1 // Copyright 2020 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 <cmath>
8 #include <functional>
9 #include <random>
10 #include <vector>
11
12 #include <benchmark/benchmark.h>
13 #include "bench/utils.h"
14 #include <xnnpack/AlignedAllocator.h>
15 #include <xnnpack/common.h>
16 #include <xnnpack/params.h>
17 #include <xnnpack/vunary.h>
18
19
f32_relu(benchmark::State & state,xnn_f32_relu_ukernel_function f32_relu,benchmark::utils::IsaCheckFunction isa_check=nullptr)20 static void f32_relu(
21 benchmark::State& state,
22 xnn_f32_relu_ukernel_function f32_relu,
23 benchmark::utils::IsaCheckFunction isa_check = nullptr)
24 {
25 if (isa_check && !isa_check(state)) {
26 return;
27 }
28
29 const size_t elements = state.range(0);
30
31 std::random_device random_device;
32 auto rng = std::mt19937(random_device());
33 auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng));
34
35 std::vector<float, AlignedAllocator<float, 64>> x(elements);
36 std::generate(x.begin(), x.end(), std::ref(f32rng));
37 std::vector<float, AlignedAllocator<float, 64>> y(elements);
38 std::generate(x.begin(), x.end(), std::ref(f32rng));
39
40 for (auto _ : state) {
41 f32_relu(elements * sizeof(float), x.data(), y.data(), NULL);
42 }
43
44 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
45 if (cpu_frequency != 0) {
46 state.counters["cpufreq"] = cpu_frequency;
47 }
48
49 const size_t elements_per_iteration = elements;
50 state.counters["elements"] =
51 benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);
52
53 const size_t bytes_per_iteration = 2 * elements * sizeof(float);
54 state.counters["bytes"] =
55 benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
56 }
57
58 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
59 BENCHMARK_CAPTURE(f32_relu, sse_x4, xnn_f32_relu_ukernel__sse_x4)
60 ->RangeMultiplier(10)
61 ->Range(1000, 100000000)
62 ->UseRealTime();
63
64 BENCHMARK_CAPTURE(f32_relu, sse_x8, xnn_f32_relu_ukernel__sse_x8)
65 ->RangeMultiplier(10)
66 ->Range(1000, 100000000)
67 ->UseRealTime();
68
69 BENCHMARK_CAPTURE(f32_relu, avx_x8, xnn_f32_relu_ukernel__avx_x8, benchmark::utils::CheckAVX)
70 ->RangeMultiplier(10)
71 ->Range(1000, 100000000)
72 ->UseRealTime();
73
74 BENCHMARK_CAPTURE(f32_relu, avx_x16, xnn_f32_relu_ukernel__avx_x16, benchmark::utils::CheckAVX)
75 ->RangeMultiplier(10)
76 ->Range(1000, 100000000)
77 ->UseRealTime();
78
79 BENCHMARK_CAPTURE(f32_relu, avx512f_x16, xnn_f32_relu_ukernel__avx512f_x16, benchmark::utils::CheckAVX512F)
80 ->RangeMultiplier(10)
81 ->Range(1000, 100000000)
82 ->UseRealTime();
83
84 BENCHMARK_CAPTURE(f32_relu, avx512f_x32, xnn_f32_relu_ukernel__avx512f_x32, benchmark::utils::CheckAVX512F)
85 ->RangeMultiplier(10)
86 ->Range(1000, 100000000)
87 ->UseRealTime();
88 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
89
90 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
91 BENCHMARK_CAPTURE(f32_relu, neon_x4, xnn_f32_relu_ukernel__neon_x4, benchmark::utils::CheckNEON)
92 ->RangeMultiplier(10)
93 ->Range(1000, 100000000)
94 ->UseRealTime();
95 BENCHMARK_CAPTURE(f32_relu, neon_x8, xnn_f32_relu_ukernel__neon_x8, benchmark::utils::CheckNEON)
96 ->RangeMultiplier(10)
97 ->Range(1000, 100000000)
98 ->UseRealTime();
99 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
100
101
102 #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD
103 BENCHMARK_CAPTURE(f32_relu, wasm_x1, xnn_f32_relu_ukernel__wasm_x1)
104 ->RangeMultiplier(10)
105 ->Range(1000, 100000000)
106 ->UseRealTime();
107
108 BENCHMARK_CAPTURE(f32_relu, wasm_x2, xnn_f32_relu_ukernel__wasm_x2)
109 ->RangeMultiplier(10)
110 ->Range(1000, 100000000)
111 ->UseRealTime();
112
113 BENCHMARK_CAPTURE(f32_relu, wasm_x4, xnn_f32_relu_ukernel__wasm_x4)
114 ->RangeMultiplier(10)
115 ->Range(1000, 100000000)
116 ->UseRealTime();
117
118 BENCHMARK_CAPTURE(f32_relu, wasm_x8, xnn_f32_relu_ukernel__wasm_x8)
119 ->RangeMultiplier(10)
120 ->Range(1000, 100000000)
121 ->UseRealTime();
122
123 BENCHMARK_CAPTURE(f32_relu, wasm32_shr_x1, xnn_f32_relu_ukernel__wasm32_shr_x1)
124 ->RangeMultiplier(10)
125 ->Range(1000, 100000000)
126 ->UseRealTime();
127
128 BENCHMARK_CAPTURE(f32_relu, wasm32_shr_x2, xnn_f32_relu_ukernel__wasm32_shr_x2)
129 ->RangeMultiplier(10)
130 ->Range(1000, 100000000)
131 ->UseRealTime();
132
133 BENCHMARK_CAPTURE(f32_relu, wasm32_shr_x4, xnn_f32_relu_ukernel__wasm32_shr_x4)
134 ->RangeMultiplier(10)
135 ->Range(1000, 100000000)
136 ->UseRealTime();
137 #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD
138
139 #if XNN_ARCH_WASMSIMD
140 BENCHMARK_CAPTURE(f32_relu, wasmsimd_x4, xnn_f32_relu_ukernel__wasmsimd_x4)
141 ->RangeMultiplier(10)
142 ->Range(1000, 100000000)
143 ->UseRealTime();
144
145 BENCHMARK_CAPTURE(f32_relu, wasmsimd_x8, xnn_f32_relu_ukernel__wasmsimd_x8)
146 ->RangeMultiplier(10)
147 ->Range(1000, 100000000)
148 ->UseRealTime();
149
150 BENCHMARK_CAPTURE(f32_relu, wasmsimd_x16, xnn_f32_relu_ukernel__wasmsimd_x16)
151 ->RangeMultiplier(10)
152 ->Range(1000, 100000000)
153 ->UseRealTime();
154 #endif // XNN_ARCH_WASMSIMD
155
156 BENCHMARK_CAPTURE(f32_relu, scalar_x1, xnn_f32_relu_ukernel__scalar_x1)
157 ->RangeMultiplier(10)
158 ->Range(1000, 100000000)
159 ->UseRealTime();
160
161 BENCHMARK_CAPTURE(f32_relu, scalar_x2, xnn_f32_relu_ukernel__scalar_x2)
162 ->RangeMultiplier(10)
163 ->Range(1000, 100000000)
164 ->UseRealTime();
165
166 BENCHMARK_CAPTURE(f32_relu, scalar_x4, xnn_f32_relu_ukernel__scalar_x4)
167 ->RangeMultiplier(10)
168 ->Range(1000, 100000000)
169 ->UseRealTime();
170
171 BENCHMARK_CAPTURE(f32_relu, scalar_x8, xnn_f32_relu_ukernel__scalar_x8)
172 ->RangeMultiplier(10)
173 ->Range(1000, 100000000)
174 ->UseRealTime();
175
176 #ifndef XNNPACK_BENCHMARK_NO_MAIN
177 BENCHMARK_MAIN();
178 #endif
179