* [dpdk-dev] [PATCH 3/3] net/bnxt: add AVX2 vector PMD
2021-05-24 18:59 [dpdk-dev] [PATCH 0/3] net/bnxt: vector mode patches Lance Richardson
2021-05-24 18:59 ` [dpdk-dev] [PATCH 1/3] net/bnxt: refactor HW ptype mapping table Lance Richardson
2021-05-24 18:59 ` [dpdk-dev] [PATCH 2/3] net/bnxt: fix Rx burst size constraint Lance Richardson
@ 2021-05-24 18:59 ` Lance Richardson
2021-05-26 18:33 ` Lance Richardson
2021-06-07 21:36 ` [dpdk-dev] [PATCH 0/3] net/bnxt: vector mode patches Ajit Khaparde
3 siblings, 1 reply; 6+ messages in thread
From: Lance Richardson @ 2021-05-24 18:59 UTC (permalink / raw)
To: Ajit Khaparde, Somnath Kotur, Bruce Richardson,
Konstantin Ananyev, Jerin Jacob, Ruifeng Wang
Cc: dev
[-- Attachment #1: Type: text/plain, Size: 40275 bytes --]
Implement AVX2 vector PMD.
Signed-off-by: Lance Richardson <lance.richardson@broadcom.com>
Reviewed-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
---
doc/guides/nics/bnxt.rst | 57 ++-
drivers/net/bnxt/bnxt_ethdev.c | 119 +++--
drivers/net/bnxt/bnxt_rxr.c | 4 +-
drivers/net/bnxt/bnxt_rxr.h | 11 +-
drivers/net/bnxt/bnxt_rxtx_vec_avx2.c | 597 ++++++++++++++++++++++++++
drivers/net/bnxt/bnxt_rxtx_vec_neon.c | 25 +-
drivers/net/bnxt/bnxt_rxtx_vec_sse.c | 31 +-
drivers/net/bnxt/bnxt_txr.h | 7 +
drivers/net/bnxt/meson.build | 17 +
9 files changed, 780 insertions(+), 88 deletions(-)
create mode 100644 drivers/net/bnxt/bnxt_rxtx_vec_avx2.c
diff --git a/doc/guides/nics/bnxt.rst b/doc/guides/nics/bnxt.rst
index 0fb2032447..feb0c6a765 100644
--- a/doc/guides/nics/bnxt.rst
+++ b/doc/guides/nics/bnxt.rst
@@ -853,23 +853,36 @@ DPDK implements a light-weight library to allow PMDs to be bonded together and p
Vector Processing
-----------------
+The BNXT PMD provides vectorized burst transmit/receive function implementations
+on x86-based platforms using SSE (Streaming SIMD Extensions) and AVX2 (Advanced
+Vector Extensions 2) instructions, and on Arm-based platforms using Arm Neon
+Advanced SIMD instructions. Vector processing support is currently implemented
+only for Intel/AMD and Arm CPU architectures.
+
Vector processing provides significantly improved performance over scalar
-processing (see Vector Processor, here).
+processing. This improved performance is derived from a number of optimizations:
+
+* Using SIMD instructions to operate on multiple packets in parallel.
+* Using SIMD instructions to do more work per instruction than is possible
+ with scalar instructions, for example by leveraging 128-bit and 256-bi
+ load/store instructions or by using SIMD shuffle and permute operations.
+* Batching
-The BNXT PMD supports the vector processing using SSE (Streaming SIMD
-Extensions) instructions on x86 platforms. It also supports NEON intrinsics for
-vector processing on ARM CPUs. The BNXT vPMD (vector mode PMD) is available for
-Intel/AMD and ARM CPU architectures.
+ * TX: transmit completions are processed in bulk.
+ * RX: bulk allocation of mbufs is used when allocating rxq buffers.
-This improved performance comes from several optimizations:
+* Simplifications enabled by not supporting chained mbufs in vector mode.
+* Simplifications enabled by not supporting some stateless offloads in vector
+ mode:
-* Batching
- * TX: processing completions in bulk
- * RX: allocating mbufs in bulk
-* Chained mbufs are *not* supported, i.e. a packet should fit a single mbuf
-* Some stateless offloads are *not* supported with vector processing
- * TX: no offloads will be supported
- * RX: reduced RX offloads (listed below) will be supported::
+ * TX: only the following reduced set of transmit offloads is supported in
+ vector mode::
+
+ DEV_TX_OFFLOAD_MBUF_FAST_FREE
+
+ * RX: only the following reduced set of receive offloads is supported in
+ vector mode (note that jumbo MTU is allowed only when the MTU setting
+ does not require `DEV_RX_OFFLOAD_SCATTER` to be enabled)::
DEV_RX_OFFLOAD_VLAN_STRIP
DEV_RX_OFFLOAD_KEEP_CRC
@@ -878,23 +891,21 @@ This improved performance comes from several optimizations:
DEV_RX_OFFLOAD_UDP_CKSUM
DEV_RX_OFFLOAD_TCP_CKSUM
DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM
+ DEV_RX_OFFLOAD_OUTER_UDP_CKSUM
DEV_RX_OFFLOAD_RSS_HASH
DEV_RX_OFFLOAD_VLAN_FILTER
-The BNXT Vector PMD is enabled in DPDK builds by default.
-
-However, a decision to enable vector mode will be made when the port transitions
-from stopped to started. Any TX offloads or some RX offloads (other than listed
-above) will disable the vector mode.
-Offload configuration changes that impact vector mode must be made when the port
-is stopped.
+The BNXT Vector PMD is enabled in DPDK builds by default. The decision to enable
+vector processing is made at run-time when the port is started; if no transmit
+offloads outside the set supported for vector mode are enabled then vector mode
+transmit will be enabled, and if no receive offloads outside the set supported
+for vector mode are enabled then vector mode receive will be enabled. Offload
+configuration changes that impact the decision to enable vector mode are allowed
+only when the port is stopped.
Note that TX (or RX) vector mode can be enabled independently from RX (or TX)
vector mode.
-Also vector mode is allowed when jumbo is enabled
-as long as the MTU setting does not require scattered Rx.
-
Appendix
--------
diff --git a/drivers/net/bnxt/bnxt_ethdev.c b/drivers/net/bnxt/bnxt_ethdev.c
index 0208795fd2..a7d056a34f 100644
--- a/drivers/net/bnxt/bnxt_ethdev.c
+++ b/drivers/net/bnxt/bnxt_ethdev.c
@@ -1178,32 +1178,57 @@ bnxt_receive_function(struct rte_eth_dev *eth_dev)
return bnxt_recv_pkts;
}
-#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
-#ifndef RTE_LIBRTE_IEEE1588
+#if (defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)) && \
+ !defined(RTE_LIBRTE_IEEE1588)
+
+ /* Vector mode receive cannot be enabled if scattered rx is in use. */
+ if (eth_dev->data->scattered_rx)
+ goto use_scalar_rx;
+
/*
- * Vector mode receive can be enabled only if scatter rx is not
- * in use and rx offloads are limited to VLAN stripping and
- * CRC stripping.
+ * Vector mode receive cannot be enabled if Truflow is enabled or if
+ * asynchronous completions and receive completions can be placed in
+ * the same completion ring.
*/
- if (!eth_dev->data->scattered_rx &&
- !(eth_dev->data->dev_conf.rxmode.offloads &
- ~(DEV_RX_OFFLOAD_VLAN_STRIP |
- DEV_RX_OFFLOAD_KEEP_CRC |
- DEV_RX_OFFLOAD_JUMBO_FRAME |
- DEV_RX_OFFLOAD_IPV4_CKSUM |
- DEV_RX_OFFLOAD_UDP_CKSUM |
- DEV_RX_OFFLOAD_TCP_CKSUM |
- DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM |
- DEV_RX_OFFLOAD_OUTER_UDP_CKSUM |
- DEV_RX_OFFLOAD_RSS_HASH |
- DEV_RX_OFFLOAD_VLAN_FILTER)) &&
- !BNXT_TRUFLOW_EN(bp) && BNXT_NUM_ASYNC_CPR(bp) &&
- rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
- PMD_DRV_LOG(INFO, "Using vector mode receive for port %d\n",
+ if (BNXT_TRUFLOW_EN(bp) || !BNXT_NUM_ASYNC_CPR(bp))
+ goto use_scalar_rx;
+
+ /*
+ * Vector mode receive cannot be enabled if any receive offloads outside
+ * a limited subset have been enabled.
+ */
+ if (eth_dev->data->dev_conf.rxmode.offloads &
+ ~(DEV_RX_OFFLOAD_VLAN_STRIP |
+ DEV_RX_OFFLOAD_KEEP_CRC |
+ DEV_RX_OFFLOAD_JUMBO_FRAME |
+ DEV_RX_OFFLOAD_IPV4_CKSUM |
+ DEV_RX_OFFLOAD_UDP_CKSUM |
+ DEV_RX_OFFLOAD_TCP_CKSUM |
+ DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM |
+ DEV_RX_OFFLOAD_OUTER_UDP_CKSUM |
+ DEV_RX_OFFLOAD_RSS_HASH |
+ DEV_RX_OFFLOAD_VLAN_FILTER))
+ goto use_scalar_rx;
+
+#if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
+ if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256 &&
+ rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1) {
+ PMD_DRV_LOG(INFO,
+ "Using AVX2 vector mode receive for port %d\n",
+ eth_dev->data->port_id);
+ bp->flags |= BNXT_FLAG_RX_VECTOR_PKT_MODE;
+ return bnxt_recv_pkts_vec_avx2;
+ }
+ #endif
+ if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
+ PMD_DRV_LOG(INFO,
+ "Using SSE vector mode receive for port %d\n",
eth_dev->data->port_id);
bp->flags |= BNXT_FLAG_RX_VECTOR_PKT_MODE;
return bnxt_recv_pkts_vec;
}
+
+use_scalar_rx:
PMD_DRV_LOG(INFO, "Vector mode receive disabled for port %d\n",
eth_dev->data->port_id);
PMD_DRV_LOG(INFO,
@@ -1211,7 +1236,6 @@ bnxt_receive_function(struct rte_eth_dev *eth_dev)
eth_dev->data->port_id,
eth_dev->data->scattered_rx,
eth_dev->data->dev_conf.rxmode.offloads);
-#endif
#endif
bp->flags &= ~BNXT_FLAG_RX_VECTOR_PKT_MODE;
return bnxt_recv_pkts;
@@ -1226,22 +1250,36 @@ bnxt_transmit_function(struct rte_eth_dev *eth_dev)
if (BNXT_CHIP_SR2(bp))
return bnxt_xmit_pkts;
-#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
-#ifndef RTE_LIBRTE_IEEE1588
+#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64) && \
+ !defined(RTE_LIBRTE_IEEE1588)
uint64_t offloads = eth_dev->data->dev_conf.txmode.offloads;
/*
* Vector mode transmit can be enabled only if not using scatter rx
* or tx offloads.
*/
- if (!eth_dev->data->scattered_rx &&
- !(offloads & ~DEV_TX_OFFLOAD_MBUF_FAST_FREE) &&
- !BNXT_TRUFLOW_EN(bp) &&
- rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
- PMD_DRV_LOG(INFO, "Using vector mode transmit for port %d\n",
+ if (eth_dev->data->scattered_rx ||
+ (offloads & ~DEV_TX_OFFLOAD_MBUF_FAST_FREE) ||
+ BNXT_TRUFLOW_EN(bp))
+ goto use_scalar_tx;
+
+#if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
+ if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256 &&
+ rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1) {
+ PMD_DRV_LOG(INFO,
+ "Using AVX2 vector mode transmit for port %d\n",
+ eth_dev->data->port_id);
+ return bnxt_xmit_pkts_vec_avx2;
+ }
+#endif
+ if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
+ PMD_DRV_LOG(INFO,
+ "Using SSE vector mode transmit for port %d\n",
eth_dev->data->port_id);
return bnxt_xmit_pkts_vec;
}
+
+use_scalar_tx:
PMD_DRV_LOG(INFO, "Vector mode transmit disabled for port %d\n",
eth_dev->data->port_id);
PMD_DRV_LOG(INFO,
@@ -1249,7 +1287,6 @@ bnxt_transmit_function(struct rte_eth_dev *eth_dev)
eth_dev->data->port_id,
eth_dev->data->scattered_rx,
offloads);
-#endif
#endif
return bnxt_xmit_pkts;
}
@@ -2859,11 +2896,15 @@ static const struct {
eth_rx_burst_t pkt_burst;
const char *info;
} bnxt_rx_burst_info[] = {
- {bnxt_recv_pkts, "Scalar"},
+ {bnxt_recv_pkts, "Scalar"},
#if defined(RTE_ARCH_X86)
- {bnxt_recv_pkts_vec, "Vector SSE"},
-#elif defined(RTE_ARCH_ARM64)
- {bnxt_recv_pkts_vec, "Vector Neon"},
+ {bnxt_recv_pkts_vec, "Vector SSE"},
+#endif
+#if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
+ {bnxt_recv_pkts_vec_avx2, "Vector AVX2"},
+#endif
+#if defined(RTE_ARCH_ARM64)
+ {bnxt_recv_pkts_vec, "Vector Neon"},
#endif
};
@@ -2889,11 +2930,15 @@ static const struct {
eth_tx_burst_t pkt_burst;
const char *info;
} bnxt_tx_burst_info[] = {
- {bnxt_xmit_pkts, "Scalar"},
+ {bnxt_xmit_pkts, "Scalar"},
#if defined(RTE_ARCH_X86)
- {bnxt_xmit_pkts_vec, "Vector SSE"},
-#elif defined(RTE_ARCH_ARM64)
- {bnxt_xmit_pkts_vec, "Vector Neon"},
+ {bnxt_xmit_pkts_vec, "Vector SSE"},
+#endif
+#if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
+ {bnxt_xmit_pkts_vec_avx2, "Vector AVX2"},
+#endif
+#if defined(RTE_ARCH_ARM64)
+ {bnxt_xmit_pkts_vec, "Vector Neon"},
#endif
};
diff --git a/drivers/net/bnxt/bnxt_rxr.c b/drivers/net/bnxt/bnxt_rxr.c
index a6a8fb213b..4eef75f6be 100644
--- a/drivers/net/bnxt/bnxt_rxr.c
+++ b/drivers/net/bnxt/bnxt_rxr.c
@@ -1147,7 +1147,7 @@ int bnxt_init_rx_ring_struct(struct bnxt_rx_queue *rxq, unsigned int socket_id)
/* Allocate extra rx ring entries for vector rx. */
ring->vmem_size = sizeof(struct rte_mbuf *) *
- (ring->ring_size + RTE_BNXT_DESCS_PER_LOOP);
+ (ring->ring_size + BNXT_RX_EXTRA_MBUF_ENTRIES);
ring->vmem = (void **)&rxr->rx_buf_ring;
ring->fw_ring_id = INVALID_HW_RING_ID;
@@ -1251,7 +1251,7 @@ int bnxt_init_one_rx_ring(struct bnxt_rx_queue *rxq)
/* Initialize dummy mbuf pointers for vector mode rx. */
for (i = ring->ring_size;
- i < ring->ring_size + RTE_BNXT_DESCS_PER_LOOP; i++) {
+ i < ring->ring_size + BNXT_RX_EXTRA_MBUF_ENTRIES; i++) {
rxr->rx_buf_ring[i] = &rxq->fake_mbuf;
}
diff --git a/drivers/net/bnxt/bnxt_rxr.h b/drivers/net/bnxt/bnxt_rxr.h
index 79f1458698..955bf3e99e 100644
--- a/drivers/net/bnxt/bnxt_rxr.h
+++ b/drivers/net/bnxt/bnxt_rxr.h
@@ -42,7 +42,12 @@ static inline uint16_t bnxt_tpa_start_agg_id(struct bnxt *bp,
RX_PKT_CMPL_AGG_BUFS_SFT)
/* Number of descriptors to process per inner loop in vector mode. */
-#define RTE_BNXT_DESCS_PER_LOOP 4U
+#define BNXT_RX_DESCS_PER_LOOP_VEC128 4U /* SSE, Neon */
+#define BNXT_RX_DESCS_PER_LOOP_VEC256 8U /* AVX2 */
+
+/* Number of extra Rx mbuf ring entries to allocate for vector mode. */
+#define BNXT_RX_EXTRA_MBUF_ENTRIES \
+ RTE_MAX(BNXT_RX_DESCS_PER_LOOP_VEC128, BNXT_RX_DESCS_PER_LOOP_VEC256)
#define BNXT_OL_FLAGS_TBL_DIM 64
#define BNXT_OL_FLAGS_ERR_TBL_DIM 32
@@ -106,6 +111,10 @@ uint16_t bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
int bnxt_rxq_vec_setup(struct bnxt_rx_queue *rxq);
#endif
+#if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
+uint16_t bnxt_recv_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
+ uint16_t nb_pkts);
+#endif
void bnxt_set_mark_in_mbuf(struct bnxt *bp,
struct rx_pkt_cmpl_hi *rxcmp1,
struct rte_mbuf *mbuf);
diff --git a/drivers/net/bnxt/bnxt_rxtx_vec_avx2.c b/drivers/net/bnxt/bnxt_rxtx_vec_avx2.c
new file mode 100644
index 0000000000..a06dfec90e
--- /dev/null
+++ b/drivers/net/bnxt/bnxt_rxtx_vec_avx2.c
@@ -0,0 +1,597 @@
+/* SPDX-License-Identifier: BSD-3-Clause */
+/* Copyright(c) 2019-2021 Broadcom All rights reserved. */
+
+#include <inttypes.h>
+#include <stdbool.h>
+
+#include <rte_bitmap.h>
+#include <rte_byteorder.h>
+#include <rte_malloc.h>
+#include <rte_memory.h>
+#include <rte_vect.h>
+
+#include "bnxt.h"
+#include "bnxt_cpr.h"
+#include "bnxt_ring.h"
+
+#include "bnxt_txq.h"
+#include "bnxt_txr.h"
+#include "bnxt_rxtx_vec_common.h"
+#include <unistd.h>
+
+static uint16_t
+recv_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
+{
+ struct bnxt_rx_queue *rxq = rx_queue;
+ const __m256i mbuf_init =
+ _mm256_set_epi64x(0, 0, 0, rxq->mbuf_initializer);
+ struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
+ struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
+ uint16_t cp_ring_size = cpr->cp_ring_struct->ring_size;
+ uint16_t rx_ring_size = rxr->rx_ring_struct->ring_size;
+ struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
+ uint64_t valid, desc_valid_mask = ~0ULL;
+ const __m256i info3_v_mask = _mm256_set1_epi32(CMPL_BASE_V);
+ uint32_t raw_cons = cpr->cp_raw_cons;
+ uint32_t cons, mbcons;
+ int nb_rx_pkts = 0;
+ int i;
+ const __m256i valid_target =
+ _mm256_set1_epi32(!!(raw_cons & cp_ring_size));
+ const __m256i dsc_shuf_msk =
+ _mm256_set_epi8(0xff, 0xff, 0xff, 0xff, /* Zeroes. */
+ 7, 6, /* metadata type */
+ 9, 8, /* flags2 low 16 */
+ 5, 4, /* vlan_tci */
+ 1, 0, /* errors_v2 */
+ 0xff, 0xff, 0xff, 0xff, /* Zeroes. */
+ 0xff, 0xff, 0xff, 0xff, /* Zeroes. */
+ 7, 6, /* metadata type */
+ 9, 8, /* flags2 low 16 */
+ 5, 4, /* vlan_tci */
+ 1, 0, /* errors_v2 */
+ 0xff, 0xff, 0xff, 0xff); /* Zeroes. */
+ const __m256i shuf_msk =
+ _mm256_set_epi8(15, 14, 13, 12, /* rss */
+ 7, 6, /* vlan_tci */
+ 3, 2, /* data_len */
+ 0xFF, 0xFF, 3, 2, /* pkt_len */
+ 0xFF, 0xFF, 0xFF, 0xFF, /* pkt_type (zeroes) */
+ 15, 14, 13, 12, /* rss */
+ 7, 6, /* vlan_tci */
+ 3, 2, /* data_len */
+ 0xFF, 0xFF, 3, 2, /* pkt_len */
+ 0xFF, 0xFF, 0xFF, 0xFF); /* pkt_type (zeroes) */
+ const __m256i flags_type_mask =
+ _mm256_set1_epi32(RX_PKT_CMPL_FLAGS_ITYPE_MASK);
+ const __m256i flags2_mask1 =
+ _mm256_set1_epi32(CMPL_FLAGS2_VLAN_TUN_MSK);
+ const __m256i flags2_mask2 =
+ _mm256_set1_epi32(RX_PKT_CMPL_FLAGS2_IP_TYPE);
+ const __m256i rss_mask =
+ _mm256_set1_epi32(RX_PKT_CMPL_FLAGS_RSS_VALID);
+ __m256i t0, t1, flags_type, flags2, index, errors;
+ __m256i ptype_idx, ptypes, is_tunnel;
+ __m256i mbuf01, mbuf23, mbuf45, mbuf67;
+ __m256i rearm0, rearm1, rearm2, rearm3, rearm4, rearm5, rearm6, rearm7;
+ __m256i ol_flags, ol_flags_hi;
+ __m256i rss_flags;
+
+ /* Validate ptype table indexing at build time. */
+ bnxt_check_ptype_constants();
+
+ /* If Rx Q was stopped return */
+ if (unlikely(!rxq->rx_started))
+ return 0;
+
+ if (rxq->rxrearm_nb >= rxq->rx_free_thresh)
+ bnxt_rxq_rearm(rxq, rxr);
+
+ nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, BNXT_RX_DESCS_PER_LOOP_VEC256);
+
+ cons = raw_cons & (cp_ring_size - 1);
+ mbcons = (raw_cons / 2) & (rx_ring_size - 1);
+
+ /* Prefetch first four descriptor pairs. */
+ rte_prefetch0(&cp_desc_ring[cons + 0]);
+ rte_prefetch0(&cp_desc_ring[cons + 4]);
+ rte_prefetch0(&cp_desc_ring[cons + 8]);
+ rte_prefetch0(&cp_desc_ring[cons + 12]);
+
+ /* Ensure that we do not go past the ends of the rings. */
+ nb_pkts = RTE_MIN(nb_pkts, RTE_MIN(rx_ring_size - mbcons,
+ (cp_ring_size - cons) / 2));
+ /*
+ * If we are at the end of the ring, ensure that descriptors after the
+ * last valid entry are not treated as valid. Otherwise, force the
+ * maximum number of packets to receive to be a multiple of the per-
+ * loop count.
+ */
+ if (nb_pkts < BNXT_RX_DESCS_PER_LOOP_VEC256) {
+ desc_valid_mask >>=
+ CHAR_BIT * (BNXT_RX_DESCS_PER_LOOP_VEC256 - nb_pkts);
+ } else {
+ nb_pkts =
+ RTE_ALIGN_FLOOR(nb_pkts, BNXT_RX_DESCS_PER_LOOP_VEC256);
+ }
+
+ /* Handle RX burst request */
+ for (i = 0; i < nb_pkts; i += BNXT_RX_DESCS_PER_LOOP_VEC256,
+ cons += BNXT_RX_DESCS_PER_LOOP_VEC256 * 2,
+ mbcons += BNXT_RX_DESCS_PER_LOOP_VEC256) {
+ __m256i desc0, desc1, desc2, desc3, desc4, desc5, desc6, desc7;
+ __m256i rxcmp0_1, rxcmp2_3, rxcmp4_5, rxcmp6_7, info3_v;
+ __m256i errors_v2;
+ uint32_t num_valid;
+
+ /* Copy eight mbuf pointers to output array. */
+ t0 = _mm256_loadu_si256((void *)&rxr->rx_buf_ring[mbcons]);
+ _mm256_storeu_si256((void *)&rx_pkts[i], t0);
+#ifdef RTE_ARCH_X86_64
+ t0 = _mm256_loadu_si256((void *)&rxr->rx_buf_ring[mbcons + 4]);
+ _mm256_storeu_si256((void *)&rx_pkts[i + 4], t0);
+#endif
+
+ /* Prefetch eight descriptor pairs for next iteration. */
+ if (i + BNXT_RX_DESCS_PER_LOOP_VEC256 < nb_pkts) {
+ rte_prefetch0(&cp_desc_ring[cons + 16]);
+ rte_prefetch0(&cp_desc_ring[cons + 20]);
+ rte_prefetch0(&cp_desc_ring[cons + 24]);
+ rte_prefetch0(&cp_desc_ring[cons + 28]);
+ }
+
+ /*
+ * Load eight receive completion descriptors into 256-bit
+ * registers. Loads are issued in reverse order in order to
+ * ensure consistent state.
+ */
+ desc7 = _mm256_load_si256((void *)&cp_desc_ring[cons + 14]);
+ rte_compiler_barrier();
+ desc6 = _mm256_load_si256((void *)&cp_desc_ring[cons + 12]);
+ rte_compiler_barrier();
+ desc5 = _mm256_load_si256((void *)&cp_desc_ring[cons + 10]);
+ rte_compiler_barrier();
+ desc4 = _mm256_load_si256((void *)&cp_desc_ring[cons + 8]);
+ rte_compiler_barrier();
+ desc3 = _mm256_load_si256((void *)&cp_desc_ring[cons + 6]);
+ rte_compiler_barrier();
+ desc2 = _mm256_load_si256((void *)&cp_desc_ring[cons + 4]);
+ rte_compiler_barrier();
+ desc1 = _mm256_load_si256((void *)&cp_desc_ring[cons + 2]);
+ rte_compiler_barrier();
+ desc0 = _mm256_load_si256((void *)&cp_desc_ring[cons + 0]);
+
+ /*
+ * Pack needed fields from each descriptor into a compressed
+ * 128-bit layout and pair two compressed descriptors into
+ * 256-bit registers. The 128-bit compressed layout is as
+ * follows:
+ * Bits 0-15: flags_type field from low completion record.
+ * Bits 16-31: len field from low completion record.
+ * Bits 32-47: flags2 (low 16 bits) from high completion.
+ * Bits 48-79: metadata from high completion record.
+ * Bits 80-95: errors_v2 from high completion record.
+ * Bits 96-127: rss hash from low completion record.
+ */
+ t0 = _mm256_permute2f128_si256(desc6, desc7, 0x20);
+ t1 = _mm256_permute2f128_si256(desc6, desc7, 0x31);
+ t1 = _mm256_shuffle_epi8(t1, dsc_shuf_msk);
+ rxcmp6_7 = _mm256_blend_epi32(t0, t1, 0x66);
+
+ t0 = _mm256_permute2f128_si256(desc4, desc5, 0x20);
+ t1 = _mm256_permute2f128_si256(desc4, desc5, 0x31);
+ t1 = _mm256_shuffle_epi8(t1, dsc_shuf_msk);
+ rxcmp4_5 = _mm256_blend_epi32(t0, t1, 0x66);
+
+ t0 = _mm256_permute2f128_si256(desc2, desc3, 0x20);
+ t1 = _mm256_permute2f128_si256(desc2, desc3, 0x31);
+ t1 = _mm256_shuffle_epi8(t1, dsc_shuf_msk);
+ rxcmp2_3 = _mm256_blend_epi32(t0, t1, 0x66);
+
+ t0 = _mm256_permute2f128_si256(desc0, desc1, 0x20);
+ t1 = _mm256_permute2f128_si256(desc0, desc1, 0x31);
+ t1 = _mm256_shuffle_epi8(t1, dsc_shuf_msk);
+ rxcmp0_1 = _mm256_blend_epi32(t0, t1, 0x66);
+
+ /* Compute packet type table indices for eight packets. */
+ t0 = _mm256_unpacklo_epi32(rxcmp0_1, rxcmp2_3);
+ t1 = _mm256_unpacklo_epi32(rxcmp4_5, rxcmp6_7);
+ flags_type = _mm256_unpacklo_epi64(t0, t1);
+ ptype_idx = _mm256_and_si256(flags_type, flags_type_mask);
+ ptype_idx = _mm256_srli_epi32(ptype_idx,
+ RX_PKT_CMPL_FLAGS_ITYPE_SFT -
+ BNXT_PTYPE_TBL_TYPE_SFT);
+
+ t0 = _mm256_unpacklo_epi32(rxcmp0_1, rxcmp2_3);
+ t1 = _mm256_unpacklo_epi32(rxcmp4_5, rxcmp6_7);
+ flags2 = _mm256_unpackhi_epi64(t0, t1);
+
+ t0 = _mm256_srli_epi32(_mm256_and_si256(flags2, flags2_mask1),
+ RX_PKT_CMPL_FLAGS2_META_FORMAT_SFT -
+ BNXT_PTYPE_TBL_VLAN_SFT);
+ ptype_idx = _mm256_or_si256(ptype_idx, t0);
+
+ t0 = _mm256_srli_epi32(_mm256_and_si256(flags2, flags2_mask2),
+ RX_PKT_CMPL_FLAGS2_IP_TYPE_SFT -
+ BNXT_PTYPE_TBL_IP_VER_SFT);
+ ptype_idx = _mm256_or_si256(ptype_idx, t0);
+
+ /*
+ * Load ptypes for eight packets using gather. Gather operations
+ * have extremely high latency (~19 cycles), execution and use
+ * of result should be separated as much as possible.
+ */
+ ptypes = _mm256_i32gather_epi32((int *)bnxt_ptype_table,
+ ptype_idx, sizeof(uint32_t));
+ /*
+ * Compute ol_flags and checksum error table indices for eight
+ * packets.
+ */
+ is_tunnel = _mm256_and_si256(flags2, _mm256_set1_epi32(4));
+ is_tunnel = _mm256_slli_epi32(is_tunnel, 3);
+ flags2 = _mm256_and_si256(flags2, _mm256_set1_epi32(0x1F));
+
+ /* Extract errors_v2 fields for eight packets. */
+ t0 = _mm256_unpackhi_epi32(rxcmp0_1, rxcmp2_3);
+ t1 = _mm256_unpackhi_epi32(rxcmp4_5, rxcmp6_7);
+ errors_v2 = _mm256_unpacklo_epi64(t0, t1);
+
+ errors = _mm256_srli_epi32(errors_v2, 4);
+ errors = _mm256_and_si256(errors, _mm256_set1_epi32(0xF));
+ errors = _mm256_and_si256(errors, flags2);
+
+ index = _mm256_andnot_si256(errors, flags2);
+ errors = _mm256_or_si256(errors,
+ _mm256_srli_epi32(is_tunnel, 1));
+ index = _mm256_or_si256(index, is_tunnel);
+
+ /*
+ * Load ol_flags for eight packets using gather. Gather
+ * operations have extremely high latency (~19 cycles),
+ * execution and use of result should be separated as much
+ * as possible.
+ */
+ ol_flags = _mm256_i32gather_epi32((int *)rxr->ol_flags_table,
+ index, sizeof(uint32_t));
+ errors = _mm256_i32gather_epi32((int *)rxr->ol_flags_err_table,
+ errors, sizeof(uint32_t));
+
+ /*
+ * Pack the 128-bit array of valid descriptor flags into 64
+ * bits and count the number of set bits in order to determine
+ * the number of valid descriptors.
+ */
+ const __m256i perm_msk =
+ _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
+ info3_v = _mm256_permutevar8x32_epi32(errors_v2, perm_msk);
+ info3_v = _mm256_and_si256(errors_v2, info3_v_mask);
+ info3_v = _mm256_xor_si256(info3_v, valid_target);
+
+ info3_v = _mm256_packs_epi32(info3_v, _mm256_setzero_si256());
+ valid = _mm_cvtsi128_si64(_mm256_extracti128_si256(info3_v, 1));
+ valid = (valid << CHAR_BIT) |
+ _mm_cvtsi128_si64(_mm256_castsi256_si128(info3_v));
+ num_valid = __builtin_popcountll(valid & desc_valid_mask);
+
+ if (num_valid == 0)
+ break;
+
+ /* Update mbuf rearm_data for eight packets. */
+ mbuf01 = _mm256_shuffle_epi8(rxcmp0_1, shuf_msk);
+ mbuf23 = _mm256_shuffle_epi8(rxcmp2_3, shuf_msk);
+ mbuf45 = _mm256_shuffle_epi8(rxcmp4_5, shuf_msk);
+ mbuf67 = _mm256_shuffle_epi8(rxcmp6_7, shuf_msk);
+
+ /* Blend in ptype field for two mbufs at a time. */
+ mbuf01 = _mm256_blend_epi32(mbuf01, ptypes, 0x11);
+ mbuf23 = _mm256_blend_epi32(mbuf23,
+ _mm256_srli_si256(ptypes, 4), 0x11);
+ mbuf45 = _mm256_blend_epi32(mbuf45,
+ _mm256_srli_si256(ptypes, 8), 0x11);
+ mbuf67 = _mm256_blend_epi32(mbuf67,
+ _mm256_srli_si256(ptypes, 12), 0x11);
+
+ /* Unpack rearm data, set fixed fields for first four mbufs. */
+ rearm0 = _mm256_permute2f128_si256(mbuf_init, mbuf01, 0x20);
+ rearm1 = _mm256_blend_epi32(mbuf_init, mbuf01, 0xF0);
+ rearm2 = _mm256_permute2f128_si256(mbuf_init, mbuf23, 0x20);
+ rearm3 = _mm256_blend_epi32(mbuf_init, mbuf23, 0xF0);
+
+ /* Compute final ol_flags values for eight packets. */
+ rss_flags = _mm256_and_si256(flags_type, rss_mask);
+ rss_flags = _mm256_srli_epi32(rss_flags, 9);
+ ol_flags = _mm256_or_si256(ol_flags, errors);
+ ol_flags = _mm256_or_si256(ol_flags, rss_flags);
+ ol_flags_hi = _mm256_permute2f128_si256(ol_flags,
+ ol_flags, 0x11);
+
+ /* Set ol_flags fields for first four packets. */
+ rearm0 = _mm256_blend_epi32(rearm0,
+ _mm256_slli_si256(ol_flags, 8),
+ 0x04);
+ rearm1 = _mm256_blend_epi32(rearm1,
+ _mm256_slli_si256(ol_flags_hi, 8),
+ 0x04);
+ rearm2 = _mm256_blend_epi32(rearm2,
+ _mm256_slli_si256(ol_flags, 4),
+ 0x04);
+ rearm3 = _mm256_blend_epi32(rearm3,
+ _mm256_slli_si256(ol_flags_hi, 4),
+ 0x04);
+
+ /* Store all mbuf fields for first four packets. */
+ _mm256_storeu_si256((void *)&rx_pkts[i + 0]->rearm_data,
+ rearm0);
+ _mm256_storeu_si256((void *)&rx_pkts[i + 1]->rearm_data,
+ rearm1);
+ _mm256_storeu_si256((void *)&rx_pkts[i + 2]->rearm_data,
+ rearm2);
+ _mm256_storeu_si256((void *)&rx_pkts[i + 3]->rearm_data,
+ rearm3);
+
+ /* Unpack rearm data, set fixed fields for final four mbufs. */
+ rearm4 = _mm256_permute2f128_si256(mbuf_init, mbuf45, 0x20);
+ rearm5 = _mm256_blend_epi32(mbuf_init, mbuf45, 0xF0);
+ rearm6 = _mm256_permute2f128_si256(mbuf_init, mbuf67, 0x20);
+ rearm7 = _mm256_blend_epi32(mbuf_init, mbuf67, 0xF0);
+
+ /* Set ol_flags fields for final four packets. */
+ rearm4 = _mm256_blend_epi32(rearm4, ol_flags, 0x04);
+ rearm5 = _mm256_blend_epi32(rearm5, ol_flags_hi, 0x04);
+ rearm6 = _mm256_blend_epi32(rearm6,
+ _mm256_srli_si256(ol_flags, 4),
+ 0x04);
+ rearm7 = _mm256_blend_epi32(rearm7,
+ _mm256_srli_si256(ol_flags_hi, 4),
+ 0x04);
+
+ /* Store all mbuf fields for final four packets. */
+ _mm256_storeu_si256((void *)&rx_pkts[i + 4]->rearm_data,
+ rearm4);
+ _mm256_storeu_si256((void *)&rx_pkts[i + 5]->rearm_data,
+ rearm5);
+ _mm256_storeu_si256((void *)&rx_pkts[i + 6]->rearm_data,
+ rearm6);
+ _mm256_storeu_si256((void *)&rx_pkts[i + 7]->rearm_data,
+ rearm7);
+
+ nb_rx_pkts += num_valid;
+ if (num_valid < BNXT_RX_DESCS_PER_LOOP_VEC256)
+ break;
+ }
+
+ if (nb_rx_pkts) {
+ rxr->rx_raw_prod = RING_ADV(rxr->rx_raw_prod, nb_rx_pkts);
+
+ rxq->rxrearm_nb += nb_rx_pkts;
+ cpr->cp_raw_cons += 2 * nb_rx_pkts;
+ bnxt_db_cq(cpr);
+ }
+
+ return nb_rx_pkts;
+}
+
+uint16_t
+bnxt_recv_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
+ uint16_t nb_pkts)
+{
+ uint16_t cnt = 0;
+
+ while (nb_pkts > RTE_BNXT_MAX_RX_BURST) {
+ uint16_t burst;
+
+ burst = recv_burst_vec_avx2(rx_queue, rx_pkts + cnt,
+ RTE_BNXT_MAX_RX_BURST);
+
+ cnt += burst;
+ nb_pkts -= burst;
+
+ if (burst < RTE_BNXT_MAX_RX_BURST)
+ return cnt;
+ }
+ return cnt + recv_burst_vec_avx2(rx_queue, rx_pkts + cnt, nb_pkts);
+}
+
+static void
+bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq)
+{
+ struct bnxt_cp_ring_info *cpr = txq->cp_ring;
+ uint32_t raw_cons = cpr->cp_raw_cons;
+ uint32_t cons;
+ uint32_t nb_tx_pkts = 0;
+ struct tx_cmpl *txcmp;
+ struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
+ struct bnxt_ring *cp_ring_struct = cpr->cp_ring_struct;
+ uint32_t ring_mask = cp_ring_struct->ring_mask;
+
+ do {
+ cons = RING_CMPL(ring_mask, raw_cons);
+ txcmp = (struct tx_cmpl *)&cp_desc_ring[cons];
+
+ if (!CMP_VALID(txcmp, raw_cons, cp_ring_struct))
+ break;
+
+ nb_tx_pkts += txcmp->opaque;
+ raw_cons = NEXT_RAW_CMP(raw_cons);
+ } while (nb_tx_pkts < ring_mask);
+
+ if (nb_tx_pkts) {
+ if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
+ bnxt_tx_cmp_vec_fast(txq, nb_tx_pkts);
+ else
+ bnxt_tx_cmp_vec(txq, nb_tx_pkts);
+ cpr->cp_raw_cons = raw_cons;
+ bnxt_db_cq(cpr);
+ }
+}
+
+static inline void
+bnxt_xmit_one(struct rte_mbuf *mbuf, struct tx_bd_long *txbd,
+ struct rte_mbuf **tx_buf)
+{
+ uint64_t dsc_hi, dsc_lo;
+ __m128i desc;
+
+ *tx_buf = mbuf;
+
+ dsc_hi = mbuf->buf_iova + mbuf->data_off;
+ dsc_lo = (mbuf->data_len << 16) |
+ bnxt_xmit_flags_len(mbuf->data_len, TX_BD_FLAGS_NOCMPL);
+
+ desc = _mm_set_epi64x(dsc_hi, dsc_lo);
+ _mm_store_si128((void *)txbd, desc);
+}
+
+static uint16_t
+bnxt_xmit_fixed_burst_vec(struct bnxt_tx_queue *txq, struct rte_mbuf **pkts,
+ uint16_t nb_pkts)
+{
+ struct bnxt_tx_ring_info *txr = txq->tx_ring;
+ uint16_t tx_prod, tx_raw_prod = txr->tx_raw_prod;
+ struct tx_bd_long *txbd;
+ struct rte_mbuf **tx_buf;
+ uint16_t to_send;
+
+ tx_prod = RING_IDX(txr->tx_ring_struct, tx_raw_prod);
+ txbd = &txr->tx_desc_ring[tx_prod];
+ tx_buf = &txr->tx_buf_ring[tx_prod];
+
+ /* Prefetch next transmit buffer descriptors. */
+ rte_prefetch0(txbd);
+ rte_prefetch0(txbd + 3);
+
+ nb_pkts = RTE_MIN(nb_pkts, bnxt_tx_avail(txq));
+
+ if (unlikely(nb_pkts == 0))
+ return 0;
+
+ /* Handle TX burst request */
+ to_send = nb_pkts;
+
+ /*
+ * If current descriptor is not on a 32-byte boundary, send one packet
+ * to align for 32-byte stores.
+ */
+ if (tx_prod & 1) {
+ bnxt_xmit_one(pkts[0], txbd++, tx_buf++);
+ to_send--;
+ pkts++;
+ }
+
+ /*
+ * Send four packets per loop, with a single store for each pair
+ * of descriptors.
+ */
+ while (to_send >= BNXT_TX_DESCS_PER_LOOP) {
+ uint64_t dsc0_hi, dsc0_lo, dsc1_hi, dsc1_lo;
+ uint64_t dsc2_hi, dsc2_lo, dsc3_hi, dsc3_lo;
+ __m256i dsc01, dsc23;
+
+ /* Prefetch next transmit buffer descriptors. */
+ rte_prefetch0(txbd + 4);
+ rte_prefetch0(txbd + 7);
+
+ /* Copy four mbuf pointers to tx buf ring. */
+#ifdef RTE_ARCH_X86_64
+ __m256i tmp = _mm256_loadu_si256((void *)pkts);
+ _mm256_storeu_si256((void *)tx_buf, tmp);
+#else
+ __m128i tmp = _mm_loadu_si128((void *)pkts);
+ _mm_storeu_si128((void *)tx_buf, tmp);
+#endif
+
+ dsc0_hi = tx_buf[0]->buf_iova + tx_buf[0]->data_off;
+ dsc0_lo = (tx_buf[0]->data_len << 16) |
+ bnxt_xmit_flags_len(tx_buf[0]->data_len,
+ TX_BD_FLAGS_NOCMPL);
+
+ dsc1_hi = tx_buf[1]->buf_iova + tx_buf[1]->data_off;
+ dsc1_lo = (tx_buf[1]->data_len << 16) |
+ bnxt_xmit_flags_len(tx_buf[1]->data_len,
+ TX_BD_FLAGS_NOCMPL);
+
+ dsc01 = _mm256_set_epi64x(dsc1_hi, dsc1_lo, dsc0_hi, dsc0_lo);
+
+ dsc2_hi = tx_buf[2]->buf_iova + tx_buf[2]->data_off;
+ dsc2_lo = (tx_buf[2]->data_len << 16) |
+ bnxt_xmit_flags_len(tx_buf[2]->data_len,
+ TX_BD_FLAGS_NOCMPL);
+
+ dsc3_hi = tx_buf[3]->buf_iova + tx_buf[3]->data_off;
+ dsc3_lo = (tx_buf[3]->data_len << 16) |
+ bnxt_xmit_flags_len(tx_buf[3]->data_len,
+ TX_BD_FLAGS_NOCMPL);
+
+ dsc23 = _mm256_set_epi64x(dsc3_hi, dsc3_lo, dsc2_hi, dsc2_lo);
+
+ _mm256_store_si256((void *)txbd, dsc01);
+ _mm256_store_si256((void *)(txbd + 2), dsc23);
+
+ to_send -= BNXT_TX_DESCS_PER_LOOP;
+ pkts += BNXT_TX_DESCS_PER_LOOP;
+ txbd += BNXT_TX_DESCS_PER_LOOP;
+ tx_buf += BNXT_TX_DESCS_PER_LOOP;
+ }
+
+ /* Send any remaining packets, writing each descriptor individually. */
+ while (to_send) {
+ bnxt_xmit_one(pkts[0], txbd++, tx_buf++);
+ to_send--;
+ pkts++;
+ }
+
+ /* Request a completion for the final packet of the burst. */
+ txbd[-1].opaque = nb_pkts;
+ txbd[-1].flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
+
+ tx_raw_prod += nb_pkts;
+ bnxt_db_write(&txr->tx_db, tx_raw_prod);
+
+ txr->tx_raw_prod = tx_raw_prod;
+
+ return nb_pkts;
+}
+
+uint16_t
+bnxt_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
+ uint16_t nb_pkts)
+{
+ int nb_sent = 0;
+ struct bnxt_tx_queue *txq = tx_queue;
+ struct bnxt_tx_ring_info *txr = txq->tx_ring;
+ uint16_t ring_size = txr->tx_ring_struct->ring_size;
+
+ /* Tx queue was stopped; wait for it to be restarted */
+ if (unlikely(!txq->tx_started)) {
+ PMD_DRV_LOG(DEBUG, "Tx q stopped;return\n");
+ return 0;
+ }
+
+ /* Handle TX completions */
+ if (bnxt_tx_bds_in_hw(txq) >= txq->tx_free_thresh)
+ bnxt_handle_tx_cp_vec(txq);
+
+ while (nb_pkts) {
+ uint16_t ret, num;
+
+ /*
+ * Ensure that no more than RTE_BNXT_MAX_TX_BURST packets
+ * are transmitted before the next completion.
+ */
+ num = RTE_MIN(nb_pkts, RTE_BNXT_MAX_TX_BURST);
+
+ /*
+ * Ensure that a ring wrap does not occur within a call to
+ * bnxt_xmit_fixed_burst_vec().
+ */
+ num = RTE_MIN(num, ring_size -
+ (txr->tx_raw_prod & (ring_size - 1)));
+ ret = bnxt_xmit_fixed_burst_vec(txq, &tx_pkts[nb_sent], num);
+ nb_sent += ret;
+ nb_pkts -= ret;
+ if (ret < num)
+ break;
+ }
+
+ return nb_sent;
+}
diff --git a/drivers/net/bnxt/bnxt_rxtx_vec_neon.c b/drivers/net/bnxt/bnxt_rxtx_vec_neon.c
index a6e630ea5e..b4e9202568 100644
--- a/drivers/net/bnxt/bnxt_rxtx_vec_neon.c
+++ b/drivers/net/bnxt/bnxt_rxtx_vec_neon.c
@@ -200,17 +200,20 @@ recv_burst_vec_neon(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
* maximum number of packets to receive to be a multiple of the per-
* loop count.
*/
- if (nb_pkts < RTE_BNXT_DESCS_PER_LOOP)
- desc_valid_mask >>= 16 * (RTE_BNXT_DESCS_PER_LOOP - nb_pkts);
- else
- nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_BNXT_DESCS_PER_LOOP);
+ if (nb_pkts < BNXT_RX_DESCS_PER_LOOP_VEC128) {
+ desc_valid_mask >>=
+ 16 * (BNXT_RX_DESCS_PER_LOOP_VEC128 - nb_pkts);
+ } else {
+ nb_pkts =
+ RTE_ALIGN_FLOOR(nb_pkts, BNXT_RX_DESCS_PER_LOOP_VEC128);
+ }
/* Handle RX burst request */
- for (i = 0; i < nb_pkts; i += RTE_BNXT_DESCS_PER_LOOP,
- cons += RTE_BNXT_DESCS_PER_LOOP * 2,
- mbcons += RTE_BNXT_DESCS_PER_LOOP) {
- uint32x4_t rxcmp1[RTE_BNXT_DESCS_PER_LOOP];
- uint32x4_t rxcmp[RTE_BNXT_DESCS_PER_LOOP];
+ for (i = 0; i < nb_pkts; i += BNXT_RX_DESCS_PER_LOOP_VEC128,
+ cons += BNXT_RX_DESCS_PER_LOOP_VEC128 * 2,
+ mbcons += BNXT_RX_DESCS_PER_LOOP_VEC128) {
+ uint32x4_t rxcmp1[BNXT_RX_DESCS_PER_LOOP_VEC128];
+ uint32x4_t rxcmp[BNXT_RX_DESCS_PER_LOOP_VEC128];
uint32x4_t info3_v;
uint64x2_t t0, t1;
uint32_t num_valid;
@@ -226,7 +229,7 @@ recv_burst_vec_neon(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
#endif
/* Prefetch four descriptor pairs for next iteration. */
- if (i + RTE_BNXT_DESCS_PER_LOOP < nb_pkts) {
+ if (i + BNXT_RX_DESCS_PER_LOOP_VEC128 < nb_pkts) {
rte_prefetch0(&cp_desc_ring[cons + 8]);
rte_prefetch0(&cp_desc_ring[cons + 12]);
}
@@ -284,7 +287,7 @@ recv_burst_vec_neon(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
rxr);
nb_rx_pkts += num_valid;
- if (num_valid < RTE_BNXT_DESCS_PER_LOOP)
+ if (num_valid < BNXT_RX_DESCS_PER_LOOP_VEC128)
break;
}
diff --git a/drivers/net/bnxt/bnxt_rxtx_vec_sse.c b/drivers/net/bnxt/bnxt_rxtx_vec_sse.c
index fe074f82cf..c479697ac0 100644
--- a/drivers/net/bnxt/bnxt_rxtx_vec_sse.c
+++ b/drivers/net/bnxt/bnxt_rxtx_vec_sse.c
@@ -191,17 +191,20 @@ recv_burst_vec_sse(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
* maximum number of packets to receive to be a multiple of the per-
* loop count.
*/
- if (nb_pkts < RTE_BNXT_DESCS_PER_LOOP)
- desc_valid_mask >>= 16 * (RTE_BNXT_DESCS_PER_LOOP - nb_pkts);
- else
- nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_BNXT_DESCS_PER_LOOP);
+ if (nb_pkts < BNXT_RX_DESCS_PER_LOOP_VEC128) {
+ desc_valid_mask >>=
+ 16 * (BNXT_RX_DESCS_PER_LOOP_VEC128 - nb_pkts);
+ } else {
+ nb_pkts =
+ RTE_ALIGN_FLOOR(nb_pkts, BNXT_RX_DESCS_PER_LOOP_VEC128);
+ }
/* Handle RX burst request */
- for (i = 0; i < nb_pkts; i += RTE_BNXT_DESCS_PER_LOOP,
- cons += RTE_BNXT_DESCS_PER_LOOP * 2,
- mbcons += RTE_BNXT_DESCS_PER_LOOP) {
- __m128i rxcmp1[RTE_BNXT_DESCS_PER_LOOP];
- __m128i rxcmp[RTE_BNXT_DESCS_PER_LOOP];
+ for (i = 0; i < nb_pkts; i += BNXT_RX_DESCS_PER_LOOP_VEC128,
+ cons += BNXT_RX_DESCS_PER_LOOP_VEC128 * 2,
+ mbcons += BNXT_RX_DESCS_PER_LOOP_VEC128) {
+ __m128i rxcmp1[BNXT_RX_DESCS_PER_LOOP_VEC128];
+ __m128i rxcmp[BNXT_RX_DESCS_PER_LOOP_VEC128];
__m128i tmp0, tmp1, info3_v;
uint32_t num_valid;
@@ -216,7 +219,7 @@ recv_burst_vec_sse(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
#endif
/* Prefetch four descriptor pairs for next iteration. */
- if (i + RTE_BNXT_DESCS_PER_LOOP < nb_pkts) {
+ if (i + BNXT_RX_DESCS_PER_LOOP_VEC128 < nb_pkts) {
rte_prefetch0(&cp_desc_ring[cons + 8]);
rte_prefetch0(&cp_desc_ring[cons + 12]);
}
@@ -265,7 +268,7 @@ recv_burst_vec_sse(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
rxr);
nb_rx_pkts += num_valid;
- if (num_valid < RTE_BNXT_DESCS_PER_LOOP)
+ if (num_valid < BNXT_RX_DESCS_PER_LOOP_VEC128)
break;
}
@@ -383,7 +386,7 @@ bnxt_xmit_fixed_burst_vec(struct bnxt_tx_queue *txq, struct rte_mbuf **tx_pkts,
/* Handle TX burst request */
to_send = nb_pkts;
- while (to_send >= RTE_BNXT_DESCS_PER_LOOP) {
+ while (to_send >= BNXT_TX_DESCS_PER_LOOP) {
/* Prefetch next transmit buffer descriptors. */
rte_prefetch0(txbd + 4);
rte_prefetch0(txbd + 7);
@@ -393,8 +396,8 @@ bnxt_xmit_fixed_burst_vec(struct bnxt_tx_queue *txq, struct rte_mbuf **tx_pkts,
bnxt_xmit_one(tx_pkts[2], txbd++, tx_buf++);
bnxt_xmit_one(tx_pkts[3], txbd++, tx_buf++);
- to_send -= RTE_BNXT_DESCS_PER_LOOP;
- tx_pkts += RTE_BNXT_DESCS_PER_LOOP;
+ to_send -= BNXT_TX_DESCS_PER_LOOP;
+ tx_pkts += BNXT_TX_DESCS_PER_LOOP;
}
while (to_send) {
diff --git a/drivers/net/bnxt/bnxt_txr.h b/drivers/net/bnxt/bnxt_txr.h
index e4bd90f883..6bfdc6d01a 100644
--- a/drivers/net/bnxt/bnxt_txr.h
+++ b/drivers/net/bnxt/bnxt_txr.h
@@ -11,6 +11,9 @@
#define BNXT_MAX_TSO_SEGS 32
#define BNXT_MIN_PKT_SIZE 52
+/* Number of transmit descriptors processed per inner loop in vector mode. */
+#define BNXT_TX_DESCS_PER_LOOP 4U
+
struct bnxt_tx_ring_info {
uint16_t tx_raw_prod;
uint16_t tx_raw_cons;
@@ -48,6 +51,10 @@ uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
uint16_t bnxt_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
uint16_t nb_pkts);
#endif
+#if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
+uint16_t bnxt_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
+ uint16_t nb_pkts);
+#endif
int bnxt_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id);
int bnxt_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id);
diff --git a/drivers/net/bnxt/meson.build b/drivers/net/bnxt/meson.build
index 117c753489..41c4796366 100644
--- a/drivers/net/bnxt/meson.build
+++ b/drivers/net/bnxt/meson.build
@@ -82,6 +82,23 @@ sources = files(
if arch_subdir == 'x86'
sources += files('bnxt_rxtx_vec_sse.c')
+ # 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('bnxt_rxtx_vec_avx2.c')
+ elif cc.has_argument('-mavx2')
+ cflags += ['-DCC_AVX2_SUPPORT']
+ bnxt_avx2_lib = static_library('bnxt_avx2_lib',
+ 'bnxt_rxtx_vec_avx2.c',
+ dependencies: [static_rte_ethdev,
+ static_rte_bus_pci,
+ static_rte_kvargs, static_rte_hash],
+ include_directories: includes,
+ c_args: [cflags, '-mavx2'])
+ objs += bnxt_avx2_lib.extract_objects('bnxt_rxtx_vec_avx2.c')
+ endif
elif arch_subdir == 'arm' and host_machine.cpu_family().startswith('aarch64')
sources += files('bnxt_rxtx_vec_neon.c')
endif
--
2.25.1
^ permalink raw reply [flat|nested] 6+ messages in thread