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
|
// 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 <algorithm>
#include <cfloat>
#include <cmath>
#include <functional>
#include <random>
#include <vector>
#include <benchmark/benchmark.h>
#include <fp16/fp16.h>
#include <xnnpack/AlignedAllocator.h>
#include <xnnpack/common.h>
#include <xnnpack/math-stubs.h>
static void SigmoidError(benchmark::State& state,
xnn_f32_unary_math_function sigmoid,
size_t tile_size)
{
// The smallest x for which sigmoidf(x) is normalized (-0x1.5D589Ep+6f).
const uint32_t min_input = 0xC2AEAC4F;
// The largest x for which sigmoidf(x) is not 1.0f (0x1.154244p+4f).
const uint32_t max_input = 0x418AA122;
// Number of tiles in one block of inputs/outputs. Combining multiple tiles in a block reduce function call overhead.
const size_t num_tiles = 100;
double max_ulp_error = 0.0;
std::vector<float, AlignedAllocator<float, 64>> x(tile_size * num_tiles);
std::vector<float, AlignedAllocator<float, 64>> y(tile_size * num_tiles);
for (auto _ : state) {
for (uint32_t n = min_input; int32_t(n) < 0; n -= tile_size * num_tiles) {
for (uint32_t i = 0; i < tile_size * num_tiles; i++) {
x[i] = fp32_from_bits(std::max<uint32_t>(n - i, 0x80000000));
}
std::fill(y.begin(), y.end(), std::nanf(""));
sigmoid(tile_size * num_tiles * sizeof(float), x.data(), y.data());
for (uint32_t i = 0; i < tile_size * num_tiles; i++) {
const double e_ref = std::exp(double(x[i]));
const double y_ref = e_ref / (e_ref + 1.0);
const double abs_error = std::abs(y_ref - double(y[i]));
const float y_abs = std::abs(y_ref);
const float y_ulp = fp32_from_bits(fp32_to_bits(y_abs) + 1) - y_abs;
max_ulp_error = std::max<double>(max_ulp_error, abs_error / y_ulp);
}
}
for (uint32_t n = 0; n < max_input; n += tile_size * num_tiles) {
for (uint32_t i = 0; i < tile_size * num_tiles; i++) {
x[i] = fp32_from_bits(std::min<uint32_t>(n + i, max_input));
}
std::fill(y.begin(), y.end(), std::nanf(""));
sigmoid(tile_size * num_tiles * sizeof(float), x.data(), y.data());
for (uint32_t i = 0; i < tile_size * num_tiles; i++) {
const double y_ref = 1.0 / (1.0 + std::exp(-double(x[i])));
const double abs_error = std::abs(y_ref - double(y[i]));
const float y_abs = std::abs(y_ref);
const float y_ulp = fp32_from_bits(fp32_to_bits(y_abs) + 1) - y_abs;
max_ulp_error = std::max<double>(max_ulp_error, abs_error / y_ulp);
}
}
}
state.counters["ULPERROR"] = benchmark::Counter(max_ulp_error);
}
#if XNN_ARCH_ARM || XNN_ARCH_ARM64
static void f32_sigmoid__neon_frac_p9_p10_nr1recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neon_frac_p9_p10_nr1recps, 4);
}
static void f32_sigmoid__neon_rr2_lut2048_p1_nr2recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neon_rr2_lut2048_p1_nr2recps, 4);
}
static void f32_sigmoid__neonfma_rr2_lut2048_p1_nr2recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr2_lut2048_p1_nr2recps, 4);
}
static void f32_sigmoid__neonfma_rr2_lut2048_p1_nr1recps1fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr2_lut2048_p1_nr1recps1fma, 4);
}
static void f32_sigmoid__neonfma_rr2_lut2048_p1_nr2fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr2_lut2048_p1_nr2fma, 4);
}
static void f32_sigmoid__neon_rr2_lut64_p2_nr2recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neon_rr2_lut64_p2_nr2recps, 4);
}
static void f32_sigmoid__neonfma_rr2_lut64_p2_nr2recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr2_lut64_p2_nr2recps, 4);
}
static void f32_sigmoid__neonfma_rr2_lut64_p2_nr1recps1fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr2_lut64_p2_nr1recps1fma, 4);
}
static void f32_sigmoid__neonfma_rr2_lut64_p2_nr2fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr2_lut64_p2_nr2fma, 4);
}
static void f32_sigmoid__neon_rr2_p5_nr2recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neon_rr2_p5_nr2recps, 4);
}
static void f32_sigmoid__neonfma_rr2_p5_nr2recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr2_p5_nr2recps, 4);
}
static void f32_sigmoid__neonfma_rr2_p5_nr1recps1fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr2_p5_nr1recps1fma, 4);
}
static void f32_sigmoid__neonfma_rr2_p5_nr2fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr2_p5_nr2fma, 4);
}
static void f32_sigmoid__neon_rr1_lut2048_p1_nr2recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neon_rr1_lut2048_p1_nr2recps, 4);
}
static void f32_sigmoid__neonfma_rr1_lut2048_p1_nr2recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr1_lut2048_p1_nr2recps, 4);
}
static void f32_sigmoid__neonfma_rr1_lut2048_p1_nr1recps1fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr1_lut2048_p1_nr1recps1fma, 4);
}
static void f32_sigmoid__neonfma_rr1_lut2048_p1_nr2fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr1_lut2048_p1_nr2fma, 4);
}
static void f32_sigmoid__neon_rr1_lut64_p2_nr2recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neon_rr1_lut64_p2_nr2recps, 4);
}
static void f32_sigmoid__neonfma_rr1_lut64_p2_nr2recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr1_lut64_p2_nr2recps, 4);
}
static void f32_sigmoid__neonfma_rr1_lut64_p2_nr1recps1fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr1_lut64_p2_nr1recps1fma, 4);
}
static void f32_sigmoid__neonfma_rr1_lut64_p2_nr2fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr1_lut64_p2_nr2fma, 4);
}
static void f32_sigmoid__neon_rr1_p5_nr2recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neon_rr1_p5_nr2recps, 4);
}
static void f32_sigmoid__neonfma_rr1_p5_nr2recps(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr1_p5_nr2recps, 4);
}
static void f32_sigmoid__neonfma_rr1_p5_nr1recps1fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr1_p5_nr1recps1fma, 4);
}
static void f32_sigmoid__neonfma_rr1_p5_nr2fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr1_p5_nr2fma, 4);
}
BENCHMARK(f32_sigmoid__neon_frac_p9_p10_nr1recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neon_rr2_lut2048_p1_nr2recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr2_lut2048_p1_nr2recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr2_lut2048_p1_nr1recps1fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr2_lut2048_p1_nr2fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neon_rr2_lut64_p2_nr2recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr2_lut64_p2_nr2recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr2_lut64_p2_nr1recps1fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr2_lut64_p2_nr2fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neon_rr2_p5_nr2recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr2_p5_nr2recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr2_p5_nr1recps1fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr2_p5_nr2fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neon_rr1_lut2048_p1_nr2recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr1_lut2048_p1_nr2recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr1_lut2048_p1_nr1recps1fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr1_lut2048_p1_nr2fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neon_rr1_lut64_p2_nr2recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr1_lut64_p2_nr2recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr1_lut64_p2_nr1recps1fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr1_lut64_p2_nr2fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neon_rr1_p5_nr2recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr1_p5_nr2recps)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr1_p5_nr1recps1fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr1_p5_nr2fma)->Unit(benchmark::kMillisecond)->Iterations(1);
#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
#if XNN_ARCH_ARM64
static void f32_sigmoid__neonfma_rr2_lut2048_p1_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr2_lut2048_p1_div, 4);
}
static void f32_sigmoid__neonfma_rr2_lut64_p2_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr2_lut64_p2_div, 4);
}
static void f32_sigmoid__neonfma_rr2_p5_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr2_p5_div, 4);
}
static void f32_sigmoid__neonfma_rr1_lut2048_p1_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr1_lut2048_p1_div, 4);
}
static void f32_sigmoid__neonfma_rr1_lut64_p2_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr1_lut64_p2_div, 4);
}
static void f32_sigmoid__neonfma_rr1_p5_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__neonfma_rr1_p5_div, 4);
}
BENCHMARK(f32_sigmoid__neonfma_rr2_lut2048_p1_div)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr2_lut64_p2_div)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr2_p5_div)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr1_lut2048_p1_div)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr1_lut64_p2_div)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__neonfma_rr1_p5_div)->Unit(benchmark::kMillisecond)->Iterations(1);
#endif // XNN_ARCH_ARM64
#if XNN_ARCH_X86 || XNN_ARCH_X86_64
static void f32_sigmoid__avx2_rr2_p5_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__avx2_rr2_p5_div, 8);
}
static void f32_sigmoid__avx2_rr2_p5_nr2fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__avx2_rr2_p5_nr2fma, 8);
}
static void f32_sigmoid__avx2_rr2_p5_nr1fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__avx2_rr2_p5_nr1fma, 8);
}
static void f32_sigmoid__avx2_rr1_p5_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__avx2_rr1_p5_div, 8);
}
static void f32_sigmoid__avx2_rr1_p5_nr2fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__avx2_rr1_p5_nr2fma, 8);
}
static void f32_sigmoid__avx2_rr1_p5_nr1fma(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__avx2_rr1_p5_nr1fma, 8);
}
static void f32_sigmoid__sse2_p5_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__sse2_p5_div, 4);
}
BENCHMARK(f32_sigmoid__avx2_rr2_p5_div)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__avx2_rr2_p5_nr2fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__avx2_rr2_p5_nr1fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__avx2_rr1_p5_div)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__avx2_rr1_p5_nr2fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__avx2_rr1_p5_nr1fma)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__sse2_p5_div)->Unit(benchmark::kMillisecond)->Iterations(1);
#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
#if !XNN_ARCH_WASM && !XNN_ARCH_ASMJS
static void f32_sigmoid__psimd_p5_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__psimd_p5_div, 4);
}
BENCHMARK(f32_sigmoid__psimd_p5_div)->Unit(benchmark::kMillisecond)->Iterations(1);
#endif // !XNN_ARCH_WASM && !XNN_ARCH_ASMJS
static void f32_sigmoid__scalar_lut2048_p1_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__scalar_lut2048_p1_div, 1);
}
static void f32_sigmoid__scalar_lut64_p2_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__scalar_lut64_p2_div, 1);
}
static void f32_sigmoid__scalar_p5_div(benchmark::State& state) {
SigmoidError(state, xnn_math_f32_sigmoid__scalar_p5_div, 1);
}
BENCHMARK(f32_sigmoid__scalar_lut2048_p1_div)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__scalar_lut64_p2_div)->Unit(benchmark::kMillisecond)->Iterations(1);
BENCHMARK(f32_sigmoid__scalar_p5_div)->Unit(benchmark::kMillisecond)->Iterations(1);
#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
#endif
|
