net/bnxt: minor performance fixes for AVX2 Rx
cover net/bnxt: minor performance fixes for AVX2 Rx
This series contains two minor performance fixes for the bnxt AVX2 vecgtorized burst receive function.
Lance Richardson (2):
net/bnxt: avoid unnecessary work in AVX2 Rx path net/bnxt: remove software prefetches from AVX2 Rx path
drivers/net/bnxt/bnxt_rxtx_vec_avx2.c | 16 +++------------- 1 file changed, 3 insertions(+), 13 deletions(-)
Signature · 1 line
2.25.1
PATCH 1/2 net/bnxt: avoid unnecessary work in AVX2 Rx path
Each call to the AVX2 vector burst receive function makes at least one pass through the function's inner loop, loading 256 bytes of completion descriptors and copying 8 rte_mbuf pointers regardless of whether there are any packets to be received.
Unidirectional forwarding performance is improved by about 3-4% if we ensure that at least one packet can be received before entering the inner loop.
--- drivers/net/bnxt/bnxt_rxtx_vec_avx2.c | 4 ++++ 1 file changed, 4 insertions(+)
drivers/net/bnxt/bnxt_rxtx_vec_avx2.c +4 -0
@@ -98,6 +98,10 @@ recv_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)rte_prefetch0 (&cp_desc_ring[cons +8 ]);rte_prefetch0 (&cp_desc_ring[cons +12 ]); +/* Return immediately if there is not at least one completed packet. */ +if (!bnxt_cpr_cmp_valid (&cp_desc_ring[cons], raw_cons, cp_ring_size)) +return 0 ; +/* 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 ));
Signature · 1 line
2.25.1
PATCH 2/2 net/bnxt: remove software prefetches from AVX2 Rx path
In reply to Lance RichardsonTesting has shown no performance benefit from software prefetching of receive completion descriptors in the AVX2 burst receive path, and slightly better performance without them on some CPU families, so this patch removes them.
--- drivers/net/bnxt/bnxt_rxtx_vec_avx2.c | 14 -------------- 1 file changed, 14 deletions(-)
drivers/net/bnxt/bnxt_rxtx_vec_avx2.c +0 -14
@@ -92,12 +92,6 @@ recv_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts) 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 ]); -/* Return immediately if there is not at least one completed packet. */ if (!bnxt_cpr_cmp_valid (&cp_desc_ring[cons], raw_cons, cp_ring_size))return 0 ;
@@ -136,14 +130,6 @@ recv_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)_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
Signature · 1 line
2.25.1
Quoted 7 lines · Lance Richardson
This series contains two minor performance fixes for the bnxt AVX2 vecgtorized burst receive function.
Lance Richardson (2):
net/bnxt: avoid unnecessary work in AVX2 Rx path net/bnxt: remove software prefetches from AVX2 Rx path
Patchset applied to dpdk-next-net-brcm. Thanks
Quoted 7 lines
drivers/net/bnxt/bnxt_rxtx_vec_avx2.c | 16 +++------------- 1 file changed, 3 insertions(+), 13 deletions(-)--
2.25.1