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/vmul.h>
21
22
qu8_vmul(benchmark::State & state,xnn_qu8_vmul_minmax_ukernel_function vmul,xnn_init_qu8_mul_minmax_params_fn init_params,benchmark::utils::IsaCheckFunction isa_check=nullptr)23 static void qu8_vmul(
24 benchmark::State& state,
25 xnn_qu8_vmul_minmax_ukernel_function vmul,
26 xnn_init_qu8_mul_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>> b(num_elements);
43 std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> product(num_elements);
44 std::generate(a.begin(), a.end(), std::ref(u8rng));
45 std::generate(b.begin(), b.end(), std::ref(u8rng));
46
47 union xnn_qu8_mul_minmax_params params;
48 init_params(¶ms,
49 127 /* a zero point */, 127 /* b zero point */, 127 /* output zero point */,
50 0.75f /* product-output scale */,
51 std::numeric_limits<uint8_t>::min() + 1, std::numeric_limits<uint8_t>::max() - 1);
52 for (auto _ : state) {
53 vmul(num_elements * sizeof(uint8_t), a.data(), b.data(), product.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(qu8_vmul, fp32__neonv8_ld64_x8,
72 xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8,
73 xnn_init_qu8_mul_minmax_fp32_neonv8_params,
74 benchmark::utils::CheckNEONV8)
75 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
76 ->UseRealTime();
77 BENCHMARK_CAPTURE(qu8_vmul, fp32__neonv8_ld64_x16,
78 xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x16,
79 xnn_init_qu8_mul_minmax_fp32_neonv8_params,
80 benchmark::utils::CheckNEONV8)
81 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
82 ->UseRealTime();
83
84 BENCHMARK_CAPTURE(qu8_vmul, fp32__neonv8_ld128_x16,
85 xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16,
86 xnn_init_qu8_mul_minmax_fp32_neonv8_params,
87 benchmark::utils::CheckNEONV8)
88 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
89 ->UseRealTime();
90
91 BENCHMARK_CAPTURE(qu8_vmul, fp32__neon_ld64_x8,
92 xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8,
93 xnn_init_qu8_mul_minmax_fp32_neon_params,
94 benchmark::utils::CheckNEON)
95 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
96 ->UseRealTime();
97 BENCHMARK_CAPTURE(qu8_vmul, fp32__neon_ld64_x16,
98 xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x16,
99 xnn_init_qu8_mul_minmax_fp32_neon_params,
100 benchmark::utils::CheckNEON)
101 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
102 ->UseRealTime();
103
104 BENCHMARK_CAPTURE(qu8_vmul, fp32__neon_ld128_x16,
105 xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld128_x16,
106 xnn_init_qu8_mul_minmax_fp32_neon_params,
107 benchmark::utils::CheckNEON)
108 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
109 ->UseRealTime();
110
111 BENCHMARK_CAPTURE(qu8_vmul, rndnu__neon_ld64_x8,
112 xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8,
113 xnn_init_qu8_mul_minmax_rndnu_neon_params,
114 benchmark::utils::CheckNEON)
115 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
116 ->UseRealTime();
117 BENCHMARK_CAPTURE(qu8_vmul, rndnu__neon_ld64_x16,
118 xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x16,
119 xnn_init_qu8_mul_minmax_rndnu_neon_params,
120 benchmark::utils::CheckNEON)
121 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
122 ->UseRealTime();
123
124 BENCHMARK_CAPTURE(qu8_vmul, rndnu__neon_ld128_x16,
125 xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld128_x16,
126 xnn_init_qu8_mul_minmax_rndnu_neon_params,
127 benchmark::utils::CheckNEON)
128 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
129 ->UseRealTime();
130 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
131
132 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
133 BENCHMARK_CAPTURE(qu8_vmul, avx_mul16_ld64_x8,
134 xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8,
135 xnn_init_qu8_mul_minmax_fp32_sse2_params,
136 benchmark::utils::CheckAVX)
137 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
138 ->UseRealTime();
139 BENCHMARK_CAPTURE(qu8_vmul, avx_mul16_ld64_x16,
140 xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16,
141 xnn_init_qu8_mul_minmax_fp32_sse2_params,
142 benchmark::utils::CheckAVX)
143 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
144 ->UseRealTime();
145
146 BENCHMARK_CAPTURE(qu8_vmul, sse41_mul16_ld64_x8,
147 xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8,
148 xnn_init_qu8_mul_minmax_fp32_sse2_params,
149 benchmark::utils::CheckSSE41)
150 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
151 ->UseRealTime();
152 BENCHMARK_CAPTURE(qu8_vmul, sse41_mul16_ld64_x16,
153 xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16,
154 xnn_init_qu8_mul_minmax_fp32_sse2_params,
155 benchmark::utils::CheckSSE41)
156 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
157 ->UseRealTime();
158
159 BENCHMARK_CAPTURE(qu8_vmul, sse2_mul16_ld64_x8,
160 xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8,
161 xnn_init_qu8_mul_minmax_fp32_sse2_params)
162 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
163 ->UseRealTime();
164 BENCHMARK_CAPTURE(qu8_vmul, sse2_mul16_ld64_x16,
165 xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16,
166 xnn_init_qu8_mul_minmax_fp32_sse2_params)
167 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
168 ->UseRealTime();
169 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
170
171 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
172 BENCHMARK_CAPTURE(qu8_vmul, wasmsimd_mul32_ld64_x8,
173 xnn_qu8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8,
174 xnn_init_qu8_mul_minmax_fp32_wasmsimd_params)
175 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
176 ->UseRealTime();
177 BENCHMARK_CAPTURE(qu8_vmul, wasmsimd_mul32_ld64_x16,
178 xnn_qu8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x16,
179 xnn_init_qu8_mul_minmax_fp32_wasmsimd_params)
180 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
181 ->UseRealTime();
182 #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
183
184 BENCHMARK_CAPTURE(qu8_vmul, scalar_x1,
185 xnn_qu8_vmul_minmax_fp32_ukernel__scalar_x1,
186 xnn_init_qu8_mul_minmax_fp32_scalar_params)
187 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
188 ->UseRealTime();
189 BENCHMARK_CAPTURE(qu8_vmul, scalar_x2,
190 xnn_qu8_vmul_minmax_fp32_ukernel__scalar_x2,
191 xnn_init_qu8_mul_minmax_fp32_scalar_params)
192 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
193 ->UseRealTime();
194 BENCHMARK_CAPTURE(qu8_vmul, scalar_x4,
195 xnn_qu8_vmul_minmax_fp32_ukernel__scalar_x4,
196 xnn_init_qu8_mul_minmax_fp32_scalar_params)
197 ->Apply(benchmark::utils::BinaryElementwiseParameters<uint8_t, uint8_t>)
198 ->UseRealTime();
199
200 #ifndef XNNPACK_BENCHMARK_NO_MAIN
201 BENCHMARK_MAIN();
202 #endif
203