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/raddstoreexpminusmax.h>
22*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/rmax.h>
23*4bdc9457SAndroid Build Coastguard Worker
24*4bdc9457SAndroid Build Coastguard Worker
f16_raddstoreexpminusmax(benchmark::State & state,xnn_f16_rmax_ukernel_function rmax,xnn_f16_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax,xnn_init_f16_expminus_params_fn init_params,benchmark::utils::IsaCheckFunction isa_check=nullptr)25*4bdc9457SAndroid Build Coastguard Worker static void f16_raddstoreexpminusmax(
26*4bdc9457SAndroid Build Coastguard Worker benchmark::State& state,
27*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel_function rmax,
28*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax,
29*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_params_fn init_params,
30*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::IsaCheckFunction isa_check = nullptr)
31*4bdc9457SAndroid Build Coastguard Worker {
32*4bdc9457SAndroid Build Coastguard Worker if (isa_check && !isa_check(state)) {
33*4bdc9457SAndroid Build Coastguard Worker return;
34*4bdc9457SAndroid Build Coastguard Worker }
35*4bdc9457SAndroid Build Coastguard Worker
36*4bdc9457SAndroid Build Coastguard Worker const size_t elements = state.range(0);
37*4bdc9457SAndroid Build Coastguard Worker const size_t cache_line_size_max = 128;
38*4bdc9457SAndroid Build Coastguard Worker const size_t packed_elements = benchmark::utils::RoundUp(elements, cache_line_size_max / sizeof(uint16_t));
39*4bdc9457SAndroid Build Coastguard Worker
40*4bdc9457SAndroid Build Coastguard Worker std::random_device random_device;
41*4bdc9457SAndroid Build Coastguard Worker auto rng = std::mt19937(random_device());
42*4bdc9457SAndroid Build Coastguard Worker auto f32rng = std::bind(std::uniform_real_distribution<float>(-100.0f, 100.0f), std::ref(rng));
43*4bdc9457SAndroid Build Coastguard Worker auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);
44*4bdc9457SAndroid Build Coastguard Worker
45*4bdc9457SAndroid Build Coastguard Worker const size_t num_buffers = 1 +
46*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), packed_elements * sizeof(uint16_t));
47*4bdc9457SAndroid Build Coastguard Worker std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> x(elements);
48*4bdc9457SAndroid Build Coastguard Worker std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> y(packed_elements * num_buffers);
49*4bdc9457SAndroid Build Coastguard Worker
50*4bdc9457SAndroid Build Coastguard Worker std::generate(x.begin(), x.end(), std::ref(f16rng));
51*4bdc9457SAndroid Build Coastguard Worker
52*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::DisableDenormals();
53*4bdc9457SAndroid Build Coastguard Worker
54*4bdc9457SAndroid Build Coastguard Worker xnn_f16_expminus_params params;
55*4bdc9457SAndroid Build Coastguard Worker init_params(¶ms);
56*4bdc9457SAndroid Build Coastguard Worker
57*4bdc9457SAndroid Build Coastguard Worker size_t buffer_index = 0;
58*4bdc9457SAndroid Build Coastguard Worker for (auto _ : state) {
59*4bdc9457SAndroid Build Coastguard Worker state.PauseTiming();
60*4bdc9457SAndroid Build Coastguard Worker uint16_t x_max = UINT16_C(0x7E00) /* NaN */;
61*4bdc9457SAndroid Build Coastguard Worker rmax(elements * sizeof(uint16_t), x.data(), &x_max);
62*4bdc9457SAndroid Build Coastguard Worker if (++buffer_index == num_buffers) {
63*4bdc9457SAndroid Build Coastguard Worker buffer_index = 0;
64*4bdc9457SAndroid Build Coastguard Worker }
65*4bdc9457SAndroid Build Coastguard Worker state.ResumeTiming();
66*4bdc9457SAndroid Build Coastguard Worker
67*4bdc9457SAndroid Build Coastguard Worker uint16_t y_sum = UINT16_C(0x7E00) /* NaN */;
68*4bdc9457SAndroid Build Coastguard Worker raddstoreexpminusmax(elements * sizeof(uint16_t), x.data(), &x_max, y.data() + buffer_index * packed_elements, &y_sum, ¶ms);
69*4bdc9457SAndroid Build Coastguard Worker }
70*4bdc9457SAndroid Build Coastguard Worker
71*4bdc9457SAndroid Build Coastguard Worker const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
72*4bdc9457SAndroid Build Coastguard Worker if (cpu_frequency != 0) {
73*4bdc9457SAndroid Build Coastguard Worker state.counters["cpufreq"] = cpu_frequency;
74*4bdc9457SAndroid Build Coastguard Worker }
75*4bdc9457SAndroid Build Coastguard Worker
76*4bdc9457SAndroid Build Coastguard Worker const size_t elements_per_iteration = elements;
77*4bdc9457SAndroid Build Coastguard Worker state.counters["elements"] =
78*4bdc9457SAndroid Build Coastguard Worker benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);
79*4bdc9457SAndroid Build Coastguard Worker
80*4bdc9457SAndroid Build Coastguard Worker const size_t bytes_per_iteration = 2 * elements * sizeof(uint16_t);
81*4bdc9457SAndroid Build Coastguard Worker state.counters["bytes"] =
82*4bdc9457SAndroid Build Coastguard Worker benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
83*4bdc9457SAndroid Build Coastguard Worker }
84*4bdc9457SAndroid Build Coastguard Worker
85*4bdc9457SAndroid Build Coastguard Worker #if XNN_ENABLE_ARM_FP16 && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
86*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x32,
87*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
88*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x32,
89*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
90*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
91*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
92*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
93*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x32_acc2,
94*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
95*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x32_acc2,
96*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
97*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
98*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
99*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
100*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x32_acc4,
101*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
102*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x32_acc4,
103*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
104*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
105*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
106*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
107*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x40,
108*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
109*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x40,
110*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
111*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
112*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
113*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
114*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x40_acc2,
115*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
116*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x40_acc2,
117*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
118*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
119*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
120*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
121*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x40_acc5,
122*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
123*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x40_acc5,
124*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
125*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
126*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
127*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
128*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x48,
129*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
130*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x48,
131*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
132*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
133*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
134*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
135*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x48_acc2,
136*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
137*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x48_acc2,
138*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_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_raddstoreexpminusmax, neonfp16arith_rr2_p2_x48_acc3,
143*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
144*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x48_acc3,
145*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
146*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
147*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
148*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
149*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x64,
150*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
151*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x64,
152*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
153*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
154*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
155*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
156*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x64_acc2,
157*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
158*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x64_acc2,
159*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
160*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
161*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
162*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
163*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x64_acc4,
164*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
165*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x64_acc4,
166*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
167*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
168*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
169*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
170*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x72,
171*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
172*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x72,
173*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
174*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
175*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
176*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
177*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x72_acc3,
178*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
179*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x72_acc3,
180*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
181*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
182*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
183*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
184*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x80,
185*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
186*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x80,
187*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_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_raddstoreexpminusmax, neonfp16arith_rr2_p2_x80_acc2,
192*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
193*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x80_acc2,
194*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
195*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
196*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
197*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
198*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x80_acc5,
199*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
200*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x80_acc5,
201*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
202*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
203*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
204*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
205*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x96,
206*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
207*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x96,
208*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
209*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
210*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
211*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
212*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x96_acc2,
213*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
214*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x96_acc2,
215*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
216*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
217*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
218*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
219*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x96_acc3,
220*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
221*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x96_acc3,
222*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
223*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
224*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
225*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
226*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, neonfp16arith_rr2_p2_x96_acc6,
227*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__neonfp16arith,
228*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x96_acc6,
229*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_neonfp16arith_rr2_p2_params,
230*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckNEONFP16ARITH)
231*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
232*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
233*4bdc9457SAndroid Build Coastguard Worker #endif // XNN_ENABLE_ARM_FP16 && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
234*4bdc9457SAndroid Build Coastguard Worker
235*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_X86 || XNN_ARCH_X86_64
236*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x32,
237*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
238*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x32,
239*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
240*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
241*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
242*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
243*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x32_acc2,
244*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
245*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x32_acc2,
246*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
247*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
248*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
249*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
250*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x32_acc4,
251*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
252*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x32_acc4,
253*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
254*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
255*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
256*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
257*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x40,
258*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
259*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x40,
260*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
261*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
262*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
263*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
264*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x40_acc2,
265*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
266*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x40_acc2,
267*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
268*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
269*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
270*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
271*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x40_acc5,
272*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
273*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x40_acc5,
274*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
275*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
276*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
277*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
278*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x48,
279*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
280*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x48,
281*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_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_raddstoreexpminusmax, avx2_rr1_p2_x48_acc2,
286*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
287*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x48_acc2,
288*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
289*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
290*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
291*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
292*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x48_acc3,
293*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
294*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x48_acc3,
295*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
296*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
297*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
298*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
299*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x64,
300*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
301*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x64,
302*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
303*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
304*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
305*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
306*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x64_acc2,
307*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
308*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x64_acc2,
309*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
310*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
311*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
312*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
313*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x64_acc4,
314*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
315*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x64_acc4,
316*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
317*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
318*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
319*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
320*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x72,
321*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
322*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x72,
323*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
324*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
325*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
326*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
327*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x72_acc3,
328*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
329*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x72_acc3,
330*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
331*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
332*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
333*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
334*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x80,
335*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
336*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x80,
337*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
338*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
339*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
340*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
341*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x80_acc2,
342*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
343*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x80_acc2,
344*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
345*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
346*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
347*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
348*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x80_acc5,
349*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
350*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x80_acc5,
351*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
352*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
353*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
354*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
355*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x96,
356*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
357*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x96,
358*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
359*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
360*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
361*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
362*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x96_acc2,
363*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
364*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x96_acc2,
365*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
366*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
367*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
368*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
369*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x96_acc3,
370*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
371*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x96_acc3,
372*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
373*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
374*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
375*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
376*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(f16_raddstoreexpminusmax, avx2_rr1_p2_x96_acc6,
377*4bdc9457SAndroid Build Coastguard Worker xnn_f16_rmax_ukernel__f16c,
378*4bdc9457SAndroid Build Coastguard Worker xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x96_acc6,
379*4bdc9457SAndroid Build Coastguard Worker xnn_init_f16_expminus_avx2_rr1_p2_params,
380*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::CheckAVX2)
381*4bdc9457SAndroid Build Coastguard Worker ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
382*4bdc9457SAndroid Build Coastguard Worker ->UseRealTime();
383*4bdc9457SAndroid Build Coastguard Worker #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
384*4bdc9457SAndroid Build Coastguard Worker
385*4bdc9457SAndroid Build Coastguard Worker #ifndef XNNPACK_BENCHMARK_NO_MAIN
386*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_MAIN();
387*4bdc9457SAndroid Build Coastguard Worker #endif
388