blob: f182c395a57676555dd620f5bd9915dbf910badf [file] [log] [blame]
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001#include <algorithm>
2#include <cfloat>
3#include <chrono>
4#include <cmath>
5#include <functional>
6#include <random>
7#include <vector>
8
9#include "bench/utils.h"
10#include <xnnpack/AlignedAllocator.h>
11#include <xnnpack/common.h>
12#include <xnnpack/params.h>
13#include <xnnpack/raddextexp.h>
14#include <xnnpack/vscaleextexp.h>
15
16#include <benchmark/benchmark.h>
17
18
19static void f32_vscaleextexp(
20 benchmark::State& state,
21 xnn_f32_raddextexp_ukernel_function raddextexp,
22 xnn_f32_vscaleextexp_ukernel_function vscaleextexp,
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 std::vector<float, AlignedAllocator<float, 64>> y(packed_n * num_buffers);
41
42 std::generate(x.begin(), x.end(), std::ref(f32rng));
43
44 benchmark::utils::DisableDenormals();
45
46 size_t buffer_index = 0;
47 for (auto _ : state) {
48 state.PauseTiming();
49 float scale[2];
50 raddextexp(n * sizeof(float), x.data(), scale);
51 const float ext_mantissa = 1.0f / scale[0];
52 const float ext_exponent = -scale[1];
53 if (++buffer_index == num_buffers) {
54 buffer_index = 0;
55 }
56 state.ResumeTiming();
57
58 vscaleextexp(n * sizeof(float), x.data(), y.data() + packed_n * buffer_index, ext_mantissa, ext_exponent);
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()) * 2 * 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_vscaleextexp, avx512f_p5_scalef_x16,
77 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
78 xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x16,
79 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
80 BENCHMARK_CAPTURE(f32_vscaleextexp, avx512f_p5_scalef_x32,
81 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
82 xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x32,
83 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
84 BENCHMARK_CAPTURE(f32_vscaleextexp, avx512f_p5_scalef_x48,
85 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
86 xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x48,
87 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
88 BENCHMARK_CAPTURE(f32_vscaleextexp, avx512f_p5_scalef_x64,
89 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
90 xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x64,
91 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
92 BENCHMARK_CAPTURE(f32_vscaleextexp, avx512f_p5_scalef_x80,
93 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
94 xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x80,
95 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
96 BENCHMARK_CAPTURE(f32_vscaleextexp, avx512f_p5_scalef_x96,
97 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
98 xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x96,
99 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
100 BENCHMARK_CAPTURE(f32_vscaleextexp, avx512f_p5_scalef_x112,
101 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
102 xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112,
103 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
104 BENCHMARK_CAPTURE(f32_vscaleextexp, avx512f_p5_scalef_x128,
105 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
106 xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128,
107 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
108 BENCHMARK_CAPTURE(f32_vscaleextexp, avx512f_p5_scalef_x144,
109 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
110 xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144,
111 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
112 BENCHMARK_CAPTURE(f32_vscaleextexp, avx512f_p5_scalef_x160,
113 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
114 xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160,
115 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
116 BENCHMARK_CAPTURE(f32_vscaleextexp, avx512f_p5_scalef_x176,
117 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
118 xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176,
119 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
120 BENCHMARK_CAPTURE(f32_vscaleextexp, avx512f_p5_scalef_x192,
121 xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
122 xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192,
123 benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
124
125 BENCHMARK_CAPTURE(f32_vscaleextexp, avx2_p5_x8,
126 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
127 xnn_f32_vscaleextexp_ukernel__avx2_p5_x8,
128 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
129 BENCHMARK_CAPTURE(f32_vscaleextexp, avx2_p5_x16,
130 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
131 xnn_f32_vscaleextexp_ukernel__avx2_p5_x16,
132 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
133 BENCHMARK_CAPTURE(f32_vscaleextexp, avx2_p5_x24,
134 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
135 xnn_f32_vscaleextexp_ukernel__avx2_p5_x24,
136 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
137 BENCHMARK_CAPTURE(f32_vscaleextexp, avx2_p5_x32,
138 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
139 xnn_f32_vscaleextexp_ukernel__avx2_p5_x32,
140 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
141 BENCHMARK_CAPTURE(f32_vscaleextexp, avx2_p5_x40,
142 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
143 xnn_f32_vscaleextexp_ukernel__avx2_p5_x40,
144 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
145 BENCHMARK_CAPTURE(f32_vscaleextexp, avx2_p5_x48,
146 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
147 xnn_f32_vscaleextexp_ukernel__avx2_p5_x48,
148 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
149 BENCHMARK_CAPTURE(f32_vscaleextexp, avx2_p5_x56,
150 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
151 xnn_f32_vscaleextexp_ukernel__avx2_p5_x56,
152 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
153 BENCHMARK_CAPTURE(f32_vscaleextexp, avx2_p5_x64,
154 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
155 xnn_f32_vscaleextexp_ukernel__avx2_p5_x64,
156 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
157 BENCHMARK_CAPTURE(f32_vscaleextexp, avx2_p5_x72,
158 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
159 xnn_f32_vscaleextexp_ukernel__avx2_p5_x72,
160 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
161 BENCHMARK_CAPTURE(f32_vscaleextexp, avx2_p5_x80,
162 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
163 xnn_f32_vscaleextexp_ukernel__avx2_p5_x80,
164 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
165 BENCHMARK_CAPTURE(f32_vscaleextexp, avx2_p5_x88,
166 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
167 xnn_f32_vscaleextexp_ukernel__avx2_p5_x88,
168 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
169 BENCHMARK_CAPTURE(f32_vscaleextexp, avx2_p5_x96,
170 xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
171 xnn_f32_vscaleextexp_ukernel__avx2_p5_x96,
172 benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
173#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
174
175#ifndef XNNPACK_BENCHMARK_NO_MAIN
176BENCHMARK_MAIN();
177#endif