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
qu8_vaddc(benchmark::State & state,xnn_qu8_vadd_minmax_ukernel_function vaddc,xnn_init_qu8_add_minmax_params_fn init_params,benchmark::utils::IsaCheckFunction isa_check=nullptr)23 static void qu8_vaddc(
24 benchmark::State& state,
25 xnn_qu8_vadd_minmax_ukernel_function vaddc,
26 xnn_init_qu8_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 u8rng = std::bind(
38 std::uniform_int_distribution<uint32_t>(std::numeric_limits<uint8_t>::min(), std::numeric_limits<uint8_t>::max()),
39 std::ref(rng));
40
41 std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> a(num_elements);
42 std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> sum(num_elements);
43 std::generate(a.begin(), a.end(), std::ref(u8rng));
44 const uint8_t b = u8rng();
45
46 union xnn_qu8_add_minmax_params params;
47 init_params(¶ms,
48 127 /* a zero point */, 127 /* b zero point */, 127 /* output zero point */,
49 0.5f /* a-output scale */, 0.75f /* b-output scale */,
50 std::numeric_limits<uint8_t>::min() + 1, std::numeric_limits<uint8_t>::max() - 1);
51 for (auto _ : state) {
52 vaddc(num_elements * sizeof(uint8_t), a.data(), &b, sum.data(), ¶ms);
53 }
54
55 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
56 if (cpu_frequency != 0) {
57 state.counters["cpufreq"] = cpu_frequency;
58 }
59
60 const size_t num_elements_per_iteration = num_elements;
61 state.counters["num_elements"] =
62 benchmark::Counter(uint64_t(state.iterations()) * num_elements_per_iteration, benchmark::Counter::kIsRate);
63
64 const size_t bytes_per_iteration = 2 * num_elements * sizeof(int8_t);
65 state.counters["bytes"] =
66 benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
67 }
68
69 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
70 BENCHMARK_CAPTURE(qu8_vaddc, neon_ld64_x8,
71 xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8,
72 xnn_init_qu8_add_minmax_neon_params,
73 benchmark::utils::CheckNEON)
74 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
75 ->UseRealTime();
76 BENCHMARK_CAPTURE(qu8_vaddc, neon_ld64_x16,
77 xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x16,
78 xnn_init_qu8_add_minmax_neon_params,
79 benchmark::utils::CheckNEON)
80 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
81 ->UseRealTime();
82 BENCHMARK_CAPTURE(qu8_vaddc, neon_ld64_x32,
83 xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x32,
84 xnn_init_qu8_add_minmax_neon_params,
85 benchmark::utils::CheckNEON)
86 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
87 ->UseRealTime();
88
89 BENCHMARK_CAPTURE(qu8_vaddc, neon_ld128_x16,
90 xnn_qu8_vaddc_minmax_ukernel__neon_ld128_x16,
91 xnn_init_qu8_add_minmax_neon_params,
92 benchmark::utils::CheckNEON)
93 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
94 ->UseRealTime();
95 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
96
97 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
98 BENCHMARK_CAPTURE(qu8_vaddc, avx512skx_mul32_ld128_x16,
99 xnn_qu8_vaddc_minmax_ukernel__avx512skx_mul32_ld128_x16,
100 xnn_init_qu8_add_minmax_avx512_params,
101 benchmark::utils::CheckAVX512SKX)
102 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
103 ->UseRealTime();
104 BENCHMARK_CAPTURE(qu8_vaddc, avx512skx_mul32_ld128_x32,
105 xnn_qu8_vaddc_minmax_ukernel__avx512skx_mul32_ld128_x32,
106 xnn_init_qu8_add_minmax_avx512_params,
107 benchmark::utils::CheckAVX512SKX)
108 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
109 ->UseRealTime();
110
111 BENCHMARK_CAPTURE(qu8_vaddc, avx2_mul32_ld64_x8,
112 xnn_qu8_vaddc_minmax_ukernel__avx2_mul32_ld64_x8,
113 xnn_init_qu8_add_minmax_avx2_params,
114 benchmark::utils::CheckAVX2)
115 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
116 ->UseRealTime();
117 BENCHMARK_CAPTURE(qu8_vaddc, avx2_mul32_ld64_x16,
118 xnn_qu8_vaddc_minmax_ukernel__avx2_mul32_ld64_x16,
119 xnn_init_qu8_add_minmax_avx2_params,
120 benchmark::utils::CheckAVX2)
121 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
122 ->UseRealTime();
123
124 BENCHMARK_CAPTURE(qu8_vaddc, xop_mul32_ld32_x8,
125 xnn_qu8_vaddc_minmax_ukernel__xop_mul32_ld32_x8,
126 xnn_init_qu8_add_minmax_sse4_params,
127 benchmark::utils::CheckXOP)
128 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
129 ->UseRealTime();
130 BENCHMARK_CAPTURE(qu8_vaddc, xop_mul32_ld32_x16,
131 xnn_qu8_vaddc_minmax_ukernel__xop_mul32_ld32_x16,
132 xnn_init_qu8_add_minmax_sse4_params,
133 benchmark::utils::CheckXOP)
134 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
135 ->UseRealTime();
136
137 BENCHMARK_CAPTURE(qu8_vaddc, avx_mul16_ld64_x8,
138 xnn_qu8_vaddc_minmax_ukernel__avx_mul16_ld64_x8,
139 xnn_init_qu8_add_minmax_sse2_params,
140 benchmark::utils::CheckAVX)
141 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
142 ->UseRealTime();
143 BENCHMARK_CAPTURE(qu8_vaddc, avx_mul16_ld64_x16,
144 xnn_qu8_vaddc_minmax_ukernel__avx_mul16_ld64_x16,
145 xnn_init_qu8_add_minmax_sse2_params,
146 benchmark::utils::CheckAVX)
147 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
148 ->UseRealTime();
149
150 BENCHMARK_CAPTURE(qu8_vaddc, avx_mul32_ld32_x8,
151 xnn_qu8_vaddc_minmax_ukernel__avx_mul32_ld32_x8,
152 xnn_init_qu8_add_minmax_sse4_params,
153 benchmark::utils::CheckAVX)
154 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
155 ->UseRealTime();
156 BENCHMARK_CAPTURE(qu8_vaddc, avx_mul32_ld32_x16,
157 xnn_qu8_vaddc_minmax_ukernel__avx_mul32_ld32_x16,
158 xnn_init_qu8_add_minmax_sse4_params,
159 benchmark::utils::CheckAVX)
160 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
161 ->UseRealTime();
162
163 BENCHMARK_CAPTURE(qu8_vaddc, sse41_mul16_ld64_x8,
164 xnn_qu8_vaddc_minmax_ukernel__sse41_mul16_ld64_x8,
165 xnn_init_qu8_add_minmax_sse2_params,
166 benchmark::utils::CheckSSE41)
167 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
168 ->UseRealTime();
169 BENCHMARK_CAPTURE(qu8_vaddc, sse41_mul16_ld64_x16,
170 xnn_qu8_vaddc_minmax_ukernel__sse41_mul16_ld64_x16,
171 xnn_init_qu8_add_minmax_sse2_params,
172 benchmark::utils::CheckSSE41)
173 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
174 ->UseRealTime();
175
176 BENCHMARK_CAPTURE(qu8_vaddc, sse41_mul32_ld32_x8,
177 xnn_qu8_vaddc_minmax_ukernel__sse41_mul32_ld32_x8,
178 xnn_init_qu8_add_minmax_sse4_params,
179 benchmark::utils::CheckSSE41)
180 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
181 ->UseRealTime();
182 BENCHMARK_CAPTURE(qu8_vaddc, sse41_mul32_ld32_x16,
183 xnn_qu8_vaddc_minmax_ukernel__sse41_mul32_ld32_x16,
184 xnn_init_qu8_add_minmax_sse4_params,
185 benchmark::utils::CheckSSE41)
186 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
187 ->UseRealTime();
188
189 BENCHMARK_CAPTURE(qu8_vaddc, sse2_mul16_ld64_x8,
190 xnn_qu8_vaddc_minmax_ukernel__sse2_mul16_ld64_x8,
191 xnn_init_qu8_add_minmax_sse2_params)
192 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
193 ->UseRealTime();
194 BENCHMARK_CAPTURE(qu8_vaddc, sse2_mul16_ld64_x16,
195 xnn_qu8_vaddc_minmax_ukernel__sse2_mul16_ld64_x16,
196 xnn_init_qu8_add_minmax_sse2_params)
197 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
198 ->UseRealTime();
199 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
200
201 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
202 BENCHMARK_CAPTURE(qu8_vaddc, wasmsimd_x8,
203 xnn_qu8_vaddc_minmax_ukernel__wasmsimd_x8,
204 xnn_init_qu8_add_minmax_wasmsimd_params)
205 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
206 ->UseRealTime();
207 BENCHMARK_CAPTURE(qu8_vaddc, wasmsimd_x16,
208 xnn_qu8_vaddc_minmax_ukernel__wasmsimd_x16,
209 xnn_init_qu8_add_minmax_wasmsimd_params)
210 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
211 ->UseRealTime();
212 #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
213
214 BENCHMARK_CAPTURE(qu8_vaddc, scalar_x1,
215 xnn_qu8_vaddc_minmax_ukernel__scalar_x1,
216 xnn_init_qu8_add_minmax_scalar_params)
217 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
218 ->UseRealTime();
219 BENCHMARK_CAPTURE(qu8_vaddc, scalar_x2,
220 xnn_qu8_vaddc_minmax_ukernel__scalar_x2,
221 xnn_init_qu8_add_minmax_scalar_params)
222 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
223 ->UseRealTime();
224 BENCHMARK_CAPTURE(qu8_vaddc, scalar_x4,
225 xnn_qu8_vaddc_minmax_ukernel__scalar_x4,
226 xnn_init_qu8_add_minmax_scalar_params)
227 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
228 ->UseRealTime();
229
230 #ifndef XNNPACK_BENCHMARK_NO_MAIN
231 BENCHMARK_MAIN();
232 #endif
233