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 "bench/utils.h"
14
15 #include <fp16/fp16.h>
16 #include <xnnpack/AlignedAllocator.h>
17 #include <xnnpack/common.h>
18 #include <xnnpack/params.h>
19 #include <xnnpack/params-init.h>
20 #include <xnnpack/vcvt.h>
21
22
qs8_f32_vcvt(benchmark::State & state,xnn_qs8_f32_vcvt_ukernel_function cvt,xnn_init_qs8_f32_cvt_params_fn init_params,benchmark::utils::IsaCheckFunction isa_check=nullptr)23 static void qs8_f32_vcvt(
24 benchmark::State& state,
25 xnn_qs8_f32_vcvt_ukernel_function cvt,
26 xnn_init_qs8_f32_cvt_params_fn init_params,
27 benchmark::utils::IsaCheckFunction isa_check = nullptr)
28 {
29 if (isa_check && !isa_check(state)) {
30 return;
31 }
32
33 const size_t num_elements = state.range(0);
34
35 std::random_device random_device;
36 auto rng = std::mt19937(random_device());
37 auto i8rng = std::bind(
38 std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()),
39 std::ref(rng));
40
41 std::vector<int8_t, AlignedAllocator<int8_t, 64>> x(num_elements + XNN_EXTRA_BYTES / sizeof(int8_t));
42 std::vector<float, AlignedAllocator<float, 64>> y(num_elements);
43 std::generate(x.begin(), x.end(), std::ref(i8rng));
44 std::fill(y.begin(), y.end(), std::nanf(""));
45
46 xnn_qs8_f32_cvt_params params;
47 init_params(¶ms,
48 0.25f /* scale */,
49 1 /* output zero point */);
50 for (auto _ : state) {
51 cvt(num_elements * sizeof(int8_t), x.data(), y.data(), ¶ms);
52 }
53
54 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
55 if (cpu_frequency != 0) {
56 state.counters["cpufreq"] = cpu_frequency;
57 }
58
59 const size_t elements_per_iteration = num_elements;
60 state.counters["elements"] =
61 benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);
62
63 const size_t bytes_per_iteration = num_elements * (sizeof(int8_t) + sizeof(float));
64 state.counters["bytes"] =
65 benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
66 }
67
68 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
69 BENCHMARK_CAPTURE(qs8_f32_vcvt, neon_x8,
70 xnn_qs8_f32_vcvt_ukernel__neon_x8,
71 xnn_init_qs8_f32_cvt_neon_params,
72 benchmark::utils::CheckNEON)
73 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
74 ->UseRealTime();
75 BENCHMARK_CAPTURE(qs8_f32_vcvt, neon_x16,
76 xnn_qs8_f32_vcvt_ukernel__neon_x16,
77 xnn_init_qs8_f32_cvt_neon_params,
78 benchmark::utils::CheckNEON)
79 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
80 ->UseRealTime();
81 BENCHMARK_CAPTURE(qs8_f32_vcvt, neon_x24,
82 xnn_qs8_f32_vcvt_ukernel__neon_x24,
83 xnn_init_qs8_f32_cvt_neon_params,
84 benchmark::utils::CheckNEON)
85 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
86 ->UseRealTime();
87 BENCHMARK_CAPTURE(qs8_f32_vcvt, neon_x32,
88 xnn_qs8_f32_vcvt_ukernel__neon_x32,
89 xnn_init_qs8_f32_cvt_neon_params,
90 benchmark::utils::CheckNEON)
91 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
92 ->UseRealTime();
93 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
94
95 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
96 BENCHMARK_CAPTURE(qs8_f32_vcvt, avx512skx_x16,
97 xnn_qs8_f32_vcvt_ukernel__avx512skx_x16,
98 xnn_init_qs8_f32_cvt_avx512_params,
99 benchmark::utils::CheckAVX512SKX)
100 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
101 ->UseRealTime();
102 BENCHMARK_CAPTURE(qs8_f32_vcvt, avx512skx_x32,
103 xnn_qs8_f32_vcvt_ukernel__avx512skx_x32,
104 xnn_init_qs8_f32_cvt_avx512_params,
105 benchmark::utils::CheckAVX512SKX)
106 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
107 ->UseRealTime();
108 BENCHMARK_CAPTURE(qs8_f32_vcvt, avx512skx_x48,
109 xnn_qs8_f32_vcvt_ukernel__avx512skx_x48,
110 xnn_init_qs8_f32_cvt_avx512_params,
111 benchmark::utils::CheckAVX512SKX)
112 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
113 ->UseRealTime();
114 BENCHMARK_CAPTURE(qs8_f32_vcvt, avx512skx_x64,
115 xnn_qs8_f32_vcvt_ukernel__avx512skx_x64,
116 xnn_init_qs8_f32_cvt_avx512_params,
117 benchmark::utils::CheckAVX512SKX)
118 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
119 ->UseRealTime();
120
121 BENCHMARK_CAPTURE(qs8_f32_vcvt, avx2_x8,
122 xnn_qs8_f32_vcvt_ukernel__avx2_x8,
123 xnn_init_qs8_f32_cvt_avx_params,
124 benchmark::utils::CheckAVX2)
125 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
126 ->UseRealTime();
127 BENCHMARK_CAPTURE(qs8_f32_vcvt, avx2_x16,
128 xnn_qs8_f32_vcvt_ukernel__avx2_x16,
129 xnn_init_qs8_f32_cvt_avx_params,
130 benchmark::utils::CheckAVX2)
131 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
132 ->UseRealTime();
133 BENCHMARK_CAPTURE(qs8_f32_vcvt, avx2_x24,
134 xnn_qs8_f32_vcvt_ukernel__avx2_x24,
135 xnn_init_qs8_f32_cvt_avx_params,
136 benchmark::utils::CheckAVX2)
137 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
138 ->UseRealTime();
139 BENCHMARK_CAPTURE(qs8_f32_vcvt, avx2_x32,
140 xnn_qs8_f32_vcvt_ukernel__avx2_x32,
141 xnn_init_qs8_f32_cvt_avx_params,
142 benchmark::utils::CheckAVX2)
143 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
144 ->UseRealTime();
145
146 BENCHMARK_CAPTURE(qs8_f32_vcvt, avx_x8,
147 xnn_qs8_f32_vcvt_ukernel__avx_x8,
148 xnn_init_qs8_f32_cvt_avx_params,
149 benchmark::utils::CheckAVX)
150 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
151 ->UseRealTime();
152 BENCHMARK_CAPTURE(qs8_f32_vcvt, avx_x16,
153 xnn_qs8_f32_vcvt_ukernel__avx_x16,
154 xnn_init_qs8_f32_cvt_avx_params,
155 benchmark::utils::CheckAVX)
156 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
157 ->UseRealTime();
158 BENCHMARK_CAPTURE(qs8_f32_vcvt, avx_x24,
159 xnn_qs8_f32_vcvt_ukernel__avx_x24,
160 xnn_init_qs8_f32_cvt_avx_params,
161 benchmark::utils::CheckAVX)
162 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
163 ->UseRealTime();
164 BENCHMARK_CAPTURE(qs8_f32_vcvt, avx_x32,
165 xnn_qs8_f32_vcvt_ukernel__avx_x32,
166 xnn_init_qs8_f32_cvt_avx_params,
167 benchmark::utils::CheckAVX)
168 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
169 ->UseRealTime();
170
171 BENCHMARK_CAPTURE(qs8_f32_vcvt, sse41_x8,
172 xnn_qs8_f32_vcvt_ukernel__sse41_x8,
173 xnn_init_qs8_f32_cvt_sse4_params,
174 benchmark::utils::CheckSSE41)
175 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
176 ->UseRealTime();
177 BENCHMARK_CAPTURE(qs8_f32_vcvt, sse41_x16,
178 xnn_qs8_f32_vcvt_ukernel__sse41_x16,
179 xnn_init_qs8_f32_cvt_sse4_params,
180 benchmark::utils::CheckSSE41)
181 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
182 ->UseRealTime();
183 BENCHMARK_CAPTURE(qs8_f32_vcvt, sse41_x24,
184 xnn_qs8_f32_vcvt_ukernel__sse41_x24,
185 xnn_init_qs8_f32_cvt_sse4_params,
186 benchmark::utils::CheckSSE41)
187 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
188 ->UseRealTime();
189 BENCHMARK_CAPTURE(qs8_f32_vcvt, sse41_x32,
190 xnn_qs8_f32_vcvt_ukernel__sse41_x32,
191 xnn_init_qs8_f32_cvt_sse4_params,
192 benchmark::utils::CheckSSE41)
193 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
194 ->UseRealTime();
195
196 BENCHMARK_CAPTURE(qs8_f32_vcvt, sse2_x8,
197 xnn_qs8_f32_vcvt_ukernel__sse2_x8,
198 xnn_init_qs8_f32_cvt_sse2_params)
199 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
200 ->UseRealTime();
201 BENCHMARK_CAPTURE(qs8_f32_vcvt, sse2_x16,
202 xnn_qs8_f32_vcvt_ukernel__sse2_x16,
203 xnn_init_qs8_f32_cvt_sse2_params)
204 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
205 ->UseRealTime();
206 BENCHMARK_CAPTURE(qs8_f32_vcvt, sse2_x24,
207 xnn_qs8_f32_vcvt_ukernel__sse2_x24,
208 xnn_init_qs8_f32_cvt_sse2_params)
209 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
210 ->UseRealTime();
211 BENCHMARK_CAPTURE(qs8_f32_vcvt, sse2_x32,
212 xnn_qs8_f32_vcvt_ukernel__sse2_x32,
213 xnn_init_qs8_f32_cvt_sse2_params)
214 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
215 ->UseRealTime();
216 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
217
218 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
219 BENCHMARK_CAPTURE(qs8_f32_vcvt, wasmsimd_x8,
220 xnn_qs8_f32_vcvt_ukernel__wasmsimd_x8,
221 xnn_init_qs8_f32_cvt_wasmsimd_params)
222 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
223 ->UseRealTime();
224 BENCHMARK_CAPTURE(qs8_f32_vcvt, wasmsimd_x16,
225 xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16,
226 xnn_init_qs8_f32_cvt_wasmsimd_params)
227 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
228 ->UseRealTime();
229 BENCHMARK_CAPTURE(qs8_f32_vcvt, wasmsimd_x24,
230 xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24,
231 xnn_init_qs8_f32_cvt_wasmsimd_params)
232 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
233 ->UseRealTime();
234 BENCHMARK_CAPTURE(qs8_f32_vcvt, wasmsimd_x32,
235 xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32,
236 xnn_init_qs8_f32_cvt_wasmsimd_params)
237 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
238 ->UseRealTime();
239 #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
240
241 BENCHMARK_CAPTURE(qs8_f32_vcvt, scalar_x1,
242 xnn_qs8_f32_vcvt_ukernel__scalar_x1,
243 xnn_init_qs8_f32_cvt_scalar_params)
244 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
245 ->UseRealTime();
246 BENCHMARK_CAPTURE(qs8_f32_vcvt, scalar_x2,
247 xnn_qs8_f32_vcvt_ukernel__scalar_x2,
248 xnn_init_qs8_f32_cvt_scalar_params)
249 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
250 ->UseRealTime();
251 BENCHMARK_CAPTURE(qs8_f32_vcvt, scalar_x3,
252 xnn_qs8_f32_vcvt_ukernel__scalar_x3,
253 xnn_init_qs8_f32_cvt_scalar_params)
254 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
255 ->UseRealTime();
256 BENCHMARK_CAPTURE(qs8_f32_vcvt, scalar_x4,
257 xnn_qs8_f32_vcvt_ukernel__scalar_x4,
258 xnn_init_qs8_f32_cvt_scalar_params)
259 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, float>)
260 ->UseRealTime();
261
262 #ifndef XNNPACK_BENCHMARK_NO_MAIN
263 BENCHMARK_MAIN();
264 #endif
265