1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2021 Intel Corporation
3 */
4
5 #include <stdio.h>
6 #include <stdint.h>
7 #include <stdlib.h>
8 #include <math.h>
9
10 #include <rte_cycles.h>
11 #include <rte_malloc.h>
12 #include <rte_random.h>
13 #include <rte_thash.h>
14
15 #include "test.h"
16
17 #define ITERATIONS (1 << 15)
18 #define BATCH_SZ (1 << 10)
19
20 #define IPV4_2_TUPLE_LEN (8)
21 #define IPV4_4_TUPLE_LEN (12)
22 #define IPV6_2_TUPLE_LEN (32)
23 #define IPV6_4_TUPLE_LEN (36)
24
25
26 static const uint8_t default_rss_key[] = {
27 0x6d, 0x5a, 0x56, 0xda, 0x25, 0x5b, 0x0e, 0xc2,
28 0x41, 0x67, 0x25, 0x3d, 0x43, 0xa3, 0x8f, 0xb0,
29 0xd0, 0xca, 0x2b, 0xcb, 0xae, 0x7b, 0x30, 0xb4,
30 0x77, 0xcb, 0x2d, 0xa3, 0x80, 0x30, 0xf2, 0x0c,
31 0x6a, 0x42, 0xb7, 0x3b, 0xbe, 0xac, 0x01, 0xfa,
32 };
33
34 enum test_rss_type {
35 TEST_SOFTRSS,
36 TEST_SOFTRSS_BE,
37 TEST_RSS_GFNI
38 };
39
40 static inline uint64_t
run_rss_calc(uint32_t * tuples[BATCH_SZ],enum test_rss_type type,int len,const void * key)41 run_rss_calc(uint32_t *tuples[BATCH_SZ], enum test_rss_type type, int len,
42 const void *key)
43 {
44 int i, j;
45 uint64_t start_tsc, end_tsc;
46 volatile uint32_t hash = 0;
47
48 start_tsc = rte_rdtsc_precise();
49 for (i = 0; i < ITERATIONS; i++) {
50 for (j = 0; j < BATCH_SZ; j++) {
51 if (type == TEST_SOFTRSS)
52 hash ^= rte_softrss(tuples[j], len /
53 sizeof(uint32_t), (const uint8_t *)key);
54 else if (type == TEST_SOFTRSS_BE)
55 hash ^= rte_softrss_be(tuples[j], len /
56 sizeof(uint32_t), (const uint8_t *)key);
57 else
58 hash ^= rte_thash_gfni((const uint64_t *)key,
59 (uint8_t *)tuples[j], len);
60 }
61 }
62 end_tsc = rte_rdtsc_precise();
63
64 /* To avoid compiler warnings set hash to used. */
65 RTE_SET_USED(hash);
66
67 return end_tsc - start_tsc;
68 }
69
70 static inline uint64_t
run_rss_calc_bulk(uint32_t * tuples[BATCH_SZ],int len,const void * key)71 run_rss_calc_bulk(uint32_t *tuples[BATCH_SZ], int len, const void *key)
72 {
73 int i;
74 uint64_t start_tsc, end_tsc;
75 uint32_t bulk_hash[BATCH_SZ] = { 0 };
76
77 start_tsc = rte_rdtsc_precise();
78 for (i = 0; i < ITERATIONS; i++)
79 rte_thash_gfni_bulk((const uint64_t *)key, len,
80 (uint8_t **)tuples, bulk_hash, BATCH_SZ);
81
82 end_tsc = rte_rdtsc_precise();
83
84 return end_tsc - start_tsc;
85 }
86
87 static void
run_thash_test(unsigned int tuple_len)88 run_thash_test(unsigned int tuple_len)
89 {
90 uint32_t *tuples[BATCH_SZ];
91 unsigned int i, j;
92 uint32_t len = RTE_ALIGN_CEIL(tuple_len, sizeof(uint32_t));
93 uint64_t tsc_diff;
94
95 for (i = 0; i < BATCH_SZ; i++) {
96 tuples[i] = rte_zmalloc(NULL, len, 0);
97 for (j = 0; j < len / sizeof(uint32_t); j++)
98 tuples[i][j] = rte_rand();
99 }
100
101 tsc_diff = run_rss_calc(tuples, TEST_SOFTRSS, len, default_rss_key);
102 printf("Average rte_softrss() takes \t\t%.1f cycles for key len %d\n",
103 (double)(tsc_diff) / (double)(ITERATIONS * BATCH_SZ), len);
104
105 tsc_diff = run_rss_calc(tuples, TEST_SOFTRSS_BE, len,
106 default_rss_key);
107 printf("Average rte_softrss_be() takes \t\t%.1f cycles for key len %d\n",
108 (double)(tsc_diff) / (double)(ITERATIONS * BATCH_SZ), len);
109
110 if (!rte_thash_gfni_supported())
111 return;
112
113 uint64_t rss_key_matrixes[RTE_DIM(default_rss_key)];
114
115 rte_thash_complete_matrix(rss_key_matrixes, default_rss_key,
116 RTE_DIM(default_rss_key));
117
118 tsc_diff = run_rss_calc(tuples, TEST_RSS_GFNI, len, rss_key_matrixes);
119 printf("Average rte_thash_gfni takes \t\t%.1f cycles for key len %d\n",
120 (double)(tsc_diff) / (double)(ITERATIONS * BATCH_SZ), len);
121
122 tsc_diff = run_rss_calc_bulk(tuples, len, rss_key_matrixes);
123 printf("Average rte_thash_gfni_bulk takes \t%.1f cycles for key len %d\n",
124 (double)(tsc_diff) / (double)(ITERATIONS * BATCH_SZ), len);
125 }
126
127 static int
test_thash_perf(void)128 test_thash_perf(void)
129 {
130 run_thash_test(IPV4_2_TUPLE_LEN);
131 run_thash_test(IPV4_4_TUPLE_LEN);
132 run_thash_test(IPV6_2_TUPLE_LEN);
133 run_thash_test(IPV6_4_TUPLE_LEN);
134
135 return 0;
136 }
137
138 REGISTER_PERF_TEST(thash_perf_autotest, test_thash_perf);
139