Gitiles
Code Review Sign In
gerrit-public.fairphone.software / platform / external / XNNPACK / 4c617797f1334f52fde4d622f5332725b7ff83f8 / . / bench / qu8-vmulc.cc
blob: dfc1066bbbad942980c6ae7bc75014af9b4e054e [file] [log] [blame]
Marat Dukhan8b024c92021-08-03 00:05:14 -0700[diff] [blame]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
22static void qu8_vmulc(
23 benchmark::State& state,
24 xnn_qu8_vmul_minmax_ukernel_function vmulc,
25 xnn_init_qu8_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 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>> product(num_elements);
42 std::generate(a.begin(), a.end(), std::ref(u8rng));
43 const uint8_t b = u8rng();
44
45 union xnn_qu8_mul_minmax_params params;
46 init_params(&params,
47 127 /* a zero point */, 127 /* b zero point */, 127 /* output zero point */,
48 0.75f /* product-output scale */,
49 std::numeric_limits<uint8_t>::min() + 1, std::numeric_limits<uint8_t>::max() - 1);
50 for (auto _ : state) {
51 vmulc(num_elements * sizeof(uint8_t), a.data(), &b, product.data(), &params);
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_vmulc, neonv8_ld64_x8,
70 xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8,
71 xnn_init_qu8_mul_minmax_fp32_neonv8_params,
72 benchmark::utils::CheckNEONV8)
73 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
74 ->UseRealTime();
75 BENCHMARK_CAPTURE(qu8_vmulc, neonv8_ld64_x16,
76 xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x16,
77 xnn_init_qu8_mul_minmax_fp32_neonv8_params,
78 benchmark::utils::CheckNEONV8)
79 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
80 ->UseRealTime();
81
82 BENCHMARK_CAPTURE(qu8_vmulc, neonv8_ld128_x16,
83 xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16,
84 xnn_init_qu8_mul_minmax_fp32_neonv8_params,
85 benchmark::utils::CheckNEONV8)
86 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
87 ->UseRealTime();
88
89BENCHMARK_CAPTURE(qu8_vmulc, neon_ld64_x8,
90 xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8,
91 xnn_init_qu8_mul_minmax_fp32_neon_params,
92 benchmark::utils::CheckNEON)
93 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
94 ->UseRealTime();
95 BENCHMARK_CAPTURE(qu8_vmulc, neon_ld64_x16,
96 xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x16,
97 xnn_init_qu8_mul_minmax_fp32_neon_params,
98 benchmark::utils::CheckNEON)
99 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
100 ->UseRealTime();
101
102 BENCHMARK_CAPTURE(qu8_vmulc, neon_ld128_x16,
103 xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16,
104 xnn_init_qu8_mul_minmax_fp32_neon_params,
105 benchmark::utils::CheckNEON)
106 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
107 ->UseRealTime();
108#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
109
110#if XNN_ARCH_X86 || XNN_ARCH_X86_64
111 BENCHMARK_CAPTURE(qu8_vmulc, avx_mul16_ld64_x8,
112 xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8,
113 xnn_init_qu8_mul_minmax_fp32_sse2_params,
114 benchmark::utils::CheckAVX)
115 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
116 ->UseRealTime();
117 BENCHMARK_CAPTURE(qu8_vmulc, avx_mul16_ld64_x16,
118 xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16,
119 xnn_init_qu8_mul_minmax_fp32_sse2_params,
120 benchmark::utils::CheckAVX)
121 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
122 ->UseRealTime();
123
124 BENCHMARK_CAPTURE(qu8_vmulc, sse41_mul16_ld64_x8,
125 xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8,
126 xnn_init_qu8_mul_minmax_fp32_sse2_params,
127 benchmark::utils::CheckSSE41)
128 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
129 ->UseRealTime();
130 BENCHMARK_CAPTURE(qu8_vmulc, sse41_mul16_ld64_x16,
131 xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16,
132 xnn_init_qu8_mul_minmax_fp32_sse2_params,
133 benchmark::utils::CheckSSE41)
134 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
135 ->UseRealTime();
136
137 BENCHMARK_CAPTURE(qu8_vmulc, sse2_mul16_ld64_x8,
138 xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8,
139 xnn_init_qu8_mul_minmax_fp32_sse2_params)
140 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
141 ->UseRealTime();
142 BENCHMARK_CAPTURE(qu8_vmulc, sse2_mul16_ld64_x16,
143 xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16,
144 xnn_init_qu8_mul_minmax_fp32_sse2_params)
145 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
146 ->UseRealTime();
147#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
148
Marat Dukhan4c617792021-12-21 15:47:58 -0800[diff] [blame^]149#if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
Marat Dukhan8b024c92021-08-03 00:05:14 -0700[diff] [blame]150 BENCHMARK_CAPTURE(qu8_vmulc, wasmsimd_mul32_ld64_x8,
151 xnn_qu8_vmulc_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8,
152 xnn_init_qu8_mul_minmax_fp32_wasmsimd_params)
153 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
154 ->UseRealTime();
155 BENCHMARK_CAPTURE(qu8_vmulc, wasmsimd_mul32_ld64_x16,
156 xnn_qu8_vmulc_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x16,
157 xnn_init_qu8_mul_minmax_fp32_wasmsimd_params)
158 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
159 ->UseRealTime();
Marat Dukhan4c617792021-12-21 15:47:58 -0800[diff] [blame^]160#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
Marat Dukhan8b024c92021-08-03 00:05:14 -0700[diff] [blame]161
162BENCHMARK_CAPTURE(qu8_vmulc, scalar_x1,
163 xnn_qu8_vmulc_minmax_fp32_ukernel__scalar_x1,
164 xnn_init_qu8_mul_minmax_fp32_scalar_params)
165 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
166 ->UseRealTime();
167BENCHMARK_CAPTURE(qu8_vmulc, scalar_x2,
168 xnn_qu8_vmulc_minmax_fp32_ukernel__scalar_x2,
169 xnn_init_qu8_mul_minmax_fp32_scalar_params)
170 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
171 ->UseRealTime();
172BENCHMARK_CAPTURE(qu8_vmulc, scalar_x4,
173 xnn_qu8_vmulc_minmax_fp32_ukernel__scalar_x4,
174 xnn_init_qu8_mul_minmax_fp32_scalar_params)
175 ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
176 ->UseRealTime();
177
178#ifndef XNNPACK_BENCHMARK_NO_MAIN
179BENCHMARK_MAIN();
180#endif