xref: /aosp_15_r20/external/XNNPACK/bench/x8-lut.cc (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Copyright 2021 Google LLC
2*4bdc9457SAndroid Build Coastguard Worker //
3*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
4*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
5*4bdc9457SAndroid Build Coastguard Worker 
6*4bdc9457SAndroid Build Coastguard Worker #include <algorithm>
7*4bdc9457SAndroid Build Coastguard Worker #include <array>
8*4bdc9457SAndroid Build Coastguard Worker #include <cmath>
9*4bdc9457SAndroid Build Coastguard Worker #include <functional>
10*4bdc9457SAndroid Build Coastguard Worker #include <random>
11*4bdc9457SAndroid Build Coastguard Worker #include <vector>
12*4bdc9457SAndroid Build Coastguard Worker 
13*4bdc9457SAndroid Build Coastguard Worker #include <benchmark/benchmark.h>
14*4bdc9457SAndroid Build Coastguard Worker #include "bench/utils.h"
15*4bdc9457SAndroid Build Coastguard Worker 
16*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack.h>
17*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/aligned-allocator.h>
18*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
19*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/lut.h>
20*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/microfnptr.h>
21*4bdc9457SAndroid Build Coastguard Worker 
22*4bdc9457SAndroid Build Coastguard Worker 
x8_lut(benchmark::State & state,xnn_x8_lut_ukernel_function lut,benchmark::utils::IsaCheckFunction isa_check=nullptr)23*4bdc9457SAndroid Build Coastguard Worker static void x8_lut(
24*4bdc9457SAndroid Build Coastguard Worker   benchmark::State& state,
25*4bdc9457SAndroid Build Coastguard Worker   xnn_x8_lut_ukernel_function lut,
26*4bdc9457SAndroid Build Coastguard Worker   benchmark::utils::IsaCheckFunction isa_check = nullptr)
27*4bdc9457SAndroid Build Coastguard Worker {
28*4bdc9457SAndroid Build Coastguard Worker   if (isa_check && !isa_check(state)) {
29*4bdc9457SAndroid Build Coastguard Worker     return;
30*4bdc9457SAndroid Build Coastguard Worker   }
31*4bdc9457SAndroid Build Coastguard Worker 
32*4bdc9457SAndroid Build Coastguard Worker   const size_t num_elements = state.range(0);
33*4bdc9457SAndroid Build Coastguard Worker   std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> input(num_elements);
34*4bdc9457SAndroid Build Coastguard Worker   std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> output(num_elements);
35*4bdc9457SAndroid Build Coastguard Worker   std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> table(256);
36*4bdc9457SAndroid Build Coastguard Worker 
37*4bdc9457SAndroid Build Coastguard Worker   std::random_device random_device;
38*4bdc9457SAndroid Build Coastguard Worker   auto rng = std::mt19937(random_device());
39*4bdc9457SAndroid Build Coastguard Worker   auto u8rng = std::bind(
40*4bdc9457SAndroid Build Coastguard Worker     std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), std::ref(rng));
41*4bdc9457SAndroid Build Coastguard Worker   std::generate(input.begin(), input.end(), std::ref(u8rng));
42*4bdc9457SAndroid Build Coastguard Worker   std::generate(table.begin(), table.end(), std::ref(u8rng));
43*4bdc9457SAndroid Build Coastguard Worker   std::fill(output.begin(), output.end(), UINT8_C(0xAA));
44*4bdc9457SAndroid Build Coastguard Worker 
45*4bdc9457SAndroid Build Coastguard Worker   for (auto _ : state) {
46*4bdc9457SAndroid Build Coastguard Worker     lut(num_elements * sizeof(uint8_t), input.data(), output.data(), table.data());
47*4bdc9457SAndroid Build Coastguard Worker   }
48*4bdc9457SAndroid Build Coastguard Worker 
49*4bdc9457SAndroid Build Coastguard Worker   const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
50*4bdc9457SAndroid Build Coastguard Worker   if (cpu_frequency != 0) {
51*4bdc9457SAndroid Build Coastguard Worker     state.counters["cpufreq"] = cpu_frequency;
52*4bdc9457SAndroid Build Coastguard Worker   }
53*4bdc9457SAndroid Build Coastguard Worker 
54*4bdc9457SAndroid Build Coastguard Worker   const size_t elements_per_iteration = num_elements;
55*4bdc9457SAndroid Build Coastguard Worker   state.counters["elements"] =
56*4bdc9457SAndroid Build Coastguard Worker     benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);
57*4bdc9457SAndroid Build Coastguard Worker 
58*4bdc9457SAndroid Build Coastguard Worker   const size_t bytes_per_iteration = 2 * num_elements * sizeof(uint8_t);
59*4bdc9457SAndroid Build Coastguard Worker   state.counters["bytes"] =
60*4bdc9457SAndroid Build Coastguard Worker     benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
61*4bdc9457SAndroid Build Coastguard Worker }
62*4bdc9457SAndroid Build Coastguard Worker 
63*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_ARM64
64*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, neon_tbx128x4_x16,
65*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__neon_tbx128x4_x16)
66*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
67*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
68*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, neon_tbx128x4_x32,
69*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__neon_tbx128x4_x32)
70*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
71*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
72*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, neon_tbx128x4_x48,
73*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__neon_tbx128x4_x48)
74*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
75*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
76*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, neon_tbx128x4_x64,
77*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__neon_tbx128x4_x64)
78*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
79*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
80*4bdc9457SAndroid Build Coastguard Worker #endif  // XNN_ARCH_ARM64
81*4bdc9457SAndroid Build Coastguard Worker 
82*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_X86 || XNN_ARCH_X86_64
83*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, avx512skx_vpshufb_x64,
84*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__avx512skx_vpshufb_x64,
85*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX512SKX)
86*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
87*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
88*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, avx512skx_vpshufb_x128,
89*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__avx512skx_vpshufb_x128,
90*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX512SKX)
91*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
92*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
93*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, avx512skx_vpshufb_x192,
94*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__avx512skx_vpshufb_x192,
95*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX512SKX)
96*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
97*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
98*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, avx512skx_vpshufb_x256,
99*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__avx512skx_vpshufb_x256,
100*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX512SKX)
101*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
102*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
103*4bdc9457SAndroid Build Coastguard Worker 
104*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, avx2_x32,
105*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__avx2_x32,
106*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
107*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
108*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
109*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, avx2_x64,
110*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__avx2_x64,
111*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
112*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
113*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
114*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, avx2_x96,
115*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__avx2_x96,
116*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
117*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
118*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
119*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, avx2_x128,
120*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__avx2_x128,
121*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX2)
122*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
123*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
124*4bdc9457SAndroid Build Coastguard Worker 
125*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, avx_x16,
126*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__avx_x16,
127*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX)
128*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
129*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
130*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, avx_x32,
131*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__avx_x32,
132*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX)
133*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
134*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
135*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, avx_x48,
136*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__avx_x48,
137*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX)
138*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
139*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
140*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, avx_x64,
141*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__avx_x64,
142*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckAVX)
143*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
144*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
145*4bdc9457SAndroid Build Coastguard Worker 
146*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, ssse3_x16,
147*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__ssse3_x16,
148*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckSSSE3)
149*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
150*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
151*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, ssse3_x32,
152*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__ssse3_x32,
153*4bdc9457SAndroid Build Coastguard Worker                     benchmark::utils::CheckSSSE3)
154*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
155*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
156*4bdc9457SAndroid Build Coastguard Worker #endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
157*4bdc9457SAndroid Build Coastguard Worker 
158*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
159*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, wasmsimd_x16,
160*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__wasmsimd_x16)
161*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
162*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
163*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, wasmsimd_x32,
164*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__wasmsimd_x32)
165*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
166*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
167*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, wasmsimd_x48,
168*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__wasmsimd_x48)
169*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
170*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
171*4bdc9457SAndroid Build Coastguard Worker   BENCHMARK_CAPTURE(x8_lut, wasmsimd_x64,
172*4bdc9457SAndroid Build Coastguard Worker                     xnn_x8_lut_ukernel__wasmsimd_x64)
173*4bdc9457SAndroid Build Coastguard Worker     ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
174*4bdc9457SAndroid Build Coastguard Worker     ->UseRealTime();
175*4bdc9457SAndroid Build Coastguard Worker #endif  // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
176*4bdc9457SAndroid Build Coastguard Worker 
177*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(x8_lut, scalar_x1,
178*4bdc9457SAndroid Build Coastguard Worker                   xnn_x8_lut_ukernel__scalar_x1)
179*4bdc9457SAndroid Build Coastguard Worker   ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
180*4bdc9457SAndroid Build Coastguard Worker   ->UseRealTime();
181*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(x8_lut, scalar_x2,
182*4bdc9457SAndroid Build Coastguard Worker                   xnn_x8_lut_ukernel__scalar_x2)
183*4bdc9457SAndroid Build Coastguard Worker   ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
184*4bdc9457SAndroid Build Coastguard Worker   ->UseRealTime();
185*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(x8_lut, scalar_x4,
186*4bdc9457SAndroid Build Coastguard Worker                   xnn_x8_lut_ukernel__scalar_x4)
187*4bdc9457SAndroid Build Coastguard Worker   ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
188*4bdc9457SAndroid Build Coastguard Worker   ->UseRealTime();
189*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(x8_lut, scalar_x8,
190*4bdc9457SAndroid Build Coastguard Worker                   xnn_x8_lut_ukernel__scalar_x8)
191*4bdc9457SAndroid Build Coastguard Worker   ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
192*4bdc9457SAndroid Build Coastguard Worker   ->UseRealTime();
193*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(x8_lut, scalar_x16,
194*4bdc9457SAndroid Build Coastguard Worker                   xnn_x8_lut_ukernel__scalar_x16)
195*4bdc9457SAndroid Build Coastguard Worker   ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>)
196*4bdc9457SAndroid Build Coastguard Worker   ->UseRealTime();
197*4bdc9457SAndroid Build Coastguard Worker 
198*4bdc9457SAndroid Build Coastguard Worker #ifndef XNNPACK_BENCHMARK_NO_MAIN
199*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_MAIN();
200*4bdc9457SAndroid Build Coastguard Worker #endif
201