DPDK patches and discussions
 help / color / mirror / Atom feed
From: Konstantin Ananyev <konstantin.ananyev@intel.com>
To: dev@dpdk.org
Cc: jerinj@marvell.com, ruifeng.wang@arm.com,
	vladimir.medvedkin@intel.com,
	Konstantin Ananyev <konstantin.ananyev@intel.com>
Subject: [dpdk-dev] [PATCH 20.11 7/7] acl: enhance AVX512 classify implementation
Date: Fri,  7 Aug 2020 17:28:29 +0100	[thread overview]
Message-ID: <20200807162829.11690-8-konstantin.ananyev@intel.com> (raw)
In-Reply-To: <20200807162829.11690-1-konstantin.ananyev@intel.com>

Add search_avx512x16x2() which uses mostly 512-bit width
registers/instructions and is able to process up to 32 flows in
parallel.

Signed-off-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
---

These patch depends on:
https://patches.dpdk.org/patch/70429/
to be applied first.

 lib/librte_acl/acl_run_avx512.c    |   3 +
 lib/librte_acl/acl_run_avx512x16.h | 635 +++++++++++++++++++++++++++++
 2 files changed, 638 insertions(+)
 create mode 100644 lib/librte_acl/acl_run_avx512x16.h

diff --git a/lib/librte_acl/acl_run_avx512.c b/lib/librte_acl/acl_run_avx512.c
index 8ee996679..332e359fb 100644
--- a/lib/librte_acl/acl_run_avx512.c
+++ b/lib/librte_acl/acl_run_avx512.c
@@ -121,11 +121,14 @@ resolve_mcgt8_avx512x1(uint32_t result[],
 }
 
 #include "acl_run_avx512x8.h"
+#include "acl_run_avx512x16.h"
 
 int
 rte_acl_classify_avx512(const struct rte_acl_ctx *ctx, const uint8_t **data,
 	uint32_t *results, uint32_t num, uint32_t categories)
 {
+	if (num >= 2 * MAX_SEARCHES_AVX16)
+		return search_avx512x16x2(ctx, data, results, num, categories);
 	if (num >= MAX_SEARCHES_AVX16)
 		return search_avx512x8x2(ctx, data, results, num, categories);
 	if (num >= MAX_SEARCHES_SSE8)
diff --git a/lib/librte_acl/acl_run_avx512x16.h b/lib/librte_acl/acl_run_avx512x16.h
new file mode 100644
index 000000000..53216bda3
--- /dev/null
+++ b/lib/librte_acl/acl_run_avx512x16.h
@@ -0,0 +1,635 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#define	MASK16_BIT	(sizeof(__mmask16) * CHAR_BIT)
+
+#define NUM_AVX512X16X2	(2 * MASK16_BIT)
+#define MSK_AVX512X16X2	(NUM_AVX512X16X2 - 1)
+
+static const __rte_x86_zmm_t zmm_match_mask = {
+	.u32 = {
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+		RTE_ACL_NODE_MATCH,
+	},
+};
+
+static const __rte_x86_zmm_t zmm_index_mask = {
+	.u32 = {
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX,
+	},
+};
+
+static const __rte_x86_zmm_t zmm_trlo_idle = {
+	.u32 = {
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+		RTE_ACL_DFA_SIZE | RTE_ACL_NODE_SINGLE,
+	},
+};
+
+static const __rte_x86_zmm_t zmm_trhi_idle = {
+	.u32 = {
+		0, 0, 0, 0,
+		0, 0, 0, 0,
+		0, 0, 0, 0,
+		0, 0, 0, 0,
+	},
+};
+
+static const __rte_x86_zmm_t zmm_shuffle_input = {
+	.u32 = {
+		0x00000000, 0x04040404, 0x08080808, 0x0c0c0c0c,
+		0x00000000, 0x04040404, 0x08080808, 0x0c0c0c0c,
+		0x00000000, 0x04040404, 0x08080808, 0x0c0c0c0c,
+		0x00000000, 0x04040404, 0x08080808, 0x0c0c0c0c,
+	},
+};
+
+static const __rte_x86_zmm_t zmm_four_32 = {
+	.u32 = {
+		4, 4, 4, 4,
+		4, 4, 4, 4,
+		4, 4, 4, 4,
+		4, 4, 4, 4,
+	},
+};
+
+static const __rte_x86_zmm_t zmm_idx_add = {
+	.u32 = {
+		0, 1, 2, 3,
+		4, 5, 6, 7,
+		8, 9, 10, 11,
+		12, 13, 14, 15,
+	},
+};
+
+static const __rte_x86_zmm_t zmm_range_base = {
+	.u32 = {
+		0xffffff00, 0xffffff04, 0xffffff08, 0xffffff0c,
+		0xffffff00, 0xffffff04, 0xffffff08, 0xffffff0c,
+		0xffffff00, 0xffffff04, 0xffffff08, 0xffffff0c,
+		0xffffff00, 0xffffff04, 0xffffff08, 0xffffff0c,
+	},
+};
+
+/*
+ * Calculate the address of the next transition for
+ * all types of nodes. Note that only DFA nodes and range
+ * nodes actually transition to another node. Match
+ * nodes not supposed to be encountered here.
+ * For quad range nodes:
+ * Calculate number of range boundaries that are less than the
+ * input value. Range boundaries for each node are in signed 8 bit,
+ * ordered from -128 to 127.
+ * This is effectively a popcnt of bytes that are greater than the
+ * input byte.
+ * Single nodes are processed in the same ways as quad range nodes.
+ */
+static __rte_always_inline __m512i
+calc_addr16(__m512i index_mask, __m512i next_input, __m512i shuffle_input,
+	__m512i four_32, __m512i range_base, __m512i tr_lo, __m512i tr_hi)
+{
+	__mmask64 qm;
+	__mmask16 dfa_msk;
+	__m512i addr, in, node_type, r, t;
+	__m512i dfa_ofs, quad_ofs;
+
+	t = _mm512_xor_si512(index_mask, index_mask);
+	in = _mm512_shuffle_epi8(next_input, shuffle_input);
+
+	/* Calc node type and node addr */
+	node_type = _mm512_andnot_si512(index_mask, tr_lo);
+	addr = _mm512_and_si512(index_mask, tr_lo);
+
+	/* mask for DFA type(0) nodes */
+	dfa_msk = _mm512_cmpeq_epi32_mask(node_type, t);
+
+	/* DFA calculations. */
+	r = _mm512_srli_epi32(in, 30);
+	r = _mm512_add_epi8(r, range_base);
+	t = _mm512_srli_epi32(in, 24);
+	r = _mm512_shuffle_epi8(tr_hi, r);
+
+	dfa_ofs = _mm512_sub_epi32(t, r);
+
+	/* QUAD/SINGLE calculations. */
+	qm = _mm512_cmpgt_epi8_mask(in, tr_hi);
+	t = _mm512_maskz_set1_epi8(qm, (uint8_t)UINT8_MAX);
+	t = _mm512_lzcnt_epi32(t);
+	t = _mm512_srli_epi32(t, 3);
+	quad_ofs = _mm512_sub_epi32(four_32, t);
+
+	/* blend DFA and QUAD/SINGLE. */
+	t = _mm512_mask_mov_epi32(quad_ofs, dfa_msk, dfa_ofs);
+
+	/* calculate address for next transitions. */
+	addr = _mm512_add_epi32(addr, t);
+	return addr;
+}
+
+/*
+ * Process 8 transitions in parallel.
+ * tr_lo contains low 32 bits for 8 transition.
+ * tr_hi contains high 32 bits for 8 transition.
+ * next_input contains up to 4 input bytes for 8 flows.
+ */
+static __rte_always_inline __m512i
+transition16(__m512i next_input, const uint64_t *trans, __m512i *tr_lo,
+	__m512i *tr_hi)
+{
+	const int32_t *tr;
+	__m512i addr;
+
+	tr = (const int32_t *)(uintptr_t)trans;
+
+	/* Calculate the address (array index) for all 8 transitions. */
+	addr = calc_addr16(zmm_index_mask.z, next_input, zmm_shuffle_input.z,
+		zmm_four_32.z, zmm_range_base.z, *tr_lo, *tr_hi);
+
+	/* load lower 32 bits of 8 transactions at once. */
+	*tr_lo = _mm512_i32gather_epi32(addr, tr, sizeof(trans[0]));
+
+	next_input = _mm512_srli_epi32(next_input, CHAR_BIT);
+
+	/* load high 32 bits of 8 transactions at once. */
+	*tr_hi = _mm512_i32gather_epi32(addr, (tr + 1), sizeof(trans[0]));
+
+	return next_input;
+}
+
+static __rte_always_inline void
+first_trans16(const struct acl_flow_avx512 *flow, __m512i next_input,
+	__mmask16 msk, __m512i *tr_lo, __m512i *tr_hi)
+{
+	const int32_t *tr;
+	__m512i addr, root;
+
+	tr = (const int32_t *)(uintptr_t)flow->trans;
+
+	addr = _mm512_set1_epi32(UINT8_MAX);
+	root = _mm512_set1_epi32(flow->root_index);
+
+	addr = _mm512_and_si512(next_input, addr);
+	addr = _mm512_add_epi32(root, addr);
+
+	/* load lower 32 bits of 8 transactions at once. */
+	*tr_lo = _mm512_mask_i32gather_epi32(*tr_lo, msk, addr, tr,
+		sizeof(flow->trans[0]));
+
+	/* load high 32 bits of 8 transactions at once. */
+	*tr_hi = _mm512_mask_i32gather_epi32(*tr_hi, msk, addr, (tr + 1),
+		sizeof(flow->trans[0]));
+}
+
+static inline __m512i
+get_next_4bytes_avx512x16(const struct acl_flow_avx512 *flow, __m512i pdata[2],
+	uint32_t msk, __m512i *di)
+{
+	const int32_t *div;
+	__m512i one, zero, t, p[2];
+	ymm_t inp[2];
+
+	static const __rte_x86_zmm_t zmm_pminp = {
+		.u32 = {
+			0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
+			0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17,
+		},
+	};
+
+	const __mmask16 pmidx_msk = 0x5555;
+
+	static const __rte_x86_zmm_t zmm_pmidx[2] = {
+		[0] = {
+			.u32 = {
+				0, 0, 1, 0, 2, 0, 3, 0,
+				4, 0, 5, 0, 6, 0, 7, 0,
+			},
+		},
+		[1] = {
+			.u32 = {
+				8, 0, 9, 0, 10, 0, 11, 0,
+				12, 0, 13, 0, 14, 0, 15, 0,
+			},
+		},
+	};
+
+	div = (const int32_t *)flow->data_index;
+
+	one = _mm512_set1_epi32(1);
+	zero = _mm512_xor_si512(one, one);
+
+	t = _mm512_mask_i32gather_epi32(zero, msk, *di, div, sizeof(div[0]));
+
+	*di = _mm512_mask_add_epi32(*di, msk, *di, one);
+
+	p[0] = _mm512_maskz_permutexvar_epi32(pmidx_msk, zmm_pmidx[0].z, t);
+	p[1] = _mm512_maskz_permutexvar_epi32(pmidx_msk, zmm_pmidx[1].z, t);
+
+	p[0] = _mm512_add_epi64(p[0], pdata[0]);
+	p[1] = _mm512_add_epi64(p[1], pdata[1]);
+
+	inp[0] = _mm512_mask_i64gather_epi32(_mm512_castsi512_si256(zero),
+		(msk & UINT8_MAX), p[0], NULL, sizeof(uint8_t));
+	inp[1] = _mm512_mask_i64gather_epi32(_mm512_castsi512_si256(zero),
+		(msk >> CHAR_BIT), p[1], NULL, sizeof(uint8_t));
+
+	return _mm512_permutex2var_epi32(_mm512_castsi256_si512(inp[0]),
+			zmm_pminp.z, _mm512_castsi256_si512(inp[1]));
+}
+
+static inline void
+start_flow16(struct acl_flow_avx512 *flow, uint32_t num, uint32_t msk,
+	__m512i pdata[2], __m512i *idx, __m512i *di)
+{
+	uint32_t n, nm[2];
+	__m512i ni, nd[2];
+
+	n = __builtin_popcount(msk & UINT8_MAX);
+	nm[0] = (1 << n) - 1;
+	nm[1] = (1 << (num - n)) - 1;
+
+	nd[0] = _mm512_maskz_loadu_epi64(nm[0],
+		flow->idata + flow->num_packets);
+	nd[1] = _mm512_maskz_loadu_epi64(nm[1],
+		flow->idata + flow->num_packets + n);
+
+	ni = _mm512_set1_epi32(flow->num_packets);
+	ni = _mm512_add_epi32(ni, zmm_idx_add.z);
+
+	pdata[0] = _mm512_mask_expand_epi64(pdata[0], (msk & UINT8_MAX), nd[0]);
+	pdata[1] = _mm512_mask_expand_epi64(pdata[1], (msk >> CHAR_BIT), nd[1]);
+
+	*idx = _mm512_mask_expand_epi32(*idx, msk, ni);
+	*di = _mm512_maskz_mov_epi32(msk ^ UINT16_MAX, *di);
+
+	flow->num_packets += num;
+}
+
+static inline uint32_t
+update_flow_mask16(const struct acl_flow_avx512 *flow, __mmask16 *fmsk,
+	__mmask16 *rmsk)
+{
+	uint32_t i, j, k, m, n;
+
+	fmsk[0] ^= rmsk[0];
+	m = rmsk[0];
+
+	k = __builtin_popcount(m);
+	n = flow->total_packets - flow->num_packets;
+
+	if (n < k) {
+		/* reduce mask */
+		for (i = k - n; i != 0; i--) {
+			j = sizeof(m) * CHAR_BIT - 1 - __builtin_clz(m);
+			m ^= 1 << j;
+		}
+	} else
+		n = k;
+
+	rmsk[0] = m;
+	fmsk[0] |= rmsk[0];
+
+	return n;
+}
+
+static inline uint32_t
+match_process_avx512x16(struct acl_flow_avx512 *flow, __mmask16 *fmsk,
+	__mmask16 *rmsk, __m512i pdata[2], __m512i *di, __m512i *idx,
+	__m512i *tr_lo, __m512i *tr_hi)
+{
+	uint32_t n;
+	__m512i res;
+
+	if (rmsk[0] == 0)
+		return 0;
+
+	/* extract match indexes */
+	res = _mm512_and_si512(tr_lo[0], zmm_index_mask.z);
+
+	/* mask  matched transitions to nop */
+	tr_lo[0] = _mm512_mask_mov_epi32(tr_lo[0], rmsk[0], zmm_trlo_idle.z);
+	tr_hi[0] = _mm512_mask_mov_epi32(tr_hi[0], rmsk[0], zmm_trhi_idle.z);
+
+	/* save found match indexes */
+	_mm512_mask_i32scatter_epi32(flow->matches, rmsk[0],
+		idx[0], res, sizeof(flow->matches[0]));
+
+	/* update masks and start new flows for matches */
+	n = update_flow_mask16(flow, fmsk, rmsk);
+	start_flow16(flow, n, rmsk[0], pdata, idx, di);
+
+	return n;
+}
+
+static inline void
+match_check_process_avx512x16x2(struct acl_flow_avx512 *flow, __mmask16 fm[2],
+	__m512i pdata[4], __m512i di[2], __m512i idx[2], __m512i inp[2],
+	__m512i tr_lo[2], __m512i tr_hi[2])
+{
+	uint32_t n[2];
+	__mmask16 rm[2];
+
+	/* check for matches */
+	rm[0] = _mm512_test_epi32_mask(tr_lo[0], zmm_match_mask.z);
+	rm[1] = _mm512_test_epi32_mask(tr_lo[1], zmm_match_mask.z);
+
+	while ((rm[0] | rm[1]) != 0) {
+
+		n[0] = match_process_avx512x16(flow, &fm[0], &rm[0], &pdata[0],
+			&di[0], &idx[0], &tr_lo[0], &tr_hi[0]);
+		n[1] = match_process_avx512x16(flow, &fm[1], &rm[1], &pdata[2],
+			&di[1], &idx[1], &tr_lo[1], &tr_hi[1]);
+
+		if (n[0] != 0) {
+			inp[0] = get_next_4bytes_avx512x16(flow, &pdata[0],
+				rm[0], &di[0]);
+			first_trans16(flow, inp[0], rm[0], &tr_lo[0],
+				&tr_hi[0]);
+			rm[0] = _mm512_test_epi32_mask(tr_lo[0],
+				zmm_match_mask.z);
+		}
+
+		if (n[1] != 0) {
+			inp[1] = get_next_4bytes_avx512x16(flow, &pdata[2],
+				rm[1], &di[1]);
+			first_trans16(flow, inp[1], rm[1], &tr_lo[1],
+				&tr_hi[1]);
+			rm[1] = _mm512_test_epi32_mask(tr_lo[1],
+				zmm_match_mask.z);
+		}
+	}
+}
+
+static inline void
+search_trie_avx512x16x2(struct acl_flow_avx512 *flow)
+{
+	__mmask16 fm[2];
+	__m512i di[2], idx[2], in[2], pdata[4], tr_lo[2], tr_hi[2];
+
+	/* first 1B load */
+	start_flow16(flow, MASK16_BIT, UINT16_MAX, &pdata[0], &idx[0], &di[0]);
+	start_flow16(flow, MASK16_BIT, UINT16_MAX, &pdata[2], &idx[1], &di[1]);
+
+	in[0] = get_next_4bytes_avx512x16(flow, &pdata[0], UINT16_MAX, &di[0]);
+	in[1] = get_next_4bytes_avx512x16(flow, &pdata[2], UINT16_MAX, &di[1]);
+
+	first_trans16(flow, in[0], UINT16_MAX, &tr_lo[0], &tr_hi[0]);
+	first_trans16(flow, in[1], UINT16_MAX, &tr_lo[1], &tr_hi[1]);
+
+	fm[0] = UINT16_MAX;
+	fm[1] = UINT16_MAX;
+
+	/* match check */
+	match_check_process_avx512x16x2(flow, fm, pdata, di, idx, in,
+		tr_lo, tr_hi);
+
+	while ((fm[0] | fm[1]) != 0) {
+
+		/* load next 4B */
+
+		in[0] = get_next_4bytes_avx512x16(flow, &pdata[0], fm[0],
+			&di[0]);
+		in[1] = get_next_4bytes_avx512x16(flow, &pdata[2], fm[1],
+			&di[1]);
+
+		/* main 4B loop */
+
+		in[0] = transition16(in[0], flow->trans, &tr_lo[0], &tr_hi[0]);
+		in[1] = transition16(in[1], flow->trans, &tr_lo[1], &tr_hi[1]);
+
+		in[0] = transition16(in[0], flow->trans, &tr_lo[0], &tr_hi[0]);
+		in[1] = transition16(in[1], flow->trans, &tr_lo[1], &tr_hi[1]);
+
+		in[0] = transition16(in[0], flow->trans, &tr_lo[0], &tr_hi[0]);
+		in[1] = transition16(in[1], flow->trans, &tr_lo[1], &tr_hi[1]);
+
+		in[0] = transition16(in[0], flow->trans, &tr_lo[0], &tr_hi[0]);
+		in[1] = transition16(in[1], flow->trans, &tr_lo[1], &tr_hi[1]);
+
+		/* check for matches */
+		match_check_process_avx512x16x2(flow, fm, pdata, di, idx, in,
+			tr_lo, tr_hi);
+	}
+}
+
+static inline __m512i
+resolve_match_idx_avx512x16(__m512i mi)
+{
+	RTE_BUILD_BUG_ON(sizeof(struct rte_acl_match_results) !=
+		1 << (match_log + 2));
+	return _mm512_slli_epi32(mi, match_log);
+}
+
+static inline __m512i
+resolve_pri_avx512x16(const int32_t res[], const int32_t pri[],
+	const uint32_t match[], __mmask16 msk, uint32_t nb_trie,
+	uint32_t nb_skip)
+{
+	uint32_t i;
+	const uint32_t *pm;
+	__mmask16 m;
+	__m512i cp, cr, np, nr, mch;
+
+	const __m512i zero = _mm512_set1_epi32(0);
+
+	mch = _mm512_maskz_loadu_epi32(msk, match);
+	mch = resolve_match_idx_avx512x16(mch);
+
+	cr = _mm512_mask_i32gather_epi32(zero, msk, mch, res, sizeof(res[0]));
+	cp = _mm512_mask_i32gather_epi32(zero, msk, mch, pri, sizeof(pri[0]));
+
+	for (i = 1, pm = match + nb_skip; i != nb_trie;
+			i++, pm += nb_skip) {
+
+		mch = _mm512_maskz_loadu_epi32(msk, pm);
+		mch = resolve_match_idx_avx512x16(mch);
+
+		nr = _mm512_mask_i32gather_epi32(zero, msk, mch, res,
+			sizeof(res[0]));
+		np = _mm512_mask_i32gather_epi32(zero, msk, mch, pri,
+			sizeof(pri[0]));
+
+		m = _mm512_cmpgt_epi32_mask(cp, np);
+		cr = _mm512_mask_mov_epi32(nr, m, cr);
+		cp = _mm512_mask_mov_epi32(np, m, cp);
+	}
+
+	return cr;
+}
+
+/*
+ * Resolve num (<= 16) matches for single category
+ */
+static inline void
+resolve_sc_avx512x16(uint32_t result[], const int32_t res[],
+	const int32_t pri[], const uint32_t match[], uint32_t nb_pkt,
+	uint32_t nb_trie, uint32_t nb_skip)
+{
+	__mmask16 msk;
+	__m512i cr;
+
+	msk = (1 << nb_pkt) - 1;
+	cr = resolve_pri_avx512x16(res, pri, match, msk, nb_trie, nb_skip);
+	_mm512_mask_storeu_epi32(result, msk, cr);
+}
+
+/*
+ * Resolve matches for single category
+ */
+static inline void
+resolve_sc_avx512x16x2(uint32_t result[],
+	const struct rte_acl_match_results pr[], const uint32_t match[],
+	uint32_t nb_pkt, uint32_t nb_trie)
+{
+	uint32_t i, j, k, n;
+	const uint32_t *pm;
+	const int32_t *res, *pri;
+	__mmask16 m[2];
+	__m512i cp[2], cr[2], np[2], nr[2], mch[2];
+
+	res = (const int32_t *)pr->results;
+	pri = pr->priority;
+
+	for (k = 0; k != (nb_pkt & ~MSK_AVX512X16X2); k += NUM_AVX512X16X2) {
+
+		j = k + MASK16_BIT;
+
+		/* load match indexes for first trie */
+		mch[0] = _mm512_loadu_si512(match + k);
+		mch[1] = _mm512_loadu_si512(match + j);
+
+		mch[0] = resolve_match_idx_avx512x16(mch[0]);
+		mch[1] = resolve_match_idx_avx512x16(mch[1]);
+
+		/* load matches and their priorities for first trie */
+
+		cr[0] = _mm512_i32gather_epi32(mch[0], res, sizeof(res[0]));
+		cr[1] = _mm512_i32gather_epi32(mch[1], res, sizeof(res[0]));
+
+		cp[0] = _mm512_i32gather_epi32(mch[0], pri, sizeof(pri[0]));
+		cp[1] = _mm512_i32gather_epi32(mch[1], pri, sizeof(pri[0]));
+
+		/* select match with highest priority */
+		for (i = 1, pm = match + nb_pkt; i != nb_trie;
+				i++, pm += nb_pkt) {
+
+			mch[0] = _mm512_loadu_si512(pm + k);
+			mch[1] = _mm512_loadu_si512(pm + j);
+
+			mch[0] = resolve_match_idx_avx512x16(mch[0]);
+			mch[1] = resolve_match_idx_avx512x16(mch[1]);
+
+			nr[0] = _mm512_i32gather_epi32(mch[0], res,
+				sizeof(res[0]));
+			nr[1] = _mm512_i32gather_epi32(mch[1], res,
+				sizeof(res[0]));
+
+			np[0] = _mm512_i32gather_epi32(mch[0], pri,
+				sizeof(pri[0]));
+			np[1] = _mm512_i32gather_epi32(mch[1], pri,
+				sizeof(pri[0]));
+
+			m[0] = _mm512_cmpgt_epi32_mask(cp[0], np[0]);
+			m[1] = _mm512_cmpgt_epi32_mask(cp[1], np[1]);
+
+			cr[0] = _mm512_mask_mov_epi32(nr[0], m[0], cr[0]);
+			cr[1] = _mm512_mask_mov_epi32(nr[1], m[1], cr[1]);
+
+			cp[0] = _mm512_mask_mov_epi32(np[0], m[0], cp[0]);
+			cp[1] = _mm512_mask_mov_epi32(np[1], m[1], cp[1]);
+		}
+
+		_mm512_storeu_si512(result + k, cr[0]);
+		_mm512_storeu_si512(result + j, cr[1]);
+	}
+
+	n = nb_pkt - k;
+	if (n != 0) {
+		if (n > MASK16_BIT) {
+			resolve_sc_avx512x16(result + k, res, pri, match + k,
+				MASK16_BIT, nb_trie, nb_pkt);
+			k += MASK16_BIT;
+			n -= MASK16_BIT;
+		}
+		resolve_sc_avx512x16(result + k, res, pri, match + k, n,
+				nb_trie, nb_pkt);
+	}
+}
+
+static inline int
+search_avx512x16x2(const struct rte_acl_ctx *ctx, const uint8_t **data,
+	uint32_t *results, uint32_t total_packets, uint32_t categories)
+{
+	uint32_t i, *pm;
+	const struct rte_acl_match_results *pr;
+	struct acl_flow_avx512 flow;
+	uint32_t match[ctx->num_tries * total_packets];
+
+	for (i = 0, pm = match; i != ctx->num_tries; i++, pm += total_packets) {
+
+		/* setup for next trie */
+		acl_set_flow_avx512(&flow, ctx, i, data, pm, total_packets);
+
+		/* process the trie */
+		search_trie_avx512x16x2(&flow);
+	}
+
+	/* resolve matches */
+	pr = (const struct rte_acl_match_results *)
+		(ctx->trans_table + ctx->match_index);
+
+	if (categories == 1)
+		resolve_sc_avx512x16x2(results, pr, match, total_packets,
+			ctx->num_tries);
+	else if (categories <= RTE_ACL_MAX_CATEGORIES / 2)
+		resolve_mcle8_avx512x1(results, pr, match, total_packets,
+			categories, ctx->num_tries);
+	else
+		resolve_mcgt8_avx512x1(results, pr, match, total_packets,
+			categories, ctx->num_tries);
+
+	return 0;
+}
-- 
2.17.1


  parent reply	other threads:[~2020-08-07 16:30 UTC|newest]

Thread overview: 70+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-08-07 16:28 [dpdk-dev] [PATCH 20.11 0/7] acl: introduce AVX512 classify method Konstantin Ananyev
2020-08-07 16:28 ` [dpdk-dev] [PATCH 20.11 1/7] acl: fix x86 build when compiler doesn't support AVX2 Konstantin Ananyev
2020-08-07 16:28 ` [dpdk-dev] [PATCH 20.11 2/7] app/acl: few small improvements Konstantin Ananyev
2020-08-07 16:28 ` [dpdk-dev] [PATCH 20.11 3/7] acl: remove of unused enum value Konstantin Ananyev
2020-08-07 16:28 ` [dpdk-dev] [PATCH 20.11 4/7] acl: add infrastructure to support AVX512 classify Konstantin Ananyev
2020-08-07 16:28 ` [dpdk-dev] [PATCH 20.11 5/7] app/acl: add AVX512 classify support Konstantin Ananyev
2020-08-07 16:28 ` [dpdk-dev] [PATCH 20.11 6/7] acl: introduce AVX512 classify implementation Konstantin Ananyev
2020-08-07 16:28 ` Konstantin Ananyev [this message]
2020-09-15 16:50 ` [dpdk-dev] [PATCH v2 00/12] acl: introduce AVX512 classify method Konstantin Ananyev
2020-09-15 16:50   ` [dpdk-dev] [PATCH v2 01/12] acl: fix x86 build when compiler doesn't support AVX2 Konstantin Ananyev
2020-09-15 16:50   ` [dpdk-dev] [PATCH v2 02/12] doc: fix mixing classify methods in ACL guide Konstantin Ananyev
2020-09-15 16:50   ` [dpdk-dev] [PATCH v2 03/12] acl: remove of unused enum value Konstantin Ananyev
2020-09-27  3:27     ` Ruifeng Wang
2020-09-15 16:50   ` [dpdk-dev] [PATCH v2 04/12] acl: remove library constructor Konstantin Ananyev
2020-09-15 16:50   ` [dpdk-dev] [PATCH v2 05/12] app/acl: few small improvements Konstantin Ananyev
2020-09-15 16:50   ` [dpdk-dev] [PATCH v2 06/12] test/acl: expand classify test coverage Konstantin Ananyev
2020-09-15 16:50   ` [dpdk-dev] [PATCH v2 07/12] acl: add infrastructure to support AVX512 classify Konstantin Ananyev
2020-09-16  9:11     ` Bruce Richardson
2020-09-16  9:36       ` Medvedkin, Vladimir
2020-09-16  9:49         ` Bruce Richardson
2020-09-16 10:06           ` Ananyev, Konstantin
2020-09-15 16:50   ` [dpdk-dev] [PATCH v2 08/12] acl: introduce AVX512 classify implementation Konstantin Ananyev
2020-09-15 16:50   ` [dpdk-dev] [PATCH v2 09/12] acl: enhance " Konstantin Ananyev
2020-09-15 16:50   ` [dpdk-dev] [PATCH v2 10/12] acl: for AVX512 classify use 4B load whenever possible Konstantin Ananyev
2020-09-15 16:50   ` [dpdk-dev] [PATCH v2 11/12] test/acl: add AVX512 classify support Konstantin Ananyev
2020-09-15 16:50   ` [dpdk-dev] [PATCH v2 12/12] app/acl: " Konstantin Ananyev
2020-10-05 18:45   ` [dpdk-dev] [PATCH v3 00/14] acl: introduce AVX512 classify methods Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 01/14] acl: fix x86 build when compiler doesn't support AVX2 Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 02/14] doc: fix missing classify methods in ACL guide Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 03/14] acl: remove of unused enum value Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 04/14] acl: remove library constructor Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 05/14] app/acl: few small improvements Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 06/14] test/acl: expand classify test coverage Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 07/14] acl: add infrastructure to support AVX512 classify Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 08/14] acl: introduce 256-bit width AVX512 classify implementation Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 09/14] acl: update default classify algorithm selection Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 10/14] acl: introduce 512-bit width AVX512 classify implementation Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 11/14] acl: for AVX512 classify use 4B load whenever possible Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 12/14] acl: deduplicate AVX512 code paths Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 13/14] test/acl: add AVX512 classify support Konstantin Ananyev
2020-10-05 18:45     ` [dpdk-dev] [PATCH v3 14/14] app/acl: " Konstantin Ananyev
2020-10-06 15:03     ` [dpdk-dev] [PATCH v4 00/14] acl: introduce AVX512 classify methods Konstantin Ananyev
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 01/14] acl: fix x86 build when compiler doesn't support AVX2 Konstantin Ananyev
2020-10-08 13:42         ` [dpdk-dev] [dpdk-stable] " David Marchand
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 02/14] doc: fix missing classify methods in ACL guide Konstantin Ananyev
2020-10-08 13:42         ` David Marchand
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 03/14] acl: remove of unused enum value Konstantin Ananyev
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 04/14] acl: remove library constructor Konstantin Ananyev
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 05/14] app/acl: few small improvements Konstantin Ananyev
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 06/14] test/acl: expand classify test coverage Konstantin Ananyev
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 07/14] acl: add infrastructure to support AVX512 classify Konstantin Ananyev
2020-10-13 19:17         ` David Marchand
2020-10-13 22:26           ` Ananyev, Konstantin
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 08/14] acl: introduce 256-bit width AVX512 classify implementation Konstantin Ananyev
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 09/14] acl: update default classify algorithm selection Konstantin Ananyev
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 10/14] acl: introduce 512-bit width AVX512 classify implementation Konstantin Ananyev
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 11/14] acl: for AVX512 classify use 4B load whenever possible Konstantin Ananyev
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 12/14] acl: deduplicate AVX512 code paths Konstantin Ananyev
2020-10-16 15:56         ` Ferruh Yigit
2020-10-16 16:20           ` Thomas Monjalon
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 13/14] test/acl: add AVX512 classify support Konstantin Ananyev
2020-10-14 10:26         ` David Marchand
2020-10-14 10:32           ` Ananyev, Konstantin
2020-10-14 10:35             ` David Marchand
2020-10-06 15:03       ` [dpdk-dev] [PATCH v4 14/14] app/acl: " Konstantin Ananyev
2020-10-14 12:40       ` [dpdk-dev] [PATCH v4 00/14] acl: introduce AVX512 classify methods David Marchand
2020-10-06 15:05     ` [dpdk-dev] [PATCH v3 " David Marchand
2020-10-06 16:07       ` Ananyev, Konstantin
2020-10-08 10:49         ` David Marchand
2020-10-14  9:23         ` Kinsella, Ray

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20200807162829.11690-8-konstantin.ananyev@intel.com \
    --to=konstantin.ananyev@intel.com \
    --cc=dev@dpdk.org \
    --cc=jerinj@marvell.com \
    --cc=ruifeng.wang@arm.com \
    --cc=vladimir.medvedkin@intel.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).