Re: [PATCH v2 2/2] net/idpf: enable AVX2 for split queue Tx

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

On Thu, Sep 25, 2025 at 02:50:20PM +0530, Shaiq Wani wrote:
> In case some CPUs don't support AVX512. Enable AVX2 for them to
> get better per-core performance.
> 
> In the single queue model, the same descriptor queue is used by SW
> to post descriptors to the device and used by device to report completed
> descriptors to SW. While as the split queue model separates them into
> different queues for parallel processing and improved performance.
> 
> Signed-off-by: Shaiq Wani <shaiq.wani@intel.com>
> ---

Hi,

see comments inline below.

/Bruce

>  drivers/net/intel/idpf/idpf_common_rxtx.h     |   3 +
>  .../net/intel/idpf/idpf_common_rxtx_avx2.c    | 202 ++++++++++++++++++
>  drivers/net/intel/idpf/idpf_rxtx.c            |   9 +
>  3 files changed, 214 insertions(+)
> 
> diff --git a/drivers/net/intel/idpf/idpf_common_rxtx.h b/drivers/net/intel/idpf/idpf_common_rxtx.h
> index 3a9af06c86..ef3199524a 100644
> --- a/drivers/net/intel/idpf/idpf_common_rxtx.h
> +++ b/drivers/net/intel/idpf/idpf_common_rxtx.h
> @@ -262,6 +262,9 @@ uint16_t idpf_dp_singleq_recv_pkts_avx2(void *rx_queue,
>  					struct rte_mbuf **rx_pkts,
>  					uint16_t nb_pkts);
>  __rte_internal
> +uint16_t idpf_dp_splitq_xmit_pkts_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
> +				       uint16_t nb_pkts);
> +__rte_internal
>  uint16_t idpf_dp_singleq_xmit_pkts_avx2(void *tx_queue,
>  					struct rte_mbuf **tx_pkts,
>  					uint16_t nb_pkts);
> diff --git a/drivers/net/intel/idpf/idpf_common_rxtx_avx2.c b/drivers/net/intel/idpf/idpf_common_rxtx_avx2.c
> index b24653f195..7b28c6b32d 100644
> --- a/drivers/net/intel/idpf/idpf_common_rxtx_avx2.c
> +++ b/drivers/net/intel/idpf/idpf_common_rxtx_avx2.c
> @@ -889,3 +889,205 @@ idpf_dp_singleq_xmit_pkts_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
>  
>  	return nb_tx;
>  }
> +
> +static __rte_always_inline void
> +idpf_splitq_scan_cq_ring(struct ci_tx_queue *cq)
> +{
> +	struct idpf_splitq_tx_compl_desc *compl_ring;
> +	struct ci_tx_queue *txq;
> +	uint16_t genid, txq_qid, cq_qid, i;
> +	uint8_t ctype;
> +
> +	cq_qid = cq->tx_tail;
> +
> +	for (i = 0; i < IDPD_TXQ_SCAN_CQ_THRESH; i++) {
> +		if (cq_qid == cq->nb_tx_desc) {
> +			cq_qid = 0;
> +			cq->expected_gen_id ^= 1;  /* toggle generation bit */
> +		}
> +
> +		compl_ring = &cq->compl_ring[cq_qid];
> +
> +		genid = (rte_le_to_cpu_16(compl_ring->qid_comptype_gen) &
> +			 IDPF_TXD_COMPLQ_GEN_M) >> IDPF_TXD_COMPLQ_GEN_S;
> +
> +		if (genid != cq->expected_gen_id)
> +			break;
> +
> +		ctype = (rte_le_to_cpu_16(compl_ring->qid_comptype_gen) &
> +			 IDPF_TXD_COMPLQ_COMPL_TYPE_M) >> IDPF_TXD_COMPLQ_COMPL_TYPE_S;
> +
> +		txq_qid = (rte_le_to_cpu_16(compl_ring->qid_comptype_gen) &
> +			   IDPF_TXD_COMPLQ_QID_M) >> IDPF_TXD_COMPLQ_QID_S;
> +
> +		txq = cq->txqs[txq_qid - cq->tx_start_qid];
> +		if (ctype == IDPF_TXD_COMPLT_RS)
> +			txq->rs_compl_count++;
> +
> +		cq_qid++;
> +	}
> +
> +	cq->tx_tail = cq_qid;
> +}
> +
> +static __rte_always_inline void
> +idpf_splitq_vtx1_avx2(volatile struct idpf_flex_tx_sched_desc *txdp,
> +				struct rte_mbuf *pkt, uint64_t flags)
> +{
> +	uint64_t high_qw =
> +	IDPF_TX_DESC_DTYPE_FLEX_FLOW_SCHE |
> +	((uint64_t)flags) |
> +	((uint64_t)pkt->data_len << IDPF_TXD_QW1_TX_BUF_SZ_S);
> +
> +	__m128i descriptor = _mm_set_epi64x(high_qw,
> +				pkt->buf_iova + pkt->data_off);

Indentation is a little off here. The high_qw line continuations should be
indented beyond the first tab-stop. For the descriptor definition, a triple
indent is a bit excessive, double indent should be sufficient.

> +	_mm_storeu_si128(RTE_CAST_PTR(__m128i *, txdp), descriptor);
> +}
> +
> +
> +static inline void
> +idpf_splitq_vtx_avx2(volatile struct idpf_flex_tx_sched_desc *txdp,
> +				struct rte_mbuf **pkt, uint16_t nb_pkts, uint64_t flags)
> +{
> +	const uint64_t hi_qw_tmpl = IDPF_TX_DESC_DTYPE_FLEX_FLOW_SCHE |
> +	((uint64_t)flags);

Again, indent more here.

> +
> +	/* align if needed */
> +	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
> +		idpf_splitq_vtx1_avx2(txdp, *pkt, flags);
> +		txdp++, pkt++, nb_pkts--;
> +	}
> +
> +	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
> +		uint64_t hi_qw3 = hi_qw_tmpl |
> +			((uint64_t)pkt[3]->data_len << IDPF_TXD_QW1_TX_BUF_SZ_S);
> +		uint64_t hi_qw2 = hi_qw_tmpl |
> +			((uint64_t)pkt[2]->data_len << IDPF_TXD_QW1_TX_BUF_SZ_S);
> +		uint64_t hi_qw1 = hi_qw_tmpl |
> +			((uint64_t)pkt[1]->data_len << IDPF_TXD_QW1_TX_BUF_SZ_S);
> +		uint64_t hi_qw0 = hi_qw_tmpl |
> +			((uint64_t)pkt[0]->data_len << IDPF_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_storeu_si256(RTE_CAST_PTR(__m256i *, txdp + 2), desc2_3);
> +		_mm256_storeu_si256(RTE_CAST_PTR(__m256i *, txdp), desc0_1);
> +	}

Rather than casting away the volatile here, did you consider casting it
away earlier and passing in txdp as non-volatile parameter to this
function?

> +
> +	while (nb_pkts--) {
> +		idpf_splitq_vtx1_avx2(txdp, *pkt, flags);
> +		txdp++;
> +		pkt++;
> +	}
> +}
> +
> +static inline uint16_t
> +idpf_splitq_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
> +						uint16_t nb_pkts)
> +{
> +	struct ci_tx_queue *txq = (struct ci_tx_queue *)tx_queue;
> +	volatile struct idpf_flex_tx_sched_desc *txdp;
> +	struct ci_tx_entry_vec *txep;
> +	uint16_t n, nb_commit, tx_id;
> +	uint64_t cmd_dtype = IDPF_TXD_FLEX_FLOW_CMD_EOP;
> +
> +	tx_id = txq->tx_tail;
> +
> +	/* restrict to max burst size */
> +	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
> +
> +	/* make sure we have enough free space */
> +	if (txq->nb_tx_free < txq->tx_free_thresh)
> +		ci_tx_free_bufs_vec(txq, idpf_tx_desc_done, false);
> +
> +	nb_commit = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
> +	nb_pkts = nb_commit;
> +	if (unlikely(nb_pkts == 0))
> +		return 0;
> +
> +	txdp = &txq->desc_ring[tx_id];

Suggestion: I would cast away the volatile here and save later casting
away.

> +	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) {
> +		ci_tx_backlog_entry_vec(txep, tx_pkts, n);
> +
> +		idpf_splitq_vtx_avx2(txdp, tx_pkts, n - 1, cmd_dtype);
> +		tx_pkts += (n - 1);
> +		txdp += (n - 1);
> +
> +		idpf_splitq_vtx1_avx2(txdp, *tx_pkts++, cmd_dtype);
> +
> +		nb_commit = (uint16_t)(nb_commit - n);
> +
> +		tx_id = 0;
> +		txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
> +
> +		txdp = &txq->desc_ring[tx_id];
> +		txep = (void *)txq->sw_ring;
> +		txep += tx_id;
> +	}
> +
> +	ci_tx_backlog_entry_vec(txep, tx_pkts, nb_commit);
> +
> +	idpf_splitq_vtx_avx2(txdp, tx_pkts, nb_commit, cmd_dtype);
> +
> +	tx_id = (uint16_t)(tx_id + nb_commit);
> +	if (tx_id > txq->tx_next_rs)
> +		txq->tx_next_rs =
> +		(uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
> +
> +	txq->tx_tail = tx_id;
> +
> +	IDPF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
> +
> +	return nb_pkts;
> +}
> +
> +static __rte_always_inline uint16_t
> +idpf_splitq_xmit_pkts_vec_avx2_cmn(void *tx_queue, struct rte_mbuf **tx_pkts,
> +						uint16_t nb_pkts)
> +{
> +	struct ci_tx_queue *txq = (struct ci_tx_queue *)tx_queue;
> +	uint16_t nb_tx = 0;
> +
> +	while (nb_pkts) {
> +		uint16_t ret, num;
> +		idpf_splitq_scan_cq_ring(txq->complq);
> +
> +		if (txq->rs_compl_count > txq->tx_free_thresh) {
> +			ci_tx_free_bufs_vec(txq, idpf_tx_desc_done, false);
> +			txq->rs_compl_count -= txq->tx_rs_thresh;
> +		}
> +
> +		num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
> +		ret = idpf_splitq_xmit_fixed_burst_vec_avx2(tx_queue,
> +								&tx_pkts[nb_tx],
> +								num);

Reduce over-indent here. If you only indent by 2 extra, you can put the
last two parameters on the one line.

> +		nb_tx += ret;
> +		nb_pkts -= ret;
> +		if (ret < num)
> +			break;
> +	}
> +
> +	return nb_tx;
> +}
> +
> +RTE_EXPORT_INTERNAL_SYMBOL(idpf_dp_splitq_xmit_pkts_avx2)
> +uint16_t
> +idpf_dp_splitq_xmit_pkts_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
> +						uint16_t nb_pkts)
> +{
> +	return idpf_splitq_xmit_pkts_vec_avx2_cmn(tx_queue, tx_pkts, nb_pkts);
> +}

Do we need a separate wrapper function here. Are there future plans for
another different wrapper around the same common function?

> diff --git a/drivers/net/intel/idpf/idpf_rxtx.c b/drivers/net/intel/idpf/idpf_rxtx.c
> index 1c725065df..6950fabb49 100644
> --- a/drivers/net/intel/idpf/idpf_rxtx.c
> +++ b/drivers/net/intel/idpf/idpf_rxtx.c
> @@ -850,6 +850,15 @@ idpf_set_tx_function(struct rte_eth_dev *dev)
>  				return;
>  			}
>  #endif /* CC_AVX512_SUPPORT */
> +			if (tx_simd_width == RTE_VECT_SIMD_256) {
> +				PMD_DRV_LOG(NOTICE,
> +					    "Using Split AVX2 Vector Tx (port %d).",
> +					    dev->data->port_id);
> +				dev->tx_pkt_burst = idpf_dp_splitq_xmit_pkts_avx2;
> +				dev->tx_pkt_prepare = idpf_dp_prep_pkts;
> +				return;
> +			}
> +
>  		}
>  		PMD_DRV_LOG(NOTICE,
>  			    "Using Split Scalar Tx (port %d).",
> -- 
> 2.34.1
>