[dpdk-dev,v3,2/3] lpm: add support for NEON

Message ID 1454040645-23864-3-git-send-email-jerin.jacob@caviumnetworks.com (mailing list archive)
State Superseded, archived
Delegated to: Thomas Monjalon
Headers

Commit Message

Jerin Jacob Jan. 29, 2016, 4:10 a.m. UTC
Enabled CONFIG_RTE_LIBRTE_LPM, CONFIG_RTE_LIBRTE_TABLE,
CONFIG_RTE_LIBRTE_PIPELINE libraries for arm and arm64

TABLE, PIPELINE libraries were disabled due to LPM library dependency.

Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Signed-off-by: Jianbo Liu <jianbo.liu@linaro.org>
---
 app/test/test_xmmt_ops.h                   |  20 ++++
 config/defconfig_arm-armv7a-linuxapp-gcc   |   3 -
 config/defconfig_arm64-armv8a-linuxapp-gcc |   3 -
 lib/librte_lpm/Makefile                    |   4 +
 lib/librte_lpm/rte_lpm.h                   |   4 +
 lib/librte_lpm/rte_lpm_neon.h              | 148 +++++++++++++++++++++++++++++
 6 files changed, 176 insertions(+), 6 deletions(-)
 create mode 100644 lib/librte_lpm/rte_lpm_neon.h
  

Comments

Thomas Monjalon Feb. 11, 2016, 11:46 a.m. UTC | #1
2016-01-29 09:40, Jerin Jacob:
> --- a/app/test/test_xmmt_ops.h
> +++ b/app/test/test_xmmt_ops.h
> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64)
[...]
> +#else
[...]
> --- a/lib/librte_lpm/Makefile
> +++ b/lib/librte_lpm/Makefile
> +ifneq ($(filter y,$(CONFIG_RTE_ARCH_ARM) $(CONFIG_RTE_ARCH_ARM64)),)
> +SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_neon.h
> +else
>  SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_sse.h
> +endif
[...]
> --- a/lib/librte_lpm/rte_lpm.h
> +++ b/lib/librte_lpm/rte_lpm.h
> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64)
> +#include "rte_lpm_neon.h"
> +#else
>  #include "rte_lpm_sse.h"
> +#endif

Instead of defaulting to x86 SSE, it would be better to replace
"else" by "elif X86/SSE".
I suggest using RTE_ARCH_X86 or RTE_CPUFLAG_SSEx.
By the way, what is the minimum SSE version required?
  
Jerin Jacob Feb. 12, 2016, 6:47 a.m. UTC | #2
On Thu, Feb 11, 2016 at 12:46:33PM +0100, Thomas Monjalon wrote:
> 2016-01-29 09:40, Jerin Jacob:
> > --- a/lib/librte_lpm/Makefile
> > +++ b/lib/librte_lpm/Makefile
> > +ifneq ($(filter y,$(CONFIG_RTE_ARCH_ARM) $(CONFIG_RTE_ARCH_ARM64)),)
> > +SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_neon.h
> > +else
> >  SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_sse.h
> > +endif
> [...]
> > --- a/lib/librte_lpm/rte_lpm.h
> > +++ b/lib/librte_lpm/rte_lpm.h
> > +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64)
> > +#include "rte_lpm_neon.h"
> > +#else
> >  #include "rte_lpm_sse.h"
> > +#endif
>
> Instead of defaulting to x86 SSE, it would be better to replace
> "else" by "elif X86/SSE".
> I suggest using RTE_ARCH_X86 or RTE_CPUFLAG_SSEx.

Some architectures(tile)[1] are planning to emulate SSE instruction used
in LPM for LPM library support.So that way it makes sense to use SSE as default.

But if anyone has any objections then I can add the check else let
keep in existing way.

[1]
http://dpdk.org/ml/archives/dev/2016-January/031147.html

Jerin
> By the way, what is the minimum SSE version required?
>
  
Thomas Monjalon Feb. 12, 2016, 8:42 a.m. UTC | #3
2016-02-12 12:17, Jerin Jacob:
> On Thu, Feb 11, 2016 at 12:46:33PM +0100, Thomas Monjalon wrote:
> > 2016-01-29 09:40, Jerin Jacob:
> > > --- a/lib/librte_lpm/Makefile
> > > +++ b/lib/librte_lpm/Makefile
> > > +ifneq ($(filter y,$(CONFIG_RTE_ARCH_ARM) $(CONFIG_RTE_ARCH_ARM64)),)
> > > +SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_neon.h
> > > +else
> > >  SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_sse.h
> > > +endif
> > [...]
> > > --- a/lib/librte_lpm/rte_lpm.h
> > > +++ b/lib/librte_lpm/rte_lpm.h
> > > +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64)
> > > +#include "rte_lpm_neon.h"
> > > +#else
> > >  #include "rte_lpm_sse.h"
> > > +#endif
> >
> > Instead of defaulting to x86 SSE, it would be better to replace
> > "else" by "elif X86/SSE".
> > I suggest using RTE_ARCH_X86 or RTE_CPUFLAG_SSEx.
> 
> Some architectures(tile)[1] are planning to emulate SSE instruction used
> in LPM for LPM library support.So that way it makes sense to use SSE as default.

Not sure it is a great idea to emulate instructions of another arch.

> But if anyone has any objections then I can add the check else let
> keep in existing way.

If Tile wants to use x86 code, it's better to do it explicitly (X86 || TILE).

> [1]
> http://dpdk.org/ml/archives/dev/2016-January/031147.html
  

Patch

diff --git a/app/test/test_xmmt_ops.h b/app/test/test_xmmt_ops.h
index c055912..c18fc12 100644
--- a/app/test/test_xmmt_ops.h
+++ b/app/test/test_xmmt_ops.h
@@ -36,6 +36,24 @@ 
 
 #include <rte_vect.h>
 
+#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64)
+
+/* vect_* abstraction implementation using NEON */
+
+/* loads the xmm_t value from address p(does not need to be 16-byte aligned)*/
+#define vect_loadu_sil128(p) vld1q_s32((const int32_t *)p)
+
+/* sets the 4 signed 32-bit integer values and returns the xmm_t variable */
+static inline xmm_t  __attribute__((always_inline))
+vect_set_epi32(int i3, int i2, int i1, int i0)
+{
+	int32_t data[4] = {i0, i1, i2, i3};
+
+	return vld1q_s32(data);
+}
+
+#else
+
 /* vect_* abstraction implementation using SSE */
 
 /* loads the xmm_t value from address p(does not need to be 16-byte aligned)*/
@@ -44,4 +62,6 @@ 
 /* sets the 4 signed 32-bit integer values and returns the xmm_t variable */
 #define vect_set_epi32(i3, i2, i1, i0) _mm_set_epi32(i3, i2, i1, i0)
 
+#endif
+
 #endif /* _TEST_XMMT_OPS_H_ */
diff --git a/config/defconfig_arm-armv7a-linuxapp-gcc b/config/defconfig_arm-armv7a-linuxapp-gcc
index cbebd64..efffa1f 100644
--- a/config/defconfig_arm-armv7a-linuxapp-gcc
+++ b/config/defconfig_arm-armv7a-linuxapp-gcc
@@ -53,9 +53,6 @@  CONFIG_RTE_LIBRTE_KNI=n
 CONFIG_RTE_EAL_IGB_UIO=n
 
 # fails to compile on ARM
-CONFIG_RTE_LIBRTE_LPM=n
-CONFIG_RTE_LIBRTE_TABLE=n
-CONFIG_RTE_LIBRTE_PIPELINE=n
 CONFIG_RTE_SCHED_VECTOR=n
 
 # cannot use those on ARM
diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc
index 504f3ed..57f7941 100644
--- a/config/defconfig_arm64-armv8a-linuxapp-gcc
+++ b/config/defconfig_arm64-armv8a-linuxapp-gcc
@@ -51,7 +51,4 @@  CONFIG_RTE_LIBRTE_IVSHMEM=n
 CONFIG_RTE_LIBRTE_FM10K_PMD=n
 CONFIG_RTE_LIBRTE_I40E_PMD=n
 
-CONFIG_RTE_LIBRTE_LPM=n
-CONFIG_RTE_LIBRTE_TABLE=n
-CONFIG_RTE_LIBRTE_PIPELINE=n
 CONFIG_RTE_SCHED_VECTOR=n
diff --git a/lib/librte_lpm/Makefile b/lib/librte_lpm/Makefile
index ce3a1d1..7f93006 100644
--- a/lib/librte_lpm/Makefile
+++ b/lib/librte_lpm/Makefile
@@ -47,7 +47,11 @@  SRCS-$(CONFIG_RTE_LIBRTE_LPM) := rte_lpm.c rte_lpm6.c
 # install this header file
 SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include := rte_lpm.h rte_lpm6.h
 
+ifneq ($(filter y,$(CONFIG_RTE_ARCH_ARM) $(CONFIG_RTE_ARCH_ARM64)),)
+SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_neon.h
+else
 SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_sse.h
+endif
 
 # this lib needs eal
 DEPDIRS-$(CONFIG_RTE_LIBRTE_LPM) += lib/librte_eal
diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h
index dfe1378..0c892de 100644
--- a/lib/librte_lpm/rte_lpm.h
+++ b/lib/librte_lpm/rte_lpm.h
@@ -384,7 +384,11 @@  static inline void
 rte_lpm_lookupx4(const struct rte_lpm *lpm, xmm_t ip, uint16_t hop[4],
 		 uint16_t defv);
 
+#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64)
+#include "rte_lpm_neon.h"
+#else
 #include "rte_lpm_sse.h"
+#endif
 
 #ifdef __cplusplus
 }
diff --git a/lib/librte_lpm/rte_lpm_neon.h b/lib/librte_lpm/rte_lpm_neon.h
new file mode 100644
index 0000000..fcd2a8a
--- /dev/null
+++ b/lib/librte_lpm/rte_lpm_neon.h
@@ -0,0 +1,148 @@ 
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright(c) 2015 Cavium Networks. All rights reserved.
+ *   All rights reserved.
+ *
+ *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
+ *   All rights reserved.
+ *
+ *   Derived rte_lpm_lookupx4 implementation from lib/librte_lpm/rte_lpm_sse.h
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Cavium Networks nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef _RTE_LPM_NEON_H_
+#define _RTE_LPM_NEON_H_
+
+#include <rte_branch_prediction.h>
+#include <rte_byteorder.h>
+#include <rte_memory.h>
+#include <rte_common.h>
+#include <rte_vect.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static inline void
+rte_lpm_lookupx4(const struct rte_lpm *lpm, xmm_t ip, uint16_t hop[4],
+		 uint16_t defv)
+{
+	uint32x4_t i24;
+	rte_xmm_t i8;
+	uint16_t tbl[4];
+	uint64_t idx, pt;
+
+	const uint32_t mask = UINT8_MAX;
+	const int32x4_t mask8 = vdupq_n_s32(mask);
+
+	/*
+	 * RTE_LPM_VALID_EXT_ENTRY_BITMASK for 4 LPM entries
+	 * as one 64-bit value (0x0300030003000300).
+	 */
+	const uint64_t mask_xv =
+		((uint64_t)RTE_LPM_VALID_EXT_ENTRY_BITMASK |
+		(uint64_t)RTE_LPM_VALID_EXT_ENTRY_BITMASK << 16 |
+		(uint64_t)RTE_LPM_VALID_EXT_ENTRY_BITMASK << 32 |
+		(uint64_t)RTE_LPM_VALID_EXT_ENTRY_BITMASK << 48);
+
+	/*
+	 * RTE_LPM_LOOKUP_SUCCESS for 4 LPM entries
+	 * as one 64-bit value (0x0100010001000100).
+	 */
+	const uint64_t mask_v =
+		((uint64_t)RTE_LPM_LOOKUP_SUCCESS |
+		(uint64_t)RTE_LPM_LOOKUP_SUCCESS << 16 |
+		(uint64_t)RTE_LPM_LOOKUP_SUCCESS << 32 |
+		(uint64_t)RTE_LPM_LOOKUP_SUCCESS << 48);
+
+	/* get 4 indexes for tbl24[]. */
+	i24 = vshrq_n_u32((uint32x4_t)ip, CHAR_BIT);
+
+	/* extract values from tbl24[] */
+	idx = vgetq_lane_u64((uint64x2_t)i24, 0);
+
+	tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx];
+	tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32];
+
+	idx = vgetq_lane_u64((uint64x2_t)i24, 1);
+
+	tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx];
+	tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32];
+
+	/* get 4 indexes for tbl8[]. */
+	i8.x = vandq_s32(ip, mask8);
+
+	pt = (uint64_t)tbl[0] |
+		(uint64_t)tbl[1] << 16 |
+		(uint64_t)tbl[2] << 32 |
+		(uint64_t)tbl[3] << 48;
+
+	/* search successfully finished for all 4 IP addresses. */
+	if (likely((pt & mask_xv) == mask_v)) {
+		uintptr_t ph = (uintptr_t)hop;
+		*(uint64_t *)ph = pt & RTE_LPM_MASKX4_RES;
+		return;
+	}
+
+	if (unlikely((pt & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
+			RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
+		i8.u32[0] = i8.u32[0] +
+			(uint8_t)tbl[0] * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
+		tbl[0] = *(const uint16_t *)&lpm->tbl8[i8.u32[0]];
+	}
+	if (unlikely((pt >> 16 & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
+			RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
+		i8.u32[1] = i8.u32[1] +
+			(uint8_t)tbl[1] * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
+		tbl[1] = *(const uint16_t *)&lpm->tbl8[i8.u32[1]];
+	}
+	if (unlikely((pt >> 32 & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
+			RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
+		i8.u32[2] = i8.u32[2] +
+			(uint8_t)tbl[2] * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
+		tbl[2] = *(const uint16_t *)&lpm->tbl8[i8.u32[2]];
+	}
+	if (unlikely((pt >> 48 & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
+			RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
+		i8.u32[3] = i8.u32[3] +
+			(uint8_t)tbl[3] * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
+		tbl[3] = *(const uint16_t *)&lpm->tbl8[i8.u32[3]];
+	}
+
+	hop[0] = (tbl[0] & RTE_LPM_LOOKUP_SUCCESS) ? (uint8_t)tbl[0] : defv;
+	hop[1] = (tbl[1] & RTE_LPM_LOOKUP_SUCCESS) ? (uint8_t)tbl[1] : defv;
+	hop[2] = (tbl[2] & RTE_LPM_LOOKUP_SUCCESS) ? (uint8_t)tbl[2] : defv;
+	hop[3] = (tbl[3] & RTE_LPM_LOOKUP_SUCCESS) ? (uint8_t)tbl[3] : defv;
+}
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _RTE_LPM_NEON_H_ */