Re: [PATCH 1/2] common/idpf: enable AVX2 for single queue Rx

[Bruce Richardson <bruce.richardson@intel.com>]

On Wed, Jan 08, 2025 at 05:47:56PM +0530, Shaiq Wani wrote:
> In case some CPUs don't support AVX512. Enable AVX2 for them to
> get better per-core performance.
> 
> Signed-off-by: Shaiq Wani <shaiq.wani@intel.com>

Hi,

some review comments inline below.

/Bruce

> ---
>  drivers/common/idpf/idpf_common_device.h    |   1 +
>  drivers/common/idpf/idpf_common_rxtx.h      |   4 +
>  drivers/common/idpf/idpf_common_rxtx_avx2.c | 590 ++++++++++++++++++++
>  drivers/common/idpf/meson.build             |  15 +
>  drivers/common/idpf/version.map             |   1 +
>  drivers/net/idpf/idpf_rxtx.c                |  12 +
>  6 files changed, 623 insertions(+)
>  create mode 100644 drivers/common/idpf/idpf_common_rxtx_avx2.c
> 
> diff --git a/drivers/common/idpf/idpf_common_device.h b/drivers/common/idpf/idpf_common_device.h
> index bfa927a5ff..734be1c88a 100644
> --- a/drivers/common/idpf/idpf_common_device.h
> +++ b/drivers/common/idpf/idpf_common_device.h
> @@ -123,6 +123,7 @@ struct idpf_vport {
>  
>  	bool rx_vec_allowed;
>  	bool tx_vec_allowed;
> +	bool rx_use_avx2;
>  	bool rx_use_avx512;
>  	bool tx_use_avx512;
>  
> diff --git a/drivers/common/idpf/idpf_common_rxtx.h b/drivers/common/idpf/idpf_common_rxtx.h
> index eeeeed12e2..f50cf5ef46 100644
> --- a/drivers/common/idpf/idpf_common_rxtx.h
> +++ b/drivers/common/idpf/idpf_common_rxtx.h
> @@ -302,5 +302,9 @@ uint16_t idpf_dp_splitq_xmit_pkts_avx512(void *tx_queue, struct rte_mbuf **tx_pk
>  __rte_internal
>  uint16_t idpf_dp_singleq_recv_scatter_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
>  			  uint16_t nb_pkts);
> +__rte_internal
> +uint16_t idpf_dp_singleq_recv_pkts_avx2(void *rx_queue,
> +					struct rte_mbuf **rx_pkts,
> +					uint16_t nb_pkts);
>  

I'm a little confused by the "singleq" part of the name here, can you
explain a little (in the commit message, perhaps) what is the "single"
referring to? Does the driver have the ability to poll multiple queues at
once or something?

>  #endif /* _IDPF_COMMON_RXTX_H_ */
> diff --git a/drivers/common/idpf/idpf_common_rxtx_avx2.c b/drivers/common/idpf/idpf_common_rxtx_avx2.c
> new file mode 100644
> index 0000000000..a05b26c68a
> --- /dev/null
> +++ b/drivers/common/idpf/idpf_common_rxtx_avx2.c
> @@ -0,0 +1,590 @@
> +/* SPDX-License-Identifier: BSD-3-Clause
> + * Copyright(c) 2023 Intel Corporation
> + */
> +
> +#include <rte_vect.h>
> +
> +#include "idpf_common_rxtx.h"
> +#include "idpf_common_device.h"
> +
> +#ifndef __INTEL_COMPILER
> +#pragma GCC diagnostic ignored "-Wcast-qual"
> +#endif

There is work ongoing to stop using this warning removal [1]. This code may
need to be rebased on top of that if it's applied soon.

[1] https://patches.dpdk.org/project/dpdk/list/?series=34390

> +
> +static __rte_always_inline void
> +idpf_singleq_rx_rearm(struct idpf_rx_queue *rxq)
> +{
> +	int i;
> +	uint16_t rx_id;
> +	volatile union virtchnl2_rx_desc *rxdp = rxq->rx_ring;
> +	struct rte_mbuf **rxep = &rxq->sw_ring[rxq->rxrearm_start];
> +
> +	rxdp += rxq->rxrearm_start;
> +
> +	/* Pull 'n' more MBUFs into the software ring */
> +	if (rte_mempool_get_bulk(rxq->mp,
> +				 (void *)rxep,
> +				 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++) {
> +				rxep[i] = &rxq->fake_mbuf;
> +				_mm_store_si128((__m128i *)&rxdp[i].read,
> +						dma_addr0);
> +			}
> +		}
> +		rte_atomic_fetch_add_explicit(&rxq->rx_stats.mbuf_alloc_failed,
> +				   IDPF_RXQ_REARM_THRESH, rte_memory_order_relaxed);
> +
> +		return;
> +	}
> +
> +	struct rte_mbuf *mb0, *mb1;
> +	__m128i dma_addr0, dma_addr1;
> +	__m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM,
> +			RTE_PKTMBUF_HEADROOM);
> +	/* Initialize the mbufs in vector, process 2 mbufs in one loop */
> +	for (i = 0; i < IDPF_RXQ_REARM_THRESH; i += 2, rxep += 2) {
> +		__m128i vaddr0, vaddr1;
> +
> +		mb0 = rxep[0];
> +		mb1 = rxep[1];
> +
> +		/* 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);
> +
> +		/* convert pa to dma_addr hdr/data */
> +		dma_addr0 = _mm_unpackhi_epi64(vaddr0, vaddr0);
> +		dma_addr1 = _mm_unpackhi_epi64(vaddr1, vaddr1);
> +
> +		/* add headroom to pa values */
> +		dma_addr0 = _mm_add_epi64(dma_addr0, hdr_room);
> +		dma_addr1 = _mm_add_epi64(dma_addr1, hdr_room);
> +
> +		/* flush desc with pa dma_addr */
> +		_mm_store_si128((__m128i *)&rxdp++->read, dma_addr0);
> +		_mm_store_si128((__m128i *)&rxdp++->read, dma_addr1);
> +	}
> +
> +	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 */
> +	IDPF_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
> +}
> +
> +static inline uint16_t
> +_idpf_singleq_recv_raw_pkts_vec_avx2(struct idpf_rx_queue *rxq, struct rte_mbuf **rx_pkts,
> +				     uint16_t nb_pkts, uint8_t *split_packet)
> +{
> +#define IDPF_DESCS_PER_LOOP_AVX 8
> +
> +	const uint32_t *ptype_tbl = rxq->adapter->ptype_tbl;
> +	const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
> +			0, rxq->mbuf_initializer);
> +	struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
> +	volatile union virtchnl2_rx_desc *rxdp = rxq->rx_ring;
> +	const int avx_aligned = ((rxq->rx_tail & 1) == 0);
> +
> +	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_rx_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;
> +
> +	/* 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,
> +			VIRTCHNL2_RX_FLEX_DESC_STATUS0_EOF_S);
> +
> +	/* mask to shuffle from desc. to mbuf (2 descriptors)*/
> +	const __m256i shuf_msk =
> +		_mm256_set_epi8
> +			(/* first descriptor */
> +			 0xFF, 0xFF,
> +			 0xFF, 0xFF,	/* rss hash parsed separately */
> +			 11, 10,	/* octet 10~11, 16 bits vlan_macip */
> +			 5, 4,		/* octet 4~5, 16 bits data_len */
> +			 0xFF, 0xFF,	/* skip hi 16 bits pkt_len, zero out */
> +			 5, 4,		/* octet 4~5, 16 bits pkt_len */
> +			 0xFF, 0xFF,	/* pkt_type set as unknown */
> +			 0xFF, 0xFF,	/*pkt_type set as unknown */
> +			 /* second descriptor */
> +			 0xFF, 0xFF,
> +			 0xFF, 0xFF,	/* rss hash parsed separately */
> +			 11, 10,	/* octet 10~11, 16 bits vlan_macip */
> +			 5, 4,		/* octet 4~5, 16 bits data_len */
> +			 0xFF, 0xFF,	/* skip hi 16 bits pkt_len, zero out */
> +			 5, 4,		/* octet 4~5, 16 bits pkt_len */
> +			 0xFF, 0xFF,	/* pkt_type set as unknown */
> +			 0xFF, 0xFF	/*pkt_type set as unknown */
> +			);
> +	/**
> +	 * compile-time check the above crc and shuffle layout is correct.
> +	 * NOTE: the first field (lowest address) is given last in set_epi
> +	 * calls above.
> +	 */
> +	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
> +			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
> +	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
> +			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
> +	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
> +			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
> +	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
> +			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
> +
> +	/* Status/Error flag masks */
> +	/**
> +	 * mask everything except Checksum Reports, RSS indication
> +	 * and VLAN indication.
> +	 * bit6:4 for IP/L4 checksum errors.
> +	 * bit12 is for RSS indication.
> +	 * bit13 is for VLAN indication.
> +	 */
> +	const __m256i flags_mask =
> +		 _mm256_set1_epi32((0xF << 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((RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 |
> +		 RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
> +		  RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_BAD  |
> +		 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_BAD  |
> +		 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
> +		 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
> +		 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
> +		 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
> +		 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		/**
> +		 * second 128-bits
> +		 * shift right 20 bits to use the low two bits to indicate
> +		 * outer checksum status
> +		 * shift right 1 bit to make sure it not exceed 255
> +		 */
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_BAD  |
> +		 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_BAD  |
> +		 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
> +		 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
> +		 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_BAD |
> +		 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
> +		 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
> +		(RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
> +		 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1);
> +	const __m256i cksum_mask =
> +		 _mm256_set1_epi32(RTE_MBUF_F_RX_IP_CKSUM_MASK |
> +				   RTE_MBUF_F_RX_L4_CKSUM_MASK |
> +				   RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
> +				   RTE_MBUF_F_RX_OUTER_L4_CKSUM_MASK);
> +	/**
> +	 * 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,
> +			RTE_MBUF_F_RX_RSS_HASH | RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED,
> +			RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED,
> +			RTE_MBUF_F_RX_RSS_HASH, 0,
> +			/* end up 128-bits */
> +			0, 0, 0, 0,
> +			0, 0, 0, 0,
> +			0, 0, 0, 0,
> +			RTE_MBUF_F_RX_RSS_HASH | RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED,
> +			RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED,
> +			RTE_MBUF_F_RX_RSS_HASH, 0);
> +
> +	RTE_SET_USED(avx_aligned); /* for 32B descriptors we don't use this */
> +

Does the driver or HW support 16B descriptors? If not, just remove this
variable completely. Don't keep it just for consistency with other drivers.

> +	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
> +
> +		__m256i raw_desc0_1, raw_desc2_3, raw_desc4_5, raw_desc6_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));
> +

Here and a number of other places, I think you can reduce the amount of
word-wrapping being done. Unlike when the first AVX2 code was being
written, we now can use up to 100 characters be line.

> +		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);
> +
> +		if (split_packet) {
> +			int j;
> +
> +			for (j = 0; j < IDPF_DESCS_PER_LOOP_AVX; j++)
> +				rte_mbuf_prefetch_part2(rx_pkts[i + j]);
> +		}
> +

I don't see any use of buffer reassembly for multi-segment packets in the
driver code. If it's not planned to add this, then you can remove this
block. If it is planned to add it, then keep this, and you can base the
implementation on the common function being extracted out of other
drivers[2].

[2] https://patches.dpdk.org/project/dpdk/patch/20250120120016.1530274-3-bruce.richardson@intel.com/


> +		/**
> +		 * convert descriptors 4-7 into mbufs, re-arrange fields.
> +		 * Then write into the mbuf.
> +		 */
> +		__m256i mb6_7 = _mm256_shuffle_epi8(raw_desc6_7, shuf_msk);
> +		__m256i mb4_5 = _mm256_shuffle_epi8(raw_desc4_5, shuf_msk);
> +
> +		/**
> +		 * to get packet types, ptype is located in bit16-25
> +		 * of each 128bits
> +		 */
> +		const __m256i ptype_mask =
> +			_mm256_set1_epi16(VIRTCHNL2_RX_FLEX_DESC_PTYPE_M);
> +		const __m256i ptypes6_7 =
> +			_mm256_and_si256(raw_desc6_7, ptype_mask);
> +		const __m256i ptypes4_5 =
> +			_mm256_and_si256(raw_desc4_5, ptype_mask);
> +		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);
> +
> +		mb6_7 = _mm256_insert_epi32(mb6_7, ptype_tbl[ptype7], 4);
> +		mb6_7 = _mm256_insert_epi32(mb6_7, ptype_tbl[ptype6], 0);
> +		mb4_5 = _mm256_insert_epi32(mb4_5, ptype_tbl[ptype5], 4);
> +		mb4_5 = _mm256_insert_epi32(mb4_5, ptype_tbl[ptype4], 0);
> +		/* merge the status bits into one register */
> +		const __m256i status4_7 = _mm256_unpackhi_epi32(raw_desc6_7,
> +				raw_desc4_5);
> +
> +		/**
> +		 * convert descriptors 0-3 into mbufs, re-arrange fields.
> +		 * Then write into the mbuf.
> +		 */
> +		__m256i mb2_3 = _mm256_shuffle_epi8(raw_desc2_3, shuf_msk);
> +		__m256i mb0_1 = _mm256_shuffle_epi8(raw_desc0_1, shuf_msk);
> +
> +		/**
> +		 * to get packet types, ptype is located in bit16-25
> +		 * of each 128bits
> +		 */
> +		const __m256i ptypes2_3 =
> +			_mm256_and_si256(raw_desc2_3, ptype_mask);
> +		const __m256i ptypes0_1 =
> +			_mm256_and_si256(raw_desc0_1, ptype_mask);
> +		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);
> +
> +		mb2_3 = _mm256_insert_epi32(mb2_3, ptype_tbl[ptype3], 4);
> +		mb2_3 = _mm256_insert_epi32(mb2_3, ptype_tbl[ptype2], 0);
> +		mb0_1 = _mm256_insert_epi32(mb0_1, ptype_tbl[ptype1], 4);
> +		mb0_1 = _mm256_insert_epi32(mb0_1, ptype_tbl[ptype0], 0);
> +		/* merge the status bits into one register */
> +		const __m256i status0_3 = _mm256_unpackhi_epi32(raw_desc2_3,
> +								raw_desc0_1);
> +
> +		/**
> +		 * take the two sets of status bits and merge to one
> +		 * After merge, the packets status flags are in the
> +		 * order (hi->lo): [1, 3, 5, 7, 0, 2, 4, 6]
> +		 */
> +		__m256i status0_7 = _mm256_unpacklo_epi64(status4_7,
> +							  status0_3);
> +
> +		/* 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);
> +
> +		__m256i l4_outer_mask = _mm256_set1_epi32(0x6);
> +		__m256i l4_outer_flags =
> +				_mm256_and_si256(l3_l4_flags, l4_outer_mask);
> +		l4_outer_flags = _mm256_slli_epi32(l4_outer_flags, 20);
> +
> +		__m256i l3_l4_mask = _mm256_set1_epi32(~0x6);
> +		l3_l4_flags = _mm256_and_si256(l3_l4_flags, l3_l4_mask);
> +		l3_l4_flags = _mm256_or_si256(l3_l4_flags, l4_outer_flags);
> +		l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
> +		/* set rss and vlan flags */
> +		const __m256i rss_vlan_flag_bits =
> +			_mm256_srli_epi32(flag_bits, 12);
> +		const __m256i rss_vlan_flags =
> +			_mm256_shuffle_epi8(rss_vlan_flags_shuf,
> +					    rss_vlan_flag_bits);
> +
> +		/* merge flags */
> +		__m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
> +				rss_vlan_flags);
> +
> +		/**
> +		 * At this point, we have the 8 sets of flags in the low 16-bits
> +		 * of each 32-bit value in vlan0.
> +		 * We want to extract these, and merge them with the mbuf init
> +		 * data so we can do a single write to the mbuf to set the flags
> +		 * and all the other initialization fields. Extracting the
> +		 * appropriate flags means that we have to do a shift and blend
> +		 * for each mbuf before we do the write. However, we can also
> +		 * add in the previously computed rx_descriptor fields to
> +		 * make a single 256-bit write per mbuf
> +		 */
> +		/* check the structure matches expectations */
> +		RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
> +				 offsetof(struct rte_mbuf, rearm_data) + 8);
> +		RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
> +				 RTE_ALIGN(offsetof(struct rte_mbuf,
> +						    rearm_data),
> +					   16));
> +		/* build up data and do writes */
> +		__m256i rearm0, rearm1, rearm2, rearm3, rearm4, rearm5,
> +			rearm6, rearm7;
> +		rearm6 = _mm256_blend_epi32(mbuf_init,
> +					    _mm256_slli_si256(mbuf_flags, 8),
> +					    0x04);
> +		rearm4 = _mm256_blend_epi32(mbuf_init,
> +					    _mm256_slli_si256(mbuf_flags, 4),
> +					    0x04);
> +		rearm2 = _mm256_blend_epi32(mbuf_init, mbuf_flags, 0x04);
> +		rearm0 = _mm256_blend_epi32(mbuf_init,
> +					    _mm256_srli_si256(mbuf_flags, 4),
> +					    0x04);
> +		/* permute to add in the rx_descriptor e.g. rss fields */
> +		rearm6 = _mm256_permute2f128_si256(rearm6, mb6_7, 0x20);
> +		rearm4 = _mm256_permute2f128_si256(rearm4, mb4_5, 0x20);
> +		rearm2 = _mm256_permute2f128_si256(rearm2, mb2_3, 0x20);
> +		rearm0 = _mm256_permute2f128_si256(rearm0, mb0_1, 0x20);
> +		/* write to mbuf */
> +		_mm256_storeu_si256((__m256i *)&rx_pkts[i + 6]->rearm_data,
> +				    rearm6);
> +		_mm256_storeu_si256((__m256i *)&rx_pkts[i + 4]->rearm_data,
> +				    rearm4);
> +		_mm256_storeu_si256((__m256i *)&rx_pkts[i + 2]->rearm_data,
> +				    rearm2);
> +		_mm256_storeu_si256((__m256i *)&rx_pkts[i + 0]->rearm_data,
> +				    rearm0);
> +
> +		/* repeat for the odd mbufs */
> +		const __m256i odd_flags =
> +			_mm256_castsi128_si256
> +				(_mm256_extracti128_si256(mbuf_flags, 1));
> +		rearm7 = _mm256_blend_epi32(mbuf_init,
> +					    _mm256_slli_si256(odd_flags, 8),
> +					    0x04);
> +		rearm5 = _mm256_blend_epi32(mbuf_init,
> +					    _mm256_slli_si256(odd_flags, 4),
> +					    0x04);
> +		rearm3 = _mm256_blend_epi32(mbuf_init, odd_flags, 0x04);
> +		rearm1 = _mm256_blend_epi32(mbuf_init,
> +					    _mm256_srli_si256(odd_flags, 4),
> +					    0x04);
> +		/* since odd mbufs are already in hi 128-bits use blend */
> +		rearm7 = _mm256_blend_epi32(rearm7, mb6_7, 0xF0);
> +		rearm5 = _mm256_blend_epi32(rearm5, mb4_5, 0xF0);
> +		rearm3 = _mm256_blend_epi32(rearm3, mb2_3, 0xF0);
> +		rearm1 = _mm256_blend_epi32(rearm1, mb0_1, 0xF0);
> +		/* again write to mbufs */
> +		_mm256_storeu_si256((__m256i *)&rx_pkts[i + 7]->rearm_data,
> +				    rearm7);
> +		_mm256_storeu_si256((__m256i *)&rx_pkts[i + 5]->rearm_data,
> +				    rearm5);
> +		_mm256_storeu_si256((__m256i *)&rx_pkts[i + 3]->rearm_data,
> +				    rearm3);
> +		_mm256_storeu_si256((__m256i *)&rx_pkts[i + 1]->rearm_data,
> +				    rearm1);
> +
> +		/* extract and record EOP bit */
> +		if (split_packet) {
> +			const __m128i eop_mask =
> +				_mm_set1_epi16(1 << VIRTCHNL2_RX_FLEX_DESC_STATUS0_EOF_S);
> +			const __m256i eop_bits256 = _mm256_and_si256(status0_7,
> +								     eop_check);
> +			/* pack status bits into a single 128-bit register */
> +			const __m128i eop_bits =
> +				_mm_packus_epi32
> +					(_mm256_castsi256_si128(eop_bits256),
> +					 _mm256_extractf128_si256(eop_bits256,
> +								  1));
> +			/**
> +			 * flip bits, and mask out the EOP bit, which is now
> +			 * a split-packet bit i.e. !EOP, rather than EOP one.
> +			 */
> +			__m128i split_bits = _mm_andnot_si128(eop_bits,
> +					eop_mask);
> +			/**
> +			 * eop bits are out of order, so we need to shuffle them
> +			 * back into order again. In doing so, only use low 8
> +			 * bits, which acts like another pack instruction
> +			 * The original order is (hi->lo): 1,3,5,7,0,2,4,6
> +			 * [Since we use epi8, the 16-bit positions are
> +			 * multiplied by 2 in the eop_shuffle value.]
> +			 */
> +			__m128i eop_shuffle =
> +				_mm_set_epi8(/* zero hi 64b */
> +					     0xFF, 0xFF, 0xFF, 0xFF,
> +					     0xFF, 0xFF, 0xFF, 0xFF,
> +					     /* move values to lo 64b */
> +					     8, 0, 10, 2,
> +					     12, 4, 14, 6);
> +			split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
> +			*(uint64_t *)split_packet =
> +				_mm_cvtsi128_si64(split_bits);
> +			split_packet += IDPF_DESCS_PER_LOOP_AVX;
> +		}

As above, if there are no plans to support multi-buffer packet reassembly,
drop this.

> +
> +		/* 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 = rte_popcount64
> +					(_mm_cvtsi128_si64
> +						(_mm256_extracti128_si256
> +							(status0_7, 1)));
> +		burst += rte_popcount64
> +				(_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 avx2 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_dp_singleq_recv_pkts_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
> +			       uint16_t nb_pkts)
> +{
> +	return _idpf_singleq_recv_raw_pkts_vec_avx2(rx_queue, rx_pkts, nb_pkts, NULL);
> +}
> diff --git a/drivers/common/idpf/meson.build b/drivers/common/idpf/meson.build
> index 46fd45c03b..4caa06a9b7 100644
> --- a/drivers/common/idpf/meson.build
> +++ b/drivers/common/idpf/meson.build
> @@ -16,6 +16,21 @@ sources = files(
>  )
>  
>  if arch_subdir == 'x86'
> +    # compile AVX2 version if either:
> +    # a. we have AVX supported in minimum instruction set baseline
> +    # b. it's not minimum instruction set, but supported by compiler
> +    if cc.get_define('__AVX2__', args: machine_args) != ''
> +        cflags += ['-DCC_AVX2_SUPPORT']
> +        sources += files('idpf_common_rxtx_avx2.c')
> +    elif cc.has_argument('-mavx2')

This logic is out-of-date, since all supported compilers have AVX2 support.
Suggest reworking using drivers/net/ice/meson.build as a reference.

> +       cflags += ['-DCC_AVX2_SUPPORT']
> +        idpf_avx2_lib = static_library('idpf_avx2_lib',
> +                'idpf_common_rxtx_avx2.c',
> +               dependencies: [static_rte_ethdev, static_rte_kvargs, static_rte_hash],
> +                include_directories: includes,
> +                c_args: [cflags, '-mavx2'])
> +       objs += idpf_avx2_lib.extract_objects('idpf_common_rxtx_avx2.c')
> +    endif
>      if cc_has_avx512
>          cflags += ['-DCC_AVX512_SUPPORT']
>          avx512_args = cflags + cc_avx512_flags
> diff --git a/drivers/common/idpf/version.map b/drivers/common/idpf/version.map
> index 0729f6b912..4510aae6b3 100644
> --- a/drivers/common/idpf/version.map
> +++ b/drivers/common/idpf/version.map
> @@ -14,6 +14,7 @@ INTERNAL {
>  	idpf_dp_splitq_recv_pkts_avx512;
>  	idpf_dp_splitq_xmit_pkts;
>  	idpf_dp_splitq_xmit_pkts_avx512;
> +	idpf_dp_singleq_recv_pkts_avx2;
>  

This list should be alphabetical, so singleq should go before splitq.

>  	idpf_qc_rx_thresh_check;
>  	idpf_qc_rx_queue_release;
> diff --git a/drivers/net/idpf/idpf_rxtx.c b/drivers/net/idpf/idpf_rxtx.c
> index 858bbefe3b..80c6c325e8 100644
> --- a/drivers/net/idpf/idpf_rxtx.c
> +++ b/drivers/net/idpf/idpf_rxtx.c
> @@ -776,6 +776,11 @@ idpf_set_rx_function(struct rte_eth_dev *dev)
>  	    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
>  		vport->rx_vec_allowed = true;
>  
> +		if ((rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
> +		     rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&

There are no CPUs that support AVX512 witout supporting AVX2 - and if there
were we probably couldn't use an AVX2 code path on them anyway. Therefore
only check the AVX2 flag and the bitwidth.

> +		    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
> +			vport->rx_use_avx2 = 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 &&
> @@ -827,6 +832,13 @@ idpf_set_rx_function(struct rte_eth_dev *dev)
>  				return;
>  			}
>  #endif /* CC_AVX512_SUPPORT */
> +			if (vport->rx_use_avx2) {
> +				PMD_DRV_LOG(NOTICE,
> +					    "Using Single AVX2 Vector Rx (port %d).",
> +					    dev->data->port_id);
> +				dev->rx_pkt_burst = idpf_dp_singleq_recv_pkts_avx2;
> +				return;
> +			}
>  		}
>  
>  		if (dev->data->scattered_rx) {
> -- 
> 2.34.1
>