blob: 38ce344fcc1d36bac873899c13d6bee4c2d2b29c [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
29 const size_t n = state.range(0);
30 const size_t cache_line_size_max = 128;
31 const size_t packed_n = benchmark::utils::RoundUp(n, cache_line_size_max / sizeof(float));
32
33 std::random_device random_device;
34 auto rng = std::mt19937(random_device());
35 auto f32rng = std::bind(std::uniform_real_distribution<float>(-1000.0f, 1000.0f), rng);
36
37 const size_t num_buffers = 1 +
38 benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), packed_n * sizeof(float));
39 std::vector<float, AlignedAllocator<float, 64>> x(n);
40
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("") };
52 raddextexp(n * sizeof(float), x.data(), y_sum);
53 }
54
55 state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency();
56 state.counters["elements"] =
57 benchmark::Counter(uint64_t(state.iterations()) * n, benchmark::Counter::kIsRate);
58 state.counters["bytes"] =
59 benchmark::Counter(uint64_t(state.iterations()) * sizeof(float) * n, benchmark::Counter::kIsRate);
60}
61
62static void CharacteristicArguments(benchmark::internal::Benchmark* b) {
63 b->ArgName("N");
64 for (int32_t n = 10000; n <= 100000000; n *= 10) {
65 b->Arg(n);
66 }
67}
68
69#if XNN_ARCH_X86 || XNN_ARCH_X86_64
70 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x128,
71 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128,
72 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
73 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x128_acc2,
74 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
75 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
76 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x128_acc4,
77 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4,
78 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
79
80 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x144,
81 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144,
82 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
83 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x144_acc3,
84 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3,
85 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
86
87 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x160,
88 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160,
89 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
90 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x160_acc2,
91 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2,
92 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
93 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x160_acc5,
94 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5,
95 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
96
97 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x192,
98 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192,
99 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
100 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x192_acc2,
101 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2,
102 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
103 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x192_acc3,
104 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3,
105 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
106 BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x192_acc6,
107 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6,
108 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
109
110 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x64,
111 xnn_f32_raddextexp_ukernel__avx2_p5_x64,
112 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
113 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x64_acc2,
114 xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2,
115 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
116 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x64_acc4,
117 xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4,
118 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
119
120 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x72,
121 xnn_f32_raddextexp_ukernel__avx2_p5_x72,
122 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
123 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x72_acc3,
124 xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3,
125 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
126
127 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x80,
128 xnn_f32_raddextexp_ukernel__avx2_p5_x80,
129 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
130 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x80_acc2,
131 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
132 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
133 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x80_acc5,
134 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5,
135 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
136
137 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x96,
138 xnn_f32_raddextexp_ukernel__avx2_p5_x96,
139 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
140 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x96_acc2,
141 xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2,
142 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
143 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x96_acc3,
144 xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3,
145 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
146 BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x96_acc6,
147 xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6,
148 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
149#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
150
151#ifndef XNNPACK_BENCHMARK_NO_MAIN
152BENCHMARK_MAIN();
153#endif