blob: 6196e0125c75f71b9d806dd36646fb1cdf898b66 [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/raddextexp.h>
18
19
20static void f32_raddextexp(
21 benchmark::State& state,
22 xnn_f32_raddextexp_ukernel_function raddextexp,
23 benchmark::utils::IsaCheckFunction isa_check = nullptr)
24{
25 if (isa_check && !isa_check(state)) {
26 return;
27 }
28
Marat Dukhand713e8a2020-12-04 14:23:12 -080029 const size_t elements = state.range(0);
Marat Dukhan4c4eb002019-12-08 21:27:49 -080030 const size_t cache_line_size_max = 128;
Marat Dukhand713e8a2020-12-04 14:23:12 -080031 const size_t packed_elements = benchmark::utils::RoundUp(elements, cache_line_size_max / sizeof(float));
Marat Dukhan4c4eb002019-12-08 21:27:49 -080032
33 std::random_device random_device;
34 auto rng = std::mt19937(random_device());
Marat Dukhan44f0ca72020-08-02 21:46:58 -070035 auto f32rng = std::bind(std::uniform_real_distribution<float>(-1000.0f, 1000.0f), std::ref(rng));
Marat Dukhan4c4eb002019-12-08 21:27:49 -080036
37 const size_t num_buffers = 1 +
Marat Dukhand713e8a2020-12-04 14:23:12 -080038 benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), packed_elements * sizeof(float));
39 std::vector<float, AlignedAllocator<float, 64>> x(elements);
Marat Dukhan4c4eb002019-12-08 21:27:49 -080040
41 std::generate(x.begin(), x.end(), std::ref(f32rng));
42
43 benchmark::utils::DisableDenormals();
44
45 size_t buffer_index = 0;
46 for (auto _ : state) {
47 if (++buffer_index == num_buffers) {
48 buffer_index = 0;
49 }
50
51 float y_sum[2] = { nanf(""), nanf("") };
Marat Dukhand713e8a2020-12-04 14:23:12 -080052 raddextexp(elements * sizeof(float), x.data(), y_sum);
Marat Dukhan4c4eb002019-12-08 21:27:49 -080053 }
54
Marat Dukhand713e8a2020-12-04 14:23:12 -080055 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
56 if (cpu_frequency != 0) {
57 state.counters["cpufreq"] = cpu_frequency;
58 }
59
60 const size_t elements_per_iteration = elements;
Marat Dukhan4c4eb002019-12-08 21:27:49 -080061 state.counters["elements"] =
Marat Dukhand713e8a2020-12-04 14:23:12 -080062 benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);
63
64 const size_t bytes_per_iteration = 2 * elements * sizeof(float);
Marat Dukhan4c4eb002019-12-08 21:27:49 -080065 state.counters["bytes"] =
Marat Dukhand713e8a2020-12-04 14:23:12 -080066 benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
Marat Dukhan4c4eb002019-12-08 21:27:49 -080067}
68
69static void CharacteristicArguments(benchmark::internal::Benchmark* b) {
70 b->ArgName("N");
71 for (int32_t n = 10000; n <= 100000000; n *= 10) {
72 b->Arg(n);
73 }
74}
75
76#if XNN_ARCH_X86 || XNN_ARCH_X86_64
77 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x128,
78 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128,
79 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
80 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x128_acc2,
81 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
82 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
83 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x128_acc4,
84 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4,
85 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
86
87 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x144,
88 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144,
89 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
90 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x144_acc3,
91 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3,
92 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
93
94 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x160,
95 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160,
96 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
97 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x160_acc2,
98 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2,
99 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
100 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x160_acc5,
101 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5,
102 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
103
104 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x192,
105 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192,
106 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
107 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x192_acc2,
108 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2,
109 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
110 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x192_acc3,
111 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3,
112 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
113 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x192_acc6,
114 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6,
115 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
116
117 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x64,
118 xnn_f32_raddextexp_ukernel__avx2_p5_x64,
119 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
120 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x64_acc2,
121 xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2,
122 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
123 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x64_acc4,
124 xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4,
125 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
126
127 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x72,
128 xnn_f32_raddextexp_ukernel__avx2_p5_x72,
129 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
130 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x72_acc3,
131 xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3,
132 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
133
134 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x80,
135 xnn_f32_raddextexp_ukernel__avx2_p5_x80,
136 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
137 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x80_acc2,
138 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
139 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
140 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x80_acc5,
141 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5,
142 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
143
144 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x96,
145 xnn_f32_raddextexp_ukernel__avx2_p5_x96,
146 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
147 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x96_acc2,
148 xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2,
149 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
150 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x96_acc3,
151 xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3,
152 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
153 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x96_acc6,
154 xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6,
155 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
156#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
157
158#ifndef XNNPACK_BENCHMARK_NO_MAIN
159BENCHMARK_MAIN();
160#endif