DPDK  22.07.0
rte_thash_x86_gfni.h
Go to the documentation of this file.
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2021 Intel Corporation
3  */
4 
5 #ifndef _RTE_THASH_X86_GFNI_H_
6 #define _RTE_THASH_X86_GFNI_H_
7 
15 #include <rte_vect.h>
16 
17 #ifdef __cplusplus
18 extern "C" {
19 #endif
20 
21 #if defined(__GFNI__) && defined(__AVX512F__)
22 #define RTE_THASH_GFNI_DEFINED
23 
24 #define RTE_THASH_FIRST_ITER_MSK 0x0f0f0f0f0f0e0c08
25 #define RTE_THASH_PERM_MSK 0x0f0f0f0f0f0f0f0f
26 #define RTE_THASH_FIRST_ITER_MSK_2 0xf0f0f0f0f0e0c080
27 #define RTE_THASH_PERM_MSK_2 0xf0f0f0f0f0f0f0f0
28 #define RTE_THASH_REWIND_MSK 0x0000000000113377
29 
30 __rte_internal
31 static inline void
32 __rte_thash_xor_reduce(__m512i xor_acc, uint32_t *val_1, uint32_t *val_2)
33 {
34  __m256i tmp_256_1, tmp_256_2;
35  __m128i tmp128_1, tmp128_2;
36 
37  tmp_256_1 = _mm512_castsi512_si256(xor_acc);
38  tmp_256_2 = _mm512_extracti32x8_epi32(xor_acc, 1);
39  tmp_256_1 = _mm256_xor_si256(tmp_256_1, tmp_256_2);
40 
41  tmp128_1 = _mm256_castsi256_si128(tmp_256_1);
42  tmp128_2 = _mm256_extracti32x4_epi32(tmp_256_1, 1);
43  tmp128_1 = _mm_xor_si128(tmp128_1, tmp128_2);
44 
45 #ifdef RTE_ARCH_X86_64
46  uint64_t tmp_1, tmp_2;
47  tmp_1 = _mm_extract_epi64(tmp128_1, 0);
48  tmp_2 = _mm_extract_epi64(tmp128_1, 1);
49  tmp_1 ^= tmp_2;
50 
51  *val_1 = (uint32_t)tmp_1;
52  *val_2 = (uint32_t)(tmp_1 >> 32);
53 #else
54  uint32_t tmp_1, tmp_2;
55  tmp_1 = _mm_extract_epi32(tmp128_1, 0);
56  tmp_2 = _mm_extract_epi32(tmp128_1, 1);
57  tmp_1 ^= _mm_extract_epi32(tmp128_1, 2);
58  tmp_2 ^= _mm_extract_epi32(tmp128_1, 3);
59 
60  *val_1 = tmp_1;
61  *val_2 = tmp_2;
62 #endif
63 }
64 
65 __rte_internal
66 static inline __m512i
67 __rte_thash_gfni(const uint64_t *mtrx, const uint8_t *tuple,
68  const uint8_t *secondary_tuple, int len)
69 {
70  __m512i permute_idx = _mm512_set_epi32(0x07060504, 0x07060504,
71  0x06050403, 0x06050403,
72  0x05040302, 0x05040302,
73  0x04030201, 0x04030201,
74  0x03020100, 0x03020100,
75  0x020100FF, 0x020100FF,
76  0x0100FFFE, 0x0100FFFE,
77  0x00FFFEFD, 0x00FFFEFD);
78  const __m512i rewind_idx = _mm512_set_epi32(0x00000000, 0x00000000,
79  0x00000000, 0x00000000,
80  0x00000000, 0x00000000,
81  0x00000000, 0x00000000,
82  0x00000000, 0x00000000,
83  0x0000003B, 0x0000003B,
84  0x00003B3A, 0x00003B3A,
85  0x003B3A39, 0x003B3A39);
86  const __mmask64 rewind_mask = RTE_THASH_REWIND_MSK;
87  const __m512i shift_8 = _mm512_set1_epi8(8);
88  __m512i xor_acc = _mm512_setzero_si512();
89  __m512i perm_bytes = _mm512_setzero_si512();
90  __m512i vals, matrixes, tuple_bytes, tuple_bytes_2;
91  __mmask64 load_mask, permute_mask, permute_mask_2;
92  int chunk_len = 0, i = 0;
93  uint8_t mtrx_msk;
94  const int prepend = 3;
95 
96  for (; len > 0; len -= 64, tuple += 64) {
97  if (i == 8)
98  perm_bytes = _mm512_maskz_permutexvar_epi8(rewind_mask,
99  rewind_idx, perm_bytes);
100 
101  permute_mask = RTE_THASH_FIRST_ITER_MSK;
102  load_mask = (len >= 64) ? UINT64_MAX : ((1ULL << len) - 1);
103  tuple_bytes = _mm512_maskz_loadu_epi8(load_mask, tuple);
104  if (secondary_tuple) {
105  permute_mask_2 = RTE_THASH_FIRST_ITER_MSK_2;
106  tuple_bytes_2 = _mm512_maskz_loadu_epi8(load_mask,
107  secondary_tuple);
108  }
109 
110  chunk_len = __builtin_popcountll(load_mask);
111  for (i = 0; i < ((chunk_len + prepend) / 8); i++, mtrx += 8) {
112  perm_bytes = _mm512_mask_permutexvar_epi8(perm_bytes,
113  permute_mask, permute_idx, tuple_bytes);
114 
115  if (secondary_tuple)
116  perm_bytes =
117  _mm512_mask_permutexvar_epi8(perm_bytes,
118  permute_mask_2, permute_idx,
119  tuple_bytes_2);
120 
121  matrixes = _mm512_maskz_loadu_epi64(UINT8_MAX, mtrx);
122  vals = _mm512_gf2p8affine_epi64_epi8(perm_bytes,
123  matrixes, 0);
124 
125  xor_acc = _mm512_xor_si512(xor_acc, vals);
126  permute_idx = _mm512_add_epi8(permute_idx, shift_8);
127  permute_mask = RTE_THASH_PERM_MSK;
128  if (secondary_tuple)
129  permute_mask_2 = RTE_THASH_PERM_MSK_2;
130  }
131  }
132 
133  int rest_len = (chunk_len + prepend) % 8;
134  if (rest_len != 0) {
135  mtrx_msk = (1 << (rest_len % 8)) - 1;
136  matrixes = _mm512_maskz_loadu_epi64(mtrx_msk, mtrx);
137  if (i == 8) {
138  perm_bytes = _mm512_maskz_permutexvar_epi8(rewind_mask,
139  rewind_idx, perm_bytes);
140  } else {
141  perm_bytes = _mm512_mask_permutexvar_epi8(perm_bytes,
142  permute_mask, permute_idx, tuple_bytes);
143 
144  if (secondary_tuple)
145  perm_bytes =
146  _mm512_mask_permutexvar_epi8(
147  perm_bytes, permute_mask_2,
148  permute_idx, tuple_bytes_2);
149  }
150 
151  vals = _mm512_gf2p8affine_epi64_epi8(perm_bytes, matrixes, 0);
152  xor_acc = _mm512_xor_si512(xor_acc, vals);
153  }
154 
155  return xor_acc;
156 }
157 
175 __rte_experimental
176 static inline uint32_t
177 rte_thash_gfni(const uint64_t *m, const uint8_t *tuple, int len)
178 {
179  uint32_t val, val_zero;
180 
181  __m512i xor_acc = __rte_thash_gfni(m, tuple, NULL, len);
182  __rte_thash_xor_reduce(xor_acc, &val, &val_zero);
183 
184  return val;
185 }
186 
207 __rte_experimental
208 static inline void
209 rte_thash_gfni_bulk(const uint64_t *mtrx, int len, uint8_t *tuple[],
210  uint32_t val[], uint32_t num)
211 {
212  uint32_t i;
213  uint32_t val_zero;
214  __m512i xor_acc;
215 
216  for (i = 0; i != (num & ~1); i += 2) {
217  xor_acc = __rte_thash_gfni(mtrx, tuple[i], tuple[i + 1], len);
218  __rte_thash_xor_reduce(xor_acc, val + i, val + i + 1);
219  }
220 
221  if (num & 1) {
222  xor_acc = __rte_thash_gfni(mtrx, tuple[i], NULL, len);
223  __rte_thash_xor_reduce(xor_acc, val + i, &val_zero);
224  }
225 }
226 
227 #endif /* __GFNI__ && __AVX512F__ */
228 
229 #ifdef __cplusplus
230 }
231 #endif
232 
233 #endif /* _RTE_THASH_X86_GFNI_H_ */
rte_vect.h
Generated by   doxygen 1.8.13