xref: /aosp_15_r20/external/XNNPACK/bench/f16-vsigmoid.cc (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Copyright 2022 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 <fp16/fp16.h>
14*4bdc9457SAndroid Build Coastguard Worker #include "bench/utils.h"
15*4bdc9457SAndroid Build Coastguard Worker 
16*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack.h>
17*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/aligned-allocator.h>
18*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
19*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/microfnptr.h>
20*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/microparams-init.h>
21*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vunary.h>
22*4bdc9457SAndroid Build Coastguard Worker 
23*4bdc9457SAndroid Build Coastguard Worker 
f16_vsigmoid(benchmark::State & state,xnn_f16_vsigmoid_ukernel_function sigmoid,xnn_init_f16_sigmoid_params_fn init_params,benchmark::utils::IsaCheckFunction isa_check=nullptr)24*4bdc9457SAndroid Build Coastguard Worker static void f16_vsigmoid(
25*4bdc9457SAndroid Build Coastguard Worker   benchmark::State& state,
26*4bdc9457SAndroid Build Coastguard Worker   xnn_f16_vsigmoid_ukernel_function sigmoid,
27*4bdc9457SAndroid Build Coastguard Worker   xnn_init_f16_sigmoid_params_fn init_params,
28*4bdc9457SAndroid Build Coastguard Worker   benchmark::utils::IsaCheckFunction isa_check = nullptr)
29*4bdc9457SAndroid Build Coastguard Worker {
30*4bdc9457SAndroid Build Coastguard Worker   if (isa_check && !isa_check(state)) {
31*4bdc9457SAndroid Build Coastguard Worker     return;
32*4bdc9457SAndroid Build Coastguard Worker   }
33*4bdc9457SAndroid Build Coastguard Worker 
34*4bdc9457SAndroid Build Coastguard Worker   const size_t num_elements = state.range(0);
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>(-10.0f, 10.0f), std::ref(rng));
39*4bdc9457SAndroid Build Coastguard Worker   auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);
40*4bdc9457SAndroid Build Coastguard Worker 
41*4bdc9457SAndroid Build Coastguard Worker   std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> x(num_elements);
42*4bdc9457SAndroid Build Coastguard Worker   std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> y(num_elements);
43*4bdc9457SAndroid Build Coastguard Worker   std::generate(x.begin(), x.end(), std::ref(f16rng));
44*4bdc9457SAndroid Build Coastguard Worker   std::fill(y.begin(), y.end(), UINT16_C(0x7E00) /* NaN */);
45*4bdc9457SAndroid Build Coastguard Worker 
46*4bdc9457SAndroid Build Coastguard Worker   xnn_f16_sigmoid_params params;
47*4bdc9457SAndroid Build Coastguard Worker   init_params(&params);
48*4bdc9457SAndroid Build Coastguard Worker   for (auto _ : state) {
49*4bdc9457SAndroid Build Coastguard Worker     sigmoid(num_elements * sizeof(uint16_t), x.data(), y.data(), &params);
50*4bdc9457SAndroid Build Coastguard Worker   }
51*4bdc9457SAndroid Build Coastguard Worker 
52*4bdc9457SAndroid Build Coastguard Worker   const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
53*4bdc9457SAndroid Build Coastguard Worker   if (cpu_frequency != 0) {
54*4bdc9457SAndroid Build Coastguard Worker     state.counters["cpufreq"] = cpu_frequency;
55*4bdc9457SAndroid Build Coastguard Worker   }
56*4bdc9457SAndroid Build Coastguard Worker 
57*4bdc9457SAndroid Build Coastguard Worker   const size_t elements_per_iteration = num_elements;
58*4bdc9457SAndroid Build Coastguard Worker   state.counters["elements"] =
59*4bdc9457SAndroid Build Coastguard Worker     benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);
60*4bdc9457SAndroid Build Coastguard Worker 
61*4bdc9457SAndroid Build Coastguard Worker   const size_t bytes_per_iteration = 2 * num_elements * sizeof(uint16_t);
62*4bdc9457SAndroid Build Coastguard Worker   state.counters["bytes"] =
63*4bdc9457SAndroid Build Coastguard Worker     benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
64*4bdc9457SAndroid Build Coastguard Worker }
65*4bdc9457SAndroid Build Coastguard Worker 
66*4bdc9457SAndroid Build Coastguard Worker #if XNN_ENABLE_ARM_FP16 && XNN_ARCH_ARM64
67*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_div_x8,
68*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x8,
69*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
70*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
71*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
72*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
73*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_div_x16,
74*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x16,
75*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
76*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
77*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
78*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
79*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_div_x24,
80*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x24,
81*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
82*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
83*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
84*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
85*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_div_x32,
86*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x32,
87*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
88*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
89*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
90*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
91*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_div_x40,
92*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x40,
93*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
94*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
95*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
96*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
97*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_div_x48,
98*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x48,
99*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
100*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
101*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
102*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
103*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_div_x56,
104*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x56,
105*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
106*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
107*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
108*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
109*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_div_x64,
110*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x64,
111*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
112*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
113*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
114*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
115*4bdc9457SAndroid Build Coastguard Worker #endif  // XNN_ENABLE_ARM_FP16 && XNN_ARCH_ARM64
116*4bdc9457SAndroid Build Coastguard Worker 
117*4bdc9457SAndroid Build Coastguard Worker #if XNN_ENABLE_ARM_FP16 && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
118*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1fma_x8,
119*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1fma_x8,
120*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
121*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
122*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
123*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
124*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1fma_x16,
125*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1fma_x16,
126*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
127*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
128*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
129*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
130*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1fma_x24,
131*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1fma_x24,
132*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
133*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
134*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
135*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
136*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1fma_x32,
137*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1fma_x32,
138*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
139*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
140*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
141*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
142*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1fma_x40,
143*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1fma_x40,
144*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
145*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
146*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
147*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
148*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1fma_x48,
149*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1fma_x48,
150*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
151*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
152*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
153*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
154*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1fma_x56,
155*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1fma_x56,
156*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
157*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
158*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
159*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
160*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1fma_x64,
161*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1fma_x64,
162*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
163*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
164*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
165*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
166*4bdc9457SAndroid Build Coastguard Worker 
167*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1recps_x8,
168*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1recps_x8,
169*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
170*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
171*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
172*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
173*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1recps_x16,
174*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1recps_x16,
175*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
176*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
177*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
178*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
179*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1recps_x24,
180*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1recps_x24,
181*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
182*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
183*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
184*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
185*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1recps_x32,
186*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1recps_x32,
187*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
188*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
189*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
190*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
191*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1recps_x40,
192*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1recps_x40,
193*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
194*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
195*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
196*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
197*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1recps_x48,
198*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1recps_x48,
199*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
200*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
201*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
202*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
203*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1recps_x56,
204*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1recps_x56,
205*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
206*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
207*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
208*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
209*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, neonfp16arith_rr2_p2_nr1recps_x64,
210*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1recps_x64,
211*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params,
212*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckNEONFP16ARITH)
213*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
214*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
215*4bdc9457SAndroid Build Coastguard Worker #endif  // XNN_ENABLE_ARM_FP16 && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
216*4bdc9457SAndroid Build Coastguard Worker 
217*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_X86 || XNN_ARCH_X86_64
218*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_div_x8,
219*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x8,
220*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
221*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
222*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
223*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
224*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_div_x16,
225*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x16,
226*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
227*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
228*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
229*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
230*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_div_x24,
231*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x24,
232*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
233*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
234*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
235*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
236*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_div_x32,
237*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x32,
238*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
239*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
240*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
241*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
242*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_div_x40,
243*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x40,
244*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
245*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
246*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
247*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
248*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_div_x48,
249*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x48,
250*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
251*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
252*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
253*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
254*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_div_x56,
255*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x56,
256*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
257*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
258*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
259*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
260*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_div_x64,
261*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x64,
262*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
263*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
264*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
265*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
266*4bdc9457SAndroid Build Coastguard Worker 
267*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_rcp_x8,
268*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x8,
269*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
270*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
271*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
272*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
273*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_rcp_x16,
274*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x16,
275*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
276*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
277*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
278*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
279*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_rcp_x24,
280*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x24,
281*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
282*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
283*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
284*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
285*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_rcp_x32,
286*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x32,
287*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
288*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
289*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
290*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
291*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_rcp_x40,
292*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x40,
293*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
294*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
295*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
296*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
297*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_rcp_x48,
298*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x48,
299*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
300*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
301*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
302*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
303*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_rcp_x56,
304*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x56,
305*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
306*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
307*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
308*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
309*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(f16_vsigmoid, avx2_rr1_p2_rcp_x64,
310*4bdc9457SAndroid Build Coastguard Worker                     xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x64,
311*4bdc9457SAndroid Build Coastguard Worker                     xnn_init_f16_sigmoid_avx2_rr1_p2_params,
312*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
313*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
314*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
315*4bdc9457SAndroid Build Coastguard Worker #endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
316*4bdc9457SAndroid Build Coastguard Worker 
317*4bdc9457SAndroid Build Coastguard Worker #ifndef XNNPACK_BENCHMARK_NO_MAIN
318*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_MAIN();
319*4bdc9457SAndroid Build Coastguard Worker #endif
320