[dpdk-dev] [PATCH v6 1/4] hash: add new toeplitz hash implementation

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

This patch add a new Toeplitz hash implementation using
Galios Fields New Instructions (GFNI).

Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
---
 app/test/test_thash.c                       | 172 ++++++++++++++++++++++++++
 doc/api/doxy-api-index.md                   |   1 +
 doc/guides/prog_guide/toeplitz_hash_lib.rst |  28 ++++-
 doc/guides/rel_notes/release_21_11.rst      |   4 +
 lib/hash/meson.build                        |   6 +-
 lib/hash/rte_thash.c                        |  29 +++++
 lib/hash/rte_thash.h                        |  35 ++++++
 lib/hash/rte_thash_gfni.h                   |  54 +++++++++
 lib/hash/rte_thash_x86_gfni.h               | 182 ++++++++++++++++++++++++++++
 lib/hash/version.map                        |   5 +
 10 files changed, 511 insertions(+), 5 deletions(-)
 create mode 100644 lib/hash/rte_thash_gfni.h
 create mode 100644 lib/hash/rte_thash_x86_gfni.h

diff --git a/app/test/test_thash.c b/app/test/test_thash.c
index d8981fb..22d784e 100644
--- a/app/test/test_thash.c
+++ b/app/test/test_thash.c
@@ -6,6 +6,7 @@
 #include <rte_eal.h>
 #include <rte_ip.h>
 #include <rte_random.h>
+#include <rte_malloc.h>
 
 #include "test.h"
 
@@ -78,6 +79,34 @@ uint8_t default_rss_key[] = {
 0x6a, 0x42, 0xb7, 0x3b, 0xbe, 0xac, 0x01, 0xfa,
 };
 
+static const uint8_t big_rss_key[] = {
+	0x6d, 0x5a, 0x56, 0xda, 0x25, 0x5b, 0x0e, 0xc2,
+	0x41, 0x67, 0x25, 0x3d, 0x43, 0xa3, 0x8f, 0xb0,
+	0xd0, 0xca, 0x2b, 0xcb, 0xae, 0x7b, 0x30, 0xb4,
+	0x77, 0xcb, 0x2d, 0xa3, 0x80, 0x30, 0xf2, 0x0c,
+	0x6a, 0x42, 0xb7, 0x3b, 0xbe, 0xac, 0x01, 0xfa,
+	0x6d, 0x5a, 0x56, 0xda, 0x25, 0x5b, 0x0e, 0xc2,
+	0x41, 0x67, 0x25, 0x3d, 0x43, 0xa3, 0x8f, 0xb0,
+	0xd0, 0xca, 0x2b, 0xcb, 0xae, 0x7b, 0x30, 0xb4,
+	0x77, 0xcb, 0x2d, 0xa3, 0x80, 0x30, 0xf2, 0x0c,
+	0x6a, 0x42, 0xb7, 0x3b, 0xbe, 0xac, 0x01, 0xfa,
+	0x6d, 0x5a, 0x56, 0xda, 0x25, 0x5b, 0x0e, 0xc2,
+	0x41, 0x67, 0x25, 0x3d, 0x43, 0xa3, 0x8f, 0xb0,
+	0xd0, 0xca, 0x2b, 0xcb, 0xae, 0x7b, 0x30, 0xb4,
+	0x77, 0xcb, 0x2d, 0xa3, 0x80, 0x30, 0xf2, 0x0c,
+	0x6a, 0x42, 0xb7, 0x3b, 0xbe, 0xac, 0x01, 0xfa,
+	0x6d, 0x5a, 0x56, 0xda, 0x25, 0x5b, 0x0e, 0xc2,
+	0x41, 0x67, 0x25, 0x3d, 0x43, 0xa3, 0x8f, 0xb0,
+	0xd0, 0xca, 0x2b, 0xcb, 0xae, 0x7b, 0x30, 0xb4,
+	0x77, 0xcb, 0x2d, 0xa3, 0x80, 0x30, 0xf2, 0x0c,
+	0x6a, 0x42, 0xb7, 0x3b, 0xbe, 0xac, 0x01, 0xfa,
+	0x6d, 0x5a, 0x56, 0xda, 0x25, 0x5b, 0x0e, 0xc2,
+	0x41, 0x67, 0x25, 0x3d, 0x43, 0xa3, 0x8f, 0xb0,
+	0xd0, 0xca, 0x2b, 0xcb, 0xae, 0x7b, 0x30, 0xb4,
+	0x77, 0xcb, 0x2d, 0xa3, 0x80, 0x30, 0xf2, 0x0c,
+	0x6a, 0x42, 0xb7, 0x3b, 0xbe, 0xac, 0x01, 0xfa,
+};
+
 static int
 test_toeplitz_hash_calc(void)
 {
@@ -145,6 +174,146 @@ test_toeplitz_hash_calc(void)
 }
 
 static int
+test_toeplitz_hash_gfni(void)
+{
+	uint32_t i, j;
+	union rte_thash_tuple tuple;
+	uint32_t rss_l3, rss_l3l4;
+	uint64_t rss_key_matrixes[RTE_DIM(default_rss_key)];
+
+	if (!rte_thash_gfni_supported())
+		return TEST_SKIPPED;
+
+	/* Convert RSS key into matrixes */
+	rte_thash_complete_matrix(rss_key_matrixes, default_rss_key,
+		RTE_DIM(default_rss_key));
+
+	for (i = 0; i < RTE_DIM(v4_tbl); i++) {
+		tuple.v4.src_addr = rte_cpu_to_be_32(v4_tbl[i].src_ip);
+		tuple.v4.dst_addr = rte_cpu_to_be_32(v4_tbl[i].dst_ip);
+		tuple.v4.sport = rte_cpu_to_be_16(v4_tbl[i].dst_port);
+		tuple.v4.dport = rte_cpu_to_be_16(v4_tbl[i].src_port);
+
+		rss_l3 = rte_thash_gfni(rss_key_matrixes, (uint8_t *)&tuple,
+				RTE_THASH_V4_L3_LEN * 4);
+		rss_l3l4 = rte_thash_gfni(rss_key_matrixes, (uint8_t *)&tuple,
+				RTE_THASH_V4_L4_LEN * 4);
+		if ((rss_l3 != v4_tbl[i].hash_l3) ||
+				(rss_l3l4 != v4_tbl[i].hash_l3l4))
+			return -TEST_FAILED;
+	}
+
+	for (i = 0; i < RTE_DIM(v6_tbl); i++) {
+		for (j = 0; j < RTE_DIM(tuple.v6.src_addr); j++)
+			tuple.v6.src_addr[j] = v6_tbl[i].src_ip[j];
+		for (j = 0; j < RTE_DIM(tuple.v6.dst_addr); j++)
+			tuple.v6.dst_addr[j] = v6_tbl[i].dst_ip[j];
+		tuple.v6.sport = rte_cpu_to_be_16(v6_tbl[i].dst_port);
+		tuple.v6.dport = rte_cpu_to_be_16(v6_tbl[i].src_port);
+		rss_l3 = rte_thash_gfni(rss_key_matrixes, (uint8_t *)&tuple,
+				RTE_THASH_V6_L3_LEN * 4);
+		rss_l3l4 = rte_thash_gfni(rss_key_matrixes, (uint8_t *)&tuple,
+				RTE_THASH_V6_L4_LEN * 4);
+		if ((rss_l3 != v6_tbl[i].hash_l3) ||
+				(rss_l3l4 != v6_tbl[i].hash_l3l4))
+			return -TEST_FAILED;
+	}
+
+	return TEST_SUCCESS;
+}
+
+#define DATA_SZ		4
+#define ITER		1000
+
+enum {
+	SCALAR_DATA_BUF_1_HASH_IDX = 0,
+	SCALAR_DATA_BUF_2_HASH_IDX,
+	GFNI_DATA_BUF_1_HASH_IDX,
+	GFNI_DATA_BUF_2_HASH_IDX,
+	HASH_IDXES
+};
+
+static int
+test_toeplitz_hash_rand_data(void)
+{
+	uint32_t data[2][DATA_SZ];
+	uint32_t scalar_data[2][DATA_SZ];
+	uint32_t hash[HASH_IDXES] = { 0 };
+	uint64_t rss_key_matrixes[RTE_DIM(default_rss_key)];
+	int i, j;
+
+	if (!rte_thash_gfni_supported())
+		return TEST_SKIPPED;
+
+	rte_thash_complete_matrix(rss_key_matrixes, default_rss_key,
+		RTE_DIM(default_rss_key));
+
+	for (i = 0; i < ITER; i++) {
+		for (j = 0; j < DATA_SZ; j++) {
+			data[0][j] = rte_rand();
+			data[1][j] = rte_rand();
+			scalar_data[0][j] = rte_cpu_to_be_32(data[0][j]);
+			scalar_data[1][j] = rte_cpu_to_be_32(data[1][j]);
+		}
+
+		hash[SCALAR_DATA_BUF_1_HASH_IDX] = rte_softrss(scalar_data[0],
+			DATA_SZ, default_rss_key);
+		hash[SCALAR_DATA_BUF_2_HASH_IDX] = rte_softrss(scalar_data[1],
+			DATA_SZ, default_rss_key);
+		hash[GFNI_DATA_BUF_1_HASH_IDX] = rte_thash_gfni(
+			rss_key_matrixes, (uint8_t *)data[0],
+			DATA_SZ * sizeof(uint32_t));
+		hash[GFNI_DATA_BUF_2_HASH_IDX] = rte_thash_gfni(
+			rss_key_matrixes, (uint8_t *)data[1],
+			DATA_SZ * sizeof(uint32_t));
+
+		if ((hash[SCALAR_DATA_BUF_1_HASH_IDX] !=
+				hash[GFNI_DATA_BUF_1_HASH_IDX]) ||
+				(hash[SCALAR_DATA_BUF_2_HASH_IDX] !=
+				hash[GFNI_DATA_BUF_2_HASH_IDX]))
+
+			return -TEST_FAILED;
+	}
+
+	return TEST_SUCCESS;
+}
+
+enum {
+	RSS_V4_IDX,
+	RSS_V6_IDX
+};
+
+static int
+test_big_tuple_gfni(void)
+{
+	uint32_t arr[16];
+	uint32_t arr_softrss[16];
+	uint32_t hash_1, hash_2;
+	uint64_t rss_key_matrixes[RTE_DIM(big_rss_key)];
+	unsigned int i, size = RTE_DIM(arr) * sizeof(uint32_t);
+
+	if (!rte_thash_gfni_supported())
+		return TEST_SKIPPED;
+
+	/* Convert RSS key into matrixes */
+	rte_thash_complete_matrix(rss_key_matrixes, big_rss_key,
+		RTE_DIM(big_rss_key));
+
+	for (i = 0; i < RTE_DIM(arr); i++) {
+		arr[i] = rte_rand();
+		arr_softrss[i] = rte_be_to_cpu_32(arr[i]);
+	}
+
+	hash_1 = rte_softrss(arr_softrss, RTE_DIM(arr), big_rss_key);
+	hash_2 = rte_thash_gfni(rss_key_matrixes, (uint8_t *)arr, size);
+
+	if (hash_1 != hash_2)
+		return -TEST_FAILED;
+
+	return TEST_SUCCESS;
+}
+
+static int
 test_create_invalid(void)
 {
 	struct rte_thash_ctx *ctx;
@@ -577,6 +746,9 @@ static struct unit_test_suite thash_tests = {
 	.teardown = NULL,
 	.unit_test_cases = {
 	TEST_CASE(test_toeplitz_hash_calc),
+	TEST_CASE(test_toeplitz_hash_gfni),
+	TEST_CASE(test_toeplitz_hash_rand_data),
+	TEST_CASE(test_big_tuple_gfni),
 	TEST_CASE(test_create_invalid),
 	TEST_CASE(test_multiple_create),
 	TEST_CASE(test_free_null),
diff --git a/doc/api/doxy-api-index.md b/doc/api/doxy-api-index.md
index 49892a3..4245b96 100644
--- a/doc/api/doxy-api-index.md
+++ b/doc/api/doxy-api-index.md
@@ -142,6 +142,7 @@ The public API headers are grouped by topics:
   [hash]               (@ref rte_hash.h),
   [jhash]              (@ref rte_jhash.h),
   [thash]              (@ref rte_thash.h),
+  [thash_gfni]         (@ref rte_thash_gfni.h),
   [FBK hash]           (@ref rte_fbk_hash.h),
   [CRC hash]           (@ref rte_hash_crc.h)
 
diff --git a/doc/guides/prog_guide/toeplitz_hash_lib.rst b/doc/guides/prog_guide/toeplitz_hash_lib.rst
index f916857..acdd8c3 100644
--- a/doc/guides/prog_guide/toeplitz_hash_lib.rst
+++ b/doc/guides/prog_guide/toeplitz_hash_lib.rst
@@ -19,24 +19,44 @@ to calculate the RSS hash sum to spread the traffic among the queues.
 Toeplitz hash function API
 --------------------------
 
-There are two functions that provide calculation of the Toeplitz hash sum:
+There are three functions that provide calculation of the Toeplitz hash sum:
 
 * ``rte_softrss()``
 * ``rte_softrss_be()``
+* ``rte_thash_gfni()``
 
-Both of these functions take the parameters:
+First two functions are scalar implementation and take the parameters:
 
 * A pointer to the tuple, containing fields extracted from the packet.
 * A length of this tuple counted in double words.
 * A pointer to the RSS hash key corresponding to the one installed on the NIC.
 
-Both functions expect the tuple to be in "host" byte order
-and a multiple of 4 bytes in length.
+Both of above mentioned _softrss_ functions expect the tuple to be in
+"host" byte order and a multiple of 4 bytes in length.
 The ``rte_softrss()`` function expects the ``rss_key``
 to be exactly the same as the one installed on the NIC.
 The ``rte_softrss_be`` function is a faster implementation,
 but it expects ``rss_key`` to be converted to the host byte order.
 
+The last function is vectorized implementation using
+Galois Fields New Instructions. Could be used if ``rte_thash_gfni_supported`` returns true.
+It expects the tuple to be in network byte order.
+
+``rte_thash_gfni()`` calculates the hash value for a single tuple
+
+``rte_thash_gfni()`` takes the parameters:
+
+* A pointer to the matrices derived from the RSS hash key using ``rte_thash_complete_matrix()``.
+* A pointer to the tuple.
+* A length of the tuple in bytes.
+
+``rte_thash_complete_matrix()`` is a function that calculates matrices required by
+GFNI implementations from the RSS hash key. It takes the parameters:
+
+* A pointer to the memory where the matrices will be written.
+* A pointer to the RSS hash key.
+* Length of the RSS hash key in bytes.
+
 
 Predictable RSS
 ---------------
diff --git a/doc/guides/rel_notes/release_21_11.rst b/doc/guides/rel_notes/release_21_11.rst
index 1ccac87..4daeb4a 100644
--- a/doc/guides/rel_notes/release_21_11.rst
+++ b/doc/guides/rel_notes/release_21_11.rst
@@ -305,6 +305,10 @@ New Features
     * Pcapng format with timestamps and meta-data.
     * Fixes packet capture with stripped VLAN tags.
 
+* **Added optimized Toeplitz hash implementation.**
+
+  Added optimized Toeplitz hash implementation using Galois Fields New Instructions.
+
 
 Removed Items
 -------------
diff --git a/lib/hash/meson.build b/lib/hash/meson.build
index 9bc5ef9..12b1afc 100644
--- a/lib/hash/meson.build
+++ b/lib/hash/meson.build
@@ -7,8 +7,12 @@ headers = files(
         'rte_hash.h',
         'rte_jhash.h',
         'rte_thash.h',
+        'rte_thash_gfni.h',
+)
+indirect_headers += files(
+        'rte_crc_arm64.h',
+        'rte_thash_x86_gfni.h',
 )
-indirect_headers += files('rte_crc_arm64.h')
 
 sources = files('rte_cuckoo_hash.c', 'rte_fbk_hash.c', 'rte_thash.c')
 deps += ['net']
diff --git a/lib/hash/rte_thash.c b/lib/hash/rte_thash.c
index 696a112..e605a6f 100644
--- a/lib/hash/rte_thash.c
+++ b/lib/hash/rte_thash.c
@@ -90,6 +90,35 @@ struct rte_thash_ctx {
 	uint8_t		hash_key[0];
 };
 
+int
+rte_thash_gfni_supported(void)
+{
+#ifdef RTE_THASH_GFNI_DEFINED
+	if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_GFNI) &&
+			(rte_vect_get_max_simd_bitwidth() >=
+			RTE_VECT_SIMD_512))
+		return 1;
+#endif
+
+	return 0;
+};
+
+void
+rte_thash_complete_matrix(uint64_t *matrixes, const uint8_t *rss_key, int size)
+{
+	int i, j;
+	uint8_t *m = (uint8_t *)matrixes;
+	uint8_t left_part, right_part;
+
+	for (i = 0; i < size; i++) {
+		for (j = 0; j < 8; j++) {
+			left_part = rss_key[i] << j;
+			right_part = (uint16_t)(rss_key[i + 1]) >> (8 - j);
+			m[i * 8 + j] = left_part|right_part;
+		}
+	}
+}
+
 static inline uint32_t
 get_bit_lfsr(struct thash_lfsr *lfsr)
 {
diff --git a/lib/hash/rte_thash.h b/lib/hash/rte_thash.h
index a26fe56..40146cf 100644
--- a/lib/hash/rte_thash.h
+++ b/lib/hash/rte_thash.h
@@ -24,6 +24,7 @@ extern "C" {
 #include <rte_config.h>
 #include <rte_ip.h>
 #include <rte_common.h>
+#include <rte_thash_gfni.h>
 
 #if defined(RTE_ARCH_X86) || defined(__ARM_NEON)
 #include <rte_vect.h>
@@ -219,6 +220,40 @@ rte_softrss_be(uint32_t *input_tuple, uint32_t input_len,
 	return ret;
 }
 
+/**
+ * Indicates if GFNI implementations of the Toeplitz hash are supported.
+ *
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice.
+ *
+ * @return
+ *  1 if GFNI is supported
+ *  0 otherwise
+ */
+__rte_experimental
+int
+rte_thash_gfni_supported(void);
+
+/**
+ * Converts Toeplitz hash key (RSS key) into matrixes required
+ * for GFNI implementation
+ *
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice.
+ *
+ * @param matrixes
+ *  pointer to the memory where matrices will be written.
+ *  Note: the size of this memory must be equal to size * 8
+ * @param rss_key
+ *  pointer to the Toeplitz hash key
+ * @param size
+ *  Size of the rss_key in bytes.
+ */
+__rte_experimental
+void
+rte_thash_complete_matrix(uint64_t *matrixes, const uint8_t *rss_key,
+	int size);
+
 /** @internal Logarithm of minimum size of the RSS ReTa */
 #define	RTE_THASH_RETA_SZ_MIN	2U
 /** @internal Logarithm of maximum size of the RSS ReTa */
diff --git a/lib/hash/rte_thash_gfni.h b/lib/hash/rte_thash_gfni.h
new file mode 100644
index 0000000..bbacd41
--- /dev/null
+++ b/lib/hash/rte_thash_gfni.h
@@ -0,0 +1,54 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2021 Intel Corporation
+ */
+
+#ifndef _RTE_THASH_GFNI_H_
+#define _RTE_THASH_GFNI_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <rte_log.h>
+
+#ifdef RTE_ARCH_X86
+
+#include <rte_thash_x86_gfni.h>
+
+#endif
+
+#ifndef RTE_THASH_GFNI_DEFINED
+
+/**
+ * Calculate Toeplitz hash.
+ * Dummy implementation.
+ *
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice.
+ *
+ * @param m
+ *  Pointer to the matrices generated from the corresponding
+ *  RSS hash key using rte_thash_complete_matrix().
+ * @param tuple
+ *  Pointer to the data to be hashed. Data must be in network byte order.
+ * @param len
+ *  Length of the data to be hashed.
+ * @return
+ *  Calculated Toeplitz hash value.
+ */
+__rte_experimental
+static inline uint32_t
+rte_thash_gfni(const uint64_t *mtrx __rte_unused,
+	const uint8_t *key __rte_unused, int len __rte_unused)
+{
+	RTE_LOG(ERR, HASH, "%s is undefined under given arch\n", __func__);
+	return 0;
+}
+
+#endif /* RTE_THASH_GFNI_DEFINED */
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _RTE_THASH_GFNI_H_ */
diff --git a/lib/hash/rte_thash_x86_gfni.h b/lib/hash/rte_thash_x86_gfni.h
new file mode 100644
index 0000000..1cb7353
--- /dev/null
+++ b/lib/hash/rte_thash_x86_gfni.h
@@ -0,0 +1,182 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2021 Intel Corporation
+ */
+
+#ifndef _RTE_THASH_X86_GFNI_H_
+#define _RTE_THASH_X86_GFNI_H_
+
+/**
+ * @file
+ *
+ * Optimized Toeplitz hash functions implementation
+ * using Galois Fields New Instructions.
+ */
+
+#include <rte_vect.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifdef __GFNI__
+#define RTE_THASH_GFNI_DEFINED
+
+#define RTE_THASH_FIRST_ITER_MSK	0x0f0f0f0f0f0e0c08
+#define RTE_THASH_PERM_MSK		0x0f0f0f0f0f0f0f0f
+#define RTE_THASH_FIRST_ITER_MSK_2	0xf0f0f0f0f0e0c080
+#define RTE_THASH_PERM_MSK_2		0xf0f0f0f0f0f0f0f0
+#define RTE_THASH_REWIND_MSK		0x0000000000113377
+
+__rte_internal
+static inline void
+__rte_thash_xor_reduce(__m512i xor_acc, uint32_t *val_1, uint32_t *val_2)
+{
+	__m256i tmp_256_1, tmp_256_2;
+	__m128i tmp128_1, tmp128_2;
+	uint64_t tmp_1, tmp_2;
+
+	tmp_256_1 = _mm512_castsi512_si256(xor_acc);
+	tmp_256_2 = _mm512_extracti32x8_epi32(xor_acc, 1);
+	tmp_256_1 = _mm256_xor_si256(tmp_256_1, tmp_256_2);
+
+	tmp128_1 = _mm256_castsi256_si128(tmp_256_1);
+	tmp128_2 = _mm256_extracti32x4_epi32(tmp_256_1, 1);
+	tmp128_1 = _mm_xor_si128(tmp128_1, tmp128_2);
+
+	tmp_1 = _mm_extract_epi64(tmp128_1, 0);
+	tmp_2 = _mm_extract_epi64(tmp128_1, 1);
+	tmp_1 ^= tmp_2;
+
+	*val_1 = (uint32_t)tmp_1;
+	*val_2 = (uint32_t)(tmp_1 >> 32);
+}
+
+__rte_internal
+static inline __m512i
+__rte_thash_gfni(const uint64_t *mtrx, const uint8_t *tuple,
+	const uint8_t *secondary_tuple, int len)
+{
+	__m512i permute_idx = _mm512_set_epi8(7, 6, 5, 4, 7, 6, 5, 4,
+						6, 5, 4, 3, 6, 5, 4, 3,
+						5, 4, 3, 2, 5, 4, 3, 2,
+						4, 3, 2, 1, 4, 3, 2, 1,
+						3, 2, 1, 0, 3, 2, 1, 0,
+						2, 1, 0, -1, 2, 1, 0, -1,
+						1, 0, -1, -2, 1, 0, -1, -2,
+						0, -1, -2, -3, 0, -1, -2, -3);
+
+	const __m512i rewind_idx = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
+						0, 0, 0, 0, 0, 0, 0, 0,
+						0, 0, 0, 0, 0, 0, 0, 0,
+						0, 0, 0, 0, 0, 0, 0, 0,
+						0, 0, 0, 0, 0, 0, 0, 0,
+						0, 0, 0, 59, 0, 0, 0, 59,
+						0, 0, 59, 58, 0, 0, 59, 58,
+						0, 59, 58, 57, 0, 59, 58, 57);
+	const __mmask64 rewind_mask = RTE_THASH_REWIND_MSK;
+	const __m512i shift_8 = _mm512_set1_epi8(8);
+	__m512i xor_acc = _mm512_setzero_si512();
+	__m512i perm_bytes = _mm512_setzero_si512();
+	__m512i vals, matrixes, tuple_bytes, tuple_bytes_2;
+	__mmask64 load_mask, permute_mask, permute_mask_2;
+	int chunk_len = 0, i = 0;
+	uint8_t mtrx_msk;
+	const int prepend = 3;
+
+	for (; len > 0; len -= 64, tuple += 64) {
+		if (i == 8)
+			perm_bytes = _mm512_maskz_permutexvar_epi8(rewind_mask,
+				rewind_idx, perm_bytes);
+
+		permute_mask = RTE_THASH_FIRST_ITER_MSK;
+		load_mask = (len >= 64) ? UINT64_MAX : ((1ULL << len) - 1);
+		tuple_bytes = _mm512_maskz_loadu_epi8(load_mask, tuple);
+		if (secondary_tuple) {
+			permute_mask_2 = RTE_THASH_FIRST_ITER_MSK_2;
+			tuple_bytes_2 = _mm512_maskz_loadu_epi8(load_mask,
+				secondary_tuple);
+		}
+
+		chunk_len = __builtin_popcountll(load_mask);
+		for (i = 0; i < ((chunk_len + prepend) / 8); i++, mtrx += 8) {
+			perm_bytes = _mm512_mask_permutexvar_epi8(perm_bytes,
+				permute_mask, permute_idx, tuple_bytes);
+
+			if (secondary_tuple)
+				perm_bytes =
+					_mm512_mask_permutexvar_epi8(perm_bytes,
+					permute_mask_2, permute_idx,
+					tuple_bytes_2);
+
+			matrixes = _mm512_maskz_loadu_epi64(UINT8_MAX, mtrx);
+			vals = _mm512_gf2p8affine_epi64_epi8(perm_bytes,
+				matrixes, 0);
+
+			xor_acc = _mm512_xor_si512(xor_acc, vals);
+			permute_idx = _mm512_add_epi8(permute_idx, shift_8);
+			permute_mask = RTE_THASH_PERM_MSK;
+			if (secondary_tuple)
+				permute_mask_2 = RTE_THASH_PERM_MSK_2;
+		}
+	}
+
+	int rest_len = (chunk_len + prepend) % 8;
+	if (rest_len != 0) {
+		mtrx_msk = (1 << (rest_len % 8)) - 1;
+		matrixes = _mm512_maskz_loadu_epi64(mtrx_msk, mtrx);
+		if (i == 8) {
+			perm_bytes = _mm512_maskz_permutexvar_epi8(rewind_mask,
+				rewind_idx, perm_bytes);
+		} else {
+			perm_bytes = _mm512_mask_permutexvar_epi8(perm_bytes,
+				permute_mask, permute_idx, tuple_bytes);
+
+			if (secondary_tuple)
+				perm_bytes =
+					_mm512_mask_permutexvar_epi8(
+					perm_bytes, permute_mask_2,
+					permute_idx, tuple_bytes_2);
+		}
+
+		vals = _mm512_gf2p8affine_epi64_epi8(perm_bytes, matrixes, 0);
+		xor_acc = _mm512_xor_si512(xor_acc, vals);
+	}
+
+	return xor_acc;
+}
+
+/**
+ * Calculate Toeplitz hash.
+ *
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice.
+ *
+ * @param m
+ *  Pointer to the matrices generated from the corresponding
+ *  RSS hash key using rte_thash_complete_matrix().
+ * @param tuple
+ *  Pointer to the data to be hashed. Data must be in network byte order.
+ * @param len
+ *  Length of the data to be hashed.
+ * @return
+ *  Calculated Toeplitz hash value.
+ */
+__rte_experimental
+static inline uint32_t
+rte_thash_gfni(const uint64_t *m, const uint8_t *tuple, int len)
+{
+	uint32_t val, val_zero;
+
+	__m512i xor_acc = __rte_thash_gfni(m, tuple, NULL, len);
+	__rte_thash_xor_reduce(xor_acc, &val, &val_zero);
+
+	return val;
+}
+
+#endif /* _GFNI_ */
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _RTE_THASH_X86_GFNI_H_ */
diff --git a/lib/hash/version.map b/lib/hash/version.map
index 8185470..153ab87 100644
--- a/lib/hash/version.map
+++ b/lib/hash/version.map
@@ -37,6 +37,7 @@ DPDK_22 {
 EXPERIMENTAL {
 	global:
 
+	#added in 21.05
 	rte_thash_add_helper;
 	rte_thash_adjust_tuple;
 	rte_thash_find_existing;
@@ -45,4 +46,8 @@ EXPERIMENTAL {
 	rte_thash_get_helper;
 	rte_thash_get_key;
 	rte_thash_init_ctx;
+
+	#added in 21.11
+	rte_thash_complete_matrix;
+	rte_thash_gfni_supported;
 };
-- 
2.7.4