RE: [PATCH v2 13/14] net/idpf: add AVX512 data path for single queue model

["Wu, Wenjun1" <wenjun1.wu@intel.com>]

> -----Original Message-----
> From: Guo, Junfeng <junfeng.guo@intel.com>
> Sent: Monday, September 5, 2022 6:58 PM
> To: Zhang, Qi Z <qi.z.zhang@intel.com>; Wu, Jingjing
> <jingjing.wu@intel.com>; Xing, Beilei <beilei.xing@intel.com>
> Cc: dev@dpdk.org; Wang, Xiao W <xiao.w.wang@intel.com>; Guo, Junfeng
> <junfeng.guo@intel.com>; Wu, Wenjun1 <wenjun1.wu@intel.com>
> Subject: [PATCH v2 13/14] net/idpf: add AVX512 data path for single queue
> model
> 
> Add support of AVX512 vector data path for single queue model.
> 
> Signed-off-by: Wenjun Wu <wenjun1.wu@intel.com>
> Signed-off-by: Junfeng Guo <junfeng.guo@intel.com>
> ---
>  drivers/net/idpf/idpf_ethdev.h          |   5 +
>  drivers/net/idpf/idpf_rxtx.c            | 137 ++++
>  drivers/net/idpf/idpf_rxtx.h            |  11 +
>  drivers/net/idpf/idpf_rxtx_vec_avx512.c | 917 ++++++++++++++++++++++++
> drivers/net/idpf/idpf_rxtx_vec_common.h |  89 +++
>  drivers/net/idpf/meson.build            |  31 +-
>  6 files changed, 1189 insertions(+), 1 deletion(-)  create mode 100644
> drivers/net/idpf/idpf_rxtx_vec_avx512.c
>  create mode 100644 drivers/net/idpf/idpf_rxtx_vec_common.h
> 
> diff --git a/drivers/net/idpf/idpf_ethdev.h b/drivers/net/idpf/idpf_ethdev.h
> index f96867f3d5..a32d5758ac 100644
> --- a/drivers/net/idpf/idpf_ethdev.h
> +++ b/drivers/net/idpf/idpf_ethdev.h
> @@ -179,6 +179,11 @@ struct idpf_adapter {
>  	uint32_t ptype_tbl[IDPF_MAX_PKT_TYPE] __rte_cache_min_aligned;
> 
>  	bool stopped;
> +
> +	bool rx_vec_allowed;
> +	bool tx_vec_allowed;
> +	bool rx_use_avx512;
> +	bool tx_use_avx512;
>  };
> 
>  TAILQ_HEAD(idpf_adapter_list, idpf_adapter); diff --git
> a/drivers/net/idpf/idpf_rxtx.c b/drivers/net/idpf/idpf_rxtx.c index
> 54d83a7c61..e31d202646 100644
> --- a/drivers/net/idpf/idpf_rxtx.c
> +++ b/drivers/net/idpf/idpf_rxtx.c
> @@ -4,9 +4,11 @@
> 
>  #include <ethdev_driver.h>
>  #include <rte_net.h>
> +#include <rte_vect.h>
> 
>  #include "idpf_ethdev.h"
>  #include "idpf_rxtx.h"
> +#include "idpf_rxtx_vec_common.h"
> 
>  const uint32_t *
>  idpf_dev_supported_ptypes_get(struct rte_eth_dev *dev __rte_unused)
> @@ -2146,15 +2148,110 @@ idpf_prep_pkts(__rte_unused void *tx_queue,
> struct rte_mbuf **tx_pkts,
>  	return i;
>  }
> 
> +static void __rte_cold
> +release_rxq_mbufs_vec(struct idpf_rx_queue *rxq) {
> +	const uint16_t mask = rxq->nb_rx_desc - 1;
> +	uint16_t i;
> +
> +	if (!rxq->sw_ring || rxq->rxrearm_nb >= rxq->nb_rx_desc)
> +		return;
> +
> +	/* free all mbufs that are valid in the ring */
> +	if (rxq->rxrearm_nb == 0) {
> +		for (i = 0; i < rxq->nb_rx_desc; i++) {
> +			if (rxq->sw_ring[i])
> +				rte_pktmbuf_free_seg(rxq->sw_ring[i]);
> +		}
> +	} else {
> +		for (i = rxq->rx_tail; i != rxq->rxrearm_start; i = (i + 1) & mask)
> {
> +			if (rxq->sw_ring[i])
> +				rte_pktmbuf_free_seg(rxq->sw_ring[i]);
> +		}
> +	}
> +
> +	rxq->rxrearm_nb = rxq->nb_rx_desc;
> +
> +	/* set all entries to NULL */
> +	memset(rxq->sw_ring, 0, sizeof(rxq->sw_ring[0]) * rxq-
> >nb_rx_desc); }
> +
> +static const struct idpf_rxq_ops def_singleq_rx_ops_vec = {
> +	.release_mbufs = release_rxq_mbufs_vec, };
> +
> +static inline int
> +idpf_singleq_rx_vec_setup_default(struct idpf_rx_queue *rxq) {
> +	uintptr_t p;
> +	struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
> +
> +	mb_def.nb_segs = 1;
> +	mb_def.data_off = RTE_PKTMBUF_HEADROOM;
> +	mb_def.port = rxq->port_id;
> +	rte_mbuf_refcnt_set(&mb_def, 1);
> +
> +	/* prevent compiler reordering: rearm_data covers previous fields */
> +	rte_compiler_barrier();
> +	p = (uintptr_t)&mb_def.rearm_data;
> +	rxq->mbuf_initializer = *(uint64_t *)p;
> +	return 0;
> +}
> +
> +int __rte_cold
> +idpf_singleq_rx_vec_setup(struct idpf_rx_queue *rxq) {
> +	rxq->ops = &def_singleq_rx_ops_vec;
> +	return idpf_singleq_rx_vec_setup_default(rxq);
> +}
> +
>  void
>  idpf_set_rx_function(struct rte_eth_dev *dev)  {
>  	struct idpf_vport *vport = dev->data->dev_private;
> +	struct idpf_adapter *ad = vport->adapter;
> +	struct idpf_rx_queue *rxq;
> +	int i;
> +
> +#ifdef RTE_ARCH_X86
> +	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
> +		if (idpf_rx_vec_dev_check_default(dev) ==
> IDPF_VECTOR_PATH &&
> +		    rte_vect_get_max_simd_bitwidth() >=
> RTE_VECT_SIMD_128) {
> +			ad->rx_vec_allowed = true;
> +
> +			if (rte_vect_get_max_simd_bitwidth() >=
> RTE_VECT_SIMD_512) #ifdef
> +CC_AVX512_SUPPORT
> +				if
> (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1 &&
> +
> rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512BW) == 1)
> +					ad->rx_use_avx512 = true;
> +#else
> +			PMD_DRV_LOG(NOTICE,
> +				"AVX512 is not supported in build env");
> #endif /*
> +CC_AVX512_SUPPORT */
> +		} else {
> +			ad->rx_vec_allowed = false;
> +		}
> +	}
> +#endif /* RTE_ARCH_X86 */
> 
>  	if (vport->rxq_model == VIRTCHNL2_QUEUE_MODEL_SPLIT) {
>  		dev->rx_pkt_burst = idpf_splitq_recv_pkts;
>  		return;
>  	} else {
> +#ifdef RTE_ARCH_X86
> +		if (ad->rx_vec_allowed) {
> +			for (i = 0; i < dev->data->nb_tx_queues; i++) {
> +				rxq = dev->data->rx_queues[i];
> +				(void)idpf_singleq_rx_vec_setup(rxq);
> +			}
> +#ifdef CC_AVX512_SUPPORT
> +			if (ad->rx_use_avx512) {
> +				dev->rx_pkt_burst =
> idpf_singleq_recv_pkts_avx512;
> +				return;
> +			}
> +#endif /* CC_AVX512_SUPPORT */
> +		}
> +#endif /* RTE_ARCH_X86 */
>  		dev->rx_pkt_burst = idpf_singleq_recv_pkts;
>  		return;
>  	}
> @@ -2164,12 +2261,52 @@ void
>  idpf_set_tx_function(struct rte_eth_dev *dev)  {
>  	struct idpf_vport *vport = dev->data->dev_private;
> +	struct idpf_adapter *ad = vport->adapter;
> +	struct idpf_tx_queue *txq;
> +	int i;
> +
> +#ifdef RTE_ARCH_X86
> +	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
> +		if (idpf_rx_vec_dev_check_default(dev) ==
> IDPF_VECTOR_PATH &&
> +		    rte_vect_get_max_simd_bitwidth() >=
> RTE_VECT_SIMD_128) {
> +			ad->tx_vec_allowed = true;
> +			if (rte_vect_get_max_simd_bitwidth() >=
> RTE_VECT_SIMD_512) #ifdef
> +CC_AVX512_SUPPORT
> +				if
> (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1 &&
> +
> rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512BW) == 1)
> +					ad->tx_use_avx512 = true;
> +#else
> +				PMD_DRV_LOG(NOTICE,
> +					"AVX512 is not supported in build
> env"); #endif /*
> +CC_AVX512_SUPPORT */
> +		} else {
> +			ad->tx_vec_allowed = false;
> +		}
> +	}
> +#endif /* RTE_ARCH_X86 */
> 
>  	if (vport->txq_model == VIRTCHNL2_QUEUE_MODEL_SPLIT) {
>  		dev->tx_pkt_burst = idpf_splitq_xmit_pkts;
>  		dev->tx_pkt_prepare = idpf_prep_pkts;
>  		return;
>  	} else {
> +#ifdef RTE_ARCH_X86
> +		if (ad->tx_vec_allowed) {
> +#ifdef CC_AVX512_SUPPORT
> +			if (ad->tx_use_avx512) {
> +				for (i = 0; i < dev->data->nb_tx_queues; i++) {
> +					txq = dev->data->tx_queues[i];
> +					if (!txq)
> +						continue;
> +
> 	idpf_singleq_tx_vec_setup_avx512(txq);
> +				}
> +				dev->tx_pkt_burst =
> idpf_singleq_xmit_pkts_avx512;
> +				dev->tx_pkt_prepare = idpf_prep_pkts;
> +				return;
> +			}
> +#endif /* CC_AVX512_SUPPORT */
> +		}
> +#endif /* RTE_ARCH_X86 */
>  		dev->tx_pkt_burst = idpf_singleq_xmit_pkts;
>  		dev->tx_pkt_prepare = idpf_prep_pkts;
>  		return;
> diff --git a/drivers/net/idpf/idpf_rxtx.h b/drivers/net/idpf/idpf_rxtx.h index
> 3ccf9efe50..decd0a98c2 100644
> --- a/drivers/net/idpf/idpf_rxtx.h
> +++ b/drivers/net/idpf/idpf_rxtx.h
> @@ -122,6 +122,10 @@ struct idpf_tx_entry {
>  	uint16_t last_id;
>  };
> 
> +struct idpf_tx_vec_entry {
> +	struct rte_mbuf *mbuf;
> +};
> +
>  /* Structure associated with each TX queue. */  struct idpf_tx_queue {
>  	const struct rte_memzone *mz;		/* memzone for Tx
> ring */
> @@ -207,12 +211,19 @@ uint16_t idpf_singleq_recv_pkts(void *rx_queue,
> struct rte_mbuf **rx_pkts,
>  				uint16_t nb_pkts);
>  uint16_t idpf_splitq_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
>  			       uint16_t nb_pkts);
> +uint16_t idpf_singleq_recv_pkts_avx512(void *rx_queue, struct rte_mbuf
> **rx_pkts,
> +				       uint16_t nb_pkts);
>  uint16_t idpf_singleq_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
>  				uint16_t nb_pkts);
>  uint16_t idpf_splitq_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
>  			       uint16_t nb_pkts);
> +uint16_t idpf_singleq_xmit_pkts_avx512(void *tx_queue, struct rte_mbuf
> **tx_pkts,
> +				       uint16_t nb_pkts);
> +int idpf_singleq_tx_vec_setup_avx512(struct idpf_tx_queue *txq);
>  uint16_t idpf_prep_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
>  			uint16_t nb_pkts);
> +int idpf_singleq_rx_vec_setup(struct idpf_rx_queue *rxq);
> +
>  void idpf_stop_queues(struct rte_eth_dev *dev);
> 
>  void idpf_set_rx_function(struct rte_eth_dev *dev); diff --git
> a/drivers/net/idpf/idpf_rxtx_vec_avx512.c
> b/drivers/net/idpf/idpf_rxtx_vec_avx512.c
> new file mode 100644
> index 0000000000..70eb5e1e7e
> --- /dev/null
> +++ b/drivers/net/idpf/idpf_rxtx_vec_avx512.c
> @@ -0,0 +1,917 @@
> +/* SPDX-License-Identifier: BSD-3-Clause
> + * Copyright(c) 2022 Intel Corporation
> + */
> +
> +#include "idpf_rxtx_vec_common.h"
> +
> +#include <rte_vect.h>
> +
> +#ifndef __INTEL_COMPILER
> +#pragma GCC diagnostic ignored "-Wcast-qual"
> +#endif
> +
> +#define IDPF_DESCS_PER_LOOP_AVX 8
> +#define PKTLEN_SHIFT 10
> +
> +/**************************************************************
> ********
> +********
> + * If user knows a specific offload is not enabled by APP,
> + * the macro can be commented to save the effort of fast path.
> + * Currently below 2 features are supported in RX path,
> + * 1, checksum offload
> + * 2, VLAN/QINQ stripping
> + * 3, RSS hash
> + * 4, packet type analysis
> + * 5, flow director ID report
> +
> +***************************************************************
> ********
> +*******/
> +#define IDPF_RX_PTYPE_OFFLOAD

The comments should be reworked since it does not fully match the code.

> +
> +static __rte_always_inline void
> +idpf_singleq_rearm_common(struct idpf_rx_queue *rxq) {
> +	struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start];
> +	volatile union virtchnl2_rx_desc *rxdp = rxq->rx_ring;
> +	uint16_t rx_id;
> +	int i;
> +
> +	rxdp += rxq->rxrearm_start;
> +
> +	/* Pull 'n' more MBUFs into the software ring */
> +	if (rte_mempool_get_bulk(rxq->mp,
> +				 (void *)rxp,
> +				 IDPF_RXQ_REARM_THRESH) < 0) {
> +		if (rxq->rxrearm_nb + IDPF_RXQ_REARM_THRESH >=
> +		    rxq->nb_rx_desc) {
> +			__m128i dma_addr0;
> +
> +			dma_addr0 = _mm_setzero_si128();
> +			for (i = 0; i < IDPF_VPMD_DESCS_PER_LOOP; i++) {
> +				rxp[i] = &rxq->fake_mbuf;
> +				_mm_store_si128((__m128i *)&rxdp[i].read,
> +						dma_addr0);
> +			}
> +		}
> +		rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
> +			IDPF_RXQ_REARM_THRESH;
> +		return;
> +	}
> +	struct rte_mbuf *mb0, *mb1, *mb2, *mb3;
> +	struct rte_mbuf *mb4, *mb5, *mb6, *mb7;
> +	__m512i dma_addr0_3, dma_addr4_7;
> +	__m512i hdr_room =
> _mm512_set1_epi64(RTE_PKTMBUF_HEADROOM);
> +	/* Initialize the mbufs in vector, process 8 mbufs in one loop */
> +	for (i = 0; i < IDPF_RXQ_REARM_THRESH;
> +			i += 8, rxp += 8, rxdp += 8) {
> +		__m128i vaddr0, vaddr1, vaddr2, vaddr3;
> +		__m128i vaddr4, vaddr5, vaddr6, vaddr7;
> +		__m256i vaddr0_1, vaddr2_3;
> +		__m256i vaddr4_5, vaddr6_7;
> +		__m512i vaddr0_3, vaddr4_7;
> +
> +		mb0 = rxp[0];
> +		mb1 = rxp[1];
> +		mb2 = rxp[2];
> +		mb3 = rxp[3];
> +		mb4 = rxp[4];
> +		mb5 = rxp[5];
> +		mb6 = rxp[6];
> +		mb7 = rxp[7];
> +
> +		/* load buf_addr(lo 64bit) and buf_iova(hi 64bit) */
> +		RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, buf_iova) !=
> +				offsetof(struct rte_mbuf, buf_addr) + 8);
> +		vaddr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
> +		vaddr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
> +		vaddr2 = _mm_loadu_si128((__m128i *)&mb2->buf_addr);
> +		vaddr3 = _mm_loadu_si128((__m128i *)&mb3->buf_addr);
> +		vaddr4 = _mm_loadu_si128((__m128i *)&mb4->buf_addr);
> +		vaddr5 = _mm_loadu_si128((__m128i *)&mb5->buf_addr);
> +		vaddr6 = _mm_loadu_si128((__m128i *)&mb6->buf_addr);
> +		vaddr7 = _mm_loadu_si128((__m128i *)&mb7->buf_addr);
> +
> +		/**
> +		 * merge 0 & 1, by casting 0 to 256-bit and inserting 1
> +		 * into the high lanes. Similarly for 2 & 3, and so on.
> +		 */
> +		vaddr0_1 =
> +
> 	_mm256_inserti128_si256(_mm256_castsi128_si256(vaddr0),
> +						vaddr1, 1);
> +		vaddr2_3 =
> +
> 	_mm256_inserti128_si256(_mm256_castsi128_si256(vaddr2),
> +						vaddr3, 1);
> +		vaddr4_5 =
> +
> 	_mm256_inserti128_si256(_mm256_castsi128_si256(vaddr4),
> +						vaddr5, 1);
> +		vaddr6_7 =
> +
> 	_mm256_inserti128_si256(_mm256_castsi128_si256(vaddr6),
> +						vaddr7, 1);
> +		vaddr0_3 =
> +
> 	_mm512_inserti64x4(_mm512_castsi256_si512(vaddr0_1),
> +						vaddr2_3, 1);
> +		vaddr4_7 =
> +
> 	_mm512_inserti64x4(_mm512_castsi256_si512(vaddr4_5),
> +						vaddr6_7, 1);
> +
> +		/* convert pa to dma_addr hdr/data */
> +		dma_addr0_3 = _mm512_unpackhi_epi64(vaddr0_3,
> vaddr0_3);
> +		dma_addr4_7 = _mm512_unpackhi_epi64(vaddr4_7,
> vaddr4_7);
> +
> +		/* add headroom to pa values */
> +		dma_addr0_3 = _mm512_add_epi64(dma_addr0_3,
> hdr_room);
> +		dma_addr4_7 = _mm512_add_epi64(dma_addr4_7,
> hdr_room);
> +
> +		/* flush desc with pa dma_addr */
> +		_mm512_store_si512((__m512i *)&rxdp->read,
> dma_addr0_3);
> +		_mm512_store_si512((__m512i *)&(rxdp + 4)->read,
> dma_addr4_7);
> +	}
> +
> +	rxq->rxrearm_start += IDPF_RXQ_REARM_THRESH;
> +	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
> +		rxq->rxrearm_start = 0;
> +
> +	rxq->rxrearm_nb -= IDPF_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 */
> +	IECM_PCI_REG_WRITE(rxq->qrx_tail, rx_id); }
> +
> +static __rte_always_inline void
> +idpf_singleq_rearm(struct idpf_rx_queue *rxq) {
> +	int i;
> +	uint16_t rx_id;
> +	volatile union virtchnl2_rx_desc *rxdp = rxq->rx_ring;
> +	struct rte_mempool_cache *cache =
> +		rte_mempool_default_cache(rxq->mp, rte_lcore_id());
> +	struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start];
> +
> +	rxdp += rxq->rxrearm_start;
> +
> +	if (unlikely(!cache))
> +		return idpf_singleq_rearm_common(rxq);
> +
> +	/* We need to pull 'n' more MBUFs into the software ring from
> mempool
> +	 * We inline the mempool function here, so we can vectorize the
> copy
> +	 * from the cache into the shadow ring.
> +	 */
> +
> +	/* Can this be satisfied from the cache? */
> +	if (cache->len < IDPF_RXQ_REARM_THRESH) {
> +		/* No. Backfill the cache first, and then fill from it */
> +		uint32_t req = IDPF_RXQ_REARM_THRESH + (cache->size -
> +							cache->len);
> +
> +		/* How many do we require i.e. number to fill the cache + the
> request */
> +		int ret = rte_mempool_ops_dequeue_bulk
> +				(rxq->mp, &cache->objs[cache->len], req);
> +		if (ret == 0) {
> +			cache->len += req;
> +		} else {
> +			if (rxq->rxrearm_nb + IDPF_RXQ_REARM_THRESH >=
> +			    rxq->nb_rx_desc) {
> +				__m128i dma_addr0;
> +
> +				dma_addr0 = _mm_setzero_si128();
> +				for (i = 0; i < IDPF_VPMD_DESCS_PER_LOOP;
> i++) {
> +					rxp[i] = &rxq->fake_mbuf;
> +					_mm_storeu_si128((__m128i
> *)&rxdp[i].read,
> +							 dma_addr0);
> +				}
> +			}
> +			rte_eth_devices[rxq->port_id].data-
> >rx_mbuf_alloc_failed +=
> +					IDPF_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_IDPF_16BYTE_RX_DESC
> +	/* to shuffle the addresses to correct slots. Values 4-7 will contain
> +	 * zeros, so use 7 for a zero-value.
> +	 */
> +	const __m512i permute_idx = _mm512_set_epi64(7, 7, 3, 1, 7, 7, 2, 0);
> +#else
> +	const __m512i permute_idx = _mm512_set_epi64(7, 3, 6, 2, 5, 1, 4, 0);
> +#endif

Conditional compilation is not necessary here.

> +
> +	/* Initialize the mbufs in vector, process 8 mbufs in one loop, taking
> +	 * from mempool cache and populating both shadow and HW rings
> +	 */
> +	for (i = 0; i < IDPF_RXQ_REARM_THRESH /
> IDPF_DESCS_PER_LOOP_AVX; i++) {
> +		const __m512i mbuf_ptrs = _mm512_loadu_si512
> +			(&cache->objs[cache->len -
> IDPF_DESCS_PER_LOOP_AVX]);
> +		_mm512_storeu_si512(rxp, mbuf_ptrs);
> +
> +		const __m512i iova_base_addrs = _mm512_i64gather_epi64
> +				(_mm512_add_epi64(mbuf_ptrs,
> iova_offsets),
> +				 0, /* base */
> +				 1  /* scale */);
> +		const __m512i iova_addrs =
> _mm512_add_epi64(iova_base_addrs,
> +				headroom);
> +#ifndef RTE_LIBRTE_IDPF_16BYTE_RX_DESC

Conditional compilation is not necessary here.

> +		const __m512i iovas0 = _mm512_castsi256_si512
> +				(_mm512_extracti64x4_epi64(iova_addrs, 0));
> +		const __m512i iovas1 = _mm512_castsi256_si512
> +				(_mm512_extracti64x4_epi64(iova_addrs, 1));
> +
> +		/* permute leaves desc 2-3 addresses in header address slots
> 0-1
> +		 * but these are ignored by driver since header split not
> +		 * enabled. Similarly for desc 6 & 7.
> +		 */
> +		const __m512i desc0_1 = _mm512_permutexvar_epi64
> +				(permute_idx,
> +				 iovas0);
> +		const __m512i desc2_3 = _mm512_bsrli_epi128(desc0_1, 8);
> +
> +		const __m512i desc4_5 = _mm512_permutexvar_epi64
> +				(permute_idx,
> +				 iovas1);
> +		const __m512i desc6_7 = _mm512_bsrli_epi128(desc4_5, 8);
> +
> +		_mm512_storeu_si512((void *)rxdp, desc0_1);
> +		_mm512_storeu_si512((void *)(rxdp + 2), desc2_3);
> +		_mm512_storeu_si512((void *)(rxdp + 4), desc4_5);
> +		_mm512_storeu_si512((void *)(rxdp + 6), desc6_7); #else
> +		/* permute leaves desc 4-7 addresses in header address slots
> 0-3
> +		 * but these are ignored by driver since header split not
> +		 * enabled.
> +		 */
> +		const __m512i desc0_3 =
> _mm512_permutexvar_epi64(permute_idx,
> +								 iova_addrs);
> +		const __m512i desc4_7 = _mm512_bsrli_epi128(desc0_3, 8);
> +
> +		_mm512_storeu_si512((void *)rxdp, desc0_3);
> +		_mm512_storeu_si512((void *)(rxdp + 4), desc4_7); #endif
> +		rxp += IDPF_DESCS_PER_LOOP_AVX;
> +		rxdp += IDPF_DESCS_PER_LOOP_AVX;
> +		cache->len -= IDPF_DESCS_PER_LOOP_AVX;
> +	}
> +
> +	rxq->rxrearm_start += IDPF_RXQ_REARM_THRESH;
> +	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
> +		rxq->rxrearm_start = 0;
> +
> +	rxq->rxrearm_nb -= IDPF_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 */
> +	IECM_PCI_REG_WRITE(rxq->qrx_tail, rx_id); }
> +
> +#define IDPF_RX_LEN_MASK 0x80808080
> +static __rte_always_inline uint16_t
> +_idpf_singleq_recv_raw_pkts_avx512(struct idpf_rx_queue *rxq,
> +				   struct rte_mbuf **rx_pkts,
> +				   uint16_t nb_pkts)
> +{
> +#ifdef IDPF_RX_PTYPE_OFFLOAD
> +	const uint32_t *type_table = rxq->adapter->ptype_tbl; #endif
> +
> +	const __m256i mbuf_init = _mm256_set_epi64x(0, 0, 0,
> +						    rxq->mbuf_initializer);
> +	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
> +	volatile union virtchnl2_rx_desc *rxdp = rxq->rx_ring;
> +
> +	rxdp += rxq->rx_tail;
> +
> +	rte_prefetch0(rxdp);
> +
> +	/* nb_pkts has to be floor-aligned to IDPF_DESCS_PER_LOOP_AVX */
> +	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IDPF_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 > IDPF_RXQ_REARM_THRESH)
> +		idpf_singleq_rearm(rxq);
> +
> +	/* Before we start moving massive data around, check to see if
> +	 * there is actually a packet available
> +	 */
> +	if (!(rxdp->flex_nic_wb.status_error0  &
> +	      rte_cpu_to_le_32(1 <<
> VIRTCHNL2_RX_FLEX_DESC_STATUS0_DD_S)))
> +		return 0;
> +
> +	/* constants used in processing loop */
> +	const __m512i crc_adjust =
> +		_mm512_set_epi32
> +			(/* 1st descriptor */
> +			 0,             /* ignore non-length fields */
> +			 -rxq->crc_len, /* sub crc on data_len */
> +			 -rxq->crc_len, /* sub crc on pkt_len */
> +			 0,             /* ignore pkt_type field */
> +			 /* 2nd descriptor */
> +			 0,             /* ignore non-length fields */
> +			 -rxq->crc_len, /* sub crc on data_len */
> +			 -rxq->crc_len, /* sub crc on pkt_len */
> +			 0,             /* ignore pkt_type field */
> +			 /* 3rd descriptor */
> +			 0,             /* ignore non-length fields */
> +			 -rxq->crc_len, /* sub crc on data_len */
> +			 -rxq->crc_len, /* sub crc on pkt_len */
> +			 0,             /* ignore pkt_type field */
> +			 /* 4th descriptor */
> +			 0,             /* ignore non-length fields */
> +			 -rxq->crc_len, /* sub crc on data_len */
> +			 -rxq->crc_len, /* sub crc on pkt_len */
> +			 0              /* ignore pkt_type field */
> +			);
> +
> +	/* 8 packets DD mask, LSB in each 32-bit value */
> +	const __m256i dd_check = _mm256_set1_epi32(1);
> +
> +	/* mask to shuffle from desc. to mbuf (4 descriptors)*/
> +	const __m512i shuf_msk =
> +		_mm512_set_epi32
> +			(/* 1st descriptor */
> +			 0xFFFFFFFF,    /* octet 4~7, 32bits rss */
> +			 0xFFFF0504,    /* octet 2~3, low 16 bits vlan_macip */
> +					/* octet 15~14, 16 bits data_len */
> +			 0xFFFF0504,    /* skip high 16 bits pkt_len, zero out
> */
> +					/* octet 15~14, low 16 bits pkt_len */
> +			 0xFFFFFFFF,    /* pkt_type set as unknown */
> +			 /* 2nd descriptor */
> +			 0xFFFFFFFF,    /* octet 4~7, 32bits rss */
> +			 0xFFFF0504,    /* octet 2~3, low 16 bits vlan_macip */
> +					/* octet 15~14, 16 bits data_len */
> +			 0xFFFF0504,    /* skip high 16 bits pkt_len, zero out
> */
> +					/* octet 15~14, low 16 bits pkt_len */
> +			 0xFFFFFFFF,    /* pkt_type set as unknown */
> +			 /* 3rd descriptor */
> +			 0xFFFFFFFF,    /* octet 4~7, 32bits rss */
> +			 0xFFFF0504,    /* octet 2~3, low 16 bits vlan_macip */
> +					/* octet 15~14, 16 bits data_len */
> +			 0xFFFF0504,    /* skip high 16 bits pkt_len, zero out
> */
> +					/* octet 15~14, low 16 bits pkt_len */
> +			 0xFFFFFFFF,    /* pkt_type set as unknown */
> +			 /* 4th descriptor */
> +			 0xFFFFFFFF,    /* octet 4~7, 32bits rss */
> +			 0xFFFF0504,    /* octet 2~3, low 16 bits vlan_macip */
> +					/* octet 15~14, 16 bits data_len */
> +			 0xFFFF0504,    /* skip high 16 bits pkt_len, zero out
> */
> +					/* octet 15~14, low 16 bits pkt_len */
> +			 0xFFFFFFFF     /* pkt_type set as unknown */
> +			);

The comments should be reworked since it does not fully match the code.

> +	/**
> +	 * 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);
> +
> +	uint16_t i, received;
> +
> +	for (i = 0, received = 0; i < nb_pkts;
> +	     i += IDPF_DESCS_PER_LOOP_AVX,
> +	     rxdp += IDPF_DESCS_PER_LOOP_AVX) {
> +		/* step 1, copy over 8 mbuf pointers to rx_pkts array */
> +		_mm256_storeu_si256((void *)&rx_pkts[i],
> +				    _mm256_loadu_si256((void *)&sw_ring[i]));
> #ifdef
> +RTE_ARCH_X86_64
> +		_mm256_storeu_si256
> +			((void *)&rx_pkts[i + 4],
> +			 _mm256_loadu_si256((void *)&sw_ring[i + 4]));
> #endif
> +
> +		__m512i raw_desc0_3, raw_desc4_7;
> +		const __m128i raw_desc7 =
> +			_mm_load_si128((void *)(rxdp + 7));
> +		rte_compiler_barrier();
> +		const __m128i raw_desc6 =
> +			_mm_load_si128((void *)(rxdp + 6));
> +		rte_compiler_barrier();
> +		const __m128i raw_desc5 =
> +			_mm_load_si128((void *)(rxdp + 5));
> +		rte_compiler_barrier();
> +		const __m128i raw_desc4 =
> +			_mm_load_si128((void *)(rxdp + 4));
> +		rte_compiler_barrier();
> +		const __m128i raw_desc3 =
> +			_mm_load_si128((void *)(rxdp + 3));
> +		rte_compiler_barrier();
> +		const __m128i raw_desc2 =
> +			_mm_load_si128((void *)(rxdp + 2));
> +		rte_compiler_barrier();
> +		const __m128i raw_desc1 =
> +			_mm_load_si128((void *)(rxdp + 1));
> +		rte_compiler_barrier();
> +		const __m128i raw_desc0 =
> +			_mm_load_si128((void *)(rxdp + 0));
> +
> +		raw_desc4_7 = _mm512_broadcast_i32x4(raw_desc4);
> +		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7,
> raw_desc5, 1);
> +		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7,
> raw_desc6, 2);
> +		raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7,
> raw_desc7, 3);
> +		raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0);
> +		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3,
> raw_desc1, 1);
> +		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3,
> raw_desc2, 2);
> +		raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3,
> raw_desc3, 3);
> +
> +		/**
> +		 * convert descriptors 4-7 into mbufs, adjusting length and
> +		 * re-arranging fields. Then write into the mbuf
> +		 */
> +		const __m512i len4_7 = _mm512_slli_epi32(raw_desc4_7,
> +							 PKTLEN_SHIFT);
> +		const __m512i desc4_7 =
> _mm512_mask_blend_epi16(IDPF_RX_LEN_MASK,
> +								raw_desc4_7,
> +								len4_7);
> +		__m512i mb4_7 = _mm512_shuffle_epi8(desc4_7, shuf_msk);
> +
> +		mb4_7 = _mm512_add_epi32(mb4_7, crc_adjust); #ifdef
> +IDPF_RX_PTYPE_OFFLOAD
> +		/**
> +		 * to get packet types, shift 64-bit values down 30 bits
> +		 * and so ptype is in lower 8-bits in each
> +		 */
> +		const __m512i ptypes4_7 = _mm512_srli_epi64(desc4_7, 16);
> +		const __m256i ptypes6_7 =
> _mm512_extracti64x4_epi64(ptypes4_7, 1);
> +		const __m256i ptypes4_5 =
> _mm512_extracti64x4_epi64(ptypes4_7, 0);
> +		const uint8_t ptype7 = _mm256_extract_epi8(ptypes6_7, 16);
> +		const uint8_t ptype6 = _mm256_extract_epi8(ptypes6_7, 0);
> +		const uint8_t ptype5 = _mm256_extract_epi8(ptypes4_5, 16);
> +		const uint8_t ptype4 = _mm256_extract_epi8(ptypes4_5, 0);
> +
> +		const __m512i ptype4_7 = _mm512_set_epi32
> +			(0, 0, 0, type_table[ptype7],
> +			 0, 0, 0, type_table[ptype6],
> +			 0, 0, 0, type_table[ptype5],
> +			 0, 0, 0, type_table[ptype4]);
> +		mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7,
> ptype4_7); #endif
> +
> +		/**
> +		 * convert descriptors 0-3 into mbufs, adjusting length and
> +		 * re-arranging fields. Then write into the mbuf
> +		 */
> +		const __m512i len0_3 = _mm512_slli_epi32(raw_desc0_3,
> +							 PKTLEN_SHIFT);
> +		const __m512i desc0_3 =
> _mm512_mask_blend_epi16(IDPF_RX_LEN_MASK,
> +								raw_desc0_3,
> +								len0_3);
> +		__m512i mb0_3 = _mm512_shuffle_epi8(desc0_3, shuf_msk);
> +
> +		mb0_3 = _mm512_add_epi32(mb0_3, crc_adjust); #ifdef
> +IDPF_RX_PTYPE_OFFLOAD
> +		/* get the packet types */
> +		const __m512i ptypes0_3 = _mm512_srli_epi64(desc0_3, 16);
> +		const __m256i ptypes2_3 =
> _mm512_extracti64x4_epi64(ptypes0_3, 1);
> +		const __m256i ptypes0_1 =
> _mm512_extracti64x4_epi64(ptypes0_3, 0);
> +		const uint8_t ptype3 = _mm256_extract_epi8(ptypes2_3, 16);
> +		const uint8_t ptype2 = _mm256_extract_epi8(ptypes2_3, 0);
> +		const uint8_t ptype1 = _mm256_extract_epi8(ptypes0_1, 16);
> +		const uint8_t ptype0 = _mm256_extract_epi8(ptypes0_1, 0);
> +
> +		const __m512i ptype0_3 = _mm512_set_epi32
> +			(0, 0, 0, type_table[ptype3],
> +			 0, 0, 0, type_table[ptype2],
> +			 0, 0, 0, type_table[ptype1],
> +			 0, 0, 0, type_table[ptype0]);
> +		mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3,
> ptype0_3); #endif
> +
> +		/**
> +		 * 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 */
> +
> +		/**
> +		 * At this point, we have the 8 sets of flags in the low 16-bits
> +		 * of each 32-bit value in vlan0.
> +		 * We want to extract these, and merge them with the mbuf
> init
> +		 * data so we can do a single write to the mbuf to set the
> flags
> +		 * and all the other initialization fields. Extracting the
> +		 * appropriate flags means that we have to do a shift and
> blend
> +		 * for each mbuf before we do the write. However, we can
> also
> +		 * add in the previously computed rx_descriptor fields to
> +		 * make a single 256-bit write per mbuf
> +		 */
> +		/* check the structure matches expectations */
> +		RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
> +				 offsetof(struct rte_mbuf, rearm_data) + 8);
> +		RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf,
> rearm_data) !=
> +				 RTE_ALIGN(offsetof(struct rte_mbuf,
> +						    rearm_data),
> +						    16));
> +		/* build up data and do writes */
> +		__m256i rearm0, rearm1, rearm2, rearm3, rearm4, rearm5,
> +			rearm6, rearm7;
> +		const __m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7,
> 0);
> +		const __m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7,
> 1);
> +		const __m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3,
> 0);
> +		const __m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3,
> 1);
> +
> +		rearm6 = _mm256_permute2f128_si256(mbuf_init, mb6_7,
> 0x20);
> +		rearm4 = _mm256_permute2f128_si256(mbuf_init, mb4_5,
> 0x20);
> +		rearm2 = _mm256_permute2f128_si256(mbuf_init, mb2_3,
> 0x20);
> +		rearm0 = _mm256_permute2f128_si256(mbuf_init, 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);
> +
> +		rearm7 = _mm256_blend_epi32(mbuf_init, mb6_7, 0xF0);
> +		rearm5 = _mm256_blend_epi32(mbuf_init, mb4_5, 0xF0);
> +		rearm3 = _mm256_blend_epi32(mbuf_init, mb2_3, 0xF0);
> +		rearm1 = _mm256_blend_epi32(mbuf_init, 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);
> +
> +		/* 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 != IDPF_DESCS_PER_LOOP_AVX)
> +			break;
> +	}
> +
> +	/* update tail pointers */
> +	rxq->rx_tail += received;
> +	rxq->rx_tail &= (rxq->nb_rx_desc - 1);
> +	if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep aligned */
> +		rxq->rx_tail--;
> +		received--;
> +	}
> +	rxq->rxrearm_nb += received;
> +	return received;
> +}
> +
> +/**
> + * Notice:
> + * - nb_pkts < IDPF_DESCS_PER_LOOP, just return no packet  */ uint16_t
> +idpf_singleq_recv_pkts_avx512(void *rx_queue, struct rte_mbuf **rx_pkts,
> +			  uint16_t nb_pkts)
> +{
> +	return _idpf_singleq_recv_raw_pkts_avx512(rx_queue, rx_pkts,
> nb_pkts);
> +}
> +
> +static __rte_always_inline int
> +idpf_tx_free_bufs_avx512(struct idpf_tx_queue *txq) {
> +	struct idpf_tx_vec_entry *txep;
> +	uint32_t n;
> +	uint32_t i;
> +	int nb_free = 0;
> +	struct rte_mbuf *m, *free[txq->rs_thresh];
> +
> +	/* check DD bits on threshold descriptor */
> +	if ((txq->tx_ring[txq->next_dd].qw1 &
> +			rte_cpu_to_le_64(IECM_TXD_QW1_DTYPE_M)) !=
> +
> 	rte_cpu_to_le_64(IECM_TX_DESC_DTYPE_DESC_DONE))
> +		return 0;
> +
> +	n = txq->rs_thresh;
> +
> +	 /* first buffer to free from S/W ring is at index
> +	  * tx_next_dd - (tx_rs_thresh-1)
> +	  */
> +	txep = (void *)txq->sw_ring;
> +	txep += txq->next_dd - (n - 1);
> +
> +	if (txq->offloads & RTE_ETH_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;
> +
> +		if (!cache || cache->len == 0)
> +			goto normal;
> +
> +		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;
> +	}
> +
> +normal:
> +	m = rte_pktmbuf_prefree_seg(txep[0].mbuf);
> +	if (likely(m)) {
> +		free[0] = m;
> +		nb_free = 1;
> +		for (i = 1; i < n; i++) {
> +			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
> +			if (likely(m)) {
> +				if (likely(m->pool == free[0]->pool)) {
> +					free[nb_free++] = m;
> +				} else {
> +					rte_mempool_put_bulk(free[0]->pool,
> +							     (void *)free,
> +							     nb_free);
> +					free[0] = m;
> +					nb_free = 1;
> +				}
> +			}
> +		}
> +		rte_mempool_put_bulk(free[0]->pool, (void **)free,
> nb_free);
> +	} else {
> +		for (i = 1; i < n; i++) {
> +			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
> +			if (m)
> +				rte_mempool_put(m->pool, m);
> +		}
> +	}
> +
> +done:
> +	/* buffers were freed, update counters */
> +	txq->nb_free = (uint16_t)(txq->nb_free + txq->rs_thresh);
> +	txq->next_dd = (uint16_t)(txq->next_dd + txq->rs_thresh);
> +	if (txq->next_dd >= txq->nb_tx_desc)
> +		txq->next_dd = (uint16_t)(txq->rs_thresh - 1);
> +
> +	return txq->rs_thresh;
> +}
> +
> +static __rte_always_inline void
> +tx_backlog_entry_avx512(struct idpf_tx_vec_entry *txep,
> +			struct rte_mbuf **tx_pkts, uint16_t nb_pkts) {
> +	int i;
> +
> +	for (i = 0; i < (int)nb_pkts; ++i)
> +		txep[i].mbuf = tx_pkts[i];
> +}
> +
> +static __rte_always_inline void
> +idpf_vtx1(volatile struct iecm_base_tx_desc *txdp,
> +	  struct rte_mbuf *pkt, uint64_t flags) {
> +	uint64_t high_qw =
> +		(IECM_TX_DESC_DTYPE_DATA |
> +		 ((uint64_t)flags  << IECM_TXD_QW1_CMD_S) |
> +		 ((uint64_t)pkt->data_len << IECM_TXD_QW1_TX_BUF_SZ_S));
> +
> +	__m128i descriptor = _mm_set_epi64x(high_qw,
> +					    pkt->buf_iova + pkt->data_off);
> +	_mm_storeu_si128((__m128i *)txdp, descriptor); }
> +
> +#define IDPF_TX_LEN_MASK 0xAA
> +#define IDPF_TX_OFF_MASK 0x55
> +static __rte_always_inline void
> +idpf_vtx(volatile struct iecm_base_tx_desc *txdp,
> +	 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags) {
> +	const uint64_t hi_qw_tmpl = (IECM_TX_DESC_DTYPE_DATA  |
> +			((uint64_t)flags  << IECM_TXD_QW1_CMD_S));
> +
> +	/* if unaligned on 32-bit boundary, do one to align */
> +	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
> +		idpf_vtx1(txdp, *pkt, flags);
> +		nb_pkts--, txdp++, pkt++;
> +	}
> +
> +	/* do 4 at a time while possible, in bursts */
> +	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
> +		uint64_t hi_qw3 =
> +			hi_qw_tmpl |
> +			((uint64_t)pkt[3]->data_len <<
> +			 IECM_TXD_QW1_TX_BUF_SZ_S);
> +		uint64_t hi_qw2 =
> +			hi_qw_tmpl |
> +			((uint64_t)pkt[2]->data_len <<
> +			 IECM_TXD_QW1_TX_BUF_SZ_S);
> +		uint64_t hi_qw1 =
> +			hi_qw_tmpl |
> +			((uint64_t)pkt[1]->data_len <<
> +			 IECM_TXD_QW1_TX_BUF_SZ_S);
> +		uint64_t hi_qw0 =
> +			hi_qw_tmpl |
> +			((uint64_t)pkt[0]->data_len <<
> +			 IECM_TXD_QW1_TX_BUF_SZ_S);
> +
> +		__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,
> +				 hi_qw1,
> +				 pkt[1]->buf_iova + pkt[1]->data_off,
> +				 hi_qw0,
> +				 pkt[0]->buf_iova + pkt[0]->data_off);
> +		_mm512_storeu_si512((void *)txdp, desc0_3);
> +	}
> +
> +	/* do any last ones */
> +	while (nb_pkts) {
> +		idpf_vtx1(txdp, *pkt, flags);
> +		txdp++, pkt++, nb_pkts--;
> +	}
> +}
> +
> +static __rte_always_inline uint16_t
> +idpf_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf
> **tx_pkts,
> +				 uint16_t nb_pkts)
> +{
> +	struct idpf_tx_queue *txq = (struct idpf_tx_queue *)tx_queue;
> +	volatile struct iecm_base_tx_desc *txdp;
> +	struct idpf_tx_vec_entry *txep;
> +	uint16_t n, nb_commit, tx_id;
> +	/* bit2 is reserved and must be set to 1 according to Spec */
> +	uint64_t flags = IECM_TX_DESC_CMD_EOP | 0x04;
> +	uint64_t rs = IECM_TX_DESC_CMD_RS | flags;
> +
> +	/* cross rx_thresh boundary is not allowed */
> +	nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
> +
> +	if (txq->nb_free < txq->free_thresh)
> +		idpf_tx_free_bufs_avx512(txq);
> +
> +	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
> +	if (unlikely(nb_pkts == 0))
> +		return 0;
> +
> +	tx_id = txq->tx_tail;
> +	txdp = &txq->tx_ring[tx_id];
> +	txep = (void *)txq->sw_ring;
> +	txep += tx_id;
> +
> +	txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
> +
> +	n = (uint16_t)(txq->nb_tx_desc - tx_id);
> +	if (nb_commit >= n) {
> +		tx_backlog_entry_avx512(txep, tx_pkts, n);
> +
> +		idpf_vtx(txdp, tx_pkts, n - 1, flags);
> +		tx_pkts += (n - 1);
> +		txdp += (n - 1);
> +
> +		idpf_vtx1(txdp, *tx_pkts++, rs);
> +
> +		nb_commit = (uint16_t)(nb_commit - n);
> +
> +		tx_id = 0;
> +		txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
> +
> +		/* avoid reach the end of ring */
> +		txdp = &txq->tx_ring[tx_id];
> +		txep = (void *)txq->sw_ring;
> +		txep += tx_id;
> +	}
> +
> +	tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
> +
> +	idpf_vtx(txdp, tx_pkts, nb_commit, flags);
> +
> +	tx_id = (uint16_t)(tx_id + nb_commit);
> +	if (tx_id > txq->next_rs) {
> +		txq->tx_ring[txq->next_rs].qw1 |=
> +
> 	rte_cpu_to_le_64(((uint64_t)IECM_TX_DESC_CMD_RS) <<
> +					 IECM_TXD_QW1_CMD_S);
> +		txq->next_rs =
> +			(uint16_t)(txq->next_rs + txq->rs_thresh);
> +	}
> +
> +	txq->tx_tail = tx_id;
> +
> +	IECM_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
> +
> +	return nb_pkts;
> +}
> +
> +static __rte_always_inline uint16_t
> +idpf_xmit_pkts_vec_avx512_cmn(void *tx_queue, struct rte_mbuf
> **tx_pkts,
> +			      uint16_t nb_pkts)
> +{
> +	uint16_t nb_tx = 0;
> +	struct idpf_tx_queue *txq = (struct idpf_tx_queue *)tx_queue;
> +
> +	while (nb_pkts) {
> +		uint16_t ret, num;
> +
> +		num = (uint16_t)RTE_MIN(nb_pkts, txq->rs_thresh);
> +		ret = idpf_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;
> +}
> +
> +uint16_t
> +idpf_singleq_xmit_pkts_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
> +			     uint16_t nb_pkts)
> +{
> +	return idpf_xmit_pkts_vec_avx512_cmn(tx_queue, tx_pkts,
> nb_pkts); }
> +
> +static inline void
> +idpf_singleq_tx_release_mbufs_avx512(struct idpf_tx_queue *txq) {
> +	unsigned int i;
> +	const uint16_t max_desc = (uint16_t)(txq->nb_tx_desc - 1);
> +	struct idpf_tx_vec_entry *swr = (void *)txq->sw_ring;
> +
> +	if (!txq->sw_ring || txq->nb_free == max_desc)
> +		return;
> +
> +	i = txq->next_dd - txq->rs_thresh + 1;
> +	if (txq->tx_tail < i) {
> +		for (; i < txq->nb_tx_desc; i++) {
> +			rte_pktmbuf_free_seg(swr[i].mbuf);
> +			swr[i].mbuf = NULL;
> +		}
> +		i = 0;
> +	}
> +}
> +
> +static const struct idpf_txq_ops avx512_singleq_tx_vec_ops = {
> +	.release_mbufs = idpf_singleq_tx_release_mbufs_avx512,
> +};
> +
> +int __rte_cold
> +idpf_singleq_tx_vec_setup_avx512(struct idpf_tx_queue *txq) {
> +	txq->ops = &avx512_singleq_tx_vec_ops;
> +	return 0;
> +}
> diff --git a/drivers/net/idpf/idpf_rxtx_vec_common.h
> b/drivers/net/idpf/idpf_rxtx_vec_common.h
> new file mode 100644
> index 0000000000..336d4c8b25
> --- /dev/null
> +++ b/drivers/net/idpf/idpf_rxtx_vec_common.h
> @@ -0,0 +1,89 @@
> +/* SPDX-License-Identifier: BSD-3-Clause
> + * Copyright(c) 2022 Intel Corporation
> + */
> +
> +#ifndef _IDPF_RXTX_VEC_COMMON_H_
> +#define _IDPF_RXTX_VEC_COMMON_H_
> +#include <stdint.h>
> +#include <ethdev_driver.h>
> +#include <rte_malloc.h>
> +
> +#include "idpf_ethdev.h"
> +#include "idpf_rxtx.h"
> +
> +#ifndef __INTEL_COMPILER
> +#pragma GCC diagnostic ignored "-Wcast-qual"
> +#endif
> +
> +#define IDPF_VECTOR_PATH		0
> +
> +static inline int
> +idpf_rx_vec_queue_default(struct idpf_rx_queue *rxq) {
> +	if (!rxq)
> +		return -1;
> +
> +	if (!rte_is_power_of_2(rxq->nb_rx_desc))
> +		return -1;
> +
> +	if (rxq->rx_free_thresh < IDPF_VPMD_RX_MAX_BURST)
> +		return -1;
> +
> +	if (rxq->nb_rx_desc % rxq->rx_free_thresh)
> +		return -1;
> +
> +	/* Currently, vector path doesn't support timestamp. */
> +	if (rxq->offloads & RTE_ETH_RX_OFFLOAD_TIMESTAMP)
> +		return -1;
> +
> +	return IDPF_VECTOR_PATH;
> +}
> +
> +static inline int
> +idpf_tx_vec_queue_default(struct idpf_tx_queue *txq) {
> +	if (!txq)
> +		return -1;
> +
> +	if (txq->rs_thresh < IDPF_VPMD_TX_MAX_BURST ||
> +	    (txq->rs_thresh & 3) != 0)
> +		return -1;
> +
> +	return IDPF_VECTOR_PATH;
> +}
> +
> +static inline int
> +idpf_rx_vec_dev_check_default(struct rte_eth_dev *dev) {
> +	int i;
> +	struct idpf_rx_queue *rxq;
> +	int ret = 0;
> +
> +	for (i = 0; i < dev->data->nb_rx_queues; i++) {
> +		rxq = dev->data->rx_queues[i];
> +		ret = (idpf_rx_vec_queue_default(rxq));
> +		if (ret < 0)
> +			return -1;
> +	}
> +
> +	return IDPF_VECTOR_PATH;
> +}
> +
> +static inline int
> +idpf_tx_vec_dev_check_default(struct rte_eth_dev *dev) {
> +	int i;
> +	struct idpf_tx_queue *txq;
> +	int ret = 0;
> +
> +	for (i = 0; i < dev->data->nb_tx_queues; i++) {
> +		txq = dev->data->tx_queues[i];
> +		ret = idpf_tx_vec_queue_default(txq);
> +		if (ret < 0)
> +			return -1;
> +	}
> +
> +	return IDPF_VECTOR_PATH;
> +}
> +
> +#endif /*_IDPF_RXTX_VEC_COMMON_H_*/
> diff --git a/drivers/net/idpf/meson.build b/drivers/net/idpf/meson.build
> index 338a39e391..ee4bb94a97 100644
> --- a/drivers/net/idpf/meson.build
> +++ b/drivers/net/idpf/meson.build
> @@ -16,4 +16,33 @@ sources = files(
>  	'idpf_vchnl.c',
>  )
> 
> -includes += include_directories('base') \ No newline at end of file
> +if arch_subdir == 'x86'
> +	idpf_avx512_cpu_support = (
> +		cc.get_define('__AVX512F__', args: machine_args) != '' and
> +		cc.get_define('__AVX512BW__', args: machine_args) != ''
> +	)
> +
> +	idpf_avx512_cc_support = (
> +		not machine_args.contains('-mno-avx512f') and
> +		cc.has_argument('-mavx512f') and
> +		cc.has_argument('-mavx512bw')
> +	)
> +
> +	if idpf_avx512_cpu_support == true or idpf_avx512_cc_support ==
> true
> +		cflags += ['-DCC_AVX512_SUPPORT']
> +		avx512_args = [cflags, '-mavx512f', '-mavx512bw']
> +		if cc.has_argument('-march=skylake-avx512')
> +			avx512_args += '-march=skylake-avx512'
> +		endif
> +		idpf_avx512_lib = static_library('idpf_avx512_lib',
> +			'idpf_rxtx_vec_avx512.c',
> +			dependencies: [static_rte_ethdev, static_rte_bus_pci,
> +			static_rte_kvargs, static_rte_hash],
> +			include_directories: includes,
> +			c_args: avx512_args)
> +		objs +=
> idpf_avx512_lib.extract_objects('idpf_rxtx_vec_avx512.c')
> +	endif
> +endif
> +
> +includes += include_directories('base')
> +

4 spaces should be used instead of a tab.

> --
> 2.25.1