1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <algorithm>
7 #include <cfloat>
8 #include <cmath>
9 #include <functional>
10 #include <memory>
11 #include <numeric>
12 #include <random>
13 #include <vector>
14
15 #include <cpuinfo.h>
16 #include <pthreadpool.h>
17
18 #include <benchmark/benchmark.h>
19 #include <fp16/fp16.h>
20
21 #include "bench/utils.h"
22 #include <xnnpack/AlignedAllocator.h>
23 #include <xnnpack/common.h>
24 #include <xnnpack/math-stubs.h>
25
26
27 struct ComputeErrorContext {
28 const float* input;
29 const float* output;
30 float* error;
31 };
32
ComputeError(struct ComputeErrorContext * context,size_t start,size_t range)33 static void ComputeError(
34 struct ComputeErrorContext* context,
35 size_t start,
36 size_t range)
37 {
38 const float* input = context->input;
39 const float* output = context->output;
40 float* error = context->error;
41 for (size_t i = start; i < start + range; i++) {
42 const double output_ref = std::exp(double(input[i]));
43 const double abs_error = std::abs(output_ref - double(output[i]));
44 const float output_abs = std::abs(output_ref);
45 const float output_ulp = fp32_from_bits(fp32_to_bits(output_abs) + 1) - output_abs;
46 error[i] = float(abs_error / output_ulp);
47 }
48 }
49
ExpError(benchmark::State & state,xnn_f32_unary_math_function exp,benchmark::utils::IsaCheckFunction isa_check=nullptr)50 static void ExpError(
51 benchmark::State& state,
52 xnn_f32_unary_math_function exp,
53 benchmark::utils::IsaCheckFunction isa_check = nullptr)
54 {
55 if (!cpuinfo_initialize()) {
56 state.SkipWithError("failed cpuinfo init");
57 return;
58 }
59 if (isa_check && !isa_check(state)) {
60 return;
61 }
62
63 // The smallest x for which expf(x) is non-zero (-0x1.9FE368p+6f).
64 const uint32_t min_input = 0xC2CFF1B4;
65 // The largest x for which expf(x) is finite (0x1.62E42Ep6f).
66 const uint32_t max_input = 0x42B17217;
67 // Number of elements in one block of inputs/outputs.
68 // Combining multiple elements in a block reduce function call overhead.
69 const size_t block_size = 16384;
70 // Number of elements in one parallelization tile. Worker threads process this many elements in each task.
71 const size_t tile_size = 64;
72
73 uint32_t num_threads = cpuinfo_get_cores_count();
74 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
75 // Use all cores except for the least performant cluster
76 if (cpuinfo_get_clusters_count() > 1) {
77 num_threads -= cpuinfo_get_cluster(cpuinfo_get_clusters_count() - 1)->core_count;
78 }
79 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
80
81 std::unique_ptr<pthreadpool, decltype(&pthreadpool_destroy)> threadpool(
82 pthreadpool_create(num_threads), pthreadpool_destroy);
83
84 std::vector<float, AlignedAllocator<float, 64>> x(block_size);
85 std::vector<float, AlignedAllocator<float, 64>> y(block_size);
86 std::vector<float> ulp_error(block_size);
87 float max_ulp_error = 0.0f;
88
89 ComputeErrorContext context;
90 context.input = x.data();
91 context.output = y.data();
92 context.error = ulp_error.data();
93 for (auto _ : state) {
94 for (uint32_t n = min_input; int32_t(n) < 0; n -= block_size) {
95 for (uint32_t i = 0; i < block_size; i++) {
96 x[i] = fp32_from_bits(std::max<uint32_t>(n - i, 0x80000000));
97 }
98 std::fill(y.begin(), y.end(), std::nanf(""));
99
100 exp(block_size * sizeof(float), x.data(), y.data());
101
102 pthreadpool_parallelize_1d_tile_1d(
103 threadpool.get(),
104 reinterpret_cast<pthreadpool_task_1d_tile_1d_t>(ComputeError),
105 static_cast<void*>(&context),
106 block_size, tile_size, 0 /* flags */);
107
108 max_ulp_error = std::accumulate(ulp_error.cbegin(), ulp_error.cend(), max_ulp_error,
109 static_cast<const float& (*)(const float&, const float&)>(std::max<float>));
110 }
111 for (uint32_t n = 0; n < max_input; n += block_size) {
112 for (uint32_t i = 0; i < block_size; i++) {
113 x[i] = fp32_from_bits(std::min<uint32_t>(n + i, max_input));
114 }
115 std::fill(y.begin(), y.end(), std::nanf(""));
116
117 exp(block_size * sizeof(float), x.data(), y.data());
118
119 pthreadpool_parallelize_1d_tile_1d(
120 threadpool.get(),
121 reinterpret_cast<pthreadpool_task_1d_tile_1d_t>(ComputeError),
122 static_cast<void*>(&context),
123 block_size, tile_size, 0 /* flags */);
124
125 max_ulp_error = std::accumulate(ulp_error.cbegin(), ulp_error.cend(), max_ulp_error,
126 static_cast<const float& (*)(const float&, const float&)>(std::max<float>));
127 }
128 }
129
130 state.counters["ULPERROR"] = benchmark::Counter(max_ulp_error);
131 }
132
133 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
134 BENCHMARK_CAPTURE(ExpError, neonfma_rr2_lut64_p2,
135 xnn_math_f32_exp__neonfma_rr2_lut64_p2,
136 benchmark::utils::CheckNEONFMA)
137 ->Unit(benchmark::kMillisecond)
138 ->Iterations(1);
139 BENCHMARK_CAPTURE(ExpError, neonfma_rr2_p5,
140 xnn_math_f32_exp__neonfma_rr2_p5,
141 benchmark::utils::CheckNEONFMA)
142 ->Unit(benchmark::kMillisecond)
143 ->Iterations(1);
144 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
145
146 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
147 BENCHMARK_CAPTURE(ExpError, avx512f_rr2_lut16_p3_perm,
148 xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm,
149 benchmark::utils::CheckAVX512F)
150 ->Unit(benchmark::kMillisecond)
151 ->Iterations(1);
152 BENCHMARK_CAPTURE(ExpError, avx512f_rr2_lut16_p3_perm_scalef,
153 xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm_scalef,
154 benchmark::utils::CheckAVX512F)
155 ->Unit(benchmark::kMillisecond)
156 ->Iterations(1);
157 BENCHMARK_CAPTURE(ExpError, avx512f_rr2_lut32_p2_perm2,
158 xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2,
159 benchmark::utils::CheckAVX512F)
160 ->Unit(benchmark::kMillisecond)
161 ->Iterations(1);
162 BENCHMARK_CAPTURE(ExpError, avx512f_rr2_lut32_p2_perm2_scalef,
163 xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2_scalef,
164 benchmark::utils::CheckAVX512F)
165 ->Unit(benchmark::kMillisecond)
166 ->Iterations(1);
167 BENCHMARK_CAPTURE(ExpError, avx512f_rr2_p5,
168 xnn_math_f32_exp__avx512f_rr2_p5,
169 benchmark::utils::CheckAVX512F)
170 ->Unit(benchmark::kMillisecond)
171 ->Iterations(1);
172 BENCHMARK_CAPTURE(ExpError, avx512f_rr2_p5_scalef,
173 xnn_math_f32_exp__avx512f_rr2_p5_scalef,
174 benchmark::utils::CheckAVX512F)
175 ->Unit(benchmark::kMillisecond)
176 ->Iterations(1);
177
178 BENCHMARK_CAPTURE(ExpError, avx2_rr2_lut8_p3_perm,
179 xnn_math_f32_exp__avx2_rr2_lut8_p3_perm,
180 benchmark::utils::CheckAVX2)
181 ->Unit(benchmark::kMillisecond)
182 ->Iterations(1);
183 BENCHMARK_CAPTURE(ExpError, avx2_rr2_lut8_p4_perm,
184 xnn_math_f32_exp__avx2_rr2_lut8_p4_perm,
185 benchmark::utils::CheckAVX2)
186 ->Unit(benchmark::kMillisecond)
187 ->Iterations(1);
188 BENCHMARK_CAPTURE(ExpError, avx2_rr2_p5,
189 xnn_math_f32_exp__avx2_rr2_p5,
190 benchmark::utils::CheckAVX2)
191 ->Unit(benchmark::kMillisecond)
192 ->Iterations(1);
193
194 BENCHMARK_CAPTURE(ExpError, avx_rr2_p5,
195 xnn_math_f32_exp__avx_rr2_p5,
196 benchmark::utils::CheckAVX)
197 ->Unit(benchmark::kMillisecond)
198 ->Iterations(1);
199
200 BENCHMARK_CAPTURE(ExpError, sse2_rr2_lut64_p2,
201 xnn_math_f32_exp__sse2_rr2_lut64_p2)
202 ->Unit(benchmark::kMillisecond)
203 ->Iterations(1);
204 BENCHMARK_CAPTURE(ExpError, sse2_rr2_p5,
205 xnn_math_f32_exp__sse2_rr2_p5)
206 ->Unit(benchmark::kMillisecond)
207 ->Iterations(1);
208 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
209
210 #ifndef XNNPACK_BENCHMARK_NO_MAIN
211 BENCHMARK_MAIN();
212 #endif
213