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 <xnnpack/AlignedAllocator.h>
7 #include <xnnpack/common.h>
8 #include <xnnpack/params.h>
9 #include <xnnpack/transpose.h>
10
11 #include <algorithm>
12 #include <cmath>
13 #include <functional>
14 #include <numeric>
15 #include <vector>
16
17 #include "bench/utils.h"
18 #include <benchmark/benchmark.h>
19
transpose(benchmark::State & state,xnn_x8_transpose_ukernel_function transpose,benchmark::utils::IsaCheckFunction isa_check=nullptr)20 void transpose(
21 benchmark::State& state,
22 xnn_x8_transpose_ukernel_function transpose,
23 benchmark::utils::IsaCheckFunction isa_check = nullptr)
24 {
25 if (isa_check && !isa_check(state)) {
26 return;
27 }
28 const size_t height = state.range(0);
29 const size_t width = state.range(1);
30 const size_t tile_hbytes = height * sizeof(uint8_t);
31 const size_t tile_wbytes = width * sizeof(uint8_t);
32
33 std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> x(
34 height * width + XNN_EXTRA_BYTES / sizeof(uint8_t));
35 std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> y(
36 height * width + XNN_EXTRA_BYTES / sizeof(uint8_t));
37 std::iota(x.begin(), x.end(), 0);
38 std::fill(y.begin(), y.end(), 0);
39
40 for (auto _ : state) {
41 transpose(x.data(), y.data(), tile_wbytes, tile_hbytes, width,
42 height);
43 }
44
45 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
46 if (cpu_frequency != 0) {
47 state.counters["cpufreq"] = cpu_frequency;
48 }
49 }
50
BenchmarkKernelSize(benchmark::internal::Benchmark * b)51 static void BenchmarkKernelSize(benchmark::internal::Benchmark* b)
52 {
53 b->ArgNames({"height", "width"});
54 b->Args({32, 32});
55 b->Args({64, 64});
56 b->Args({117, 117});
57 b->Args({1024, 1024});
58 }
59
60 BENCHMARK_CAPTURE(transpose, 1x2_scalar_int, xnn_x8_transpose_ukernel__1x2_scalar_int)
61 ->Apply(BenchmarkKernelSize)->UseRealTime();
62 BENCHMARK_CAPTURE(transpose, 1x4_scalar_int, xnn_x8_transpose_ukernel__1x4_scalar_int)
63 ->Apply(BenchmarkKernelSize)->UseRealTime();
64 BENCHMARK_CAPTURE(transpose, 2x1_scalar_int, xnn_x8_transpose_ukernel__2x1_scalar_int)
65 ->Apply(BenchmarkKernelSize)->UseRealTime();
66 BENCHMARK_CAPTURE(transpose, 2x2_scalar_int, xnn_x8_transpose_ukernel__2x2_scalar_int)
67 ->Apply(BenchmarkKernelSize)->UseRealTime();
68 BENCHMARK_CAPTURE(transpose, 2x4_scalar_int, xnn_x8_transpose_ukernel__2x4_scalar_int)
69 ->Apply(BenchmarkKernelSize)->UseRealTime();
70 BENCHMARK_CAPTURE(transpose, 4x1_scalar_int, xnn_x8_transpose_ukernel__4x1_scalar_int)
71 ->Apply(BenchmarkKernelSize)->UseRealTime();
72 BENCHMARK_CAPTURE(transpose, 4x2_scalar_int, xnn_x8_transpose_ukernel__4x2_scalar_int)
73 ->Apply(BenchmarkKernelSize)->UseRealTime();
74 BENCHMARK_CAPTURE(transpose, 4x4_scalar_int, xnn_x8_transpose_ukernel__4x4_scalar_int)
75 ->Apply(BenchmarkKernelSize)->UseRealTime();
76
77 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
78 BENCHMARK_CAPTURE(transpose, 16x16_reuse_mov_zip_neon, xnn_x8_transpose_ukernel__16x16_reuse_mov_zip_neon)
79 ->Apply(BenchmarkKernelSize)->UseRealTime();
80 BENCHMARK_CAPTURE(transpose, 16x16_reuse_switch_neon, xnn_x8_transpose_ukernel__16x16_reuse_switch_zip_neon)
81 ->Apply(BenchmarkKernelSize)->UseRealTime();
82
83 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
84
85 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
86 BENCHMARK_CAPTURE(transpose, 16x16_reuse_mov_sse2, xnn_x8_transpose_ukernel__16x16_reuse_mov_sse2)
87 ->Apply(BenchmarkKernelSize)->UseRealTime();
88 BENCHMARK_CAPTURE(transpose, 16x16_reuse_switch_sse2, xnn_x8_transpose_ukernel__16x16_reuse_switch_sse2)
89 ->Apply(BenchmarkKernelSize)->UseRealTime();
90 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
91
92
93 #ifndef XNNPACK_BENCHMARK_NO_MAIN
94 BENCHMARK_MAIN();
95 #endif
96