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/vmul.h>
20
21
qs8_vmulc(benchmark::State & state,xnn_qs8_vmul_minmax_ukernel_function vmulc,xnn_init_qs8_mul_minmax_params_fn init_params,benchmark::utils::IsaCheckFunction isa_check=nullptr)22 static void qs8_vmulc(
23 benchmark::State& state,
24 xnn_qs8_vmul_minmax_ukernel_function vmulc,
25 xnn_init_qs8_mul_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 i8rng = std::bind(
37 std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()),
38 std::ref(rng));
39
40 std::vector<int8_t, AlignedAllocator<int8_t, 64>> a(num_elements);
41 std::vector<int8_t, AlignedAllocator<int8_t, 64>> product(num_elements);
42 std::generate(a.begin(), a.end(), std::ref(i8rng));
43 const int8_t b = i8rng();
44
45 union xnn_qs8_mul_minmax_params params;
46 init_params(¶ms,
47 1 /* a zero point */, 1 /* b zero point */, 1 /* output zero point */,
48 0.75f /* product-output scale */,
49 std::numeric_limits<int8_t>::min() + 1, std::numeric_limits<int8_t>::max() - 1);
50 for (auto _ : state) {
51 vmulc(num_elements * sizeof(int8_t), a.data(), &b, product.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(qs8_vmulc, fp32__neonv8_ld64_x8,
70 xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8,
71 xnn_init_qs8_mul_minmax_fp32_neonv8_params,
72 benchmark::utils::CheckNEONV8)
73 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
74 ->UseRealTime();
75 BENCHMARK_CAPTURE(qs8_vmulc, fp32__neonv8_ld64_x16,
76 xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x16,
77 xnn_init_qs8_mul_minmax_fp32_neonv8_params,
78 benchmark::utils::CheckNEONV8)
79 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
80 ->UseRealTime();
81
82 BENCHMARK_CAPTURE(qs8_vmulc, fp32__neonv8_ld128_x16,
83 xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16,
84 xnn_init_qs8_mul_minmax_fp32_neonv8_params,
85 benchmark::utils::CheckNEONV8)
86 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
87 ->UseRealTime();
88
89 BENCHMARK_CAPTURE(qs8_vmulc, fp32__neon_ld64_x8,
90 xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8,
91 xnn_init_qs8_mul_minmax_fp32_neon_params,
92 benchmark::utils::CheckNEON)
93 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
94 ->UseRealTime();
95 BENCHMARK_CAPTURE(qs8_vmulc, fp32__neon_ld64_x16,
96 xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x16,
97 xnn_init_qs8_mul_minmax_fp32_neon_params,
98 benchmark::utils::CheckNEON)
99 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
100 ->UseRealTime();
101
102 BENCHMARK_CAPTURE(qs8_vmulc, fp32__neon_ld128_x16,
103 xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld128_x16,
104 xnn_init_qs8_mul_minmax_fp32_neon_params,
105 benchmark::utils::CheckNEON)
106 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
107 ->UseRealTime();
108
109 BENCHMARK_CAPTURE(qs8_vmulc, rndnu__neon_ld64_x8,
110 xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8,
111 xnn_init_qs8_mul_minmax_rndnu_neon_params,
112 benchmark::utils::CheckNEON)
113 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
114 ->UseRealTime();
115 BENCHMARK_CAPTURE(qs8_vmulc, rndnu__neon_ld64_x16,
116 xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16,
117 xnn_init_qs8_mul_minmax_rndnu_neon_params,
118 benchmark::utils::CheckNEON)
119 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
120 ->UseRealTime();
121
122 BENCHMARK_CAPTURE(qs8_vmulc, rndnu__neon_ld128_x16,
123 xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16,
124 xnn_init_qs8_mul_minmax_rndnu_neon_params,
125 benchmark::utils::CheckNEON)
126 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
127 ->UseRealTime();
128 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
129
130 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
131 BENCHMARK_CAPTURE(qs8_vmulc, avx_mul16_ld64_x8,
132 xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8,
133 xnn_init_qs8_mul_minmax_fp32_sse4_params,
134 benchmark::utils::CheckAVX)
135 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
136 ->UseRealTime();
137 BENCHMARK_CAPTURE(qs8_vmulc, avx_mul16_ld64_x16,
138 xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16,
139 xnn_init_qs8_mul_minmax_fp32_sse4_params,
140 benchmark::utils::CheckAVX)
141 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
142 ->UseRealTime();
143
144 BENCHMARK_CAPTURE(qs8_vmulc, sse41_mul16_ld64_x8,
145 xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8,
146 xnn_init_qs8_mul_minmax_fp32_sse4_params,
147 benchmark::utils::CheckSSE41)
148 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
149 ->UseRealTime();
150 BENCHMARK_CAPTURE(qs8_vmulc, sse41_mul16_ld64_x16,
151 xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16,
152 xnn_init_qs8_mul_minmax_fp32_sse4_params,
153 benchmark::utils::CheckSSE41)
154 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
155 ->UseRealTime();
156
157 BENCHMARK_CAPTURE(qs8_vmulc, sse2_mul16_ld64_x8,
158 xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8,
159 xnn_init_qs8_mul_minmax_fp32_sse2_params)
160 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
161 ->UseRealTime();
162 BENCHMARK_CAPTURE(qs8_vmulc, sse2_mul16_ld64_x16,
163 xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16,
164 xnn_init_qs8_mul_minmax_fp32_sse2_params)
165 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
166 ->UseRealTime();
167 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
168
169 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
170 BENCHMARK_CAPTURE(qs8_vmulc, wasmsimd_mul32_ld64_x8,
171 xnn_qs8_vmulc_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8,
172 xnn_init_qs8_mul_minmax_fp32_wasmsimd_params)
173 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
174 ->UseRealTime();
175 BENCHMARK_CAPTURE(qs8_vmulc, wasmsimd_mul32_ld64_x16,
176 xnn_qs8_vmulc_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x16,
177 xnn_init_qs8_mul_minmax_fp32_wasmsimd_params)
178 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
179 ->UseRealTime();
180 #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
181
182 BENCHMARK_CAPTURE(qs8_vmulc, scalar_x1,
183 xnn_qs8_vmulc_minmax_fp32_ukernel__scalar_x1,
184 xnn_init_qs8_mul_minmax_fp32_scalar_params)
185 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
186 ->UseRealTime();
187 BENCHMARK_CAPTURE(qs8_vmulc, scalar_x2,
188 xnn_qs8_vmulc_minmax_fp32_ukernel__scalar_x2,
189 xnn_init_qs8_mul_minmax_fp32_scalar_params)
190 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
191 ->UseRealTime();
192 BENCHMARK_CAPTURE(qs8_vmulc, scalar_x4,
193 xnn_qs8_vmulc_minmax_fp32_ukernel__scalar_x4,
194 xnn_init_qs8_mul_minmax_fp32_scalar_params)
195 ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>)
196 ->UseRealTime();
197
198 #ifndef XNNPACK_BENCHMARK_NO_MAIN
199 BENCHMARK_MAIN();
200 #endif
201