xref: /aosp_15_r20/external/XNNPACK/bench/x8-lut.cc (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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 <array>
8 #include <cmath>
9 #include <functional>
10 #include <random>
11 #include <vector>
12 
13 #include <benchmark/benchmark.h>
14 #include "bench/utils.h"
15 
16 #include <xnnpack.h>
17 #include <xnnpack/aligned-allocator.h>
18 #include <xnnpack/common.h>
19 #include <xnnpack/lut.h>
20 #include <xnnpack/microfnptr.h>
21 
22 
x8_lut(benchmark::State & state,xnn_x8_lut_ukernel_function lut,benchmark::utils::IsaCheckFunction isa_check=nullptr)23 static void x8_lut(
24   benchmark::State& state,
25   xnn_x8_lut_ukernel_function lut,
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   std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> input(num_elements);
34   std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> output(num_elements);
35   std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> table(256);
36 
37   std::random_device random_device;
38   auto rng = std::mt19937(random_device());
39   auto u8rng = std::bind(
40     std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), std::ref(rng));
41   std::generate(input.begin(), input.end(), std::ref(u8rng));
42   std::generate(table.begin(), table.end(), std::ref(u8rng));
43   std::fill(output.begin(), output.end(), UINT8_C(0xAA));
44 
45   for (auto _ : state) {
46     lut(num_elements * sizeof(uint8_t), input.data(), output.data(), table.data());
47   }
48 
49   const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
50   if (cpu_frequency != 0) {
51     state.counters["cpufreq"] = cpu_frequency;
52   }
53 
54   const size_t elements_per_iteration = num_elements;
55   state.counters["elements"] =
56     benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);
57 
58   const size_t bytes_per_iteration = 2 * num_elements * sizeof(uint8_t);
59   state.counters["bytes"] =
60     benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
61 }
62 
63 #if XNN_ARCH_ARM64
64   BENCHMARK_CAPTURE(x8_lut, neon_tbx128x4_x16,
65                     xnn_x8_lut_ukernel__neon_tbx128x4_x16)
66     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
67     ->UseRealTime();
68   BENCHMARK_CAPTURE(x8_lut, neon_tbx128x4_x32,
69                     xnn_x8_lut_ukernel__neon_tbx128x4_x32)
70     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
71     ->UseRealTime();
72   BENCHMARK_CAPTURE(x8_lut, neon_tbx128x4_x48,
73                     xnn_x8_lut_ukernel__neon_tbx128x4_x48)
74     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
75     ->UseRealTime();
76   BENCHMARK_CAPTURE(x8_lut, neon_tbx128x4_x64,
77                     xnn_x8_lut_ukernel__neon_tbx128x4_x64)
78     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
79     ->UseRealTime();
80 #endif  // XNN_ARCH_ARM64
81 
82 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
83   BENCHMARK_CAPTURE(x8_lut, avx512skx_vpshufb_x64,
84                     xnn_x8_lut_ukernel__avx512skx_vpshufb_x64,
85                     benchmark::utils::CheckAVX512SKX)
86     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
87     ->UseRealTime();
88   BENCHMARK_CAPTURE(x8_lut, avx512skx_vpshufb_x128,
89                     xnn_x8_lut_ukernel__avx512skx_vpshufb_x128,
90                     benchmark::utils::CheckAVX512SKX)
91     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
92     ->UseRealTime();
93   BENCHMARK_CAPTURE(x8_lut, avx512skx_vpshufb_x192,
94                     xnn_x8_lut_ukernel__avx512skx_vpshufb_x192,
95                     benchmark::utils::CheckAVX512SKX)
96     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
97     ->UseRealTime();
98   BENCHMARK_CAPTURE(x8_lut, avx512skx_vpshufb_x256,
99                     xnn_x8_lut_ukernel__avx512skx_vpshufb_x256,
100                     benchmark::utils::CheckAVX512SKX)
101     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
102     ->UseRealTime();
103 
104   BENCHMARK_CAPTURE(x8_lut, avx2_x32,
105                     xnn_x8_lut_ukernel__avx2_x32,
106                     benchmark::utils::CheckAVX2)
107     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
108     ->UseRealTime();
109   BENCHMARK_CAPTURE(x8_lut, avx2_x64,
110                     xnn_x8_lut_ukernel__avx2_x64,
111                     benchmark::utils::CheckAVX2)
112     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
113     ->UseRealTime();
114   BENCHMARK_CAPTURE(x8_lut, avx2_x96,
115                     xnn_x8_lut_ukernel__avx2_x96,
116                     benchmark::utils::CheckAVX2)
117     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
118     ->UseRealTime();
119   BENCHMARK_CAPTURE(x8_lut, avx2_x128,
120                     xnn_x8_lut_ukernel__avx2_x128,
121                     benchmark::utils::CheckAVX2)
122     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
123     ->UseRealTime();
124 
125   BENCHMARK_CAPTURE(x8_lut, avx_x16,
126                     xnn_x8_lut_ukernel__avx_x16,
127                     benchmark::utils::CheckAVX)
128     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
129     ->UseRealTime();
130   BENCHMARK_CAPTURE(x8_lut, avx_x32,
131                     xnn_x8_lut_ukernel__avx_x32,
132                     benchmark::utils::CheckAVX)
133     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
134     ->UseRealTime();
135   BENCHMARK_CAPTURE(x8_lut, avx_x48,
136                     xnn_x8_lut_ukernel__avx_x48,
137                     benchmark::utils::CheckAVX)
138     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
139     ->UseRealTime();
140   BENCHMARK_CAPTURE(x8_lut, avx_x64,
141                     xnn_x8_lut_ukernel__avx_x64,
142                     benchmark::utils::CheckAVX)
143     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
144     ->UseRealTime();
145 
146   BENCHMARK_CAPTURE(x8_lut, ssse3_x16,
147                     xnn_x8_lut_ukernel__ssse3_x16,
148                     benchmark::utils::CheckSSSE3)
149     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
150     ->UseRealTime();
151   BENCHMARK_CAPTURE(x8_lut, ssse3_x32,
152                     xnn_x8_lut_ukernel__ssse3_x32,
153                     benchmark::utils::CheckSSSE3)
154     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
155     ->UseRealTime();
156 #endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
157 
158 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
159   BENCHMARK_CAPTURE(x8_lut, wasmsimd_x16,
160                     xnn_x8_lut_ukernel__wasmsimd_x16)
161     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
162     ->UseRealTime();
163   BENCHMARK_CAPTURE(x8_lut, wasmsimd_x32,
164                     xnn_x8_lut_ukernel__wasmsimd_x32)
165     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
166     ->UseRealTime();
167   BENCHMARK_CAPTURE(x8_lut, wasmsimd_x48,
168                     xnn_x8_lut_ukernel__wasmsimd_x48)
169     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
170     ->UseRealTime();
171   BENCHMARK_CAPTURE(x8_lut, wasmsimd_x64,
172                     xnn_x8_lut_ukernel__wasmsimd_x64)
173     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
174     ->UseRealTime();
175 #endif  // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
176 
177 BENCHMARK_CAPTURE(x8_lut, scalar_x1,
178                   xnn_x8_lut_ukernel__scalar_x1)
179   ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
180   ->UseRealTime();
181 BENCHMARK_CAPTURE(x8_lut, scalar_x2,
182                   xnn_x8_lut_ukernel__scalar_x2)
183   ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
184   ->UseRealTime();
185 BENCHMARK_CAPTURE(x8_lut, scalar_x4,
186                   xnn_x8_lut_ukernel__scalar_x4)
187   ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
188   ->UseRealTime();
189 BENCHMARK_CAPTURE(x8_lut, scalar_x8,
190                   xnn_x8_lut_ukernel__scalar_x8)
191   ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
192   ->UseRealTime();
193 BENCHMARK_CAPTURE(x8_lut, scalar_x16,
194                   xnn_x8_lut_ukernel__scalar_x16)
195   ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
196   ->UseRealTime();
197 
198 #ifndef XNNPACK_BENCHMARK_NO_MAIN
199 BENCHMARK_MAIN();
200 #endif
201