[dpdk-dev] [PATCH v7 8/8] net/ice: support vector AVX2 in TX

[Wenzhuo Lu <wenzhuo.lu@intel.com>]

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
---
 doc/guides/nics/ice.rst                |  18 ++++
 doc/guides/rel_notes/release_19_05.rst |   4 +
 drivers/net/ice/ice_rxtx.c             |  13 ++-
 drivers/net/ice/ice_rxtx.h             |   2 +
 drivers/net/ice/ice_rxtx_vec_avx2.c    | 158 +++++++++++++++++++++++++++++++++
 5 files changed, 193 insertions(+), 2 deletions(-)

diff --git a/doc/guides/nics/ice.rst b/doc/guides/nics/ice.rst
index 3998d5e..fdbc02e 100644
--- a/doc/guides/nics/ice.rst
+++ b/doc/guides/nics/ice.rst
@@ -64,6 +64,24 @@ Driver compilation and testing
 Refer to the document :ref:`compiling and testing a PMD for a NIC <pmd_build_and_test>`
 for details.
 
+Features
+--------
+
+Vector PMD
+~~~~~~~~~~
+
+Vector PMD for RX and TX path are selected automatically. The paths
+are chosen based on 2 conditions.
+
+- ``CPU``
+  On the X86 platform, the driver checks if the CPU supports AVX2.
+  If it's supported, AVX2 paths will be chosen. If not, SSE is chosen.
+
+- ``Offload features``
+  The supported HW offload features are described in the document ice_vec.ini.
+  If any not supported features are used, ICE vector PMD is disabled and the
+  normal paths are chosen.
+
 Sample Application Notes
 ------------------------
 
diff --git a/doc/guides/rel_notes/release_19_05.rst b/doc/guides/rel_notes/release_19_05.rst
index 6f76de3..fbea42f 100644
--- a/doc/guides/rel_notes/release_19_05.rst
+++ b/doc/guides/rel_notes/release_19_05.rst
@@ -96,6 +96,10 @@ New Features
   Improved testpmd application performance on ARM platform. For ``macswap``
   forwarding mode, NEON intrinsics were used to do swap to save CPU cycles.
 
+* **Added support of vector instructions on ICE.**
+
+   Added support of SSE and AVX2 instructions in ICE RX and TX path.
+
 
 Removed Items
 -------------
diff --git a/drivers/net/ice/ice_rxtx.c b/drivers/net/ice/ice_rxtx.c
index 860155f..5264055 100644
--- a/drivers/net/ice/ice_rxtx.c
+++ b/drivers/net/ice/ice_rxtx.c
@@ -2356,15 +2356,24 @@ void __attribute__((cold))
 #ifdef RTE_ARCH_X86
 	struct ice_tx_queue *txq;
 	int i;
+	bool use_avx2 = false;
 
 	if (!ice_tx_vec_dev_check(dev)) {
 		for (i = 0; i < dev->data->nb_tx_queues; i++) {
 			txq = dev->data->tx_queues[i];
 			(void)ice_txq_vec_setup(txq);
 		}
-		PMD_DRV_LOG(DEBUG, "Using Vector Tx (port %d).",
+
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
+		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
+			use_avx2 = true;
+
+		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
+			    use_avx2 ? "avx2 " : "",
 			    dev->data->port_id);
-		dev->tx_pkt_burst = ice_xmit_pkts_vec;
+		dev->tx_pkt_burst = use_avx2 ?
+				    ice_xmit_pkts_vec_avx2 :
+				    ice_xmit_pkts_vec;
 		dev->tx_pkt_prepare = NULL;
 
 		return;
diff --git a/drivers/net/ice/ice_rxtx.h b/drivers/net/ice/ice_rxtx.h
index dfc3224..64e9f20 100644
--- a/drivers/net/ice/ice_rxtx.h
+++ b/drivers/net/ice/ice_rxtx.h
@@ -184,4 +184,6 @@ uint16_t ice_recv_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
 uint16_t ice_recv_scattered_pkts_vec_avx2(void *rx_queue,
 					  struct rte_mbuf **rx_pkts,
 					  uint16_t nb_pkts);
+uint16_t ice_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
+				uint16_t nb_pkts);
 #endif /* _ICE_RXTX_H_ */
diff --git a/drivers/net/ice/ice_rxtx_vec_avx2.c b/drivers/net/ice/ice_rxtx_vec_avx2.c
index 2459ff3..fac869a 100644
--- a/drivers/net/ice/ice_rxtx_vec_avx2.c
+++ b/drivers/net/ice/ice_rxtx_vec_avx2.c
@@ -684,3 +684,161 @@
 	return retval + ice_recv_scattered_burst_vec_avx2(rx_queue,
 				rx_pkts + retval, nb_pkts);
 }
+
+static inline void
+ice_vtx1(volatile struct ice_tx_desc *txdp,
+	 struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(ICE_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << ICE_TXD_QW1_CMD_S) |
+		 ((uint64_t)pkt->data_len << ICE_TXD_QW1_TX_BUF_SZ_S));
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+				pkt->buf_physaddr + pkt->data_off);
+	_mm_store_si128((__m128i *)txdp, descriptor);
+}
+
+static inline void
+ice_vtx(volatile struct ice_tx_desc *txdp,
+	struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (ICE_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << ICE_TXD_QW1_CMD_S));
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		ice_vtx1(txdp, *pkt, flags);
+		nb_pkts--, txdp++, pkt++;
+	}
+
+	/* do two at a time while possible, in bursts */
+	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
+		uint64_t hi_qw3 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[3]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw2 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[2]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw1 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[1]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		uint64_t hi_qw0 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[0]->data_len <<
+			 ICE_TXD_QW1_TX_BUF_SZ_S);
+
+		__m256i desc2_3 =
+			_mm256_set_epi64x
+				(hi_qw3,
+				 pkt[3]->buf_physaddr + pkt[3]->data_off,
+				 hi_qw2,
+				 pkt[2]->buf_physaddr + pkt[2]->data_off);
+		__m256i desc0_1 =
+			_mm256_set_epi64x
+				(hi_qw1,
+				 pkt[1]->buf_physaddr + pkt[1]->data_off,
+				 hi_qw0,
+				 pkt[0]->buf_physaddr + pkt[0]->data_off);
+		_mm256_store_si256((void *)(txdp + 2), desc2_3);
+		_mm256_store_si256((void *)txdp, desc0_1);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		ice_vtx1(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+ice_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
+			      uint16_t nb_pkts)
+{
+	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
+	volatile struct ice_tx_desc *txdp;
+	struct ice_tx_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	uint64_t flags = ICE_TD_CMD;
+	uint64_t rs = ICE_TX_DESC_CMD_RS | ICE_TD_CMD;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+
+	if (txq->nb_tx_free < txq->tx_free_thresh)
+		ice_tx_free_bufs(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = &txq->sw_ring[tx_id];
+
+	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		ice_tx_backlog_entry(txep, tx_pkts, n);
+
+		ice_vtx(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		ice_vtx1(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = &txq->sw_ring[tx_id];
+	}
+
+	ice_tx_backlog_entry(txep, tx_pkts, nb_commit);
+
+	ice_vtx(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->tx_next_rs) {
+		txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)ICE_TX_DESC_CMD_RS) <<
+					 ICE_TXD_QW1_CMD_S);
+		txq->tx_next_rs =
+			(uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	ICE_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+ice_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
+		       uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+		ret = ice_xmit_fixed_burst_vec_avx2(tx_queue, &tx_pkts[nb_tx],
+						    num);
+		nb_tx += ret;
+		nb_pkts -= ret;
+		if (ret < num)
+			break;
+	}
+
+	return nb_tx;
+}
-- 
1.9.3