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

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

On Wed, Jan 08, 2025 at 05:47:57PM +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

>  doc/guides/rel_notes/release_25_03.rst      |   3 +
>  drivers/common/idpf/idpf_common_device.h    |   1 +
>  drivers/common/idpf/idpf_common_rxtx.h      |   4 +
>  drivers/common/idpf/idpf_common_rxtx_avx2.c | 225 ++++++++++++++++++++
>  drivers/common/idpf/version.map             |   1 +
>  drivers/net/idpf/idpf_rxtx.c                |  14 ++
>  6 files changed, 248 insertions(+)
> 
> diff --git a/doc/guides/rel_notes/release_25_03.rst b/doc/guides/rel_notes/release_25_03.rst
> index 426dfcd982..7ded85dac4 100644
> --- a/doc/guides/rel_notes/release_25_03.rst
> +++ b/doc/guides/rel_notes/release_25_03.rst
> @@ -55,6 +55,9 @@ New Features
>       Also, make sure to start the actual text at the margin.
>       =======================================================
>  
> +   * **Added support of vector instructions on IDPF.**
> +
> +     Added support of AVX2 instructions in IDPF single queue RX and TX path.
>  

Driver already had vector instructions so title is a little misleading.
Clarify the title to be AVX2-specific. For the body, please clarify singleq
vs splitq and what the differences are and when one might get the benefit
of the AVX2 code path.

>  Removed Items
>  -------------
> diff --git a/drivers/common/idpf/idpf_common_device.h b/drivers/common/idpf/idpf_common_device.h
> index 734be1c88a..5f3e4a4fcf 100644
> --- a/drivers/common/idpf/idpf_common_device.h
> +++ b/drivers/common/idpf/idpf_common_device.h
> @@ -124,6 +124,7 @@ struct idpf_vport {
>  	bool rx_vec_allowed;
>  	bool tx_vec_allowed;

Do we have vector paths other than the 2 AVX ones below. If not, why do we
need this flag?

>  	bool rx_use_avx2;
> +	bool tx_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 f50cf5ef46..e19e1878f3 100644
> --- a/drivers/common/idpf/idpf_common_rxtx.h
> +++ b/drivers/common/idpf/idpf_common_rxtx.h
> @@ -306,5 +306,9 @@ __rte_internal
>  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_singleq_xmit_pkts_avx2(void *tx_queue,
> +					struct rte_mbuf **tx_pkts,
> +					uint16_t nb_pkts);
>  
>  #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
> index a05b26c68a..a4bc8e2bef 100644
> --- a/drivers/common/idpf/idpf_common_rxtx_avx2.c
> +++ b/drivers/common/idpf/idpf_common_rxtx_avx2.c
> @@ -588,3 +588,228 @@ idpf_dp_singleq_recv_pkts_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
>  {
>  	return _idpf_singleq_recv_raw_pkts_vec_avx2(rx_queue, rx_pkts, nb_pkts, NULL);
>  }
> +
> +static __rte_always_inline void
> +idpf_tx_backlog_entry(struct idpf_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 __rte_always_inline int
> +idpf_singleq_tx_free_bufs_vec(struct idpf_tx_queue *txq)
> +{
> +	struct idpf_tx_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(IDPF_TXD_QW1_DTYPE_M)) !=
> +			rte_cpu_to_le_64(IDPF_TX_DESC_DTYPE_DESC_DONE))
> +		return 0;
> +
> +	n = txq->rs_thresh;
> +
> +	 /* first buffer to free from S/W ring is at index
> +	  * next_dd - (rs_thresh-1)
> +	  */
> +	txep = &txq->sw_ring[txq->next_dd - (n - 1)];
> +	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);
> +		}
> +	}
> +
> +	/* 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;
> +}
> +

If/when patchset [1] is merged, this code should be reworked to use the
common functions.

[1] https://patches.dpdk.org/project/dpdk/list/?series=34398

> +static inline void
> +idpf_singleq_vtx1(volatile struct idpf_base_tx_desc *txdp,
> +		  struct rte_mbuf *pkt, uint64_t flags)
> +{
> +	uint64_t high_qw =
> +		(IDPF_TX_DESC_DTYPE_DATA |
> +		 ((uint64_t)flags  << IDPF_TXD_QW1_CMD_S) |
> +		 ((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);
> +	_mm_store_si128((__m128i *)txdp, descriptor);
> +}
> +
> +static inline void
> +idpf_singleq_vtx(volatile struct idpf_base_tx_desc *txdp,
> +		 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
> +{
> +	const uint64_t hi_qw_tmpl = (IDPF_TX_DESC_DTYPE_DATA |
> +			((uint64_t)flags  << IDPF_TXD_QW1_CMD_S));
> +
> +	/* if unaligned on 32-bit boundary, do one to align */
> +	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
> +		idpf_singleq_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 <<
> +			 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_store_si256((void *)(txdp + 2), desc2_3);
> +		_mm256_store_si256((void *)txdp, desc0_1);
> +	}
> +
> +	/* do any last ones */
> +	while (nb_pkts) {
> +		idpf_singleq_vtx1(txdp, *pkt, flags);
> +		txdp++, pkt++, nb_pkts--;
> +	}
> +}
> +
> +static inline uint16_t
> +idpf_singleq_xmit_fixed_burst_vec_avx2(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 idpf_base_tx_desc *txdp;
> +	struct idpf_tx_entry *txep;
> +	uint16_t n, nb_commit, tx_id;
> +	uint64_t flags = IDPF_TX_DESC_CMD_EOP;
> +	uint64_t rs = IDPF_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_singleq_tx_free_bufs_vec(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 = &txq->sw_ring[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) {
> +		idpf_tx_backlog_entry(txep, tx_pkts, n);
> +
> +		idpf_singleq_vtx(txdp, tx_pkts, n - 1, flags);
> +		tx_pkts += (n - 1);
> +		txdp += (n - 1);
> +
> +		idpf_singleq_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 = &txq->sw_ring[tx_id];
> +	}
> +
> +	idpf_tx_backlog_entry(txep, tx_pkts, nb_commit);
> +
> +	idpf_singleq_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)IDPF_TX_DESC_CMD_RS) <<
> +					 IDPF_TXD_QW1_CMD_S);
> +		txq->next_rs =
> +			(uint16_t)(txq->next_rs + txq->rs_thresh);
> +	}
> +
> +	txq->tx_tail = tx_id;
> +
> +	IDPF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
> +
> +	return nb_pkts;
> +}
> +
> +uint16_t
> +idpf_dp_singleq_xmit_pkts_avx2(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_singleq_xmit_fixed_burst_vec_avx2(tx_queue, &tx_pkts[nb_tx],
> +						    num);
> +		nb_tx += ret;
> +		nb_pkts -= ret;
> +		if (ret < num)
> +			break;
> +	}
> +
> +	return nb_tx;
> +}
> diff --git a/drivers/common/idpf/version.map b/drivers/common/idpf/version.map
> index 4510aae6b3..eadcb9a2cf 100644
> --- a/drivers/common/idpf/version.map
> +++ b/drivers/common/idpf/version.map
> @@ -15,6 +15,7 @@ INTERNAL {
>  	idpf_dp_splitq_xmit_pkts;
>  	idpf_dp_splitq_xmit_pkts_avx512;
>  	idpf_dp_singleq_recv_pkts_avx2;
> +	idpf_dp_singleq_xmit_pkts_avx2;
>  
>  	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 80c6c325e8..579293b2e8 100644
> --- a/drivers/net/idpf/idpf_rxtx.c
> +++ b/drivers/net/idpf/idpf_rxtx.c
> @@ -888,6 +888,12 @@ idpf_set_tx_function(struct rte_eth_dev *dev)
>  	if (idpf_tx_vec_dev_check_default(dev) == IDPF_VECTOR_PATH &&
>  	    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
>  		vport->tx_vec_allowed = true;
> +
> +		if ((rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
> +		     rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&

As with the Rx path, only check the AVX2 flag here.

> +		    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
> +			vport->tx_use_avx2 = true;
> +
>  		if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512)
>  #ifdef CC_AVX512_SUPPORT
>  		{
> @@ -947,6 +953,14 @@ idpf_set_tx_function(struct rte_eth_dev *dev)
>  				return;
>  			}
>  #endif /* CC_AVX512_SUPPORT */
> +			if (vport->tx_use_avx2) {
> +				PMD_DRV_LOG(NOTICE,
> +					    "Using Single AVX2 Vector Tx (port %d).",
> +					    dev->data->port_id);
> +				dev->tx_pkt_burst = idpf_dp_singleq_xmit_pkts_avx2;
> +				dev->tx_pkt_prepare = idpf_dp_prep_pkts;
> +				return;
> +			}
>  		}
>  		PMD_DRV_LOG(NOTICE,
>  			    "Using Single Scalar Tx (port %d).",
> -- 
> 2.34.1
>