1 // Copyright 2021 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
15 #include <xnnpack.h>
16 #include <xnnpack/aligned-allocator.h>
17 #include <xnnpack/common.h>
18 #include <xnnpack/microfnptr.h>
19 #include <xnnpack/microparams-init.h>
20 #include <xnnpack/vunary.h>
21
22
f32_vlrelu(benchmark::State & state,xnn_f32_vlrelu_ukernel_function vlrelu,xnn_init_f32_lrelu_params_fn init_params,benchmark::utils::IsaCheckFunction isa_check=nullptr)23 static void f32_vlrelu(
24 benchmark::State& state,
25 xnn_f32_vlrelu_ukernel_function vlrelu,
26 xnn_init_f32_lrelu_params_fn init_params,
27 benchmark::utils::IsaCheckFunction isa_check = nullptr)
28 {
29 if (isa_check && !isa_check(state)) {
30 return;
31 }
32
33 const size_t elements = state.range(0);
34 std::vector<float, AlignedAllocator<float, 64>> input(elements);
35 std::vector<float, AlignedAllocator<float, 64>> output(elements);
36
37 std::random_device random_device;
38 auto rng = std::mt19937(random_device());
39 auto f32rng = std::bind(std::uniform_real_distribution<float>(-5.0f, 5.0f), std::ref(rng));
40 std::generate(input.begin(), input.end(), std::ref(f32rng));
41 std::fill(output.begin(), output.end(), std::nanf(""));
42
43 union xnn_f32_lrelu_params params;
44 init_params(¶ms, 0.01f);
45 for (auto _ : state) {
46 vlrelu(elements * sizeof(float), input.data(), output.data(), ¶ms);
47 }
48
49 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
50 if (cpu_frequency != 0) {
51 state.counters["cpufreq"] = cpu_frequency;
52 }
53
54 const size_t elements_per_iteration = elements;
55 state.counters["elements"] =
56 benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);
57
58 const size_t bytes_per_iteration = 2 * elements * sizeof(float);
59 state.counters["bytes"] =
60 benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
61 }
62
63 #if XNN_ARCH_ARM64 || XNN_ARCH_ARM64
64 BENCHMARK_CAPTURE(f32_vlrelu, neon_x4,
65 xnn_f32_vlrelu_ukernel__neon_x4,
66 xnn_init_f32_lrelu_scalar_params,
67 benchmark::utils::CheckNEON)
68 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
69 ->UseRealTime();
70 BENCHMARK_CAPTURE(f32_vlrelu, neon_x8,
71 xnn_f32_vlrelu_ukernel__neon_x8,
72 xnn_init_f32_lrelu_scalar_params,
73 benchmark::utils::CheckNEON)
74 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
75 ->UseRealTime();
76 #endif // XNN_ARCH_ARM64 || XNN_ARCH_ARM64
77
78 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
79 BENCHMARK_CAPTURE(f32_vlrelu, sse_x4,
80 xnn_f32_vlrelu_ukernel__sse_x4,
81 xnn_init_f32_lrelu_sse_params)
82 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
83 ->UseRealTime();
84 BENCHMARK_CAPTURE(f32_vlrelu, sse_x8,
85 xnn_f32_vlrelu_ukernel__sse_x8,
86 xnn_init_f32_lrelu_sse_params)
87 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
88 ->UseRealTime();
89
90 BENCHMARK_CAPTURE(f32_vlrelu, sse2_x4,
91 xnn_f32_vlrelu_ukernel__sse2_x4,
92 xnn_init_f32_lrelu_sse_params)
93 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
94 ->UseRealTime();
95 BENCHMARK_CAPTURE(f32_vlrelu, sse2_x8,
96 xnn_f32_vlrelu_ukernel__sse2_x8,
97 xnn_init_f32_lrelu_sse_params)
98 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
99 ->UseRealTime();
100
101 BENCHMARK_CAPTURE(f32_vlrelu, sse41_x4,
102 xnn_f32_vlrelu_ukernel__sse41_x4,
103 xnn_init_f32_lrelu_sse_params,
104 benchmark::utils::CheckSSE41)
105 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
106 ->UseRealTime();
107 BENCHMARK_CAPTURE(f32_vlrelu, sse41_x8,
108 xnn_f32_vlrelu_ukernel__sse41_x8,
109 xnn_init_f32_lrelu_sse_params,
110 benchmark::utils::CheckSSE41)
111 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
112 ->UseRealTime();
113
114 BENCHMARK_CAPTURE(f32_vlrelu, avx_x8,
115 xnn_f32_vlrelu_ukernel__avx_x8,
116 xnn_init_f32_lrelu_avx_params,
117 benchmark::utils::CheckAVX)
118 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
119 ->UseRealTime();
120 BENCHMARK_CAPTURE(f32_vlrelu, avx_x16,
121 xnn_f32_vlrelu_ukernel__avx_x16,
122 xnn_init_f32_lrelu_avx_params,
123 benchmark::utils::CheckAVX)
124 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
125 ->UseRealTime();
126
127 BENCHMARK_CAPTURE(f32_vlrelu, avx512f_x16,
128 xnn_f32_vlrelu_ukernel__avx512f_x16,
129 xnn_init_f32_lrelu_scalar_params,
130 benchmark::utils::CheckAVX512F)
131 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
132 ->UseRealTime();
133 BENCHMARK_CAPTURE(f32_vlrelu, avx512f_x32,
134 xnn_f32_vlrelu_ukernel__avx512f_x32,
135 xnn_init_f32_lrelu_scalar_params,
136 benchmark::utils::CheckAVX512F)
137 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
138 ->UseRealTime();
139 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
140
141 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
142 BENCHMARK_CAPTURE(f32_vlrelu, wasmsimd_bitselect_x4,
143 xnn_f32_vlrelu_ukernel__wasmsimd_bitselect_x4,
144 xnn_init_f32_lrelu_wasmsimd_params)
145 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
146 ->UseRealTime();
147 BENCHMARK_CAPTURE(f32_vlrelu, wasmsimd_bitselect_x8,
148 xnn_f32_vlrelu_ukernel__wasmsimd_bitselect_x8,
149 xnn_init_f32_lrelu_wasmsimd_params)
150 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
151 ->UseRealTime();
152
153 BENCHMARK_CAPTURE(f32_vlrelu, wasmsimd_minmax_x4,
154 xnn_f32_vlrelu_ukernel__wasmsimd_minmax_x4,
155 xnn_init_f32_lrelu_wasmsimd_params)
156 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
157 ->UseRealTime();
158 BENCHMARK_CAPTURE(f32_vlrelu, wasmsimd_minmax_x8,
159 xnn_f32_vlrelu_ukernel__wasmsimd_minmax_x8,
160 xnn_init_f32_lrelu_wasmsimd_params)
161 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
162 ->UseRealTime();
163 #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
164
165 #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
166 BENCHMARK_CAPTURE(f32_vlrelu, wasm_x1,
167 xnn_f32_vlrelu_ukernel__wasm_x1,
168 xnn_init_f32_lrelu_scalar_params)
169 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
170 ->UseRealTime();
171 BENCHMARK_CAPTURE(f32_vlrelu, wasm_x2,
172 xnn_f32_vlrelu_ukernel__wasm_x2,
173 xnn_init_f32_lrelu_scalar_params)
174 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
175 ->UseRealTime();
176 BENCHMARK_CAPTURE(f32_vlrelu, wasm_x4,
177 xnn_f32_vlrelu_ukernel__wasm_x4,
178 xnn_init_f32_lrelu_scalar_params)
179 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
180 ->UseRealTime();
181 #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
182
183 BENCHMARK_CAPTURE(f32_vlrelu, scalar_x1,
184 xnn_f32_vlrelu_ukernel__scalar_x1,
185 xnn_init_f32_lrelu_scalar_params)
186 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
187 ->UseRealTime();
188 BENCHMARK_CAPTURE(f32_vlrelu, scalar_x2,
189 xnn_f32_vlrelu_ukernel__scalar_x2,
190 xnn_init_f32_lrelu_scalar_params)
191 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
192 ->UseRealTime();
193 BENCHMARK_CAPTURE(f32_vlrelu, scalar_x4,
194 xnn_f32_vlrelu_ukernel__scalar_x4,
195 xnn_init_f32_lrelu_scalar_params)
196 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
197 ->UseRealTime();
198
199 #ifndef XNNPACK_BENCHMARK_NO_MAIN
200 BENCHMARK_MAIN();
201 #endif
202