Interleaving SSE SIMD load, shuffle, and store, helps to
improve the overall mac-swapp Mpps for both RX and TX.
Test Result:
* Platform: AMD EPYC 9554 @3.1GHz, no boost
* Test scenarios: TEST-PMD 64B IO vs MAC-SWAP
* NIC: broadcom P2100: loopback 2*100Gbps
<mode : Mpps Ingress: Mpps Egress>
------------------------------------------------
- MAC-SWAP original: 45.75 : 43.8
- MAC-SWAP register mod: 45.73 : 44.83
- MAC-SWAP register+ofl mod: 46.36 : 44.79
- MAC-SWAP register+ofl+interleave mod: 46.0 : 45.1
Signed-off-by: Vipin Varghese <vipin.vargh...@amd.com>
---
app/test-pmd/macswap_sse.h | 12 +++++++-----
1 file changed, 7 insertions(+), 5 deletions(-)
diff --git a/app/test-pmd/macswap_sse.h b/app/test-pmd/macswap_sse.h
index 67ff7fdfbb..1f547388b7 100644
--- a/app/test-pmd/macswap_sse.h
+++ b/app/test-pmd/macswap_sse.h
@@ -52,23 +52,25 @@ do_macswap(struct rte_mbuf *pkts[], uint16_t nb,
addr1 = _mm_loadu_si128((__m128i *)eth_hdr[1]);
mbuf_field_set(mb[1], ol_flags);
+ addr0 = _mm_shuffle_epi8(addr0, shfl_msk);
+
mb[2] = pkts[i++];
eth_hdr[2] = rte_pktmbuf_mtod(mb[2], struct rte_ether_hdr *);
addr2 = _mm_loadu_si128((__m128i *)eth_hdr[2]);
mbuf_field_set(mb[2], ol_flags);
+ addr1 = _mm_shuffle_epi8(addr1, shfl_msk);
+ _mm_storeu_si128((__m128i *)eth_hdr[0], addr0);
+
mb[3] = pkts[i++];
eth_hdr[3] = rte_pktmbuf_mtod(mb[3], struct rte_ether_hdr *);
addr3 = _mm_loadu_si128((__m128i *)eth_hdr[3]);
mbuf_field_set(mb[3], ol_flags);
- addr0 = _mm_shuffle_epi8(addr0, shfl_msk);
- addr1 = _mm_shuffle_epi8(addr1, shfl_msk);
addr2 = _mm_shuffle_epi8(addr2, shfl_msk);
- addr3 = _mm_shuffle_epi8(addr3, shfl_msk);
-
- _mm_storeu_si128((__m128i *)eth_hdr[0], addr0);
_mm_storeu_si128((__m128i *)eth_hdr[1], addr1);
+
+ addr3 = _mm_shuffle_epi8(addr3, shfl_msk);
_mm_storeu_si128((__m128i *)eth_hdr[2], addr2);
_mm_storeu_si128((__m128i *)eth_hdr[3], addr3);
--
2.34.1
