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 <xnnpack.h>
16 #include <xnnpack/aligned-allocator.h>
17 #include <xnnpack/common.h>
18 #include <xnnpack/microfnptr.h>
19 #include <xnnpack/microparams-init.h>
20 #include <xnnpack/vadd.h>
21
22
qs8_vadd(benchmark::State & state,xnn_qs8_vadd_minmax_ukernel_function vadd,xnn_init_qs8_add_minmax_params_fn init_params,benchmark::utils::IsaCheckFunction isa_check=nullptr)23 static void qs8_vadd(
24 benchmark::State& state,
25 xnn_qs8_vadd_minmax_ukernel_function vadd,
26 xnn_init_qs8_add_minmax_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>> a(num_elements);
42 std::vector<int8_t, AlignedAllocator<int8_t, 64>> b(num_elements);
43 std::vector<int8_t, AlignedAllocator<int8_t, 64>> sum(num_elements);
44 std::generate(a.begin(), a.end(), std::ref(i8rng));
45 std::generate(b.begin(), b.end(), std::ref(i8rng));
46
47 union xnn_qs8_add_minmax_params params;
48 init_params(¶ms,
49 1 /* a zero point */, 1 /* b zero point */, 1 /* output zero point */,
50 0.5f /* a-output scale */, 0.75f /* b-output scale */,
51 std::numeric_limits<int8_t>::min() + 1, std::numeric_limits<int8_t>::max() - 1);
52 for (auto _ : state) {
53 vadd(num_elements * sizeof(int8_t), a.data(), b.data(), sum.data(), ¶ms);
54 }
55
56 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
57 if (cpu_frequency != 0) {
58 state.counters["cpufreq"] = cpu_frequency;
59 }
60
61 const size_t num_elements_per_iteration = num_elements;
62 state.counters["num_elements"] =
63 benchmark::Counter(uint64_t(state.iterations()) * num_elements_per_iteration, benchmark::Counter::kIsRate);
64
65 const size_t bytes_per_iteration = 3 * num_elements * sizeof(int8_t);
66 state.counters["bytes"] =
67 benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
68 }
69
70 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
71 BENCHMARK_CAPTURE(qs8_vadd, neon_ld64_x8,
72 xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8,
73 xnn_init_qs8_add_minmax_neon_params,
74 benchmark::utils::CheckNEON)
75 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
76 ->UseRealTime();
77 BENCHMARK_CAPTURE(qs8_vadd, neon_ld64_x16,
78 xnn_qs8_vadd_minmax_ukernel__neon_ld64_x16,
79 xnn_init_qs8_add_minmax_neon_params,
80 benchmark::utils::CheckNEON)
81 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
82 ->UseRealTime();
83 BENCHMARK_CAPTURE(qs8_vadd, neon_ld64_x24,
84 xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24,
85 xnn_init_qs8_add_minmax_neon_params,
86 benchmark::utils::CheckNEON)
87 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
88 ->UseRealTime();
89 BENCHMARK_CAPTURE(qs8_vadd, neon_ld64_x32,
90 xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32,
91 xnn_init_qs8_add_minmax_neon_params,
92 benchmark::utils::CheckNEON)
93 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
94 ->UseRealTime();
95
96 BENCHMARK_CAPTURE(qs8_vadd, neon_ld128_x16,
97 xnn_qs8_vadd_minmax_ukernel__neon_ld128_x16,
98 xnn_init_qs8_add_minmax_neon_params,
99 benchmark::utils::CheckNEON)
100 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
101 ->UseRealTime();
102 BENCHMARK_CAPTURE(qs8_vadd, neon_ld128_x32,
103 xnn_qs8_vadd_minmax_ukernel__neon_ld128_x32,
104 xnn_init_qs8_add_minmax_neon_params,
105 benchmark::utils::CheckNEON)
106 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
107 ->UseRealTime();
108 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
109
110 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
111 BENCHMARK_CAPTURE(qs8_vadd, avx512skx_mul32_ld128_x16,
112 xnn_qs8_vadd_minmax_ukernel__avx512skx_mul32_ld128_x16,
113 xnn_init_qs8_add_minmax_avx512_params,
114 benchmark::utils::CheckAVX512SKX)
115 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
116 ->UseRealTime();
117 BENCHMARK_CAPTURE(qs8_vadd, avx512skx_mul32_ld128_x32,
118 xnn_qs8_vadd_minmax_ukernel__avx512skx_mul32_ld128_x32,
119 xnn_init_qs8_add_minmax_avx512_params,
120 benchmark::utils::CheckAVX512SKX)
121 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
122 ->UseRealTime();
123
124 BENCHMARK_CAPTURE(qs8_vadd, avx2_mul32_ld64_x8,
125 xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x8,
126 xnn_init_qs8_add_minmax_avx2_params,
127 benchmark::utils::CheckAVX2)
128 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
129 ->UseRealTime();
130 BENCHMARK_CAPTURE(qs8_vadd, avx2_mul32_ld64_x16,
131 xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x16,
132 xnn_init_qs8_add_minmax_avx2_params,
133 benchmark::utils::CheckAVX2)
134 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
135 ->UseRealTime();
136 BENCHMARK_CAPTURE(qs8_vadd, avx2_mul32_ld64_x24,
137 xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x24,
138 xnn_init_qs8_add_minmax_avx2_params,
139 benchmark::utils::CheckAVX2)
140 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
141 ->UseRealTime();
142 BENCHMARK_CAPTURE(qs8_vadd, avx2_mul32_ld64_x32,
143 xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x32,
144 xnn_init_qs8_add_minmax_avx2_params,
145 benchmark::utils::CheckAVX2)
146 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
147 ->UseRealTime();
148
149 BENCHMARK_CAPTURE(qs8_vadd, xop_mul32_ld32_x8,
150 xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8,
151 xnn_init_qs8_add_minmax_sse4_mul32_params,
152 benchmark::utils::CheckXOP)
153 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
154 ->UseRealTime();
155 BENCHMARK_CAPTURE(qs8_vadd, xop_mul32_ld32_x16,
156 xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x16,
157 xnn_init_qs8_add_minmax_sse4_mul32_params,
158 benchmark::utils::CheckXOP)
159 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
160 ->UseRealTime();
161 BENCHMARK_CAPTURE(qs8_vadd, xop_mul32_ld32_x24,
162 xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x24,
163 xnn_init_qs8_add_minmax_sse4_mul32_params,
164 benchmark::utils::CheckXOP)
165 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
166 ->UseRealTime();
167 BENCHMARK_CAPTURE(qs8_vadd, xop_mul32_ld32_x32,
168 xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x32,
169 xnn_init_qs8_add_minmax_sse4_mul32_params,
170 benchmark::utils::CheckXOP)
171 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
172 ->UseRealTime();
173
174 BENCHMARK_CAPTURE(qs8_vadd, avx_mul16_ld64_x8,
175 xnn_qs8_vadd_minmax_ukernel__avx_mul16_ld64_x8,
176 xnn_init_qs8_add_minmax_sse4_mul16_params,
177 benchmark::utils::CheckAVX)
178 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
179 ->UseRealTime();
180 BENCHMARK_CAPTURE(qs8_vadd, avx_mul16_ld64_x16,
181 xnn_qs8_vadd_minmax_ukernel__avx_mul16_ld64_x16,
182 xnn_init_qs8_add_minmax_sse4_mul16_params,
183 benchmark::utils::CheckAVX)
184 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
185 ->UseRealTime();
186 BENCHMARK_CAPTURE(qs8_vadd, avx_mul16_ld64_x24,
187 xnn_qs8_vadd_minmax_ukernel__avx_mul16_ld64_x24,
188 xnn_init_qs8_add_minmax_sse4_mul16_params,
189 benchmark::utils::CheckAVX)
190 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
191 ->UseRealTime();
192 BENCHMARK_CAPTURE(qs8_vadd, avx_mul16_ld64_x32,
193 xnn_qs8_vadd_minmax_ukernel__avx_mul16_ld64_x32,
194 xnn_init_qs8_add_minmax_sse4_mul16_params,
195 benchmark::utils::CheckAVX)
196 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
197 ->UseRealTime();
198
199 BENCHMARK_CAPTURE(qs8_vadd, avx_mul32_ld32_x8,
200 xnn_qs8_vadd_minmax_ukernel__avx_mul32_ld32_x8,
201 xnn_init_qs8_add_minmax_sse4_mul32_params,
202 benchmark::utils::CheckAVX)
203 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
204 ->UseRealTime();
205 BENCHMARK_CAPTURE(qs8_vadd, avx_mul32_ld32_x16,
206 xnn_qs8_vadd_minmax_ukernel__avx_mul32_ld32_x16,
207 xnn_init_qs8_add_minmax_sse4_mul32_params,
208 benchmark::utils::CheckAVX)
209 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
210 ->UseRealTime();
211 BENCHMARK_CAPTURE(qs8_vadd, avx_mul32_ld32_x24,
212 xnn_qs8_vadd_minmax_ukernel__avx_mul32_ld32_x24,
213 xnn_init_qs8_add_minmax_sse4_mul32_params,
214 benchmark::utils::CheckAVX)
215 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
216 ->UseRealTime();
217 BENCHMARK_CAPTURE(qs8_vadd, avx_mul32_ld32_x32,
218 xnn_qs8_vadd_minmax_ukernel__avx_mul32_ld32_x32,
219 xnn_init_qs8_add_minmax_sse4_mul32_params,
220 benchmark::utils::CheckAVX)
221 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
222 ->UseRealTime();
223
224 BENCHMARK_CAPTURE(qs8_vadd, sse41_mul16_ld64_x8,
225 xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x8,
226 xnn_init_qs8_add_minmax_sse4_mul16_params,
227 benchmark::utils::CheckSSE41)
228 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
229 ->UseRealTime();
230 BENCHMARK_CAPTURE(qs8_vadd, sse41_mul16_ld64_x16,
231 xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x16,
232 xnn_init_qs8_add_minmax_sse4_mul16_params,
233 benchmark::utils::CheckSSE41)
234 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
235 ->UseRealTime();
236 BENCHMARK_CAPTURE(qs8_vadd, sse41_mul16_ld64_x24,
237 xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x24,
238 xnn_init_qs8_add_minmax_sse4_mul16_params,
239 benchmark::utils::CheckSSE41)
240 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
241 ->UseRealTime();
242 BENCHMARK_CAPTURE(qs8_vadd, sse41_mul16_ld64_x32,
243 xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32,
244 xnn_init_qs8_add_minmax_sse4_mul16_params,
245 benchmark::utils::CheckSSE41)
246 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
247 ->UseRealTime();
248
249 BENCHMARK_CAPTURE(qs8_vadd, sse41_mul32_ld32_x8,
250 xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x8,
251 xnn_init_qs8_add_minmax_sse4_mul32_params,
252 benchmark::utils::CheckSSE41)
253 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
254 ->UseRealTime();
255 BENCHMARK_CAPTURE(qs8_vadd, sse41_mul32_ld32_x16,
256 xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16,
257 xnn_init_qs8_add_minmax_sse4_mul32_params,
258 benchmark::utils::CheckSSE41)
259 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
260 ->UseRealTime();
261 BENCHMARK_CAPTURE(qs8_vadd, sse41_mul32_ld32_x24,
262 xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x24,
263 xnn_init_qs8_add_minmax_sse4_mul32_params,
264 benchmark::utils::CheckSSE41)
265 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
266 ->UseRealTime();
267 BENCHMARK_CAPTURE(qs8_vadd, sse41_mul32_ld32_x32,
268 xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32,
269 xnn_init_qs8_add_minmax_sse4_mul32_params,
270 benchmark::utils::CheckSSE41)
271 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
272 ->UseRealTime();
273
274 BENCHMARK_CAPTURE(qs8_vadd, sse2_mul16_ld64_x8,
275 xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x8,
276 xnn_init_qs8_add_minmax_sse2_params)
277 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
278 ->UseRealTime();
279 BENCHMARK_CAPTURE(qs8_vadd, sse2_mul16_ld64_x16,
280 xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x16,
281 xnn_init_qs8_add_minmax_sse2_params)
282 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
283 ->UseRealTime();
284 BENCHMARK_CAPTURE(qs8_vadd, sse2_mul16_ld64_x24,
285 xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x24,
286 xnn_init_qs8_add_minmax_sse2_params)
287 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
288 ->UseRealTime();
289 BENCHMARK_CAPTURE(qs8_vadd, sse2_mul16_ld64_x32,
290 xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32,
291 xnn_init_qs8_add_minmax_sse2_params)
292 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
293 ->UseRealTime();
294 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
295
296 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
297 BENCHMARK_CAPTURE(qs8_vadd, wasmsimd_x8,
298 xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8,
299 xnn_init_qs8_add_minmax_wasmsimd_params)
300 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
301 ->UseRealTime();
302 BENCHMARK_CAPTURE(qs8_vadd, wasmsimd_x16,
303 xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16,
304 xnn_init_qs8_add_minmax_wasmsimd_params)
305 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
306 ->UseRealTime();
307 BENCHMARK_CAPTURE(qs8_vadd, wasmsimd_x24,
308 xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24,
309 xnn_init_qs8_add_minmax_wasmsimd_params)
310 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
311 ->UseRealTime();
312 BENCHMARK_CAPTURE(qs8_vadd, wasmsimd_x32,
313 xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32,
314 xnn_init_qs8_add_minmax_wasmsimd_params)
315 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
316 ->UseRealTime();
317 #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
318
319 BENCHMARK_CAPTURE(qs8_vadd, scalar_x1,
320 xnn_qs8_vadd_minmax_ukernel__scalar_x1,
321 xnn_init_qs8_add_minmax_scalar_params)
322 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
323 ->UseRealTime();
324 BENCHMARK_CAPTURE(qs8_vadd, scalar_x2,
325 xnn_qs8_vadd_minmax_ukernel__scalar_x2,
326 xnn_init_qs8_add_minmax_scalar_params)
327 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
328 ->UseRealTime();
329 BENCHMARK_CAPTURE(qs8_vadd, scalar_x4,
330 xnn_qs8_vadd_minmax_ukernel__scalar_x4,
331 xnn_init_qs8_add_minmax_scalar_params)
332 ->Apply(benchmark::utils::BinaryElementwiseParameters<int8_t, int8_t>)
333 ->UseRealTime();
334
335 #ifndef XNNPACK_BENCHMARK_NO_MAIN
336 BENCHMARK_MAIN();
337 #endif
338