[dpdk-dev] [PATCH v4 7/8] fib6: introduce AVX512 lookup

[Vladimir Medvedkin <vladimir.medvedkin@intel.com>]

Add new lookup implementation for FIB6 trie algorithm using
AVX512 instruction set

Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
---
 lib/librte_fib/Makefile      |  10 ++
 lib/librte_fib/meson.build   |  13 ++
 lib/librte_fib/rte_fib6.h    |   3 +-
 lib/librte_fib/trie.c        |  21 +++
 lib/librte_fib/trie_avx512.c | 269 +++++++++++++++++++++++++++++++++++
 lib/librte_fib/trie_avx512.h |  20 +++
 6 files changed, 335 insertions(+), 1 deletion(-)
 create mode 100644 lib/librte_fib/trie_avx512.c
 create mode 100644 lib/librte_fib/trie_avx512.h

diff --git a/lib/librte_fib/Makefile b/lib/librte_fib/Makefile
index 3958da106..761c7c847 100644
--- a/lib/librte_fib/Makefile
+++ b/lib/librte_fib/Makefile
@@ -25,12 +25,22 @@ grep -q __AVX512F__ && echo 1)
 CC_AVX512DQ_SUPPORT=$(shell $(CC) -mavx512dq -dM -E - </dev/null 2>&1 | \
 grep -q __AVX512DQ__ && echo 1)
 
+CC_AVX512BW_SUPPORT=$(shell $(CC) -mavx512bw -dM -E - </dev/null 2>&1 | \
+grep -q __AVX512BW__ && echo 1)
+
 ifeq ($(CC_AVX512F_SUPPORT), 1)
 	ifeq ($(CC_AVX512DQ_SUPPORT), 1)
 		SRCS-$(CONFIG_RTE_LIBRTE_FIB) += dir24_8_avx512.c
 		CFLAGS_dir24_8_avx512.o += -mavx512f
 		CFLAGS_dir24_8_avx512.o += -mavx512dq
 		CFLAGS_dir24_8.o += -DCC_DIR24_8_AVX512_SUPPORT
+		ifeq ($(CC_AVX512BW_SUPPORT), 1)
+			SRCS-$(CONFIG_RTE_LIBRTE_FIB) += trie_avx512.c
+			CFLAGS_trie_avx512.o += -mavx512f
+			CFLAGS_trie_avx512.o += -mavx512dq
+			CFLAGS_trie_avx512.o += -mavx512bw
+			CFLAGS_trie.o += -DCC_TRIE_AVX512_SUPPORT
+		endif
 	endif
 endif
 include $(RTE_SDK)/mk/rte.lib.mk
diff --git a/lib/librte_fib/meson.build b/lib/librte_fib/meson.build
index d96ff0288..98c8752be 100644
--- a/lib/librte_fib/meson.build
+++ b/lib/librte_fib/meson.build
@@ -13,6 +13,8 @@ if arch_subdir == 'x86' and not machine_args.contains('-mno-avx512f')
 	if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F')
 		cflags += ['-DCC_DIR24_8_AVX512_SUPPORT']
 		sources += files('dir24_8_avx512.c')
+		cflags += ['-DCC_TRIE_AVX512_SUPPORT']
+		sources += files('trie_avx512.c')
 	elif cc.has_multi_arguments('-mavx512f', '-mavx512dq')
 		dir24_8_avx512_tmp = static_library('dir24_8_avx512_tmp',
 				'dir24_8_avx512.c',
@@ -20,6 +22,17 @@ if arch_subdir == 'x86' and not machine_args.contains('-mno-avx512f')
 				c_args: cflags + ['-mavx512f', '-mavx512dq'])
 		objs += dir24_8_avx512_tmp.extract_objects('dir24_8_avx512.c')
 		cflags += '-DCC_DIR24_8_AVX512_SUPPORT'
+		# TRIE AVX512 implementation uses avx512bw intrinsics along with
+		# avx512f and avx512dq
+		if cc.has_argument('-mavx512bw')
+			trie_avx512_tmp = static_library('trie_avx512_tmp',
+				'trie_avx512.c',
+				dependencies: static_rte_eal,
+				c_args: cflags + ['-mavx512f', \
+					'-mavx512dq', '-mavx512bw'])
+			objs += trie_avx512_tmp.extract_objects('trie_avx512.c')
+			cflags += '-DCC_TRIE_AVX512_SUPPORT'
+		endif
 	endif
 endif
 
diff --git a/lib/librte_fib/rte_fib6.h b/lib/librte_fib/rte_fib6.h
index e029c7624..303be55c1 100644
--- a/lib/librte_fib/rte_fib6.h
+++ b/lib/librte_fib/rte_fib6.h
@@ -60,7 +60,8 @@ enum rte_fib_trie_nh_sz {
 };
 
 enum rte_fib_trie_lookup_type {
-	RTE_FIB6_TRIE_SCALAR
+	RTE_FIB6_TRIE_SCALAR,
+	RTE_FIB6_TRIE_VECTOR_AVX512
 };
 
 /** FIB configuration structure */
diff --git a/lib/librte_fib/trie.c b/lib/librte_fib/trie.c
index 136e938df..d0233ad01 100644
--- a/lib/librte_fib/trie.c
+++ b/lib/librte_fib/trie.c
@@ -18,6 +18,12 @@
 #include <rte_fib6.h>
 #include "trie.h"
 
+#ifdef CC_TRIE_AVX512_SUPPORT
+
+#include "trie_avx512.h"
+
+#endif /* CC_TRIE_AVX512_SUPPORT */
+
 #define TRIE_NAMESIZE		64
 
 enum edge {
@@ -48,6 +54,21 @@ trie_get_lookup_fn(void *p, enum rte_fib_trie_lookup_type type)
 		default:
 			return NULL;
 		}
+#ifdef CC_TRIE_AVX512_SUPPORT
+	case RTE_FIB6_TRIE_VECTOR_AVX512:
+		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) <= 0)
+			return NULL;
+		switch (nh_sz) {
+		case RTE_FIB6_TRIE_2B:
+			return rte_trie_vec_lookup_bulk_2b;
+		case RTE_FIB6_TRIE_4B:
+			return rte_trie_vec_lookup_bulk_4b;
+		case RTE_FIB6_TRIE_8B:
+			return rte_trie_vec_lookup_bulk_8b;
+		default:
+			return NULL;
+		}
+#endif
 	default:
 		return NULL;
 	}
diff --git a/lib/librte_fib/trie_avx512.c b/lib/librte_fib/trie_avx512.c
new file mode 100644
index 000000000..b1c9e4ede
--- /dev/null
+++ b/lib/librte_fib/trie_avx512.c
@@ -0,0 +1,269 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include <rte_vect.h>
+#include <rte_fib6.h>
+
+#include "trie.h"
+#include "trie_avx512.h"
+
+static __rte_always_inline void
+transpose_x16(uint8_t ips[16][RTE_FIB6_IPV6_ADDR_SIZE],
+	__m512i *first, __m512i *second, __m512i *third, __m512i *fourth)
+{
+	__m512i tmp1, tmp2, tmp3, tmp4;
+	__m512i tmp5, tmp6, tmp7, tmp8;
+	const __rte_x86_zmm_t perm_idxes = {
+		.u32 = { 0, 4, 8, 12, 2, 6, 10, 14,
+			1, 5, 9, 13, 3, 7, 11, 15
+		},
+	};
+
+	/* load all ip addresses */
+	tmp1 = _mm512_loadu_si512(&ips[0][0]);
+	tmp2 = _mm512_loadu_si512(&ips[4][0]);
+	tmp3 = _mm512_loadu_si512(&ips[8][0]);
+	tmp4 = _mm512_loadu_si512(&ips[12][0]);
+
+	/* transpose 4 byte chunks of 16 ips */
+	tmp5 = _mm512_unpacklo_epi32(tmp1, tmp2);
+	tmp7 = _mm512_unpackhi_epi32(tmp1, tmp2);
+	tmp6 = _mm512_unpacklo_epi32(tmp3, tmp4);
+	tmp8 = _mm512_unpackhi_epi32(tmp3, tmp4);
+
+	tmp1 = _mm512_unpacklo_epi32(tmp5, tmp6);
+	tmp3 = _mm512_unpackhi_epi32(tmp5, tmp6);
+	tmp2 = _mm512_unpacklo_epi32(tmp7, tmp8);
+	tmp4 = _mm512_unpackhi_epi32(tmp7, tmp8);
+
+	/* first 4-byte chunks of ips[] */
+	*first = _mm512_permutexvar_epi32(perm_idxes.z, tmp1);
+	/* second 4-byte chunks of ips[] */
+	*second = _mm512_permutexvar_epi32(perm_idxes.z, tmp3);
+	/* third 4-byte chunks of ips[] */
+	*third = _mm512_permutexvar_epi32(perm_idxes.z, tmp2);
+	/* fourth 4-byte chunks of ips[] */
+	*fourth = _mm512_permutexvar_epi32(perm_idxes.z, tmp4);
+}
+
+static __rte_always_inline void
+transpose_x8(uint8_t ips[8][RTE_FIB6_IPV6_ADDR_SIZE],
+	__m512i *first, __m512i *second)
+{
+	__m512i tmp1, tmp2, tmp3, tmp4;
+	const __rte_x86_zmm_t perm_idxes = {
+		.u64 = { 0, 2, 4, 6, 1, 3, 5, 7
+		},
+	};
+
+	tmp1 = _mm512_loadu_si512(&ips[0][0]);
+	tmp2 = _mm512_loadu_si512(&ips[4][0]);
+
+	tmp3 = _mm512_unpacklo_epi64(tmp1, tmp2);
+	*first = _mm512_permutexvar_epi64(perm_idxes.z, tmp3);
+	tmp4 = _mm512_unpackhi_epi64(tmp1, tmp2);
+	*second = _mm512_permutexvar_epi64(perm_idxes.z, tmp4);
+}
+
+static __rte_always_inline void
+trie_vec_lookup_x16(void *p, uint8_t ips[16][RTE_FIB6_IPV6_ADDR_SIZE],
+	uint64_t *next_hops, int size)
+{
+	struct rte_trie_tbl *dp = (struct rte_trie_tbl *)p;
+	const __m512i zero = _mm512_set1_epi32(0);
+	const __m512i lsb = _mm512_set1_epi32(1);
+	const __m512i two_lsb = _mm512_set1_epi32(3);
+	__m512i first, second, third, fourth; /*< IPv6 four byte chunks */
+	__m512i idxes, res, shuf_idxes;
+	__m512i tmp, tmp2, bytes, byte_chunk, base_idxes;
+	/* used to mask gather values if size is 2 (16 bit next hops) */
+	const __m512i res_msk = _mm512_set1_epi32(UINT16_MAX);
+	const __rte_x86_zmm_t bswap = {
+		.u8 = { 2, 1, 0, 255, 6, 5, 4, 255,
+			10, 9, 8, 255, 14, 13, 12, 255,
+			2, 1, 0, 255, 6, 5, 4, 255,
+			10, 9, 8, 255, 14, 13, 12, 255,
+			2, 1, 0, 255, 6, 5, 4, 255,
+			10, 9, 8, 255, 14, 13, 12, 255,
+			2, 1, 0, 255, 6, 5, 4, 255,
+			10, 9, 8, 255, 14, 13, 12, 255
+			},
+	};
+	const __mmask64 k = 0x1111111111111111;
+	int i = 3;
+	__mmask16 msk_ext, new_msk;
+	__mmask16 exp_msk = 0x5555;
+
+	transpose_x16(ips, &first, &second, &third, &fourth);
+
+	/* get_tbl24_idx() for every 4 byte chunk */
+	idxes = _mm512_shuffle_epi8(first, bswap.z);
+
+	/**
+	 * lookup in tbl24
+	 * Put it inside branch to make compiller happy with -O0
+	 */
+	if (size == sizeof(uint16_t)) {
+		res = _mm512_i32gather_epi32(idxes, (const int *)dp->tbl24, 2);
+		res = _mm512_and_epi32(res, res_msk);
+	} else
+		res = _mm512_i32gather_epi32(idxes, (const int *)dp->tbl24, 4);
+
+
+	/* get extended entries indexes */
+	msk_ext = _mm512_test_epi32_mask(res, lsb);
+
+	tmp = _mm512_srli_epi32(res, 1);
+
+	/* idxes to retrieve bytes */
+	shuf_idxes = _mm512_setr_epi32(3, 7, 11, 15,
+				19, 23, 27, 31,
+				35, 39, 43, 47,
+				51, 55, 59, 63);
+
+	base_idxes = _mm512_setr_epi32(0, 4, 8, 12,
+				16, 20, 24, 28,
+				32, 36, 40, 44,
+				48, 52, 56, 60);
+
+	/* traverse down the trie */
+	while (msk_ext) {
+		idxes = _mm512_maskz_slli_epi32(msk_ext, tmp, 8);
+		byte_chunk = (i < 8) ?
+			((i >= 4) ? second : first) :
+			((i >= 12) ? fourth : third);
+		bytes = _mm512_maskz_shuffle_epi8(k, byte_chunk, shuf_idxes);
+		idxes = _mm512_maskz_add_epi32(msk_ext, idxes, bytes);
+		if (size == sizeof(uint16_t)) {
+			tmp = _mm512_mask_i32gather_epi32(zero, msk_ext,
+				idxes, (const int *)dp->tbl8, 2);
+			tmp = _mm512_and_epi32(tmp, res_msk);
+		} else
+			tmp = _mm512_mask_i32gather_epi32(zero, msk_ext,
+				idxes, (const int *)dp->tbl8, 4);
+		new_msk = _mm512_test_epi32_mask(tmp, lsb);
+		res = _mm512_mask_blend_epi32(msk_ext ^ new_msk, res, tmp);
+		tmp = _mm512_srli_epi32(tmp, 1);
+		msk_ext = new_msk;
+
+		shuf_idxes = _mm512_maskz_add_epi8(k, shuf_idxes, lsb);
+		shuf_idxes = _mm512_and_epi32(shuf_idxes, two_lsb);
+		shuf_idxes = _mm512_maskz_add_epi8(k, shuf_idxes, base_idxes);
+		i++;
+	}
+
+	res = _mm512_srli_epi32(res, 1);
+	tmp = _mm512_maskz_expand_epi32(exp_msk, res);
+	__m256i tmp256;
+	tmp256 = _mm512_extracti32x8_epi32(res, 1);
+	tmp2 = _mm512_maskz_expand_epi32(exp_msk,
+		_mm512_castsi256_si512(tmp256));
+	_mm512_storeu_si512(next_hops, tmp);
+	_mm512_storeu_si512(next_hops + 8, tmp2);
+}
+
+static void
+trie_vec_lookup_x8_8b(void *p, uint8_t ips[8][RTE_FIB6_IPV6_ADDR_SIZE],
+	uint64_t *next_hops)
+{
+	struct rte_trie_tbl *dp = (struct rte_trie_tbl *)p;
+	const __m512i zero = _mm512_set1_epi32(0);
+	const __m512i lsb = _mm512_set1_epi32(1);
+	const __m512i three_lsb = _mm512_set1_epi32(7);
+	__m512i first, second; /*< IPv6 eight byte chunks */
+	__m512i idxes, res, shuf_idxes;
+	__m512i tmp, bytes, byte_chunk, base_idxes;
+	const __rte_x86_zmm_t bswap = {
+		.u8 = { 2, 1, 0, 255, 255, 255, 255, 255,
+			10, 9, 8, 255, 255, 255, 255, 255,
+			2, 1, 0, 255, 255, 255, 255, 255,
+			10, 9, 8, 255, 255, 255, 255, 255,
+			2, 1, 0, 255, 255, 255, 255, 255,
+			10, 9, 8, 255, 255, 255, 255, 255,
+			2, 1, 0, 255, 255, 255, 255, 255,
+			10, 9, 8, 255, 255, 255, 255, 255
+			},
+	};
+	const __mmask64 k = 0x101010101010101;
+	int i = 3;
+	__mmask8 msk_ext, new_msk;
+
+	transpose_x8(ips, &first, &second);
+
+	/* get_tbl24_idx() for every 4 byte chunk */
+	idxes = _mm512_shuffle_epi8(first, bswap.z);
+
+	/* lookup in tbl24 */
+	res = _mm512_i64gather_epi64(idxes, (const void *)dp->tbl24, 8);
+	/* get extended entries indexes */
+	msk_ext = _mm512_test_epi64_mask(res, lsb);
+
+	tmp = _mm512_srli_epi64(res, 1);
+
+	/* idxes to retrieve bytes */
+	shuf_idxes = _mm512_setr_epi64(3, 11, 19, 27, 35, 43, 51, 59);
+
+	base_idxes = _mm512_setr_epi64(0, 8, 16, 24, 32, 40, 48, 56);
+
+	/* traverse down the trie */
+	while (msk_ext) {
+		idxes = _mm512_maskz_slli_epi64(msk_ext, tmp, 8);
+		byte_chunk = (i < 8) ? first : second;
+		bytes = _mm512_maskz_shuffle_epi8(k, byte_chunk, shuf_idxes);
+		idxes = _mm512_maskz_add_epi64(msk_ext, idxes, bytes);
+		tmp = _mm512_mask_i64gather_epi64(zero, msk_ext,
+				idxes, (const void *)dp->tbl8, 8);
+		new_msk = _mm512_test_epi64_mask(tmp, lsb);
+		res = _mm512_mask_blend_epi64(msk_ext ^ new_msk, res, tmp);
+		tmp = _mm512_srli_epi64(tmp, 1);
+		msk_ext = new_msk;
+
+		shuf_idxes = _mm512_maskz_add_epi8(k, shuf_idxes, lsb);
+		shuf_idxes = _mm512_and_epi64(shuf_idxes, three_lsb);
+		shuf_idxes = _mm512_maskz_add_epi8(k, shuf_idxes, base_idxes);
+		i++;
+	}
+
+	res = _mm512_srli_epi64(res, 1);
+	_mm512_storeu_si512(next_hops, res);
+}
+
+void
+rte_trie_vec_lookup_bulk_2b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],
+	uint64_t *next_hops, const unsigned int n)
+{
+	uint32_t i;
+	for (i = 0; i < (n / 16); i++) {
+		trie_vec_lookup_x16(p, (uint8_t (*)[16])&ips[i * 16][0],
+				next_hops + i * 16, sizeof(uint16_t));
+	}
+	rte_trie_lookup_bulk_2b(p, (uint8_t (*)[16])&ips[i * 16][0],
+			next_hops + i * 16, n - i * 16);
+}
+
+void
+rte_trie_vec_lookup_bulk_4b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],
+	uint64_t *next_hops, const unsigned int n)
+{
+	uint32_t i;
+	for (i = 0; i < (n / 16); i++) {
+		trie_vec_lookup_x16(p, (uint8_t (*)[16])&ips[i * 16][0],
+				next_hops + i * 16, sizeof(uint32_t));
+	}
+	rte_trie_lookup_bulk_4b(p, (uint8_t (*)[16])&ips[i * 16][0],
+			next_hops + i * 16, n - i * 16);
+}
+
+void
+rte_trie_vec_lookup_bulk_8b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],
+	uint64_t *next_hops, const unsigned int n)
+{
+	uint32_t i;
+	for (i = 0; i < (n / 8); i++) {
+		trie_vec_lookup_x8_8b(p, (uint8_t (*)[16])&ips[i * 8][0],
+				next_hops + i * 8);
+	}
+	rte_trie_lookup_bulk_8b(p, (uint8_t (*)[16])&ips[i * 8][0],
+			next_hops + i * 8, n - i * 8);
+}
diff --git a/lib/librte_fib/trie_avx512.h b/lib/librte_fib/trie_avx512.h
new file mode 100644
index 000000000..ef8c7f0e3
--- /dev/null
+++ b/lib/librte_fib/trie_avx512.h
@@ -0,0 +1,20 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#ifndef _TRIE_AVX512_H_
+#define _TRIE_AVX512_H_
+
+void
+rte_trie_vec_lookup_bulk_2b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],
+	uint64_t *next_hops, const unsigned int n);
+
+void
+rte_trie_vec_lookup_bulk_4b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],
+	uint64_t *next_hops, const unsigned int n);
+
+void
+rte_trie_vec_lookup_bulk_8b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],
+	uint64_t *next_hops, const unsigned int n);
+
+#endif /* _TRIE_AVX512_H_ */
-- 
2.17.1