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
f16_f32_vcvt(benchmark::State & state,xnn_f16_f32_vcvt_ukernel_function cvt,xnn_init_f16_f32_cvt_params_fn init_params=nullptr,benchmark::utils::IsaCheckFunction isa_check=nullptr)24 static void f16_f32_vcvt(
25 benchmark::State& state,
26 xnn_f16_f32_vcvt_ukernel_function cvt,
27 xnn_init_f16_f32_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 auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);
40
41 std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> x(num_elements + XNN_EXTRA_BYTES / sizeof(uint16_t));
42 std::vector<float, AlignedAllocator<float, 64>> y(num_elements);
43 std::generate(x.begin(), x.end(), std::ref(f16rng));
44 std::fill(y.begin(), y.end(), std::nanf(""));
45
46 xnn_f16_f32_cvt_params params;
47 if (init_params != nullptr) {
48 init_params(¶ms);
49 }
50 for (auto _ : state) {
51 cvt(num_elements * sizeof(uint16_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(uint16_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(f16_f32_vcvt, neonfp16_x8,
70 xnn_f16_f32_vcvt_ukernel__neonfp16_x8,
71 nullptr /* init params */,
72 benchmark::utils::CheckNEONFP16)
73 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
74 ->UseRealTime();
75 BENCHMARK_CAPTURE(f16_f32_vcvt, neonfp16_x16,
76 xnn_f16_f32_vcvt_ukernel__neonfp16_x16,
77 nullptr /* init params */,
78 benchmark::utils::CheckNEONFP16)
79 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
80 ->UseRealTime();
81
82 BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int16_x8,
83 xnn_f16_f32_vcvt_ukernel__neon_int16_x8,
84 xnn_init_f16_f32_cvt_neon_params,
85 benchmark::utils::CheckNEON)
86 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
87 ->UseRealTime();
88 BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int16_x16,
89 xnn_f16_f32_vcvt_ukernel__neon_int16_x16,
90 xnn_init_f16_f32_cvt_neon_params,
91 benchmark::utils::CheckNEON)
92 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
93 ->UseRealTime();
94 BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int16_x24,
95 xnn_f16_f32_vcvt_ukernel__neon_int16_x24,
96 xnn_init_f16_f32_cvt_neon_params,
97 benchmark::utils::CheckNEON)
98 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
99 ->UseRealTime();
100 BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int16_x32,
101 xnn_f16_f32_vcvt_ukernel__neon_int16_x32,
102 xnn_init_f16_f32_cvt_neon_params,
103 benchmark::utils::CheckNEON)
104 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
105 ->UseRealTime();
106
107 BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int32_x8,
108 xnn_f16_f32_vcvt_ukernel__neon_int32_x8,
109 xnn_init_f16_f32_cvt_neon_params,
110 benchmark::utils::CheckNEON)
111 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
112 ->UseRealTime();
113 BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int32_x16,
114 xnn_f16_f32_vcvt_ukernel__neon_int32_x16,
115 xnn_init_f16_f32_cvt_neon_params,
116 benchmark::utils::CheckNEON)
117 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
118 ->UseRealTime();
119 BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int32_x24,
120 xnn_f16_f32_vcvt_ukernel__neon_int32_x24,
121 xnn_init_f16_f32_cvt_neon_params,
122 benchmark::utils::CheckNEON)
123 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
124 ->UseRealTime();
125 BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int32_x32,
126 xnn_f16_f32_vcvt_ukernel__neon_int32_x32,
127 xnn_init_f16_f32_cvt_neon_params,
128 benchmark::utils::CheckNEON)
129 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
130 ->UseRealTime();
131 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
132
133 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
134 BENCHMARK_CAPTURE(f16_f32_vcvt, avx512skx_x16,
135 xnn_f16_f32_vcvt_ukernel__avx512skx_x16,
136 nullptr /* init params */,
137 benchmark::utils::CheckAVX512SKX)
138 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
139 ->UseRealTime();
140 BENCHMARK_CAPTURE(f16_f32_vcvt, avx512skx_x32,
141 xnn_f16_f32_vcvt_ukernel__avx512skx_x32,
142 nullptr /* init params */,
143 benchmark::utils::CheckAVX512SKX)
144 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
145 ->UseRealTime();
146
147 BENCHMARK_CAPTURE(f16_f32_vcvt, f16c_x8,
148 xnn_f16_f32_vcvt_ukernel__f16c_x8,
149 nullptr /* init params */,
150 benchmark::utils::CheckF16C)
151 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
152 ->UseRealTime();
153 BENCHMARK_CAPTURE(f16_f32_vcvt, f16c_x16,
154 xnn_f16_f32_vcvt_ukernel__f16c_x16,
155 nullptr /* init params */,
156 benchmark::utils::CheckF16C)
157 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
158 ->UseRealTime();
159
160 BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int16_x8,
161 xnn_f16_f32_vcvt_ukernel__avx_int16_x8,
162 xnn_init_f16_f32_cvt_sse_int16_params,
163 benchmark::utils::CheckAVX)
164 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
165 ->UseRealTime();
166 BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int16_x16,
167 xnn_f16_f32_vcvt_ukernel__avx_int16_x16,
168 xnn_init_f16_f32_cvt_sse_int16_params,
169 benchmark::utils::CheckAVX)
170 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
171 ->UseRealTime();
172 BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int16_x24,
173 xnn_f16_f32_vcvt_ukernel__avx_int16_x24,
174 xnn_init_f16_f32_cvt_sse_int16_params,
175 benchmark::utils::CheckAVX)
176 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
177 ->UseRealTime();
178 BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int16_x32,
179 xnn_f16_f32_vcvt_ukernel__avx_int16_x32,
180 xnn_init_f16_f32_cvt_sse_int16_params,
181 benchmark::utils::CheckAVX)
182 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
183 ->UseRealTime();
184
185 BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int32_x8,
186 xnn_f16_f32_vcvt_ukernel__avx_int32_x8,
187 xnn_init_f16_f32_cvt_sse_int32_params,
188 benchmark::utils::CheckAVX)
189 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
190 ->UseRealTime();
191 BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int32_x16,
192 xnn_f16_f32_vcvt_ukernel__avx_int32_x16,
193 xnn_init_f16_f32_cvt_sse_int32_params,
194 benchmark::utils::CheckAVX)
195 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
196 ->UseRealTime();
197 BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int32_x24,
198 xnn_f16_f32_vcvt_ukernel__avx_int32_x24,
199 xnn_init_f16_f32_cvt_sse_int32_params,
200 benchmark::utils::CheckAVX)
201 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
202 ->UseRealTime();
203 BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int32_x32,
204 xnn_f16_f32_vcvt_ukernel__avx_int32_x32,
205 xnn_init_f16_f32_cvt_sse_int32_params,
206 benchmark::utils::CheckAVX)
207 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
208 ->UseRealTime();
209
210 BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int16_x8,
211 xnn_f16_f32_vcvt_ukernel__sse41_int16_x8,
212 xnn_init_f16_f32_cvt_sse_int16_params,
213 benchmark::utils::CheckSSE41)
214 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
215 ->UseRealTime();
216 BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int16_x16,
217 xnn_f16_f32_vcvt_ukernel__sse41_int16_x16,
218 xnn_init_f16_f32_cvt_sse_int16_params,
219 benchmark::utils::CheckSSE41)
220 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
221 ->UseRealTime();
222 BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int16_x24,
223 xnn_f16_f32_vcvt_ukernel__sse41_int16_x24,
224 xnn_init_f16_f32_cvt_sse_int16_params,
225 benchmark::utils::CheckSSE41)
226 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
227 ->UseRealTime();
228 BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int16_x32,
229 xnn_f16_f32_vcvt_ukernel__sse41_int16_x32,
230 xnn_init_f16_f32_cvt_sse_int16_params,
231 benchmark::utils::CheckSSE41)
232 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
233 ->UseRealTime();
234
235 BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int32_x8,
236 xnn_f16_f32_vcvt_ukernel__sse41_int32_x8,
237 xnn_init_f16_f32_cvt_sse_int32_params,
238 benchmark::utils::CheckSSE41)
239 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
240 ->UseRealTime();
241 BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int32_x16,
242 xnn_f16_f32_vcvt_ukernel__sse41_int32_x16,
243 xnn_init_f16_f32_cvt_sse_int32_params,
244 benchmark::utils::CheckSSE41)
245 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
246 ->UseRealTime();
247 BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int32_x24,
248 xnn_f16_f32_vcvt_ukernel__sse41_int32_x24,
249 xnn_init_f16_f32_cvt_sse_int32_params,
250 benchmark::utils::CheckSSE41)
251 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
252 ->UseRealTime();
253 BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int32_x32,
254 xnn_f16_f32_vcvt_ukernel__sse41_int32_x32,
255 xnn_init_f16_f32_cvt_sse_int32_params,
256 benchmark::utils::CheckSSE41)
257 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
258 ->UseRealTime();
259
260 BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int16_x8,
261 xnn_f16_f32_vcvt_ukernel__sse2_int16_x8,
262 xnn_init_f16_f32_cvt_sse_int16_params)
263 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
264 ->UseRealTime();
265 BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int16_x16,
266 xnn_f16_f32_vcvt_ukernel__sse2_int16_x16,
267 xnn_init_f16_f32_cvt_sse_int16_params)
268 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
269 ->UseRealTime();
270 BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int16_x24,
271 xnn_f16_f32_vcvt_ukernel__sse2_int16_x24,
272 xnn_init_f16_f32_cvt_sse_int16_params)
273 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
274 ->UseRealTime();
275 BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int16_x32,
276 xnn_f16_f32_vcvt_ukernel__sse2_int16_x32,
277 xnn_init_f16_f32_cvt_sse_int16_params)
278 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
279 ->UseRealTime();
280
281 BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int32_x8,
282 xnn_f16_f32_vcvt_ukernel__sse2_int32_x8,
283 xnn_init_f16_f32_cvt_sse_int32_params)
284 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
285 ->UseRealTime();
286 BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int32_x16,
287 xnn_f16_f32_vcvt_ukernel__sse2_int32_x16,
288 xnn_init_f16_f32_cvt_sse_int32_params)
289 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
290 ->UseRealTime();
291 BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int32_x24,
292 xnn_f16_f32_vcvt_ukernel__sse2_int32_x24,
293 xnn_init_f16_f32_cvt_sse_int32_params)
294 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
295 ->UseRealTime();
296 BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int32_x32,
297 xnn_f16_f32_vcvt_ukernel__sse2_int32_x32,
298 xnn_init_f16_f32_cvt_sse_int32_params)
299 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
300 ->UseRealTime();
301 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
302
303 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
304 BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int16_x8,
305 xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x8,
306 xnn_init_f16_f32_cvt_wasmsimd_int16_params)
307 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
308 ->UseRealTime();
309 BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int16_x16,
310 xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16,
311 xnn_init_f16_f32_cvt_wasmsimd_int16_params)
312 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
313 ->UseRealTime();
314 BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int16_x24,
315 xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x24,
316 xnn_init_f16_f32_cvt_wasmsimd_int16_params)
317 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
318 ->UseRealTime();
319 BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int16_x32,
320 xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x32,
321 xnn_init_f16_f32_cvt_wasmsimd_int16_params)
322 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
323 ->UseRealTime();
324
325 BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int32_x8,
326 xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8,
327 xnn_init_f16_f32_cvt_wasmsimd_int32_params)
328 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
329 ->UseRealTime();
330 BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int32_x16,
331 xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16,
332 xnn_init_f16_f32_cvt_wasmsimd_int32_params)
333 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
334 ->UseRealTime();
335 BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int32_x24,
336 xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24,
337 xnn_init_f16_f32_cvt_wasmsimd_int32_params)
338 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
339 ->UseRealTime();
340 BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int32_x32,
341 xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32,
342 xnn_init_f16_f32_cvt_wasmsimd_int32_params)
343 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
344 ->UseRealTime();
345 #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
346
347 BENCHMARK_CAPTURE(f16_f32_vcvt, scalar_x1,
348 xnn_f16_f32_vcvt_ukernel__scalar_x1,
349 xnn_init_f16_f32_cvt_scalar_params)
350 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
351 ->UseRealTime();
352 BENCHMARK_CAPTURE(f16_f32_vcvt, scalar_x2,
353 xnn_f16_f32_vcvt_ukernel__scalar_x2,
354 xnn_init_f16_f32_cvt_scalar_params)
355 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
356 ->UseRealTime();
357 BENCHMARK_CAPTURE(f16_f32_vcvt, scalar_x3,
358 xnn_f16_f32_vcvt_ukernel__scalar_x3,
359 xnn_init_f16_f32_cvt_scalar_params)
360 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
361 ->UseRealTime();
362 BENCHMARK_CAPTURE(f16_f32_vcvt, scalar_x4,
363 xnn_f16_f32_vcvt_ukernel__scalar_x4,
364 xnn_init_f16_f32_cvt_scalar_params)
365 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>)
366 ->UseRealTime();
367
368 #ifndef XNNPACK_BENCHMARK_NO_MAIN
369 BENCHMARK_MAIN();
370 #endif
371