xref: /aosp_15_r20/external/XNNPACK/bench/f32-raddextexp.cc (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Copyright 2019 Google LLC
2*4bdc9457SAndroid Build Coastguard Worker //
3*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
4*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
5*4bdc9457SAndroid Build Coastguard Worker 
6*4bdc9457SAndroid Build Coastguard Worker #include <algorithm>
7*4bdc9457SAndroid Build Coastguard Worker #include <cmath>
8*4bdc9457SAndroid Build Coastguard Worker #include <functional>
9*4bdc9457SAndroid Build Coastguard Worker #include <random>
10*4bdc9457SAndroid Build Coastguard Worker #include <vector>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker #include <benchmark/benchmark.h>
13*4bdc9457SAndroid Build Coastguard Worker #include "bench/utils.h"
14*4bdc9457SAndroid Build Coastguard Worker 
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack.h>
16*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/aligned-allocator.h>
17*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
18*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/microfnptr.h>
19*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/raddextexp.h>
20*4bdc9457SAndroid Build Coastguard Worker 
21*4bdc9457SAndroid Build Coastguard Worker 
f32_raddextexp(benchmark::State & state,xnn_f32_raddextexp_ukernel_function raddextexp,benchmark::utils::IsaCheckFunction isa_check=nullptr)22*4bdc9457SAndroid Build Coastguard Worker static void f32_raddextexp(
23*4bdc9457SAndroid Build Coastguard Worker   benchmark::State& state,
24*4bdc9457SAndroid Build Coastguard Worker   xnn_f32_raddextexp_ukernel_function raddextexp,
25*4bdc9457SAndroid Build Coastguard Worker   benchmark::utils::IsaCheckFunction isa_check = nullptr)
26*4bdc9457SAndroid Build Coastguard Worker {
27*4bdc9457SAndroid Build Coastguard Worker   if (isa_check && !isa_check(state)) {
28*4bdc9457SAndroid Build Coastguard Worker     return;
29*4bdc9457SAndroid Build Coastguard Worker   }
30*4bdc9457SAndroid Build Coastguard Worker 
31*4bdc9457SAndroid Build Coastguard Worker   const size_t elements = state.range(0);
32*4bdc9457SAndroid Build Coastguard Worker   const size_t cache_line_size_max = 128;
33*4bdc9457SAndroid Build Coastguard Worker   const size_t packed_elements = benchmark::utils::RoundUp(elements, cache_line_size_max / sizeof(float));
34*4bdc9457SAndroid Build Coastguard Worker 
35*4bdc9457SAndroid Build Coastguard Worker   std::random_device random_device;
36*4bdc9457SAndroid Build Coastguard Worker   auto rng = std::mt19937(random_device());
37*4bdc9457SAndroid Build Coastguard Worker   auto f32rng = std::bind(std::uniform_real_distribution<float>(-1000.0f, 1000.0f), std::ref(rng));
38*4bdc9457SAndroid Build Coastguard Worker 
39*4bdc9457SAndroid Build Coastguard Worker   const size_t num_buffers = 1 +
40*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), packed_elements * sizeof(float));
41*4bdc9457SAndroid Build Coastguard Worker   std::vector<float, AlignedAllocator<float, 64>> x(elements);
42*4bdc9457SAndroid Build Coastguard Worker 
43*4bdc9457SAndroid Build Coastguard Worker   std::generate(x.begin(), x.end(), std::ref(f32rng));
44*4bdc9457SAndroid Build Coastguard Worker 
45*4bdc9457SAndroid Build Coastguard Worker   benchmark::utils::DisableDenormals();
46*4bdc9457SAndroid Build Coastguard Worker 
47*4bdc9457SAndroid Build Coastguard Worker   size_t buffer_index = 0;
48*4bdc9457SAndroid Build Coastguard Worker   for (auto _ : state) {
49*4bdc9457SAndroid Build Coastguard Worker     if (++buffer_index == num_buffers) {
50*4bdc9457SAndroid Build Coastguard Worker       buffer_index = 0;
51*4bdc9457SAndroid Build Coastguard Worker     }
52*4bdc9457SAndroid Build Coastguard Worker 
53*4bdc9457SAndroid Build Coastguard Worker     float y_sum[2] = { nanf(""), nanf("") };
54*4bdc9457SAndroid Build Coastguard Worker     raddextexp(elements * sizeof(float), x.data(), y_sum);
55*4bdc9457SAndroid Build Coastguard Worker   }
56*4bdc9457SAndroid Build Coastguard Worker 
57*4bdc9457SAndroid Build Coastguard Worker   const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
58*4bdc9457SAndroid Build Coastguard Worker   if (cpu_frequency != 0) {
59*4bdc9457SAndroid Build Coastguard Worker     state.counters["cpufreq"] = cpu_frequency;
60*4bdc9457SAndroid Build Coastguard Worker   }
61*4bdc9457SAndroid Build Coastguard Worker 
62*4bdc9457SAndroid Build Coastguard Worker   const size_t elements_per_iteration = elements;
63*4bdc9457SAndroid Build Coastguard Worker   state.counters["elements"] =
64*4bdc9457SAndroid Build Coastguard Worker     benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);
65*4bdc9457SAndroid Build Coastguard Worker 
66*4bdc9457SAndroid Build Coastguard Worker   const size_t bytes_per_iteration = 2 * elements * sizeof(float);
67*4bdc9457SAndroid Build Coastguard Worker   state.counters["bytes"] =
68*4bdc9457SAndroid Build Coastguard Worker     benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
69*4bdc9457SAndroid Build Coastguard Worker }
70*4bdc9457SAndroid Build Coastguard Worker 
CharacteristicArguments(benchmark::internal::Benchmark * b)71*4bdc9457SAndroid Build Coastguard Worker static void CharacteristicArguments(benchmark::internal::Benchmark* b) {
72*4bdc9457SAndroid Build Coastguard Worker   b->ArgName("N");
73*4bdc9457SAndroid Build Coastguard Worker   for (int32_t n = 10000; n <= 100000000; n *= 10) {
74*4bdc9457SAndroid Build Coastguard Worker     b->Arg(n);
75*4bdc9457SAndroid Build Coastguard Worker   }
76*4bdc9457SAndroid Build Coastguard Worker }
77*4bdc9457SAndroid Build Coastguard Worker 
78*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_X86 || XNN_ARCH_X86_64
79*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x128,
80*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128,
81*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
82*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x128_acc2,
83*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2,
84*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
85*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x128_acc4,
86*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4,
87*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
88*4bdc9457SAndroid Build Coastguard Worker 
89*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x144,
90*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144,
91*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
92*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x144_acc3,
93*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3,
94*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
95*4bdc9457SAndroid Build Coastguard Worker 
96*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x160,
97*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160,
98*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
99*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x160_acc2,
100*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2,
101*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
102*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x160_acc5,
103*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5,
104*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
105*4bdc9457SAndroid Build Coastguard Worker 
106*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x192,
107*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192,
108*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
109*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x192_acc2,
110*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2,
111*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
112*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x192_acc3,
113*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3,
114*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
115*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx512f_p5_scalef_x192_acc6,
116*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6,
117*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
118*4bdc9457SAndroid Build Coastguard Worker 
119*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x64,
120*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx2_p5_x64,
121*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
122*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x64_acc2,
123*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2,
124*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
125*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x64_acc4,
126*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4,
127*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
128*4bdc9457SAndroid Build Coastguard Worker 
129*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x72,
130*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx2_p5_x72,
131*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
132*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x72_acc3,
133*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3,
134*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
135*4bdc9457SAndroid Build Coastguard Worker 
136*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x80,
137*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx2_p5_x80,
138*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
139*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x80_acc2,
140*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2,
141*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
142*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x80_acc5,
143*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5,
144*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
145*4bdc9457SAndroid Build Coastguard Worker 
146*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x96,
147*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx2_p5_x96,
148*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
149*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x96_acc2,
150*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2,
151*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
152*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x96_acc3,
153*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3,
154*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
155*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_raddextexp, avx2_p5_x96_acc6,
156*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6,
157*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
158*4bdc9457SAndroid Build Coastguard Worker #endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
159*4bdc9457SAndroid Build Coastguard Worker 
160*4bdc9457SAndroid Build Coastguard Worker #ifndef XNNPACK_BENCHMARK_NO_MAIN
161*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_MAIN();
162*4bdc9457SAndroid Build Coastguard Worker #endif
163