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 <numeric>
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/transpose.h>
20
21
transpose(benchmark::State & state,xnn_x32_transposec_ukernel_function transpose,benchmark::utils::IsaCheckFunction isa_check=nullptr)22 void transpose(
23 benchmark::State& state,
24 xnn_x32_transposec_ukernel_function transpose,
25 benchmark::utils::IsaCheckFunction isa_check = nullptr)
26 {
27 if (isa_check && !isa_check(state)) {
28 return;
29 }
30 const size_t height = state.range(0);
31 const size_t width = state.range(1);
32 const size_t tile_hbytes = height * sizeof(uint32_t);
33 const size_t tile_wbytes = width * sizeof(uint32_t);
34
35 std::vector<uint32_t, AlignedAllocator<uint32_t, 64>> x(
36 height * width + XNN_EXTRA_BYTES / sizeof(uint32_t));
37 std::vector<uint32_t, AlignedAllocator<uint32_t, 64>> y(
38 height * width + XNN_EXTRA_BYTES / sizeof(uint32_t));
39 std::iota(x.begin(), x.end(), 0);
40 std::fill(y.begin(), y.end(), 0);
41
42 for (auto _ : state) {
43 transpose(x.data(), y.data(), tile_wbytes, tile_hbytes, width,
44 height);
45 }
46
47 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
48 if (cpu_frequency != 0) {
49 state.counters["cpufreq"] = cpu_frequency;
50 }
51 }
52
BenchmarkKernelSize(benchmark::internal::Benchmark * b)53 static void BenchmarkKernelSize(benchmark::internal::Benchmark* b)
54 {
55 b->ArgNames({"height", "width"});
56 b->Args({32, 32});
57 b->Args({64, 64});
58 b->Args({117, 117});
59 b->Args({1024, 1024});
60 }
61
62 BENCHMARK_CAPTURE(transpose, 1x2_scalar_int, xnn_x32_transposec_ukernel__1x2_scalar_int)
63 ->Apply(BenchmarkKernelSize)->UseRealTime();
64 BENCHMARK_CAPTURE(transpose, 1x4_scalar_int, xnn_x32_transposec_ukernel__1x4_scalar_int)
65 ->Apply(BenchmarkKernelSize)->UseRealTime();
66 BENCHMARK_CAPTURE(transpose, 2x1_scalar_int, xnn_x32_transposec_ukernel__2x1_scalar_int)
67 ->Apply(BenchmarkKernelSize)->UseRealTime();
68 BENCHMARK_CAPTURE(transpose, 2x2_scalar_int, xnn_x32_transposec_ukernel__2x2_scalar_int)
69 ->Apply(BenchmarkKernelSize)->UseRealTime();
70 BENCHMARK_CAPTURE(transpose, 2x4_scalar_int, xnn_x32_transposec_ukernel__2x4_scalar_int)
71 ->Apply(BenchmarkKernelSize)->UseRealTime();
72 BENCHMARK_CAPTURE(transpose, 4x1_scalar_int, xnn_x32_transposec_ukernel__4x1_scalar_int)
73 ->Apply(BenchmarkKernelSize)->UseRealTime();
74 BENCHMARK_CAPTURE(transpose, 4x2_scalar_int, xnn_x32_transposec_ukernel__4x2_scalar_int)
75 ->Apply(BenchmarkKernelSize)->UseRealTime();
76 BENCHMARK_CAPTURE(transpose, 4x4_scalar_int, xnn_x32_transposec_ukernel__4x4_scalar_int)
77 ->Apply(BenchmarkKernelSize)->UseRealTime();
78 BENCHMARK_CAPTURE(transpose, 1x2_scalar_float, xnn_x32_transposec_ukernel__1x2_scalar_float)
79 ->Apply(BenchmarkKernelSize)->UseRealTime();
80 BENCHMARK_CAPTURE(transpose, 1x4_scalar_float, xnn_x32_transposec_ukernel__1x4_scalar_float)
81 ->Apply(BenchmarkKernelSize)->UseRealTime();
82 BENCHMARK_CAPTURE(transpose, 2x1_scalar_float, xnn_x32_transposec_ukernel__2x1_scalar_float)
83 ->Apply(BenchmarkKernelSize)->UseRealTime();
84 BENCHMARK_CAPTURE(transpose, 2x2_scalar_float, xnn_x32_transposec_ukernel__2x2_scalar_float)
85 ->Apply(BenchmarkKernelSize)->UseRealTime();
86 BENCHMARK_CAPTURE(transpose, 2x4_scalar_float, xnn_x32_transposec_ukernel__2x4_scalar_float)
87 ->Apply(BenchmarkKernelSize)->UseRealTime();
88 BENCHMARK_CAPTURE(transpose, 4x1_scalar_float, xnn_x32_transposec_ukernel__4x1_scalar_float)
89 ->Apply(BenchmarkKernelSize)->UseRealTime();
90 BENCHMARK_CAPTURE(transpose, 4x2_scalar_float, xnn_x32_transposec_ukernel__4x2_scalar_float)
91 ->Apply(BenchmarkKernelSize)->UseRealTime();
92 BENCHMARK_CAPTURE(transpose, 4x4_scalar_float, xnn_x32_transposec_ukernel__4x4_scalar_float)
93 ->Apply(BenchmarkKernelSize)->UseRealTime();
94
95 #if XNN_ARCH_ARM64
96 BENCHMARK_CAPTURE(transpose, 4x4_aarch64_neon_tbl, xnn_x32_transposec_ukernel__4x4_aarch64_neon_tbl)
97 ->Apply(BenchmarkKernelSize)->UseRealTime();
98 #endif // XNN_ARCH_ARM64
99
100 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
101 BENCHMARK_CAPTURE(transpose, 2x2_multi_dec_neon, xnn_x32_transposec_ukernel__2x2_multi_dec_zip_neon)
102 ->Apply(BenchmarkKernelSize)->UseRealTime();
103 BENCHMARK_CAPTURE(transpose, 2x2_multi_mov_neon, xnn_x32_transposec_ukernel__2x2_multi_mov_zip_neon)
104 ->Apply(BenchmarkKernelSize)->UseRealTime();
105 BENCHMARK_CAPTURE(transpose, 2x2_multi_multi_neon, xnn_x32_transposec_ukernel__2x2_multi_multi_zip_neon)
106 ->Apply(BenchmarkKernelSize)->UseRealTime();
107 BENCHMARK_CAPTURE(transpose, 2x2_multi_switch_neon, xnn_x32_transposec_ukernel__2x2_multi_switch_zip_neon)
108 ->Apply(BenchmarkKernelSize)->UseRealTime();
109 BENCHMARK_CAPTURE(transpose, 2x2_multi_dec_neon, xnn_x32_transposec_ukernel__2x2_reuse_dec_zip_neon)
110 ->Apply(BenchmarkKernelSize)->UseRealTime();
111 BENCHMARK_CAPTURE(transpose, 2x2_multi_mov_neon, xnn_x32_transposec_ukernel__2x2_reuse_mov_zip_neon)
112 ->Apply(BenchmarkKernelSize)->UseRealTime();
113 BENCHMARK_CAPTURE(transpose, 2x2_multi_multi_neon, xnn_x32_transposec_ukernel__2x2_reuse_multi_zip_neon)
114 ->Apply(BenchmarkKernelSize)->UseRealTime();
115 BENCHMARK_CAPTURE(transpose, 2x2_multi_switch_neon, xnn_x32_transposec_ukernel__2x2_reuse_switch_zip_neon)
116 ->Apply(BenchmarkKernelSize)->UseRealTime();
117 BENCHMARK_CAPTURE(transpose, 4x4_multi_dec_zip_neon, xnn_x32_transposec_ukernel__4x4_multi_dec_zip_neon)
118 ->Apply(BenchmarkKernelSize)->UseRealTime();
119 BENCHMARK_CAPTURE(transpose, 4x4_multi_mov_zip_neon, xnn_x32_transposec_ukernel__4x4_multi_mov_zip_neon)
120 ->Apply(BenchmarkKernelSize)->UseRealTime();
121 BENCHMARK_CAPTURE(transpose, 4x4_multi_multi_zip_neon, xnn_x32_transposec_ukernel__4x4_multi_multi_zip_neon)
122 ->Apply(BenchmarkKernelSize)->UseRealTime();
123 BENCHMARK_CAPTURE(transpose, 4x4_multi_switch_zip_neon, xnn_x32_transposec_ukernel__4x4_multi_switch_zip_neon)
124 ->Apply(BenchmarkKernelSize)->UseRealTime();
125 BENCHMARK_CAPTURE(transpose, 4x4_reuse_dec_zip_neon, xnn_x32_transposec_ukernel__4x4_reuse_dec_zip_neon)
126 ->Apply(BenchmarkKernelSize)->UseRealTime();
127 BENCHMARK_CAPTURE(transpose, 4x4_reuse_mov_zip_neon, xnn_x32_transposec_ukernel__4x4_reuse_mov_zip_neon)
128 ->Apply(BenchmarkKernelSize)->UseRealTime();
129 BENCHMARK_CAPTURE(transpose, 4x4_reuse_multi_zip_neon, xnn_x32_transposec_ukernel__4x4_reuse_multi_zip_neon)
130 ->Apply(BenchmarkKernelSize)->UseRealTime();
131 BENCHMARK_CAPTURE(transpose, 4x4_reuse_switch_zip_neon, xnn_x32_transposec_ukernel__4x4_reuse_switch_zip_neon)
132 ->Apply(BenchmarkKernelSize)->UseRealTime();
133 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
134
135 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
136 BENCHMARK_CAPTURE(transpose, 4x4_multi_mov_wasmsimd, xnn_x32_transposec_ukernel__4x4_multi_mov_wasmsimd)
137 ->Apply(BenchmarkKernelSize)->UseRealTime();
138 BENCHMARK_CAPTURE(transpose, 4x4_multi_multi_wasmsimd, xnn_x32_transposec_ukernel__4x4_multi_multi_wasmsimd)
139 ->Apply(BenchmarkKernelSize)->UseRealTime();
140 BENCHMARK_CAPTURE(transpose, 4x4_multi_switch_wasmsimd, xnn_x32_transposec_ukernel__4x4_multi_switch_wasmsimd)
141 ->Apply(BenchmarkKernelSize)->UseRealTime();
142 BENCHMARK_CAPTURE(transpose, 4x4_reuse_mov_wasmsimd, xnn_x32_transposec_ukernel__4x4_reuse_mov_wasmsimd)
143 ->Apply(BenchmarkKernelSize)->UseRealTime();
144 BENCHMARK_CAPTURE(transpose, 4x4_reuse_multi_wasmsimd, xnn_x32_transposec_ukernel__4x4_reuse_multi_wasmsimd)
145 ->Apply(BenchmarkKernelSize)->UseRealTime();
146 BENCHMARK_CAPTURE(transpose, 4x4_reuse_switch_wasmsimd, xnn_x32_transposec_ukernel__4x4_reuse_switch_wasmsimd)
147 ->Apply(BenchmarkKernelSize)->UseRealTime();
148 #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
149
150 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
151 BENCHMARK_CAPTURE(transpose, 4x4_sse, xnn_x32_transposec_ukernel__4x4_sse)
152 ->Apply(BenchmarkKernelSize)->UseRealTime();
153 BENCHMARK_CAPTURE(transpose, 4x4_multi_mov_sse2, xnn_x32_transposec_ukernel__4x4_multi_mov_sse2)
154 ->Apply(BenchmarkKernelSize)->UseRealTime();
155 BENCHMARK_CAPTURE(transpose, 4x4_multi_multi_sse2, xnn_x32_transposec_ukernel__4x4_multi_multi_sse2)
156 ->Apply(BenchmarkKernelSize)->UseRealTime();
157 BENCHMARK_CAPTURE(transpose, 4x4_multi_switch_sse2, xnn_x32_transposec_ukernel__4x4_multi_switch_sse2)
158 ->Apply(BenchmarkKernelSize)->UseRealTime();
159 BENCHMARK_CAPTURE(transpose, 4x4_reuse_mov_sse2, xnn_x32_transposec_ukernel__4x4_reuse_mov_sse2)
160 ->Apply(BenchmarkKernelSize)->UseRealTime();
161 BENCHMARK_CAPTURE(transpose, 4x4_reuse_multi_sse2, xnn_x32_transposec_ukernel__4x4_reuse_multi_sse2)
162 ->Apply(BenchmarkKernelSize)->UseRealTime();
163 BENCHMARK_CAPTURE(transpose, 4x4_reuse_switch_sse2, xnn_x32_transposec_ukernel__4x4_reuse_switch_sse2)
164 ->Apply(BenchmarkKernelSize)->UseRealTime();
165 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
166
167
168 #ifndef XNNPACK_BENCHMARK_NO_MAIN
169 BENCHMARK_MAIN();
170 #endif
171