1 // Copyright 2021 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <algorithm>
7 #include <cmath>
8 #include <functional>
9 #include <random>
10 #include <vector>
11
12 #include <benchmark/benchmark.h>
13 #include <fp16/fp16.h>
14 #include "bench/utils.h"
15
16 #include <xnnpack.h>
17 #include <xnnpack/aligned-allocator.h>
18 #include <xnnpack/common.h>
19 #include <xnnpack/microfnptr.h>
20 #include <xnnpack/microparams-init.h>
21 #include <xnnpack/vcvt.h>
22
23
f32_f16_vcvt(benchmark::State & state,xnn_f32_f16_vcvt_ukernel_function cvt,xnn_init_f32_f16_cvt_params_fn init_params=nullptr,benchmark::utils::IsaCheckFunction isa_check=nullptr)24 static void f32_f16_vcvt(
25 benchmark::State& state,
26 xnn_f32_f16_vcvt_ukernel_function cvt,
27 xnn_init_f32_f16_cvt_params_fn init_params = nullptr,
28 benchmark::utils::IsaCheckFunction isa_check = nullptr)
29 {
30 if (isa_check && !isa_check(state)) {
31 return;
32 }
33
34 const size_t num_elements = state.range(0);
35
36 std::random_device random_device;
37 auto rng = std::mt19937(random_device());
38 auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng));
39
40 std::vector<float, AlignedAllocator<float, 64>> x(num_elements + XNN_EXTRA_BYTES / sizeof(float));
41 std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> y(num_elements);
42 std::generate(x.begin(), x.end(), std::ref(f32rng));
43 std::fill(y.begin(), y.end(), UINT16_C(0x7E00));
44
45 xnn_f32_f16_cvt_params params;
46 if (init_params != nullptr) {
47 init_params(¶ms);
48 }
49 for (auto _ : state) {
50 cvt(num_elements * sizeof(uint16_t), x.data(), y.data(), ¶ms);
51 }
52
53 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
54 if (cpu_frequency != 0) {
55 state.counters["cpufreq"] = cpu_frequency;
56 }
57
58 const size_t elements_per_iteration = num_elements;
59 state.counters["elements"] =
60 benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);
61
62 const size_t bytes_per_iteration = num_elements * (sizeof(uint16_t) + sizeof(float));
63 state.counters["bytes"] =
64 benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
65 }
66
67 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
68 BENCHMARK_CAPTURE(f32_f16_vcvt, neonfp16_x8,
69 xnn_f32_f16_vcvt_ukernel__neonfp16_x8,
70 nullptr /* init params */,
71 benchmark::utils::CheckNEONFP16)
72 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
73 ->UseRealTime();
74 BENCHMARK_CAPTURE(f32_f16_vcvt, neonfp16_x16,
75 xnn_f32_f16_vcvt_ukernel__neonfp16_x16,
76 nullptr /* init params */,
77 benchmark::utils::CheckNEONFP16)
78 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
79 ->UseRealTime();
80
81 BENCHMARK_CAPTURE(f32_f16_vcvt, neon_x8,
82 xnn_f32_f16_vcvt_ukernel__neon_x8,
83 xnn_init_f32_f16_cvt_neon_params,
84 benchmark::utils::CheckNEON)
85 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
86 ->UseRealTime();
87 BENCHMARK_CAPTURE(f32_f16_vcvt, neon_x16,
88 xnn_f32_f16_vcvt_ukernel__neon_x16,
89 xnn_init_f32_f16_cvt_neon_params,
90 benchmark::utils::CheckNEON)
91 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
92 ->UseRealTime();
93 BENCHMARK_CAPTURE(f32_f16_vcvt, neon_x24,
94 xnn_f32_f16_vcvt_ukernel__neon_x24,
95 xnn_init_f32_f16_cvt_neon_params,
96 benchmark::utils::CheckNEON)
97 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
98 ->UseRealTime();
99 BENCHMARK_CAPTURE(f32_f16_vcvt, neon_x32,
100 xnn_f32_f16_vcvt_ukernel__neon_x32,
101 xnn_init_f32_f16_cvt_neon_params,
102 benchmark::utils::CheckNEON)
103 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
104 ->UseRealTime();
105 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
106
107 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
108 BENCHMARK_CAPTURE(f32_f16_vcvt, avx512skx_x16,
109 xnn_f32_f16_vcvt_ukernel__avx512skx_x16,
110 nullptr /* init params */,
111 benchmark::utils::CheckAVX512SKX)
112 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
113 ->UseRealTime();
114 BENCHMARK_CAPTURE(f32_f16_vcvt, avx512skx_x32,
115 xnn_f32_f16_vcvt_ukernel__avx512skx_x32,
116 nullptr /* init params */,
117 benchmark::utils::CheckAVX512SKX)
118 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
119 ->UseRealTime();
120
121 BENCHMARK_CAPTURE(f32_f16_vcvt, f16c_x8,
122 xnn_f32_f16_vcvt_ukernel__f16c_x8,
123 xnn_init_f32_f16_cvt_f16c_params,
124 benchmark::utils::CheckF16C)
125 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
126 ->UseRealTime();
127 BENCHMARK_CAPTURE(f32_f16_vcvt, f16c_x16,
128 xnn_f32_f16_vcvt_ukernel__f16c_x16,
129 xnn_init_f32_f16_cvt_f16c_params,
130 benchmark::utils::CheckF16C)
131 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
132 ->UseRealTime();
133
134 BENCHMARK_CAPTURE(f32_f16_vcvt, avx_x8,
135 xnn_f32_f16_vcvt_ukernel__avx_x8,
136 xnn_init_f32_f16_cvt_sse2_params,
137 benchmark::utils::CheckAVX)
138 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
139 ->UseRealTime();
140 BENCHMARK_CAPTURE(f32_f16_vcvt, avx_x16,
141 xnn_f32_f16_vcvt_ukernel__avx_x16,
142 xnn_init_f32_f16_cvt_sse2_params,
143 benchmark::utils::CheckAVX)
144 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
145 ->UseRealTime();
146 BENCHMARK_CAPTURE(f32_f16_vcvt, avx_x24,
147 xnn_f32_f16_vcvt_ukernel__avx_x24,
148 xnn_init_f32_f16_cvt_sse2_params,
149 benchmark::utils::CheckAVX)
150 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
151 ->UseRealTime();
152 BENCHMARK_CAPTURE(f32_f16_vcvt, avx_x32,
153 xnn_f32_f16_vcvt_ukernel__avx_x32,
154 xnn_init_f32_f16_cvt_sse2_params,
155 benchmark::utils::CheckAVX)
156 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
157 ->UseRealTime();
158
159 BENCHMARK_CAPTURE(f32_f16_vcvt, sse41_x8,
160 xnn_f32_f16_vcvt_ukernel__sse41_x8,
161 xnn_init_f32_f16_cvt_sse2_params,
162 benchmark::utils::CheckSSE41)
163 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
164 ->UseRealTime();
165 BENCHMARK_CAPTURE(f32_f16_vcvt, sse41_x16,
166 xnn_f32_f16_vcvt_ukernel__sse41_x16,
167 xnn_init_f32_f16_cvt_sse2_params,
168 benchmark::utils::CheckSSE41)
169 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
170 ->UseRealTime();
171 BENCHMARK_CAPTURE(f32_f16_vcvt, sse41_x24,
172 xnn_f32_f16_vcvt_ukernel__sse41_x24,
173 xnn_init_f32_f16_cvt_sse2_params,
174 benchmark::utils::CheckSSE41)
175 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
176 ->UseRealTime();
177 BENCHMARK_CAPTURE(f32_f16_vcvt, sse41_x32,
178 xnn_f32_f16_vcvt_ukernel__sse41_x32,
179 xnn_init_f32_f16_cvt_sse2_params,
180 benchmark::utils::CheckSSE41)
181 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
182 ->UseRealTime();
183
184 BENCHMARK_CAPTURE(f32_f16_vcvt, sse2_x8,
185 xnn_f32_f16_vcvt_ukernel__sse2_x8,
186 xnn_init_f32_f16_cvt_sse2_params)
187 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
188 ->UseRealTime();
189 BENCHMARK_CAPTURE(f32_f16_vcvt, sse2_x16,
190 xnn_f32_f16_vcvt_ukernel__sse2_x16,
191 xnn_init_f32_f16_cvt_sse2_params)
192 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
193 ->UseRealTime();
194 BENCHMARK_CAPTURE(f32_f16_vcvt, sse2_x24,
195 xnn_f32_f16_vcvt_ukernel__sse2_x24,
196 xnn_init_f32_f16_cvt_sse2_params)
197 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
198 ->UseRealTime();
199 BENCHMARK_CAPTURE(f32_f16_vcvt, sse2_x32,
200 xnn_f32_f16_vcvt_ukernel__sse2_x32,
201 xnn_init_f32_f16_cvt_sse2_params)
202 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
203 ->UseRealTime();
204 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
205
206 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
207 BENCHMARK_CAPTURE(f32_f16_vcvt, wasmsimd_x8,
208 xnn_f32_f16_vcvt_ukernel__wasmsimd_x8,
209 xnn_init_f32_f16_cvt_wasmsimd_params)
210 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
211 ->UseRealTime();
212 BENCHMARK_CAPTURE(f32_f16_vcvt, wasmsimd_x16,
213 xnn_f32_f16_vcvt_ukernel__wasmsimd_x16,
214 xnn_init_f32_f16_cvt_wasmsimd_params)
215 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
216 ->UseRealTime();
217 BENCHMARK_CAPTURE(f32_f16_vcvt, wasmsimd_x24,
218 xnn_f32_f16_vcvt_ukernel__wasmsimd_x24,
219 xnn_init_f32_f16_cvt_wasmsimd_params)
220 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
221 ->UseRealTime();
222 BENCHMARK_CAPTURE(f32_f16_vcvt, wasmsimd_x32,
223 xnn_f32_f16_vcvt_ukernel__wasmsimd_x32,
224 xnn_init_f32_f16_cvt_wasmsimd_params)
225 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
226 ->UseRealTime();
227 #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
228
229 BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_bitcast_x1,
230 xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x1,
231 xnn_init_f32_f16_cvt_scalar_bitcast_params)
232 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
233 ->UseRealTime();
234 BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_bitcast_x2,
235 xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x2,
236 xnn_init_f32_f16_cvt_scalar_bitcast_params)
237 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
238 ->UseRealTime();
239 BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_bitcast_x3,
240 xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x3,
241 xnn_init_f32_f16_cvt_scalar_bitcast_params)
242 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
243 ->UseRealTime();
244 BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_bitcast_x4,
245 xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x4,
246 xnn_init_f32_f16_cvt_scalar_bitcast_params)
247 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
248 ->UseRealTime();
249
250 BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_fabsf_x1,
251 xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x1,
252 xnn_init_f32_f16_cvt_scalar_fabsf_params)
253 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
254 ->UseRealTime();
255 BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_fabsf_x2,
256 xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x2,
257 xnn_init_f32_f16_cvt_scalar_fabsf_params)
258 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
259 ->UseRealTime();
260 BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_fabsf_x3,
261 xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x3,
262 xnn_init_f32_f16_cvt_scalar_fabsf_params)
263 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
264 ->UseRealTime();
265 BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_fabsf_x4,
266 xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x4,
267 xnn_init_f32_f16_cvt_scalar_fabsf_params)
268 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
269 ->UseRealTime();
270
271 #ifndef XNNPACK_BENCHMARK_NO_MAIN
272 BENCHMARK_MAIN();
273 #endif
274