blob: e7c7470cd22fcb9645e6f0f675196f6d3fb32ad3 [file] [log] [blame]
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001// 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 <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/raddexpminusmax.h>
18#include <xnnpack/rmax.h>
19
20
21static void f32_raddexpminusmax(
22 benchmark::State& state,
23 xnn_f32_rmax_ukernel_function rmax,
24 xnn_f32_raddexpminusmax_ukernel_function raddexpminusmax,
25 benchmark::utils::IsaCheckFunction isa_check = nullptr)
26{
27 if (isa_check && !isa_check(state)) {
28 return;
29 }
30
31 const size_t n = state.range(0);
32 const size_t cache_line_size_max = 128;
33 const size_t packed_n = benchmark::utils::RoundUp(n, cache_line_size_max / sizeof(float));
34
35 std::random_device random_device;
36 auto rng = std::mt19937(random_device());
37 auto f32rng = std::bind(std::uniform_real_distribution<float>(-1000.0f, 1000.0f), rng);
38
39 const size_t num_buffers = 1 +
40 benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), packed_n * sizeof(float));
41 std::vector<float, AlignedAllocator<float, 64>> x(n);
42
43 std::generate(x.begin(), x.end(), std::ref(f32rng));
44
45 benchmark::utils::DisableDenormals();
46
47 size_t buffer_index = 0;
48 for (auto _ : state) {
49 state.PauseTiming();
50 float x_max = nanf("");
51 rmax(n * sizeof(float), x.data(), &x_max);
52 if (++buffer_index == num_buffers) {
53 buffer_index = 0;
54 }
55 state.ResumeTiming();
56
57 float y_sum = nanf("");
58 raddexpminusmax(n * sizeof(float), x.data(), &y_sum, x_max);
59 }
60
61 state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency();
62 state.counters["elements"] =
63 benchmark::Counter(uint64_t(state.iterations()) * n, benchmark::Counter::kIsRate);
64 state.counters["bytes"] =
65 benchmark::Counter(uint64_t(state.iterations()) * sizeof(float) * n, benchmark::Counter::kIsRate);
66}
67
68static void CharacteristicArguments(benchmark::internal::Benchmark* b) {
69 b->ArgName("N");
70 for (int32_t n = 10000; n <= 100000000; n *= 10) {
71 b->Arg(n);
72 }
73}
74
75#if XNN_ARCH_X86 || XNN_ARCH_X86_64
76 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx2_p5_x64,
77 xnn_f32_rmax_ukernel__avx,
78 xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64,
79 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
80 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx2_p5_x64_acc2,
81 xnn_f32_rmax_ukernel__avx,
82 xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2,
83 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
84 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx2_p5_x64_acc4,
85 xnn_f32_rmax_ukernel__avx,
86 xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4,
87 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
88
89 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx2_p5_x72,
90 xnn_f32_rmax_ukernel__avx,
91 xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72,
92 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
93 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx2_p5_x72_acc3,
94 xnn_f32_rmax_ukernel__avx,
95 xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3,
96 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
97
98 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx2_p5_x80,
99 xnn_f32_rmax_ukernel__avx,
100 xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80,
101 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
102 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx2_p5_x80_acc2,
103 xnn_f32_rmax_ukernel__avx,
104 xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
105 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
106 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx2_p5_x80_acc5,
107 xnn_f32_rmax_ukernel__avx,
108 xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc5,
109 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
110
111 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx2_p5_x96,
112 xnn_f32_rmax_ukernel__avx,
113 xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96,
114 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
115 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx2_p5_x96_acc2,
116 xnn_f32_rmax_ukernel__avx,
117 xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2,
118 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
119 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx2_p5_x96_acc3,
120 xnn_f32_rmax_ukernel__avx,
121 xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3,
122 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
123 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx2_p5_x96_acc6,
124 xnn_f32_rmax_ukernel__avx,
125 xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6,
126 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
127
128 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx512f_p5_scalef_x128,
129 xnn_f32_rmax_ukernel__avx,
130 xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128,
131 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
132 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx512f_p5_scalef_x128_acc2,
133 xnn_f32_rmax_ukernel__avx,
134 xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
135 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
136 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx512f_p5_scalef_x128_acc4,
137 xnn_f32_rmax_ukernel__avx,
138 xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4,
139 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
140
141 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx512f_p5_scalef_x144,
142 xnn_f32_rmax_ukernel__avx,
143 xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144,
144 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
145 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx512f_p5_scalef_x144_acc3,
146 xnn_f32_rmax_ukernel__avx,
147 xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3,
148 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
149
150 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx512f_p5_scalef_x160,
151 xnn_f32_rmax_ukernel__avx,
152 xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160,
153 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
154 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx512f_p5_scalef_x160_acc2,
155 xnn_f32_rmax_ukernel__avx,
156 xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2,
157 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
158 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx512f_p5_scalef_x160_acc5,
159 xnn_f32_rmax_ukernel__avx,
160 xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5,
161 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
162
163 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx512f_p5_scalef_x192,
164 xnn_f32_rmax_ukernel__avx,
165 xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192,
166 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
167 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx512f_p5_scalef_x192_acc2,
168 xnn_f32_rmax_ukernel__avx,
169 xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2,
170 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
171 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx512f_p5_scalef_x192_acc3,
172 xnn_f32_rmax_ukernel__avx,
173 xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3,
174 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
175 BENCHMARK_CAPTURE(f32_raddexpminusmax, avx512f_p5_scalef_x192_acc6,
176 xnn_f32_rmax_ukernel__avx,
177 xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6,
178 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
179#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
180
181#ifndef XNNPACK_BENCHMARK_NO_MAIN
182BENCHMARK_MAIN();
183#endif