From: Sun Yuechi <sunyue...@iscas.ac.cn>
The initialization of vtbl_entry is not fully vectorized here because
doing so would require __riscv_vluxei32_v_u32m1, which is slower
than the scalar approach in this small-scale scenario.
- Test: app/test/lpm_perf_autotest
- Platform: Banana Pi(BPI-F3)
- SoC: Spacemit X60 (8 cores with Vector extension)
- CPU Frequency: up to 1.6 GHz
- Cache: 256 KiB L1d ×8, 256 KiB L1i ×8, 1 MiB L2 ×2
- Memory: 16 GiB
- Kernel: Linux 6.6.36
- Compiler: GCC 14.2.0 (with RVV intrinsic support)
Test results(LPM LookupX4):
scalar: 5.7 cycles
rvv: 4.6 cycles
Signed-off-by: Sun Yuechi <sunyue...@iscas.ac.cn>
---
MAINTAINERS | 2 ++
lib/lpm/meson.build | 1 +
lib/lpm/rte_lpm.h | 2 ++
lib/lpm/rte_lpm_rvv.h | 62 +++++++++++++++++++++++++++++++++++++++++++
4 files changed, 67 insertions(+)
create mode 100644 lib/lpm/rte_lpm_rvv.h
diff --git a/MAINTAINERS b/MAINTAINERS
index 3e16789250..0f207ac129 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -340,6 +340,8 @@ M: Stanislaw Kardach <stanislaw.kard...@gmail.com>
F: config/riscv/
F: doc/guides/linux_gsg/cross_build_dpdk_for_riscv.rst
F: lib/eal/riscv/
+M: sunyuechi <sunyue...@iscas.ac.cn>
+F: lib/**/*rvv*
Intel x86
M: Bruce Richardson <bruce.richard...@intel.com>
diff --git a/lib/lpm/meson.build b/lib/lpm/meson.build
index fae4f79fb9..09133061e5 100644
--- a/lib/lpm/meson.build
+++ b/lib/lpm/meson.build
@@ -17,6 +17,7 @@ indirect_headers += files(
'rte_lpm_scalar.h',
'rte_lpm_sse.h',
'rte_lpm_sve.h',
+ 'rte_lpm_rvv.h',
)
deps += ['hash']
deps += ['rcu']
diff --git a/lib/lpm/rte_lpm.h b/lib/lpm/rte_lpm.h
index 7df64f06b1..b06517206f 100644
--- a/lib/lpm/rte_lpm.h
+++ b/lib/lpm/rte_lpm.h
@@ -408,6 +408,8 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, xmm_t ip,
uint32_t hop[4],
#include "rte_lpm_altivec.h"
#elif defined(RTE_ARCH_X86)
#include "rte_lpm_sse.h"
+#elif defined(RTE_ARCH_RISCV) && defined(RTE_RISCV_FEATURE_V)
+#include "rte_lpm_rvv.h"
#else
#include "rte_lpm_scalar.h"
#endif
diff --git a/lib/lpm/rte_lpm_rvv.h b/lib/lpm/rte_lpm_rvv.h
new file mode 100644
index 0000000000..5f48fb2b32
--- /dev/null
+++ b/lib/lpm/rte_lpm_rvv.h
@@ -0,0 +1,62 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright (c) 2025 Institute of Software Chinese Academy of Sciences
(ISCAS).
+ */
+
+#ifndef _RTE_LPM_RVV_H_
+#define _RTE_LPM_RVV_H_
+
+#include <rte_vect.h>
+
+#include <rte_cpuflags.h>
+#include <riscv_vector.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define RTE_LPM_LOOKUP_SUCCESS 0x01000000
+#define RTE_LPM_VALID_EXT_ENTRY_BITMASK 0x03000000
+
+static inline void rte_lpm_lookupx4(
+ const struct rte_lpm *lpm, xmm_t ip, uint32_t hop[4], uint32_t defv)
+{
+ size_t vl = 4;
+
+ const uint32_t *tbl24_p = (const uint32_t *)lpm->tbl24;
+ uint32_t tbl_entries[4] = {
+ tbl24_p[((uint32_t)ip[0]) >> 8],
+ tbl24_p[((uint32_t)ip[1]) >> 8],
+ tbl24_p[((uint32_t)ip[2]) >> 8],
+ tbl24_p[((uint32_t)ip[3]) >> 8],
+ };
+ vuint32m1_t vtbl_entry = __riscv_vle32_v_u32m1(tbl_entries, vl);
+
+ vbool32_t mask = __riscv_vmseq_vx_u32m1_b32(
+ __riscv_vand_vx_u32m1(vtbl_entry, RTE_LPM_VALID_EXT_ENTRY_BITMASK,
vl),
+ RTE_LPM_VALID_EXT_ENTRY_BITMASK, vl);
+
+ vuint32m1_t vtbl8_index = __riscv_vsll_vx_u32m1(
+ __riscv_vadd_vv_u32m1(
+ __riscv_vsll_vx_u32m1(__riscv_vand_vx_u32m1(vtbl_entry,
0x00FFFFFF, vl), 8, vl),
+ __riscv_vand_vx_u32m1(
+ __riscv_vle32_v_u32m1((const uint32_t *)&ip, vl),
0x000000FF, vl),
+ vl),
+ 2, vl);
+
+ vtbl_entry = __riscv_vluxei32_v_u32m1_mu(
+ mask, vtbl_entry, (const uint32_t *)(lpm->tbl8), vtbl8_index, vl);
+
+ vuint32m1_t vnext_hop = __riscv_vand_vx_u32m1(vtbl_entry, 0x00FFFFFF,
vl);
+ mask = __riscv_vmseq_vx_u32m1_b32(
+ __riscv_vand_vx_u32m1(vtbl_entry, RTE_LPM_LOOKUP_SUCCESS, vl), 0,
vl);
+
+ vnext_hop = __riscv_vmerge_vxm_u32m1(vnext_hop, defv, mask, vl);
+
+ __riscv_vse32_v_u32m1(hop, vnext_hop, vl);
+}
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _RTE_LPM_RVV_H_ */
