forked from google/XNNPACK
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathutils.cc
276 lines (241 loc) · 6.97 KB
/
utils.cc
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
// Copyright 2019 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <mutex>
#ifdef __linux__
#include <sched.h>
#endif
#if defined(__ANDROID__) || defined(_WIN32) || defined(__CYGWIN__)
#include <malloc.h>
#endif
#if defined(__SSE__) || defined(__x86_64__)
#include <xmmintrin.h>
#endif
#include <cpuinfo.h>
#include "bench/utils.h"
static void* wipe_buffer = nullptr;
static size_t wipe_buffer_size = 0;
static std::once_flag wipe_buffer_guard;
static void InitWipeBuffer() {
// Default: the largest know cache size (128 MB Intel Crystalwell L4 cache).
wipe_buffer_size = 128 * 1024 * 1024;
if (cpuinfo_initialize()) {
wipe_buffer_size = benchmark::utils::GetMaxCacheSize();
}
#if defined(_WIN32)
wipe_buffer = _aligned_malloc(wipe_buffer_size, 128);
#elif defined(__ANDROID__) || defined(__CYGWIN__)
// memalign is obsolete, but it is the only option on Android until API level 17.
wipe_buffer = memalign(128, wipe_buffer_size);
#else
(void) posix_memalign((void**) &wipe_buffer, 128, wipe_buffer_size);
#endif
if (wipe_buffer != nullptr) {
memset(wipe_buffer, 0xA5, wipe_buffer_size);
}
}
namespace benchmark {
namespace utils {
uint32_t PrefetchToL1(const void* ptr, size_t size) {
uint32_t step = 16;
if (cpuinfo_initialize()) {
step = cpuinfo_get_l1d_cache(0)->line_size;
}
const uint8_t* u8_ptr = static_cast<const uint8_t*>(ptr);
// Compute and return sum of data to prevent compiler from removing data reads.
uint32_t sum = 0;
while (size >= step) {
sum += uint32_t(*u8_ptr);
u8_ptr += step;
size -= step;
}
return sum;
}
uint32_t WipeCache() {
std::call_once(wipe_buffer_guard, InitWipeBuffer);
return PrefetchToL1(wipe_buffer, wipe_buffer_size);
}
void DisableDenormals() {
#if defined(__SSE__) || defined(__x86_64__)
_mm_setcsr(_mm_getcsr() | 0x8040);
#elif defined(__arm__) && defined(__ARM_FP) && (__ARM_FP != 0)
uint32_t fpscr;
#if defined(__thumb__) && !defined(__thumb2__)
__asm__ __volatile__(
"VMRS %[fpscr], fpscr\n"
"ORRS %[fpscr], %[bitmask]\n"
"VMSR fpscr, %[fpscr]\n"
: [fpscr] "=l" (fpscr)
: [bitmask] "l" (0x1000000)
: "cc");
#else
__asm__ __volatile__(
"VMRS %[fpscr], fpscr\n"
"ORR %[fpscr], #0x1000000\n"
"VMSR fpscr, %[fpscr]\n"
: [fpscr] "=r" (fpscr));
#endif
#elif defined(__aarch64__)
uint64_t fpcr;
__asm__ __volatile__(
"MRS %[fpcr], fpcr\n"
"ORR %w[fpcr], %w[fpcr], 0x1000000\n"
"ORR %w[fpcr], %w[fpcr], 0x80000\n"
"MSR fpcr, %[fpcr]\n"
: [fpcr] "=r" (fpcr));
#endif
}
// Return clockrate in Hz
uint64_t GetCurrentCpuFrequency() {
#ifdef __linux__
int freq = 0;
char cpuinfo_name[512];
int cpu = sched_getcpu();
snprintf(cpuinfo_name, sizeof(cpuinfo_name),
"/sys/devices/system/cpu/cpu%d/cpufreq/scaling_cur_freq", cpu);
FILE* f = fopen(cpuinfo_name, "r");
if (f) {
if (fscanf(f, "%d", &freq)) {
fclose(f);
return uint64_t(freq) * 1000;
}
fclose(f);
}
#endif // __linux__
return 0;
}
size_t GetMaxCacheSize() {
if (!cpuinfo_initialize()) {
#if CPUINFO_ARCH_ARM || CPUINFO_ARCH_ARM64
// DynamIQ max: 4 MB
return 4 * 1024 * 1024;
#else
// Intel eDRAM max: 128 MB
return 128 * 1024 * 1024;
#endif
}
return cpuinfo_get_max_cache_size();
}
void MultiThreadingParameters(benchmark::internal::Benchmark* benchmark) {
benchmark->ArgName("T");
// Disabled thread pool (execution on the caller thread only).
benchmark->Arg(1);
if (cpuinfo_initialize()) {
// All cores except the little ones.
uint32_t max_cores = cpuinfo_get_cores_count();
if (cpuinfo_get_clusters_count() > 1) {
max_cores -= cpuinfo_get_cluster(cpuinfo_get_clusters_count() - 1)->core_count;
}
for (uint32_t t = 2; t <= max_cores; t++) {
benchmark->Arg(t);
}
// All cores (if more than one cluster).
if (cpuinfo_get_cores_count() > max_cores) {
benchmark->Arg(cpuinfo_get_cores_count());
}
// All cores + hyperthreads (only if hyperthreading supported).
if (cpuinfo_get_processors_count() > cpuinfo_get_cores_count()) {
benchmark->Arg(cpuinfo_get_processors_count());
}
}
}
bool CheckVFP(benchmark::State& state) {
if (!cpuinfo_initialize() || !(cpuinfo_has_arm_vfpv2() || cpuinfo_has_arm_vfpv3())) {
state.SkipWithError("no VFP extension");
return false;
}
return true;
}
bool CheckNEONFP16ARITH(benchmark::State& state) {
if (!cpuinfo_initialize() || !cpuinfo_has_arm_neon_fp16_arith()) {
state.SkipWithError("no NEON-FP16-ARITH extension");
return false;
}
return true;
}
bool CheckNEON(benchmark::State& state) {
if (!cpuinfo_initialize() || !cpuinfo_has_arm_neon()) {
state.SkipWithError("no NEON extension");
return false;
}
return true;
}
bool CheckNEONFMA(benchmark::State& state) {
if (!cpuinfo_initialize() || !cpuinfo_has_arm_neon_fma()) {
state.SkipWithError("no NEON-FMA extension");
return false;
}
return true;
}
bool CheckNEONDOT(benchmark::State& state) {
if (!cpuinfo_initialize() || !cpuinfo_has_arm_neon_dot()) {
state.SkipWithError("no NEON-DOT extension");
return false;
}
return true;
}
bool CheckSSSE3(benchmark::State& state) {
if (!cpuinfo_initialize() || !cpuinfo_has_x86_ssse3()) {
state.SkipWithError("no SSSE3 extension");
return false;
}
return true;
}
bool CheckSSE41(benchmark::State& state) {
if (!cpuinfo_initialize() || !cpuinfo_has_x86_sse4_1()) {
state.SkipWithError("no SSE4.1 extension");
return false;
}
return true;
}
bool CheckAVX(benchmark::State& state) {
if (!cpuinfo_initialize() || !cpuinfo_has_x86_avx()) {
state.SkipWithError("no AVX extension");
return false;
}
return true;
}
bool CheckXOP(benchmark::State& state) {
if (!cpuinfo_initialize() || !cpuinfo_has_x86_xop()) {
state.SkipWithError("no XOP extension");
return false;
}
return true;
}
bool CheckFMA3(benchmark::State& state) {
if (!cpuinfo_initialize() || !cpuinfo_has_x86_fma3()) {
state.SkipWithError("no FMA3 extension");
return false;
}
return true;
}
bool CheckAVX2(benchmark::State& state) {
if (!cpuinfo_initialize() || !cpuinfo_has_x86_avx2()) {
state.SkipWithError("no AVX2 extension");
return false;
}
return true;
}
bool CheckAVX512F(benchmark::State& state) {
if (!cpuinfo_initialize() || !cpuinfo_has_x86_avx512f()) {
state.SkipWithError("no AVX512F extension");
return false;
}
return true;
}
bool CheckAVX512SKX(benchmark::State& state) {
if (!cpuinfo_initialize() || !cpuinfo_has_x86_avx512f() ||
!cpuinfo_has_x86_avx512cd() || !cpuinfo_has_x86_avx512bw() ||
!cpuinfo_has_x86_avx512dq() || !cpuinfo_has_x86_avx512vl())
{
state.SkipWithError("no AVX512 SKX extensions");
return false;
}
return true;
}
} // namespace utils
} // namespace benchmark