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
f32_qu8_vcvt(benchmark::State & state,xnn_f32_qu8_vcvt_ukernel_function cvt,xnn_init_f32_qu8_cvt_params_fn init_params,benchmark::utils::IsaCheckFunction isa_check=nullptr)23 static void f32_qu8_vcvt(
24 benchmark::State& state,
25 xnn_f32_qu8_vcvt_ukernel_function cvt,
26 xnn_init_f32_qu8_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 f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng));
38
39 std::vector<float, AlignedAllocator<float, 64>> x(num_elements + XNN_EXTRA_BYTES / sizeof(float));
40 std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> y(num_elements);
41 std::generate(x.begin(), x.end(), std::ref(f32rng));
42 std::fill(y.begin(), y.end(), UINT8_C(0xA5));
43
44 xnn_f32_qu8_cvt_params params;
45 init_params(¶ms,
46 25.0f /* scale */,
47 127 /* output zero point */,
48 std::numeric_limits<uint8_t>::min() + 1 /* output min */,
49 std::numeric_limits<uint8_t>::max() - 1 /* output max */);
50 for (auto _ : state) {
51 cvt(num_elements * sizeof(uint8_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(uint8_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(f32_qu8_vcvt, neonv8_x8,
70 xnn_f32_qu8_vcvt_ukernel__neonv8_x8,
71 xnn_init_f32_qu8_cvt_neonv8_params,
72 benchmark::utils::CheckNEONV8)
73 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
74 ->UseRealTime();
75 BENCHMARK_CAPTURE(f32_qu8_vcvt, neonv8_x16,
76 xnn_f32_qu8_vcvt_ukernel__neonv8_x16,
77 xnn_init_f32_qu8_cvt_neonv8_params,
78 benchmark::utils::CheckNEONV8)
79 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
80 ->UseRealTime();
81 BENCHMARK_CAPTURE(f32_qu8_vcvt, neonv8_x24,
82 xnn_f32_qu8_vcvt_ukernel__neonv8_x24,
83 xnn_init_f32_qu8_cvt_neonv8_params,
84 benchmark::utils::CheckNEONV8)
85 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
86 ->UseRealTime();
87 BENCHMARK_CAPTURE(f32_qu8_vcvt, neonv8_x32,
88 xnn_f32_qu8_vcvt_ukernel__neonv8_x32,
89 xnn_init_f32_qu8_cvt_neonv8_params,
90 benchmark::utils::CheckNEONV8)
91 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
92 ->UseRealTime();
93
94 BENCHMARK_CAPTURE(f32_qu8_vcvt, neon_x8,
95 xnn_f32_qu8_vcvt_ukernel__neon_x8,
96 xnn_init_f32_qu8_cvt_neon_params,
97 benchmark::utils::CheckNEON)
98 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
99 ->UseRealTime();
100 BENCHMARK_CAPTURE(f32_qu8_vcvt, neon_x16,
101 xnn_f32_qu8_vcvt_ukernel__neon_x16,
102 xnn_init_f32_qu8_cvt_neon_params,
103 benchmark::utils::CheckNEON)
104 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
105 ->UseRealTime();
106 BENCHMARK_CAPTURE(f32_qu8_vcvt, neon_x24,
107 xnn_f32_qu8_vcvt_ukernel__neon_x24,
108 xnn_init_f32_qu8_cvt_neon_params,
109 benchmark::utils::CheckNEON)
110 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
111 ->UseRealTime();
112 BENCHMARK_CAPTURE(f32_qu8_vcvt, neon_x32,
113 xnn_f32_qu8_vcvt_ukernel__neon_x32,
114 xnn_init_f32_qu8_cvt_neon_params,
115 benchmark::utils::CheckNEON)
116 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
117 ->UseRealTime();
118 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
119
120 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
121 BENCHMARK_CAPTURE(f32_qu8_vcvt, avx512skx_x32,
122 xnn_f32_qu8_vcvt_ukernel__avx512skx_x32,
123 xnn_init_f32_qu8_cvt_avx512_params,
124 benchmark::utils::CheckAVX512SKX)
125 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, int8_t>)
126 ->UseRealTime();
127 BENCHMARK_CAPTURE(f32_qu8_vcvt, avx512skx_x64,
128 xnn_f32_qu8_vcvt_ukernel__avx512skx_x64,
129 xnn_init_f32_qu8_cvt_avx512_params,
130 benchmark::utils::CheckAVX512SKX)
131 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, int8_t>)
132 ->UseRealTime();
133 BENCHMARK_CAPTURE(f32_qu8_vcvt, avx512skx_x96,
134 xnn_f32_qu8_vcvt_ukernel__avx512skx_x96,
135 xnn_init_f32_qu8_cvt_avx512_params,
136 benchmark::utils::CheckAVX512SKX)
137 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, int8_t>)
138 ->UseRealTime();
139 BENCHMARK_CAPTURE(f32_qu8_vcvt, avx512skx_x128,
140 xnn_f32_qu8_vcvt_ukernel__avx512skx_x128,
141 xnn_init_f32_qu8_cvt_avx512_params,
142 benchmark::utils::CheckAVX512SKX)
143 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, int8_t>)
144 ->UseRealTime();
145
146 BENCHMARK_CAPTURE(f32_qu8_vcvt, avx2_x16,
147 xnn_f32_qu8_vcvt_ukernel__avx2_x16,
148 xnn_init_f32_qu8_cvt_avx2_params,
149 benchmark::utils::CheckAVX2)
150 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
151 ->UseRealTime();
152 BENCHMARK_CAPTURE(f32_qu8_vcvt, avx2_x32,
153 xnn_f32_qu8_vcvt_ukernel__avx2_x32,
154 xnn_init_f32_qu8_cvt_avx2_params,
155 benchmark::utils::CheckAVX2)
156 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
157 ->UseRealTime();
158 BENCHMARK_CAPTURE(f32_qu8_vcvt, avx2_x48,
159 xnn_f32_qu8_vcvt_ukernel__avx2_x48,
160 xnn_init_f32_qu8_cvt_avx2_params,
161 benchmark::utils::CheckAVX2)
162 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
163 ->UseRealTime();
164 BENCHMARK_CAPTURE(f32_qu8_vcvt, avx2_x64,
165 xnn_f32_qu8_vcvt_ukernel__avx2_x64,
166 xnn_init_f32_qu8_cvt_avx2_params,
167 benchmark::utils::CheckAVX2)
168 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
169 ->UseRealTime();
170
171 BENCHMARK_CAPTURE(f32_qu8_vcvt, avx_x8,
172 xnn_f32_qu8_vcvt_ukernel__avx_x8,
173 xnn_init_f32_qu8_cvt_avx_params,
174 benchmark::utils::CheckAVX)
175 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
176 ->UseRealTime();
177 BENCHMARK_CAPTURE(f32_qu8_vcvt, avx_x16,
178 xnn_f32_qu8_vcvt_ukernel__avx_x16,
179 xnn_init_f32_qu8_cvt_avx_params,
180 benchmark::utils::CheckAVX)
181 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
182 ->UseRealTime();
183 BENCHMARK_CAPTURE(f32_qu8_vcvt, avx_x24,
184 xnn_f32_qu8_vcvt_ukernel__avx_x24,
185 xnn_init_f32_qu8_cvt_avx_params,
186 benchmark::utils::CheckAVX)
187 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
188 ->UseRealTime();
189 BENCHMARK_CAPTURE(f32_qu8_vcvt, avx_x32,
190 xnn_f32_qu8_vcvt_ukernel__avx_x32,
191 xnn_init_f32_qu8_cvt_avx_params,
192 benchmark::utils::CheckAVX)
193 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
194 ->UseRealTime();
195
196 BENCHMARK_CAPTURE(f32_qu8_vcvt, sse2_x8,
197 xnn_f32_qu8_vcvt_ukernel__sse2_x8,
198 xnn_init_f32_qu8_cvt_sse2_params)
199 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
200 ->UseRealTime();
201 BENCHMARK_CAPTURE(f32_qu8_vcvt, sse2_x16,
202 xnn_f32_qu8_vcvt_ukernel__sse2_x16,
203 xnn_init_f32_qu8_cvt_sse2_params)
204 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
205 ->UseRealTime();
206 BENCHMARK_CAPTURE(f32_qu8_vcvt, sse2_x24,
207 xnn_f32_qu8_vcvt_ukernel__sse2_x24,
208 xnn_init_f32_qu8_cvt_sse2_params)
209 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
210 ->UseRealTime();
211 BENCHMARK_CAPTURE(f32_qu8_vcvt, sse2_x32,
212 xnn_f32_qu8_vcvt_ukernel__sse2_x32,
213 xnn_init_f32_qu8_cvt_sse2_params)
214 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
215 ->UseRealTime();
216 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
217
218 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
219 BENCHMARK_CAPTURE(f32_qu8_vcvt, wasmsimd_cvt_x8,
220 xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x8,
221 xnn_init_f32_qu8_cvt_wasmsimd_cvt_params)
222 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
223 ->UseRealTime();
224 BENCHMARK_CAPTURE(f32_qu8_vcvt, wasmsimd_cvt_x16,
225 xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x16,
226 xnn_init_f32_qu8_cvt_wasmsimd_cvt_params)
227 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
228 ->UseRealTime();
229 BENCHMARK_CAPTURE(f32_qu8_vcvt, wasmsimd_cvt_x24,
230 xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x24,
231 xnn_init_f32_qu8_cvt_wasmsimd_cvt_params)
232 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
233 ->UseRealTime();
234 BENCHMARK_CAPTURE(f32_qu8_vcvt, wasmsimd_cvt_x32,
235 xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32,
236 xnn_init_f32_qu8_cvt_wasmsimd_cvt_params)
237 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
238 ->UseRealTime();
239
240 BENCHMARK_CAPTURE(f32_qu8_vcvt, wasmsimd_magic_x8,
241 xnn_f32_qu8_vcvt_ukernel__wasmsimd_magic_x8,
242 xnn_init_f32_qu8_cvt_wasmsimd_magic_params)
243 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
244 ->UseRealTime();
245 BENCHMARK_CAPTURE(f32_qu8_vcvt, wasmsimd_magic_x16,
246 xnn_f32_qu8_vcvt_ukernel__wasmsimd_magic_x16,
247 xnn_init_f32_qu8_cvt_wasmsimd_magic_params)
248 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
249 ->UseRealTime();
250 BENCHMARK_CAPTURE(f32_qu8_vcvt, wasmsimd_magic_x24,
251 xnn_f32_qu8_vcvt_ukernel__wasmsimd_magic_x24,
252 xnn_init_f32_qu8_cvt_wasmsimd_magic_params)
253 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
254 ->UseRealTime();
255 BENCHMARK_CAPTURE(f32_qu8_vcvt, wasmsimd_magic_x32,
256 xnn_f32_qu8_vcvt_ukernel__wasmsimd_magic_x32,
257 xnn_init_f32_qu8_cvt_wasmsimd_magic_params)
258 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
259 ->UseRealTime();
260 #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
261
262 #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
263 BENCHMARK_CAPTURE(f32_qu8_vcvt, wasm_fmagic_x1,
264 xnn_f32_qu8_vcvt_ukernel__wasm_fmagic_x1,
265 xnn_init_f32_qu8_cvt_scalar_fmagic_params)
266 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
267 ->UseRealTime();
268 BENCHMARK_CAPTURE(f32_qu8_vcvt, wasm_fmagic_x2,
269 xnn_f32_qu8_vcvt_ukernel__wasm_fmagic_x2,
270 xnn_init_f32_qu8_cvt_scalar_fmagic_params)
271 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
272 ->UseRealTime();
273 BENCHMARK_CAPTURE(f32_qu8_vcvt, wasm_fmagic_x3,
274 xnn_f32_qu8_vcvt_ukernel__wasm_fmagic_x3,
275 xnn_init_f32_qu8_cvt_scalar_fmagic_params)
276 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
277 ->UseRealTime();
278 BENCHMARK_CAPTURE(f32_qu8_vcvt, wasm_fmagic_x4,
279 xnn_f32_qu8_vcvt_ukernel__wasm_fmagic_x4,
280 xnn_init_f32_qu8_cvt_scalar_fmagic_params)
281 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
282 ->UseRealTime();
283 #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
284
285 BENCHMARK_CAPTURE(f32_qu8_vcvt, scalar_fmagic_x1,
286 xnn_f32_qu8_vcvt_ukernel__scalar_fmagic_x1,
287 xnn_init_f32_qu8_cvt_scalar_fmagic_params)
288 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
289 ->UseRealTime();
290 BENCHMARK_CAPTURE(f32_qu8_vcvt, scalar_fmagic_x2,
291 xnn_f32_qu8_vcvt_ukernel__scalar_fmagic_x2,
292 xnn_init_f32_qu8_cvt_scalar_fmagic_params)
293 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
294 ->UseRealTime();
295 BENCHMARK_CAPTURE(f32_qu8_vcvt, scalar_fmagic_x3,
296 xnn_f32_qu8_vcvt_ukernel__scalar_fmagic_x3,
297 xnn_init_f32_qu8_cvt_scalar_fmagic_params)
298 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
299 ->UseRealTime();
300 BENCHMARK_CAPTURE(f32_qu8_vcvt, scalar_fmagic_x4,
301 xnn_f32_qu8_vcvt_ukernel__scalar_fmagic_x4,
302 xnn_init_f32_qu8_cvt_scalar_fmagic_params)
303 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
304 ->UseRealTime();
305
306 BENCHMARK_CAPTURE(f32_qu8_vcvt, scalar_imagic_x1,
307 xnn_f32_qu8_vcvt_ukernel__scalar_imagic_x1,
308 xnn_init_f32_qu8_cvt_scalar_imagic_params)
309 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
310 ->UseRealTime();
311 BENCHMARK_CAPTURE(f32_qu8_vcvt, scalar_imagic_x2,
312 xnn_f32_qu8_vcvt_ukernel__scalar_imagic_x2,
313 xnn_init_f32_qu8_cvt_scalar_imagic_params)
314 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
315 ->UseRealTime();
316 BENCHMARK_CAPTURE(f32_qu8_vcvt, scalar_imagic_x3,
317 xnn_f32_qu8_vcvt_ukernel__scalar_imagic_x3,
318 xnn_init_f32_qu8_cvt_scalar_imagic_params)
319 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
320 ->UseRealTime();
321 BENCHMARK_CAPTURE(f32_qu8_vcvt, scalar_imagic_x4,
322 xnn_f32_qu8_vcvt_ukernel__scalar_imagic_x4,
323 xnn_init_f32_qu8_cvt_scalar_imagic_params)
324 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
325 ->UseRealTime();
326
327 BENCHMARK_CAPTURE(f32_qu8_vcvt, scalar_lrintf_x1,
328 xnn_f32_qu8_vcvt_ukernel__scalar_lrintf_x1,
329 xnn_init_f32_qu8_cvt_scalar_lrintf_params)
330 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
331 ->UseRealTime();
332 BENCHMARK_CAPTURE(f32_qu8_vcvt, scalar_lrintf_x2,
333 xnn_f32_qu8_vcvt_ukernel__scalar_lrintf_x2,
334 xnn_init_f32_qu8_cvt_scalar_lrintf_params)
335 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
336 ->UseRealTime();
337 BENCHMARK_CAPTURE(f32_qu8_vcvt, scalar_lrintf_x3,
338 xnn_f32_qu8_vcvt_ukernel__scalar_lrintf_x3,
339 xnn_init_f32_qu8_cvt_scalar_lrintf_params)
340 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
341 ->UseRealTime();
342 BENCHMARK_CAPTURE(f32_qu8_vcvt, scalar_lrintf_x4,
343 xnn_f32_qu8_vcvt_ukernel__scalar_lrintf_x4,
344 xnn_init_f32_qu8_cvt_scalar_lrintf_params)
345 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, uint8_t>)
346 ->UseRealTime();
347
348 #ifndef XNNPACK_BENCHMARK_NO_MAIN
349 BENCHMARK_MAIN();
350 #endif
351