DPDK patches and discussions
 help / color / mirror / Atom feed
* [dpdk-dev] [PATCH v1 0/2] AVX512 vPMD on ice
@ 2020-09-10  6:55 Leyi Rong
  2020-09-10  6:55 ` [dpdk-dev] [PATCH v1 1/2] net/ice: add AVX512 vector path Leyi Rong
                   ` (4 more replies)
  0 siblings, 5 replies; 25+ messages in thread
From: Leyi Rong @ 2020-09-10  6:55 UTC (permalink / raw)
  To: bruce.richardson, wenzhuo.lu, qi.z.zhang; +Cc: dev, Leyi Rong

This patchset aims to support AVX512 vPMD on ice.

Has dependency on http://patches.dpdk.org/cover/76096/
which support SIMD bitwidth selection properly.

Leyi Rong (2):
  net/ice: add AVX512 vector path
  net/ice: optimize Tx path on AVX512 vPMD

 drivers/net/ice/ice_rxtx.c            |  88 ++-
 drivers/net/ice/ice_rxtx.h            |  11 +
 drivers/net/ice/ice_rxtx_vec_avx512.c | 927 ++++++++++++++++++++++++++
 drivers/net/ice/ice_rxtx_vec_common.h |  36 +-
 drivers/net/ice/meson.build           |  13 +
 5 files changed, 1050 insertions(+), 25 deletions(-)
 create mode 100644 drivers/net/ice/ice_rxtx_vec_avx512.c

-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v1 1/2] net/ice: add AVX512 vector path
  2020-09-10  6:55 [dpdk-dev] [PATCH v1 0/2] AVX512 vPMD on ice Leyi Rong
@ 2020-09-10  6:55 ` Leyi Rong
  2020-09-10  9:32   ` Bruce Richardson
  2020-09-10  6:55 ` [dpdk-dev] [PATCH v1 2/2] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
                   ` (3 subsequent siblings)
  4 siblings, 1 reply; 25+ messages in thread
From: Leyi Rong @ 2020-09-10  6:55 UTC (permalink / raw)
  To: bruce.richardson, wenzhuo.lu, qi.z.zhang; +Cc: dev, Leyi Rong

Add AVX512 support for ice PMD. This patch adds ice_rxtx_vec_avx512.c
to support ice AVX512 vPMD.

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

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 drivers/net/ice/ice_rxtx.c            |  88 ++-
 drivers/net/ice/ice_rxtx.h            |   7 +
 drivers/net/ice/ice_rxtx_vec_avx512.c | 824 ++++++++++++++++++++++++++
 drivers/net/ice/meson.build           |  13 +
 4 files changed, 914 insertions(+), 18 deletions(-)
 create mode 100644 drivers/net/ice/ice_rxtx_vec_avx512.c

diff --git a/drivers/net/ice/ice_rxtx.c b/drivers/net/ice/ice_rxtx.c
index ccddae687..096bf098f 100644
--- a/drivers/net/ice/ice_rxtx.c
+++ b/drivers/net/ice/ice_rxtx.c
@@ -1852,6 +1852,10 @@ ice_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	if (dev->rx_pkt_burst == ice_recv_pkts_vec ||
 	    dev->rx_pkt_burst == ice_recv_scattered_pkts_vec ||
+#ifdef CC_AVX512_SUPPORT
+	    dev->rx_pkt_burst == ice_recv_pkts_vec_avx512 ||
+	    dev->rx_pkt_burst == ice_recv_scattered_pkts_vec_avx512 ||
+#endif
 	    dev->rx_pkt_burst == ice_recv_pkts_vec_avx2 ||
 	    dev->rx_pkt_burst == ice_recv_scattered_pkts_vec_avx2)
 		return ptypes;
@@ -2909,6 +2913,7 @@ ice_set_rx_function(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	struct ice_rx_queue *rxq;
 	int i;
+	bool use_avx512 = false;
 	bool use_avx2 = false;
 
 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
@@ -2924,10 +2929,18 @@ ice_set_rx_function(struct rte_eth_dev *dev)
 				}
 			}
 
-			if ((rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
+			if (rte_get_max_simd_bitwidth() >= RTE_MAX_512_SIMD &&
+			rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+#ifdef CC_AVX512_SUPPORT
+				use_avx512 = true;
+#else
+			PMD_DRV_LOG(NOTICE,
+				"AVX512 is not supported in build env");
+#endif
+			if (!use_avx512 &&
+			(rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 			rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&
-					rte_get_max_simd_bitwidth()
-					>= RTE_MAX_256_SIMD)
+			rte_get_max_simd_bitwidth() >= RTE_MAX_256_SIMD)
 				use_avx2 = true;
 
 		} else {
@@ -2937,20 +2950,37 @@ ice_set_rx_function(struct rte_eth_dev *dev)
 
 	if (ad->rx_vec_allowed) {
 		if (dev->data->scattered_rx) {
-			PMD_DRV_LOG(DEBUG,
+			if (use_avx512) {
+				PMD_DRV_LOG(NOTICE,
+					"Using AVX512 Vector Scattered Rx (port %d).",
+					dev->data->port_id);
+				dev->rx_pkt_burst =
+					ice_recv_scattered_pkts_vec_avx512;
+			} else {
+				PMD_DRV_LOG(DEBUG,
 					"Using %sVector Scattered Rx (port %d).",
 					use_avx2 ? "avx2 " : "",
 					dev->data->port_id);
-			dev->rx_pkt_burst = use_avx2 ?
+				dev->rx_pkt_burst = use_avx2 ?
 					ice_recv_scattered_pkts_vec_avx2 :
 					ice_recv_scattered_pkts_vec;
+			}
 		} else {
-			PMD_DRV_LOG(DEBUG, "Using %sVector Rx (port %d).",
+			if (use_avx512) {
+				PMD_DRV_LOG(NOTICE,
+					"Using AVX512 Vector Rx (port %d).",
+					dev->data->port_id);
+				dev->rx_pkt_burst =
+					ice_recv_pkts_vec_avx512;
+			} else {
+				PMD_DRV_LOG(DEBUG,
+					"Using %sVector Rx (port %d).",
 					use_avx2 ? "avx2 " : "",
 					dev->data->port_id);
-			dev->rx_pkt_burst = use_avx2 ?
-						ice_recv_pkts_vec_avx2 :
-						ice_recv_pkts_vec;
+				dev->rx_pkt_burst = use_avx2 ?
+					ice_recv_pkts_vec_avx2 :
+					ice_recv_pkts_vec;
+			}
 		}
 		return;
 	}
@@ -2987,6 +3017,10 @@ static const struct {
 	{ ice_recv_pkts_bulk_alloc,         "Scalar Bulk Alloc" },
 	{ ice_recv_pkts,                    "Scalar" },
 #ifdef RTE_ARCH_X86
+#ifdef CC_AVX512_SUPPORT
+	{ ice_recv_scattered_pkts_vec_avx512, "Vector AVX512 Scattered" },
+	{ ice_recv_pkts_vec_avx512,           "Vector AVX512" },
+#endif
 	{ ice_recv_scattered_pkts_vec_avx2, "Vector AVX2 Scattered" },
 	{ ice_recv_pkts_vec_avx2,           "Vector AVX2" },
 	{ ice_recv_scattered_pkts_vec,      "Vector SSE Scattered" },
@@ -3091,6 +3125,7 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	struct ice_tx_queue *txq;
 	int i;
+	bool use_avx512 = false;
 	bool use_avx2 = false;
 
 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
@@ -3106,10 +3141,18 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 				}
 			}
 
-			if ((rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
+			if (rte_get_max_simd_bitwidth() >= RTE_MAX_512_SIMD &&
+			rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+#ifdef CC_AVX512_SUPPORT
+				use_avx512 = true;
+#else
+			PMD_DRV_LOG(NOTICE,
+				"AVX512 is not supported in build env");
+#endif
+			if (!use_avx512 &&
+			(rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 			rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&
-					rte_get_max_simd_bitwidth()
-					>= RTE_MAX_256_SIMD)
+			rte_get_max_simd_bitwidth() >= RTE_MAX_256_SIMD)
 				use_avx2 = true;
 
 		} else {
@@ -3118,12 +3161,18 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 	}
 
 	if (ad->tx_vec_allowed) {
-		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
-			    use_avx2 ? "avx2 " : "",
-			    dev->data->port_id);
-		dev->tx_pkt_burst = use_avx2 ?
-				    ice_xmit_pkts_vec_avx2 :
-				    ice_xmit_pkts_vec;
+		if (use_avx512) {
+			PMD_DRV_LOG(NOTICE, "Using AVX512 Vector Tx (port %d).",
+				    dev->data->port_id);
+			dev->tx_pkt_burst = ice_xmit_pkts_vec_avx512;
+		} else {
+			PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
+				    use_avx2 ? "avx2 " : "",
+				    dev->data->port_id);
+			dev->tx_pkt_burst = use_avx2 ?
+					    ice_xmit_pkts_vec_avx2 :
+					    ice_xmit_pkts_vec;
+		}
 		dev->tx_pkt_prepare = NULL;
 
 		return;
@@ -3148,6 +3197,9 @@ static const struct {
 	{ ice_xmit_pkts_simple,   "Scalar Simple" },
 	{ ice_xmit_pkts,          "Scalar" },
 #ifdef RTE_ARCH_X86
+#ifdef CC_AVX512_SUPPORT
+	{ ice_xmit_pkts_vec_avx512, "Vector AVX512" },
+#endif
 	{ ice_xmit_pkts_vec_avx2, "Vector AVX2" },
 	{ ice_xmit_pkts_vec,      "Vector SSE" },
 #endif
diff --git a/drivers/net/ice/ice_rxtx.h b/drivers/net/ice/ice_rxtx.h
index 2fdcfb7d0..a39b41c05 100644
--- a/drivers/net/ice/ice_rxtx.h
+++ b/drivers/net/ice/ice_rxtx.h
@@ -200,6 +200,13 @@ uint16_t ice_recv_scattered_pkts_vec_avx2(void *rx_queue,
 					  uint16_t nb_pkts);
 uint16_t ice_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 				uint16_t nb_pkts);
+uint16_t ice_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				  uint16_t nb_pkts);
+uint16_t ice_recv_scattered_pkts_vec_avx512(void *rx_queue,
+					    struct rte_mbuf **rx_pkts,
+					    uint16_t nb_pkts);
+uint16_t ice_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				  uint16_t nb_pkts);
 int ice_fdir_programming(struct ice_pf *pf, struct ice_fltr_desc *fdir_desc);
 int ice_tx_done_cleanup(void *txq, uint32_t free_cnt);
 
diff --git a/drivers/net/ice/ice_rxtx_vec_avx512.c b/drivers/net/ice/ice_rxtx_vec_avx512.c
new file mode 100644
index 000000000..6a9d0a8ea
--- /dev/null
+++ b/drivers/net/ice/ice_rxtx_vec_avx512.c
@@ -0,0 +1,824 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2019 Intel Corporation
+ */
+
+#include "ice_rxtx_vec_common.h"
+
+#include <x86intrin.h>
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+#define ICE_DESCS_PER_LOOP_AVX 8
+
+static inline void
+ice_rxq_rearm(struct ice_rx_queue *rxq)
+{
+	int i;
+	uint16_t rx_id;
+	volatile union ice_rx_flex_desc *rxdp;
+	struct ice_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
+	struct rte_mempool_cache *cache = rte_mempool_default_cache(rxq->mp,
+			rte_lcore_id());
+
+	rxdp = rxq->rx_ring + rxq->rxrearm_start;
+
+	/* We need to pull 'n' more MBUFs into the software ring */
+	if (cache->len < ICE_RXQ_REARM_THRESH) {
+		uint32_t req = ICE_RXQ_REARM_THRESH + (cache->size -
+				cache->len);
+
+		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 + ICE_RXQ_REARM_THRESH >=
+			    rxq->nb_rx_desc) {
+				__m128i dma_addr0;
+
+				dma_addr0 = _mm_setzero_si128();
+				for (i = 0; i < ICE_DESCS_PER_LOOP; i++) {
+					rxep[i].mbuf = &rxq->fake_mbuf;
+					_mm_store_si128
+						((__m128i *)&rxdp[i].read,
+							dma_addr0);
+				}
+			}
+			rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+				ICE_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_ICE_16BYTE_RX_DESC
+	/* shuffle the iova into 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
+
+	/* fill up the rxd in vector, process 8 mbufs in one loop */
+	for (i = 0; i < ICE_RXQ_REARM_THRESH / 8; i++) {
+		const __m512i mbuf_ptrs = _mm512_loadu_si512
+			(&cache->objs[cache->len - 8]);
+		_mm512_store_si512(rxep, mbuf_ptrs);
+
+		/* gather iova of mbuf0-7 into one zmm reg */
+		const __m512i iova_base_addrs = _mm512_i64gather_epi64
+			(_mm512_add_epi64(mbuf_ptrs, iova_offsets),
+				0, /* base */
+				1  /* scale */);
+		const __m512i iova_addrs = _mm512_add_epi64(iova_base_addrs,
+				headroom);
+#ifndef RTE_LIBRTE_ICE_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 iova 2-3 in hdr_addr of desc 0-1
+		 * but these are ignored by driver since header split not
+		 * enabled. Similarly for desc 4 & 5.
+		 */
+		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_store_si512((void *)rxdp, desc0_1);
+		_mm512_store_si512((void *)(rxdp + 2), desc2_3);
+		_mm512_store_si512((void *)(rxdp + 4), desc4_5);
+		_mm512_store_si512((void *)(rxdp + 6), desc6_7);
+#else
+		/* permute leaves iova 4-7 in hdr_addr of desc 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_store_si512((void *)rxdp, desc0_3);
+		_mm512_store_si512((void *)(rxdp + 4), desc4_7);
+#endif
+		rxep += 8, rxdp += 8, cache->len -= 8;
+	}
+
+	rxq->rxrearm_start += ICE_RXQ_REARM_THRESH;
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= ICE_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 */
+	ICE_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
+}
+
+static inline uint16_t
+_ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
+			      struct rte_mbuf **rx_pkts,
+			      uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
+			0, rxq->mbuf_initializer);
+	struct ice_rx_entry *sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union ice_rx_flex_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to ICE_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, ICE_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 > ICE_RXQ_REARM_THRESH)
+		ice_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 << ICE_RX_FLEX_DESC_STATUS0_DD_S)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set4_epi32
+			(0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore non-length fields */
+			);
+
+	/* 8 packets DD mask, LSB in each 32-bit value */
+	const __m256i dd_check = _mm256_set1_epi32(1);
+
+	/* 8 packets EOP mask, second-LSB in each 32-bit value */
+	const __m256i eop_check = _mm256_slli_epi32(dd_check,
+			ICE_RX_DESC_STATUS_EOF_S);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set4_epi32
+			(/* octet 12~15, 32 bits rss */
+			 15 << 24 | 14 << 16 | 13 << 8 | 12,
+			 /* octet 10~11, 16 bits vlan_macip */
+			 /* octet 4~5, 16 bits data_len */
+			 11 << 24 | 10 << 16 | 5 << 8 | 4,
+			 /* skip hi 16 bits pkt_len, zero out */
+			 /* octet 4~5, 16 bits pkt_len */
+			 0xFFFF << 16 | 5 << 8 | 4,
+			 /* pkt_type set as unknown */
+			 0xFFFFFFFF
+			);
+
+	/**
+	 * compile-time check the above crc and shuffle layout is correct.
+	 * NOTE: the first field (lowest address) is given last in set_epi
+	 * calls above.
+	 */
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
+
+	/* Status/Error flag masks */
+	/**
+	 * mask everything except 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,
+			/* 2nd 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,
+			/* 2nd 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 += ICE_DESCS_PER_LOOP_AVX,
+	     rxdp += ICE_DESCS_PER_LOOP_AVX) {
+		/* step 1, copy over 8 mbuf pointers to rx_pkts array */
+		_mm256_storeu_si256((void *)&rx_pkts[i],
+				    _mm256_loadu_si256((void *)&sw_ring[i]));
+#ifdef RTE_ARCH_X86_64
+		_mm256_storeu_si256
+			((void *)&rx_pkts[i + 4],
+			 _mm256_loadu_si256((void *)&sw_ring[i + 4]));
+#endif
+
+		__m512i raw_desc0_3, raw_desc4_7;
+		__m256i raw_desc0_1, raw_desc2_3, raw_desc4_5, raw_desc6_7;
+
+		/* load in descriptors, in reverse order */
+		const __m128i raw_desc7 =
+			_mm_load_si128((void *)(rxdp + 7));
+		rte_compiler_barrier();
+		const __m128i raw_desc6 =
+			_mm_load_si128((void *)(rxdp + 6));
+		rte_compiler_barrier();
+		const __m128i raw_desc5 =
+			_mm_load_si128((void *)(rxdp + 5));
+		rte_compiler_barrier();
+		const __m128i raw_desc4 =
+			_mm_load_si128((void *)(rxdp + 4));
+		rte_compiler_barrier();
+		const __m128i raw_desc3 =
+			_mm_load_si128((void *)(rxdp + 3));
+		rte_compiler_barrier();
+		const __m128i raw_desc2 =
+			_mm_load_si128((void *)(rxdp + 2));
+		rte_compiler_barrier();
+		const __m128i raw_desc1 =
+			_mm_load_si128((void *)(rxdp + 1));
+		rte_compiler_barrier();
+		const __m128i raw_desc0 =
+			_mm_load_si128((void *)(rxdp + 0));
+
+		raw_desc6_7 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc6),
+				 raw_desc7, 1);
+		raw_desc4_5 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc4),
+				 raw_desc5, 1);
+		raw_desc2_3 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc2),
+				 raw_desc3, 1);
+		raw_desc0_1 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc0),
+				 raw_desc1, 1);
+
+		raw_desc4_7 =
+			_mm512_inserti64x4
+				(_mm512_castsi256_si512(raw_desc4_5),
+				 raw_desc6_7, 1);
+		raw_desc0_3 =
+			_mm512_inserti64x4
+				(_mm512_castsi256_si512(raw_desc0_1),
+				 raw_desc2_3, 1);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < ICE_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 0-7 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb4_7 = _mm512_shuffle_epi8(raw_desc4_7, shuf_msk);
+		__m512i mb0_3 = _mm512_shuffle_epi8(raw_desc0_3, shuf_msk);
+
+		mb4_7 = _mm512_add_epi32(mb4_7, crc_adjust);
+		mb0_3 = _mm512_add_epi32(mb0_3, crc_adjust);
+
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptype_mask =
+			_mm512_set1_epi16(ICE_RX_FLEX_DESC_PTYPE_M);
+
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptypes4_7 =
+			_mm512_and_si512(raw_desc4_7, ptype_mask);
+		const __m512i ptypes0_3 =
+			_mm512_and_si512(raw_desc0_3, ptype_mask);
+
+		const __m256i ptypes6_7 =
+			_mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 =
+			_mm512_extracti64x4_epi64(ptypes4_7, 0);
+		const __m256i ptypes2_3 =
+			_mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 =
+			_mm512_extracti64x4_epi64(ptypes0_3, 0);
+		const 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 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 ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, ptype_tbl[ptype7],
+			 0, 0, 0, ptype_tbl[ptype6],
+			 0, 0, 0, ptype_tbl[ptype5],
+			 0, 0, 0, ptype_tbl[ptype4]);
+		const __m512i ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, ptype_tbl[ptype3],
+			 0, 0, 0, ptype_tbl[ptype2],
+			 0, 0, 0, ptype_tbl[ptype1],
+			 0, 0, 0, ptype_tbl[ptype0]);
+
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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 */
+		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				rss_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;
+
+		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);
+
+		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);
+
+		/* 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 << ICE_RX_DESC_STATUS_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 += ICE_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 != ICE_DESCS_PER_LOOP_AVX)
+			break;
+	}
+
+	/* update tail pointers */
+	rxq->rx_tail += received;
+	rxq->rx_tail &= (rxq->nb_rx_desc - 1);
+	if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep avx2 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
+/**
+ * Notice:
+ * - nb_pkts < ICE_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+ice_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+			 uint16_t nb_pkts)
+{
+	return _ice_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 < ICE_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+ice_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				    uint16_t nb_pkts)
+{
+	struct ice_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[ICE_VPMD_RX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _ice_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 + ice_rx_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 < ICE_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+ice_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > ICE_VPMD_RX_BURST) {
+		uint16_t burst = ice_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, ICE_VPMD_RX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < ICE_VPMD_RX_BURST)
+			return retval;
+	}
+	return retval + ice_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
+
+static inline void
+ice_vtx1(volatile struct ice_tx_desc *txdp,
+	 struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(ICE_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << ICE_TXD_QW1_CMD_S) |
+		 ((uint64_t)pkt->data_len << ICE_TXD_QW1_TX_BUF_SZ_S));
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+				pkt->buf_physaddr + pkt->data_off);
+	_mm_store_si128((__m128i *)txdp, descriptor);
+}
+
+static inline void
+ice_vtx(volatile struct ice_tx_desc *txdp,
+	struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (ICE_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << ICE_TXD_QW1_CMD_S));
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		ice_vtx1(txdp, *pkt, flags);
+		nb_pkts--, txdp++, pkt++;
+	}
+
+	/* do two at a time while possible, in bursts */
+	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
+		uint64_t hi_qw3 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[3]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw2 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[2]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw1 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[1]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw0 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[0]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+
+		__m256i desc2_3 =
+			_mm256_set_epi64x
+				(hi_qw3,
+				 pkt[3]->buf_physaddr + pkt[3]->data_off,
+				 hi_qw2,
+				 pkt[2]->buf_physaddr + pkt[2]->data_off);
+		__m256i desc0_1 =
+			_mm256_set_epi64x
+				(hi_qw1,
+				 pkt[1]->buf_physaddr + pkt[1]->data_off,
+				 hi_qw0,
+				 pkt[0]->buf_physaddr + pkt[0]->data_off);
+		_mm256_store_si256((void *)(txdp + 2), desc2_3);
+		_mm256_store_si256((void *)txdp, desc0_1);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		ice_vtx1(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				uint16_t nb_pkts)
+{
+	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
+	volatile struct ice_tx_desc *txdp;
+	struct ice_tx_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	uint64_t flags = ICE_TD_CMD;
+	uint64_t rs = ICE_TX_DESC_CMD_RS | ICE_TD_CMD;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+
+	if (txq->nb_tx_free < txq->tx_free_thresh)
+		ice_tx_free_bufs(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = &txq->sw_ring[tx_id];
+
+	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		ice_tx_backlog_entry(txep, tx_pkts, n);
+
+		ice_vtx(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		ice_vtx1(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = &txq->sw_ring[tx_id];
+	}
+
+	ice_tx_backlog_entry(txep, tx_pkts, nb_commit);
+
+	ice_vtx(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->tx_next_rs) {
+		txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)ICE_TX_DESC_CMD_RS) <<
+					 ICE_TXD_QW1_CMD_S);
+		txq->tx_next_rs =
+			(uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	ICE_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+ice_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+			 uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+		ret = ice_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;
+}
diff --git a/drivers/net/ice/meson.build b/drivers/net/ice/meson.build
index e6fe74487..c0b23015e 100644
--- a/drivers/net/ice/meson.build
+++ b/drivers/net/ice/meson.build
@@ -33,6 +33,19 @@ if arch_subdir == 'x86'
 				c_args: [cflags, '-mavx2'])
 		objs += ice_avx2_lib.extract_objects('ice_rxtx_vec_avx2.c')
 	endif
+
+	if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F') or not machine_args.contains('-mno-avx512f')
+		if cc.has_argument('-mavx512f') and cc.has_argument('-mavx512bw')
+			cflags += ['-DCC_AVX512_SUPPORT']
+			ice_avx512_lib = static_library('ice_avx512_lib',
+					      'ice_rxtx_vec_avx512.c',
+					      dependencies: [static_rte_ethdev,
+						static_rte_kvargs, static_rte_hash],
+					      include_directories: includes,
+					      c_args: [cflags, '-march=skylake-avx512', '-mavx512f'])
+			objs += ice_avx512_lib.extract_objects('ice_rxtx_vec_avx512.c')
+		endif
+	endif
 endif
 
 sources += files('ice_dcf.c',
-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v1 2/2] net/ice: optimize Tx path on AVX512 vPMD
  2020-09-10  6:55 [dpdk-dev] [PATCH v1 0/2] AVX512 vPMD on ice Leyi Rong
  2020-09-10  6:55 ` [dpdk-dev] [PATCH v1 1/2] net/ice: add AVX512 vector path Leyi Rong
@ 2020-09-10  6:55 ` Leyi Rong
  2020-09-15  1:17   ` Wang, Haiyue
  2020-09-18  3:35 ` [dpdk-dev] [PATCH v2 0/3] AVX512 vPMD on ice Leyi Rong
                   ` (2 subsequent siblings)
  4 siblings, 1 reply; 25+ messages in thread
From: Leyi Rong @ 2020-09-10  6:55 UTC (permalink / raw)
  To: bruce.richardson, wenzhuo.lu, qi.z.zhang; +Cc: dev, Leyi Rong

Optimize Tx path by using AVX512 instructions and vectorize the
tx free bufs process.

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 drivers/net/ice/ice_rxtx.h            |   4 +
 drivers/net/ice/ice_rxtx_vec_avx512.c | 147 ++++++++++++++++++++++----
 drivers/net/ice/ice_rxtx_vec_common.h |  36 +++++--
 3 files changed, 158 insertions(+), 29 deletions(-)

diff --git a/drivers/net/ice/ice_rxtx.h b/drivers/net/ice/ice_rxtx.h
index a39b41c05..08084f5c5 100644
--- a/drivers/net/ice/ice_rxtx.h
+++ b/drivers/net/ice/ice_rxtx.h
@@ -88,6 +88,10 @@ struct ice_tx_entry {
 	uint16_t last_id;
 };
 
+struct ice_vec_tx_entry {
+	struct rte_mbuf *mbuf;
+};
+
 struct ice_tx_queue {
 	uint16_t nb_tx_desc; /* number of TX descriptors */
 	rte_iova_t tx_ring_dma; /* TX ring DMA address */
diff --git a/drivers/net/ice/ice_rxtx_vec_avx512.c b/drivers/net/ice/ice_rxtx_vec_avx512.c
index 6a9d0a8ea..1bc1191d0 100644
--- a/drivers/net/ice/ice_rxtx_vec_avx512.c
+++ b/drivers/net/ice/ice_rxtx_vec_avx512.c
@@ -665,6 +665,108 @@ ice_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
 				rx_pkts + retval, nb_pkts);
 }
 
+static __rte_always_inline int
+ice_tx_free_bufs_avx512(struct ice_tx_queue *txq)
+{
+	struct ice_vec_tx_entry *txep;
+	uint32_t n;
+	uint32_t i;
+	int nb_free = 0;
+	struct rte_mbuf *m, *free[ICE_TX_MAX_FREE_BUF_SZ];
+
+	/* check DD bits on threshold descriptor */
+	if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
+			rte_cpu_to_le_64(ICE_TXD_QW1_DTYPE_M)) !=
+			rte_cpu_to_le_64(ICE_TX_DESC_DTYPE_DESC_DONE))
+		return 0;
+
+	n = txq->tx_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->tx_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_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
+	txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
+	if (txq->tx_next_dd >= txq->nb_tx_desc)
+		txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
+
+	return txq->tx_rs_thresh;
+}
+
 static inline void
 ice_vtx1(volatile struct ice_tx_desc *txdp,
 	 struct rte_mbuf *pkt, uint64_t flags)
@@ -686,13 +788,6 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 	const uint64_t hi_qw_tmpl = (ICE_TX_DESC_DTYPE_DATA |
 			((uint64_t)flags  << ICE_TXD_QW1_CMD_S));
 
-	/* if unaligned on 32-bit boundary, do one to align */
-	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
-		ice_vtx1(txdp, *pkt, flags);
-		nb_pkts--, txdp++, pkt++;
-	}
-
-	/* do two at a time while possible, in bursts */
 	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
 		uint64_t hi_qw3 =
 			hi_qw_tmpl |
@@ -711,20 +806,17 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 			((uint64_t)pkt[0]->data_len <<
 			 ICE_TXD_QW1_TX_BUF_SZ_S);
 
-		__m256i desc2_3 =
-			_mm256_set_epi64x
+		__m512i desc0_3 =
+			_mm512_set_epi64
 				(hi_qw3,
 				 pkt[3]->buf_physaddr + pkt[3]->data_off,
 				 hi_qw2,
-				 pkt[2]->buf_physaddr + pkt[2]->data_off);
-		__m256i desc0_1 =
-			_mm256_set_epi64x
-				(hi_qw1,
+				 pkt[2]->buf_physaddr + pkt[2]->data_off,
+				 hi_qw1,
 				 pkt[1]->buf_physaddr + pkt[1]->data_off,
 				 hi_qw0,
 				 pkt[0]->buf_physaddr + pkt[0]->data_off);
-		_mm256_store_si256((void *)(txdp + 2), desc2_3);
-		_mm256_store_si256((void *)txdp, desc0_1);
+		_mm512_storeu_si512((void *)txdp, desc0_3);
 	}
 
 	/* do any last ones */
@@ -734,13 +826,23 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 	}
 }
 
+static __rte_always_inline void
+ice_tx_backlog_entry_avx512(struct ice_vec_tx_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 uint16_t
 ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 				uint16_t nb_pkts)
 {
 	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
 	volatile struct ice_tx_desc *txdp;
-	struct ice_tx_entry *txep;
+	struct ice_vec_tx_entry *txep;
 	uint16_t n, nb_commit, tx_id;
 	uint64_t flags = ICE_TD_CMD;
 	uint64_t rs = ICE_TX_DESC_CMD_RS | ICE_TD_CMD;
@@ -749,7 +851,7 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
 
 	if (txq->nb_tx_free < txq->tx_free_thresh)
-		ice_tx_free_bufs(txq);
+		ice_tx_free_bufs_avx512(txq);
 
 	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
 	if (unlikely(nb_pkts == 0))
@@ -757,13 +859,14 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 
 	tx_id = txq->tx_tail;
 	txdp = &txq->tx_ring[tx_id];
-	txep = &txq->sw_ring[tx_id];
+	txep = (void *)txq->sw_ring;
+	txep += tx_id;
 
 	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
 
 	n = (uint16_t)(txq->nb_tx_desc - tx_id);
 	if (nb_commit >= n) {
-		ice_tx_backlog_entry(txep, tx_pkts, n);
+		ice_tx_backlog_entry_avx512(txep, tx_pkts, n);
 
 		ice_vtx(txdp, tx_pkts, n - 1, flags);
 		tx_pkts += (n - 1);
@@ -777,11 +880,11 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 		txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
 
 		/* avoid reach the end of ring */
-		txdp = &txq->tx_ring[tx_id];
-		txep = &txq->sw_ring[tx_id];
+		txdp = txq->tx_ring;
+		txep = (void *)txq->sw_ring;
 	}
 
-	ice_tx_backlog_entry(txep, tx_pkts, nb_commit);
+	ice_tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
 
 	ice_vtx(txdp, tx_pkts, nb_commit, flags);
 
diff --git a/drivers/net/ice/ice_rxtx_vec_common.h b/drivers/net/ice/ice_rxtx_vec_common.h
index 46e3be98a..ee0bb1798 100644
--- a/drivers/net/ice/ice_rxtx_vec_common.h
+++ b/drivers/net/ice/ice_rxtx_vec_common.h
@@ -189,16 +189,38 @@ _ice_tx_queue_release_mbufs_vec(struct ice_tx_queue *txq)
 	 *  so need to free remains more carefully.
 	 */
 	i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
-	if (txq->tx_tail < i) {
-		for (; i < txq->nb_tx_desc; i++) {
+
+#ifdef CC_AVX512_SUPPORT
+	struct rte_eth_dev *dev = txq->vsi->adapter->eth_dev;
+
+	if (dev->tx_pkt_burst == ice_xmit_pkts_vec_avx512) {
+		struct ice_vec_tx_entry *swr = (void *)txq->sw_ring;
+
+		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;
+		}
+		for (; i < txq->tx_tail; i++) {
+			rte_pktmbuf_free_seg(swr[i].mbuf);
+			swr[i].mbuf = NULL;
+		}
+	} else
+#endif
+	{
+		if (txq->tx_tail < i) {
+			for (; i < txq->nb_tx_desc; i++) {
+				rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
+				txq->sw_ring[i].mbuf = NULL;
+			}
+			i = 0;
+		}
+		for (; i < txq->tx_tail; i++) {
 			rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
 			txq->sw_ring[i].mbuf = NULL;
 		}
-		i = 0;
-	}
-	for (; i < txq->tx_tail; i++) {
-		rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
-		txq->sw_ring[i].mbuf = NULL;
 	}
 }
 
-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* Re: [dpdk-dev] [PATCH v1 1/2] net/ice: add AVX512 vector path
  2020-09-10  6:55 ` [dpdk-dev] [PATCH v1 1/2] net/ice: add AVX512 vector path Leyi Rong
@ 2020-09-10  9:32   ` Bruce Richardson
  0 siblings, 0 replies; 25+ messages in thread
From: Bruce Richardson @ 2020-09-10  9:32 UTC (permalink / raw)
  To: Leyi Rong; +Cc: wenzhuo.lu, qi.z.zhang, dev

On Thu, Sep 10, 2020 at 02:55:03PM +0800, Leyi Rong wrote:
> Add AVX512 support for ice PMD. This patch adds ice_rxtx_vec_avx512.c
> to support ice AVX512 vPMD.
> 
> This patch aims to enable AVX512 on ice vPMD. Main changes are focus
> on Rx path compared with AVX2 vPMD.
> 
> Signed-off-by: Leyi Rong <leyi.rong@intel.com>
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> ---
>  drivers/net/ice/ice_rxtx.c            |  88 ++-
>  drivers/net/ice/ice_rxtx.h            |   7 +
>  drivers/net/ice/ice_rxtx_vec_avx512.c | 824 ++++++++++++++++++++++++++
>  drivers/net/ice/meson.build           |  13 +
>  4 files changed, 914 insertions(+), 18 deletions(-)
>  create mode 100644 drivers/net/ice/ice_rxtx_vec_avx512.c
<snip>
> +	if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F') or not machine_args.contains('-mno-avx512f')

I think you need to split these conditions:
If the cpu flag is set, you just want to unconditionally add the file to
the build list. However, if no-avx512 flag is set, you want to skip the
whole block completely.

> +		if cc.has_argument('-mavx512f') and cc.has_argument('-mavx512bw')
> +			cflags += ['-DCC_AVX512_SUPPORT']
> +			ice_avx512_lib = static_library('ice_avx512_lib',
> +					      'ice_rxtx_vec_avx512.c',
> +					      dependencies: [static_rte_ethdev,
> +						static_rte_kvargs, static_rte_hash],
> +					      include_directories: includes,
> +					      c_args: [cflags, '-march=skylake-avx512', '-mavx512f'])
> +			objs += ice_avx512_lib.extract_objects('ice_rxtx_vec_avx512.c')
> +		endif
> +	endif
>  endif
>  
>  sources += files('ice_dcf.c',
> -- 
> 2.17.1
> 

^ permalink raw reply	[flat|nested] 25+ messages in thread

* Re: [dpdk-dev] [PATCH v1 2/2] net/ice: optimize Tx path on AVX512 vPMD
  2020-09-10  6:55 ` [dpdk-dev] [PATCH v1 2/2] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
@ 2020-09-15  1:17   ` Wang, Haiyue
  0 siblings, 0 replies; 25+ messages in thread
From: Wang, Haiyue @ 2020-09-15  1:17 UTC (permalink / raw)
  To: Rong, Leyi, Richardson, Bruce, Lu, Wenzhuo, Zhang, Qi Z; +Cc: dev, Rong, Leyi

> -----Original Message-----
> From: dev <dev-bounces@dpdk.org> On Behalf Of Leyi Rong
> Sent: Thursday, September 10, 2020 14:55
> To: Richardson, Bruce <bruce.richardson@intel.com>; Lu, Wenzhuo <wenzhuo.lu@intel.com>; Zhang, Qi Z
> <qi.z.zhang@intel.com>
> Cc: dev@dpdk.org; Rong, Leyi <leyi.rong@intel.com>
> Subject: [dpdk-dev] [PATCH v1 2/2] net/ice: optimize Tx path on AVX512 vPMD
> 
> Optimize Tx path by using AVX512 instructions and vectorize the
> tx free bufs process.
> 
> Signed-off-by: Leyi Rong <leyi.rong@intel.com>
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> ---
>  drivers/net/ice/ice_rxtx.h            |   4 +
>  drivers/net/ice/ice_rxtx_vec_avx512.c | 147 ++++++++++++++++++++++----
>  drivers/net/ice/ice_rxtx_vec_common.h |  36 +++++--
>  3 files changed, 158 insertions(+), 29 deletions(-)
> 


> -			_mm256_set_epi64x
> +		__m512i desc0_3 =
> +			_mm512_set_epi64
>  				(hi_qw3,
>  				 pkt[3]->buf_physaddr + pkt[3]->data_off,
>  				 hi_qw2,
> -				 pkt[2]->buf_physaddr + pkt[2]->data_off);
> -		__m256i desc0_1 =
> -			_mm256_set_epi64x
> -				(hi_qw1,
> +				 pkt[2]->buf_physaddr + pkt[2]->data_off,
> +				 hi_qw1,
>  				 pkt[1]->buf_physaddr + pkt[1]->data_off,
>  				 hi_qw0,
>  				 pkt[0]->buf_physaddr + pkt[0]->data_off);

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

> --
> 2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v2 0/3] AVX512 vPMD on ice
  2020-09-10  6:55 [dpdk-dev] [PATCH v1 0/2] AVX512 vPMD on ice Leyi Rong
  2020-09-10  6:55 ` [dpdk-dev] [PATCH v1 1/2] net/ice: add AVX512 vector path Leyi Rong
  2020-09-10  6:55 ` [dpdk-dev] [PATCH v1 2/2] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
@ 2020-09-18  3:35 ` Leyi Rong
  2020-09-18  3:35   ` [dpdk-dev] [PATCH v2 1/3] net/ice: add AVX512 vector path Leyi Rong
                     ` (2 more replies)
  2020-10-20 10:51 ` [dpdk-dev] [PATCH v3 0/3] AVX512 vPMD on ice Leyi Rong
  2020-10-23  4:14 ` [dpdk-dev] [PATCH v4 0/3] AVX512 vPMD on ice Leyi Rong
  4 siblings, 3 replies; 25+ messages in thread
From: Leyi Rong @ 2020-09-18  3:35 UTC (permalink / raw)
  To: qi.z.zhang, wenzhuo.lu, burce.richardson; +Cc: dev, Leyi Rong

This patchset aims to support AVX512 vPMD on ice.

Has dependency on http://patches.dpdk.org/cover/76096/
which support SIMD bitwidth selection properly.

---
v2:
- No internal judgement when RTE_MACHINE_CPUFLAG_AVX512F is set in
  meson.build.
- Add RSS hash parsing as default RXDID is set to #22.
- Use buf_iova instead of buf_physaddr, as buf_physaddr will be removed.

Leyi Rong (3):
  net/ice: add AVX512 vector path
  net/ice: add RSS hash parsing in AVX512 path
  net/ice: optimize Tx path on AVX512 vPMD

 drivers/net/ice/ice_rxtx.c            |   88 ++-
 drivers/net/ice/ice_rxtx.h            |   11 +
 drivers/net/ice/ice_rxtx_vec_avx512.c | 1018 +++++++++++++++++++++++++
 drivers/net/ice/ice_rxtx_vec_common.h |   36 +-
 drivers/net/ice/meson.build           |   11 +
 5 files changed, 1139 insertions(+), 25 deletions(-)
 create mode 100644 drivers/net/ice/ice_rxtx_vec_avx512.c

-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v2 1/3] net/ice: add AVX512 vector path
  2020-09-18  3:35 ` [dpdk-dev] [PATCH v2 0/3] AVX512 vPMD on ice Leyi Rong
@ 2020-09-18  3:35   ` Leyi Rong
  2020-09-18  3:35   ` [dpdk-dev] [PATCH v2 2/3] net/ice: add RSS hash parsing in AVX512 path Leyi Rong
  2020-09-18  3:35   ` [dpdk-dev] [PATCH v2 3/3] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
  2 siblings, 0 replies; 25+ messages in thread
From: Leyi Rong @ 2020-09-18  3:35 UTC (permalink / raw)
  To: qi.z.zhang, wenzhuo.lu, burce.richardson; +Cc: dev, Leyi Rong, Bruce Richardson

Add AVX512 support for ice PMD. This patch adds ice_rxtx_vec_avx512.c
to support ice AVX512 vPMD.

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

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 drivers/net/ice/ice_rxtx.c            |  88 ++-
 drivers/net/ice/ice_rxtx.h            |   7 +
 drivers/net/ice/ice_rxtx_vec_avx512.c | 824 ++++++++++++++++++++++++++
 drivers/net/ice/meson.build           |  11 +
 4 files changed, 912 insertions(+), 18 deletions(-)
 create mode 100644 drivers/net/ice/ice_rxtx_vec_avx512.c

diff --git a/drivers/net/ice/ice_rxtx.c b/drivers/net/ice/ice_rxtx.c
index 5a29af743c..cc7946c0b7 100644
--- a/drivers/net/ice/ice_rxtx.c
+++ b/drivers/net/ice/ice_rxtx.c
@@ -1876,6 +1876,10 @@ ice_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	if (dev->rx_pkt_burst == ice_recv_pkts_vec ||
 	    dev->rx_pkt_burst == ice_recv_scattered_pkts_vec ||
+#ifdef CC_AVX512_SUPPORT
+	    dev->rx_pkt_burst == ice_recv_pkts_vec_avx512 ||
+	    dev->rx_pkt_burst == ice_recv_scattered_pkts_vec_avx512 ||
+#endif
 	    dev->rx_pkt_burst == ice_recv_pkts_vec_avx2 ||
 	    dev->rx_pkt_burst == ice_recv_scattered_pkts_vec_avx2)
 		return ptypes;
@@ -2933,6 +2937,7 @@ ice_set_rx_function(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	struct ice_rx_queue *rxq;
 	int i;
+	bool use_avx512 = false;
 	bool use_avx2 = false;
 
 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
@@ -2948,10 +2953,18 @@ ice_set_rx_function(struct rte_eth_dev *dev)
 				}
 			}
 
-			if ((rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
+			if (rte_get_max_simd_bitwidth() >= RTE_MAX_512_SIMD &&
+			rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+#ifdef CC_AVX512_SUPPORT
+				use_avx512 = true;
+#else
+			PMD_DRV_LOG(NOTICE,
+				"AVX512 is not supported in build env");
+#endif
+			if (!use_avx512 &&
+			(rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 			rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&
-					rte_get_max_simd_bitwidth()
-					>= RTE_MAX_256_SIMD)
+			rte_get_max_simd_bitwidth() >= RTE_MAX_256_SIMD)
 				use_avx2 = true;
 
 		} else {
@@ -2961,20 +2974,37 @@ ice_set_rx_function(struct rte_eth_dev *dev)
 
 	if (ad->rx_vec_allowed) {
 		if (dev->data->scattered_rx) {
-			PMD_DRV_LOG(DEBUG,
+			if (use_avx512) {
+				PMD_DRV_LOG(NOTICE,
+					"Using AVX512 Vector Scattered Rx (port %d).",
+					dev->data->port_id);
+				dev->rx_pkt_burst =
+					ice_recv_scattered_pkts_vec_avx512;
+			} else {
+				PMD_DRV_LOG(DEBUG,
 					"Using %sVector Scattered Rx (port %d).",
 					use_avx2 ? "avx2 " : "",
 					dev->data->port_id);
-			dev->rx_pkt_burst = use_avx2 ?
+				dev->rx_pkt_burst = use_avx2 ?
 					ice_recv_scattered_pkts_vec_avx2 :
 					ice_recv_scattered_pkts_vec;
+			}
 		} else {
-			PMD_DRV_LOG(DEBUG, "Using %sVector Rx (port %d).",
+			if (use_avx512) {
+				PMD_DRV_LOG(NOTICE,
+					"Using AVX512 Vector Rx (port %d).",
+					dev->data->port_id);
+				dev->rx_pkt_burst =
+					ice_recv_pkts_vec_avx512;
+			} else {
+				PMD_DRV_LOG(DEBUG,
+					"Using %sVector Rx (port %d).",
 					use_avx2 ? "avx2 " : "",
 					dev->data->port_id);
-			dev->rx_pkt_burst = use_avx2 ?
-						ice_recv_pkts_vec_avx2 :
-						ice_recv_pkts_vec;
+				dev->rx_pkt_burst = use_avx2 ?
+					ice_recv_pkts_vec_avx2 :
+					ice_recv_pkts_vec;
+			}
 		}
 		return;
 	}
@@ -3011,6 +3041,10 @@ static const struct {
 	{ ice_recv_pkts_bulk_alloc,         "Scalar Bulk Alloc" },
 	{ ice_recv_pkts,                    "Scalar" },
 #ifdef RTE_ARCH_X86
+#ifdef CC_AVX512_SUPPORT
+	{ ice_recv_scattered_pkts_vec_avx512, "Vector AVX512 Scattered" },
+	{ ice_recv_pkts_vec_avx512,           "Vector AVX512" },
+#endif
 	{ ice_recv_scattered_pkts_vec_avx2, "Vector AVX2 Scattered" },
 	{ ice_recv_pkts_vec_avx2,           "Vector AVX2" },
 	{ ice_recv_scattered_pkts_vec,      "Vector SSE Scattered" },
@@ -3115,6 +3149,7 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	struct ice_tx_queue *txq;
 	int i;
+	bool use_avx512 = false;
 	bool use_avx2 = false;
 
 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
@@ -3130,10 +3165,18 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 				}
 			}
 
-			if ((rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
+			if (rte_get_max_simd_bitwidth() >= RTE_MAX_512_SIMD &&
+			rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+#ifdef CC_AVX512_SUPPORT
+				use_avx512 = true;
+#else
+			PMD_DRV_LOG(NOTICE,
+				"AVX512 is not supported in build env");
+#endif
+			if (!use_avx512 &&
+			(rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
 			rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&
-					rte_get_max_simd_bitwidth()
-					>= RTE_MAX_256_SIMD)
+			rte_get_max_simd_bitwidth() >= RTE_MAX_256_SIMD)
 				use_avx2 = true;
 
 		} else {
@@ -3142,12 +3185,18 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 	}
 
 	if (ad->tx_vec_allowed) {
-		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
-			    use_avx2 ? "avx2 " : "",
-			    dev->data->port_id);
-		dev->tx_pkt_burst = use_avx2 ?
-				    ice_xmit_pkts_vec_avx2 :
-				    ice_xmit_pkts_vec;
+		if (use_avx512) {
+			PMD_DRV_LOG(NOTICE, "Using AVX512 Vector Tx (port %d).",
+				    dev->data->port_id);
+			dev->tx_pkt_burst = ice_xmit_pkts_vec_avx512;
+		} else {
+			PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
+				    use_avx2 ? "avx2 " : "",
+				    dev->data->port_id);
+			dev->tx_pkt_burst = use_avx2 ?
+					    ice_xmit_pkts_vec_avx2 :
+					    ice_xmit_pkts_vec;
+		}
 		dev->tx_pkt_prepare = NULL;
 
 		return;
@@ -3172,6 +3221,9 @@ static const struct {
 	{ ice_xmit_pkts_simple,   "Scalar Simple" },
 	{ ice_xmit_pkts,          "Scalar" },
 #ifdef RTE_ARCH_X86
+#ifdef CC_AVX512_SUPPORT
+	{ ice_xmit_pkts_vec_avx512, "Vector AVX512" },
+#endif
 	{ ice_xmit_pkts_vec_avx2, "Vector AVX2" },
 	{ ice_xmit_pkts_vec,      "Vector SSE" },
 #endif
diff --git a/drivers/net/ice/ice_rxtx.h b/drivers/net/ice/ice_rxtx.h
index 9fa57b3b27..c13e54b251 100644
--- a/drivers/net/ice/ice_rxtx.h
+++ b/drivers/net/ice/ice_rxtx.h
@@ -243,6 +243,13 @@ uint16_t ice_recv_scattered_pkts_vec_avx2(void *rx_queue,
 					  uint16_t nb_pkts);
 uint16_t ice_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 				uint16_t nb_pkts);
+uint16_t ice_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				  uint16_t nb_pkts);
+uint16_t ice_recv_scattered_pkts_vec_avx512(void *rx_queue,
+					    struct rte_mbuf **rx_pkts,
+					    uint16_t nb_pkts);
+uint16_t ice_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				  uint16_t nb_pkts);
 int ice_fdir_programming(struct ice_pf *pf, struct ice_fltr_desc *fdir_desc);
 int ice_tx_done_cleanup(void *txq, uint32_t free_cnt);
 
diff --git a/drivers/net/ice/ice_rxtx_vec_avx512.c b/drivers/net/ice/ice_rxtx_vec_avx512.c
new file mode 100644
index 0000000000..6a9d0a8eaa
--- /dev/null
+++ b/drivers/net/ice/ice_rxtx_vec_avx512.c
@@ -0,0 +1,824 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2019 Intel Corporation
+ */
+
+#include "ice_rxtx_vec_common.h"
+
+#include <x86intrin.h>
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+#define ICE_DESCS_PER_LOOP_AVX 8
+
+static inline void
+ice_rxq_rearm(struct ice_rx_queue *rxq)
+{
+	int i;
+	uint16_t rx_id;
+	volatile union ice_rx_flex_desc *rxdp;
+	struct ice_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
+	struct rte_mempool_cache *cache = rte_mempool_default_cache(rxq->mp,
+			rte_lcore_id());
+
+	rxdp = rxq->rx_ring + rxq->rxrearm_start;
+
+	/* We need to pull 'n' more MBUFs into the software ring */
+	if (cache->len < ICE_RXQ_REARM_THRESH) {
+		uint32_t req = ICE_RXQ_REARM_THRESH + (cache->size -
+				cache->len);
+
+		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 + ICE_RXQ_REARM_THRESH >=
+			    rxq->nb_rx_desc) {
+				__m128i dma_addr0;
+
+				dma_addr0 = _mm_setzero_si128();
+				for (i = 0; i < ICE_DESCS_PER_LOOP; i++) {
+					rxep[i].mbuf = &rxq->fake_mbuf;
+					_mm_store_si128
+						((__m128i *)&rxdp[i].read,
+							dma_addr0);
+				}
+			}
+			rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+				ICE_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_ICE_16BYTE_RX_DESC
+	/* shuffle the iova into 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
+
+	/* fill up the rxd in vector, process 8 mbufs in one loop */
+	for (i = 0; i < ICE_RXQ_REARM_THRESH / 8; i++) {
+		const __m512i mbuf_ptrs = _mm512_loadu_si512
+			(&cache->objs[cache->len - 8]);
+		_mm512_store_si512(rxep, mbuf_ptrs);
+
+		/* gather iova of mbuf0-7 into one zmm reg */
+		const __m512i iova_base_addrs = _mm512_i64gather_epi64
+			(_mm512_add_epi64(mbuf_ptrs, iova_offsets),
+				0, /* base */
+				1  /* scale */);
+		const __m512i iova_addrs = _mm512_add_epi64(iova_base_addrs,
+				headroom);
+#ifndef RTE_LIBRTE_ICE_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 iova 2-3 in hdr_addr of desc 0-1
+		 * but these are ignored by driver since header split not
+		 * enabled. Similarly for desc 4 & 5.
+		 */
+		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_store_si512((void *)rxdp, desc0_1);
+		_mm512_store_si512((void *)(rxdp + 2), desc2_3);
+		_mm512_store_si512((void *)(rxdp + 4), desc4_5);
+		_mm512_store_si512((void *)(rxdp + 6), desc6_7);
+#else
+		/* permute leaves iova 4-7 in hdr_addr of desc 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_store_si512((void *)rxdp, desc0_3);
+		_mm512_store_si512((void *)(rxdp + 4), desc4_7);
+#endif
+		rxep += 8, rxdp += 8, cache->len -= 8;
+	}
+
+	rxq->rxrearm_start += ICE_RXQ_REARM_THRESH;
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= ICE_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 */
+	ICE_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
+}
+
+static inline uint16_t
+_ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
+			      struct rte_mbuf **rx_pkts,
+			      uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
+			0, rxq->mbuf_initializer);
+	struct ice_rx_entry *sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union ice_rx_flex_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to ICE_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, ICE_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 > ICE_RXQ_REARM_THRESH)
+		ice_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 << ICE_RX_FLEX_DESC_STATUS0_DD_S)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set4_epi32
+			(0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore non-length fields */
+			);
+
+	/* 8 packets DD mask, LSB in each 32-bit value */
+	const __m256i dd_check = _mm256_set1_epi32(1);
+
+	/* 8 packets EOP mask, second-LSB in each 32-bit value */
+	const __m256i eop_check = _mm256_slli_epi32(dd_check,
+			ICE_RX_DESC_STATUS_EOF_S);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set4_epi32
+			(/* octet 12~15, 32 bits rss */
+			 15 << 24 | 14 << 16 | 13 << 8 | 12,
+			 /* octet 10~11, 16 bits vlan_macip */
+			 /* octet 4~5, 16 bits data_len */
+			 11 << 24 | 10 << 16 | 5 << 8 | 4,
+			 /* skip hi 16 bits pkt_len, zero out */
+			 /* octet 4~5, 16 bits pkt_len */
+			 0xFFFF << 16 | 5 << 8 | 4,
+			 /* pkt_type set as unknown */
+			 0xFFFFFFFF
+			);
+
+	/**
+	 * compile-time check the above crc and shuffle layout is correct.
+	 * NOTE: the first field (lowest address) is given last in set_epi
+	 * calls above.
+	 */
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
+
+	/* Status/Error flag masks */
+	/**
+	 * mask everything except 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,
+			/* 2nd 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,
+			/* 2nd 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 += ICE_DESCS_PER_LOOP_AVX,
+	     rxdp += ICE_DESCS_PER_LOOP_AVX) {
+		/* step 1, copy over 8 mbuf pointers to rx_pkts array */
+		_mm256_storeu_si256((void *)&rx_pkts[i],
+				    _mm256_loadu_si256((void *)&sw_ring[i]));
+#ifdef RTE_ARCH_X86_64
+		_mm256_storeu_si256
+			((void *)&rx_pkts[i + 4],
+			 _mm256_loadu_si256((void *)&sw_ring[i + 4]));
+#endif
+
+		__m512i raw_desc0_3, raw_desc4_7;
+		__m256i raw_desc0_1, raw_desc2_3, raw_desc4_5, raw_desc6_7;
+
+		/* load in descriptors, in reverse order */
+		const __m128i raw_desc7 =
+			_mm_load_si128((void *)(rxdp + 7));
+		rte_compiler_barrier();
+		const __m128i raw_desc6 =
+			_mm_load_si128((void *)(rxdp + 6));
+		rte_compiler_barrier();
+		const __m128i raw_desc5 =
+			_mm_load_si128((void *)(rxdp + 5));
+		rte_compiler_barrier();
+		const __m128i raw_desc4 =
+			_mm_load_si128((void *)(rxdp + 4));
+		rte_compiler_barrier();
+		const __m128i raw_desc3 =
+			_mm_load_si128((void *)(rxdp + 3));
+		rte_compiler_barrier();
+		const __m128i raw_desc2 =
+			_mm_load_si128((void *)(rxdp + 2));
+		rte_compiler_barrier();
+		const __m128i raw_desc1 =
+			_mm_load_si128((void *)(rxdp + 1));
+		rte_compiler_barrier();
+		const __m128i raw_desc0 =
+			_mm_load_si128((void *)(rxdp + 0));
+
+		raw_desc6_7 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc6),
+				 raw_desc7, 1);
+		raw_desc4_5 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc4),
+				 raw_desc5, 1);
+		raw_desc2_3 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc2),
+				 raw_desc3, 1);
+		raw_desc0_1 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc0),
+				 raw_desc1, 1);
+
+		raw_desc4_7 =
+			_mm512_inserti64x4
+				(_mm512_castsi256_si512(raw_desc4_5),
+				 raw_desc6_7, 1);
+		raw_desc0_3 =
+			_mm512_inserti64x4
+				(_mm512_castsi256_si512(raw_desc0_1),
+				 raw_desc2_3, 1);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < ICE_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 0-7 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb4_7 = _mm512_shuffle_epi8(raw_desc4_7, shuf_msk);
+		__m512i mb0_3 = _mm512_shuffle_epi8(raw_desc0_3, shuf_msk);
+
+		mb4_7 = _mm512_add_epi32(mb4_7, crc_adjust);
+		mb0_3 = _mm512_add_epi32(mb0_3, crc_adjust);
+
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptype_mask =
+			_mm512_set1_epi16(ICE_RX_FLEX_DESC_PTYPE_M);
+
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptypes4_7 =
+			_mm512_and_si512(raw_desc4_7, ptype_mask);
+		const __m512i ptypes0_3 =
+			_mm512_and_si512(raw_desc0_3, ptype_mask);
+
+		const __m256i ptypes6_7 =
+			_mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 =
+			_mm512_extracti64x4_epi64(ptypes4_7, 0);
+		const __m256i ptypes2_3 =
+			_mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 =
+			_mm512_extracti64x4_epi64(ptypes0_3, 0);
+		const 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 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 ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, ptype_tbl[ptype7],
+			 0, 0, 0, ptype_tbl[ptype6],
+			 0, 0, 0, ptype_tbl[ptype5],
+			 0, 0, 0, ptype_tbl[ptype4]);
+		const __m512i ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, ptype_tbl[ptype3],
+			 0, 0, 0, ptype_tbl[ptype2],
+			 0, 0, 0, ptype_tbl[ptype1],
+			 0, 0, 0, ptype_tbl[ptype0]);
+
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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 */
+		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				rss_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;
+
+		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);
+
+		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);
+
+		/* 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 << ICE_RX_DESC_STATUS_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 += ICE_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 != ICE_DESCS_PER_LOOP_AVX)
+			break;
+	}
+
+	/* update tail pointers */
+	rxq->rx_tail += received;
+	rxq->rx_tail &= (rxq->nb_rx_desc - 1);
+	if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep avx2 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
+/**
+ * Notice:
+ * - nb_pkts < ICE_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+ice_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+			 uint16_t nb_pkts)
+{
+	return _ice_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 < ICE_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+ice_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				    uint16_t nb_pkts)
+{
+	struct ice_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[ICE_VPMD_RX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _ice_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 + ice_rx_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 < ICE_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+ice_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > ICE_VPMD_RX_BURST) {
+		uint16_t burst = ice_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, ICE_VPMD_RX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < ICE_VPMD_RX_BURST)
+			return retval;
+	}
+	return retval + ice_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
+
+static inline void
+ice_vtx1(volatile struct ice_tx_desc *txdp,
+	 struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(ICE_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << ICE_TXD_QW1_CMD_S) |
+		 ((uint64_t)pkt->data_len << ICE_TXD_QW1_TX_BUF_SZ_S));
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+				pkt->buf_physaddr + pkt->data_off);
+	_mm_store_si128((__m128i *)txdp, descriptor);
+}
+
+static inline void
+ice_vtx(volatile struct ice_tx_desc *txdp,
+	struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (ICE_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << ICE_TXD_QW1_CMD_S));
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		ice_vtx1(txdp, *pkt, flags);
+		nb_pkts--, txdp++, pkt++;
+	}
+
+	/* do two at a time while possible, in bursts */
+	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
+		uint64_t hi_qw3 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[3]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw2 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[2]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw1 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[1]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw0 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[0]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+
+		__m256i desc2_3 =
+			_mm256_set_epi64x
+				(hi_qw3,
+				 pkt[3]->buf_physaddr + pkt[3]->data_off,
+				 hi_qw2,
+				 pkt[2]->buf_physaddr + pkt[2]->data_off);
+		__m256i desc0_1 =
+			_mm256_set_epi64x
+				(hi_qw1,
+				 pkt[1]->buf_physaddr + pkt[1]->data_off,
+				 hi_qw0,
+				 pkt[0]->buf_physaddr + pkt[0]->data_off);
+		_mm256_store_si256((void *)(txdp + 2), desc2_3);
+		_mm256_store_si256((void *)txdp, desc0_1);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		ice_vtx1(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				uint16_t nb_pkts)
+{
+	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
+	volatile struct ice_tx_desc *txdp;
+	struct ice_tx_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	uint64_t flags = ICE_TD_CMD;
+	uint64_t rs = ICE_TX_DESC_CMD_RS | ICE_TD_CMD;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+
+	if (txq->nb_tx_free < txq->tx_free_thresh)
+		ice_tx_free_bufs(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = &txq->sw_ring[tx_id];
+
+	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		ice_tx_backlog_entry(txep, tx_pkts, n);
+
+		ice_vtx(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		ice_vtx1(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = &txq->sw_ring[tx_id];
+	}
+
+	ice_tx_backlog_entry(txep, tx_pkts, nb_commit);
+
+	ice_vtx(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->tx_next_rs) {
+		txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)ICE_TX_DESC_CMD_RS) <<
+					 ICE_TXD_QW1_CMD_S);
+		txq->tx_next_rs =
+			(uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	ICE_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+ice_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+			 uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+		ret = ice_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;
+}
diff --git a/drivers/net/ice/meson.build b/drivers/net/ice/meson.build
index e6fe744879..a972ed668e 100644
--- a/drivers/net/ice/meson.build
+++ b/drivers/net/ice/meson.build
@@ -33,6 +33,17 @@ if arch_subdir == 'x86'
 				c_args: [cflags, '-mavx2'])
 		objs += ice_avx2_lib.extract_objects('ice_rxtx_vec_avx2.c')
 	endif
+
+	if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F') or (not machine_args.contains('-mno-avx512f') and cc.has_argument('-mavx512f'))
+		cflags += ['-DCC_AVX512_SUPPORT']
+		ice_avx512_lib = static_library('ice_avx512_lib',
+				      'ice_rxtx_vec_avx512.c',
+				      dependencies: [static_rte_ethdev,
+					static_rte_kvargs, static_rte_hash],
+				      include_directories: includes,
+				      c_args: [cflags, '-march=skylake-avx512', '-mavx512f'])
+		objs += ice_avx512_lib.extract_objects('ice_rxtx_vec_avx512.c')
+	endif
 endif
 
 sources += files('ice_dcf.c',
-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v2 2/3] net/ice: add RSS hash parsing in AVX512 path
  2020-09-18  3:35 ` [dpdk-dev] [PATCH v2 0/3] AVX512 vPMD on ice Leyi Rong
  2020-09-18  3:35   ` [dpdk-dev] [PATCH v2 1/3] net/ice: add AVX512 vector path Leyi Rong
@ 2020-09-18  3:35   ` Leyi Rong
  2020-09-18  3:35   ` [dpdk-dev] [PATCH v2 3/3] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
  2 siblings, 0 replies; 25+ messages in thread
From: Leyi Rong @ 2020-09-18  3:35 UTC (permalink / raw)
  To: qi.z.zhang, wenzhuo.lu, burce.richardson; +Cc: dev, Leyi Rong

Support RSS hash parsing in AVX512 data path as the default
RXDID is set to #22, that means the RSS hash field locates
in the 2nd 16B of each Flex Rx descriptor.

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/ice/ice_rxtx_vec_avx512.c | 105 ++++++++++++++++++++++++--
 1 file changed, 98 insertions(+), 7 deletions(-)

diff --git a/drivers/net/ice/ice_rxtx_vec_avx512.c b/drivers/net/ice/ice_rxtx_vec_avx512.c
index 6a9d0a8eaa..a2a5d9987a 100644
--- a/drivers/net/ice/ice_rxtx_vec_avx512.c
+++ b/drivers/net/ice/ice_rxtx_vec_avx512.c
@@ -176,8 +176,8 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
 	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
 	const __m512i shuf_msk =
 		_mm512_set4_epi32
-			(/* octet 12~15, 32 bits rss */
-			 15 << 24 | 14 << 16 | 13 << 8 | 12,
+			(/* rss hash parsed separately */
+			 0xFFFFFFFF,
 			 /* octet 10~11, 16 bits vlan_macip */
 			 /* octet 4~5, 16 bits data_len */
 			 11 << 24 | 10 << 16 | 5 << 8 | 4,
@@ -399,6 +399,11 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
 		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
 		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
 
+		__m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7, 0);
+		__m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7, 1);
+		__m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3, 0);
+		__m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3, 1);
+
 		/**
 		 * use permute/extract to get status content
 		 * After the operations, the packets status flags are in the
@@ -438,6 +443,97 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
 		/* merge flags */
 		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
 				rss_vlan_flags);
+
+#ifndef RTE_LIBRTE_ICE_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.
@@ -471,11 +567,6 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
 					    _mm256_srli_si256(mbuf_flags, 4),
 					    0x04);
 
-		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);
-
 		/* 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);
-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v2 3/3] net/ice: optimize Tx path on AVX512 vPMD
  2020-09-18  3:35 ` [dpdk-dev] [PATCH v2 0/3] AVX512 vPMD on ice Leyi Rong
  2020-09-18  3:35   ` [dpdk-dev] [PATCH v2 1/3] net/ice: add AVX512 vector path Leyi Rong
  2020-09-18  3:35   ` [dpdk-dev] [PATCH v2 2/3] net/ice: add RSS hash parsing in AVX512 path Leyi Rong
@ 2020-09-18  3:35   ` Leyi Rong
  2 siblings, 0 replies; 25+ messages in thread
From: Leyi Rong @ 2020-09-18  3:35 UTC (permalink / raw)
  To: qi.z.zhang, wenzhuo.lu, burce.richardson; +Cc: dev, Leyi Rong, Bruce Richardson

Optimize Tx path by using AVX512 instructions and vectorize the
tx free bufs process.

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 drivers/net/ice/ice_rxtx.h            |   4 +
 drivers/net/ice/ice_rxtx_vec_avx512.c | 155 +++++++++++++++++++++-----
 drivers/net/ice/ice_rxtx_vec_common.h |  36 ++++--
 3 files changed, 162 insertions(+), 33 deletions(-)

diff --git a/drivers/net/ice/ice_rxtx.h b/drivers/net/ice/ice_rxtx.h
index c13e54b251..6eb3447d26 100644
--- a/drivers/net/ice/ice_rxtx.h
+++ b/drivers/net/ice/ice_rxtx.h
@@ -91,6 +91,10 @@ struct ice_tx_entry {
 	uint16_t last_id;
 };
 
+struct ice_vec_tx_entry {
+	struct rte_mbuf *mbuf;
+};
+
 struct ice_tx_queue {
 	uint16_t nb_tx_desc; /* number of TX descriptors */
 	rte_iova_t tx_ring_dma; /* TX ring DMA address */
diff --git a/drivers/net/ice/ice_rxtx_vec_avx512.c b/drivers/net/ice/ice_rxtx_vec_avx512.c
index a2a5d9987a..e5e7cc1482 100644
--- a/drivers/net/ice/ice_rxtx_vec_avx512.c
+++ b/drivers/net/ice/ice_rxtx_vec_avx512.c
@@ -756,6 +756,108 @@ ice_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
 				rx_pkts + retval, nb_pkts);
 }
 
+static __rte_always_inline int
+ice_tx_free_bufs_avx512(struct ice_tx_queue *txq)
+{
+	struct ice_vec_tx_entry *txep;
+	uint32_t n;
+	uint32_t i;
+	int nb_free = 0;
+	struct rte_mbuf *m, *free[ICE_TX_MAX_FREE_BUF_SZ];
+
+	/* check DD bits on threshold descriptor */
+	if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
+			rte_cpu_to_le_64(ICE_TXD_QW1_DTYPE_M)) !=
+			rte_cpu_to_le_64(ICE_TX_DESC_DTYPE_DESC_DONE))
+		return 0;
+
+	n = txq->tx_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->tx_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_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
+	txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
+	if (txq->tx_next_dd >= txq->nb_tx_desc)
+		txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
+
+	return txq->tx_rs_thresh;
+}
+
 static inline void
 ice_vtx1(volatile struct ice_tx_desc *txdp,
 	 struct rte_mbuf *pkt, uint64_t flags)
@@ -766,7 +868,7 @@ ice_vtx1(volatile struct ice_tx_desc *txdp,
 		 ((uint64_t)pkt->data_len << ICE_TXD_QW1_TX_BUF_SZ_S));
 
 	__m128i descriptor = _mm_set_epi64x(high_qw,
-				pkt->buf_physaddr + pkt->data_off);
+				pkt->buf_iova + pkt->data_off);
 	_mm_store_si128((__m128i *)txdp, descriptor);
 }
 
@@ -777,13 +879,6 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 	const uint64_t hi_qw_tmpl = (ICE_TX_DESC_DTYPE_DATA |
 			((uint64_t)flags  << ICE_TXD_QW1_CMD_S));
 
-	/* if unaligned on 32-bit boundary, do one to align */
-	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
-		ice_vtx1(txdp, *pkt, flags);
-		nb_pkts--, txdp++, pkt++;
-	}
-
-	/* do two at a time while possible, in bursts */
 	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
 		uint64_t hi_qw3 =
 			hi_qw_tmpl |
@@ -802,20 +897,17 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 			((uint64_t)pkt[0]->data_len <<
 			 ICE_TXD_QW1_TX_BUF_SZ_S);
 
-		__m256i desc2_3 =
-			_mm256_set_epi64x
+		__m512i desc0_3 =
+			_mm512_set_epi64
 				(hi_qw3,
-				 pkt[3]->buf_physaddr + pkt[3]->data_off,
+				 pkt[3]->buf_iova + pkt[3]->data_off,
 				 hi_qw2,
-				 pkt[2]->buf_physaddr + pkt[2]->data_off);
-		__m256i desc0_1 =
-			_mm256_set_epi64x
-				(hi_qw1,
-				 pkt[1]->buf_physaddr + pkt[1]->data_off,
+				 pkt[2]->buf_iova + pkt[2]->data_off,
+				 hi_qw1,
+				 pkt[1]->buf_iova + pkt[1]->data_off,
 				 hi_qw0,
-				 pkt[0]->buf_physaddr + pkt[0]->data_off);
-		_mm256_store_si256((void *)(txdp + 2), desc2_3);
-		_mm256_store_si256((void *)txdp, desc0_1);
+				 pkt[0]->buf_iova + pkt[0]->data_off);
+		_mm512_storeu_si512((void *)txdp, desc0_3);
 	}
 
 	/* do any last ones */
@@ -825,13 +917,23 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 	}
 }
 
+static __rte_always_inline void
+ice_tx_backlog_entry_avx512(struct ice_vec_tx_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 uint16_t
 ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 				uint16_t nb_pkts)
 {
 	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
 	volatile struct ice_tx_desc *txdp;
-	struct ice_tx_entry *txep;
+	struct ice_vec_tx_entry *txep;
 	uint16_t n, nb_commit, tx_id;
 	uint64_t flags = ICE_TD_CMD;
 	uint64_t rs = ICE_TX_DESC_CMD_RS | ICE_TD_CMD;
@@ -840,7 +942,7 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
 
 	if (txq->nb_tx_free < txq->tx_free_thresh)
-		ice_tx_free_bufs(txq);
+		ice_tx_free_bufs_avx512(txq);
 
 	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
 	if (unlikely(nb_pkts == 0))
@@ -848,13 +950,14 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 
 	tx_id = txq->tx_tail;
 	txdp = &txq->tx_ring[tx_id];
-	txep = &txq->sw_ring[tx_id];
+	txep = (void *)txq->sw_ring;
+	txep += tx_id;
 
 	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
 
 	n = (uint16_t)(txq->nb_tx_desc - tx_id);
 	if (nb_commit >= n) {
-		ice_tx_backlog_entry(txep, tx_pkts, n);
+		ice_tx_backlog_entry_avx512(txep, tx_pkts, n);
 
 		ice_vtx(txdp, tx_pkts, n - 1, flags);
 		tx_pkts += (n - 1);
@@ -868,11 +971,11 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 		txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
 
 		/* avoid reach the end of ring */
-		txdp = &txq->tx_ring[tx_id];
-		txep = &txq->sw_ring[tx_id];
+		txdp = txq->tx_ring;
+		txep = (void *)txq->sw_ring;
 	}
 
-	ice_tx_backlog_entry(txep, tx_pkts, nb_commit);
+	ice_tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
 
 	ice_vtx(txdp, tx_pkts, nb_commit, flags);
 
diff --git a/drivers/net/ice/ice_rxtx_vec_common.h b/drivers/net/ice/ice_rxtx_vec_common.h
index e2019c8d6a..ae2ac29f2a 100644
--- a/drivers/net/ice/ice_rxtx_vec_common.h
+++ b/drivers/net/ice/ice_rxtx_vec_common.h
@@ -189,16 +189,38 @@ _ice_tx_queue_release_mbufs_vec(struct ice_tx_queue *txq)
 	 *  so need to free remains more carefully.
 	 */
 	i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
-	if (txq->tx_tail < i) {
-		for (; i < txq->nb_tx_desc; i++) {
+
+#ifdef CC_AVX512_SUPPORT
+	struct rte_eth_dev *dev = txq->vsi->adapter->eth_dev;
+
+	if (dev->tx_pkt_burst == ice_xmit_pkts_vec_avx512) {
+		struct ice_vec_tx_entry *swr = (void *)txq->sw_ring;
+
+		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;
+		}
+		for (; i < txq->tx_tail; i++) {
+			rte_pktmbuf_free_seg(swr[i].mbuf);
+			swr[i].mbuf = NULL;
+		}
+	} else
+#endif
+	{
+		if (txq->tx_tail < i) {
+			for (; i < txq->nb_tx_desc; i++) {
+				rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
+				txq->sw_ring[i].mbuf = NULL;
+			}
+			i = 0;
+		}
+		for (; i < txq->tx_tail; i++) {
 			rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
 			txq->sw_ring[i].mbuf = NULL;
 		}
-		i = 0;
-	}
-	for (; i < txq->tx_tail; i++) {
-		rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
-		txq->sw_ring[i].mbuf = NULL;
 	}
 }
 
-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v3 0/3] AVX512 vPMD on ice
  2020-09-10  6:55 [dpdk-dev] [PATCH v1 0/2] AVX512 vPMD on ice Leyi Rong
                   ` (2 preceding siblings ...)
  2020-09-18  3:35 ` [dpdk-dev] [PATCH v2 0/3] AVX512 vPMD on ice Leyi Rong
@ 2020-10-20 10:51 ` Leyi Rong
  2020-10-20 10:51   ` [dpdk-dev] [PATCH v3 1/3] net/ice: add AVX512 vector path Leyi Rong
                     ` (2 more replies)
  2020-10-23  4:14 ` [dpdk-dev] [PATCH v4 0/3] AVX512 vPMD on ice Leyi Rong
  4 siblings, 3 replies; 25+ messages in thread
From: Leyi Rong @ 2020-10-20 10:51 UTC (permalink / raw)
  To: bruce.richardson, wenzhuo.lu, qi.z.zhang; +Cc: dev, Leyi Rong

This patchset aims to support AVX512 vPMD on ice.

---
v3:
- Code rebased.

v2:
- No internal judgement when RTE_MACHINE_CPUFLAG_AVX512F is set in
  meson.build.
- Add RSS hash parsing as default RXDID is set to #22.
- Use buf_iova instead of buf_physaddr, as buf_physaddr will be removed.

Leyi Rong (3):
  net/ice: add AVX512 vector path
  net/ice: add RSS hash parsing in AVX512 path
  net/ice: optimize Tx path on AVX512 vPMD

 drivers/net/ice/ice_rxtx.c            |   90 ++-
 drivers/net/ice/ice_rxtx.h            |   11 +
 drivers/net/ice/ice_rxtx_vec_avx512.c | 1018 +++++++++++++++++++++++++
 drivers/net/ice/ice_rxtx_vec_common.h |   36 +-
 drivers/net/ice/meson.build           |   11 +
 5 files changed, 1141 insertions(+), 25 deletions(-)
 create mode 100644 drivers/net/ice/ice_rxtx_vec_avx512.c

-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v3 1/3] net/ice: add AVX512 vector path
  2020-10-20 10:51 ` [dpdk-dev] [PATCH v3 0/3] AVX512 vPMD on ice Leyi Rong
@ 2020-10-20 10:51   ` Leyi Rong
  2020-10-20 10:51   ` [dpdk-dev] [PATCH v3 2/3] net/ice: add RSS hash parsing in AVX512 path Leyi Rong
  2020-10-20 10:51   ` [dpdk-dev] [PATCH v3 3/3] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
  2 siblings, 0 replies; 25+ messages in thread
From: Leyi Rong @ 2020-10-20 10:51 UTC (permalink / raw)
  To: bruce.richardson, wenzhuo.lu, qi.z.zhang; +Cc: dev, Leyi Rong

Add AVX512 support for ice PMD. This patch adds ice_rxtx_vec_avx512.c
to support ice AVX512 vPMD.

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

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 drivers/net/ice/ice_rxtx.c            |  90 ++-
 drivers/net/ice/ice_rxtx.h            |   7 +
 drivers/net/ice/ice_rxtx_vec_avx512.c | 824 ++++++++++++++++++++++++++
 drivers/net/ice/meson.build           |  11 +
 4 files changed, 914 insertions(+), 18 deletions(-)
 create mode 100644 drivers/net/ice/ice_rxtx_vec_avx512.c

diff --git a/drivers/net/ice/ice_rxtx.c b/drivers/net/ice/ice_rxtx.c
index ee576c362a..f6291894cd 100644
--- a/drivers/net/ice/ice_rxtx.c
+++ b/drivers/net/ice/ice_rxtx.c
@@ -1930,6 +1930,10 @@ ice_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	if (dev->rx_pkt_burst == ice_recv_pkts_vec ||
 	    dev->rx_pkt_burst == ice_recv_scattered_pkts_vec ||
+#ifdef CC_AVX512_SUPPORT
+	    dev->rx_pkt_burst == ice_recv_pkts_vec_avx512 ||
+	    dev->rx_pkt_burst == ice_recv_scattered_pkts_vec_avx512 ||
+#endif
 	    dev->rx_pkt_burst == ice_recv_pkts_vec_avx2 ||
 	    dev->rx_pkt_burst == ice_recv_scattered_pkts_vec_avx2)
 		return ptypes;
@@ -2987,6 +2991,7 @@ ice_set_rx_function(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	struct ice_rx_queue *rxq;
 	int i;
+	bool use_avx512 = false;
 	bool use_avx2 = false;
 
 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
@@ -3001,9 +3006,18 @@ ice_set_rx_function(struct rte_eth_dev *dev)
 				}
 			}
 
-			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)
+			if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512 &&
+			rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+#ifdef CC_AVX512_SUPPORT
+				use_avx512 = true;
+#else
+			PMD_DRV_LOG(NOTICE,
+				"AVX512 is not supported in build env");
+#endif
+			if (!use_avx512 &&
+			(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;
 
 		} else {
@@ -3013,20 +3027,37 @@ ice_set_rx_function(struct rte_eth_dev *dev)
 
 	if (ad->rx_vec_allowed) {
 		if (dev->data->scattered_rx) {
-			PMD_DRV_LOG(DEBUG,
+			if (use_avx512) {
+				PMD_DRV_LOG(NOTICE,
+					"Using AVX512 Vector Scattered Rx (port %d).",
+					dev->data->port_id);
+				dev->rx_pkt_burst =
+					ice_recv_scattered_pkts_vec_avx512;
+			} else {
+				PMD_DRV_LOG(DEBUG,
 					"Using %sVector Scattered Rx (port %d).",
 					use_avx2 ? "avx2 " : "",
 					dev->data->port_id);
-			dev->rx_pkt_burst = use_avx2 ?
+				dev->rx_pkt_burst = use_avx2 ?
 					ice_recv_scattered_pkts_vec_avx2 :
 					ice_recv_scattered_pkts_vec;
+			}
 		} else {
-			PMD_DRV_LOG(DEBUG, "Using %sVector Rx (port %d).",
+			if (use_avx512) {
+				PMD_DRV_LOG(NOTICE,
+					"Using AVX512 Vector Rx (port %d).",
+					dev->data->port_id);
+				dev->rx_pkt_burst =
+					ice_recv_pkts_vec_avx512;
+			} else {
+				PMD_DRV_LOG(DEBUG,
+					"Using %sVector Rx (port %d).",
 					use_avx2 ? "avx2 " : "",
 					dev->data->port_id);
-			dev->rx_pkt_burst = use_avx2 ?
-						ice_recv_pkts_vec_avx2 :
-						ice_recv_pkts_vec;
+				dev->rx_pkt_burst = use_avx2 ?
+					ice_recv_pkts_vec_avx2 :
+					ice_recv_pkts_vec;
+			}
 		}
 		return;
 	}
@@ -3063,6 +3094,10 @@ static const struct {
 	{ ice_recv_pkts_bulk_alloc,         "Scalar Bulk Alloc" },
 	{ ice_recv_pkts,                    "Scalar" },
 #ifdef RTE_ARCH_X86
+#ifdef CC_AVX512_SUPPORT
+	{ ice_recv_scattered_pkts_vec_avx512, "Vector AVX512 Scattered" },
+	{ ice_recv_pkts_vec_avx512,           "Vector AVX512" },
+#endif
 	{ ice_recv_scattered_pkts_vec_avx2, "Vector AVX2 Scattered" },
 	{ ice_recv_pkts_vec_avx2,           "Vector AVX2" },
 	{ ice_recv_scattered_pkts_vec,      "Vector SSE Scattered" },
@@ -3167,6 +3202,7 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	struct ice_tx_queue *txq;
 	int i;
+	bool use_avx512 = false;
 	bool use_avx2 = false;
 
 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
@@ -3181,9 +3217,18 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 				}
 			}
 
-			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)
+			if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512 &&
+			rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+#ifdef CC_AVX512_SUPPORT
+				use_avx512 = true;
+#else
+			PMD_DRV_LOG(NOTICE,
+				"AVX512 is not supported in build env");
+#endif
+			if (!use_avx512 &&
+			(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;
 
 		} else {
@@ -3192,12 +3237,18 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 	}
 
 	if (ad->tx_vec_allowed) {
-		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
-			    use_avx2 ? "avx2 " : "",
-			    dev->data->port_id);
-		dev->tx_pkt_burst = use_avx2 ?
-				    ice_xmit_pkts_vec_avx2 :
-				    ice_xmit_pkts_vec;
+		if (use_avx512) {
+			PMD_DRV_LOG(NOTICE, "Using AVX512 Vector Tx (port %d).",
+				    dev->data->port_id);
+			dev->tx_pkt_burst = ice_xmit_pkts_vec_avx512;
+		} else {
+			PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
+				    use_avx2 ? "avx2 " : "",
+				    dev->data->port_id);
+			dev->tx_pkt_burst = use_avx2 ?
+					    ice_xmit_pkts_vec_avx2 :
+					    ice_xmit_pkts_vec;
+		}
 		dev->tx_pkt_prepare = NULL;
 
 		return;
@@ -3222,6 +3273,9 @@ static const struct {
 	{ ice_xmit_pkts_simple,   "Scalar Simple" },
 	{ ice_xmit_pkts,          "Scalar" },
 #ifdef RTE_ARCH_X86
+#ifdef CC_AVX512_SUPPORT
+	{ ice_xmit_pkts_vec_avx512, "Vector AVX512" },
+#endif
 	{ ice_xmit_pkts_vec_avx2, "Vector AVX2" },
 	{ ice_xmit_pkts_vec,      "Vector SSE" },
 #endif
diff --git a/drivers/net/ice/ice_rxtx.h b/drivers/net/ice/ice_rxtx.h
index 1c23c7541e..7cebbc2916 100644
--- a/drivers/net/ice/ice_rxtx.h
+++ b/drivers/net/ice/ice_rxtx.h
@@ -248,6 +248,13 @@ uint16_t ice_recv_scattered_pkts_vec_avx2(void *rx_queue,
 					  uint16_t nb_pkts);
 uint16_t ice_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 				uint16_t nb_pkts);
+uint16_t ice_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				  uint16_t nb_pkts);
+uint16_t ice_recv_scattered_pkts_vec_avx512(void *rx_queue,
+					    struct rte_mbuf **rx_pkts,
+					    uint16_t nb_pkts);
+uint16_t ice_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				  uint16_t nb_pkts);
 int ice_fdir_programming(struct ice_pf *pf, struct ice_fltr_desc *fdir_desc);
 int ice_tx_done_cleanup(void *txq, uint32_t free_cnt);
 
diff --git a/drivers/net/ice/ice_rxtx_vec_avx512.c b/drivers/net/ice/ice_rxtx_vec_avx512.c
new file mode 100644
index 0000000000..6a9d0a8eaa
--- /dev/null
+++ b/drivers/net/ice/ice_rxtx_vec_avx512.c
@@ -0,0 +1,824 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2019 Intel Corporation
+ */
+
+#include "ice_rxtx_vec_common.h"
+
+#include <x86intrin.h>
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+#define ICE_DESCS_PER_LOOP_AVX 8
+
+static inline void
+ice_rxq_rearm(struct ice_rx_queue *rxq)
+{
+	int i;
+	uint16_t rx_id;
+	volatile union ice_rx_flex_desc *rxdp;
+	struct ice_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
+	struct rte_mempool_cache *cache = rte_mempool_default_cache(rxq->mp,
+			rte_lcore_id());
+
+	rxdp = rxq->rx_ring + rxq->rxrearm_start;
+
+	/* We need to pull 'n' more MBUFs into the software ring */
+	if (cache->len < ICE_RXQ_REARM_THRESH) {
+		uint32_t req = ICE_RXQ_REARM_THRESH + (cache->size -
+				cache->len);
+
+		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 + ICE_RXQ_REARM_THRESH >=
+			    rxq->nb_rx_desc) {
+				__m128i dma_addr0;
+
+				dma_addr0 = _mm_setzero_si128();
+				for (i = 0; i < ICE_DESCS_PER_LOOP; i++) {
+					rxep[i].mbuf = &rxq->fake_mbuf;
+					_mm_store_si128
+						((__m128i *)&rxdp[i].read,
+							dma_addr0);
+				}
+			}
+			rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+				ICE_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_ICE_16BYTE_RX_DESC
+	/* shuffle the iova into 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
+
+	/* fill up the rxd in vector, process 8 mbufs in one loop */
+	for (i = 0; i < ICE_RXQ_REARM_THRESH / 8; i++) {
+		const __m512i mbuf_ptrs = _mm512_loadu_si512
+			(&cache->objs[cache->len - 8]);
+		_mm512_store_si512(rxep, mbuf_ptrs);
+
+		/* gather iova of mbuf0-7 into one zmm reg */
+		const __m512i iova_base_addrs = _mm512_i64gather_epi64
+			(_mm512_add_epi64(mbuf_ptrs, iova_offsets),
+				0, /* base */
+				1  /* scale */);
+		const __m512i iova_addrs = _mm512_add_epi64(iova_base_addrs,
+				headroom);
+#ifndef RTE_LIBRTE_ICE_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 iova 2-3 in hdr_addr of desc 0-1
+		 * but these are ignored by driver since header split not
+		 * enabled. Similarly for desc 4 & 5.
+		 */
+		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_store_si512((void *)rxdp, desc0_1);
+		_mm512_store_si512((void *)(rxdp + 2), desc2_3);
+		_mm512_store_si512((void *)(rxdp + 4), desc4_5);
+		_mm512_store_si512((void *)(rxdp + 6), desc6_7);
+#else
+		/* permute leaves iova 4-7 in hdr_addr of desc 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_store_si512((void *)rxdp, desc0_3);
+		_mm512_store_si512((void *)(rxdp + 4), desc4_7);
+#endif
+		rxep += 8, rxdp += 8, cache->len -= 8;
+	}
+
+	rxq->rxrearm_start += ICE_RXQ_REARM_THRESH;
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= ICE_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 */
+	ICE_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
+}
+
+static inline uint16_t
+_ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
+			      struct rte_mbuf **rx_pkts,
+			      uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
+			0, rxq->mbuf_initializer);
+	struct ice_rx_entry *sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union ice_rx_flex_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to ICE_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, ICE_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 > ICE_RXQ_REARM_THRESH)
+		ice_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 << ICE_RX_FLEX_DESC_STATUS0_DD_S)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set4_epi32
+			(0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore non-length fields */
+			);
+
+	/* 8 packets DD mask, LSB in each 32-bit value */
+	const __m256i dd_check = _mm256_set1_epi32(1);
+
+	/* 8 packets EOP mask, second-LSB in each 32-bit value */
+	const __m256i eop_check = _mm256_slli_epi32(dd_check,
+			ICE_RX_DESC_STATUS_EOF_S);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set4_epi32
+			(/* octet 12~15, 32 bits rss */
+			 15 << 24 | 14 << 16 | 13 << 8 | 12,
+			 /* octet 10~11, 16 bits vlan_macip */
+			 /* octet 4~5, 16 bits data_len */
+			 11 << 24 | 10 << 16 | 5 << 8 | 4,
+			 /* skip hi 16 bits pkt_len, zero out */
+			 /* octet 4~5, 16 bits pkt_len */
+			 0xFFFF << 16 | 5 << 8 | 4,
+			 /* pkt_type set as unknown */
+			 0xFFFFFFFF
+			);
+
+	/**
+	 * compile-time check the above crc and shuffle layout is correct.
+	 * NOTE: the first field (lowest address) is given last in set_epi
+	 * calls above.
+	 */
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
+
+	/* Status/Error flag masks */
+	/**
+	 * mask everything except 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,
+			/* 2nd 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,
+			/* 2nd 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 += ICE_DESCS_PER_LOOP_AVX,
+	     rxdp += ICE_DESCS_PER_LOOP_AVX) {
+		/* step 1, copy over 8 mbuf pointers to rx_pkts array */
+		_mm256_storeu_si256((void *)&rx_pkts[i],
+				    _mm256_loadu_si256((void *)&sw_ring[i]));
+#ifdef RTE_ARCH_X86_64
+		_mm256_storeu_si256
+			((void *)&rx_pkts[i + 4],
+			 _mm256_loadu_si256((void *)&sw_ring[i + 4]));
+#endif
+
+		__m512i raw_desc0_3, raw_desc4_7;
+		__m256i raw_desc0_1, raw_desc2_3, raw_desc4_5, raw_desc6_7;
+
+		/* load in descriptors, in reverse order */
+		const __m128i raw_desc7 =
+			_mm_load_si128((void *)(rxdp + 7));
+		rte_compiler_barrier();
+		const __m128i raw_desc6 =
+			_mm_load_si128((void *)(rxdp + 6));
+		rte_compiler_barrier();
+		const __m128i raw_desc5 =
+			_mm_load_si128((void *)(rxdp + 5));
+		rte_compiler_barrier();
+		const __m128i raw_desc4 =
+			_mm_load_si128((void *)(rxdp + 4));
+		rte_compiler_barrier();
+		const __m128i raw_desc3 =
+			_mm_load_si128((void *)(rxdp + 3));
+		rte_compiler_barrier();
+		const __m128i raw_desc2 =
+			_mm_load_si128((void *)(rxdp + 2));
+		rte_compiler_barrier();
+		const __m128i raw_desc1 =
+			_mm_load_si128((void *)(rxdp + 1));
+		rte_compiler_barrier();
+		const __m128i raw_desc0 =
+			_mm_load_si128((void *)(rxdp + 0));
+
+		raw_desc6_7 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc6),
+				 raw_desc7, 1);
+		raw_desc4_5 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc4),
+				 raw_desc5, 1);
+		raw_desc2_3 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc2),
+				 raw_desc3, 1);
+		raw_desc0_1 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc0),
+				 raw_desc1, 1);
+
+		raw_desc4_7 =
+			_mm512_inserti64x4
+				(_mm512_castsi256_si512(raw_desc4_5),
+				 raw_desc6_7, 1);
+		raw_desc0_3 =
+			_mm512_inserti64x4
+				(_mm512_castsi256_si512(raw_desc0_1),
+				 raw_desc2_3, 1);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < ICE_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 0-7 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb4_7 = _mm512_shuffle_epi8(raw_desc4_7, shuf_msk);
+		__m512i mb0_3 = _mm512_shuffle_epi8(raw_desc0_3, shuf_msk);
+
+		mb4_7 = _mm512_add_epi32(mb4_7, crc_adjust);
+		mb0_3 = _mm512_add_epi32(mb0_3, crc_adjust);
+
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptype_mask =
+			_mm512_set1_epi16(ICE_RX_FLEX_DESC_PTYPE_M);
+
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptypes4_7 =
+			_mm512_and_si512(raw_desc4_7, ptype_mask);
+		const __m512i ptypes0_3 =
+			_mm512_and_si512(raw_desc0_3, ptype_mask);
+
+		const __m256i ptypes6_7 =
+			_mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 =
+			_mm512_extracti64x4_epi64(ptypes4_7, 0);
+		const __m256i ptypes2_3 =
+			_mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 =
+			_mm512_extracti64x4_epi64(ptypes0_3, 0);
+		const 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 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 ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, ptype_tbl[ptype7],
+			 0, 0, 0, ptype_tbl[ptype6],
+			 0, 0, 0, ptype_tbl[ptype5],
+			 0, 0, 0, ptype_tbl[ptype4]);
+		const __m512i ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, ptype_tbl[ptype3],
+			 0, 0, 0, ptype_tbl[ptype2],
+			 0, 0, 0, ptype_tbl[ptype1],
+			 0, 0, 0, ptype_tbl[ptype0]);
+
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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 */
+		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				rss_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;
+
+		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);
+
+		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);
+
+		/* 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 << ICE_RX_DESC_STATUS_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 += ICE_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 != ICE_DESCS_PER_LOOP_AVX)
+			break;
+	}
+
+	/* update tail pointers */
+	rxq->rx_tail += received;
+	rxq->rx_tail &= (rxq->nb_rx_desc - 1);
+	if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep avx2 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
+/**
+ * Notice:
+ * - nb_pkts < ICE_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+ice_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+			 uint16_t nb_pkts)
+{
+	return _ice_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 < ICE_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+ice_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				    uint16_t nb_pkts)
+{
+	struct ice_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[ICE_VPMD_RX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _ice_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 + ice_rx_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 < ICE_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+ice_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > ICE_VPMD_RX_BURST) {
+		uint16_t burst = ice_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, ICE_VPMD_RX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < ICE_VPMD_RX_BURST)
+			return retval;
+	}
+	return retval + ice_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
+
+static inline void
+ice_vtx1(volatile struct ice_tx_desc *txdp,
+	 struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(ICE_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << ICE_TXD_QW1_CMD_S) |
+		 ((uint64_t)pkt->data_len << ICE_TXD_QW1_TX_BUF_SZ_S));
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+				pkt->buf_physaddr + pkt->data_off);
+	_mm_store_si128((__m128i *)txdp, descriptor);
+}
+
+static inline void
+ice_vtx(volatile struct ice_tx_desc *txdp,
+	struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (ICE_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << ICE_TXD_QW1_CMD_S));
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		ice_vtx1(txdp, *pkt, flags);
+		nb_pkts--, txdp++, pkt++;
+	}
+
+	/* do two at a time while possible, in bursts */
+	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
+		uint64_t hi_qw3 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[3]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw2 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[2]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw1 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[1]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw0 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[0]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+
+		__m256i desc2_3 =
+			_mm256_set_epi64x
+				(hi_qw3,
+				 pkt[3]->buf_physaddr + pkt[3]->data_off,
+				 hi_qw2,
+				 pkt[2]->buf_physaddr + pkt[2]->data_off);
+		__m256i desc0_1 =
+			_mm256_set_epi64x
+				(hi_qw1,
+				 pkt[1]->buf_physaddr + pkt[1]->data_off,
+				 hi_qw0,
+				 pkt[0]->buf_physaddr + pkt[0]->data_off);
+		_mm256_store_si256((void *)(txdp + 2), desc2_3);
+		_mm256_store_si256((void *)txdp, desc0_1);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		ice_vtx1(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				uint16_t nb_pkts)
+{
+	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
+	volatile struct ice_tx_desc *txdp;
+	struct ice_tx_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	uint64_t flags = ICE_TD_CMD;
+	uint64_t rs = ICE_TX_DESC_CMD_RS | ICE_TD_CMD;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+
+	if (txq->nb_tx_free < txq->tx_free_thresh)
+		ice_tx_free_bufs(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = &txq->sw_ring[tx_id];
+
+	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		ice_tx_backlog_entry(txep, tx_pkts, n);
+
+		ice_vtx(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		ice_vtx1(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = &txq->sw_ring[tx_id];
+	}
+
+	ice_tx_backlog_entry(txep, tx_pkts, nb_commit);
+
+	ice_vtx(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->tx_next_rs) {
+		txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)ICE_TX_DESC_CMD_RS) <<
+					 ICE_TXD_QW1_CMD_S);
+		txq->tx_next_rs =
+			(uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	ICE_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+ice_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+			 uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+		ret = ice_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;
+}
diff --git a/drivers/net/ice/meson.build b/drivers/net/ice/meson.build
index 99e1b773a3..077e98f5da 100644
--- a/drivers/net/ice/meson.build
+++ b/drivers/net/ice/meson.build
@@ -33,6 +33,17 @@ if arch_subdir == 'x86'
 				c_args: [cflags, '-mavx2'])
 		objs += ice_avx2_lib.extract_objects('ice_rxtx_vec_avx2.c')
 	endif
+
+	if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F') or (not machine_args.contains('-mno-avx512f') and cc.has_argument('-mavx512f'))
+		cflags += ['-DCC_AVX512_SUPPORT']
+		ice_avx512_lib = static_library('ice_avx512_lib',
+				      'ice_rxtx_vec_avx512.c',
+				      dependencies: [static_rte_ethdev,
+					static_rte_kvargs, static_rte_hash],
+				      include_directories: includes,
+				      c_args: [cflags, '-march=skylake-avx512', '-mavx512f'])
+		objs += ice_avx512_lib.extract_objects('ice_rxtx_vec_avx512.c')
+	endif
 endif
 
 sources += files('ice_dcf.c',
-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v3 2/3] net/ice: add RSS hash parsing in AVX512 path
  2020-10-20 10:51 ` [dpdk-dev] [PATCH v3 0/3] AVX512 vPMD on ice Leyi Rong
  2020-10-20 10:51   ` [dpdk-dev] [PATCH v3 1/3] net/ice: add AVX512 vector path Leyi Rong
@ 2020-10-20 10:51   ` Leyi Rong
  2020-10-20 10:51   ` [dpdk-dev] [PATCH v3 3/3] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
  2 siblings, 0 replies; 25+ messages in thread
From: Leyi Rong @ 2020-10-20 10:51 UTC (permalink / raw)
  To: bruce.richardson, wenzhuo.lu, qi.z.zhang; +Cc: dev, Leyi Rong

Support RSS hash parsing in AVX512 data path as the default
RXDID is set to #22, that means the RSS hash field locates
in the 2nd 16B of each Flex Rx descriptor.

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/ice/ice_rxtx_vec_avx512.c | 105 ++++++++++++++++++++++++--
 1 file changed, 98 insertions(+), 7 deletions(-)

diff --git a/drivers/net/ice/ice_rxtx_vec_avx512.c b/drivers/net/ice/ice_rxtx_vec_avx512.c
index 6a9d0a8eaa..a2a5d9987a 100644
--- a/drivers/net/ice/ice_rxtx_vec_avx512.c
+++ b/drivers/net/ice/ice_rxtx_vec_avx512.c
@@ -176,8 +176,8 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
 	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
 	const __m512i shuf_msk =
 		_mm512_set4_epi32
-			(/* octet 12~15, 32 bits rss */
-			 15 << 24 | 14 << 16 | 13 << 8 | 12,
+			(/* rss hash parsed separately */
+			 0xFFFFFFFF,
 			 /* octet 10~11, 16 bits vlan_macip */
 			 /* octet 4~5, 16 bits data_len */
 			 11 << 24 | 10 << 16 | 5 << 8 | 4,
@@ -399,6 +399,11 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
 		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
 		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
 
+		__m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7, 0);
+		__m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7, 1);
+		__m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3, 0);
+		__m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3, 1);
+
 		/**
 		 * use permute/extract to get status content
 		 * After the operations, the packets status flags are in the
@@ -438,6 +443,97 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
 		/* merge flags */
 		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
 				rss_vlan_flags);
+
+#ifndef RTE_LIBRTE_ICE_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.
@@ -471,11 +567,6 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
 					    _mm256_srli_si256(mbuf_flags, 4),
 					    0x04);
 
-		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);
-
 		/* 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);
-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v3 3/3] net/ice: optimize Tx path on AVX512 vPMD
  2020-10-20 10:51 ` [dpdk-dev] [PATCH v3 0/3] AVX512 vPMD on ice Leyi Rong
  2020-10-20 10:51   ` [dpdk-dev] [PATCH v3 1/3] net/ice: add AVX512 vector path Leyi Rong
  2020-10-20 10:51   ` [dpdk-dev] [PATCH v3 2/3] net/ice: add RSS hash parsing in AVX512 path Leyi Rong
@ 2020-10-20 10:51   ` Leyi Rong
  2 siblings, 0 replies; 25+ messages in thread
From: Leyi Rong @ 2020-10-20 10:51 UTC (permalink / raw)
  To: bruce.richardson, wenzhuo.lu, qi.z.zhang; +Cc: dev, Leyi Rong

Optimize Tx path by using AVX512 instructions and vectorize the
tx free bufs process.

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 drivers/net/ice/ice_rxtx.h            |   4 +
 drivers/net/ice/ice_rxtx_vec_avx512.c | 155 +++++++++++++++++++++-----
 drivers/net/ice/ice_rxtx_vec_common.h |  36 ++++--
 3 files changed, 162 insertions(+), 33 deletions(-)

diff --git a/drivers/net/ice/ice_rxtx.h b/drivers/net/ice/ice_rxtx.h
index 7cebbc2916..23409d479a 100644
--- a/drivers/net/ice/ice_rxtx.h
+++ b/drivers/net/ice/ice_rxtx.h
@@ -96,6 +96,10 @@ struct ice_tx_entry {
 	uint16_t last_id;
 };
 
+struct ice_vec_tx_entry {
+	struct rte_mbuf *mbuf;
+};
+
 struct ice_tx_queue {
 	uint16_t nb_tx_desc; /* number of TX descriptors */
 	rte_iova_t tx_ring_dma; /* TX ring DMA address */
diff --git a/drivers/net/ice/ice_rxtx_vec_avx512.c b/drivers/net/ice/ice_rxtx_vec_avx512.c
index a2a5d9987a..e5e7cc1482 100644
--- a/drivers/net/ice/ice_rxtx_vec_avx512.c
+++ b/drivers/net/ice/ice_rxtx_vec_avx512.c
@@ -756,6 +756,108 @@ ice_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
 				rx_pkts + retval, nb_pkts);
 }
 
+static __rte_always_inline int
+ice_tx_free_bufs_avx512(struct ice_tx_queue *txq)
+{
+	struct ice_vec_tx_entry *txep;
+	uint32_t n;
+	uint32_t i;
+	int nb_free = 0;
+	struct rte_mbuf *m, *free[ICE_TX_MAX_FREE_BUF_SZ];
+
+	/* check DD bits on threshold descriptor */
+	if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
+			rte_cpu_to_le_64(ICE_TXD_QW1_DTYPE_M)) !=
+			rte_cpu_to_le_64(ICE_TX_DESC_DTYPE_DESC_DONE))
+		return 0;
+
+	n = txq->tx_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->tx_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_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
+	txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
+	if (txq->tx_next_dd >= txq->nb_tx_desc)
+		txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
+
+	return txq->tx_rs_thresh;
+}
+
 static inline void
 ice_vtx1(volatile struct ice_tx_desc *txdp,
 	 struct rte_mbuf *pkt, uint64_t flags)
@@ -766,7 +868,7 @@ ice_vtx1(volatile struct ice_tx_desc *txdp,
 		 ((uint64_t)pkt->data_len << ICE_TXD_QW1_TX_BUF_SZ_S));
 
 	__m128i descriptor = _mm_set_epi64x(high_qw,
-				pkt->buf_physaddr + pkt->data_off);
+				pkt->buf_iova + pkt->data_off);
 	_mm_store_si128((__m128i *)txdp, descriptor);
 }
 
@@ -777,13 +879,6 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 	const uint64_t hi_qw_tmpl = (ICE_TX_DESC_DTYPE_DATA |
 			((uint64_t)flags  << ICE_TXD_QW1_CMD_S));
 
-	/* if unaligned on 32-bit boundary, do one to align */
-	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
-		ice_vtx1(txdp, *pkt, flags);
-		nb_pkts--, txdp++, pkt++;
-	}
-
-	/* do two at a time while possible, in bursts */
 	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
 		uint64_t hi_qw3 =
 			hi_qw_tmpl |
@@ -802,20 +897,17 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 			((uint64_t)pkt[0]->data_len <<
 			 ICE_TXD_QW1_TX_BUF_SZ_S);
 
-		__m256i desc2_3 =
-			_mm256_set_epi64x
+		__m512i desc0_3 =
+			_mm512_set_epi64
 				(hi_qw3,
-				 pkt[3]->buf_physaddr + pkt[3]->data_off,
+				 pkt[3]->buf_iova + pkt[3]->data_off,
 				 hi_qw2,
-				 pkt[2]->buf_physaddr + pkt[2]->data_off);
-		__m256i desc0_1 =
-			_mm256_set_epi64x
-				(hi_qw1,
-				 pkt[1]->buf_physaddr + pkt[1]->data_off,
+				 pkt[2]->buf_iova + pkt[2]->data_off,
+				 hi_qw1,
+				 pkt[1]->buf_iova + pkt[1]->data_off,
 				 hi_qw0,
-				 pkt[0]->buf_physaddr + pkt[0]->data_off);
-		_mm256_store_si256((void *)(txdp + 2), desc2_3);
-		_mm256_store_si256((void *)txdp, desc0_1);
+				 pkt[0]->buf_iova + pkt[0]->data_off);
+		_mm512_storeu_si512((void *)txdp, desc0_3);
 	}
 
 	/* do any last ones */
@@ -825,13 +917,23 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 	}
 }
 
+static __rte_always_inline void
+ice_tx_backlog_entry_avx512(struct ice_vec_tx_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 uint16_t
 ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 				uint16_t nb_pkts)
 {
 	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
 	volatile struct ice_tx_desc *txdp;
-	struct ice_tx_entry *txep;
+	struct ice_vec_tx_entry *txep;
 	uint16_t n, nb_commit, tx_id;
 	uint64_t flags = ICE_TD_CMD;
 	uint64_t rs = ICE_TX_DESC_CMD_RS | ICE_TD_CMD;
@@ -840,7 +942,7 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
 
 	if (txq->nb_tx_free < txq->tx_free_thresh)
-		ice_tx_free_bufs(txq);
+		ice_tx_free_bufs_avx512(txq);
 
 	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
 	if (unlikely(nb_pkts == 0))
@@ -848,13 +950,14 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 
 	tx_id = txq->tx_tail;
 	txdp = &txq->tx_ring[tx_id];
-	txep = &txq->sw_ring[tx_id];
+	txep = (void *)txq->sw_ring;
+	txep += tx_id;
 
 	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
 
 	n = (uint16_t)(txq->nb_tx_desc - tx_id);
 	if (nb_commit >= n) {
-		ice_tx_backlog_entry(txep, tx_pkts, n);
+		ice_tx_backlog_entry_avx512(txep, tx_pkts, n);
 
 		ice_vtx(txdp, tx_pkts, n - 1, flags);
 		tx_pkts += (n - 1);
@@ -868,11 +971,11 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 		txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
 
 		/* avoid reach the end of ring */
-		txdp = &txq->tx_ring[tx_id];
-		txep = &txq->sw_ring[tx_id];
+		txdp = txq->tx_ring;
+		txep = (void *)txq->sw_ring;
 	}
 
-	ice_tx_backlog_entry(txep, tx_pkts, nb_commit);
+	ice_tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
 
 	ice_vtx(txdp, tx_pkts, nb_commit, flags);
 
diff --git a/drivers/net/ice/ice_rxtx_vec_common.h b/drivers/net/ice/ice_rxtx_vec_common.h
index e2019c8d6a..ae2ac29f2a 100644
--- a/drivers/net/ice/ice_rxtx_vec_common.h
+++ b/drivers/net/ice/ice_rxtx_vec_common.h
@@ -189,16 +189,38 @@ _ice_tx_queue_release_mbufs_vec(struct ice_tx_queue *txq)
 	 *  so need to free remains more carefully.
 	 */
 	i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
-	if (txq->tx_tail < i) {
-		for (; i < txq->nb_tx_desc; i++) {
+
+#ifdef CC_AVX512_SUPPORT
+	struct rte_eth_dev *dev = txq->vsi->adapter->eth_dev;
+
+	if (dev->tx_pkt_burst == ice_xmit_pkts_vec_avx512) {
+		struct ice_vec_tx_entry *swr = (void *)txq->sw_ring;
+
+		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;
+		}
+		for (; i < txq->tx_tail; i++) {
+			rte_pktmbuf_free_seg(swr[i].mbuf);
+			swr[i].mbuf = NULL;
+		}
+	} else
+#endif
+	{
+		if (txq->tx_tail < i) {
+			for (; i < txq->nb_tx_desc; i++) {
+				rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
+				txq->sw_ring[i].mbuf = NULL;
+			}
+			i = 0;
+		}
+		for (; i < txq->tx_tail; i++) {
 			rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
 			txq->sw_ring[i].mbuf = NULL;
 		}
-		i = 0;
-	}
-	for (; i < txq->tx_tail; i++) {
-		rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
-		txq->sw_ring[i].mbuf = NULL;
 	}
 }
 
-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v4 0/3] AVX512 vPMD on ice
  2020-09-10  6:55 [dpdk-dev] [PATCH v1 0/2] AVX512 vPMD on ice Leyi Rong
                   ` (3 preceding siblings ...)
  2020-10-20 10:51 ` [dpdk-dev] [PATCH v3 0/3] AVX512 vPMD on ice Leyi Rong
@ 2020-10-23  4:14 ` Leyi Rong
  2020-10-23  4:14   ` [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path Leyi Rong
                     ` (3 more replies)
  4 siblings, 4 replies; 25+ messages in thread
From: Leyi Rong @ 2020-10-23  4:14 UTC (permalink / raw)
  To: qi.z.zhang, wenzhuo.lu, burce.richardson; +Cc: dev, Leyi Rong

This patchset aims to support AVX512 vPMD on ice.

---
v4:
- Replace buf_physaddr by buf_iova in the 1st patch to eliminate build
  error.

v3:
- Code rebased.

v2:
- No internal judgement when RTE_MACHINE_CPUFLAG_AVX512F is set in
  meson.build.
- Add RSS hash parsing as default RXDID is set to #22.
- Use buf_iova instead of buf_physaddr, as buf_physaddr will be removed.

Leyi Rong (3):
  net/ice: add AVX512 vector path
  net/ice: add RSS hash parsing in AVX512 path
  net/ice: optimize Tx path on AVX512 vPMD

 drivers/net/ice/ice_rxtx.c            |   90 ++-
 drivers/net/ice/ice_rxtx.h            |   11 +
 drivers/net/ice/ice_rxtx_vec_avx512.c | 1018 +++++++++++++++++++++++++
 drivers/net/ice/ice_rxtx_vec_common.h |   36 +-
 drivers/net/ice/meson.build           |   11 +
 5 files changed, 1141 insertions(+), 25 deletions(-)
 create mode 100644 drivers/net/ice/ice_rxtx_vec_avx512.c

-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
  2020-10-23  4:14 ` [dpdk-dev] [PATCH v4 0/3] AVX512 vPMD on ice Leyi Rong
@ 2020-10-23  4:14   ` Leyi Rong
  2020-10-25 16:23     ` David Marchand
  2020-10-27  8:32     ` Ali Alnubani
  2020-10-23  4:14   ` [dpdk-dev] [PATCH v4 2/3] net/ice: add RSS hash parsing in AVX512 path Leyi Rong
                     ` (2 subsequent siblings)
  3 siblings, 2 replies; 25+ messages in thread
From: Leyi Rong @ 2020-10-23  4:14 UTC (permalink / raw)
  To: qi.z.zhang, wenzhuo.lu, burce.richardson; +Cc: dev, Leyi Rong, Bruce Richardson

Add AVX512 support for ice PMD. This patch adds ice_rxtx_vec_avx512.c
to support ice AVX512 vPMD.

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

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 drivers/net/ice/ice_rxtx.c            |  90 ++-
 drivers/net/ice/ice_rxtx.h            |   7 +
 drivers/net/ice/ice_rxtx_vec_avx512.c | 824 ++++++++++++++++++++++++++
 drivers/net/ice/meson.build           |  11 +
 4 files changed, 914 insertions(+), 18 deletions(-)
 create mode 100644 drivers/net/ice/ice_rxtx_vec_avx512.c

diff --git a/drivers/net/ice/ice_rxtx.c b/drivers/net/ice/ice_rxtx.c
index ee576c362a..f6291894cd 100644
--- a/drivers/net/ice/ice_rxtx.c
+++ b/drivers/net/ice/ice_rxtx.c
@@ -1930,6 +1930,10 @@ ice_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	if (dev->rx_pkt_burst == ice_recv_pkts_vec ||
 	    dev->rx_pkt_burst == ice_recv_scattered_pkts_vec ||
+#ifdef CC_AVX512_SUPPORT
+	    dev->rx_pkt_burst == ice_recv_pkts_vec_avx512 ||
+	    dev->rx_pkt_burst == ice_recv_scattered_pkts_vec_avx512 ||
+#endif
 	    dev->rx_pkt_burst == ice_recv_pkts_vec_avx2 ||
 	    dev->rx_pkt_burst == ice_recv_scattered_pkts_vec_avx2)
 		return ptypes;
@@ -2987,6 +2991,7 @@ ice_set_rx_function(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	struct ice_rx_queue *rxq;
 	int i;
+	bool use_avx512 = false;
 	bool use_avx2 = false;
 
 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
@@ -3001,9 +3006,18 @@ ice_set_rx_function(struct rte_eth_dev *dev)
 				}
 			}
 
-			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)
+			if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512 &&
+			rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+#ifdef CC_AVX512_SUPPORT
+				use_avx512 = true;
+#else
+			PMD_DRV_LOG(NOTICE,
+				"AVX512 is not supported in build env");
+#endif
+			if (!use_avx512 &&
+			(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;
 
 		} else {
@@ -3013,20 +3027,37 @@ ice_set_rx_function(struct rte_eth_dev *dev)
 
 	if (ad->rx_vec_allowed) {
 		if (dev->data->scattered_rx) {
-			PMD_DRV_LOG(DEBUG,
+			if (use_avx512) {
+				PMD_DRV_LOG(NOTICE,
+					"Using AVX512 Vector Scattered Rx (port %d).",
+					dev->data->port_id);
+				dev->rx_pkt_burst =
+					ice_recv_scattered_pkts_vec_avx512;
+			} else {
+				PMD_DRV_LOG(DEBUG,
 					"Using %sVector Scattered Rx (port %d).",
 					use_avx2 ? "avx2 " : "",
 					dev->data->port_id);
-			dev->rx_pkt_burst = use_avx2 ?
+				dev->rx_pkt_burst = use_avx2 ?
 					ice_recv_scattered_pkts_vec_avx2 :
 					ice_recv_scattered_pkts_vec;
+			}
 		} else {
-			PMD_DRV_LOG(DEBUG, "Using %sVector Rx (port %d).",
+			if (use_avx512) {
+				PMD_DRV_LOG(NOTICE,
+					"Using AVX512 Vector Rx (port %d).",
+					dev->data->port_id);
+				dev->rx_pkt_burst =
+					ice_recv_pkts_vec_avx512;
+			} else {
+				PMD_DRV_LOG(DEBUG,
+					"Using %sVector Rx (port %d).",
 					use_avx2 ? "avx2 " : "",
 					dev->data->port_id);
-			dev->rx_pkt_burst = use_avx2 ?
-						ice_recv_pkts_vec_avx2 :
-						ice_recv_pkts_vec;
+				dev->rx_pkt_burst = use_avx2 ?
+					ice_recv_pkts_vec_avx2 :
+					ice_recv_pkts_vec;
+			}
 		}
 		return;
 	}
@@ -3063,6 +3094,10 @@ static const struct {
 	{ ice_recv_pkts_bulk_alloc,         "Scalar Bulk Alloc" },
 	{ ice_recv_pkts,                    "Scalar" },
 #ifdef RTE_ARCH_X86
+#ifdef CC_AVX512_SUPPORT
+	{ ice_recv_scattered_pkts_vec_avx512, "Vector AVX512 Scattered" },
+	{ ice_recv_pkts_vec_avx512,           "Vector AVX512" },
+#endif
 	{ ice_recv_scattered_pkts_vec_avx2, "Vector AVX2 Scattered" },
 	{ ice_recv_pkts_vec_avx2,           "Vector AVX2" },
 	{ ice_recv_scattered_pkts_vec,      "Vector SSE Scattered" },
@@ -3167,6 +3202,7 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	struct ice_tx_queue *txq;
 	int i;
+	bool use_avx512 = false;
 	bool use_avx2 = false;
 
 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
@@ -3181,9 +3217,18 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 				}
 			}
 
-			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)
+			if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512 &&
+			rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+#ifdef CC_AVX512_SUPPORT
+				use_avx512 = true;
+#else
+			PMD_DRV_LOG(NOTICE,
+				"AVX512 is not supported in build env");
+#endif
+			if (!use_avx512 &&
+			(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;
 
 		} else {
@@ -3192,12 +3237,18 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 	}
 
 	if (ad->tx_vec_allowed) {
-		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
-			    use_avx2 ? "avx2 " : "",
-			    dev->data->port_id);
-		dev->tx_pkt_burst = use_avx2 ?
-				    ice_xmit_pkts_vec_avx2 :
-				    ice_xmit_pkts_vec;
+		if (use_avx512) {
+			PMD_DRV_LOG(NOTICE, "Using AVX512 Vector Tx (port %d).",
+				    dev->data->port_id);
+			dev->tx_pkt_burst = ice_xmit_pkts_vec_avx512;
+		} else {
+			PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
+				    use_avx2 ? "avx2 " : "",
+				    dev->data->port_id);
+			dev->tx_pkt_burst = use_avx2 ?
+					    ice_xmit_pkts_vec_avx2 :
+					    ice_xmit_pkts_vec;
+		}
 		dev->tx_pkt_prepare = NULL;
 
 		return;
@@ -3222,6 +3273,9 @@ static const struct {
 	{ ice_xmit_pkts_simple,   "Scalar Simple" },
 	{ ice_xmit_pkts,          "Scalar" },
 #ifdef RTE_ARCH_X86
+#ifdef CC_AVX512_SUPPORT
+	{ ice_xmit_pkts_vec_avx512, "Vector AVX512" },
+#endif
 	{ ice_xmit_pkts_vec_avx2, "Vector AVX2" },
 	{ ice_xmit_pkts_vec,      "Vector SSE" },
 #endif
diff --git a/drivers/net/ice/ice_rxtx.h b/drivers/net/ice/ice_rxtx.h
index 1c23c7541e..7cebbc2916 100644
--- a/drivers/net/ice/ice_rxtx.h
+++ b/drivers/net/ice/ice_rxtx.h
@@ -248,6 +248,13 @@ uint16_t ice_recv_scattered_pkts_vec_avx2(void *rx_queue,
 					  uint16_t nb_pkts);
 uint16_t ice_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 				uint16_t nb_pkts);
+uint16_t ice_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				  uint16_t nb_pkts);
+uint16_t ice_recv_scattered_pkts_vec_avx512(void *rx_queue,
+					    struct rte_mbuf **rx_pkts,
+					    uint16_t nb_pkts);
+uint16_t ice_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				  uint16_t nb_pkts);
 int ice_fdir_programming(struct ice_pf *pf, struct ice_fltr_desc *fdir_desc);
 int ice_tx_done_cleanup(void *txq, uint32_t free_cnt);
 
diff --git a/drivers/net/ice/ice_rxtx_vec_avx512.c b/drivers/net/ice/ice_rxtx_vec_avx512.c
new file mode 100644
index 0000000000..73c4ffce99
--- /dev/null
+++ b/drivers/net/ice/ice_rxtx_vec_avx512.c
@@ -0,0 +1,824 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2019 Intel Corporation
+ */
+
+#include "ice_rxtx_vec_common.h"
+
+#include <x86intrin.h>
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+#define ICE_DESCS_PER_LOOP_AVX 8
+
+static inline void
+ice_rxq_rearm(struct ice_rx_queue *rxq)
+{
+	int i;
+	uint16_t rx_id;
+	volatile union ice_rx_flex_desc *rxdp;
+	struct ice_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
+	struct rte_mempool_cache *cache = rte_mempool_default_cache(rxq->mp,
+			rte_lcore_id());
+
+	rxdp = rxq->rx_ring + rxq->rxrearm_start;
+
+	/* We need to pull 'n' more MBUFs into the software ring */
+	if (cache->len < ICE_RXQ_REARM_THRESH) {
+		uint32_t req = ICE_RXQ_REARM_THRESH + (cache->size -
+				cache->len);
+
+		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 + ICE_RXQ_REARM_THRESH >=
+			    rxq->nb_rx_desc) {
+				__m128i dma_addr0;
+
+				dma_addr0 = _mm_setzero_si128();
+				for (i = 0; i < ICE_DESCS_PER_LOOP; i++) {
+					rxep[i].mbuf = &rxq->fake_mbuf;
+					_mm_store_si128
+						((__m128i *)&rxdp[i].read,
+							dma_addr0);
+				}
+			}
+			rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+				ICE_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_ICE_16BYTE_RX_DESC
+	/* shuffle the iova into 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
+
+	/* fill up the rxd in vector, process 8 mbufs in one loop */
+	for (i = 0; i < ICE_RXQ_REARM_THRESH / 8; i++) {
+		const __m512i mbuf_ptrs = _mm512_loadu_si512
+			(&cache->objs[cache->len - 8]);
+		_mm512_store_si512(rxep, mbuf_ptrs);
+
+		/* gather iova of mbuf0-7 into one zmm reg */
+		const __m512i iova_base_addrs = _mm512_i64gather_epi64
+			(_mm512_add_epi64(mbuf_ptrs, iova_offsets),
+				0, /* base */
+				1  /* scale */);
+		const __m512i iova_addrs = _mm512_add_epi64(iova_base_addrs,
+				headroom);
+#ifndef RTE_LIBRTE_ICE_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 iova 2-3 in hdr_addr of desc 0-1
+		 * but these are ignored by driver since header split not
+		 * enabled. Similarly for desc 4 & 5.
+		 */
+		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_store_si512((void *)rxdp, desc0_1);
+		_mm512_store_si512((void *)(rxdp + 2), desc2_3);
+		_mm512_store_si512((void *)(rxdp + 4), desc4_5);
+		_mm512_store_si512((void *)(rxdp + 6), desc6_7);
+#else
+		/* permute leaves iova 4-7 in hdr_addr of desc 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_store_si512((void *)rxdp, desc0_3);
+		_mm512_store_si512((void *)(rxdp + 4), desc4_7);
+#endif
+		rxep += 8, rxdp += 8, cache->len -= 8;
+	}
+
+	rxq->rxrearm_start += ICE_RXQ_REARM_THRESH;
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= ICE_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 */
+	ICE_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
+}
+
+static inline uint16_t
+_ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
+			      struct rte_mbuf **rx_pkts,
+			      uint16_t nb_pkts, uint8_t *split_packet)
+{
+	const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
+	const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
+			0, rxq->mbuf_initializer);
+	struct ice_rx_entry *sw_ring = &rxq->sw_ring[rxq->rx_tail];
+	volatile union ice_rx_flex_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
+
+	rte_prefetch0(rxdp);
+
+	/* nb_pkts has to be floor-aligned to ICE_DESCS_PER_LOOP_AVX */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, ICE_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 > ICE_RXQ_REARM_THRESH)
+		ice_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 << ICE_RX_FLEX_DESC_STATUS0_DD_S)))
+		return 0;
+
+	/* constants used in processing loop */
+	const __m512i crc_adjust =
+		_mm512_set4_epi32
+			(0,             /* ignore non-length fields */
+			 -rxq->crc_len, /* sub crc on data_len */
+			 -rxq->crc_len, /* sub crc on pkt_len */
+			 0              /* ignore non-length fields */
+			);
+
+	/* 8 packets DD mask, LSB in each 32-bit value */
+	const __m256i dd_check = _mm256_set1_epi32(1);
+
+	/* 8 packets EOP mask, second-LSB in each 32-bit value */
+	const __m256i eop_check = _mm256_slli_epi32(dd_check,
+			ICE_RX_DESC_STATUS_EOF_S);
+
+	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
+	const __m512i shuf_msk =
+		_mm512_set4_epi32
+			(/* octet 12~15, 32 bits rss */
+			 15 << 24 | 14 << 16 | 13 << 8 | 12,
+			 /* octet 10~11, 16 bits vlan_macip */
+			 /* octet 4~5, 16 bits data_len */
+			 11 << 24 | 10 << 16 | 5 << 8 | 4,
+			 /* skip hi 16 bits pkt_len, zero out */
+			 /* octet 4~5, 16 bits pkt_len */
+			 0xFFFF << 16 | 5 << 8 | 4,
+			 /* pkt_type set as unknown */
+			 0xFFFFFFFF
+			);
+
+	/**
+	 * compile-time check the above crc and shuffle layout is correct.
+	 * NOTE: the first field (lowest address) is given last in set_epi
+	 * calls above.
+	 */
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
+			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
+
+	/* Status/Error flag masks */
+	/**
+	 * mask everything except 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,
+			/* 2nd 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,
+			/* 2nd 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 += ICE_DESCS_PER_LOOP_AVX,
+	     rxdp += ICE_DESCS_PER_LOOP_AVX) {
+		/* step 1, copy over 8 mbuf pointers to rx_pkts array */
+		_mm256_storeu_si256((void *)&rx_pkts[i],
+				    _mm256_loadu_si256((void *)&sw_ring[i]));
+#ifdef RTE_ARCH_X86_64
+		_mm256_storeu_si256
+			((void *)&rx_pkts[i + 4],
+			 _mm256_loadu_si256((void *)&sw_ring[i + 4]));
+#endif
+
+		__m512i raw_desc0_3, raw_desc4_7;
+		__m256i raw_desc0_1, raw_desc2_3, raw_desc4_5, raw_desc6_7;
+
+		/* load in descriptors, in reverse order */
+		const __m128i raw_desc7 =
+			_mm_load_si128((void *)(rxdp + 7));
+		rte_compiler_barrier();
+		const __m128i raw_desc6 =
+			_mm_load_si128((void *)(rxdp + 6));
+		rte_compiler_barrier();
+		const __m128i raw_desc5 =
+			_mm_load_si128((void *)(rxdp + 5));
+		rte_compiler_barrier();
+		const __m128i raw_desc4 =
+			_mm_load_si128((void *)(rxdp + 4));
+		rte_compiler_barrier();
+		const __m128i raw_desc3 =
+			_mm_load_si128((void *)(rxdp + 3));
+		rte_compiler_barrier();
+		const __m128i raw_desc2 =
+			_mm_load_si128((void *)(rxdp + 2));
+		rte_compiler_barrier();
+		const __m128i raw_desc1 =
+			_mm_load_si128((void *)(rxdp + 1));
+		rte_compiler_barrier();
+		const __m128i raw_desc0 =
+			_mm_load_si128((void *)(rxdp + 0));
+
+		raw_desc6_7 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc6),
+				 raw_desc7, 1);
+		raw_desc4_5 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc4),
+				 raw_desc5, 1);
+		raw_desc2_3 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc2),
+				 raw_desc3, 1);
+		raw_desc0_1 =
+			_mm256_inserti128_si256
+				(_mm256_castsi128_si256(raw_desc0),
+				 raw_desc1, 1);
+
+		raw_desc4_7 =
+			_mm512_inserti64x4
+				(_mm512_castsi256_si512(raw_desc4_5),
+				 raw_desc6_7, 1);
+		raw_desc0_3 =
+			_mm512_inserti64x4
+				(_mm512_castsi256_si512(raw_desc0_1),
+				 raw_desc2_3, 1);
+
+		if (split_packet) {
+			int j;
+
+			for (j = 0; j < ICE_DESCS_PER_LOOP_AVX; j++)
+				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
+		}
+
+		/**
+		 * convert descriptors 0-7 into mbufs, re-arrange fields.
+		 * Then write into the mbuf.
+		 */
+		__m512i mb4_7 = _mm512_shuffle_epi8(raw_desc4_7, shuf_msk);
+		__m512i mb0_3 = _mm512_shuffle_epi8(raw_desc0_3, shuf_msk);
+
+		mb4_7 = _mm512_add_epi32(mb4_7, crc_adjust);
+		mb0_3 = _mm512_add_epi32(mb0_3, crc_adjust);
+
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptype_mask =
+			_mm512_set1_epi16(ICE_RX_FLEX_DESC_PTYPE_M);
+
+		/**
+		 * to get packet types, ptype is located in bit16-25
+		 * of each 128bits
+		 */
+		const __m512i ptypes4_7 =
+			_mm512_and_si512(raw_desc4_7, ptype_mask);
+		const __m512i ptypes0_3 =
+			_mm512_and_si512(raw_desc0_3, ptype_mask);
+
+		const __m256i ptypes6_7 =
+			_mm512_extracti64x4_epi64(ptypes4_7, 1);
+		const __m256i ptypes4_5 =
+			_mm512_extracti64x4_epi64(ptypes4_7, 0);
+		const __m256i ptypes2_3 =
+			_mm512_extracti64x4_epi64(ptypes0_3, 1);
+		const __m256i ptypes0_1 =
+			_mm512_extracti64x4_epi64(ptypes0_3, 0);
+		const 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 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 ptype4_7 = _mm512_set_epi32
+			(0, 0, 0, ptype_tbl[ptype7],
+			 0, 0, 0, ptype_tbl[ptype6],
+			 0, 0, 0, ptype_tbl[ptype5],
+			 0, 0, 0, ptype_tbl[ptype4]);
+		const __m512i ptype0_3 = _mm512_set_epi32
+			(0, 0, 0, ptype_tbl[ptype3],
+			 0, 0, 0, ptype_tbl[ptype2],
+			 0, 0, 0, ptype_tbl[ptype1],
+			 0, 0, 0, ptype_tbl[ptype0]);
+
+		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
+		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
+
+		/**
+		 * 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 */
+		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
+				rss_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;
+
+		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);
+
+		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);
+
+		/* 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 << ICE_RX_DESC_STATUS_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 += ICE_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 != ICE_DESCS_PER_LOOP_AVX)
+			break;
+	}
+
+	/* update tail pointers */
+	rxq->rx_tail += received;
+	rxq->rx_tail &= (rxq->nb_rx_desc - 1);
+	if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep avx2 aligned */
+		rxq->rx_tail--;
+		received--;
+	}
+	rxq->rxrearm_nb += received;
+	return received;
+}
+
+/**
+ * Notice:
+ * - nb_pkts < ICE_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+ice_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+			 uint16_t nb_pkts)
+{
+	return _ice_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 < ICE_DESCS_PER_LOOP, just return no packet
+ */
+static uint16_t
+ice_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				    uint16_t nb_pkts)
+{
+	struct ice_rx_queue *rxq = rx_queue;
+	uint8_t split_flags[ICE_VPMD_RX_BURST] = {0};
+
+	/* get some new buffers */
+	uint16_t nb_bufs = _ice_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 + ice_rx_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 < ICE_DESCS_PER_LOOP, just return no packet
+ */
+uint16_t
+ice_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
+				   uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > ICE_VPMD_RX_BURST) {
+		uint16_t burst = ice_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, ICE_VPMD_RX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < ICE_VPMD_RX_BURST)
+			return retval;
+	}
+	return retval + ice_recv_scattered_burst_vec_avx512(rx_queue,
+				rx_pkts + retval, nb_pkts);
+}
+
+static inline void
+ice_vtx1(volatile struct ice_tx_desc *txdp,
+	 struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(ICE_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << ICE_TXD_QW1_CMD_S) |
+		 ((uint64_t)pkt->data_len << ICE_TXD_QW1_TX_BUF_SZ_S));
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+				pkt->buf_iova + pkt->data_off);
+	_mm_store_si128((__m128i *)txdp, descriptor);
+}
+
+static inline void
+ice_vtx(volatile struct ice_tx_desc *txdp,
+	struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (ICE_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << ICE_TXD_QW1_CMD_S));
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		ice_vtx1(txdp, *pkt, flags);
+		nb_pkts--, txdp++, pkt++;
+	}
+
+	/* do two at a time while possible, in bursts */
+	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
+		uint64_t hi_qw3 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[3]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw2 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[2]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw1 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[1]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw0 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[0]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+
+		__m256i desc2_3 =
+			_mm256_set_epi64x
+				(hi_qw3,
+				 pkt[3]->buf_iova + pkt[3]->data_off,
+				 hi_qw2,
+				 pkt[2]->buf_iova + pkt[2]->data_off);
+		__m256i desc0_1 =
+			_mm256_set_epi64x
+				(hi_qw1,
+				 pkt[1]->buf_iova + pkt[1]->data_off,
+				 hi_qw0,
+				 pkt[0]->buf_iova + pkt[0]->data_off);
+		_mm256_store_si256((void *)(txdp + 2), desc2_3);
+		_mm256_store_si256((void *)txdp, desc0_1);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		ice_vtx1(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+				uint16_t nb_pkts)
+{
+	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
+	volatile struct ice_tx_desc *txdp;
+	struct ice_tx_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	uint64_t flags = ICE_TD_CMD;
+	uint64_t rs = ICE_TX_DESC_CMD_RS | ICE_TD_CMD;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+
+	if (txq->nb_tx_free < txq->tx_free_thresh)
+		ice_tx_free_bufs(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = &txq->sw_ring[tx_id];
+
+	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		ice_tx_backlog_entry(txep, tx_pkts, n);
+
+		ice_vtx(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		ice_vtx1(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = &txq->sw_ring[tx_id];
+	}
+
+	ice_tx_backlog_entry(txep, tx_pkts, nb_commit);
+
+	ice_vtx(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->tx_next_rs) {
+		txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)ICE_TX_DESC_CMD_RS) <<
+					 ICE_TXD_QW1_CMD_S);
+		txq->tx_next_rs =
+			(uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	ICE_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+ice_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
+			 uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+		ret = ice_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;
+}
diff --git a/drivers/net/ice/meson.build b/drivers/net/ice/meson.build
index 254595af85..85e8baf912 100644
--- a/drivers/net/ice/meson.build
+++ b/drivers/net/ice/meson.build
@@ -34,6 +34,17 @@ if arch_subdir == 'x86'
 				c_args: [cflags, '-mavx2'])
 		objs += ice_avx2_lib.extract_objects('ice_rxtx_vec_avx2.c')
 	endif
+
+	if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F') or (not machine_args.contains('-mno-avx512f') and cc.has_argument('-mavx512f'))
+		cflags += ['-DCC_AVX512_SUPPORT']
+		ice_avx512_lib = static_library('ice_avx512_lib',
+				      'ice_rxtx_vec_avx512.c',
+				      dependencies: [static_rte_ethdev,
+					static_rte_kvargs, static_rte_hash],
+				      include_directories: includes,
+				      c_args: [cflags, '-march=skylake-avx512', '-mavx512f'])
+		objs += ice_avx512_lib.extract_objects('ice_rxtx_vec_avx512.c')
+	endif
 endif
 
 sources += files('ice_dcf.c',
-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v4 2/3] net/ice: add RSS hash parsing in AVX512 path
  2020-10-23  4:14 ` [dpdk-dev] [PATCH v4 0/3] AVX512 vPMD on ice Leyi Rong
  2020-10-23  4:14   ` [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path Leyi Rong
@ 2020-10-23  4:14   ` Leyi Rong
  2020-10-23  4:14   ` [dpdk-dev] [PATCH v4 3/3] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
  2020-10-23  9:39   ` [dpdk-dev] [PATCH v4 0/3] AVX512 vPMD on ice Zhang, Qi Z
  3 siblings, 0 replies; 25+ messages in thread
From: Leyi Rong @ 2020-10-23  4:14 UTC (permalink / raw)
  To: qi.z.zhang, wenzhuo.lu, burce.richardson; +Cc: dev, Leyi Rong

Support RSS hash parsing in AVX512 data path as the default
RXDID is set to #22, that means the RSS hash field locates
in the 2nd 16B of each Flex Rx descriptor.

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
 drivers/net/ice/ice_rxtx_vec_avx512.c | 105 ++++++++++++++++++++++++--
 1 file changed, 98 insertions(+), 7 deletions(-)

diff --git a/drivers/net/ice/ice_rxtx_vec_avx512.c b/drivers/net/ice/ice_rxtx_vec_avx512.c
index 73c4ffce99..bcef7fc800 100644
--- a/drivers/net/ice/ice_rxtx_vec_avx512.c
+++ b/drivers/net/ice/ice_rxtx_vec_avx512.c
@@ -176,8 +176,8 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
 	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
 	const __m512i shuf_msk =
 		_mm512_set4_epi32
-			(/* octet 12~15, 32 bits rss */
-			 15 << 24 | 14 << 16 | 13 << 8 | 12,
+			(/* rss hash parsed separately */
+			 0xFFFFFFFF,
 			 /* octet 10~11, 16 bits vlan_macip */
 			 /* octet 4~5, 16 bits data_len */
 			 11 << 24 | 10 << 16 | 5 << 8 | 4,
@@ -399,6 +399,11 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
 		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7);
 		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3);
 
+		__m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7, 0);
+		__m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7, 1);
+		__m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3, 0);
+		__m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3, 1);
+
 		/**
 		 * use permute/extract to get status content
 		 * After the operations, the packets status flags are in the
@@ -438,6 +443,97 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
 		/* merge flags */
 		const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
 				rss_vlan_flags);
+
+#ifndef RTE_LIBRTE_ICE_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.
@@ -471,11 +567,6 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq,
 					    _mm256_srli_si256(mbuf_flags, 4),
 					    0x04);
 
-		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);
-
 		/* 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);
-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* [dpdk-dev] [PATCH v4 3/3] net/ice: optimize Tx path on AVX512 vPMD
  2020-10-23  4:14 ` [dpdk-dev] [PATCH v4 0/3] AVX512 vPMD on ice Leyi Rong
  2020-10-23  4:14   ` [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path Leyi Rong
  2020-10-23  4:14   ` [dpdk-dev] [PATCH v4 2/3] net/ice: add RSS hash parsing in AVX512 path Leyi Rong
@ 2020-10-23  4:14   ` Leyi Rong
  2020-10-23  9:39   ` [dpdk-dev] [PATCH v4 0/3] AVX512 vPMD on ice Zhang, Qi Z
  3 siblings, 0 replies; 25+ messages in thread
From: Leyi Rong @ 2020-10-23  4:14 UTC (permalink / raw)
  To: qi.z.zhang, wenzhuo.lu, burce.richardson; +Cc: dev, Leyi Rong, Bruce Richardson

Optimize Tx path by using AVX512 instructions and vectorize the
tx free bufs process.

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 drivers/net/ice/ice_rxtx.h            |   4 +
 drivers/net/ice/ice_rxtx_vec_avx512.c | 147 ++++++++++++++++++++++----
 drivers/net/ice/ice_rxtx_vec_common.h |  36 +++++--
 3 files changed, 158 insertions(+), 29 deletions(-)

diff --git a/drivers/net/ice/ice_rxtx.h b/drivers/net/ice/ice_rxtx.h
index 7cebbc2916..23409d479a 100644
--- a/drivers/net/ice/ice_rxtx.h
+++ b/drivers/net/ice/ice_rxtx.h
@@ -96,6 +96,10 @@ struct ice_tx_entry {
 	uint16_t last_id;
 };
 
+struct ice_vec_tx_entry {
+	struct rte_mbuf *mbuf;
+};
+
 struct ice_tx_queue {
 	uint16_t nb_tx_desc; /* number of TX descriptors */
 	rte_iova_t tx_ring_dma; /* TX ring DMA address */
diff --git a/drivers/net/ice/ice_rxtx_vec_avx512.c b/drivers/net/ice/ice_rxtx_vec_avx512.c
index bcef7fc800..e5e7cc1482 100644
--- a/drivers/net/ice/ice_rxtx_vec_avx512.c
+++ b/drivers/net/ice/ice_rxtx_vec_avx512.c
@@ -756,6 +756,108 @@ ice_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
 				rx_pkts + retval, nb_pkts);
 }
 
+static __rte_always_inline int
+ice_tx_free_bufs_avx512(struct ice_tx_queue *txq)
+{
+	struct ice_vec_tx_entry *txep;
+	uint32_t n;
+	uint32_t i;
+	int nb_free = 0;
+	struct rte_mbuf *m, *free[ICE_TX_MAX_FREE_BUF_SZ];
+
+	/* check DD bits on threshold descriptor */
+	if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
+			rte_cpu_to_le_64(ICE_TXD_QW1_DTYPE_M)) !=
+			rte_cpu_to_le_64(ICE_TX_DESC_DTYPE_DESC_DONE))
+		return 0;
+
+	n = txq->tx_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->tx_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_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
+	txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
+	if (txq->tx_next_dd >= txq->nb_tx_desc)
+		txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
+
+	return txq->tx_rs_thresh;
+}
+
 static inline void
 ice_vtx1(volatile struct ice_tx_desc *txdp,
 	 struct rte_mbuf *pkt, uint64_t flags)
@@ -777,13 +879,6 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 	const uint64_t hi_qw_tmpl = (ICE_TX_DESC_DTYPE_DATA |
 			((uint64_t)flags  << ICE_TXD_QW1_CMD_S));
 
-	/* if unaligned on 32-bit boundary, do one to align */
-	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
-		ice_vtx1(txdp, *pkt, flags);
-		nb_pkts--, txdp++, pkt++;
-	}
-
-	/* do two at a time while possible, in bursts */
 	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
 		uint64_t hi_qw3 =
 			hi_qw_tmpl |
@@ -802,20 +897,17 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 			((uint64_t)pkt[0]->data_len <<
 			 ICE_TXD_QW1_TX_BUF_SZ_S);
 
-		__m256i desc2_3 =
-			_mm256_set_epi64x
+		__m512i desc0_3 =
+			_mm512_set_epi64
 				(hi_qw3,
 				 pkt[3]->buf_iova + pkt[3]->data_off,
 				 hi_qw2,
-				 pkt[2]->buf_iova + pkt[2]->data_off);
-		__m256i desc0_1 =
-			_mm256_set_epi64x
-				(hi_qw1,
+				 pkt[2]->buf_iova + pkt[2]->data_off,
+				 hi_qw1,
 				 pkt[1]->buf_iova + pkt[1]->data_off,
 				 hi_qw0,
 				 pkt[0]->buf_iova + pkt[0]->data_off);
-		_mm256_store_si256((void *)(txdp + 2), desc2_3);
-		_mm256_store_si256((void *)txdp, desc0_1);
+		_mm512_storeu_si512((void *)txdp, desc0_3);
 	}
 
 	/* do any last ones */
@@ -825,13 +917,23 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 	}
 }
 
+static __rte_always_inline void
+ice_tx_backlog_entry_avx512(struct ice_vec_tx_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 uint16_t
 ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 				uint16_t nb_pkts)
 {
 	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
 	volatile struct ice_tx_desc *txdp;
-	struct ice_tx_entry *txep;
+	struct ice_vec_tx_entry *txep;
 	uint16_t n, nb_commit, tx_id;
 	uint64_t flags = ICE_TD_CMD;
 	uint64_t rs = ICE_TX_DESC_CMD_RS | ICE_TD_CMD;
@@ -840,7 +942,7 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
 
 	if (txq->nb_tx_free < txq->tx_free_thresh)
-		ice_tx_free_bufs(txq);
+		ice_tx_free_bufs_avx512(txq);
 
 	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
 	if (unlikely(nb_pkts == 0))
@@ -848,13 +950,14 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 
 	tx_id = txq->tx_tail;
 	txdp = &txq->tx_ring[tx_id];
-	txep = &txq->sw_ring[tx_id];
+	txep = (void *)txq->sw_ring;
+	txep += tx_id;
 
 	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
 
 	n = (uint16_t)(txq->nb_tx_desc - tx_id);
 	if (nb_commit >= n) {
-		ice_tx_backlog_entry(txep, tx_pkts, n);
+		ice_tx_backlog_entry_avx512(txep, tx_pkts, n);
 
 		ice_vtx(txdp, tx_pkts, n - 1, flags);
 		tx_pkts += (n - 1);
@@ -868,11 +971,11 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 		txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
 
 		/* avoid reach the end of ring */
-		txdp = &txq->tx_ring[tx_id];
-		txep = &txq->sw_ring[tx_id];
+		txdp = txq->tx_ring;
+		txep = (void *)txq->sw_ring;
 	}
 
-	ice_tx_backlog_entry(txep, tx_pkts, nb_commit);
+	ice_tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
 
 	ice_vtx(txdp, tx_pkts, nb_commit, flags);
 
diff --git a/drivers/net/ice/ice_rxtx_vec_common.h b/drivers/net/ice/ice_rxtx_vec_common.h
index e2019c8d6a..ae2ac29f2a 100644
--- a/drivers/net/ice/ice_rxtx_vec_common.h
+++ b/drivers/net/ice/ice_rxtx_vec_common.h
@@ -189,16 +189,38 @@ _ice_tx_queue_release_mbufs_vec(struct ice_tx_queue *txq)
 	 *  so need to free remains more carefully.
 	 */
 	i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
-	if (txq->tx_tail < i) {
-		for (; i < txq->nb_tx_desc; i++) {
+
+#ifdef CC_AVX512_SUPPORT
+	struct rte_eth_dev *dev = txq->vsi->adapter->eth_dev;
+
+	if (dev->tx_pkt_burst == ice_xmit_pkts_vec_avx512) {
+		struct ice_vec_tx_entry *swr = (void *)txq->sw_ring;
+
+		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;
+		}
+		for (; i < txq->tx_tail; i++) {
+			rte_pktmbuf_free_seg(swr[i].mbuf);
+			swr[i].mbuf = NULL;
+		}
+	} else
+#endif
+	{
+		if (txq->tx_tail < i) {
+			for (; i < txq->nb_tx_desc; i++) {
+				rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
+				txq->sw_ring[i].mbuf = NULL;
+			}
+			i = 0;
+		}
+		for (; i < txq->tx_tail; i++) {
 			rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
 			txq->sw_ring[i].mbuf = NULL;
 		}
-		i = 0;
-	}
-	for (; i < txq->tx_tail; i++) {
-		rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
-		txq->sw_ring[i].mbuf = NULL;
 	}
 }
 
-- 
2.17.1


^ permalink raw reply	[flat|nested] 25+ messages in thread

* Re: [dpdk-dev] [PATCH v4 0/3] AVX512 vPMD on ice
  2020-10-23  4:14 ` [dpdk-dev] [PATCH v4 0/3] AVX512 vPMD on ice Leyi Rong
                     ` (2 preceding siblings ...)
  2020-10-23  4:14   ` [dpdk-dev] [PATCH v4 3/3] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
@ 2020-10-23  9:39   ` Zhang, Qi Z
  3 siblings, 0 replies; 25+ messages in thread
From: Zhang, Qi Z @ 2020-10-23  9:39 UTC (permalink / raw)
  To: Rong, Leyi, Lu, Wenzhuo, burce.richardson; +Cc: dev



> -----Original Message-----
> From: Rong, Leyi <leyi.rong@intel.com>
> Sent: Friday, October 23, 2020 12:14 PM
> To: Zhang, Qi Z <qi.z.zhang@intel.com>; Lu, Wenzhuo
> <wenzhuo.lu@intel.com>; burce.richardson@intel.com
> Cc: dev@dpdk.org; Rong, Leyi <leyi.rong@intel.com>
> Subject: [PATCH v4 0/3] AVX512 vPMD on ice
> 
> This patchset aims to support AVX512 vPMD on ice.
> 
> ---
> v4:
> - Replace buf_physaddr by buf_iova in the 1st patch to eliminate build
>   error.
> 
> v3:
> - Code rebased.
> 
> v2:
> - No internal judgement when RTE_MACHINE_CPUFLAG_AVX512F is set in
>   meson.build.
> - Add RSS hash parsing as default RXDID is set to #22.
> - Use buf_iova instead of buf_physaddr, as buf_physaddr will be removed.
> 
> Leyi Rong (3):
>   net/ice: add AVX512 vector path
>   net/ice: add RSS hash parsing in AVX512 path
>   net/ice: optimize Tx path on AVX512 vPMD
> 
>  drivers/net/ice/ice_rxtx.c            |   90 ++-
>  drivers/net/ice/ice_rxtx.h            |   11 +
>  drivers/net/ice/ice_rxtx_vec_avx512.c | 1018 +++++++++++++++++++++++++
>  drivers/net/ice/ice_rxtx_vec_common.h |   36 +-
>  drivers/net/ice/meson.build           |   11 +
>  5 files changed, 1141 insertions(+), 25 deletions(-)  create mode 100644
> drivers/net/ice/ice_rxtx_vec_avx512.c
> 
> --
> 2.17.1

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

Applied to dpdk-next-net-intel.

Thanks
Qi


^ permalink raw reply	[flat|nested] 25+ messages in thread

* Re: [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
  2020-10-23  4:14   ` [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path Leyi Rong
@ 2020-10-25 16:23     ` David Marchand
  2020-10-26  7:12       ` Rong, Leyi
  2020-10-27  8:32     ` Ali Alnubani
  1 sibling, 1 reply; 25+ messages in thread
From: David Marchand @ 2020-10-25 16:23 UTC (permalink / raw)
  To: Leyi Rong, Qi Zhang, Yigit, Ferruh
  Cc: Wenzhuo Lu, dev, Bruce Richardson, Thomas Monjalon

Hello Leyi, Qi, Ferruh,

On Fri, Oct 23, 2020 at 6:37 AM Leyi Rong <leyi.rong@intel.com> wrote:
> diff --git a/drivers/net/ice/meson.build b/drivers/net/ice/meson.build
> index 254595af85..85e8baf912 100644
> --- a/drivers/net/ice/meson.build
> +++ b/drivers/net/ice/meson.build
> @@ -34,6 +34,17 @@ if arch_subdir == 'x86'
>                                 c_args: [cflags, '-mavx2'])
>                 objs += ice_avx2_lib.extract_objects('ice_rxtx_vec_avx2.c')
>         endif
> +
> +       if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F') or (not machine_args.contains('-mno-avx512f') and cc.has_argument('-mavx512f'))
> +               cflags += ['-DCC_AVX512_SUPPORT']
> +               ice_avx512_lib = static_library('ice_avx512_lib',
> +                                     'ice_rxtx_vec_avx512.c',
> +                                     dependencies: [static_rte_ethdev,
> +                                       static_rte_kvargs, static_rte_hash],
> +                                     include_directories: includes,
> +                                     c_args: [cflags, '-march=skylake-avx512', '-mavx512f'])
> +               objs += ice_avx512_lib.extract_objects('ice_rxtx_vec_avx512.c')
> +       endif
>  endif
>
>  sources += files('ice_dcf.c',
> --
> 2.17.1
>

RTE_MACHINE_CPUFLAG_AVX512F can be removed.
Worth fixing before hitting the main tree.

Thanks.

-- 
David Marchand


^ permalink raw reply	[flat|nested] 25+ messages in thread

* Re: [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
  2020-10-25 16:23     ` David Marchand
@ 2020-10-26  7:12       ` Rong, Leyi
  2020-10-26  8:09         ` David Marchand
  2020-10-27 10:22         ` Ferruh Yigit
  0 siblings, 2 replies; 25+ messages in thread
From: Rong, Leyi @ 2020-10-26  7:12 UTC (permalink / raw)
  To: David Marchand, Zhang, Qi Z, Yigit, Ferruh
  Cc: Lu, Wenzhuo, dev, Richardson, Bruce, Thomas Monjalon


> -----Original Message-----
> From: David Marchand <david.marchand@redhat.com>
> Sent: Monday, October 26, 2020 12:24 AM
> To: Rong, Leyi <leyi.rong@intel.com>; Zhang, Qi Z <qi.z.zhang@intel.com>; Yigit,
> Ferruh <ferruh.yigit@intel.com>
> Cc: Lu, Wenzhuo <wenzhuo.lu@intel.com>; dev <dev@dpdk.org>; Richardson,
> Bruce <bruce.richardson@intel.com>; Thomas Monjalon
> <thomas@monjalon.net>
> Subject: Re: [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
> 
> Hello Leyi, Qi, Ferruh,
> 
> On Fri, Oct 23, 2020 at 6:37 AM Leyi Rong <leyi.rong@intel.com> wrote:
> > diff --git a/drivers/net/ice/meson.build b/drivers/net/ice/meson.build
> > index 254595af85..85e8baf912 100644
> > --- a/drivers/net/ice/meson.build
> > +++ b/drivers/net/ice/meson.build
> > @@ -34,6 +34,17 @@ if arch_subdir == 'x86'
> >                                 c_args: [cflags, '-mavx2'])
> >                 objs += ice_avx2_lib.extract_objects('ice_rxtx_vec_avx2.c')
> >         endif
> > +
> > +       if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F') or (not
> machine_args.contains('-mno-avx512f') and cc.has_argument('-mavx512f'))
> > +               cflags += ['-DCC_AVX512_SUPPORT']
> > +               ice_avx512_lib = static_library('ice_avx512_lib',
> > +                                     'ice_rxtx_vec_avx512.c',
> > +                                     dependencies: [static_rte_ethdev,
> > +                                       static_rte_kvargs, static_rte_hash],
> > +                                     include_directories: includes,
> > +                                     c_args: [cflags, '-march=skylake-avx512', '-mavx512f'])
> > +               objs += ice_avx512_lib.extract_objects('ice_rxtx_vec_avx512.c')
> > +       endif
> >  endif
> >
> >  sources += files('ice_dcf.c',
> > --
> > 2.17.1
> >
> 
> RTE_MACHINE_CPUFLAG_AVX512F can be removed.
> Worth fixing before hitting the main tree.
> 
> Thanks.
> 
> --
> David Marchand

Hello David,

Would prefer using __AVX512F__ instead of RTE_MACHINE_CPUFLAG_AVX512F here rather than remove the RTE_MACHINE_CPUFLAG_ macro directly to check the CPU capability.
So the judgment statement will be 
if cc.get_define('__AVX512F__', args: machine_args) != '' or (not machine_args.contains('-mno-avx512f') and cc.has_argument('-mavx512f'))

what do you think?


Hello Ferruh,

As the patchset is already merged into dpdk-next-net, I'm going to make another patch for this if it's accepted?

^ permalink raw reply	[flat|nested] 25+ messages in thread

* Re: [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
  2020-10-26  7:12       ` Rong, Leyi
@ 2020-10-26  8:09         ` David Marchand
  2020-10-27 10:19           ` Bruce Richardson
  2020-10-27 10:22         ` Ferruh Yigit
  1 sibling, 1 reply; 25+ messages in thread
From: David Marchand @ 2020-10-26  8:09 UTC (permalink / raw)
  To: Rong, Leyi, Richardson, Bruce
  Cc: Zhang, Qi Z, Yigit, Ferruh, Lu, Wenzhuo, dev, Thomas Monjalon

On Mon, Oct 26, 2020 at 8:13 AM Rong, Leyi <leyi.rong@intel.com> wrote:
> Would prefer using __AVX512F__ instead of RTE_MACHINE_CPUFLAG_AVX512F here rather than remove the RTE_MACHINE_CPUFLAG_ macro directly to check the CPU capability.
> So the judgment statement will be
> if cc.get_define('__AVX512F__', args: machine_args) != '' or (not machine_args.contains('-mno-avx512f') and cc.has_argument('-mavx512f'))
>
> what do you think?

No opinion as I have yet to understand the subtleties to control
enablement of avx stuff.
Bruce?


-- 
David Marchand


^ permalink raw reply	[flat|nested] 25+ messages in thread

* Re: [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
  2020-10-23  4:14   ` [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path Leyi Rong
  2020-10-25 16:23     ` David Marchand
@ 2020-10-27  8:32     ` Ali Alnubani
  2020-10-27  8:42       ` Ali Alnubani
  1 sibling, 1 reply; 25+ messages in thread
From: Ali Alnubani @ 2020-10-27  8:32 UTC (permalink / raw)
  To: Leyi Rong, qi.z.zhang, wenzhuo.lu, burce.richardson; +Cc: dev, Bruce Richardson

Hi,

> -----Original Message-----
> From: dev <dev-bounces@dpdk.org> On Behalf Of Leyi Rong
> Sent: Friday, October 23, 2020 7:14 AM
> To: qi.z.zhang@intel.com; wenzhuo.lu@intel.com;
> burce.richardson@intel.com
> Cc: dev@dpdk.org; Leyi Rong <leyi.rong@intel.com>; Bruce Richardson
> <bruce.richardson@intel.com>
> Subject: [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
> 
> Add AVX512 support for ice PMD. This patch adds ice_rxtx_vec_avx512.c to
> support ice AVX512 vPMD.
> 
> This patch aims to enable AVX512 on ice vPMD. Main changes are focus on Rx
> path compared with AVX2 vPMD.
> 
> Signed-off-by: Leyi Rong <leyi.rong@intel.com>
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> ---

This patch is causing the following build failure on CentOS 7 with gcc 4.8.5:

"""
drivers/librte_net_ice.a(net_ice_ice_rxtx.c.o): In function `ice_set_rx_function':
drivers/net/ice/ice_rxtx.c:3034: undefined reference to `ice_recv_scattered_pkts_vec_avx512'
drivers/net/ice/ice_rxtx.c:3050: undefined reference to `ice_recv_pkts_vec_avx512'
drivers/librte_net_ice.a(net_ice_ice_rxtx.c.o): In function `ice_set_tx_function':
drivers/net/ice/ice_rxtx.c:3243: undefined reference to `ice_xmit_pkts_vec_avx512'
"""

It reproduces when building with "meson --werror --buildtype=debug build && ninja -C build".

Regards,
Ali

^ permalink raw reply	[flat|nested] 25+ messages in thread

* Re: [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
  2020-10-27  8:32     ` Ali Alnubani
@ 2020-10-27  8:42       ` Ali Alnubani
  0 siblings, 0 replies; 25+ messages in thread
From: Ali Alnubani @ 2020-10-27  8:42 UTC (permalink / raw)
  To: Leyi Rong, qi.z.zhang, wenzhuo.lu, burce.richardson; +Cc: dev, Bruce Richardson

> -----Original Message-----
> From: dev <dev-bounces@dpdk.org> On Behalf Of Ali Alnubani
> Sent: Tuesday, October 27, 2020 10:33 AM
> To: Leyi Rong <leyi.rong@intel.com>; qi.z.zhang@intel.com;
> wenzhuo.lu@intel.com; burce.richardson@intel.com
> Cc: dev@dpdk.org; Bruce Richardson <bruce.richardson@intel.com>
> Subject: Re: [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
> 
> Hi,
> 
> > -----Original Message-----
> > From: dev <dev-bounces@dpdk.org> On Behalf Of Leyi Rong
> > Sent: Friday, October 23, 2020 7:14 AM
> > To: qi.z.zhang@intel.com; wenzhuo.lu@intel.com;
> > burce.richardson@intel.com
> > Cc: dev@dpdk.org; Leyi Rong <leyi.rong@intel.com>; Bruce Richardson
> > <bruce.richardson@intel.com>
> > Subject: [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
> >
> > Add AVX512 support for ice PMD. This patch adds ice_rxtx_vec_avx512.c
> > to support ice AVX512 vPMD.
> >
> > This patch aims to enable AVX512 on ice vPMD. Main changes are focus
> > on Rx path compared with AVX2 vPMD.
> >
> > Signed-off-by: Leyi Rong <leyi.rong@intel.com>
> > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > ---
> 
> This patch is causing the following build failure on CentOS 7 with gcc 4.8.5:
> 
> """
> drivers/librte_net_ice.a(net_ice_ice_rxtx.c.o): In function
> `ice_set_rx_function':
> drivers/net/ice/ice_rxtx.c:3034: undefined reference to
> `ice_recv_scattered_pkts_vec_avx512'
> drivers/net/ice/ice_rxtx.c:3050: undefined reference to
> `ice_recv_pkts_vec_avx512'
> drivers/librte_net_ice.a(net_ice_ice_rxtx.c.o): In function
> `ice_set_tx_function':
> drivers/net/ice/ice_rxtx.c:3243: undefined reference to
> `ice_xmit_pkts_vec_avx512'
> """
> 
> It reproduces when building with "meson --werror --buildtype=debug build
> && ninja -C build".
> 

It's also causing the following build failure on Ubuntu 18.04 (gcc 7.5.0) and on OpenSUSE Leap 15.2 (gcc 7.5.0):

"""
drivers/net/ice/ice_rxtx_vec_avx512.c: In function '_ice_recv_raw_pkts_vec_avx512':
/usr/lib/gcc/x86_64-linux-gnu/7/include/avx512bwintrin.h:1831:1: error: inlining failed in call to always_inline '_mm512_shuffle_epi8': target specific option mismatch
 _mm512_shuffle_epi8 (__m512i __A, __m512i __B)
 ^~~~~~~~~~~~~~~~~~~
drivers/net/ice/ice_rxtx_vec_avx512.c:350:11: note: called from here
   __m512i mb0_3 = _mm512_shuffle_epi8(raw_desc0_3, shuf_msk);
           ^~~~~
"""

Regards,
Ali

^ permalink raw reply	[flat|nested] 25+ messages in thread

* Re: [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
  2020-10-26  8:09         ` David Marchand
@ 2020-10-27 10:19           ` Bruce Richardson
  0 siblings, 0 replies; 25+ messages in thread
From: Bruce Richardson @ 2020-10-27 10:19 UTC (permalink / raw)
  To: David Marchand
  Cc: Rong, Leyi, Zhang, Qi Z, Yigit, Ferruh, Lu, Wenzhuo, dev,
	Thomas Monjalon

On Mon, Oct 26, 2020 at 09:09:31AM +0100, David Marchand wrote:
> On Mon, Oct 26, 2020 at 8:13 AM Rong, Leyi <leyi.rong@intel.com> wrote:
> > Would prefer using __AVX512F__ instead of RTE_MACHINE_CPUFLAG_AVX512F here rather than remove the RTE_MACHINE_CPUFLAG_ macro directly to check the CPU capability.
> > So the judgment statement will be
> > if cc.get_define('__AVX512F__', args: machine_args) != '' or (not machine_args.contains('-mno-avx512f') and cc.has_argument('-mavx512f'))
> >
> > what do you think?
> 
> No opinion as I have yet to understand the subtleties to control
> enablement of avx stuff.
> Bruce?
> 
The above proposal is correct, in that we now check the compiler flags
directly rather than using our own custom flags.

/Bruce

^ permalink raw reply	[flat|nested] 25+ messages in thread

* Re: [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
  2020-10-26  7:12       ` Rong, Leyi
  2020-10-26  8:09         ` David Marchand
@ 2020-10-27 10:22         ` Ferruh Yigit
  1 sibling, 0 replies; 25+ messages in thread
From: Ferruh Yigit @ 2020-10-27 10:22 UTC (permalink / raw)
  To: Rong, Leyi, David Marchand, Zhang, Qi Z
  Cc: Lu, Wenzhuo, dev, Richardson, Bruce, Thomas Monjalon

On 10/26/2020 7:12 AM, Rong, Leyi wrote:
> 
>> -----Original Message-----
>> From: David Marchand <david.marchand@redhat.com>
>> Sent: Monday, October 26, 2020 12:24 AM
>> To: Rong, Leyi <leyi.rong@intel.com>; Zhang, Qi Z <qi.z.zhang@intel.com>; Yigit,
>> Ferruh <ferruh.yigit@intel.com>
>> Cc: Lu, Wenzhuo <wenzhuo.lu@intel.com>; dev <dev@dpdk.org>; Richardson,
>> Bruce <bruce.richardson@intel.com>; Thomas Monjalon
>> <thomas@monjalon.net>
>> Subject: Re: [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path
>>
>> Hello Leyi, Qi, Ferruh,
>>
>> On Fri, Oct 23, 2020 at 6:37 AM Leyi Rong <leyi.rong@intel.com> wrote:
>>> diff --git a/drivers/net/ice/meson.build b/drivers/net/ice/meson.build
>>> index 254595af85..85e8baf912 100644
>>> --- a/drivers/net/ice/meson.build
>>> +++ b/drivers/net/ice/meson.build
>>> @@ -34,6 +34,17 @@ if arch_subdir == 'x86'
>>>                                  c_args: [cflags, '-mavx2'])
>>>                  objs += ice_avx2_lib.extract_objects('ice_rxtx_vec_avx2.c')
>>>          endif
>>> +
>>> +       if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F') or (not
>> machine_args.contains('-mno-avx512f') and cc.has_argument('-mavx512f'))
>>> +               cflags += ['-DCC_AVX512_SUPPORT']
>>> +               ice_avx512_lib = static_library('ice_avx512_lib',
>>> +                                     'ice_rxtx_vec_avx512.c',
>>> +                                     dependencies: [static_rte_ethdev,
>>> +                                       static_rte_kvargs, static_rte_hash],
>>> +                                     include_directories: includes,
>>> +                                     c_args: [cflags, '-march=skylake-avx512', '-mavx512f'])
>>> +               objs += ice_avx512_lib.extract_objects('ice_rxtx_vec_avx512.c')
>>> +       endif
>>>   endif
>>>
>>>   sources += files('ice_dcf.c',
>>> --
>>> 2.17.1
>>>
>>
>> RTE_MACHINE_CPUFLAG_AVX512F can be removed.
>> Worth fixing before hitting the main tree.
>>
>> Thanks.
>>
>> --
>> David Marchand
> 
> Hello David,
> 
> Would prefer using __AVX512F__ instead of RTE_MACHINE_CPUFLAG_AVX512F here rather than remove the RTE_MACHINE_CPUFLAG_ macro directly to check the CPU capability.
> So the judgment statement will be
> if cc.get_define('__AVX512F__', args: machine_args) != '' or (not machine_args.contains('-mno-avx512f') and cc.has_argument('-mavx512f'))
> 
> what do you think?
> 
> 
> Hello Ferruh,
> 
> As the patchset is already merged into dpdk-next-net, I'm going to make another patch for this if it's accepted?
> 

Hi Leyi,

Please make another patch, I can squash it in the next-net.

Thanks,
ferruh

^ permalink raw reply	[flat|nested] 25+ messages in thread

end of thread, other threads:[~2020-10-27 10:22 UTC | newest]

Thread overview: 25+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-09-10  6:55 [dpdk-dev] [PATCH v1 0/2] AVX512 vPMD on ice Leyi Rong
2020-09-10  6:55 ` [dpdk-dev] [PATCH v1 1/2] net/ice: add AVX512 vector path Leyi Rong
2020-09-10  9:32   ` Bruce Richardson
2020-09-10  6:55 ` [dpdk-dev] [PATCH v1 2/2] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
2020-09-15  1:17   ` Wang, Haiyue
2020-09-18  3:35 ` [dpdk-dev] [PATCH v2 0/3] AVX512 vPMD on ice Leyi Rong
2020-09-18  3:35   ` [dpdk-dev] [PATCH v2 1/3] net/ice: add AVX512 vector path Leyi Rong
2020-09-18  3:35   ` [dpdk-dev] [PATCH v2 2/3] net/ice: add RSS hash parsing in AVX512 path Leyi Rong
2020-09-18  3:35   ` [dpdk-dev] [PATCH v2 3/3] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
2020-10-20 10:51 ` [dpdk-dev] [PATCH v3 0/3] AVX512 vPMD on ice Leyi Rong
2020-10-20 10:51   ` [dpdk-dev] [PATCH v3 1/3] net/ice: add AVX512 vector path Leyi Rong
2020-10-20 10:51   ` [dpdk-dev] [PATCH v3 2/3] net/ice: add RSS hash parsing in AVX512 path Leyi Rong
2020-10-20 10:51   ` [dpdk-dev] [PATCH v3 3/3] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
2020-10-23  4:14 ` [dpdk-dev] [PATCH v4 0/3] AVX512 vPMD on ice Leyi Rong
2020-10-23  4:14   ` [dpdk-dev] [PATCH v4 1/3] net/ice: add AVX512 vector path Leyi Rong
2020-10-25 16:23     ` David Marchand
2020-10-26  7:12       ` Rong, Leyi
2020-10-26  8:09         ` David Marchand
2020-10-27 10:19           ` Bruce Richardson
2020-10-27 10:22         ` Ferruh Yigit
2020-10-27  8:32     ` Ali Alnubani
2020-10-27  8:42       ` Ali Alnubani
2020-10-23  4:14   ` [dpdk-dev] [PATCH v4 2/3] net/ice: add RSS hash parsing in AVX512 path Leyi Rong
2020-10-23  4:14   ` [dpdk-dev] [PATCH v4 3/3] net/ice: optimize Tx path on AVX512 vPMD Leyi Rong
2020-10-23  9:39   ` [dpdk-dev] [PATCH v4 0/3] AVX512 vPMD on ice 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).