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/AlignedAllocator.h>
16 #include <xnnpack/common.h>
17 #include <xnnpack/params.h>
18 #include <xnnpack/params-init.h>
19 #include <xnnpack/vaddsub.h>
20
21
qu8_vaddc(benchmark::State & state,xnn_qu8_vaddsub_minmax_ukernel_function vaddc,xnn_init_qu8_addsub_minmax_params_fn init_params,benchmark::utils::IsaCheckFunction isa_check=nullptr)22 static void qu8_vaddc(
23 benchmark::State& state,
24 xnn_qu8_vaddsub_minmax_ukernel_function vaddc,
25 xnn_init_qu8_addsub_minmax_params_fn init_params,
26 benchmark::utils::IsaCheckFunction isa_check = nullptr)
27 {
28 if (isa_check && !isa_check(state)) {
29 return;
30 }
31
32 const size_t num_elements = state.range(0);
33
34 std::random_device random_device;
35 auto rng = std::mt19937(random_device());
36 auto u8rng = std::bind(
37 std::uniform_int_distribution<uint32_t>(std::numeric_limits<uint8_t>::min(), std::numeric_limits<uint8_t>::max()),
38 std::ref(rng));
39
40 std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> a(num_elements);
41 std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> sum(num_elements);
42 std::generate(a.begin(), a.end(), std::ref(u8rng));
43 const uint8_t b = u8rng();
44
45 union xnn_qu8_addsub_minmax_params params;
46 init_params(¶ms,
47 127 /* a zero point */, 127 /* b zero point */, 127 /* output zero point */,
48 0.5f /* a-output scale */, 0.75f /* b-output scale */,
49 std::numeric_limits<uint8_t>::min() + 1, std::numeric_limits<uint8_t>::max() - 1);
50 for (auto _ : state) {
51 vaddc(num_elements * sizeof(uint8_t), a.data(), &b, sum.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 num_elements_per_iteration = num_elements;
60 state.counters["num_elements"] =
61 benchmark::Counter(uint64_t(state.iterations()) * num_elements_per_iteration, benchmark::Counter::kIsRate);
62
63 const size_t bytes_per_iteration = 2 * num_elements * sizeof(int8_t);
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(qu8_vaddc, neon_ld64_x8,
70 xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8,
71 xnn_init_qu8_add_minmax_neon_params,
72 benchmark::utils::CheckNEON)
73 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
74 ->UseRealTime();
75 BENCHMARK_CAPTURE(qu8_vaddc, neon_ld64_x16,
76 xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x16,
77 xnn_init_qu8_add_minmax_neon_params,
78 benchmark::utils::CheckNEON)
79 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
80 ->UseRealTime();
81 BENCHMARK_CAPTURE(qu8_vaddc, neon_ld64_x32,
82 xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x32,
83 xnn_init_qu8_add_minmax_neon_params,
84 benchmark::utils::CheckNEON)
85 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
86 ->UseRealTime();
87
88 BENCHMARK_CAPTURE(qu8_vaddc, neon_ld128_x16,
89 xnn_qu8_vaddc_minmax_ukernel__neon_ld128_x16,
90 xnn_init_qu8_add_minmax_neon_params,
91 benchmark::utils::CheckNEON)
92 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
93 ->UseRealTime();
94 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
95
96 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
97 BENCHMARK_CAPTURE(qu8_vaddc, avx512skx_mul32_ld128_x16,
98 xnn_qu8_vaddc_minmax_ukernel__avx512skx_mul32_ld128_x16,
99 xnn_init_qu8_add_minmax_avx512_params,
100 benchmark::utils::CheckAVX512SKX)
101 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
102 ->UseRealTime();
103 BENCHMARK_CAPTURE(qu8_vaddc, avx512skx_mul32_ld128_x32,
104 xnn_qu8_vaddc_minmax_ukernel__avx512skx_mul32_ld128_x32,
105 xnn_init_qu8_add_minmax_avx512_params,
106 benchmark::utils::CheckAVX512SKX)
107 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
108 ->UseRealTime();
109
110 BENCHMARK_CAPTURE(qu8_vaddc, avx2_mul32_ld64_x8,
111 xnn_qu8_vaddc_minmax_ukernel__avx2_mul32_ld64_x8,
112 xnn_init_qu8_add_minmax_avx2_params,
113 benchmark::utils::CheckAVX2)
114 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
115 ->UseRealTime();
116 BENCHMARK_CAPTURE(qu8_vaddc, avx2_mul32_ld64_x16,
117 xnn_qu8_vaddc_minmax_ukernel__avx2_mul32_ld64_x16,
118 xnn_init_qu8_add_minmax_avx2_params,
119 benchmark::utils::CheckAVX2)
120 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
121 ->UseRealTime();
122
123 BENCHMARK_CAPTURE(qu8_vaddc, xop_mul32_ld32_x8,
124 xnn_qu8_vaddc_minmax_ukernel__xop_mul32_ld32_x8,
125 xnn_init_qu8_add_minmax_sse4_params,
126 benchmark::utils::CheckXOP)
127 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
128 ->UseRealTime();
129 BENCHMARK_CAPTURE(qu8_vaddc, xop_mul32_ld32_x16,
130 xnn_qu8_vaddc_minmax_ukernel__xop_mul32_ld32_x16,
131 xnn_init_qu8_add_minmax_sse4_params,
132 benchmark::utils::CheckXOP)
133 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
134 ->UseRealTime();
135
136 BENCHMARK_CAPTURE(qu8_vaddc, avx_mul16_ld64_x8,
137 xnn_qu8_vaddc_minmax_ukernel__avx_mul16_ld64_x8,
138 xnn_init_qu8_add_minmax_sse2_params,
139 benchmark::utils::CheckAVX)
140 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
141 ->UseRealTime();
142 BENCHMARK_CAPTURE(qu8_vaddc, avx_mul16_ld64_x16,
143 xnn_qu8_vaddc_minmax_ukernel__avx_mul16_ld64_x16,
144 xnn_init_qu8_add_minmax_sse2_params,
145 benchmark::utils::CheckAVX)
146 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
147 ->UseRealTime();
148
149 BENCHMARK_CAPTURE(qu8_vaddc, avx_mul32_ld32_x8,
150 xnn_qu8_vaddc_minmax_ukernel__avx_mul32_ld32_x8,
151 xnn_init_qu8_add_minmax_sse4_params,
152 benchmark::utils::CheckAVX)
153 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
154 ->UseRealTime();
155 BENCHMARK_CAPTURE(qu8_vaddc, avx_mul32_ld32_x16,
156 xnn_qu8_vaddc_minmax_ukernel__avx_mul32_ld32_x16,
157 xnn_init_qu8_add_minmax_sse4_params,
158 benchmark::utils::CheckAVX)
159 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
160 ->UseRealTime();
161
162 BENCHMARK_CAPTURE(qu8_vaddc, sse41_mul16_ld64_x8,
163 xnn_qu8_vaddc_minmax_ukernel__sse41_mul16_ld64_x8,
164 xnn_init_qu8_add_minmax_sse2_params,
165 benchmark::utils::CheckSSE41)
166 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
167 ->UseRealTime();
168 BENCHMARK_CAPTURE(qu8_vaddc, sse41_mul16_ld64_x16,
169 xnn_qu8_vaddc_minmax_ukernel__sse41_mul16_ld64_x16,
170 xnn_init_qu8_add_minmax_sse2_params,
171 benchmark::utils::CheckSSE41)
172 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
173 ->UseRealTime();
174
175 BENCHMARK_CAPTURE(qu8_vaddc, sse41_mul32_ld32_x8,
176 xnn_qu8_vaddc_minmax_ukernel__sse41_mul32_ld32_x8,
177 xnn_init_qu8_add_minmax_sse4_params,
178 benchmark::utils::CheckSSE41)
179 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
180 ->UseRealTime();
181 BENCHMARK_CAPTURE(qu8_vaddc, sse41_mul32_ld32_x16,
182 xnn_qu8_vaddc_minmax_ukernel__sse41_mul32_ld32_x16,
183 xnn_init_qu8_add_minmax_sse4_params,
184 benchmark::utils::CheckSSE41)
185 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
186 ->UseRealTime();
187
188 BENCHMARK_CAPTURE(qu8_vaddc, sse2_mul16_ld64_x8,
189 xnn_qu8_vaddc_minmax_ukernel__sse2_mul16_ld64_x8,
190 xnn_init_qu8_add_minmax_sse2_params)
191 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
192 ->UseRealTime();
193 BENCHMARK_CAPTURE(qu8_vaddc, sse2_mul16_ld64_x16,
194 xnn_qu8_vaddc_minmax_ukernel__sse2_mul16_ld64_x16,
195 xnn_init_qu8_add_minmax_sse2_params)
196 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
197 ->UseRealTime();
198 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
199
200 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
201 BENCHMARK_CAPTURE(qu8_vaddc, wasmsimd_x8,
202 xnn_qu8_vaddc_minmax_ukernel__wasmsimd_x8,
203 xnn_init_qu8_add_minmax_wasmsimd_params)
204 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
205 ->UseRealTime();
206 BENCHMARK_CAPTURE(qu8_vaddc, wasmsimd_x16,
207 xnn_qu8_vaddc_minmax_ukernel__wasmsimd_x16,
208 xnn_init_qu8_add_minmax_wasmsimd_params)
209 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
210 ->UseRealTime();
211 #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
212
213 BENCHMARK_CAPTURE(qu8_vaddc, scalar_x1,
214 xnn_qu8_vaddc_minmax_ukernel__scalar_x1,
215 xnn_init_qu8_add_minmax_scalar_params)
216 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
217 ->UseRealTime();
218 BENCHMARK_CAPTURE(qu8_vaddc, scalar_x2,
219 xnn_qu8_vaddc_minmax_ukernel__scalar_x2,
220 xnn_init_qu8_add_minmax_scalar_params)
221 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
222 ->UseRealTime();
223 BENCHMARK_CAPTURE(qu8_vaddc, scalar_x4,
224 xnn_qu8_vaddc_minmax_ukernel__scalar_x4,
225 xnn_init_qu8_add_minmax_scalar_params)
226 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
227 ->UseRealTime();
228
229 #ifndef XNNPACK_BENCHMARK_NO_MAIN
230 BENCHMARK_MAIN();
231 #endif
232