xref: /aosp_15_r20/external/XNNPACK/bench/f32-vscaleexpminusmax.cc (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker #include <algorithm>
2*4bdc9457SAndroid Build Coastguard Worker #include <cfloat>
3*4bdc9457SAndroid Build Coastguard Worker #include <chrono>
4*4bdc9457SAndroid Build Coastguard Worker #include <cmath>
5*4bdc9457SAndroid Build Coastguard Worker #include <functional>
6*4bdc9457SAndroid Build Coastguard Worker #include <random>
7*4bdc9457SAndroid Build Coastguard Worker #include <vector>
8*4bdc9457SAndroid Build Coastguard Worker 
9*4bdc9457SAndroid Build Coastguard Worker #include <benchmark/benchmark.h>
10*4bdc9457SAndroid Build Coastguard Worker #include "bench/utils.h"
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack.h>
13*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/aligned-allocator.h>
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/microfnptr.h>
16*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/rmax.h>
17*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/raddexpminusmax.h>
18*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vscaleexpminusmax.h>
19*4bdc9457SAndroid Build Coastguard Worker 
20*4bdc9457SAndroid Build Coastguard Worker 
f32_vscaleexpminusmax(benchmark::State & state,xnn_f32_rmax_ukernel_function rmax,xnn_f32_raddexpminusmax_ukernel_function raddexpminusmax,xnn_f32_vscaleexpminusmax_ukernel_function vscaleexpminusmax,benchmark::utils::IsaCheckFunction isa_check=nullptr)21*4bdc9457SAndroid Build Coastguard Worker static void f32_vscaleexpminusmax(
22*4bdc9457SAndroid Build Coastguard Worker   benchmark::State& state,
23*4bdc9457SAndroid Build Coastguard Worker   xnn_f32_rmax_ukernel_function rmax,
24*4bdc9457SAndroid Build Coastguard Worker   xnn_f32_raddexpminusmax_ukernel_function raddexpminusmax,
25*4bdc9457SAndroid Build Coastguard Worker   xnn_f32_vscaleexpminusmax_ukernel_function vscaleexpminusmax,
26*4bdc9457SAndroid Build Coastguard Worker   benchmark::utils::IsaCheckFunction isa_check = nullptr)
27*4bdc9457SAndroid Build Coastguard Worker {
28*4bdc9457SAndroid Build Coastguard Worker   if (isa_check && !isa_check(state)) {
29*4bdc9457SAndroid Build Coastguard Worker     return;
30*4bdc9457SAndroid Build Coastguard Worker   }
31*4bdc9457SAndroid Build Coastguard Worker 
32*4bdc9457SAndroid Build Coastguard Worker   const size_t elements = state.range(0);
33*4bdc9457SAndroid Build Coastguard Worker   const size_t cache_line_size_max = 128;
34*4bdc9457SAndroid Build Coastguard Worker   const size_t packed_elements = benchmark::utils::RoundUp(elements, cache_line_size_max / sizeof(float));
35*4bdc9457SAndroid Build Coastguard Worker 
36*4bdc9457SAndroid Build Coastguard Worker   std::random_device random_device;
37*4bdc9457SAndroid Build Coastguard Worker   auto rng = std::mt19937(random_device());
38*4bdc9457SAndroid Build Coastguard Worker   auto f32rng = std::bind(std::uniform_real_distribution<float>(-1000.0f, 1000.0f), std::ref(rng));
39*4bdc9457SAndroid Build Coastguard Worker 
40*4bdc9457SAndroid Build Coastguard Worker   const size_t num_buffers = 1 +
41*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), packed_elements * sizeof(float));
42*4bdc9457SAndroid Build Coastguard Worker   std::vector<float, AlignedAllocator<float, 64>> x(elements);
43*4bdc9457SAndroid Build Coastguard Worker   std::vector<float, AlignedAllocator<float, 64>> y(packed_elements * num_buffers);
44*4bdc9457SAndroid Build Coastguard Worker 
45*4bdc9457SAndroid Build Coastguard Worker   std::generate(x.begin(), x.end(), std::ref(f32rng));
46*4bdc9457SAndroid Build Coastguard Worker 
47*4bdc9457SAndroid Build Coastguard Worker   benchmark::utils::DisableDenormals();
48*4bdc9457SAndroid Build Coastguard Worker 
49*4bdc9457SAndroid Build Coastguard Worker   size_t buffer_index = 0;
50*4bdc9457SAndroid Build Coastguard Worker   for (auto _ : state) {
51*4bdc9457SAndroid Build Coastguard Worker     state.PauseTiming();
52*4bdc9457SAndroid Build Coastguard Worker     float x_max = nanf("");
53*4bdc9457SAndroid Build Coastguard Worker     rmax(elements * sizeof(float), x.data(), &x_max);
54*4bdc9457SAndroid Build Coastguard Worker     float y_sum = nanf("");
55*4bdc9457SAndroid Build Coastguard Worker     raddexpminusmax(elements * sizeof(float), x.data(), &y_sum, x_max);
56*4bdc9457SAndroid Build Coastguard Worker     if (++buffer_index == num_buffers) {
57*4bdc9457SAndroid Build Coastguard Worker       buffer_index = 0;
58*4bdc9457SAndroid Build Coastguard Worker     }
59*4bdc9457SAndroid Build Coastguard Worker     state.ResumeTiming();
60*4bdc9457SAndroid Build Coastguard Worker 
61*4bdc9457SAndroid Build Coastguard Worker     vscaleexpminusmax(elements * sizeof(float), x.data(), y.data() + packed_elements * buffer_index, x_max, 1.0f / y_sum);
62*4bdc9457SAndroid Build Coastguard Worker   }
63*4bdc9457SAndroid Build Coastguard Worker 
64*4bdc9457SAndroid Build Coastguard Worker   const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
65*4bdc9457SAndroid Build Coastguard Worker   if (cpu_frequency != 0) {
66*4bdc9457SAndroid Build Coastguard Worker     state.counters["cpufreq"] = cpu_frequency;
67*4bdc9457SAndroid Build Coastguard Worker   }
68*4bdc9457SAndroid Build Coastguard Worker 
69*4bdc9457SAndroid Build Coastguard Worker   const size_t elements_per_iteration = elements;
70*4bdc9457SAndroid Build Coastguard Worker   state.counters["elements"] =
71*4bdc9457SAndroid Build Coastguard Worker     benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);
72*4bdc9457SAndroid Build Coastguard Worker 
73*4bdc9457SAndroid Build Coastguard Worker   const size_t bytes_per_iteration = 2 * elements * sizeof(float);
74*4bdc9457SAndroid Build Coastguard Worker   state.counters["bytes"] =
75*4bdc9457SAndroid Build Coastguard Worker     benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
76*4bdc9457SAndroid Build Coastguard Worker }
77*4bdc9457SAndroid Build Coastguard Worker 
CharacteristicArguments(benchmark::internal::Benchmark * b)78*4bdc9457SAndroid Build Coastguard Worker static void CharacteristicArguments(benchmark::internal::Benchmark* b) {
79*4bdc9457SAndroid Build Coastguard Worker   b->ArgName("N");
80*4bdc9457SAndroid Build Coastguard Worker   for (int32_t n = 10000; n <= 100000000; n *= 10) {
81*4bdc9457SAndroid Build Coastguard Worker     b->Arg(n);
82*4bdc9457SAndroid Build Coastguard Worker   }
83*4bdc9457SAndroid Build Coastguard Worker }
84*4bdc9457SAndroid Build Coastguard Worker 
85*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_X86 || XNN_ARCH_X86_64
86*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx512f_p5_scalef_x16,
87*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx512f,
88*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
89*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16,
90*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
91*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx512f_p5_scalef_x32,
92*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx512f,
93*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
94*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32,
95*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
96*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx512f_p5_scalef_x48,
97*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx512f,
98*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
99*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x48,
100*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
101*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx512f_p5_scalef_x64,
102*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx512f,
103*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
104*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x64,
105*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
106*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx512f_p5_scalef_x80,
107*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx512f,
108*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
109*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x80,
110*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
111*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx512f_p5_scalef_x96,
112*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx512f,
113*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
114*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96,
115*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
116*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx512f_p5_scalef_x112,
117*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx512f,
118*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
119*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112,
120*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
121*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx512f_p5_scalef_x128,
122*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx512f,
123*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
124*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128,
125*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
126*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx512f_p5_scalef_x144,
127*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx512f,
128*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
129*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144,
130*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
131*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx512f_p5_scalef_x160,
132*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx512f,
133*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
134*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160,
135*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
136*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx512f_p5_scalef_x176,
137*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx512f,
138*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
139*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176,
140*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
141*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx512f_p5_scalef_x192,
142*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx512f,
143*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2,
144*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192,
145*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime();
146*4bdc9457SAndroid Build Coastguard Worker 
147*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx2_p5_x8,
148*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx,
149*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
150*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x8,
151*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
152*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx2_p5_x16,
153*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx,
154*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
155*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x16,
156*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
157*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx2_p5_x24,
158*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx,
159*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
160*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x24,
161*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
162*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx2_p5_x32,
163*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx,
164*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
165*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x32,
166*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
167*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx2_p5_x40,
168*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx,
169*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
170*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x40,
171*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
172*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx2_p5_x48,
173*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx,
174*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
175*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48,
176*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
177*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx2_p5_x56,
178*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx,
179*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
180*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56,
181*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
182*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx2_p5_x64,
183*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx,
184*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
185*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64,
186*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
187*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx2_p5_x72,
188*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx,
189*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
190*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72,
191*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
192*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx2_p5_x80,
193*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx,
194*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
195*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80,
196*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
197*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx2_p5_x88,
198*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx,
199*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
200*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88,
201*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
202*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f32_vscaleexpminusmax, avx2_p5_x96,
203*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_rmax_ukernel__avx,
204*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2,
205*4bdc9457SAndroid Build Coastguard Worker     xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96,
206*4bdc9457SAndroid Build Coastguard Worker     benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime();
207*4bdc9457SAndroid Build Coastguard Worker #endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
208*4bdc9457SAndroid Build Coastguard Worker 
209*4bdc9457SAndroid Build Coastguard Worker #ifndef XNNPACK_BENCHMARK_NO_MAIN
210*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_MAIN();
211*4bdc9457SAndroid Build Coastguard Worker #endif
212