net/i40e: reduce L1 cache misses in NEON Rx

For N1 platform, packet mbuf load and descs load are hot spots to limit
the performance for "desc_to_ptype_v" and "desc_to_olflags_v" functions
in i40e rx NEON path. This is because packet mbuf and descs are evicted
from l1d-cache to l2d-cache.

To reduce l1d-cache-misses and improve the performance, change the code
order and move "desc_to_ptype_v" and "desc_to_olflags_v" functions
forward to the location, where packet mbuf and descs are just loaded.

Test Result:
dpdk:21.08-rc1
gcc-9
For n1sdp, the patch improves the performance by 1.8%.
For thunderx2, no performance changes.

Signed-off-by: Feifei Wang <feifei.wang2@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
This commit is contained in:
Feifei Wang 2021-07-23 11:10:49 +08:00 committed by Qi Zhang
parent decc3b6aa5
commit 319df9f9bf

View File

@ -301,18 +301,6 @@ _recv_raw_pkts_vec(struct i40e_rx_queue *__rte_restrict rxq,
rte_mbuf_prefetch_part2(rx_pkts[pos + 3]);
}
/* C.1 4=>2 filter staterr info only */
sterr_tmp2 = vzipq_u16(vreinterpretq_u16_u64(descs[1]),
vreinterpretq_u16_u64(descs[3]));
sterr_tmp1 = vzipq_u16(vreinterpretq_u16_u64(descs[0]),
vreinterpretq_u16_u64(descs[2]));
/* C.2 get 4 pkts staterr value */
staterr = vzipq_u16(sterr_tmp1.val[1],
sterr_tmp2.val[1]).val[0];
desc_to_olflags_v(rxq, descs, &rx_pkts[pos]);
/* pkts shift the pktlen field to be 16-bit aligned*/
uint32x4_t len3 = vshlq_u32(vreinterpretq_u32_u64(descs[3]),
len_shl);
@ -367,10 +355,22 @@ _recv_raw_pkts_vec(struct i40e_rx_queue *__rte_restrict rxq,
desc_to_ptype_v(descs, &rx_pkts[pos], ptype_tbl);
desc_to_olflags_v(rxq, descs, &rx_pkts[pos]);
if (likely(pos + RTE_I40E_DESCS_PER_LOOP < nb_pkts)) {
rte_prefetch_non_temporal(rxdp + RTE_I40E_DESCS_PER_LOOP);
}
/* C.1 4=>2 filter staterr info only */
sterr_tmp2 = vzipq_u16(vreinterpretq_u16_u64(descs[1]),
vreinterpretq_u16_u64(descs[3]));
sterr_tmp1 = vzipq_u16(vreinterpretq_u16_u64(descs[0]),
vreinterpretq_u16_u64(descs[2]));
/* C.2 get 4 pkts staterr value */
staterr = vzipq_u16(sterr_tmp1.val[1],
sterr_tmp2.val[1]).val[0];
/* C* extract and record EOP bit */
if (split_packet) {
uint8x16_t eop_shuf_mask = {