app/testpmd: optimize MAC swap for Arm
authorRuifeng Wang <ruifeng.wang@arm.com>
Tue, 12 Mar 2019 05:35:27 +0000 (13:35 +0800)
committerFerruh Yigit <ferruh.yigit@intel.com>
Wed, 20 Mar 2019 17:15:42 +0000 (18:15 +0100)
Improved MAC swap performance for ARM platform.
The improvement was achieved by using neon intrinsics
to save CPU cycles and doing swap for four packets
at a time.
The optimization had 15% - 20% throughput boost
in testpmd MAC swap mode.

Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
Reviewed-by: Gavin Hu <gavin.hu@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
app/test-pmd/macswap.c
app/test-pmd/macswap_neon.h [new file with mode: 0644]
doc/guides/rel_notes/release_19_05.rst

index cbb41b7..71af916 100644 (file)
 #include <rte_flow.h>
 
 #include "testpmd.h"
-#ifdef RTE_ARCH_X86
+#if defined(RTE_ARCH_X86)
 #include "macswap_sse.h"
+#elif defined(RTE_MACHINE_CPUFLAG_NEON)
+#include "macswap_neon.h"
 #else
 #include "macswap.h"
 #endif
diff --git a/app/test-pmd/macswap_neon.h b/app/test-pmd/macswap_neon.h
new file mode 100644 (file)
index 0000000..bdf416a
--- /dev/null
@@ -0,0 +1,97 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2019 Arm Limited
+ *
+ * Copyright(c) 2019 Intel Corporation
+ *
+ * Derived do_macswap implementation from app/test-pmd/macswap_sse.h
+ */
+
+#ifndef _MACSWAP_NEON_H_
+#define _MACSWAP_NEON_H_
+
+#include "macswap_common.h"
+#include "rte_vect.h"
+
+static inline void
+do_macswap(struct rte_mbuf *pkts[], uint16_t nb,
+               struct rte_port *txp)
+{
+       struct ether_hdr *eth_hdr[4];
+       struct rte_mbuf *mb[4];
+       uint64_t ol_flags;
+       int i;
+       int r;
+       uint8x16_t v0, v1, v2, v3;
+       /**
+        * Index map be used to shuffle the 16 bytes.
+        * byte 0-5 will be swapped with byte 6-11.
+        * byte 12-15 will keep unchanged.
+        */
+       const uint8x16_t idx_map = {6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5,
+                               12, 13, 14, 15};
+
+       ol_flags = ol_flags_init(txp->dev_conf.txmode.offloads);
+       vlan_qinq_set(pkts, nb, ol_flags,
+                       txp->tx_vlan_id, txp->tx_vlan_id_outer);
+
+       i = 0;
+       r = nb;
+
+       while (r >= 4) {
+               if (r >= 8) {
+                       rte_prefetch0(rte_pktmbuf_mtod(pkts[i + 4], void *));
+                       rte_prefetch0(rte_pktmbuf_mtod(pkts[i + 5], void *));
+                       rte_prefetch0(rte_pktmbuf_mtod(pkts[i + 6], void *));
+                       rte_prefetch0(rte_pktmbuf_mtod(pkts[i + 7], void *));
+               }
+
+               mb[0] = pkts[i++];
+               eth_hdr[0] = rte_pktmbuf_mtod(mb[0], struct ether_hdr *);
+
+               mb[1] = pkts[i++];
+               eth_hdr[1] = rte_pktmbuf_mtod(mb[1], struct ether_hdr *);
+
+               mb[2] = pkts[i++];
+               eth_hdr[2] = rte_pktmbuf_mtod(mb[2], struct ether_hdr *);
+
+               mb[3] = pkts[i++];
+               eth_hdr[3] = rte_pktmbuf_mtod(mb[3], struct ether_hdr *);
+
+               v0 = vld1q_u8((uint8_t const *)eth_hdr[0]);
+               v1 = vld1q_u8((uint8_t const *)eth_hdr[1]);
+               v2 = vld1q_u8((uint8_t const *)eth_hdr[2]);
+               v3 = vld1q_u8((uint8_t const *)eth_hdr[3]);
+
+               v0 = vqtbl1q_u8(v0, idx_map);
+               v1 = vqtbl1q_u8(v1, idx_map);
+               v2 = vqtbl1q_u8(v2, idx_map);
+               v3 = vqtbl1q_u8(v3, idx_map);
+
+               vst1q_u8((uint8_t *)eth_hdr[0], v0);
+               vst1q_u8((uint8_t *)eth_hdr[1], v1);
+               vst1q_u8((uint8_t *)eth_hdr[2], v2);
+               vst1q_u8((uint8_t *)eth_hdr[3], v3);
+
+               mbuf_field_set(mb[0], ol_flags);
+               mbuf_field_set(mb[1], ol_flags);
+               mbuf_field_set(mb[2], ol_flags);
+               mbuf_field_set(mb[3], ol_flags);
+               r -= 4;
+       }
+
+       for ( ; i < nb; i++) {
+               if (i < nb - 1)
+                       rte_prefetch0(rte_pktmbuf_mtod(pkts[i+1], void *));
+               mb[0] = pkts[i];
+               eth_hdr[0] = rte_pktmbuf_mtod(mb[0], struct ether_hdr *);
+
+               /* Swap dest and src mac addresses. */
+               v0 = vld1q_u8((uint8_t const *)eth_hdr[0]);
+               v0 = vqtbl1q_u8(v0, idx_map);
+               vst1q_u8((uint8_t *)eth_hdr[0], v0);
+
+               mbuf_field_set(mb[0], ol_flags);
+       }
+}
+
+#endif /* _MACSWAP_NEON_H_ */
index 61a2c73..6f76de3 100644 (file)
@@ -91,6 +91,11 @@ New Features
 
   * Added promiscuous mode support.
 
+* **Updated the testpmd application.**
+
+  Improved testpmd application performance on ARM platform. For ``macswap``
+  forwarding mode, NEON intrinsics were used to do swap to save CPU cycles.
+
 
 Removed Items
 -------------