1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2020 Intel Corporation
6 * Defines required by "acl_run_avx512_common.h".
7 * Note that all of them has to be undefined by the end
8 * of this file, as "acl_run_avx512_common.h" can be included several
9 * times from different *.h files for the same *.c.
13 * This implementation uses 256-bit registers(ymm) and intrinsics.
14 * So our main SIMD type is 256-bit width and each such variable can
15 * process sizeof(__m256i) / sizeof(uint32_t) == 8 entries in parallel.
17 #define _T_simd __m256i
18 #define _T_mask __mmask8
20 /* Naming convention for static const variables. */
21 #define _SC_(x) ymm_##x
22 #define _SV_(x) (ymm_##x.y)
24 /* Naming convention for internal functions. */
25 #define _F_(x) x##_avx512x8
28 * Same intrinsics have different syntaxes (depending on the bit-width),
29 * so to overcome that few macros need to be defined.
32 /* Naming convention for generic epi(packed integers) type intrinsics. */
33 #define _M_I_(x) _mm256_##x
35 /* Naming convention for si(whole simd integer) type intrinsics. */
36 #define _M_SI_(x) _mm256_##x##_si256
38 /* Naming convention for masked gather type intrinsics. */
39 #define _M_MGI_(x) _mm256_m##x
41 /* Naming convention for gather type intrinsics. */
42 #define _M_GI_(name, idx, base, scale) _mm256_##name(base, idx, scale)
44 /* num/mask of transitions per SIMD regs */
45 #define _SIMD_MASK_BIT_ (sizeof(_T_simd) / sizeof(uint32_t))
46 #define _SIMD_MASK_MAX_ RTE_LEN2MASK(_SIMD_MASK_BIT_, uint32_t)
48 #define _SIMD_FLOW_NUM_ (2 * _SIMD_MASK_BIT_)
49 #define _SIMD_FLOW_MSK_ (_SIMD_FLOW_NUM_ - 1)
51 /* num/mask of pointers per SIMD regs */
52 #define _SIMD_PTR_NUM_ (sizeof(_T_simd) / sizeof(uintptr_t))
53 #define _SIMD_PTR_MSK_ RTE_LEN2MASK(_SIMD_PTR_NUM_, uint32_t)
55 static const rte_ymm_t _SC_(match_mask) = {
68 static const rte_ymm_t _SC_(index_mask) = {
81 static const rte_ymm_t _SC_(trlo_idle) = {
94 static const rte_ymm_t _SC_(trhi_idle) = {
101 static const rte_ymm_t _SC_(shuffle_input) = {
103 0x00000000, 0x04040404, 0x08080808, 0x0c0c0c0c,
104 0x00000000, 0x04040404, 0x08080808, 0x0c0c0c0c,
108 static const rte_ymm_t _SC_(four_32) = {
115 static const rte_ymm_t _SC_(idx_add) = {
122 static const rte_ymm_t _SC_(range_base) = {
124 0xffffff00, 0xffffff04, 0xffffff08, 0xffffff0c,
125 0xffffff00, 0xffffff04, 0xffffff08, 0xffffff0c,
129 static const rte_ymm_t _SC_(pminp) = {
131 0x00, 0x01, 0x02, 0x03,
132 0x08, 0x09, 0x0a, 0x0b,
136 static const __mmask16 _SC_(pmidx_msk) = 0x55;
138 static const rte_ymm_t _SC_(pmidx[2]) = {
141 0, 0, 1, 0, 2, 0, 3, 0,
146 4, 0, 5, 0, 6, 0, 7, 0,
152 * unfortunately current AVX512 ISA doesn't provide ability for
153 * gather load on a byte quantity. So we have to mimic it in SW,
154 * by doing 4x1B scalar loads.
156 static inline __m128i
157 _m256_mask_gather_epi8x4(__m256i pdata, __mmask8 mask)
162 static const uint32_t zero;
164 p.y = _mm256_mask_set1_epi64(pdata, mask ^ _SIMD_PTR_MSK_,
167 v.u32[0] = *(uint8_t *)p.u64[0];
168 v.u32[1] = *(uint8_t *)p.u64[1];
169 v.u32[2] = *(uint8_t *)p.u64[2];
170 v.u32[3] = *(uint8_t *)p.u64[3];
176 * Gather 4/1 input bytes for up to 8 (2*8) locations in parallel.
178 static __rte_always_inline __m256i
179 _F_(gather_bytes)(__m256i zero, const __m256i p[2], const uint32_t m[2],
184 if (bnum == sizeof(uint8_t)) {
185 inp[0] = _m256_mask_gather_epi8x4(p[0], m[0]);
186 inp[1] = _m256_mask_gather_epi8x4(p[1], m[1]);
188 inp[0] = _mm256_mmask_i64gather_epi32(
189 _mm256_castsi256_si128(zero),
190 m[0], p[0], NULL, sizeof(uint8_t));
191 inp[1] = _mm256_mmask_i64gather_epi32(
192 _mm256_castsi256_si128(zero),
193 m[1], p[1], NULL, sizeof(uint8_t));
196 /* squeeze input into one 256-bit register */
197 return _mm256_permutex2var_epi32(_mm256_castsi128_si256(inp[0]),
198 _SV_(pminp), _mm256_castsi128_si256(inp[1]));
201 #include "acl_run_avx512_common.h"
204 * Perform search for up to (2 * 8) flows in parallel.
205 * Use two sets of metadata, each serves 8 flows max.
208 search_avx512x8x2(const struct rte_acl_ctx *ctx, const uint8_t **data,
209 uint32_t *results, uint32_t total_packets, uint32_t categories)
212 const struct rte_acl_match_results *pr;
213 struct acl_flow_avx512 flow;
214 uint32_t match[ctx->num_tries * total_packets];
216 for (i = 0, pm = match; i != ctx->num_tries; i++, pm += total_packets) {
218 /* setup for next trie */
219 acl_set_flow_avx512(&flow, ctx, i, data, pm, total_packets);
221 /* process the trie */
222 _F_(search_trie)(&flow);
225 /* resolve matches */
226 pr = (const struct rte_acl_match_results *)
227 (ctx->trans_table + ctx->match_index);
230 _F_(resolve_single_cat)(results, pr, match, total_packets,
233 resolve_mcle8_avx512x1(results, pr, match, total_packets,
234 categories, ctx->num_tries);
239 #undef _SIMD_PTR_MSK_
240 #undef _SIMD_PTR_NUM_
241 #undef _SIMD_FLOW_MSK_
242 #undef _SIMD_FLOW_NUM_
243 #undef _SIMD_MASK_MAX_
244 #undef _SIMD_MASK_BIT_