From: <pbhagavatula@marvell.com>
To: <jerinj@marvell.com>, Ruifeng Wang <ruifeng.wang@arm.com>,
Vamsi Attunuru <vattunuru@marvell.com>
Cc: <dev@dpdk.org>, Pavan Nikhilesh <pbhagavatula@marvell.com>
Subject: [PATCH v5 2/2] net/octeon_ep: add Rx NEON routine
Date: Fri, 2 Feb 2024 20:36:38 +0530 [thread overview]
Message-ID: <20240202150638.15222-2-pbhagavatula@marvell.com> (raw)
In-Reply-To: <20240202150638.15222-1-pbhagavatula@marvell.com>
From: Pavan Nikhilesh <pbhagavatula@marvell.com>
Add Rx ARM NEON SIMD routine.
Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---
doc/guides/rel_notes/release_24_03.rst | 1 +
drivers/net/octeon_ep/cnxk_ep_rx_neon.c | 148 ++++++++++++++++++++++++
drivers/net/octeon_ep/meson.build | 6 +-
drivers/net/octeon_ep/otx_ep_ethdev.c | 5 +-
drivers/net/octeon_ep/otx_ep_rxtx.h | 6 +
5 files changed, 164 insertions(+), 2 deletions(-)
create mode 100644 drivers/net/octeon_ep/cnxk_ep_rx_neon.c
diff --git a/doc/guides/rel_notes/release_24_03.rst b/doc/guides/rel_notes/release_24_03.rst
index c8fcaaad6d..7a83b545cc 100644
--- a/doc/guides/rel_notes/release_24_03.rst
+++ b/doc/guides/rel_notes/release_24_03.rst
@@ -63,6 +63,7 @@ New Features
to avoid mbuf write commits.
* Added optimized SSE Rx routines.
* Added optimized AVX2 Rx routines.
+ * Added optimized NEON Rx routines.
* **Updated Marvell cnxk net driver.**
diff --git a/drivers/net/octeon_ep/cnxk_ep_rx_neon.c b/drivers/net/octeon_ep/cnxk_ep_rx_neon.c
new file mode 100644
index 0000000000..8abd8711e1
--- /dev/null
+++ b/drivers/net/octeon_ep/cnxk_ep_rx_neon.c
@@ -0,0 +1,148 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(C) 2023 Marvell.
+ */
+
+#include "cnxk_ep_rx.h"
+
+static __rte_always_inline void
+cnxk_ep_process_pkts_vec_neon(struct rte_mbuf **rx_pkts, struct otx_ep_droq *droq,
+ uint16_t new_pkts)
+{
+ const uint8x16_t mask0 = {0, 1, 0xff, 0xff, 0, 1, 0xff, 0xff,
+ 4, 5, 0xff, 0xff, 4, 5, 0xff, 0xff};
+ const uint8x16_t mask1 = {8, 9, 0xff, 0xff, 8, 9, 0xff, 0xff,
+ 12, 13, 0xff, 0xff, 12, 13, 0xff, 0xff};
+ struct rte_mbuf **recv_buf_list = droq->recv_buf_list;
+ uint32_t pidx0, pidx1, pidx2, pidx3;
+ struct rte_mbuf *m0, *m1, *m2, *m3;
+ uint32_t read_idx = droq->read_idx;
+ uint16_t nb_desc = droq->nb_desc;
+ uint32_t idx0, idx1, idx2, idx3;
+ uint64x2_t s01, s23;
+ uint32x4_t bytes;
+ uint16_t pkts = 0;
+
+ idx0 = read_idx;
+ s01 = vdupq_n_u64(0);
+ bytes = vdupq_n_u32(0);
+ while (pkts < new_pkts) {
+
+ idx1 = otx_ep_incr_index(idx0, 1, nb_desc);
+ idx2 = otx_ep_incr_index(idx1, 1, nb_desc);
+ idx3 = otx_ep_incr_index(idx2, 1, nb_desc);
+
+ if (new_pkts - pkts > 4) {
+ pidx0 = otx_ep_incr_index(idx3, 1, nb_desc);
+ pidx1 = otx_ep_incr_index(pidx0, 1, nb_desc);
+ pidx2 = otx_ep_incr_index(pidx1, 1, nb_desc);
+ pidx3 = otx_ep_incr_index(pidx2, 1, nb_desc);
+
+ rte_prefetch_non_temporal(cnxk_pktmbuf_mtod(recv_buf_list[pidx0], void *));
+ rte_prefetch_non_temporal(cnxk_pktmbuf_mtod(recv_buf_list[pidx1], void *));
+ rte_prefetch_non_temporal(cnxk_pktmbuf_mtod(recv_buf_list[pidx2], void *));
+ rte_prefetch_non_temporal(cnxk_pktmbuf_mtod(recv_buf_list[pidx3], void *));
+ }
+
+ m0 = recv_buf_list[idx0];
+ m1 = recv_buf_list[idx1];
+ m2 = recv_buf_list[idx2];
+ m3 = recv_buf_list[idx3];
+
+ /* Load packet size big-endian. */
+ s01 = vsetq_lane_u32(cnxk_pktmbuf_mtod(m0, struct otx_ep_droq_info *)->length >> 48,
+ s01, 0);
+ s01 = vsetq_lane_u32(cnxk_pktmbuf_mtod(m1, struct otx_ep_droq_info *)->length >> 48,
+ s01, 1);
+ s01 = vsetq_lane_u32(cnxk_pktmbuf_mtod(m2, struct otx_ep_droq_info *)->length >> 48,
+ s01, 2);
+ s01 = vsetq_lane_u32(cnxk_pktmbuf_mtod(m3, struct otx_ep_droq_info *)->length >> 48,
+ s01, 3);
+ /* Convert to little-endian. */
+ s01 = vrev16q_u8(s01);
+
+ /* Vertical add, consolidate outside the loop. */
+ bytes += vaddq_u32(bytes, s01);
+ /* Segregate to packet length and data length. */
+ s23 = vqtbl1q_u8(s01, mask1);
+ s01 = vqtbl1q_u8(s01, mask0);
+
+ /* Store packet length and data length to mbuf. */
+ *(uint64_t *)&m0->pkt_len = vgetq_lane_u64(s01, 0);
+ *(uint64_t *)&m1->pkt_len = vgetq_lane_u64(s01, 1);
+ *(uint64_t *)&m2->pkt_len = vgetq_lane_u64(s23, 0);
+ *(uint64_t *)&m3->pkt_len = vgetq_lane_u64(s23, 1);
+
+ /* Reset rearm data. */
+ *(uint64_t *)&m0->rearm_data = droq->rearm_data;
+ *(uint64_t *)&m1->rearm_data = droq->rearm_data;
+ *(uint64_t *)&m2->rearm_data = droq->rearm_data;
+ *(uint64_t *)&m3->rearm_data = droq->rearm_data;
+
+ rx_pkts[pkts++] = m0;
+ rx_pkts[pkts++] = m1;
+ rx_pkts[pkts++] = m2;
+ rx_pkts[pkts++] = m3;
+ idx0 = otx_ep_incr_index(idx3, 1, nb_desc);
+ }
+ droq->read_idx = idx0;
+
+ droq->refill_count += new_pkts;
+ droq->pkts_pending -= new_pkts;
+ /* Stats */
+ droq->stats.pkts_received += new_pkts;
+#if defined(RTE_ARCH_32)
+ droq->stats.bytes_received += vgetq_lane_u32(bytes, 0);
+ droq->stats.bytes_received += vgetq_lane_u32(bytes, 1);
+ droq->stats.bytes_received += vgetq_lane_u32(bytes, 2);
+ droq->stats.bytes_received += vgetq_lane_u32(bytes, 3);
+#else
+ droq->stats.bytes_received += vaddvq_u32(bytes);
+#endif
+}
+
+uint16_t __rte_noinline __rte_hot
+cnxk_ep_recv_pkts_neon(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
+{
+ struct otx_ep_droq *droq = (struct otx_ep_droq *)rx_queue;
+ uint16_t new_pkts, vpkts;
+
+ /* Refill RX buffers */
+ if (droq->refill_count >= DROQ_REFILL_THRESHOLD)
+ cnxk_ep_rx_refill(droq);
+
+ new_pkts = cnxk_ep_rx_pkts_to_process(droq, nb_pkts);
+ vpkts = RTE_ALIGN_FLOOR(new_pkts, CNXK_EP_OQ_DESC_PER_LOOP_SSE);
+ cnxk_ep_process_pkts_vec_neon(rx_pkts, droq, vpkts);
+ cnxk_ep_process_pkts_scalar(&rx_pkts[vpkts], droq, new_pkts - vpkts);
+
+ return new_pkts;
+}
+
+uint16_t __rte_noinline __rte_hot
+cn9k_ep_recv_pkts_neon(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
+{
+ struct otx_ep_droq *droq = (struct otx_ep_droq *)rx_queue;
+ uint16_t new_pkts, vpkts;
+
+ /* Refill RX buffers */
+ if (droq->refill_count >= DROQ_REFILL_THRESHOLD) {
+ cnxk_ep_rx_refill(droq);
+ } else {
+ /* SDP output goes into DROP state when output doorbell count
+ * goes below drop count. When door bell count is written with
+ * a value greater than drop count SDP output should come out
+ * of DROP state. Due to a race condition this is not happening.
+ * Writing doorbell register with 0 again may make SDP output
+ * come out of this state.
+ */
+
+ rte_write32(0, droq->pkts_credit_reg);
+ }
+
+ new_pkts = cnxk_ep_rx_pkts_to_process(droq, nb_pkts);
+ vpkts = RTE_ALIGN_FLOOR(new_pkts, CNXK_EP_OQ_DESC_PER_LOOP_SSE);
+ cnxk_ep_process_pkts_vec_neon(rx_pkts, droq, vpkts);
+ cnxk_ep_process_pkts_scalar(&rx_pkts[vpkts], droq, new_pkts - vpkts);
+
+ return new_pkts;
+}
diff --git a/drivers/net/octeon_ep/meson.build b/drivers/net/octeon_ep/meson.build
index e8ae56018d..d5d40b23a1 100644
--- a/drivers/net/octeon_ep/meson.build
+++ b/drivers/net/octeon_ep/meson.build
@@ -29,7 +29,11 @@ if arch_subdir == 'x86'
endif
endif
-extra_flags = ['-Wno-strict-aliasing']
+if arch_subdir == 'arm'
+ sources += files('cnxk_ep_rx_neon.c')
+endif
+
+extra_flags = ['-Wno-strict-aliasing', '-flax-vector-conversions']
foreach flag: extra_flags
if cc.has_argument(flag)
cflags += flag
diff --git a/drivers/net/octeon_ep/otx_ep_ethdev.c b/drivers/net/octeon_ep/otx_ep_ethdev.c
index 42a97ea110..8daa7d225c 100644
--- a/drivers/net/octeon_ep/otx_ep_ethdev.c
+++ b/drivers/net/octeon_ep/otx_ep_ethdev.c
@@ -59,6 +59,8 @@ otx_ep_set_rx_func(struct rte_eth_dev *eth_dev)
rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1)
eth_dev->rx_pkt_burst = &cnxk_ep_recv_pkts_avx;
#endif
+#elif defined(RTE_ARCH_ARM64)
+ eth_dev->rx_pkt_burst = &cnxk_ep_recv_pkts_neon;
#endif
if (otx_epvf->rx_offloads & RTE_ETH_RX_OFFLOAD_SCATTER)
eth_dev->rx_pkt_burst = &cnxk_ep_recv_pkts_mseg;
@@ -71,8 +73,9 @@ otx_ep_set_rx_func(struct rte_eth_dev *eth_dev)
rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1)
eth_dev->rx_pkt_burst = &cn9k_ep_recv_pkts_avx;
#endif
+#elif defined(RTE_ARCH_ARM64)
+ eth_dev->rx_pkt_burst = &cn9k_ep_recv_pkts_neon;
#endif
-
if (otx_epvf->rx_offloads & RTE_ETH_RX_OFFLOAD_SCATTER)
eth_dev->rx_pkt_burst = &cn9k_ep_recv_pkts_mseg;
} else {
diff --git a/drivers/net/octeon_ep/otx_ep_rxtx.h b/drivers/net/octeon_ep/otx_ep_rxtx.h
index 8f306bd94e..f5bc807dc0 100644
--- a/drivers/net/octeon_ep/otx_ep_rxtx.h
+++ b/drivers/net/octeon_ep/otx_ep_rxtx.h
@@ -60,12 +60,18 @@ cnxk_ep_recv_pkts_mseg(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t budge
uint16_t
cn9k_ep_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t budget);
+uint16_t
+cnxk_ep_recv_pkts_neon(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts);
+
uint16_t
cn9k_ep_recv_pkts_sse(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t budget);
uint16_t
cn9k_ep_recv_pkts_avx(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t budget);
+uint16_t
+cn9k_ep_recv_pkts_neon(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts);
+
uint16_t
cn9k_ep_recv_pkts_mseg(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t budget);
#endif /* _OTX_EP_RXTX_H_ */
--
2.25.1
next prev parent reply other threads:[~2024-02-02 15:06 UTC|newest]
Thread overview: 16+ messages / expand[flat|nested] mbox.gz Atom feed top
2024-01-21 16:43 [PATCH v3 1/2] net/octeon_ep: improve Rx performance pbhagavatula
2024-01-21 16:43 ` [PATCH v3 2/2] net/octeon_ep: add Rx NEON routine pbhagavatula
2024-02-01 16:38 ` Jerin Jacob
2024-02-01 22:23 ` [PATCH v2 1/2] net/octeon_ep: improve Rx performance pbhagavatula
2024-02-01 22:23 ` [PATCH v2 2/2] net/octeon_ep: add Rx NEON routine pbhagavatula
2024-02-02 8:11 ` Jerin Jacob
2024-02-02 8:43 ` [PATCH v3 1/2] net/octeon_ep: improve Rx performance pbhagavatula
2024-02-02 8:43 ` [PATCH v3 2/2] net/octeon_ep: add Rx NEON routine pbhagavatula
2024-02-02 11:09 ` Jerin Jacob
2024-02-02 11:31 ` [PATCH v4 1/2] net/octeon_ep: improve Rx performance pbhagavatula
2024-02-02 11:31 ` [PATCH v4 2/2] net/octeon_ep: add Rx NEON routine pbhagavatula
2024-02-02 15:06 ` [PATCH v5 1/2] net/octeon_ep: improve Rx performance pbhagavatula
2024-02-02 15:06 ` pbhagavatula [this message]
2024-02-05 6:16 ` [PATCH v5 2/2] net/octeon_ep: add Rx NEON routine Jerin Jacob
2024-02-02 15:09 ` [PATCH v6 1/2] net/octeon_ep: improve Rx performance pbhagavatula
2024-02-02 15:09 ` [PATCH v6 2/2] net/octeon_ep: add Rx NEON routine pbhagavatula
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20240202150638.15222-2-pbhagavatula@marvell.com \
--to=pbhagavatula@marvell.com \
--cc=dev@dpdk.org \
--cc=jerinj@marvell.com \
--cc=ruifeng.wang@arm.com \
--cc=vattunuru@marvell.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).