Patch Detail
get:
Show a patch.
patch:
Update a patch.
put:
Update a patch.
GET /api/patches/73546/?format=api
{ "id": 73546, "url": "https://patches.dpdk.org/api/patches/73546/?format=api", "web_url": "https://patches.dpdk.org/project/dpdk/patch/33b37ee64d3b511024f8ae423275ead2e6feb3ac.1594238610.git.vladimir.medvedkin@intel.com/", "project": { "id": 1, "url": "https://patches.dpdk.org/api/projects/1/?format=api", "name": "DPDK", "link_name": "dpdk", "list_id": "dev.dpdk.org", "list_email": "dev@dpdk.org", "web_url": "http://core.dpdk.org", "scm_url": "git://dpdk.org/dpdk", "webscm_url": "http://git.dpdk.org/dpdk", "list_archive_url": "https://inbox.dpdk.org/dev", "list_archive_url_format": "https://inbox.dpdk.org/dev/{}", "commit_url_format": "" }, "msgid": "<33b37ee64d3b511024f8ae423275ead2e6feb3ac.1594238610.git.vladimir.medvedkin@intel.com>", "list_archive_url": "https://inbox.dpdk.org/dev/33b37ee64d3b511024f8ae423275ead2e6feb3ac.1594238610.git.vladimir.medvedkin@intel.com", "date": "2020-07-08T20:16:12", "name": "[v4,7/8] fib6: introduce AVX512 lookup", "commit_ref": null, "pull_url": null, "state": "superseded", "archived": true, "hash": "8d240b4320fa5cdcb13f7605e02768a4dc70d01d", "submitter": { "id": 1216, "url": "https://patches.dpdk.org/api/people/1216/?format=api", "name": "Vladimir Medvedkin", "email": "vladimir.medvedkin@intel.com" }, "delegate": { "id": 1, "url": "https://patches.dpdk.org/api/users/1/?format=api", "username": "tmonjalo", "first_name": "Thomas", "last_name": "Monjalon", "email": "thomas@monjalon.net" }, "mbox": "https://patches.dpdk.org/project/dpdk/patch/33b37ee64d3b511024f8ae423275ead2e6feb3ac.1594238610.git.vladimir.medvedkin@intel.com/mbox/", "series": [ { "id": 10896, "url": "https://patches.dpdk.org/api/series/10896/?format=api", "web_url": "https://patches.dpdk.org/project/dpdk/list/?series=10896", "date": "2020-07-08T20:16:05", "name": "fib: implement AVX512 vector lookup", "version": 4, "mbox": "https://patches.dpdk.org/series/10896/mbox/" } ], "comments": "https://patches.dpdk.org/api/patches/73546/comments/", "check": "success", "checks": "https://patches.dpdk.org/api/patches/73546/checks/", "tags": {}, "related": [], "headers": { "Return-Path": "<dev-bounces@dpdk.org>", "X-Original-To": "patchwork@inbox.dpdk.org", "Delivered-To": "patchwork@inbox.dpdk.org", "Received": [ "from dpdk.org (dpdk.org [92.243.14.124])\n\tby inbox.dpdk.org (Postfix) with ESMTP id E7F16A0526;\n\tWed, 8 Jul 2020 22:17:54 +0200 (CEST)", "from [92.243.14.124] (localhost [127.0.0.1])\n\tby dpdk.org (Postfix) with ESMTP id 50D601DEEC;\n\tWed, 8 Jul 2020 22:17:03 +0200 (CEST)", "from mga18.intel.com (mga18.intel.com [134.134.136.126])\n by dpdk.org (Postfix) with ESMTP id 466321DCDD\n for <dev@dpdk.org>; Wed, 8 Jul 2020 22:16:58 +0200 (CEST)", "from fmsmga002.fm.intel.com ([10.253.24.26])\n by orsmga106.jf.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384;\n 08 Jul 2020 13:16:57 -0700", "from silpixa00400322.ir.intel.com ([10.237.214.86])\n by fmsmga002.fm.intel.com with ESMTP; 08 Jul 2020 13:16:56 -0700" ], "IronPort-SDR": [ "\n P3zgWuZKyx8eO5XyJVFyMrXjfF8ic/c+xmvLnQz43Pzc1FQVdjPYO1XUv+P4F+ZSIqwg2yFdsy\n ssRLTdF3vm9A==", "\n O+B2du0Q3aQ/Dsz7WMxCpWKCugEETAtJPgsuQyAdqzAuHHwIFOlgwtVG7+5i2UoRakr1wx5+Hb\n o4sLOy9AZxQg==" ], "X-IronPort-AV": [ "E=McAfee;i=\"6000,8403,9676\"; a=\"135346092\"", "E=Sophos;i=\"5.75,329,1589266800\"; d=\"scan'208\";a=\"135346092\"", "E=Sophos;i=\"5.75,329,1589266800\"; d=\"scan'208\";a=\"315988623\"" ], "X-Amp-Result": "SKIPPED(no attachment in message)", "X-Amp-File-Uploaded": "False", "X-ExtLoop1": "1", "From": "Vladimir Medvedkin <vladimir.medvedkin@intel.com>", "To": "dev@dpdk.org", "Cc": "konstantin.ananyev@intel.com,\n\tbruce.richardson@intel.com", "Date": "Wed, 8 Jul 2020 21:16:12 +0100", "Message-Id": "\n <33b37ee64d3b511024f8ae423275ead2e6feb3ac.1594238610.git.vladimir.medvedkin@intel.com>", "X-Mailer": "git-send-email 2.17.1", "In-Reply-To": [ "<cover.1594238609.git.vladimir.medvedkin@intel.com>", "<cover.1594238609.git.vladimir.medvedkin@intel.com>" ], "References": [ "<cover.1594238609.git.vladimir.medvedkin@intel.com>", "<cover.1589890262.git.vladimir.medvedkin@intel.com>\n <cover.1594238609.git.vladimir.medvedkin@intel.com>" ], "Subject": "[dpdk-dev] [PATCH v4 7/8] fib6: introduce AVX512 lookup", "X-BeenThere": "dev@dpdk.org", "X-Mailman-Version": "2.1.15", "Precedence": "list", "List-Id": "DPDK patches and discussions <dev.dpdk.org>", "List-Unsubscribe": "<https://mails.dpdk.org/options/dev>,\n <mailto:dev-request@dpdk.org?subject=unsubscribe>", "List-Archive": "<http://mails.dpdk.org/archives/dev/>", "List-Post": "<mailto:dev@dpdk.org>", "List-Help": "<mailto:dev-request@dpdk.org?subject=help>", "List-Subscribe": "<https://mails.dpdk.org/listinfo/dev>,\n <mailto:dev-request@dpdk.org?subject=subscribe>", "Errors-To": "dev-bounces@dpdk.org", "Sender": "\"dev\" <dev-bounces@dpdk.org>" }, "content": "Add new lookup implementation for FIB6 trie algorithm using\nAVX512 instruction set\n\nSigned-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>\nAcked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>\n---\n lib/librte_fib/Makefile | 10 ++\n lib/librte_fib/meson.build | 13 ++\n lib/librte_fib/rte_fib6.h | 3 +-\n lib/librte_fib/trie.c | 21 +++\n lib/librte_fib/trie_avx512.c | 269 +++++++++++++++++++++++++++++++++++\n lib/librte_fib/trie_avx512.h | 20 +++\n 6 files changed, 335 insertions(+), 1 deletion(-)\n create mode 100644 lib/librte_fib/trie_avx512.c\n create mode 100644 lib/librte_fib/trie_avx512.h", "diff": "diff --git a/lib/librte_fib/Makefile b/lib/librte_fib/Makefile\nindex 3958da106..761c7c847 100644\n--- a/lib/librte_fib/Makefile\n+++ b/lib/librte_fib/Makefile\n@@ -25,12 +25,22 @@ grep -q __AVX512F__ && echo 1)\n CC_AVX512DQ_SUPPORT=$(shell $(CC) -mavx512dq -dM -E - </dev/null 2>&1 | \\\n grep -q __AVX512DQ__ && echo 1)\n \n+CC_AVX512BW_SUPPORT=$(shell $(CC) -mavx512bw -dM -E - </dev/null 2>&1 | \\\n+grep -q __AVX512BW__ && echo 1)\n+\n ifeq ($(CC_AVX512F_SUPPORT), 1)\n \tifeq ($(CC_AVX512DQ_SUPPORT), 1)\n \t\tSRCS-$(CONFIG_RTE_LIBRTE_FIB) += dir24_8_avx512.c\n \t\tCFLAGS_dir24_8_avx512.o += -mavx512f\n \t\tCFLAGS_dir24_8_avx512.o += -mavx512dq\n \t\tCFLAGS_dir24_8.o += -DCC_DIR24_8_AVX512_SUPPORT\n+\t\tifeq ($(CC_AVX512BW_SUPPORT), 1)\n+\t\t\tSRCS-$(CONFIG_RTE_LIBRTE_FIB) += trie_avx512.c\n+\t\t\tCFLAGS_trie_avx512.o += -mavx512f\n+\t\t\tCFLAGS_trie_avx512.o += -mavx512dq\n+\t\t\tCFLAGS_trie_avx512.o += -mavx512bw\n+\t\t\tCFLAGS_trie.o += -DCC_TRIE_AVX512_SUPPORT\n+\t\tendif\n \tendif\n endif\n include $(RTE_SDK)/mk/rte.lib.mk\ndiff --git a/lib/librte_fib/meson.build b/lib/librte_fib/meson.build\nindex d96ff0288..98c8752be 100644\n--- a/lib/librte_fib/meson.build\n+++ b/lib/librte_fib/meson.build\n@@ -13,6 +13,8 @@ if arch_subdir == 'x86' and not machine_args.contains('-mno-avx512f')\n \tif dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F')\n \t\tcflags += ['-DCC_DIR24_8_AVX512_SUPPORT']\n \t\tsources += files('dir24_8_avx512.c')\n+\t\tcflags += ['-DCC_TRIE_AVX512_SUPPORT']\n+\t\tsources += files('trie_avx512.c')\n \telif cc.has_multi_arguments('-mavx512f', '-mavx512dq')\n \t\tdir24_8_avx512_tmp = static_library('dir24_8_avx512_tmp',\n \t\t\t\t'dir24_8_avx512.c',\n@@ -20,6 +22,17 @@ if arch_subdir == 'x86' and not machine_args.contains('-mno-avx512f')\n \t\t\t\tc_args: cflags + ['-mavx512f', '-mavx512dq'])\n \t\tobjs += dir24_8_avx512_tmp.extract_objects('dir24_8_avx512.c')\n \t\tcflags += '-DCC_DIR24_8_AVX512_SUPPORT'\n+\t\t# TRIE AVX512 implementation uses avx512bw intrinsics along with\n+\t\t# avx512f and avx512dq\n+\t\tif cc.has_argument('-mavx512bw')\n+\t\t\ttrie_avx512_tmp = static_library('trie_avx512_tmp',\n+\t\t\t\t'trie_avx512.c',\n+\t\t\t\tdependencies: static_rte_eal,\n+\t\t\t\tc_args: cflags + ['-mavx512f', \\\n+\t\t\t\t\t'-mavx512dq', '-mavx512bw'])\n+\t\t\tobjs += trie_avx512_tmp.extract_objects('trie_avx512.c')\n+\t\t\tcflags += '-DCC_TRIE_AVX512_SUPPORT'\n+\t\tendif\n \tendif\n endif\n \ndiff --git a/lib/librte_fib/rte_fib6.h b/lib/librte_fib/rte_fib6.h\nindex e029c7624..303be55c1 100644\n--- a/lib/librte_fib/rte_fib6.h\n+++ b/lib/librte_fib/rte_fib6.h\n@@ -60,7 +60,8 @@ enum rte_fib_trie_nh_sz {\n };\n \n enum rte_fib_trie_lookup_type {\n-\tRTE_FIB6_TRIE_SCALAR\n+\tRTE_FIB6_TRIE_SCALAR,\n+\tRTE_FIB6_TRIE_VECTOR_AVX512\n };\n \n /** FIB configuration structure */\ndiff --git a/lib/librte_fib/trie.c b/lib/librte_fib/trie.c\nindex 136e938df..d0233ad01 100644\n--- a/lib/librte_fib/trie.c\n+++ b/lib/librte_fib/trie.c\n@@ -18,6 +18,12 @@\n #include <rte_fib6.h>\n #include \"trie.h\"\n \n+#ifdef CC_TRIE_AVX512_SUPPORT\n+\n+#include \"trie_avx512.h\"\n+\n+#endif /* CC_TRIE_AVX512_SUPPORT */\n+\n #define TRIE_NAMESIZE\t\t64\n \n enum edge {\n@@ -48,6 +54,21 @@ trie_get_lookup_fn(void *p, enum rte_fib_trie_lookup_type type)\n \t\tdefault:\n \t\t\treturn NULL;\n \t\t}\n+#ifdef CC_TRIE_AVX512_SUPPORT\n+\tcase RTE_FIB6_TRIE_VECTOR_AVX512:\n+\t\tif (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) <= 0)\n+\t\t\treturn NULL;\n+\t\tswitch (nh_sz) {\n+\t\tcase RTE_FIB6_TRIE_2B:\n+\t\t\treturn rte_trie_vec_lookup_bulk_2b;\n+\t\tcase RTE_FIB6_TRIE_4B:\n+\t\t\treturn rte_trie_vec_lookup_bulk_4b;\n+\t\tcase RTE_FIB6_TRIE_8B:\n+\t\t\treturn rte_trie_vec_lookup_bulk_8b;\n+\t\tdefault:\n+\t\t\treturn NULL;\n+\t\t}\n+#endif\n \tdefault:\n \t\treturn NULL;\n \t}\ndiff --git a/lib/librte_fib/trie_avx512.c b/lib/librte_fib/trie_avx512.c\nnew file mode 100644\nindex 000000000..b1c9e4ede\n--- /dev/null\n+++ b/lib/librte_fib/trie_avx512.c\n@@ -0,0 +1,269 @@\n+/* SPDX-License-Identifier: BSD-3-Clause\n+ * Copyright(c) 2020 Intel Corporation\n+ */\n+\n+#include <rte_vect.h>\n+#include <rte_fib6.h>\n+\n+#include \"trie.h\"\n+#include \"trie_avx512.h\"\n+\n+static __rte_always_inline void\n+transpose_x16(uint8_t ips[16][RTE_FIB6_IPV6_ADDR_SIZE],\n+\t__m512i *first, __m512i *second, __m512i *third, __m512i *fourth)\n+{\n+\t__m512i tmp1, tmp2, tmp3, tmp4;\n+\t__m512i tmp5, tmp6, tmp7, tmp8;\n+\tconst __rte_x86_zmm_t perm_idxes = {\n+\t\t.u32 = { 0, 4, 8, 12, 2, 6, 10, 14,\n+\t\t\t1, 5, 9, 13, 3, 7, 11, 15\n+\t\t},\n+\t};\n+\n+\t/* load all ip addresses */\n+\ttmp1 = _mm512_loadu_si512(&ips[0][0]);\n+\ttmp2 = _mm512_loadu_si512(&ips[4][0]);\n+\ttmp3 = _mm512_loadu_si512(&ips[8][0]);\n+\ttmp4 = _mm512_loadu_si512(&ips[12][0]);\n+\n+\t/* transpose 4 byte chunks of 16 ips */\n+\ttmp5 = _mm512_unpacklo_epi32(tmp1, tmp2);\n+\ttmp7 = _mm512_unpackhi_epi32(tmp1, tmp2);\n+\ttmp6 = _mm512_unpacklo_epi32(tmp3, tmp4);\n+\ttmp8 = _mm512_unpackhi_epi32(tmp3, tmp4);\n+\n+\ttmp1 = _mm512_unpacklo_epi32(tmp5, tmp6);\n+\ttmp3 = _mm512_unpackhi_epi32(tmp5, tmp6);\n+\ttmp2 = _mm512_unpacklo_epi32(tmp7, tmp8);\n+\ttmp4 = _mm512_unpackhi_epi32(tmp7, tmp8);\n+\n+\t/* first 4-byte chunks of ips[] */\n+\t*first = _mm512_permutexvar_epi32(perm_idxes.z, tmp1);\n+\t/* second 4-byte chunks of ips[] */\n+\t*second = _mm512_permutexvar_epi32(perm_idxes.z, tmp3);\n+\t/* third 4-byte chunks of ips[] */\n+\t*third = _mm512_permutexvar_epi32(perm_idxes.z, tmp2);\n+\t/* fourth 4-byte chunks of ips[] */\n+\t*fourth = _mm512_permutexvar_epi32(perm_idxes.z, tmp4);\n+}\n+\n+static __rte_always_inline void\n+transpose_x8(uint8_t ips[8][RTE_FIB6_IPV6_ADDR_SIZE],\n+\t__m512i *first, __m512i *second)\n+{\n+\t__m512i tmp1, tmp2, tmp3, tmp4;\n+\tconst __rte_x86_zmm_t perm_idxes = {\n+\t\t.u64 = { 0, 2, 4, 6, 1, 3, 5, 7\n+\t\t},\n+\t};\n+\n+\ttmp1 = _mm512_loadu_si512(&ips[0][0]);\n+\ttmp2 = _mm512_loadu_si512(&ips[4][0]);\n+\n+\ttmp3 = _mm512_unpacklo_epi64(tmp1, tmp2);\n+\t*first = _mm512_permutexvar_epi64(perm_idxes.z, tmp3);\n+\ttmp4 = _mm512_unpackhi_epi64(tmp1, tmp2);\n+\t*second = _mm512_permutexvar_epi64(perm_idxes.z, tmp4);\n+}\n+\n+static __rte_always_inline void\n+trie_vec_lookup_x16(void *p, uint8_t ips[16][RTE_FIB6_IPV6_ADDR_SIZE],\n+\tuint64_t *next_hops, int size)\n+{\n+\tstruct rte_trie_tbl *dp = (struct rte_trie_tbl *)p;\n+\tconst __m512i zero = _mm512_set1_epi32(0);\n+\tconst __m512i lsb = _mm512_set1_epi32(1);\n+\tconst __m512i two_lsb = _mm512_set1_epi32(3);\n+\t__m512i first, second, third, fourth; /*< IPv6 four byte chunks */\n+\t__m512i idxes, res, shuf_idxes;\n+\t__m512i tmp, tmp2, bytes, byte_chunk, base_idxes;\n+\t/* used to mask gather values if size is 2 (16 bit next hops) */\n+\tconst __m512i res_msk = _mm512_set1_epi32(UINT16_MAX);\n+\tconst __rte_x86_zmm_t bswap = {\n+\t\t.u8 = { 2, 1, 0, 255, 6, 5, 4, 255,\n+\t\t\t10, 9, 8, 255, 14, 13, 12, 255,\n+\t\t\t2, 1, 0, 255, 6, 5, 4, 255,\n+\t\t\t10, 9, 8, 255, 14, 13, 12, 255,\n+\t\t\t2, 1, 0, 255, 6, 5, 4, 255,\n+\t\t\t10, 9, 8, 255, 14, 13, 12, 255,\n+\t\t\t2, 1, 0, 255, 6, 5, 4, 255,\n+\t\t\t10, 9, 8, 255, 14, 13, 12, 255\n+\t\t\t},\n+\t};\n+\tconst __mmask64 k = 0x1111111111111111;\n+\tint i = 3;\n+\t__mmask16 msk_ext, new_msk;\n+\t__mmask16 exp_msk = 0x5555;\n+\n+\ttranspose_x16(ips, &first, &second, &third, &fourth);\n+\n+\t/* get_tbl24_idx() for every 4 byte chunk */\n+\tidxes = _mm512_shuffle_epi8(first, bswap.z);\n+\n+\t/**\n+\t * lookup in tbl24\n+\t * Put it inside branch to make compiller happy with -O0\n+\t */\n+\tif (size == sizeof(uint16_t)) {\n+\t\tres = _mm512_i32gather_epi32(idxes, (const int *)dp->tbl24, 2);\n+\t\tres = _mm512_and_epi32(res, res_msk);\n+\t} else\n+\t\tres = _mm512_i32gather_epi32(idxes, (const int *)dp->tbl24, 4);\n+\n+\n+\t/* get extended entries indexes */\n+\tmsk_ext = _mm512_test_epi32_mask(res, lsb);\n+\n+\ttmp = _mm512_srli_epi32(res, 1);\n+\n+\t/* idxes to retrieve bytes */\n+\tshuf_idxes = _mm512_setr_epi32(3, 7, 11, 15,\n+\t\t\t\t19, 23, 27, 31,\n+\t\t\t\t35, 39, 43, 47,\n+\t\t\t\t51, 55, 59, 63);\n+\n+\tbase_idxes = _mm512_setr_epi32(0, 4, 8, 12,\n+\t\t\t\t16, 20, 24, 28,\n+\t\t\t\t32, 36, 40, 44,\n+\t\t\t\t48, 52, 56, 60);\n+\n+\t/* traverse down the trie */\n+\twhile (msk_ext) {\n+\t\tidxes = _mm512_maskz_slli_epi32(msk_ext, tmp, 8);\n+\t\tbyte_chunk = (i < 8) ?\n+\t\t\t((i >= 4) ? second : first) :\n+\t\t\t((i >= 12) ? fourth : third);\n+\t\tbytes = _mm512_maskz_shuffle_epi8(k, byte_chunk, shuf_idxes);\n+\t\tidxes = _mm512_maskz_add_epi32(msk_ext, idxes, bytes);\n+\t\tif (size == sizeof(uint16_t)) {\n+\t\t\ttmp = _mm512_mask_i32gather_epi32(zero, msk_ext,\n+\t\t\t\tidxes, (const int *)dp->tbl8, 2);\n+\t\t\ttmp = _mm512_and_epi32(tmp, res_msk);\n+\t\t} else\n+\t\t\ttmp = _mm512_mask_i32gather_epi32(zero, msk_ext,\n+\t\t\t\tidxes, (const int *)dp->tbl8, 4);\n+\t\tnew_msk = _mm512_test_epi32_mask(tmp, lsb);\n+\t\tres = _mm512_mask_blend_epi32(msk_ext ^ new_msk, res, tmp);\n+\t\ttmp = _mm512_srli_epi32(tmp, 1);\n+\t\tmsk_ext = new_msk;\n+\n+\t\tshuf_idxes = _mm512_maskz_add_epi8(k, shuf_idxes, lsb);\n+\t\tshuf_idxes = _mm512_and_epi32(shuf_idxes, two_lsb);\n+\t\tshuf_idxes = _mm512_maskz_add_epi8(k, shuf_idxes, base_idxes);\n+\t\ti++;\n+\t}\n+\n+\tres = _mm512_srli_epi32(res, 1);\n+\ttmp = _mm512_maskz_expand_epi32(exp_msk, res);\n+\t__m256i tmp256;\n+\ttmp256 = _mm512_extracti32x8_epi32(res, 1);\n+\ttmp2 = _mm512_maskz_expand_epi32(exp_msk,\n+\t\t_mm512_castsi256_si512(tmp256));\n+\t_mm512_storeu_si512(next_hops, tmp);\n+\t_mm512_storeu_si512(next_hops + 8, tmp2);\n+}\n+\n+static void\n+trie_vec_lookup_x8_8b(void *p, uint8_t ips[8][RTE_FIB6_IPV6_ADDR_SIZE],\n+\tuint64_t *next_hops)\n+{\n+\tstruct rte_trie_tbl *dp = (struct rte_trie_tbl *)p;\n+\tconst __m512i zero = _mm512_set1_epi32(0);\n+\tconst __m512i lsb = _mm512_set1_epi32(1);\n+\tconst __m512i three_lsb = _mm512_set1_epi32(7);\n+\t__m512i first, second; /*< IPv6 eight byte chunks */\n+\t__m512i idxes, res, shuf_idxes;\n+\t__m512i tmp, bytes, byte_chunk, base_idxes;\n+\tconst __rte_x86_zmm_t bswap = {\n+\t\t.u8 = { 2, 1, 0, 255, 255, 255, 255, 255,\n+\t\t\t10, 9, 8, 255, 255, 255, 255, 255,\n+\t\t\t2, 1, 0, 255, 255, 255, 255, 255,\n+\t\t\t10, 9, 8, 255, 255, 255, 255, 255,\n+\t\t\t2, 1, 0, 255, 255, 255, 255, 255,\n+\t\t\t10, 9, 8, 255, 255, 255, 255, 255,\n+\t\t\t2, 1, 0, 255, 255, 255, 255, 255,\n+\t\t\t10, 9, 8, 255, 255, 255, 255, 255\n+\t\t\t},\n+\t};\n+\tconst __mmask64 k = 0x101010101010101;\n+\tint i = 3;\n+\t__mmask8 msk_ext, new_msk;\n+\n+\ttranspose_x8(ips, &first, &second);\n+\n+\t/* get_tbl24_idx() for every 4 byte chunk */\n+\tidxes = _mm512_shuffle_epi8(first, bswap.z);\n+\n+\t/* lookup in tbl24 */\n+\tres = _mm512_i64gather_epi64(idxes, (const void *)dp->tbl24, 8);\n+\t/* get extended entries indexes */\n+\tmsk_ext = _mm512_test_epi64_mask(res, lsb);\n+\n+\ttmp = _mm512_srli_epi64(res, 1);\n+\n+\t/* idxes to retrieve bytes */\n+\tshuf_idxes = _mm512_setr_epi64(3, 11, 19, 27, 35, 43, 51, 59);\n+\n+\tbase_idxes = _mm512_setr_epi64(0, 8, 16, 24, 32, 40, 48, 56);\n+\n+\t/* traverse down the trie */\n+\twhile (msk_ext) {\n+\t\tidxes = _mm512_maskz_slli_epi64(msk_ext, tmp, 8);\n+\t\tbyte_chunk = (i < 8) ? first : second;\n+\t\tbytes = _mm512_maskz_shuffle_epi8(k, byte_chunk, shuf_idxes);\n+\t\tidxes = _mm512_maskz_add_epi64(msk_ext, idxes, bytes);\n+\t\ttmp = _mm512_mask_i64gather_epi64(zero, msk_ext,\n+\t\t\t\tidxes, (const void *)dp->tbl8, 8);\n+\t\tnew_msk = _mm512_test_epi64_mask(tmp, lsb);\n+\t\tres = _mm512_mask_blend_epi64(msk_ext ^ new_msk, res, tmp);\n+\t\ttmp = _mm512_srli_epi64(tmp, 1);\n+\t\tmsk_ext = new_msk;\n+\n+\t\tshuf_idxes = _mm512_maskz_add_epi8(k, shuf_idxes, lsb);\n+\t\tshuf_idxes = _mm512_and_epi64(shuf_idxes, three_lsb);\n+\t\tshuf_idxes = _mm512_maskz_add_epi8(k, shuf_idxes, base_idxes);\n+\t\ti++;\n+\t}\n+\n+\tres = _mm512_srli_epi64(res, 1);\n+\t_mm512_storeu_si512(next_hops, res);\n+}\n+\n+void\n+rte_trie_vec_lookup_bulk_2b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],\n+\tuint64_t *next_hops, const unsigned int n)\n+{\n+\tuint32_t i;\n+\tfor (i = 0; i < (n / 16); i++) {\n+\t\ttrie_vec_lookup_x16(p, (uint8_t (*)[16])&ips[i * 16][0],\n+\t\t\t\tnext_hops + i * 16, sizeof(uint16_t));\n+\t}\n+\trte_trie_lookup_bulk_2b(p, (uint8_t (*)[16])&ips[i * 16][0],\n+\t\t\tnext_hops + i * 16, n - i * 16);\n+}\n+\n+void\n+rte_trie_vec_lookup_bulk_4b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],\n+\tuint64_t *next_hops, const unsigned int n)\n+{\n+\tuint32_t i;\n+\tfor (i = 0; i < (n / 16); i++) {\n+\t\ttrie_vec_lookup_x16(p, (uint8_t (*)[16])&ips[i * 16][0],\n+\t\t\t\tnext_hops + i * 16, sizeof(uint32_t));\n+\t}\n+\trte_trie_lookup_bulk_4b(p, (uint8_t (*)[16])&ips[i * 16][0],\n+\t\t\tnext_hops + i * 16, n - i * 16);\n+}\n+\n+void\n+rte_trie_vec_lookup_bulk_8b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],\n+\tuint64_t *next_hops, const unsigned int n)\n+{\n+\tuint32_t i;\n+\tfor (i = 0; i < (n / 8); i++) {\n+\t\ttrie_vec_lookup_x8_8b(p, (uint8_t (*)[16])&ips[i * 8][0],\n+\t\t\t\tnext_hops + i * 8);\n+\t}\n+\trte_trie_lookup_bulk_8b(p, (uint8_t (*)[16])&ips[i * 8][0],\n+\t\t\tnext_hops + i * 8, n - i * 8);\n+}\ndiff --git a/lib/librte_fib/trie_avx512.h b/lib/librte_fib/trie_avx512.h\nnew file mode 100644\nindex 000000000..ef8c7f0e3\n--- /dev/null\n+++ b/lib/librte_fib/trie_avx512.h\n@@ -0,0 +1,20 @@\n+/* SPDX-License-Identifier: BSD-3-Clause\n+ * Copyright(c) 2020 Intel Corporation\n+ */\n+\n+#ifndef _TRIE_AVX512_H_\n+#define _TRIE_AVX512_H_\n+\n+void\n+rte_trie_vec_lookup_bulk_2b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],\n+\tuint64_t *next_hops, const unsigned int n);\n+\n+void\n+rte_trie_vec_lookup_bulk_4b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],\n+\tuint64_t *next_hops, const unsigned int n);\n+\n+void\n+rte_trie_vec_lookup_bulk_8b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],\n+\tuint64_t *next_hops, const unsigned int n);\n+\n+#endif /* _TRIE_AVX512_H_ */\n", "prefixes": [ "v4", "7/8" ] }