_vec_avx512.c',
+					dependencies: [static_rte_ethdev,
+						static_rte_kvargs, static_rte_hash],
+					include_directories: includes,
+					c_args: [cflags, '-mavx512f', '-march=skylake-avx512'])
+			objs += iavf_avx512_lib.extract_objects('iavf_rxtx_vec_avx512.c')
+		endif
+	endif
 endif
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v5 2/3] net/iavf: enable AVX512 for flexible RX
  2020-10-21  7:47 ` [dpdk-dev] [PATCH v5 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-10-21  7:47   ` [dpdk-dev] [PATCH v5 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
@ 2020-10-21  7:47   ` Wenzhuo Lu
  2020-10-21  7:47   ` [dpdk-dev] [PATCH v5 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
  2 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-10-21  7:47 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the flexible RX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  10 +
 drivers/net/iavf/iavf_rxtx.h            |   6 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 707 ++++++++++++++++++++++++++++++++
 3 files changed, 723 insertions(+)

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index 0067b64..fbcddd3 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2135,6 +2135,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
@@ -2154,6 +2159,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 7c1f05f..03b095d 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -440,9 +440,15 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
 uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
 				   uint16_t nb_pkts);
+uint16_t iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					    struct rte_mbuf **rx_pkts,
+					    uint16_t nb_pkts);
 uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 					     struct rte_mbuf **rx_pkts,
 					     uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+						      struct rte_mbuf **rx_pkts,
+						      uint16_t nb_pkts);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index 959067c..df0f43b 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -614,6 +614,631 @@
 	return received;
 }
 
+static inline __m256i
+flex_rxd_to_fdir_flags_vec_avx512(const __m256i fdir_id0_7)
+{
+#define FDID_MIS_MAGIC 0xFFFFFFFF
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR != (1 << 2));
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
+	const __m256i pkt_fdir_bit = _mm256_set1_epi32(PKT_RX_FDIR |
+						       PKT_RX_FDIR_ID);
+	/* desc->flow_id field == 0xFFFFFFFF means fdir mismatch */
+	const __m256i fdir_mis_mask = _mm256_set1_epi32(FDID_MIS_MAGIC);
+	__m256i fdir_mask = _mm256_cmpeq_epi32(fdir_id0_7,
+					       fdir_mis_mask);
+	/* this XOR op results to bit-reverse the fdir_mask */
+	fdir_mask = _mm256_xor_si256(fdir_mask, fdir_mis_mask);
+	const __m256i fdir_flags = _mm256_and_si256(fdir_mask, pkt_fdir_bit);
+
+	return fdir_flags;
+}
+
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512_flex_rxd(struct iavf_rx_queue *rxq,
+					struct rte_mbuf **rx_pkts,
+					uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0, 0,
+						    rxq->mbuf_initializer);
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_flex_desc *rxdp =
+		(union iavf_rx_flex_desc *)rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_rxq_rearm(rxq);
+
+	/* Before we start moving massive data around, check to see if
+	 * there is actually a packet available
+	 */
+	if (!(rxdp->wb.status_error0 &
+	      rte_cpu_to_le_32(1 << IAVF_RX_FLEX_DESC_STATUS0_DD_S)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 4th descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF     /* pkt_type set as unknown */
+			);
+	/**
+	 * 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 Checksum Reports, RSS indication
+	 * and VLAN indication.
+	 * bit6:4 for IP/L4 checksum errors.
+	 * bit12 is for RSS indication.
+	 * bit13 is for VLAN indication.
+	 */
+	const __m256i flags_mask =
+		_mm256_set1_epi32((7 << 4) | (1 << 12) | (1 << 13));
+	/**
+	 * data to be shuffled by the result of the flags mask shifted by 4
+	 * 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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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);
+	/**
+	 * data to be shuffled by result of flag mask, shifted down 12.
+	 * If RSS(bit12)/VLAN(bit13) are set,
+	 * shuffle moves appropriate flags in place.
+	 */
+	const __m256i rss_vlan_flags_shuf = _mm256_set_epi8(0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0,
+			/* end up 128-bits */
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0);
+
+	uint16_t i, received;
+
+	for (i = 0, received = 0; i < nb_pkts;
+	     i += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb4_7 = _mm512_shuffle_epi8(raw_desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptype_mask =
+			_mm512_set1_epi16(IAVF_RX_FLEX_DESC_PTYPE_M);
+		const __m512i ptypes4_7 =
+			_mm512_and_si512(raw_desc4_7, ptype_mask);
+		const __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 0);
+		const uint16_t ptype7 = _mm256_extract_epi16(ptypes6_7, 9);
+		const uint16_t ptype6 = _mm256_extract_epi16(ptypes6_7, 1);
+		const uint16_t ptype5 = _mm256_extract_epi16(ptypes4_5, 9);
+		const uint16_t ptype4 = _mm256_extract_epi16(ptypes4_5, 1);
+
+		const __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb0_3 = _mm512_shuffle_epi8(raw_desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptypes0_3 =
+			_mm512_and_si512(raw_desc0_3, ptype_mask);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		const uint16_t ptype3 = _mm256_extract_epi16(ptypes2_3, 9);
+		const uint16_t ptype2 = _mm256_extract_epi16(ptypes2_3, 1);
+		const uint16_t ptype1 = _mm256_extract_epi16(ptypes0_1, 9);
+		const uint16_t ptype0 = _mm256_extract_epi16(ptypes0_1, 1);
+
+		const __m512i ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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);
+		/**
+		 * 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, 4));
+		l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
+		l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
+		/* set rss and vlan flags */
+		const __m256i rss_vlan_flag_bits =
+			_mm256_srli_epi32(flag_bits, 12);
+		const __m256i rss_vlan_flags =
+			_mm256_shuffle_epi8(rss_vlan_flags_shuf,
+					    rss_vlan_flag_bits);
+
+		/* merge flags */
+		__m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+						     rss_vlan_flags);
+
+		if (rxq->fdir_enabled) {
+			const __m512i fdir_permute_mask = _mm512_set_epi32
+				(0, 0, 0, 0,
+				 0, 0, 0, 0,
+				 7, 15, 23, 31,
+				 3, 11, 19, 27);
+			__m512i fdir_tmp = _mm512_permutex2var_epi32
+				(raw_desc0_3, fdir_permute_mask, raw_desc4_7);
+			const __m256i fdir_id0_7 = _mm512_extracti64x4_epi64
+				(fdir_tmp, 0);
+			const __m256i fdir_flags =
+				flex_rxd_to_fdir_flags_vec_avx512(fdir_id0_7);
+
+			/* merge with fdir_flags */
+			mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_flags);
+
+			/* write to mbuf: have to use scalar store here */
+			rx_pkts[i + 0]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 3);
+
+			rx_pkts[i + 1]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 7);
+
+			rx_pkts[i + 2]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 2);
+
+			rx_pkts[i + 3]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 6);
+
+			rx_pkts[i + 4]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 1);
+
+			rx_pkts[i + 5]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 5);
+
+			rx_pkts[i + 6]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 0);
+
+			rx_pkts[i + 7]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 4);
+		} /* if() on fdir_enabled */
+
+		__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);
+
+#ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
+		/**
+		 * needs to load 2nd 16B of each desc for RSS hash parsing,
+		 * will cause performance drop to get into this context.
+		 */
+		if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
+		    DEV_RX_OFFLOAD_RSS_HASH) {
+			/* load bottom half of every 32B desc */
+			const __m128i raw_desc_bh7 =
+				_mm_load_si128
+					((void *)(&rxdp[7].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh6 =
+				_mm_load_si128
+					((void *)(&rxdp[6].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh5 =
+				_mm_load_si128
+					((void *)(&rxdp[5].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh4 =
+				_mm_load_si128
+					((void *)(&rxdp[4].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh3 =
+				_mm_load_si128
+					((void *)(&rxdp[3].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh2 =
+				_mm_load_si128
+					((void *)(&rxdp[2].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh1 =
+				_mm_load_si128
+					((void *)(&rxdp[1].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh0 =
+				_mm_load_si128
+					((void *)(&rxdp[0].wb.status_error1));
+
+			__m256i raw_desc_bh6_7 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh6),
+					 raw_desc_bh7, 1);
+			__m256i raw_desc_bh4_5 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh4),
+					 raw_desc_bh5, 1);
+			__m256i raw_desc_bh2_3 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh2),
+					 raw_desc_bh3, 1);
+			__m256i raw_desc_bh0_1 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh0),
+					 raw_desc_bh1, 1);
+
+			/**
+			 * to shift the 32b RSS hash value to the
+			 * highest 32b of each 128b before mask
+			 */
+			__m256i rss_hash6_7 =
+				_mm256_slli_epi64(raw_desc_bh6_7, 32);
+			__m256i rss_hash4_5 =
+				_mm256_slli_epi64(raw_desc_bh4_5, 32);
+			__m256i rss_hash2_3 =
+				_mm256_slli_epi64(raw_desc_bh2_3, 32);
+			__m256i rss_hash0_1 =
+				_mm256_slli_epi64(raw_desc_bh0_1, 32);
+
+			__m256i rss_hash_msk =
+				_mm256_set_epi32(0xFFFFFFFF, 0, 0, 0,
+						 0xFFFFFFFF, 0, 0, 0);
+
+			rss_hash6_7 = _mm256_and_si256
+					(rss_hash6_7, rss_hash_msk);
+			rss_hash4_5 = _mm256_and_si256
+					(rss_hash4_5, rss_hash_msk);
+			rss_hash2_3 = _mm256_and_si256
+					(rss_hash2_3, rss_hash_msk);
+			rss_hash0_1 = _mm256_and_si256
+					(rss_hash0_1, rss_hash_msk);
+
+			mb6_7 = _mm256_or_si256(mb6_7, rss_hash6_7);
+			mb4_5 = _mm256_or_si256(mb4_5, rss_hash4_5);
+			mb2_3 = _mm256_or_si256(mb2_3, rss_hash2_3);
+			mb0_1 = _mm256_or_si256(mb0_1, rss_hash0_1);
+		} /* if() on RSS hash parsing */
+#endif
+
+		/**
+		 * 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 <<
+					       IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+			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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
 /**
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -626,6 +1251,18 @@
 }
 
 /**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts)
+{
+	return _iavf_recv_raw_pkts_vec_avx512_flex_rxd(rx_queue, rx_pkts,
+						       nb_pkts, NULL);
+}
+
+/**
  * vPMD receive routine that reassembles single burst of 32 scattered packets
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -689,3 +1326,73 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+/**
+ * vPMD receive routine that reassembles single burst of
+ * 32 scattered packets for flex RxD
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512_flex_rxd(void *rx_queue,
+					      struct rte_mbuf **rx_pkts,
+					      uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_recv_raw_pkts_vec_avx512_flex_rxd(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 for flex RxD.
+ * Main receive routine that can handle arbitrary burst sizes
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst =
+			iavf_recv_scattered_burst_vec_avx512_flex_rxd
+				(rx_queue, rx_pkts + retval,
+				 IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v5 3/3] net/iavf: enable AVX512 for TX
  2020-10-21  7:47 ` [dpdk-dev] [PATCH v5 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-10-21  7:47   ` [dpdk-dev] [PATCH v5 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
  2020-10-21  7:47   ` [dpdk-dev] [PATCH v5 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
@ 2020-10-21  7:47   ` Wenzhuo Lu
  2 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-10-21  7:47 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the TX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 doc/guides/rel_notes/release_20_11.rst  |   3 +
 drivers/net/iavf/iavf_ethdev.c          |   3 +-
 drivers/net/iavf/iavf_rxtx.c            |  33 +++-
 drivers/net/iavf/iavf_rxtx.h            |   7 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 304 ++++++++++++++++++++++++++++++++
 5 files changed, 342 insertions(+), 8 deletions(-)

diff --git a/doc/guides/rel_notes/release_20_11.rst b/doc/guides/rel_notes/release_20_11.rst
index 0d45b50..c981b64 100644
--- a/doc/guides/rel_notes/release_20_11.rst
+++ b/doc/guides/rel_notes/release_20_11.rst
@@ -20,6 +20,9 @@ DPDK Release 20.11
       make doc-guides-html
       xdg-open build/doc/html/guides/rel_notes/release_20_11.html
 
+   * **Added support of vector instructions on IAVF.**
+
+     Added support of AVX512 instructions in IAVF RX and TX path.
 
 New Features
 ------------
diff --git a/drivers/net/iavf/iavf_ethdev.c b/drivers/net/iavf/iavf_ethdev.c
index 0ef023c..fe6c8cb 100644
--- a/drivers/net/iavf/iavf_ethdev.c
+++ b/drivers/net/iavf/iavf_ethdev.c
@@ -606,7 +606,8 @@ static int iavf_config_rx_queues_irqs(struct rte_eth_dev *dev,
 		DEV_TX_OFFLOAD_GRE_TNL_TSO |
 		DEV_TX_OFFLOAD_IPIP_TNL_TSO |
 		DEV_TX_OFFLOAD_GENEVE_TNL_TSO |
-		DEV_TX_OFFLOAD_MULTI_SEGS;
+		DEV_TX_OFFLOAD_MULTI_SEGS |
+		DEV_TX_OFFLOAD_MBUF_FAST_FREE;
 
 	dev_info->default_rxconf = (struct rte_eth_rxconf) {
 		.rx_free_thresh = IAVF_DEFAULT_RX_FREE_THRESH,
diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index fbcddd3..a94f646 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2209,20 +2209,21 @@
 	struct iavf_tx_queue *txq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_tx_vec_dev_check(dev) &&
 			rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
-		for (i = 0; i < dev->data->nb_tx_queues; i++) {
-			txq = dev->data->tx_queues[i];
-			if (!txq)
-				continue;
-			iavf_txq_vec_setup(txq);
-		}
-
 		if ((rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 		     rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&
 				rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1 &&
+		    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512)
+			use_avx512 = true;
+#endif
 
 		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
 			    use_avx2 ? "avx2 " : "",
@@ -2230,8 +2231,26 @@
 		dev->tx_pkt_burst = use_avx2 ?
 				    iavf_xmit_pkts_vec_avx2 :
 				    iavf_xmit_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+		if (use_avx512)
+			dev->tx_pkt_burst = iavf_xmit_pkts_vec_avx512;
+#endif
 		dev->tx_pkt_prepare = NULL;
 
+		for (i = 0; i < dev->data->nb_tx_queues; i++) {
+			txq = dev->data->tx_queues[i];
+			if (!txq)
+				continue;
+#ifdef CC_AVX512_SUPPORT
+			if (use_avx512)
+				iavf_txq_vec_setup_avx512(txq);
+			else
+				iavf_txq_vec_setup(txq);
+#else
+			iavf_txq_vec_setup(txq);
+#endif
+		}
+
 		return;
 	}
 #endif
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 03b095d..b22ccc4 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -122,6 +122,10 @@ struct iavf_tx_entry {
 	uint16_t last_id;
 };
 
+struct iavf_tx_vec_entry {
+	struct rte_mbuf *mbuf;
+};
+
 /* Structure associated with each TX queue. */
 struct iavf_tx_queue {
 	const struct rte_memzone *mz;  /* memzone for Tx ring */
@@ -449,6 +453,9 @@ uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
 						      struct rte_mbuf **rx_pkts,
 						      uint16_t nb_pkts);
+uint16_t iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				   uint16_t nb_pkts);
+int iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index df0f43b..8680734 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -1396,3 +1396,307 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+static __rte_always_inline int
+iavf_tx_free_bufs_avx512(struct iavf_tx_queue *txq)
+{
+	struct iavf_tx_vec_entry *txep;
+	uint32_t n;
+	uint32_t i;
+	int nb_free = 0;
+	struct rte_mbuf *m, *free[IAVF_VPMD_TX_MAX_FREE_BUF];
+
+	/* check DD bits on threshold descriptor */
+	if ((txq->tx_ring[txq->next_dd].cmd_type_offset_bsz &
+	     rte_cpu_to_le_64(IAVF_TXD_QW1_DTYPE_MASK)) !=
+	    rte_cpu_to_le_64(IAVF_TX_DESC_DTYPE_DESC_DONE))
+		return 0;
+
+	n = txq->rs_thresh;
+
+	 /* first buffer to free from S/W ring is at index
+	  * tx_next_dd - (tx_rs_thresh-1)
+	  */
+	txep = (void *)txq->sw_ring;
+	txep += txq->next_dd - (n - 1);
+
+	if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE && (n & 31) == 0) {
+		struct rte_mempool *mp = txep[0].mbuf->pool;
+		struct rte_mempool_cache *cache = rte_mempool_default_cache(mp,
+								rte_lcore_id());
+		void **cache_objs = &cache->objs[cache->len];
+
+		if (n > RTE_MEMPOOL_CACHE_MAX_SIZE) {
+			rte_mempool_ops_enqueue_bulk(mp, (void *)txep, n);
+			goto done;
+		}
+
+		/* The cache follows the following algorithm
+		 *   1. Add the objects to the cache
+		 *   2. Anything greater than the cache min value (if it crosses the
+		 *   cache flush threshold) is flushed to the ring.
+		 */
+		/* Add elements back into the cache */
+		uint32_t copied = 0;
+		/* n is multiple of 32 */
+		while (copied < n) {
+			const __m512i a = _mm512_loadu_si512(&txep[copied]);
+			const __m512i b = _mm512_loadu_si512(&txep[copied + 8]);
+			const __m512i c = _mm512_loadu_si512(&txep[copied + 16]);
+			const __m512i d = _mm512_loadu_si512(&txep[copied + 24]);
+
+			_mm512_storeu_si512(&cache_objs[copied], a);
+			_mm512_storeu_si512(&cache_objs[copied + 8], b);
+			_mm512_storeu_si512(&cache_objs[copied + 16], c);
+			_mm512_storeu_si512(&cache_objs[copied + 24], d);
+			copied += 32;
+		}
+		cache->len += n;
+
+		if (cache->len >= cache->flushthresh) {
+			rte_mempool_ops_enqueue_bulk(mp,
+						     &cache->objs[cache->size],
+						     cache->len - cache->size);
+			cache->len = cache->size;
+		}
+		goto done;
+	}
+
+	m = rte_pktmbuf_prefree_seg(txep[0].mbuf);
+	if (likely(m)) {
+		free[0] = m;
+		nb_free = 1;
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (likely(m)) {
+				if (likely(m->pool == free[0]->pool)) {
+					free[nb_free++] = m;
+				} else {
+					rte_mempool_put_bulk(free[0]->pool,
+							     (void *)free,
+							     nb_free);
+					free[0] = m;
+					nb_free = 1;
+				}
+			}
+		}
+		rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
+	} else {
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (m)
+				rte_mempool_put(m->pool, m);
+		}
+	}
+
+done:
+	/* buffers were freed, update counters */
+	txq->nb_free = (uint16_t)(txq->nb_free + txq->rs_thresh);
+	txq->next_dd = (uint16_t)(txq->next_dd + txq->rs_thresh);
+	if (txq->next_dd >= txq->nb_tx_desc)
+		txq->next_dd = (uint16_t)(txq->rs_thresh - 1);
+
+	return txq->rs_thresh;
+}
+
+static __rte_always_inline void
+tx_backlog_entry_avx512(struct iavf_tx_vec_entry *txep,
+			struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	int i;
+
+	for (i = 0; i < (int)nb_pkts; ++i)
+		txep[i].mbuf = tx_pkts[i];
+}
+
+static inline void
+iavf_vtx1(volatile struct iavf_tx_desc *txdp,
+	  struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(IAVF_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT) |
+		 ((uint64_t)pkt->data_len << IAVF_TXD_QW1_TX_BUF_SZ_SHIFT));
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+					    pkt->buf_iova + pkt->data_off);
+	_mm_storeu_si128((__m128i *)txdp, descriptor);
+}
+
+#define IAVF_TX_LEN_MASK 0xAA
+#define IAVF_TX_OFF_MASK 0x55
+static inline void
+iavf_vtx(volatile struct iavf_tx_desc *txdp,
+	 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (IAVF_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT));
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		iavf_vtx1(txdp, *pkt, flags);
+		nb_pkts--, txdp++, pkt++;
+	}
+
+	/* do 4 at a time while possible, in bursts */
+	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
+		__m512i desc4 =
+			_mm512_set_epi64
+				((uint64_t)pkt[3]->data_len,
+				 pkt[3]->buf_iova,
+				 (uint64_t)pkt[2]->data_len,
+				 pkt[2]->buf_iova,
+				 (uint64_t)pkt[1]->data_len,
+				 pkt[1]->buf_iova,
+				 (uint64_t)pkt[0]->data_len,
+				 pkt[0]->buf_iova);
+		__m512i hi_qw_tmpl_4 = _mm512_set1_epi64(hi_qw_tmpl);
+		__m512i data_off_4 =
+			_mm512_set_epi64
+				(0,
+				 pkt[3]->data_off,
+				 0,
+				 pkt[2]->data_off,
+				 0,
+				 pkt[1]->data_off,
+				 0,
+				 pkt[0]->data_off);
+
+		desc4 = _mm512_mask_slli_epi64(desc4, IAVF_TX_LEN_MASK, desc4,
+					       IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
+		desc4 = _mm512_mask_or_epi64(desc4, IAVF_TX_LEN_MASK, desc4,
+					     hi_qw_tmpl_4);
+		desc4 = _mm512_mask_add_epi64(desc4, IAVF_TX_OFF_MASK, desc4,
+					      data_off_4);
+		_mm512_storeu_si512((void *)txdp, desc4);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		iavf_vtx1(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+iavf_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				 uint16_t nb_pkts)
+{
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+	volatile struct iavf_tx_desc *txdp;
+	struct iavf_tx_vec_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	/* bit2 is reserved and must be set to 1 according to Spec */
+	uint64_t flags = IAVF_TX_DESC_CMD_EOP | IAVF_TX_DESC_CMD_ICRC;
+	uint64_t rs = IAVF_TX_DESC_CMD_RS | flags;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
+
+	if (txq->nb_free < txq->free_thresh)
+		iavf_tx_free_bufs_avx512(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = (void *)txq->sw_ring;
+	txep += tx_id;
+
+	txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		tx_backlog_entry_avx512(txep, tx_pkts, n);
+
+		iavf_vtx(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		iavf_vtx1(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = (void *)txq->sw_ring;
+		txep += tx_id;
+	}
+
+	tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
+
+	iavf_vtx(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->next_rs) {
+		txq->tx_ring[txq->next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)IAVF_TX_DESC_CMD_RS) <<
+					 IAVF_TXD_QW1_CMD_SHIFT);
+		txq->next_rs =
+			(uint16_t)(txq->next_rs + txq->rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	IAVF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+			  uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->rs_thresh);
+		ret = iavf_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;
+}
+
+static inline void
+iavf_tx_queue_release_mbufs_avx512(struct iavf_tx_queue *txq)
+{
+	unsigned int i;
+	const uint16_t max_desc = (uint16_t)(txq->nb_tx_desc - 1);
+	struct iavf_tx_vec_entry *swr = (void *)txq->sw_ring;
+
+	if (!txq->sw_ring || txq->nb_free == max_desc)
+		return;
+
+	i = txq->next_dd - txq->rs_thresh + 1;
+	if (txq->tx_tail < i) {
+		for (; i < txq->nb_tx_desc; i++) {
+			rte_pktmbuf_free_seg(swr[i].mbuf);
+			swr[i].mbuf = NULL;
+		}
+		i = 0;
+	}
+}
+
+static const struct iavf_txq_ops avx512_vec_txq_ops = {
+	.release_mbufs = iavf_tx_queue_release_mbufs_avx512,
+};
+
+int __rte_cold
+iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq)
+{
+	txq->ops = &avx512_vec_txq_ops;
+	return 0;
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v6 0/3] enable AVX512 for iavf
  2020-09-10  5:59 [dpdk-dev] [PATCH 0/3] enable AVX512 for iavf Wenzhuo Lu
                   ` (6 preceding siblings ...)
  2020-10-21  7:47 ` [dpdk-dev] [PATCH v5 0/3] enable AVX512 for iavf Wenzhuo Lu
@ 2020-10-28  5:14 ` Wenzhuo Lu
  2020-10-28  5:14   ` [dpdk-dev] [PATCH v6 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
                     ` (2 more replies)
  2020-10-29  1:24 ` [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf Wenzhuo Lu
  8 siblings, 3 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-10-28  5:14 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu

AVX512 instructions is supported by more and more platforms. These instructions can be used in the data path to enhance the per-core performance of packet processing.
Comparing with the existing implementation, this path set introduces some AVX512 instructions into the iavf data path, and we get a better per-code throughput.

v2:
Update meson.build.
Repalce the deprecated 'buf_physaddr' by 'buf_iova'.

v3:
Fix compile errors.

v4:
Fix wrong info in commnets.
Trivial adjustment of the arrangement.

v5:
Support "max SIMD bitwidth".

v6:
Rework meson build to fix compile issue for AVX512BW.

Wenzhuo Lu (3):
  net/iavf: enable AVX512 for legacy RX
  net/iavf: enable AVX512 for flexible RX
  net/iavf: enable AVX512 for TX

 doc/guides/rel_notes/release_20_11.rst  |    3 +
 drivers/net/iavf/iavf_ethdev.c          |    3 +-
 drivers/net/iavf/iavf_rxtx.c            |   73 +-
 drivers/net/iavf/iavf_rxtx.h            |   18 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 1702 +++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |   20 +
 6 files changed, 1807 insertions(+), 12 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v6 1/3] net/iavf: enable AVX512 for legacy RX
  2020-10-28  5:14 ` [dpdk-dev] [PATCH v6 0/3] enable AVX512 for iavf Wenzhuo Lu
@ 2020-10-28  5:14   ` Wenzhuo Lu
  2020-10-28  5:14   ` [dpdk-dev] [PATCH v6 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
  2020-10-28  5:15   ` [dpdk-dev] [PATCH v6 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
  2 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-10-28  5:14 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the legacy RX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  29 +-
 drivers/net/iavf/iavf_rxtx.h            |   5 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 691 ++++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |  20 +
 4 files changed, 741 insertions(+), 4 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index edb2dc3..6eedb12 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2104,6 +2104,9 @@
 	struct iavf_rx_queue *rxq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_rx_vec_dev_check(dev) &&
 			rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
@@ -2116,6 +2119,12 @@
 		     rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&
 				rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1 &&
+		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512BW) == 1 &&
+		    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512)
+			use_avx512 = true;
+#endif
 
 		if (dev->data->scattered_rx) {
 			PMD_DRV_LOG(DEBUG,
@@ -2123,27 +2132,39 @@
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
 					iavf_recv_scattered_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512;
+#endif
+			}
 		} else {
 			PMD_DRV_LOG(DEBUG, "Using %sVector Rx (port %d).",
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
 					iavf_recv_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512;
+#endif
+			}
 		}
 
 		return;
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 3d02c65..7c1f05f 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -438,6 +438,11 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_tx_vec_dev_check(struct rte_eth_dev *dev);
 int iavf_rxq_vec_setup(struct iavf_rx_queue *rxq);
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
+uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
new file mode 100644
index 0000000..959067c
--- /dev/null
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -0,0 +1,691 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include "iavf_rxtx_vec_common.h"
+
+#include <x86intrin.h>
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+#define IAVF_DESCS_PER_LOOP_AVX 8
+#define PKTLEN_SHIFT 10
+
+static inline void
+iavf_rxq_rearm(struct iavf_rx_queue *rxq)
+{
+	int i;
+	uint16_t rx_id;
+	volatile union iavf_rx_desc *rxdp;
+	struct rte_mempool_cache *cache =
+		rte_mempool_default_cache(rxq->mp, rte_lcore_id());
+	struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start];
+
+	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.
+	 */
+
+	/* Can this be satisfied from the cache? */
+	if (cache->len < IAVF_RXQ_REARM_THRESH) {
+		/* No. Backfill the cache first, and then fill from it */
+		uint32_t req = IAVF_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 + IAVF_RXQ_REARM_THRESH >=
+			    rxq->nb_rx_desc) {
+				__m128i dma_addr0;
+
+				dma_addr0 = _mm_setzero_si128();
+				for (i = 0; i < IAVF_VPMD_DESCS_PER_LOOP; i++) {
+					rxp[i] = &rxq->fake_mbuf;
+					_mm_storeu_si128((__m128i *)&rxdp[i].read,
+							 dma_addr0);
+				}
+			}
+			rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+					IAVF_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_IAVF_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 < IAVF_RXQ_REARM_THRESH / IAVF_DESCS_PER_LOOP_AVX; i++) {
+		const __m512i mbuf_ptrs = _mm512_loadu_si512
+			(&cache->objs[cache->len - IAVF_DESCS_PER_LOOP_AVX]);
+		_mm512_storeu_si512(rxp, mbuf_ptrs);
+
+		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_IAVF_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 6 & 7.
+		 */
+		const __m512i desc0_1 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas0);
+		const __m512i desc2_3 = _mm512_bsrli_epi128(desc0_1, 8);
+
+		const __m512i desc4_5 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas1);
+		const __m512i desc6_7 = _mm512_bsrli_epi128(desc4_5, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_1);
+		_mm512_storeu_si512((void *)(rxdp + 2), desc2_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_5);
+		_mm512_storeu_si512((void *)(rxdp + 6), desc6_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 desc0_3 = _mm512_permutexvar_epi64(permute_idx,
+								 iova_addrs);
+		const __m512i desc4_7 = _mm512_bsrli_epi128(desc0_3, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_7);
+#endif
+		rxp += IAVF_DESCS_PER_LOOP_AVX;
+		rxdp += IAVF_DESCS_PER_LOOP_AVX;
+		cache->len -= IAVF_DESCS_PER_LOOP_AVX;
+	}
+
+	rxq->rxrearm_start += IAVF_RXQ_REARM_THRESH;
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= IAVF_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 */
+	IAVF_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
+}
+
+#define IAVF_RX_LEN_MASK 0x80808080
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512(struct iavf_rx_queue *rxq,
+			       struct rte_mbuf **rx_pkts,
+			       uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0, 0,
+						    rxq->mbuf_initializer);
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_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 << IAVF_RX_DESC_STATUS_DD_SHIFT)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_DESC_STATUS_EOF_SHIFT);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 4th descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF     /* pkt_type set as unknown */
+			);
+	/**
+	 * 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 += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 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 desc4_7 = _mm512_mask_blend_epi16(IAVF_RX_LEN_MASK,
+								raw_desc4_7,
+								len4_7);
+		__m512i mb4_7 = _mm512_shuffle_epi8(desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, 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 __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 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 __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, adjusting length and
+		 * re-arranging fields. Then write into the mbuf
+		 */
+		const __m512i len0_3 = _mm512_slli_epi32(raw_desc0_3,
+							 PKTLEN_SHIFT);
+		const __m512i desc0_3 = _mm512_mask_blend_epi16(IAVF_RX_LEN_MASK,
+								raw_desc0_3,
+								len0_3);
+		__m512i mb0_3 = _mm512_shuffle_epi8(desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/* get the packet types */
+		const __m512i ptypes0_3 = _mm512_srli_epi64(desc0_3, 30);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		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 ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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_flags =
+			_mm256_shuffle_epi8(rss_flags_shuf,
+					    _mm256_srli_epi32(flag_bits, 11));
+		/**
+		 * 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 */
+		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				_mm256_or_si256(rss_flags, vlan_flags));
+		/**
+		 * 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;
+		const __m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7, 0);
+		const __m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7, 1);
+		const __m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3, 0);
+		const __m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3, 1);
+
+		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 << IAVF_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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
+/**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+			  uint16_t nb_pkts)
+{
+	return _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				     uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				    uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst = iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
diff --git a/drivers/net/iavf/meson.build b/drivers/net/iavf/meson.build
index 33407c5..3388cdf 100644
--- a/drivers/net/iavf/meson.build
+++ b/drivers/net/iavf/meson.build
@@ -34,4 +34,24 @@ if arch_subdir == 'x86'
 				c_args: [cflags, '-mavx2'])
 		objs += iavf_avx2_lib.extract_objects('iavf_rxtx_vec_avx2.c')
 	endif
+
+	iavf_avx512_cpu_support = (
+		cc.get_define('__AVX512F__', args: machine_args) != '' and
+		cc.get_define('__AVX512BW__', args: machine_args) != '')
+
+	iavf_avx512_cc_support = (
+		not machine_args.contains('-mno-avx512f') and
+		cc.has_argument('-mavx512f') and
+		cc.has_argument('-mavx512bw'))
+
+	if iavf_avx512_cpu_support == true or iavf_avx512_cc_support == true
+		cflags += ['-DCC_AVX512_SUPPORT']
+		iavf_avx512_lib = static_library('iavf_avx512_lib',
+				'iavf_rxtx_vec_avx512.c',
+				dependencies: [static_rte_ethdev,
+					static_rte_kvargs, static_rte_hash],
+				include_directories: includes,
+				c_args: [cflags, '-mavx512f', '-mavx512bw', '-march=skylake-avx512'])
+		objs += iavf_avx512_lib.extract_objects('iavf_rxtx_vec_avx512.c')
+	endif
 endif
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v6 2/3] net/iavf: enable AVX512 for flexible RX
  2020-10-28  5:14 ` [dpdk-dev] [PATCH v6 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-10-28  5:14   ` [dpdk-dev] [PATCH v6 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
@ 2020-10-28  5:14   ` Wenzhuo Lu
  2020-10-28  5:15   ` [dpdk-dev] [PATCH v6 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
  2 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-10-28  5:14 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the flexible RX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  10 +
 drivers/net/iavf/iavf_rxtx.h            |   6 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 707 ++++++++++++++++++++++++++++++++
 3 files changed, 723 insertions(+)

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index 6eedb12..69a4c3e 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2136,6 +2136,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
@@ -2155,6 +2160,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 7c1f05f..03b095d 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -440,9 +440,15 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
 uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
 				   uint16_t nb_pkts);
+uint16_t iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					    struct rte_mbuf **rx_pkts,
+					    uint16_t nb_pkts);
 uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 					     struct rte_mbuf **rx_pkts,
 					     uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+						      struct rte_mbuf **rx_pkts,
+						      uint16_t nb_pkts);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index 959067c..df0f43b 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -614,6 +614,631 @@
 	return received;
 }
 
+static inline __m256i
+flex_rxd_to_fdir_flags_vec_avx512(const __m256i fdir_id0_7)
+{
+#define FDID_MIS_MAGIC 0xFFFFFFFF
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR != (1 << 2));
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
+	const __m256i pkt_fdir_bit = _mm256_set1_epi32(PKT_RX_FDIR |
+						       PKT_RX_FDIR_ID);
+	/* desc->flow_id field == 0xFFFFFFFF means fdir mismatch */
+	const __m256i fdir_mis_mask = _mm256_set1_epi32(FDID_MIS_MAGIC);
+	__m256i fdir_mask = _mm256_cmpeq_epi32(fdir_id0_7,
+					       fdir_mis_mask);
+	/* this XOR op results to bit-reverse the fdir_mask */
+	fdir_mask = _mm256_xor_si256(fdir_mask, fdir_mis_mask);
+	const __m256i fdir_flags = _mm256_and_si256(fdir_mask, pkt_fdir_bit);
+
+	return fdir_flags;
+}
+
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512_flex_rxd(struct iavf_rx_queue *rxq,
+					struct rte_mbuf **rx_pkts,
+					uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0, 0,
+						    rxq->mbuf_initializer);
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_flex_desc *rxdp =
+		(union iavf_rx_flex_desc *)rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_rxq_rearm(rxq);
+
+	/* Before we start moving massive data around, check to see if
+	 * there is actually a packet available
+	 */
+	if (!(rxdp->wb.status_error0 &
+	      rte_cpu_to_le_32(1 << IAVF_RX_FLEX_DESC_STATUS0_DD_S)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 4th descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF     /* pkt_type set as unknown */
+			);
+	/**
+	 * 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 Checksum Reports, RSS indication
+	 * and VLAN indication.
+	 * bit6:4 for IP/L4 checksum errors.
+	 * bit12 is for RSS indication.
+	 * bit13 is for VLAN indication.
+	 */
+	const __m256i flags_mask =
+		_mm256_set1_epi32((7 << 4) | (1 << 12) | (1 << 13));
+	/**
+	 * data to be shuffled by the result of the flags mask shifted by 4
+	 * 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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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);
+	/**
+	 * data to be shuffled by result of flag mask, shifted down 12.
+	 * If RSS(bit12)/VLAN(bit13) are set,
+	 * shuffle moves appropriate flags in place.
+	 */
+	const __m256i rss_vlan_flags_shuf = _mm256_set_epi8(0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0,
+			/* end up 128-bits */
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0);
+
+	uint16_t i, received;
+
+	for (i = 0, received = 0; i < nb_pkts;
+	     i += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb4_7 = _mm512_shuffle_epi8(raw_desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptype_mask =
+			_mm512_set1_epi16(IAVF_RX_FLEX_DESC_PTYPE_M);
+		const __m512i ptypes4_7 =
+			_mm512_and_si512(raw_desc4_7, ptype_mask);
+		const __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 0);
+		const uint16_t ptype7 = _mm256_extract_epi16(ptypes6_7, 9);
+		const uint16_t ptype6 = _mm256_extract_epi16(ptypes6_7, 1);
+		const uint16_t ptype5 = _mm256_extract_epi16(ptypes4_5, 9);
+		const uint16_t ptype4 = _mm256_extract_epi16(ptypes4_5, 1);
+
+		const __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb0_3 = _mm512_shuffle_epi8(raw_desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptypes0_3 =
+			_mm512_and_si512(raw_desc0_3, ptype_mask);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		const uint16_t ptype3 = _mm256_extract_epi16(ptypes2_3, 9);
+		const uint16_t ptype2 = _mm256_extract_epi16(ptypes2_3, 1);
+		const uint16_t ptype1 = _mm256_extract_epi16(ptypes0_1, 9);
+		const uint16_t ptype0 = _mm256_extract_epi16(ptypes0_1, 1);
+
+		const __m512i ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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);
+		/**
+		 * 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, 4));
+		l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
+		l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
+		/* set rss and vlan flags */
+		const __m256i rss_vlan_flag_bits =
+			_mm256_srli_epi32(flag_bits, 12);
+		const __m256i rss_vlan_flags =
+			_mm256_shuffle_epi8(rss_vlan_flags_shuf,
+					    rss_vlan_flag_bits);
+
+		/* merge flags */
+		__m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+						     rss_vlan_flags);
+
+		if (rxq->fdir_enabled) {
+			const __m512i fdir_permute_mask = _mm512_set_epi32
+				(0, 0, 0, 0,
+				 0, 0, 0, 0,
+				 7, 15, 23, 31,
+				 3, 11, 19, 27);
+			__m512i fdir_tmp = _mm512_permutex2var_epi32
+				(raw_desc0_3, fdir_permute_mask, raw_desc4_7);
+			const __m256i fdir_id0_7 = _mm512_extracti64x4_epi64
+				(fdir_tmp, 0);
+			const __m256i fdir_flags =
+				flex_rxd_to_fdir_flags_vec_avx512(fdir_id0_7);
+
+			/* merge with fdir_flags */
+			mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_flags);
+
+			/* write to mbuf: have to use scalar store here */
+			rx_pkts[i + 0]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 3);
+
+			rx_pkts[i + 1]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 7);
+
+			rx_pkts[i + 2]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 2);
+
+			rx_pkts[i + 3]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 6);
+
+			rx_pkts[i + 4]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 1);
+
+			rx_pkts[i + 5]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 5);
+
+			rx_pkts[i + 6]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 0);
+
+			rx_pkts[i + 7]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 4);
+		} /* if() on fdir_enabled */
+
+		__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);
+
+#ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
+		/**
+		 * needs to load 2nd 16B of each desc for RSS hash parsing,
+		 * will cause performance drop to get into this context.
+		 */
+		if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
+		    DEV_RX_OFFLOAD_RSS_HASH) {
+			/* load bottom half of every 32B desc */
+			const __m128i raw_desc_bh7 =
+				_mm_load_si128
+					((void *)(&rxdp[7].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh6 =
+				_mm_load_si128
+					((void *)(&rxdp[6].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh5 =
+				_mm_load_si128
+					((void *)(&rxdp[5].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh4 =
+				_mm_load_si128
+					((void *)(&rxdp[4].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh3 =
+				_mm_load_si128
+					((void *)(&rxdp[3].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh2 =
+				_mm_load_si128
+					((void *)(&rxdp[2].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh1 =
+				_mm_load_si128
+					((void *)(&rxdp[1].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh0 =
+				_mm_load_si128
+					((void *)(&rxdp[0].wb.status_error1));
+
+			__m256i raw_desc_bh6_7 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh6),
+					 raw_desc_bh7, 1);
+			__m256i raw_desc_bh4_5 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh4),
+					 raw_desc_bh5, 1);
+			__m256i raw_desc_bh2_3 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh2),
+					 raw_desc_bh3, 1);
+			__m256i raw_desc_bh0_1 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh0),
+					 raw_desc_bh1, 1);
+
+			/**
+			 * to shift the 32b RSS hash value to the
+			 * highest 32b of each 128b before mask
+			 */
+			__m256i rss_hash6_7 =
+				_mm256_slli_epi64(raw_desc_bh6_7, 32);
+			__m256i rss_hash4_5 =
+				_mm256_slli_epi64(raw_desc_bh4_5, 32);
+			__m256i rss_hash2_3 =
+				_mm256_slli_epi64(raw_desc_bh2_3, 32);
+			__m256i rss_hash0_1 =
+				_mm256_slli_epi64(raw_desc_bh0_1, 32);
+
+			__m256i rss_hash_msk =
+				_mm256_set_epi32(0xFFFFFFFF, 0, 0, 0,
+						 0xFFFFFFFF, 0, 0, 0);
+
+			rss_hash6_7 = _mm256_and_si256
+					(rss_hash6_7, rss_hash_msk);
+			rss_hash4_5 = _mm256_and_si256
+					(rss_hash4_5, rss_hash_msk);
+			rss_hash2_3 = _mm256_and_si256
+					(rss_hash2_3, rss_hash_msk);
+			rss_hash0_1 = _mm256_and_si256
+					(rss_hash0_1, rss_hash_msk);
+
+			mb6_7 = _mm256_or_si256(mb6_7, rss_hash6_7);
+			mb4_5 = _mm256_or_si256(mb4_5, rss_hash4_5);
+			mb2_3 = _mm256_or_si256(mb2_3, rss_hash2_3);
+			mb0_1 = _mm256_or_si256(mb0_1, rss_hash0_1);
+		} /* if() on RSS hash parsing */
+#endif
+
+		/**
+		 * 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 <<
+					       IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+			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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
 /**
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -626,6 +1251,18 @@
 }
 
 /**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts)
+{
+	return _iavf_recv_raw_pkts_vec_avx512_flex_rxd(rx_queue, rx_pkts,
+						       nb_pkts, NULL);
+}
+
+/**
  * vPMD receive routine that reassembles single burst of 32 scattered packets
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -689,3 +1326,73 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+/**
+ * vPMD receive routine that reassembles single burst of
+ * 32 scattered packets for flex RxD
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512_flex_rxd(void *rx_queue,
+					      struct rte_mbuf **rx_pkts,
+					      uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_recv_raw_pkts_vec_avx512_flex_rxd(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 for flex RxD.
+ * Main receive routine that can handle arbitrary burst sizes
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst =
+			iavf_recv_scattered_burst_vec_avx512_flex_rxd
+				(rx_queue, rx_pkts + retval,
+				 IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v6 3/3] net/iavf: enable AVX512 for TX
  2020-10-28  5:14 ` [dpdk-dev] [PATCH v6 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-10-28  5:14   ` [dpdk-dev] [PATCH v6 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
  2020-10-28  5:14   ` [dpdk-dev] [PATCH v6 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
@ 2020-10-28  5:15   ` Wenzhuo Lu
  2 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-10-28  5:15 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the TX descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 doc/guides/rel_notes/release_20_11.rst  |   3 +
 drivers/net/iavf/iavf_ethdev.c          |   3 +-
 drivers/net/iavf/iavf_rxtx.c            |  34 +++-
 drivers/net/iavf/iavf_rxtx.h            |   7 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 304 ++++++++++++++++++++++++++++++++
 5 files changed, 343 insertions(+), 8 deletions(-)

diff --git a/doc/guides/rel_notes/release_20_11.rst b/doc/guides/rel_notes/release_20_11.rst
index 0d45b50..c981b64 100644
--- a/doc/guides/rel_notes/release_20_11.rst
+++ b/doc/guides/rel_notes/release_20_11.rst
@@ -20,6 +20,9 @@ DPDK Release 20.11
       make doc-guides-html
       xdg-open build/doc/html/guides/rel_notes/release_20_11.html
 
+   * **Added support of vector instructions on IAVF.**
+
+     Added support of AVX512 instructions in IAVF RX and TX path.
 
 New Features
 ------------
diff --git a/drivers/net/iavf/iavf_ethdev.c b/drivers/net/iavf/iavf_ethdev.c
index 0ef023c..fe6c8cb 100644
--- a/drivers/net/iavf/iavf_ethdev.c
+++ b/drivers/net/iavf/iavf_ethdev.c
@@ -606,7 +606,8 @@ static int iavf_config_rx_queues_irqs(struct rte_eth_dev *dev,
 		DEV_TX_OFFLOAD_GRE_TNL_TSO |
 		DEV_TX_OFFLOAD_IPIP_TNL_TSO |
 		DEV_TX_OFFLOAD_GENEVE_TNL_TSO |
-		DEV_TX_OFFLOAD_MULTI_SEGS;
+		DEV_TX_OFFLOAD_MULTI_SEGS |
+		DEV_TX_OFFLOAD_MBUF_FAST_FREE;
 
 	dev_info->default_rxconf = (struct rte_eth_rxconf) {
 		.rx_free_thresh = IAVF_DEFAULT_RX_FREE_THRESH,
diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index 69a4c3e..582afe7 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2210,20 +2210,22 @@
 	struct iavf_tx_queue *txq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_tx_vec_dev_check(dev) &&
 			rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
-		for (i = 0; i < dev->data->nb_tx_queues; i++) {
-			txq = dev->data->tx_queues[i];
-			if (!txq)
-				continue;
-			iavf_txq_vec_setup(txq);
-		}
-
 		if ((rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 		     rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&
 				rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1 &&
+		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512BW) == 1 &&
+		    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512)
+			use_avx512 = true;
+#endif
 
 		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
 			    use_avx2 ? "avx2 " : "",
@@ -2231,8 +2233,26 @@
 		dev->tx_pkt_burst = use_avx2 ?
 				    iavf_xmit_pkts_vec_avx2 :
 				    iavf_xmit_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+		if (use_avx512)
+			dev->tx_pkt_burst = iavf_xmit_pkts_vec_avx512;
+#endif
 		dev->tx_pkt_prepare = NULL;
 
+		for (i = 0; i < dev->data->nb_tx_queues; i++) {
+			txq = dev->data->tx_queues[i];
+			if (!txq)
+				continue;
+#ifdef CC_AVX512_SUPPORT
+			if (use_avx512)
+				iavf_txq_vec_setup_avx512(txq);
+			else
+				iavf_txq_vec_setup(txq);
+#else
+			iavf_txq_vec_setup(txq);
+#endif
+		}
+
 		return;
 	}
 #endif
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 03b095d..b22ccc4 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -122,6 +122,10 @@ struct iavf_tx_entry {
 	uint16_t last_id;
 };
 
+struct iavf_tx_vec_entry {
+	struct rte_mbuf *mbuf;
+};
+
 /* Structure associated with each TX queue. */
 struct iavf_tx_queue {
 	const struct rte_memzone *mz;  /* memzone for Tx ring */
@@ -449,6 +453,9 @@ uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
 						      struct rte_mbuf **rx_pkts,
 						      uint16_t nb_pkts);
+uint16_t iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				   uint16_t nb_pkts);
+int iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq);
 
 const uint32_t *iavf_get_default_ptype_table(void);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index df0f43b..8680734 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -1396,3 +1396,307 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+static __rte_always_inline int
+iavf_tx_free_bufs_avx512(struct iavf_tx_queue *txq)
+{
+	struct iavf_tx_vec_entry *txep;
+	uint32_t n;
+	uint32_t i;
+	int nb_free = 0;
+	struct rte_mbuf *m, *free[IAVF_VPMD_TX_MAX_FREE_BUF];
+
+	/* check DD bits on threshold descriptor */
+	if ((txq->tx_ring[txq->next_dd].cmd_type_offset_bsz &
+	     rte_cpu_to_le_64(IAVF_TXD_QW1_DTYPE_MASK)) !=
+	    rte_cpu_to_le_64(IAVF_TX_DESC_DTYPE_DESC_DONE))
+		return 0;
+
+	n = txq->rs_thresh;
+
+	 /* first buffer to free from S/W ring is at index
+	  * tx_next_dd - (tx_rs_thresh-1)
+	  */
+	txep = (void *)txq->sw_ring;
+	txep += txq->next_dd - (n - 1);
+
+	if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE && (n & 31) == 0) {
+		struct rte_mempool *mp = txep[0].mbuf->pool;
+		struct rte_mempool_cache *cache = rte_mempool_default_cache(mp,
+								rte_lcore_id());
+		void **cache_objs = &cache->objs[cache->len];
+
+		if (n > RTE_MEMPOOL_CACHE_MAX_SIZE) {
+			rte_mempool_ops_enqueue_bulk(mp, (void *)txep, n);
+			goto done;
+		}
+
+		/* The cache follows the following algorithm
+		 *   1. Add the objects to the cache
+		 *   2. Anything greater than the cache min value (if it crosses the
+		 *   cache flush threshold) is flushed to the ring.
+		 */
+		/* Add elements back into the cache */
+		uint32_t copied = 0;
+		/* n is multiple of 32 */
+		while (copied < n) {
+			const __m512i a = _mm512_loadu_si512(&txep[copied]);
+			const __m512i b = _mm512_loadu_si512(&txep[copied + 8]);
+			const __m512i c = _mm512_loadu_si512(&txep[copied + 16]);
+			const __m512i d = _mm512_loadu_si512(&txep[copied + 24]);
+
+			_mm512_storeu_si512(&cache_objs[copied], a);
+			_mm512_storeu_si512(&cache_objs[copied + 8], b);
+			_mm512_storeu_si512(&cache_objs[copied + 16], c);
+			_mm512_storeu_si512(&cache_objs[copied + 24], d);
+			copied += 32;
+		}
+		cache->len += n;
+
+		if (cache->len >= cache->flushthresh) {
+			rte_mempool_ops_enqueue_bulk(mp,
+						     &cache->objs[cache->size],
+						     cache->len - cache->size);
+			cache->len = cache->size;
+		}
+		goto done;
+	}
+
+	m = rte_pktmbuf_prefree_seg(txep[0].mbuf);
+	if (likely(m)) {
+		free[0] = m;
+		nb_free = 1;
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (likely(m)) {
+				if (likely(m->pool == free[0]->pool)) {
+					free[nb_free++] = m;
+				} else {
+					rte_mempool_put_bulk(free[0]->pool,
+							     (void *)free,
+							     nb_free);
+					free[0] = m;
+					nb_free = 1;
+				}
+			}
+		}
+		rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
+	} else {
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (m)
+				rte_mempool_put(m->pool, m);
+		}
+	}
+
+done:
+	/* buffers were freed, update counters */
+	txq->nb_free = (uint16_t)(txq->nb_free + txq->rs_thresh);
+	txq->next_dd = (uint16_t)(txq->next_dd + txq->rs_thresh);
+	if (txq->next_dd >= txq->nb_tx_desc)
+		txq->next_dd = (uint16_t)(txq->rs_thresh - 1);
+
+	return txq->rs_thresh;
+}
+
+static __rte_always_inline void
+tx_backlog_entry_avx512(struct iavf_tx_vec_entry *txep,
+			struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	int i;
+
+	for (i = 0; i < (int)nb_pkts; ++i)
+		txep[i].mbuf = tx_pkts[i];
+}
+
+static inline void
+iavf_vtx1(volatile struct iavf_tx_desc *txdp,
+	  struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(IAVF_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT) |
+		 ((uint64_t)pkt->data_len << IAVF_TXD_QW1_TX_BUF_SZ_SHIFT));
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+					    pkt->buf_iova + pkt->data_off);
+	_mm_storeu_si128((__m128i *)txdp, descriptor);
+}
+
+#define IAVF_TX_LEN_MASK 0xAA
+#define IAVF_TX_OFF_MASK 0x55
+static inline void
+iavf_vtx(volatile struct iavf_tx_desc *txdp,
+	 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (IAVF_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT));
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		iavf_vtx1(txdp, *pkt, flags);
+		nb_pkts--, txdp++, pkt++;
+	}
+
+	/* do 4 at a time while possible, in bursts */
+	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
+		__m512i desc4 =
+			_mm512_set_epi64
+				((uint64_t)pkt[3]->data_len,
+				 pkt[3]->buf_iova,
+				 (uint64_t)pkt[2]->data_len,
+				 pkt[2]->buf_iova,
+				 (uint64_t)pkt[1]->data_len,
+				 pkt[1]->buf_iova,
+				 (uint64_t)pkt[0]->data_len,
+				 pkt[0]->buf_iova);
+		__m512i hi_qw_tmpl_4 = _mm512_set1_epi64(hi_qw_tmpl);
+		__m512i data_off_4 =
+			_mm512_set_epi64
+				(0,
+				 pkt[3]->data_off,
+				 0,
+				 pkt[2]->data_off,
+				 0,
+				 pkt[1]->data_off,
+				 0,
+				 pkt[0]->data_off);
+
+		desc4 = _mm512_mask_slli_epi64(desc4, IAVF_TX_LEN_MASK, desc4,
+					       IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
+		desc4 = _mm512_mask_or_epi64(desc4, IAVF_TX_LEN_MASK, desc4,
+					     hi_qw_tmpl_4);
+		desc4 = _mm512_mask_add_epi64(desc4, IAVF_TX_OFF_MASK, desc4,
+					      data_off_4);
+		_mm512_storeu_si512((void *)txdp, desc4);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		iavf_vtx1(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+iavf_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				 uint16_t nb_pkts)
+{
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+	volatile struct iavf_tx_desc *txdp;
+	struct iavf_tx_vec_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	/* bit2 is reserved and must be set to 1 according to Spec */
+	uint64_t flags = IAVF_TX_DESC_CMD_EOP | IAVF_TX_DESC_CMD_ICRC;
+	uint64_t rs = IAVF_TX_DESC_CMD_RS | flags;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
+
+	if (txq->nb_free < txq->free_thresh)
+		iavf_tx_free_bufs_avx512(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = (void *)txq->sw_ring;
+	txep += tx_id;
+
+	txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		tx_backlog_entry_avx512(txep, tx_pkts, n);
+
+		iavf_vtx(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		iavf_vtx1(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = (void *)txq->sw_ring;
+		txep += tx_id;
+	}
+
+	tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
+
+	iavf_vtx(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->next_rs) {
+		txq->tx_ring[txq->next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)IAVF_TX_DESC_CMD_RS) <<
+					 IAVF_TXD_QW1_CMD_SHIFT);
+		txq->next_rs =
+			(uint16_t)(txq->next_rs + txq->rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	IAVF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+			  uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->rs_thresh);
+		ret = iavf_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;
+}
+
+static inline void
+iavf_tx_queue_release_mbufs_avx512(struct iavf_tx_queue *txq)
+{
+	unsigned int i;
+	const uint16_t max_desc = (uint16_t)(txq->nb_tx_desc - 1);
+	struct iavf_tx_vec_entry *swr = (void *)txq->sw_ring;
+
+	if (!txq->sw_ring || txq->nb_free == max_desc)
+		return;
+
+	i = txq->next_dd - txq->rs_thresh + 1;
+	if (txq->tx_tail < i) {
+		for (; i < txq->nb_tx_desc; i++) {
+			rte_pktmbuf_free_seg(swr[i].mbuf);
+			swr[i].mbuf = NULL;
+		}
+		i = 0;
+	}
+}
+
+static const struct iavf_txq_ops avx512_vec_txq_ops = {
+	.release_mbufs = iavf_tx_queue_release_mbufs_avx512,
+};
+
+int __rte_cold
+iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq)
+{
+	txq->ops = &avx512_vec_txq_ops;
+	return 0;
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf
  2020-09-10  5:59 [dpdk-dev] [PATCH 0/3] enable AVX512 for iavf Wenzhuo Lu
                   ` (7 preceding siblings ...)
  2020-10-28  5:14 ` [dpdk-dev] [PATCH v6 0/3] enable AVX512 for iavf Wenzhuo Lu
@ 2020-10-29  1:24 ` Wenzhuo Lu
  2020-10-29  1:24   ` [dpdk-dev] [PATCH v7 1/3] net/iavf: enable AVX512 for legacy Rx Wenzhuo Lu
                     ` (3 more replies)
  8 siblings, 4 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-10-29  1:24 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu

AVX512 instructions is supported by more and more platforms. These instructions can be used in the data path to enhance the per-core performance of packet processing.
Comparing with the existing implementation, this path set introduces some AVX512 instructions into the iavf data path, and we get a better per-code throughput.

v2:
Update meson.build.
Repalce the deprecated 'buf_physaddr' by 'buf_iova'.

v3:
Fix compile errors.

v4:
Fix wrong info in commnets.
Trivial adjustment of the arrangement.

v5:
Support "max SIMD bitwidth".

v6:
Rework meson build to fix compile issue for AVX512BW.

v7:
rebased on next-net_intel.

Wenzhuo Lu (3):
  net/iavf: enable AVX512 for legacy Rx
  net/iavf: enable AVX512 for flexible Rx
  net/iavf: enable AVX512 for Tx

 doc/guides/rel_notes/release_20_11.rst  |    3 +
 drivers/net/iavf/iavf_ethdev.c          |    3 +-
 drivers/net/iavf/iavf_rxtx.c            |   73 +-
 drivers/net/iavf/iavf_rxtx.h            |   18 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 1702 +++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |   20 +
 6 files changed, 1807 insertions(+), 12 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v7 1/3] net/iavf: enable AVX512 for legacy Rx
  2020-10-29  1:24 ` [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf Wenzhuo Lu
@ 2020-10-29  1:24   ` Wenzhuo Lu
  2020-10-29  1:24   ` [dpdk-dev] [PATCH v7 2/3] net/iavf: enable AVX512 for flexible Rx Wenzhuo Lu
                     ` (2 subsequent siblings)
  3 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-10-29  1:24 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the legacy Rx descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  29 +-
 drivers/net/iavf/iavf_rxtx.h            |   5 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 691 ++++++++++++++++++++++++++++++++
 drivers/net/iavf/meson.build            |  20 +
 4 files changed, 741 insertions(+), 4 deletions(-)
 create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index d30aaf8..abadf0a 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2311,6 +2311,9 @@
 	struct iavf_rx_queue *rxq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_rx_vec_dev_check(dev) &&
 			rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
@@ -2323,6 +2326,12 @@
 		     rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&
 				rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1 &&
+		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512BW) == 1 &&
+		    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512)
+			use_avx512 = true;
+#endif
 
 		if (dev->data->scattered_rx) {
 			PMD_DRV_LOG(DEBUG,
@@ -2330,27 +2339,39 @@
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
 					iavf_recv_scattered_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512;
+#endif
+			}
 		} else {
 			PMD_DRV_LOG(DEBUG, "Using %sVector Rx (port %d).",
 				    use_avx2 ? "avx2 " : "",
 				    dev->data->port_id);
 			if (vf->vf_res->vf_cap_flags &
-				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC)
+				VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
-			else
+			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
 					iavf_recv_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512;
+#endif
+			}
 		}
 
 		return;
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 02945b8..3da7189 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -463,6 +463,11 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_tx_vec_dev_check(struct rte_eth_dev *dev);
 int iavf_rxq_vec_setup(struct iavf_rx_queue *rxq);
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
+uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts);
 
 uint8_t iavf_proto_xtr_type_to_rxdid(uint8_t xtr_type);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
new file mode 100644
index 0000000..959067c
--- /dev/null
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -0,0 +1,691 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include "iavf_rxtx_vec_common.h"
+
+#include <x86intrin.h>
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+#define IAVF_DESCS_PER_LOOP_AVX 8
+#define PKTLEN_SHIFT 10
+
+static inline void
+iavf_rxq_rearm(struct iavf_rx_queue *rxq)
+{
+	int i;
+	uint16_t rx_id;
+	volatile union iavf_rx_desc *rxdp;
+	struct rte_mempool_cache *cache =
+		rte_mempool_default_cache(rxq->mp, rte_lcore_id());
+	struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start];
+
+	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.
+	 */
+
+	/* Can this be satisfied from the cache? */
+	if (cache->len < IAVF_RXQ_REARM_THRESH) {
+		/* No. Backfill the cache first, and then fill from it */
+		uint32_t req = IAVF_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 + IAVF_RXQ_REARM_THRESH >=
+			    rxq->nb_rx_desc) {
+				__m128i dma_addr0;
+
+				dma_addr0 = _mm_setzero_si128();
+				for (i = 0; i < IAVF_VPMD_DESCS_PER_LOOP; i++) {
+					rxp[i] = &rxq->fake_mbuf;
+					_mm_storeu_si128((__m128i *)&rxdp[i].read,
+							 dma_addr0);
+				}
+			}
+			rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+					IAVF_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_IAVF_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 < IAVF_RXQ_REARM_THRESH / IAVF_DESCS_PER_LOOP_AVX; i++) {
+		const __m512i mbuf_ptrs = _mm512_loadu_si512
+			(&cache->objs[cache->len - IAVF_DESCS_PER_LOOP_AVX]);
+		_mm512_storeu_si512(rxp, mbuf_ptrs);
+
+		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_IAVF_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 6 & 7.
+		 */
+		const __m512i desc0_1 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas0);
+		const __m512i desc2_3 = _mm512_bsrli_epi128(desc0_1, 8);
+
+		const __m512i desc4_5 = _mm512_permutexvar_epi64
+				(permute_idx,
+				 iovas1);
+		const __m512i desc6_7 = _mm512_bsrli_epi128(desc4_5, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_1);
+		_mm512_storeu_si512((void *)(rxdp + 2), desc2_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_5);
+		_mm512_storeu_si512((void *)(rxdp + 6), desc6_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 desc0_3 = _mm512_permutexvar_epi64(permute_idx,
+								 iova_addrs);
+		const __m512i desc4_7 = _mm512_bsrli_epi128(desc0_3, 8);
+
+		_mm512_storeu_si512((void *)rxdp, desc0_3);
+		_mm512_storeu_si512((void *)(rxdp + 4), desc4_7);
+#endif
+		rxp += IAVF_DESCS_PER_LOOP_AVX;
+		rxdp += IAVF_DESCS_PER_LOOP_AVX;
+		cache->len -= IAVF_DESCS_PER_LOOP_AVX;
+	}
+
+	rxq->rxrearm_start += IAVF_RXQ_REARM_THRESH;
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= IAVF_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 */
+	IAVF_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
+}
+
+#define IAVF_RX_LEN_MASK 0x80808080
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512(struct iavf_rx_queue *rxq,
+			       struct rte_mbuf **rx_pkts,
+			       uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0, 0,
+						    rxq->mbuf_initializer);
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_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 << IAVF_RX_DESC_STATUS_DD_SHIFT)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_DESC_STATUS_EOF_SHIFT);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 4th descriptor */
+			 0x07060504,    /* octet 4~7, 32bits rss */
+			 0x03020F0E,    /* octet 2~3, low 16 bits vlan_macip */
+					/* octet 15~14, 16 bits data_len */
+			 0xFFFF0F0E,    /* skip high 16 bits pkt_len, zero out */
+					/* octet 15~14, low 16 bits pkt_len */
+			 0xFFFFFFFF     /* pkt_type set as unknown */
+			);
+	/**
+	 * 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 += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 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 desc4_7 = _mm512_mask_blend_epi16(IAVF_RX_LEN_MASK,
+								raw_desc4_7,
+								len4_7);
+		__m512i mb4_7 = _mm512_shuffle_epi8(desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, 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 __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 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 __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, adjusting length and
+		 * re-arranging fields. Then write into the mbuf
+		 */
+		const __m512i len0_3 = _mm512_slli_epi32(raw_desc0_3,
+							 PKTLEN_SHIFT);
+		const __m512i desc0_3 = _mm512_mask_blend_epi16(IAVF_RX_LEN_MASK,
+								raw_desc0_3,
+								len0_3);
+		__m512i mb0_3 = _mm512_shuffle_epi8(desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/* get the packet types */
+		const __m512i ptypes0_3 = _mm512_srli_epi64(desc0_3, 30);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		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 ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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_flags =
+			_mm256_shuffle_epi8(rss_flags_shuf,
+					    _mm256_srli_epi32(flag_bits, 11));
+		/**
+		 * 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 */
+		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				_mm256_or_si256(rss_flags, vlan_flags));
+		/**
+		 * 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;
+		const __m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7, 0);
+		const __m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7, 1);
+		const __m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3, 0);
+		const __m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3, 1);
+
+		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 << IAVF_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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
+/**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+			  uint16_t nb_pkts)
+{
+	return _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				     uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_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 < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				    uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst = iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
diff --git a/drivers/net/iavf/meson.build b/drivers/net/iavf/meson.build
index 306faaf..e257f5a 100644
--- a/drivers/net/iavf/meson.build
+++ b/drivers/net/iavf/meson.build
@@ -34,6 +34,26 @@ if arch_subdir == 'x86'
 				c_args: [cflags, '-mavx2'])
 		objs += iavf_avx2_lib.extract_objects('iavf_rxtx_vec_avx2.c')
 	endif
+
+	iavf_avx512_cpu_support = (
+		cc.get_define('__AVX512F__', args: machine_args) != '' and
+		cc.get_define('__AVX512BW__', args: machine_args) != '')
+
+	iavf_avx512_cc_support = (
+		not machine_args.contains('-mno-avx512f') and
+		cc.has_argument('-mavx512f') and
+		cc.has_argument('-mavx512bw'))
+
+	if iavf_avx512_cpu_support == true or iavf_avx512_cc_support == true
+		cflags += ['-DCC_AVX512_SUPPORT']
+		iavf_avx512_lib = static_library('iavf_avx512_lib',
+				'iavf_rxtx_vec_avx512.c',
+				dependencies: [static_rte_ethdev,
+					static_rte_kvargs, static_rte_hash],
+				include_directories: includes,
+				c_args: [cflags, '-mavx512f', '-mavx512bw', '-march=skylake-avx512'])
+		objs += iavf_avx512_lib.extract_objects('iavf_rxtx_vec_avx512.c')
+	endif
 endif
 
 headers = files('rte_pmd_iavf.h')
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v7 2/3] net/iavf: enable AVX512 for flexible Rx
  2020-10-29  1:24 ` [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-10-29  1:24   ` [dpdk-dev] [PATCH v7 1/3] net/iavf: enable AVX512 for legacy Rx Wenzhuo Lu
@ 2020-10-29  1:24   ` Wenzhuo Lu
  2020-10-29  1:24   ` [dpdk-dev] [PATCH v7 3/3] net/iavf: enable AVX512 for Tx Wenzhuo Lu
  2020-10-29  4:00   ` [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf Zhang, Qi Z
  3 siblings, 0 replies; 39+ messages in thread
From: Wenzhuo Lu @ 2020-10-29  1:24 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the flexible Rx descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  10 +
 drivers/net/iavf/iavf_rxtx.h            |   6 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 707 ++++++++++++++++++++++++++++++++
 3 files changed, 723 insertions(+)

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index abadf0a..f51471f 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2343,6 +2343,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2_flex_rxd :
 					iavf_recv_scattered_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_scattered_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_scattered_pkts_vec_avx2 :
@@ -2362,6 +2367,11 @@
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2_flex_rxd :
 					iavf_recv_pkts_vec_flex_rxd;
+#ifdef CC_AVX512_SUPPORT
+				if (use_avx512)
+					dev->rx_pkt_burst =
+						iavf_recv_pkts_vec_avx512_flex_rxd;
+#endif
 			} else {
 				dev->rx_pkt_burst = use_avx2 ?
 					iavf_recv_pkts_vec_avx2 :
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 3da7189..5bf91df 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -465,9 +465,15 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 int iavf_txq_vec_setup(struct iavf_tx_queue *txq);
 uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
 				   uint16_t nb_pkts);
+uint16_t iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					    struct rte_mbuf **rx_pkts,
+					    uint16_t nb_pkts);
 uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 					     struct rte_mbuf **rx_pkts,
 					     uint16_t nb_pkts);
+uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+						      struct rte_mbuf **rx_pkts,
+						      uint16_t nb_pkts);
 
 uint8_t iavf_proto_xtr_type_to_rxdid(uint8_t xtr_type);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index 959067c..df0f43b 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -614,6 +614,631 @@
 	return received;
 }
 
+static inline __m256i
+flex_rxd_to_fdir_flags_vec_avx512(const __m256i fdir_id0_7)
+{
+#define FDID_MIS_MAGIC 0xFFFFFFFF
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR != (1 << 2));
+	RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
+	const __m256i pkt_fdir_bit = _mm256_set1_epi32(PKT_RX_FDIR |
+						       PKT_RX_FDIR_ID);
+	/* desc->flow_id field == 0xFFFFFFFF means fdir mismatch */
+	const __m256i fdir_mis_mask = _mm256_set1_epi32(FDID_MIS_MAGIC);
+	__m256i fdir_mask = _mm256_cmpeq_epi32(fdir_id0_7,
+					       fdir_mis_mask);
+	/* this XOR op results to bit-reverse the fdir_mask */
+	fdir_mask = _mm256_xor_si256(fdir_mask, fdir_mis_mask);
+	const __m256i fdir_flags = _mm256_and_si256(fdir_mask, pkt_fdir_bit);
+
+	return fdir_flags;
+}
+
+static inline uint16_t
+_iavf_recv_raw_pkts_vec_avx512_flex_rxd(struct iavf_rx_queue *rxq,
+					struct rte_mbuf **rx_pkts,
+					uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0, 0,
+						    rxq->mbuf_initializer);
+	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union iavf_rx_flex_desc *rxdp =
+		(union iavf_rx_flex_desc *)rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_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 > IAVF_RXQ_REARM_THRESH)
+		iavf_rxq_rearm(rxq);
+
+	/* Before we start moving massive data around, check to see if
+	 * there is actually a packet available
+	 */
+	if (!(rxdp->wb.status_error0 &
+	      rte_cpu_to_le_32(1 << IAVF_RX_FLEX_DESC_STATUS0_DD_S)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 2nd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 3rd descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0,             /* ignore pkt_type field */
+			 /* 4th descriptor */
+			 0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore pkt_type field */
+			);
+
+	/* 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,
+			IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set_epi32
+			(/* 1st descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 2nd descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 3rd descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF,    /* pkt_type set as unknown */
+			 /* 4th descriptor */
+			 0xFFFFFFFF,    /* rss hash parsed separately */
+			 0x0B0A0504,    /* octet 10~11, 16 bits vlan_macip */
+					/* octet 4~5, 16 bits data_len */
+			 0xFFFF0504,    /* skip hi 16 bits pkt_len, zero out */
+					/* octet 4~5, 16 bits pkt_len */
+			 0xFFFFFFFF     /* pkt_type set as unknown */
+			);
+	/**
+	 * 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 Checksum Reports, RSS indication
+	 * and VLAN indication.
+	 * bit6:4 for IP/L4 checksum errors.
+	 * bit12 is for RSS indication.
+	 * bit13 is for VLAN indication.
+	 */
+	const __m256i flags_mask =
+		_mm256_set1_epi32((7 << 4) | (1 << 12) | (1 << 13));
+	/**
+	 * data to be shuffled by the result of the flags mask shifted by 4
+	 * 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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
+			 PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
+			(PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_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);
+	/**
+	 * data to be shuffled by result of flag mask, shifted down 12.
+	 * If RSS(bit12)/VLAN(bit13) are set,
+	 * shuffle moves appropriate flags in place.
+	 */
+	const __m256i rss_vlan_flags_shuf = _mm256_set_epi8(0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0,
+			/* end up 128-bits */
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			0, 0, 0, 0,
+			PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+			PKT_RX_RSS_HASH, 0);
+
+	uint16_t i, received;
+
+	for (i = 0, received = 0; i < nb_pkts;
+	     i += IAVF_DESCS_PER_LOOP_AVX,
+	     rxdp += IAVF_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;
+
+		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_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2);
+		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3);
+		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2);
+		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 4-7 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb4_7 = _mm512_shuffle_epi8(raw_desc4_7, shuf_msk);
+
+		mb4_7 = _mm512_add_epi16(mb4_7, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptype_mask =
+			_mm512_set1_epi16(IAVF_RX_FLEX_DESC_PTYPE_M);
+		const __m512i ptypes4_7 =
+			_mm512_and_si512(raw_desc4_7, ptype_mask);
+		const __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 0);
+		const uint16_t ptype7 = _mm256_extract_epi16(ptypes6_7, 9);
+		const uint16_t ptype6 = _mm256_extract_epi16(ptypes6_7, 1);
+		const uint16_t ptype5 = _mm256_extract_epi16(ptypes4_5, 9);
+		const uint16_t ptype4 = _mm256_extract_epi16(ptypes4_5, 1);
+
+		const __m512i ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype7],
+			 0, 0, 0, type_table[ptype6],
+			 0, 0, 0, type_table[ptype5],
+			 0, 0, 0, type_table[ptype4]);
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+
+		/**
+		 * convert descriptors 0-3 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb0_3 = _mm512_shuffle_epi8(raw_desc0_3, shuf_msk);
+
+		mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust);
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptypes0_3 =
+			_mm512_and_si512(raw_desc0_3, ptype_mask);
+		const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0);
+		const uint16_t ptype3 = _mm256_extract_epi16(ptypes2_3, 9);
+		const uint16_t ptype2 = _mm256_extract_epi16(ptypes2_3, 1);
+		const uint16_t ptype1 = _mm256_extract_epi16(ptypes0_1, 9);
+		const uint16_t ptype0 = _mm256_extract_epi16(ptypes0_1, 1);
+
+		const __m512i ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, type_table[ptype3],
+			 0, 0, 0, type_table[ptype2],
+			 0, 0, 0, type_table[ptype1],
+			 0, 0, 0, type_table[ptype0]);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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
+			(raw_desc4_7, status_permute_msk, raw_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);
+		/**
+		 * 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, 4));
+		l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
+		l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
+		/* set rss and vlan flags */
+		const __m256i rss_vlan_flag_bits =
+			_mm256_srli_epi32(flag_bits, 12);
+		const __m256i rss_vlan_flags =
+			_mm256_shuffle_epi8(rss_vlan_flags_shuf,
+					    rss_vlan_flag_bits);
+
+		/* merge flags */
+		__m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+						     rss_vlan_flags);
+
+		if (rxq->fdir_enabled) {
+			const __m512i fdir_permute_mask = _mm512_set_epi32
+				(0, 0, 0, 0,
+				 0, 0, 0, 0,
+				 7, 15, 23, 31,
+				 3, 11, 19, 27);
+			__m512i fdir_tmp = _mm512_permutex2var_epi32
+				(raw_desc0_3, fdir_permute_mask, raw_desc4_7);
+			const __m256i fdir_id0_7 = _mm512_extracti64x4_epi64
+				(fdir_tmp, 0);
+			const __m256i fdir_flags =
+				flex_rxd_to_fdir_flags_vec_avx512(fdir_id0_7);
+
+			/* merge with fdir_flags */
+			mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_flags);
+
+			/* write to mbuf: have to use scalar store here */
+			rx_pkts[i + 0]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 3);
+
+			rx_pkts[i + 1]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 7);
+
+			rx_pkts[i + 2]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 2);
+
+			rx_pkts[i + 3]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 6);
+
+			rx_pkts[i + 4]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 1);
+
+			rx_pkts[i + 5]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 5);
+
+			rx_pkts[i + 6]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 0);
+
+			rx_pkts[i + 7]->hash.fdir.hi =
+				_mm256_extract_epi32(fdir_id0_7, 4);
+		} /* if() on fdir_enabled */
+
+		__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);
+
+#ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
+		/**
+		 * needs to load 2nd 16B of each desc for RSS hash parsing,
+		 * will cause performance drop to get into this context.
+		 */
+		if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
+		    DEV_RX_OFFLOAD_RSS_HASH) {
+			/* load bottom half of every 32B desc */
+			const __m128i raw_desc_bh7 =
+				_mm_load_si128
+					((void *)(&rxdp[7].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh6 =
+				_mm_load_si128
+					((void *)(&rxdp[6].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh5 =
+				_mm_load_si128
+					((void *)(&rxdp[5].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh4 =
+				_mm_load_si128
+					((void *)(&rxdp[4].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh3 =
+				_mm_load_si128
+					((void *)(&rxdp[3].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh2 =
+				_mm_load_si128
+					((void *)(&rxdp[2].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh1 =
+				_mm_load_si128
+					((void *)(&rxdp[1].wb.status_error1));
+			rte_compiler_barrier();
+			const __m128i raw_desc_bh0 =
+				_mm_load_si128
+					((void *)(&rxdp[0].wb.status_error1));
+
+			__m256i raw_desc_bh6_7 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh6),
+					 raw_desc_bh7, 1);
+			__m256i raw_desc_bh4_5 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh4),
+					 raw_desc_bh5, 1);
+			__m256i raw_desc_bh2_3 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh2),
+					 raw_desc_bh3, 1);
+			__m256i raw_desc_bh0_1 =
+				_mm256_inserti128_si256
+					(_mm256_castsi128_si256(raw_desc_bh0),
+					 raw_desc_bh1, 1);
+
+			/**
+			 * to shift the 32b RSS hash value to the
+			 * highest 32b of each 128b before mask
+			 */
+			__m256i rss_hash6_7 =
+				_mm256_slli_epi64(raw_desc_bh6_7, 32);
+			__m256i rss_hash4_5 =
+				_mm256_slli_epi64(raw_desc_bh4_5, 32);
+			__m256i rss_hash2_3 =
+				_mm256_slli_epi64(raw_desc_bh2_3, 32);
+			__m256i rss_hash0_1 =
+				_mm256_slli_epi64(raw_desc_bh0_1, 32);
+
+			__m256i rss_hash_msk =
+				_mm256_set_epi32(0xFFFFFFFF, 0, 0, 0,
+						 0xFFFFFFFF, 0, 0, 0);
+
+			rss_hash6_7 = _mm256_and_si256
+					(rss_hash6_7, rss_hash_msk);
+			rss_hash4_5 = _mm256_and_si256
+					(rss_hash4_5, rss_hash_msk);
+			rss_hash2_3 = _mm256_and_si256
+					(rss_hash2_3, rss_hash_msk);
+			rss_hash0_1 = _mm256_and_si256
+					(rss_hash0_1, rss_hash_msk);
+
+			mb6_7 = _mm256_or_si256(mb6_7, rss_hash6_7);
+			mb4_5 = _mm256_or_si256(mb4_5, rss_hash4_5);
+			mb2_3 = _mm256_or_si256(mb2_3, rss_hash2_3);
+			mb0_1 = _mm256_or_si256(mb0_1, rss_hash0_1);
+		} /* if() on RSS hash parsing */
+#endif
+
+		/**
+		 * 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 <<
+					       IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
+			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(/* zero hi 64b */
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     0xFF, 0xFF, 0xFF, 0xFF,
+					     /* move values to lo 64b */
+					     8, 0, 10, 2,
+					     12, 4, 14, 6);
+			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
+			*(uint64_t *)split_packet =
+				_mm_cvtsi128_si64(split_bits);
+			split_packet += IAVF_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 != IAVF_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 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
 /**
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -626,6 +1251,18 @@
 }
 
 /**
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_pkts_vec_avx512_flex_rxd(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts)
+{
+	return _iavf_recv_raw_pkts_vec_avx512_flex_rxd(rx_queue, rx_pkts,
+						       nb_pkts, NULL);
+}
+
+/**
  * vPMD receive routine that reassembles single burst of 32 scattered packets
  * Notice:
  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
@@ -689,3 +1326,73 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+/**
+ * vPMD receive routine that reassembles single burst of
+ * 32 scattered packets for flex RxD
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+iavf_recv_scattered_burst_vec_avx512_flex_rxd(void *rx_queue,
+					      struct rte_mbuf **rx_pkts,
+					      uint16_t nb_pkts)
+{
+	struct iavf_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _iavf_recv_raw_pkts_vec_avx512_flex_rxd(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 for flex RxD.
+ * Main receive routine that can handle arbitrary burst sizes
+ * Notice:
+ * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
+					     struct rte_mbuf **rx_pkts,
+					     uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
+		uint16_t burst =
+			iavf_recv_scattered_burst_vec_avx512_flex_rxd
+				(rx_queue, rx_pkts + retval,
+				 IAVF_VPMD_RX_MAX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < IAVF_VPMD_RX_MAX_BURST)
+			return retval;
+	}
+	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* [dpdk-dev] [PATCH v7 3/3] net/iavf: enable AVX512 for Tx
  2020-10-29  1:24 ` [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf Wenzhuo Lu
  2020-10-29  1:24   ` [dpdk-dev] [PATCH v7 1/3] net/iavf: enable AVX512 for legacy Rx Wenzhuo Lu
  2020-10-29  1:24   ` [dpdk-dev] [PATCH v7 2/3] net/iavf: enable AVX512 for flexible Rx Wenzhuo Lu
@ 2020-10-29  1:24   ` Wenzhuo Lu
  2020-10-30 23:29     ` Ferruh Yigit
  2020-10-29  4:00   ` [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf Zhang, Qi Z
  3 siblings, 1 reply; 39+ messages in thread
From: Wenzhuo Lu @ 2020-10-29  1:24 UTC (permalink / raw)
  To: dev; +Cc: Wenzhuo Lu, Bruce Richardson, Leyi Rong

To enhance the per-core performance, this patch adds some AVX512
instructions to the data path to handle the Tx descriptors.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 doc/guides/rel_notes/release_20_11.rst  |   3 +
 drivers/net/iavf/iavf_ethdev.c          |   3 +-
 drivers/net/iavf/iavf_rxtx.c            |  34 +++-
 drivers/net/iavf/iavf_rxtx.h            |   7 +
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 304 ++++++++++++++++++++++++++++++++
 5 files changed, 343 insertions(+), 8 deletions(-)

diff --git a/doc/guides/rel_notes/release_20_11.rst b/doc/guides/rel_notes/release_20_11.rst
index 89e0959..6d7c59d 100644
--- a/doc/guides/rel_notes/release_20_11.rst
+++ b/doc/guides/rel_notes/release_20_11.rst
@@ -20,6 +20,9 @@ DPDK Release 20.11
       make doc-guides-html
       xdg-open build/doc/html/guides/rel_notes/release_20_11.html
 
+   * **Added support of vector instructions on IAVF.**
+
+     Added support of AVX512 instructions in IAVF RX and TX path.
 
 New Features
 ------------
diff --git a/drivers/net/iavf/iavf_ethdev.c b/drivers/net/iavf/iavf_ethdev.c
index 6a67990..7e3c26a 100644
--- a/drivers/net/iavf/iavf_ethdev.c
+++ b/drivers/net/iavf/iavf_ethdev.c
@@ -792,7 +792,8 @@ static int iavf_config_rx_queues_irqs(struct rte_eth_dev *dev,
 		DEV_TX_OFFLOAD_GRE_TNL_TSO |
 		DEV_TX_OFFLOAD_IPIP_TNL_TSO |
 		DEV_TX_OFFLOAD_GENEVE_TNL_TSO |
-		DEV_TX_OFFLOAD_MULTI_SEGS;
+		DEV_TX_OFFLOAD_MULTI_SEGS |
+		DEV_TX_OFFLOAD_MBUF_FAST_FREE;
 
 	dev_info->default_rxconf = (struct rte_eth_rxconf) {
 		.rx_free_thresh = IAVF_DEFAULT_RX_FREE_THRESH,
diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index f51471f..baac5d6 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -2417,20 +2417,22 @@
 	struct iavf_tx_queue *txq;
 	int i;
 	bool use_avx2 = false;
+#ifdef CC_AVX512_SUPPORT
+	bool use_avx512 = false;
+#endif
 
 	if (!iavf_tx_vec_dev_check(dev) &&
 			rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
-		for (i = 0; i < dev->data->nb_tx_queues; i++) {
-			txq = dev->data->tx_queues[i];
-			if (!txq)
-				continue;
-			iavf_txq_vec_setup(txq);
-		}
-
 		if ((rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 		     rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&
 				rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
 			use_avx2 = true;
+#ifdef CC_AVX512_SUPPORT
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1 &&
+		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512BW) == 1 &&
+		    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512)
+			use_avx512 = true;
+#endif
 
 		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
 			    use_avx2 ? "avx2 " : "",
@@ -2438,8 +2440,26 @@
 		dev->tx_pkt_burst = use_avx2 ?
 				    iavf_xmit_pkts_vec_avx2 :
 				    iavf_xmit_pkts_vec;
+#ifdef CC_AVX512_SUPPORT
+		if (use_avx512)
+			dev->tx_pkt_burst = iavf_xmit_pkts_vec_avx512;
+#endif
 		dev->tx_pkt_prepare = NULL;
 
+		for (i = 0; i < dev->data->nb_tx_queues; i++) {
+			txq = dev->data->tx_queues[i];
+			if (!txq)
+				continue;
+#ifdef CC_AVX512_SUPPORT
+			if (use_avx512)
+				iavf_txq_vec_setup_avx512(txq);
+			else
+				iavf_txq_vec_setup(txq);
+#else
+			iavf_txq_vec_setup(txq);
+#endif
+		}
+
 		return;
 	}
 #endif
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 5bf91df..d4b4935 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -203,6 +203,10 @@ struct iavf_tx_entry {
 	uint16_t last_id;
 };
 
+struct iavf_tx_vec_entry {
+	struct rte_mbuf *mbuf;
+};
+
 /* Structure associated with each TX queue. */
 struct iavf_tx_queue {
 	const struct rte_memzone *mz;  /* memzone for Tx ring */
@@ -474,6 +478,9 @@ uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue,
 uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
 						      struct rte_mbuf **rx_pkts,
 						      uint16_t nb_pkts);
+uint16_t iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				   uint16_t nb_pkts);
+int iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq);
 
 uint8_t iavf_proto_xtr_type_to_rxdid(uint8_t xtr_type);
 
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index df0f43b..8680734 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -1396,3 +1396,307 @@
 	return retval + iavf_recv_scattered_burst_vec_avx512_flex_rxd(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+static __rte_always_inline int
+iavf_tx_free_bufs_avx512(struct iavf_tx_queue *txq)
+{
+	struct iavf_tx_vec_entry *txep;
+	uint32_t n;
+	uint32_t i;
+	int nb_free = 0;
+	struct rte_mbuf *m, *free[IAVF_VPMD_TX_MAX_FREE_BUF];
+
+	/* check DD bits on threshold descriptor */
+	if ((txq->tx_ring[txq->next_dd].cmd_type_offset_bsz &
+	     rte_cpu_to_le_64(IAVF_TXD_QW1_DTYPE_MASK)) !=
+	    rte_cpu_to_le_64(IAVF_TX_DESC_DTYPE_DESC_DONE))
+		return 0;
+
+	n = txq->rs_thresh;
+
+	 /* first buffer to free from S/W ring is at index
+	  * tx_next_dd - (tx_rs_thresh-1)
+	  */
+	txep = (void *)txq->sw_ring;
+	txep += txq->next_dd - (n - 1);
+
+	if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE && (n & 31) == 0) {
+		struct rte_mempool *mp = txep[0].mbuf->pool;
+		struct rte_mempool_cache *cache = rte_mempool_default_cache(mp,
+								rte_lcore_id());
+		void **cache_objs = &cache->objs[cache->len];
+
+		if (n > RTE_MEMPOOL_CACHE_MAX_SIZE) {
+			rte_mempool_ops_enqueue_bulk(mp, (void *)txep, n);
+			goto done;
+		}
+
+		/* The cache follows the following algorithm
+		 *   1. Add the objects to the cache
+		 *   2. Anything greater than the cache min value (if it crosses the
+		 *   cache flush threshold) is flushed to the ring.
+		 */
+		/* Add elements back into the cache */
+		uint32_t copied = 0;
+		/* n is multiple of 32 */
+		while (copied < n) {
+			const __m512i a = _mm512_loadu_si512(&txep[copied]);
+			const __m512i b = _mm512_loadu_si512(&txep[copied + 8]);
+			const __m512i c = _mm512_loadu_si512(&txep[copied + 16]);
+			const __m512i d = _mm512_loadu_si512(&txep[copied + 24]);
+
+			_mm512_storeu_si512(&cache_objs[copied], a);
+			_mm512_storeu_si512(&cache_objs[copied + 8], b);
+			_mm512_storeu_si512(&cache_objs[copied + 16], c);
+			_mm512_storeu_si512(&cache_objs[copied + 24], d);
+			copied += 32;
+		}
+		cache->len += n;
+
+		if (cache->len >= cache->flushthresh) {
+			rte_mempool_ops_enqueue_bulk(mp,
+						     &cache->objs[cache->size],
+						     cache->len - cache->size);
+			cache->len = cache->size;
+		}
+		goto done;
+	}
+
+	m = rte_pktmbuf_prefree_seg(txep[0].mbuf);
+	if (likely(m)) {
+		free[0] = m;
+		nb_free = 1;
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (likely(m)) {
+				if (likely(m->pool == free[0]->pool)) {
+					free[nb_free++] = m;
+				} else {
+					rte_mempool_put_bulk(free[0]->pool,
+							     (void *)free,
+							     nb_free);
+					free[0] = m;
+					nb_free = 1;
+				}
+			}
+		}
+		rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
+	} else {
+		for (i = 1; i < n; i++) {
+			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
+			if (m)
+				rte_mempool_put(m->pool, m);
+		}
+	}
+
+done:
+	/* buffers were freed, update counters */
+	txq->nb_free = (uint16_t)(txq->nb_free + txq->rs_thresh);
+	txq->next_dd = (uint16_t)(txq->next_dd + txq->rs_thresh);
+	if (txq->next_dd >= txq->nb_tx_desc)
+		txq->next_dd = (uint16_t)(txq->rs_thresh - 1);
+
+	return txq->rs_thresh;
+}
+
+static __rte_always_inline void
+tx_backlog_entry_avx512(struct iavf_tx_vec_entry *txep,
+			struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	int i;
+
+	for (i = 0; i < (int)nb_pkts; ++i)
+		txep[i].mbuf = tx_pkts[i];
+}
+
+static inline void
+iavf_vtx1(volatile struct iavf_tx_desc *txdp,
+	  struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(IAVF_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT) |
+		 ((uint64_t)pkt->data_len << IAVF_TXD_QW1_TX_BUF_SZ_SHIFT));
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+					    pkt->buf_iova + pkt->data_off);
+	_mm_storeu_si128((__m128i *)txdp, descriptor);
+}
+
+#define IAVF_TX_LEN_MASK 0xAA
+#define IAVF_TX_OFF_MASK 0x55
+static inline void
+iavf_vtx(volatile struct iavf_tx_desc *txdp,
+	 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (IAVF_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT));
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		iavf_vtx1(txdp, *pkt, flags);
+		nb_pkts--, txdp++, pkt++;
+	}
+
+	/* do 4 at a time while possible, in bursts */
+	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
+		__m512i desc4 =
+			_mm512_set_epi64
+				((uint64_t)pkt[3]->data_len,
+				 pkt[3]->buf_iova,
+				 (uint64_t)pkt[2]->data_len,
+				 pkt[2]->buf_iova,
+				 (uint64_t)pkt[1]->data_len,
+				 pkt[1]->buf_iova,
+				 (uint64_t)pkt[0]->data_len,
+				 pkt[0]->buf_iova);
+		__m512i hi_qw_tmpl_4 = _mm512_set1_epi64(hi_qw_tmpl);
+		__m512i data_off_4 =
+			_mm512_set_epi64
+				(0,
+				 pkt[3]->data_off,
+				 0,
+				 pkt[2]->data_off,
+				 0,
+				 pkt[1]->data_off,
+				 0,
+				 pkt[0]->data_off);
+
+		desc4 = _mm512_mask_slli_epi64(desc4, IAVF_TX_LEN_MASK, desc4,
+					       IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
+		desc4 = _mm512_mask_or_epi64(desc4, IAVF_TX_LEN_MASK, desc4,
+					     hi_qw_tmpl_4);
+		desc4 = _mm512_mask_add_epi64(desc4, IAVF_TX_OFF_MASK, desc4,
+					      data_off_4);
+		_mm512_storeu_si512((void *)txdp, desc4);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		iavf_vtx1(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+iavf_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				 uint16_t nb_pkts)
+{
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+	volatile struct iavf_tx_desc *txdp;
+	struct iavf_tx_vec_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	/* bit2 is reserved and must be set to 1 according to Spec */
+	uint64_t flags = IAVF_TX_DESC_CMD_EOP | IAVF_TX_DESC_CMD_ICRC;
+	uint64_t rs = IAVF_TX_DESC_CMD_RS | flags;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
+
+	if (txq->nb_free < txq->free_thresh)
+		iavf_tx_free_bufs_avx512(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = (void *)txq->sw_ring;
+	txep += tx_id;
+
+	txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		tx_backlog_entry_avx512(txep, tx_pkts, n);
+
+		iavf_vtx(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		iavf_vtx1(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = (void *)txq->sw_ring;
+		txep += tx_id;
+	}
+
+	tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
+
+	iavf_vtx(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->next_rs) {
+		txq->tx_ring[txq->next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)IAVF_TX_DESC_CMD_RS) <<
+					 IAVF_TXD_QW1_CMD_SHIFT);
+		txq->next_rs =
+			(uint16_t)(txq->next_rs + txq->rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	IAVF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+			  uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->rs_thresh);
+		ret = iavf_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;
+}
+
+static inline void
+iavf_tx_queue_release_mbufs_avx512(struct iavf_tx_queue *txq)
+{
+	unsigned int i;
+	const uint16_t max_desc = (uint16_t)(txq->nb_tx_desc - 1);
+	struct iavf_tx_vec_entry *swr = (void *)txq->sw_ring;
+
+	if (!txq->sw_ring || txq->nb_free == max_desc)
+		return;
+
+	i = txq->next_dd - txq->rs_thresh + 1;
+	if (txq->tx_tail < i) {
+		for (; i < txq->nb_tx_desc; i++) {
+			rte_pktmbuf_free_seg(swr[i].mbuf);
+			swr[i].mbuf = NULL;
+		}
+		i = 0;
+	}
+}
+
+static const struct iavf_txq_ops avx512_vec_txq_ops = {
+	.release_mbufs = iavf_tx_queue_release_mbufs_avx512,
+};
+
+int __rte_cold
+iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq)
+{
+	txq->ops = &avx512_vec_txq_ops;
+	return 0;
+}
-- 
1.9.3


^ permalink raw reply	[flat|nested] 39+ messages in thread

* Re: [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf
  2020-10-29  1:24 ` [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf Wenzhuo Lu
                     ` (2 preceding siblings ...)
  2020-10-29  1:24   ` [dpdk-dev] [PATCH v7 3/3] net/iavf: enable AVX512 for Tx Wenzhuo Lu
@ 2020-10-29  4:00   ` Zhang, Qi Z
  3 siblings, 0 replies; 39+ messages in thread
From: Zhang, Qi Z @ 2020-10-29  4:00 UTC (permalink / raw)
  To: Lu, Wenzhuo, dev; +Cc: Lu, Wenzhuo



> -----Original Message-----
> From: dev <dev-bounces@dpdk.org> On Behalf Of Wenzhuo Lu
> Sent: Thursday, October 29, 2020 9:24 AM
> To: dev@dpdk.org
> Cc: Lu, Wenzhuo <wenzhuo.lu@intel.com>
> Subject: [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf
> 
> AVX512 instructions is supported by more and more platforms. These
> instructions can be used in the data path to enhance the per-core performance
> of packet processing.
> Comparing with the existing implementation, this path set introduces some
> AVX512 instructions into the iavf data path, and we get a better per-code
> throughput.
> 
> v2:
> Update meson.build.
> Repalce the deprecated 'buf_physaddr' by 'buf_iova'.
> 
> v3:
> Fix compile errors.
> 
> v4:
> Fix wrong info in commnets.
> Trivial adjustment of the arrangement.
> 
> v5:
> Support "max SIMD bitwidth".
> 
> v6:
> Rework meson build to fix compile issue for AVX512BW.
> 
> v7:
> rebased on next-net_intel.
> 
> Wenzhuo Lu (3):
>   net/iavf: enable AVX512 for legacy Rx
>   net/iavf: enable AVX512 for flexible Rx
>   net/iavf: enable AVX512 for Tx
> 
>  doc/guides/rel_notes/release_20_11.rst  |    3 +
>  drivers/net/iavf/iavf_ethdev.c          |    3 +-
>  drivers/net/iavf/iavf_rxtx.c            |   73 +-
>  drivers/net/iavf/iavf_rxtx.h            |   18 +
>  drivers/net/iavf/iavf_rxtx_vec_avx512.c | 1702
> +++++++++++++++++++++++++++++++
>  drivers/net/iavf/meson.build            |   20 +
>  6 files changed, 1807 insertions(+), 12 deletions(-)  create mode 100644
> drivers/net/iavf/iavf_rxtx_vec_avx512.c
> 
> --
> 1.9.3

Acked-by: Qi Zhang <qi.z.zhang@intel.com>

Applied to dpdk-next-net-intel.

Thanks
Qi


^ permalink raw reply	[flat|nested] 39+ messages in thread

* Re: [dpdk-dev] [PATCH v7 3/3] net/iavf: enable AVX512 for Tx
  2020-10-29  1:24   ` [dpdk-dev] [PATCH v7 3/3] net/iavf: enable AVX512 for Tx Wenzhuo Lu
@ 2020-10-30 23:29     ` Ferruh Yigit
  0 siblings, 0 replies; 39+ messages in thread
From: Ferruh Yigit @ 2020-10-30 23:29 UTC (permalink / raw)
  To: Wenzhuo Lu, dev; +Cc: Bruce Richardson, Leyi Rong

On 10/29/2020 1:24 AM, Wenzhuo Lu wrote:
> To enhance the per-core performance, this patch adds some AVX512
> instructions to the data path to handle the Tx descriptors.
> 
> Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> Signed-off-by: Leyi Rong <leyi.rong@intel.com>
> ---
>   doc/guides/rel_notes/release_20_11.rst  |   3 +
>   drivers/net/iavf/iavf_ethdev.c          |   3 +-
>   drivers/net/iavf/iavf_rxtx.c            |  34 +++-
>   drivers/net/iavf/iavf_rxtx.h            |   7 +
>   drivers/net/iavf/iavf_rxtx_vec_avx512.c | 304 ++++++++++++++++++++++++++++++++
>   5 files changed, 343 insertions(+), 8 deletions(-)
> 
> diff --git a/doc/guides/rel_notes/release_20_11.rst b/doc/guides/rel_notes/release_20_11.rst
> index 89e0959..6d7c59d 100644
> --- a/doc/guides/rel_notes/release_20_11.rst
> +++ b/doc/guides/rel_notes/release_20_11.rst
> @@ -20,6 +20,9 @@ DPDK Release 20.11
>         make doc-guides-html
>         xdg-open build/doc/html/guides/rel_notes/release_20_11.html
>   
> +   * **Added support of vector instructions on IAVF.**
> +
> +     Added support of AVX512 instructions in IAVF RX and TX path.
>   

The documentation update is in the comment section, fixed in next-net.


^ permalink raw reply	[flat|nested] 39+ messages in thread

end of thread, other threads:[~2020-10-30 23:29 UTC | newest]

Thread overview: 39+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-09-10  5:59 [dpdk-dev] [PATCH 0/3] enable AVX512 for iavf Wenzhuo Lu
2020-09-10  5:59 ` [dpdk-dev] [PATCH 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
2020-09-10  9:29   ` Bruce Richardson
2020-09-11  3:06     ` Lu, Wenzhuo
2020-09-10  5:59 ` [dpdk-dev] [PATCH 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
2020-09-10  5:59 ` [dpdk-dev] [PATCH 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
2020-09-15  1:17   ` Wang, Haiyue
2020-09-17  1:29     ` Lu, Wenzhuo
2020-09-17  1:39 ` [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf Wenzhuo Lu
2020-09-17  1:39   ` [dpdk-dev] [PATCH v2 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
2020-09-17  1:39   ` [dpdk-dev] [PATCH v2 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
2020-09-17  1:39   ` [dpdk-dev] [PATCH v2 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
2020-09-17  7:37   ` [dpdk-dev] [PATCH v2 0/3] enable AVX512 for iavf Morten Brørup
2020-09-17  9:13     ` Bruce Richardson
2020-09-17  9:35       ` Morten Brørup
2020-09-21  8:13 ` [dpdk-dev] [PATCH v3 " Wenzhuo Lu
2020-09-21  8:13   ` [dpdk-dev] [PATCH v3 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
2020-09-21  8:13   ` [dpdk-dev] [PATCH v3 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
2020-09-21  8:13   ` [dpdk-dev] [PATCH v3 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
2020-09-21 19:10     ` Morten Brørup
2020-09-22  1:34       ` Lu, Wenzhuo
2020-09-27  1:30 ` [dpdk-dev] [PATCH v4 0/3] enable AVX512 for iavf Wenzhuo Lu
2020-09-27  1:30   ` [dpdk-dev] [PATCH v4 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
2020-09-27  1:30   ` [dpdk-dev] [PATCH v4 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
2020-09-27  1:30   ` [dpdk-dev] [PATCH v4 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
2020-10-21  7:47 ` [dpdk-dev] [PATCH v5 0/3] enable AVX512 for iavf Wenzhuo Lu
2020-10-21  7:47   ` [dpdk-dev] [PATCH v5 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
2020-10-21  7:47   ` [dpdk-dev] [PATCH v5 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
2020-10-21  7:47   ` [dpdk-dev] [PATCH v5 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
2020-10-28  5:14 ` [dpdk-dev] [PATCH v6 0/3] enable AVX512 for iavf Wenzhuo Lu
2020-10-28  5:14   ` [dpdk-dev] [PATCH v6 1/3] net/iavf: enable AVX512 for legacy RX Wenzhuo Lu
2020-10-28  5:14   ` [dpdk-dev] [PATCH v6 2/3] net/iavf: enable AVX512 for flexible RX Wenzhuo Lu
2020-10-28  5:15   ` [dpdk-dev] [PATCH v6 3/3] net/iavf: enable AVX512 for TX Wenzhuo Lu
2020-10-29  1:24 ` [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf Wenzhuo Lu
2020-10-29  1:24   ` [dpdk-dev] [PATCH v7 1/3] net/iavf: enable AVX512 for legacy Rx Wenzhuo Lu
2020-10-29  1:24   ` [dpdk-dev] [PATCH v7 2/3] net/iavf: enable AVX512 for flexible Rx Wenzhuo Lu
2020-10-29  1:24   ` [dpdk-dev] [PATCH v7 3/3] net/iavf: enable AVX512 for Tx Wenzhuo Lu
2020-10-30 23:29     ` Ferruh Yigit
2020-10-29  4:00   ` [dpdk-dev] [PATCH v7 0/3] enable AVX512 for iavf Zhang, Qi Z

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